From 9285adb5ac8e28a9cc82ce708bb2975dc5a074dd Mon Sep 17 00:00:00 2001 From: Gian Marco Iodice Date: Thu, 5 Sep 2019 16:10:27 +0100 Subject: COMPMID-2599: Implement a new and generic depthwise convolution on OpenCL (Fp32/FP16-NHWC) Part 1 Change-Id: I5e1d27a7006199e9229e455a1df9bfc2ed4e8341 Signed-off-by: Gian Marco Iodice Reviewed-on: https://review.mlplatform.org/c/1898 Comments-Addressed: Arm Jenkins Reviewed-by: Giorgio Arena Tested-by: Arm Jenkins --- .../CL/DepthwiseConvolutionLayerNative.cpp | 305 +++++++++++++++++++++ .../fixtures/DepthwiseConvolutionLayerFixture.h | 121 ++++++++ 2 files changed, 426 insertions(+) create mode 100644 tests/validation/CL/DepthwiseConvolutionLayerNative.cpp (limited to 'tests') diff --git a/tests/validation/CL/DepthwiseConvolutionLayerNative.cpp b/tests/validation/CL/DepthwiseConvolutionLayerNative.cpp new file mode 100644 index 0000000000..bbcded9267 --- /dev/null +++ b/tests/validation/CL/DepthwiseConvolutionLayerNative.cpp @@ -0,0 +1,305 @@ +/* + * Copyright (c) 2019 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/CLDepthwiseConvolutionLayerNativeKernel.h" +#include "arm_compute/core/KernelDescriptors.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "arm_compute/runtime/CL/CLTensor.h" +#include "arm_compute/runtime/CL/CLTensorAllocator.h" +#include "tests/CL/CLAccessor.h" +#include "tests/CL/Helper.h" +#include "tests/PaddingCalculator.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/DepthwiseConvolutionLayerFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +using namespace arm_compute::misc::shape_calculator; + +// Create function for CLDepthwiseConvolutionLayerNativeKernel +using CLDepthwiseConvolutionLayerNative = CLSynthetizeFunction; + +// Fixture for CLDepthwiseConvolutionLayerNative +template +using CLDepthwiseConvolutionLayerNativeFixture = DepthwiseConvolutionLayerNativeConfigurableValidationFixture; + +namespace +{ +// *INDENT-OFF* +// clang-format off +RelativeTolerance rel_tolerance_f32(0.001f); +constexpr float abs_tolerance_f32(0.0001f); + +RelativeTolerance tolerance_f16(half_float::half(0.01)); + +/** Width values to test - Precommit */ +const auto width_values_precommit = framework::dataset::make("width", { 37U } ); + +/** Width values to test - Nightly */ +const auto width_values_nightly = framework::dataset::make("width", { 53U, 47U } ); + +/** Height values to test - Precommit */ +const auto height_values_precommit = framework::dataset::make("height", { 19U } ); + +/** Height values to test - Nightly */ +const auto height_values_nightly = framework::dataset::make("height", { 39U, 43U } ); + +/** Channel values to test - Precommit */ +const auto channel_values_precommit = framework::dataset::make("channels", { 15U }); + +/** Channel values to test - Nightly */ +const auto channel_values_nightly = framework::dataset::make("channels", { 33U, 19U }); + +/** Batch values to test - Precommit */ +const auto batch_values_precommit = framework::dataset::make("batch", { 1U, 2U }); + +/** Batch values to test - Nightly */ +const auto batch_values_nightly = framework::dataset::make("batch", { 1U, 3U }); + +/** Kernel size values to test - Precommit */ +const auto kernel_sz_values_precommit = framework::dataset::make("kernel_size", { Size2D(1U, 1U), Size2D(1U, 3U), Size2D(5U, 5U) }); + +/** Kernel size values to test - Nightly */ +const auto kernel_sz_values_nightly = framework::dataset::make("kernel_size", { Size2D(3U, 5U), Size2D(5U, 1U), Size2D(1U, 7U), Size2D(9U, 7U) }); + +/** Depth multiplier values to test - All */ +const auto depth_multiplier_values = framework::dataset::make("depth_multiplier", {3U}); + +/** Dilation values to test - All */ +const auto dilation_values = framework::dataset::make("dilation", { Size2D(1U, 1U), Size2D(3U, 3U) }); + +/** Stride values to test - All */ +const auto stride_values = framework::dataset::make("stride", { Size2D(1U, 1U), Size2D(3U, 2U) }); + +/** Padding values to test - All */ +const auto padding_valid_values = framework::dataset::make("padding_valid", { true, false }); + +/** Data type values to test - All */ +const auto data_type_values = framework::dataset::make("data_type", { DataType::F32, DataType::F16 }); + +/** Data layout values to test - All */ +const auto data_layout_values = framework::dataset::make("data_layout", { DataLayout::NHWC }); + +/** N0 values to test - Precommit */ +const auto n0_values_precommit = framework::dataset::make("N0", {2, 4}); + +/** N0 values to test - Nightly */ +const auto n0_values_nightly = framework::dataset::make("N0", {3, 8}); + +/** Activation values to test */ +const auto act_values = framework::dataset::make("Activation", +{ + ActivationLayerInfo(), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 8.f, 2.f), +}); + +} // namespace + +TEST_SUITE(CL) +TEST_SUITE(DepthwiseConvolutionLayerNative) +TEST_SUITE(Float) +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerNativeFixture, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + width_values_precommit, + height_values_precommit), + channel_values_precommit), + batch_values_precommit), + kernel_sz_values_precommit), + framework::dataset::make("depth_multiplier", 1)), + dilation_values), + stride_values), + padding_valid_values), + framework::dataset::make("DataType", DataType::F32)), + data_layout_values), + act_values), + n0_values_precommit)) +{ + // Validate output + validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthwiseConvolutionLayerNativeFixture, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + width_values_nightly, + height_values_nightly), + channel_values_nightly), + batch_values_nightly), + kernel_sz_values_nightly), + framework::dataset::make("depth_multiplier", 1)), + dilation_values), + stride_values), + padding_valid_values), + framework::dataset::make("DataType", DataType::F32)), + data_layout_values), + act_values), + n0_values_nightly)) +{ + // Validate output + validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); +} +TEST_SUITE_END() // FP32 + +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerNativeFixture, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + width_values_precommit, + height_values_precommit), + channel_values_precommit), + batch_values_precommit), + kernel_sz_values_precommit), + framework::dataset::make("depth_multiplier", 1)), + dilation_values), + stride_values), + padding_valid_values), + framework::dataset::make("DataType", DataType::F16)), + data_layout_values), + act_values), + n0_values_precommit)) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthwiseConvolutionLayerNativeFixture, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + width_values_nightly, + height_values_nightly), + channel_values_nightly), + batch_values_nightly), + kernel_sz_values_nightly), + framework::dataset::make("depth_multiplier", 1)), + dilation_values), + stride_values), + padding_valid_values), + framework::dataset::make("DataType", DataType::F16)), + data_layout_values), + act_values), + n0_values_nightly)) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} +TEST_SUITE_END() // FP16 +TEST_SUITE_END() // Float +TEST_SUITE(DepthMultiplier) +TEST_SUITE(Float) +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerNativeFixture, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + width_values_precommit, + height_values_precommit), + channel_values_precommit), + batch_values_precommit), + kernel_sz_values_precommit), + depth_multiplier_values), + dilation_values), + stride_values), + padding_valid_values), + framework::dataset::make("DataType", DataType::F32)), + data_layout_values), + act_values), + framework::dataset::make("N0", 1))) +{ + // Validate output + validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthwiseConvolutionLayerNativeFixture, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + width_values_nightly, + height_values_nightly), + channel_values_nightly), + batch_values_nightly), + kernel_sz_values_nightly), + depth_multiplier_values), + dilation_values), + stride_values), + padding_valid_values), + framework::dataset::make("DataType", DataType::F32)), + data_layout_values), + act_values), + framework::dataset::make("N0", 1))) +{ + // Validate output + validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); +} +TEST_SUITE_END() // FP32 + +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerNativeFixture, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + width_values_precommit, + height_values_precommit), + channel_values_precommit), + batch_values_precommit), + kernel_sz_values_precommit), + depth_multiplier_values), + dilation_values), + stride_values), + padding_valid_values), + framework::dataset::make("DataType", DataType::F16)), + data_layout_values), + act_values), + framework::dataset::make("N0", 1))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthwiseConvolutionLayerNativeFixture, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + width_values_nightly, + height_values_nightly), + channel_values_nightly), + batch_values_nightly), + kernel_sz_values_nightly), + depth_multiplier_values), + dilation_values), + stride_values), + padding_valid_values), + framework::dataset::make("DataType", DataType::F16)), + data_layout_values), + act_values), + framework::dataset::make("N0", 1))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} +TEST_SUITE_END() // FP16 +TEST_SUITE_END() // Float +TEST_SUITE_END() // DepthMultiplier +TEST_SUITE_END() // DepthwiseConvolutionLayerNative +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h b/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h index a3ac49eef1..2c9b31866b 100644 --- a/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h +++ b/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h @@ -301,6 +301,127 @@ protected: SimpleTensor _reference{}; }; +template +class DepthwiseConvolutionLayerNativeConfigurableValidationFixture : public DepthwiseConvolutionLayerValidationGenericFixture +{ +public: + template + void setup(size_t width, size_t height, size_t channel, size_t batch, Size2D kernel_size, size_t depth_multiplier, Size2D dilation, Size2D stride, bool padding_valid, DataType data_type, + DataLayout data_layout, const ActivationLayerInfo &act_info, unsigned int n0) + { + const TensorShape src_shape(width, height, channel, batch); + const TensorShape weights_shape(kernel_size.width, kernel_size.height, channel * depth_multiplier); + const TensorShape biases_shape(weights_shape.z()); + + PadStrideInfo conv_info; + if(padding_valid) + { + conv_info = PadStrideInfo(); + } + else + { + conv_info = calculate_same_pad(src_shape, weights_shape, PadStrideInfo(stride.width, stride.height), DataLayout::NCHW, dilation); + } + + _target = compute_target(src_shape, weights_shape, biases_shape, conv_info, dilation, depth_multiplier, data_type, data_layout, act_info, n0); + _reference = compute_reference(src_shape, weights_shape, biases_shape, conv_info, dilation, depth_multiplier, data_type, act_info); + } + +protected: + template + void fill(U &&tensor, int i) + { + switch(tensor.data_type()) + { + case DataType::F32: + { + std::uniform_real_distribution<> distribution(-1.0f, 1.0f); + library->fill(tensor, distribution, i); + break; + } + case DataType::F16: + { + std::uniform_real_distribution<> distribution(-1.0f, 1.0f); + library->fill(tensor, distribution, i); + break; + } + default: + library->fill_tensor_uniform(tensor, i); + } + } + + TensorType compute_target(TensorShape input_shape, TensorShape weights_shape, TensorShape biases_shape, PadStrideInfo &conv_info, Size2D dilation, + unsigned int depth_multiplier, const DataType data_type, const DataLayout data_layout, const ActivationLayerInfo &act_info, unsigned int n0) + { + if(data_layout == DataLayout::NHWC) + { + permute(input_shape, PermutationVector(2U, 0U, 1U)); + permute(weights_shape, PermutationVector(2U, 0U, 1U)); + } + + // Create tensors + TensorType src = create_tensor(input_shape, data_type, 1, QuantizationInfo(), data_layout); + TensorType weights = create_tensor(weights_shape, data_type, 1, QuantizationInfo(), data_layout); + TensorType biases = create_tensor(biases_shape, data_type, 1, QuantizationInfo(), data_layout); + TensorType dst = create_tensor(TensorShape(), data_type, 1, QuantizationInfo(), data_layout); + + DWCWeightsKernelInfo dwc_weights_info; + dwc_weights_info.n0 = n0; + + DWCKernelInfo dwc_info; + dwc_info.activation_info = act_info; + + // Create Depthwise Convolution configure function + FunctionType dwc; + dwc.configure(&src, &weights, &biases, &dst, dwc_weights_info, dwc_info, conv_info, depth_multiplier, dilation); + + ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(weights.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(biases.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Allocate tensors + src.allocator()->allocate(); + weights.allocator()->allocate(); + biases.allocator()->allocate(); + dst.allocator()->allocate(); + + ARM_COMPUTE_EXPECT(!src.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!weights.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!biases.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Fill tensors + fill(AccessorType(src), 0); + fill(AccessorType(weights), 1); + fill(AccessorType(biases), 2); + + // Compute function + dwc.run(); + + return dst; + } + + SimpleTensor compute_reference(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &biases_shape, const PadStrideInfo &conv_info, + const Size2D &dilation, unsigned int depth_multiplier, const DataType data_type, const ActivationLayerInfo &act_info) + { + SimpleTensor src{ input_shape, data_type }; + SimpleTensor weights{ weights_shape, data_type }; + SimpleTensor biases{ biases_shape, data_type }; + + fill(src, 0); + fill(weights, 1); + fill(biases, 2); + + const TensorShape dst_shape = compute_depthwise_convolution_shape(TensorInfo(input_shape, 1, data_type), TensorInfo(weights_shape, 1, data_type), conv_info, + depth_multiplier, dilation); + return reference::activation_layer(reference::depthwise_convolution(src, weights, biases, dst_shape, conv_info, depth_multiplier, dilation), act_info); + } + + TensorType _target{}; + SimpleTensor _reference{}; +}; + template class DepthwiseConvolutionLayerValidationQuantizedFixture : public DepthwiseConvolutionLayerValidationGenericFixture { -- cgit v1.2.1