!4895 benchmark add fp16 flag for gate

Merge pull request !4895 from zhaozhenlong/lite/issue/gate_benchmark_comment_failed

mindspore/lite/include/context.h

@@ -65,7 +65,7 @@ class MS_API Context {
   virtual ~Context();
 
  public:
-  bool float16_priority = true; /**< allow priority select float16 kernel */
+  bool float16_priority = false; /**< allow priority select float16 kernel */
   DeviceContext device_ctx_{DT_CPU};
   int thread_num_ = 2; /**< thread number config for thread pool */
   std::shared_ptr<Allocator> allocator = nullptr;

mindspore/lite/src/runtime/kernel/arm/fp16/arithmetic_fp16.cc

@@ -446,21 +446,21 @@ kernel::LiteKernel *CpuArithmeticFp16KernelCreator(const std::vector<lite::tenso
   return kernel;
 }
 
-REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Mul, CpuArithmeticFp16KernelCreator)
-REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Add, CpuArithmeticFp16KernelCreator)
-REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Sub, CpuArithmeticFp16KernelCreator)
-REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Div, CpuArithmeticFp16KernelCreator)
-REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_FloorMod, CpuArithmeticFp16KernelCreator)
-REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_FloorDiv, CpuArithmeticFp16KernelCreator)
-REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_LogicalAnd, CpuArithmeticFp16KernelCreator)
-REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_LogicalOr, CpuArithmeticFp16KernelCreator)
-REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Maximum, CpuArithmeticFp16KernelCreator)
-REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Minimum, CpuArithmeticFp16KernelCreator)
-REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_NotEqual, CpuArithmeticFp16KernelCreator)
-REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Equal, CpuArithmeticFp16KernelCreator)
-REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Less, CpuArithmeticFp16KernelCreator)
-REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_LessEqual, CpuArithmeticFp16KernelCreator)
-REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Greater, CpuArithmeticFp16KernelCreator)
-REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_GreaterEqual, CpuArithmeticFp16KernelCreator)
-REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Eltwise, CpuArithmeticFp16KernelCreator)
+// REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Mul, CpuArithmeticFp16KernelCreator)
+// REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Add, CpuArithmeticFp16KernelCreator)
+// REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Sub, CpuArithmeticFp16KernelCreator)
+// REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Div, CpuArithmeticFp16KernelCreator)
+// REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_FloorMod, CpuArithmeticFp16KernelCreator)
+// REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_FloorDiv, CpuArithmeticFp16KernelCreator)
+// REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_LogicalAnd, CpuArithmeticFp16KernelCreator)
+// REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_LogicalOr, CpuArithmeticFp16KernelCreator)
+// REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Maximum, CpuArithmeticFp16KernelCreator)
+// REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Minimum, CpuArithmeticFp16KernelCreator)
+// REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_NotEqual, CpuArithmeticFp16KernelCreator)
+// REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Equal, CpuArithmeticFp16KernelCreator)
+// REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Less, CpuArithmeticFp16KernelCreator)
+// REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_LessEqual, CpuArithmeticFp16KernelCreator)
+// REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Greater, CpuArithmeticFp16KernelCreator)
+// REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_GreaterEqual, CpuArithmeticFp16KernelCreator)
+// REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Eltwise, CpuArithmeticFp16KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/src/runtime/kernel/arm/fp16/convolution_fp16.cc

@@ -273,5 +273,5 @@ kernel::LiteKernel *CpuConvFp16KernelCreator(const std::vector<lite::tensor::Ten
   }
   return kernel;
 }
-REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Conv2D, CpuConvFp16KernelCreator)
+// REG_KERNEL(kCPU, kNumberTypeFloat16, PrimitiveType_Conv2D, CpuConvFp16KernelCreator)
 }  // namespace mindspore::kernel

mindspore/lite/tools/benchmark/benchmark.cc

@@ -396,6 +396,7 @@ int Benchmark::RunBenchmark(const std::string &deviceType) {
     context->cpu_bind_mode_ = NO_BIND;
   }
   context->thread_num_ = _flags->numThreads;
+  context->float16_priority = _flags->fp16Priority;
   session = session::LiteSession::CreateSession(context);
   delete (context);
   if (session == nullptr) {
@@ -503,6 +504,7 @@ int Benchmark::Init() {
   MS_LOG(INFO) << "AccuracyThreshold = " << this->_flags->accuracyThreshold;
   MS_LOG(INFO) << "WarmUpLoopCount = " << this->_flags->warmUpLoopCount;
   MS_LOG(INFO) << "NumThreads = " << this->_flags->numThreads;
+  MS_LOG(INFO) << "Fp16Priority = " << this->_flags->fp16Priority;
   MS_LOG(INFO) << "calibDataPath = " << this->_flags->calibDataPath;
 
   if (this->_flags->loopCount < 1) {

mindspore/lite/tools/benchmark/benchmark.h

@@ -63,6 +63,7 @@ class MS_API BenchmarkFlags : public virtual FlagParser {
     // MarkPerformance
     AddFlag(&BenchmarkFlags::loopCount, "loopCount", "Run loop count", 10);
     AddFlag(&BenchmarkFlags::numThreads, "numThreads", "Run threads number", 2);
+    AddFlag(&BenchmarkFlags::fp16Priority, "fp16Priority", "Priority float16", false);
     AddFlag(&BenchmarkFlags::warmUpLoopCount, "warmUpLoopCount", "Run warm up loop", 3);
     // MarkAccuracy
     AddFlag(&BenchmarkFlags::calibDataPath, "calibDataPath", "Calibration data file path", "");
@@ -88,6 +89,7 @@ class MS_API BenchmarkFlags : public virtual FlagParser {
   // MarkPerformance
   int loopCount;
   int numThreads;
+  bool fp16Priority;
   int warmUpLoopCount;
   // MarkAccuracy
   std::string calibDataPath;