Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Revert "disable gemm f16 on CPU (#19744)" #26

Draft
wants to merge 1 commit into
base: main
Choose a base branch
from
Draft

Revert "disable gemm f16 on CPU (#19744)" #26

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 0 additions & 14 deletions onnxruntime/core/optimizer/layer_norm_fusion.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -462,13 +462,6 @@ Status LayerNormFusion::ApplyImpl(Graph& graph, bool& modified, int graph_level,

NodeArg* x_input = has_leading_cast ? graph.GetNode(p_reduce_mean_input_node->Index())->MutableInputDefs()[0]
: reduce_mean_node.MutableInputDefs()[0];

// CPU doesn't support fp16
if (reduce_mean_node.GetExecutionProviderType() == kCpuExecutionProvider &&
x_input->TypeAsProto()->tensor_type().elem_type() == ONNX_NAMESPACE::TensorProto_DataType_FLOAT16) {
continue;
}

InlinedVector<NodeArg*> layer_norm_input_defs{x_input, scale, bias};
Node& layer_norm_node = graph.AddNode(graph.GenerateNodeName(mul_node.Name() + "/LayerNormFusion/"),
"LayerNormalization",
Expand Down Expand Up @@ -711,13 +704,6 @@ Status SimplifiedLayerNormFusion::ApplyImpl(Graph& graph, bool& modified, int gr

NodeArg* x_input = has_leading_cast ? graph.GetNode(p_pow_input_node->Index())->MutableInputDefs()[0]
: pow_node.MutableInputDefs()[0];

// CPU doesn't support fp16
if (reduce_mean_node.GetExecutionProviderType() == kCpuExecutionProvider &&
x_input->TypeAsProto()->tensor_type().elem_type() == ONNX_NAMESPACE::TensorProto_DataType_FLOAT16) {
continue;
}

InlinedVector<NodeArg*> layer_norm_input_defs{x_input, scale};
Node& layer_norm_node =
graph.AddNode(graph.GenerateNodeName(mul_node.Name() + "/SimplifiedLayerNormFusion/"), "SimplifiedLayerNormalization",
Expand Down
21 changes: 21 additions & 0 deletions onnxruntime/core/providers/cpu/cpu_execution_provider.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -144,6 +144,9 @@ class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 7, Aco
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 7, Atan);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 7, 8, float, Gemm);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 7, 8, double, Gemm);
#ifdef MLAS_F16VEC_INTRINSICS_SUPPORTED
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 7, 8, MLFloat16, Gemm);
#endif
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 1, 10, Hardmax);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 1, 10, float, LogSoftmax);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 1, 10, double, LogSoftmax);
Expand Down Expand Up @@ -344,6 +347,9 @@ class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOn
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 9, 10, Flatten);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 9, 10, float, Gemm);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 9, 10, double, Gemm);
#ifdef MLAS_F16VEC_INTRINSICS_SUPPORTED
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 9, 10, MLFloat16, Gemm);
#endif
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 9, 12, float, MatMul);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 9, 12, double, MatMul);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 9, 12, int32_t, MatMul);
Expand Down Expand Up @@ -514,6 +520,9 @@ class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 11, Sp
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 11, 12, ScatterND);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 11, 12, float, Gemm);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 11, 12, double, Gemm);
#ifdef MLAS_F16VEC_INTRINSICS_SUPPORTED
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 11, 12, MLFloat16, Gemm);
#endif
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 11, 12, GatherElements);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 11, uint8_t, BitShift);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 11, uint32_t, BitShift);
Expand Down Expand Up @@ -620,6 +629,9 @@ class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain,
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 13, string, Expand);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 13, float, Gemm);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 13, double, Gemm);
#ifdef MLAS_F16VEC_INTRINSICS_SUPPORTED
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 13, MLFloat16, Gemm);
#endif
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 13, float, MatMul);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 13, double, MatMul);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 13, int32_t, MatMul);
Expand Down Expand Up @@ -2853,6 +2865,15 @@ Status RegisterFp16Kernels(KernelRegistry& kernel_registry) {
MLFloat16, LeakyRelu)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 16, MLFloat16,
LeakyRelu)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 7, 8,
MLFloat16, Gemm)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 9, 10,
MLFloat16, Gemm)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 11, 12,
MLFloat16, Gemm)>,

BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCpuExecutionProvider, kOnnxDomain, 13, MLFloat16,
Gemm)>,
};

for (auto& function_table_entry : function_table) {
Expand Down
2 changes: 1 addition & 1 deletion onnxruntime/test/providers/cpu/math/gemm_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -383,7 +383,7 @@ class GemmOpTypedTests : public ::testing::Test {
// On CPUs without fp16 instructions the tests will output a warning:
// "registered execution providers CPUExecutionProvider were unable to run the model"
// , then they will still pass.
using GemmOpTypedTestsTypes = ::testing::Types<float, double>;
using GemmOpTypedTestsTypes = ::testing::Types<float, double, MLFloat16>;
TYPED_TEST_SUITE(GemmOpTypedTests, GemmOpTypedTestsTypes);

TYPED_TEST(GemmOpTypedTests, TestGemmScalarBroadcast) {
Expand Down
Loading