提交 661826a7 编写于 作者: T tensor-tang

enable MKLDNN inference test

上级 48ac9780
...@@ -21,6 +21,7 @@ DEFINE_string(fp16_dirname, "", "Directory of the float16 inference model."); ...@@ -21,6 +21,7 @@ DEFINE_string(fp16_dirname, "", "Directory of the float16 inference model.");
DEFINE_int32(batch_size, 1, "Batch size of input data"); DEFINE_int32(batch_size, 1, "Batch size of input data");
DEFINE_int32(repeat, 1, "Running the inference program repeat times"); DEFINE_int32(repeat, 1, "Running the inference program repeat times");
DEFINE_bool(skip_cpu, false, "Skip the cpu test"); DEFINE_bool(skip_cpu, false, "Skip the cpu test");
DEFINE_bool(use_mkldnn, false, "Use MKLDNN to run inference");
TEST(inference, image_classification) { TEST(inference, image_classification) {
if (FLAGS_dirname.empty() || FLAGS_batch_size < 1 || FLAGS_repeat < 1) { if (FLAGS_dirname.empty() || FLAGS_batch_size < 1 || FLAGS_repeat < 1) {
...@@ -58,8 +59,10 @@ TEST(inference, image_classification) { ...@@ -58,8 +59,10 @@ TEST(inference, image_classification) {
// Run inference on CPU // Run inference on CPU
LOG(INFO) << "--- CPU Runs: ---"; LOG(INFO) << "--- CPU Runs: ---";
LOG(INFO) << "Batch size is " << FLAGS_batch_size; LOG(INFO) << "Batch size is " << FLAGS_batch_size;
LOG(INFO) << "FLAGS_use_mkldnn: " << FLAGS_use_mkldnn;
TestInference<paddle::platform::CPUPlace, false, true>( TestInference<paddle::platform::CPUPlace, false, true>(
dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat, is_combined); dirname, cpu_feeds, cpu_fetchs1, FLAGS_repeat, is_combined,
FLAGS_use_mkldnn);
LOG(INFO) << output1.dims(); LOG(INFO) << output1.dims();
} }
......
...@@ -133,11 +133,24 @@ std::vector<std::vector<int64_t>> GetFeedTargetShapes( ...@@ -133,11 +133,24 @@ std::vector<std::vector<int64_t>> GetFeedTargetShapes(
return feed_target_shapes; return feed_target_shapes;
} }
void EnableMKLDNN(
const std::unique_ptr<paddle::framework::ProgramDesc>& 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 <typename Place, bool CreateVars = true, bool PrepareContext = false> template <typename Place, bool CreateVars = true, bool PrepareContext = false>
void TestInference(const std::string& dirname, void TestInference(const std::string& dirname,
const std::vector<paddle::framework::LoDTensor*>& cpu_feeds, const std::vector<paddle::framework::LoDTensor*>& cpu_feeds,
const std::vector<paddle::framework::LoDTensor*>& cpu_fetchs, const std::vector<paddle::framework::LoDTensor*>& cpu_fetchs,
const int repeat = 1, const bool is_combined = false) { const int repeat = 1, const bool is_combined = false,
const bool use_mkldnn = false) {
// 1. Define place, executor, scope // 1. Define place, executor, scope
auto place = Place(); auto place = Place();
auto executor = paddle::framework::Executor(place); auto executor = paddle::framework::Executor(place);
...@@ -169,6 +182,9 @@ void TestInference(const std::string& dirname, ...@@ -169,6 +182,9 @@ void TestInference(const std::string& dirname,
"init_program", "init_program",
paddle::platform::DeviceContextPool::Instance().Get(place)); paddle::platform::DeviceContextPool::Instance().Get(place));
inference_program = InitProgram(&executor, scope, dirname, is_combined); inference_program = InitProgram(&executor, scope, dirname, is_combined);
if (use_mkldnn) {
EnableMKLDNN(inference_program);
}
} }
// Disable the profiler and print the timing information // Disable the profiler and print the timing information
paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kDefault, paddle::platform::DisableProfiler(paddle::platform::EventSortingKey::kDefault,
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册