diff --git a/paddle/fluid/framework/executor.cc b/paddle/fluid/framework/executor.cc index 3d68c5fb870d5b575f97eeb286528544402b8ed9..d4d6c34108b9f1e457d8eb0c36d10339b03330bd 100644 --- a/paddle/fluid/framework/executor.cc +++ b/paddle/fluid/framework/executor.cc @@ -24,6 +24,7 @@ limitations under the License. */ #include "paddle/fluid/platform/profiler.h" DECLARE_bool(benchmark); +DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run"); namespace paddle { namespace framework { @@ -115,6 +116,7 @@ void Executor::CreateVariables(const ProgramDesc& pdesc, Scope* scope, void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id, bool create_local_scope, bool create_vars) { platform::RecordBlock b(block_id); + if (FLAGS_use_mkldnn) EnableMKLDNN(pdesc); auto ctx = Prepare(pdesc, block_id); RunPreparedContext(ctx.get(), scope, create_local_scope, create_vars); } @@ -214,6 +216,7 @@ void Executor::Run(const ProgramDesc& program, Scope* scope, const std::string& feed_holder_name, const std::string& fetch_holder_name) { platform::RecordBlock b(kProgramId); + if (FLAGS_use_mkldnn) EnableMKLDNN(program); bool has_feed_ops = has_feed_operators(program.Block(0), *feed_targets, feed_holder_name); bool has_fetch_ops = @@ -225,7 +228,6 @@ void Executor::Run(const ProgramDesc& program, Scope* scope, unique_ptr_of_copy_program.reset(new ProgramDesc(program)); copy_program = unique_ptr_of_copy_program.get(); } - auto* global_block = copy_program->MutableBlock(0); if (!has_feed_ops) { @@ -378,5 +380,19 @@ void Executor::RunPreparedContext( } } +void Executor::EnableMKLDNN(const ProgramDesc& program) { +#ifdef PADDLE_WITH_MKLDNN + VLOG(3) << "use_mkldnn=True"; + for (size_t bid = 0; bid < program.Size(); ++bid) { + auto* block = const_cast(program).MutableBlock(bid); + for (auto* op : block->AllOps()) { + if (op->HasAttr("use_mkldnn")) { + op->SetAttr("use_mkldnn", true); + } + } + } +#endif +} + } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/executor.h b/paddle/fluid/framework/executor.h index 0c3c23611d95e0da67cabfb8fb2755a4a52c991b..e6f9c3d31c18f762ef2de269977e0642a79fb174 100644 --- a/paddle/fluid/framework/executor.h +++ b/paddle/fluid/framework/executor.h @@ -81,6 +81,8 @@ class Executor { const std::string& feed_holder_name = "feed", const std::string& fetch_holder_name = "fetch"); + void EnableMKLDNN(const ProgramDesc& program); + private: const platform::Place place_; }; diff --git a/paddle/fluid/inference/tests/book/test_inference_image_classification.cc b/paddle/fluid/inference/tests/book/test_inference_image_classification.cc index 987da18116cc6f4902bd66ae317f2470a8bc5057..60c761c5281e2f535aab0200c93fb738addcdb87 100644 --- a/paddle/fluid/inference/tests/book/test_inference_image_classification.cc +++ b/paddle/fluid/inference/tests/book/test_inference_image_classification.cc @@ -21,7 +21,6 @@ DEFINE_string(fp16_dirname, "", "Directory of the float16 inference model."); DEFINE_int32(batch_size, 1, "Batch size of input data"); DEFINE_int32(repeat, 1, "Running the inference program repeat times"); DEFINE_bool(skip_cpu, false, "Skip the cpu test"); -DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run inference"); TEST(inference, image_classification) { if (FLAGS_dirname.empty() || FLAGS_batch_size < 1 || FLAGS_repeat < 1) { @@ -59,10 +58,8 @@ TEST(inference, image_classification) { // Run inference on CPU LOG(INFO) << "--- CPU Runs: ---"; LOG(INFO) << "Batch size is " << FLAGS_batch_size; - LOG(INFO) << "FLAGS_use_mkldnn: " << FLAGS_use_mkldnn; TestInference( - dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat, is_combined, - FLAGS_use_mkldnn); + dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat, is_combined); LOG(INFO) << output1.dims(); } diff --git a/paddle/fluid/inference/tests/book/test_inference_nlp.cc b/paddle/fluid/inference/tests/book/test_inference_nlp.cc index a0e83a17058a4edcb8f23f23ce155e644ae0cf3b..9dcd79c3bb9ed713ff0f12024969cc5798750988 100644 --- a/paddle/fluid/inference/tests/book/test_inference_nlp.cc +++ b/paddle/fluid/inference/tests/book/test_inference_nlp.cc @@ -27,7 +27,6 @@ limitations under the License. */ DEFINE_string(model_path, "", "Directory of the inference model."); DEFINE_string(data_file, "", "File of input index data."); DEFINE_int32(repeat, 100, "Running the inference program repeat times"); -DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run inference"); DEFINE_bool(prepare_vars, true, "Prepare variables before executor"); DEFINE_int32(num_threads, 1, "Number of threads should be used"); @@ -190,9 +189,6 @@ TEST(inference, nlp) { std::unique_ptr inference_program; inference_program = InitProgram(&executor, scope.get(), FLAGS_model_path, /*model combined*/ false); - if (FLAGS_use_mkldnn) { - EnableMKLDNN(inference_program); - } // always prepare context std::unique_ptr ctx; ctx = executor.Prepare(*inference_program, 0); diff --git a/paddle/fluid/inference/tests/test_helper.h b/paddle/fluid/inference/tests/test_helper.h index 01b8dc0be662da22fe15a79cd9abfe5fa92c9577..44c36b1683b037832a218df02184e7cd2ba143e9 100644 --- a/paddle/fluid/inference/tests/test_helper.h +++ b/paddle/fluid/inference/tests/test_helper.h @@ -22,6 +22,8 @@ limitations under the License. */ #include "paddle/fluid/inference/io.h" #include "paddle/fluid/platform/profiler.h" +DECLARE_bool(use_mkldnn); + template void SetupTensor(paddle::framework::LoDTensor* input, paddle::framework::DDim dims, T lower, T upper) { @@ -133,24 +135,11 @@ std::vector> GetFeedTargetShapes( return feed_target_shapes; } -void EnableMKLDNN( - const std::unique_ptr& program) { - for (size_t bid = 0; bid < program->Size(); ++bid) { - auto* block = program->MutableBlock(bid); - for (auto* op : block->AllOps()) { - if (op->HasAttr("use_mkldnn")) { - op->SetAttr("use_mkldnn", true); - } - } - } -} - template void TestInference(const std::string& dirname, const std::vector& cpu_feeds, const std::vector& cpu_fetchs, - const int repeat = 1, const bool is_combined = false, - const bool use_mkldnn = false) { + const int repeat = 1, const bool is_combined = false) { // 1. Define place, executor, scope auto place = Place(); auto executor = paddle::framework::Executor(place); @@ -182,9 +171,6 @@ void TestInference(const std::string& dirname, "init_program", paddle::platform::DeviceContextPool::Instance().Get(place)); inference_program = InitProgram(&executor, scope, dirname, is_combined); - if (use_mkldnn) { - EnableMKLDNN(inference_program); - } } // Disable the profiler and print the timing information paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kDefault, @@ -210,7 +196,10 @@ void TestInference(const std::string& dirname, fetch_targets[fetch_target_names[i]] = cpu_fetchs[i]; } - // 6. Run the inference program + // 6. If export Flags_use_mkldnn=True, use mkldnn related ops. + if (FLAGS_use_mkldnn) executor.EnableMKLDNN(*inference_program); + + // 7. Run the inference program { if (!CreateVars) { // If users don't want to create and destroy variables every time they diff --git a/paddle/testing/paddle_gtest_main.cc b/paddle/testing/paddle_gtest_main.cc index 586ec48477f085a14d2f15b265a95d596705694f..507479c8622c8d33722e08bba018ad1ba5452e15 100644 --- a/paddle/testing/paddle_gtest_main.cc +++ b/paddle/testing/paddle_gtest_main.cc @@ -30,7 +30,7 @@ int main(int argc, char** argv) { new_argv.push_back( strdup("--tryfromenv=fraction_of_gpu_memory_to_use,use_pinned_memory")); #else - new_argv.push_back(strdup("--tryfromenv=use_pinned_memory")); + new_argv.push_back(strdup("--tryfromenv=use_pinned_memory,use_mkldnn")); #endif int new_argc = static_cast(new_argv.size()); char** new_argv_address = new_argv.data(); diff --git a/python/paddle/fluid/__init__.py b/python/paddle/fluid/__init__.py index c4fad620f0c49bb6b0ad3be22a564c16619efb0b..49147192799b228a77c0ac4b9944eb317881a353 100644 --- a/python/paddle/fluid/__init__.py +++ b/python/paddle/fluid/__init__.py @@ -116,7 +116,7 @@ def __bootstrap__(): read_env_flags = [ 'use_pinned_memory', 'check_nan_inf', 'benchmark', 'warpctc_dir', - 'eager_delete_scope' + 'eager_delete_scope', 'use_mkldnn' ] if core.is_compiled_with_cuda(): read_env_flags += [