saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-07-30 20:18:08 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

fix bugs in cpuid_info (#10334)

Browse source 
* fix serveral bugs in cpuid_info
This commit is contained in:
Yufeng Li 2022-01-20 16:30:18 -08:00 committed by GitHub
parent 2dcb69685e
commit d2b1424968
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
18 changed files with 64 additions and 69 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

15
onnxruntime/core/common/cpuid_info.cc
View file

@ -1,13 +1,9 @@
// Copyright (c) Microsoft Corporation. All rights reserved.
// Licensed under the MIT License.
#if defined(_M_IX86) || (defined(_M_X64) && !defined(_M_ARM64EC)) || defined(__i386__) || defined(__x86_64__)
#define CPUIDINFO_ARCH_X86
#endif
#if defined(_M_ARM64) || defined(__aarch64__) || defined(_M_ARM) || defined(__arm__)
#define CPUIDINFO_ARCH_ARM
#endif
#include "core/common/cpuid_info.h"
#include "core/common/logging/logging.h"
#include "core/common/logging/severity.h"
#if defined(CPUIDINFO_ARCH_X86)
#include <memory>
@ -30,9 +26,6 @@
#endif
#include <mutex>
#include "core/common/cpuid_info.h"
#include "core/common/logging/logging.h"
#include "core/common/logging/severity.h"
#if _WIN32
#define HAS_WINDOWS_DESKTOP WINAPI_FAMILY_PARTITION(WINAPI_PARTITION_DESKTOP)
@ -65,8 +58,6 @@ static inline int XGETBV() {
}
#endif // CPUIDINFO_ARCH_X86
CPUIDInfo CPUIDInfo::instance_;
CPUIDInfo::CPUIDInfo() {
#if (defined(CPUIDINFO_ARCH_X86) || defined(CPUIDINFO_ARCH_ARM)) && defined(CPUINFO_SUPPORTED)

10
onnxruntime/core/common/cpuid_info.h
View file

@ -5,6 +5,14 @@
#include "core/common/common.h"
#if defined(_M_IX86) || (defined(_M_X64) && !defined(_M_ARM64EC)) || defined(__i386__) || defined(__x86_64__)
#define CPUIDINFO_ARCH_X86
#endif
#if defined(_M_ARM64) || defined(__aarch64__) || defined(_M_ARM) || defined(__arm__)
#define CPUIDINFO_ARCH_ARM
#endif
namespace onnxruntime {
class CPUIDInfo {
@ -41,8 +49,6 @@ class CPUIDInfo {
bool pytorch_cpuinfo_init_{false};
#endif
bool has_arm_neon_dot_{false};
static CPUIDInfo instance_;
};
} // namespace onnxruntime

12
onnxruntime/core/mlas/lib/compute.cpp
View file

@ -271,7 +271,7 @@ Return Value:
--*/
{
#if defined(MLAS_TARGET_AMD64)
MlasPlatform.ComputeExpF32Kernel(Input, Output, N);
GetMlasPlatform().ComputeExpF32Kernel(Input, Output, N);
#else
MlasComputeExpF32Kernel(Input, Output, N);
#endif
@ -850,7 +850,7 @@ Return Value:
//
#if defined(MLAS_TARGET_AMD64)
float Maximum = MlasPlatform.ReduceMaximumF32Kernel(Input, D);
float Maximum = GetMlasPlatform().ReduceMaximumF32Kernel(Input, D);
#else
float Maximum = MlasReduceMaximumF32Kernel(Input, D);
#endif
@ -863,7 +863,7 @@ Return Value:
//
#if defined(MLAS_TARGET_AMD64)
float Accumulation = MlasPlatform.ComputeSumExpF32Kernel(Input, nullptr, D, &NegativeMaximum);
float Accumulation = GetMlasPlatform().ComputeSumExpF32Kernel(Input, nullptr, D, &NegativeMaximum);
#else
float Accumulation = MlasComputeSumExpF32Kernel(Input, nullptr, D, &NegativeMaximum);
#endif
@ -875,7 +875,7 @@ Return Value:
float Parameters[] = { NegativeMaximum, std::log(Accumulation)};
#if defined(MLAS_TARGET_AMD64)
MlasPlatform.ComputeLogSoftmaxOutputF32Kernel(Input, Output, D, Parameters);
GetMlasPlatform().ComputeLogSoftmaxOutputF32Kernel(Input, Output, D, Parameters);
#else
MlasComputeLogSoftmaxOutputF32Kernel(Input, Output, D, Parameters);
#endif
@ -888,7 +888,7 @@ Return Value:
//
#if defined(MLAS_TARGET_AMD64)
float Accumulation = MlasPlatform.ComputeSumExpF32Kernel(Input, Output, D, &NegativeMaximum);
float Accumulation = GetMlasPlatform().ComputeSumExpF32Kernel(Input, Output, D, &NegativeMaximum);
#else
float Accumulation = MlasComputeSumExpF32Kernel(Input, Output, D, &NegativeMaximum);
#endif
@ -900,7 +900,7 @@ Return Value:
float Parameters[] = { 1.0f / Accumulation };
#if defined(MLAS_TARGET_AMD64)
MlasPlatform.ComputeSoftmaxOutputF32Kernel(Output, D, Parameters);
GetMlasPlatform().ComputeSoftmaxOutputF32Kernel(Output, D, Parameters);
#else
MlasComputeSoftmaxOutputF32Kernel(Output, D, Parameters);
#endif

2
onnxruntime/core/mlas/lib/convsym.cpp
View file

@ -205,7 +205,7 @@ MLAS_FORCEINLINE
const
MLAS_CONV_SYM_DISPATCH*
GetConvSymDispatch(bool InputIsSigned){
return InputIsSigned ? MlasPlatform.ConvSymS8S8Dispatch : MlasPlatform.ConvSymU8S8Dispatch;
return InputIsSigned ? GetMlasPlatform().ConvSymS8S8Dispatch : GetMlasPlatform().ConvSymU8S8Dispatch;
}
size_t

6
onnxruntime/core/mlas/lib/dgemm.cpp
View file

@ -531,7 +531,7 @@ Return Value:
size_t RowsHandled;
#if defined(MLAS_TARGET_AMD64_IX86) || defined (MLAS_TARGET_POWER)
RowsHandled = MlasPlatform.GemmDoubleKernel(A, B, C, CountK, CountM, CountN, lda, ldc, alpha, ZeroMode);
RowsHandled = GetMlasPlatform().GemmDoubleKernel(A, B, C, CountK, CountM, CountN, lda, ldc, alpha, ZeroMode);
#else
if (ZeroMode) {
RowsHandled = MlasDgemmKernelZero(A, B, C, CountK, CountM, CountN, lda, ldc, alpha);
@ -830,10 +830,10 @@ MlasGemmBatch(
ptrdiff_t TargetThreadCount;
if (Complexity < double(MLAS_DGEMM_THREAD_COMPLEXITY * MlasPlatform.MaximumThreadCount)) {
if (Complexity < double(MLAS_DGEMM_THREAD_COMPLEXITY * GetMlasPlatform().MaximumThreadCount)) {
TargetThreadCount = ptrdiff_t(Complexity / double(MLAS_DGEMM_THREAD_COMPLEXITY)) + 1;
} else {
TargetThreadCount = MlasPlatform.MaximumThreadCount;
TargetThreadCount = GetMlasPlatform().MaximumThreadCount;
}
ptrdiff_t MaximumThreadCount = MlasGetMaximumThreadCount(ThreadPool);

2
onnxruntime/core/mlas/lib/erf.cpp
View file

@ -262,7 +262,7 @@ Return Value:
--*/
{
#if defined(MLAS_TARGET_AMD64)
MlasPlatform.ErfKernelRoutine(Input, Output, N);
GetMlasPlatform().ErfKernelRoutine(Input, Output, N);
#else
MlasErfKernel(Input, Output, N);
#endif

2
onnxruntime/core/mlas/lib/logistic.cpp
View file

@ -179,7 +179,7 @@ Return Value:
--*/
{
#if defined(MLAS_TARGET_AMD64)
MlasPlatform.LogisticKernelRoutine(Input, Output, N);
GetMlasPlatform().LogisticKernelRoutine(Input, Output, N);
#else
MlasLogisticKernel(Input, Output, N);
#endif

6
onnxruntime/core/mlas/lib/mlasi.h
View file

@ -841,7 +841,11 @@ struct MLAS_PLATFORM {
};
extern MLAS_PLATFORM MlasPlatform;
inline
MLAS_PLATFORM& GetMlasPlatform(){
static MLAS_PLATFORM MlasPlatform;
return MlasPlatform;
}
//
// Threading support.

8
onnxruntime/core/mlas/lib/platform.cpp
View file

@ -43,12 +43,6 @@ MLASCPUIDInfo::MLASCPUIDInfo() { has_arm_neon_dot_ = ((getauxval(AT_HWCAP) & HWC
#endif
#endif // MLAS_TARGET_ARM64
//
// Stores the platform information.
//
MLAS_PLATFORM MlasPlatform;
#ifdef MLAS_TARGET_AMD64_IX86
//
@ -422,7 +416,7 @@ Return Value:
--*/
{
#if defined(MLAS_TARGET_AMD64)
return MlasPlatform.PreferredBufferAlignment;
return GetMlasPlatform().PreferredBufferAlignment;
#else
return MLAS_DEFAULT_PREFERRED_BUFFER_ALIGNMENT;
#endif

8
onnxruntime/core/mlas/lib/qdwconv.cpp
View file

@ -288,14 +288,14 @@ Return Value:
if (InputIsSigned) {
if (FilterIsSigned) {
MlasPlatform.ConvDepthwiseS8S8Kernel(
GetMlasPlatform().ConvDepthwiseS8S8Kernel(
reinterpret_cast<const int8_t* const*>(Input), static_cast<int8_t>(InputZeroPoint),
reinterpret_cast<const int8_t*>(Filter), static_cast<int8_t>(FilterZeroPoint),
Output, Channels, OutputCount, KernelSize
);
} else {
MlasPlatform.ConvDepthwiseS8U8Kernel(
GetMlasPlatform().ConvDepthwiseS8U8Kernel(
reinterpret_cast<const int8_t* const*>(Input), static_cast<int8_t>(InputZeroPoint),
reinterpret_cast<const uint8_t*>(Filter), static_cast<uint8_t>(FilterZeroPoint),
Output, Channels, OutputCount, KernelSize
@ -304,14 +304,14 @@ Return Value:
} else {
if (FilterIsSigned) {
MlasPlatform.ConvDepthwiseU8S8Kernel(
GetMlasPlatform().ConvDepthwiseU8S8Kernel(
reinterpret_cast<const uint8_t* const*>(Input), static_cast<uint8_t>(InputZeroPoint),
reinterpret_cast<const int8_t*>(Filter), static_cast<int8_t>(FilterZeroPoint),
Output, Channels, OutputCount, KernelSize
);
} else {
MlasPlatform.ConvDepthwiseU8U8Kernel(
GetMlasPlatform().ConvDepthwiseU8U8Kernel(
reinterpret_cast<const uint8_t* const*>(Input), static_cast<uint8_t>(InputZeroPoint),
reinterpret_cast<const uint8_t*>(Filter), static_cast<uint8_t>(FilterZeroPoint),
Output, Channels, OutputCount, KernelSize

4
onnxruntime/core/mlas/lib/qgemm.cpp
View file

@ -165,10 +165,10 @@ MlasGemmBatch(
ptrdiff_t TargetThreadCount;
if (Complexity < double(MLAS_QGEMM_THREAD_COMPLEXITY * MlasPlatform.MaximumThreadCount)) {
if (Complexity < double(MLAS_QGEMM_THREAD_COMPLEXITY * GetMlasPlatform().MaximumThreadCount)) {
TargetThreadCount = ptrdiff_t(Complexity / double(MLAS_QGEMM_THREAD_COMPLEXITY)) + 1;
} else {
TargetThreadCount = MlasPlatform.MaximumThreadCount;
TargetThreadCount = GetMlasPlatform().MaximumThreadCount;
}
ptrdiff_t MaximumThreadCount = MlasGetMaximumThreadCount(ThreadPool);

8
onnxruntime/core/mlas/lib/qgemm.h
View file

@ -710,21 +710,21 @@ MlasGemmQuantGetDispatch(
#if defined(MLAS_TARGET_AMD64_IX86)
if (!AIsSigned) {
if (BIsSigned) {
GemmQuantDispatch = MlasPlatform.GemmU8S8Dispatch;
GemmQuantDispatch = GetMlasPlatform().GemmU8S8Dispatch;
}
else {
GemmQuantDispatch = MlasPlatform.GemmU8U8Dispatch;
GemmQuantDispatch = GetMlasPlatform().GemmU8U8Dispatch;
}
}
#elif defined(MLAS_TARGET_ARM64)
if(BIsSigned) {
if(MlasPlatform.GemmU8X8Dispatch == &MlasGemmU8X8DispatchNeon) {
if(GetMlasPlatform().GemmU8X8Dispatch == &MlasGemmU8X8DispatchNeon) {
GemmQuantDispatch = &MlasGemmX8S8DispatchNeon;
} else {
GemmQuantDispatch = AIsSigned? &MlasGemmS8S8DispatchSdot : &MlasGemmU8X8DispatchUdot;
}
} else if(!AIsSigned) {
GemmQuantDispatch = MlasPlatform.GemmU8X8Dispatch;
GemmQuantDispatch = GetMlasPlatform().GemmU8X8Dispatch;
}
#elif defined(MLAS_TARGET_ARM64EC) || (defined(MLAS_TARGET_ARM) && !defined(_MSC_VER))
if(BIsSigned || !AIsSigned) {

10
onnxruntime/core/mlas/lib/qgemm_kernel_avx2.cpp
View file

@ -107,7 +107,7 @@ MlasGemmQuantTryGemvKernel<MLAS_GEMM_U8S8_KERNEL_AVX2>(
)
{
if (!AIsSigned && BIsSigned) {
MlasPlatform.GemvU8S8Kernel(A, B, C, CountK, CountN, ldb);
GetMlasPlatform().GemvU8S8Kernel(A, B, C, CountK, CountN, ldb);
return true;
}
@ -179,8 +179,8 @@ MlasGemmQuantKernel<MLAS_GEMM_U8S8_KERNEL_AVX2>(
bool ZeroMode
)
{
return MlasPlatform.GemmU8S8Kernel(A, B, C, PackedCountK, CountM, CountN, ldc,
RowSumBuffer, ColumnSumBuffer, ZeroPointB, ZeroMode);
return GetMlasPlatform().GemmU8S8Kernel(A, B, C, PackedCountK, CountM, CountN, ldc,
RowSumBuffer, ColumnSumBuffer, ZeroPointB, ZeroMode);
}
const MLAS_GEMM_QUANT_DISPATCH MlasGemmU8S8DispatchAvx2 = {
@ -260,8 +260,8 @@ MlasGemmQuantKernel<MLAS_GEMM_U8U8_KERNEL_AVX2>(
bool ZeroMode
)
{
return MlasPlatform.GemmU8U8Kernel(A, B, C, PackedCountK, CountM, CountN, ldc,
RowSumBuffer, ColumnSumBuffer, ZeroPointB, ZeroMode);
return GetMlasPlatform().GemmU8U8Kernel(A, B, C, PackedCountK, CountM, CountN, ldc,
RowSumBuffer, ColumnSumBuffer, ZeroPointB, ZeroMode);
}
const MLAS_GEMM_QUANT_DISPATCH MlasGemmU8U8DispatchAvx2 = {

4
onnxruntime/core/mlas/lib/qladd.cpp
View file

@ -520,7 +520,7 @@ MlasQLinearAdd<int8_t>(
)
{
#if defined(MLAS_TARGET_AMD64)
MlasPlatform.QLinearAddS8Kernel(
GetMlasPlatform().QLinearAddS8Kernel(
#else
MlasQLinearAddKernel<int8_t>(
#endif
@ -545,7 +545,7 @@ MlasQLinearAdd<uint8_t>(
)
{
#if defined(MLAS_TARGET_AMD64)
MlasPlatform.QLinearAddU8Kernel(
GetMlasPlatform().QLinearAddU8Kernel(
#else
MlasQLinearAddKernel<uint8_t>(
#endif

6
onnxruntime/core/mlas/lib/quantize.cpp
View file

@ -248,7 +248,7 @@ MlasQuantizeLinear<int8_t>(
)
{
#if defined(MLAS_TARGET_AMD64)
MlasPlatform.QuantizeLinearS8Kernel(
GetMlasPlatform().QuantizeLinearS8Kernel(
#else
MlasQuantizeLinearS8Kernel(
#endif
@ -267,7 +267,7 @@ MlasQuantizeLinear<uint8_t>(
)
{
#if defined(MLAS_TARGET_AMD64)
MlasPlatform.QuantizeLinearU8Kernel(
GetMlasPlatform().QuantizeLinearU8Kernel(
#else
MlasQuantizeLinearU8Kernel(
#endif
@ -989,7 +989,7 @@ Return Value:
--*/
{
#if defined(MLAS_TARGET_AMD64)
MlasPlatform.ReduceMinimumMaximumF32Kernel(Input, Min, Max, N);
GetMlasPlatform().ReduceMinimumMaximumF32Kernel(Input, Min, Max, N);
#else
MlasReduceMinimumMaximumF32Kernel(Input, Min, Max, N);
#endif

16
onnxruntime/core/mlas/lib/sgemm.cpp
View file

@ -475,7 +475,7 @@ Return Value:
#if defined(MLAS_TARGET_AMD64)
MLAS_SGEMM_TRANSPOSE_PACKB_BLOCK_ROUTINE* SgemmTransposePackB16x4Routine =
MlasPlatform.TransposePackB16x4Routine;
GetMlasPlatform().TransposePackB16x4Routine;
while (x >= 4) {
@ -1062,7 +1062,7 @@ Return Value:
size_t RowsHandled;
#if defined(MLAS_TARGET_AMD64_IX86) || defined(MLAS_TARGET_POWER)
RowsHandled = MlasPlatform.GemmFloatKernel(A, B, C, CountK, CountM, CountN, lda, ldc, alpha, ZeroMode);
RowsHandled = GetMlasPlatform().GemmFloatKernel(A, B, C, CountK, CountM, CountN, lda, ldc, alpha, ZeroMode);
#else
if (ZeroMode) {
RowsHandled = MlasSgemmKernelZero(A, B, C, CountK, CountM, CountN, lda, ldc, alpha);
@ -1163,9 +1163,9 @@ Return Value:
MLAS_SGEMM_KERNEL_M1_ROUTINE* SgemmKernelM1Routine;
if (TransB == CblasNoTrans) {
SgemmKernelM1Routine = MlasPlatform.KernelM1Routine;
SgemmKernelM1Routine = GetMlasPlatform().KernelM1Routine;
} else {
SgemmKernelM1Routine = MlasPlatform.KernelM1TransposeBRoutine;
SgemmKernelM1Routine = GetMlasPlatform().KernelM1TransposeBRoutine;
}
if (SgemmKernelM1Routine != nullptr) {
@ -1198,9 +1198,9 @@ Return Value:
MLAS_SGEMM_KERNEL_M1_ROUTINE* SgemmKernelM1Routine;
if (TransA == CblasNoTrans) {
SgemmKernelM1Routine = MlasPlatform.KernelM1TransposeBRoutine;
SgemmKernelM1Routine = GetMlasPlatform().KernelM1TransposeBRoutine;
} else {
SgemmKernelM1Routine = MlasPlatform.KernelM1Routine;
SgemmKernelM1Routine = GetMlasPlatform().KernelM1Routine;
}
if (SgemmKernelM1Routine != nullptr) {
@ -1580,10 +1580,10 @@ MlasGemmBatch(
ptrdiff_t TargetThreadCount;
if (Complexity < double(MLAS_SGEMM_THREAD_COMPLEXITY * MlasPlatform.MaximumThreadCount)) {
if (Complexity < double(MLAS_SGEMM_THREAD_COMPLEXITY * GetMlasPlatform().MaximumThreadCount)) {
TargetThreadCount = ptrdiff_t(Complexity / double(MLAS_SGEMM_THREAD_COMPLEXITY)) + 1;
} else {
TargetThreadCount = MlasPlatform.MaximumThreadCount;
TargetThreadCount = GetMlasPlatform().MaximumThreadCount;
}
ptrdiff_t MaximumThreadCount = MlasGetMaximumThreadCount(ThreadPool);

12
onnxruntime/core/mlas/lib/snchwc.cpp
View file

@ -102,7 +102,7 @@ Return Value:
--*/
{
#if defined(MLAS_TARGET_AMD64)
return MlasPlatform.NchwcBlockSize;
return GetMlasPlatform().NchwcBlockSize;
#else
return 1;
#endif
@ -675,7 +675,7 @@ struct MLAS_NCHWC_CONV_NCHWC_ALGORITHM : MLAS_NCHWC_GROUPED_CONV_ALGORITHM
const size_t BlockedOutputWidth = BlockSize * OutputWidth;
#if defined(MLAS_TARGET_AMD64)
MLAS_CONV_FLOAT_KERNEL* Kernel = MlasPlatform.ConvNchwcFloatKernel;
MLAS_CONV_FLOAT_KERNEL* Kernel = GetMlasPlatform().ConvNchwcFloatKernel;
#else
MLAS_CONV_FLOAT_KERNEL* Kernel = MlasConvNchwcFloatKernel;
#endif
@ -785,7 +785,7 @@ struct MLAS_NCHWC_CONV_NCHW_ALGORITHM : MLAS_NCHWC_GROUPED_CONV_ALGORITHM
const size_t BlockedOutputWidth = BlockSize * OutputWidth;
#if defined(MLAS_TARGET_AMD64)
MLAS_CONV_FLOAT_KERNEL* Kernel = MlasPlatform.ConvNchwFloatKernel;
MLAS_CONV_FLOAT_KERNEL* Kernel = GetMlasPlatform().ConvNchwFloatKernel;
#else
MLAS_CONV_FLOAT_KERNEL* Kernel = MlasConvNchwFloatKernel;
#endif
@ -880,7 +880,7 @@ struct MLAS_NCHWC_CONV_POINTWISE_ALGORITHM : MLAS_NCHWC_GROUPED_CONV_ALGORITHM
const size_t OutputStrideBytes = BlockSize * OutputSize * sizeof(float);
#if defined(MLAS_TARGET_AMD64)
MLAS_CONV_POINTWISE_FLOAT_KERNEL* Kernel = MlasPlatform.ConvPointwiseFloatKernel;
MLAS_CONV_POINTWISE_FLOAT_KERNEL* Kernel = GetMlasPlatform().ConvPointwiseFloatKernel;
#else
MLAS_CONV_POINTWISE_FLOAT_KERNEL* Kernel = MlasConvPointwiseFloatKernel;
#endif
@ -1017,7 +1017,7 @@ struct MLAS_NCHWC_CONV_DEPTHWISE_ALGORITHM : MLAS_NCHWC_CONV_ALGORITHM
const size_t BlockedOutputWidth = BlockSize * OutputWidth;
#if defined(MLAS_TARGET_AMD64)
MLAS_CONV_DEPTHWISE_FLOAT_KERNEL* Kernel = MlasPlatform.ConvDepthwiseFloatKernel;
MLAS_CONV_DEPTHWISE_FLOAT_KERNEL* Kernel = GetMlasPlatform().ConvDepthwiseFloatKernel;
#else
MLAS_CONV_DEPTHWISE_FLOAT_KERNEL* Kernel = MlasConvDepthwiseFloatKernel;
#endif
@ -1132,7 +1132,7 @@ struct MLAS_NCHWC_POOL_ALGORITHM : MLAS_NCHWC_NN_ALGORITHM
const size_t InputStrideBytes = DilatedInputWidthBytes - KernelWidth * DilationWidthBytes;
#if defined(MLAS_TARGET_AMD64)
MLAS_POOL_FLOAT_KERNEL* Kernel = MlasPlatform.PoolFloatKernel[WorkBlock->PoolingKind];
MLAS_POOL_FLOAT_KERNEL* Kernel = GetMlasPlatform().PoolFloatKernel[WorkBlock->PoolingKind];
#else
MLAS_POOL_FLOAT_KERNEL* Kernel = PoolKernels[WorkBlock->PoolingKind];
#endif

2
onnxruntime/core/mlas/lib/tanh.cpp
View file

@ -177,7 +177,7 @@ Return Value:
--*/
{
#if defined(MLAS_TARGET_AMD64)
MlasPlatform.TanhKernelRoutine(Input, Output, N);
GetMlasPlatform().TanhKernelRoutine(Input, Output, N);
#else
MlasTanhKernel(Input, Output, N);
#endif