saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-07-07 17:15:29 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Enable conv/conv_transpose for opset 11 in cuda execution provider. (#2401)

Browse source 
Enable conv/conv_transpose and existing pooling for opset 11 in cuda execution provider.

They are of spec dilates/strides change related cuda pooling ops for op set 11.
This commit is contained in:
Zhang Lei 2019-11-26 11:35:28 -08:00 committed by GitHub
parent 1ca0e0866e
commit ca8ff8c91c
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
4 changed files with 84 additions and 32 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

72
onnxruntime/core/providers/cuda/cuda_execution_provider.cc
View file

@ -389,12 +389,12 @@ class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain,
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, LRN);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, LRN);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, LRN);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, Conv);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, Conv);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, Conv);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ConvTranspose);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ConvTranspose);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ConvTranspose);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, Conv);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, Conv);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, Conv);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ConvTranspose);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ConvTranspose);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ConvTranspose);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, float, AveragePool);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, double, AveragePool);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, MLFloat16, AveragePool);
@ -549,13 +549,13 @@ class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDom
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 9, 10, Scan);
// opset 10
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, float, AveragePool);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, double, AveragePool);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, MLFloat16, AveragePool);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, 10, float, AveragePool);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, 10, double, AveragePool);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, 10, MLFloat16, AveragePool);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, Dropout);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, float, MaxPool);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, double, MaxPool);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, MLFloat16, MaxPool);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, 10, float, MaxPool);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, 10, double, MaxPool);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, 10, MLFloat16, MaxPool);
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, 10, NonMaxSuppression);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, float, Resize);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, double, Resize);
@ -640,6 +640,18 @@ class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, S
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, Squeeze);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, TopK);
class ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, Unsqueeze);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, Conv);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, Conv);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, Conv);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ConvTranspose);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ConvTranspose);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ConvTranspose);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, AveragePool);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, AveragePool);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, AveragePool);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, MaxPool);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, MaxPool);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, MaxPool);
static void RegisterCudaKernels(KernelRegistry& kernel_registry) {
static const BuildKernelCreateInfoFn function_table[] = {
@ -824,12 +836,12 @@ static void RegisterCudaKernels(KernelRegistry& kernel_registry) {
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, LRN)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, LRN)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, LRN)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, Conv)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, Conv)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, Conv)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, float, ConvTranspose)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, double, ConvTranspose)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, MLFloat16, ConvTranspose)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, Conv)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, Conv)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, Conv)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, float, ConvTranspose)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, double, ConvTranspose)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 1, 10, MLFloat16, ConvTranspose)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, float, AveragePool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, double, AveragePool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 7, 9, MLFloat16, AveragePool)>,
@ -983,13 +995,13 @@ static void RegisterCudaKernels(KernelRegistry& kernel_registry) {
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 9, 10, Scan)>,
// opset 10
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, float, AveragePool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, double, AveragePool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, MLFloat16, AveragePool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, 10, float, AveragePool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, 10, double, AveragePool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, 10, MLFloat16, AveragePool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, Dropout)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, float, MaxPool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, double, MaxPool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, MLFloat16, MaxPool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, 10, float, MaxPool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, 10, double, MaxPool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, 10, MLFloat16, MaxPool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, 10, NonMaxSuppression)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, float, Resize)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 10, double, Resize)>,
@ -1075,6 +1087,18 @@ static void RegisterCudaKernels(KernelRegistry& kernel_registry) {
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, Squeeze)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, TopK)>,
BuildKernelCreateInfo<ONNX_OPERATOR_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, Unsqueeze)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, Conv)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, Conv)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, Conv)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, ConvTranspose)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, ConvTranspose)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, ConvTranspose)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, AveragePool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, AveragePool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, AveragePool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, float, MaxPool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, double, MaxPool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 11, MLFloat16, MaxPool)>,
};
for (auto& function_table_entry : function_table) {

12
onnxruntime/core/providers/cuda/nn/conv.cc
View file

@ -9,11 +9,21 @@
namespace onnxruntime {
namespace cuda {
// Op Set 11 for Conv only update document to clearify default dilations and strides value.
// which are already convered by op set 11 cpu versoin, so simply add declaration.
#define REGISTER_KERNEL_TYPED(T) \
ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX( \
Conv, \
kOnnxDomain, \
1, 10, \
T, \
kCudaExecutionProvider, \
KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
Conv<T>); \
ONNX_OPERATOR_TYPED_KERNEL_EX( \
Conv, \
kOnnxDomain, \
1, \
11, \
T, \
kCudaExecutionProvider, \
KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \

12
onnxruntime/core/providers/cuda/nn/conv_transpose.cc
View file

@ -6,11 +6,21 @@
namespace onnxruntime {
namespace cuda {
// Op Set 11 for ConvTranspose only update document to clearify default dilations and strides value.
// which are already covered by op set 11 cpu version, so simply add declaration.
#define REGISTER_KERNEL_TYPED(T) \
ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_EX( \
ConvTranspose, \
kOnnxDomain, \
1, 10, \
T, \
kCudaExecutionProvider, \
KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \
ConvTranspose<T>); \
ONNX_OPERATOR_TYPED_KERNEL_EX( \
ConvTranspose, \
kOnnxDomain, \
1, \
11, \
T, \
kCudaExecutionProvider, \
KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<T>()), \

20
onnxruntime/core/providers/cuda/nn/pool.cc
View file

@ -30,12 +30,17 @@ namespace cuda {
KernelDefBuilder().TypeConstraint("T", DataTypeImpl::GetTensorType<data_type>()).TypeConstraint("I", DataTypeImpl::GetTensorType<int64_t>()), \
Pool<data_type, pool_type>);
POOLING_KERNEL_VERSIONED(AveragePool, float, AveragePool, 7, 9)
POOLING_KERNEL_VERSIONED(AveragePool, double, AveragePool, 7, 9)
POOLING_KERNEL_VERSIONED(AveragePool, MLFloat16, AveragePool, 7, 9)
POOLING_KERNEL(AveragePool, float, AveragePool, 10)
POOLING_KERNEL(AveragePool, double, AveragePool, 10)
POOLING_KERNEL(AveragePool, MLFloat16, AveragePool, 10)
POOLING_KERNEL_VERSIONED(AveragePool, float, AveragePool, 10, 10)
POOLING_KERNEL_VERSIONED(AveragePool, double, AveragePool, 10, 10)
POOLING_KERNEL_VERSIONED(AveragePool, MLFloat16, AveragePool, 10, 10)
//AveragePool and MaxPool op set 11 only update spec document on default value for dilations and strides.
POOLING_KERNEL(AveragePool, float, AveragePool, 11)
POOLING_KERNEL(AveragePool, double, AveragePool, 11)
POOLING_KERNEL(AveragePool, MLFloat16, AveragePool, 11)
POOLING_KERNEL(GlobalAveragePool, float, AveragePool, 1)
POOLING_KERNEL(GlobalAveragePool, double, AveragePool, 1)
POOLING_KERNEL(GlobalAveragePool, MLFloat16, AveragePool, 1)
@ -45,9 +50,12 @@ POOLING_KERNEL_VERSIONED(MaxPool, MLFloat16, MaxPool<1>, 1, 7)
POOLING_KERNEL_VERSIONED(MaxPool, float, MaxPool<8>, 8, 9)
POOLING_KERNEL_VERSIONED(MaxPool, double, MaxPool<8>, 8, 9)
POOLING_KERNEL_VERSIONED(MaxPool, MLFloat16, MaxPool<8>, 8, 9)
POOLING_KERNEL(MaxPool, float, MaxPool<8>, 10)
POOLING_KERNEL(MaxPool, double, MaxPool<8>, 10)
POOLING_KERNEL(MaxPool, MLFloat16, MaxPool<8>, 10)
POOLING_KERNEL_VERSIONED(MaxPool, float, MaxPool<8>, 10, 10)
POOLING_KERNEL_VERSIONED(MaxPool, double, MaxPool<8>, 10, 10)
POOLING_KERNEL_VERSIONED(MaxPool, MLFloat16, MaxPool<8>, 10, 10)
POOLING_KERNEL(MaxPool, float, MaxPool<8>, 11)
POOLING_KERNEL(MaxPool, double, MaxPool<8>, 11)
POOLING_KERNEL(MaxPool, MLFloat16, MaxPool<8>, 11)
POOLING_KERNEL(GlobalMaxPool, float, MaxPool<1>, 1)
POOLING_KERNEL(GlobalMaxPool, double, MaxPool<1>, 1)
POOLING_KERNEL(GlobalMaxPool, MLFloat16, MaxPool<1>, 1)