כתיבת בדיקות יחידה למעברים של HLO

TEST_F(PriorityFusionTest, FuseBroadcastIntoBitcastConsumers) {
  absl::string_view kHlo = R"(
    HloModule test_module

    // CHECK: ENTRY main
    ENTRY main {
      // CHECK-NEXT: %[[PARAM:.*]] = f32[96]{0} parameter(0)
      param_0 = f32[96]{0} parameter(0)
      broadcast = f32[8,96,128,7]{3,2,1,0} broadcast(param_0), dimensions={1}
      bitcast.6079.2 = f32[8,24,4,128,7]{4,3,2,1,0} bitcast(broadcast)
      // CHECK-NEXT: ROOT %{ {.*} } fusion(%[[PARAM]]) { {.*} }
      ROOT transpose.1990.2 = f32[8,24,128,7,4]{4,3,2,1,0} transpose(bitcast.6079.2), dimensions={0,1,3,4,2}
    }
  )";
  RunAndFilecheckHloRewrite(kHlo, std::move(priority_fusion_));
}

// RUN: hlo-opt %s --platform=gpu --stage=llvm-before-optimizations --xla_gpu_target_config_filename=%S/../../../backends/gpu/specs/%{GPU}.txtpb | FileCheck --check-prefixes=CHECK-%{PTX} %s

HloModule Test, is_scheduled=true
fused_computation {
  param_0 = f32[100,200]{1,0} parameter(0)
  ROOT b.1 = f32[200,100]{1,0} transpose(f32[100,200]{1,0} param_0), dimensions={1,0}
}
ENTRY main {
  a = f32[100, 200]{1,0} parameter(0)
  // CHECK-PTX:         call void @llvm.nvvm.barrier0
  // CHECK-GCN:         call void @llvm.amdgcn.s.barrier
  ROOT wrapped_b = f32[200,100]{1,0} fusion(f32[100,200]{1,0} a), kind=kInput, calls=fused_computation
}