From 54afcb7ec663b4f87ee806072eb52d0693c93755 Mon Sep 17 00:00:00 2001 From: tensor-tang Date: Wed, 9 Jan 2019 13:01:34 +0000 Subject: [PATCH] add compare zerocopy test with native result test=develop --- .../tests/api/analyzer_seq_pool1_tester.cc | 53 +++++++++++++------ 1 file changed, 36 insertions(+), 17 deletions(-) diff --git a/paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc b/paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc index cd0fcedb9ae..1cf326fc89d 100644 --- a/paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc +++ b/paddle/fluid/inference/tests/api/analyzer_seq_pool1_tester.cc @@ -168,15 +168,13 @@ TEST(Analyzer_seq_pool1, compare) { reinterpret_cast(&cfg), input_slots_all); } -// Check the fuse status -TEST(Analyzer_seq_pool1, fuse_statis) { +void analysis_fuse_statis(bool use_zerocopy) { AnalysisConfig cfg; SetConfig(&cfg); + cfg.SwitchUseFeedFetchOps(!use_zerocopy); int num_ops; auto predictor = CreatePaddlePredictor(cfg); - auto fuse_statis = GetFuseStatis( - static_cast(predictor.get()), &num_ops); - + auto fuse_statis = GetFuseStatis(predictor.get(), &num_ops); ASSERT_TRUE(fuse_statis.count("fc_fuse")); ASSERT_EQ(fuse_statis.at("fc_fuse"), 10); ASSERT_TRUE(fuse_statis.count("seqpool_concat_fuse")); @@ -185,6 +183,9 @@ TEST(Analyzer_seq_pool1, fuse_statis) { EXPECT_EQ(num_ops, 195); } +// Check the fuse status +TEST(Analyzer_seq_pool1, fuse_statis) { analysis_fuse_statis(false); } + void PrepareZeroCopyInputs( const std::unique_ptr &predictor, std::vector> *inputs) { @@ -202,7 +203,8 @@ void PrepareZeroCopyInputs( } } -std::unique_ptr zerocopy_profile(int repeat_times) { +// return the output values +std::vector zerocopy_profile(int repeat_times) { AnalysisConfig config; SetConfig(&config); config.SwitchUseFeedFetchOps(false); @@ -225,23 +227,40 @@ std::unique_ptr zerocopy_profile(int repeat_times) { } PrintTime(FLAGS_batch_size, repeat_times, 1, 0, timer.toc() / repeat_times, 1); - return output_tensor; + + VLOG(3) << "ZeroCopy output: " << DescribeZeroCopyTensor(*output_tensor); + PaddlePlace place; + int output_size{0}; + auto *pdata = output_tensor->data(&place, &output_size); + std::vector res(output_size); + for (int i = 0; i < output_size; ++i) { + res[i] = pdata[i]; + } + return res; } TEST(Analyzer_seq_pool1, zerocopy_profile) { zerocopy_profile(FLAGS_repeat); } -TEST(Analyzer_seq_pool1, zerocopy_fuse_statis) { +TEST(Analyzer_seq_pool1, zerocopy_fuse_statis) { analysis_fuse_statis(true); } + +TEST(Analyzer_seq_pool1, zerocopy_compare_native) { AnalysisConfig config; SetConfig(&config); - config.SwitchUseFeedFetchOps(false); - auto predictor = CreatePaddlePredictor(config); - int num_ops; - auto fuse_statis = GetFuseStatis(predictor.get(), &num_ops); - ASSERT_TRUE(fuse_statis.count("fc_fuse")); - ASSERT_EQ(fuse_statis.at("fc_fuse"), 10); - ASSERT_TRUE(fuse_statis.count("seqpool_concat_fuse")); - EXPECT_EQ(fuse_statis.at("seqpool_concat_fuse"), 2); - ASSERT_EQ(num_ops, 195); + config.SwitchUseFeedFetchOps(true); + auto predictor = CreatePaddlePredictor(config.ToNativeConfig()); + std::vector native_outputs; + std::vector> input_slots_all; + SetInput(&input_slots_all); + ASSERT_TRUE(predictor->Run(input_slots_all[0], &native_outputs)); + EXPECT_EQ(native_outputs.size(), 1UL); + + auto zerocopy_output = zerocopy_profile(1); + EXPECT_EQ(zerocopy_output.size() * sizeof(float), + native_outputs.front().data.length()); + auto *native_data = static_cast(native_outputs.front().data.data()); + for (size_t i = 0; i < zerocopy_output.size(); ++i) { + EXPECT_NEAR(zerocopy_output[i], native_data[i], 1e-3); + } } } // namespace analysis -- GitLab