StableHLO স্পেসিফিকেশন

StableHLO হল মেশিন লার্নিং (ML) মডেলে উচ্চ-স্তরের অপারেশনের (HLO) জন্য একটি অপারেশন সেট। StableHLO বিভিন্ন ML ফ্রেমওয়ার্ক এবং ML কম্পাইলারগুলির মধ্যে একটি পোর্টেবিলিটি স্তর হিসাবে কাজ করে: ML ফ্রেমওয়ার্কগুলি যেগুলি StableHLO প্রোগ্রামগুলি তৈরি করে সেগুলি ML কম্পাইলারগুলির সাথে সামঞ্জস্যপূর্ণ যেগুলি StableHLO প্রোগ্রামগুলি ব্যবহার করে৷

আমাদের লক্ষ্য হল বিভিন্ন ML ফ্রেমওয়ার্ক (যেমন TensorFlow, JAX এবং PyTorch) এবং ML কম্পাইলার (যেমন XLA এবং IREE) এর মধ্যে আরও আন্তঃকার্যযোগ্যতা তৈরি করে ML বিকাশকে সহজ করা এবং ত্বরান্বিত করা। সেই শেষের দিকে, এই নথিটি StableHLO প্রোগ্রামিং ভাষার জন্য একটি স্পেসিফিকেশন প্রদান করে।

এই স্পেসিফিকেশন তিনটি প্রধান বিভাগ রয়েছে. প্রথমত, প্রোগ্রাম বিভাগটি StableHLO প্রোগ্রামগুলির গঠন বর্ণনা করে যা StableHLO ফাংশন নিয়ে গঠিত যা নিজেরাই StableHLO ops নিয়ে গঠিত। সেই কাঠামোর মধ্যে, Ops বিভাগটি পৃথক অপ্সের শব্দার্থবিদ্যা নির্দিষ্ট করে। এক্সিকিউশন বিভাগটি একটি প্রোগ্রামের মধ্যে একসাথে এই সমস্ত অপারেশন চালানোর জন্য শব্দার্থবিদ্যা প্রদান করে। অবশেষে, স্বরলিপি বিভাগটি স্পেসিফিকেশন জুড়ে ব্যবহৃত স্বরলিপি নিয়ে আলোচনা করে।

StableHLO এর পূর্ববর্তী রিলিজ থেকে স্পেকটি দেখতে, আগ্রহের ট্যাগ করা রিলিজে রেপো খুলুন। উদাহরণস্বরূপ, StableHLO v0.19.0 Spec । StableHLO-এর প্রতিটি ছোট সংস্করণ বাম্পে ঘটে যাওয়া পরিবর্তনগুলি দেখতে, VhloDialect.td- এ সংস্করণ লগ পড়ুন।

প্রোগ্রাম

Program ::= {Func}

StableHLO প্রোগ্রামগুলি একটি নির্বিচারে StableHLO ফাংশন নিয়ে গঠিত। নিচে @main ফাংশন সহ একটি উদাহরণ প্রোগ্রাম রয়েছে যাতে 3টি ইনপুট ( %image , %weights এবং %bias ) এবং 1টি আউটপুট রয়েছে। ফাংশনের শরীরে 6টি অপ্স আছে।

func.func @main(
  %image: tensor<28x28xf32>,
  %weights: tensor<784x10xf32>,
  %bias: tensor<1x10xf32>
) -> tensor<1x10xf32> {
  %0 = "stablehlo.reshape"(%image) : (tensor<28x28xf32>) -> tensor<1x784xf32>
  %1 = "stablehlo.dot"(%0, %weights) : (tensor<1x784xf32>, tensor<784x10xf32>) -> tensor<1x10xf32>
  %2 = "stablehlo.add"(%1, %bias) : (tensor<1x10xf32>, tensor<1x10xf32>) -> tensor<1x10xf32>
  %3 = "stablehlo.constant"() {value = dense<0.0> : tensor<1x10xf32>} : () -> tensor<1x10xf32>
  %4 = "stablehlo.maximum"(%2, %3) : (tensor<1x10xf32>, tensor<1x10xf32>) -> tensor<1x10xf32>
  "func.return"(%4): (tensor<1x10xf32>) -> ()
}

ফাংশন

Func        ::= 'func' '.' 'func' FuncId FuncInputs FuncOutputs '{' FuncBody '}'
FuncInputs  ::= '(' [FuncInput {',' FuncInput}] `)`
FuncInput   ::= ValueId ':' ValueType
FuncOutputs ::= ['->' FuncOutput, {',' FuncOutput}]
FuncOutput  ::= ValueType
FuncBody    ::= {Op}

StableHLO ফাংশন (যাকে ফাংশন নামেও ডাকা হয়) এর একটি শনাক্তকারী, ইনপুট/আউটপুট এবং একটি বডি থাকে। ভবিষ্যতে, আমরা HLO ( #425 , #626 , #740 , #744 ) এর সাথে আরও ভাল সামঞ্জস্য অর্জনের জন্য ফাংশনের জন্য অতিরিক্ত মেটাডেটা চালু করার পরিকল্পনা করছি।

শনাক্তকারী

FuncId  ::= '@' letter {letter | digit}
ValueId ::= '%' digit {digit}
          | '%' letter {letter | digit}
letter  ::= 'a' | ... | 'z' | 'A' | ... | 'Z' | '_'
digit   ::= '0' | ... | '9'

StableHLO শনাক্তকারী অনেক প্রোগ্রামিং ভাষায় শনাক্তকারীর মতো, দুটি বিশেষত্ব সহ: 1) সমস্ত শনাক্তকারীর সিগিল থাকে যা বিভিন্ন ধরণের শনাক্তকারীকে আলাদা করে, 2) স্ট্যাবলএইচএলও প্রোগ্রামের প্রজন্মকে সহজ করার জন্য মান শনাক্তকারী সম্পূর্ণরূপে সংখ্যাসূচক হতে পারে।

প্রকারভেদ

Type         ::= ValueType | NonValueType
ValueType    ::= TensorType | QuantizedTensorType | TokenType | TupleType
NonValueType ::= TensorElementType | QuantizedTensorElementType | FunctionType | StringType

StableHLO প্রকারগুলিকে মান প্রকারে শ্রেণীবদ্ধ করা হয় (যাকে প্রথম-শ্রেণীর প্রকারও বলা হয়) যা StableHLO মান এবং অ-মূল্য প্রকারের প্রতিনিধিত্ব করে যা অন্যান্য প্রোগ্রাম উপাদানগুলিকে বর্ণনা করে। StableHLO প্রকারগুলি অনেক প্রোগ্রামিং ল্যাঙ্গুয়েজের প্রকারের মতই, যার প্রধান বিশেষত্ব হল StableHLO-এর ডোমেন-নির্দিষ্ট প্রকৃতি যার ফলে কিছু অস্বাভাবিক ফলাফল দেখা যায় (যেমন স্কেলার প্রকারগুলি মান প্রকার নয়)।

TensorType ::= 'tensor' '<' Shape TensorElementType '>'
Shape ::= {DimensionSize 'x'}
DimensionSize ::= digit {digit} | '?'

টেনসরের ধরনগুলি টেনসরকে প্রতিনিধিত্ব করে, যেমন বহুমাত্রিক অ্যারে। তাদের একটি আকৃতি এবং একটি উপাদানের ধরন রয়েছে, যেখানে একটি আকৃতি 0 থেকে R-1 পর্যন্ত সংখ্যাযুক্ত সংশ্লিষ্ট মাত্রাগুলির (যাকে অক্ষও বলা হয়) এর আরোহী ক্রমে নন-নেতিবাচক বা অজানা মাত্রার আকারগুলিকে প্রতিনিধিত্ব করে। R মাত্রার সংখ্যাকে র‌্যাঙ্ক বলা হয়। উদাহরণস্বরূপ, tensor<2x3xf32> আকৃতি 2x3 এবং উপাদান প্রকার f32 সহ একটি টেনসর প্রকার। এর দুটি মাত্রা রয়েছে (বা অন্য কথায়, দুটি অক্ষ) - 0 তম মাত্রা এবং 1 ম মাত্রা - যার আকার 2 এবং 3৷ এর র্যাঙ্ক হল 2৷

আকৃতি আংশিক বা সম্পূর্ণ অজানা (গতিশীল) হতে পারে, যেমন tensor<?x2xf64> আংশিকভাবে অজানা এবং tensor<?x?xf64> সম্পূর্ণ অজানা। একটি ? . আকারগুলিকে র‌্যাঙ্ক করা যাবে না।

ভবিষ্যতে, আমরা ডাইমেনশন সাইজ এবং এলিমেন্টের প্রকারের বাইরে টেনসরের প্রকারগুলিকে অন্বেষণ করার পরিকল্পনা করছি, উদাহরণস্বরূপ, লেআউট ( #629 ) এবং স্পারসিটি ( #1078 ) অন্তর্ভুক্ত করা।

QuantizedTensorType ::= 'tensor' '<' Shape QuantizedTensorElementType '>'
QuantizedTensorElementType ::= '!quant.uniform' '<'
                  QuantizationStorageType
                  ['<' QuantizationStorageMin ':' QuantizationStorageMax '>']
                  ':' QuantizationExpressedType
                  [':' QuantizationDimension]
                  ',' QuantizationParameters '>'
QuantizationStorageType ::= IntegerType
QuantizationStorageMin ::= IntegerLiteral
QuantizationStorageMax ::= IntegerLiteral
QuantizationExpressedType ::= FloatType
QuantizationDimension ::= IntegerLiteral
QuantizationParameters ::= QuantizationParameter
                         | '{' QuantizationParameter {',' QuantizationParameter} '}'
QuantizationParameter ::= QuantizationScale [':' QuantizationZeroPoint]
QuantizationScale ::= FloatLiteral
QuantizationZeroPoint ::= IntegerLiteral
নাম টাইপ সীমাবদ্ধতা
storage_type পূর্ণসংখ্যার ধরন (C1-C3), (C8)
storage_min পূর্ণসংখ্যা ধ্রুবক (C1), (C3), (C7)
storage_max পূর্ণসংখ্যা ধ্রুবক (C2), (C3), (C7)
expressed_type ভাসমান-বিন্দু টাইপ (C4)
quantization_dimension ঐচ্ছিক পূর্ণসংখ্যা ধ্রুবক (C10-C12)
scales ভাসমান-বিন্দু ধ্রুবকের বৈচিত্র্যময় সংখ্যা (C4-C6), (C9), (C10), (C13)
zero_points পূর্ণসংখ্যা ধ্রুবকের বৈচিত্র্যময় সংখ্যা (C7-C9)

কোয়ান্টাইজড এলিমেন্টের ধরন storage_min থেকে storage_max (অন্তর্ভুক্ত) পরিসরে একটি স্টোরেজ প্রকারের পূর্ণসংখ্যার মানগুলিকে উপস্থাপন করে যা একটি প্রকাশিত ধরণের ফ্লোটিং-পয়েন্ট মানগুলির সাথে মিলে যায়। একটি প্রদত্ত পূর্ণসংখ্যার মান i জন্য, সংশ্লিষ্ট ফ্লোটিং-পয়েন্ট মান f f = (i - zero_point) * scale হিসাবে গণনা করা যেতে পারে, যেখানে scale এবং zero_point কে কোয়ান্টাইজেশন প্যারামিটার বলা হয়। storage_min এবং storage_max ব্যাকরণে ঐচ্ছিক, কিন্তু যথাক্রমে min_value(storage_type) এবং max_value(storage_type) এর ডিফল্ট মান রয়েছে। কোয়ান্টাইজড উপাদানের প্রকারের নিম্নলিখিত সীমাবদ্ধতা রয়েছে:

  • (C1) type(storage_min) = storage_type
  • (C2) type(storage_max) = storage_type
  • (C3) min_value(storage_type) <= storage_min < storage_max <= max_value(storage_type)
  • (C4) type(scales...) = expressed_type
  • (C5) 0 < scales
  • (C6) is_finite(scales...)
  • (C7) storage_min <= zero_points <= storage_max
  • (C8) type(zero_points...) = storage_type
  • (C9) size(scales) = size(zero_points)
  • (C10) যদি_ is_empty(quantization_dimension) , তাহলে size(scales) = 1
  • (C11) 0 <= quantization_dimension

এই মুহুর্তে, QuantizationScale একটি ফ্লোটিং-পয়েন্ট ধ্রুবক, কিন্তু পূর্ণসংখ্যা-ভিত্তিক স্কেলগুলিতে প্রবল আগ্রহ রয়েছে, যা গুণক এবং স্থানান্তর দ্বারা উপস্থাপিত হয়। আমরা নিকট ভবিষ্যতে এটি অন্বেষণ করার পরিকল্পনা করছি ( #1404 )।

QuantizationZeroPoint শব্দার্থবিদ্যার উপর একটি চলমান আলোচনা রয়েছে, যার মধ্যে রয়েছে ধরন, মান এবং একটি কোয়ান্টাইজড টেনসরের ধরনে শুধুমাত্র একটি বা সম্ভাব্য একাধিক শূন্য পয়েন্ট থাকতে পারে কিনা। এই আলোচনার ফলাফলের উপর ভিত্তি করে, শূন্য পয়েন্টের কাছাকাছি স্পেসিফিকেশন ভবিষ্যতে পরিবর্তিত হতে পারে ( #1405 )।

আর একটি চলমান আলোচনায় QuantizationStorageMin এবং QuantizationStorageMax শব্দার্থবিদ্যা জড়িত যে এই মানগুলিতে এবং কোয়ান্টাইজড টেনসরগুলির মানগুলির উপর কোন সীমাবদ্ধতা আরোপ করা উচিত কিনা তা নির্ধারণ করতে ( #1406 )।

অবশেষে, আমরা অজানা স্কেল এবং শূন্য বিন্দুর প্রতিনিধিত্বকারী অন্বেষণ করার পরিকল্পনা করছি, একইভাবে আমরা অজানা মাত্রার আকার ( #1407 ) প্রতিনিধিত্ব করে অন্বেষণ করার পরিকল্পনা করছি।

কোয়ান্টাইজড টেনসর প্রকারগুলি কোয়ান্টাইজড উপাদান সহ টেনসরগুলিকে উপস্থাপন করে। এই টেনসরগুলি রেগুলার টেনসরের মতো হুবহু একই, তাদের উপাদানগুলির রেগুলার এলিমেন্ট টাইপের পরিবর্তে কোয়ান্টাইজড এলিমেন্ট টাইপ আছে।

কোয়ান্টাইজেশন টেনসরে, কোয়ান্টাইজেশন প্রতি-টেনসর হতে পারে, যার অর্থ, সমগ্র টেনসরের জন্য একটি scale এবং zero_point থাকতে পারে বা প্রতি-অক্ষ হতে পারে, যার অর্থ, একাধিক scales এবং zero_points থাকতে পারে, একটি নির্দিষ্ট মাত্রা quantization_dimension প্রতি স্লাইসে এক জোড়া। আরও আনুষ্ঠানিকভাবে, প্রতি-অক্ষ পরিমাপ সহ একটি টেনসর t এ, quantization_dimension dim(t, quantization_dimension) স্লাইস রয়েছে : t[:, ..., 0, ..., :], t[:, ..., 1, ..., :] , ইত্যাদি। i তম স্লাইসের সমস্ত উপাদান তাদের পরিমাপ হিসাবে scales[i] এবং zero_points[i] ব্যবহার করে পরামিতি কোয়ান্টাইজড টেনসর প্রকারের নিম্নলিখিত সীমাবদ্ধতা রয়েছে:

  • প্রতি-টেনসর পরিমাপের জন্য:
    • কোন অতিরিক্ত সীমাবদ্ধতা.
  • প্রতি-অক্ষ পরিমাপের জন্য:
    • (C12) quantization_dimension < rank(self) .
    • (C13) dim(self, quantization_dimension) = size(scales)
TokenType ::= 'token'

টোকেন প্রকারগুলি টোকেনগুলিকে প্রতিনিধিত্ব করে, যেমন অস্বচ্ছ মানগুলি কিছু ক্রিয়াকলাপের দ্বারা উত্পাদিত এবং সেবন করে৷ এক্সিকিউশন বিভাগে বর্ণিত ক্রিয়াকলাপের উপর কার্যকর আদেশ আরোপ করার জন্য টোকেন ব্যবহার করা হয়।

TupleType ::= 'tuple' '<' TupleElementTypes '>'
TupleElementTypes ::= [ValueType {',' ValueType}]

Tuple প্রকারগুলি tuples, অর্থাৎ ভিন্নধর্মী তালিকাগুলিকে প্রতিনিধিত্ব করে। Tuples হল একটি উত্তরাধিকার বৈশিষ্ট্য যা শুধুমাত্র HLO এর সাথে সামঞ্জস্যের জন্য বিদ্যমান। HLO-তে, টিপলগুলি বৈচিত্র্যময় ইনপুট এবং আউটপুটগুলিকে উপস্থাপন করতে ব্যবহৃত হয়। StableHLO-তে, বৈচিত্র্যময় ইনপুট এবং আউটপুটগুলি স্থানীয়ভাবে সমর্থিত, এবং StableHLO-তে টিপলের একমাত্র ব্যবহার হল HLO ABI-কে ব্যাপকভাবে উপস্থাপন করা যেখানে যেমন T , tuple<T> এবং tuple<tuple<T>> একটি নির্দিষ্ট বাস্তবায়নের উপর নির্ভর করে বস্তুগতভাবে ভিন্ন হতে পারে। . ভবিষ্যতে, আমরা HLO ABI-তে পরিবর্তন করার পরিকল্পনা করছি যা আমাদের StableHLO ( #598 ) থেকে টিপল ধরনের অপসারণ করতে পারে।

TensorElementType ::= BooleanType | IntegerType | FloatType | ComplexType
BooleanType ::= 'i1'
IntegerType ::= SignedIntegerType | UnsignedIntegerType
SignedIntegerType ::= 'si2' | 'si4' | 'si8' | 'si16' | 'si32' | 'si64'
UnsignedIntegerType ::= 'ui2' | 'ui4' | 'ui8' | 'ui16' | 'ui32' | 'ui64'
FloatType ::= 'f4E2M1FN' | 'f6E2M3FN' | 'f6E3M2FN' | 'f8E3M4' | 'f8E4M3'
            | 'f8E4M3FN' | 'f8E4M3FNUZ' | 'f8E4M3B11FNUZ' | 'f8E5M2'
            | 'f8E5M2FNUZ' | 'f8E8M0FNU' | 'bf16' | 'f16' | 'f32' | 'f64'
TensorFloat32 ::= 'tf32'
ComplexType ::= 'complex' '<' ComplexElementType '>'
ComplexElementType ::= 'f32' | 'f64'

উপাদান প্রকারগুলি টেনসর ধরণের উপাদানগুলির প্রতিনিধিত্ব করে। অনেক প্রোগ্রামিং ভাষার বিপরীতে, এই প্রকারগুলি StableHLO-তে প্রথম শ্রেণীর নয়। এর মানে হল যে StableHLO প্রোগ্রামগুলি এই ধরণের মানগুলিকে সরাসরি উপস্থাপন করতে পারে না (ফলে, টাইপ tensor<T> > এর 0-মাত্রিক টেনসর মানগুলির সাথে T টাইপের স্কেলার মানগুলিকে উপস্থাপন করা বাহাদুরিপূর্ণ)।

  • বুলিয়ান টাইপ বুলিয়ান মান true এবং false উপস্থাপন করে।
  • পূর্ণসংখ্যার প্রকারগুলি হয় স্বাক্ষরিত ( si ) বা স্বাক্ষরবিহীন ( ui ) হতে পারে এবং সমর্থিত বিট প্রস্থগুলির মধ্যে একটি থাকতে পারে ( 2 , 4 , 8 , 16 , 32 বা 64 )৷ স্বাক্ষরিত siN প্রকারগুলি -2^(N-1) থেকে 2^(N-1)-1 সহ পূর্ণসংখ্যার মানগুলিকে উপস্থাপন করে, এবং স্বাক্ষরবিহীন uiN প্রকারগুলি 0 থেকে 2^N-1 সহ পূর্ণসংখ্যার মানগুলিকে উপস্থাপন করে।
  • ফ্লোটিং-পয়েন্ট প্রকারগুলি নিম্নলিখিতগুলির মধ্যে একটি হতে পারে:
  • জটিল প্রকারগুলি জটিল মানগুলিকে উপস্থাপন করে যেগুলির একটি বাস্তব অংশ এবং একই উপাদান প্রকারের একটি কাল্পনিক অংশ রয়েছে৷ সমর্থিত জটিল প্রকারগুলি হল complex<f32> (উভয় অংশই f32 প্রকারের) এবং complex<f64> (উভয় অংশই f64 প্রকার)।
FunctionType ::= '(' InputTypes ')' '->' '(' OutputTypes ')'
InputTypes ::= [ValueType {',' ValueType}]
OutputTypes ::= [ValueType {',' ValueType}]

ফাংশন প্রকারগুলি নামযুক্ত এবং বেনামী উভয় ফাংশন উপস্থাপন করে। তাদের ইনপুট প্রকার রয়েছে ( -> এর বাম দিকের প্রকারের তালিকা) এবং আউটপুট প্রকারগুলি ( -> এর ডানদিকের প্রকারের তালিকা)। অনেক প্রোগ্রামিং ভাষায়, ফাংশনের ধরন প্রথম শ্রেণীর, কিন্তু StableHLO তে নয়।

StringType ::= 'string'

স্ট্রিং টাইপ বাইটের ক্রম উপস্থাপন করে। অনেক প্রোগ্রামিং ভাষার বিপরীতে, স্ট্রিং টাইপ StableHLO-তে প্রথম শ্রেণীর নয় এবং শুধুমাত্র প্রোগ্রাম উপাদানগুলির জন্য স্ট্যাটিক মেটাডেটা নির্দিষ্ট করতে ব্যবহৃত হয়।

অপারেশন

StableHLO অপারেশনগুলি (যাকে ops ও বলা হয়) মেশিন লার্নিং মডেলগুলিতে উচ্চ-স্তরের অপারেশনগুলির একটি বন্ধ সেট উপস্থাপন করে। উপরে যেমন আলোচনা করা হয়েছে, StableHLO সিনট্যাক্স MLIR দ্বারা প্রচণ্ডভাবে অনুপ্রাণিত, যা অগত্যা সবচেয়ে ergonomic বিকল্প নয়, কিন্তু ML ফ্রেমওয়ার্ক এবং ML কম্পাইলারগুলির মধ্যে আরও আন্তঃকার্যযোগ্যতা তৈরি করার জন্য StableHLO-এর লক্ষ্যের জন্য যুক্তিযুক্তভাবে উপযুক্ত।

Op            ::= [OpOutputs] OpName OpInputs ':' OpSignature
OpName        ::= '"' 'stablehlo' '.' OpMnemonic '"'
OpMnemonic    ::= 'abs' | 'add' | ...

StableHLO অপারেশনগুলির (যাকে অপসও বলা হয়) একটি নাম, ইনপুট/আউটপুট এবং একটি স্বাক্ষর থাকে। নাম stablehlo. উপসর্গ এবং একটি স্মারক যা স্বতন্ত্রভাবে সমর্থিত অপ্সগুলির একটিকে চিহ্নিত করে। সমস্ত সমর্থিত অপারেশনগুলির একটি বিস্তৃত তালিকার জন্য নীচে দেখুন৷

OpInputs        ::= OpInputValues OpInputFuncs OpInputAttrs
OpInputValues   ::= '(' [OpInputValue {',' OpInputValue}] ')'
OpInputValue    ::= ValueId
OpInputFuncs    ::= ['(' OpInputFunc {',' OpInputFunc} ')']
OpInputAttrs    ::= ['{' OpInputAttr {',' OpInputAttr} '}']
OpOutputs       ::= [OpOutput {',' OpOutput} '=']
OpOutput        ::= ValueId

অপস ইনপুট গ্রহণ করে এবং আউটপুট উত্পাদন করে। ইনপুটগুলি ইনপুট মানগুলিতে শ্রেণীবদ্ধ করা হয় (নির্বাহের সময় গণনা করা হয়), ইনপুট ফাংশন (স্ট্যাবলিকভাবে প্রদান করা হয়, কারণ স্টেবলএইচএলও ফাংশনগুলি প্রথম শ্রেণীর মান নয়) এবং ইনপুট বৈশিষ্ট্যগুলি (এছাড়াও স্থিরভাবে প্রদান করা হয়)। একটি অপ দ্বারা খাওয়া এবং উত্পাদিত ইনপুট এবং আউটপুট তার স্মৃতির উপর নির্ভর করে। উদাহরণস্বরূপ, add অপটি 2টি ইনপুট মান গ্রহণ করে এবং 1টি আউটপুট মান তৈরি করে। তুলনায়, select_and_scatter op 3টি ইনপুট মান, 2টি ইনপুট ফাংশন এবং 3টি ইনপুট বৈশিষ্ট্য ব্যবহার করে৷

OpInputFunc ::= '{' Unused FuncInputs ':' FuncBody '}'
Unused      ::= '^' digit {digit}
              | '^' letter {letter | digit}

ইনপুট ফাংশনগুলি (যাকে বেনামী ফাংশনও বলা হয়) নামযুক্ত ফাংশনগুলির সাথে খুব মিল: 1) তাদের কোনও শনাক্তকারী নেই (তাই নাম "বেনামী"), 2) তারা আউটপুট প্রকারগুলি ঘোষণা করে না (আউটপুট প্রকারগুলি হল ফাংশনের মধ্যে return অপ থেকে অনুমান করা হয়েছে)।

ইনপুট ফাংশনের জন্য সিনট্যাক্সে বর্তমানে একটি অব্যবহৃত অংশ রয়েছে (উপরে Unused উত্পাদন দেখুন) যা MLIR-এর সাথে সামঞ্জস্যের জন্য রয়েছে। এমএলআইআর-এ, "অঞ্চল" এর আরও সাধারণ ধারণা রয়েছে যেখানে জাম্প অপ্সের মাধ্যমে একসাথে সংযুক্ত একাধিক "ব্লক" থাকতে পারে। এই ব্লকগুলিতে Unused প্রোডাকশনের সাথে সঙ্গতিপূর্ণ আইডি রয়েছে, যাতে তাদের একে অপরের থেকে আলাদা করা যায়। StableHLO এর জাম্প অপ্স নেই, তাই MLIR সিনট্যাক্সের সংশ্লিষ্ট অংশটি অব্যবহৃত (কিন্তু এখনও আছে)।

OpInputAttr      ::= OpInputAttrName '=' OpInputAttrValue
OpInputAttrName  ::= letter {letter | digit}
OpInputAttrValue ::= Constant

ইনপুট বৈশিষ্ট্যগুলির একটি নাম এবং একটি মান রয়েছে যা সমর্থিত ধ্রুবকগুলির মধ্যে একটি৷ তারা প্রোগ্রাম উপাদানগুলির জন্য স্ট্যাটিক মেটাডেটা নির্দিষ্ট করার প্রাথমিক উপায়। উদাহরণ স্বরূপ, concatenate op বৈশিষ্ট্যের dimension ব্যবহার করে তার ইনপুট মানগুলি যে মাত্রার সাথে সংযুক্ত করা হয়েছে তা নির্দিষ্ট করতে। একইভাবে, slice অপটি ইনপুট মান স্লাইস করতে ব্যবহৃত সীমানা নির্দিষ্ট করতে start_indices এবং limit_indices মতো একাধিক বৈশিষ্ট্য ব্যবহার করে।

এই মুহুর্তে, বন্যের StableHLO প্রোগ্রামে মাঝে মাঝে এমন বৈশিষ্ট্য থাকে যা এই নথিতে বর্ণনা করা হয়নি। ভবিষ্যতে, আমরা হয় StableHLO অপসেটে এই বৈশিষ্ট্যগুলি শোষণ করার বা StableHLO প্রোগ্রামগুলিতে উপস্থিত হওয়া থেকে নিষিদ্ধ করার পরিকল্পনা করছি৷ ইতিমধ্যে, এখানে এই বৈশিষ্ট্যগুলির তালিকা রয়েছে:

  • layout ( #629 )।
  • mhlo.frontend_attributes ( #628 )।
  • mhlo.sharding ( #619 )।
  • output_operand_aliases ( #740 )।
  • অবস্থান মেটাডেটা ( #594 )।
OpSignature ::= '(' [ValueType {',' ValueType}] ')' '->' '(' [ValueType {',' ValueType}] ')'

অপ স্বাক্ষরে সমস্ত ইনপুট মান ( -> এর বাম দিকের প্রকারের তালিকা) এবং সমস্ত আউটপুট মানগুলির প্রকার ( -> এর ডানদিকের প্রকারের তালিকা) থাকে। কঠোরভাবে বলতে গেলে, ইনপুট প্রকারগুলি অপ্রয়োজনীয়, এবং আউটপুট প্রকারগুলি প্রায় সবসময়ই অপ্রয়োজনীয় (কারণ বেশিরভাগ StableHLO অপ্সের জন্য, আউটপুট প্রকারগুলি ইনপুট থেকে অনুমান করা যেতে পারে)। তবুও, MLIR এর সাথে সামঞ্জস্যের জন্য op স্বাক্ষর ইচ্ছাকৃতভাবে StableHLO সিনট্যাক্সের অংশ।

নীচে একটি উদাহরণ op যার স্মৃতিচিহ্ন হল select_and_scatter । এটি 3টি ইনপুট মান ( %operand , %source এবং %init_value ), 2টি ইনপুট ফাংশন এবং 3টি ইনপুট অ্যাট্রিবিউট ( window_dimensions , window_strides এবং padding ) ব্যবহার করে। নোট করুন কিভাবে op-এর স্বাক্ষর শুধুমাত্র এর ইনপুট মানগুলির প্রকারগুলি অন্তর্ভুক্ত করে (কিন্তু ইনপুট ফাংশন এবং বৈশিষ্ট্যগুলির প্রকারগুলি নয় যা ইনলাইনে দেওয়া হয়)৷

%result = "stablehlo.select_and_scatter"(%operand, %source, %init_value) ({
  ^bb0(%arg0: tensor<i32>, %arg1: tensor<i32>):
    %0 = "stablehlo.compare"(%arg0, %arg1) {
      comparison_direction = #stablehlo<comparison_direction GE>
    } : (tensor<i32>, tensor<i32>) -> tensor<i1>
    "stablehlo.return"(%0) : (tensor<i1>) -> ()
}, {
  ^bb0(%arg0: tensor<i32>, %arg1: tensor<i32>):
    %0 = "stablehlo.add"(%arg0, %arg1) : (tensor<i32>, tensor<i32>) -> tensor<i32>
    "stablehlo.return"(%0) : (tensor<i32>) -> ()
}) {
  window_dimensions = dense<[3, 1]> : tensor<2xi64>,
  window_strides = dense<[2, 1]> : tensor<2xi64>,
  padding = dense<[[0, 1], [0, 0]]> : tensor<2x2xi64>
} : (tensor<4x2xi32>, tensor<2x2xi32>, tensor<i32>) -> tensor<4x2xi32>

ধ্রুবক

Constant ::= BooleanConstant
           | IntegerConstant
           | FloatConstant
           | ComplexConstant
           | TensorConstant
           | QuantizedTensorConstant
           | StringConstant
           | EnumConstant

StableHLO ধ্রুবকগুলির একটি আক্ষরিক এবং একটি প্রকার রয়েছে যা একসাথে একটি StableHLO মান উপস্থাপন করে। সাধারণত, টাইপটি ধ্রুবক সিনট্যাক্সের অংশ, এটি দ্ব্যর্থহীন হলে (যেমন একটি বুলিয়ান ধ্রুবকের দ্ব্যর্থহীনভাবে i1 টাইপ থাকে, যেখানে একটি পূর্ণসংখ্যা ধ্রুবকের একাধিক সম্ভাব্য প্রকার থাকতে পারে)।

BooleanConstant ::= BooleanLiteral
BooleanLiteral  ::= 'true' | 'false'

বুলিয়ান ধ্রুবক বুলিয়ান মান true এবং false উপস্থাপন করে। বুলিয়ান ধ্রুবকের টাইপ i1 আছে।

IntegerConstant   ::= IntegerLiteral ':' IntegerType
IntegerLiteral    ::= ['-' | '+'] DecimalDigits
                    | ['-' | '+'] '0x' HexadecimalDigits
DecimalDigits     ::= decimalDigit {decimalDigit}
HexadecimalDigits ::= hexadecimalDigit {hexadecimalDigit}
decimalDigit      ::= '0' | ... | '9'
hexadecimalDigit  ::= decimalDigit | 'a' | ... | 'f' | 'A' | ... | 'F'

পূর্ণসংখ্যার ধ্রুবকগুলি দশমিক বা হেক্সাডেসিমেল স্বরলিপি ব্যবহার করে এমন স্ট্রিংগুলির মাধ্যমে পূর্ণসংখ্যার মানগুলিকে উপস্থাপন করে। অন্যান্য বেস, যেমন বাইনারি বা অক্টাল, সমর্থিত নয়। পূর্ণসংখ্যার ধ্রুবকগুলির নিম্নলিখিত সীমাবদ্ধতা রয়েছে:

  • (C1) is_wellformed(integer_literal, integer_type)
FloatConstant  ::= FloatLiteral ':' FloatType
FloatLiteral   ::= SignPart IntegerPart FractionalPart ScientificPart
                 | '0x' [HexadecimalDigits]
SignPart       ::= ['-' | '+']
IntegerPart    ::= DecimalDigits
FractionalPart ::= ['.' [DecimalDigits]]
ScientificPart ::= [('e' | 'E') ['-' | '+'] DecimalDigits]

ফ্লোটিং-পয়েন্ট ধ্রুবকগুলি স্ট্রিংগুলির মাধ্যমে ফ্লোটিং-পয়েন্ট মানগুলি উপস্থাপন করে যা দশমিক বা বৈজ্ঞানিক নোটেশন ব্যবহার করে। অতিরিক্তভাবে, হেক্সাডেসিমেল নোটেশনটি সংশ্লিষ্ট প্রকারের ফ্লোটিং-পয়েন্ট বিন্যাসে অন্তর্নিহিত বিটগুলিকে সরাসরি নির্দিষ্ট করতে ব্যবহার করা যেতে পারে। ফ্লোটিং-পয়েন্ট ধ্রুবকগুলির নিম্নলিখিত সীমাবদ্ধতা রয়েছে:

  • (C1) নন-হেক্সাডেসিমেল নোটেশন ব্যবহার করা হলে, is_wellformed(float_literal, float_type)
  • (C2) যদি হেক্সাডেসিমাল নোটেশন ব্যবহার করা হয়, size(hexadecimal_digits) = num_bits(float_type) / 4
ComplexConstant ::= ComplexLiteral ':' ComplexType
ComplexLiteral  ::= '(' RealPart ',' ImaginaryPart ')'
RealPart        ::= FloatLiteral
ImaginaryPart   ::= FloatLiteral

জটিল ধ্রুবকগুলি একটি বাস্তব অংশ (প্রথমে আসে) এবং একটি কাল্পনিক অংশ (দ্বিতীয় আসে) তালিকা ব্যবহার করে জটিল মানগুলি উপস্থাপন করে। উদাহরণস্বরূপ, (1.0, 0.0) : complex<f32> 1.0 + 0.0i , এবং (0.0, 1.0) : complex<f32> 0.0 + 1.0i প্রতিনিধিত্ব করে। যে ক্রমে এই অংশগুলি মেমরিতে সংরক্ষণ করা হয় তা বাস্তবায়ন-সংজ্ঞায়িত। জটিল ধ্রুবকগুলির নিম্নলিখিত সীমাবদ্ধতা রয়েছে:

  • (C1) is_wellformed(real_part, complex_element_type(complex_type))
  • (C2) is_wellformed(imaginary_part, complex_element_type(complex_type))
TensorConstant ::= TensorLiteral ':' TensorType
TensorLiteral  ::= 'dense' '<' (DenseLiteral | ElementLiteral) '>'
DenseLiteral   ::= DenseDimension | DenseElements
DenseDimension ::= '[' [DenseLiteral {',' DenseLiteral}] ']'
DenseElements  ::= [ElementLiteral {',' ElementLiteral}]
ElementLiteral ::= BooleanLiteral | IntegerLiteral | FloatLiteral | ComplexLiteral

টেনসর ধ্রুবকগুলি NumPy স্বরলিপির মাধ্যমে নির্দিষ্ট নেস্টেড তালিকা ব্যবহার করে টেনসর মানগুলি উপস্থাপন করে। উদাহরণস্বরূপ, dense<[[1, 2, 3], [4, 5, 6]]> : tensor<2x3xi32> সূচকগুলি থেকে উপাদানগুলিতে নিম্নলিখিত ম্যাপিং সহ একটি টেনসর মান উপস্থাপন করে: {0, 0} => 1 , {0, 1} => 2 , {0, 2} => 3 , {1, 0} => 4 , {1, 1} => 5 , {1, 2} => 6 । যে ক্রমে এই উপাদানগুলি মেমরিতে সংরক্ষণ করা হয় তা বাস্তবায়ন-সংজ্ঞায়িত। টেনসর ধ্রুবকগুলির নিম্নলিখিত সীমাবদ্ধতা রয়েছে:

  • (C1) has_syntax(tensor_literal, element_type(tensor_type)) , যেখানে:
    • has_syntax(element_literal: Syntax, element_type: Type) = is_wellformed(element_literal, type)
    • has_syntax(tensor_literal: List, element_type: Type) = has_syntax(tensor_literal..., element_type)
  • (C2) has_shape(tensor_literal, shape(tensor_type)) , যেখানে:
    • has_shape(element_literal: Syntax, []) = true
    • has_shape(tensor_literal: List, shape: List) = size(tensor_literal) = shape[0] and has_shape(tensor_literal..., shape[1:])
    • অন্যথায়, false
QuantizedTensorConstant ::= QuantizedTensorLiteral ':' QuantizedTensorType
QuantizedTensorLiteral  ::= 'dense' '<' (DenseLiteral | ElementLiteral) '>'

কোয়ান্টাইজড টেনসর ধ্রুবকগুলি তাদের স্টোরেজ প্রকারের ধ্রুবক হিসাবে নির্দিষ্ট উপাদানগুলির সাথে টেনসর ধ্রুবকের মতো একই স্বরলিপি ব্যবহার করে কোয়ান্টাইজড টেনসর মানগুলি উপস্থাপন করে। কোয়ান্টাইজড টেনসর ধ্রুবকগুলির নিম্নলিখিত সীমাবদ্ধতা রয়েছে:

  • (C1) has_syntax(quantized_tensor_literal, storage_type(quantized_tensor_type))
  • (C2) has_shape(quantized_tensor_literal, shape(quantized_tensor_type))
StringConstant  ::= StringLiteral
StringLiteral   ::= '"' {stringCharacter | escapeSequence} '"'
stringCharacter ::= all ASCII characters except '\00', '\01', ... '\1f' and '"'
escapeSequence  ::= '\' ('"' | '\' | 'n' | 't' | (hexadecimalDigit hexadecimalDigit))

স্ট্রিং লিটারেলে ASCII অক্ষর এবং এস্কেপ সিকোয়েন্স ব্যবহার করে নির্দিষ্ট করা বাইট থাকে। তারা এনকোডিং-অজ্ঞেয়বাদী, তাই এই বাইটের ব্যাখ্যা বাস্তবায়ন-সংজ্ঞায়িত। স্ট্রিং লিটারেলে টাইপ string থাকে।

অপ্স

abs

শব্দার্থবিদ্যা

operand টেনসরে উপাদান-ভিত্তিক abs অপারেশন করে এবং result টেনসর তৈরি করে। উপাদান প্রকারের উপর নির্ভর করে, নিম্নলিখিতগুলি করে:

  • স্বাক্ষরিত পূর্ণসংখ্যার জন্য: পূর্ণসংখ্যা মডুলাস।
  • ফ্লোটের জন্য: IEEE-754 থেকে abs
  • জটিল সংখ্যার জন্য: জটিল মডুলাস।
  • কোয়ান্টাইজড প্রকারের জন্য: dequantize_op_quantize(abs, operand, type(result))

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand স্বাক্ষরিত পূর্ণসংখ্যার টেনসর, ভাসমান-বিন্দু, বা জটিল প্রকার বা প্রতি-টেনসর কোয়ান্টাইজড টেনসর (C1-C2)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result স্বাক্ষরিত পূর্ণসংখ্যা বা ফ্লোটিং-পয়েন্ট টাইপ বা প্রতি-টেনসর কোয়ান্টাইজড টেনসরের টেনসর (C1-C2)

সীমাবদ্ধতা

  • (C1) shape(result) = shape(operand)
  • (C2) baseline_element_type(result) এইভাবে সংজ্ঞায়িত করা হয়েছে:
    • complex_element_type(element_type(operand)) is_complex(operand) হলে।
    • baseline_element_type(operand) অন্যথায়।

উদাহরণ

// %operand: [-2, 0, 2]
%result = "stablehlo.abs"(%operand) : (tensor<3xi32>) -> tensor<3xi32>
// %result: [2, 0, 2]

আরো উদাহরণ

যোগ করুন

শব্দার্থবিদ্যা

দুটি টেনসর lhs এবং rhs এর উপাদান-ভিত্তিক সংযোজন সম্পাদন করে এবং একটি result টেনসর তৈরি করে। উপাদান প্রকারের উপর নির্ভর করে, নিম্নলিখিতগুলি করে:

  • বুলিয়ানদের জন্য: লজিক্যাল OR।
  • পূর্ণসংখ্যার জন্য: পূর্ণসংখ্যা সংযোজন।
  • ফ্লোটের জন্য: IEEE-754 থেকে addition
  • জটিল সংখ্যার জন্য: জটিল যোগ।
  • quantized প্রকারের জন্য: dequantize_op_quantize(add, lhs, rhs, type(result))

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs টেনসর বা কোয়ান্টাইজড টেনসর (C1-C6)
(I2) rhs টেনসর বা কোয়ান্টাইজড টেনসর (C1-C5), (C7)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result টেনসর বা কোয়ান্টাইজড টেনসর (C1-C7)

সীমাবদ্ধতা

  • যদি অপারেশনটি নন-কোয়ান্টাইজড টেনসর ব্যবহার করে:
    • (C1) type(lhs) = type(rhs) = type(result)
  • যদি অপারেশন কোয়ান্টাইজড টেনসর ব্যবহার করে:
    • (C2) is_quantized(lhs) and is_quantized(rhs) and is_quantized(result)
    • (C3) storage_type(lhs) = storage_type(rhs) = storage_type(result)
    • (C4) expressed_type(lhs) = expressed_type(rhs) = expressed_type(result)
    • (C5) (is_per_axis_quantized(lhs) or is_per_axis_quantized(rhs)) = is_per_axis_quantized(result)
    • (C6) যদি_ is_per_axis_quantized(lhs) , তাহলে quantization_dimension(lhs) = quantization_dimension(result)
    • (C7) যদি_ is_per_axis_quantized(rhs) , তাহলে quantization_dimension(rhs) = quantization_dimension(result)

উদাহরণ

// %lhs: [[1, 2], [3, 4]]
// %rhs: [[5, 6], [7, 8]]
%result = "stablehlo.add"(%lhs, %rhs) : (tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32>
// %result: [[6, 8], [10, 12]]

আরো উদাহরণ

পরে_সব

শব্দার্থবিদ্যা

নিশ্চিত করে যে inputs উৎপাদনকারী ক্রিয়াকলাপগুলি result উপর নির্ভর করে এমন কোনও অপারেশনের আগে সম্পাদিত হয়। এই ক্রিয়াকলাপটি কার্যকর করা কিছুই করে না, এটি শুধুমাত্র result থেকে inputs ডেটা নির্ভরতা স্থাপনের জন্য বিদ্যমান।

ইনপুট

লেবেল নাম টাইপ
(I1) inputs token বৈচিত্র্যময় সংখ্যা

আউটপুট

নাম টাইপ
result token

উদাহরণ

// %input0: !stablehlo.token
// %input1: !stablehlo.token
%result = "stablehlo.after_all"(%input0, %input1) : (!stablehlo.token, !stablehlo.token) -> !stablehlo.token

আরো উদাহরণ

সব_জড়ো করা

শব্দার্থবিদ্যা

StableHLO প্রসেস গ্রিডে প্রতিটি প্রসেস গ্রুপের মধ্যে, all_gather_dim বরাবর প্রতিটি প্রসেস থেকে operands টেনসরের মান একত্রিত করে এবং results টেনসর তৈরি করে।

অপারেশনটি StableHLO প্রসেস গ্রিডকে process_groups এ বিভক্ত করে যা নিম্নরূপ সংজ্ঞায়িত করা হয়েছে:

  • cross_replica(replica_groups) যদি channel_id <= 0 and use_global_device_ids = false
  • cross_replica_and_partition(replica_groups) যদি channel_id > 0 and use_global_device_ids = false
  • flattened_ids(replica_groups) যদি channel_id > 0 and use_global_device_ids = true

পরে, প্রতিটি process_group মধ্যে:

  • operands...@receiver = [operand@sender for sender in process_group] process_group এ সমস্ত receiver জন্য।
  • results...@process = concatenate(operands...@process, all_gather_dim) process_group এ সমস্ত process জন্য।

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operands টেনসরের বিভিন্ন সংখ্যা বা প্রতি-টেনসর কোয়ান্টাইজড টেনসর (C1), (C6)
(I2) all_gather_dim si64 টাইপের ধ্রুবক (C1), (C6)
(I3) replica_groups si64 টাইপের 2-মাত্রিক টেনসর ধ্রুবক (C2-C4)
(I4) channel_id si64 টাইপের ধ্রুবক (C5)
(I5) use_global_device_ids টাইপ i1 এর ধ্রুবক (C5)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
results টেনসরের বিভিন্ন সংখ্যা বা প্রতি-টেনসর কোয়ান্টাইজড টেনসর (C6)

সীমাবদ্ধতা

  • (C1) 0 <= all_gather_dim < rank(operands...)
  • (C2) is_unique(replica_groups)
  • (C3) size(replica_groups) সংজ্ঞায়িত করা হয়েছে এভাবে:
    • num_replicas যদি cross_replica ব্যবহার করা হয়।
    • num_replicas যদি cross_replica_and_partition ব্যবহার করা হয়।
    • num_processes যদি flattened_ids ব্যবহার করা হয়।
  • (C4) 0 <= replica_groups < size(replica_groups)
  • (C5) যদি use_global_device_ids = true হয়, তাহলে channel_id > 0
  • (C6) type(results...) = type(operands...) ছাড়া:
    • dim(results..., all_gather_dim) = dim(operands..., all_gather_dim) * dim(process_groups, 1)

উদাহরণ

// num_replicas: 2
// num_partitions: 1
// %operand0@(0, 0): [[1, 2], [3, 4]]
// %operand0@(1, 0): [[5, 6], [7, 8]]
// %operand1@(0, 0): [[11, 12], [13, 14]]
// %operand1@(1, 0): [[15, 16], [17, 18]]
%result:2 = "stablehlo.all_gather"(%operand0, %operand1) {
  all_gather_dim = 1 : i64,
  replica_groups = dense<[[0, 1]]> : tensor<1x2xi64>,
  // channel_id = 0
  channel_handle = #stablehlo.channel_handle<handle = 0, type = 0>
  // use_global_device_ids = false
} : (tensor<2x2xi64>, tensor<2x2xi64>) -> (tensor<2x4xi64>, tensor<2x4xi64>)
// %result0@(0, 0): [[1, 2, 5, 6], [3, 4, 7, 8]]
// %result0@(1, 0): [[1, 2, 5, 6], [3, 4, 7, 8]]
// %result1@(0, 0): [[11, 12, 15, 16], [13, 14, 17, 18]]
// %result1@(1, 0): [[11, 12, 15, 16], [13, 14, 17, 18]]

আরো উদাহরণ

all_reduce

শব্দার্থবিদ্যা

StableHLO প্রসেস গ্রিডের প্রতিটি প্রক্রিয়া গোষ্ঠীর মধ্যে, প্রতিটি প্রক্রিয়া থেকে operands টেনসরের মানগুলিতে একটি হ্রাস ফাংশন computation প্রয়োগ করে এবং results টেনসর তৈরি করে।

অপারেশনটি StableHLO প্রসেস গ্রিডকে process_groups এ বিভক্ত করে যা নিম্নরূপ সংজ্ঞায়িত করা হয়েছে:

  • cross_replica(replica_groups) যদি channel_id <= 0 and use_global_device_ids = false
  • cross_replica_and_partition(replica_groups) যদি channel_id > 0 and use_global_device_ids = false
  • flattened_ids(replica_groups) যদি channel_id > 0 and use_global_device_ids = true

পরে, প্রতিটি process_group মধ্যে:

  • results...@process[result_index] = exec(schedule) কিছু বাইনারি ট্রি schedule জন্য exec(সূচি) যেখানে:
    • exec(node) = computation(exec(node.left), exec(node.right))
    • exec(leaf) = leaf.value .
  • schedule হল একটি বাস্তবায়ন-সংজ্ঞায়িত বাইনারি ট্রি যার ইন-অর্ডার ট্রাভার্সাল হল to_destination_type(operands...@process_group...[result_index], type(func_inputs(computation)[0]))

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operands টেনসরের বিভিন্ন সংখ্যা বা প্রতি-টেনসর কোয়ান্টাইজড টেনসর (C5), (C6)
(I2) replica_groups si64 টাইপের 1-মাত্রিক টেনসর ধ্রুবকের বৈচিত্র্যিক সংখ্যা (C1-C3)
(I3) channel_id si64 টাইপের ধ্রুবক (C4)
(I4) use_global_device_ids টাইপ i1 এর ধ্রুবক (C4)
(I5) computation ফাংশন (C5)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
results টেনসরের বিভিন্ন সংখ্যা বা প্রতি-টেনসর কোয়ান্টাইজড টেনসর (C6-C7)

সীমাবদ্ধতা

  • (C1) is_unique(replica_groups)
  • (C2) size(replica_groups) এইভাবে সংজ্ঞায়িত করা হয়েছে:
    • num_replicas যদি cross_replica ব্যবহার করা হয়।
    • num_replicas যদি cross_replica_and_partition ব্যবহার করা হয়।
    • num_processes যদি flattened_ids ব্যবহার করা হয়।
  • (C3) 0 <= replica_groups < size(replica_groups)
  • (C4) যদি use_global_device_ids = true হয়, তাহলে channel_id > 0
  • (C5) computation টাইপ আছে (tensor<E>, tensor<E>) -> (tensor<E>) যেখানে is_promotable(element_type(operand), E)
  • (C6) shape(results...) = shape(operands...)
  • (C7) element_type(results...) = E

উদাহরণ

// num_replicas: 2
// num_partitions: 1
// %operand0@(0, 0): [1, 2, 3, 4]
// %operand0@(1, 0): [5, 6, 7, 8]
// %operand1@(0, 0): [9, 10, 11, 12]
// %operand1@(1, 0): [13, 14, 15, 16]
%result:2 = "stablehlo.all_reduce"(%operand0, %operand0) ({
  ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>):
    %0 = "stablehlo.add"(%arg0, %arg1) : (tensor<i64>, tensor<i64>) -> tensor<i64>
    "stablehlo.return"(%0) : (tensor<i64>) -> ()
}) {
  replica_groups = dense<[[0, 1]]> : tensor<1x2xi64>,
  // channel_id = 0
  channel_handle = #stablehlo.channel_handle<handle = 0, type = 0>
  // use_global_device_ids = false
} : (tensor<4xi64>, tensor<4xi64>) -> (tensor<4xi64>, tensor<4xi64>)
// %result0@(0, 0): [6, 8, 10, 12]
// %result0@(1, 0): [6, 8, 10, 12]
// %result1@(0, 0): [22, 24, 26, 28]
// %result1@(1, 0): [22, 24, 26, 28]

আরো উদাহরণ

সব_থেকে_সমস্ত

শব্দার্থবিদ্যা

সব_থেকে_সমস্ত

StableHLO প্রসেস গ্রিডে প্রতিটি প্রসেস গ্রুপের মধ্যে, split_dimension বরাবর operands টেনসরের মানকে অংশে বিভক্ত করে, প্রসেসের মধ্যে বিভক্ত অংশগুলিকে ছড়িয়ে দেয়, concat_dimension বরাবর বিক্ষিপ্ত অংশগুলিকে সংযুক্ত করে এবং results টেনসর তৈরি করে। অপারেশনটি StableHLO প্রসেস গ্রিডকে process_groups এ বিভক্ত করে যা নিম্নরূপ সংজ্ঞায়িত করা হয়েছে:

  • cross_replica(replica_groups) যদি channel_id <= 0 হয়।
  • cross_partition(replica_groups) যদি channel_id > 0 হয়।

পরে, প্রতিটি process_group মধ্যে:

  • split_parts...@sender = split(operands...@sender, split_count, split_dimension) process_group সকল sender জন্য।
  • scattered_parts...@receiver = [split_parts...@sender[receiver_index] for sender in process_group] যেখানে receiver_index = process_group.index(receiver)
  • results...@process = concatenate(scattered_parts...@process, concat_dimension)

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operands টেনসরের বিভিন্ন সংখ্যা বা প্রতি-টেনসর কোয়ান্টাইজড টেনসর (C1-C3), (C9)
(I2) split_dimension si64 টাইপের ধ্রুবক (C1), (C2), (C9)
(I3) concat_dimension si64 টাইপের ধ্রুবক (C3), (C9)
(I4) split_count si64 টাইপের ধ্রুবক (C2), (C4), (C8), (C9)
(I5) replica_groups si64 টাইপের 2-মাত্রিক টেনসর ধ্রুবক (C5-C8)
(I6) channel_id si64 টাইপের ধ্রুবক

আউটপুট

নাম টাইপ সীমাবদ্ধতা
results টেনসরের বিভিন্ন সংখ্যা বা প্রতি-টেনসর কোয়ান্টাইজড টেনসর (C9)

সীমাবদ্ধতা

  • (C1) 0 <= split_dimension < rank(operands...)
  • (C2) dim(operands..., split_dimension) % split_count = 0
  • (C3) 0 <= concat_dimension < rank(operands...)
  • (C4) 0 < split_count
  • (C5) is_unique(replica_groups)
  • (C6) size(replica_groups) এইভাবে সংজ্ঞায়িত করা হয়েছে:
    • num_replicas যদি cross_replica ব্যবহার করা হয়।
    • num_partitions যদি cross_partition ব্যবহার করা হয়।
  • (C7) 0 <= replica_groups < size(replica_groups)
  • (C8) dim(replica_groups, 1) = split_count
  • (C9) type(results...) = type(operands...) ব্যতীত, যদি split_dimension != concat_dimension :
    • dim(results..., split_dimension) = dim(operands..., split_dimension) / split_count
    • dim(results..., concat_dimension) = dim(operands..., concat_dimension) * split_count

উদাহরণ

// num_replicas: 2
// num_partitions: 1
// %operand1@(0, 0): [[1, 2, 3, 4],
//                    [5, 6, 7, 8]]
// %operand1@(1, 0): [[9, 10, 11, 12],
//                    [13, 14, 15, 16]]
// %operand2@(0, 0): [[17, 18, 19, 20],
//                    [21, 22, 23, 24]]
// %operand2@(1, 0): [[25, 26, 27, 28],
//                    [29, 30, 31, 32]]
%result:2 = "stablehlo.all_to_all"(%operand1, %operand2) {
  split_dimension = 1 : i64,
  concat_dimension = 0 : i64,
  split_count = 2 : i64,
  replica_groups = dense<[[0, 1]]> : tensor<1x2xi64>
  // channel_id = 0
} : (tensor<2x4xi64>, tensor<2x4xi64>) -> (tensor<4x2xi64>, tensor<4x2xi64>)
// %result#0@(0, 0): [[1, 2], [5, 6], [9, 10], [13, 14]]
// %result#0@(1, 0): [[3, 4], [7, 8], [11, 12], [15, 16]]
// %result#1@(0, 0): [[17, 18], [21, 22], [25, 26], [29, 30]]
// %result#1@(1, 0): [[19, 20], [23, 24], [27, 28], [31, 32]]

আরো উদাহরণ

এবং

শব্দার্থবিদ্যা

দুটি টেনসর lhs এবং rhs এর উপাদান-ভিত্তিক AND সম্পাদন করে এবং একটি result টেনসর তৈরি করে। উপাদান প্রকারের উপর নির্ভর করে, নিম্নলিখিতগুলি করে:

  • বুলিয়ানদের জন্য: যৌক্তিক এবং।
  • পূর্ণসংখ্যার জন্য: bitwise AND.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs বুলিয়ান বা পূর্ণসংখ্যা ধরনের টেনসর (C1)
(I2) rhs বুলিয়ান বা পূর্ণসংখ্যা ধরনের টেনসর (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result বুলিয়ান বা পূর্ণসংখ্যা ধরনের টেনসর (C1)

সীমাবদ্ধতা

  • (C1) type(lhs) = type(rhs) = type(result)

উদাহরণ

// %lhs: [[1, 2], [3, 4]]
// %rhs: [[5, 6], [7, 8]]
%result = "stablehlo.and"(%lhs, %rhs) : (tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32>
// %result: [[1, 2], [3, 0]]

আরো উদাহরণ

atan2

শব্দার্থবিদ্যা

lhs এবং rhs টেনসরে উপাদান-ভিত্তিক atan2 অপারেশন করে এবং result টেনসর তৈরি করে। উপাদান প্রকারের উপর নির্ভর করে, নিম্নলিখিতগুলি করে:

  • ফ্লোটের জন্য: IEEE-754 থেকে atan2
  • জটিল সংখ্যার জন্য: জটিল atan2।
  • কোয়ান্টাইজড ধরনের জন্য: dequantize_op_quantize(atan2, lhs, rhs, type(result))

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs ফ্লোটিং-পয়েন্ট বা জটিল প্রকারের টেনসর বা প্রতি-টেনসর কোয়ান্টাইজড টেনসর (C1)
(I2) rhs ফ্লোটিং-পয়েন্ট বা জটিল প্রকারের টেনসর বা প্রতি-টেনসর কোয়ান্টাইজড টেনসর (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result ফ্লোটিং-পয়েন্ট বা জটিল প্রকারের টেনসর বা প্রতি-টেনসর কোয়ান্টাইজড টেনসর (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(lhs) = baseline_type(rhs) = baseline_type(result)

উদাহরণ

// %lhs: [0.0, 1.0, -1.0]
// %rhs: [0.0, 0.0, 0.0]
%result = "stablehlo.atan2"(%lhs, %rhs) : (tensor<3xf64>, tensor<3xf64>) -> tensor<3xf64>
// %result: [0.0, 1.57079637, -1.57079637] // [0.0, pi/2, -pi/2]

আরো উদাহরণ

ব্যাচ_নর্ম_গ্র্যাড

শব্দার্থবিদ্যা

grad_output থেকে batch_norm_training ব্যাকপ্রোপাগেটিং এর বিভিন্ন ইনপুটের গ্রেডিয়েন্ট গণনা করে এবং grad_operand , grad_scale এবং grad_offset টেনসর তৈরি করে। আরও আনুষ্ঠানিকভাবে, এই ক্রিয়াকলাপটি নিম্নরূপ পাইথন সিনট্যাক্স ব্যবহার করে বিদ্যমান StableHLO ক্রিয়াকলাপগুলির একটি পচন হিসাবে প্রকাশ করা যেতে পারে:

def compute_sum(operand, feature_index):
  (sum,) = reduce(
      inputs=[operand],
      init_values=[constant(0, element_type(operand))],
      dimensions=[i for i in range(rank(operand)) if i != feature_index],
      body=lambda x, y: add(x, y))
  return sum

def compute_mean(operand, feature_index):
  sum = compute_sum(operand, feature_index)
  divisor = constant(size(operand) / dim(operand, feature_index),
                     element_type(operand))
  divisor_bcast = broadcast_in_dim(divisor, [], type(sum))
  return divide(sum, divisor_bcast)

def batch_norm_grad(operand, scale, mean, variance, grad_output, epsilon, feature_index):
  # Broadcast inputs to type(operand)
  scale_bcast = broadcast_in_dim(scale, [feature_index], type(operand))
  mean_bcast = broadcast_in_dim(mean, [feature_index], type(operand))
  variance_bcast = broadcast_in_dim(variance, [feature_index], type(operand))
  epsilon_bcast = broadcast_in_dim(constant(epsilon, element_type(operand)), [],
                                   type(operand))

  # Perform normalization using the provided `mean` and `variance`
  # Intermediate values will be useful for computing gradients
  centered_operand = subtract(operand, mean_bcast)
  stddev = sqrt(add(variance_bcast, epsilon_bcast))
  normalized_operand = divide(centered_operand, stddev)

  # Use the implementation from batchnorm_expander.cc in XLA
  # Temporary variables have exactly the same names as in the C++ code
  elements_per_feature = broadcast_in_dim(
      constant(divide(size(operand), dim(operand, feature_index)),
               element_type(grad_output)),
      [], type(operand))
  i1 = multiply(grad_output, elements_per_feature)
  i2 = broadcast_in_dim(
      compute_sum(grad_output, feature_index), [feature_index], type(operand))
  i3 = broadcast_in_dim(
      compute_sum(multiply(grad_output, centered_operand), feature_index),
      [feature_index], type(operand))
  i4 = multiply(i3, centered_operand)
  i5 = divide(i4, add(variance_bcast, epsilon_bcast))
  i6 = subtract(subtract(i1, i2), i5)

  grad_operand =
      multiply(divide(divide(scale_bcast, stddev), elements_per_feature), i6)
  grad_scale =
      compute_sum(multiply(grad_output, normalized_operand), feature_index)
  grad_offset = compute_sum(grad_output, feature_index)

  return grad_operand, grad_scale, grad_offset

quantized প্রকারের জন্য, dequantize_batch_norm_grad_or_training_quantize(lambda operand, scale, mean, variance, grad_output: batch_norm_grad(operand, scale, mean, variance, grad_output, epsilon, feature_index), operand, scale, mean, variance, grad_output, type(grad_operand), type(grad_scale), type(feature_index))

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand ফ্লোটিং-পয়েন্ট টাইপের টেনসর বা প্রতি-টেনসর কোয়ান্টাইজড টেনসর (C1-C3), (C5)
(I2) scale ফ্লোটিং-পয়েন্ট বা প্রতি-টেনসর কোয়ান্টাইজড টাইপের 1-মাত্রিক টেনসর (C2), (C4), (C5)
(I3) mean ফ্লোটিং-পয়েন্ট বা প্রতি-টেনসর কোয়ান্টাইজড টাইপের 1-মাত্রিক টেনসর (C2), (C4)
(I4) variance ফ্লোটিং-পয়েন্ট বা প্রতি-টেনসর কোয়ান্টাইজড টাইপের 1-মাত্রিক টেনসর (C2), (C4)
(I5) grad_output ফ্লোটিং-পয়েন্ট টাইপের টেনসর বা প্রতি-টেনসর কোয়ান্টাইজড টেনসর (C2), (C3)
(I6) epsilon f32 টাইপের ধ্রুবক
(I7) feature_index si64 টাইপের ধ্রুবক (C1), (C5)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
grad_operand ফ্লোটিং-পয়েন্ট টাইপের টেনসর বা প্রতি-টেনসর কোয়ান্টাইজড টেনসর (C2), (C3)
grad_scale ফ্লোটিং-পয়েন্ট বা প্রতি-টেনসর কোয়ান্টাইজড টাইপের 1-মাত্রিক টেনসর (C2), (C4)
grad_offset ফ্লোটিং-পয়েন্ট বা প্রতি-টেনসর কোয়ান্টাইজড টাইপের 1-মাত্রিক টেনসর (C2), (C4)

সীমাবদ্ধতা

  • (C1) 0 <= feature_index < rank(operand)
  • (C2) operand , scale , mean , variance , grad_output , grad_operand , grad_scale এবং grad_offset একই baseline_element_type আছে।
  • (C3) operand , grad_output এবং grad_operand আকৃতি একই।
  • (C4) scale , mean , variance , grad_scale এবং grad_offset আকৃতি একই।
  • (C5) size(scale) = dim(operand, feature_index)

উদাহরণ

// %operand: [
//            [[1.0, 2.0], [3.0, 4.0]],
//            [[3.0, 4.0], [1.0, 2.0]]
//           ]
// %scale: [1.0, 1.0]
// %mean: [2.0, 3.0]
// %variance: [1.0, 1.0]
// %grad_output: [
//                [[0.1, 0.1], [0.1, 0.1]],
//                [[0.1, 0.1], [0.1, 0.1]]
//               ]
%grad_operand, %grad_scale, %grad_offset =
"stablehlo.batch_norm_grad"(%operand, %scale, %mean, %variance, %grad_output) {
  epsilon = 0.0 : f32,
  feature_index = 2 : i64
} : (tensor<2x2x2xf64>, tensor<2xf64>, tensor<2xf64>, tensor<2xf64>,
     tensor<2x2x2xf64>) -> (tensor<2x2x2xf64>, tensor<2xf64>, tensor<2xf64>)
// %grad_operand: [
//                 [[0.0, 0.0], [0.0, 0.0]],
//                 [[0.0, 0.0], [0.0, 0.0]]
//                ]
// %grad_scale:  [0.0, 0.0]
// %grad_offset: [0.4, 0.4]

ব্যাচ_নর্ম_ইনফারেন্স

শব্দার্থবিদ্যা

feature_index মাত্রা ব্যতীত সমস্ত মাত্রা জুড়ে operand টেনসরকে স্বাভাবিক করে এবং একটি result টেনসর তৈরি করে। আরও আনুষ্ঠানিকভাবে, এই ক্রিয়াকলাপটি নিম্নরূপ পাইথন সিনট্যাক্স ব্যবহার করে বিদ্যমান StableHLO ক্রিয়াকলাপগুলির একটি পচন হিসাবে প্রকাশ করা যেতে পারে:

def batch_norm_inference(operand, scale, offset, mean, variance, epsilon, feature_index):
  # Broadcast inputs to shape(operand)
  scale_bcast = broadcast_in_dim(scale, [feature_index], type(operand))
  offset_bcast = broadcast_in_dim(offset, [feature_index], type(operand))
  mean_bcast = broadcast_in_dim(mean, [feature_index], type(operand))
  variance_bcast = broadcast_in_dim(variance, [feature_index], type(operand))
  epsilon_bcast = broadcast_in_dim(constant(epsilon, element_type(operand)), [],
                                   type(operand))

  # Perform normalization using the provided `mean` and `variance` instead of
  # computing them like `batch_norm_training` does.
  centered_operand = subtract(operand, mean_bcast)
  stddev = sqrt(add(variance_bcast, epsilon_bcast))
  normalized_operand = divide(centered_operand, stddev)
  return add(multiply(scale_bcast, normalized_operand), offset_bcast)

quantized প্রকারের জন্য, dequantize_op_quantize(lambda operand, scale, offset, mean, variance: batch_norm_inference(operand, scale, offset, mean, variance, epsilon, feature_index), operand, scale, offset, mean, variance, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand ফ্লোটিং-পয়েন্ট টাইপের টেনসর বা প্রতি-টেনসর কোয়ান্টাইজড টেনসর (C1-C7)
(I2) scale ফ্লোটিং-পয়েন্ট বা প্রতি-টেনসর কোয়ান্টাইজড টাইপের 1-মাত্রিক টেনসর (C2), (C3)
(I3) offset ফ্লোটিং-পয়েন্ট বা প্রতি-টেনসর কোয়ান্টাইজড টাইপের 1-মাত্রিক টেনসর (C2), (C4)
(I4) mean ফ্লোটিং-পয়েন্ট বা প্রতি-টেনসর কোয়ান্টাইজড টাইপের 1-মাত্রিক টেনসর (C5)
(I5) variance ফ্লোটিং-পয়েন্ট বা প্রতি-টেনসর কোয়ান্টাইজড টাইপের 1-মাত্রিক টেনসর (C2), (C6)
(I6) epsilon f32 টাইপের ধ্রুবক
(I7) feature_index si64 টাইপের ধ্রুবক (C1), (C3-C6)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result ফ্লোটিং-পয়েন্ট টাইপের টেনসর বা প্রতি-টেনসর কোয়ান্টাইজড টেনসর (C2), (C7)

সীমাবদ্ধতা

  • (C1) 0 <= feature_index < rank(operand)
  • (C2) operand , scale , offset , mean , variance এবং result একই baseline_element_type আছে।
  • (C3) size(scale) = dim(operand, feature_index)
  • (C4) size(offset) = dim(operand, feature_index)
  • (C5) size(mean) = dim(operand, feature_index)
  • (C6) size(variance) = dim(operand, feature_index)
  • (C7) baseline_type(operand) = baseline_type(result)

উদাহরণ

// %operand: [
//            [[1.0, 2.0], [3.0, 4.0]],
//            [[3.0, 4.0], [1.0, 2.0]]
//           ]
// %scale: [1.0, 1.0]
// %offset: [1.0, 1.0]
// %mean: [2.0, 3.0]
// %variance: [1.0, 1.0]
%result = "stablehlo.batch_norm_inference"(%operand, %scale, %offset, %mean, %variance) {
  epsilon = 0.0 : f32,
  feature_index = 2 : i64
} : (tensor<2x2x2xf64>, tensor<2xf64>, tensor<2xf64>, tensor<2xf64>, tensor<2xf64>) -> tensor<2x2x2xf64>
// %result: [
//           [[0.0, 0.0], [2.0, 2.0]],
//           [[2.0, 2.0], [0.0, 0.0]]
//          ]

ব্যাচ_নর্ম_প্রশিক্ষণ

শব্দার্থবিদ্যা

feature_index ডাইমেনশন ব্যতীত সমস্ত মাত্রা জুড়ে কম্পিউট মানে এবং ভ্যারিয়েন্স এবং output , batch_mean এবং batch_var টেনসর উত্পাদনকারী operand টেনসরকে স্বাভাবিক করে। আরও আনুষ্ঠানিকভাবে, এই ক্রিয়াকলাপটি নিম্নরূপ পাইথন সিনট্যাক্স ব্যবহার করে বিদ্যমান StableHLO ক্রিয়াকলাপগুলির একটি পচন হিসাবে প্রকাশ করা যেতে পারে:

def compute_mean(operand, feature_index):
  (sum,) = reduce(
      inputs=[operand],
      init_values=[constant(0, element_type(operand))],
      dimensions=[i for i in range(rank(operand)) if i != feature_index],
      body=lambda x, y: add(x, y))
  divisor = constant(size(operand) / dim(operand, feature_index),
                     element_type(operand))
  divisor_bcast = broadcast_in_dim(divisor, [], type(sum))
  return divide(sum, divisor_bcast)

def compute_variance(operand, feature_index):
  mean = compute_mean(operand, feature_index)
  mean_bcast = broadcast_in_dim(mean, [feature_index], type(operand))
  centered_operand = subtract(operand, mean_bcast)
  return compute_mean(mul(centered_operand, centered_operand), feature_index)

def batch_norm_training(operand, scale, offset, epsilon, feature_index):
  mean = compute_mean(operand, feature_index)
  variance = compute_variance(operand, feature_index)
  return batch_norm_inference(operand, scale, offset, mean, variance, epsilon,
                              feature_index),
         mean, variance

quantized ধরনের জন্য, dequantize_batch_norm_grad_or_training_quantize(lambda operand, scale, offset: batch_norm_training(operand, scale, offset, epsilon, feature_index), operand, scale, offset, type(output), type(batch_mean), type(batch_var))

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand ফ্লোটিং-পয়েন্ট টাইপের টেনসর বা প্রতি-টেনসর কোয়ান্টাইজড টেনসর (C1)
(I2) scale ফ্লোটিং-পয়েন্টের 1-মাত্রিক টেনসর বা প্রতি-টেনসর কোয়ান্টাইজড (C2), (C3)
(I3) offset ফ্লোটিং-পয়েন্টের 1-মাত্রিক টেনসর বা প্রতি-টেনসর কোয়ান্টাইজড (C2), (C4)
(I4) epsilon f32 টাইপের ধ্রুবক (C1), (C3-C6)
(I5) feature_index si64 টাইপের ধ্রুবক (C1), (C3-C6)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
output ফ্লোটিং-পয়েন্ট টাইপের টেনসর বা প্রতি-টেনসর কোয়ান্টাইজড টেনসর (C7)
batch_mean ফ্লোটিং-পয়েন্টের 1-মাত্রিক টেনসর বা প্রতি-টেনসর কোয়ান্টাইজড (C2), (C5)
batch_var ফ্লোটিং-পয়েন্টের 1-মাত্রিক টেনসর বা প্রতি-টেনসর কোয়ান্টাইজড (C2), (C6)

সীমাবদ্ধতা

  • (C1) 0 <= feature_index < rank(operand)
  • (C2) operand , scale , offset , batch_mean , batch_var এবং output একই baseline_element_type আছে।
  • (C3) size(scale) = dim(operand, feature_index)
  • (C4) size(offset) = dim(operand, feature_index)
  • (C5) size(batch_mean) = dim(operand, feature_index)
  • (C6) size(batch_var) = dim(operand, feature_index)
  • (C7) baseline_type(output) = baseline_type(operand)

উদাহরণ

// %operand: [
//            [[1.0, 2.0], [3.0, 4.0]],
//            [[3.0, 4.0], [1.0, 2.0]]
//           ]
// %scale: [1.0, 1.0]
// %offset: [1.0, 1.0]
%output, %batch_mean, %batch_var = "stablehlo.batch_norm_training"(%operand, %scale, %offset) {
  epsilon = 0.0 : f32,
  feature_index = 2 : i64
} : (tensor<2x2x2xf64>, tensor<2xf64>, tensor<2xf64>) ->
    (tensor<2x2x2xf64>, tensor<2xf64>, tensor<2xf64>)
// %output: [
//           [[0.0, 0.0], [2.0, 2.0]],
//           [[2.0, 2.0], [0.0, 0.0]]
//          ]
// %batch_mean: [2.0, 3.0]
// %batch_var: [1.0, 1.0]

bitcast_convert

শব্দার্থবিদ্যা

operand টেনসরে একটি বিটকাস্ট অপারেশন সম্পাদন করে এবং একটি result টেনসর তৈরি করে যেখানে result টেনসরের ধরন ব্যবহার করে সম্পূর্ণ operand টেনসরের বিটগুলি পুনরায় ব্যাখ্যা করা হয়।

আরো আনুষ্ঠানিকভাবে, দেওয়া হল E = element_type(operand) , E' = element_type(result) , এবং R = rank(operand) :

  • যদি num_bits(E') < num_bits(E) , bits(result[i0, ..., iR-1, :]) = bits(operand[i0, ..., iR-1])
  • যদি num_bits(E') > num_bits(E) , bits(result[i0, ..., iR-2]) = bits(operand[i0, ..., iR-2, :])
  • যদি num_bits(E') = num_bits(E) , bits(result[i0, ..., iR-1]) = bits(operand[i0, ..., iR-1])

bits একটি প্রদত্ত মানের ইন-মেমরি উপস্থাপনা প্রদান করে, এবং এর আচরণটি বাস্তবায়ন-সংজ্ঞায়িত কারণ টেনসরগুলির সঠিক উপস্থাপনাটি বাস্তবায়ন-সংজ্ঞায়িত, এবং উপাদান প্রকারের সঠিক উপস্থাপনাটি বাস্তবায়ন-সংজ্ঞায়িত।

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand টেনসর বা কোয়ান্টাইজড টেনসর (C1-C2)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result টেনসর বা কোয়ান্টাইজড টেনসর (C1-C2)

সীমাবদ্ধতা

  • (C1) দেওয়া E = is_quantized(operand) ? storage_type(operand) : element_type(operand) , E' = is_quantized(result) ? storage_type(result) : element_type(result) , এবং R = rank(operand) :
    • যদি num_bits(E') = num_bits(E) , shape(result) = shape(operand)
    • যদি num_bits(E') < num_bits(E) :
    • rank(result) = R + 1
    • dim(result, i) = dim(operand, i) সব 0 <= i < R জন্য।
    • dim(result, R) * num_bits(E') = num_bits(E)
    • যদি num_bits(E') > num_bits(E) :
    • rank(result) = R - 1
    • dim(result, i) = dim(operand, i) সব 0 <= i < R জন্য।
    • dim(operand, R - 1) * num_bits(E) = num_bits(E')
  • (C2) যদি is_complex(operand) or is_complex(result) হয়, তাহলে is_complex(operand) and is_complex(result)

উদাহরণ

// %operand: 0x0123456789ABCDEF
%result = "stablehlo.bitcast_convert"(%operand) : (tensor<f64>) -> tensor<4xf16>
// %result: [0xCDEF, 0x89AB, 0x4567, 0x0123] // little-endian representation

আরো উদাহরণ

broadcast_in_dim

শব্দার্থবিদ্যা

operand টেনসরে ডেটা নকল করে একটি ইনপুট টেনসরের মাত্রা এবং/অথবা র‌্যাঙ্ক প্রসারিত করে এবং result টেনসর তৈরি করে। আরও আনুষ্ঠানিকভাবে, result[result_index] = operand[operand_index] যেখানে সমস্ত d axes(operand) :

  • operand_index[d] = 0 যদি dim(operand, d) = 1
  • operand_index[d] = result_index[broadcast_dimensions[d]] অন্যথায়।

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand টেনসর বা কোয়ান্টাইজড টেনসর (সি 1-সি 2), (সি 5-সি 6)
(I2) broadcast_dimensions টাইপ si64 এর 1-মাত্রিক টেনসর ধ্রুবক (সি 2-সি 6)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result টেনসর বা কোয়ান্টাইজড টেনসর (সি 1), (সি 3), (সি 5-সি 6)

সীমাবদ্ধতা

  • (সি 1) element_type(result) দ্বারা দেওয়া হয়েছে:
    • element_type(operand) , যদি !is_per_axis_quantized(operand)
    • element_type(operand) ব্যতীত quantization_dimension(operand) , scales(operand) , এবং zero_points(operand) quantization_dimension(result) , scales(result) , এবং zero_points(result) রেসপন্স থেকে পৃথক হতে পারে, অন্যথায়।
  • (সি 2) size(broadcast_dimensions) = rank(operand)
  • (সি 3) 0 <= broadcast_dimensions < rank(result)
  • (সি 4) is_unique(broadcast_dimensions)
  • (সি 5) সমস্ত d axes(operand) :
    • dim(operand, d) = 1 বা
    • dim(operand, d) = dim(result, broadcast_dimensions[d])
  • (সি 6) যদি is_per_axis_quantized(result) :
    • quantization_dimension(result) = broadcast_dimensions[quantization_dimension(operand)]
    • যদি dim(operand, quantization_dimension(operand)) = 1 , তবে scales(result)[i] = scales(operand)[0] and zero_points(result)[i] = zero_points(operand)[0] for i in range(dim(result, quantization_dimension(result)))

উদাহরণ

// %operand: [
//            [1, 2, 3]
//           ]
%result = "stablehlo.broadcast_in_dim"(%operand) {
  broadcast_dimensions = array<i64: 2, 1>
} : (tensor<1x3xi32>) -> tensor<2x3x2xi32>
// %result: [
//            [
//             [1, 1],
//             [2, 2],
//             [3, 3]
//            ],
//            [
//             [1, 1],
//             [2, 2],
//             [3, 3]
//            ]
//          ]

আরো উদাহরণ

মামলা

শব্দার্থবিদ্যা

index মানের উপর নির্ভর করে branches থেকে ঠিক একটি ফাংশন সম্পাদন করে আউটপুট উত্পাদন করে। আরও আনুষ্ঠানিকভাবে, result = selected_branch() যেখানে:

  • selected_branch = branches[index] যদি 0 <= index < size(branches)
  • selected_branch = branches[-1] অন্যথায়।

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) index টাইপ si32 এর 0-মাত্রিক টেনসর
(I2) branches ফাংশন ভেরিয়াদিক সংখ্যা (সি 1-সি 4)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
results টেনসর, কোয়ান্টাইজড টেনসর বা টোকেনগুলির ভেরিয়াদিক সংখ্যা (C4)

সীমাবদ্ধতা

  • (সি 1) 0 < size(branches)
  • (সি 2) input_types(branches...) = []
  • (সি 3) same(output_types(branches...))
  • (সি 4) type(results...) = output_types(branches[0])

উদাহরণ

// %index: -1
// %result_branch0: [0, 0]
// %result_branch1: [1, 1]
%result0, %result1 = "stablehlo.case"(%index) ({
  "stablehlo.return"(%result_branch0, %result_branch0) : (tensor<2xi64>, tensor<2xi64>) -> ()
}, {
  "stablehlo.return"(%result_branch1, %result_branch1) : (tensor<2xi64>, tensor<2xi64>) -> ()
}) : (tensor<i32>) -> (tensor<2xi64>, tensor<2xi64>)
// %result0: [1, 1]
// %result1: [1, 1]

আরো উদাহরণ

সিবিআরটি

শব্দার্থবিদ্যা

operand টেনসরে উপাদান-ভিত্তিক কিউবিক রুট অপারেশন সম্পাদন করে এবং result টেনসর তৈরি করে। উপাদান ধরণের উপর নির্ভর করে নিম্নলিখিতগুলি করে:

  • ভাসমানগুলির জন্য: আইইইই -754 থেকে rootn(x, 3)
  • জটিল সংখ্যার জন্য: জটিল কিউবিক রুট।
  • কোয়ান্টাইজড প্রকারের জন্য: dequantize_op_quantize(cbrt, operand, type(result))

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand ভাসমান-পয়েন্ট বা জটিল প্রকার বা প্রতি টেনসর কোয়ান্টাইজড টেনসর অফ টেনসর (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result ভাসমান-পয়েন্ট বা জটিল প্রকার বা প্রতি টেনসর কোয়ান্টাইজড টেনসর অফ টেনসর (C1)

সীমাবদ্ধতা

  • (সি 1) baseline_type(operand) = baseline_type(result)

উদাহরণ

// %operand: [0.0, 1.0, 8.0, 27.0]
%result = "stablehlo.cbrt"(%operand) : (tensor<4xf64>) -> tensor<4xf64>
// %result: [0.0, 1.0, 2.0, 3.0]

আরো উদাহরণ

ছাদ

শব্দার্থবিদ্যা

operand টেনসারের উপাদান-ভিত্তিক সিল সম্পাদন করে এবং একটি result টেনসর তৈরি করে। আইইইইই -754 স্পেসিফিকেশন থেকে roundToIntegralTowardPositive অপারেশন প্রয়োগ করে। কোয়ান্টাইজড প্রকারের জন্য, dequantize_op_quantize(ceil, operand, type(result)) সম্পাদন করে।

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand ভাসমান-পয়েন্ট টাইপ বা প্রতি টেনসর কোয়ান্টাইজড টেনসর অফ টেনসর (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result ভাসমান-পয়েন্ট টাইপ বা প্রতি টেনসর কোয়ান্টাইজড টেনসর অফ টেনসর (C1)

সীমাবদ্ধতা

  • (সি 1) baseline_type(operand) = baseline_type(result)

উদাহরণ

// %operand: [-0.8166, -0.2530, 0.2530, 0.8166, 2.0]
%result = "stablehlo.ceil"(%operand) : (tensor<5xf32>) -> tensor<5xf32>
// %result: [-0.0, -0.0, 1.0, 1.0, 2.0]

আরো উদাহরণ

চোলেস্কি

শব্দার্থবিদ্যা

ম্যাট্রিক্সের একটি ব্যাচের কোলেস্কি পচন গণনা করে।

আরও আনুষ্ঠানিকভাবে, i index_space(result) জন্য, result[i0, ..., iR-3, :, :] a[i0, ..., iR-3, :, :] , এর একটি কোলেস্কি পচন নিম্ন-ত্রিভুজাকার উভয় আকারে (যদি lower true হয়) বা উপরের-ত্রিভুজাকার (যদি lower false হয়) ম্যাট্রিক্স। বিপরীত ত্রিভুজের আউটপুট মানগুলি, অর্থাত্ উচ্চতর ত্রিভুজ বা কঠোর নিম্ন ত্রিভুজ অনুসারে, বাস্তবায়ন-সংজ্ঞায়িত।

যদি সেখানে উপস্থিত থাকে i ইনপুট ম্যাট্রিক্স কোনও হার্মিটিয়ান পজিটিভ-ডিফিনেট ম্যাট্রিক্স নয়, তবে আচরণটি অপরিজ্ঞাত।

কোয়ান্টাইজড প্রকারের জন্য, dequantize_op_quantize(lambda operand: cholesky(operand, lower), a, type(result)) সম্পাদন করে।

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) a ভাসমান-পয়েন্ট বা জটিল প্রকার বা প্রতি টেনসর কোয়ান্টাইজড টেনসর অফ টেনসর (সি 1-সি 3)
(I2) lower টাইপ i1 এর 0-মাত্রিক টেনসর ধ্রুবক

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result ভাসমান-পয়েন্ট বা জটিল প্রকার বা প্রতি টেনসর কোয়ান্টাইজড টেনসর অফ টেনসর (C1)

সীমাবদ্ধতা

  • (সি 1) baseline_type(a) = baseline_type(result)
  • (সি 2) 2 <= rank(a)
  • (সি 3) dim(a, -2) = dim(a, -1)

উদাহরণ

// %a: [
//      [1.0, 2.0, 3.0],
//      [2.0, 20.0, 26.0],
//      [3.0, 26.0, 70.0]
//     ]
%result = "stablehlo.cholesky"(%a) {
  lower = true
} : (tensor<3x3xf32>) -> tensor<3x3xf64>
// %result: [
//           [1.0, 0.0, 0.0],
//           [2.0, 4.0, 0.0],
//           [3.0, 5.0, 6.0]
//          ]

বাতা

শব্দার্থবিদ্যা

ন্যূনতম এবং সর্বোচ্চ মানের মধ্যে operand টেনসারের প্রতিটি উপাদানকে ক্ল্যাম্প করে এবং result টেনসর তৈরি করে। আরও আনুষ্ঠানিকভাবে, result[result_index] = minimum(maximum(operand[result_index], min_element), max_element) , যেখানে min_element = rank(min) = 0 ? min[] : min[result_index] , max_element = rank(max) = 0 ? max[] : max[result_index] । কোয়ান্টাইজড প্রকারের জন্য, dequantize_op_quantize(clamp, min, operand, max, type(result)) সম্পাদন করে।

জটিল সংখ্যার উপর অর্ডার দেওয়ার ক্ষেত্রে অবাক করা শব্দার্থবিজ্ঞানের সাথে জড়িত, তাই ভবিষ্যতে আমরা এই অপারেশনের জন্য জটিল সংখ্যার জন্য সমর্থন অপসারণের পরিকল্পনা করছি ( #560 )।

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) min টেনসর বা প্রতি টেনসর কোয়ান্টাইজড টেনসর (সি 1), (সি 3)
(I2) operand টেনসর বা প্রতি টেনসর কোয়ান্টাইজড টেনসর (সি 1-সি 4)
(I3) max টেনসর বা প্রতি টেনসর কোয়ান্টাইজড টেনসর (সি 2), (সি 3)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result টেনসর বা প্রতি টেনসর কোয়ান্টাইজড টেনসর (C4)

সীমাবদ্ধতা

  • (সি 1) rank(min) = 0 or shape(min) = shape(operand)
  • (সি 2) rank(max) = 0 or shape(max) = shape(operand)
  • baseline_element_type(min) = baseline_element_type(operand) = baseline_element_type(max)
  • (সি 4) baseline_type(operand) = baseline_type(result)

উদাহরণ

// %min: [5, 10, 15]
// %operand: [3, 13, 23]
// %max: [10, 15, 20]
%result = "stablehlo.clamp"(%min, %operand, %max) : (tensor<3xi32>, tensor<3xi32>, tensor<3xi32>) -> tensor<3xi32>
// %result: [5, 13, 20]

আরো উদাহরণ

সমষ্টিগত_ব্রোডকাস্ট

শব্দার্থবিদ্যা

স্থিতিশীলতা প্রক্রিয়া গ্রিডে প্রতিটি প্রক্রিয়া গোষ্ঠীর মধ্যে, উত্স প্রক্রিয়া থেকে operand টেনসারের মান লক্ষ্য প্রক্রিয়াগুলিতে প্রেরণ করুন এবং result টেনসর উত্পাদন করুন।

অপারেশনটি স্টেবলহ্লো প্রক্রিয়া গ্রিডকে process_groups গ্রুপগুলিতে বিভক্ত করে যা নিম্নলিখিত হিসাবে সংজ্ঞায়িত করা হয়েছে:

  • cross_replica(replica_groups) যদি channel_id <= 0
  • cross_partition(replica_groups) যদি channel_id > 0

এরপরে, result@process দেওয়া হয়:

  • operand@process_groups[i, 0] যদি সেখানে উপস্থিত থাকে i প্রক্রিয়াটি process_groups[i]
  • broadcast_in_dim(constant(is_quantized(result) ? quantize(0, element_type(result)) : 0, element_type(result)), [], type(result)) অন্যথায়।

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand টেনসর বা প্রতি টেনসর কোয়ান্টাইজড টেনসর (C3)
(I2) replica_groups টাইপ si64 এর 1-মাত্রিক টেনসর ধ্রুবকের ভেরিয়াদিক সংখ্যা (সি 1), (সি 2)
(I3) channel_id টাইপ si64 এর ধ্রুবক

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result টেনসর বা প্রতি টেনসর কোয়ান্টাইজড টেনসর (C3)

সীমাবদ্ধতা

  • (সি 1) is_unique(replica_groups)
  • (সি 2) 0 <= replica_groups < N যেখানে N হিসাবে সংজ্ঞায়িত করা হয়েছে:
    • num_replicas যদি cross_replica ব্যবহার করা হয়।
    • num_partitions যদি cross_partition ব্যবহার করা হয়।
  • (সি 3) type(result) = type(operand)

উদাহরণ

// num_replicas: 4
// num_partitions: 1
// %operand@(0, 0): [[1, 2]]
// %operand@(1, 0): [[3, 4]]
// %operand@(2, 0): [[5, 6]]
// %operand@(3, 0): [[7, 8]]
%result = "stablehlo.collective_broadcast"(%operand) {
  replica_groups = dense<[[2, 1]]> : tensor<1x2xi64>,
  channel_handle = #stablehlo.channel_handle<handle = 0, type = 0>
} : (tensor1x2xi64>) -> tensor<1x2xi64>
// %result@(0, 0): [[0, 0]]
// %result@(1, 0): [[5, 6]]
// %result@(2, 0): [[5, 6]]
// %result@(3, 0): [[0, 0]]

সমষ্টিগত_প্রেমুট

শব্দার্থবিদ্যা

স্থিতিশীলতা প্রক্রিয়া গ্রিডের প্রতিটি প্রক্রিয়া গোষ্ঠীর মধ্যে, উত্স প্রক্রিয়া থেকে operand টেনসারের মানটি লক্ষ্য প্রক্রিয়াতে প্রেরণ করে এবং result টেনসর তৈরি করে।

অপারেশনটি স্টেবলহ্লো প্রক্রিয়া গ্রিডকে process_groups গ্রুপগুলিতে বিভক্ত করে যা নিম্নলিখিত হিসাবে সংজ্ঞায়িত করা হয়েছে:

  • cross_replica(source_target_pairs) যদি channel_id <= 0
  • cross_partition(source_target_pairs) যদি channel_id > 0

এরপরে, result@process দেওয়া হয়:

  • operand@process_groups[i, 0] , যদি সেখানে উপস্থিত থাকে যে i যেমন process_groups[i, 1] = process
  • broadcast_in_dim(constant(is_quantized(result) ? quantize(0, element_type(result)) : 0, element_type(result)), [], type(result)) অন্যথায়।

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand টেনসর বা প্রতি টেনসর কোয়ান্টাইজড টেনসর (C5)
(I2) source_target_pairs টাইপ si64 এর 2-মাত্রিক টেনসর ধ্রুবক (সি 1-সি 4)
(I3) channel_id টাইপ si64 এর ধ্রুবক

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result টেনসর বা প্রতি টেনসর কোয়ান্টাইজড টেনসর (C1)

সীমাবদ্ধতা

  • (সি 1) dim(source_target_pairs, 1) = 2
  • (সি 2) is_unique(source_target_pairs[:, 0])
  • (সি 3) is_unique(source_target_pairs[:, 1])
  • (সি 4) 0 <= source_target_pairs < N , যেখানে N হিসাবে সংজ্ঞায়িত করা হয়েছে:
    • num_replicas যদি cross_replica ব্যবহার করা হয়।
    • num_partitions যদি cross_partition ব্যবহার করা হয়।
  • (সি 5) type(result) = type(operand)

উদাহরণ

// num_replicas: 3
// num_partitions: 1
// %operand@(0, 0): [[1, 2], [3, 4]]
// %operand@(1, 0): [[5, 6], [7, 8]]
// %operand@(2, 0): [[9, 10], [11, 12]]
%result = "stablehlo.collective_permute"(%operand) {
  source_target_pairs = dense<[[0, 1], [1, 2]]> : tensor<2x2xi64>,
  channel_handle = #stablehlo.channel_handle<handle = 0, type = 0>
} : (tensor<2x2xi64>) -> tensor<2x2xi64>
//
// %result@(0, 0): [[0, 0], [0, 0]]
// %result@(1, 0): [[1, 2], [3, 4]]
// %result@(2, 0): [[5, 6], [7, 8]]

আরো উদাহরণ

তুলনা

শব্দার্থবিদ্যা

comparison_direction এবং compare_type অনুসারে lhs এবং rhs টেনারগুলির উপাদান-ভিত্তিক তুলনা সম্পাদন করে এবং result টেনসর তৈরি করে।

comparison_direction এবং compare_type মানগুলি নিম্নলিখিত শব্দার্থবিজ্ঞান রয়েছে:

বুলিয়ান এবং পূর্ণসংখ্যার উপাদানগুলির জন্য:

  • EQ : lhs = rhs
  • NE : lhs != rhs
  • GE : lhs >= rhs
  • GT : lhs > rhs
  • LE : lhs <= rhs
  • LT : lhs < rhs

compare_type = FLOAT সহ ভাসমান-পয়েন্ট উপাদান প্রকারের জন্য, ওপি নিম্নলিখিত আইইইই -754 অপারেশনগুলি প্রয়োগ করে:

  • EQ : compareQuietEqual
  • NE : compareQuietNotEqual
  • GE : compareQuietGreaterEqual
  • GT : compareQuietGreater
  • LE : compareQuietLessEqual
  • LT : compareQuietLess

compare_type = TOTALORDER সহ ভাসমান-পয়েন্ট উপাদান প্রকারের জন্য, ওপি আইইইই -754 থেকে totalOrder এবং compareQuietEqual ক্রিয়াকলাপের সংমিশ্রণ ব্যবহার করে।

জটিল উপাদানগুলির জন্য, (real, imag) জোড়গুলির লিক্সোগ্রাফিক তুলনা প্রদত্ত comparison_direction এবং compare_type ব্যবহার করে সঞ্চালিত হয়। জটিল সংখ্যার উপর একটি ক্রম চাপানোতে আশ্চর্যজনক শব্দার্থবিজ্ঞানের সাথে জড়িত, সুতরাং ভবিষ্যতে আমরা comparison_direction GE , GT , LE বা LT ( #560 ) হলে জটিল সংখ্যার জন্য সমর্থন অপসারণের পরিকল্পনা করছি।

কোয়ান্টাইজড ধরণের জন্য। dequantize_compare(lhs, rhs, comparison_direction)

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs টেনসর বা প্রতি টেনসর কোয়ান্টাইজড টেনসর (সি 1-সি 3)
(I2) rhs টেনসর বা প্রতি টেনসর কোয়ান্টাইজড টেনসর (সি 1-সি 2)
(I3) comparison_direction EQ , NE , NE, GE , GT , LE , এবং LT এর এনাম
(আই 4) compare_type FLOAT , TOTALORDER , SIGNED এবং UNSIGNED এনাম (C3)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result বুলিয়ান প্রকারের টেনসর (C2)

সীমাবদ্ধতা

  • (সি 1) baseline_element_type(lhs) = baseline_element_type(rhs)
  • (সি 2) shape(lhs) = shape(rhs) = shape(result)
  • (সি 3) compare_type হিসাবে সংজ্ঞায়িত করা হয়েছে:
    • SIGNED যদি is_signed_integer(element_type(lhs))
    • UNSIGNED যদি is_unsigned_integer(element_type(lhs)) or is_boolean(element_type(lhs))
    • FLOAT বা TOTALORDER যদি is_float(element_type(lhs))
    • FLOAT যদি is_complex(element_type(lhs))

উদাহরণ

// %lhs: [1.0, 3.0]
// %rhs: [1.1, 2.9]
%result = "stablehlo.compare"(%lhs, %rhs) {
  comparison_direction = #stablehlo<comparison_direction LT>,
  compare_type = #stablehlo<comparison_type FLOAT>
} : (tensor<2xf32>, tensor<2xf32>) -> tensor<2xi1>
// %result: [true, false]

আরো উদাহরণ

জটিল

শব্দার্থবিদ্যা

বাস্তব এবং কাল্পনিক মান, lhs এবং rhs একজোড়া থেকে একটি জটিল মানতে উপাদান-ভিত্তিক রূপান্তর সম্পাদন করে এবং result টেনসর তৈরি করে।

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs টাইপ f32 বা f64 এর টেনসর (সি 1-সি 3)
(I2) rhs টাইপ f32 বা f64 এর টেনসর (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result জটিল প্রকারের টেনসর (সি 2), (সি 3)

সীমাবদ্ধতা

  • (সি 1) type(lhs) = type(rhs)
  • (সি 2) shape(result) = shape(lhs)
  • (সি 3) element_type(result) টাইপ complex<E> যেখানে E = element_type(lhs) রয়েছে।

উদাহরণ

// %lhs: [1.0, 3.0]
// %rhs: [2.0, 4.0]
%result = "stablehlo.complex"(%lhs, %rhs) : (tensor<2xf64>, tensor<2xf64>) -> tensor<2xcomplex<f64>>
// %result: [(1.0, 2.0), (3.0, 4.0)]

আরো উদাহরণ

যৌগিক

শব্দার্থবিদ্যা

inputs এবং composite_attributes গ্রহণ এবং results উত্পাদন করে এমন অন্যান্য স্থিতিশীল অপারেশনগুলির তৈরি (রচিত) একটি অপারেশনকে এনক্যাপসুলেট করে। ওপি -র শব্দার্থবিজ্ঞানগুলি decomposition বৈশিষ্ট্য দ্বারা প্রয়োগ করা হয়। composite ওপি প্রোগ্রামের শব্দার্থবিজ্ঞান পরিবর্তন না করে তার পচন দিয়ে প্রতিস্থাপন করা যেতে পারে। যেসব ক্ষেত্রে পচনকে অন্তর্ভুক্ত করা একই ওপি শব্দার্থবিজ্ঞান সরবরাহ করে না, custom_call ব্যবহার করা পছন্দ করে।

version ক্ষেত্রটি ( 0 এ ডিফল্ট) যখন কোনও সংমিশ্রণের শব্দার্থবিজ্ঞানের পরিবর্তন হয় তখন বোঝাতে ব্যবহৃত হয়।

ইনপুট

লেবেল নাম টাইপ
(I1) inputs মানগুলির ভেরিয়াদিক সংখ্যা
(I2) name টাইপ string ধ্রুবক
(I3) composite_attributes বৈশিষ্ট্য অভিধান
(আই 4) decomposition টাইপ string ধ্রুবক
(আই 5) version টাইপ si32 এর ধ্রুবক

আউটপুট

নাম টাইপ
results মানগুলির ভেরিয়াদিক সংখ্যা

সীমাবদ্ধতা

  • (সি 1) is_namespaced_op_name(name)
  • (সি 2) is_defined_in_parent_scope(decomposition)
  • (সি 3) types(inputs...) == input_types(decomposition)
  • (সি 4) types(results...) == output_types(decomposition)

উদাহরণ

%results = "stablehlo.composite"(%input0, %input1) {
  name = "my_namespace.my_op",
  composite_attributes = {
    my_attribute = "my_value"
  },
  decomposition = @my_op,
  version = 1 : i32
} : (tensor<f32>, tensor<f32>) -> tensor<f32>

আরো উদাহরণ

সংযুক্ত করা

শব্দার্থবিদ্যা

প্রদত্ত আর্গুমেন্ট হিসাবে একই ক্রমে dimension মাত্রা বরাবর inputs কনটেনটেটস এবং result টেনসর তৈরি করে। আরও আনুষ্ঠানিকভাবে, result[i0, ..., id, ..., iR-1] = inputs[k][i0, ..., kd, ..., iR-1] , যেখানে:

  1. id = d0 + ... + dk-1 + kd
  2. d dimension সমান, এবং d0 , ... d inputs মাত্রা মাপ।

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) inputs টেনার বা প্রতি টেনসর কোয়ান্টাইজড টেনারগুলির ভেরিয়াদিক সংখ্যা (সি 1-সি 6)
(I2) dimension টাইপ si64 এর ধ্রুবক (সি 2), (সি 4), (সি 6)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result টেনসর বা প্রতি টেনসর কোয়ান্টাইজড টেনসর (C5-C6)

সীমাবদ্ধতা

  • (সি 1) same(element_type(inputs...))
  • (সি 2) same(shape(inputs...)) dim(inputs..., dimension) ব্যতীত।
  • (সি 3) 0 < size(inputs)
  • (সি 4) 0 <= dimension < rank(inputs[0])
  • (সি 5) element_type(result) = element_type(inputs[0])
  • (সি 6) shape(result) = shape(inputs[0]) ব্যতীত:
    • dim(result, dimension) = dim(inputs[0], dimension) + ...

উদাহরণ

// %input0: [[1, 2], [3, 4], [5, 6]]
// %input1: [[7, 8]]
%result = "stablehlo.concatenate"(%input0, %input1) {
  dimension = 0 : i64
} : (tensor<3x2xi64>, tensor<1x2xi64>) -> tensor<4x2xi64>
// %result: [[1, 2], [3, 4], [5, 6], [7, 8]]

আরো উদাহরণ

ধ্রুবক

শব্দার্থবিদ্যা

একটি ধ্রুবক value থেকে একটি output টেনসর উত্পাদন করে।

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) value ধ্রুবক (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
output টেনসর বা কোয়ান্টাইজড টেনসর (C1)

সীমাবদ্ধতা

  • (সি 1) type(value) = type(output)

উদাহরণ

%output = "stablehlo.constant"() {
  value = dense<[[0.0, 1.0], [2.0, 3.0]]> : tensor<2x2xf32>
} : () -> tensor<2x2xf32>
// %output: [[0.0, 1.0], [2.0, 3.0]]

আরো উদাহরণ

রূপান্তর

শব্দার্থবিদ্যা

operand টেনসারে একটি উপাদান থেকে অন্য উপাদান থেকে অন্যটিতে একটি উপাদান-ভিত্তিক রূপান্তর সম্পাদন করে এবং result টেনসর তৈরি করে।

বুলিয়ান থেকে যে-কোনও-সমর্থিত-টাইপ রূপান্তরগুলির জন্য, মান false শূন্যে রূপান্তরিত হয় এবং true মানটি একটিতে রূপান্তরিত হয়। যে কোনও সমর্থিত-টাইপ-টু-বুলিয়ান রূপান্তরগুলির জন্য, একটি শূন্য মানকে false রূপান্তরিত করা হয়, এবং অ-শূন্য মানগুলি true রূপান্তরিত হয়। জটিল ধরণের জন্য এটি কীভাবে কাজ করে তার জন্য নীচে দেখুন।

পূর্ণসংখ্যা-থেকে-বুদ্ধিমান , পূর্ণসংখ্যা-থেকে-ভাসমান-পয়েন্ট বা ভাসমান-পয়েন্ট-থেকে-ভাসমান-পয়েন্ট জড়িত রূপান্তরগুলির জন্য, যদি উত্সের মানটি গন্তব্য প্রকারের মধ্যে হুবহু উপস্থাপিত হতে পারে তবে ফলাফলের মানটি সঠিক উপস্থাপনা। অন্যথায়, আচরণটি টিবিডি ( #180 )।

ভাসমান-পয়েন্ট-টু-ইনজিগারে জড়িত রূপান্তরগুলির জন্য, ভগ্নাংশের অংশটি কাটা হয়। যদি কাটা মানটি গন্তব্য প্রকারে প্রতিনিধিত্ব করা যায় না তবে আচরণটি টিবিডি ( #180 )।

জটিল থেকে জটিল জড়িত রূপান্তর বাস্তব এবং কল্পিত অংশগুলিকে রূপান্তর করার জন্য ভাসমান-পয়েন্ট-থেকে-ভাসমান-পয়েন্ট রূপান্তরগুলির একই আচরণ অনুসরণ করে।

জটিল-থেকে-অন্য-অন্য ধরণের এবং যে কোনও ধরণের থেকে জটিল রূপান্তরগুলির জন্য, উত্স কাল্পনিক মানটিকে উপেক্ষা করা হয় বা গন্তব্য কাল্পনিক মান যথাক্রমে শূন্য হয়। আসল অংশটির রূপান্তর ভাসমান-পয়েন্ট রূপান্তর অনুসরণ করে।

নীতিগতভাবে, এই অপারেশনটি ডেকুয়ান্টিজেশন (কোয়ান্টাইজড টেনসর থেকে নিয়মিত টেনারগুলিতে রূপান্তর), কোয়ান্টাইজেশন (নিয়মিত টেনারস থেকে কোয়ান্টাইজড টেনসরগুলিতে রূপান্তর) এবং প্রয়োজনীয়করণ (কোয়ান্টাইজড টেনসারের মধ্যে রূপান্তর) প্রকাশ করতে পারে, তবে এই মুহুর্তে আমরা সেই জন্য উত্সর্গীকৃত অপারেশনগুলি - uniform_dequantize প্রথম ব্যবহারের কেস এবং দ্বিতীয় এবং তৃতীয় ব্যবহারের ক্ষেত্রে uniform_quantize । ভবিষ্যতে, এই দুটি ওপি convert ( #1576 ) এ একীভূত হতে পারে।

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand টেনসর (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result টেনসর (C1)

সীমাবদ্ধতা

  • (সি 1) shape(operand) = shape(result)

উদাহরণ

// %operand: [-1, 0, 1]
%result = "stablehlo.convert"(%operand) : (tensor<3xi64>) -> tensor<3xcomplex<f64>>
// %result: [(-1.0, 0.0), (0.0, 0.0), (1.0, 0.0)]

আরো উদাহরণ

আবর্তন

শব্দার্থবিদ্যা

lhs উইন্ডো এবং rhs স্লাইসগুলির মধ্যে ডট পণ্য গণনা করে এবং result উত্পাদন করে। নিম্নলিখিত চিত্রটি দেখায় যে কীভাবে result উপাদানগুলি lhs এবং rhs থেকে একটি কংক্রিট উদাহরণ ব্যবহার করে গণনা করা হয়।

আবর্তন

আরও আনুষ্ঠানিকভাবে, lhs উইন্ডোজ প্রকাশ করতে সক্ষম হওয়ার জন্য lhs ক্ষেত্রে ইনপুটগুলির নিম্নলিখিত পুনর্নির্মাণের বিষয়টি বিবেচনা করুন:

  • lhs_window_dimensions = lhs_shape(dim(lhs, input_batch_dimension), dim(rhs, kernel_spatial_dimensions), dim(lhs, input_feature_dimension))
  • lhs_window_strides = lhs_shape(1, window_strides, 1)
  • lhs_padding = lhs_shape([0, 0], padding, [0, 0])
  • lhs_base_dilations = lhs_shape(1, lhs_dilation, 1)
  • lhs_window_dilations = lhs_shape(1, rhs_dilation, 1)

এই রিফ্রেমিংটি নিম্নলিখিত সহায়ক ফাংশনগুলি ব্যবহার করে:

  • lhs_shape(n, hw, c) = permute([n] + hw + [c], [input_batch_dimension] + input_spatial_dimensions + [input_feature_dimension])
  • result_shape(n1, hw, c1) = permute([n1] + hw + [c1], [output_batch_dimension] + output_spatial_dimensions + [output_feature_dimension])
  • permute([j0, j1, ..., jR-1], permutation) = [i0, i1, ..., iR-1] যেখানে j[d] = i[permutation[d]]

যদি feature_group_count = 1 এবং batch_group_count = 1 হয়, তবে সূচক_স্পেসে সমস্ত output_spatial_index জন্য index_space(dim(result, output_spatial_dimensions...)) , result[result_shape(:, output_spatial_index, :)] = dot_product যেখানে:

  • padding_value = constant(0, element_type(lhs))
  • padded_lhs = pad(lhs, padding_value, lhs_padding[:, 0], lhs_padding[:, 1], lhs_base_dilations - 1)
  • lhs_window_start = lhs_shape(0, output_spatial_index, 0) * lhs_window_strides
  • lhs_window = slice(padded_lhs, lhs_window_start, lhs_window_start + lhs_window_dimensions, lhs_window_dilations)
  • reversed_lhs_window = reverse(lhs_window, [input_spatial_dimensions[dim] for dim in range(size(window_reversal)) if window_reversal[dim] = true]) । এই বৈশিষ্ট্যটি অব্যবহৃত বলে মনে হচ্ছে, সুতরাং ভবিষ্যতে আমরা এটি অপসারণের পরিকল্পনা করছি ( #1181 )।
  • dot_product = dot_general(reversed_lhs_window, rhs, lhs_batching_dimensions=[], lhs_contracting_dimensions=input_spatial_dimensions + [input_feature_dimension], rhs_batching_dimensions=[], rhs_contracting_dimensions=kernel_spatial_dimensions + [kernel_input_feature_dimension])

যদি feature_group_count > 1 :

  • lhses = split(lhs, feature_group_count, input_feature_dimension)
  • rhses = split(rhs, feature_group_count, kernel_output_feature_dimension)
  • results... = convolution(lhses..., rhses..., ..., feature_group_count=1, ...)
  • result = concatenate(results, output_feature_dimension)

যদি batch_group_count > 1 :

  • lhses = split(lhs, batch_group_count, input_batch_dimension)
  • rhses = split(rhs, batch_group_count, kernel_output_feature_dimension)
  • results... = convolution(lhses..., rhses..., ..., batch_group_count=1, ...)
  • result = concatenate(results, output_feature_dimension)

কোয়ান্টাইজড প্রকারের জন্য, dequantize_op_quantize( lambda lhs, rhs: convolution(lhs, rhs, window_strides, padding, lhs_dilation, rhs_dilation, window_reversal, input_batch_dimension, input_feature_dimension, input_spatial_dimensions, kernel_input_feature_dimension, kernel_output_feature_dimension, kernel_spatial_dimensions, output_batch_dimension, output_feature_dimension, output_spatial_dimensions, feature_group_count, batch_group_count, precision_config), lhs, rhs, type(result))

হাইব্রিড কোয়ান্টাইজড প্রকারের জন্য, hybrid_dequantize_then_op( lambda lhs, rhs: convolution(lhs, rhs, window_strides, padding, lhs_dilation, rhs_dilation, window_reversal, input_batch_dimension, input_feature_dimension, input_spatial_dimensions, kernel_input_feature_dimension, kernel_output_feature_dimension, kernel_spatial_dimensions, output_batch_dimension, output_feature_dimension, output_spatial_dimensions, feature_group_count, batch_group_count, precision_config), lhs, rhs) সম্পাদন করে hybrid_dequantize_then_op( lambda lhs, rhs: convolution(lhs, rhs, window_strides, padding, lhs_dilation, rhs_dilation, window_reversal, input_batch_dimension, input_feature_dimension, input_spatial_dimensions, kernel_input_feature_dimension, kernel_output_feature_dimension, kernel_spatial_dimensions, output_batch_dimension, output_feature_dimension, output_spatial_dimensions, feature_group_count, batch_group_count, precision_config), lhs, rhs)

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs টেনসর বা প্রতি টেনসর কোয়ান্টাইজড টেনসর (সি 1), (সি 10-সি 11), (সি 14) (সি 25), (সি 27-সি 28), (সি 31-সি 32), (সি 34)
(I2) rhs টেনসর বা কোয়ান্টাইজড টেনসর (সি 1), (সি 14-সি 16), (সি 25), (সি 27-সি 29), (সি 31-সি 34)
(I3) window_strides টাইপ si64 এর 1-মাত্রিক টেনসর ধ্রুবক (সি 2-সি 3), (সি 25)
(আই 4) padding টাইপ si64 এর 2-মাত্রিক টেনসর ধ্রুবক (সি 4), (সি 25)
(আই 5) lhs_dilation টাইপ si64 এর 1-মাত্রিক টেনসর ধ্রুবক (সি 5-সি 6), (সি 25)
(আই 6) rhs_dilation টাইপ si64 এর 1-মাত্রিক টেনসর ধ্রুবক (সি 7-সি 8), (সি 25)
(I7) window_reversal টাইপ i1 এর 1-মাত্রিক টেনসর ধ্রুবক (সি 9)
(আই 8) input_batch_dimension টাইপ si64 এর ধ্রুবক (সি 10), (সি 13), (সি 25)
(আই 9) input_feature_dimension টাইপ si64 এর ধ্রুবক (সি 11), (সি 13-সি 14)
(আই 10) input_spatial_dimensions টাইপ si64 এর 1-মাত্রিক টেনসর ধ্রুবক (সি 12), (সি 13), (সি 25)
(আই 11) kernel_input_feature_dimension টাইপ si64 এর ধ্রুবক (সি 14), (সি 18)
(আই 12) kernel_output_feature_dimension টাইপ si64 এর ধ্রুবক (সি 15-সি 16), (সি 18), (সি 25), (সি 29)
(আই 13) kernel_spatial_dimensions টাইপ si64 এর 1-মাত্রিক টেনসর ধ্রুবক (সি 17-সি 18), (সি 25)
(আই 14) output_batch_dimension টাইপ si64 এর ধ্রুবক (সি 20), (সি 25)
(আই 15) output_feature_dimension টাইপ si64 এর ধ্রুবক (সি 20), (সি 25), (সি 30)
(আই 16) output_spatial_dimensions টাইপ si64 এর 1-মাত্রিক টেনসর ধ্রুবক (সি 19-সি 20), (সি 25)
(আই 17) feature_group_count টাইপ si64 এর ধ্রুবক (সি 11), (সি 14), (সি 16), (সি 21), (সি 23)
(আই 18) batch_group_count টাইপ si64 এর ধ্রুবক (সি 10), (সি 15), (সি 22), (সি 23), (সি 25)
(আই 19) precision_config DEFAULT , HIGH এবং HIGHEST এনামগুলির ভেরিয়াডিক সংখ্যা (সি 24)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result টেনসর বা কোয়ান্টাইজড টেনসর (সি 25-সি 28), (সি 30), (সি 32-34)

সীমাবদ্ধতা

  • (সি 1) N = rank(lhs) = rank(rhs)
  • (সি 2) size(window_strides) = N - 2
  • (সি 3) 0 < window_strides
  • (সি 4) shape(padding) = [N - 2, 2]
  • (সি 5) size(lhs_dilation) = N - 2
  • (সি 6) 0 < lhs_dilation
  • (সি 7) size(rhs_dilation) = N - 2
  • (সি 8) 0 < rhs_dilation
  • (সি 9) size(window_reversal) = N - 2
  • (সি 10) dim(lhs, input_batch_dimension) % batch_group_count = 0
  • (সি 11) dim(lhs, input_feature_dimension) % feature_group_count = 0
  • (সি 12) size(input_spatial_dimensions) = N - 2
  • input_dimensions = [input_batch_dimension] + input_spatial_dimensions + [input_feature_dimension]
    • is_unique(input_dimensions)
    • 0 <= input_dimensions < N
  • (C14) dim(rhs, kernel_input_feature_dimension) = dim(lhs, input_feature_dimension) / feature_group_count
  • (সি 15) dim(rhs, kernel_output_feature_dimension) % batch_group_count = 0
  • (সি 16) dim(rhs, kernel_output_feature_dimension) % feature_group_count = 0
  • (সি 17) size(kernel_spatial_dimensions) = N - 2
  • (C18) প্রদত্ত kernel_dimensions = kernel_spatial_dimensions + [kernel_input_feature_dimension] + [kernel_output_feature_dimension] :
    • is_unique(kernel_dimensions)
    • 0 <= kernel_dimensions < N
  • (সি 19) size(output_spatial_dimensions) = N - 2
  • (সি 20) প্রদত্ত output_dimensions = [output_batch_dimension] + output_spatial_dimensions + [output_feature_dimension] :
    • is_unique(output_dimensions)
    • 0 <= output_dimensions < N
  • (সি 21) 0 < feature_group_count
  • (সি 22) 0 < batch_group_count
  • (সি 23) feature_group_count = 1 or batch_group_count = 1
  • (সি 24) size(precision_config) = 2
  • (সি 25) dim(result, result_dim) হিসাবে সংজ্ঞায়িত করা হয়েছে:
    • dim(lhs, input_batch_dimension) / batch_group_count যদি result_dim = output_batch_dimension
    • dim(rhs, kernel_output_feature_dimension) যদি result_dim = output_feature_dimension
    • num_windows অন্যথায়, যেখানে:
    • output_spatial_dimensions[spatial_dim] = result_dim
    • lhs_dim = input_spatial_dimensions[spatial_dim]
    • rhs_dim = kernel_spatial_dimensions[spatial_dim]
    • dilated_input_shape[lhs_dim] = dim(lhs, lhs_dim) = 0 ? 0 : (dim(lhs, lhs_dim) - 1) * lhs_dilation[spatial_dim] + 1
    • padded_input_shape[lhs_dim] = padding[spatial_dim, 0] + dilated_input_shape[lhs_dim] + padding[spatial_dim, 1]
    • dilated_window_shape[lhs_dim] = dim(rhs, rhs_dim) = 0 ? 0 : (dim(rhs, rhs_dim) - 1) * rhs_dilation[spatial_dim] + 1
    • is_empty_window[lhs_dim] = padded_input_shape[lhs_dim] = 0 || dilated_window_shape[lhs_dim] > padded_input_shape[lhs_dim]
    • num_windows = is_empty_window[lhs_dim] ? 0 : floor((padded_input_shape[lhs_dim] - dilated_window_shape[lhs_dim]) / window_strides[spatial_dim]) + 1
  • (সি 26) rank(result) = N
  • যদি অপারেশনটি অ-কোয়ান্টাইজড টেনার ব্যবহার করে:
    • (সি 27) element_type(lhs) = element_type(rhs) = element_type(result)
  • যদি অপারেশনটি কোয়ান্টাইজড টেনার ব্যবহার করে:
    • (সি 28) is_quantized(lhs) = is_quantized(result) and is_quantized(rhs)
    • (C29) যদি is_per_axis_quantized(rhs) , তবে quantization_dimension(rhs) = kernel_output_feature_dimension
    • (সি 30) যদি is_per_axis_quantized(result) হয়, তবে quantization_dimension(result) = output_feature_dimension
    • যদি is_quantized(lhs) :
    • (সি 31) storage_type(lhs) = storage_type(rhs)
    • (সি 32) expressed_type(lhs) = expressed_type(rhs) = expressed_type(result)
    • (C33) যদি is_per_tensor_quantized(rhs) , তবে is_per_tensor_quantized(result)
    • যদি !is_quantized(lhs) :
    • (সি 34) element_type(lhs) = expressed_type(rhs) = element_type(result)

উদাহরণ

// %lhs: [[
//        [
//          [1], [2], [5], [6]
//        ],
//        [
//          [3], [4], [7], [8]
//        ],
//        [
//          [10], [11], [14], [15]
//        ],
//        [
//          [12], [13], [16], [17]
//        ]
//      ]]
//
// %rhs: [
//        [[[1]], [[1]], [[1]]],
//        [[[1]], [[1]], [[1]]],
//        [[[1]], [[1]], [[1]]]
//       ]
%result = "stablehlo.convolution"(%lhs, %rhs) {
  window_strides = array<i64: 4, 4>,
  padding = dense<0> : tensor<2x2xi64>,
  lhs_dilation = array<i64: 2, 2>,
  rhs_dilation = array<i64: 1, 1>,
  window_reversal = array<i1: false, false>,
  // In the StableHLO dialect, dimension numbers are encoded via:
  // `[<input dimensions>]x[<kernel dimensions>]->[output dimensions]`.
  // "b" is batch dimension, "f" is feature dimension,
  // "i" is input feature dimension, "o" is output feature dimension,
  // "0/1/etc" are spatial dimensions.
  dimension_numbers = #stablehlo.conv<[b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f]>,
  batch_group_count = 1 : i64,
  feature_group_count = 1 : i64,
  precision_config = [#stablehlo<precision DEFAULT>, #stablehlo<precision DEFAULT>]
} : (tensor<1x4x4x1xi64>, tensor<3x3x1x1xi64>) -> tensor<1x2x2x1xi64>
// %result: [[
//            [[10], [26]],
//            [[46], [62]]
//          ]]

আরো উদাহরণ

কোসাইন

শব্দার্থবিদ্যা

operand টেনসারে উপাদান-ভিত্তিক কোসাইন অপারেশন সম্পাদন করে এবং result টেনসর তৈরি করে। উপাদান ধরণের উপর নির্ভর করে নিম্নলিখিতগুলি করে:

  • ভাসমানের জন্য: আইইইই -754 থেকে cos
  • জটিল সংখ্যার জন্য: জটিল কোসাইন।
  • কোয়ান্টাইজড প্রকারের জন্য: dequantize_op_quantize(cosine, operand, type(result))

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand ভাসমান-পয়েন্ট বা জটিল প্রকার বা প্রতি টেনসর কোয়ান্টাইজড টেনসর অফ টেনসর (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result ভাসমান-পয়েন্ট বা জটিল প্রকার বা প্রতি টেনসর কোয়ান্টাইজড টেনসর অফ টেনসর (C1)

সীমাবদ্ধতা

  • (সি 1) baseline_type(operand) = baseline_type(result)

উদাহরণ

// %operand: [
//            [0.0, 1.57079632],       // [0, pi/2]
//            [3.14159265, 4.71238898] // [pi, 3pi/2]
//           ]
%result = "stablehlo.cosine"(%operand) : (tensor<2x2xf32>) -> tensor<2x2xf32>
// %result: [[1.0, 0.0], [-1.0, 0.0]]

আরো উদাহরণ

COUNT_LEADING_ZEROS

শব্দার্থবিদ্যা

operand টেনসারে শীর্ষস্থানীয় শূন্য বিটগুলির সংখ্যার উপাদান-ভিত্তিক গণনা সম্পাদন করে এবং result টেনসর তৈরি করে।

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand পূর্ণসংখ্যার প্রকারের টেনসর (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result পূর্ণসংখ্যার প্রকারের টেনসর (C1)

সীমাবদ্ধতা

  • (সি 1) type(operand) = type(result)

উদাহরণ

// %operand: [[0, 1], [128, -1]]
%result = "stablehlo.count_leading_zeros"(%operand) : (tensor<2x2xi64>) -> tensor<2x2xi64>
// %result: [[64, 63], [56, 0]]

আরো উদাহরণ

কাস্টম_ক্যাল

শব্দার্থবিদ্যা

একটি বাস্তবায়ন-সংজ্ঞায়িত অপারেশন call_target_name এনক্যাপসুলেট করে যা inputs নেয় এবং called_computations গ্রহণ করে এবং results দেয়। has_side_effect , backend_config এবং api_version অতিরিক্ত বাস্তবায়ন-সংজ্ঞায়িত মেটাডেটা সরবরাহ করতে ব্যবহৃত হতে পারে।

এই মুহুর্তে, এই অপারেশনে মেটাডেটার মোটামুটি বিশৃঙ্খলাযুক্ত সংগ্রহ রয়েছে যা এক্সএলএ সংকলকটিতে এর সমকক্ষ অপারেশনের জৈব বিবর্তনকে প্রতিফলিত করে। ভবিষ্যতে, আমরা এই মেটাডেটা ( #741 ) একত্রিত করার পরিকল্পনা করছি।

ইনপুট

লেবেল নাম টাইপ
(I1) inputs মানগুলির ভেরিয়াদিক সংখ্যা
(I2) call_target_name টাইপ string ধ্রুবক
(I3) has_side_effect টাইপ i1 এর ধ্রুবক
(আই 4) backend_config টাইপ string বা অ্যাট্রিবিউট অভিধানের ধ্রুবক
(আই 5) api_version টাইপ si32 এর ধ্রুবক
(আই 6) called_computations টাইপ string ধ্রুবকগুলির ভেরিয়াডিক সংখ্যা

আউটপুট

নাম টাইপ
results মানগুলির ভেরিয়াদিক সংখ্যা

উদাহরণ

%results = "stablehlo.custom_call"(%input0) {
  call_target_name = "foo",
  has_side_effect = false,
  backend_config = {bar = 42 : i32},
  api_version = 4 : i32,
  called_computations = [@foo]
} : (tensor<f64>) -> tensor<f64>

ভাগ করা

শব্দার্থবিদ্যা

লভ্যাংশ lhs এবং বিভাজক rhs টেনারগুলির উপাদান-ভিত্তিক বিভাগ সম্পাদন করে এবং result টেনসর তৈরি করে। উপাদান ধরণের উপর নির্ভর করে নিম্নলিখিতগুলি করে:

  • পূর্ণসংখ্যার জন্য: পূর্ণসংখ্যা বিভাগ যা কোনও ভগ্নাংশের অংশটি ফেলে দেওয়া বীজগণিতের ভাগফল উত্পাদন করে।
  • ভাসমানের জন্য: আইইইই -754 থেকে division
  • জটিল সংখ্যার জন্য: জটিল বিভাগ।
  • কোয়ান্টাইজড ধরণের জন্য:
    • dequantize_op_quantize(divide, lhs, rhs, type(result))

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs পূর্ণসংখ্যার টেনসর, ভাসমান-পয়েন্ট বা জটিল প্রকার বা প্রতি টেনসর কোয়ান্টাইজড টেনসর (C1)
(I2) rhs পূর্ণসংখ্যার টেনসর, ভাসমান-পয়েন্ট বা জটিল প্রকার বা প্রতি টেনসর কোয়ান্টাইজড টেনসর (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result পূর্ণসংখ্যার টেনসর, ভাসমান-পয়েন্ট, বা জটিল প্রকার বা প্রতি টেনসর কোয়ান্টাইজড টেনসর (C1)

সীমাবদ্ধতা

  • (সি 1) baseline_type(lhs) = baseline_type(rhs) = baseline_type(result)

উদাহরণ

// %lhs: [17.1, -17.1, 17.1, -17.1]
// %rhs: [3.0, 3.0, -3.0, -3.0]
%result = "stablehlo.divide"(%lhs, %rhs) : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
// %result: [5.66666651, -5.66666651, -5.66666651, 5.66666651]

আরো উদাহরণ

ডট_ জেনারাল

শব্দার্থবিদ্যা

lhs স্লাইস এবং rhs টুকরোগুলির মধ্যে ডট পণ্যগুলি গণনা করে এবং result টেনসর তৈরি করে।

আরও আনুষ্ঠানিকভাবে, result[result_index] = dot_product , যেখানে:

  • lhs_result_dimensions = [d for d in axes(lhs) and d not in lhs_batching_dimensions and d not in lhs_contracting_dimensions]
  • rhs_result_dimensions = [d for d in axes(rhs) and d not in rhs_batching_dimensions and d not in rhs_contracting_dimensions]
  • result_batching_index + result_lhs_index + result_rhs_index = result_index যেখানে size(result_batching_index) = size(lhs_batching_dimensions) size(result_lhs_index) = size(lhs_result_dimensions) size(result_rhs_index) = size(rhs_result_dimensions)
  • transposed_lhs = transpose(lhs, lhs_batching_dimensions + lhs_result_dimensions + lhs_contracting_dimensions)
  • transposed_lhs_slice = slice(transposed_lhs, result_batching_index + result_lhs_index + [:, ..., :])
  • reshaped_lhs_slice = reshape(transposed_lhs_slice, dims(lhs, lhs_contracting_dimensions))
  • transposed_rhs = transpose(rhs, rhs_batching_dimensions + rhs_result_dimensions + rhs_contracting_dimensions)
  • transposed_rhs_slice = slice(transposed_rhs, result_batching_index + result_rhs_index + [:, ..., :])
  • reshaped_rhs_slice = reshape(transposed_rhs_slice, dims(rhs, rhs_contracting_dimensions))
  • dot_product = reduce( inputs=[multiply(reshaped_lhs_slice, reshaped_rhs_slice)], init_values=[constant(0, element_type(result))], dimensions=range(size(lhs_contracting_dimensions)), body=lambda x, y: add(x, y))

কোয়ান্টাইজড প্রকারের জন্য, dequantize_op_quantize( lambda lhs, rhs: dot_general(lhs, rhs, lhs_batching_dimensions, rhs_batching_dimensions, lhs_contracting_dimensions, rhs_contracting_dimensions, precision_config), lhs, rhs, type(result))

হাইব্রিড কোয়ান্টাইজড প্রকারের জন্য, hybrid_dequantize_then_op( lambda lhs, rhs: dot_general(lhs, rhs, lhs_batching_dimensions, rhs_batching_dimensions, lhs_contracting_dimensions, rhs_contracting_dimensions, precision_config), lhs, rhs)

precision_config এক্সিলারেটর ব্যাকেন্ডগুলিতে গণনার জন্য গতি এবং নির্ভুলতার মধ্যে ট্রেড অফকে নিয়ন্ত্রণ করে। এটি নিম্নলিখিতগুলির মধ্যে একটি হতে পারে (এই মুহুর্তে, এই এনাম মানগুলির শব্দার্থবিজ্ঞানগুলি আন্ডারস্পাইফাইড করা হয়েছে, তবে আমরা এটিকে #755 এ সম্বোধন করার পরিকল্পনা করছি):

  • DEFAULT : দ্রুততম গণনা, তবে মূল সংখ্যায় কমপক্ষে সঠিক অনুমান।
  • HIGH : ধীর গণনা, তবে মূল সংখ্যার আরও সঠিক অনুমান।
  • HIGHEST : ধীরতম গণনা, তবে মূল সংখ্যার সর্বাধিক সঠিক অনুমান।

একটি DotAlgorithm ডট অপারেশন বাস্তবায়নের জন্য ব্যবহৃত অ্যালগরিদমের প্রধান বৈশিষ্ট্যগুলি সংজ্ঞায়িত করে, যা নির্ভুলতাও সংজ্ঞায়িত করে। যদি অ্যালগরিদম অ্যাট্রিবিউট ক্ষেত্রগুলি সেট করা থাকে তবে precision_config অবশ্যই DEFAULT হতে হবে। ডিফল্ট প্যারামিটারগুলি বাস্তবায়ন সংজ্ঞায়িত হওয়ায় DotAlgorithms কোনও ডিফল্ট মান নেই। এই হিসাবে, সমস্ত ডট অ্যালগরিদম ক্ষেত্রগুলি খালি ডট অ্যালগরিদম নির্দিষ্ট করার জন্য None সেট করা যেতে পারে, যা পরিবর্তে precision_config মানটি ব্যবহার করবে।

DotAlgorithm ক্ষেত্রগুলির মধ্যে রয়েছে:

  • lhs_precision_type এবং rhs_precision_type , অপারেশনের LHS এবং RHS এর যথাযথতাগুলি গোলাকার হয়। যথার্থ প্রকারগুলি ইনপুটগুলির স্টোরেজ প্রকার এবং আউটপুট থেকে স্বতন্ত্র।
  • জমা হওয়ার জন্য ব্যবহৃত নির্ভুলতা accumulation_type
  • lhs_component_count , rhs_component_count এবং num_primitive_operations প্রয়োগ করা হয় যখন আমরা এমন একটি অ্যালগরিদম করছি যা এলএইচএস এবং/বা আরএইচএসকে একাধিক উপাদানগুলিতে পচে যায় এবং এই মানগুলিতে একাধিক "আদিম" ডট অপারেশন করে - সাধারণত একটি উচ্চতর নির্ভুলতা অনুকরণ করার জন্য (যেমন একটি উচ্চতর নির্ভুলতা অনুকরণ করে। জন্য উচ্চ-নির্ভুলতা গণনা : BF16_6X TF32_3X, ইত্যাদি)। কোনও পচে যাওয়া অ্যালগরিদমের জন্য, এই মানগুলি 1 এ সেট করা উচিত।
  • কিছু পদক্ষেপের জন্য নিম্ন নির্ভুলতায় জমে থাকা অনুমোদিত কিনা তা নির্দিষ্ট করার জন্য allow_imprecise_accumulation (যেমন CUBLASLT_MATMUL_DESC_FAST_ACCUM )।

উদাহরণস্বরূপ DotAlgorithm বৈশিষ্ট্য:

// Inputs are casted to tf32, and then accumulated in f32:
{lhs_precision_type = tf32,
 rhs_precision_type = tf32,
 accumulation_type = f32,
 lhs_component_count = 1,
 rhs_component_count = 1,
 num_primitive_operations = 1,
 allow_imprecise_accumulation = false}


// bf16_6x: each input is decomposed to 3 bf16 components, then 6 dot operations are done on those components, and the result is accumulated in f32.
{lhs_precision_type = bf16,
 rhs_precision_type = bf16,
 accumulation_type = f32,
 lhs_component_count = 3,
 rhs_component_count = 3,
 num_primitive_operations = 6,
 allow_imprecise_accumulation = false}


// Inputs are (casted to) f8e5m2, and we accumulate in f32, but for some steps we may accumulate in lower precision.
{lhs_precision_type = f8e5m2,
 rhs_precision_type = f8e5m2,
 accumulation_type = f32,
 lhs_component_count = 1,
 rhs_component_count = 1,
 num_primitive_operations = 1,
 allow_imprecise_accumulation = true}

কোন সংমিশ্রণগুলি সমর্থিত তা সিদ্ধান্ত নেওয়া বাস্তবায়নগুলির উপর নির্ভর করে। সাধারণভাবে, এটি গ্যারান্টিযুক্ত নয় যে প্রতিটি অ্যালগরিদম স্থিতিশীলতার গ্রাহক দ্বারা প্রতিটি এক্সিলারেটর প্রকারে সমর্থিত। যদি কোনও প্রদত্ত অ্যালগরিদম সমর্থিত না হয় তবে বিকল্পের দিকে ফিরে যাওয়ার বিরোধিতা হিসাবে একটি ত্রুটি উত্থাপন করা উচিত। স্থিতিশীলতা যাচাইকরণটি সর্বোত্তম প্রচেষ্টা যাচাইকরণ সরবরাহ করবে, অ্যালগরিদমগুলি প্রতিরোধ করবে যা কোনও হার্ডওয়্যারটিতে সমর্থিত বলে জানা যায় না।

কিছু সমর্থিত অ্যালগরিদম মানগুলির জন্য xla_data.proto > Algorithm দেখুন। টিকিট #2483 ব্যাকএন্ড দ্বারা সমর্থিত অ্যালগরিদমগুলিতে একটি কেন্দ্রীভূত ডক তৈরির পরিকল্পনাটি ক্যাপচার করে।

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs টেনসর বা প্রতি টেনসর কোয়ান্টাইজড টেনসর (সি 5-সি 6), (সি 9-সি 10), (সি 12-সি 14), (সি 17-সি 18), (সি 20)
(I2) rhs টেনসর বা কোয়ান্টাইজড টেনসর (সি 7-সি 10), (সি 12-সি 20)
(I3) lhs_batching_dimensions টাইপ si64 এর 1-মাত্রিক টেনসর ধ্রুবক (সি 1), (সি 3), (সি 5), (সি 9), (সি 12)
(আই 4) rhs_batching_dimensions টাইপ si64 এর 1-মাত্রিক টেনসর ধ্রুবক (সি 1), (সি 4), (সি 7), (সি 9)
(আই 5) lhs_contracting_dimensions টাইপ si64 এর 1-মাত্রিক টেনসর ধ্রুবক (সি 2), (সি 3), (সি 6), (সি 10)
(আই 6) rhs_contracting_dimensions টাইপ si64 এর 1-মাত্রিক টেনসর ধ্রুবক (সি 2), (সি 4), (সি 8), (সি 10), (সি 16)
(I7) precision_config DEFAULT , HIGH এবং HIGHEST এনামগুলির ভেরিয়াডিক সংখ্যা (সি 11), (সি 21)
(আই 8) lhs_precision_type ফ্লোটটাইপ বা টেনসরফ্লোট 32 (সি 21)
(আই 9) rhs_precision_type ফ্লোটটাইপ বা টেনসরফ্লোট 32 (সি 21)
(আই 10) accumulation_type ফ্লোটটাইপ বা টেনসরফ্লোট 32 (সি 21)
(আই 11) lhs_component_count টাইপ si32 এর ধ্রুবক (সি 21), (সি 22)
(আই 12) rhs_component_count টাইপ si32 এর ধ্রুবক (সি 21), (সি 23)
(আই 13) num_primitive_operations টাইপ si32 এর ধ্রুবক (সি 21), (সি 24)
(আই 14) allow_imprecise_accumulation টাইপ bool ধ্রুবক (সি 21)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result টেনসর বা কোয়ান্টাইজড টেনসর (সি 12), (সি 14), (সি 18-সি 20)

সীমাবদ্ধতা

  • (সি 1) size(lhs_batching_dimensions) = size(rhs_batching_dimensions)
  • (সি 2) size(lhs_contracting_dimensions) = size(rhs_contracting_dimensions)
  • (সি 3) is_unique(lhs_batching_dimensions + lhs_contracting_dimensions)
  • (সি 4) is_unique(rhs_batching_dimensions + rhs_contracting_dimensions)
  • (সি 5) 0 <= lhs_batching_dimensions < rank(lhs)
  • (সি 6) 0 <= lhs_contracting_dimensions < rank(lhs)
  • (সি 7) 0 <= rhs_batching_dimensions < rank(rhs)
  • (সি 8) 0 <= rhs_contracting_dimensions < rank(rhs)
  • (সি 9) dim(lhs, lhs_batching_dimensions...) = dim(rhs, rhs_batching_dimensions...)
  • (সি 10) dim(lhs, lhs_contracting_dimensions...) = dim(rhs, rhs_contracting_dimensions...)
  • (সি 11) size(precision_config) = 2
  • (সি 12) shape(result) = dim(lhs, lhs_batching_dimensions) + dim(lhs, lhs_result_dimensions) + dim(rhs, rhs_result_dimensions)
  • যদি অপারেশনটি অ-কোয়ান্টাইজড টেনার ব্যবহার করে:
    • (সি 13) element_type(lhs) = element_type(rhs)
  • যদি অপারেশনটি কোয়ান্টাইজড টেনার ব্যবহার করে:
    • (সি 14) is_quantized(lhs) = is_quantized(result) and is_quantized(rhs)
    • (সি 15) zero_points(rhs) = 0
    • (সি 16) যদি is_per_axis_quantized(rhs) হয়, তবে quantization_dimension(rhs) rhs_contracting_dimensions নয়।
    • যদি is_quantized(lhs) :
    • (সি 17) storage_type(lhs) = storage_type(rhs)
    • (সি 18) expressed_type(lhs) = expressed_type(rhs) = expressed_type(result)
    • (C19) যদি is_per_tensor_quantized(rhs) , তবে is_per_tensor_quantized(result)
    • যদি !is_quantized(lhs) :
    • (সি 20) element_type(lhs) = expressed_type(rhs) = element_type(result)
  • যদি !is_empty_algorithm(lhs_precision_type, rhs_precision_type, accumulation_type, lhs_component_count, rhs_component_count, num_primitive_operations allow_imprecise_accumulation) :
    • (সি 21) precision_config... = DEFAULT
    • (C22) 0 < lhs_component_count
    • (C23) 0 < rhs_component_count
    • (C24) 0 < num_primitive_operations

উদাহরণ

// %lhs: [
//        [[1, 2],
//         [3, 4]],
//        [[5, 6],
//         [7, 8]]
//       ]
// %rhs: [
//        [[1, 0],
//         [0, 1]],
//        [[1, 0],
//         [0, 1]]
//       ]
%result = "stablehlo.dot_general"(%lhs, %rhs) {
  dot_dimension_numbers = #stablehlo.dot<
    lhs_batching_dimensions = [0],
    rhs_batching_dimensions = [0],
    lhs_contracting_dimensions = [2],
    rhs_contracting_dimensions = [1]
  >,
  precision_config = [#stablehlo<precision DEFAULT>, #stablehlo<precision DEFAULT>],
  algorithm = #stablehlo.dot_algorithm<
    lhs_precision_type = tf32,
    rhs_precision_type = tf32,
    accumulation_type = f32,
    lhs_component_count = 1,
    rhs_component_count = 1,
    num_primitive_operations = 1,
    allow_imprecise_accumulation = false
  >
} : (tensor<2x2x2xi64>, tensor<2x2x2xi64>) -> tensor<2x2x2xi64>
// %result: [
//           [[1, 2],
//            [3, 4]],
//           [[5, 6],
//            [7, 8]]
//          ]

আরো উদাহরণ

ডায়নামিক_ব্রোডকাস্ট_ইন_ডিম

শব্দার্থবিদ্যা

এই অপারেশনটি ব্রডকাস্ট_ইন_ডিম ওপি -র সাথে কার্যত অভিন্ন, তবে ফলাফলের আকারটি output_dimensions মাধ্যমে গতিশীলভাবে নির্দিষ্ট করা হয়।

অপারেশনটি মাত্রাগুলির প্রসারিত আচরণ সম্পর্কে স্থির জ্ঞান প্রকাশ করার জন্য known_expanding_dimensions , known_nonexpanding_dimensions চ্ছিক বৈশিষ্ট্যগুলিও গ্রহণ করে। যদি নির্দিষ্ট না করা হয় তবে সমস্ত মাত্রা সম্ভবত প্রসারিত বলে ধরে নেওয়া হয়।

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand টেনসর বা কোয়ান্টাইজড টেনসর (সি 1-সি 2), (সি 5-সি 6), (সি 9)
(I2) output_dimensions পূর্ণসংখ্যার প্রকারের 1-মাত্রিক টেনসর (C7)
(I3) broadcast_dimensions পূর্ণসংখ্যার ধরণের 1-মাত্রিক ধ্রুবক টেনসর (সি 2-সি 6)
(আই 4) known_expanding_dimensions পূর্ণসংখ্যার ধরণের 1-মাত্রিক ধ্রুবক টেনসর (সি 8-সি 9)
(আই 5) known_nonexpanding_dimensions পূর্ণসংখ্যার ধরণের 1-মাত্রিক ধ্রুবক টেনসর (সি 8-সি 9)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result টেনসর বা কোয়ান্টাইজড টেনসর (সি 1), (সি 3), (সি 5-সি 7)

সীমাবদ্ধতা

  • (সি 1) element_type(result) দ্বারা দেওয়া হয়েছে:
    • element_type(operand) , যদি !is_per_axis_quantized(operand)
    • element_type(operand) except that quantization_dimension(operand) , scales(operand) , and zero_points(operand) may differ from quantization_dimension(result) , scales(result) , and zero_points(result) resp., otherwise.
  • (C2) size(broadcast_dimensions) = rank(operand) .
  • (C3) 0 <= broadcast_dimensions < rank(result) .
  • (C4) is_unique(broadcast_dimensions) .
  • (C5) For all d in axes(operand) :
    • dim(operand, d) = 1 or
    • dim(operand, d) = dim(result, broadcast_dimensions[d]) .
  • (C6) If is_per_axis_quantized(result) :
    • quantization_dimension(result) = broadcast_dimensions[quantization_dimension(operand)] .
    • If dim(operand, quantization_dimension(operand)) = 1 , then scales(result)[i] = scales(operand)[0] and zero_points(result)[i] = zero_points(operand)[0] for i in range(dim(result, quantization_dimension(result))) .
  • (C7) size(output_dimensions) = rank(result) .
  • (C8) is_unique(known_expanding_dimensions + known_nonexpanding_dimensions) .
  • (C9) 0 <= known_expanding_dimensions < rank(operand) .
  • (C10) 0 <= known_nonexpanding_dimensions < rank(operand) .

উদাহরণ

// %operand: [
//            [1, 2, 3]
//           ]
%operand = stablehlo.constant dense<[[1, 2, 3]]> : tensor<1x3xi64>
%output_dimensions = stablehlo.constant dense<[2, 3, 2]> : tensor<3xi64>
%result = "stablehlo.dynamic_broadcast_in_dim"(%operand, %output_dimensions) {
  broadcast_dimensions = array<i64: 2, 1>,
  known_expanding_dimensions = array<i64: 0>,
  known_nonexpanding_dimensions = array<i64: 1>
} : (tensor<1x3xi64>, tensor<3xi64>) -> tensor<2x3x2xi64>
// %result: [
//            [
//             [1, 1],
//             [2, 2],
//             [3, 3]
//            ],
//            [
//             [1, 1],
//             [2, 2],
//             [3, 3]
//            ]
//          ]

আরো উদাহরণ

dynamic_conv

শব্দার্থবিদ্যা

This operation is functionally identical to convolution op, but the padding is specified dynamically via padding .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor or per-tensor quantized tensor (C1), (C10-C11), (C14) (C25), (C26-C27), (C30-C31), (C33)
(I2) rhs tensor or quantized tensor (C1), (C14-C16), (C26-C28), (C30-C33)
(I3) padding 2-dimensional tensor of integer type (C4)
(I4) window_strides 1-dimensional tensor constant of type si64 (C2-C3)
(I5) lhs_dilation 1-dimensional tensor constant of type si64 (C5-C6)
(I6) rhs_dilation 1-dimensional tensor constant of type si64 (C7-C8)
(I7) window_reversal 1-dimensional tensor constant of type i1 (C9)
(I8) input_batch_dimension constant of type si64 (C10), (C13)
(I9) input_feature_dimension constant of type si64 (C11), (C13-C14)
(I10) input_spatial_dimensions 1-dimensional tensor constant of type si64 (C12), (C13)
(I11) kernel_input_feature_dimension constant of type si64 (C14), (C18)
(I12) kernel_output_feature_dimension constant of type si64 (C15-C16), (C18), (C28)
(I13) kernel_spatial_dimensions 1-dimensional tensor constant of type si64 (C17-C18)
(I14) output_batch_dimension constant of type si64 (C20)
(I15) output_feature_dimension constant of type si64 (C20), (C29)
(I16) output_spatial_dimensions 1-dimensional tensor constant of type si64 (C19-C20)
(I17) feature_group_count constant of type si64 (C11), (C14), (C16), (C21), (C23)
(I18) batch_group_count constant of type si64 (C10), (C15), (C22), (C23)
(I19) precision_config variadic number of enums of DEFAULT , HIGH , and HIGHEST (C24)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or quantized tensor (C25-C27), (C29), (C31-C33)

সীমাবদ্ধতা

  • (C1) N = rank(lhs) = rank(rhs) .
  • (C2) size(window_strides) = N - 2 .
  • (C3) 0 < window_strides .
  • (C4) shape(padding) = [N - 2, 2] .
  • (C5) size(lhs_dilation) = N - 2 .
  • (C6) 0 < lhs_dilation .
  • (C7) size(rhs_dilation) = N - 2 .
  • (C8) 0 < rhs_dilation .
  • (C9) size(window_reversal) = N - 2 .
  • (C10) dim(lhs, input_batch_dimension) % batch_group_count = 0 .
  • (C11) dim(lhs, input_feature_dimension) % feature_group_count = 0 .
  • (C12) size(input_spatial_dimensions) = N - 2 .
  • (C13) Given input_dimensions = [input_batch_dimension] + input_spatial_dimensions + [input_feature_dimension] :
    • is_unique(input_dimensions) .
    • 0 <= input_dimensions < N .
  • (C14) dim(rhs, kernel_input_feature_dimension) = dim(lhs, input_feature_dimension) / feature_group_count .
  • (C15) dim(rhs, kernel_output_feature_dimension) % batch_group_count = 0 .
  • (C16) dim(rhs, kernel_output_feature_dimension) % feature_group_count = 0 .
  • (C17) size(kernel_spatial_dimensions) = N - 2 .
  • (C18) Given kernel_dimensions = kernel_spatial_dimensions + [kernel_input_feature_dimension] + [kernel_output_feature_dimension] :
    • is_unique(kernel_dimensions) .
    • 0 <= kernel_dimensions < N .
  • (C19) size(output_spatial_dimensions) = N - 2 .
  • (C20) Given output_dimensions = [output_batch_dimension] + output_spatial_dimensions + [output_feature_dimension] :
    • is_unique(output_dimensions) .
    • 0 <= output_dimensions < N .
  • (C21) 0 < feature_group_count .
  • (C22) 0 < batch_group_count .
  • (C23) feature_group_count = 1 or batch_group_count = 1 .
  • (C24) size(precision_config) = 2 .
  • (C25) dim(result, result_dim) is defined as:
    • dim(lhs, input_batch_dimension) / batch_group_count if result_dim = output_batch_dimension .
    • dim(rhs, kernel_output_feature_dimension) if result_dim = output_feature_dimension .
    • num_windows otherwise, where:
    • output_spatial_dimensions[spatial_dim] = result_dim .
    • lhs_dim = input_spatial_dimensions[spatial_dim] .
    • rhs_dim = kernel_spatial_dimensions[spatial_dim] .
    • dilated_input_shape[lhs_dim] = dim(lhs, lhs_dim) = 0 ? 0 : (dim(lhs, lhs_dim) - 1) * lhs_dilation[spatial_dim] + 1 .
    • padded_input_shape[lhs_dim] = padding[spatial_dim, 0] + dilated_input_shape[lhs_dim] + padding[spatial_dim, 1] .
    • dilated_window_shape[lhs_dim] = dim(rhs, rhs_dim) = 0 ? 0 : (dim(rhs, rhs_dim) - 1) * rhs_dilation[spatial_dim] + 1 .
    • is_empty_window[lhs_dim] = padded_input_shape[lhs_dim] = 0 || dilated_window_shape[lhs_dim] > padded_input_shape[lhs_dim] .
    • num_windows = is_empty_window[lhs_dim] ? 0 : floor((padded_input_shape[lhs_dim] - dilated_window_shape[lhs_dim]) / window_strides[spatial_dim]) + 1 .
  • (C26) rank(result) = N .
  • If the operation uses non-quantized tensors:
    • (C27) element_type(lhs) = element_type(rhs) = element_type(result) .
  • If the operation uses quantized tensors:
    • (C28) is_quantized(lhs) = is_quantized(result) and is_quantized(rhs) .
    • (C29) If is_per_axis_quantized(rhs) , then quantization_dimension(rhs) = kernel_output_feature_dimension .
    • (C30) If is_per_axis_quantized(result) , then quantization_dimension(result) = output_feature_dimension .
    • If is_quantized(lhs) :
    • (C31) storage_type(lhs) = storage_type(rhs) .
    • (C32) expressed_type(lhs) = expressed_type(rhs) = expressed_type(result) .
    • (C33) If is_per_tensor_quantized(rhs) , then is_per_tensor_quantized(result) .
    • If !is_quantized(lhs) :
    • (C34) element_type(lhs) = expressed_type(rhs) = element_type(result) .

উদাহরণ

// %lhs: [[
//        [[1], [2], [5], [6]],
//        [[3], [4], [7], [8]],
//        [[10], [11], [14], [15]],
//        [[12], [13], [16], [17]]
//      ]]
//
// %rhs: [
//         [[[1]], [[1]], [[1]]],
//         [[[1]], [[1]], [[1]]],
//         [[[1]], [[1]], [[1]]]
//        ]
// %padding: [[1, 1],
//            [1, 1]]
%result = "stablehlo.dynamic_conv"(%lhs, %rhs, %padding) {
  window_strides = array<i64: 4, 4>,
  lhs_dilation = array<i64: 2, 2>,
  rhs_dilation = array<i64: 1, 1>,
  window_reversal = array<i1: false, false>,
  dimension_numbers = #stablehlo.conv<raw
    input_batch_dimension = 0,
    input_feature_dimension = 3,
    input_spatial_dimensions = [0, 1],
    kernel_input_feature_dimension = 2,
    kernel_output_feature_dimension = 3,
    kernel_spatial_dimensions = [0, 1],
    output_batch_dimension = 0,
    output_feature_dimension = 3,
    output_spatial_dimensions = [1, 2]
  >,
  feature_group_count = 1 : i64,
  batch_group_count = 1 : i64,
  precision_config = [#stablehlo<precision DEFAULT>, #stablehlo<precision DEFAULT>]
} : (tensor<1x4x4x1xi64>, tensor<3x3x1x1xi64>, tensor<2x2xi64>) -> tensor<1x2x2x1xi64>
// %result: [[
//            [[1], [5]],
//            [[10], [14]]
//          ]]

আরো উদাহরণ

dynamic_gather

শব্দার্থবিদ্যা

This operation is functionally identical to gather op, with the slice_sizes specified dynamically as a value.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or per-tensor quantized tensor (C1), (C7), (C10-C12), (C14)
(I2) start_indices tensor of integer type (C2), (C3), (C13)
(I3) slice_sizes 1-dimensional tensor of integer type (C8), (C11-C13)
(I4) offset_dims 1-dimensional tensor constant of type si64 (C1), (C4-C5), (C13)
(I5) collapsed_slice_dims 1-dimensional tensor constant of type si64 (C1), (C6-C8), (C13)
(I6) start_index_map 1-dimensional tensor constant of type si64 (C3), (C9), (C10)
(I7) index_vector_dim constant of type si64 (C2), (C3), (C13)
(I8) indices_are_sorted constant of type i1

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C5), (C13-C14)

সীমাবদ্ধতা

  • (C1) rank(operand) = size(offset_dims) + size(collapsed_slice_dims) .
  • (C2) 0 <= index_vector_dim <= rank(start_indices) .
  • (C3) size(start_index_map) = index_vector_dim < rank(start_indices) ? dim(start_indices, index_vector_dim) : 1 .
  • (C4) is_unique(offset_dims) and is_sorted(offset_dims) .
  • (C5) 0 <= offset_dims < rank(result) .
  • (C6) is_unique(collapsed_slice_dims) and is_sorted(collapsed_slice_dims) .
  • (C7) 0 <= collapsed_slice_dims < rank(operand) .
  • (C8) slice_sizes[collapsed_slice_dims...] <= 1 .
  • (C9) is_unique(start_index_map) .
  • (C10) 0 <= start_index_map < rank(operand) .
  • (C11) size(slice_sizes) = rank(operand) .
  • (C12) 0 <= slice_sizes <= shape(operand) .
  • (C13) shape(result) = combine(batch_dim_sizes, offset_dim_sizes) where:
    • batch_dim_sizes = shape(start_indices) except that the dimension size of start_indices corresponding to index_vector_dim is not included.
    • offset_dim_sizes = shape(slice_sizes) except that the dimension sizes in slice_sizes corresponding to collapsed_slice_dims are not included.
    • combine puts batch_dim_sizes at axes corresponding to batch_dims and offset_dim_sizes at axes corresponding to offset_dims .
  • (C14) element_type(operand) = element_type(result) .

উদাহরণ

// %operand: [
//            [[1, 2], [3, 4], [5, 6], [7, 8]],
//            [[9, 10],[11, 12], [13, 14], [15, 16]],
//            [[17, 18], [19, 20], [21, 22], [23, 24]]
//           ]
// %start_indices: [
//                  [[0, 0], [1, 0], [2, 1]],
//                  [[0, 1], [1, 1], [0, 2]]
//                 ]
// %slize_sizes: [1, 2, 2]
%result = "stablehlo.dynamic_gather"(%operand, %start_indices, %slize_sizes) {
  dimension_numbers = #stablehlo.gather<
    offset_dims = [2, 3],
    collapsed_slice_dims = [0],
    start_index_map = [1, 0],
    index_vector_dim = 2>,
  indices_are_sorted = false
} : (tensor<3x4x2xi64>, tensor<2x3x2xi64>, tensor<3xi64>) -> tensor<2x3x2x2xi64>
// %result: [
//            [
//              [[1, 2], [3, 4]],
//              [[3, 4], [5, 6]],
//              [[13, 14], [15, 16]]
//            ],
//            [
//              [[9, 10], [11, 12]],
//              [[11, 12], [13, 14]],
//              [[17, 18], [19, 20]]
//            ]
//          ]

আরো উদাহরণ

dynamic_iota

শব্দার্থবিদ্যা

This operation is functionally identical to iota op, but the result shape is specified dynamically via output_shape .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) output_shape 1-dimensional tensor of integer type (C1), (C2)
(I2) iota_dimension si64 (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of integer, floating-point, or complex type or per-tensor quantized tensor (C2)

সীমাবদ্ধতা

  • (C1) 0 <= iota_dimension < size(output_shape) .
  • (C2) rank(result) = size(output_shape) .

উদাহরণ

%output_shape = stablehlo.constant dense<[4, 5]> : tensor<2xi64>
%result = "stablehlo.dynamic_iota"(%output_shape) {
  iota_dimension = 0 : i64
} : (tensor<2xi64>) -> tensor<4x5xi64>
// %result: [
//           [0, 0, 0, 0, 0],
//           [1, 1, 1, 1, 1],
//           [2, 2, 2, 2, 2],
//           [3, 3, 3, 3, 3]
//          ]

আরো উদাহরণ

dynamic_pad

শব্দার্থবিদ্যা

This operation is functionally identical to pad op, but with edge_padding_low , edge_padding_high , and interior_padding specified dynamically as values.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or per-tensor quantized tensor (C1), (C2), (C4)
(I2) padding_value 0-dimensional tensor or per-tensor quantized tensor (C1)
(I3) edge_padding_low 1-dimensional tensor of integer type (C1), (C4)
(I4) edge_padding_high 1-dimensional tensor of integer type (C1), (C4)
(I5) interior_padding 1-dimensional tensor of integer type (C2-C4)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C3-C6)

সীমাবদ্ধতা

  • (C1) element_type(operand) = element_type(padding_value) = element_type(result) .
  • (C2) size(edge_padding_low) = size(edge_padding_high) = size(interior_padding) = rank(operand) .
  • (C3) 0 <= interior_padding .
  • (C4) shape(result) = shape(operand) + edge_padding_low + max(shape(operand) - 1, 0) * interior_padding + edge_padding_high .

উদাহরণ

// %operand: [
//            [1, 2, 3],
//            [4, 5, 6]
//           ]
// %padding_value: 0
// %edge_padding_low: [0, 1]
// %edge_padding_high: [2, 1]
// %interior_padding: [1, 2]
%result = "stablehlo.dynamic_pad"(%operand, %padding_value,
  %edge_padding_low, %edge_padding_high, %interior_padding
) : (tensor<2x3xi64>, tensor<i64>, tensor<2xi64>, tensor<2xi64>, tensor<2xi64>) -> tensor<5x9xi64>
// %result: [
//           [0, 1, 0, 0, 2, 0, 0, 3, 0],
//           [0, 0, 0, 0, 0, 0, 0, 0, 0],
//           [0, 4, 0, 0, 5, 0, 0, 6, 0],
//           [0, 0, 0, 0, 0, 0, 0, 0, 0],
//           [0, 0, 0, 0, 0, 0, 0, 0, 0]
//          ]

আরো উদাহরণ

dynamic_reshape

শব্দার্থবিদ্যা

This operation is functionally identical to reshape op, but the result shape is specified dynamically via output_shape .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or quantized tensor (C1-C3)
(I2) output_shape 1-dimensional tensor of integer type (C4)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or quantized tensor (C1-C4)

সীমাবদ্ধতা

  • (C1) element_type(result) is given by:
    • element_type(operand) , if !is_per_axis_quantized(operand) .
    • element_type(operand) except that quantization_dimension(operand) and quantization_dimension(result) may differ, otherwise.
  • (C2) size(operand) = size(result) .
  • (C3) If is_per_axis_quantized(operand) :
    • reduce(dims(operand, [0, 1, ..., quantization_dimension(operand) - 1]), init_values=1, dimensions=[0], body=lambda x, y: x * y) = reduce(dims(result, [0, 1, ..., quantization_dimension(result) - 1]), init_values=1, dimensions=[0], body=lambda x, y: x * y) .
    • dim(operand, quantization_dimension(operand)) = dim(result, quantization_dimension(result)) .
    • reduce(dims(operand, [quantization_dimension(operand) + 1, ..., rank(operand) - 1]), init_values=1, dimensions=[0], body=lambda x, y: x * y) = reduce(dims(result, [quantization_dimension(result) + 1, ..., rank(result) - 1]), init_values=1, dimensions=[0], body=lambda x, y: x * y) .
  • (C4) size(output_shape) = rank(result) .

উদাহরণ

// %operand: [[1, 2, 3], [4, 5, 6]]
// %output_shape: [3, 2]
%result = "stablehlo.dynamic_reshape"(%operand, %output_shape) : (tensor<2x3xi64>, tensor<2xi64>) -> tensor<3x2xi64>
// %result: [[1, 2], [3, 4], [5, 6]]

আরো উদাহরণ

dynamic_slice

শব্দার্থবিদ্যা

Extracts a slice from the operand using dynamically-computed starting indices and produces a result tensor. start_indices contain the starting indices of the slice for each dimension subject to potential adjustment, and slice_sizes contain the sizes of the slice for each dimension. More formally, result[result_index] = operand[operand_index] where:

  • adjusted_start_indices = clamp(0, start_indices, shape(operand) - slice_sizes) .
  • operand_index = adjusted_start_indices + result_index .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or per-tensor quantized tensor (C1), (C2), (C4)
(I2) start_indices variadic number of 0-dimensional tensors of integer type (C2), (C3)
(I3) slice_sizes 1-dimensional tensor constant of type si64 (C2), (C4), (C5)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C1), (C5)

সীমাবদ্ধতা

  • (C1) element_type(operand) = element_type(result) .
  • (C2) size(start_indices) = size(slice_sizes) = rank(operand) .
  • (C3) same(type(start_indices...)) .
  • (C4) 0 <= slice_sizes <= shape(operand) .
  • (C5) shape(result) = slice_sizes .

উদাহরণ

// %operand: [
//            [0, 0, 1, 1],
//            [0, 0, 1, 1],
//            [0, 0, 0, 0],
//            [0, 0, 0, 0]
//           ]
// %start_indices0: -1
// %start_indices1: 3
%result = "stablehlo.dynamic_slice"(%operand, %start_indices0, %start_indices1) {
  slice_sizes = array<i64: 2, 2>
} : (tensor<4x4xi32>, tensor<i64>, tensor<i64>) -> tensor<2x2xi32>
// %result: [
//           [1, 1],
//           [1, 1]
//          ]

আরো উদাহরণ

dynamic_update_slice

শব্দার্থবিদ্যা

Produces a result tensor which is equal to the operand tensor except that the slice starting at start_indices is updated with the values in update . More formally, result[result_index] is defined as:

  • update[update_index] if 0 <= update_index < shape(update) where:
    • adjusted_start_indices = clamp(0, start_indices, shape(operand) - shape(update)) .
    • update_index = result_index - adjusted_start_indices .
  • operand[result_index] otherwise.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or per-tensor quantized tensor (C1-C4), (C6)
(I2) update tensor or per-tensor quantized tensor (C2), (C3), (C6)
(I3) start_indices variadic number of 0-dimensional tensors of integer type (C4), (C5)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) type(operand) = type(result) .
  • (C2) element_type(update) = element_type(operand) .
  • (C3) rank(update) = rank(operand) .
  • (C4) size(start_indices) = rank(operand) .
  • (C5) same(type(start_indices...)) .
  • (C6) 0 <= shape(update) <= shape(operand) .

উদাহরণ

// %operand: [
//            [1, 1, 0, 0],
//            [1, 1, 0, 0],
//            [1, 1, 1, 1],
//            [1, 1, 1, 1]
//           ]
// %update: [
//           [1, 1],
//           [1, 1]
//          ]
// %start_indices0: -1
// %start_indices1: 3
%result = "stablehlo.dynamic_update_slice"(%operand, %update, %start_indices0, %start_indices1)
  : (tensor<4x4xi32>, tensor<2x2xi32>, tensor<i64>, tensor<i64>) -> tensor<4x4xi32>
// %result: [
//           [1, 1, 1, 1],
//           [1, 1, 1, 1],
//           [1, 1, 1, 1],
//           [1, 1, 1, 1]
//          ]

আরো উদাহরণ

সূচকীয়

শব্দার্থবিদ্যা

Performs element-wise exponential operation on operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For floats: exp from IEEE-754.
  • For complex numbers: complex exponential.
  • For quantized types: dequantize_op_quantize(exponential, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [[0.0, 1.0], [2.0, 3.0]]
%result = "stablehlo.exponential"(%operand) : (tensor<2x2xf64>) -> tensor<2x2xf64>
// %result: [[1.0, 2.7182818284590451], [7.3890560989306504, 20.085536923187668]]

আরো উদাহরণ

exponential_minus_one

শব্দার্থবিদ্যা

Performs element-wise exponential minus one operation on operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For floats: expm1 from IEEE-754.
  • For complex numbers: complex exponential minus one.
  • For quantized types: dequantize_op_quantize(exponential_minus_one, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [0.0, 1.0]
%result = "stablehlo.exponential_minus_one"(%operand) : (tensor<2xf64>) -> tensor<2xf64>
// %result: [0.0, 1.71828187]

আরো উদাহরণ

fft

শব্দার্থবিদ্যা

Performs the forward and inverse Fourier transforms for real and complex inputs/outputs.

fft_type is one of the following:

  • FFT : Forward complex-to-complex FFT.
  • IFFT : Inverse complex-to-complex FFT.
  • RFFT : Forward real-to-complex FFT.
  • IRFFT : Inverse real-to-complex FFT (ie takes complex, returns real).

More formally, given the function fft which takes 1-dimensional tensors of complex types as input, produces 1-dimensional tensors of same types as output and computes the discrete Fourier transform:

For fft_type = FFT , result is defined as the final result of a series of L computations where L = size(fft_length) . For example, for L = 3 :

  • result1[i0, ..., :] = fft(operand[i0, ..., :]) .
  • result2[i0, ..., :, iR-1] = fft(result1[i0, ..., :, iR-1]) .
  • result[i0, ..., :, iR-2, iR-1] = fft(result2[i0, ..., :, iR-2, iR-1]) .

Furthermore, given the function ifft which has the same type signature and computes the inverse of fft :

For fft_type = IFFT , result is defined as the inverse of the computations for fft_type = FFT . For example, for L = 3 :

  • result1[i0, ..., :, iR-2, iR-1] = ifft(operand[i0, ..., :, iR-2, iR-1]) .
  • result2[i0, ..., :, iR-1] = ifft(result1[i0, ..., :, iR-1]) .
  • result[i0, ..., :] = ifft(result2[i0, ..., :]) .

Furthermore, given the function rfft which takes 1-dimensional tensors of floating-point types, produces 1-dimensional tensors of complex types of the same floating-point semantics and works as follows:

  • rfft(real_operand) = truncated_result where
  • complex_operand... = (real_operand..., 0.0) .
  • complex_result = fft(complex_operand) .
  • truncated_result = complex_result[:(rank(complex_result) / 2 + 1)] .

(When the discrete Fourier transform is computed for real operands, the first N/2 + 1 elements of the result unambiguously define the rest of the result, so the result of rfft is truncated to avoid computing redundant elements).

For fft_type = RFFT , result is defined as the final result of a series of L computations where L = size(fft_length) . For example, for L = 3 :

  • result1[i0, ..., :] = rfft(operand[i0, ..., :]) .
  • result2[i0, ..., :, iR-1] = fft(result1[i0, ..., :, iR-1]) .
  • result[i0, ..., :, iR-2, iR-1] = fft(result2[i0, ..., :, iR-2, iR-1]) .

Finally, given the function irfft which has the same type signature and computes the inverse of rfft :

For fft_type = IRFFT , result is defined as the inverse of the computations for fft_type = RFFT . For example, for L = 3 :

  • result1[i0, ..., :, iR-2, iR-1] = ifft(operand[i0, ..., :, iR-2, iR-1]) .
  • result2[i0, ..., :, iR-1] = ifft(result1[i0, ..., :, iR-1]) .
  • result[i0, ..., :] = irfft(result2[i0, ..., :]) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type (C1), (C2), (C4), (C5)
(I2) fft_type enum of FFT , IFFT , RFFT , and IRFFT (C2), (C5)
(I3) fft_length 1-dimensional tensor constant of type si64 (C1), (C3), (C4)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type (C2), (C4), (C5)

সীমাবদ্ধতা

  • (C1) size(fft_length) <= rank(operand) .
  • (C2) The relationship between operand and result element types varies:
    • If fft_type = FFT , element_type(operand) and element_type(result) have the same complex type.
    • If fft_type = IFFT , element_type(operand) and element_type(result) have the same complex type.
    • If fft_type = RFFT , element_type(operand) is a floating-point type and element_type(result) is a complex type of the same floating-point semantics.
    • If fft_type = IRFFT , element_type(operand) is a complex type and element_type(result) is a floating-point type of the same floating-point semantics.
  • (C3) 1 <= size(fft_length) <= 3 .
  • (C4) If among operand and result , there is a tensor real of a floating-point type, then shape(real)[-size(fft_length):] = fft_length .
  • (C5) shape(result) = shape(operand) except for:
    • If fft_type = RFFT , dim(result, -1) = dim(operand, -1) = 0 ? 0 : dim(operand, -1) / 2 + 1 .
    • If fft_type = IRFFT , dim(operand, -1) = dim(result, -1) = 0 ? 0 : dim(result, -1) / 2 + 1 .

উদাহরণ

// %operand: [(1.0, 0.0), (0.0, 0.0), (0.0, 0.0), (0.0, 0.0)]
%result = "stablehlo.fft"(%operand) {
  fft_type = #stablehlo<fft_type FFT>,
  fft_length = array<i64: 4>
} : (tensor<4xcomplex<f32>>) -> tensor<4xcomplex<f32>>
// %result: [(1.0, 0.0), (1.0, 0.0), (1.0, 0.0), (1.0, 0.0)]

মেঝে

শব্দার্থবিদ্যা

Performs element-wise floor of operand tensor and produces a result tensor. Implements the roundToIntegralTowardNegative operation from the IEEE-754 specification. For quantized types, performs dequantize_op_quantize(floor, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [-0.8166, -0.2530, 0.2530, 0.8166, 2.0]
%result = "stablehlo.floor"(%operand) : (tensor<5xf32>) -> tensor<5xf32>
// %result: [-1.0, -1.0, 0.0, 0.0, 2.0]

আরো উদাহরণ

জড়ো করা

শব্দার্থবিদ্যা

Gathers slices from operand tensor from offsets specified in start_indices and produces a result tensor.

The following diagram shows how elements in result map on elements in operand using a concrete example. The diagram picks a few example result indices and explains in detail which operand indices they correspond to.

জড়ো করা

More formally, result[result_index] = operand[operand_index] where:

  • batch_dims = [d for d in axes(result) and d not in offset_dims] .
  • batch_index = result_index[batch_dims...] .
  • start_index is defined as:
    • start_indices[bi0, ..., :, ..., biN] where bi are individual elements in batch_index and : is inserted at the index_vector_dim index, if index_vector_dim < rank(start_indices) .
    • [start_indices[batch_index]] otherwise.
  • For d_operand in axes(operand) ,
    • full_start_index[d_operand] = clamp(start_index[d_start], 0, dim(operand, d_operand) - slice_sizes[d_operand]) if d_operand = start_index_map[d_start] .
    • full_start_index[d_operand] = 0 otherwise.
  • For d_operand in axes(operand) ,
    • full_batching_index[d_operand] = batch_index[d_start - (d_start < index_vector_dim ? 0 : 1)] if d_operand = operand_batching_dims[i_batching] and d_start = start_indices_batching_dims[i_batching] .
    • full_batching_index[d_operand] = 0 otherwise.
  • offset_index = result_index[offset_dims...] .
  • full_offset_index = [oi0, ..., 0, ..., oiN] where oi are individual elements in offset_index , and 0 is inserted at indices from collapsed_slice_dims and operand_batching_dims .
  • operand_index = full_start_index + full_batching_index + full_offset_index .

If indices_are_sorted is true then the implementation can assume that start_indices are sorted with respect to start_index_map , otherwise the behavior is undefined. More formally, for all i1 < i2 from indices(result) , full_start_index(i1) <= full_start_index(i2) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or per-tensor quantized tensor (C1), (C8), (C11), (C17), (C19-C21), (C23)
(I2) start_indices tensor of integer type (C2-C3), (C14), (C17), (C22)
(I3) offset_dims 1-dimensional tensor constant of type si64 (C1), (C4-C5), (C22)
(I4) collapsed_slice_dims 1-dimensional tensor constant of type si64 (C1), (C6-C9), (C22)
(I5) operand_batching_dims 1-dimensional tensor constant of type si64 (C1), (C6), (C10-C12), (C16-C18), (C22)
(I6) start_indices_batching_dims 1-dimensional tensor constant of type si64 (C13-C17)
(I7) start_index_map 1-dimensional tensor constant of type si64 (C3), (C18-C19)
(I8) index_vector_dim constant of type si64 (C2-C3), (C15), (C22)
(I9) slice_sizes 1-dimensional tensor constant of type si64 (C9), (C12), (C20-C22)
(I10) indices_are_sorted constant of type i1

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C5), (C22-C23)

সীমাবদ্ধতা

  • (C1) rank(operand) = size(offset_dims) + size(collapsed_slice_dims) + size(operand_batching_dims) .
  • (C2) 0 <= index_vector_dim <= rank(start_indices) .
  • (C3) size(start_index_map) = index_vector_dim < rank(start_indices) ? dim(start_indices, index_vector_dim) : 1 .
  • (C4) is_unique(offset_dims) and is_sorted(offset_dims) .
  • (C5) 0 <= offset_dims < rank(result) .
  • (C6) is_unique(concatenate(collapsed_slice_dims, operand_batching_dims))
  • (C7) is_sorted(collapsed_slice_dims) .
  • (C8) 0 <= collapsed_slice_dims < rank(operand) .
  • (C9) slice_sizes[collapsed_slice_dims...] <= 1 .
  • (C10) is_sorted(operand_batching_dims) .
  • (C11) 0 <= operand_batching_dims < rank(operand) .
  • (C12) slice_sizes[operand_batching_dims...] <= 1 .
  • (C13) is_unique(start_indices_batching_dims) .
  • (C14) 0 <= start_indices_batching_dims < rank(start_indices) .
  • (C15) index_vector_dim not in start_indices_batching_dims .
  • (C16) size(operand_batching_dims) == size(start_indices_batching_dims) .
  • (C17) dim(operand, operand_batching_dims...) = dim(start_indices, start_indices_batching_dims...) .
  • (C18) is_unique(concatenate(start_index_map, operand_batching_dims)) .
  • (C19) 0 <= start_index_map < rank(operand) .
  • (C20) size(slice_sizes) = rank(operand) .
  • (C21) 0 <= slice_sizes <= shape(operand) .
  • (C22) shape(result) = combine(batch_dim_sizes, offset_dim_sizes) where:
    • batch_dim_sizes = shape(start_indices) except that the dimension size of start_indices corresponding to index_vector_dim is not included.
    • offset_dim_sizes = slice_sizes except that the dimension sizes in slice_sizes corresponding to collapsed_slice_dims and operand_batching_dims are not included.
    • combine puts batch_dim_sizes at axes corresponding to batch_dims and offset_dim_sizes at axes corresponding to offset_dims .
  • (C23) element_type(operand) = element_type(result) .

উদাহরণ

// %operand: [
//            [
//             [[1, 2], [3, 4], [5, 6], [7, 8]],
//             [[9, 10],[11, 12], [13, 14], [15, 16]],
//             [[17, 18], [19, 20], [21, 22], [23, 24]]
//            ],
//            [
//             [[25, 26], [27, 28], [29, 30], [31, 32]],
//             [[33, 34], [35, 36], [37, 38], [39, 40]],
//             [[41, 42], [43, 44], [45, 46], [47, 48]]
//            ]
//           ]
// %start_indices: [
//                  [
//                   [[0, 0], [1, 0], [2, 1]],
//                   [[0, 1], [1, 1], [0, 9]]
//                  ],
//                  [
//                   [[0, 0], [2, 1], [2, 2]],
//                   [[1, 2], [0, 1], [1, 0]]
//                  ]
//                 ]
%result = "stablehlo.gather"(%operand, %start_indices) {
  dimension_numbers = #stablehlo.gather<
    offset_dims = [3, 4],
    collapsed_slice_dims = [1],
    operand_batching_dims = [0],
    start_indices_batching_dims = [1],
    start_index_map = [2, 1],
    index_vector_dim = 3>,
  slice_sizes = array<i64: 1, 1, 2, 2>,
  indices_are_sorted = false
} : (tensor<2x3x4x2xi32>, tensor<2x2x3x2xi64>) -> tensor<2x2x3x2x2xi32>
// %result: [
//           [
//            [
//             [[1, 2], [3, 4]],
//             [[3, 4], [5, 6]],
//             [[13, 14], [15, 16]]
//            ],
//            [
//             [[33, 34], [35, 36]],
//             [[35, 36], [37, 38]],
//             [[41, 42], [43, 44]]
//            ]
//           ],
//           [
//            [
//             [[1, 2], [3, 4]],
//             [[13, 14], [15, 16]],
//             [[21, 22], [23, 24]]
//            ],
//            [
//             [[43, 44], [45, 46]],
//             [[33, 34], [35, 36]],
//             [[27, 28], [29, 30]]
//            ]
//           ]
//          ]

আরো উদাহরণ

get_dimension_size

শব্দার্থবিদ্যা

Produces the size of the given dimension of the operand . More formally, result = dim(operand, dimension) . The Semantics concerns only with the shape component of the type. The element-type could be anything.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or quantized tensor (C1)
(I2) dimension constant of type si64 (C1)

আউটপুট

নাম টাইপ
result 0-dimensional tensor of type si32

সীমাবদ্ধতা

  • (C1) 0 <= dimension < rank(operand) .

উদাহরণ

// %operand: [[1, 2, 3], [4, 5, 6]]
%result = "stablehlo.get_dimension_size"(%operand) {
  dimension = 1 : i64
} : (tensor<2x3xi64>) -> tensor<i32>
// %result: 3

আরো উদাহরণ

get_tuple_element

শব্দার্থবিদ্যা

Extracts element at index position of the operand tuple and produces a result . More formally, result = operand[index] .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand টিপল (C1), (C2)
(I2) index constant of type si32 (C1), (C2)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result any supported type (C2)

সীমাবদ্ধতা

  • (C1) 0 <= index < size(operand) .
  • (C2) type(result) = tuple_element_types(operand)[index] .

উদাহরণ

// %operand: ([1.0, 2.0], (3))
  index = 0 : i32
} : (tuple<tensor<2xf32>, tuple<tensor<i32>>>) -> tensor<2xf32>
// %result: [1.0, 2.0]

আরো উদাহরণ

যদি

শব্দার্থবিদ্যা

Produces the output from executing exactly one function from true_branch or false_branch depending on the value of pred . More formally, result = pred ? true_branch() : false_branch() .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) pred 0-dimensional tensor of type i1
(I2) true_branch ফাংশন (C1-C3)
(I3) false_branch ফাংশন (C1), (C2)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
results variadic number of tensors, quantized tensors or tokens (C3)

সীমাবদ্ধতা

  • (C1) input_types(true_branch) = input_types(false_branch) = [] .
  • (C2) output_types(true_branch) = output_types(false_branch) .
  • (C3) type(results...) = output_types(true_branch) .

উদাহরণ

// %result_true_branch: 10
// %result_false_branch: 11
// %pred: true
%result = "stablehlo.if"(%pred) ({
  "stablehlo.return"(%result_true_branch) : (tensor<i32>) -> ()
}, {
  "stablehlo.return"(%result_false_branch) : (tensor<i32>) -> ()
}) : (tensor<i1>) -> tensor<i32>
// %result: 10

আরো উদাহরণ

ইমেজ

শব্দার্থবিদ্যা

Extracts the imaginary part, element-wise, from the operand and produces a result tensor. More formally, for each element x : imag(x) = is_complex(x) ? imaginary_part(x) : constant(0, element_type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type (C1), (C2)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point type (C1), (C2)

সীমাবদ্ধতা

  • (C1) shape(result) = shape(operand) .
  • (C2) element_type(result) is defined as:
    • complex_element_type(element_type(operand)) if is_complex(operand) .
    • element_type(operand) otherwise.

উদাহরণ

// %operand: [(1.0, 2.0), (3.0, 4.0)]
%result = "stablehlo.imag"(%operand) : (tensor<2xcomplex<f32>>) -> tensor<2xf32>
// %result: [2.0, 4.0]

আরো উদাহরণ

infeed

শব্দার্থবিদ্যা

Reads data from the infeed and produces results .

Semantics of infeed_config is implementation-defined.

results consist of payload values which come first and a token which comes last. In the future, we are planning to split the payload and the token into two separate outputs to improve clarity ( #670 ).

ইনপুট

লেবেল নাম টাইপ
(I1) token token
(I2) infeed_config constant of type string

আউটপুট

নাম টাইপ সীমাবদ্ধতা
results variadic number of tensors, quantized tensors or tokens (C1-C3)

সীমাবদ্ধতা

  • (C1) 0 < size(results) .
  • (C2) is_empty(result[:-1]) or is_tensor(type(results[:-1])) .
  • (C3) is_token(type(results[-1])) .

উদাহরণ

// %token: !stablehlo.token
// infeed_queue[0]: [[1, 2], [3, 4]]
// infeed_queue[1]: [[5, 6], [7, 8]]
%results0:2 = "stablehlo.infeed"(%token) {
  infeed_config = ""
} : (!stablehlo.token) -> (tensor<2x2xi64>, !stablehlo.token)
// results0#0: [[1, 2], [3, 4]]
%results1:2 = "stablehlo.infeed"(%token) {
  infeed_config = ""
} : (!stablehlo.token) -> (tensor<2x2xi64>, !stablehlo.token)
// results1#0: [[5, 6], [7, 8]]

আরো উদাহরণ

iota

শব্দার্থবিদ্যা

Fills an output tensor with values in increasing order starting from zero along the iota_dimension dimension. আরো আনুষ্ঠানিকভাবে,

output[output_index] = constant(is_quantized(output) ? quantize(output_index[iota_dimension], element_type(output)) : output_index[iota_dimension], element_type(output)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) iota_dimension si64 (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
output tensor of integer, floating-point, or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) 0 <= iota_dimension < rank(output) .

উদাহরণ

%output = "stablehlo.iota"() {
  iota_dimension = 0 : i64
} : () -> tensor<4x5xi32>
// %output: [
//           [0, 0, 0, 0, 0],
//           [1, 1, 1, 1, 1],
//           [2, 2, 2, 2, 2],
//           [3, 3, 3, 3, 3]
//          ]

%output = "stablehlo.iota"() {
  iota_dimension = 1 : i64
} : () -> tensor<4x5xi32>
// %output: [
//           [0, 1, 2, 3, 4],
//           [0, 1, 2, 3, 4],
//           [0, 1, 2, 3, 4],
//           [0, 1, 2, 3, 4]
//          ]

আরো উদাহরণ

is_finite

শব্দার্থবিদ্যা

Performs element-wise check whether the value in x is finite (ie is neither +Inf, -Inf, nor NaN) and produces a y tensor. Implements the isFinite operation from the IEEE-754 specification. For quantized types, the result is always true .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) x tensor of floating-point type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
y tensor of boolean type (C1)

সীমাবদ্ধতা

  • (C1) shape(x) = shape(y) .

উদাহরণ

// Logical values: -Inf, +Inf, NaN, ...
// %x: [0xFFF0000000000000, 0x7FF0000000000000, 0x7FF8000000000000, -10.0, -0.0, 0.0, 10.0]
%y = "stablehlo.is_finite"(%x) : (tensor<7xf64) -> tensor<7xi1>
// %y: [false, false, false, true, true, true, true]

আরো উদাহরণ

লগ

শব্দার্থবিদ্যা

Performs element-wise logarithm operation on operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For floats: log from IEEE-754.
  • For complex numbers: complex logarithm.
  • For quantized types: dequantize_op_quantize(log, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [[1.0, 2.0], [3.0, 4.0]]
%result = "stablehlo.log"(%operand) : (tensor<2x2xf64>) -> tensor<2x2xf64>
// %result: [[0.0, 0.69314718055994529], [1.0986122886681098, 1.3862943611198906]]

আরো উদাহরণ

log_plus_one

শব্দার্থবিদ্যা

Performs element-wise logarithm plus one operation on operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For floats: logp1 from IEEE-754.
  • For complex numbers: complex(log(hypot(real(x) + 1, imag(x))), atan2(imag(x), real(x) + 1))
  • For quantized types: dequantize_op_quantize(log_plus_one, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [0.0, -0.999, 7.0, 6.38905621, 15.0]
%result = "stablehlo.log_plus_one"(%operand) : (tensor<5xf64>) -> tensor<5xf64>
// %result: [0.0, -6.90776825, 2.07944155, 2.0, 2.77258873]

আরো উদাহরণ

লজিস্টিক

শব্দার্থবিদ্যা

Performs element-wise logistic operation on operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For floats: division(1, addition(1, exp(-x))) from IEEE-754.
  • For complex numbers: complex logistic.
  • For quantized types: dequantize_op_quantize(logistic, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [[0.0, 1.0], [2.0, 3.0]]
%result = "stablehlo.logistic"(%operand) : (tensor<2x2xf64>) -> tensor<2x2xf64>
// %result: [[0.5, 0.73105858], [0.88079708, 0.95257413]]

আরো উদাহরণ

মানচিত্র

শব্দার্থবিদ্যা

Applies a map function computation to inputs along the dimensions and produces a result tensor.

More formally, result[result_index] = computation(inputs...[result_index]) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) inputs variadic number of tensors or per-tensor quantized tensors (C1-C4)
(I2) dimensions 1-dimensional tensor constant of type si64 (C3)
(I3) computation ফাংশন (C4)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C1), (C4)

সীমাবদ্ধতা

  • (C1) shape(inputs...) = shape(result) .
  • (C2) 0 < size(inputs) = N .
  • (C3) dimensions = range(rank(inputs[0])) .
  • (C4) computation has type (tensor<E0>, ..., tensor<EN-1>) -> tensor<E'> where Ei = element_type(inputs[i]) and E' = element_type(result) .

উদাহরণ

// %input0: [[0, 1], [2, 3]]
// %input1: [[4, 5], [6, 7]]
%result = "stablehlo.map"(%input0, %input1) ({
  ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>):
    %0 = stablehlo.multiply %arg0, %arg1 : tensor<i64>
    stablehlo.return %0 : tensor<i64>
}) {
  dimensions = array<i64: 0, 1>
} : (tensor<2x2xi64>, tensor<2x2xi64>) -> tensor<2x2xi64>
// %result: [[0, 5], [12, 21]]

আরো উদাহরণ

সর্বোচ্চ

শব্দার্থবিদ্যা

Performs element-wise max operation on tensors lhs and rhs and produces a result tensor. Depending on the element type, does the following:

  • For booleans: logical OR.
  • For integers: integer maximum.
  • For floats: maximum from IEEE-754.
  • For complex numbers: lexicographic maximum for the (real, imaginary) pair. Imposing an ordering on complex numbers involves surprising semantics, so in the future we are planning to remove support for complex numbers for this operation ( #560 ).
  • For quantized types:
    • dequantize_op_quantize(maximum, lhs, rhs, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor or per-tensor quantized tensor (C1)
(I2) rhs tensor or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(lhs) = baseline_type(rhs) = baseline_type(result) .

উদাহরণ

// %lhs: [[1, 2], [7, 8]]
// %rhs: [[5, 6], [3, 4]]
%result = "stablehlo.maximum"(%lhs, %rhs) : (tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32>
// %result: [[5, 6], [7, 8]]

আরো উদাহরণ

সর্বনিম্ন

শব্দার্থবিদ্যা

Performs element-wise min operation on tensors lhs and rhs and produces a result tensor. Depending on the element type, does the following:

  • For booleans: logical AND.
  • For integers: integer minimum.
  • For floats: minimum from IEEE-754.
  • For complex numbers: lexicographic minimum for the (real, imaginary) pair. Imposing an ordering on complex numbers involves surprising semantics, so in the future we are planning to remove support for complex numbers for this operation ( #560 ).
  • For quantized types:
    • dequantize_op_quantize(minimum, lhs, rhs, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor or per-tensor quantized tensor (C1)
(I2) rhs tensor or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(lhs) = baseline_type(rhs) = baseline_type(result) .

উদাহরণ

// %lhs: [[1, 2], [7, 8]]
// %rhs: [[5, 6], [3, 4]]
%result = "stablehlo.minimum"(%lhs, %rhs) : (tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32>
// %result: [[1, 2], [3, 4]]

আরো উদাহরণ

গুণ

শব্দার্থবিদ্যা

Performs element-wise product of two tensors lhs and rhs and produces a result tensor. Depending on the element type, does the following:

  • For booleans: logical AND.
  • For integers: integer multiplication.
  • For floats: multiplication from IEEE-754.
  • For complex numbers: complex multiplication.
  • For quantized types:
    • dequantize_op_quantize(multiply, lhs, rhs, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor or per-tensor quantized tensor (C1)
(I2) rhs tensor or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %lhs: [[1, 2], [3, 4]]
// %rhs: [[5, 6], [7, 8]]
%result = "stablehlo.multiply"(%lhs, %rhs) : (tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32>
// %result: [[5, 12], [21, 32]]

আরো উদাহরণ

অস্বীকার

শব্দার্থবিদ্যা

Performs element-wise negation of operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For signed integers: integer negation.
  • For unsigned integers: bitcast to signed integer, integer negation, bitcast back to unsigned integer.
  • For floats: negate from IEEE-754.
  • For complex numbers: complex negation.
  • For quantized types: dequantize_op_quantize(negate, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of integer, floating-point, or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of integer, floating-point, or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// Negation operation with integer Tensors
// %operand: [0, -2]
%result = "stablehlo.negate"(%operand) : (tensor<2xi32>) -> tensor<2xi32>
// %result: [0, 2]

// Negation operation with with complex tensors
// %operand: (2.5, 0.0)
%result = "stablehlo.negate"(%operand) : (tensor<1xcomplex<f32>>) -> tensor<1xcomplex<f32>>
// %result: [-2.5, -0.0]

আরো উদাহরণ

না

শব্দার্থবিদ্যা

Performs element-wise NOT of tensor operand and produces a result tensor. Depending on the element type, does the following:

  • For booleans: logical NOT.
  • For integers: bitwise NOT.

যুক্তি

নাম টাইপ সীমাবদ্ধতা
operand tensor of boolean or integer type (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of boolean or integer type (C1)

সীমাবদ্ধতা

  • (C1) type(operand) = type(result) .

উদাহরণ

// Bitwise operation with with integer tensors
// %operand: [[1, 2], [3, 4]]
%result = "stablehlo.not"(%operand) : (tensor<2x2xi32>) -> tensor<2x2xi32>
// %result: [[-2, -3], [-4, -5]]

// Bitwise operation with with boolean tensors
// %operand: [true, false]
%result = "stablehlo.not"(%operand) : (tensor<2xi1>) -> tensor<2xi1>
// %result: [false, true]

আরো উদাহরণ

optimization_barrier

শব্দার্থবিদ্যা

Ensures that the operations that produce the operand are executed before any operations that depend on the result and prevents compiler transformations from moving operations across the barrier. Other than that, the operation is an identity, ie result = operand .

যুক্তি

নাম টাইপ সীমাবদ্ধতা
operand variadic number of tensors, per-tensor quantized tensors or tokens (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result variadic number of tensors, per-tensor quantized tensors or tokens (C1)

সীমাবদ্ধতা

  • (C1) type(operand...) = type(result...) .

উদাহরণ

// %operand0: 0.0
// %operand1: 1.0
%result0, %result1 = "stablehlo.optimization_barrier"(%operand0, %operand1) : (tensor<f32>, tensor<f32>) -> (tensor<f32>, tensor<f32>)
// %result0: 0.0
// %result1: 1.0

আরো উদাহরণ

বা

শব্দার্থবিদ্যা

Performs element-wise OR of two tensors lhs and rhs and produces a result tensor. Depending on the element type, does the following:

  • For booleans: logical OR.
  • For integers: bitwise OR.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor of integer or boolean type (C1)
(I2) rhs tensor of integer or boolean type (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of integer or boolean type (C1)

সীমাবদ্ধতা

  • (C1) type(lhs) = type(rhs) = type(result) .

উদাহরণ

// Bitwise operation with with integer tensors
// %lhs: [[1, 2], [3, 4]]
// %rhs: [[5, 6], [7, 8]]
%result = "stablehlo.or"(%lhs, %rhs) : (tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32>
// %result: [[5, 6], [7, 12]]

// Logical operation with with boolean tensors
// %lhs: [[false, false], [true, true]]
// %rhs: [[false, true], [false, true]]
%result = "stablehlo.or"(%lhs, %rhs) : (tensor<2x2xi1>, tensor<2x2xi1>) -> tensor<2x2xi1>
// %result: [[false, true], [true, true]]

আরো উদাহরণ

outfeed

শব্দার্থবিদ্যা

Writes inputs to the outfeed and produces a result token.

Semantics of outfeed_config is implementation-defined.

ইনপুট

লেবেল নাম টাইপ
(I1) inputs variadic number of tensors or quantized tensors
(I2) token token
(I3) outfeed_config constant of type string

আউটপুট

নাম টাইপ
result token

উদাহরণ

%result = "stablehlo.outfeed"(%input0, %token) {
  outfeed_config = ""
} : (tensor<2x2x2xi64>, !stablehlo.token) -> !stablehlo.token

আরো উদাহরণ

প্যাড

শব্দার্থবিদ্যা

Expands operand by padding around the tensor as well as between the elements of the tensor with the given padding_value .

edge_padding_low and edge_padding_high specify the amount of padding added at the low-end (next to index 0) and the high-end (next to the highest index) of each dimension respectively. The amount of padding can be negative, where the absolute value of negative padding indicates the number of elements to remove from the specified dimension.

interior_padding specifies the amount of padding added between any two elements in each dimension which may not be negative. Interior padding occurs before edge padding such that negative edge padding will remove elements from the interior-padded operand.

More formally, result[result_index] is defined as:

  • operand[operand_index] if result_index = edge_padding_low + operand_index * (interior_padding + 1) .
  • padding_value otherwise.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or per-tensor quantized tensor (C1), (C2), (C4)
(I2) padding_value 0-dimensional tensor or per-tensor quantized tensor (C1)
(I3) edge_padding_low 1-dimensional tensor constant of type si64 (C1), (C4)
(I4) edge_padding_high 1-dimensional tensor constant of type si64 (C1), (C4)
(I5) interior_padding 1-dimensional tensor constant of type si64 (C2-C4)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C3-C6)

সীমাবদ্ধতা

  • (C1) element_type(operand) = element_type(padding_value) = element_type(result) .
  • (C2) size(edge_padding_low) = size(edge_padding_high) = size(interior_padding) = rank(operand) .
  • (C3) 0 <= interior_padding .
  • (C4) shape(result) = shape(operand) + edge_padding_low + max(shape(operand) - 1, 0) * interior_padding + edge_padding_high .

উদাহরণ

// %operand: [
//            [1, 2, 3],
//            [4, 5, 6]
//           ]
// %padding_value: 0
%result = "stablehlo.pad"(%operand, %padding_value) {
  edge_padding_low = array<i64: 0, 1>,
  edge_padding_high = array<i64: 2, 1>,
  interior_padding = array<i64: 1, 2>
} : (tensor<2x3xi32>, tensor<i32>) -> tensor<5x9xi32>
// %result: [
//           [0, 1, 0, 0, 2, 0, 0, 3, 0],
//           [0, 0, 0, 0, 0, 0, 0, 0, 0],
//           [0, 4, 0, 0, 5, 0, 0, 6, 0],
//           [0, 0, 0, 0, 0, 0, 0, 0, 0],
//           [0, 0, 0, 0, 0, 0, 0, 0, 0]
//          ]

আরো উদাহরণ

partition_id

শব্দার্থবিদ্যা

Produces partition_id of the current process.

আউটপুট

নাম টাইপ
result 0-dimensional tensor of type ui32

উদাহরণ

%result = "stablehlo.partition_id"() : () -> tensor<ui32>

আরো উদাহরণ

popcnt

শব্দার্থবিদ্যা

Performs element-wise count of the number of bits set in the operand tensor and produces a result tensor.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of integer type (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of integer type (C1)

সীমাবদ্ধতা

  • (C1) type(operand) = type(result) .

উদাহরণ

// %operand: [0, 1, 2, 127]
%result = "stablehlo.popcnt"(%operand) : (tensor<4xi64>) -> tensor<4xi64>
// %result: [0, 1, 1, 7]

আরো উদাহরণ

ক্ষমতা

শব্দার্থবিদ্যা

Performs element-wise exponentiation of lhs tensor by rhs tensor and produces a result tensor. Depending on the element type, does the following:

  • For integers: integer exponentiation.
  • For floats: pow from IEEE-754.
  • For complex numbers: complex exponentiation.
  • For quantized types: dequantize_op_quantize(power, lhs, rhs, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor of integer, floating-point, or complex type or per-tensor quantized tensor (C1)
(I2) rhs tensor of integer, floating-point, or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of integer, floating-point, or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %lhs: [-2.0, -0.0, -36.0, 5.0, 3.0, 10000.0]
// %rhs: [2.0, 2.0, 1.1, 2.0, -1.0, 10.0]
%result = "stablehlo.power"(%lhs, %rhs) : (tensor<6xf64>, tensor<6xf64>) -> tensor<6xf64>
// %result: [4.0, 0.0, -nan, 25.0, 0.333333343, inf]

আরো উদাহরণ

বাস্তব

শব্দার্থবিদ্যা

Extracts the real part, element-wise, from the operand and produces a result tensor. More formally, for each element x : real(x) = is_complex(x) ? real_part(x) : x .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type (C1), (C2)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point type (C1), (C2)

সীমাবদ্ধতা

  • (C1) shape(result) = shape(operand) .
  • (C2) element_type(result) is defined as:
    • complex_element_type(element_type(operand)) if is_complex(operand) .
    • element_type(operand) otherwise.

উদাহরণ

// %operand: [(1.0, 2.0), (3.0, 4.0)]
%result = "stablehlo.real"(%operand) : (tensor<2xcomplex<f32>>) -> tensor<2xf32>
// %result: [1.0, 3.0]

আরো উদাহরণ

recv

শব্দার্থবিদ্যা

Receives data from a channel with channel_id and produces results .

If is_host_transfer is true , then the operation transfers data from the host. Otherwise, it transfers data from another device. What this means is implementation-defined. This flag duplicates the information provided in channel_type , so in the future we are planning to only keep one of them ( #666 ).

results consist of payload values which come first and a token which comes last. In the future, we are planning to split the payload and the token into two separate outputs to improve clarity ( #670 ).

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) token token (C4)
(I2) channel_id constant of type si64
(I3) channel_type enum of DEVICE_TO_DEVICE and HOST_TO_DEVICE (C1)
(I4) is_host_transfer constant of type i1 (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
results variadic number of tensors, quantized tensors or tokens (C2-C4)

সীমাবদ্ধতা

  • (C1) channel_type is defined as:
    • HOST_TO_DEVICE if is_host_transfer = true ,
    • DEVICE_TO_DEVICE otherwise.
  • (C2) 0 < size(results) .
  • (C3) is_empty(result[:-1]) or is_tensor(type(results[:-1])) .
  • (C4) is_token(type(results[-1])) .

উদাহরণ

%results0, %results1 = "stablehlo.recv"(%token) {
  channel_handle = #stablehlo.channel_handle<handle = 1, type = 3>,
  is_host_transfer = true
} : (!stablehlo.token) -> (tensor<2x2xi64>, !stablehlo.token)

আরো উদাহরণ

হ্রাস করা

শব্দার্থবিদ্যা

Applies a reduction function body to inputs and init_values along the dimensions and produces results tensors.

The order of reductions is implementation-defined, which means that body and init_values must form a monoid to guarantee that the operation produces the same results for all inputs on all implementations. However, this condition doesn't hold for many popular reductions. Eg floating-point addition for body and zero for init_values don't actually form a monoid because floating-point addition is not associative.

More formally, results...[j0, ..., jR-1] = reduce(input_slices_converted) where:

  • input_slices = inputs...[j0, ..., :, ..., jR-1] , where : are inserted at dimensions .
  • input_slices_converted = to_destination_type(input_slices..., type(func_inputs(body)[:len(func_inputs(body))//2])...) .
  • init_values_converted = to_destination_type(init_values..., type(func_inputs(body)[len(func_inputs(body))//2:])...) .
  • reduce(input_slices_converted) = exec(schedule) for some binary tree schedule where:
    • exec(node) = body(exec(node.left), exec(node.right)) .
    • exec(leaf) = leaf.value .
  • schedule is an implementation-defined full binary tree whose in-order traversal consists of:
    • input_slices_converted...[index] values, for all index in index_space(input_slices_converted) in the ascending lexicographic order of index .
    • Interspersed with an implementation-defined amount of init_values_converted at implementation-defined positions.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) inputs variadic number of tensors or per-tensor quantized tensors (C1-C4), (C6), (C7)
(I2) init_values variadic number of 0-dimensional tensors or per-tensor quantized tensors (C2), (C3)
(I3) dimensions 1-dimensional tensor constant of type si64 (C4), (C5), (C7)
(I4) body ফাংশন (C6)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
results variadic number of tensors or per-tensor quantized tensors (C3), (C7), (C8)

সীমাবদ্ধতা

  • (C1) same(shape(inputs...)) .
  • (C2) element_type(inputs...) = element_type(init_values...) .
  • (C3) 0 < size(inputs) = size(init_values) = size(results) = N .
  • (C4) 0 <= dimensions < rank(inputs[0]) .
  • (C5) is_unique(dimensions) .
  • (C6) body has type (tensor<E0>, ..., tensor<EN-1>, tensor<E0>, ..., tensor<EN-1>) -> (tensor<E0>, ..., tensor<EN-1>) where is_promotable(element_type(inputs[i]), Ei) .
  • (C7) shape(results...) = shape(inputs...) except that the dimension sizes of inputs... corresponding to dimensions are not included.
  • (C8) element_type(results[i]) = Ei for all i in [0,N) .

উদাহরণ

// %input = [[0, 1, 2, 3, 4, 5]]
// %init_value = 0
%result = "stablehlo.reduce"(%input, %init_value) ({
  ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>):
    %0 = "stablehlo.add"(%arg0, %arg1) : (tensor<i64>, tensor<i64>) -> tensor<i64>
    "stablehlo.return"(%0) : (tensor<i64>) -> ()
}) {
  dimensions = array<i64: 1>
} : (tensor<1x6xi64>, tensor<i64>) -> tensor<1xi64>
// %result = [15]

আরো উদাহরণ

reduce_precision

শব্দার্থবিদ্যা

Performs element-wise conversion of operand to another floating-point type that uses exponent_bits and mantissa_bits and back to the original floating-point type and produces an output tensor.

আরো আনুষ্ঠানিকভাবে:

  • The mantissa bits of the original value are updated to round the original value to the nearest value representable with mantissa_bits using roundToIntegralTiesToEven semantics.
  • Then, if mantissa_bits are smaller than the number of mantissa bits of the original value, the mantissa bits are truncated to mantissa_bits .
  • Then, if the exponent bits of the intermediate result don't fit into the range provided by exponent_bits , the intermediate result overflows to infinity using the original sign or underflows to zero using the original sign.
  • For quantized types, performs dequantize_op_quantize( lambda operand: reduce_precision(operand, exponent_bits, mantissa_bits), operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point type or per-tensor quantized tensor (C1)
(I2) exponent_bits constant of type si32 (C2)
(I3) mantissa_bits constant of type si32 (C3)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
output tensor of floating-point type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(output) .
  • (C2) 1 <= exponent_bits .
  • (C3) 0 <= mantissa_bits .

উদাহরণ

// Logical values: +Inf, NaN, +Denormal, 0.0, 65519.0, 65520.0
// %operand: [0x7FF0000000000000, 0x7FFFFFFFFFFFFFFF, 0x0000000000000001, 0.0, 65519.0, 65520.0]
%output = "stablehlo.reduce_precision"(%operand) {
  exponent_bits = 5 : i32,
  mantissa_bits = 10 : i32
} : (tensor<6xf64>) -> tensor<6xf64>
// Logical values: +Inf, NaN, 0.0, 0.0, 65504.0, +Inf
// %output: [0x7FF0000000000000, 0x7FFFFFFFFFFFFFFF, 0.0, 0.0, 65504.0, 0x7FF0000000000000]

আরো উদাহরণ

reduce_scatter

শব্দার্থবিদ্যা

reduce_scatter

Within each process group in the StableHLO process grid, performs reduction, using computations , over the values of the operand tensor from each process, splits the reduction result along scatter_dimension into parts, and scatters the split parts between the processes to produce the result .

The operation splits the StableHLO process grid into process_groups which is defined as follows:

  • cross_replica(replica_groups) if channel_id <= 0 and use_global_device_ids = false .
  • cross_replica_and_partition(replica_groups) if channel_id > 0 and use_global_device_ids = false .
  • flattened_ids(replica_groups) if channel_id > 0 and use_global_device_ids = true .

Afterwards, within each process_group :

  • reduced_value = all_reduce(operand, replica_groups, channel_id, use_global_device_ids, computation) .
  • parts@sender = split(reduced_value@sender, dim(process_groups, 1), scatter_dimension) .
  • result@receiver = parts@sender[receiver_index] for all sender in process_group , where receiver_index = process_group.index(receiver) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or per-tensor quantized tensor (C1), (C2), (C7), (C8)
(I2) scatter_dimension constant of type si64 (C1), (C2), (C8)
(I3) replica_groups 2-dimensional tensor constant of type si64 (C3-C5)
(I4) channel_id constant of type si64 (C6)
(I5) use_global_device_ids constant of type i1 (C6)
(I6) computation ফাংশন (C7)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C8-C9)

সীমাবদ্ধতা

  • (C1) dim(operand, scatter_dimension) % dim(process_groups, 1) = 0 .
  • (C2) 0 <= scatter_dimension < rank(operand) .
  • (C3) is_unique(replica_groups) .
  • (C4) size(replica_groups) is defined as:
    • num_replicas if cross_replica is used.
    • num_replicas if cross_replica_and_partition is used.
    • num_processes if flattened_ids is used.
  • (C5) 0 <= replica_groups < size(replica_groups) .
  • (C6) If use_global_device_ids = true , then channel_id > 0 .
  • (C7) computation has type (tensor<E>, tensor<E>) -> (tensor<E>) where is_promotable(element_type(operand), E) .
  • (C8) shape(result) = shape(operand) except:
    • dim(result, scatter_dimension) = dim(operand, scatter_dimension) / dim(process_groups, 1) .
  • (C9) element_type(result) = E .

উদাহরণ

// num_replicas: 2
// num_partitions: 1
// %operand@(0, 0): [[1, 2, 3, 4],
//                   [5, 6, 7, 8]]
// %operand@(1, 0): [[9, 10, 11, 12],
//                   [13, 14, 15, 16]]
%result = "stablehlo.reduce_scatter"(%operand) ({
  ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>):
  %0 = "stablehlo.add"(%arg0, %arg1) : (tensor<i64>, tensor<i64>) -> tensor<i64>
  "stablehlo.return"(%0) : (tensor<i64>) -> ()
}) {
  scatter_dimension = 1 : i64,
  replica_groups = dense<[[0, 1]]> : tensor<1x2xi64>,
  channel_handle = #stablehlo.channel_handle<handle = 0, type = 0>
} : (tensor<2x4xi64>) -> tensor<2x2xi64>
//
// %result@(0, 0): [[10, 12],
//                  [18, 20]]
// %result@(1, 0): [[14, 16],
//                  [22, 24]]

আরো উদাহরণ

reduce_window

শব্দার্থবিদ্যা

Applies a reduction function body to windows of inputs and init_values and produces results .

The following diagram shows how elements in results... are computed from inputs... using a concrete example.

reduce_window

More formally, results...[result_index] = reduce(windows, init_values, axes(inputs...), body) (see reduce ) where:

  • padded_inputs = pad(inputs..., init_values..., padding[:, 0], padding[:, 1], base_dilations - 1) .
  • window_start = result_index * window_strides .
  • window_end = window_start + (window_dimensions - 1) * window_dilations + 1 .
  • windows = slice(padded_inputs..., window_start, window_end, window_dilations) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) inputs variadic number of tensors or per-tensor quantized tensors (C1-C4), (C6), (C8), (C10), (C12), (C13), (C15)
(I2) init_values variadic number of 0-dimensional tensors or per-tensor quantized tensors (C1), (C13)
(I3) window_dimensions 1-dimensional tensor constant of type si64 (C4), (C5), (C15)
(I4) window_strides 1-dimensional tensor constant of type si64 (C6), (C7), (C15)
(I5) base_dilations 1-dimensional tensor constant of type si64 (C8), (C9), (C15)
(I6) window_dilations 1-dimensional tensor constant of type si64 (C10), (C11), (C15)
(I7) padding 2-dimensional tensor constant of type si64 (C12), (C15)
(I8) body ফাংশন (C13)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
results variadic number of tensors or per-tensor quantized tensors (C1), (C14-C16)

সীমাবদ্ধতা

  • (C1) 0 < size(inputs) = size(init_values) = size(results) = N .
  • (C2) same(shape(inputs...)) .
  • (C3) element_type(inputs...) = element_type(init_values...) .
  • (C4) size(window_dimensions) = rank(inputs[0]) .
  • (C5) 0 < window_dimensions .
  • (C6) size(window_strides) = rank(inputs[0]) .
  • (C7) 0 < window_strides .
  • (C8) size(base_dilations) = rank(inputs[0]) .
  • (C9) 0 < base_dilations .
  • (C10) size(window_dilations) = rank(inputs[0]) .
  • (C11) 0 < window_dilations .
  • (C12) shape(padding) = [rank(inputs[0]), 2] .
  • (C13) body has type (tensor<E0>, ..., tensor<EN-1>, tensor<E0>, ..., tensor<EN-1>) -> (tensor<E0>, ..., tensor<EN-1>) where is_promotable(element_type(inputs[i]), Ei) .
  • (C14) same(shape(results...)) .
  • (C15) shape(results[0]) = num_windows where:
    • dilated_input_shape = shape(inputs[0]) = 0 ? 0 : (shape(inputs[0]) - 1) * base_dilations + 1 .
    • padded_input_shape = padding[:, 0] + dilated_input_shape + padding[:, 1] .
    • dilated_window_shape = (window_dimensions - 1) * window_dilations + 1 .
    • is_empty_window = padded_input_shape = 0 || dilated_window_shape > padded_input_shape .
    • num_windows = is_empty_window ? 0 : floor((padded_input_shape - dilated_window_shape) / window_strides) + 1 .
  • (C16) element_type(results[i]) = Ei for all i in [0,N) .

উদাহরণ

// %input = [[1, 2], [3, 4], [5, 6]]
// %init_value = 0
%result = "stablehlo.reduce_window"(%input, %init_value) ({
  ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>):
    %0 = "stablehlo.add"(%arg0, %arg1) : (tensor<i64>, tensor<i64>) -> tensor<i64>
    "stablehlo.return"(%0) : (tensor<i64>) -> ()
}) {
  window_dimensions = array<i64: 2, 1>,
  window_strides = array<i64: 4, 1>,
  base_dilations = array<i64: 2, 1>,
  window_dilations = array<i64: 3, 1>,
  padding = dense<[[2, 1], [0, 0]]> : tensor<2x2xi64>
} : (tensor<3x2xi64>, tensor<i64>) -> tensor<2x2xi64>
// %result = [[0, 0], [3, 4]]

আরো উদাহরণ

অবশিষ্ট

শব্দার্থবিদ্যা

Performs element-wise remainder of dividend lhs and divisor rhs tensors and produces a result tensor.

More formally, the sign of the result is taken from the dividend, and the absolute value of the result is always less than the divisor's absolute value. The remainder is calculated as lhs - d * rhs , where d is given by:

  • For integers: stablehlo.divide(lhs, rhs) .
  • For floats: division(lhs, rhs) from IEEE-754 with rounding attribute roundTowardZero .
  • For complex numbers: TBD ( #997 ).
  • For quantized types:
    • dequantize_op_quantize(remainder, lhs, rhs, type(result)) .

For floating-point element types, this operation is in contrast with the remainder operation from IEEE-754 specification where d is an integral value nearest to the exact value of lhs/rhs with ties to even.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor of integer, floating-point or complex type or per-tensor quantized tensor (C1)
(I2) rhs tensor of integer, floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of integer, floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %lhs: [17, -17, 17, -17]
// %rhs: [3, 3, -3, -3]
%result = "stablehlo.remainder"(%lhs, %rhs) : (tensor<4xi64>, tensor<4xi64>) -> tensor<4xi64>
// %result: [2, -2, 2, -2]

আরো উদাহরণ

প্রতিরূপ_আইডি

শব্দার্থবিদ্যা

Produces replica_id of the current process.

আউটপুট

নাম টাইপ
result 0-dimensional tensor of type ui32

উদাহরণ

%result = "stablehlo.replica_id"() : () -> tensor<ui32>

আরো উদাহরণ

নতুন আকার দেওয়া

শব্দার্থবিদ্যা

Performs reshape of operand tensor to a result tensor. Conceptually, it amounts to keeping the same canonical representation but potentially changing the shape, eg from tensor<2x3xf32> to tensor<3x2xf32> or tensor<6xf32> .

More formally, result[result_index] = operand[operand_index] where result_index and operand_index have the same position in the lexicographic ordering of index_space(result) and index_space(operand) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or quantized tensor (C1-C3)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or quantized tensor (C1-C3)

সীমাবদ্ধতা

  • (C1) element_type(result) is given by:
    • element_type(operand) , if !is_per_axis_quantized(operand) .
    • element_type(operand) except that quantization_dimension(operand) and quantization_dimension(result) may differ, otherwise.
  • (C2) size(operand) = size(result) .
  • (C3) If is_per_axis_quantized(operand) :
    • reduce(dims(operand, [0, 1, ..., quantization_dimension(operand) - 1]), init_values=1, dimensions=[0], body=lambda x, y: x * y) = reduce(dims(result, [0, 1, ..., quantization_dimension(result) - 1]), init_values=1, dimensions=[0], body=lambda x, y: x * y) .
    • dim(operand, quantization_dimension(operand)) = dim(result, quantization_dimension(result)) .
    • reduce(dims(operand, [quantization_dimension(operand) + 1, ..., rank(operand) - 1]), init_values=1, dimensions=[0], body=lambda x, y: x * y) = reduce(dims(result, [quantization_dimension(result) + 1, ..., rank(result) - 1]), init_values=1, dimensions=[0], body=lambda x, y: x * y) .

উদাহরণ

// %operand: [[1, 2, 3], [4, 5, 6]]
%result = "stablehlo.reshape"(%operand) : (tensor<2x3xi32>) -> tensor<3x2xi32>
// %result: [[1, 2], [3, 4], [5, 6]]

আরো উদাহরণ

বিপরীত

শব্দার্থবিদ্যা

Reverses the order of elements in the operand along the specified dimensions and produces a result tensor. More formally, result[result_index] = operand[operand_index] where:

  • operand_index[d] = dim(result, d) - result_index[d] - 1 if d in dimensions .
  • operand_index[d] = result_index[d] otherwise.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or per-tensor quantized tensor (C1), (C3)
(I2) dimensions 1-dimensional tensor constant of type si64 (C2), (C3)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C1), (C3)

সীমাবদ্ধতা

  • (C1) type(operand) = type(result) .
  • (C2) is_unique(dimensions) .
  • (C3) 0 <= dimensions < rank(result) .

উদাহরণ

// %operand = [[1, 2], [3, 4], [5, 6]]
%result = "stablehlo.reverse"(%operand) {
  dimensions = array<i64: 1>
} : (tensor<3x2xi32>) -> tensor<3x2xi32>
// %result: [[2, 1], [4, 3], [6, 5]]

আরো উদাহরণ

আরএনজি

শব্দার্থবিদ্যা

Generates random numbers using the rng_distribution algorithm and produces a result tensor of a given shape shape .

If rng_distribution = UNIFORM , then the random numbers are generated following the uniform distribution over the interval [a, b) . If a >= b , the behavior is undefined.

If rng_distribution = NORMAL , then the random numbers are generated following the normal distribution with mean = a and standard deviation = b . If b < 0 , the behavior is undefined.

The exact way how random numbers are generated is implementation-defined. For example, they may or may not be deterministic, and they may or may not use hidden state.

In conversations with many stakeholders, this op has come up as effectively deprecated, so in the future we are planning to explore removing it ( #597 ).

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) a 0-dimensional tensor of integer, boolean, or floating-point type (C1), (C2)
(I2) b 0-dimensional tensor of integer, boolean, or floating-point type (C1), (C2)
(I3) shape 1-dimensional tensor constant of type si64 (C3)
(I4) rng_distribution enum of UNIFORM and NORMAL (C2)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of integer, boolean, or floating-point type (C1-C3)

সীমাবদ্ধতা

  • (C1) element_type(a) = element_type(b) = element_type(result) .
  • (C2) If rng_distribution = NORMAL , then is_float(a) .
  • (C3) shape(result) = shape .

উদাহরণ

// %a = 0
// %b = 2
// %shape = [3, 3]
%result = "stablehlo.rng"(%a, %b, %shape) {
  rng_distribution = #stablehlo<rng_distribution UNIFORM>
} : (tensor<i32>, tensor<i32>, tensor<2xi64>) -> tensor<3x3xi32>
// %result: [
//           [1, 0, 1],
//           [1, 1, 1],
//           [0, 0, 0]
//          ]

rng_bit_generator

শব্দার্থবিদ্যা

Returns an output filled with uniform random bits and an updated output state output_state using the pseudorandom number generator algorithm rng_algorithm given an initial state initial_state . The output is guaranteed to be deterministic function of initial_state , but it is not guaranteed to be deterministic between implementations.

rng_algorithm is one of the following:

  • DEFAULT : Implementation-defined algorithm.
  • THREE_FRY : Implementation-defined variant of the Threefry algorithm.*
  • PHILOX : Implementation-defined variant of the Philox algorithm.*

* See: Salmon et al. SC 2011. Parallel random numbers: as easy as 1, 2, 3.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) rng_algorithm enum of DEFAULT , THREE_FRY , and PHILOX (C2)
(I2) initial_state 1-dimensional tensor of type ui64 (C1), (C2)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
output_state 1-dimensional tensor of type ui64 (C1)
output tensor of integer or floating-point type

সীমাবদ্ধতা

  • (C1) type(initial_state) = type(output_state) .
  • (C2) size(initial_state) is defined as:
    • implementation-defined if rng_algorithm = DEFAULT .
    • 2 if rng_algorithm = THREE_FRY .
    • 2 or 3 if rng_algorithm = PHILOX .

উদাহরণ

// %initial_state: [1, 2]
%output_state, %output = "stablehlo.rng_bit_generator"(%initial_state) {
  rng_algorithm = #stablehlo<rng_algorithm THREE_FRY>
} : (tensor<2xui64>) -> (tensor<2xui64>, tensor<2x2xui64>)
// %output_state: [1, 6]
// %output: [
//           [9236835810183407956, 16087790271692313299],
//           [18212823393184779219, 2658481902456610144]
//          ]

round_nearest_afz

শব্দার্থবিদ্যা

Performs element-wise rounding towards the nearest integer, breaking ties away from zero, on the operand tensor and produces a result tensor. Implements the roundToIntegralTiesToAway operation from the IEEE-754 specification. For quantized types, performs dequantize_op_quantize(round_nearest_afz, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand = [-2.5, 0.4, 0.5, 0.6, 2.5]
%result = "stablehlo.round_nearest_afz"(%operand) : (tensor<5xf64>) -> tensor<5xf64>
// %result: [-3.0, 0.0, 1.0, 1.0, 3.0]

আরো উদাহরণ

round_nearest_even

শব্দার্থবিদ্যা

Performs element-wise rounding towards the nearest integer, breaking ties towards the even integer, on the operand tensor and produces a result tensor. Implements the roundToIntegralTiesToEven operation from the IEEE-754 specification. For quantized types, performs dequantize_op_quantize(round_nearest_even, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand = [-2.5, 0.4, 0.5, 0.6, 2.5]
%result = "stablehlo.round_nearest_even"(%operand) : (tensor<5xf64>) -> tensor<5xf64>
// %result: [-2.0, 0.0, 0.0, 1.0, 2.0]

আরো উদাহরণ

rsqrt

শব্দার্থবিদ্যা

Performs element-wise reciprocal square root operation on operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For floats: rSqrt from IEEE-754.
  • For complex numbers: complex reciprocal square root.
  • For quantized types: dequantize_op_quantize(rsqrt, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [[1.0, 4.0], [9.0, 25.0]]
%result = "stablehlo.rsqrt"(%operand) : (tensor<2x2xf32>) -> tensor<2x2xf32>
// %result: [[1.0, 0.5], [0.33333343, 0.2]]

আরো উদাহরণ

বিক্ষিপ্ত

শব্দার্থবিদ্যা

Produces results tensors which are equal to inputs tensors except that several slices specified by scatter_indices are updated with the values updates using update_computation .

The following diagram shows how elements in updates... map on elements in results... using a concrete example. The diagram picks a few example updates... indices and explains in detail which results... indices they correspond to.

বিক্ষিপ্ত

More formally, for all update_index in index_space(updates[0]) :

  • update_scatter_dims = [d for d in axes(updates[0]) and d not in update_window_dims] .
  • update_scatter_index = update_index[update_scatter_dims...] .
  • start_index is defined as:
    • scatter_indices[si0, ..., :, ..., siN] where si are individual elements in update_scatter_index and : is inserted at the index_vector_dim index, if index_vector_dim < rank(scatter_indices) .
    • [scatter_indices[update_scatter_index]] otherwise.
  • For d_input in axes(inputs[0]) ,
    • full_start_index[d_input] = start_index[d_start] if d_input = scatter_dims_to_operand_dims[d_start] .
    • full_start_index[d_input] = 0 otherwise.
  • For d_input in axes(inputs[0]) ,
    • full_batching_index[d_input] = update_scatter_index[d_start - (d_start < index_vector_dim ? 0 : 1)] if d_input = input_batching_dims[i_batching] and d_start = scatter_indices_batching_dims[i_batching] .
    • full_batching_index[d_input] = 0 otherwise.
  • update_window_index = update_index[update_window_dims...] .
  • full_window_index = [wi0, ..., 0, ..., wiN] where wi are individual elements in update_window_index , and 0 is inserted at indices from inserted_window_dims and input_batching_dims .
  • result_index = full_start_index + full_batching_index + full_window_index .

Given that, results = exec(schedule, inputs) , where:

  • schedule is an implementation-defined permutation of index_space(updates[0]) .
  • exec([update_index, ...], results) = exec([...], updated_results) where:
    • If result_index is in bounds for shape(results...)
    • updates_converted = to_destination_type( updates...[update_index], type(func_inputs(update_computation) [len(func_inputs(update_computation))//2:])... )
    • updated_values = update_computation(results...[result_index], updates_converted)
    • updated_results is a copy of results with results...[result_index] set to updated_values... .
    • অন্যথায়
    • updated_results = results .
  • exec([], results) = results .

If indices_are_sorted is true then the implementation can assume that scatter_indices are sorted with respect to scatter_dims_to_operand_dims , otherwise the behavior is undefined. More formally, for all i1 < i2 from indices(result) , full_start_index(i1) <= full_start_index(i2) .

If unique_indices is true then the implementation can assume that all result_index indices being scattered to are unique. If unique_indices is true but the indices being scattered to are not unique then the behavior is undefined.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) inputs variadic number of tensors or per-tensor quantized tensors (C1), (C2), (C4-C6), (C11), (C13), (C18), (C21), (C23-C24)
(I2) scatter_indices tensor of integer type (C4), (C15), (C19), (C22)
(I3) updates variadic number of tensors or per-tensor quantized tensors (C3-C6), (C8)
(I4) update_window_dims 1-dimensional tensor constant of type si64 (C2), (C4), (C7-C8)
(I5) inserted_window_dims 1-dimensional tensor constant of type si64 (C2), (C4), (C9-C11)
(I6) input_batching_dims 1-dimensional tensor constant of type si64 (C2), (C4), (C9), (C12-13), (C17-18), (C20)
(I7) scatter_indices_batching_dims 1-dimensional tensor constant of type si64 (C14-C18)
(I8) scatter_dims_to_operand_dims 1-dimensional tensor constant of type si64 (C19-C21)
(I9) index_vector_dim constant of type si64 (C4), (C16), (C19), (C22)
(I10) indices_are_sorted constant of type i1
(I11) unique_indices constant of type i1
(I12) update_computation ফাংশন (C23)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
results variadic number of tensors or per-tensor quantized tensors (C24-C25)

সীমাবদ্ধতা

  • (C1) same(shape(inputs...)) .
  • (C2) `rank(inputs[0]) = size(update_window_dims) + size(inserted_window_dims)
    • size(input_batching_dims)`.
  • (C3) same(shape(updates...)) .
  • (C4) shape(updates[0]) = combine(update_scatter_dim_sizes, update_window_dim_sizes) where:
    • update_scatter_dim_sizes = shape(scatter_indices) except that the dimension size of scatter_indices corresponding to index_vector_dim is not included.
    • update_window_dim_sizes <= shape(inputs[0]) except that the dimension sizes in inputs[0] corresponding to inserted_window_dims and input_batching_dims are not included.
    • combine puts update_scatter_dim_sizes at axes corresponding to update_scatter_dims and update_window_dim_sizes at axes corresponding to update_window_dims .
  • (C5) 0 < size(inputs) = size(updates) = N .
  • (C6) element_type(updates...) = element_type(inputs...) .
  • (C7) is_unique(update_window_dims) and is_sorted(update_window_dims) .
  • (C8) 0 <= update_window_dims < rank(updates[0]) .
  • (C9) is_unique(concatenate(inserted_window_dims, input_batching_dims))
  • (C10) is_sorted(inserted_window_dims) .
  • (C11) 0 <= inserted_window_dims < rank(inputs[0]) .
  • (C12) is_sorted(input_batching_dims) .
  • (C13) 0 <= input_batching_dims < rank(inputs[0])) .
  • (C14) is_unique(scatter_indices_batching_dims) .
  • (C15) 0 <= scatter_indices_batching_dims < rank(scatter_indices) .
  • (C16) index_vector_dim not in scatter_indices_batching_dims .
  • (C17) size(input_batching_dims) == size(scatter_indices_batching_dims) .
  • (C18) dim(inputs[0], input_batching_dims...) = dim(scatter_indices, scatter_indices_batching_dims...) .
  • (C19) size(scatter_dims_to_operand_dims) = index_vector_dim < rank(scatter_indices) ? dim(scatter_indices, index_vector_dim) : 1 .
  • (C20) is_unique(concatenate(scatter_dims_to_operand_dims, input_batching_dims)) .
  • (C21) 0 <= scatter_dims_to_operand_dims < rank(inputs[0]) .
  • (C22) 0 <= index_vector_dim <= rank(scatter_indices) .
  • (C23) update_computation has type (tensor<E0>, ..., tensor<EN-1>, tensor<E0>, ..., tensor<EN-1>) -> (tensor<E0>, ..., tensor<EN-1>) , where is_promotable(element_type(inputs[i]), Ei) .
  • (C24) shape(inputs...) = shape(results...) .
  • (C25) element_type(results[i]) = Ei for all i in [0,N) .

উদাহরণ

// %input: [
//          [
//           [[1, 2], [3, 4], [5, 6], [7, 8]],
//           [[9, 10],[11, 12], [13, 14], [15, 16]],
//           [[17, 18], [19, 20], [21, 22], [23, 24]]
//          ],
//          [
//           [[25, 26], [27, 28], [29, 30], [31, 32]],
//           [[33, 34], [35, 36], [37, 38], [39, 40]],
//           [[41, 42], [43, 44], [45, 46], [47, 48]]
//          ]
//         ]
// %scatter_indices: [
//                    [
//                     [[0, 0], [1, 0], [2, 1]],
//                     [[0, 1], [1, 1], [0, 9]]
//                    ],
//                    [
//                     [[0, 0], [2, 1], [2, 2]],
//                     [[1, 2], [0, 1], [1, 0]]
//                    ]
//                   ]
// %update: [
//           [
//            [[1, 1], [1, 1], [1, 1]],
//            [[1, 1], [1, 1], [1, 1]]
//           ],
//           [
//            [[1, 1], [1, 1], [1, 1]],
//            [[1, 1], [1, 1], [1, 1]]
//           ]
//          ]
%result = "stablehlo.scatter"(%input, %scatter_indices, %update) ({
  ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>):
    %0 = "stablehlo.add"(%arg0, %arg1) : (tensor<i64>, tensor<i64>) -> tensor<i64>
    "stablehlo.return"(%0) : (tensor<i64>) -> ()
}) {
  scatter_dimension_numbers = #stablehlo.scatter<
    update_window_dims = [3, 4],
    inserted_window_dims = [1],
    input_batching_dims = [0],
    scatter_indices_batching_dims = [1],
    scatter_dims_to_operand_dims = [2, 1],
    index_vector_dim = 3>,
  indices_are_sorted = false,
  unique_indices = false
} : (tensor<2x3x4x2xi64>, tensor<2x2x3x2xi64>, tensor<2x2x3x2x2xi64>) -> tensor<2x3x4x2xi64>
// %result: [
//           [
//            [[3, 4], [6, 7], [6, 7], [7, 8]],
//            [[9, 10],[11, 12], [15, 16], [17, 18]],
//            [[17, 18], [19, 20], [22, 23], [24, 25]]
//           ],
//           [
//            [[25, 26], [28, 29], [30, 31], [31, 32]],
//            [[35, 36], [38, 39], [38, 39], [39, 40]],
//            [[41, 42], [44, 45], [46, 47], [47, 48]]
//           ]
//          ]

আরো উদাহরণ

নির্বাচন করুন

শব্দার্থবিদ্যা

Produces a result tensor where each element is selected from on_true or on_false tensor based on the value of the corresponding element of pred . More formally, result[result_index] = pred_element ? on_true[result_index] : on_false[result_index] , where pred_element = rank(pred) = 0 ? pred[] : pred[result_index] . For quantized types, performs dequantize_select_quantize(pred, on_true, on_false, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) pred tensor of type i1 (C1)
(I2) on_true tensor or per-tensor quantized tensor (C1-C2)
(I3) on_false tensor or per-tensor quantized tensor (C2)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C2)

সীমাবদ্ধতা

  • (C1) rank(pred) = 0 or shape(pred) = shape(on_true) .
  • (C2) baseline_type(on_true) = baseline_type(on_false) = baseline_type(result) .

উদাহরণ

// %pred: [[false, true], [true, false]]
// %on_true: [[1, 2], [3, 4]]
// %on_false: [[5, 6], [7, 8]]
%result = "stablehlo.select"(%pred, %on_true, %on_false) : (tensor<2x2xi1>, tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32>
// %result: [[5, 2], [3, 8]]

আরো উদাহরণ

select_and_scatter

শব্দার্থবিদ্যা

Scatters the values from the source tensor using scatter based on the outcome of reduce_window of the input tensor using select and produces a result tensor.

The following diagram shows how elements in result are computed from operand and source using a concrete example.

select_and_scatter

আরো আনুষ্ঠানিকভাবে:

  • selected_values = reduce_window_without_init(...) with the following inputs:

    • inputs = [operand].
    • window_dimensions , window_strides , and padding which are used as is.
    • base_dilations = windows_dilations = 1 .
    • body is defined as:
    def body(arg0: tensor<E>, arg1: tensor<E>) -> tensor<E>:
      return select(arg0, arg1) ? arg0 : arg1;
    

    where E = element_type(operand) , and reduce_window_without_init works exactly like reduce_window , except that the schedule of the underlying reduce (see reduce ) doesn't include init values. It is currently unspecified what happens if the corresponding window doesn't have values ( #731 ).

  • result[result_index] = reduce([source_values], [init_value], [0], scatter) where:

    • source_values = [source[source_index] for source_index in source_indices] .
    • selected_index(source_index) = operand_index if selected_values[source_index] has the operand element from operand_index .
    • source_indices = [source_index for source_index in indices(source) if selected_index(source_index) = result_index] .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or per-tensor quantized tensor (C1-C4), (C6), (C8-C11)
(I2) source tensor or per-tensor quantized tensor (C1), (C2)
(I3) init_value 0-dimensional tensor or per-tensor quantized tensor (C3)
(I4) window_dimensions 1-dimensional tensor constant of type si64 (C2), (C4), (C5)
(I5) window_strides 1-dimensional tensor constant of type si64 (C2), (C6), (C7)
(I6) padding 2-dimensional tensor constant of type si64 (C2), (C8)
(I7) select ফাংশন (C9)
(I8) scatter ফাংশন (C10)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C11-C12)

সীমাবদ্ধতা

  • (C1) element_type(operand) = element_type(source) .
  • (C2) shape(source) = num_windows where:
    • padded_operand_shape = padding[:, 0] + shape(operand) + padding[:, 1] .
    • is_empty_window = padded_operand_shape = 0 || window_dimensions > padded_operand_shape .
    • num_windows = is_empty_window ? 0 : floor((padded_operand_shape - window_dimensions) / window_strides) + 1 .
  • (C3) element_type(init_value) = element_type(operand) .
  • (C4) size(window_dimensions) = rank(operand) .
  • (C5) 0 < window_dimensions .
  • (C6) size(window_strides) = rank(operand) .
  • (C7) 0 < window_strides .
  • (C8) shape(padding) = [rank(operand), 2] .
  • (C9) select has type (tensor<E>, tensor<E>) -> tensor<i1> where E = element_type(operand) .
  • (C10) scatter has type (tensor<E>, tensor<E>) -> tensor<E> where is_promotable(element_type(operand), E) .
  • (C11) shape(operand) = shape(result) .
  • (C12) element_type(result) = E .

উদাহরণ

// %operand: [[1, 5], [2, 5], [3, 6], [4, 4]]
// %source: [[5, 6], [7, 8]]
// %init_value: 0
%result = "stablehlo.select_and_scatter"(%operand, %source, %init_value) ({
  ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>):
    %0 = "stablehlo.compare"(%arg0, %arg1) {
      comparison_direction = #stablehlo<comparison_direction GE>
    } : (tensor<i64>, tensor<i64>) -> tensor<i1>
    "stablehlo.return"(%0) : (tensor<i1>) -> ()
}, {
  ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>):
    %0 = "stablehlo.add"(%arg0, %arg1) : (tensor<i64>, tensor<i64>) -> tensor<i64>
    "stablehlo.return"(%0) : (tensor<i64>) -> ()
}) {
  window_dimensions = array<i64: 3, 1>,
  window_strides = array<i64: 2, 1>,
  padding = dense<[[0, 1], [0, 0]]> : tensor<2x2xi64>
} : (tensor<4x2xi64>, tensor<2x2xi64>, tensor<i64>) -> tensor<4x2xi64>
// %result: [[0, 0], [0, 0], [5, 14], [7, 0]]

আরো উদাহরণ

পাঠান

শব্দার্থবিদ্যা

Sends inputs to a channel channel_id and produces a result token.

If is_host_transfer is true , then the operation transfers data to the host. Otherwise, it transfers data to another device. What this means is implementation-defined. This flag duplicates the information provided in channel_type , so in the future we are planning to only keep one of them ( #666 ).

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) inputs variadic number of tensors or quantized tensors
(I2) token token
(I3) channel_id constant of type si64
(I4) channel_type enum of DEVICE_TO_DEVICE and DEVICE_TO_HOST (C1)
(I5) is_host_transfer constant of type i1 (C1)

আউটপুট

নাম টাইপ
result token

সীমাবদ্ধতা

  • (C1) channel_type is defined as:
    • DEVICE_TO_HOST if is_host_transfer = true ,
    • DEVICE_TO_DEVICE otherwise.

উদাহরণ

%result = "stablehlo.send"(%operand, %token) {
  channel_handle = #stablehlo.channel_handle<handle = 1, type = 2>,
  is_host_transfer = true
} : (tensor<2x2xi64>, !stablehlo.token) -> !stablehlo.token

আরো উদাহরণ

shift_left

শব্দার্থবিদ্যা

Performs element-wise left-shift operation on the lhs tensor by rhs number of bits and produces a result tensor.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor of integer type (C1)
(I2) rhs tensor of integer type (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of integer type (C1)

সীমাবদ্ধতা

  • (C1) type(lhs) = type(rhs) = type(result) .

উদাহরণ

// %lhs: [-1, 0, 1]
// %rhs: [1, 2, 3]
%result = "stablehlo.shift_left"(%lhs, %rhs): (tensor<3xi64>, tensor<3xi64>) -> tensor<3xi64>
// %result: [-2, 0, 8]

আরো উদাহরণ

shift_right_arithmetic

শব্দার্থবিদ্যা

Performs element-wise arithmetic right-shift operation on the lhs tensor by rhs number of bits and produces a result tensor.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor of integer type (C1)
(I2) rhs tensor of integer type (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of integer type (C1)

সীমাবদ্ধতা

  • (C1) type(lhs) = type(rhs) = type(result) .

উদাহরণ

// %lhs: [-1, 0, 8]
// %rhs: [1, 2, 3]
%result = "stablehlo.shift_right_arithmetic"(%lhs, %rhs): (tensor<3xi64>, tensor<3xi64>) -> tensor<3xi64>
// %result: [-1, 0, 1]

আরো উদাহরণ

shift_right_logical

শব্দার্থবিদ্যা

Performs element-wise logical right-shift operation on the lhs tensor by rhs number of bits and produces a result tensor.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor of integer type (C1)
(I2) rhs tensor of integer type (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of integer type (C1)

সীমাবদ্ধতা

  • (C1) type(lhs) = type(rhs) = type(result) .

উদাহরণ

// %lhs: [-1, 0, 8]
// %rhs: [1, 2, 3]
%result = "stablehlo.shift_right_logical"(%lhs, %rhs): (tensor<3xi64>, tensor<3xi64>) -> tensor<3xi64>
// %result: [9223372036854775807, 0, 1]

আরো উদাহরণ

চিহ্ন

শব্দার্থবিদ্যা

Returns the sign of the operand element-wise and produces a result tensor. More formally, for each element x , the semantics can be expressed using Python syntax as follows:

def sign(x):
  if is_integer(x):
    if compare(x, 0, LT, SIGNED): return -1
    if compare(x, 0, EQ, SIGNED): return 0
    return 1
  elif is_float(x):
    if is_nan(x): return NaN
    if compare(x, -0.0, EQ, FLOAT): return -0.0
    if compare(x, +0.0, EQ, FLOAT): return +0.0
    if compare(x, 0.0, LT, FLOAT): return -1.0
    return 1.0
  elif is_complex(x):
    if is_nan(real(x)) or is_nan(imag(x)): return (NaN, NaN)
    if compare(x, (0.0, 0.0), EQ, FLOAT): return (0.0, 0.0)
    return divide(x, convert(abs(x), type(x)))

For quantized types, performs dequantize_op_quantize(sign, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of signed integer, floating-point, or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of signed integer, floating-point, or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// Logical values: +NaN, -1.0, -0.0, +0.0, 1.0
// operand: [0x7FFFFFFFFFFFFFFF, -1.0, -0.0, 0.0, 1.0]
%result = "stablehlo.sign"(%operand) : (tensor<5xf64>) -> tensor<5xf64>
// Logical values: +NaN, -1.0, -0.0, +0.0, 1.0
// %result: [0x7FFFFFFFFFFFFFFF, -1.0, -0.0, 0.0, 1.0]

আরো উদাহরণ

সাইন

শব্দার্থবিদ্যা

Performs element-wise sine operation on operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For floats: sin from IEEE-754.
  • For complex numbers: complex sine.
  • For quantized types: dequantize_op_quantize(sine, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [
//            [0.0, 1.57079632],       // [0, pi/2]
//            [3.14159265, 4.71238898] // [pi, 3pi/2]
//           ]
%result = "stablehlo.sine"(%operand) : (tensor<2x2xf32>) -> tensor<2x2xf32>
// %result: [[0.0, 1.0], [0.0, -1.0]]

আরো উদাহরণ

টুকরা

শব্দার্থবিদ্যা

Extracts a slice from the operand using statically-computed starting indices and produces a result tensor. start_indices contain the starting indices of the slice for each dimension, limit_indices contain the ending indices (exclusive) for the slice for each dimension, and strides contain the strides for each dimension.

More formally, result[result_index] = operand[operand_index] where operand_index = start_indices + result_index * strides .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or per-tensor quantized tensor (C1-C3), (C5)
(I2) start_indices 1-dimensional tensor constant of type si64 (C2), (C3), (C5)
(I3) limit_indices 1-dimensional tensor constant of type si64 (C2), (C3), (C5)
(I4) strides 1-dimensional tensor constant of type si64 (C2), (C4)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C1), (C5)

সীমাবদ্ধতা

  • (C1) element_type(operand) = element_type(result) .
  • (C2) size(start_indices) = size(limit_indices) = size(strides) = rank(operand) .
  • (C3) 0 <= start_indices <= limit_indices <= shape(operand) .
  • (C4) 0 < strides .
  • (C5) shape(result) = ceil((limit_indices - start_indices) / strides) .

উদাহরণ

// %operand: [
//            [0, 0, 0, 0],
//            [0, 0, 1, 1],
//            [0, 0, 1, 1]
//           ]
%result = "stablehlo.slice"(%operand) {
  start_indices = array<i64: 1, 2>,
  limit_indices = array<i64: 3, 4>,
  strides = array<i64: 1, 1>
} : (tensor<3x4xi64>) -> tensor<2x2xi64>
// % result: [
//            [1, 1],
//            [1, 1]
//           ]

আরো উদাহরণ

সাজান

শব্দার্থবিদ্যা

Sorts 1-dimensional slices of inputs along the dimension dimension together, according to a comparator and produces results .

Unlike similar inputs in other operations, dimension allows negative values, with the semantics described below. In the future, this may be disallowed for consistency reasons ( #1377 ).

If is_stable is true, then the sorting is stable, that is, relative order of elements considered to be equal by the comparator is preserved. For the case where there is a single input, two elements e1 and e2 are considered to be equal by the comparator if and only if comparator(e1, e2) = comparator(e2, e1) = false . See the formalization below for how this generalizes to multiple inputs.

More formally, for all result_index in index_space(results[0]) :

  • adjusted_dimension = dimension >= 0 ? dimension : rank(inputs[0]) + dimension .
  • result_slice = [ri0, ..., :, ..., riR-1] where riN are individual elements in result_index , and : is inserted at adjusted_dimension .
  • inputs_together = (inputs[0]..., ..., inputs[N-1]...) .
  • results_together[result_slice] = sort(inputs_together[result_slice], comparator_together) .
  • where sort sorts a 1-dimensional slice in non-descending order expecting that comparator_together returns true if the left-hand side argument is less than the right-hand second argument.
  • def comparator_together(lhs_together, rhs_together):
      args = []
      for (lhs_el, rhs_el) in zip(lhs_together, rhs_together):
        args.append(lhs_el)
        args.append(rhs_el)
      return comparator(*args)
    
  • (results[0]..., ..., results[N-1]...) = results_together .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) inputs variadic number of tensors or per-tensor quantized tensors (C1-C5)
(I2) dimension constant of type si64 (C4)
(I3) is_stable constant of type i1
(I4) comparator ফাংশন (C5)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
results variadic number of tensors or per-tensor quantized tensors (C2), (C3)

সীমাবদ্ধতা

  • (C1) 0 < size(inputs) .
  • (C2) type(inputs...) = type(results...) .
  • (C3) same(shape(inputs...) + shape(results...)) .
  • (C4) -R <= dimension < R , where R = rank(inputs[0]) .
  • (C5) comparator has type (tensor<E1>, tensor<E1>, ..., tensor<EN-1>, tensor<EN-1>) -> tensor<i1> , where Ei = element_type(inputs[i])

উদাহরণ

// %input0 = [[1, 2, 3], [3, 2, 1]]
// %input1 = [[3, 2, 1], [1, 2, 3]]
%result0, %result1 = "stablehlo.sort"(%input0, %input1) ({
  ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>, %arg2: tensor<i64>, %arg3: tensor<i64>):
    %predicate = "stablehlo.compare"(%arg0, %arg1) {
      comparison_direction = #stablehlo<comparison_direction GT>
    } : (tensor<i64>, tensor<i64>) -> tensor<i1>
    "stablehlo.return"(%predicate) : (tensor<i1>) -> ()
}) {
  dimension = 0 : i64,
  is_stable = true
} : (tensor<2x3xi64>, tensor<2x3xi64>) -> (tensor<2x3xi64>, tensor<2x3xi64>)
// %result0 = [[3, 2, 3], [1, 2, 1]]
// %result1 = [[1, 2, 1], [3, 2, 3]]

আরো উদাহরণ

sqrt

শব্দার্থবিদ্যা

Performs element-wise square root operation on operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For floats: squareRoot from IEEE-754.
  • For complex numbers: complex square root.
  • For quantized types: dequantize_op_quantize(sqrt, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [[0.0, 1.0], [4.0, 9.0]]
%result = "stablehlo.sqrt"(%operand) : (tensor<2x2xf32>) -> tensor<2x2xf32>
// %result: [[0.0, 1.0], [2.0, 3.0]]

আরো উদাহরণ

বিয়োগ

শব্দার্থবিদ্যা

Performs element-wise subtraction of two tensors lhs and rhs and produces a result tensor. Depending on the element type, does the following:

  • For integers: integer subtraction.
  • For floats: subtraction from IEEE-754.
  • For complex numbers: complex subtraction.
  • For quantized types:
    • dequantize_op_quantize(subtract, lhs, rhs, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor of integer, floating-point, or complex type or per-tensor quantized tensor (C1)
(I2) rhs tensor of integer, floating-point, or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of integer, floating-point, or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(lhs) = baseline_type(rhs) = baseline_type(result) .

উদাহরণ

// %lhs: [[6, 8], [10, 12]]
// %rhs: [[5, 6], [7, 8]]
%result = "stablehlo.subtract"(%lhs, %rhs) : (tensor<2x2xf32>, tensor<2x2xf32>) -> (tensor<2x2xf32>)
// %result: [[1, 2], [3, 4]]

আরো উদাহরণ

ট্যান

শব্দার্থবিদ্যা

Performs element-wise tangent operation on the operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For floats: tan from IEEE-754.
  • For complex numbers: complex tangent.
  • For quantized types: dequantize_op_quantize(tan, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [
//            [0.0, 1.57079632],       // [0, pi/2]
//            [3.14159265, 4.71238898] // [pi, 3pi/2]
//           ]
%result = "stablehlo.tan"(%operand) : (tensor<2x2xf64>) -> tensor<2x2xf64>
// %result: [
//           [0.0, 1.63312e+16],
//           [0.0, 5.44375e+15]
//          ]

আরো উদাহরণ

তানহ

শব্দার্থবিদ্যা

Performs element-wise hyperbolic tangent operation on operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For floats: tanh from IEEE-754.
  • For complex numbers: complex hyperbolic tangent.
  • For quantized types:
    • dequantize_op_quantize(tanh, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [-1.0, 0.0, 1.0]
%result = "stablehlo.tanh"(%operand) : (tensor<3xf32>) -> tensor<3xf32>
// %result: [-0.76159416, 0.0, 0.76159416]

আরো উদাহরণ

স্থানান্তর

শব্দার্থবিদ্যা

Permutes the dimensions of operand tensor using permutation and produces a result tensor. More formally, result[result_index] = operand[operand_index] where result_index[d] = operand_index[permutation[d]] .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or quantized tensor (C1-C4)
(I2) permutation 1-dimensional tensor constant of type si64 (C2-C4)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or quantized tensor (C1), (C3-C4)

সীমাবদ্ধতা

  • (C1) element_type(result) is given by:
    • element_type(operand) , if !is_per_axis_quantized(operand) .
    • element_type(operand) except that quantization_dimension(operand) and quantization_dimension(result) may differ, otherwise.
  • (C2) permutation is a permutation of range(rank(operand)) .
  • (C3) shape(result) = dim(operand, permutation...) .
  • (C4) If is_per_axis_quantized(result) , then quantization_dimension(operand) = permutation(quantization_dimension(result)) .

উদাহরণ

// %operand: [
//            [[1,2], [3,4], [5,6]],
//            [[7,8], [9,10], [11,12]]
//           ]
%result = "stablehlo.transpose"(%operand) {
  permutation = array<i64: 2, 1, 0>
} : (tensor<2x3x2xi32>) -> tensor<2x3x2xi32>
// %result: [
//           [[1,7], [3,9], [5,11]],
//           [[2,8], [4,10], [6,12]]
//          ]

আরো উদাহরণ

triangular_solve

শব্দার্থবিদ্যা

Solves batches of systems of linear equations with lower or upper triangular coefficient matrices.

More formally, given a and b , result[i0, ..., iR-3, :, :] is the solution to op(a[i0, ..., iR-3, :, :]) * x = b[i0, ..., iR-3, :, :] when left_side is true or x * op(a[i0, ..., iR-3, :, :]) = b[i0, ..., iR-3, :, :] when left_side is false , solving for the variable x where op(a) is determined by transpose_a , which can be one of the following:

  • NO_TRANSPOSE : Perform operation using a as-is.
  • TRANSPOSE : Perform operation on transpose of a .
  • ADJOINT : Perform operation on conjugate transpose of a .

Input data is read only from the lower triangle of a , if lower is true or upper triangle of a , otherwise. Output data is returned in the same triangle; the values in the other triangle are implementation-defined.

If unit_diagonal is true, then the implementation can assume that the diagonal elements of a are equal to 1, otherwise the behavior is undefined.

For quantized types, performs dequantize_op_quantize(lambda x, y: triangular_solve(x, y, left_side, lower, unit_diagonal, transpose_a), a, b, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) a tensor of floating-point or complex type or per-tensor quantized tensor (C1-C3)
(I2) b tensor of floating-point or complex type or per-tensor quantized tensor (C1-C4)
(I3) left_side constant of type i1 (C3)
(I4) lower constant of type i1
(I5) unit_diagonal constant of type i1
(I6) transpose_a enum of NO_TRANSPOSE , TRANSPOSE , and ADJOINT

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_element_type(a) = baseline_element_type(b) .
  • (C2) 2 <= rank(a) = rank(b) = R .
  • (C3) The relationship between shape(a) and shape(b) is defined as follows:
    • shape(a)[:-3] = shape(b)[:-3] .
    • dim(a, -2) = dim(a, -1) = dim(b, left_side ? -2 : -1) .
  • (C4) baseline_type(b) = baseline_type(result) .

উদাহরণ

// %a = [
//       [1.0, 0.0, 0.0],
//       [2.0, 4.0, 0.0],
//       [3.0, 5.0, 6.0]
//      ]
// %b = [
//       [2.0, 0.0, 0.0],
//       [4.0, 8.0, 0.0],
//       [6.0, 10.0, 12.0]
//      ]
%result = "stablehlo.triangular_solve"(%a, %b) {
  left_side = true,
  lower = true,
  unit_diagonal = false,
  transpose_a = #stablehlo<transpose NO_TRANSPOSE>
} : (tensor<3x3xf32>, tensor<3x3xf32>) -> tensor<3x3xf32>
// %result: [
//           [2.0, 0.0, 0.0],
//           [0.0, 2.0, 0.0],
//           [0.0, 0.0, 2.0]
//          ]

টিপল

শব্দার্থবিদ্যা

Produces a result tuple from values val .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) val variadic number of values (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result টিপল (C1)

সীমাবদ্ধতা

  • (C1) result has type tuple<E0, ..., EN-1> where Ei = type(val[i]) .

উদাহরণ

// %val0: [1.0, 2.0]
// %val1: (3)
%result = "stablehlo.tuple"(%val0, %val1) : (tensor<2xf32>, tuple<tensor<i32>>) -> tuple<tensor<2xf32>, tuple<tensor<i32>>>
// %result: ([1.0, 2.0], (3))

আরো উদাহরণ

uniform_dequantize

শব্দার্থবিদ্যা

Performs element-wise conversion of quantized tensor operand to a floating-point tensor result according to the quantization parameters defined by the operand type.

More formally, result = dequantize(operand) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand quantized tensor (C1), (C2)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point type (C1), (C2)

সীমাবদ্ধতা

  • (C1) shape(operand) = shape(result) .
  • (C2) element_type(result) = expressed_type(operand) .

উদাহরণ

// %operand: [10, 10]
%result = "stablehlo.uniform_dequantize"(%operand) : (tensor<2x!quant.uniform<i8:f32:0, {0.1:-30,0.5:-20}>>) -> tensor<2xf32>
// %result: [4.0, 15.0]

uniform_quantize

শব্দার্থবিদ্যা

Performs element-wise conversion of floating-point tensor or quantized tensor operand to a quantized tensor result according to the quantization parameters defined by the result type.

আরো আনুষ্ঠানিকভাবে,

  • If is_float(operand) :
    • result = quantize(operand, type(result)) .
  • If is_quantized(operand) :
    • float_result = dequantize(operand) .
    • result = quantize(float_result, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or quantized type (C1), (C2)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result quantized tensor (C1), (C2)

সীমাবদ্ধতা

  • (C1) shape(operand) = shape(result) .
  • (C2) expressed_type(result) = is_float(operand) ? element_type(operand) : expressed_type(operand) .

উদাহরণ

// %operand: [4.0, 15.0]
%result = "stablehlo.uniform_quantize"(%operand) : (tensor<2xf32>) -> tensor<2x!quant.uniform<i8:f32:0, {0.1:-30,0.5:-20}>>
// %result: [10, 10]

// %operand: [10, 10]
%result = "stablehlo.uniform_quantize"(%operand) : (tensor<2x!quant.uniform<i8:f32:0, {0.1:-30,0.5:-20}>>) -> tensor<2x!quant.uniform<i8:f32:0, {0.1:-20,0.2:-30}>>
// %result: [20, 45]

যখন

শব্দার্থবিদ্যা

Produces the output from executing body function 0 or more times while the cond function outputs true . More formally, the semantics can be expressed using Python syntax as follows:

internal_state = operand
while cond(*internal_state):
  internal_state = body(*internal_state)
results = internal_state

The behavior of an infinite loop is TBD ( #383 ).

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand variadic number of tensors, quantized tensors or tokens (C1-C3)
(I2) cond ফাংশন (C1)
(I3) body ফাংশন (C2)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
results variadic number of tensors, quantized tensors or tokens (C3)

সীমাবদ্ধতা

  • (C1) cond has type (T0, ..., TN-1) -> tensor<i1> , where Ti = type(operand[i]) .
  • (C2) body has type (T0, ..., TN-1) -> (T0, ..., TN-1) , where Ti = type(operand[i]) .
  • (C3) type(results...) = type(operand...) .

উদাহরণ

// %init_i: 1
// %init_sum: 0
// %one: 1
// %ten: 10
%results0, %results1 = "stablehlo.while"(%init_i, %init_sum) ({
  ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>):
    %cond = "stablehlo.compare"(%arg0, %ten) {
      comparison_direction = #stablehlo<comparison_direction LT>
    } : (tensor<i64>, tensor<i64>) -> tensor<i1>
    stablehlo.return %cond : tensor<i1>
  }, {
  ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>):
    %new_sum = stablehlo.add %arg1, %one : tensor<i64>
    %new_i = stablehlo.add %arg0, %one : tensor<i64>
    stablehlo.return %new_i, %new_sum : tensor<i64>, tensor<i64>
}) : (tensor<i64>, tensor<i64>) -> (tensor<i64>, tensor<i64>)
// %results0: 10
// %results1: 10

আরো উদাহরণ

xor

শব্দার্থবিদ্যা

Performs element-wise XOR of two tensors lhs and rhs and produces a result tensor. Depending on the element type, does the following:

  • For booleans: logical XOR.
  • For integers: bitwise XOR.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor of boolean or integer type (C1)
(I2) rhs tensor of boolean or integer type (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of boolean or integer type (C1)

সীমাবদ্ধতা

  • (C1) type(lhs) = type(rhs) = type(result) .

উদাহরণ

// Bitwise operation with with integer tensors
// %lhs: [[1, 2], [3, 4]]
// %rhs: [[5, 6], [7, 8]]
%result = "stablehlo.xor"(%lhs, %rhs) : (tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32>
// %result: [[4, 4], [4, 12]]

// Logical operation with with boolean tensors
// %lhs: [[false, false], [true, true]]
// %rhs: [[false, true], [false, true]]
%result = "stablehlo.xor"(%lhs, %rhs) : (tensor<2x2xi1>, tensor<2x2xi1>) -> tensor<2x2xi1>
// %result: [[false, true], [true, false]]

আরো উদাহরণ

Dialect Interop

At the moment, StableHLO programs in the wild sometimes contain operations that are not defined by StableHLO.

Module, Function, Call and Return

StableHLO uses upstream MLIR operations for ModuleOp, FuncOp, CallOp, and ReturnOp. This was done for better interop with existing MLIR machinery, as many useful passes are written targeting FuncOp and ModuleOp, and many compilation pipelines expect these ops to be present. Full compatibility guarantees are applied to these ops. If anything ever changes about these ops in an incompatible way (ie removal), StableHLO equivalents will be added to preserve compatibility.

CHLO

The CHLO opset contains higher level operations that decompose to StableHLO. Currently there are no compatibility guarantees for CHLO. For compatibility guarantees, the chlo-legalize-to-stablehlo pass must be used prior to serialization.

Shape Operations

It is a common use case in the community to use certain operations from core MLIR dialects in dynamic StableHLO programs to perform shape computations. Most commonly, these include shape dialect ops like shape_of or num_elements , tensor dialect ops like dim or from_elements , and the builtin index type.

The Dynamism RFC > O2 denotes these as out of scope, however some support for index types is included for interop purposes. There are no compatibility guarantees for these ops or types. The shape-legalize-to-stablehlo pass can be used to convert these operations to fully supported StableHLO ops.

Deprecated Operations

There are several StableHLO operations that were inherited from MHLO which are deprecated and on the way out of StableHLO. The full details on these removals can be found in the StableHLO v1.0 Cleanup #2283 . The tracker issue for these deprecations is #2340 .

These operations fall into a few categories:

  • "Not in HLO" category of StableHLO operations - they were initially part of the StableHLO opset but have been later deemed to not fit it well: broadcast , create_token , cross-replica-sum , dot , einsum , torch_index_select , unary_einsum ( #3 ) .
  • Unused ops - These operations may have been useful at some point, but the ops were either underdeveloped, or the pipelines using these ops have been refactored to not require them anymore. This includes map , tuple ( #598 ), get_tuple_element , rng , complex comparisons #560 , and convolution window_reversal ( #1181 ).

Some of these ops can be removed easily given that they can be expressed using existing ops ( broadcast , create_token , cross-replica-sum , dot , unary_einsum ) and will be removed after the existing compatibilty window passes (6 months). Others are still being explored for removal ( einsum , get_tuple_element , map , rng torch_index_select , tuple , complex comparisons, window_reversal ). Pending community feedback, these ops will either be removed, or added to the spec with full support. Until these ops futures are known, they are only guaranteed 6 months of compatibility.

মৃত্যুদন্ড

অনুক্রমিক মৃত্যুদন্ড

A StableHLO program is executed by providing input values to the main function and computing output values. Output values of a function are computed by executing the graph of ops rooted in the corresponding return op.

The execution order is implementation-defined as long as it is aligned with dataflow, ie if ops are executed before their uses. In StableHLO, all side-effecting ops consume one token and produce one token (multiple tokens can be multiplexed into one token via after_all ), so the execution order of side effects is also aligned with dataflow. For example, in the below program there are two possible execution orders: %0%1%2return and %1%0%2return .

func.func @main() -> tensor<f64> {
  %0 = stablehlo.constant dense<1.0> : tensor<f64>
  %1 = stablehlo.constant dense<2.0> : tensor<f64>
  %2 = stablehlo.add %0, %1 : tensor<f64>
  return %2 : tensor<f64>
}

More formally, a StableHLO process is a combination of: 1) a StableHLO program, 2) operation statuses (not executed yet, already executed), and 3) intermediate values that the process is working on. The process starts with input values to the main function, progresses through the graph of ops updating operation statuses and intermediate values and finishes with output values. Further formalization is TBD ( #484 ).

Parallel execution

StableHLO programs can be executed in parallel, organized into a 2D process grid of num_replicas by num_partitions which both have type ui32 .

In the StableHLO process grid , num_replicas * num_partitions of StableHLO processes are executing at the same time. Each process has a unique process_id = (replica_id, partition_id) , where replica_id in replica_ids = range(num_replicas) and partition_id in partition_ids = range(num_partitions) which both have type ui32 .

The size of the process grid is known statically for every program (in the future, we are planning to make it an explicit part of StableHLO programs #650 ), and the position within the process grid is known statically for every process. Each process has access to its position within the process grid via the replica_id and partition_id ops.

Within the process grid, the programs can all be the same (in the "Single Program, Multiple Data" style), can all be different (in the "Multiple Program, Multiple Data" style) or something in between. In the future, we are planning to introduce support for other idioms of defining parallel StableHLO programs, including GSPMD ( #619 ).

Within the process grid, the processes are mostly independent from each other - they have separate operation statuses, separate input/intermediate/output values and most of the ops are executed separately between processes, with the exception of a small number of collective ops described below .

Given that execution of most of the ops is only using values from the same process, it is usually unambiguous to refer to these values by their names. However, when describing semantics of collective ops, that is insufficient, and that gives rise to the notation name@process_id to refer to the value name within a particular process. (From that perspective, unqualified name can be viewed as a shorthand for name@(replica_id(), partition_id()) ).

The execution order across processes is implementation-defined, except for the synchronization introduced by point-to-point communication and collective ops as described below.

পয়েন্ট টু পয়েন্ট যোগাযোগ

StableHLO processes can communicate with each other through StableHLO channels . A channel is represented by a positive id of type si64 . Through various ops, it is possible to send values to channels and receive them from channels.

Further formalization, eg where these channel ids are coming from, how processes programs become aware of them and what kind of synchronization is introduced by them, is TBD ( #484 ).

Streaming communication

Every StableHLO process has access to two streaming interfaces:

  • Infeed that can be read from.
  • Outfeed that can be written to.

Unlike channels, which are used to communicate between processes and therefore have processes at both of their ends, infeeds and outfeeds have their other end implementation-defined.

Further formalization, eg how streaming communication influences execution order and what kind of synchronization is introduced by it, is TBD ( #484 ).

Collective ops

There are six collective ops in StableHLO: all_gather , all_reduce , all_to_all , collective_broadcast , collective_permute , and reduce_scatter . All these ops split the processes in the StableHLO process grid into StableHLO process groups and execute a joint computation within each process group, independently from other process groups.

Within each process group, collective ops may introduce a synchronization barrier. Further formalization, eg elaborating on when exactly this synchronization happens, how exactly the processes arrive at this barrier, and what happens if they don't, is TBD ( #484 ).

If the process group involves cross-partition communication, ie there are processes in the process group whose partition ids are different, then execution of the collective op needs a channel, and the collective op must provide a positive channel_id of type si64 . Cross-replica communication doesn't need channels.

The computations performed by the collective ops are specific to individual ops and are described in individual op sections above. However, the strategies by which the process grid is split into process groups are shared between these ops and are described in this section. More formally, StableHLO supports the following four strategies.

cross_replica

Only cross-replica communications happen within each process group. This strategy takes replica_groups - a list of lists of replica ids - and computes a Cartesian product of replica_groups by partition_ids . replica_groups must have unique elements and cover all replica_ids . More formally, using Python syntax:

def cross_replica(replica_groups: List[List[ReplicaId]]) -> List[List[ProcessId]]:
  for replica_group in replica_groups:
    for partition_id in partition_ids:
      process_group = []
      for replica_id in replica_group:
        process_group.append((replica_id, partition_id))
      yield process_group

For example, for replica_groups = [[0, 1], [2, 3]] and num_partitions = 2 , cross_replica will produce [[(0, 0), (1, 0)], [(0, 1), (1, 1)], [(2, 0), (3, 0)], [(2, 1), (3, 1)]] .

cross_partition

Only cross-partition communications happen within each process group. This strategy takes partition_groups - a list of lists of partition ids - and computes a Cartesian product of partition_groups by replica_ids . partition_groups must have unique elements and cover all partition_ids . More formally, using Python syntax:

def cross_partition(partition_groups: List[List[PartitionId]]) -> List[List[ProcessId]]:
  for partition_group in partition_groups:
    for replica_id in replica_ids:
      process_group = []
      for partition_id in partition_group:
        process_group.append((replica_id, partition_id))
      yield process_group

For example, for partition_groups = [[0, 1]] and num_replicas = 4 , cross_partition will produce [[(0, 0), (0, 1)], [(1, 0), (1, 1)], [(2, 0), (2, 1)], [(3, 0), (3, 1)]] .

cross_replica_and_partition

Both cross-replica and cross-partition communications may happen within each process group. This strategy takes replica_groups - a list of lists of replica ids - and computes Cartesian products of each replica_group by partition_ids . replica_groups must have unique elements and cover all replica_ids . More formally, using Python syntax:

def cross_replica_and_partition(replica_groups: List[List[ReplicaId]]) -> List[List[ProcessId]]:
  for replica_group in replica_groups:
    process_group = []
    for partition_id in partition_ids:
      for replica_id in replica_group:
        process_group.append((replica_id, partition_id))
    yield process_group

For example, for replica_groups = [[0, 1], [2, 3]] and num_partitions = 2 , cross_replica_and_partition will produce [[(0, 0), (1, 0), (0, 1), (1, 1)], [(2, 0), (3, 0), (2, 1), (3, 1)]] .

flattened_ids

This strategy takes flattened_id_groups - a list of lists of "flattened" process ids in the form of replica_id * num_partitions + partition_id - and turns them into process ids. flattened_id_groups must have unique elements and cover all process_ids . More formally, using Python syntax:

def flattened_ids(flattened_id_groups: List[List[ui32]]) -> List[List[ProcessId]]:
  for flattened_id_group in flattened_id_groups:
    process_group = []
    for flattened_id in flattened_id_group:
      replica_id = flattened_id // num_partitions
      partition_id = flattened_id % num_partitions
      process_group.append((replica_id, partition_id))
    yield process_group

For example, for flattened_id_groups = [[0, 1, 2, 3], [4, 5, 6, 7]] , num_replicas = 4 and num_partitions = 2 , flattened_ids will produce [[(0, 0), (0, 1), (1, 0), (1, 1)], [(2, 0), (2, 1), (3, 0), (3, 1)]] .

নির্ভুলতা

At the moment, StableHLO does not provide guarantees about numerical accuracy, but this may change in the future ( #1156 ).

Execution semantics of quantized operation

The interpretation of quantized StableHLO operations may vary depending on the hardware requirements and capabilities. For instance, some hardware may opt to interpret quantized operations using a "dequantize, perform floating-point operation, and finally quantize" strategy. Others may perform the entire computation with integer arithmetic. Consequently, the interpretation of quantized StableHLO operations is exclusively determined by the specific implementation. The interpretation of hybrid quantization ( #1575 ) should be based on the it's semantics as prescribed in the specification (via 1792 ).

ত্রুটি

StableHLO programs are validated through an extensive set of constraints for individual ops, which rules out many classes of errors prior to run time. However, error conditions are still possible, eg through integer overflows, out-of-bounds accesses, etc. Unless explicitly called out, all these errors result in implementation-defined behavior, but this may change in the future ( #1157 ).

ফ্লোটিং পয়েন্ট ব্যতিক্রম

As an exception to this rule, floating-point exceptions in StableHLO programs have well-defined behavior. Operations which result in exceptions defined by the IEEE-754 standard (invalid operation, division-by-zero, overflow, underflow, or inexact exceptions) produce default results (as defined in the standard) and continue execution without raising the corresponding status flag; similar to raiseNoFlag exception handling from the standard. Exceptions for nonstandard operations (eg complex arithmetic and certain transcendental functions) are implementation-defined.

Shape mismatches

StableHLO supports dynamically-shaped tensors. However, shapes have to agree at runtime, otherwise the behavior is undefined. StableHLO does not explicitly provide an op that can assert that a tensor has a given shape at runtime. Generating correct code is the responsibility of the producer.

As a specific example, the below program is valid. However, at runtime, the exact shapes of %arg0 and %arg1 will have to be the same, otherwise the behavior of the program is undefined:

func.func @foo(%arg0: tensor<?xi32>, %arg1: tensor<?xi32>) -> tensor<?xi32> {
    %0 = stablehlo.add %arg0, %arg1 : tensor<?xi32>
    return %0 : tensor<?xi32>
}

স্বরলিপি

For describing syntax, this document is using the modified ISO flavor of EBNF syntax ( ISO/IEC 14977:1996 , Wikipedia ), with two modifications: 1) rules are defined using ::= rather than = ,

2) concatenation is expressed using juxtaposition rather than , .

For describing semantics (ie within "Types", "Constants" and "Ops" sections), we are using formulas which are based on Python syntax extended with support for concisely expressing array operations as described below. This works well for small snippets of code, but in rare cases when larger snippets of code are needed, we use vanilla Python syntax which is always introduced explicitly.

সূত্র

Let's explore how formulas work based on an example from the dot_general specification. One of the constraints for this operation looks as follows: dim(lhs, lhs_batching_dimensions...) = dim(rhs, rhs_batching_dimensions...) .

The names used in this formula come from two sources: 1) global functions, ie dim , 2) member definitions of the corresponding program element, ie lhs , lhs_batching_dimensions , rhs and rhs_batching_dimensions inputs defined in the "Inputs" section of dot_general .

As mentioned above, the syntax of this formula is Python-based with some conciseness-oriented extensions. To make sense of the formula, let's transform it into vanilla Python syntax.

A) In these formulas, we are using = to represent equality, so the first step towards obtaining Python syntax is replacing = with == , as follows: dim(lhs, lhs_batching_dimensions...) == dim(rhs, rhs_batching_dimensions...)

B) Also, these formulas support ellipses ( ... ) which turn scalar expressions into tensor expressions. In a nutshell, f(xs...) roughly means "for each scalar x in the tensor xs , compute a scalar f(x) and then return all these scalar results together as a tensor result". In vanilla Python syntax, our example formula turns into: [dim(lhs, dim1) for dim1 in lhs_batching_dimensions] == [dim(rhs, dim2) for dim2 in rhs_batching_dimensions] .

Thanks to ellipses, it is often possible to avoid working at the level of individual scalars. However, in some tricky cases, lower-level semi-informal syntax may be used like in the start_indices[bi0, ..., :, ..., biN] formula from the gather specification. In the service of conciseness, we don't provide an exact formalism for translating such syntax to vanilla Python, in hopes that it is still intuitively understandable on case-by-case basis. Please let us know if some specific formulas look opaque, and we'll try to improve them.

Also, you will notice that formulas use ellipses to expand all sorts of lists, including tensors, lists of tensors (which eg can arise from a variadic number of tensors), etc. This is another area where we don't provide an exact formalism (eg lists are not even part of the StableHLO type system) and instead rely on intuitive understandability.

C) The final noteworthy notational vehicle that we employ is implicit broadcasting. While the StableHLO opset doesn't support implicit broadcasting, the formulas do, also in the service of conciseness. In a nutshell, if a scalar is used in a context where a tensor is expected, the scalar is broadcasted to the expected shape.

To continue the dot_general example, here's another constraint: 0 <= lhs_batching_dimensions < rank(lhs) . As defined in the dot_general specification, lhs_batching_dimensions is a tensor, however both 0 and rank(lhs) are scalars. After we apply implicit broadcasting, the formula will become [0, ..., 0] <= lhs_batching_dimensions < [rank(lhs), ..., rank(lhs)] .

When applied to a particular dot_general operation, this formula will evaluate to a tensor of booleans. When formulas are used as constraints, the constraint holds if the formula evaluates to either true or to a tensor which only has true elements.

নাম

In formulas, lexical scope includes: 1) global functions, 2) member definitions,

3) local definitions. The list of global functions is provided below. The list of element definitions depends on the program element that the notation is applied to:

  • For operations, member definitions include names introduced in "Inputs" and "Outputs" sections.
  • For everything else, member definitions include structural parts of the program element, named after the corresponding EBNF non-terminals. Most of the time, the names of these structural parts are obtained by converting the names of the non-terminals to snake case (eg IntegerLiteral => integer_literal ), but sometimes names get abbreviated in the process (eg QuantizationStorageType => storage_type ) in which case the names are introduced explicitly similarly to "Inputs" / "Outputs" sections in operation specifications.
  • Additionally, member definitions always include self to refer to the corresponding program element.

মূল্যবোধ

When formulas are evaluated, they work with the following types of values: 1) Value (actual values, eg dense<[[1, 2], [3, 4]]> : tensor<2x2xi32> ; they always know their types), 2) Placeholder (future values, eg lhs , rhs or result ; their actual values are not known yet, only their types are known), 3) Type (types as defined in the "Types" section), 4) Function (global functions as defined in the "Functions" section).

Depending on the context, names may be referring to different values. More specifically, the "Semantics" section for ops (and equivalents for other program elements) defines runtime logic, so all inputs are available as Value . In contrast, the "Constraints" section for ops (and equivalents) defines "compile-time" logic, ie something that is typically executed before runtime, so only constant inputs are available as Value and other inputs are available only as Placeholder .

নাম In "Semantics" In "Constraints"
Global functions Function Function
Constant inputs Value Value
Non-constant inputs Value Placeholder
আউটপুট Value Placeholder
স্থানীয় সংজ্ঞা Depends on the definition Depends on the definition

Let's consider an example transpose operation:

%result = "stablehlo.transpose"(%operand) {
  permutation = dense<[2, 1, 0]> : tensor<3xi64>
} : (tensor<2x3x2xi32>) -> tensor<2x3x2xi32>

For this operation, permutation is a constant, so it's available as a Value in both semantics and constraints. In contrast, operand and result are available as a Value in semantics but only as a Placeholder in constraints.

ফাংশন

Construction of types

There are no functions that can be used to construct types. Instead, we directly use type syntax because it's typically more concise. Eg (tensor<E>, tensor<E>) -> (tensor<E>) rather than function_type( [tensor_type([], E), tensor_type([], E)], [tensor_type([], E)])

Functions on types

  • element_type is defined on tensor types and quantized tensor types and returns, respectively, the TensorElementType or QuantizedTensorElementType part of the corresponding TensorType or QuantizedTensorType .
def element_type(x: Value | Placeholder | Type):
 if type(x) == TensorType:
    return tensor_element_type(x)
  if type(x) == QuantizedTensorType:
    return quantized_tensor_element_type(x)
  if type(x) is not Type:
    return element_type(type(x))
  • is_per_axis_quantized(x: Value | Placeholder | Type) -> Value is a shortcut for is_quantized(x) and quantization_dimension(x) is not None .

  • is_per_tensor_quantized(x: Value | Placeholder | Type) -> Value is a shortcut for is_quantized(x) and quantization_dimension(x) is None .

  • is_promotable(x: Type, y: Type) -> bool checks if type x can be promoted to type y . When x and y are QuantizedTensorElementType s, the promotion is applied only to the storage_type . This specific version of promotion is currently used in context of reduction computation (refer to RFC for more details).

def is_promotable(x: Type, y: Type) -> Value:
  is_same_type = (is_bool(x) and is_bool(y)) or
    (is_integer(x) and is_integer(y)) or (is_float(x) and is_float(y)) or
    (is_complex(x) and is_complex(y)) or
    (is_quantized(x) and is_quantized(y) and expressed_type(x) = expressed_type(y))

  if is_same_type == False:
    return False

  if is_integer(x) or is_float(x):
    return bitwidth(x) <= bitwidth(y)

  if is_complex(x):
    return bitwidth(element_type(x)) <= bitwidth(element_type(y))

  if is_quantized(x):
    return bitwidth(storage_type(x)) <= bitwidth(storage_type(y))

  return false
  • is_quantized(x: Value | Placeholder | Type) -> Value is a shortcut for is_quantized_tensor_element_type(x) .

  • is_type_name(x: Value | Placeholder | Type) -> Value . Available for all types. For example, is_float(x) returns true if x is a FloatType . If x is a value or placeholder, this function is a shortcut for is_type_name(type(x)) .

  • max_value(x: Type) -> Value returns the maximum value of an TensorElementType . If x is not an TensorElementType , returns None .

  • min_value(x: Type) -> Value returns the minimum possible value of an TensorElementType . If x is not an TensorElementType , returns None .

  • member_name(x: Value | Placeholder | Type) -> Any . Available for all member definitions member_name of all types. For example, tensor_element_type(x) returns the TensorElementType part of a corresponding TensorType . If x is a value or placeholder, this function is a shortcut for member_name(type(x)) . If x is not a type that has an appropriate member, or a value or a placeholder of such a type, returns None .

  • is_empty_algorithm(*args: Type) checks if all dot algorithm fields are set to None . This is needed since dot algorithms have implementation defined default behaviors, so specifying a default value would be incorrect.

Construction of values

  • operation_name(*xs: Value | Type) -> Value . Available for all operations. For example, add(lhs, rhs) takes two tensor values lhs and rhs and returns the output of evaluating the add operation with these inputs. For some operations eg broadcast_in_dim , types of their outputs are "load-bearing", ie needed to evaluate an operation. In this case, the function takes these types as arguments.

Functions on values

  • All Python's operators and functions are available. Eg both subscription and slicing notations from Python are available to index into tensors, quantized tensors and tuples.

  • to_destination_type(x: Value, destination_type: Type) -> Value is defined on tensors and returns the converted value of x based on the type(x) and destination_type as follows:

def to_destination_type(x: Value, destination_type: Type) -> Value:
  if type(x) == destination_type:
    return x

  if is_quantized(destination_type):
    if is_quantized(type(x)):
      return quantize(x, destination_type)
    assert is_float(type(x))
    return quantize(x, destination_type)

  if is_quantized(type(x)):
    assert destination_type = expressed_type(type(x))
    return dequantize(type(x))

  return convert(x, destination_type)

There is early discussion on merging convert , uniform_quantize and uniform_dequantize operations ( #1576 ). After the merge we do not need the above function and can use the operation name for convert instead.

  • is_nan(x: Value) -> Value is defined on tensors and returns true if all elements of x are NaN or false otherwise. If x is not a tensor, returns None .

  • is_sorted(x: Value) -> Value is defined on tensors and returns true if elements of x are sorted in ascending order with respect to the ascending lexicographical order of their indices or false otherwise. If x is not a tensor, returns None .

  • is_unique(x: Value) -> Value is defined on tensors and returns true if x doesn't have duplicate elements or false otherwise. If x is not a tensor, returns None .

  • member_name(x: Value) -> Any is defined for all member definitions member_name of all values. For example, real_part(x) returns the RealPart part of a corresponding ComplexConstant . If x is not a value that has an appropriate member, returns None .

  • same(x: Value) -> Value is defined on tensors and returns true if elements of x are all equal to each other or false otherwise. If the tensor doesn't have elements, that counts as "all equal to each other", ie the function returns true . If x is not a tensor, returns None .

  • split(x: Value, num_results: Value, axis: Value) -> Value is defined on tensors and returns num_results slices of x along the axis axis . If x is not a tensor or dim(x, axis) % num_results != 0 , returns None .

  • is_defined_in_parent_scope(x: Value) -> Value is defined on strings and returns true if x is the name of a function defined in the same scope as the parent function of the relevant op.

  • is_namespaced_op_name(x: Value) -> Value is defined on strings and returns true if x is a valid op name, that is it respects the following regular expression: [a-zA-Z][a-zA-Z0-9_]*([.][a-zA-Z0-9_$]+)+

Shape computations

  • axes(x: Value | Placeholder | Type) -> Value is a shortcut for range(rank(x)) .

  • dim(x: Value | Placeholder | Type, axis: Value) -> Value is a shortcut for shape(x)[axis] .

  • dims(x: Value | Placeholder | Type, axes: List) -> List is a shortcut for list(map(lambda axis: dim(x, axis), axes)) .

  • index_space(x: Value | Placeholder | Type) -> Value is defined on tensors and returns size(x) indices for the corresponding TensorType sorted in ascending lexicographical order, ie [0, ..., 0] , [0, ..., 1] , ..., shape(x) - 1 . If x is not a tensor type, a quantized tensor type, or a value or a placeholder of one of these types, returns None .

  • rank(x: Value | Placeholder | Type) -> Value is a shortcut for size(shape(x)) .

  • shape(x: Value | Placeholder | Type) -> Value is defined in the "Functions on types" section via member_name .

  • size(x: Value | Placeholder | Type) -> Value is a shortcut for reduce(lambda x, y: x * y, shape(x)) .

Quantization computations

  • def baseline_element_type(x: Value | Placeholder | Type) -> Type is a shortcut for element_type(baseline_type(x)) .

  • baseline_type is defined on tensor types and quantized tensor types and transforms them to a "baseline", ie a type with the same shape but with the quantization parameters of the element type reset to default values. This is used as a handy trick to compare both tensor and quantized tensor types uniformly, which is needed quite often. For quantized types, this enables comparing types ignoring the quantization parameters, that is, shape , storage_type , expressed_type , storage_min , storage_max , and quantization_dimension (for per-axis quantized type) must all match, but scales and zero points may differ.

def baseline_type(x: Value | Placeholder | Type) -> Type:
  if type(x) == TensorType:
    return x
  if type(x) == QuantizedTensorType:
    element_type = quantized_tensor_element_type(x)
    baseline_element_type = QuantizedTensorElementType(
      storage_type = storage_type(element_type),
      storage_min = storage_min(element_type),
      storage_max = storage_max(element_type),
      expressed_type = expressed_type(element_type),
      quantization_dimension = quantization_dimension(element_type),
      scales = [constant(1.0, expressed_type(element_type))] * dim(x, quantization_dimension(element_type)),
      zero_points = [constant(0, storage_type(element_type))] * dim(x, quantization_dimension(element_type)))
    return QuantizedTensorType(shape(x), baseline_element_type)
  if type(x) is not Type:
    return baseline_element_type(type(x))
  • dequantize is defined on quantized tensor types and turns them into floating-point tensor types. This happens via converting quantized elements which represent integer values of the storage type into corresponding floating-point values of the expressed type using the zero point and scale associated with the quantized element type.
def compute_zero_points(quantized_type, result_type):
  if is_per_tensor_quantized(quantized_type):
    return broadcast_in_dim(constant(zero_point(quantized_type), storage_type(quantized_type)), [], result_type)
  if is_per_axis_quantized(quantized_type):
    for i in index_space(result_type):
      d = quantization_dimension(quantized_type)
      zero_points[i] = zero_points(quantized_type)[i[d]]
    return zero_points

def compute_scales(quantized_type, result_type):
  if is_per_tensor_quantized(quantized_type):
    return broadcast_in_dim(constant(scale(quantized_type), expressed_type(quantized_type)), [],
            type(result_type))
  if is_per_axis_quantized(quantized_type):
    for i in index_space(result_type):
      d = quantization_dimension(quantized_type)
      scales[i] = scales(quantized_type)[i[d]]
    return scales

def dequantize(x: Value) -> Value:
  assert is_quantized(x)
  x_storage = bitcast_convert(x, storage_type(x))
  x_storage_sub = x_storage - compute_zero_points(type(x), type(x_storage))
  x_expressed_sub = convert(x_storage_sub, expressed_type(x))
  return x_expressed_sub * compute_scales(type(x), type(x_expressed_sub))
  • quantize is defined on floating-point tensor types and turns them into quantized tensor types. This happens via converting floating-point values of the expressed type into corresponding integer values of the storage type using the zero point and scale associated with the quantized element type.
def quantize(x: Value, result_type: Type) -> Value:
  assert is_float(x) and is_quantized(result_type)
  zero_points = compute_zero_points(result_type, TensorType(shape(x), storage_type(result_type)))
  converted_zero_points = convert(zero_points, expressed_type(result_type))
  converted_min = convert(storage_min(result_type), expressed_type(result_type))
  converted_max = convert(storage_max(result_type), expressed_type(result_type))

  x_scaled = x / compute_scales(result_type, type(x))
  x_scaled_add_zp = x_scaled + converted_zero_points
  x_clamped = clamp(converted_min, x_scaled_add_zp, converted_max)
  x_rounded = round_nearest_even(x_clamped)
  return convert(x_rounded, result_type)
  • dequantize_op_quantize is used to specify element-wise computations on quantized tensors. It dequantizes, ie turns quantized elements into their expressed types, then performs an operation, and then quantizes, ie turns the results back into their storage types. At the moment, this function only works for per-tensor quantization. Per-axis quantization is work in progress ( #1574 ).
def dequantize_op_quantize(op, *inputs_and_output_type):
  inputs = inputs_and_output_type[:-1]
  output_type = inputs_and_output_type[-1]

  float_inputs = map(dequantize, inputs)
  float_result = op(*float_inputs)
  return quantize(float_result, output_type)

def dequantize_batch_norm_grad_or_training_quantize(op, *inputs_and_output_types):
  inputs = inputs_and_output_type[:-3]
  float_inputs = map(dequantize, inputs)
  float_results = op(*float_inputs)
  return map(quantize, float_results, inputs_and_output_type[-3:])

def dequantize_compare(lhs, rhs, comparison_direction):
  float_lhs = dequantize(lhs)
  float_rhs = dequantize(rhs)
  return compare(float_lhs, float_rhs, comparison_direction, FLOAT)

def dequantize_select_quantize(pred, on_true, on_false, output_type):
  float_on_true = dequantize(on_true)
  float_on_false = dequantize(on_false)
  float_result = select(pred, float_on_true, float_on_false)
  return quantize(float_result, output_type)
  • hybrid_dequantize_then_op is used to specify weight-only quantization for hybrid op which accepts lhs in floating-point and rhs in quantized types. It dequantizes quantized inputs into their expressed types and performs computation in float. Element type of float lhs tensor and expressed type of quantized rhs tensor should be identical.
def hybrid_dequantize_then_op(op, lhs, rhs):
  assert(is_float(lhs) and is_quantized(rhs) and element_type(lhs) == expressed_type(rhs))
  return op(lhs, dequantize(rhs))

Grid computations

  • cross_partition(replica_groups: Value) -> Value . See the "cross_replica" section above.

  • cross_replica(replica_groups: Value) -> Value . See the "cross_replica" section above.

  • cross_replica_and_partition(replica_groups: Value) -> Value . See the "cross_replica_and_partition" section above.

  • flattened_ids(replica_groups: Value) -> Value . See the "flattened_ids" section above.

গতিশীলতা

StableHLO values can have dynamic dimension sizes, eg tensor<?xi64> . However, StableHLO values cannot have a dynamic number of dimensions (unranked dynamism, eg tensor<*xi64> ). Operands and results are allowed to use dynamic dimension sizes, even if there are constraints on the sizes. Constraints will be verified statically if possible, otherwise they are deferred to runtime and mismatches will result in undefined behavior. উদাহরণ জন্য নীচে দেখুন.

Shape mismatches for unary elementwise operations

Consider the following toy program:

func.func @foo(%arg0: tensor<?xf64>) {
  %0 = stablehlo.abs %arg0 : (tensor<?xf64>) -> tensor<2xf64>
  return
}

Such a program is unusual, because it is not common to know the shape of the result but not the shape of the input. Nonetheless, this is a valid StableHLO program. It is not possible to statically validate the abs operation in this program, because the exact shape of the operand is unknown. However, the shapes are certainly compatible, and this can be checked statically: ? could turn out to be 2 at runtime, and there would be no issue. However, ? could also turn out to be some other integer, in which case the behavior is undefined.

Note that if a dimension size is dynamic in the result, there cannot be undefined behavior. Indeed, there is no "expected" size, so there cannot be a mismatch.

Shape mismatches for binary elementwise operations

Consider the following toy program:

func.func @foo(%arg0: tensor<?xf64>, %arg1: tensor<?xf64>) {
  %0 = stablehlo.add %arg0, %arg0 : (tensor<?xf64>, tensor<?xf64>) -> tensor<?xf64>
  return
}

When it comes to binary elementwise operations, the shapes of the inputs and the result must agree at runtime. At compile time, static dimensions must be equal, otherwise they merely need to be compatible. If any dimension is dynamic in the inputs, then there could be undefined behavior at runtime, because the dynamic size may not match the corresponding size in the other operand (be it static or dynamic). If all the inputs are static, then whether the result is dynamic or not does not matter: statically known dimensions will be checked statically, and dynamic dimensions do not impose any constraints.

Shape mismatches for ops that take their output shape as an operand

Consider the following toy program:

func.func @foo(%arg0: tensor<2xi32>) {
  %0 = stablehlo.dynamic_iota %arg0, dim = 0 : (tensor<2xi32>) -> tensor<3x4xi64>
  return
}

The values in the shape operand at runtime must match the shape of the result, otherwise the behavior is undefined. That is, at runtime %arg0 must have a value of dense<[3, 4]> : tensor<2xi32> . If the shape operand is constant, this can be verified statically. If the result shape is fully dynamic, then there cannot be a mismatch.

,

StableHLO is an operation set for high-level operations (HLO) in machine learning (ML) models. StableHLO works as a portability layer between different ML frameworks and ML compilers: ML frameworks that produce StableHLO programs are compatible with ML compilers that consume StableHLO programs.

Our goal is to simplify and accelerate ML development by creating more interoperability between various ML frameworks (such as TensorFlow, JAX and PyTorch) and ML compilers (such as XLA and IREE). Towards that end, this document provides a specification for the StableHLO programming language.

This specification contains three major sections. First, the Programs section describes the structure of StableHLO programs which consist of StableHLO functions which themselves consist of StableHLO ops. Within that structure, the Ops section specifies the semantics of individual ops. The Execution section provides semantics for all these ops executing together within a program. Finally, the Notation section discusses the notation used throughout the specification.

To view the spec from a previous release of StableHLO, open the repo at the tagged release of interest. For example, the StableHLO v0.19.0 Spec . To view changes that occurred at each minor version bump of StableHLO, refer to the version log in VhloDialect.td .

প্রোগ্রাম

Program ::= {Func}

StableHLO programs consist of an arbitrary number of StableHLO functions. Below is an example program with a function @main which has 3 inputs ( %image , %weights and %bias ) and 1 output. The body of the function has 6 ops.

func.func @main(
  %image: tensor<28x28xf32>,
  %weights: tensor<784x10xf32>,
  %bias: tensor<1x10xf32>
) -> tensor<1x10xf32> {
  %0 = "stablehlo.reshape"(%image) : (tensor<28x28xf32>) -> tensor<1x784xf32>
  %1 = "stablehlo.dot"(%0, %weights) : (tensor<1x784xf32>, tensor<784x10xf32>) -> tensor<1x10xf32>
  %2 = "stablehlo.add"(%1, %bias) : (tensor<1x10xf32>, tensor<1x10xf32>) -> tensor<1x10xf32>
  %3 = "stablehlo.constant"() {value = dense<0.0> : tensor<1x10xf32>} : () -> tensor<1x10xf32>
  %4 = "stablehlo.maximum"(%2, %3) : (tensor<1x10xf32>, tensor<1x10xf32>) -> tensor<1x10xf32>
  "func.return"(%4): (tensor<1x10xf32>) -> ()
}

ফাংশন

Func        ::= 'func' '.' 'func' FuncId FuncInputs FuncOutputs '{' FuncBody '}'
FuncInputs  ::= '(' [FuncInput {',' FuncInput}] `)`
FuncInput   ::= ValueId ':' ValueType
FuncOutputs ::= ['->' FuncOutput, {',' FuncOutput}]
FuncOutput  ::= ValueType
FuncBody    ::= {Op}

StableHLO functions (which are also called named functions ) have an identifier, inputs/outputs and a body. In the future, we are planning to introduce additional metadata for functions to achieve better compatibility with HLO ( #425 , #626 , #740 , #744 ).

শনাক্তকারী

FuncId  ::= '@' letter {letter | digit}
ValueId ::= '%' digit {digit}
          | '%' letter {letter | digit}
letter  ::= 'a' | ... | 'z' | 'A' | ... | 'Z' | '_'
digit   ::= '0' | ... | '9'

StableHLO identifiers are similar to identifiers in many programming languages, with two peculiarities: 1) all identifiers have sigils which distinguish different kinds of identifiers, 2) value identifiers can be completely numeric to simplify generation of StableHLO programs.

প্রকারভেদ

Type         ::= ValueType | NonValueType
ValueType    ::= TensorType | QuantizedTensorType | TokenType | TupleType
NonValueType ::= TensorElementType | QuantizedTensorElementType | FunctionType | StringType

StableHLO types are categorized into value types (which are also called first-class types ) which represent StableHLO values and non-value types which describe other program elements. StableHLO types are similar to types in many programming languages, with the main peculiarity being StableHLO's domain-specific nature which results in some unusual outcomes (eg scalar types are not value types).

TensorType ::= 'tensor' '<' Shape TensorElementType '>'
Shape ::= {DimensionSize 'x'}
DimensionSize ::= digit {digit} | '?'

Tensor types represent tensors, ie multidimensional arrays. They have a shape and an element type , where a shape represents non-negative or unknown dimension sizes in the ascending order of the corresponding dimensions (which are also called axes ) numbered from 0 to R-1 . The number of dimensions R is called rank . For example, tensor<2x3xf32> is a tensor type with shape 2x3 and element type f32 . It has two dimensions (or, in other words, two axes) - 0th dimension and 1st dimension - whose sizes are 2 and 3. Its rank is 2.

Shapes can be partially or completely unknown (dynamic), eg tensor<?x2xf64> is partially unknown and tensor<?x?xf64> is completely unknown. Dynamic dimension sizes are represented using a ? . Shapes cannot be unranked.

In the future, we are planning to explore extending tensor types beyond dimension sizes and element types, for example, to include layouts ( #629 ) and sparsity ( #1078 ).

QuantizedTensorType ::= 'tensor' '<' Shape QuantizedTensorElementType '>'
QuantizedTensorElementType ::= '!quant.uniform' '<'
                  QuantizationStorageType
                  ['<' QuantizationStorageMin ':' QuantizationStorageMax '>']
                  ':' QuantizationExpressedType
                  [':' QuantizationDimension]
                  ',' QuantizationParameters '>'
QuantizationStorageType ::= IntegerType
QuantizationStorageMin ::= IntegerLiteral
QuantizationStorageMax ::= IntegerLiteral
QuantizationExpressedType ::= FloatType
QuantizationDimension ::= IntegerLiteral
QuantizationParameters ::= QuantizationParameter
                         | '{' QuantizationParameter {',' QuantizationParameter} '}'
QuantizationParameter ::= QuantizationScale [':' QuantizationZeroPoint]
QuantizationScale ::= FloatLiteral
QuantizationZeroPoint ::= IntegerLiteral
নাম টাইপ সীমাবদ্ধতা
storage_type integer type (C1-C3), (C8)
storage_min integer constant (C1), (C3), (C7)
storage_max integer constant (C2), (C3), (C7)
expressed_type floating-point type (C4)
quantization_dimension optional integer constant (C10-C12)
scales variadic number of floating-point constants (C4-C6), (C9), (C10), (C13)
zero_points variadic number of integer constants (C7-C9)

Quantized element types represent integer values of a storage type in the range from storage_min to storage_max (inclusive) that correspond to floating-point values of an expressed type . For a given integer value i , the corresponding floating-point value f can be computed as f = (i - zero_point) * scale , where scale and zero_point are called quantization parameters . The storage_min and storage_max are optional in the grammar, but have default values of min_value(storage_type) and max_value(storage_type) respectively. Quantized element types have the following constraints:

  • (C1) type(storage_min) = storage_type .
  • (C2) type(storage_max) = storage_type .
  • (C3) min_value(storage_type) <= storage_min < storage_max <= max_value(storage_type) .
  • (C4) type(scales...) = expressed_type .
  • (C5) 0 < scales .
  • (C6) is_finite(scales...) .
  • (C7) storage_min <= zero_points <= storage_max .
  • (C8) type(zero_points...) = storage_type .
  • (C9) size(scales) = size(zero_points) .
  • (C10) If is_empty(quantization_dimension) , then size(scales) = 1 .
  • (C11) 0 <= quantization_dimension .

At the moment, QuantizationScale is a floating-point constant, but there is strong interest in integer-based scales, represented with multipliers and shifts. We are planning to explore this in the near future ( #1404 ).

There is an ongoing discussion on the semantics of QuantizationZeroPoint , including the type, the values and whether there can be just one or potentially multiple zero points in a quantized tensor type. Based on the results of this discussion, the specification around zero points may change in the future ( #1405 ).

Another ongoing discussion involves the semantics of QuantizationStorageMin and QuantizationStorageMax to determine whether any constraints should be imposed on these values and on the values of quantized tensors ( #1406 ).

Finally, we are planning to explore representing unknown scales and zero points, similarly to how we are planning to explore representing unknown dimension sizes ( #1407 ).

Quantized tensor types represent tensors with quantized elements. These tensors are exactly the same as regular tensors, except that their elements have quantized element types, instead of regular element types.

In quantized tensors, quantization can be per-tensor , meaning, having one scale and zero_point for the entire tensor or can be per-axis , meaning, having multiple scales and zero_points , one pair per slice of a particular dimension quantization_dimension . More formally, in a tensor t with per-axis quantization, there are dim(t, quantization_dimension) slices of the quantization_dimension : t[:, ..., 0, ..., :], t[:, ..., 1, ..., :] , etc. All elements in the i th slice use scales[i] and zero_points[i] as their quantization পরামিতি Quantized tensor types have the following constraints:

  • For per-tensor quantization:
    • No additional constraints.
  • For per-axis quantization:
    • (C12) quantization_dimension < rank(self) .
    • (C13) dim(self, quantization_dimension) = size(scales) .
TokenType ::= 'token'

Token types represent tokens, ie opaque values produced and consumed by some operations. Tokens are used for imposing execution order on operations as described in the Execution section.

TupleType ::= 'tuple' '<' TupleElementTypes '>'
TupleElementTypes ::= [ValueType {',' ValueType}]

Tuple types represent tuples, ie heterogeneous lists. Tuples are a legacy feature which only exists for compatibility with HLO. In HLO, tuples are used to represent variadic inputs and outputs. In StableHLO, variadic inputs and outputs are supported natively, and the only use of tuples in StableHLO is to comprehensively represent HLO ABI where eg T , tuple<T> and tuple<tuple<T>> may be materially different depending on a particular implementation . In the future, we are planning to make changes to HLO ABI which may allow us to remove tuple types from StableHLO ( #598 ).

TensorElementType ::= BooleanType | IntegerType | FloatType | ComplexType
BooleanType ::= 'i1'
IntegerType ::= SignedIntegerType | UnsignedIntegerType
SignedIntegerType ::= 'si2' | 'si4' | 'si8' | 'si16' | 'si32' | 'si64'
UnsignedIntegerType ::= 'ui2' | 'ui4' | 'ui8' | 'ui16' | 'ui32' | 'ui64'
FloatType ::= 'f4E2M1FN' | 'f6E2M3FN' | 'f6E3M2FN' | 'f8E3M4' | 'f8E4M3'
            | 'f8E4M3FN' | 'f8E4M3FNUZ' | 'f8E4M3B11FNUZ' | 'f8E5M2'
            | 'f8E5M2FNUZ' | 'f8E8M0FNU' | 'bf16' | 'f16' | 'f32' | 'f64'
TensorFloat32 ::= 'tf32'
ComplexType ::= 'complex' '<' ComplexElementType '>'
ComplexElementType ::= 'f32' | 'f64'

Element types represent elements of tensor types. Unlike in many programming languages, these types are not first class in StableHLO. This means that StableHLO programs cannot directly represent values of these types (as a result, it is idiomatic to represent scalar values of type T with 0-dimensional tensor values of type tensor<T> ).

  • Boolean type represents boolean values true and false .
  • Integer types can be either signed ( si ) or unsigned ( ui ) and have one of the supported bit widths ( 2 , 4 , 8 , 16 , 32 or 64 ). Signed siN types represent integer values from -2^(N-1) to 2^(N-1)-1 inclusive, and unsigned uiN types represent integer values from 0 to 2^N-1 inclusive.
  • Floating-point types can be one of the following:
  • Complex types represent complex values that have a real part and an imaginary part of the same element type . Supported complex types are complex<f32> (both parts are of type f32 ) and complex<f64> (both parts are of type f64 ).
FunctionType ::= '(' InputTypes ')' '->' '(' OutputTypes ')'
InputTypes ::= [ValueType {',' ValueType}]
OutputTypes ::= [ValueType {',' ValueType}]

Function types represent both named and anonymous functions. They have input types (the list of types on the left-hand side of -> ) and output types (the list of types on the right-hand side of -> ). In many programming languages, function types are first class, but not in StableHLO.

StringType ::= 'string'

String type represents sequences of bytes. Unlike in many programming languages, string type is not first class in StableHLO and is only used to specify static metadata for program elements.

অপারেশন

StableHLO operations (which are also called ops ) represent a closed set of high-level operations in machine learning models. As discussed above, StableHLO syntax is heavily inspired by MLIR, which is not necessarily the most ergonomic alternative, but is arguably the best fit for StableHLO's goal of creating more interoperability between ML frameworks and ML compilers.

Op            ::= [OpOutputs] OpName OpInputs ':' OpSignature
OpName        ::= '"' 'stablehlo' '.' OpMnemonic '"'
OpMnemonic    ::= 'abs' | 'add' | ...

StableHLO operations (which are also called ops ) have a name, inputs/outputs and a signature. The name consists of the stablehlo. prefix and a mnemonic which uniquely identifies one of the supported ops. See below for a comprehensive list of all supported ops.

OpInputs        ::= OpInputValues OpInputFuncs OpInputAttrs
OpInputValues   ::= '(' [OpInputValue {',' OpInputValue}] ')'
OpInputValue    ::= ValueId
OpInputFuncs    ::= ['(' OpInputFunc {',' OpInputFunc} ')']
OpInputAttrs    ::= ['{' OpInputAttr {',' OpInputAttr} '}']
OpOutputs       ::= [OpOutput {',' OpOutput} '=']
OpOutput        ::= ValueId

Ops consume inputs and produce outputs . Inputs are categorized into input values (computed during execution), input functions (provided statically, because in StableHLO functions are not first-class values) and input attributes (also provided statically). The kind of inputs and outputs consumed and produced by an op depends on its mnemonic. For example, the add op consumes 2 input values and produces 1 output value. In comparison, the select_and_scatter op consumes 3 input values, 2 input functions and 3 input attributes.

OpInputFunc ::= '{' Unused FuncInputs ':' FuncBody '}'
Unused      ::= '^' digit {digit}
              | '^' letter {letter | digit}

Input functions (which are also called anonymous functions ) are very similar to named functions except that: 1) they don't have an identifier (hence the name "anonymous"), 2) they don't declare output types (output types are inferred from the return op within the function).

The syntax for input functions includes a currently unused part (see the Unused production above) which is there for compatibility with MLIR. In MLIR, there is a more general concept of "regions" which can have multiple "blocks" of ops connected together via jump ops. These blocks have ids which correspond to the Unused production, so that they can be distinguished from each other. StableHLO doesn't have jump ops, so the corresponding part of MLIR syntax is unused (but is still there).

OpInputAttr      ::= OpInputAttrName '=' OpInputAttrValue
OpInputAttrName  ::= letter {letter | digit}
OpInputAttrValue ::= Constant

Input attributes have a name and a value which is one of the supported constants. They are the primary way to specify static metadata for program elements. For example, the concatenate op uses the attribute dimension to specify the dimension along which its input values are concatenated. Similarly, the slice op uses multiple attributes like start_indices and limit_indices to specify the bounds that are used to slice the input value.

At the moment, StableHLO programs in the wild sometimes contain attributes which are not described in this document. In the future, we are planning to either absorb these attributes into the StableHLO opset or prohibit them from appearing in StableHLO programs. In the meanwhile, here is the list of these attributes:

  • layout ( #629 ).
  • mhlo.frontend_attributes ( #628 ).
  • mhlo.sharding ( #619 ).
  • output_operand_aliases ( #740 ).
  • Location metadata ( #594 ).
OpSignature ::= '(' [ValueType {',' ValueType}] ')' '->' '(' [ValueType {',' ValueType}] ')'

Op signature consists of the types of all input values (the list of types on the left-hand side of -> ) and the types of all output values (the list of types on the right-hand side of -> ). Strictly speaking, input types are redundant, and output types are almost always redundant as well (because for most StableHLO ops, output types can be inferred from inputs). Nonetheless, op signature is deliberately part of StableHLO syntax for compatibility with MLIR.

Below is an example op whose mnemonic is select_and_scatter . It consumes 3 input values ( %operand , %source and %init_value ), 2 input functions and 3 input attributes ( window_dimensions , window_strides and padding ). Note how the signature of the op only includes the types of its input values (but not the types of input functions and attributes which are provided inline).

%result = "stablehlo.select_and_scatter"(%operand, %source, %init_value) ({
  ^bb0(%arg0: tensor<i32>, %arg1: tensor<i32>):
    %0 = "stablehlo.compare"(%arg0, %arg1) {
      comparison_direction = #stablehlo<comparison_direction GE>
    } : (tensor<i32>, tensor<i32>) -> tensor<i1>
    "stablehlo.return"(%0) : (tensor<i1>) -> ()
}, {
  ^bb0(%arg0: tensor<i32>, %arg1: tensor<i32>):
    %0 = "stablehlo.add"(%arg0, %arg1) : (tensor<i32>, tensor<i32>) -> tensor<i32>
    "stablehlo.return"(%0) : (tensor<i32>) -> ()
}) {
  window_dimensions = dense<[3, 1]> : tensor<2xi64>,
  window_strides = dense<[2, 1]> : tensor<2xi64>,
  padding = dense<[[0, 1], [0, 0]]> : tensor<2x2xi64>
} : (tensor<4x2xi32>, tensor<2x2xi32>, tensor<i32>) -> tensor<4x2xi32>

ধ্রুবক

Constant ::= BooleanConstant
           | IntegerConstant
           | FloatConstant
           | ComplexConstant
           | TensorConstant
           | QuantizedTensorConstant
           | StringConstant
           | EnumConstant

StableHLO constants have a literal and a type which together represent a StableHLO value. Generally, the type is part of the constant syntax, except when it's unambiguous (eg a boolean constant unambiguously has type i1 , whereas an integer constant can have multiple possible types).

BooleanConstant ::= BooleanLiteral
BooleanLiteral  ::= 'true' | 'false'

Boolean constants represent boolean values true and false . Boolean constants have type i1 .

IntegerConstant   ::= IntegerLiteral ':' IntegerType
IntegerLiteral    ::= ['-' | '+'] DecimalDigits
                    | ['-' | '+'] '0x' HexadecimalDigits
DecimalDigits     ::= decimalDigit {decimalDigit}
HexadecimalDigits ::= hexadecimalDigit {hexadecimalDigit}
decimalDigit      ::= '0' | ... | '9'
hexadecimalDigit  ::= decimalDigit | 'a' | ... | 'f' | 'A' | ... | 'F'

Integer constants represent integer values via strings that use decimal or hexadecimal notation. Other bases, eg binary or octal, are not supported. Integer constants have the following constraints:

  • (C1) is_wellformed(integer_literal, integer_type) .
FloatConstant  ::= FloatLiteral ':' FloatType
FloatLiteral   ::= SignPart IntegerPart FractionalPart ScientificPart
                 | '0x' [HexadecimalDigits]
SignPart       ::= ['-' | '+']
IntegerPart    ::= DecimalDigits
FractionalPart ::= ['.' [DecimalDigits]]
ScientificPart ::= [('e' | 'E') ['-' | '+'] DecimalDigits]

Floating-point constants represent floating-point values via strings that use decimal or scientific notation. Additionally, hexadecimal notation can be used to directly specify the underlying bits in the floating-point format of the corresponding type. Floating-point constants have the following constraints:

  • (C1) If non-hexadecimal notation is used, is_wellformed(float_literal, float_type) .
  • (C2) If hexadecimal notation is used, size(hexadecimal_digits) = num_bits(float_type) / 4 .
ComplexConstant ::= ComplexLiteral ':' ComplexType
ComplexLiteral  ::= '(' RealPart ',' ImaginaryPart ')'
RealPart        ::= FloatLiteral
ImaginaryPart   ::= FloatLiteral

Complex constants represent complex values using lists of a real part (comes first) and an imaginary part (comes second). For example, (1.0, 0.0) : complex<f32> represents 1.0 + 0.0i , and (0.0, 1.0) : complex<f32> represents 0.0 + 1.0i . The order in which these parts are then stored in memory is implementation-defined. Complex constants have the following constraints:

  • (C1) is_wellformed(real_part, complex_element_type(complex_type)) .
  • (C2) is_wellformed(imaginary_part, complex_element_type(complex_type)) .
TensorConstant ::= TensorLiteral ':' TensorType
TensorLiteral  ::= 'dense' '<' (DenseLiteral | ElementLiteral) '>'
DenseLiteral   ::= DenseDimension | DenseElements
DenseDimension ::= '[' [DenseLiteral {',' DenseLiteral}] ']'
DenseElements  ::= [ElementLiteral {',' ElementLiteral}]
ElementLiteral ::= BooleanLiteral | IntegerLiteral | FloatLiteral | ComplexLiteral

Tensor constants represent tensor values using nested lists specified via NumPy notation. For example, dense<[[1, 2, 3], [4, 5, 6]]> : tensor<2x3xi32> represents a tensor value with the following mapping from indices to elements: {0, 0} => 1 , {0, 1} => 2 , {0, 2} => 3 , {1, 0} => 4 , {1, 1} => 5 , {1, 2} => 6 . The order in which these elements are then stored in memory is implementation-defined. Tensor constants have the following constraints:

  • (C1) has_syntax(tensor_literal, element_type(tensor_type)) , where:
    • has_syntax(element_literal: Syntax, element_type: Type) = is_wellformed(element_literal, type) .
    • has_syntax(tensor_literal: List, element_type: Type) = has_syntax(tensor_literal..., element_type) .
  • (C2) has_shape(tensor_literal, shape(tensor_type)) , where:
    • has_shape(element_literal: Syntax, []) = true .
    • has_shape(tensor_literal: List, shape: List) = size(tensor_literal) = shape[0] and has_shape(tensor_literal..., shape[1:]) .
    • অন্যথায়, false
QuantizedTensorConstant ::= QuantizedTensorLiteral ':' QuantizedTensorType
QuantizedTensorLiteral  ::= 'dense' '<' (DenseLiteral | ElementLiteral) '>'

Quantized tensor constants represent quantized tensor values using the same notation as tensor constants, with elements specified as constants of their storage type. Quantized tensor constants have the following constraints:

  • (C1) has_syntax(quantized_tensor_literal, storage_type(quantized_tensor_type)) .
  • (C2) has_shape(quantized_tensor_literal, shape(quantized_tensor_type)) .
StringConstant  ::= StringLiteral
StringLiteral   ::= '"' {stringCharacter | escapeSequence} '"'
stringCharacter ::= all ASCII characters except '\00', '\01', ... '\1f' and '"'
escapeSequence  ::= '\' ('"' | '\' | 'n' | 't' | (hexadecimalDigit hexadecimalDigit))

String literals consist of bytes specified using ASCII characters and escape sequences. They are encoding-agnostic, so the interpretation of these bytes is implementation-defined. String literals have type string .

অপ্স

abs

শব্দার্থবিদ্যা

Performs element-wise abs operation on operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For signed integers: integer modulus.
  • For floats: abs from IEEE-754.
  • For complex numbers: complex modulus.
  • For quantized types: dequantize_op_quantize(abs, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of signed integer, floating-point, or complex type or per-tensor quantized tensor (C1-C2)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of signed integer or floating-point type or per-tensor quantized tensor (C1-C2)

সীমাবদ্ধতা

  • (C1) shape(result) = shape(operand) .
  • (C2) baseline_element_type(result) is defined as:
    • complex_element_type(element_type(operand)) if is_complex(operand) .
    • baseline_element_type(operand) otherwise.

উদাহরণ

// %operand: [-2, 0, 2]
%result = "stablehlo.abs"(%operand) : (tensor<3xi32>) -> tensor<3xi32>
// %result: [2, 0, 2]

আরো উদাহরণ

যোগ করুন

শব্দার্থবিদ্যা

Performs element-wise addition of two tensors lhs and rhs and produces a result tensor. Depending on the element type, does the following:

  • For booleans: logical OR.
  • For integers: integer addition.
  • For floats: addition from IEEE-754.
  • For complex numbers: complex addition.
  • For quantized types: dequantize_op_quantize(add, lhs, rhs, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor or quantized tensor (C1-C6)
(I2) rhs tensor or quantized tensor (C1-C5), (C7)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or quantized tensor (C1-C7)

সীমাবদ্ধতা

  • If the operation uses non-quantized tensors:
    • (C1) type(lhs) = type(rhs) = type(result) .
  • If the operation uses quantized tensors:
    • (C2) is_quantized(lhs) and is_quantized(rhs) and is_quantized(result) .
    • (C3) storage_type(lhs) = storage_type(rhs) = storage_type(result) .
    • (C4) expressed_type(lhs) = expressed_type(rhs) = expressed_type(result) .
    • (C5) (is_per_axis_quantized(lhs) or is_per_axis_quantized(rhs)) = is_per_axis_quantized(result) .
    • (C6) If is_per_axis_quantized(lhs) , then quantization_dimension(lhs) = quantization_dimension(result) .
    • (C7) If is_per_axis_quantized(rhs) , then quantization_dimension(rhs) = quantization_dimension(result) .

উদাহরণ

// %lhs: [[1, 2], [3, 4]]
// %rhs: [[5, 6], [7, 8]]
%result = "stablehlo.add"(%lhs, %rhs) : (tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32>
// %result: [[6, 8], [10, 12]]

আরো উদাহরণ

after_all

শব্দার্থবিদ্যা

Ensures that the operations producing the inputs are executed before any operations that depend on result . Execution of this operation does nothing, it only exists to establish data dependencies from result to inputs .

ইনপুট

লেবেল নাম টাইপ
(I1) inputs variadic number of token

আউটপুট

নাম টাইপ
result token

উদাহরণ

// %input0: !stablehlo.token
// %input1: !stablehlo.token
%result = "stablehlo.after_all"(%input0, %input1) : (!stablehlo.token, !stablehlo.token) -> !stablehlo.token

আরো উদাহরণ

all_gather

শব্দার্থবিদ্যা

Within each process group in the StableHLO process grid, concatenates the values of the operands tensors from each process along all_gather_dim and produces results tensors.

The operation splits the StableHLO process grid into process_groups which is defined as follows:

  • cross_replica(replica_groups) if channel_id <= 0 and use_global_device_ids = false .
  • cross_replica_and_partition(replica_groups) if channel_id > 0 and use_global_device_ids = false .
  • flattened_ids(replica_groups) if channel_id > 0 and use_global_device_ids = true .

Afterwards, within each process_group :

  • operands...@receiver = [operand@sender for sender in process_group] for all receiver in process_group .
  • results...@process = concatenate(operands...@process, all_gather_dim) for all process in process_group .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operands variadic number of tensors or per-tensor quantized tensors (C1), (C6)
(I2) all_gather_dim constant of type si64 (C1), (C6)
(I3) replica_groups 2-dimensional tensor constant of type si64 (C2-C4)
(I4) channel_id constant of type si64 (C5)
(I5) use_global_device_ids constant of type i1 (C5)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
results variadic number of tensors or per-tensor quantized tensors (C6)

সীমাবদ্ধতা

  • (C1) 0 <= all_gather_dim < rank(operands...) .
  • (C2) is_unique(replica_groups) .
  • (C3) size(replica_groups) is defined as:
    • num_replicas if cross_replica is used.
    • num_replicas if cross_replica_and_partition is used.
    • num_processes if flattened_ids is used.
  • (C4) 0 <= replica_groups < size(replica_groups) .
  • (C5) If use_global_device_ids = true , then channel_id > 0 .
  • (C6) type(results...) = type(operands...) except:
    • dim(results..., all_gather_dim) = dim(operands..., all_gather_dim) * dim(process_groups, 1) .

উদাহরণ

// num_replicas: 2
// num_partitions: 1
// %operand0@(0, 0): [[1, 2], [3, 4]]
// %operand0@(1, 0): [[5, 6], [7, 8]]
// %operand1@(0, 0): [[11, 12], [13, 14]]
// %operand1@(1, 0): [[15, 16], [17, 18]]
%result:2 = "stablehlo.all_gather"(%operand0, %operand1) {
  all_gather_dim = 1 : i64,
  replica_groups = dense<[[0, 1]]> : tensor<1x2xi64>,
  // channel_id = 0
  channel_handle = #stablehlo.channel_handle<handle = 0, type = 0>
  // use_global_device_ids = false
} : (tensor<2x2xi64>, tensor<2x2xi64>) -> (tensor<2x4xi64>, tensor<2x4xi64>)
// %result0@(0, 0): [[1, 2, 5, 6], [3, 4, 7, 8]]
// %result0@(1, 0): [[1, 2, 5, 6], [3, 4, 7, 8]]
// %result1@(0, 0): [[11, 12, 15, 16], [13, 14, 17, 18]]
// %result1@(1, 0): [[11, 12, 15, 16], [13, 14, 17, 18]]

আরো উদাহরণ

all_reduce

শব্দার্থবিদ্যা

Within each process group in the StableHLO process grid, applies a reduction function computation to the values of the operands tensors from each process and produces results tensors.

The operation splits the StableHLO process grid into process_groups which is defined as follows:

  • cross_replica(replica_groups) if channel_id <= 0 and use_global_device_ids = false .
  • cross_replica_and_partition(replica_groups) if channel_id > 0 and use_global_device_ids = false .
  • flattened_ids(replica_groups) if channel_id > 0 and use_global_device_ids = true .

Afterwards, within each process_group :

  • results...@process[result_index] = exec(schedule) for some binary tree schedule where:
    • exec(node) = computation(exec(node.left), exec(node.right)) .
    • exec(leaf) = leaf.value .
  • schedule is an implementation-defined binary tree whose in-order traversal is to_destination_type(operands...@process_group...[result_index], type(func_inputs(computation)[0])) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operands variadic number of tensors or per-tensor quantized tensors (C5), (C6)
(I2) replica_groups variadic number of 1-dimensional tensor constants of type si64 (C1-C3)
(I3) channel_id constant of type si64 (C4)
(I4) use_global_device_ids constant of type i1 (C4)
(I5) computation ফাংশন (C5)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
results variadic number of tensors or per-tensor quantized tensors (C6-C7)

সীমাবদ্ধতা

  • (C1) is_unique(replica_groups) .
  • (C2) size(replica_groups) is defined as:
    • num_replicas if cross_replica is used.
    • num_replicas if cross_replica_and_partition is used.
    • num_processes if flattened_ids is used.
  • (C3) 0 <= replica_groups < size(replica_groups) .
  • (C4) If use_global_device_ids = true , then channel_id > 0 .
  • (C5) computation has type (tensor<E>, tensor<E>) -> (tensor<E>) where is_promotable(element_type(operand), E) .
  • (C6) shape(results...) = shape(operands...) .
  • (C7) element_type(results...) = E .

উদাহরণ

// num_replicas: 2
// num_partitions: 1
// %operand0@(0, 0): [1, 2, 3, 4]
// %operand0@(1, 0): [5, 6, 7, 8]
// %operand1@(0, 0): [9, 10, 11, 12]
// %operand1@(1, 0): [13, 14, 15, 16]
%result:2 = "stablehlo.all_reduce"(%operand0, %operand0) ({
  ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>):
    %0 = "stablehlo.add"(%arg0, %arg1) : (tensor<i64>, tensor<i64>) -> tensor<i64>
    "stablehlo.return"(%0) : (tensor<i64>) -> ()
}) {
  replica_groups = dense<[[0, 1]]> : tensor<1x2xi64>,
  // channel_id = 0
  channel_handle = #stablehlo.channel_handle<handle = 0, type = 0>
  // use_global_device_ids = false
} : (tensor<4xi64>, tensor<4xi64>) -> (tensor<4xi64>, tensor<4xi64>)
// %result0@(0, 0): [6, 8, 10, 12]
// %result0@(1, 0): [6, 8, 10, 12]
// %result1@(0, 0): [22, 24, 26, 28]
// %result1@(1, 0): [22, 24, 26, 28]

আরো উদাহরণ

all_to_all

শব্দার্থবিদ্যা

all_to_all

Within each process group in the StableHLO process grid, splits the values of the operands tensors along split_dimension into parts, scatters the split parts between the processes, concatenates the scattered parts along concat_dimension and produces results tensors. The operation splits the StableHLO process grid into process_groups which is defined as follows:

  • cross_replica(replica_groups) if channel_id <= 0 .
  • cross_partition(replica_groups) if channel_id > 0 .

Afterwards, within each process_group :

  • split_parts...@sender = split(operands...@sender, split_count, split_dimension) for all sender in process_group .
  • scattered_parts...@receiver = [split_parts...@sender[receiver_index] for sender in process_group] where receiver_index = process_group.index(receiver) .
  • results...@process = concatenate(scattered_parts...@process, concat_dimension) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operands variadic number of tensors or per-tensor quantized tensors (C1-C3), (C9)
(I2) split_dimension constant of type si64 (C1), (C2), (C9)
(I3) concat_dimension constant of type si64 (C3), (C9)
(I4) split_count constant of type si64 (C2), (C4), (C8), (C9)
(I5) replica_groups 2-dimensional tensor constant of type si64 (C5-C8)
(I6) channel_id constant of type si64

আউটপুট

নাম টাইপ সীমাবদ্ধতা
results variadic number of tensors or per-tensor quantized tensors (C9)

সীমাবদ্ধতা

  • (C1) 0 <= split_dimension < rank(operands...) .
  • (C2) dim(operands..., split_dimension) % split_count = 0 .
  • (C3) 0 <= concat_dimension < rank(operands...) .
  • (C4) 0 < split_count .
  • (C5) is_unique(replica_groups) .
  • (C6) size(replica_groups) is defined as:
    • num_replicas if cross_replica is used.
    • num_partitions if cross_partition is used.
  • (C7) 0 <= replica_groups < size(replica_groups) .
  • (C8) dim(replica_groups, 1) = split_count .
  • (C9) type(results...) = type(operands...) except, if split_dimension != concat_dimension :
    • dim(results..., split_dimension) = dim(operands..., split_dimension) / split_count .
    • dim(results..., concat_dimension) = dim(operands..., concat_dimension) * split_count .

উদাহরণ

// num_replicas: 2
// num_partitions: 1
// %operand1@(0, 0): [[1, 2, 3, 4],
//                    [5, 6, 7, 8]]
// %operand1@(1, 0): [[9, 10, 11, 12],
//                    [13, 14, 15, 16]]
// %operand2@(0, 0): [[17, 18, 19, 20],
//                    [21, 22, 23, 24]]
// %operand2@(1, 0): [[25, 26, 27, 28],
//                    [29, 30, 31, 32]]
%result:2 = "stablehlo.all_to_all"(%operand1, %operand2) {
  split_dimension = 1 : i64,
  concat_dimension = 0 : i64,
  split_count = 2 : i64,
  replica_groups = dense<[[0, 1]]> : tensor<1x2xi64>
  // channel_id = 0
} : (tensor<2x4xi64>, tensor<2x4xi64>) -> (tensor<4x2xi64>, tensor<4x2xi64>)
// %result#0@(0, 0): [[1, 2], [5, 6], [9, 10], [13, 14]]
// %result#0@(1, 0): [[3, 4], [7, 8], [11, 12], [15, 16]]
// %result#1@(0, 0): [[17, 18], [21, 22], [25, 26], [29, 30]]
// %result#1@(1, 0): [[19, 20], [23, 24], [27, 28], [31, 32]]

আরো উদাহরণ

এবং

শব্দার্থবিদ্যা

Performs element-wise AND of two tensors lhs and rhs and produces a result tensor. Depending on the element type, does the following:

  • For booleans: logical AND.
  • For integers: bitwise AND.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor of boolean or integer type (C1)
(I2) rhs tensor of boolean or integer type (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of boolean or integer type (C1)

সীমাবদ্ধতা

  • (C1) type(lhs) = type(rhs) = type(result) .

উদাহরণ

// %lhs: [[1, 2], [3, 4]]
// %rhs: [[5, 6], [7, 8]]
%result = "stablehlo.and"(%lhs, %rhs) : (tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32>
// %result: [[1, 2], [3, 0]]

আরো উদাহরণ

atan2

শব্দার্থবিদ্যা

Performs element-wise atan2 operation on lhs and rhs tensor and produces a result tensor. Depending on the element type, does the following:

  • For floats: atan2 from IEEE-754.
  • For complex numbers: complex atan2.
  • For quantized types: dequantize_op_quantize(atan2, lhs, rhs, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor of floating-point or complex type or per-tensor quantized tensor (C1)
(I2) rhs tensor of floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(lhs) = baseline_type(rhs) = baseline_type(result) .

উদাহরণ

// %lhs: [0.0, 1.0, -1.0]
// %rhs: [0.0, 0.0, 0.0]
%result = "stablehlo.atan2"(%lhs, %rhs) : (tensor<3xf64>, tensor<3xf64>) -> tensor<3xf64>
// %result: [0.0, 1.57079637, -1.57079637] // [0.0, pi/2, -pi/2]

আরো উদাহরণ

batch_norm_grad

শব্দার্থবিদ্যা

Computes gradients of several inputs of batch_norm_training backpropagating from grad_output , and produces grad_operand , grad_scale and grad_offset tensors. More formally, this operation can be expressed as a decomposition to existing StableHLO operations using Python syntax as follows:

def compute_sum(operand, feature_index):
  (sum,) = reduce(
      inputs=[operand],
      init_values=[constant(0, element_type(operand))],
      dimensions=[i for i in range(rank(operand)) if i != feature_index],
      body=lambda x, y: add(x, y))
  return sum

def compute_mean(operand, feature_index):
  sum = compute_sum(operand, feature_index)
  divisor = constant(size(operand) / dim(operand, feature_index),
                     element_type(operand))
  divisor_bcast = broadcast_in_dim(divisor, [], type(sum))
  return divide(sum, divisor_bcast)

def batch_norm_grad(operand, scale, mean, variance, grad_output, epsilon, feature_index):
  # Broadcast inputs to type(operand)
  scale_bcast = broadcast_in_dim(scale, [feature_index], type(operand))
  mean_bcast = broadcast_in_dim(mean, [feature_index], type(operand))
  variance_bcast = broadcast_in_dim(variance, [feature_index], type(operand))
  epsilon_bcast = broadcast_in_dim(constant(epsilon, element_type(operand)), [],
                                   type(operand))

  # Perform normalization using the provided `mean` and `variance`
  # Intermediate values will be useful for computing gradients
  centered_operand = subtract(operand, mean_bcast)
  stddev = sqrt(add(variance_bcast, epsilon_bcast))
  normalized_operand = divide(centered_operand, stddev)

  # Use the implementation from batchnorm_expander.cc in XLA
  # Temporary variables have exactly the same names as in the C++ code
  elements_per_feature = broadcast_in_dim(
      constant(divide(size(operand), dim(operand, feature_index)),
               element_type(grad_output)),
      [], type(operand))
  i1 = multiply(grad_output, elements_per_feature)
  i2 = broadcast_in_dim(
      compute_sum(grad_output, feature_index), [feature_index], type(operand))
  i3 = broadcast_in_dim(
      compute_sum(multiply(grad_output, centered_operand), feature_index),
      [feature_index], type(operand))
  i4 = multiply(i3, centered_operand)
  i5 = divide(i4, add(variance_bcast, epsilon_bcast))
  i6 = subtract(subtract(i1, i2), i5)

  grad_operand =
      multiply(divide(divide(scale_bcast, stddev), elements_per_feature), i6)
  grad_scale =
      compute_sum(multiply(grad_output, normalized_operand), feature_index)
  grad_offset = compute_sum(grad_output, feature_index)

  return grad_operand, grad_scale, grad_offset

For quantized types, performs dequantize_batch_norm_grad_or_training_quantize(lambda operand, scale, mean, variance, grad_output: batch_norm_grad(operand, scale, mean, variance, grad_output, epsilon, feature_index), operand, scale, mean, variance, grad_output, type(grad_operand), type(grad_scale), type(feature_index)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point type or per-tensor quantized tensor (C1-C3), (C5)
(I2) scale 1-dimensional tensor of floating-point or per-tensor quantized type (C2), (C4), (C5)
(I3) mean 1-dimensional tensor of floating-point or per-tensor quantized type (C2), (C4)
(I4) variance 1-dimensional tensor of floating-point or per-tensor quantized type (C2), (C4)
(I5) grad_output tensor of floating-point type or per-tensor quantized tensor (C2), (C3)
(I6) epsilon constant of type f32
(I7) feature_index constant of type si64 (C1), (C5)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
grad_operand tensor of floating-point type or per-tensor quantized tensor (C2), (C3)
grad_scale 1-dimensional tensor of floating-point or per-tensor quantized type (C2), (C4)
grad_offset 1-dimensional tensor of floating-point or per-tensor quantized type (C2), (C4)

সীমাবদ্ধতা

  • (C1) 0 <= feature_index < rank(operand) .
  • (C2) operand , scale , mean , variance , grad_output , grad_operand , grad_scale and grad_offset have the same baseline_element_type .
  • (C3) operand , grad_output and grad_operand have the same shape.
  • (C4) scale , mean , variance , grad_scale and grad_offset have the same shape.
  • (C5) size(scale) = dim(operand, feature_index) .

উদাহরণ

// %operand: [
//            [[1.0, 2.0], [3.0, 4.0]],
//            [[3.0, 4.0], [1.0, 2.0]]
//           ]
// %scale: [1.0, 1.0]
// %mean: [2.0, 3.0]
// %variance: [1.0, 1.0]
// %grad_output: [
//                [[0.1, 0.1], [0.1, 0.1]],
//                [[0.1, 0.1], [0.1, 0.1]]
//               ]
%grad_operand, %grad_scale, %grad_offset =
"stablehlo.batch_norm_grad"(%operand, %scale, %mean, %variance, %grad_output) {
  epsilon = 0.0 : f32,
  feature_index = 2 : i64
} : (tensor<2x2x2xf64>, tensor<2xf64>, tensor<2xf64>, tensor<2xf64>,
     tensor<2x2x2xf64>) -> (tensor<2x2x2xf64>, tensor<2xf64>, tensor<2xf64>)
// %grad_operand: [
//                 [[0.0, 0.0], [0.0, 0.0]],
//                 [[0.0, 0.0], [0.0, 0.0]]
//                ]
// %grad_scale:  [0.0, 0.0]
// %grad_offset: [0.4, 0.4]

batch_norm_inference

শব্দার্থবিদ্যা

Normalizes the operand tensor across all dimensions except for the feature_index dimension and produces a result tensor. More formally, this operation can be expressed as a decomposition to existing StableHLO operations using Python syntax as follows:

def batch_norm_inference(operand, scale, offset, mean, variance, epsilon, feature_index):
  # Broadcast inputs to shape(operand)
  scale_bcast = broadcast_in_dim(scale, [feature_index], type(operand))
  offset_bcast = broadcast_in_dim(offset, [feature_index], type(operand))
  mean_bcast = broadcast_in_dim(mean, [feature_index], type(operand))
  variance_bcast = broadcast_in_dim(variance, [feature_index], type(operand))
  epsilon_bcast = broadcast_in_dim(constant(epsilon, element_type(operand)), [],
                                   type(operand))

  # Perform normalization using the provided `mean` and `variance` instead of
  # computing them like `batch_norm_training` does.
  centered_operand = subtract(operand, mean_bcast)
  stddev = sqrt(add(variance_bcast, epsilon_bcast))
  normalized_operand = divide(centered_operand, stddev)
  return add(multiply(scale_bcast, normalized_operand), offset_bcast)

For quantized types, performs dequantize_op_quantize(lambda operand, scale, offset, mean, variance: batch_norm_inference(operand, scale, offset, mean, variance, epsilon, feature_index), operand, scale, offset, mean, variance, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point type or per-tensor quantized tensor (C1-C7)
(I2) scale 1-dimensional tensor of floating-point or per-tensor quantized type (C2), (C3)
(I3) offset 1-dimensional tensor of floating-point or per-tensor quantized type (C2), (C4)
(I4) mean 1-dimensional tensor of floating-point or per-tensor quantized type (C5)
(I5) variance 1-dimensional tensor of floating-point or per-tensor quantized type (C2), (C6)
(I6) epsilon constant of type f32
(I7) feature_index constant of type si64 (C1), (C3-C6)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point type or per-tensor quantized tensor (C2), (C7)

সীমাবদ্ধতা

  • (C1) 0 <= feature_index < rank(operand) .
  • (C2) operand , scale , offset , mean , variance and result have the same baseline_element_type .
  • (C3) size(scale) = dim(operand, feature_index) .
  • (C4) size(offset) = dim(operand, feature_index) .
  • (C5) size(mean) = dim(operand, feature_index) .
  • (C6) size(variance) = dim(operand, feature_index) .
  • (C7) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [
//            [[1.0, 2.0], [3.0, 4.0]],
//            [[3.0, 4.0], [1.0, 2.0]]
//           ]
// %scale: [1.0, 1.0]
// %offset: [1.0, 1.0]
// %mean: [2.0, 3.0]
// %variance: [1.0, 1.0]
%result = "stablehlo.batch_norm_inference"(%operand, %scale, %offset, %mean, %variance) {
  epsilon = 0.0 : f32,
  feature_index = 2 : i64
} : (tensor<2x2x2xf64>, tensor<2xf64>, tensor<2xf64>, tensor<2xf64>, tensor<2xf64>) -> tensor<2x2x2xf64>
// %result: [
//           [[0.0, 0.0], [2.0, 2.0]],
//           [[2.0, 2.0], [0.0, 0.0]]
//          ]

batch_norm_training

শব্দার্থবিদ্যা

Computes mean and variance across all dimensions except for the feature_index dimension and normalizes the operand tensor producing output , batch_mean and batch_var tensors. More formally, this operation can be expressed as a decomposition to existing StableHLO operations using Python syntax as follows:

def compute_mean(operand, feature_index):
  (sum,) = reduce(
      inputs=[operand],
      init_values=[constant(0, element_type(operand))],
      dimensions=[i for i in range(rank(operand)) if i != feature_index],
      body=lambda x, y: add(x, y))
  divisor = constant(size(operand) / dim(operand, feature_index),
                     element_type(operand))
  divisor_bcast = broadcast_in_dim(divisor, [], type(sum))
  return divide(sum, divisor_bcast)

def compute_variance(operand, feature_index):
  mean = compute_mean(operand, feature_index)
  mean_bcast = broadcast_in_dim(mean, [feature_index], type(operand))
  centered_operand = subtract(operand, mean_bcast)
  return compute_mean(mul(centered_operand, centered_operand), feature_index)

def batch_norm_training(operand, scale, offset, epsilon, feature_index):
  mean = compute_mean(operand, feature_index)
  variance = compute_variance(operand, feature_index)
  return batch_norm_inference(operand, scale, offset, mean, variance, epsilon,
                              feature_index),
         mean, variance

For quantized types, performs dequantize_batch_norm_grad_or_training_quantize(lambda operand, scale, offset: batch_norm_training(operand, scale, offset, epsilon, feature_index), operand, scale, offset, type(output), type(batch_mean), type(batch_var)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point type or per-tensor quantized tensor (C1)
(I2) scale 1-dimensional tensor of floating-point or per-tensor quantized (C2), (C3)
(I3) offset 1-dimensional tensor of floating-point or per-tensor quantized (C2), (C4)
(I4) epsilon constant of type f32 (C1), (C3-C6)
(I5) feature_index constant of type si64 (C1), (C3-C6)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
output tensor of floating-point type or per-tensor quantized tensor (C7)
batch_mean 1-dimensional tensor of floating-point or per-tensor quantized (C2), (C5)
batch_var 1-dimensional tensor of floating-point or per-tensor quantized (C2), (C6)

সীমাবদ্ধতা

  • (C1) 0 <= feature_index < rank(operand) .
  • (C2) operand , scale , offset , batch_mean , batch_var and output have the same baseline_element_type .
  • (C3) size(scale) = dim(operand, feature_index) .
  • (C4) size(offset) = dim(operand, feature_index) .
  • (C5) size(batch_mean) = dim(operand, feature_index) .
  • (C6) size(batch_var) = dim(operand, feature_index) .
  • (C7) baseline_type(output) = baseline_type(operand) .

উদাহরণ

// %operand: [
//            [[1.0, 2.0], [3.0, 4.0]],
//            [[3.0, 4.0], [1.0, 2.0]]
//           ]
// %scale: [1.0, 1.0]
// %offset: [1.0, 1.0]
%output, %batch_mean, %batch_var = "stablehlo.batch_norm_training"(%operand, %scale, %offset) {
  epsilon = 0.0 : f32,
  feature_index = 2 : i64
} : (tensor<2x2x2xf64>, tensor<2xf64>, tensor<2xf64>) ->
    (tensor<2x2x2xf64>, tensor<2xf64>, tensor<2xf64>)
// %output: [
//           [[0.0, 0.0], [2.0, 2.0]],
//           [[2.0, 2.0], [0.0, 0.0]]
//          ]
// %batch_mean: [2.0, 3.0]
// %batch_var: [1.0, 1.0]

bitcast_convert

শব্দার্থবিদ্যা

Performs a bitcast operation on operand tensor and produces a result tensor where the bits of the entire operand tensor are reinterpreted using the type of the result tensor.

More formally, given E = element_type(operand) , E' = element_type(result) , and R = rank(operand) :

  • If num_bits(E') < num_bits(E) , bits(result[i0, ..., iR-1, :]) = bits(operand[i0, ..., iR-1]) .
  • If num_bits(E') > num_bits(E) , bits(result[i0, ..., iR-2]) = bits(operand[i0, ..., iR-2, :]) .
  • If num_bits(E') = num_bits(E) , bits(result[i0, ..., iR-1]) = bits(operand[i0, ..., iR-1]) .

bits returns in-memory representation of a given value, and its behavior is implementation-defined because the exact representation of tensors is implementation-defined, and the exact representation of element types is implementation-defined as well.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or quantized tensor (C1-C2)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or quantized tensor (C1-C2)

সীমাবদ্ধতা

  • (C1) Given E = is_quantized(operand) ? storage_type(operand) : element_type(operand) , E' = is_quantized(result) ? storage_type(result) : element_type(result) , and R = rank(operand) :
    • If num_bits(E') = num_bits(E) , shape(result) = shape(operand) .
    • If num_bits(E') < num_bits(E) :
    • rank(result) = R + 1 .
    • dim(result, i) = dim(operand, i) for all 0 <= i < R .
    • dim(result, R) * num_bits(E') = num_bits(E) .
    • If num_bits(E') > num_bits(E) :
    • rank(result) = R - 1 .
    • dim(result, i) = dim(operand, i) for all 0 <= i < R .
    • dim(operand, R - 1) * num_bits(E) = num_bits(E') .
  • (C2) If is_complex(operand) or is_complex(result) , then is_complex(operand) and is_complex(result) .

উদাহরণ

// %operand: 0x0123456789ABCDEF
%result = "stablehlo.bitcast_convert"(%operand) : (tensor<f64>) -> tensor<4xf16>
// %result: [0xCDEF, 0x89AB, 0x4567, 0x0123] // little-endian representation

আরো উদাহরণ

broadcast_in_dim

শব্দার্থবিদ্যা

Expands the dimensions and/or rank of an input tensor by duplicating the data in the operand tensor and produces a result tensor. More formally, result[result_index] = operand[operand_index] where for all d in axes(operand) :

  • operand_index[d] = 0 if dim(operand, d) = 1 .
  • operand_index[d] = result_index[broadcast_dimensions[d]] otherwise.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or quantized tensor (C1-C2), (C5-C6)
(I2) broadcast_dimensions 1-dimensional tensor constant of type si64 (C2-C6)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or quantized tensor (C1), (C3), (C5-C6)

সীমাবদ্ধতা

  • (C1) element_type(result) is given by:
    • element_type(operand) , if !is_per_axis_quantized(operand) .
    • element_type(operand) except that quantization_dimension(operand) , scales(operand) , and zero_points(operand) may differ from quantization_dimension(result) , scales(result) , and zero_points(result) resp., otherwise.
  • (C2) size(broadcast_dimensions) = rank(operand) .
  • (C3) 0 <= broadcast_dimensions < rank(result) .
  • (C4) is_unique(broadcast_dimensions) .
  • (C5) For all d in axes(operand) :
    • dim(operand, d) = 1 or
    • dim(operand, d) = dim(result, broadcast_dimensions[d]) .
  • (C6) If is_per_axis_quantized(result) :
    • quantization_dimension(result) = broadcast_dimensions[quantization_dimension(operand)] .
    • If dim(operand, quantization_dimension(operand)) = 1 , then scales(result)[i] = scales(operand)[0] and zero_points(result)[i] = zero_points(operand)[0] for i in range(dim(result, quantization_dimension(result))) .

উদাহরণ

// %operand: [
//            [1, 2, 3]
//           ]
%result = "stablehlo.broadcast_in_dim"(%operand) {
  broadcast_dimensions = array<i64: 2, 1>
} : (tensor<1x3xi32>) -> tensor<2x3x2xi32>
// %result: [
//            [
//             [1, 1],
//             [2, 2],
//             [3, 3]
//            ],
//            [
//             [1, 1],
//             [2, 2],
//             [3, 3]
//            ]
//          ]

আরো উদাহরণ

মামলা

শব্দার্থবিদ্যা

Produces the output from executing exactly one function from branches depending on the value of index . More formally, result = selected_branch() where:

  • selected_branch = branches[index] if 0 <= index < size(branches) .
  • selected_branch = branches[-1] otherwise.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) index 0-dimensional tensor of type si32
(I2) branches variadic number of functions (C1-C4)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
results variadic number of tensors, quantized tensors or tokens (C4)

সীমাবদ্ধতা

  • (C1) 0 < size(branches) .
  • (C2) input_types(branches...) = [] .
  • (C3) same(output_types(branches...)) .
  • (C4) type(results...) = output_types(branches[0]) .

উদাহরণ

// %index: -1
// %result_branch0: [0, 0]
// %result_branch1: [1, 1]
%result0, %result1 = "stablehlo.case"(%index) ({
  "stablehlo.return"(%result_branch0, %result_branch0) : (tensor<2xi64>, tensor<2xi64>) -> ()
}, {
  "stablehlo.return"(%result_branch1, %result_branch1) : (tensor<2xi64>, tensor<2xi64>) -> ()
}) : (tensor<i32>) -> (tensor<2xi64>, tensor<2xi64>)
// %result0: [1, 1]
// %result1: [1, 1]

আরো উদাহরণ

cbrt

শব্দার্থবিদ্যা

Performs element-wise cubic root operation on operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For floats: rootn(x, 3) from IEEE-754.
  • For complex numbers: complex cubic root.
  • For quantized types: dequantize_op_quantize(cbrt, operand, type(result))

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [0.0, 1.0, 8.0, 27.0]
%result = "stablehlo.cbrt"(%operand) : (tensor<4xf64>) -> tensor<4xf64>
// %result: [0.0, 1.0, 2.0, 3.0]

আরো উদাহরণ

ছাদ

শব্দার্থবিদ্যা

Performs element-wise ceil of operand tensor and produces a result tensor. Implements the roundToIntegralTowardPositive operation from the IEEE-754 specification. For quantized types, performs dequantize_op_quantize(ceil, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [-0.8166, -0.2530, 0.2530, 0.8166, 2.0]
%result = "stablehlo.ceil"(%operand) : (tensor<5xf32>) -> tensor<5xf32>
// %result: [-0.0, -0.0, 1.0, 1.0, 2.0]

আরো উদাহরণ

cholesky

শব্দার্থবিদ্যা

Computes the Cholesky decomposition of a batch of matrices.

More formally, for all i in index_space(result) , result[i0, ..., iR-3, :, :] is a Cholesky decomposition of a[i0, ..., iR-3, :, :] , in the form of either of a lower-triangular (if lower is true ) or upper-triangular (if lower is false ) matrix. The output values in the opposite triangle, ie the strict upper triangle or strict lower triangle correspondingly, are implementation-defined.

If there exists i where the input matrix is not an Hermitian positive-definite matrix, then the behavior is undefined.

For quantized types, performs dequantize_op_quantize(lambda operand: cholesky(operand, lower), a, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) a tensor of floating-point or complex type or per-tensor quantized tensor (C1-C3)
(I2) lower 0-dimensional tensor constant of type i1

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(a) = baseline_type(result) .
  • (C2) 2 <= rank(a) .
  • (C3) dim(a, -2) = dim(a, -1) .

উদাহরণ

// %a: [
//      [1.0, 2.0, 3.0],
//      [2.0, 20.0, 26.0],
//      [3.0, 26.0, 70.0]
//     ]
%result = "stablehlo.cholesky"(%a) {
  lower = true
} : (tensor<3x3xf32>) -> tensor<3x3xf64>
// %result: [
//           [1.0, 0.0, 0.0],
//           [2.0, 4.0, 0.0],
//           [3.0, 5.0, 6.0]
//          ]

clamp

শব্দার্থবিদ্যা

Clamps every element of the operand tensor between a minimum and maximum value and produces a result tensor. More formally, result[result_index] = minimum(maximum(operand[result_index], min_element), max_element) , where min_element = rank(min) = 0 ? min[] : min[result_index] , max_element = rank(max) = 0 ? max[] : max[result_index] . For quantized types, performs dequantize_op_quantize(clamp, min, operand, max, type(result)) .

Imposing an ordering on complex numbers involves surprising semantics, so in the future we are planning to remove support for complex numbers for this operation ( #560 ).

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) min tensor or per-tensor quantized tensor (C1), (C3)
(I2) operand tensor or per-tensor quantized tensor (C1-C4)
(I3) max tensor or per-tensor quantized tensor (C2), (C3)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C4)

সীমাবদ্ধতা

  • (C1) rank(min) = 0 or shape(min) = shape(operand) .
  • (C2) rank(max) = 0 or shape(max) = shape(operand) .
  • (C3) baseline_element_type(min) = baseline_element_type(operand) = baseline_element_type(max) .
  • (C4) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %min: [5, 10, 15]
// %operand: [3, 13, 23]
// %max: [10, 15, 20]
%result = "stablehlo.clamp"(%min, %operand, %max) : (tensor<3xi32>, tensor<3xi32>, tensor<3xi32>) -> tensor<3xi32>
// %result: [5, 13, 20]

আরো উদাহরণ

collective_broadcast

শব্দার্থবিদ্যা

Within each process group in the StableHLO process grid, send the value of the operand tensor from the source process to the target processes and produce a result tensor.

The operation splits the StableHLO process grid into process_groups which is defined as follows:

  • cross_replica(replica_groups) if channel_id <= 0 .
  • cross_partition(replica_groups) if channel_id > 0 .

Afterwards, result@process is given by:

  • operand@process_groups[i, 0] if there exists an i such that the process is in process_groups[i] .
  • broadcast_in_dim(constant(is_quantized(result) ? quantize(0, element_type(result)) : 0, element_type(result)), [], type(result)) otherwise.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or per-tensor quantized tensor (C3)
(I2) replica_groups variadic number of 1-dimensional tensor constants of type si64 (C1), (C2)
(I3) channel_id constant of type si64

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C3)

সীমাবদ্ধতা

  • (C1) is_unique(replica_groups) .
  • (C2) 0 <= replica_groups < N where N is defined as:
    • num_replicas if cross_replica is used.
    • num_partitions if cross_partition is used.
  • (C3) type(result) = type(operand) .

উদাহরণ

// num_replicas: 4
// num_partitions: 1
// %operand@(0, 0): [[1, 2]]
// %operand@(1, 0): [[3, 4]]
// %operand@(2, 0): [[5, 6]]
// %operand@(3, 0): [[7, 8]]
%result = "stablehlo.collective_broadcast"(%operand) {
  replica_groups = dense<[[2, 1]]> : tensor<1x2xi64>,
  channel_handle = #stablehlo.channel_handle<handle = 0, type = 0>
} : (tensor1x2xi64>) -> tensor<1x2xi64>
// %result@(0, 0): [[0, 0]]
// %result@(1, 0): [[5, 6]]
// %result@(2, 0): [[5, 6]]
// %result@(3, 0): [[0, 0]]

collective_permute

শব্দার্থবিদ্যা

Within each process group in the StableHLO process grid, sends the value of the operand tensor from the source process to the target process and produces a result tensor.

The operation splits the StableHLO process grid into process_groups which is defined as follows:

  • cross_replica(source_target_pairs) if channel_id <= 0 .
  • cross_partition(source_target_pairs) if channel_id > 0 .

Afterwards, result@process is given by:

  • operand@process_groups[i, 0] , if there exists an i such that process_groups[i, 1] = process .
  • broadcast_in_dim(constant(is_quantized(result) ? quantize(0, element_type(result)) : 0, element_type(result)), [], type(result)) otherwise.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or per-tensor quantized tensor (C5)
(I2) source_target_pairs 2-dimensional tensor constant of type si64 (C1-C4)
(I3) channel_id constant of type si64

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) dim(source_target_pairs, 1) = 2 .
  • (C2) is_unique(source_target_pairs[:, 0]) .
  • (C3) is_unique(source_target_pairs[:, 1]) .
  • (C4) 0 <= source_target_pairs < N , where N is defined as:
    • num_replicas if cross_replica is used.
    • num_partitions if cross_partition is used.
  • (C5) type(result) = type(operand) .

উদাহরণ

// num_replicas: 3
// num_partitions: 1
// %operand@(0, 0): [[1, 2], [3, 4]]
// %operand@(1, 0): [[5, 6], [7, 8]]
// %operand@(2, 0): [[9, 10], [11, 12]]
%result = "stablehlo.collective_permute"(%operand) {
  source_target_pairs = dense<[[0, 1], [1, 2]]> : tensor<2x2xi64>,
  channel_handle = #stablehlo.channel_handle<handle = 0, type = 0>
} : (tensor<2x2xi64>) -> tensor<2x2xi64>
//
// %result@(0, 0): [[0, 0], [0, 0]]
// %result@(1, 0): [[1, 2], [3, 4]]
// %result@(2, 0): [[5, 6], [7, 8]]

আরো উদাহরণ

তুলনা

শব্দার্থবিদ্যা

Performs element-wise comparison of lhs and rhs tensors according to comparison_direction and compare_type , and produces a result tensor.

The values of comparison_direction and compare_type have the following semantics:

For boolean and integer element types:

  • EQ : lhs = rhs .
  • NE : lhs != rhs .
  • GE : lhs >= rhs .
  • GT : lhs > rhs .
  • LE : lhs <= rhs .
  • LT : lhs < rhs .

For floating-point element types with compare_type = FLOAT , the op implements the following IEEE-754 operations:

  • EQ : compareQuietEqual .
  • NE : compareQuietNotEqual .
  • GE : compareQuietGreaterEqual .
  • GT : compareQuietGreater .
  • LE : compareQuietLessEqual .
  • LT : compareQuietLess .

For floating-point element types with compare_type = TOTALORDER , the op uses the combination of totalOrder and compareQuietEqual operations from IEEE-754.

For complex element types, lexicographic comparison of (real, imag) pairs is performed using the provided comparison_direction and compare_type . Imposing an ordering on complex numbers involves surprising semantics, so in the future we are planning to remove support for complex numbers when comparison_direction is GE , GT , LE or LT ( #560 ).

For quantized types. performs dequantize_compare(lhs, rhs, comparison_direction) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor or per-tensor quantized tensor (C1-C3)
(I2) rhs tensor or per-tensor quantized tensor (C1-C2)
(I3) comparison_direction enum of EQ , NE , GE , GT , LE , and LT
(I4) compare_type enum of FLOAT , TOTALORDER , SIGNED , and UNSIGNED (C3)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of boolean type (C2)

সীমাবদ্ধতা

  • (C1) baseline_element_type(lhs) = baseline_element_type(rhs) .
  • (C2) shape(lhs) = shape(rhs) = shape(result) .
  • (C3) compare_type is defined as:
    • SIGNED if is_signed_integer(element_type(lhs)) .
    • UNSIGNED if is_unsigned_integer(element_type(lhs)) or is_boolean(element_type(lhs)) .
    • FLOAT or TOTALORDER if is_float(element_type(lhs)) .
    • FLOAT if is_complex(element_type(lhs)) .

উদাহরণ

// %lhs: [1.0, 3.0]
// %rhs: [1.1, 2.9]
%result = "stablehlo.compare"(%lhs, %rhs) {
  comparison_direction = #stablehlo<comparison_direction LT>,
  compare_type = #stablehlo<comparison_type FLOAT>
} : (tensor<2xf32>, tensor<2xf32>) -> tensor<2xi1>
// %result: [true, false]

আরো উদাহরণ

জটিল

শব্দার্থবিদ্যা

Performs element-wise conversion to a complex value from a pair of real and imaginary values, lhs and rhs , and produces a result tensor.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor of type f32 or f64 (C1-C3)
(I2) rhs tensor of type f32 or f64 (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of complex type (C2), (C3)

সীমাবদ্ধতা

  • (C1) type(lhs) = type(rhs) .
  • (C2) shape(result) = shape(lhs) .
  • (C3) element_type(result) has type complex<E> where E = element_type(lhs) .

উদাহরণ

// %lhs: [1.0, 3.0]
// %rhs: [2.0, 4.0]
%result = "stablehlo.complex"(%lhs, %rhs) : (tensor<2xf64>, tensor<2xf64>) -> tensor<2xcomplex<f64>>
// %result: [(1.0, 2.0), (3.0, 4.0)]

আরো উদাহরণ

যৌগিক

শব্দার্থবিদ্যা

Encapsulates an operation made up (composed) of other StableHLO operations, taking inputs and composite_attributes and producing results . The semantics of the op are implemented by the decomposition attribute. The composite op can be replaced with its decomposition without changing program semantics. In cases where inlining the decomposition does not provide the same op semantics, prefer using custom_call .

The version field (defaults to 0 ) is used to denote when a composite's semantics change.

ইনপুট

লেবেল নাম টাইপ
(I1) inputs variadic number of values
(I2) name constant of type string
(I3) composite_attributes attribute dictionary
(I4) decomposition constant of type string
(I5) version constant of type si32

আউটপুট

নাম টাইপ
results variadic number of values

সীমাবদ্ধতা

  • (C1) is_namespaced_op_name(name)
  • (C2) is_defined_in_parent_scope(decomposition)
  • (C3) types(inputs...) == input_types(decomposition)
  • (C4) types(results...) == output_types(decomposition)

উদাহরণ

%results = "stablehlo.composite"(%input0, %input1) {
  name = "my_namespace.my_op",
  composite_attributes = {
    my_attribute = "my_value"
  },
  decomposition = @my_op,
  version = 1 : i32
} : (tensor<f32>, tensor<f32>) -> tensor<f32>

আরো উদাহরণ

সংযুক্ত করা

শব্দার্থবিদ্যা

Concatenates inputs along dimension dimension in the same order as the given arguments and produces a result tensor. More formally, result[i0, ..., id, ..., iR-1] = inputs[k][i0, ..., kd, ..., iR-1] , where:

  1. id = d0 + ... + dk-1 + kd .
  2. d is equal to dimension , and d0 , ... are d th dimension sizes of inputs .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) inputs variadic number of tensors or per-tensor quantized tensors (C1-C6)
(I2) dimension constant of type si64 (C2), (C4), (C6)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C5-C6)

সীমাবদ্ধতা

  • (C1) same(element_type(inputs...)) .
  • (C2) same(shape(inputs...)) except for dim(inputs..., dimension) .
  • (C3) 0 < size(inputs) .
  • (C4) 0 <= dimension < rank(inputs[0]) .
  • (C5) element_type(result) = element_type(inputs[0]) .
  • (C6) shape(result) = shape(inputs[0]) except for:
    • dim(result, dimension) = dim(inputs[0], dimension) + ... .

উদাহরণ

// %input0: [[1, 2], [3, 4], [5, 6]]
// %input1: [[7, 8]]
%result = "stablehlo.concatenate"(%input0, %input1) {
  dimension = 0 : i64
} : (tensor<3x2xi64>, tensor<1x2xi64>) -> tensor<4x2xi64>
// %result: [[1, 2], [3, 4], [5, 6], [7, 8]]

আরো উদাহরণ

ধ্রুবক

শব্দার্থবিদ্যা

Produces an output tensor from a constant value .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) value ধ্রুবক (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
output tensor or quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) type(value) = type(output) .

উদাহরণ

%output = "stablehlo.constant"() {
  value = dense<[[0.0, 1.0], [2.0, 3.0]]> : tensor<2x2xf32>
} : () -> tensor<2x2xf32>
// %output: [[0.0, 1.0], [2.0, 3.0]]

আরো উদাহরণ

রূপান্তর

শব্দার্থবিদ্যা

Performs an element-wise conversion from one element type to another on operand tensor and produces a result tensor.

For boolean-to-any-supported-type conversions, the value false is converted to zero, and the value true is converted to one. For any-supported-type-to-boolean conversions, a zero value is converted to false , and non-zero values are converted to true . See below for how this work for complex types.

For conversions involving integer-to-integer , integer-to-floating-point or floating-point-to-floating-point , if the source value can be exactly represented in the destination type, the result value is that exact representation. Otherwise, the behavior is TBD ( #180 ).

For conversions involving floating-point-to-integer , the fractional part is truncated. If the truncated value cannot be represented in the destination type, the behavior is TBD ( #180 ).

Conversion involving complex-to-complex follow the same behavior of floating-point-to-floating-point conversions for converting real and imaginary parts.

For complex-to-any-other-type and any-other-type-to-complex conversions, the source imaginary value is ignored or the destination imaginary value is zeroed, respectively. The conversion of the real part follows the floating-point conversions.

In principle, this operation could express dequantization (conversion from quantized tensors to regular tensors), quantization (conversion from regular tensors to quantized tensors) and requantization (conversion between quantized tensors), but at the moment we have dedicated operations for that - uniform_dequantize for the first use case and uniform_quantize for the second and the third use cases. In the future, these two ops may be merged into convert ( #1576 ).

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand টেনসর (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result টেনসর (C1)

সীমাবদ্ধতা

  • (C1) shape(operand) = shape(result) .

উদাহরণ

// %operand: [-1, 0, 1]
%result = "stablehlo.convert"(%operand) : (tensor<3xi64>) -> tensor<3xcomplex<f64>>
// %result: [(-1.0, 0.0), (0.0, 0.0), (1.0, 0.0)]

আরো উদাহরণ

আবর্তন

শব্দার্থবিদ্যা

Computes dot products between windows of lhs and slices of rhs and produces result . The following diagram shows how elements in result are computed from lhs and rhs using a concrete example.

আবর্তন

More formally, consider the following reframing of the inputs in terms of lhs in order to be able to express windows of lhs :

  • lhs_window_dimensions = lhs_shape(dim(lhs, input_batch_dimension), dim(rhs, kernel_spatial_dimensions), dim(lhs, input_feature_dimension)) .
  • lhs_window_strides = lhs_shape(1, window_strides, 1) .
  • lhs_padding = lhs_shape([0, 0], padding, [0, 0]) .
  • lhs_base_dilations = lhs_shape(1, lhs_dilation, 1) .
  • lhs_window_dilations = lhs_shape(1, rhs_dilation, 1) .

This reframing uses the following helper functions:

  • lhs_shape(n, hw, c) = permute([n] + hw + [c], [input_batch_dimension] + input_spatial_dimensions + [input_feature_dimension]) .
  • result_shape(n1, hw, c1) = permute([n1] + hw + [c1], [output_batch_dimension] + output_spatial_dimensions + [output_feature_dimension]) .
  • permute([j0, j1, ..., jR-1], permutation) = [i0, i1, ..., iR-1] where j[d] = i[permutation[d]] .

If feature_group_count = 1 and batch_group_count = 1 , then for all output_spatial_index in index_space(dim(result, output_spatial_dimensions...)) , result[result_shape(:, output_spatial_index, :)] = dot_product where:

  • padding_value = constant(0, element_type(lhs)) .
  • padded_lhs = pad(lhs, padding_value, lhs_padding[:, 0], lhs_padding[:, 1], lhs_base_dilations - 1) .
  • lhs_window_start = lhs_shape(0, output_spatial_index, 0) * lhs_window_strides .
  • lhs_window = slice(padded_lhs, lhs_window_start, lhs_window_start + lhs_window_dimensions, lhs_window_dilations) .
  • reversed_lhs_window = reverse(lhs_window, [input_spatial_dimensions[dim] for dim in range(size(window_reversal)) if window_reversal[dim] = true]) . This feature appears to be unused, so in the future we are planning to remove it ( #1181 ).
  • dot_product = dot_general(reversed_lhs_window, rhs, lhs_batching_dimensions=[], lhs_contracting_dimensions=input_spatial_dimensions + [input_feature_dimension], rhs_batching_dimensions=[], rhs_contracting_dimensions=kernel_spatial_dimensions + [kernel_input_feature_dimension]) .

If feature_group_count > 1 :

  • lhses = split(lhs, feature_group_count, input_feature_dimension) .
  • rhses = split(rhs, feature_group_count, kernel_output_feature_dimension) .
  • results... = convolution(lhses..., rhses..., ..., feature_group_count=1, ...) .
  • result = concatenate(results, output_feature_dimension) .

If batch_group_count > 1 :

  • lhses = split(lhs, batch_group_count, input_batch_dimension) .
  • rhses = split(rhs, batch_group_count, kernel_output_feature_dimension) .
  • results... = convolution(lhses..., rhses..., ..., batch_group_count=1, ...) .
  • result = concatenate(results, output_feature_dimension) .

For quantized types, performs dequantize_op_quantize( lambda lhs, rhs: convolution(lhs, rhs, window_strides, padding, lhs_dilation, rhs_dilation, window_reversal, input_batch_dimension, input_feature_dimension, input_spatial_dimensions, kernel_input_feature_dimension, kernel_output_feature_dimension, kernel_spatial_dimensions, output_batch_dimension, output_feature_dimension, output_spatial_dimensions, feature_group_count, batch_group_count, precision_config), lhs, rhs, type(result)) .

For hybrid quantized types, performs hybrid_dequantize_then_op( lambda lhs, rhs: convolution(lhs, rhs, window_strides, padding, lhs_dilation, rhs_dilation, window_reversal, input_batch_dimension, input_feature_dimension, input_spatial_dimensions, kernel_input_feature_dimension, kernel_output_feature_dimension, kernel_spatial_dimensions, output_batch_dimension, output_feature_dimension, output_spatial_dimensions, feature_group_count, batch_group_count, precision_config), lhs, rhs) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor or per-tensor quantized tensor (C1), (C10-C11), (C14) (C25), (C27-C28), (C31-C32), (C34)
(I2) rhs tensor or quantized tensor (C1), (C14-C16), (C25), (C27-C29), (C31-C34)
(I3) window_strides 1-dimensional tensor constant of type si64 (C2-C3), (C25)
(I4) padding 2-dimensional tensor constant of type si64 (C4), (C25)
(I5) lhs_dilation 1-dimensional tensor constant of type si64 (C5-C6), (C25)
(I6) rhs_dilation 1-dimensional tensor constant of type si64 (C7-C8), (C25)
(I7) window_reversal 1-dimensional tensor constant of type i1 (C9)
(I8) input_batch_dimension constant of type si64 (C10), (C13), (C25)
(I9) input_feature_dimension constant of type si64 (C11), (C13-C14)
(I10) input_spatial_dimensions 1-dimensional tensor constant of type si64 (C12), (C13), (C25)
(I11) kernel_input_feature_dimension constant of type si64 (C14), (C18)
(I12) kernel_output_feature_dimension constant of type si64 (C15-C16), (C18), (C25), (C29)
(I13) kernel_spatial_dimensions 1-dimensional tensor constant of type si64 (C17-C18), (C25)
(I14) output_batch_dimension constant of type si64 (C20), (C25)
(I15) output_feature_dimension constant of type si64 (C20), (C25), (C30)
(I16) output_spatial_dimensions 1-dimensional tensor constant of type si64 (C19-C20), (C25)
(I17) feature_group_count constant of type si64 (C11), (C14), (C16), (C21), (C23)
(I18) batch_group_count constant of type si64 (C10), (C15), (C22), (C23), (C25)
(I19) precision_config variadic number of enums of DEFAULT , HIGH , and HIGHEST (C24)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or quantized tensor (C25-C28), (C30), (C32-34)

সীমাবদ্ধতা

  • (C1) N = rank(lhs) = rank(rhs) .
  • (C2) size(window_strides) = N - 2 .
  • (C3) 0 < window_strides .
  • (C4) shape(padding) = [N - 2, 2] .
  • (C5) size(lhs_dilation) = N - 2 .
  • (C6) 0 < lhs_dilation .
  • (C7) size(rhs_dilation) = N - 2 .
  • (C8) 0 < rhs_dilation .
  • (C9) size(window_reversal) = N - 2 .
  • (C10) dim(lhs, input_batch_dimension) % batch_group_count = 0 .
  • (C11) dim(lhs, input_feature_dimension) % feature_group_count = 0 .
  • (C12) size(input_spatial_dimensions) = N - 2 .
  • (C13) Given input_dimensions = [input_batch_dimension] + input_spatial_dimensions + [input_feature_dimension] :
    • is_unique(input_dimensions) .
    • 0 <= input_dimensions < N .
  • (C14) dim(rhs, kernel_input_feature_dimension) = dim(lhs, input_feature_dimension) / feature_group_count .
  • (C15) dim(rhs, kernel_output_feature_dimension) % batch_group_count = 0 .
  • (C16) dim(rhs, kernel_output_feature_dimension) % feature_group_count = 0 .
  • (C17) size(kernel_spatial_dimensions) = N - 2 .
  • (C18) Given kernel_dimensions = kernel_spatial_dimensions + [kernel_input_feature_dimension] + [kernel_output_feature_dimension] :
    • is_unique(kernel_dimensions) .
    • 0 <= kernel_dimensions < N .
  • (C19) size(output_spatial_dimensions) = N - 2 .
  • (C20) Given output_dimensions = [output_batch_dimension] + output_spatial_dimensions + [output_feature_dimension] :
    • is_unique(output_dimensions) .
    • 0 <= output_dimensions < N .
  • (C21) 0 < feature_group_count .
  • (C22) 0 < batch_group_count .
  • (C23) feature_group_count = 1 or batch_group_count = 1 .
  • (C24) size(precision_config) = 2 .
  • (C25) dim(result, result_dim) is defined as:
    • dim(lhs, input_batch_dimension) / batch_group_count if result_dim = output_batch_dimension .
    • dim(rhs, kernel_output_feature_dimension) if result_dim = output_feature_dimension .
    • num_windows otherwise, where:
    • output_spatial_dimensions[spatial_dim] = result_dim .
    • lhs_dim = input_spatial_dimensions[spatial_dim] .
    • rhs_dim = kernel_spatial_dimensions[spatial_dim] .
    • dilated_input_shape[lhs_dim] = dim(lhs, lhs_dim) = 0 ? 0 : (dim(lhs, lhs_dim) - 1) * lhs_dilation[spatial_dim] + 1 .
    • padded_input_shape[lhs_dim] = padding[spatial_dim, 0] + dilated_input_shape[lhs_dim] + padding[spatial_dim, 1] .
    • dilated_window_shape[lhs_dim] = dim(rhs, rhs_dim) = 0 ? 0 : (dim(rhs, rhs_dim) - 1) * rhs_dilation[spatial_dim] + 1 .
    • is_empty_window[lhs_dim] = padded_input_shape[lhs_dim] = 0 || dilated_window_shape[lhs_dim] > padded_input_shape[lhs_dim] .
    • num_windows = is_empty_window[lhs_dim] ? 0 : floor((padded_input_shape[lhs_dim] - dilated_window_shape[lhs_dim]) / window_strides[spatial_dim]) + 1 .
  • (C26) rank(result) = N .
  • If the operation uses non-quantized tensors:
    • (C27) element_type(lhs) = element_type(rhs) = element_type(result) .
  • If the operation uses quantized tensors:
    • (C28) is_quantized(lhs) = is_quantized(result) and is_quantized(rhs) .
    • (C29) If is_per_axis_quantized(rhs) , then quantization_dimension(rhs) = kernel_output_feature_dimension .
    • (C30) If is_per_axis_quantized(result) , then quantization_dimension(result) = output_feature_dimension .
    • If is_quantized(lhs) :
    • (C31) storage_type(lhs) = storage_type(rhs) .
    • (C32) expressed_type(lhs) = expressed_type(rhs) = expressed_type(result) .
    • (C33) If is_per_tensor_quantized(rhs) , then is_per_tensor_quantized(result) .
    • If !is_quantized(lhs) :
    • (C34) element_type(lhs) = expressed_type(rhs) = element_type(result) .

উদাহরণ

// %lhs: [[
//        [
//          [1], [2], [5], [6]
//        ],
//        [
//          [3], [4], [7], [8]
//        ],
//        [
//          [10], [11], [14], [15]
//        ],
//        [
//          [12], [13], [16], [17]
//        ]
//      ]]
//
// %rhs: [
//        [[[1]], [[1]], [[1]]],
//        [[[1]], [[1]], [[1]]],
//        [[[1]], [[1]], [[1]]]
//       ]
%result = "stablehlo.convolution"(%lhs, %rhs) {
  window_strides = array<i64: 4, 4>,
  padding = dense<0> : tensor<2x2xi64>,
  lhs_dilation = array<i64: 2, 2>,
  rhs_dilation = array<i64: 1, 1>,
  window_reversal = array<i1: false, false>,
  // In the StableHLO dialect, dimension numbers are encoded via:
  // `[<input dimensions>]x[<kernel dimensions>]->[output dimensions]`.
  // "b" is batch dimension, "f" is feature dimension,
  // "i" is input feature dimension, "o" is output feature dimension,
  // "0/1/etc" are spatial dimensions.
  dimension_numbers = #stablehlo.conv<[b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f]>,
  batch_group_count = 1 : i64,
  feature_group_count = 1 : i64,
  precision_config = [#stablehlo<precision DEFAULT>, #stablehlo<precision DEFAULT>]
} : (tensor<1x4x4x1xi64>, tensor<3x3x1x1xi64>) -> tensor<1x2x2x1xi64>
// %result: [[
//            [[10], [26]],
//            [[46], [62]]
//          ]]

আরো উদাহরণ

কোসাইন

শব্দার্থবিদ্যা

Performs element-wise cosine operation on operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For floats: cos from IEEE-754.
  • For complex numbers: complex cosine.
  • For quantized types: dequantize_op_quantize(cosine, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [
//            [0.0, 1.57079632],       // [0, pi/2]
//            [3.14159265, 4.71238898] // [pi, 3pi/2]
//           ]
%result = "stablehlo.cosine"(%operand) : (tensor<2x2xf32>) -> tensor<2x2xf32>
// %result: [[1.0, 0.0], [-1.0, 0.0]]

আরো উদাহরণ

count_leading_zeros

শব্দার্থবিদ্যা

Performs element-wise count of the number of leading zero bits in the operand tensor and produces a result tensor.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of integer type (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of integer type (C1)

সীমাবদ্ধতা

  • (C1) type(operand) = type(result) .

উদাহরণ

// %operand: [[0, 1], [128, -1]]
%result = "stablehlo.count_leading_zeros"(%operand) : (tensor<2x2xi64>) -> tensor<2x2xi64>
// %result: [[64, 63], [56, 0]]

আরো উদাহরণ

custom_call

শব্দার্থবিদ্যা

Encapsulates an implementation-defined operation call_target_name that takes inputs and called_computations and produces results . has_side_effect , backend_config and api_version may be used to provide additional implementation-defined metadata.

At the moment, this operation contains a fairly disorganized collection of metadata which reflects organic evolution of its counterpart operation in the XLA compiler. In the future, we are planning to unify this metadata ( #741 ).

ইনপুট

লেবেল নাম টাইপ
(I1) inputs variadic number of values
(I2) call_target_name constant of type string
(I3) has_side_effect constant of type i1
(I4) backend_config constant of type string or attribute dictionary
(I5) api_version constant of type si32
(I6) called_computations variadic number of constants of type string

আউটপুট

নাম টাইপ
results variadic number of values

উদাহরণ

%results = "stablehlo.custom_call"(%input0) {
  call_target_name = "foo",
  has_side_effect = false,
  backend_config = {bar = 42 : i32},
  api_version = 4 : i32,
  called_computations = [@foo]
} : (tensor<f64>) -> tensor<f64>

ভাগ করা

শব্দার্থবিদ্যা

Performs element-wise division of dividend lhs and divisor rhs tensors and produces a result tensor. Depending on the element type, does the following:

  • For integers: integer division which produces the algebraic quotient with any fractional part discarded.
  • For floats: division from IEEE-754.
  • For complex numbers: complex division.
  • For quantized types:
    • dequantize_op_quantize(divide, lhs, rhs, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor of integer, floating-point or complex type or per-tensor quantized tensor (C1)
(I2) rhs tensor of integer, floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of integer, floating-point, or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(lhs) = baseline_type(rhs) = baseline_type(result) .

উদাহরণ

// %lhs: [17.1, -17.1, 17.1, -17.1]
// %rhs: [3.0, 3.0, -3.0, -3.0]
%result = "stablehlo.divide"(%lhs, %rhs) : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
// %result: [5.66666651, -5.66666651, -5.66666651, 5.66666651]

আরো উদাহরণ

dot_general

শব্দার্থবিদ্যা

Computes dot products between slices of lhs and slices of rhs and produces a result tensor.

More formally, result[result_index] = dot_product , where:

  • lhs_result_dimensions = [d for d in axes(lhs) and d not in lhs_batching_dimensions and d not in lhs_contracting_dimensions] .
  • rhs_result_dimensions = [d for d in axes(rhs) and d not in rhs_batching_dimensions and d not in rhs_contracting_dimensions] .
  • result_batching_index + result_lhs_index + result_rhs_index = result_index where size(result_batching_index) = size(lhs_batching_dimensions) , size(result_lhs_index) = size(lhs_result_dimensions) and size(result_rhs_index) = size(rhs_result_dimensions) .
  • transposed_lhs = transpose(lhs, lhs_batching_dimensions + lhs_result_dimensions + lhs_contracting_dimensions) .
  • transposed_lhs_slice = slice(transposed_lhs, result_batching_index + result_lhs_index + [:, ..., :]) .
  • reshaped_lhs_slice = reshape(transposed_lhs_slice, dims(lhs, lhs_contracting_dimensions)) .
  • transposed_rhs = transpose(rhs, rhs_batching_dimensions + rhs_result_dimensions + rhs_contracting_dimensions) .
  • transposed_rhs_slice = slice(transposed_rhs, result_batching_index + result_rhs_index + [:, ..., :]) .
  • reshaped_rhs_slice = reshape(transposed_rhs_slice, dims(rhs, rhs_contracting_dimensions)) .
  • dot_product = reduce( inputs=[multiply(reshaped_lhs_slice, reshaped_rhs_slice)], init_values=[constant(0, element_type(result))], dimensions=range(size(lhs_contracting_dimensions)), body=lambda x, y: add(x, y)) .

For quantized types, performs dequantize_op_quantize( lambda lhs, rhs: dot_general(lhs, rhs, lhs_batching_dimensions, rhs_batching_dimensions, lhs_contracting_dimensions, rhs_contracting_dimensions, precision_config), lhs, rhs, type(result)) .

For hybrid quantized types, performs hybrid_dequantize_then_op( lambda lhs, rhs: dot_general(lhs, rhs, lhs_batching_dimensions, rhs_batching_dimensions, lhs_contracting_dimensions, rhs_contracting_dimensions, precision_config), lhs, rhs) .

precision_config controls the tradeoff between speed and accuracy for computations on accelerator backends. This can be one of the following (at the moment, the semantics of these enum values is underspecified, but we are planning to address this in #755 ):

  • DEFAULT : Fastest calculation, but least accurate approximation to the original number.
  • HIGH : Slower calculation, but more accurate approximation to the original number.
  • HIGHEST : Slowest calculation, but most accurate approximation to the original number.

A DotAlgorithm defines the main properties of the algorithm used to implement the dot operation, which also defines the precision. If the algorithm attribute fields are set, then the precision_config must be DEFAULT . DotAlgorithms do not have a default value, as the default parameters are implementation defined. As such, all dot algorithm fields may be set to None to specify an empty dot algorithm, which will instead use the precision_config value.

DotAlgorithm fields include:

  • lhs_precision_type and rhs_precision_type , the precisions that the LHS and RHS of the operation are rounded to. Precision types are independent from the storage types of the inputs and the output.
  • accumulation_type the precision used for accumulation.
  • lhs_component_count , rhs_component_count , and num_primitive_operations apply when we are doing an algorithm which decomposes the LHS and/or RHS into multiple components and does multiple "primitive" dot operations on those values - usually to emulate a higher precision (eg Leveraging the bfloat16 Artificial Intelligence Datatype জন্য Higher-Precision Computations : bf16_6x tf32_3x, etc). For algorithms with no decomposition, these values should be set to 1 .
  • allow_imprecise_accumulation to specify if accumulation in lower precision is permitted for some steps (eg CUBLASLT_MATMUL_DESC_FAST_ACCUM ).

Example DotAlgorithm attributes:

// Inputs are casted to tf32, and then accumulated in f32:
{lhs_precision_type = tf32,
 rhs_precision_type = tf32,
 accumulation_type = f32,
 lhs_component_count = 1,
 rhs_component_count = 1,
 num_primitive_operations = 1,
 allow_imprecise_accumulation = false}


// bf16_6x: each input is decomposed to 3 bf16 components, then 6 dot operations are done on those components, and the result is accumulated in f32.
{lhs_precision_type = bf16,
 rhs_precision_type = bf16,
 accumulation_type = f32,
 lhs_component_count = 3,
 rhs_component_count = 3,
 num_primitive_operations = 6,
 allow_imprecise_accumulation = false}


// Inputs are (casted to) f8e5m2, and we accumulate in f32, but for some steps we may accumulate in lower precision.
{lhs_precision_type = f8e5m2,
 rhs_precision_type = f8e5m2,
 accumulation_type = f32,
 lhs_component_count = 1,
 rhs_component_count = 1,
 num_primitive_operations = 1,
 allow_imprecise_accumulation = true}

It is up to the implementations to decide which combinations are supported. In general, it is not guaranteed that each algorithm is supported on each accelerator type by the consumer of the StableHLO. If a given algorithm is not supported, an error should be raised as opposed to falling back to an alternative. StableHLO verification will provide best effort verification, preventing algorithms that are not known to be supported on any hardware.

See xla_data.proto > Algorithm for some supported algorithm values. Ticket #2483 captures the plan to create a centralized doc on supported algorithms by backend.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor or per-tensor quantized tensor (C5-C6), (C9-C10), (C12-C14), (C17-C18), (C20)
(I2) rhs tensor or quantized tensor (C7-C10), (C12-C20)
(I3) lhs_batching_dimensions 1-dimensional tensor constant of type si64 (C1), (C3), (C5), (C9), (C12)
(I4) rhs_batching_dimensions 1-dimensional tensor constant of type si64 (C1), (C4), (C7), (C9)
(I5) lhs_contracting_dimensions 1-dimensional tensor constant of type si64 (C2), (C3), (C6), (C10)
(I6) rhs_contracting_dimensions 1-dimensional tensor constant of type si64 (C2), (C4), (C8), (C10), (C16)
(I7) precision_config variadic number of enums of DEFAULT , HIGH , and HIGHEST (C11), (C21)
(I8) lhs_precision_type FloatType or TensorFloat32 (C21)
(I9) rhs_precision_type FloatType or TensorFloat32 (C21)
(I10) accumulation_type FloatType or TensorFloat32 (C21)
(I11) lhs_component_count constant of type si32 (C21), (C22)
(I12) rhs_component_count constant of type si32 (C21), (C23)
(I13) num_primitive_operations constant of type si32 (C21), (C24)
(I14) allow_imprecise_accumulation constant of type bool (C21)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or quantized tensor (C12), (C14), (C18-C20)

সীমাবদ্ধতা

  • (C1) size(lhs_batching_dimensions) = size(rhs_batching_dimensions) .
  • (C2) size(lhs_contracting_dimensions) = size(rhs_contracting_dimensions) .
  • (C3) is_unique(lhs_batching_dimensions + lhs_contracting_dimensions) .
  • (C4) is_unique(rhs_batching_dimensions + rhs_contracting_dimensions) .
  • (C5) 0 <= lhs_batching_dimensions < rank(lhs) .
  • (C6) 0 <= lhs_contracting_dimensions < rank(lhs) .
  • (C7) 0 <= rhs_batching_dimensions < rank(rhs) .
  • (C8) 0 <= rhs_contracting_dimensions < rank(rhs) .
  • (C9) dim(lhs, lhs_batching_dimensions...) = dim(rhs, rhs_batching_dimensions...) .
  • (C10) dim(lhs, lhs_contracting_dimensions...) = dim(rhs, rhs_contracting_dimensions...) .
  • (C11) size(precision_config) = 2 .
  • (C12) shape(result) = dim(lhs, lhs_batching_dimensions) + dim(lhs, lhs_result_dimensions) + dim(rhs, rhs_result_dimensions) .
  • If the operation uses non-quantized tensors:
    • (C13) element_type(lhs) = element_type(rhs) .
  • If the operation uses quantized tensors:
    • (C14) is_quantized(lhs) = is_quantized(result) and is_quantized(rhs) .
    • (C15) zero_points(rhs) = 0 .
    • (C16) If is_per_axis_quantized(rhs) , then quantization_dimension(rhs) not in rhs_contracting_dimensions .
    • If is_quantized(lhs) :
    • (C17) storage_type(lhs) = storage_type(rhs) .
    • (C18) expressed_type(lhs) = expressed_type(rhs) = expressed_type(result) .
    • (C19) If is_per_tensor_quantized(rhs) , then is_per_tensor_quantized(result) .
    • If !is_quantized(lhs) :
    • (C20) element_type(lhs) = expressed_type(rhs) = element_type(result) .
  • If !is_empty_algorithm(lhs_precision_type, rhs_precision_type, accumulation_type, lhs_component_count, rhs_component_count, num_primitive_operations allow_imprecise_accumulation) :
    • (C21) precision_config... = DEFAULT .
    • (C22) 0 < lhs_component_count .
    • (C23) 0 < rhs_component_count .
    • (C24) 0 < num_primitive_operations .

উদাহরণ

// %lhs: [
//        [[1, 2],
//         [3, 4]],
//        [[5, 6],
//         [7, 8]]
//       ]
// %rhs: [
//        [[1, 0],
//         [0, 1]],
//        [[1, 0],
//         [0, 1]]
//       ]
%result = "stablehlo.dot_general"(%lhs, %rhs) {
  dot_dimension_numbers = #stablehlo.dot<
    lhs_batching_dimensions = [0],
    rhs_batching_dimensions = [0],
    lhs_contracting_dimensions = [2],
    rhs_contracting_dimensions = [1]
  >,
  precision_config = [#stablehlo<precision DEFAULT>, #stablehlo<precision DEFAULT>],
  algorithm = #stablehlo.dot_algorithm<
    lhs_precision_type = tf32,
    rhs_precision_type = tf32,
    accumulation_type = f32,
    lhs_component_count = 1,
    rhs_component_count = 1,
    num_primitive_operations = 1,
    allow_imprecise_accumulation = false
  >
} : (tensor<2x2x2xi64>, tensor<2x2x2xi64>) -> tensor<2x2x2xi64>
// %result: [
//           [[1, 2],
//            [3, 4]],
//           [[5, 6],
//            [7, 8]]
//          ]

আরো উদাহরণ

dynamic_broadcast_in_dim

শব্দার্থবিদ্যা

This operation is functionally identical to broadcast_in_dim op, but the result shape is specified dynamically via output_dimensions .

The operation also accepts optional attributes known_expanding_dimensions , known_nonexpanding_dimensions to express static knowledge about the expanding behavior of dimensions. If not specified, all dimensions are assumed to be possibly expanding.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or quantized tensor (C1-C2), (C5-C6), (C9)
(I2) output_dimensions 1-dimensional tensor of integer type (C7)
(I3) broadcast_dimensions 1-dimensional constant tensor of integer type (C2-C6)
(I4) known_expanding_dimensions 1-dimensional constant tensor of integer type (C8-C9)
(I5) known_nonexpanding_dimensions 1-dimensional constant tensor of integer type (C8-C9)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or quantized tensor (C1), (C3), (C5-C7)

সীমাবদ্ধতা

  • (C1) element_type(result) is given by:
    • element_type(operand) , if !is_per_axis_quantized(operand) .
    • element_type(operand) except that quantization_dimension(operand) , scales(operand) , and zero_points(operand) may differ from quantization_dimension(result) , scales(result) , and zero_points(result) resp., otherwise.
  • (C2) size(broadcast_dimensions) = rank(operand) .
  • (C3) 0 <= broadcast_dimensions < rank(result) .
  • (C4) is_unique(broadcast_dimensions) .
  • (C5) For all d in axes(operand) :
    • dim(operand, d) = 1 or
    • dim(operand, d) = dim(result, broadcast_dimensions[d]) .
  • (C6) If is_per_axis_quantized(result) :
    • quantization_dimension(result) = broadcast_dimensions[quantization_dimension(operand)] .
    • If dim(operand, quantization_dimension(operand)) = 1 , then scales(result)[i] = scales(operand)[0] and zero_points(result)[i] = zero_points(operand)[0] for i in range(dim(result, quantization_dimension(result))) .
  • (C7) size(output_dimensions) = rank(result) .
  • (C8) is_unique(known_expanding_dimensions + known_nonexpanding_dimensions) .
  • (C9) 0 <= known_expanding_dimensions < rank(operand) .
  • (C10) 0 <= known_nonexpanding_dimensions < rank(operand) .

উদাহরণ

// %operand: [
//            [1, 2, 3]
//           ]
%operand = stablehlo.constant dense<[[1, 2, 3]]> : tensor<1x3xi64>
%output_dimensions = stablehlo.constant dense<[2, 3, 2]> : tensor<3xi64>
%result = "stablehlo.dynamic_broadcast_in_dim"(%operand, %output_dimensions) {
  broadcast_dimensions = array<i64: 2, 1>,
  known_expanding_dimensions = array<i64: 0>,
  known_nonexpanding_dimensions = array<i64: 1>
} : (tensor<1x3xi64>, tensor<3xi64>) -> tensor<2x3x2xi64>
// %result: [
//            [
//             [1, 1],
//             [2, 2],
//             [3, 3]
//            ],
//            [
//             [1, 1],
//             [2, 2],
//             [3, 3]
//            ]
//          ]

আরো উদাহরণ

dynamic_conv

শব্দার্থবিদ্যা

This operation is functionally identical to convolution op, but the padding is specified dynamically via padding .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor or per-tensor quantized tensor (C1), (C10-C11), (C14) (C25), (C26-C27), (C30-C31), (C33)
(I2) rhs tensor or quantized tensor (C1), (C14-C16), (C26-C28), (C30-C33)
(I3) padding 2-dimensional tensor of integer type (C4)
(I4) window_strides 1-dimensional tensor constant of type si64 (C2-C3)
(I5) lhs_dilation 1-dimensional tensor constant of type si64 (C5-C6)
(I6) rhs_dilation 1-dimensional tensor constant of type si64 (C7-C8)
(I7) window_reversal 1-dimensional tensor constant of type i1 (C9)
(I8) input_batch_dimension constant of type si64 (C10), (C13)
(I9) input_feature_dimension constant of type si64 (C11), (C13-C14)
(I10) input_spatial_dimensions 1-dimensional tensor constant of type si64 (C12), (C13)
(I11) kernel_input_feature_dimension constant of type si64 (C14), (C18)
(I12) kernel_output_feature_dimension constant of type si64 (C15-C16), (C18), (C28)
(I13) kernel_spatial_dimensions 1-dimensional tensor constant of type si64 (C17-C18)
(I14) output_batch_dimension constant of type si64 (C20)
(I15) output_feature_dimension constant of type si64 (C20), (C29)
(I16) output_spatial_dimensions 1-dimensional tensor constant of type si64 (C19-C20)
(I17) feature_group_count constant of type si64 (C11), (C14), (C16), (C21), (C23)
(I18) batch_group_count constant of type si64 (C10), (C15), (C22), (C23)
(I19) precision_config variadic number of enums of DEFAULT , HIGH , and HIGHEST (C24)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or quantized tensor (C25-C27), (C29), (C31-C33)

সীমাবদ্ধতা

  • (C1) N = rank(lhs) = rank(rhs) .
  • (C2) size(window_strides) = N - 2 .
  • (C3) 0 < window_strides .
  • (C4) shape(padding) = [N - 2, 2] .
  • (C5) size(lhs_dilation) = N - 2 .
  • (C6) 0 < lhs_dilation .
  • (C7) size(rhs_dilation) = N - 2 .
  • (C8) 0 < rhs_dilation .
  • (C9) size(window_reversal) = N - 2 .
  • (C10) dim(lhs, input_batch_dimension) % batch_group_count = 0 .
  • (C11) dim(lhs, input_feature_dimension) % feature_group_count = 0 .
  • (C12) size(input_spatial_dimensions) = N - 2 .
  • (C13) Given input_dimensions = [input_batch_dimension] + input_spatial_dimensions + [input_feature_dimension] :
    • is_unique(input_dimensions) .
    • 0 <= input_dimensions < N .
  • (C14) dim(rhs, kernel_input_feature_dimension) = dim(lhs, input_feature_dimension) / feature_group_count .
  • (C15) dim(rhs, kernel_output_feature_dimension) % batch_group_count = 0 .
  • (C16) dim(rhs, kernel_output_feature_dimension) % feature_group_count = 0 .
  • (C17) size(kernel_spatial_dimensions) = N - 2 .
  • (C18) Given kernel_dimensions = kernel_spatial_dimensions + [kernel_input_feature_dimension] + [kernel_output_feature_dimension] :
    • is_unique(kernel_dimensions) .
    • 0 <= kernel_dimensions < N .
  • (C19) size(output_spatial_dimensions) = N - 2 .
  • (C20) Given output_dimensions = [output_batch_dimension] + output_spatial_dimensions + [output_feature_dimension] :
    • is_unique(output_dimensions) .
    • 0 <= output_dimensions < N .
  • (C21) 0 < feature_group_count .
  • (C22) 0 < batch_group_count .
  • (C23) feature_group_count = 1 or batch_group_count = 1 .
  • (C24) size(precision_config) = 2 .
  • (C25) dim(result, result_dim) is defined as:
    • dim(lhs, input_batch_dimension) / batch_group_count if result_dim = output_batch_dimension .
    • dim(rhs, kernel_output_feature_dimension) if result_dim = output_feature_dimension .
    • num_windows otherwise, where:
    • output_spatial_dimensions[spatial_dim] = result_dim .
    • lhs_dim = input_spatial_dimensions[spatial_dim] .
    • rhs_dim = kernel_spatial_dimensions[spatial_dim] .
    • dilated_input_shape[lhs_dim] = dim(lhs, lhs_dim) = 0 ? 0 : (dim(lhs, lhs_dim) - 1) * lhs_dilation[spatial_dim] + 1 .
    • padded_input_shape[lhs_dim] = padding[spatial_dim, 0] + dilated_input_shape[lhs_dim] + padding[spatial_dim, 1] .
    • dilated_window_shape[lhs_dim] = dim(rhs, rhs_dim) = 0 ? 0 : (dim(rhs, rhs_dim) - 1) * rhs_dilation[spatial_dim] + 1 .
    • is_empty_window[lhs_dim] = padded_input_shape[lhs_dim] = 0 || dilated_window_shape[lhs_dim] > padded_input_shape[lhs_dim] .
    • num_windows = is_empty_window[lhs_dim] ? 0 : floor((padded_input_shape[lhs_dim] - dilated_window_shape[lhs_dim]) / window_strides[spatial_dim]) + 1 .
  • (C26) rank(result) = N .
  • If the operation uses non-quantized tensors:
    • (C27) element_type(lhs) = element_type(rhs) = element_type(result) .
  • If the operation uses quantized tensors:
    • (C28) is_quantized(lhs) = is_quantized(result) and is_quantized(rhs) .
    • (C29) If is_per_axis_quantized(rhs) , then quantization_dimension(rhs) = kernel_output_feature_dimension .
    • (C30) If is_per_axis_quantized(result) , then quantization_dimension(result) = output_feature_dimension .
    • If is_quantized(lhs) :
    • (C31) storage_type(lhs) = storage_type(rhs) .
    • (C32) expressed_type(lhs) = expressed_type(rhs) = expressed_type(result) .
    • (C33) If is_per_tensor_quantized(rhs) , then is_per_tensor_quantized(result) .
    • If !is_quantized(lhs) :
    • (C34) element_type(lhs) = expressed_type(rhs) = element_type(result) .

উদাহরণ

// %lhs: [[
//        [[1], [2], [5], [6]],
//        [[3], [4], [7], [8]],
//        [[10], [11], [14], [15]],
//        [[12], [13], [16], [17]]
//      ]]
//
// %rhs: [
//         [[[1]], [[1]], [[1]]],
//         [[[1]], [[1]], [[1]]],
//         [[[1]], [[1]], [[1]]]
//        ]
// %padding: [[1, 1],
//            [1, 1]]
%result = "stablehlo.dynamic_conv"(%lhs, %rhs, %padding) {
  window_strides = array<i64: 4, 4>,
  lhs_dilation = array<i64: 2, 2>,
  rhs_dilation = array<i64: 1, 1>,
  window_reversal = array<i1: false, false>,
  dimension_numbers = #stablehlo.conv<raw
    input_batch_dimension = 0,
    input_feature_dimension = 3,
    input_spatial_dimensions = [0, 1],
    kernel_input_feature_dimension = 2,
    kernel_output_feature_dimension = 3,
    kernel_spatial_dimensions = [0, 1],
    output_batch_dimension = 0,
    output_feature_dimension = 3,
    output_spatial_dimensions = [1, 2]
  >,
  feature_group_count = 1 : i64,
  batch_group_count = 1 : i64,
  precision_config = [#stablehlo<precision DEFAULT>, #stablehlo<precision DEFAULT>]
} : (tensor<1x4x4x1xi64>, tensor<3x3x1x1xi64>, tensor<2x2xi64>) -> tensor<1x2x2x1xi64>
// %result: [[
//            [[1], [5]],
//            [[10], [14]]
//          ]]

আরো উদাহরণ

dynamic_gather

শব্দার্থবিদ্যা

This operation is functionally identical to gather op, with the slice_sizes specified dynamically as a value.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or per-tensor quantized tensor (C1), (C7), (C10-C12), (C14)
(I2) start_indices tensor of integer type (C2), (C3), (C13)
(I3) slice_sizes 1-dimensional tensor of integer type (C8), (C11-C13)
(I4) offset_dims 1-dimensional tensor constant of type si64 (C1), (C4-C5), (C13)
(I5) collapsed_slice_dims 1-dimensional tensor constant of type si64 (C1), (C6-C8), (C13)
(I6) start_index_map 1-dimensional tensor constant of type si64 (C3), (C9), (C10)
(I7) index_vector_dim constant of type si64 (C2), (C3), (C13)
(I8) indices_are_sorted constant of type i1

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C5), (C13-C14)

সীমাবদ্ধতা

  • (C1) rank(operand) = size(offset_dims) + size(collapsed_slice_dims) .
  • (C2) 0 <= index_vector_dim <= rank(start_indices) .
  • (C3) size(start_index_map) = index_vector_dim < rank(start_indices) ? dim(start_indices, index_vector_dim) : 1 .
  • (C4) is_unique(offset_dims) and is_sorted(offset_dims) .
  • (C5) 0 <= offset_dims < rank(result) .
  • (C6) is_unique(collapsed_slice_dims) and is_sorted(collapsed_slice_dims) .
  • (C7) 0 <= collapsed_slice_dims < rank(operand) .
  • (C8) slice_sizes[collapsed_slice_dims...] <= 1 .
  • (C9) is_unique(start_index_map) .
  • (C10) 0 <= start_index_map < rank(operand) .
  • (C11) size(slice_sizes) = rank(operand) .
  • (C12) 0 <= slice_sizes <= shape(operand) .
  • (C13) shape(result) = combine(batch_dim_sizes, offset_dim_sizes) where:
    • batch_dim_sizes = shape(start_indices) except that the dimension size of start_indices corresponding to index_vector_dim is not included.
    • offset_dim_sizes = shape(slice_sizes) except that the dimension sizes in slice_sizes corresponding to collapsed_slice_dims are not included.
    • combine puts batch_dim_sizes at axes corresponding to batch_dims and offset_dim_sizes at axes corresponding to offset_dims .
  • (C14) element_type(operand) = element_type(result) .

উদাহরণ

// %operand: [
//            [[1, 2], [3, 4], [5, 6], [7, 8]],
//            [[9, 10],[11, 12], [13, 14], [15, 16]],
//            [[17, 18], [19, 20], [21, 22], [23, 24]]
//           ]
// %start_indices: [
//                  [[0, 0], [1, 0], [2, 1]],
//                  [[0, 1], [1, 1], [0, 2]]
//                 ]
// %slize_sizes: [1, 2, 2]
%result = "stablehlo.dynamic_gather"(%operand, %start_indices, %slize_sizes) {
  dimension_numbers = #stablehlo.gather<
    offset_dims = [2, 3],
    collapsed_slice_dims = [0],
    start_index_map = [1, 0],
    index_vector_dim = 2>,
  indices_are_sorted = false
} : (tensor<3x4x2xi64>, tensor<2x3x2xi64>, tensor<3xi64>) -> tensor<2x3x2x2xi64>
// %result: [
//            [
//              [[1, 2], [3, 4]],
//              [[3, 4], [5, 6]],
//              [[13, 14], [15, 16]]
//            ],
//            [
//              [[9, 10], [11, 12]],
//              [[11, 12], [13, 14]],
//              [[17, 18], [19, 20]]
//            ]
//          ]

আরো উদাহরণ

dynamic_iota

শব্দার্থবিদ্যা

This operation is functionally identical to iota op, but the result shape is specified dynamically via output_shape .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) output_shape 1-dimensional tensor of integer type (C1), (C2)
(I2) iota_dimension si64 (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of integer, floating-point, or complex type or per-tensor quantized tensor (C2)

সীমাবদ্ধতা

  • (C1) 0 <= iota_dimension < size(output_shape) .
  • (C2) rank(result) = size(output_shape) .

উদাহরণ

%output_shape = stablehlo.constant dense<[4, 5]> : tensor<2xi64>
%result = "stablehlo.dynamic_iota"(%output_shape) {
  iota_dimension = 0 : i64
} : (tensor<2xi64>) -> tensor<4x5xi64>
// %result: [
//           [0, 0, 0, 0, 0],
//           [1, 1, 1, 1, 1],
//           [2, 2, 2, 2, 2],
//           [3, 3, 3, 3, 3]
//          ]

আরো উদাহরণ

dynamic_pad

শব্দার্থবিদ্যা

This operation is functionally identical to pad op, but with edge_padding_low , edge_padding_high , and interior_padding specified dynamically as values.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or per-tensor quantized tensor (C1), (C2), (C4)
(I2) padding_value 0-dimensional tensor or per-tensor quantized tensor (C1)
(I3) edge_padding_low 1-dimensional tensor of integer type (C1), (C4)
(I4) edge_padding_high 1-dimensional tensor of integer type (C1), (C4)
(I5) interior_padding 1-dimensional tensor of integer type (C2-C4)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C3-C6)

সীমাবদ্ধতা

  • (C1) element_type(operand) = element_type(padding_value) = element_type(result) .
  • (C2) size(edge_padding_low) = size(edge_padding_high) = size(interior_padding) = rank(operand) .
  • (C3) 0 <= interior_padding .
  • (C4) shape(result) = shape(operand) + edge_padding_low + max(shape(operand) - 1, 0) * interior_padding + edge_padding_high .

উদাহরণ

// %operand: [
//            [1, 2, 3],
//            [4, 5, 6]
//           ]
// %padding_value: 0
// %edge_padding_low: [0, 1]
// %edge_padding_high: [2, 1]
// %interior_padding: [1, 2]
%result = "stablehlo.dynamic_pad"(%operand, %padding_value,
  %edge_padding_low, %edge_padding_high, %interior_padding
) : (tensor<2x3xi64>, tensor<i64>, tensor<2xi64>, tensor<2xi64>, tensor<2xi64>) -> tensor<5x9xi64>
// %result: [
//           [0, 1, 0, 0, 2, 0, 0, 3, 0],
//           [0, 0, 0, 0, 0, 0, 0, 0, 0],
//           [0, 4, 0, 0, 5, 0, 0, 6, 0],
//           [0, 0, 0, 0, 0, 0, 0, 0, 0],
//           [0, 0, 0, 0, 0, 0, 0, 0, 0]
//          ]

আরো উদাহরণ

dynamic_reshape

শব্দার্থবিদ্যা

This operation is functionally identical to reshape op, but the result shape is specified dynamically via output_shape .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or quantized tensor (C1-C3)
(I2) output_shape 1-dimensional tensor of integer type (C4)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or quantized tensor (C1-C4)

সীমাবদ্ধতা

  • (C1) element_type(result) is given by:
    • element_type(operand) , if !is_per_axis_quantized(operand) .
    • element_type(operand) except that quantization_dimension(operand) and quantization_dimension(result) may differ, otherwise.
  • (C2) size(operand) = size(result) .
  • (C3) If is_per_axis_quantized(operand) :
    • reduce(dims(operand, [0, 1, ..., quantization_dimension(operand) - 1]), init_values=1, dimensions=[0], body=lambda x, y: x * y) = reduce(dims(result, [0, 1, ..., quantization_dimension(result) - 1]), init_values=1, dimensions=[0], body=lambda x, y: x * y) .
    • dim(operand, quantization_dimension(operand)) = dim(result, quantization_dimension(result)) .
    • reduce(dims(operand, [quantization_dimension(operand) + 1, ..., rank(operand) - 1]), init_values=1, dimensions=[0], body=lambda x, y: x * y) = reduce(dims(result, [quantization_dimension(result) + 1, ..., rank(result) - 1]), init_values=1, dimensions=[0], body=lambda x, y: x * y) .
  • (C4) size(output_shape) = rank(result) .

উদাহরণ

// %operand: [[1, 2, 3], [4, 5, 6]]
// %output_shape: [3, 2]
%result = "stablehlo.dynamic_reshape"(%operand, %output_shape) : (tensor<2x3xi64>, tensor<2xi64>) -> tensor<3x2xi64>
// %result: [[1, 2], [3, 4], [5, 6]]

আরো উদাহরণ

dynamic_slice

শব্দার্থবিদ্যা

Extracts a slice from the operand using dynamically-computed starting indices and produces a result tensor. start_indices contain the starting indices of the slice for each dimension subject to potential adjustment, and slice_sizes contain the sizes of the slice for each dimension. More formally, result[result_index] = operand[operand_index] where:

  • adjusted_start_indices = clamp(0, start_indices, shape(operand) - slice_sizes) .
  • operand_index = adjusted_start_indices + result_index .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or per-tensor quantized tensor (C1), (C2), (C4)
(I2) start_indices variadic number of 0-dimensional tensors of integer type (C2), (C3)
(I3) slice_sizes 1-dimensional tensor constant of type si64 (C2), (C4), (C5)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C1), (C5)

সীমাবদ্ধতা

  • (C1) element_type(operand) = element_type(result) .
  • (C2) size(start_indices) = size(slice_sizes) = rank(operand) .
  • (C3) same(type(start_indices...)) .
  • (C4) 0 <= slice_sizes <= shape(operand) .
  • (C5) shape(result) = slice_sizes .

উদাহরণ

// %operand: [
//            [0, 0, 1, 1],
//            [0, 0, 1, 1],
//            [0, 0, 0, 0],
//            [0, 0, 0, 0]
//           ]
// %start_indices0: -1
// %start_indices1: 3
%result = "stablehlo.dynamic_slice"(%operand, %start_indices0, %start_indices1) {
  slice_sizes = array<i64: 2, 2>
} : (tensor<4x4xi32>, tensor<i64>, tensor<i64>) -> tensor<2x2xi32>
// %result: [
//           [1, 1],
//           [1, 1]
//          ]

আরো উদাহরণ

dynamic_update_slice

শব্দার্থবিদ্যা

Produces a result tensor which is equal to the operand tensor except that the slice starting at start_indices is updated with the values in update . More formally, result[result_index] is defined as:

  • update[update_index] if 0 <= update_index < shape(update) where:
    • adjusted_start_indices = clamp(0, start_indices, shape(operand) - shape(update)) .
    • update_index = result_index - adjusted_start_indices .
  • operand[result_index] otherwise.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or per-tensor quantized tensor (C1-C4), (C6)
(I2) update tensor or per-tensor quantized tensor (C2), (C3), (C6)
(I3) start_indices variadic number of 0-dimensional tensors of integer type (C4), (C5)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) type(operand) = type(result) .
  • (C2) element_type(update) = element_type(operand) .
  • (C3) rank(update) = rank(operand) .
  • (C4) size(start_indices) = rank(operand) .
  • (C5) same(type(start_indices...)) .
  • (C6) 0 <= shape(update) <= shape(operand) .

উদাহরণ

// %operand: [
//            [1, 1, 0, 0],
//            [1, 1, 0, 0],
//            [1, 1, 1, 1],
//            [1, 1, 1, 1]
//           ]
// %update: [
//           [1, 1],
//           [1, 1]
//          ]
// %start_indices0: -1
// %start_indices1: 3
%result = "stablehlo.dynamic_update_slice"(%operand, %update, %start_indices0, %start_indices1)
  : (tensor<4x4xi32>, tensor<2x2xi32>, tensor<i64>, tensor<i64>) -> tensor<4x4xi32>
// %result: [
//           [1, 1, 1, 1],
//           [1, 1, 1, 1],
//           [1, 1, 1, 1],
//           [1, 1, 1, 1]
//          ]

আরো উদাহরণ

সূচকীয়

শব্দার্থবিদ্যা

Performs element-wise exponential operation on operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For floats: exp from IEEE-754.
  • For complex numbers: complex exponential.
  • For quantized types: dequantize_op_quantize(exponential, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [[0.0, 1.0], [2.0, 3.0]]
%result = "stablehlo.exponential"(%operand) : (tensor<2x2xf64>) -> tensor<2x2xf64>
// %result: [[1.0, 2.7182818284590451], [7.3890560989306504, 20.085536923187668]]

আরো উদাহরণ

exponential_minus_one

শব্দার্থবিদ্যা

Performs element-wise exponential minus one operation on operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For floats: expm1 from IEEE-754.
  • For complex numbers: complex exponential minus one.
  • For quantized types: dequantize_op_quantize(exponential_minus_one, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [0.0, 1.0]
%result = "stablehlo.exponential_minus_one"(%operand) : (tensor<2xf64>) -> tensor<2xf64>
// %result: [0.0, 1.71828187]

আরো উদাহরণ

fft

শব্দার্থবিদ্যা

Performs the forward and inverse Fourier transforms for real and complex inputs/outputs.

fft_type is one of the following:

  • FFT : Forward complex-to-complex FFT.
  • IFFT : Inverse complex-to-complex FFT.
  • RFFT : Forward real-to-complex FFT.
  • IRFFT : Inverse real-to-complex FFT (ie takes complex, returns real).

More formally, given the function fft which takes 1-dimensional tensors of complex types as input, produces 1-dimensional tensors of same types as output and computes the discrete Fourier transform:

For fft_type = FFT , result is defined as the final result of a series of L computations where L = size(fft_length) . For example, for L = 3 :

  • result1[i0, ..., :] = fft(operand[i0, ..., :]) .
  • result2[i0, ..., :, iR-1] = fft(result1[i0, ..., :, iR-1]) .
  • result[i0, ..., :, iR-2, iR-1] = fft(result2[i0, ..., :, iR-2, iR-1]) .

Furthermore, given the function ifft which has the same type signature and computes the inverse of fft :

For fft_type = IFFT , result is defined as the inverse of the computations for fft_type = FFT . For example, for L = 3 :

  • result1[i0, ..., :, iR-2, iR-1] = ifft(operand[i0, ..., :, iR-2, iR-1]) .
  • result2[i0, ..., :, iR-1] = ifft(result1[i0, ..., :, iR-1]) .
  • result[i0, ..., :] = ifft(result2[i0, ..., :]) .

Furthermore, given the function rfft which takes 1-dimensional tensors of floating-point types, produces 1-dimensional tensors of complex types of the same floating-point semantics and works as follows:

  • rfft(real_operand) = truncated_result where
  • complex_operand... = (real_operand..., 0.0) .
  • complex_result = fft(complex_operand) .
  • truncated_result = complex_result[:(rank(complex_result) / 2 + 1)] .

(When the discrete Fourier transform is computed for real operands, the first N/2 + 1 elements of the result unambiguously define the rest of the result, so the result of rfft is truncated to avoid computing redundant elements).

For fft_type = RFFT , result is defined as the final result of a series of L computations where L = size(fft_length) . For example, for L = 3 :

  • result1[i0, ..., :] = rfft(operand[i0, ..., :]) .
  • result2[i0, ..., :, iR-1] = fft(result1[i0, ..., :, iR-1]) .
  • result[i0, ..., :, iR-2, iR-1] = fft(result2[i0, ..., :, iR-2, iR-1]) .

Finally, given the function irfft which has the same type signature and computes the inverse of rfft :

For fft_type = IRFFT , result is defined as the inverse of the computations for fft_type = RFFT . For example, for L = 3 :

  • result1[i0, ..., :, iR-2, iR-1] = ifft(operand[i0, ..., :, iR-2, iR-1]) .
  • result2[i0, ..., :, iR-1] = ifft(result1[i0, ..., :, iR-1]) .
  • result[i0, ..., :] = irfft(result2[i0, ..., :]) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type (C1), (C2), (C4), (C5)
(I2) fft_type enum of FFT , IFFT , RFFT , and IRFFT (C2), (C5)
(I3) fft_length 1-dimensional tensor constant of type si64 (C1), (C3), (C4)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type (C2), (C4), (C5)

সীমাবদ্ধতা

  • (C1) size(fft_length) <= rank(operand) .
  • (C2) The relationship between operand and result element types varies:
    • If fft_type = FFT , element_type(operand) and element_type(result) have the same complex type.
    • If fft_type = IFFT , element_type(operand) and element_type(result) have the same complex type.
    • If fft_type = RFFT , element_type(operand) is a floating-point type and element_type(result) is a complex type of the same floating-point semantics.
    • If fft_type = IRFFT , element_type(operand) is a complex type and element_type(result) is a floating-point type of the same floating-point semantics.
  • (C3) 1 <= size(fft_length) <= 3 .
  • (C4) If among operand and result , there is a tensor real of a floating-point type, then shape(real)[-size(fft_length):] = fft_length .
  • (C5) shape(result) = shape(operand) except for:
    • If fft_type = RFFT , dim(result, -1) = dim(operand, -1) = 0 ? 0 : dim(operand, -1) / 2 + 1 .
    • If fft_type = IRFFT , dim(operand, -1) = dim(result, -1) = 0 ? 0 : dim(result, -1) / 2 + 1 .

উদাহরণ

// %operand: [(1.0, 0.0), (0.0, 0.0), (0.0, 0.0), (0.0, 0.0)]
%result = "stablehlo.fft"(%operand) {
  fft_type = #stablehlo<fft_type FFT>,
  fft_length = array<i64: 4>
} : (tensor<4xcomplex<f32>>) -> tensor<4xcomplex<f32>>
// %result: [(1.0, 0.0), (1.0, 0.0), (1.0, 0.0), (1.0, 0.0)]

মেঝে

শব্দার্থবিদ্যা

Performs element-wise floor of operand tensor and produces a result tensor. Implements the roundToIntegralTowardNegative operation from the IEEE-754 specification. For quantized types, performs dequantize_op_quantize(floor, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [-0.8166, -0.2530, 0.2530, 0.8166, 2.0]
%result = "stablehlo.floor"(%operand) : (tensor<5xf32>) -> tensor<5xf32>
// %result: [-1.0, -1.0, 0.0, 0.0, 2.0]

আরো উদাহরণ

জড়ো করা

শব্দার্থবিদ্যা

Gathers slices from operand tensor from offsets specified in start_indices and produces a result tensor.

The following diagram shows how elements in result map on elements in operand using a concrete example. The diagram picks a few example result indices and explains in detail which operand indices they correspond to.

জড়ো করা

More formally, result[result_index] = operand[operand_index] where:

  • batch_dims = [d for d in axes(result) and d not in offset_dims] .
  • batch_index = result_index[batch_dims...] .
  • start_index is defined as:
    • start_indices[bi0, ..., :, ..., biN] where bi are individual elements in batch_index and : is inserted at the index_vector_dim index, if index_vector_dim < rank(start_indices) .
    • [start_indices[batch_index]] otherwise.
  • For d_operand in axes(operand) ,
    • full_start_index[d_operand] = clamp(start_index[d_start], 0, dim(operand, d_operand) - slice_sizes[d_operand]) if d_operand = start_index_map[d_start] .
    • full_start_index[d_operand] = 0 otherwise.
  • For d_operand in axes(operand) ,
    • full_batching_index[d_operand] = batch_index[d_start - (d_start < index_vector_dim ? 0 : 1)] if d_operand = operand_batching_dims[i_batching] and d_start = start_indices_batching_dims[i_batching] .
    • full_batching_index[d_operand] = 0 otherwise.
  • offset_index = result_index[offset_dims...] .
  • full_offset_index = [oi0, ..., 0, ..., oiN] where oi are individual elements in offset_index , and 0 is inserted at indices from collapsed_slice_dims and operand_batching_dims .
  • operand_index = full_start_index + full_batching_index + full_offset_index .

If indices_are_sorted is true then the implementation can assume that start_indices are sorted with respect to start_index_map , otherwise the behavior is undefined. More formally, for all i1 < i2 from indices(result) , full_start_index(i1) <= full_start_index(i2) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or per-tensor quantized tensor (C1), (C8), (C11), (C17), (C19-C21), (C23)
(I2) start_indices tensor of integer type (C2-C3), (C14), (C17), (C22)
(I3) offset_dims 1-dimensional tensor constant of type si64 (C1), (C4-C5), (C22)
(I4) collapsed_slice_dims 1-dimensional tensor constant of type si64 (C1), (C6-C9), (C22)
(I5) operand_batching_dims 1-dimensional tensor constant of type si64 (C1), (C6), (C10-C12), (C16-C18), (C22)
(I6) start_indices_batching_dims 1-dimensional tensor constant of type si64 (C13-C17)
(I7) start_index_map 1-dimensional tensor constant of type si64 (C3), (C18-C19)
(I8) index_vector_dim constant of type si64 (C2-C3), (C15), (C22)
(I9) slice_sizes 1-dimensional tensor constant of type si64 (C9), (C12), (C20-C22)
(I10) indices_are_sorted constant of type i1

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C5), (C22-C23)

সীমাবদ্ধতা

  • (C1) rank(operand) = size(offset_dims) + size(collapsed_slice_dims) + size(operand_batching_dims) .
  • (C2) 0 <= index_vector_dim <= rank(start_indices) .
  • (C3) size(start_index_map) = index_vector_dim < rank(start_indices) ? dim(start_indices, index_vector_dim) : 1 .
  • (C4) is_unique(offset_dims) and is_sorted(offset_dims) .
  • (C5) 0 <= offset_dims < rank(result) .
  • (C6) is_unique(concatenate(collapsed_slice_dims, operand_batching_dims))
  • (C7) is_sorted(collapsed_slice_dims) .
  • (C8) 0 <= collapsed_slice_dims < rank(operand) .
  • (C9) slice_sizes[collapsed_slice_dims...] <= 1 .
  • (C10) is_sorted(operand_batching_dims) .
  • (C11) 0 <= operand_batching_dims < rank(operand) .
  • (C12) slice_sizes[operand_batching_dims...] <= 1 .
  • (C13) is_unique(start_indices_batching_dims) .
  • (C14) 0 <= start_indices_batching_dims < rank(start_indices) .
  • (C15) index_vector_dim not in start_indices_batching_dims .
  • (C16) size(operand_batching_dims) == size(start_indices_batching_dims) .
  • (C17) dim(operand, operand_batching_dims...) = dim(start_indices, start_indices_batching_dims...) .
  • (C18) is_unique(concatenate(start_index_map, operand_batching_dims)) .
  • (C19) 0 <= start_index_map < rank(operand) .
  • (C20) size(slice_sizes) = rank(operand) .
  • (C21) 0 <= slice_sizes <= shape(operand) .
  • (C22) shape(result) = combine(batch_dim_sizes, offset_dim_sizes) where:
    • batch_dim_sizes = shape(start_indices) except that the dimension size of start_indices corresponding to index_vector_dim is not included.
    • offset_dim_sizes = slice_sizes except that the dimension sizes in slice_sizes corresponding to collapsed_slice_dims and operand_batching_dims are not included.
    • combine puts batch_dim_sizes at axes corresponding to batch_dims and offset_dim_sizes at axes corresponding to offset_dims .
  • (C23) element_type(operand) = element_type(result) .

উদাহরণ

// %operand: [
//            [
//             [[1, 2], [3, 4], [5, 6], [7, 8]],
//             [[9, 10],[11, 12], [13, 14], [15, 16]],
//             [[17, 18], [19, 20], [21, 22], [23, 24]]
//            ],
//            [
//             [[25, 26], [27, 28], [29, 30], [31, 32]],
//             [[33, 34], [35, 36], [37, 38], [39, 40]],
//             [[41, 42], [43, 44], [45, 46], [47, 48]]
//            ]
//           ]
// %start_indices: [
//                  [
//                   [[0, 0], [1, 0], [2, 1]],
//                   [[0, 1], [1, 1], [0, 9]]
//                  ],
//                  [
//                   [[0, 0], [2, 1], [2, 2]],
//                   [[1, 2], [0, 1], [1, 0]]
//                  ]
//                 ]
%result = "stablehlo.gather"(%operand, %start_indices) {
  dimension_numbers = #stablehlo.gather<
    offset_dims = [3, 4],
    collapsed_slice_dims = [1],
    operand_batching_dims = [0],
    start_indices_batching_dims = [1],
    start_index_map = [2, 1],
    index_vector_dim = 3>,
  slice_sizes = array<i64: 1, 1, 2, 2>,
  indices_are_sorted = false
} : (tensor<2x3x4x2xi32>, tensor<2x2x3x2xi64>) -> tensor<2x2x3x2x2xi32>
// %result: [
//           [
//            [
//             [[1, 2], [3, 4]],
//             [[3, 4], [5, 6]],
//             [[13, 14], [15, 16]]
//            ],
//            [
//             [[33, 34], [35, 36]],
//             [[35, 36], [37, 38]],
//             [[41, 42], [43, 44]]
//            ]
//           ],
//           [
//            [
//             [[1, 2], [3, 4]],
//             [[13, 14], [15, 16]],
//             [[21, 22], [23, 24]]
//            ],
//            [
//             [[43, 44], [45, 46]],
//             [[33, 34], [35, 36]],
//             [[27, 28], [29, 30]]
//            ]
//           ]
//          ]

আরো উদাহরণ

get_dimension_size

শব্দার্থবিদ্যা

Produces the size of the given dimension of the operand . More formally, result = dim(operand, dimension) . The Semantics concerns only with the shape component of the type. The element-type could be anything.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or quantized tensor (C1)
(I2) dimension constant of type si64 (C1)

আউটপুট

নাম টাইপ
result 0-dimensional tensor of type si32

সীমাবদ্ধতা

  • (C1) 0 <= dimension < rank(operand) .

উদাহরণ

// %operand: [[1, 2, 3], [4, 5, 6]]
%result = "stablehlo.get_dimension_size"(%operand) {
  dimension = 1 : i64
} : (tensor<2x3xi64>) -> tensor<i32>
// %result: 3

আরো উদাহরণ

get_tuple_element

শব্দার্থবিদ্যা

Extracts element at index position of the operand tuple and produces a result . More formally, result = operand[index] .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand টিপল (C1), (C2)
(I2) index constant of type si32 (C1), (C2)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result any supported type (C2)

সীমাবদ্ধতা

  • (C1) 0 <= index < size(operand) .
  • (C2) type(result) = tuple_element_types(operand)[index] .

উদাহরণ

// %operand: ([1.0, 2.0], (3))
  index = 0 : i32
} : (tuple<tensor<2xf32>, tuple<tensor<i32>>>) -> tensor<2xf32>
// %result: [1.0, 2.0]

আরো উদাহরণ

যদি

শব্দার্থবিদ্যা

Produces the output from executing exactly one function from true_branch or false_branch depending on the value of pred . More formally, result = pred ? true_branch() : false_branch() .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) pred 0-dimensional tensor of type i1
(I2) true_branch ফাংশন (C1-C3)
(I3) false_branch ফাংশন (C1), (C2)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
results variadic number of tensors, quantized tensors or tokens (C3)

সীমাবদ্ধতা

  • (C1) input_types(true_branch) = input_types(false_branch) = [] .
  • (C2) output_types(true_branch) = output_types(false_branch) .
  • (C3) type(results...) = output_types(true_branch) .

উদাহরণ

// %result_true_branch: 10
// %result_false_branch: 11
// %pred: true
%result = "stablehlo.if"(%pred) ({
  "stablehlo.return"(%result_true_branch) : (tensor<i32>) -> ()
}, {
  "stablehlo.return"(%result_false_branch) : (tensor<i32>) -> ()
}) : (tensor<i1>) -> tensor<i32>
// %result: 10

আরো উদাহরণ

ইমেজ

শব্দার্থবিদ্যা

Extracts the imaginary part, element-wise, from the operand and produces a result tensor. More formally, for each element x : imag(x) = is_complex(x) ? imaginary_part(x) : constant(0, element_type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type (C1), (C2)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point type (C1), (C2)

সীমাবদ্ধতা

  • (C1) shape(result) = shape(operand) .
  • (C2) element_type(result) is defined as:
    • complex_element_type(element_type(operand)) if is_complex(operand) .
    • element_type(operand) otherwise.

উদাহরণ

// %operand: [(1.0, 2.0), (3.0, 4.0)]
%result = "stablehlo.imag"(%operand) : (tensor<2xcomplex<f32>>) -> tensor<2xf32>
// %result: [2.0, 4.0]

আরো উদাহরণ

infeed

শব্দার্থবিদ্যা

Reads data from the infeed and produces results .

Semantics of infeed_config is implementation-defined.

results consist of payload values which come first and a token which comes last. In the future, we are planning to split the payload and the token into two separate outputs to improve clarity ( #670 ).

ইনপুট

লেবেল নাম টাইপ
(I1) token token
(I2) infeed_config constant of type string

আউটপুট

নাম টাইপ সীমাবদ্ধতা
results variadic number of tensors, quantized tensors or tokens (C1-C3)

সীমাবদ্ধতা

  • (C1) 0 < size(results) .
  • (C2) is_empty(result[:-1]) or is_tensor(type(results[:-1])) .
  • (C3) is_token(type(results[-1])) .

উদাহরণ

// %token: !stablehlo.token
// infeed_queue[0]: [[1, 2], [3, 4]]
// infeed_queue[1]: [[5, 6], [7, 8]]
%results0:2 = "stablehlo.infeed"(%token) {
  infeed_config = ""
} : (!stablehlo.token) -> (tensor<2x2xi64>, !stablehlo.token)
// results0#0: [[1, 2], [3, 4]]
%results1:2 = "stablehlo.infeed"(%token) {
  infeed_config = ""
} : (!stablehlo.token) -> (tensor<2x2xi64>, !stablehlo.token)
// results1#0: [[5, 6], [7, 8]]

আরো উদাহরণ

iota

শব্দার্থবিদ্যা

Fills an output tensor with values in increasing order starting from zero along the iota_dimension dimension. আরো আনুষ্ঠানিকভাবে,

output[output_index] = constant(is_quantized(output) ? quantize(output_index[iota_dimension], element_type(output)) : output_index[iota_dimension], element_type(output)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) iota_dimension si64 (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
output tensor of integer, floating-point, or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) 0 <= iota_dimension < rank(output) .

উদাহরণ

%output = "stablehlo.iota"() {
  iota_dimension = 0 : i64
} : () -> tensor<4x5xi32>
// %output: [
//           [0, 0, 0, 0, 0],
//           [1, 1, 1, 1, 1],
//           [2, 2, 2, 2, 2],
//           [3, 3, 3, 3, 3]
//          ]

%output = "stablehlo.iota"() {
  iota_dimension = 1 : i64
} : () -> tensor<4x5xi32>
// %output: [
//           [0, 1, 2, 3, 4],
//           [0, 1, 2, 3, 4],
//           [0, 1, 2, 3, 4],
//           [0, 1, 2, 3, 4]
//          ]

আরো উদাহরণ

is_finite

শব্দার্থবিদ্যা

Performs element-wise check whether the value in x is finite (ie is neither +Inf, -Inf, nor NaN) and produces a y tensor. Implements the isFinite operation from the IEEE-754 specification. For quantized types, the result is always true .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) x tensor of floating-point type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
y tensor of boolean type (C1)

সীমাবদ্ধতা

  • (C1) shape(x) = shape(y) .

উদাহরণ

// Logical values: -Inf, +Inf, NaN, ...
// %x: [0xFFF0000000000000, 0x7FF0000000000000, 0x7FF8000000000000, -10.0, -0.0, 0.0, 10.0]
%y = "stablehlo.is_finite"(%x) : (tensor<7xf64) -> tensor<7xi1>
// %y: [false, false, false, true, true, true, true]

আরো উদাহরণ

লগ

শব্দার্থবিদ্যা

Performs element-wise logarithm operation on operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For floats: log from IEEE-754.
  • For complex numbers: complex logarithm.
  • For quantized types: dequantize_op_quantize(log, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [[1.0, 2.0], [3.0, 4.0]]
%result = "stablehlo.log"(%operand) : (tensor<2x2xf64>) -> tensor<2x2xf64>
// %result: [[0.0, 0.69314718055994529], [1.0986122886681098, 1.3862943611198906]]

আরো উদাহরণ

log_plus_one

শব্দার্থবিদ্যা

Performs element-wise logarithm plus one operation on operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For floats: logp1 from IEEE-754.
  • For complex numbers: complex(log(hypot(real(x) + 1, imag(x))), atan2(imag(x), real(x) + 1))
  • For quantized types: dequantize_op_quantize(log_plus_one, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [0.0, -0.999, 7.0, 6.38905621, 15.0]
%result = "stablehlo.log_plus_one"(%operand) : (tensor<5xf64>) -> tensor<5xf64>
// %result: [0.0, -6.90776825, 2.07944155, 2.0, 2.77258873]

আরো উদাহরণ

লজিস্টিক

শব্দার্থবিদ্যা

Performs element-wise logistic operation on operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For floats: division(1, addition(1, exp(-x))) from IEEE-754.
  • For complex numbers: complex logistic.
  • For quantized types: dequantize_op_quantize(logistic, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [[0.0, 1.0], [2.0, 3.0]]
%result = "stablehlo.logistic"(%operand) : (tensor<2x2xf64>) -> tensor<2x2xf64>
// %result: [[0.5, 0.73105858], [0.88079708, 0.95257413]]

আরো উদাহরণ

মানচিত্র

শব্দার্থবিদ্যা

Applies a map function computation to inputs along the dimensions and produces a result tensor.

More formally, result[result_index] = computation(inputs...[result_index]) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) inputs variadic number of tensors or per-tensor quantized tensors (C1-C4)
(I2) dimensions 1-dimensional tensor constant of type si64 (C3)
(I3) computation ফাংশন (C4)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C1), (C4)

সীমাবদ্ধতা

  • (C1) shape(inputs...) = shape(result) .
  • (C2) 0 < size(inputs) = N .
  • (C3) dimensions = range(rank(inputs[0])) .
  • (C4) computation has type (tensor<E0>, ..., tensor<EN-1>) -> tensor<E'> where Ei = element_type(inputs[i]) and E' = element_type(result) .

উদাহরণ

// %input0: [[0, 1], [2, 3]]
// %input1: [[4, 5], [6, 7]]
%result = "stablehlo.map"(%input0, %input1) ({
  ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>):
    %0 = stablehlo.multiply %arg0, %arg1 : tensor<i64>
    stablehlo.return %0 : tensor<i64>
}) {
  dimensions = array<i64: 0, 1>
} : (tensor<2x2xi64>, tensor<2x2xi64>) -> tensor<2x2xi64>
// %result: [[0, 5], [12, 21]]

আরো উদাহরণ

সর্বোচ্চ

শব্দার্থবিদ্যা

Performs element-wise max operation on tensors lhs and rhs and produces a result tensor. Depending on the element type, does the following:

  • For booleans: logical OR.
  • For integers: integer maximum.
  • For floats: maximum from IEEE-754.
  • For complex numbers: lexicographic maximum for the (real, imaginary) pair. Imposing an ordering on complex numbers involves surprising semantics, so in the future we are planning to remove support for complex numbers for this operation ( #560 ).
  • For quantized types:
    • dequantize_op_quantize(maximum, lhs, rhs, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor or per-tensor quantized tensor (C1)
(I2) rhs tensor or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(lhs) = baseline_type(rhs) = baseline_type(result) .

উদাহরণ

// %lhs: [[1, 2], [7, 8]]
// %rhs: [[5, 6], [3, 4]]
%result = "stablehlo.maximum"(%lhs, %rhs) : (tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32>
// %result: [[5, 6], [7, 8]]

আরো উদাহরণ

সর্বনিম্ন

শব্দার্থবিদ্যা

Performs element-wise min operation on tensors lhs and rhs and produces a result tensor. Depending on the element type, does the following:

  • For booleans: logical AND.
  • For integers: integer minimum.
  • For floats: minimum from IEEE-754.
  • For complex numbers: lexicographic minimum for the (real, imaginary) pair. Imposing an ordering on complex numbers involves surprising semantics, so in the future we are planning to remove support for complex numbers for this operation ( #560 ).
  • For quantized types:
    • dequantize_op_quantize(minimum, lhs, rhs, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor or per-tensor quantized tensor (C1)
(I2) rhs tensor or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(lhs) = baseline_type(rhs) = baseline_type(result) .

উদাহরণ

// %lhs: [[1, 2], [7, 8]]
// %rhs: [[5, 6], [3, 4]]
%result = "stablehlo.minimum"(%lhs, %rhs) : (tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32>
// %result: [[1, 2], [3, 4]]

আরো উদাহরণ

গুণ

শব্দার্থবিদ্যা

Performs element-wise product of two tensors lhs and rhs and produces a result tensor. Depending on the element type, does the following:

  • For booleans: logical AND.
  • For integers: integer multiplication.
  • For floats: multiplication from IEEE-754.
  • For complex numbers: complex multiplication.
  • For quantized types:
    • dequantize_op_quantize(multiply, lhs, rhs, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor or per-tensor quantized tensor (C1)
(I2) rhs tensor or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %lhs: [[1, 2], [3, 4]]
// %rhs: [[5, 6], [7, 8]]
%result = "stablehlo.multiply"(%lhs, %rhs) : (tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32>
// %result: [[5, 12], [21, 32]]

আরো উদাহরণ

অস্বীকার

শব্দার্থবিদ্যা

Performs element-wise negation of operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For signed integers: integer negation.
  • For unsigned integers: bitcast to signed integer, integer negation, bitcast back to unsigned integer.
  • For floats: negate from IEEE-754.
  • For complex numbers: complex negation.
  • For quantized types: dequantize_op_quantize(negate, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of integer, floating-point, or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of integer, floating-point, or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// Negation operation with integer Tensors
// %operand: [0, -2]
%result = "stablehlo.negate"(%operand) : (tensor<2xi32>) -> tensor<2xi32>
// %result: [0, 2]

// Negation operation with with complex tensors
// %operand: (2.5, 0.0)
%result = "stablehlo.negate"(%operand) : (tensor<1xcomplex<f32>>) -> tensor<1xcomplex<f32>>
// %result: [-2.5, -0.0]

আরো উদাহরণ

না

শব্দার্থবিদ্যা

Performs element-wise NOT of tensor operand and produces a result tensor. Depending on the element type, does the following:

  • For booleans: logical NOT.
  • For integers: bitwise NOT.

যুক্তি

নাম টাইপ সীমাবদ্ধতা
operand tensor of boolean or integer type (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of boolean or integer type (C1)

সীমাবদ্ধতা

  • (C1) type(operand) = type(result) .

উদাহরণ

// Bitwise operation with with integer tensors
// %operand: [[1, 2], [3, 4]]
%result = "stablehlo.not"(%operand) : (tensor<2x2xi32>) -> tensor<2x2xi32>
// %result: [[-2, -3], [-4, -5]]

// Bitwise operation with with boolean tensors
// %operand: [true, false]
%result = "stablehlo.not"(%operand) : (tensor<2xi1>) -> tensor<2xi1>
// %result: [false, true]

আরো উদাহরণ

optimization_barrier

শব্দার্থবিদ্যা

Ensures that the operations that produce the operand are executed before any operations that depend on the result and prevents compiler transformations from moving operations across the barrier. Other than that, the operation is an identity, ie result = operand .

যুক্তি

নাম টাইপ সীমাবদ্ধতা
operand variadic number of tensors, per-tensor quantized tensors or tokens (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result variadic number of tensors, per-tensor quantized tensors or tokens (C1)

সীমাবদ্ধতা

  • (C1) type(operand...) = type(result...) .

উদাহরণ

// %operand0: 0.0
// %operand1: 1.0
%result0, %result1 = "stablehlo.optimization_barrier"(%operand0, %operand1) : (tensor<f32>, tensor<f32>) -> (tensor<f32>, tensor<f32>)
// %result0: 0.0
// %result1: 1.0

আরো উদাহরণ

বা

শব্দার্থবিদ্যা

Performs element-wise OR of two tensors lhs and rhs and produces a result tensor. Depending on the element type, does the following:

  • For booleans: logical OR.
  • For integers: bitwise OR.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor of integer or boolean type (C1)
(I2) rhs tensor of integer or boolean type (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of integer or boolean type (C1)

সীমাবদ্ধতা

  • (C1) type(lhs) = type(rhs) = type(result) .

উদাহরণ

// Bitwise operation with with integer tensors
// %lhs: [[1, 2], [3, 4]]
// %rhs: [[5, 6], [7, 8]]
%result = "stablehlo.or"(%lhs, %rhs) : (tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32>
// %result: [[5, 6], [7, 12]]

// Logical operation with with boolean tensors
// %lhs: [[false, false], [true, true]]
// %rhs: [[false, true], [false, true]]
%result = "stablehlo.or"(%lhs, %rhs) : (tensor<2x2xi1>, tensor<2x2xi1>) -> tensor<2x2xi1>
// %result: [[false, true], [true, true]]

আরো উদাহরণ

outfeed

শব্দার্থবিদ্যা

Writes inputs to the outfeed and produces a result token.

Semantics of outfeed_config is implementation-defined.

ইনপুট

লেবেল নাম টাইপ
(I1) inputs variadic number of tensors or quantized tensors
(I2) token token
(I3) outfeed_config constant of type string

আউটপুট

নাম টাইপ
result token

উদাহরণ

%result = "stablehlo.outfeed"(%input0, %token) {
  outfeed_config = ""
} : (tensor<2x2x2xi64>, !stablehlo.token) -> !stablehlo.token

আরো উদাহরণ

প্যাড

শব্দার্থবিদ্যা

Expands operand by padding around the tensor as well as between the elements of the tensor with the given padding_value .

edge_padding_low and edge_padding_high specify the amount of padding added at the low-end (next to index 0) and the high-end (next to the highest index) of each dimension respectively. The amount of padding can be negative, where the absolute value of negative padding indicates the number of elements to remove from the specified dimension.

interior_padding specifies the amount of padding added between any two elements in each dimension which may not be negative. Interior padding occurs before edge padding such that negative edge padding will remove elements from the interior-padded operand.

More formally, result[result_index] is defined as:

  • operand[operand_index] if result_index = edge_padding_low + operand_index * (interior_padding + 1) .
  • padding_value otherwise.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or per-tensor quantized tensor (C1), (C2), (C4)
(I2) padding_value 0-dimensional tensor or per-tensor quantized tensor (C1)
(I3) edge_padding_low 1-dimensional tensor constant of type si64 (C1), (C4)
(I4) edge_padding_high 1-dimensional tensor constant of type si64 (C1), (C4)
(I5) interior_padding 1-dimensional tensor constant of type si64 (C2-C4)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C3-C6)

সীমাবদ্ধতা

  • (C1) element_type(operand) = element_type(padding_value) = element_type(result) .
  • (C2) size(edge_padding_low) = size(edge_padding_high) = size(interior_padding) = rank(operand) .
  • (C3) 0 <= interior_padding .
  • (C4) shape(result) = shape(operand) + edge_padding_low + max(shape(operand) - 1, 0) * interior_padding + edge_padding_high .

উদাহরণ

// %operand: [
//            [1, 2, 3],
//            [4, 5, 6]
//           ]
// %padding_value: 0
%result = "stablehlo.pad"(%operand, %padding_value) {
  edge_padding_low = array<i64: 0, 1>,
  edge_padding_high = array<i64: 2, 1>,
  interior_padding = array<i64: 1, 2>
} : (tensor<2x3xi32>, tensor<i32>) -> tensor<5x9xi32>
// %result: [
//           [0, 1, 0, 0, 2, 0, 0, 3, 0],
//           [0, 0, 0, 0, 0, 0, 0, 0, 0],
//           [0, 4, 0, 0, 5, 0, 0, 6, 0],
//           [0, 0, 0, 0, 0, 0, 0, 0, 0],
//           [0, 0, 0, 0, 0, 0, 0, 0, 0]
//          ]

আরো উদাহরণ

partition_id

শব্দার্থবিদ্যা

Produces partition_id of the current process.

আউটপুট

নাম টাইপ
result 0-dimensional tensor of type ui32

উদাহরণ

%result = "stablehlo.partition_id"() : () -> tensor<ui32>

আরো উদাহরণ

popcnt

শব্দার্থবিদ্যা

Performs element-wise count of the number of bits set in the operand tensor and produces a result tensor.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of integer type (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of integer type (C1)

সীমাবদ্ধতা

  • (C1) type(operand) = type(result) .

উদাহরণ

// %operand: [0, 1, 2, 127]
%result = "stablehlo.popcnt"(%operand) : (tensor<4xi64>) -> tensor<4xi64>
// %result: [0, 1, 1, 7]

আরো উদাহরণ

ক্ষমতা

শব্দার্থবিদ্যা

Performs element-wise exponentiation of lhs tensor by rhs tensor and produces a result tensor. Depending on the element type, does the following:

  • For integers: integer exponentiation.
  • For floats: pow from IEEE-754.
  • For complex numbers: complex exponentiation.
  • For quantized types: dequantize_op_quantize(power, lhs, rhs, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor of integer, floating-point, or complex type or per-tensor quantized tensor (C1)
(I2) rhs tensor of integer, floating-point, or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of integer, floating-point, or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %lhs: [-2.0, -0.0, -36.0, 5.0, 3.0, 10000.0]
// %rhs: [2.0, 2.0, 1.1, 2.0, -1.0, 10.0]
%result = "stablehlo.power"(%lhs, %rhs) : (tensor<6xf64>, tensor<6xf64>) -> tensor<6xf64>
// %result: [4.0, 0.0, -nan, 25.0, 0.333333343, inf]

আরো উদাহরণ

বাস্তব

শব্দার্থবিদ্যা

Extracts the real part, element-wise, from the operand and produces a result tensor. More formally, for each element x : real(x) = is_complex(x) ? real_part(x) : x .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type (C1), (C2)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point type (C1), (C2)

সীমাবদ্ধতা

  • (C1) shape(result) = shape(operand) .
  • (C2) element_type(result) is defined as:
    • complex_element_type(element_type(operand)) if is_complex(operand) .
    • element_type(operand) otherwise.

উদাহরণ

// %operand: [(1.0, 2.0), (3.0, 4.0)]
%result = "stablehlo.real"(%operand) : (tensor<2xcomplex<f32>>) -> tensor<2xf32>
// %result: [1.0, 3.0]

আরো উদাহরণ

recv

শব্দার্থবিদ্যা

Receives data from a channel with channel_id and produces results .

If is_host_transfer is true , then the operation transfers data from the host. Otherwise, it transfers data from another device. What this means is implementation-defined. This flag duplicates the information provided in channel_type , so in the future we are planning to only keep one of them ( #666 ).

results consist of payload values which come first and a token which comes last. In the future, we are planning to split the payload and the token into two separate outputs to improve clarity ( #670 ).

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) token token (C4)
(I2) channel_id constant of type si64
(I3) channel_type enum of DEVICE_TO_DEVICE and HOST_TO_DEVICE (C1)
(I4) is_host_transfer constant of type i1 (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
results variadic number of tensors, quantized tensors or tokens (C2-C4)

সীমাবদ্ধতা

  • (C1) channel_type is defined as:
    • HOST_TO_DEVICE if is_host_transfer = true ,
    • DEVICE_TO_DEVICE otherwise.
  • (C2) 0 < size(results) .
  • (C3) is_empty(result[:-1]) or is_tensor(type(results[:-1])) .
  • (C4) is_token(type(results[-1])) .

উদাহরণ

%results0, %results1 = "stablehlo.recv"(%token) {
  channel_handle = #stablehlo.channel_handle<handle = 1, type = 3>,
  is_host_transfer = true
} : (!stablehlo.token) -> (tensor<2x2xi64>, !stablehlo.token)

আরো উদাহরণ

হ্রাস করা

শব্দার্থবিদ্যা

Applies a reduction function body to inputs and init_values along the dimensions and produces results tensors.

The order of reductions is implementation-defined, which means that body and init_values must form a monoid to guarantee that the operation produces the same results for all inputs on all implementations. However, this condition doesn't hold for many popular reductions. Eg floating-point addition for body and zero for init_values don't actually form a monoid because floating-point addition is not associative.

More formally, results...[j0, ..., jR-1] = reduce(input_slices_converted) where:

  • input_slices = inputs...[j0, ..., :, ..., jR-1] , where : are inserted at dimensions .
  • input_slices_converted = to_destination_type(input_slices..., type(func_inputs(body)[:len(func_inputs(body))//2])...) .
  • init_values_converted = to_destination_type(init_values..., type(func_inputs(body)[len(func_inputs(body))//2:])...) .
  • reduce(input_slices_converted) = exec(schedule) for some binary tree schedule where:
    • exec(node) = body(exec(node.left), exec(node.right)) .
    • exec(leaf) = leaf.value .
  • schedule is an implementation-defined full binary tree whose in-order traversal consists of:
    • input_slices_converted...[index] values, for all index in index_space(input_slices_converted) in the ascending lexicographic order of index .
    • Interspersed with an implementation-defined amount of init_values_converted at implementation-defined positions.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) inputs variadic number of tensors or per-tensor quantized tensors (C1-C4), (C6), (C7)
(I2) init_values variadic number of 0-dimensional tensors or per-tensor quantized tensors (C2), (C3)
(I3) dimensions 1-dimensional tensor constant of type si64 (C4), (C5), (C7)
(I4) body ফাংশন (C6)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
results variadic number of tensors or per-tensor quantized tensors (C3), (C7), (C8)

সীমাবদ্ধতা

  • (C1) same(shape(inputs...)) .
  • (C2) element_type(inputs...) = element_type(init_values...) .
  • (C3) 0 < size(inputs) = size(init_values) = size(results) = N .
  • (C4) 0 <= dimensions < rank(inputs[0]) .
  • (C5) is_unique(dimensions) .
  • (C6) body has type (tensor<E0>, ..., tensor<EN-1>, tensor<E0>, ..., tensor<EN-1>) -> (tensor<E0>, ..., tensor<EN-1>) where is_promotable(element_type(inputs[i]), Ei) .
  • (C7) shape(results...) = shape(inputs...) except that the dimension sizes of inputs... corresponding to dimensions are not included.
  • (C8) element_type(results[i]) = Ei for all i in [0,N) .

উদাহরণ

// %input = [[0, 1, 2, 3, 4, 5]]
// %init_value = 0
%result = "stablehlo.reduce"(%input, %init_value) ({
  ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>):
    %0 = "stablehlo.add"(%arg0, %arg1) : (tensor<i64>, tensor<i64>) -> tensor<i64>
    "stablehlo.return"(%0) : (tensor<i64>) -> ()
}) {
  dimensions = array<i64: 1>
} : (tensor<1x6xi64>, tensor<i64>) -> tensor<1xi64>
// %result = [15]

আরো উদাহরণ

reduce_precision

শব্দার্থবিদ্যা

Performs element-wise conversion of operand to another floating-point type that uses exponent_bits and mantissa_bits and back to the original floating-point type and produces an output tensor.

আরো আনুষ্ঠানিকভাবে:

  • The mantissa bits of the original value are updated to round the original value to the nearest value representable with mantissa_bits using roundToIntegralTiesToEven semantics.
  • Then, if mantissa_bits are smaller than the number of mantissa bits of the original value, the mantissa bits are truncated to mantissa_bits .
  • Then, if the exponent bits of the intermediate result don't fit into the range provided by exponent_bits , the intermediate result overflows to infinity using the original sign or underflows to zero using the original sign.
  • For quantized types, performs dequantize_op_quantize( lambda operand: reduce_precision(operand, exponent_bits, mantissa_bits), operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point type or per-tensor quantized tensor (C1)
(I2) exponent_bits constant of type si32 (C2)
(I3) mantissa_bits constant of type si32 (C3)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
output tensor of floating-point type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(output) .
  • (C2) 1 <= exponent_bits .
  • (C3) 0 <= mantissa_bits .

উদাহরণ

// Logical values: +Inf, NaN, +Denormal, 0.0, 65519.0, 65520.0
// %operand: [0x7FF0000000000000, 0x7FFFFFFFFFFFFFFF, 0x0000000000000001, 0.0, 65519.0, 65520.0]
%output = "stablehlo.reduce_precision"(%operand) {
  exponent_bits = 5 : i32,
  mantissa_bits = 10 : i32
} : (tensor<6xf64>) -> tensor<6xf64>
// Logical values: +Inf, NaN, 0.0, 0.0, 65504.0, +Inf
// %output: [0x7FF0000000000000, 0x7FFFFFFFFFFFFFFF, 0.0, 0.0, 65504.0, 0x7FF0000000000000]

আরো উদাহরণ

reduce_scatter

শব্দার্থবিদ্যা

reduce_scatter

Within each process group in the StableHLO process grid, performs reduction, using computations , over the values of the operand tensor from each process, splits the reduction result along scatter_dimension into parts, and scatters the split parts between the processes to produce the result .

The operation splits the StableHLO process grid into process_groups which is defined as follows:

  • cross_replica(replica_groups) if channel_id <= 0 and use_global_device_ids = false .
  • cross_replica_and_partition(replica_groups) if channel_id > 0 and use_global_device_ids = false .
  • flattened_ids(replica_groups) if channel_id > 0 and use_global_device_ids = true .

Afterwards, within each process_group :

  • reduced_value = all_reduce(operand, replica_groups, channel_id, use_global_device_ids, computation) .
  • parts@sender = split(reduced_value@sender, dim(process_groups, 1), scatter_dimension) .
  • result@receiver = parts@sender[receiver_index] for all sender in process_group , where receiver_index = process_group.index(receiver) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or per-tensor quantized tensor (C1), (C2), (C7), (C8)
(I2) scatter_dimension constant of type si64 (C1), (C2), (C8)
(I3) replica_groups 2-dimensional tensor constant of type si64 (C3-C5)
(I4) channel_id constant of type si64 (C6)
(I5) use_global_device_ids constant of type i1 (C6)
(I6) computation ফাংশন (C7)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C8-C9)

সীমাবদ্ধতা

  • (C1) dim(operand, scatter_dimension) % dim(process_groups, 1) = 0 .
  • (C2) 0 <= scatter_dimension < rank(operand) .
  • (C3) is_unique(replica_groups) .
  • (C4) size(replica_groups) is defined as:
    • num_replicas if cross_replica is used.
    • num_replicas if cross_replica_and_partition is used.
    • num_processes if flattened_ids is used.
  • (C5) 0 <= replica_groups < size(replica_groups) .
  • (C6) If use_global_device_ids = true , then channel_id > 0 .
  • (C7) computation has type (tensor<E>, tensor<E>) -> (tensor<E>) where is_promotable(element_type(operand), E) .
  • (C8) shape(result) = shape(operand) except:
    • dim(result, scatter_dimension) = dim(operand, scatter_dimension) / dim(process_groups, 1) .
  • (C9) element_type(result) = E .

উদাহরণ

// num_replicas: 2
// num_partitions: 1
// %operand@(0, 0): [[1, 2, 3, 4],
//                   [5, 6, 7, 8]]
// %operand@(1, 0): [[9, 10, 11, 12],
//                   [13, 14, 15, 16]]
%result = "stablehlo.reduce_scatter"(%operand) ({
  ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>):
  %0 = "stablehlo.add"(%arg0, %arg1) : (tensor<i64>, tensor<i64>) -> tensor<i64>
  "stablehlo.return"(%0) : (tensor<i64>) -> ()
}) {
  scatter_dimension = 1 : i64,
  replica_groups = dense<[[0, 1]]> : tensor<1x2xi64>,
  channel_handle = #stablehlo.channel_handle<handle = 0, type = 0>
} : (tensor<2x4xi64>) -> tensor<2x2xi64>
//
// %result@(0, 0): [[10, 12],
//                  [18, 20]]
// %result@(1, 0): [[14, 16],
//                  [22, 24]]

আরো উদাহরণ

reduce_window

শব্দার্থবিদ্যা

Applies a reduction function body to windows of inputs and init_values and produces results .

The following diagram shows how elements in results... are computed from inputs... using a concrete example.

reduce_window

More formally, results...[result_index] = reduce(windows, init_values, axes(inputs...), body) (see reduce ) where:

  • padded_inputs = pad(inputs..., init_values..., padding[:, 0], padding[:, 1], base_dilations - 1) .
  • window_start = result_index * window_strides .
  • window_end = window_start + (window_dimensions - 1) * window_dilations + 1 .
  • windows = slice(padded_inputs..., window_start, window_end, window_dilations) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) inputs variadic number of tensors or per-tensor quantized tensors (C1-C4), (C6), (C8), (C10), (C12), (C13), (C15)
(I2) init_values variadic number of 0-dimensional tensors or per-tensor quantized tensors (C1), (C13)
(I3) window_dimensions 1-dimensional tensor constant of type si64 (C4), (C5), (C15)
(I4) window_strides 1-dimensional tensor constant of type si64 (C6), (C7), (C15)
(I5) base_dilations 1-dimensional tensor constant of type si64 (C8), (C9), (C15)
(I6) window_dilations 1-dimensional tensor constant of type si64 (C10), (C11), (C15)
(I7) padding 2-dimensional tensor constant of type si64 (C12), (C15)
(I8) body ফাংশন (C13)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
results variadic number of tensors or per-tensor quantized tensors (C1), (C14-C16)

সীমাবদ্ধতা

  • (C1) 0 < size(inputs) = size(init_values) = size(results) = N .
  • (C2) same(shape(inputs...)) .
  • (C3) element_type(inputs...) = element_type(init_values...) .
  • (C4) size(window_dimensions) = rank(inputs[0]) .
  • (C5) 0 < window_dimensions .
  • (C6) size(window_strides) = rank(inputs[0]) .
  • (C7) 0 < window_strides .
  • (C8) size(base_dilations) = rank(inputs[0]) .
  • (C9) 0 < base_dilations .
  • (C10) size(window_dilations) = rank(inputs[0]) .
  • (C11) 0 < window_dilations .
  • (C12) shape(padding) = [rank(inputs[0]), 2] .
  • (C13) body has type (tensor<E0>, ..., tensor<EN-1>, tensor<E0>, ..., tensor<EN-1>) -> (tensor<E0>, ..., tensor<EN-1>) where is_promotable(element_type(inputs[i]), Ei) .
  • (C14) same(shape(results...)) .
  • (C15) shape(results[0]) = num_windows where:
    • dilated_input_shape = shape(inputs[0]) = 0 ? 0 : (shape(inputs[0]) - 1) * base_dilations + 1 .
    • padded_input_shape = padding[:, 0] + dilated_input_shape + padding[:, 1] .
    • dilated_window_shape = (window_dimensions - 1) * window_dilations + 1 .
    • is_empty_window = padded_input_shape = 0 || dilated_window_shape > padded_input_shape .
    • num_windows = is_empty_window ? 0 : floor((padded_input_shape - dilated_window_shape) / window_strides) + 1 .
  • (C16) element_type(results[i]) = Ei for all i in [0,N) .

উদাহরণ

// %input = [[1, 2], [3, 4], [5, 6]]
// %init_value = 0
%result = "stablehlo.reduce_window"(%input, %init_value) ({
  ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>):
    %0 = "stablehlo.add"(%arg0, %arg1) : (tensor<i64>, tensor<i64>) -> tensor<i64>
    "stablehlo.return"(%0) : (tensor<i64>) -> ()
}) {
  window_dimensions = array<i64: 2, 1>,
  window_strides = array<i64: 4, 1>,
  base_dilations = array<i64: 2, 1>,
  window_dilations = array<i64: 3, 1>,
  padding = dense<[[2, 1], [0, 0]]> : tensor<2x2xi64>
} : (tensor<3x2xi64>, tensor<i64>) -> tensor<2x2xi64>
// %result = [[0, 0], [3, 4]]

আরো উদাহরণ

অবশিষ্ট

শব্দার্থবিদ্যা

Performs element-wise remainder of dividend lhs and divisor rhs tensors and produces a result tensor.

More formally, the sign of the result is taken from the dividend, and the absolute value of the result is always less than the divisor's absolute value. The remainder is calculated as lhs - d * rhs , where d is given by:

  • For integers: stablehlo.divide(lhs, rhs) .
  • For floats: division(lhs, rhs) from IEEE-754 with rounding attribute roundTowardZero .
  • For complex numbers: TBD ( #997 ).
  • For quantized types:
    • dequantize_op_quantize(remainder, lhs, rhs, type(result)) .

For floating-point element types, this operation is in contrast with the remainder operation from IEEE-754 specification where d is an integral value nearest to the exact value of lhs/rhs with ties to even.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor of integer, floating-point or complex type or per-tensor quantized tensor (C1)
(I2) rhs tensor of integer, floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of integer, floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %lhs: [17, -17, 17, -17]
// %rhs: [3, 3, -3, -3]
%result = "stablehlo.remainder"(%lhs, %rhs) : (tensor<4xi64>, tensor<4xi64>) -> tensor<4xi64>
// %result: [2, -2, 2, -2]

আরো উদাহরণ

প্রতিরূপ_আইডি

শব্দার্থবিদ্যা

Produces replica_id of the current process.

আউটপুট

নাম টাইপ
result 0-dimensional tensor of type ui32

উদাহরণ

%result = "stablehlo.replica_id"() : () -> tensor<ui32>

আরো উদাহরণ

নতুন আকার দেওয়া

শব্দার্থবিদ্যা

Performs reshape of operand tensor to a result tensor. Conceptually, it amounts to keeping the same canonical representation but potentially changing the shape, eg from tensor<2x3xf32> to tensor<3x2xf32> or tensor<6xf32> .

More formally, result[result_index] = operand[operand_index] where result_index and operand_index have the same position in the lexicographic ordering of index_space(result) and index_space(operand) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or quantized tensor (C1-C3)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or quantized tensor (C1-C3)

সীমাবদ্ধতা

  • (C1) element_type(result) is given by:
    • element_type(operand) , if !is_per_axis_quantized(operand) .
    • element_type(operand) except that quantization_dimension(operand) and quantization_dimension(result) may differ, otherwise.
  • (C2) size(operand) = size(result) .
  • (C3) If is_per_axis_quantized(operand) :
    • reduce(dims(operand, [0, 1, ..., quantization_dimension(operand) - 1]), init_values=1, dimensions=[0], body=lambda x, y: x * y) = reduce(dims(result, [0, 1, ..., quantization_dimension(result) - 1]), init_values=1, dimensions=[0], body=lambda x, y: x * y) .
    • dim(operand, quantization_dimension(operand)) = dim(result, quantization_dimension(result)) .
    • reduce(dims(operand, [quantization_dimension(operand) + 1, ..., rank(operand) - 1]), init_values=1, dimensions=[0], body=lambda x, y: x * y) = reduce(dims(result, [quantization_dimension(result) + 1, ..., rank(result) - 1]), init_values=1, dimensions=[0], body=lambda x, y: x * y) .

উদাহরণ

// %operand: [[1, 2, 3], [4, 5, 6]]
%result = "stablehlo.reshape"(%operand) : (tensor<2x3xi32>) -> tensor<3x2xi32>
// %result: [[1, 2], [3, 4], [5, 6]]

আরো উদাহরণ

বিপরীত

শব্দার্থবিদ্যা

Reverses the order of elements in the operand along the specified dimensions and produces a result tensor. More formally, result[result_index] = operand[operand_index] where:

  • operand_index[d] = dim(result, d) - result_index[d] - 1 if d in dimensions .
  • operand_index[d] = result_index[d] otherwise.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or per-tensor quantized tensor (C1), (C3)
(I2) dimensions 1-dimensional tensor constant of type si64 (C2), (C3)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C1), (C3)

সীমাবদ্ধতা

  • (C1) type(operand) = type(result) .
  • (C2) is_unique(dimensions) .
  • (C3) 0 <= dimensions < rank(result) .

উদাহরণ

// %operand = [[1, 2], [3, 4], [5, 6]]
%result = "stablehlo.reverse"(%operand) {
  dimensions = array<i64: 1>
} : (tensor<3x2xi32>) -> tensor<3x2xi32>
// %result: [[2, 1], [4, 3], [6, 5]]

আরো উদাহরণ

আরএনজি

শব্দার্থবিদ্যা

Generates random numbers using the rng_distribution algorithm and produces a result tensor of a given shape shape .

If rng_distribution = UNIFORM , then the random numbers are generated following the uniform distribution over the interval [a, b) . If a >= b , the behavior is undefined.

If rng_distribution = NORMAL , then the random numbers are generated following the normal distribution with mean = a and standard deviation = b . If b < 0 , the behavior is undefined.

The exact way how random numbers are generated is implementation-defined. For example, they may or may not be deterministic, and they may or may not use hidden state.

In conversations with many stakeholders, this op has come up as effectively deprecated, so in the future we are planning to explore removing it ( #597 ).

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) a 0-dimensional tensor of integer, boolean, or floating-point type (C1), (C2)
(I2) b 0-dimensional tensor of integer, boolean, or floating-point type (C1), (C2)
(I3) shape 1-dimensional tensor constant of type si64 (C3)
(I4) rng_distribution enum of UNIFORM and NORMAL (C2)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of integer, boolean, or floating-point type (C1-C3)

সীমাবদ্ধতা

  • (C1) element_type(a) = element_type(b) = element_type(result) .
  • (C2) If rng_distribution = NORMAL , then is_float(a) .
  • (C3) shape(result) = shape .

উদাহরণ

// %a = 0
// %b = 2
// %shape = [3, 3]
%result = "stablehlo.rng"(%a, %b, %shape) {
  rng_distribution = #stablehlo<rng_distribution UNIFORM>
} : (tensor<i32>, tensor<i32>, tensor<2xi64>) -> tensor<3x3xi32>
// %result: [
//           [1, 0, 1],
//           [1, 1, 1],
//           [0, 0, 0]
//          ]

rng_bit_generator

শব্দার্থবিদ্যা

Returns an output filled with uniform random bits and an updated output state output_state using the pseudorandom number generator algorithm rng_algorithm given an initial state initial_state . The output is guaranteed to be deterministic function of initial_state , but it is not guaranteed to be deterministic between implementations.

rng_algorithm is one of the following:

  • DEFAULT : Implementation-defined algorithm.
  • THREE_FRY : Implementation-defined variant of the Threefry algorithm.*
  • PHILOX : Implementation-defined variant of the Philox algorithm.*

* See: Salmon et al. SC 2011. Parallel random numbers: as easy as 1, 2, 3.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) rng_algorithm enum of DEFAULT , THREE_FRY , and PHILOX (C2)
(I2) initial_state 1-dimensional tensor of type ui64 (C1), (C2)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
output_state 1-dimensional tensor of type ui64 (C1)
output tensor of integer or floating-point type

সীমাবদ্ধতা

  • (C1) type(initial_state) = type(output_state) .
  • (C2) size(initial_state) is defined as:
    • implementation-defined if rng_algorithm = DEFAULT .
    • 2 if rng_algorithm = THREE_FRY .
    • 2 or 3 if rng_algorithm = PHILOX .

উদাহরণ

// %initial_state: [1, 2]
%output_state, %output = "stablehlo.rng_bit_generator"(%initial_state) {
  rng_algorithm = #stablehlo<rng_algorithm THREE_FRY>
} : (tensor<2xui64>) -> (tensor<2xui64>, tensor<2x2xui64>)
// %output_state: [1, 6]
// %output: [
//           [9236835810183407956, 16087790271692313299],
//           [18212823393184779219, 2658481902456610144]
//          ]

round_nearest_afz

শব্দার্থবিদ্যা

Performs element-wise rounding towards the nearest integer, breaking ties away from zero, on the operand tensor and produces a result tensor. Implements the roundToIntegralTiesToAway operation from the IEEE-754 specification. For quantized types, performs dequantize_op_quantize(round_nearest_afz, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand = [-2.5, 0.4, 0.5, 0.6, 2.5]
%result = "stablehlo.round_nearest_afz"(%operand) : (tensor<5xf64>) -> tensor<5xf64>
// %result: [-3.0, 0.0, 1.0, 1.0, 3.0]

আরো উদাহরণ

round_nearest_even

শব্দার্থবিদ্যা

Performs element-wise rounding towards the nearest integer, breaking ties towards the even integer, on the operand tensor and produces a result tensor. Implements the roundToIntegralTiesToEven operation from the IEEE-754 specification. For quantized types, performs dequantize_op_quantize(round_nearest_even, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand = [-2.5, 0.4, 0.5, 0.6, 2.5]
%result = "stablehlo.round_nearest_even"(%operand) : (tensor<5xf64>) -> tensor<5xf64>
// %result: [-2.0, 0.0, 0.0, 1.0, 2.0]

আরো উদাহরণ

rsqrt

শব্দার্থবিদ্যা

Performs element-wise reciprocal square root operation on operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For floats: rSqrt from IEEE-754.
  • For complex numbers: complex reciprocal square root.
  • For quantized types: dequantize_op_quantize(rsqrt, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [[1.0, 4.0], [9.0, 25.0]]
%result = "stablehlo.rsqrt"(%operand) : (tensor<2x2xf32>) -> tensor<2x2xf32>
// %result: [[1.0, 0.5], [0.33333343, 0.2]]

আরো উদাহরণ

বিক্ষিপ্ত

শব্দার্থবিদ্যা

Produces results tensors which are equal to inputs tensors except that several slices specified by scatter_indices are updated with the values updates using update_computation .

The following diagram shows how elements in updates... map on elements in results... using a concrete example. The diagram picks a few example updates... indices and explains in detail which results... indices they correspond to.

বিক্ষিপ্ত

More formally, for all update_index in index_space(updates[0]) :

  • update_scatter_dims = [d for d in axes(updates[0]) and d not in update_window_dims] .
  • update_scatter_index = update_index[update_scatter_dims...] .
  • start_index is defined as:
    • scatter_indices[si0, ..., :, ..., siN] where si are individual elements in update_scatter_index and : is inserted at the index_vector_dim index, if index_vector_dim < rank(scatter_indices) .
    • [scatter_indices[update_scatter_index]] otherwise.
  • For d_input in axes(inputs[0]) ,
    • full_start_index[d_input] = start_index[d_start] if d_input = scatter_dims_to_operand_dims[d_start] .
    • full_start_index[d_input] = 0 otherwise.
  • For d_input in axes(inputs[0]) ,
    • full_batching_index[d_input] = update_scatter_index[d_start - (d_start < index_vector_dim ? 0 : 1)] if d_input = input_batching_dims[i_batching] and d_start = scatter_indices_batching_dims[i_batching] .
    • full_batching_index[d_input] = 0 otherwise.
  • update_window_index = update_index[update_window_dims...] .
  • full_window_index = [wi0, ..., 0, ..., wiN] where wi are individual elements in update_window_index , and 0 is inserted at indices from inserted_window_dims and input_batching_dims .
  • result_index = full_start_index + full_batching_index + full_window_index .

Given that, results = exec(schedule, inputs) , where:

  • schedule is an implementation-defined permutation of index_space(updates[0]) .
  • exec([update_index, ...], results) = exec([...], updated_results) where:
    • If result_index is in bounds for shape(results...)
    • updates_converted = to_destination_type( updates...[update_index], type(func_inputs(update_computation) [len(func_inputs(update_computation))//2:])... )
    • updated_values = update_computation(results...[result_index], updates_converted)
    • updated_results is a copy of results with results...[result_index] set to updated_values... .
    • অন্যথায়
    • updated_results = results .
  • exec([], results) = results .

If indices_are_sorted is true then the implementation can assume that scatter_indices are sorted with respect to scatter_dims_to_operand_dims , otherwise the behavior is undefined. More formally, for all i1 < i2 from indices(result) , full_start_index(i1) <= full_start_index(i2) .

If unique_indices is true then the implementation can assume that all result_index indices being scattered to are unique. If unique_indices is true but the indices being scattered to are not unique then the behavior is undefined.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) inputs variadic number of tensors or per-tensor quantized tensors (C1), (C2), (C4-C6), (C11), (C13), (C18), (C21), (C23-C24)
(I2) scatter_indices tensor of integer type (C4), (C15), (C19), (C22)
(I3) updates variadic number of tensors or per-tensor quantized tensors (C3-C6), (C8)
(I4) update_window_dims 1-dimensional tensor constant of type si64 (C2), (C4), (C7-C8)
(I5) inserted_window_dims 1-dimensional tensor constant of type si64 (C2), (C4), (C9-C11)
(I6) input_batching_dims 1-dimensional tensor constant of type si64 (C2), (C4), (C9), (C12-13), (C17-18), (C20)
(I7) scatter_indices_batching_dims 1-dimensional tensor constant of type si64 (C14-C18)
(I8) scatter_dims_to_operand_dims 1-dimensional tensor constant of type si64 (C19-C21)
(I9) index_vector_dim constant of type si64 (C4), (C16), (C19), (C22)
(I10) indices_are_sorted constant of type i1
(I11) unique_indices constant of type i1
(I12) update_computation ফাংশন (C23)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
results variadic number of tensors or per-tensor quantized tensors (C24-C25)

সীমাবদ্ধতা

  • (C1) same(shape(inputs...)) .
  • (C2) `rank(inputs[0]) = size(update_window_dims) + size(inserted_window_dims)
    • size(input_batching_dims)`.
  • (C3) same(shape(updates...)) .
  • (C4) shape(updates[0]) = combine(update_scatter_dim_sizes, update_window_dim_sizes) where:
    • update_scatter_dim_sizes = shape(scatter_indices) except that the dimension size of scatter_indices corresponding to index_vector_dim is not included.
    • update_window_dim_sizes <= shape(inputs[0]) except that the dimension sizes in inputs[0] corresponding to inserted_window_dims and input_batching_dims are not included.
    • combine puts update_scatter_dim_sizes at axes corresponding to update_scatter_dims and update_window_dim_sizes at axes corresponding to update_window_dims .
  • (C5) 0 < size(inputs) = size(updates) = N .
  • (C6) element_type(updates...) = element_type(inputs...) .
  • (C7) is_unique(update_window_dims) and is_sorted(update_window_dims) .
  • (C8) 0 <= update_window_dims < rank(updates[0]) .
  • (C9) is_unique(concatenate(inserted_window_dims, input_batching_dims))
  • (C10) is_sorted(inserted_window_dims) .
  • (C11) 0 <= inserted_window_dims < rank(inputs[0]) .
  • (C12) is_sorted(input_batching_dims) .
  • (C13) 0 <= input_batching_dims < rank(inputs[0])) .
  • (C14) is_unique(scatter_indices_batching_dims) .
  • (C15) 0 <= scatter_indices_batching_dims < rank(scatter_indices) .
  • (C16) index_vector_dim not in scatter_indices_batching_dims .
  • (C17) size(input_batching_dims) == size(scatter_indices_batching_dims) .
  • (C18) dim(inputs[0], input_batching_dims...) = dim(scatter_indices, scatter_indices_batching_dims...) .
  • (C19) size(scatter_dims_to_operand_dims) = index_vector_dim < rank(scatter_indices) ? dim(scatter_indices, index_vector_dim) : 1 .
  • (C20) is_unique(concatenate(scatter_dims_to_operand_dims, input_batching_dims)) .
  • (C21) 0 <= scatter_dims_to_operand_dims < rank(inputs[0]) .
  • (C22) 0 <= index_vector_dim <= rank(scatter_indices) .
  • (C23) update_computation has type (tensor<E0>, ..., tensor<EN-1>, tensor<E0>, ..., tensor<EN-1>) -> (tensor<E0>, ..., tensor<EN-1>) , where is_promotable(element_type(inputs[i]), Ei) .
  • (C24) shape(inputs...) = shape(results...) .
  • (C25) element_type(results[i]) = Ei for all i in [0,N) .

উদাহরণ

// %input: [
//          [
//           [[1, 2], [3, 4], [5, 6], [7, 8]],
//           [[9, 10],[11, 12], [13, 14], [15, 16]],
//           [[17, 18], [19, 20], [21, 22], [23, 24]]
//          ],
//          [
//           [[25, 26], [27, 28], [29, 30], [31, 32]],
//           [[33, 34], [35, 36], [37, 38], [39, 40]],
//           [[41, 42], [43, 44], [45, 46], [47, 48]]
//          ]
//         ]
// %scatter_indices: [
//                    [
//                     [[0, 0], [1, 0], [2, 1]],
//                     [[0, 1], [1, 1], [0, 9]]
//                    ],
//                    [
//                     [[0, 0], [2, 1], [2, 2]],
//                     [[1, 2], [0, 1], [1, 0]]
//                    ]
//                   ]
// %update: [
//           [
//            [[1, 1], [1, 1], [1, 1]],
//            [[1, 1], [1, 1], [1, 1]]
//           ],
//           [
//            [[1, 1], [1, 1], [1, 1]],
//            [[1, 1], [1, 1], [1, 1]]
//           ]
//          ]
%result = "stablehlo.scatter"(%input, %scatter_indices, %update) ({
  ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>):
    %0 = "stablehlo.add"(%arg0, %arg1) : (tensor<i64>, tensor<i64>) -> tensor<i64>
    "stablehlo.return"(%0) : (tensor<i64>) -> ()
}) {
  scatter_dimension_numbers = #stablehlo.scatter<
    update_window_dims = [3, 4],
    inserted_window_dims = [1],
    input_batching_dims = [0],
    scatter_indices_batching_dims = [1],
    scatter_dims_to_operand_dims = [2, 1],
    index_vector_dim = 3>,
  indices_are_sorted = false,
  unique_indices = false
} : (tensor<2x3x4x2xi64>, tensor<2x2x3x2xi64>, tensor<2x2x3x2x2xi64>) -> tensor<2x3x4x2xi64>
// %result: [
//           [
//            [[3, 4], [6, 7], [6, 7], [7, 8]],
//            [[9, 10],[11, 12], [15, 16], [17, 18]],
//            [[17, 18], [19, 20], [22, 23], [24, 25]]
//           ],
//           [
//            [[25, 26], [28, 29], [30, 31], [31, 32]],
//            [[35, 36], [38, 39], [38, 39], [39, 40]],
//            [[41, 42], [44, 45], [46, 47], [47, 48]]
//           ]
//          ]

আরো উদাহরণ

নির্বাচন করুন

শব্দার্থবিদ্যা

Produces a result tensor where each element is selected from on_true or on_false tensor based on the value of the corresponding element of pred . More formally, result[result_index] = pred_element ? on_true[result_index] : on_false[result_index] , where pred_element = rank(pred) = 0 ? pred[] : pred[result_index] . For quantized types, performs dequantize_select_quantize(pred, on_true, on_false, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) pred tensor of type i1 (C1)
(I2) on_true tensor or per-tensor quantized tensor (C1-C2)
(I3) on_false tensor or per-tensor quantized tensor (C2)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C2)

সীমাবদ্ধতা

  • (C1) rank(pred) = 0 or shape(pred) = shape(on_true) .
  • (C2) baseline_type(on_true) = baseline_type(on_false) = baseline_type(result) .

উদাহরণ

// %pred: [[false, true], [true, false]]
// %on_true: [[1, 2], [3, 4]]
// %on_false: [[5, 6], [7, 8]]
%result = "stablehlo.select"(%pred, %on_true, %on_false) : (tensor<2x2xi1>, tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32>
// %result: [[5, 2], [3, 8]]

আরো উদাহরণ

select_and_scatter

শব্দার্থবিদ্যা

Scatters the values from the source tensor using scatter based on the outcome of reduce_window of the input tensor using select and produces a result tensor.

The following diagram shows how elements in result are computed from operand and source using a concrete example.

select_and_scatter

আরো আনুষ্ঠানিকভাবে:

  • selected_values = reduce_window_without_init(...) with the following inputs:

    • inputs = [operand].
    • window_dimensions , window_strides , and padding which are used as is.
    • base_dilations = windows_dilations = 1 .
    • body is defined as:
    def body(arg0: tensor<E>, arg1: tensor<E>) -> tensor<E>:
      return select(arg0, arg1) ? arg0 : arg1;
    

    where E = element_type(operand) , and reduce_window_without_init works exactly like reduce_window , except that the schedule of the underlying reduce (see reduce ) doesn't include init values. It is currently unspecified what happens if the corresponding window doesn't have values ( #731 ).

  • result[result_index] = reduce([source_values], [init_value], [0], scatter) where:

    • source_values = [source[source_index] for source_index in source_indices] .
    • selected_index(source_index) = operand_index if selected_values[source_index] has the operand element from operand_index .
    • source_indices = [source_index for source_index in indices(source) if selected_index(source_index) = result_index] .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or per-tensor quantized tensor (C1-C4), (C6), (C8-C11)
(I2) source tensor or per-tensor quantized tensor (C1), (C2)
(I3) init_value 0-dimensional tensor or per-tensor quantized tensor (C3)
(I4) window_dimensions 1-dimensional tensor constant of type si64 (C2), (C4), (C5)
(I5) window_strides 1-dimensional tensor constant of type si64 (C2), (C6), (C7)
(I6) padding 2-dimensional tensor constant of type si64 (C2), (C8)
(I7) select ফাংশন (C9)
(I8) scatter ফাংশন (C10)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C11-C12)

সীমাবদ্ধতা

  • (C1) element_type(operand) = element_type(source) .
  • (C2) shape(source) = num_windows where:
    • padded_operand_shape = padding[:, 0] + shape(operand) + padding[:, 1] .
    • is_empty_window = padded_operand_shape = 0 || window_dimensions > padded_operand_shape .
    • num_windows = is_empty_window ? 0 : floor((padded_operand_shape - window_dimensions) / window_strides) + 1 .
  • (C3) element_type(init_value) = element_type(operand) .
  • (C4) size(window_dimensions) = rank(operand) .
  • (C5) 0 < window_dimensions .
  • (C6) size(window_strides) = rank(operand) .
  • (C7) 0 < window_strides .
  • (C8) shape(padding) = [rank(operand), 2] .
  • (C9) select has type (tensor<E>, tensor<E>) -> tensor<i1> where E = element_type(operand) .
  • (C10) scatter has type (tensor<E>, tensor<E>) -> tensor<E> where is_promotable(element_type(operand), E) .
  • (C11) shape(operand) = shape(result) .
  • (C12) element_type(result) = E .

উদাহরণ

// %operand: [[1, 5], [2, 5], [3, 6], [4, 4]]
// %source: [[5, 6], [7, 8]]
// %init_value: 0
%result = "stablehlo.select_and_scatter"(%operand, %source, %init_value) ({
  ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>):
    %0 = "stablehlo.compare"(%arg0, %arg1) {
      comparison_direction = #stablehlo<comparison_direction GE>
    } : (tensor<i64>, tensor<i64>) -> tensor<i1>
    "stablehlo.return"(%0) : (tensor<i1>) -> ()
}, {
  ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>):
    %0 = "stablehlo.add"(%arg0, %arg1) : (tensor<i64>, tensor<i64>) -> tensor<i64>
    "stablehlo.return"(%0) : (tensor<i64>) -> ()
}) {
  window_dimensions = array<i64: 3, 1>,
  window_strides = array<i64: 2, 1>,
  padding = dense<[[0, 1], [0, 0]]> : tensor<2x2xi64>
} : (tensor<4x2xi64>, tensor<2x2xi64>, tensor<i64>) -> tensor<4x2xi64>
// %result: [[0, 0], [0, 0], [5, 14], [7, 0]]

আরো উদাহরণ

পাঠান

শব্দার্থবিদ্যা

Sends inputs to a channel channel_id and produces a result token.

If is_host_transfer is true , then the operation transfers data to the host. Otherwise, it transfers data to another device. What this means is implementation-defined. This flag duplicates the information provided in channel_type , so in the future we are planning to only keep one of them ( #666 ).

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) inputs variadic number of tensors or quantized tensors
(I2) token token
(I3) channel_id constant of type si64
(I4) channel_type enum of DEVICE_TO_DEVICE and DEVICE_TO_HOST (C1)
(I5) is_host_transfer constant of type i1 (C1)

আউটপুট

নাম টাইপ
result token

সীমাবদ্ধতা

  • (C1) channel_type is defined as:
    • DEVICE_TO_HOST if is_host_transfer = true ,
    • DEVICE_TO_DEVICE otherwise.

উদাহরণ

%result = "stablehlo.send"(%operand, %token) {
  channel_handle = #stablehlo.channel_handle<handle = 1, type = 2>,
  is_host_transfer = true
} : (tensor<2x2xi64>, !stablehlo.token) -> !stablehlo.token

আরো উদাহরণ

shift_left

শব্দার্থবিদ্যা

Performs element-wise left-shift operation on the lhs tensor by rhs number of bits and produces a result tensor.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor of integer type (C1)
(I2) rhs tensor of integer type (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of integer type (C1)

সীমাবদ্ধতা

  • (C1) type(lhs) = type(rhs) = type(result) .

উদাহরণ

// %lhs: [-1, 0, 1]
// %rhs: [1, 2, 3]
%result = "stablehlo.shift_left"(%lhs, %rhs): (tensor<3xi64>, tensor<3xi64>) -> tensor<3xi64>
// %result: [-2, 0, 8]

আরো উদাহরণ

shift_right_arithmetic

শব্দার্থবিদ্যা

Performs element-wise arithmetic right-shift operation on the lhs tensor by rhs number of bits and produces a result tensor.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor of integer type (C1)
(I2) rhs tensor of integer type (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of integer type (C1)

সীমাবদ্ধতা

  • (C1) type(lhs) = type(rhs) = type(result) .

উদাহরণ

// %lhs: [-1, 0, 8]
// %rhs: [1, 2, 3]
%result = "stablehlo.shift_right_arithmetic"(%lhs, %rhs): (tensor<3xi64>, tensor<3xi64>) -> tensor<3xi64>
// %result: [-1, 0, 1]

আরো উদাহরণ

shift_right_logical

শব্দার্থবিদ্যা

Performs element-wise logical right-shift operation on the lhs tensor by rhs number of bits and produces a result tensor.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor of integer type (C1)
(I2) rhs tensor of integer type (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of integer type (C1)

সীমাবদ্ধতা

  • (C1) type(lhs) = type(rhs) = type(result) .

উদাহরণ

// %lhs: [-1, 0, 8]
// %rhs: [1, 2, 3]
%result = "stablehlo.shift_right_logical"(%lhs, %rhs): (tensor<3xi64>, tensor<3xi64>) -> tensor<3xi64>
// %result: [9223372036854775807, 0, 1]

আরো উদাহরণ

চিহ্ন

শব্দার্থবিদ্যা

Returns the sign of the operand element-wise and produces a result tensor. More formally, for each element x , the semantics can be expressed using Python syntax as follows:

def sign(x):
  if is_integer(x):
    if compare(x, 0, LT, SIGNED): return -1
    if compare(x, 0, EQ, SIGNED): return 0
    return 1
  elif is_float(x):
    if is_nan(x): return NaN
    if compare(x, -0.0, EQ, FLOAT): return -0.0
    if compare(x, +0.0, EQ, FLOAT): return +0.0
    if compare(x, 0.0, LT, FLOAT): return -1.0
    return 1.0
  elif is_complex(x):
    if is_nan(real(x)) or is_nan(imag(x)): return (NaN, NaN)
    if compare(x, (0.0, 0.0), EQ, FLOAT): return (0.0, 0.0)
    return divide(x, convert(abs(x), type(x)))

For quantized types, performs dequantize_op_quantize(sign, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of signed integer, floating-point, or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of signed integer, floating-point, or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// Logical values: +NaN, -1.0, -0.0, +0.0, 1.0
// operand: [0x7FFFFFFFFFFFFFFF, -1.0, -0.0, 0.0, 1.0]
%result = "stablehlo.sign"(%operand) : (tensor<5xf64>) -> tensor<5xf64>
// Logical values: +NaN, -1.0, -0.0, +0.0, 1.0
// %result: [0x7FFFFFFFFFFFFFFF, -1.0, -0.0, 0.0, 1.0]

আরো উদাহরণ

সাইন

শব্দার্থবিদ্যা

Performs element-wise sine operation on operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For floats: sin from IEEE-754.
  • For complex numbers: complex sine.
  • For quantized types: dequantize_op_quantize(sine, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [
//            [0.0, 1.57079632],       // [0, pi/2]
//            [3.14159265, 4.71238898] // [pi, 3pi/2]
//           ]
%result = "stablehlo.sine"(%operand) : (tensor<2x2xf32>) -> tensor<2x2xf32>
// %result: [[0.0, 1.0], [0.0, -1.0]]

আরো উদাহরণ

টুকরা

শব্দার্থবিদ্যা

Extracts a slice from the operand using statically-computed starting indices and produces a result tensor. start_indices contain the starting indices of the slice for each dimension, limit_indices contain the ending indices (exclusive) for the slice for each dimension, and strides contain the strides for each dimension.

More formally, result[result_index] = operand[operand_index] where operand_index = start_indices + result_index * strides .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or per-tensor quantized tensor (C1-C3), (C5)
(I2) start_indices 1-dimensional tensor constant of type si64 (C2), (C3), (C5)
(I3) limit_indices 1-dimensional tensor constant of type si64 (C2), (C3), (C5)
(I4) strides 1-dimensional tensor constant of type si64 (C2), (C4)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or per-tensor quantized tensor (C1), (C5)

সীমাবদ্ধতা

  • (C1) element_type(operand) = element_type(result) .
  • (C2) size(start_indices) = size(limit_indices) = size(strides) = rank(operand) .
  • (C3) 0 <= start_indices <= limit_indices <= shape(operand) .
  • (C4) 0 < strides .
  • (C5) shape(result) = ceil((limit_indices - start_indices) / strides) .

উদাহরণ

// %operand: [
//            [0, 0, 0, 0],
//            [0, 0, 1, 1],
//            [0, 0, 1, 1]
//           ]
%result = "stablehlo.slice"(%operand) {
  start_indices = array<i64: 1, 2>,
  limit_indices = array<i64: 3, 4>,
  strides = array<i64: 1, 1>
} : (tensor<3x4xi64>) -> tensor<2x2xi64>
// % result: [
//            [1, 1],
//            [1, 1]
//           ]

আরো উদাহরণ

সাজান

শব্দার্থবিদ্যা

Sorts 1-dimensional slices of inputs along the dimension dimension together, according to a comparator and produces results .

Unlike similar inputs in other operations, dimension allows negative values, with the semantics described below. In the future, this may be disallowed for consistency reasons ( #1377 ).

If is_stable is true, then the sorting is stable, that is, relative order of elements considered to be equal by the comparator is preserved. For the case where there is a single input, two elements e1 and e2 are considered to be equal by the comparator if and only if comparator(e1, e2) = comparator(e2, e1) = false . See the formalization below for how this generalizes to multiple inputs.

More formally, for all result_index in index_space(results[0]) :

  • adjusted_dimension = dimension >= 0 ? dimension : rank(inputs[0]) + dimension .
  • result_slice = [ri0, ..., :, ..., riR-1] where riN are individual elements in result_index , and : is inserted at adjusted_dimension .
  • inputs_together = (inputs[0]..., ..., inputs[N-1]...) .
  • results_together[result_slice] = sort(inputs_together[result_slice], comparator_together) .
  • where sort sorts a 1-dimensional slice in non-descending order expecting that comparator_together returns true if the left-hand side argument is less than the right-hand second argument.
  • def comparator_together(lhs_together, rhs_together):
      args = []
      for (lhs_el, rhs_el) in zip(lhs_together, rhs_together):
        args.append(lhs_el)
        args.append(rhs_el)
      return comparator(*args)
    
  • (results[0]..., ..., results[N-1]...) = results_together .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) inputs variadic number of tensors or per-tensor quantized tensors (C1-C5)
(I2) dimension constant of type si64 (C4)
(I3) is_stable constant of type i1
(I4) comparator ফাংশন (C5)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
results variadic number of tensors or per-tensor quantized tensors (C2), (C3)

সীমাবদ্ধতা

  • (C1) 0 < size(inputs) .
  • (C2) type(inputs...) = type(results...) .
  • (C3) same(shape(inputs...) + shape(results...)) .
  • (C4) -R <= dimension < R , where R = rank(inputs[0]) .
  • (C5) comparator has type (tensor<E1>, tensor<E1>, ..., tensor<EN-1>, tensor<EN-1>) -> tensor<i1> , where Ei = element_type(inputs[i])

উদাহরণ

// %input0 = [[1, 2, 3], [3, 2, 1]]
// %input1 = [[3, 2, 1], [1, 2, 3]]
%result0, %result1 = "stablehlo.sort"(%input0, %input1) ({
  ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>, %arg2: tensor<i64>, %arg3: tensor<i64>):
    %predicate = "stablehlo.compare"(%arg0, %arg1) {
      comparison_direction = #stablehlo<comparison_direction GT>
    } : (tensor<i64>, tensor<i64>) -> tensor<i1>
    "stablehlo.return"(%predicate) : (tensor<i1>) -> ()
}) {
  dimension = 0 : i64,
  is_stable = true
} : (tensor<2x3xi64>, tensor<2x3xi64>) -> (tensor<2x3xi64>, tensor<2x3xi64>)
// %result0 = [[3, 2, 3], [1, 2, 1]]
// %result1 = [[1, 2, 1], [3, 2, 3]]

আরো উদাহরণ

sqrt

শব্দার্থবিদ্যা

Performs element-wise square root operation on operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For floats: squareRoot from IEEE-754.
  • For complex numbers: complex square root.
  • For quantized types: dequantize_op_quantize(sqrt, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [[0.0, 1.0], [4.0, 9.0]]
%result = "stablehlo.sqrt"(%operand) : (tensor<2x2xf32>) -> tensor<2x2xf32>
// %result: [[0.0, 1.0], [2.0, 3.0]]

আরো উদাহরণ

বিয়োগ

শব্দার্থবিদ্যা

Performs element-wise subtraction of two tensors lhs and rhs and produces a result tensor. Depending on the element type, does the following:

  • For integers: integer subtraction.
  • For floats: subtraction from IEEE-754.
  • For complex numbers: complex subtraction.
  • For quantized types:
    • dequantize_op_quantize(subtract, lhs, rhs, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor of integer, floating-point, or complex type or per-tensor quantized tensor (C1)
(I2) rhs tensor of integer, floating-point, or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of integer, floating-point, or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(lhs) = baseline_type(rhs) = baseline_type(result) .

উদাহরণ

// %lhs: [[6, 8], [10, 12]]
// %rhs: [[5, 6], [7, 8]]
%result = "stablehlo.subtract"(%lhs, %rhs) : (tensor<2x2xf32>, tensor<2x2xf32>) -> (tensor<2x2xf32>)
// %result: [[1, 2], [3, 4]]

আরো উদাহরণ

ট্যান

শব্দার্থবিদ্যা

Performs element-wise tangent operation on the operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For floats: tan from IEEE-754.
  • For complex numbers: complex tangent.
  • For quantized types: dequantize_op_quantize(tan, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [
//            [0.0, 1.57079632],       // [0, pi/2]
//            [3.14159265, 4.71238898] // [pi, 3pi/2]
//           ]
%result = "stablehlo.tan"(%operand) : (tensor<2x2xf64>) -> tensor<2x2xf64>
// %result: [
//           [0.0, 1.63312e+16],
//           [0.0, 5.44375e+15]
//          ]

আরো উদাহরণ

তানহ

শব্দার্থবিদ্যা

Performs element-wise hyperbolic tangent operation on operand tensor and produces a result tensor. Depending on the element type, does the following:

  • For floats: tanh from IEEE-754.
  • For complex numbers: complex hyperbolic tangent.
  • For quantized types:
    • dequantize_op_quantize(tanh, operand, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or complex type or per-tensor quantized tensor (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_type(operand) = baseline_type(result) .

উদাহরণ

// %operand: [-1.0, 0.0, 1.0]
%result = "stablehlo.tanh"(%operand) : (tensor<3xf32>) -> tensor<3xf32>
// %result: [-0.76159416, 0.0, 0.76159416]

আরো উদাহরণ

স্থানান্তর

শব্দার্থবিদ্যা

Permutes the dimensions of operand tensor using permutation and produces a result tensor. More formally, result[result_index] = operand[operand_index] where result_index[d] = operand_index[permutation[d]] .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor or quantized tensor (C1-C4)
(I2) permutation 1-dimensional tensor constant of type si64 (C2-C4)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor or quantized tensor (C1), (C3-C4)

সীমাবদ্ধতা

  • (C1) element_type(result) is given by:
    • element_type(operand) , if !is_per_axis_quantized(operand) .
    • element_type(operand) except that quantization_dimension(operand) and quantization_dimension(result) may differ, otherwise.
  • (C2) permutation is a permutation of range(rank(operand)) .
  • (C3) shape(result) = dim(operand, permutation...) .
  • (C4) If is_per_axis_quantized(result) , then quantization_dimension(operand) = permutation(quantization_dimension(result)) .

উদাহরণ

// %operand: [
//            [[1,2], [3,4], [5,6]],
//            [[7,8], [9,10], [11,12]]
//           ]
%result = "stablehlo.transpose"(%operand) {
  permutation = array<i64: 2, 1, 0>
} : (tensor<2x3x2xi32>) -> tensor<2x3x2xi32>
// %result: [
//           [[1,7], [3,9], [5,11]],
//           [[2,8], [4,10], [6,12]]
//          ]

আরো উদাহরণ

triangular_solve

শব্দার্থবিদ্যা

Solves batches of systems of linear equations with lower or upper triangular coefficient matrices.

More formally, given a and b , result[i0, ..., iR-3, :, :] is the solution to op(a[i0, ..., iR-3, :, :]) * x = b[i0, ..., iR-3, :, :] when left_side is true or x * op(a[i0, ..., iR-3, :, :]) = b[i0, ..., iR-3, :, :] when left_side is false , solving for the variable x where op(a) is determined by transpose_a , which can be one of the following:

  • NO_TRANSPOSE : Perform operation using a as-is.
  • TRANSPOSE : Perform operation on transpose of a .
  • ADJOINT : Perform operation on conjugate transpose of a .

Input data is read only from the lower triangle of a , if lower is true or upper triangle of a , otherwise. Output data is returned in the same triangle; the values in the other triangle are implementation-defined.

If unit_diagonal is true, then the implementation can assume that the diagonal elements of a are equal to 1, otherwise the behavior is undefined.

For quantized types, performs dequantize_op_quantize(lambda x, y: triangular_solve(x, y, left_side, lower, unit_diagonal, transpose_a), a, b, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) a tensor of floating-point or complex type or per-tensor quantized tensor (C1-C3)
(I2) b tensor of floating-point or complex type or per-tensor quantized tensor (C1-C4)
(I3) left_side constant of type i1 (C3)
(I4) lower constant of type i1
(I5) unit_diagonal constant of type i1
(I6) transpose_a enum of NO_TRANSPOSE , TRANSPOSE , and ADJOINT

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point or complex type or per-tensor quantized tensor (C1)

সীমাবদ্ধতা

  • (C1) baseline_element_type(a) = baseline_element_type(b) .
  • (C2) 2 <= rank(a) = rank(b) = R .
  • (C3) The relationship between shape(a) and shape(b) is defined as follows:
    • shape(a)[:-3] = shape(b)[:-3] .
    • dim(a, -2) = dim(a, -1) = dim(b, left_side ? -2 : -1) .
  • (C4) baseline_type(b) = baseline_type(result) .

উদাহরণ

// %a = [
//       [1.0, 0.0, 0.0],
//       [2.0, 4.0, 0.0],
//       [3.0, 5.0, 6.0]
//      ]
// %b = [
//       [2.0, 0.0, 0.0],
//       [4.0, 8.0, 0.0],
//       [6.0, 10.0, 12.0]
//      ]
%result = "stablehlo.triangular_solve"(%a, %b) {
  left_side = true,
  lower = true,
  unit_diagonal = false,
  transpose_a = #stablehlo<transpose NO_TRANSPOSE>
} : (tensor<3x3xf32>, tensor<3x3xf32>) -> tensor<3x3xf32>
// %result: [
//           [2.0, 0.0, 0.0],
//           [0.0, 2.0, 0.0],
//           [0.0, 0.0, 2.0]
//          ]

টিপল

শব্দার্থবিদ্যা

Produces a result tuple from values val .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) val variadic number of values (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result টিপল (C1)

সীমাবদ্ধতা

  • (C1) result has type tuple<E0, ..., EN-1> where Ei = type(val[i]) .

উদাহরণ

// %val0: [1.0, 2.0]
// %val1: (3)
%result = "stablehlo.tuple"(%val0, %val1) : (tensor<2xf32>, tuple<tensor<i32>>) -> tuple<tensor<2xf32>, tuple<tensor<i32>>>
// %result: ([1.0, 2.0], (3))

আরো উদাহরণ

uniform_dequantize

শব্দার্থবিদ্যা

Performs element-wise conversion of quantized tensor operand to a floating-point tensor result according to the quantization parameters defined by the operand type.

More formally, result = dequantize(operand) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand quantized tensor (C1), (C2)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of floating-point type (C1), (C2)

সীমাবদ্ধতা

  • (C1) shape(operand) = shape(result) .
  • (C2) element_type(result) = expressed_type(operand) .

উদাহরণ

// %operand: [10, 10]
%result = "stablehlo.uniform_dequantize"(%operand) : (tensor<2x!quant.uniform<i8:f32:0, {0.1:-30,0.5:-20}>>) -> tensor<2xf32>
// %result: [4.0, 15.0]

uniform_quantize

শব্দার্থবিদ্যা

Performs element-wise conversion of floating-point tensor or quantized tensor operand to a quantized tensor result according to the quantization parameters defined by the result type.

আরো আনুষ্ঠানিকভাবে,

  • If is_float(operand) :
    • result = quantize(operand, type(result)) .
  • If is_quantized(operand) :
    • float_result = dequantize(operand) .
    • result = quantize(float_result, type(result)) .

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand tensor of floating-point or quantized type (C1), (C2)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result quantized tensor (C1), (C2)

সীমাবদ্ধতা

  • (C1) shape(operand) = shape(result) .
  • (C2) expressed_type(result) = is_float(operand) ? element_type(operand) : expressed_type(operand) .

উদাহরণ

// %operand: [4.0, 15.0]
%result = "stablehlo.uniform_quantize"(%operand) : (tensor<2xf32>) -> tensor<2x!quant.uniform<i8:f32:0, {0.1:-30,0.5:-20}>>
// %result: [10, 10]

// %operand: [10, 10]
%result = "stablehlo.uniform_quantize"(%operand) : (tensor<2x!quant.uniform<i8:f32:0, {0.1:-30,0.5:-20}>>) -> tensor<2x!quant.uniform<i8:f32:0, {0.1:-20,0.2:-30}>>
// %result: [20, 45]

যখন

শব্দার্থবিদ্যা

Produces the output from executing body function 0 or more times while the cond function outputs true . More formally, the semantics can be expressed using Python syntax as follows:

internal_state = operand
while cond(*internal_state):
  internal_state = body(*internal_state)
results = internal_state

The behavior of an infinite loop is TBD ( #383 ).

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) operand variadic number of tensors, quantized tensors or tokens (C1-C3)
(I2) cond ফাংশন (C1)
(I3) body ফাংশন (C2)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
results variadic number of tensors, quantized tensors or tokens (C3)

সীমাবদ্ধতা

  • (C1) cond has type (T0, ..., TN-1) -> tensor<i1> , where Ti = type(operand[i]) .
  • (C2) body has type (T0, ..., TN-1) -> (T0, ..., TN-1) , where Ti = type(operand[i]) .
  • (C3) type(results...) = type(operand...) .

উদাহরণ

// %init_i: 1
// %init_sum: 0
// %one: 1
// %ten: 10
%results0, %results1 = "stablehlo.while"(%init_i, %init_sum) ({
  ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>):
    %cond = "stablehlo.compare"(%arg0, %ten) {
      comparison_direction = #stablehlo<comparison_direction LT>
    } : (tensor<i64>, tensor<i64>) -> tensor<i1>
    stablehlo.return %cond : tensor<i1>
  }, {
  ^bb0(%arg0: tensor<i64>, %arg1: tensor<i64>):
    %new_sum = stablehlo.add %arg1, %one : tensor<i64>
    %new_i = stablehlo.add %arg0, %one : tensor<i64>
    stablehlo.return %new_i, %new_sum : tensor<i64>, tensor<i64>
}) : (tensor<i64>, tensor<i64>) -> (tensor<i64>, tensor<i64>)
// %results0: 10
// %results1: 10

আরো উদাহরণ

xor

শব্দার্থবিদ্যা

Performs element-wise XOR of two tensors lhs and rhs and produces a result tensor. Depending on the element type, does the following:

  • For booleans: logical XOR.
  • For integers: bitwise XOR.

ইনপুট

লেবেল নাম টাইপ সীমাবদ্ধতা
(I1) lhs tensor of boolean or integer type (C1)
(I2) rhs tensor of boolean or integer type (C1)

আউটপুট

নাম টাইপ সীমাবদ্ধতা
result tensor of boolean or integer type (C1)

সীমাবদ্ধতা

  • (C1) type(lhs) = type(rhs) = type(result) .

উদাহরণ

// Bitwise operation with with integer tensors
// %lhs: [[1, 2], [3, 4]]
// %rhs: [[5, 6], [7, 8]]
%result = "stablehlo.xor"(%lhs, %rhs) : (tensor<2x2xi32>, tensor<2x2xi32>) -> tensor<2x2xi32>
// %result: [[4, 4], [4, 12]]

// Logical operation with with boolean tensors
// %lhs: [[false, false], [true, true]]
// %rhs: [[false, true], [false, true]]
%result = "stablehlo.xor"(%lhs, %rhs) : (tensor<2x2xi1>, tensor<2x2xi1>) -> tensor<2x2xi1>
// %result: [[false, true], [true, false]]

আরো উদাহরণ

Dialect Interop

At the moment, StableHLO programs in the wild sometimes contain operations that are not defined by StableHLO.

Module, Function, Call and Return

StableHLO uses upstream MLIR operations for ModuleOp, FuncOp, CallOp, and ReturnOp. This was done for better interop with existing MLIR machinery, as many useful passes are written targeting FuncOp and ModuleOp, and many compilation pipelines expect these ops to be present. Full compatibility guarantees are applied to these ops. If anything ever changes about these ops in an incompatible way (ie removal), StableHLO equivalents will be added to preserve compatibility.

CHLO

The CHLO opset contains higher level operations that decompose to StableHLO. Currently there are no compatibility guarantees for CHLO. For compatibility guarantees, the chlo-legalize-to-stablehlo pass must be used prior to serialization.

Shape Operations

It is a common use case in the community to use certain operations from core MLIR dialects in dynamic StableHLO programs to perform shape computations. Most commonly, these include shape dialect ops like shape_of or num_elements , tensor dialect ops like dim or from_elements , and the builtin index type.

The Dynamism RFC > O2 denotes these as out of scope, however some support for index types is included for interop purposes. There are no compatibility guarantees for these ops or types. The shape-legalize-to-stablehlo pass can be used to convert these operations to fully supported StableHLO ops.

Deprecated Operations

There are several StableHLO operations that were inherited from MHLO which are deprecated and on the way out of StableHLO. The full details on these removals can be found in the StableHLO v1.0 Cleanup #2283 . The tracker issue for these deprecations is #2340 .

These operations fall into a few categories:

  • "Not in HLO" category of StableHLO operations - they were initially part of the StableHLO opset but have been later deemed to not fit it well: broadcast , create_token , cross-replica-sum , dot , einsum , torch_index_select , unary_einsum ( #3 ) .
  • Unused ops - These operations may have been useful at some point, but the ops were either underdeveloped, or the pipelines using these ops have been refactored to not require them anymore. This includes map , tuple ( #598 ), get_tuple_element , rng , complex comparisons #560 , and convolution window_reversal ( #1181 ).

Some of these ops can be removed easily given that they can be expressed using existing ops ( broadcast , create_token , cross-replica-sum , dot , unary_einsum ) and will be removed after the existing compatibilty window passes (6 months). Others are still being explored for removal ( einsum , get_tuple_element , map , rng torch_index_select , tuple , complex comparisons, window_reversal ). Pending community feedback, these ops will either be removed, or added to the spec with full support. Until these ops futures are known, they are only guaranteed 6 months of compatibility.

মৃত্যুদন্ড

অনুক্রমিক মৃত্যুদন্ড

A StableHLO program is executed by providing input values to the main function and computing output values. Output values of a function are computed by executing the graph of ops rooted in the corresponding return op.

The execution order is implementation-defined as long as it is aligned with dataflow, ie if ops are executed before their uses. In StableHLO, all side-effecting ops consume one token and produce one token (multiple tokens can be multiplexed into one token via after_all ), so the execution order of side effects is also aligned with dataflow. For example, in the below program there are two possible execution orders: %0%1%2return and %1%0%2return .

func.func @main() -> tensor<f64> {
  %0 = stablehlo.constant dense<1.0> : tensor<f64>
  %1 = stablehlo.constant dense<2.0> : tensor<f64>
  %2 = stablehlo.add %0, %1 : tensor<f64>
  return %2 : tensor<f64>
}

More formally, a StableHLO process is a combination of: 1) a StableHLO program, 2) operation statuses (not executed yet, already executed), and 3) intermediate values that the process is working on. The process starts with input values to the main function, progresses through the graph of ops updating operation statuses and intermediate values and finishes with output values. Further formalization is TBD ( #484 ).

Parallel execution

StableHLO programs can be executed in parallel, organized into a 2D process grid of num_replicas by num_partitions which both have type ui32 .

In the StableHLO process grid , num_replicas * num_partitions of StableHLO processes are executing at the same time. Each process has a unique process_id = (replica_id, partition_id) , where replica_id in replica_ids = range(num_replicas) and partition_id in partition_ids = range(num_partitions) which both have type ui32 .

The size of the process grid is known statically for every program (in the future, we are planning to make it an explicit part of StableHLO programs #650 ), and the position within the process grid is known statically for every process. Each process has access to its position within the process grid via the replica_id and partition_id ops.

Within the process grid, the programs can all be the same (in the "Single Program, Multiple Data" style), can all be different (in the "Multiple Program, Multiple Data" style) or something in between. In the future, we are planning to introduce support for other idioms of defining parallel StableHLO programs, including GSPMD ( #619 ).

Within the process grid, the processes are mostly independent from each other - they have separate operation statuses, separate input/intermediate/output values and most of the ops are executed separately between processes, with the exception of a small number of collective ops described below .

Given that execution of most of the ops is only using values from the same process, it is usually unambiguous to refer to these values by their names. However, when describing semantics of collective ops, that is insufficient, and that gives rise to the notation name@process_id to refer to the value name within a particular process. (From that perspective, unqualified name can be viewed as a shorthand for name@(replica_id(), partition_id()) ).

The execution order across processes is implementation-defined, except for the synchronization introduced by point-to-point communication and collective ops as described below.

পয়েন্ট টু পয়েন্ট যোগাযোগ

StableHLO processes can communicate with each other through StableHLO channels . A channel is represented by a positive id of type si64 . Through various ops, it is possible to send values to channels and receive them from channels.

Further formalization, eg where these channel ids are coming from, how processes programs become aware of them and what kind of synchronization is introduced by them, is TBD ( #484 ).

Streaming communication

Every StableHLO process has access to two streaming interfaces:

  • Infeed that can be read from.
  • Outfeed that can be written to.

Unlike channels, which are used to communicate between processes and therefore have processes at both of their ends, infeeds and outfeeds have their other end implementation-defined.

Further formalization, eg how streaming communication influences execution order and what kind of synchronization is introduced by it, is TBD ( #484 ).

Collective ops

There are six collective ops in StableHLO: all_gather , all_reduce , all_to_all , collective_broadcast , collective_permute , and reduce_scatter . All these ops split the processes in the StableHLO process grid into StableHLO process groups and execute a joint computation within each process group, independently from other process groups.

Within each process group, collective ops may introduce a synchronization barrier. Further formalization, eg elaborating on when exactly this synchronization happens, how exactly the processes arrive at this barrier, and what happens if they don't, is TBD ( #484 ).

If the process group involves cross-partition communication, ie there are processes in the process group whose partition ids are different, then execution of the collective op needs a channel, and the collective op must provide a positive channel_id of type si64 . Cross-replica communication doesn't need channels.

The computations performed by the collective ops are specific to individual ops and are described in individual op sections above. However, the strategies by which the process grid is split into process groups are shared between these ops and are described in this section. More formally, StableHLO supports the following four strategies.

cross_replica

Only cross-replica communications happen within each process group. This strategy takes replica_groups - a list of lists of replica ids - and computes a Cartesian product of replica_groups by partition_ids . replica_groups must have unique elements and cover all replica_ids . More formally, using Python syntax:

def cross_replica(replica_groups: List[List[ReplicaId]]) -> List[List[ProcessId]]:
  for replica_group in replica_groups:
    for partition_id in partition_ids:
      process_group = []
      for replica_id in replica_group:
        process_group.append((replica_id, partition_id))
      yield process_group

For example, for replica_groups = [[0, 1], [2, 3]] and num_partitions = 2 , cross_replica will produce [[(0, 0), (1, 0)], [(0, 1), (1, 1)], [(2, 0), (3, 0)], [(2, 1), (3, 1)]] .

cross_partition

Only cross-partition communications happen within each process group. This strategy takes partition_groups - a list of lists of partition ids - and computes a Cartesian product of partition_groups by replica_ids . partition_groups must have unique elements and cover all partition_ids . More formally, using Python syntax:

def cross_partition(partition_groups: List[List[PartitionId]]) -> List[List[ProcessId]]:
  for partition_group in partition_groups:
    for replica_id in replica_ids:
      process_group = []
      for partition_id in partition_group:
        process_group.append((replica_id, partition_id))
      yield process_group

For example, for partition_groups = [[0, 1]] and num_replicas = 4 , cross_partition will produce [[(0, 0), (0, 1)], [(1, 0), (1, 1)], [(2, 0), (2, 1)], [(3, 0), (3, 1)]] .

cross_replica_and_partition

Both cross-replica and cross-partition communications may happen within each process group. This strategy takes replica_groups - a list of lists of replica ids - and computes Cartesian products of each replica_group by partition_ids . replica_groups must have unique elements and cover all replica_ids . More formally, using Python syntax:

def cross_replica_and_partition(replica_groups: List[List[ReplicaId]]) -> List[List[ProcessId]]:
  for replica_group in replica_groups:
    process_group = []
    for partition_id in partition_ids:
      for replica_id in replica_group:
        process_group.append((replica_id, partition_id))
    yield process_group

For example, for replica_groups = [[0, 1], [2, 3]] and num_partitions = 2 , cross_replica_and_partition will produce [[(0, 0), (1, 0), (0, 1), (1, 1)], [(2, 0), (3, 0), (2, 1), (3, 1)]] .

flattened_ids

This strategy takes flattened_id_groups - a list of lists of "flattened" process ids in the form of replica_id * num_partitions + partition_id - and turns them into process ids. flattened_id_groups must have unique elements and cover all process_ids . More formally, using Python syntax:

def flattened_ids(flattened_id_groups: List[List[ui32]]) -> List[List[ProcessId]]:
  for flattened_id_group in flattened_id_groups:
    process_group = []
    for flattened_id in flattened_id_group:
      replica_id = flattened_id // num_partitions
      partition_id = flattened_id % num_partitions
      process_group.append((replica_id, partition_id))
    yield process_group

For example, for flattened_id_groups = [[0, 1, 2, 3], [4, 5, 6, 7]] , num_replicas = 4 and num_partitions = 2 , flattened_ids will produce [[(0, 0), (0, 1), (1, 0), (1, 1)], [(2, 0), (2, 1), (3, 0), (3, 1)]] .

নির্ভুলতা

At the moment, StableHLO does not provide guarantees about numerical accuracy, but this may change in the future ( #1156 ).

Execution semantics of quantized operation

The interpretation of quantized StableHLO operations may vary depending on the hardware requirements and capabilities. For instance, some hardware may opt to interpret quantized operations using a "dequantize, perform floating-point operation, and finally quantize" strategy. Others may perform the entire computation with integer arithmetic. Consequently, the interpretation of quantized StableHLO operations is exclusively determined by the specific implementation. The interpretation of hybrid quantization ( #1575 ) should be based on the it's semantics as prescribed in the specification (via 1792 ).

ত্রুটি

StableHLO programs are validated through an extensive set of constraints for individual ops, which rules out many classes of errors prior to run time. However, error conditions are still possible, eg through integer overflows, out-of-bounds accesses, etc. Unless explicitly called out, all these errors result in implementation-defined behavior, but this may change in the future ( #1157 ).

ফ্লোটিং পয়েন্ট ব্যতিক্রম

As an exception to this rule, floating-point exceptions in StableHLO programs have well-defined behavior. Operations which result in exceptions defined by the IEEE-754 standard (invalid operation, division-by-zero, overflow, underflow, or inexact exceptions) produce default results (as defined in the standard) and continue execution without raising the corresponding status flag; similar to raiseNoFlag exception handling from the standard. Exceptions for nonstandard operations (eg complex arithmetic and certain transcendental functions) are implementation-defined.

Shape mismatches

StableHLO supports dynamically-shaped tensors. However, shapes have to agree at runtime, otherwise the behavior is undefined. StableHLO does not explicitly provide an op that can assert that a tensor has a given shape at runtime. Generating correct code is the responsibility of the producer.

As a specific example, the below program is valid. However, at runtime, the exact shapes of %arg0 and %arg1 will have to be the same, otherwise the behavior of the program is undefined:

func.func @foo(%arg0: tensor<?xi32>, %arg1: tensor<?xi32>) -> tensor<?xi32> {
    %0 = stablehlo.add %arg0, %arg1 : tensor<?xi32>
    return %0 : tensor<?xi32>
}

স্বরলিপি

For describing syntax, this document is using the modified ISO flavor of EBNF syntax ( ISO/IEC 14977:1996 , Wikipedia ), with two modifications: 1) rules are defined using ::= rather than = ,

2) concatenation is expressed using juxtaposition rather than , .

For describing semantics (ie within "Types", "Constants" and "Ops" sections), we are using formulas which are based on Python syntax extended with support for concisely expressing array operations as described below. This works well for small snippets of code, but in rare cases when larger snippets of code are needed, we use vanilla Python syntax which is always introduced explicitly.

সূত্র

Let's explore how formulas work based on an example from the dot_general specification. One of the constraints for this operation looks as follows: dim(lhs, lhs_batching_dimensions...) = dim(rhs, rhs_batching_dimensions...) .

The names used in this formula come from two sources: 1) global functions, ie dim , 2) member definitions of the corresponding program element, ie lhs , lhs_batching_dimensions , rhs and rhs_batching_dimensions inputs defined in the "Inputs" section of dot_general .

As mentioned above, the syntax of this formula is Python-based with some conciseness-oriented extensions. To make sense of the formula, let's transform it into vanilla Python syntax.

A) In these formulas, we are using = to represent equality, so the first step towards obtaining Python syntax is replacing = with == , as follows: dim(lhs, lhs_batching_dimensions...) == dim(rhs, rhs_batching_dimensions...)

B) Also, these formulas support ellipses ( ... ) which turn scalar expressions into tensor expressions. In a nutshell, f(xs...) roughly means "for each scalar x in the tensor xs , compute a scalar f(x) and then return all these scalar results together as a tensor result". In vanilla Python syntax, our example formula turns into: [dim(lhs, dim1) for dim1 in lhs_batching_dimensions] == [dim(rhs, dim2) for dim2 in rhs_batching_dimensions] .

Thanks to ellipses, it is often possible to avoid working at the level of individual scalars. However, in some tricky cases, lower-level semi-informal syntax may be used like in the start_indices[bi0, ..., :, ..., biN] formula from the gather specification. In the service of conciseness, we don't provide an exact formalism for translating such syntax to vanilla Python, in hopes that it is still intuitively understandable on case-by-case basis. Please let us know if some specific formulas look opaque, and we'll try to improve them.

Also, you will notice that formulas use ellipses to expand all sorts of lists, including tensors, lists of tensors (which eg can arise from a variadic number of tensors), etc. This is another area where we don't provide an exact formalism (eg lists are not even part of the StableHLO type system) and instead rely on intuitive understandability.

C) The final noteworthy notational vehicle that we employ is implicit broadcasting. While the StableHLO opset doesn't support implicit broadcasting, the formulas do, also in the service of conciseness. In a nutshell, if a scalar is used in a context where a tensor is expected, the scalar is broadcasted to the expected shape.

To continue the dot_general example, here's another constraint: 0 <= lhs_batching_dimensions < rank(lhs) . As defined in the dot_general specification, lhs_batching_dimensions is a tensor, however both 0 and rank(lhs) are scalars. After we apply implicit broadcasting, the formula will become [0, ..., 0] <= lhs_batching_dimensions < [rank(lhs), ..., rank(lhs)] .

When applied to a particular dot_general operation, this formula will evaluate to a tensor of booleans. When formulas are used as constraints, the constraint holds if the formula evaluates to either true or to a tensor which only has true elements.

নাম

In formulas, lexical scope includes: 1) global functions, 2) member definitions,

3) local definitions. The list of global functions is provided below. The list of element definitions depends on the program element that the notation is applied to:

  • For operations, member definitions include names introduced in "Inputs" and "Outputs" sections.
  • For everything else, member definitions include structural parts of the program element, named after the corresponding EBNF non-terminals. Most of the time, the names of these structural parts are obtained by converting the names of the non-terminals to snake case (eg IntegerLiteral => integer_literal ), but sometimes names get abbreviated in the process (eg QuantizationStorageType => storage_type ) in which case the names are introduced explicitly similarly to "Inputs" / "Outputs" sections in operation specifications.
  • Additionally, member definitions always include self to refer to the corresponding program element.

মূল্যবোধ

When formulas are evaluated, they work with the following types of values: 1) Value (actual values, eg dense<[[1, 2], [3, 4]]> : tensor<2x2xi32> ; they always know their types), 2) Placeholder (future values, eg lhs , rhs or result ; their actual values are not known yet, only their types are known), 3) Type (types as defined in the "Types" section), 4) Function (global functions as defined in the "Functions" section).

Depending on the context, names may be referring to different values. More specifically, the "Semantics" section for ops (and equivalents for other program elements) defines runtime logic, so all inputs are available as Value . In contrast, the "Constraints" section for ops (and equivalents) defines "compile-time" logic, ie something that is typically executed before runtime, so only constant inputs are available as Value and other inputs are available only as Placeholder .

নাম In "Semantics" In "Constraints"
Global functions Function Function
Constant inputs Value Value
Non-constant inputs Value Placeholder
আউটপুট Value Placeholder
স্থানীয় সংজ্ঞা Depends on the definition Depends on the definition

Let's consider an example transpose operation:

%result = "stablehlo.transpose"(%operand) {
  permutation = dense<[2, 1, 0]> : tensor<3xi64>
} : (tensor<2x3x2xi32>) -> tensor<2x3x2xi32>

For this operation, permutation is a constant, so it's available as a Value in both semantics and constraints. In contrast, operand and result are available as a Value in semantics but only as a Placeholder in constraints.

ফাংশন

Construction of types

There are no functions that can be used to construct types. Instead, we directly use type syntax because it's typically more concise. Eg (tensor<E>, tensor<E>) -> (tensor<E>) rather than function_type( [tensor_type([], E), tensor_type([], E)], [tensor_type([], E)])

Functions on types

  • element_type is defined on tensor types and quantized tensor types and returns, respectively, the TensorElementType or QuantizedTensorElementType part of the corresponding TensorType or QuantizedTensorType .
def element_type(x: Value | Placeholder | Type):
 if type(x) == TensorType:
    return tensor_element_type(x)
  if type(x) == QuantizedTensorType:
    return quantized_tensor_element_type(x)
  if type(x) is not Type:
    return element_type(type(x))
  • is_per_axis_quantized(x: Value | Placeholder | Type) -> Value is a shortcut for is_quantized(x) and quantization_dimension(x) is not None .

  • is_per_tensor_quantized(x: Value | Placeholder | Type) -> Value is a shortcut for is_quantized(x) and quantization_dimension(x) is None .

  • is_promotable(x: Type, y: Type) -> bool checks if type x can be promoted to type y . When x and y are QuantizedTensorElementType s, the promotion is applied only to the storage_type . This specific version of promotion is currently used in context of reduction computation (refer to RFC for more details).

def is_promotable(x: Type, y: Type) -> Value:
  is_same_type = (is_bool(x) and is_bool(y)) or
    (is_integer(x) and is_integer(y)) or (is_float(x) and is_float(y)) or
    (is_complex(x) and is_complex(y)) or
    (is_quantized(x) and is_quantized(y) and expressed_type(x) = expressed_type(y))

  if is_same_type == False:
    return False

  if is_integer(x) or is_float(x):
    return bitwidth(x) <= bitwidth(y)

  if is_complex(x):
    return bitwidth(element_type(x)) <= bitwidth(element_type(y))

  if is_quantized(x):
    return bitwidth(storage_type(x)) <= bitwidth(storage_type(y))

  return false
  • is_quantized(x: Value | Placeholder | Type) -> Value is a shortcut for is_quantized_tensor_element_type(x) .

  • is_type_name(x: Value | Placeholder | Type) -> Value . Available for all types. For example, is_float(x) returns true if x is a FloatType . If x is a value or placeholder, this function is a shortcut for is_type_name(type(x)) .

  • max_value(x: Type) -> Value returns the maximum value of an TensorElementType . If x is not an TensorElementType , returns None .

  • min_value(x: Type) -> Value returns the minimum possible value of an TensorElementType . If x is not an TensorElementType , returns None .

  • member_name(x: Value | Placeholder | Type) -> Any . Available for all member definitions member_name of all types. For example, tensor_element_type(x) returns the TensorElementType part of a corresponding TensorType . If x is a value or placeholder, this function is a shortcut for member_name(type(x)) . If x is not a type that has an appropriate member, or a value or a placeholder of such a type, returns None .

  • is_empty_algorithm(*args: Type) checks if all dot algorithm fields are set to None . This is needed since dot algorithms have implementation defined default behaviors, so specifying a default value would be incorrect.

Construction of values

  • operation_name(*xs: Value | Type) -> Value . Available for all operations. For example, add(lhs, rhs) takes two tensor values lhs and rhs and returns the output of evaluating the add operation with these inputs. For some operations eg broadcast_in_dim , types of their outputs are "load-bearing", ie needed to evaluate an operation. In this case, the function takes these types as arguments.

Functions on values

  • All Python's operators and functions are available. Eg both subscription and slicing notations from Python are available to index into tensors, quantized tensors and tuples.

  • to_destination_type(x: Value, destination_type: Type) -> Value is defined on tensors and returns the converted value of x based on the type(x) and destination_type as follows:

def to_destination_type(x: Value, destination_type: Type) -> Value:
  if type(x) == destination_type:
    return x

  if is_quantized(destination_type):
    if is_quantized(type(x)):
      return quantize(x, destination_type)
    assert is_float(type(x))
    return quantize(x, destination_type)

  if is_quantized(type(x)):
    assert destination_type = expressed_type(type(x))
    return dequantize(type(x))

  return convert(x, destination_type)

There is early discussion on merging convert , uniform_quantize and uniform_dequantize operations ( #1576 ). After the merge we do not need the above function and can use the operation name for convert instead.

  • is_nan(x: Value) -> Value is defined on tensors and returns true if all elements of x are NaN or false otherwise. If x is not a tensor, returns None .

  • is_sorted(x: Value) -> Value is defined on tensors and returns true if elements of x are sorted in ascending order with respect to the ascending lexicographical order of their indices or false otherwise. If x is not a tensor, returns None .

  • is_unique(x: Value) -> Value is defined on tensors and returns true if x doesn't have duplicate elements or false otherwise. If x is not a tensor, returns None .

  • member_name(x: Value) -> Any is defined for all member definitions member_name of all values. For example, real_part(x) returns the RealPart part of a corresponding ComplexConstant . If x is not a value that has an appropriate member, returns None .

  • same(x: Value) -> Value is defined on tensors and returns true if elements of x are all equal to each other or false otherwise. If the tensor doesn't have elements, that counts as "all equal to each other", ie the function returns true . If x is not a tensor, returns None .

  • split(x: Value, num_results: Value, axis: Value) -> Value is defined on tensors and returns num_results slices of x along the axis axis . If x is not a tensor or dim(x, axis) % num_results != 0 , returns None .

  • is_defined_in_parent_scope(x: Value) -> Value is defined on strings and returns true if x is the name of a function defined in the same scope as the parent function of the relevant op.

  • is_namespaced_op_name(x: Value) -> Value is defined on strings and returns true if x is a valid op name, that is it respects the following regular expression: [a-zA-Z][a-zA-Z0-9_]*([.][a-zA-Z0-9_$]+)+

Shape computations

  • axes(x: Value | Placeholder | Type) -> Value is a shortcut for range(rank(x)) .

  • dim(x: Value | Placeholder | Type, axis: Value) -> Value is a shortcut for shape(x)[axis] .

  • dims(x: Value | Placeholder | Type, axes: List) -> List is a shortcut for list(map(lambda axis: dim(x, axis), axes)) .

  • index_space(x: Value | Placeholder | Type) -> Value is defined on tensors and returns size(x) indices for the corresponding TensorType sorted in ascending lexicographical order, ie [0, ..., 0] , [0, ..., 1] , ..., shape(x) - 1 . If x is not a tensor type, a quantized tensor type, or a value or a placeholder of one of these types, returns None .

  • rank(x: Value | Placeholder | Type) -> Value is a shortcut for size(shape(x)) .

  • shape(x: Value | Placeholder | Type) -> Value is defined in the "Functions on types" section via member_name .

  • size(x: Value | Placeholder | Type) -> Value is a shortcut for reduce(lambda x, y: x * y, shape(x)) .

Quantization computations

  • def baseline_element_type(x: Value | Placeholder | Type) -> Type is a shortcut for element_type(baseline_type(x)) .

  • baseline_type is defined on tensor types and quantized tensor types and transforms them to a "baseline", ie a type with the same shape but with the quantization parameters of the element type reset to default values. This is used as a handy trick to compare both tensor and quantized tensor types uniformly, which is needed quite often. For quantized types, this enables comparing types ignoring the quantization parameters, that is, shape , storage_type , expressed_type , storage_min , storage_max , and quantization_dimension (for per-axis quantized type) must all match, but scales and zero points may differ.

def baseline_type(x: Value | Placeholder | Type) -> Type:
  if type(x) == TensorType:
    return x
  if type(x) == QuantizedTensorType:
    element_type = quantized_tensor_element_type(x)
    baseline_element_type = QuantizedTensorElementType(
      storage_type = storage_type(element_type),
      storage_min = storage_min(element_type),
      storage_max = storage_max(element_type),
      expressed_type = expressed_type(element_type),
      quantization_dimension = quantization_dimension(element_type),
      scales = [constant(1.0, expressed_type(element_type))] * dim(x, quantization_dimension(element_type)),
      zero_points = [constant(0, storage_type(element_type))] * dim(x, quantization_dimension(element_type)))
    return QuantizedTensorType(shape(x), baseline_element_type)
  if type(x) is not Type:
    return baseline_element_type(type(x))
  • dequantize is defined on quantized tensor types and turns them into floating-point tensor types. This happens via converting quantized elements which represent integer values of the storage type into corresponding floating-point values of the expressed type using the zero point and scale associated with the quantized element type.
def compute_zero_points(quantized_type, result_type):
  if is_per_tensor_quantized(quantized_type):
    return broadcast_in_dim(constant(zero_point(quantized_type), storage_type(quantized_type)), [], result_type)
  if is_per_axis_quantized(quantized_type):
    for i in index_space(result_type):
      d = quantization_dimension(quantized_type)
      zero_points[i] = zero_points(quantized_type)[i[d]]
    return zero_points

def compute_scales(quantized_type, result_type):
  if is_per_tensor_quantized(quantized_type):
    return broadcast_in_dim(constant(scale(quantized_type), expressed_type(quantized_type)), [],
            type(result_type))
  if is_per_axis_quantized(quantized_type):
    for i in index_space(result_type):
      d = quantization_dimension(quantized_type)
      scales[i] = scales(quantized_type)[i[d]]
    return scales

def dequantize(x: Value) -> Value:
  assert is_quantized(x)
  x_storage = bitcast_convert(x, storage_type(x))
  x_storage_sub = x_storage - compute_zero_points(type(x), type(x_storage))
  x_expressed_sub = convert(x_storage_sub, expressed_type(x))
  return x_expressed_sub * compute_scales(type(x), type(x_expressed_sub))
  • quantize is defined on floating-point tensor types and turns them into quantized tensor types. This happens via converting floating-point values of the expressed type into corresponding integer values of the storage type using the zero point and scale associated with the quantized element type.
def quantize(x: Value, result_type: Type) -> Value:
  assert is_float(x) and is_quantized(result_type)
  zero_points = compute_zero_points(result_type, TensorType(shape(x), storage_type(result_type)))
  converted_zero_points = convert(zero_points, expressed_type(result_type))
  converted_min = convert(storage_min(result_type), expressed_type(result_type))
  converted_max = convert(storage_max(result_type), expressed_type(result_type))

  x_scaled = x / compute_scales(result_type, type(x))
  x_scaled_add_zp = x_scaled + converted_zero_points
  x_clamped = clamp(converted_min, x_scaled_add_zp, converted_max)
  x_rounded = round_nearest_even(x_clamped)
  return convert(x_rounded, result_type)
  • dequantize_op_quantize is used to specify element-wise computations on quantized tensors. It dequantizes, ie turns quantized elements into their expressed types, then performs an operation, and then quantizes, ie turns the results back into their storage types. At the moment, this function only works for per-tensor quantization. Per-axis quantization is work in progress ( #1574 ).
def dequantize_op_quantize(op, *inputs_and_output_type):
  inputs = inputs_and_output_type[:-1]
  output_type = inputs_and_output_type[-1]

  float_inputs = map(dequantize, inputs)
  float_result = op(*float_inputs)
  return quantize(float_result, output_type)

def dequantize_batch_norm_grad_or_training_quantize(op, *inputs_and_output_types):
  inputs = inputs_and_output_type[:-3]
  float_inputs = map(dequantize, inputs)
  float_results = op(*float_inputs)
  return map(quantize, float_results, inputs_and_output_type[-3:])

def dequantize_compare(lhs, rhs, comparison_direction):
  float_lhs = dequantize(lhs)
  float_rhs = dequantize(rhs)
  return compare(float_lhs, float_rhs, comparison_direction, FLOAT)

def dequantize_select_quantize(pred, on_true, on_false, output_type):
  float_on_true = dequantize(on_true)
  float_on_false = dequantize(on_false)
  float_result = select(pred, float_on_true, float_on_false)
  return quantize(float_result, output_type)
  • hybrid_dequantize_then_op is used to specify weight-only quantization for hybrid op which accepts lhs in floating-point and rhs in quantized types. It dequantizes quantized inputs into their expressed types and performs computation in float. Element type of float lhs tensor and expressed type of quantized rhs tensor should be identical.
def hybrid_dequantize_then_op(op, lhs, rhs):
  assert(is_float(lhs) and is_quantized(rhs) and element_type(lhs) == expressed_type(rhs))
  return op(lhs, dequantize(rhs))

Grid computations

  • cross_partition(replica_groups: Value) -> Value . See the "cross_replica" section above.

  • cross_replica(replica_groups: Value) -> Value . See the "cross_replica" section above.

  • cross_replica_and_partition(replica_groups: Value) -> Value . See the "cross_replica_and_partition" section above.

  • flattened_ids(replica_groups: Value) -> Value . See the "flattened_ids" section above.

গতিশীলতা

StableHLO values can have dynamic dimension sizes, eg tensor<?xi64> . However, StableHLO values cannot have a dynamic number of dimensions (unranked dynamism, eg tensor<*xi64> ). Operands and results are allowed to use dynamic dimension sizes, even if there are constraints on the sizes. Constraints will be verified statically if possible, otherwise they are deferred to runtime and mismatches will result in undefined behavior. উদাহরণ জন্য নীচে দেখুন.

Shape mismatches for unary elementwise operations

Consider the following toy program:

func.func @foo(%arg0: tensor<?xf64>) {
  %0 = stablehlo.abs %arg0 : (tensor<?xf64>) -> tensor<2xf64>
  return
}

Such a program is unusual, because it is not common to know the shape of the result but not the shape of the input. Nonetheless, this is a valid StableHLO program. It is not possible to statically validate the abs operation in this program, because the exact shape of the operand is unknown. However, the shapes are certainly compatible, and this can be checked statically: ? could turn out to be 2 at runtime, and there would be no issue. However, ? could also turn out to be some other integer, in which case the behavior is undefined.

Note that if a dimension size is dynamic in the result, there cannot be undefined behavior. Indeed, there is no "expected" size, so there cannot be a mismatch.

Shape mismatches for binary elementwise operations

Consider the following toy program:

func.func @foo(%arg0: tensor<?xf64>, %arg1: tensor<?xf64>) {
  %0 = stablehlo.add %arg0, %arg0 : (tensor<?xf64>, tensor<?xf64>) -> tensor<?xf64>
  return
}

When it comes to binary elementwise operations, the shapes of the inputs and the result must agree at runtime. At compile time, static dimensions must be equal, otherwise they merely need to be compatible. If any dimension is dynamic in the inputs, then there could be undefined behavior at runtime, because the dynamic size may not match the corresponding size in the other operand (be it static or dynamic). If all the inputs are static, then whether the result is dynamic or not does not matter: statically known dimensions will be checked statically, and dynamic dimensions do not impose any constraints.

Shape mismatches for ops that take their output shape as an operand

Consider the following toy program:

func.func @foo(%arg0: tensor<2xi32>) {
  %0 = stablehlo.dynamic_iota %arg0, dim = 0 : (tensor<2xi32>) -> tensor<3x4xi64>
  return
}

The values in the shape operand at runtime must match the shape of the result, otherwise the behavior is undefined. That is, at runtime %arg0 must have a value of dense<[3, 4]> : tensor<2xi32> . If the shape operand is constant, this can be verified statically. If the result shape is fully dynamic, then there cannot be a mismatch.