Skip to content

[CUDA] MaxPool-22 #27715

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

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

[CUDA] MaxPool-22 #27715

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
30 changes: 20 additions & 10 deletions onnxruntime/core/providers/cuda/cuda_execution_provider.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1037,11 +1037,11 @@ class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDom

// OpSet 12
class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, 12, Clip);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, float, MaxPool);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, double, MaxPool);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, MLFloat16, MaxPool);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, int8_t, MaxPool);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, uint8_t, MaxPool);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, 21, float, MaxPool);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, 21, double, MaxPool);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, 21, MLFloat16, MaxPool);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, 21, int8_t, MaxPool);
class ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, 21, uint8_t, MaxPool);

class ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, 12, Pow);

Expand Down Expand Up @@ -1579,6 +1579,11 @@ class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain,
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 22, double, AveragePool);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 22, MLFloat16, AveragePool);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 22, BFloat16, AveragePool);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 22, float, MaxPool);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 22, double, MaxPool);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 22, MLFloat16, MaxPool);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 22, int8_t, MaxPool);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 22, uint8_t, MaxPool);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 22, float, Conv);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 22, double, Conv);
class ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 22, MLFloat16, Conv);
Expand Down Expand Up @@ -2123,11 +2128,11 @@ static Status RegisterCudaKernels(KernelRegistry& kernel_registry) {
// OpSet 12
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, 12, Clip)>,

BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, float, MaxPool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, double, MaxPool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, MLFloat16, MaxPool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, int8_t, MaxPool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, uint8_t, MaxPool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, 21, float, MaxPool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, 21, double, MaxPool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, 21, MLFloat16, MaxPool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, 21, int8_t, MaxPool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, 21, uint8_t, MaxPool)>,

BuildKernelCreateInfo<ONNX_OPERATOR_VERSIONED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 12, 12, Pow)>,

Expand Down Expand Up @@ -2663,6 +2668,11 @@ static Status RegisterCudaKernels(KernelRegistry& kernel_registry) {
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 22, double, AveragePool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 22, MLFloat16, AveragePool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 22, BFloat16, AveragePool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 22, float, MaxPool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 22, double, MaxPool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 22, MLFloat16, MaxPool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 22, int8_t, MaxPool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 22, uint8_t, MaxPool)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 22, float, Conv)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 22, double, Conv)>,
BuildKernelCreateInfo<ONNX_OPERATOR_TYPED_KERNEL_CLASS_NAME(kCudaExecutionProvider, kOnnxDomain, 22, MLFloat16, Conv)>,
Expand Down
24 changes: 16 additions & 8 deletions onnxruntime/core/providers/cuda/cuda_nhwc_kernels.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -65,10 +65,14 @@ class CUDA_NHWC_OP_TYPED_CLASS_NAME(22, float, AveragePool);
class CUDA_NHWC_OP_TYPED_CLASS_NAME(22, MLFloat16, AveragePool);
class CUDA_NHWC_OP_VERSIONED_TYPED_CLASS_NAME(11, 11, float, MaxPool);
class CUDA_NHWC_OP_VERSIONED_TYPED_CLASS_NAME(11, 11, MLFloat16, MaxPool);
class CUDA_NHWC_OP_TYPED_CLASS_NAME(12, float, MaxPool);
class CUDA_NHWC_OP_TYPED_CLASS_NAME(12, MLFloat16, MaxPool);
class CUDA_NHWC_OP_TYPED_CLASS_NAME(12, int8_t, MaxPool);
class CUDA_NHWC_OP_TYPED_CLASS_NAME(12, uint8_t, MaxPool);
class CUDA_NHWC_OP_VERSIONED_TYPED_CLASS_NAME(12, 21, float, MaxPool);
class CUDA_NHWC_OP_VERSIONED_TYPED_CLASS_NAME(12, 21, MLFloat16, MaxPool);
class CUDA_NHWC_OP_VERSIONED_TYPED_CLASS_NAME(12, 21, int8_t, MaxPool);
class CUDA_NHWC_OP_VERSIONED_TYPED_CLASS_NAME(12, 21, uint8_t, MaxPool);
class CUDA_NHWC_OP_TYPED_CLASS_NAME(22, float, MaxPool);
class CUDA_NHWC_OP_TYPED_CLASS_NAME(22, MLFloat16, MaxPool);
class CUDA_NHWC_OP_TYPED_CLASS_NAME(22, int8_t, MaxPool);
class CUDA_NHWC_OP_TYPED_CLASS_NAME(22, uint8_t, MaxPool);
class CUDA_NHWC_OP_VERSIONED_TYPED_CLASS_NAME(14, 14, float, BatchNormalization);
class CUDA_NHWC_OP_VERSIONED_TYPED_CLASS_NAME(14, 14, double, BatchNormalization);
class CUDA_NHWC_OP_VERSIONED_TYPED_CLASS_NAME(14, 14, MLFloat16, BatchNormalization);
Expand Down Expand Up @@ -130,10 +134,14 @@ Status RegisterCudaNhwcKernels(KernelRegistry& kernel_registry) {
BuildKernelCreateInfo<CUDA_NHWC_OP_VERSIONED_TYPED_CLASS_NAME(10, 10, MLFloat16, MaxPool)>,
BuildKernelCreateInfo<CUDA_NHWC_OP_VERSIONED_TYPED_CLASS_NAME(11, 11, float, MaxPool)>,
BuildKernelCreateInfo<CUDA_NHWC_OP_VERSIONED_TYPED_CLASS_NAME(11, 11, MLFloat16, MaxPool)>,
BuildKernelCreateInfo<CUDA_NHWC_OP_TYPED_CLASS_NAME(12, float, MaxPool)>,
BuildKernelCreateInfo<CUDA_NHWC_OP_TYPED_CLASS_NAME(12, MLFloat16, MaxPool)>,
BuildKernelCreateInfo<CUDA_NHWC_OP_TYPED_CLASS_NAME(12, int8_t, MaxPool)>,
BuildKernelCreateInfo<CUDA_NHWC_OP_TYPED_CLASS_NAME(12, uint8_t, MaxPool)>,
BuildKernelCreateInfo<CUDA_NHWC_OP_VERSIONED_TYPED_CLASS_NAME(12, 21, float, MaxPool)>,
BuildKernelCreateInfo<CUDA_NHWC_OP_VERSIONED_TYPED_CLASS_NAME(12, 21, MLFloat16, MaxPool)>,
BuildKernelCreateInfo<CUDA_NHWC_OP_VERSIONED_TYPED_CLASS_NAME(12, 21, int8_t, MaxPool)>,
BuildKernelCreateInfo<CUDA_NHWC_OP_VERSIONED_TYPED_CLASS_NAME(12, 21, uint8_t, MaxPool)>,
BuildKernelCreateInfo<CUDA_NHWC_OP_TYPED_CLASS_NAME(22, float, MaxPool)>,
BuildKernelCreateInfo<CUDA_NHWC_OP_TYPED_CLASS_NAME(22, MLFloat16, MaxPool)>,
BuildKernelCreateInfo<CUDA_NHWC_OP_TYPED_CLASS_NAME(22, int8_t, MaxPool)>,
BuildKernelCreateInfo<CUDA_NHWC_OP_TYPED_CLASS_NAME(22, uint8_t, MaxPool)>,
BuildKernelCreateInfo<CUDA_NHWC_OP_TYPED_CLASS_NAME(11, float, ConvTranspose)>,
BuildKernelCreateInfo<CUDA_NHWC_OP_TYPED_CLASS_NAME(11, MLFloat16, ConvTranspose)>,
BuildKernelCreateInfo<CUDA_NHWC_OP_VERSIONED_TYPED_CLASS_NAME(1, 10, float, ConvTranspose)>,
Expand Down
28 changes: 19 additions & 9 deletions onnxruntime/core/providers/cuda/nn/pool.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -72,11 +72,17 @@ POOLING_KERNEL_VERSIONED_WITH_INDICES(MaxPool, MLFloat16, MaxPool<8>, 10, 10, kO
POOLING_KERNEL_VERSIONED_WITH_INDICES(MaxPool, float, MaxPool<8>, 11, 11, kOnnxDomain, false)
POOLING_KERNEL_VERSIONED_WITH_INDICES(MaxPool, double, MaxPool<8>, 11, 11, kOnnxDomain, false)
POOLING_KERNEL_VERSIONED_WITH_INDICES(MaxPool, MLFloat16, MaxPool<8>, 11, 11, kOnnxDomain, false)
POOLING_KERNEL_WITH_INDICES(MaxPool, float, MaxPool<8>, 12, kOnnxDomain, false)
POOLING_KERNEL_WITH_INDICES(MaxPool, double, MaxPool<8>, 12, kOnnxDomain, false)
POOLING_KERNEL_WITH_INDICES(MaxPool, MLFloat16, MaxPool<8>, 12, kOnnxDomain, false)
POOLING_KERNEL_WITH_INDICES(MaxPool, int8_t, MaxPool<8>, 12, kOnnxDomain, false)
POOLING_KERNEL_WITH_INDICES(MaxPool, uint8_t, MaxPool<8>, 12, kOnnxDomain, false)
// MaxPool opsets 12-22 share the same CUDA implementation for the currently supported types.
POOLING_KERNEL_VERSIONED_WITH_INDICES(MaxPool, float, MaxPool<8>, 12, 21, kOnnxDomain, false)
POOLING_KERNEL_VERSIONED_WITH_INDICES(MaxPool, double, MaxPool<8>, 12, 21, kOnnxDomain, false)
POOLING_KERNEL_VERSIONED_WITH_INDICES(MaxPool, MLFloat16, MaxPool<8>, 12, 21, kOnnxDomain, false)
POOLING_KERNEL_VERSIONED_WITH_INDICES(MaxPool, int8_t, MaxPool<8>, 12, 21, kOnnxDomain, false)
POOLING_KERNEL_VERSIONED_WITH_INDICES(MaxPool, uint8_t, MaxPool<8>, 12, 21, kOnnxDomain, false)
POOLING_KERNEL_WITH_INDICES(MaxPool, float, MaxPool<8>, 22, kOnnxDomain, false)
POOLING_KERNEL_WITH_INDICES(MaxPool, double, MaxPool<8>, 22, kOnnxDomain, false)
POOLING_KERNEL_WITH_INDICES(MaxPool, MLFloat16, MaxPool<8>, 22, kOnnxDomain, false)
POOLING_KERNEL_WITH_INDICES(MaxPool, int8_t, MaxPool<8>, 22, kOnnxDomain, false)
POOLING_KERNEL_WITH_INDICES(MaxPool, uint8_t, MaxPool<8>, 22, kOnnxDomain, false)

POOLING_KERNEL(GlobalMaxPool, float, MaxPool<1>, 1, kOnnxDomain, false)
POOLING_KERNEL(GlobalMaxPool, double, MaxPool<1>, 1, kOnnxDomain, false)
Expand All @@ -92,10 +98,14 @@ POOLING_KERNEL_VERSIONED_WITH_INDICES(MaxPool, float, MaxPool<8>, 10, 10, kMSInt
POOLING_KERNEL_VERSIONED_WITH_INDICES(MaxPool, MLFloat16, MaxPool<8>, 10, 10, kMSInternalNHWCDomain, true)
POOLING_KERNEL_VERSIONED_WITH_INDICES(MaxPool, float, MaxPool<8>, 11, 11, kMSInternalNHWCDomain, true)
POOLING_KERNEL_VERSIONED_WITH_INDICES(MaxPool, MLFloat16, MaxPool<8>, 11, 11, kMSInternalNHWCDomain, true)
POOLING_KERNEL_WITH_INDICES(MaxPool, float, MaxPool<8>, 12, kMSInternalNHWCDomain, true)
POOLING_KERNEL_WITH_INDICES(MaxPool, MLFloat16, MaxPool<8>, 12, kMSInternalNHWCDomain, true)
POOLING_KERNEL_WITH_INDICES(MaxPool, int8_t, MaxPool<8>, 12, kMSInternalNHWCDomain, true)
POOLING_KERNEL_WITH_INDICES(MaxPool, uint8_t, MaxPool<8>, 12, kMSInternalNHWCDomain, true)
POOLING_KERNEL_VERSIONED_WITH_INDICES(MaxPool, float, MaxPool<8>, 12, 21, kMSInternalNHWCDomain, true)
POOLING_KERNEL_VERSIONED_WITH_INDICES(MaxPool, MLFloat16, MaxPool<8>, 12, 21, kMSInternalNHWCDomain, true)
POOLING_KERNEL_VERSIONED_WITH_INDICES(MaxPool, int8_t, MaxPool<8>, 12, 21, kMSInternalNHWCDomain, true)
POOLING_KERNEL_VERSIONED_WITH_INDICES(MaxPool, uint8_t, MaxPool<8>, 12, 21, kMSInternalNHWCDomain, true)
POOLING_KERNEL_WITH_INDICES(MaxPool, float, MaxPool<8>, 22, kMSInternalNHWCDomain, true)
POOLING_KERNEL_WITH_INDICES(MaxPool, MLFloat16, MaxPool<8>, 22, kMSInternalNHWCDomain, true)
POOLING_KERNEL_WITH_INDICES(MaxPool, int8_t, MaxPool<8>, 22, kMSInternalNHWCDomain, true)
POOLING_KERNEL_WITH_INDICES(MaxPool, uint8_t, MaxPool<8>, 22, kMSInternalNHWCDomain, true)

POOLING_KERNEL(GlobalMaxPool, float, MaxPool<1>, 1, kMSInternalNHWCDomain, true)
POOLING_KERNEL(GlobalMaxPool, MLFloat16, MaxPool<1>, 1, kMSInternalNHWCDomain, true)
Expand Down
30 changes: 30 additions & 0 deletions onnxruntime/test/providers/cpu/nn/pool_op_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -195,6 +195,36 @@ TEST(PoolTest, MaxPool_8_With_Index) {
MaxPool_8_WithIndexTest(true, 1 /*storage_order*/); // col major
}

TEST(PoolTest, MaxPool_22_With_Index_CUDA) {
auto cuda_ep = DefaultCudaExecutionProvider();
if (!cuda_ep) {
return;
}

OpTester test("MaxPool", 22);
test.AddAttribute("strides", std::vector<int64_t>{2, 2});
test.AddAttribute("pads", vector<int64_t>{0, 0, 0, 0});
test.AddAttribute("kernel_shape", vector<int64_t>{2, 2});

std::vector<float> x_vals = {
1.0f, 2.0f, 3.0f, 4.0f,
5.0f, 6.0f, 7.0f, 8.0f,
9.0f, 10.0f, 11.0f, 12.0f,
13.0f, 14.0f, 15.0f, 16.0f};
std::vector<int64_t> x_dims = {1, 1, 4, 4};
std::vector<int64_t> expected_dims = {1, 1, 2, 2};
std::vector<float> expected_vals = {6.0f, 8.0f, 14.0f, 16.0f};
std::vector<int64_t> expected_indices = {5, 7, 13, 15};

test.AddInput<float>("X", x_dims, x_vals);
test.AddOutput<float>("Y", expected_dims, expected_vals);
test.AddOutput<int64_t>("Indices", expected_dims, expected_indices);

std::vector<std::unique_ptr<IExecutionProvider>> execution_providers;
execution_providers.push_back(std::move(cuda_ep));
test.Run(OpTester::ExpectResult::kExpectSuccess, "", {}, nullptr, &execution_providers);
}

TEST(PoolTest, MaxPool1D_case1) {
OpTester test("MaxPool");

Expand Down
4 changes: 3 additions & 1 deletion onnxruntime/test/providers/cuda/nhwc/pool_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,6 +10,7 @@ namespace test {
template <typename T>
struct PoolOp {
std::string pooling_type;
int opset = 14;
std::vector<int64_t> input_dims;
std::vector<int64_t> kernel_shape;
int64_t channels;
Expand All @@ -20,7 +21,7 @@ struct PoolOp {
std::unique_ptr<CompareOpTester> get_test() {
RandomValueGenerator random{};

auto test = std::make_unique<CompareOpTester>(pooling_type.c_str(), 14);
auto test = std::make_unique<CompareOpTester>(pooling_type.c_str(), opset);
std::vector<T> input_data = random.Uniform<T>(input_dims, 0.0f, 0.3f);

test->AddInput<T>("X", input_dims, input_data);
Expand Down Expand Up @@ -53,6 +54,7 @@ TYPED_TEST(CudaNhwcTypedTest, AveragePoolNhwc) {
TYPED_TEST(CudaNhwcTypedTest, MaxPoolNhwc) {
auto op = PoolOp<TypeParam>{};
op.pooling_type = "MaxPool";
op.opset = 22;
op.input_dims = {1, 16, 64, 64};
op.kernel_shape = {3, 3};
op.channels = 16;
Expand Down
Loading