From 4a626a7d52e9c4759bdc16b65401a53779dd975f Mon Sep 17 00:00:00 2001 From: Pablo Tello Date: Wed, 4 Apr 2018 10:01:14 +0100 Subject: COMPMID-801: NHWC support in CLIm2Col. And extended tests coverage adding kernel shapes 3x1, 1x5 and 7x7 Change-Id: Ia7c1d4da2368d5f5fbc1a41187f4ac1aca5f150f Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/127727 Tested-by: Jenkins Reviewed-by: Gian Marco Iodice --- tests/CL/Helper.h | 9 ++ tests/validation/CL/Im2Col.cpp | 143 ++++++++++++++++++++++++++++++ tests/validation/NEON/Im2Col.cpp | 25 +++++- tests/validation/fixtures/Im2ColFixture.h | 3 +- tests/validation/reference/Im2Col.cpp | 17 +++- 5 files changed, 188 insertions(+), 9 deletions(-) create mode 100644 tests/validation/CL/Im2Col.cpp (limited to 'tests') diff --git a/tests/CL/Helper.h b/tests/CL/Helper.h index 30fbe568f4..32f9ca00e3 100644 --- a/tests/CL/Helper.h +++ b/tests/CL/Helper.h @@ -47,6 +47,15 @@ public: k->configure(std::forward(args)...); _kernel = std::move(k); } + /** Validate input arguments + * + * @param[in] args Configuration arguments. + */ + template + static Status validate(Args &&... args) + { + return K::validate(std::forward(args)...); + } }; /** As above but this also setups a Zero border on the input tensor of the specified bordersize */ diff --git a/tests/validation/CL/Im2Col.cpp b/tests/validation/CL/Im2Col.cpp new file mode 100644 index 0000000000..bfe0665fa9 --- /dev/null +++ b/tests/validation/CL/Im2Col.cpp @@ -0,0 +1,143 @@ +/* + * Copyright (c) 2018 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/CL/kernels/CLIm2ColKernel.h" +#include "arm_compute/core/Types.h" +#include "tests/CL/Helper.h" + +#include "tests/CL/CLAccessor.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/Im2ColFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +const auto conv_filter_sizes = framework::dataset::make("KernelDims", { Size2D(3U, 3U), Size2D(3U, 1U), Size2D(1U, 5U), Size2D(5U, 5U), Size2D(7U, 7U) }); +const auto padstrides = framework::dataset::make("PadStride", { PadStrideInfo(1U, 1U, 0U, 0U), PadStrideInfo(1U, 1U, 1U, 1U), PadStrideInfo(2U, 2U, 0U, 2U) }); +const auto conv_args = combine(combine(combine(conv_filter_sizes, padstrides), + framework::dataset::make("QuantizationInfo", QuantizationInfo(0.5f, 10))), + framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })); + +} // namespace +TEST_SUITE(CL) +TEST_SUITE(Im2Col) + +using CLIm2Col = CLSynthetizeFunction; + +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( + framework::dataset::make("InputInfo", { TensorInfo(TensorShape(10U, 12U, 2U), 1, DataType::U8), // Unsupported data type + TensorInfo(TensorShape(10U, 12U, 2U), 1, DataType::F32), // Mismatching data type + TensorInfo(TensorShape(10U, 12U, 2U), 1, DataType::QS8, 2), // Mismatching fixed point + TensorInfo(TensorShape(10U, 12U, 2U), 1, DataType::QASYMM8), // Bias not supported with QASYMM8 + TensorInfo(TensorShape(10U, 12U, 2U), 1, DataType::QASYMM8), // Mismatching shapes + TensorInfo(TensorShape(10U, 12U, 2U, 2U), 1, DataType::QASYMM8), + }), + framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(3U, 4U, 10U, 2U), 1, DataType::F16), + TensorInfo(TensorShape(3U, 4U, 10U, 2U), 1, DataType::F16), + TensorInfo(TensorShape(3U, 4U, 10U, 2U), 1, DataType::QS8, 3), + TensorInfo(TensorShape(3U, 3U, 10U, 2U), 1, DataType::QASYMM8), + TensorInfo(TensorShape(3U, 4U, 10U, 2U), 1, DataType::QASYMM8), + TensorInfo(TensorShape(18U, 80U, 1U, 2U), 1, DataType::QASYMM8), + })), + framework::dataset::make("HasBias", { true, true, true, true, false, false })), + framework::dataset::make("Expected", { false, false, false, false, true, true })), + input_info, output_info, has_bias, expected) +{ + + bool status = bool(CLIm2Col::validate(&input_info, &output_info, Size2D(3U, 3U), PadStrideInfo(), has_bias)); + ARM_COMPUTE_EXPECT(status == expected, framework::LogLevel::ERRORS); +} +// clang-format on +// *INDENT-ON* + +template +using CLIm2ColFixture = Im2ColValidationFixture; +TEST_SUITE(Float) +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunSmall, CLIm2ColFixture, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::F32)), + conv_args)) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +TEST_SUITE_END() + +FIXTURE_DATA_TEST_CASE(RunLarge, CLIm2ColFixture, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", DataType::F32)), + conv_args)) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} + +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(RunSmall, CLIm2ColFixture, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::F16)), + conv_args)) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunLarge, CLIm2ColFixture, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", DataType::F16)), + conv_args)) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +TEST_SUITE_END() + +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + +TEST_SUITE_END() + +TEST_SUITE(QASYMM8) +FIXTURE_DATA_TEST_CASE(RunSmall, CLIm2ColFixture, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::QASYMM8)), + conv_args)) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunLarge, CLIm2ColFixture, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", DataType::QASYMM8)), + conv_args)) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +TEST_SUITE_END() + +TEST_SUITE_END() +TEST_SUITE_END() +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/NEON/Im2Col.cpp b/tests/validation/NEON/Im2Col.cpp index 50081f07b0..3a45fa7ae4 100644 --- a/tests/validation/NEON/Im2Col.cpp +++ b/tests/validation/NEON/Im2Col.cpp @@ -39,9 +39,10 @@ namespace validation { namespace { -const auto conv_args = combine(combine(combine(framework::dataset::make("KernelDims", { Size2D(3U, 3U), Size2D(5U, 5U) }), framework::dataset::make("PadStride", { PadStrideInfo(1U, 1U, 0U, 0U), PadStrideInfo(1U, 1U, 1U, 1U), PadStrideInfo(2U, 2U, 0U, 2U) })), - framework::dataset::make("QuantizationInfo", QuantizationInfo(0.5f, 10))), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })); +const auto conv_filter_sizes = framework::dataset::make("KernelDims", { Size2D(3U, 3U), Size2D(3U, 1U), Size2D(1U, 5U), Size2D(5U, 5U), Size2D(7U, 7U) }); +const auto conv_args = combine(combine(combine(conv_filter_sizes, framework::dataset::make("PadStride", { PadStrideInfo(1U, 1U, 0U, 0U), PadStrideInfo(1U, 1U, 1U, 1U), PadStrideInfo(2U, 2U, 0U, 2U) })), + framework::dataset::make("QuantizationInfo", QuantizationInfo(0.5f, 10))), + framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })); } // namespace TEST_SUITE(NEON) TEST_SUITE(Im2Col) @@ -84,6 +85,12 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEIm2ColFixture, framework::DatasetMode: // Validate output validate(Accessor(_target), _reference); } +FIXTURE_DATA_TEST_CASE(RunLarge, NEIm2ColFixture, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", DataType::F32)), + conv_args)) +{ + // Validate output + validate(Accessor(_target), _reference); +} TEST_SUITE_END() #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC @@ -95,6 +102,12 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEIm2ColFixture, framework::DatasetMode:: // Validate output validate(Accessor(_target), _reference); } +FIXTURE_DATA_TEST_CASE(RunLarge, NEIm2ColFixture, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", DataType::F16)), + conv_args)) +{ + // Validate output + validate(Accessor(_target), _reference); +} TEST_SUITE_END() #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ @@ -108,6 +121,12 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEIm2ColFixture, framework::DatasetMod // Validate output validate(Accessor(_target), _reference); } +FIXTURE_DATA_TEST_CASE(RunLarge, NEIm2ColFixture, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", DataType::QASYMM8)), + conv_args)) +{ + // Validate output + validate(Accessor(_target), _reference); +} TEST_SUITE_END() TEST_SUITE_END() diff --git a/tests/validation/fixtures/Im2ColFixture.h b/tests/validation/fixtures/Im2ColFixture.h index 7ef3cdcdcd..6e532e7803 100644 --- a/tests/validation/fixtures/Im2ColFixture.h +++ b/tests/validation/fixtures/Im2ColFixture.h @@ -66,8 +66,7 @@ public: input_info.set_data_layout(_data_layout); const TensorShape output_shape = compute_im2col_conv_shape(&input_info, _kernel_dims, _conv_info, _has_bias, Size2D(1U, 1U)); - - _target = compute_target(input_shape, output_shape, data_type); + _target = compute_target(input_shape, output_shape, data_type); compute_reference(input_shape, output_shape, data_type); } diff --git a/tests/validation/reference/Im2Col.cpp b/tests/validation/reference/Im2Col.cpp index 5685b60026..83ef8b40a5 100644 --- a/tests/validation/reference/Im2Col.cpp +++ b/tests/validation/reference/Im2Col.cpp @@ -55,11 +55,16 @@ void im2col_nchw(const SimpleTensor &src, SimpleTensor &dst, const Size2D const int pad_val = is_data_type_quantized_asymmetric(src.data_type()) ? src.quantization_info().offset : 0; int dst_idx = 0; + // dst[dst_idx++] will write out of bounds if kernel_height == kernel_width == 1 because lasty will be the bottom padding row + // and this is not present in the dst buffer + const int lasty = src_height + (kernel_height > 1 ? pad_y : 0) - kernel_height; + const int lastx = src_width + (kernel_width > 1 ? pad_x : 0) - kernel_width; + for(int b = 0; b < batches; ++b) { - for(int y = -pad_y; y <= (src_height + pad_y - kernel_height); y += stride_y) + for(int y = -pad_y; y <= lasty; y += stride_y) { - for(int x = -pad_x; x <= (src_width + pad_x - kernel_width); x += stride_x) + for(int x = -pad_x; x <= lastx; x += stride_x) { for(int z = 0; z < src_depth; ++z) { @@ -97,11 +102,15 @@ void im2col_nhwc(const SimpleTensor &src, SimpleTensor &dst, const Size2D const int batches = src.shape().total_size_upper(3); const int pad_val = is_data_type_quantized_asymmetric(src.data_type()) ? src.quantization_info().offset : 0; int dst_idx = 0; + + const int lasty = src_height + (kernel_height > 1 ? pad_y : 0) - kernel_height; + const int lastx = src_width + (kernel_width > 1 ? pad_x : 0) - kernel_width; + for(int b = 0; b < batches; ++b) { - for(int y = -pad_y; y <= (src_height + pad_y - kernel_height); y += stride_y) + for(int y = -pad_y; y <= lasty; y += stride_y) { - for(int x = -pad_x; x <= (src_width + pad_x - kernel_width); x += stride_x) + for(int x = -pad_x; x <= lastx; x += stride_x) { for(int z = 0; z < src_depth; ++z) { -- cgit v1.2.1