OSSForks/pytorch
1
0
Fork 0
mirror of https://github.com/zebrajr/pytorch.git synced 2025-12-06 00:20:18 +01:00
Code Issues Actions 38 Packages Projects Releases Wiki Activity

[cuDNN V8 API] Enable cuDNN v8 API by default (#75466)

Browse Source 
Testing via CI
Pull Request resolved: https://github.com/pytorch/pytorch/pull/75466
Approved by: https://github.com/ngimel
This commit is contained in:
Eddie Yan 2022-05-17 21:54:17 +00:00 committed by PyTorch MergeBot
parent 068d35a648
commit 14ab3ff484
5 changed files with 5 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
CMakeLists.txt
View File

@ -212,7 +212,7 @@ cmake_dependent_option(
BUILD_NVFUSER_BENCHMARK "Build C++ binaries for nvfuser benchmarks" ON
"USE_CUDA;BUILD_TEST" OFF)
cmake_dependent_option(
USE_EXPERIMENTAL_CUDNN_V8_API "Use experimental cuDNN v8 API" OFF
USE_EXPERIMENTAL_CUDNN_V8_API "Use experimental cuDNN v8 API" ON
"USE_CUDNN" OFF)
option(USE_FBGEMM "Use FBGEMM (quantized 8-bit server operators)" ON)
option(USE_KINETO "Use Kineto profiling library" ON)

2
aten/src/ATen/native/cudnn/Conv_v8.cpp
View File

@ -269,7 +269,7 @@ auto build_opgraph_fused(const cudnnHandle_t handle, const Tensor & x, const Ten
return opGraph;
}
const auto get_generator_sources(const cudnnBackendDescriptorType_t& desc, const Tensor& x, const bool deterministic, const bool allow_tf32, const cudnnBackendHeurMode_t heur_mode) {
auto get_generator_sources(const cudnnBackendDescriptorType_t& desc, const Tensor& x, const bool deterministic, const bool allow_tf32, const cudnnBackendHeurMode_t heur_mode) {
// Method for engine config generator based on heuristics
auto heurgen_method = [/*&desc,*/ &x, deterministic, allow_tf32, heur_mode](cudnn_frontend::OperationGraph &opGraph) -> cudnn_frontend::EngineConfigList {
auto heuristics = cudnn_frontend::EngineHeuristicsBuilder()

1
aten/src/ATen/native/quantized/cudnn/Pooling.cpp
View File

@ -189,7 +189,6 @@ Tensor quantized_max_pool2d_cudnn(
stride[0], // vertical stride
stride[1])); // horizontal stride
auto dataType = getCudnnDataType(input);
float one{1};
float zero{0.0};
TensorDescriptor xDesc;

5
aten/src/ATen/native/quantized/cudnn/utils.h
View File

@ -19,7 +19,7 @@ This file contains some of the auxiliary functions used by both Conv.cpp & Linea
#include <c10/util/ArrayRef.h>
#include <cudnn_frontend.h>
struct TORCH_API PackedLinearWeightCudnn : public LinearPackedParamsBase {
struct PackedLinearWeightCudnn : public LinearPackedParamsBase {
PackedLinearWeightCudnn(
at::Tensor orig_weight,
c10::optional<at::Tensor> bias,
@ -77,7 +77,7 @@ struct TORCH_API PackedLinearWeightCudnn : public LinearPackedParamsBase {
};
template <int kSpatialDim = 2>
struct TORCH_API PackedConvWeightCudnn : public ConvPackedParamsBase<kSpatialDim> {
struct PackedConvWeightCudnn : public ConvPackedParamsBase<kSpatialDim> {
PackedConvWeightCudnn(
at::Tensor orig_weight,
c10::optional<at::Tensor> bias,
@ -317,7 +317,6 @@ cudnn_frontend::ExecutionPlan get_execplan_from_heuristics_else_fall_back(cudnn_
}
{
auto total_engines = opGraph.getEngineCount();
// std::cout << opGraph.describe() << " has " << total_engines << " engines." << std::endl;
auto engine = cudnn_frontend::EngineBuilder().setGlobalEngineIdx(0).setOperationGraph(opGraph).build();
// std::cout << engine.describe() << std::endl;

2
third_party/cudnn_frontend vendored

@ -1 +1 @@
Subproject commit fa611998a360cbabaa2dcc7c9859748144114fc0
Subproject commit 43709ab96c47e26eebcdac72f93f946d44ceffa8