diff options
Diffstat (limited to 'inference-engine/thirdparty/mkl-dnn/tests/gtests/test_pooling_forward.cpp')
| -rw-r--r-- | inference-engine/thirdparty/mkl-dnn/tests/gtests/test_pooling_forward.cpp | 253 |
1 files changed, 153 insertions, 100 deletions
diff --git a/inference-engine/thirdparty/mkl-dnn/tests/gtests/test_pooling_forward.cpp b/inference-engine/thirdparty/mkl-dnn/tests/gtests/test_pooling_forward.cpp index 855bb80b0..99867ffb5 100644 --- a/inference-engine/thirdparty/mkl-dnn/tests/gtests/test_pooling_forward.cpp +++ b/inference-engine/thirdparty/mkl-dnn/tests/gtests/test_pooling_forward.cpp @@ -65,96 +65,89 @@ void check_pool_fwd(const pool_test_params &p, const memory &src, auto pd = p.test_pd; size_t padded_c = src_d.data.layout_desc.blocking.padding_dims[1]; -#pragma omp parallel for collapse(4) schedule(static) - for (int n = 0; n < pd.mb; n++) { - for (int c = 0; c < pd.c; c++) { - for (int od = 0; od < pd.od; od++) { - for (int oh = 0; oh < pd.oh; oh++) { - for (int ow = 0; ow < pd.ow; ow++) { - size_t oidx = (size_t)n * padded_c * pd.od * pd.oh * pd.ow - + (size_t)c * pd.od * pd.oh * pd.ow - + (size_t)od * pd.oh * pd.ow - + (size_t)oh * pd.ow + ow; - data_t out = dst_data[map_index(dst_d, oidx)]; - int out_index = -1; - if(p.aalgorithm == pooling_max - && p.aprop_kind == prop_kind::forward_training) { - out_index = ws_data(map_index(ws_d, oidx)); - } - // match implementation for pooling_max: padding - // is done with lowest value and not zero, it - // affects the case when kernel slips into - // the padding area entirely - typename acc_t<data_t>::type acc_ref - = (p.aalgorithm == pooling_max) ? - std::numeric_limits<data_t>::lowest() : - data_t(0); - int out_ref_index = 0; - bool is_initialized = false; - int num_summands = 0; - - for (int kd = 0; kd < pd.kd; ++kd) - for (int kh = 0; kh < pd.kh; ++kh) - for (int kw = 0; kw < pd.kw; ++kw) - { - const int id = od * pd.strd - pd.padf + kd; - const int ih = oh * pd.strh - pd.padt + kh; - const int iw = ow * pd.strw - pd.padl + kw; - - if (id < 0 || id >= pd.id) continue; - if (ih < 0 || ih >= pd.ih) continue; - if (iw < 0 || iw >= pd.iw) continue; - - size_t iidx - = (size_t)n * padded_c * pd.id * pd.ih * pd.iw - + (size_t)c * pd.id * pd.ih * pd.iw - + (size_t)id * pd.ih * pd.iw - + (size_t)ih * pd.iw + iw; - - data_t d = src_data[map_index(src_d, iidx)]; - if (p.aalgorithm == pooling_max) { - if (!is_initialized) { - acc_ref = d; - out_ref_index = kd * pd.kw * pd.kh - + kh * pd.kw + kw; - is_initialized = true; - } else { - if (acc_ref < d) { - acc_ref = d; - out_ref_index = kd * pd.kw * pd.kh - + kh * pd.kw + kw; - } - } - } else if (p.aalgorithm == pooling_avg_include_padding || - p.aalgorithm == pooling_avg_exclude_padding) { - acc_ref += d; - num_summands++; - } + mkldnn::impl::parallel_nd(pd.mb, pd.c, pd.od, pd.oh, pd.ow, + [&](int n, int c, int od, int oh, int ow) { + size_t oidx = (size_t)n * padded_c * pd.od * pd.oh * pd.ow + + (size_t)c * pd.od * pd.oh * pd.ow + + (size_t)od * pd.oh * pd.ow + + (size_t)oh * pd.ow + ow; + data_t out = dst_data[map_index(dst_d, oidx)]; + int out_index = -1; + if(p.aalgorithm == pooling_max + && p.aprop_kind == prop_kind::forward_training) { + out_index = ws_data(map_index(ws_d, oidx)); + } + // match implementation for pooling_max: padding + // is done with lowest value and not zero, it + // affects the case when kernel slips into + // the padding area entirely + typename acc_t<data_t>::type acc_ref + = (p.aalgorithm == pooling_max) ? + std::numeric_limits<data_t>::lowest() : + data_t(0); + int out_ref_index = 0; + bool is_initialized = false; + int num_summands = 0; + + for (int kd = 0; kd < pd.kd; ++kd) + for (int kh = 0; kh < pd.kh; ++kh) + for (int kw = 0; kw < pd.kw; ++kw) + { + const int id = od * pd.strd - pd.padf + kd; + const int ih = oh * pd.strh - pd.padt + kh; + const int iw = ow * pd.strw - pd.padl + kw; + + if (id < 0 || id >= pd.id) continue; + if (ih < 0 || ih >= pd.ih) continue; + if (iw < 0 || iw >= pd.iw) continue; + + size_t iidx + = (size_t)n * padded_c * pd.id * pd.ih * pd.iw + + (size_t)c * pd.id * pd.ih * pd.iw + + (size_t)id * pd.ih * pd.iw + + (size_t)ih * pd.iw + iw; + + data_t d = src_data[map_index(src_d, iidx)]; + if (p.aalgorithm == pooling_max) { + if (!is_initialized) { + acc_ref = d; + out_ref_index = kd * pd.kw * pd.kh + + kh * pd.kw + kw; + is_initialized = true; + } else { + if (acc_ref < d) { + acc_ref = d; + out_ref_index = kd * pd.kw * pd.kh + + kh * pd.kw + kw; } + } + } else if (p.aalgorithm == pooling_avg_include_padding || + p.aalgorithm == pooling_avg_exclude_padding) { + acc_ref += d; + num_summands++; + } + } - if (p.aalgorithm == pooling_avg_include_padding) { - num_summands = pd.kw * pd.kh * pd.kd; - } + if (p.aalgorithm == pooling_avg_include_padding) { + num_summands = pd.kw * pd.kh * pd.kd; + } - if (p.aalgorithm == pooling_avg_include_padding || - p.aalgorithm == pooling_avg_exclude_padding) { - acc_ref = out_round<data_t>( - (float)acc_ref / num_summands); - } + if (p.aalgorithm == pooling_avg_include_padding || + p.aalgorithm == pooling_avg_exclude_padding) { + acc_ref = out_round<data_t>( + (float)acc_ref / num_summands); + } - const data_t out_ref = (data_t)acc_ref; - EXPECT_NEAR(out, out_ref, 1e-6); - if(p.aalgorithm == pooling_max - && p.aprop_kind == forward_training) { - EXPECT_EQ(out_index, out_ref_index) << " n = " << n - << " c = " << c << " od = " << od << " oh = " << oh - << " ow = " << ow; - } - } - } + const data_t out_ref = (data_t)acc_ref; + EXPECT_NEAR(out, out_ref, 1e-6); + if(p.aalgorithm == pooling_max + && p.aprop_kind == forward_training) { + EXPECT_EQ(out_index, out_ref_index) << " n = " << n + << " c = " << c << " od = " << od << " oh = " << oh + << " ow = " << ow; } } - } + ); } template <typename data_t> @@ -486,6 +479,35 @@ TEST_P(pooling_test_float, TestsPooling) } INSTANTIATE_TEST_CASE_P( + TestPoolingForwardZeroDim, pooling_test_float, ::testing::Values( + pool_test_params_float{ prop_kind::forward_training, + engine::kind::cpu, algorithm::pooling_max, memory::format::nchw, + memory::format::nchw, EXPAND_SIZES_2D( 2, 0, 4, 4, 4, 4, 3, 3, 1, 1, 1, 1 )}, + pool_test_params_float{ prop_kind::forward_training, + engine::kind::cpu, algorithm::pooling_max, memory::format::nhwc, + memory::format::nhwc, EXPAND_SIZES_2D( 0, 4, 4, 4, 4, 4, 3, 3, 1, 1, 1, 1 )}, + pool_test_params_float{ prop_kind::forward_training, + engine::kind::cpu, algorithm::pooling_max, memory::format::nchw, + memory::format::nchw, EXPAND_SIZES_2D( 2, 4, 0, 4, 4, 4, 3, 3, 1, 1, 1, 1 )} + )); + +INSTANTIATE_TEST_CASE_P( + TestPoolingForwardEF, pooling_test_float, ::testing::Values( + pool_test_params_float{ prop_kind::forward_training, + engine::kind::cpu, algorithm::pooling_max, memory::format::nchw, + memory::format::nchw, EXPAND_SIZES_2D( 2, -4, 4, 4, 4, 4, 3, 3, 1, 1, 1, 1 ), + true, mkldnn_invalid_arguments}, + pool_test_params_float{ prop_kind::forward_training, + engine::kind::cpu, algorithm::pooling_max, memory::format::nchw, + memory::format::nchw, EXPAND_SIZES_2D( -1, 4, 4, 4, 4, 4, 3, 3, 1, 1, 1, 1 ), + true, mkldnn_invalid_arguments}, + pool_test_params_float{ prop_kind::forward_training, + engine::kind::cpu, algorithm::eltwise_square, memory::format::nchw, + memory::format::nchw, EXPAND_SIZES_2D( 2, 4, 4, 4, 4, 4, 3, 3, 1, 1, 1, 1 ), + true, mkldnn_invalid_arguments} + )); + +INSTANTIATE_TEST_CASE_P( TestPooling_nChw16c_with_padded, pooling_test_float, ::testing::Values( pool_test_params{ prop_kind::forward_training, engine::kind::cpu, algorithm::pooling_max, memory::format::nChw16c, @@ -511,6 +533,31 @@ INSTANTIATE_TEST_CASE_P( )); INSTANTIATE_TEST_CASE_P( + TestPooling_nChw8c_with_padded, pooling_test_float, ::testing::Values( + pool_test_params{ prop_kind::forward_training, + engine::kind::cpu, algorithm::pooling_max, memory::format::nChw8c, + memory::format::nChw8c, EXPAND_SIZES_2D(4, 5, 6, 6, 7, 7, 2, 2, 1, 1, 1, 1) }, + pool_test_params{ prop_kind::forward_training, + engine::kind::cpu, algorithm::pooling_max, memory::format::nChw8c, + memory::format::nChw8c, EXPAND_SIZES_2D(4, 9, 60, 60, 31, 31, 3, 4, 1, 1, 2, 2) }, + pool_test_params{ prop_kind::forward_training, + engine::kind::cpu, algorithm::pooling_avg_exclude_padding, memory::format::nChw8c, + memory::format::nChw8c, EXPAND_SIZES_2D(4, 14, 60, 60, 31, 31, 3, 2, 1, 1, 2, 2) }, + pool_test_params{ prop_kind::forward_training, + engine::kind::cpu, algorithm::pooling_avg_include_padding, memory::format::nChw8c, + memory::format::nChw8c, EXPAND_SIZES_2D(4, 17, 60, 60, 31, 31, 4, 3, 1, 1, 2, 2) }, + pool_test_params{ prop_kind::forward_inference, + engine::kind::cpu, algorithm::pooling_max, memory::format::nChw8c, + memory::format::nChw8c, EXPAND_SIZES_2D(4, 14, 60, 60, 31, 31, 2, 3, 1, 1, 2, 2) }, + pool_test_params{ prop_kind::forward_inference, + engine::kind::cpu, algorithm::pooling_avg_exclude_padding, memory::format::nChw8c, + memory::format::nChw8c, EXPAND_SIZES_2D(4, 25, 60, 60, 31, 31, 2, 4, 1, 1, 2, 2) }, + pool_test_params{ prop_kind::forward_inference, + engine::kind::cpu, algorithm::pooling_avg_include_padding, memory::format::nChw8c, + memory::format::nChw8c, EXPAND_SIZES_2D(4, 28, 60, 60, 31, 31, 4, 2, 1, 1, 2, 2) } + )); + +INSTANTIATE_TEST_CASE_P( TestPoolingForwardMaxKernelSlipsToPadding, pooling_test_float, ::testing::Values( pool_test_params{ prop_kind::forward_training, engine::kind::cpu, algorithm::pooling_max, memory::format::nchw, @@ -549,6 +596,28 @@ INSTANTIATE_TEST_CASE_P( )); INSTANTIATE_TEST_CASE_P( + TestPooling3D_nCdhw8c, pooling_test_float, ::testing::Values( + pool_test_params{ prop_kind::forward_training, + engine::kind::cpu, algorithm::pooling_max, memory::format::nCdhw8c, + memory::format::nCdhw8c, EXPAND_SIZES_3D(2, 32, 60, 60, 60, 31, 31, 31, 2, 3, 4, 1, 1, 1, 2, 2, 2) }, + pool_test_params{ prop_kind::forward_training, + engine::kind::cpu, algorithm::pooling_avg_exclude_padding, memory::format::nCdhw8c, + memory::format::nCdhw8c, EXPAND_SIZES_3D(2, 32, 60, 60, 60, 31, 31, 31, 4, 3, 2, 1, 1, 1, 2, 2, 2) }, + pool_test_params{ prop_kind::forward_training, + engine::kind::cpu, algorithm::pooling_avg_include_padding, memory::format::nCdhw8c, + memory::format::nCdhw8c, EXPAND_SIZES_3D(2, 32, 60, 60, 60, 31, 31, 31, 2, 4, 3, 1, 1, 1, 2, 2, 2) }, + pool_test_params{ prop_kind::forward_inference, + engine::kind::cpu, algorithm::pooling_max, memory::format::nCdhw8c, + memory::format::nCdhw8c, EXPAND_SIZES_3D(2, 32, 60, 60, 60, 31, 31, 31, 4, 2, 3, 1, 1, 1, 2, 2, 2) }, + pool_test_params{ prop_kind::forward_inference, + engine::kind::cpu, algorithm::pooling_avg_exclude_padding, memory::format::nCdhw8c, + memory::format::nCdhw8c, EXPAND_SIZES_3D(2, 32, 60, 60, 60, 31, 31, 31, 3, 2, 4, 1, 1, 1, 2, 2, 2) }, + pool_test_params{ prop_kind::forward_inference, + engine::kind::cpu, algorithm::pooling_avg_include_padding, memory::format::nCdhw8c, + memory::format::nCdhw8c, EXPAND_SIZES_3D(2, 32, 60, 60, 60, 31, 31, 31, 3, 4, 2, 1, 1, 1, 2, 2, 2) } + )); + +INSTANTIATE_TEST_CASE_P( TestPooling3D_ndhwc, pooling_test_float, ::testing::Values( pool_test_params{ prop_kind::forward_training, engine::kind::cpu, algorithm::pooling_max, memory::format::ndhwc, @@ -632,22 +701,6 @@ INSTANTIATE_TEST_CASE_P( )); INSTANTIATE_TEST_CASE_P( - TestPoolingForwardEF, pooling_test_float, ::testing::Values( - pool_test_params_float{ prop_kind::forward_training, - engine::kind::cpu, algorithm::pooling_max, memory::format::nchw, - memory::format::nchw, EXPAND_SIZES_2D( 2, 0, 4, 4, 4, 4, 3, 3, 1, 1, 1, 1 ), - true, mkldnn_invalid_arguments}, - pool_test_params_float{ prop_kind::forward_training, - engine::kind::cpu, algorithm::pooling_max, memory::format::nchw, - memory::format::nchw, EXPAND_SIZES_2D( 0, 4, 4, 4, 4, 4, 3, 3, 1, 1, 1, 1 ), - true, mkldnn_invalid_arguments}, - pool_test_params_float{ prop_kind::forward_training, - engine::kind::cpu, algorithm::pooling_max, memory::format::nchw, - memory::format::nchw, EXPAND_SIZES_2D( 2, 4, 0, 4, 4, 4, 3, 3, 1, 1, 1, 1 ), - true, mkldnn_invalid_arguments} - )); - -INSTANTIATE_TEST_CASE_P( TestPoolingForwardMax, pooling_test_float, ::testing::Values( pool_test_params_float{ prop_kind::forward_training, engine::kind::cpu, algorithm::pooling_max, memory::format::nchw, |