Skip to content
This repository has been archived by the owner on Nov 17, 2023. It is now read-only.

[MXNET-1179] Enforce deterministic algorithms in convolution layers (v1.3.x) #13152

Merged
merged 7 commits into from
Nov 7, 2018
Merged

[MXNET-1179] Enforce deterministic algorithms in convolution layers (v1.3.x) #13152

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
3a2dcc2
add env variable to choose deterministic cudnn alg
apeforest Oct 26, 2018
2c56d7f
set default value to false
apeforest Oct 29, 2018
ec18ebc
fix build failure in Windows GPU
apeforest Oct 29, 2018
2a15893
revert the previous change
apeforest Oct 29, 2018
a1ff714
only check determinism in CUDNN 7.x release
apeforest Oct 29, 2018
89af0d2
Add cudnn version check
apeforest Oct 29, 2018
a28c285
fix lint error
apeforest Oct 29, 2018
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
4 changes: 4 additions & 0 deletions src/operator/nn/cudnn/cudnn_convolution-inl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -854,13 +854,17 @@ class CuDNNConvolutionOp {
size_t workspace_byte, CuDNNAlgo<AlgoType> *algo) {
// Determine the fastest acceptable algo that matches the algo_preference (-1 = any),
// regardless of mathType.
bool enforce_determinism = dmlc::GetEnv("MXNET_ENFORCE_DETERMINISM", false);
for (decltype(perf_results.size()) i = 0; i != perf_results.size(); ++i) {
const auto &result = perf_results[i];
bool algo_is_tensor_core = false;
#if CUDNN_MAJOR >= 7
algo_is_tensor_core = result.mathType == CUDNN_TENSOR_OP_MATH;
#endif
if (result.status == CUDNN_STATUS_SUCCESS &&
#if CUDNN_MAJOR >= 7
(!enforce_determinism || result.determinism == cudnnDeterminism_t::CUDNN_DETERMINISTIC) &&
#endif
(param_.cudnn_tune.value() != conv::kLimited || result.memory <= workspace_byte)) {
algo->Set(result.algo, algo_is_tensor_core);
return;
Expand Down
4 changes: 4 additions & 0 deletions src/operator/nn/cudnn/cudnn_deconvolution-inl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -800,13 +800,17 @@ class CuDNNDeconvolutionOp {
void AlgoFinalSelect(const std::vector<PerfType> &perf_results, std::string kernel_name,
size_t workspace_byte, CuDNNAlgo<AlgoType> *algo) {
// Determine the fastest acceptable algo regardless of mathType.
bool enforce_determinism = dmlc::GetEnv("MXNET_ENFORCE_DETERMINISM", false);
for (decltype(perf_results.size()) i = 0; i != perf_results.size(); ++i) {
const auto &result = perf_results[i];
bool algo_is_tensor_core = false;
#if CUDNN_MAJOR >= 7
algo_is_tensor_core = result.mathType == CUDNN_TENSOR_OP_MATH;
#endif
if (result.status == CUDNN_STATUS_SUCCESS &&
#if CUDNN_MAJOR >= 7
(!enforce_determinism || result.determinism == cudnnDeterminism_t::CUDNN_DETERMINISTIC) &&
#endif
(param_.cudnn_tune.value() != conv::kLimited || result.memory <= workspace_byte)) {
algo->Set(result.algo, algo_is_tensor_core);
return;
Expand Down
5 changes: 5 additions & 0 deletions src/operator/nn/cudnn/cudnn_pooling-inl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -48,7 +48,12 @@ class CuDNNPoolingOp {
param_ = p;
switch (param_.pool_type) {
case pool_enum::kMaxPooling:
#if CUDNN_MAJOR >= 7
mode_ = dmlc::GetEnv("MXNET_ENFORCE_DETERMINISM", false) ?
CUDNN_POOLING_MAX_DETERMINISTIC : CUDNN_POOLING_MAX;
#else
mode_ = CUDNN_POOLING_MAX;
#endif
break;
case pool_enum::kAvgPooling:
if (param_.count_include_pad.has_value() && !param_.count_include_pad.value()) {
Expand Down