From 93a690edc48ad78cc9d7a580638475c6e77b32c2 Mon Sep 17 00:00:00 2001 From: Giorgio Arena Date: Tue, 1 Aug 2017 16:09:33 +0100 Subject: COMPMID-452 CL Depthwise Separable Convolution Layer kernel implementation, validation and benchmarking for 3x3xC depthwise filter and DataType::F32. Change-Id: I95c0c87709763cdbf58d0de66025eac86e30791b Reviewed-on: http://mpd-gerrit.cambridge.arm.com/82768 Tested-by: Kaizen Reviewed-by: Steven Niu --- tests/fixtures_new/DepthwiseConvolutionFixture.h | 88 ++++++++++++++++++ .../DepthwiseSeparableConvolutionLayerFixture.h | 103 +++++++++++++++++++++ 2 files changed, 191 insertions(+) create mode 100644 tests/fixtures_new/DepthwiseConvolutionFixture.h create mode 100644 tests/fixtures_new/DepthwiseSeparableConvolutionLayerFixture.h (limited to 'tests/fixtures_new') diff --git a/tests/fixtures_new/DepthwiseConvolutionFixture.h b/tests/fixtures_new/DepthwiseConvolutionFixture.h new file mode 100644 index 0000000000..8223c210d4 --- /dev/null +++ b/tests/fixtures_new/DepthwiseConvolutionFixture.h @@ -0,0 +1,88 @@ +/* + * Copyright (c) 2017 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. + */ +#ifndef ARM_COMPUTE_TEST_DEPTHWISECONVOLUTIONFIXTURE +#define ARM_COMPUTE_TEST_DEPTHWISECONVOLUTIONFIXTURE + +#include "arm_compute/core/TensorShape.h" +#include "arm_compute/core/Types.h" +#include "framework/Fixture.h" +#include "tests/Globals.h" +#include "tests/Utils.h" + +namespace arm_compute +{ +namespace test +{ +/** Fixture that can be used for NEON and CL */ +template +class DepthwiseConvolutionFixture : public framework::Fixture +{ +public: + template + void setup(TensorShape src_shape, TensorShape weights_shape, TensorShape dst_shape, PadStrideInfo info, DataType data_type, int batches) + { + // Set batched in source and destination shapes + const unsigned int fixed_point_position = 4; + src_shape.set(3 /* batch */, batches); + dst_shape.set(3 /* batch */, batches); + + // Create tensors + src = create_tensor(src_shape, data_type, 1, fixed_point_position); + weights = create_tensor(weights_shape, data_type, 1, fixed_point_position); + dst = create_tensor(dst_shape, data_type, 1, fixed_point_position); + + // Create and configure function + depth_conv.configure(&src, &dst, &weights, info); + + // Allocate tensors + src.allocator()->allocate(); + weights.allocator()->allocate(); + dst.allocator()->allocate(); + + // Fill tensors + library->fill_tensor_uniform(Accessor(src), 0); + library->fill_tensor_uniform(Accessor(weights), 1); + } + + void run() + { + depth_conv.run(); + } + + void teardown() + { + src.allocator()->free(); + weights.allocator()->free(); + dst.allocator()->free(); + } + +private: + TensorType src{}; + TensorType weights{}; + TensorType dst{}; + Function depth_conv{}; +}; +} // namespace test +} // namespace arm_compute +#endif /* ARM_COMPUTE_TEST_DEPTHWISECONVOLUTIONFIXTURE */ diff --git a/tests/fixtures_new/DepthwiseSeparableConvolutionLayerFixture.h b/tests/fixtures_new/DepthwiseSeparableConvolutionLayerFixture.h new file mode 100644 index 0000000000..c4ace2ba1c --- /dev/null +++ b/tests/fixtures_new/DepthwiseSeparableConvolutionLayerFixture.h @@ -0,0 +1,103 @@ +/* + * Copyright (c) 2017 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. + */ +#ifndef ARM_COMPUTE_TEST_DEPTHWISESEPARABLECONVOLUTIONLAYERFIXTURE +#define ARM_COMPUTE_TEST_DEPTHWISESEPARABLECONVOLUTIONLAYERFIXTURE + +#include "arm_compute/core/TensorShape.h" +#include "arm_compute/core/Types.h" +#include "framework/Fixture.h" +#include "tests/Globals.h" +#include "tests/Utils.h" + +namespace arm_compute +{ +namespace test +{ +/** Fixture that can be used for NEON and CL */ +template +class DepthwiseSeparableConvolutionLayerFixture : public framework::Fixture +{ +public: + template + void setup(TensorShape src_shape, TensorShape depthwise_weights_shape, TensorShape depthwise_out_shape, TensorShape pointwise_weights_shape, TensorShape biases_shape, TensorShape dst_shape, + PadStrideInfo pad_stride_depthwise_info, PadStrideInfo pad_stride_pointwise_info, DataType data_type, int batches) + { + // Set batched in source and destination shapes + const unsigned int fixed_point_position = 4; + src_shape.set(3 /* batch */, batches); + depthwise_out_shape.set(3 /* batch */, batches); + dst_shape.set(3 /* batch */, batches); + + src = create_tensor(src_shape, data_type, 1, fixed_point_position); + depthwise_weights = create_tensor(depthwise_weights_shape, data_type, 1, fixed_point_position); + depthwise_out = create_tensor(depthwise_out_shape, data_type, 1, fixed_point_position); + pointwise_weights = create_tensor(pointwise_weights_shape, data_type, 1, fixed_point_position); + biases = create_tensor(biases_shape, data_type, 1, fixed_point_position); + dst = create_tensor(dst_shape, data_type, 1, fixed_point_position); + + // Create and configure function + depth_sep_conv_layer.configure(&src, &depthwise_weights, &depthwise_out, &pointwise_weights, &biases, &dst, pad_stride_depthwise_info, pad_stride_pointwise_info); + + // Allocate tensors + src.allocator()->allocate(); + depthwise_weights.allocator()->allocate(); + depthwise_out.allocator()->allocate(); + pointwise_weights.allocator()->allocate(); + biases.allocator()->allocate(); + dst.allocator()->allocate(); + + // Fill tensors + library->fill_tensor_uniform(Accessor(src), 0); + library->fill_tensor_uniform(Accessor(depthwise_weights), 1); + library->fill_tensor_uniform(Accessor(pointwise_weights), 2); + library->fill_tensor_uniform(Accessor(biases), 3); + } + + void run() + { + depth_sep_conv_layer.run(); + } + + void teardown() + { + src.allocator()->free(); + depthwise_weights.allocator()->free(); + depthwise_out.allocator()->free(); + pointwise_weights.allocator()->free(); + biases.allocator()->free(); + dst.allocator()->free(); + } + +private: + TensorType src{}; + TensorType depthwise_weights{}; + TensorType depthwise_out{}; + TensorType pointwise_weights{}; + TensorType biases{}; + TensorType dst{}; + Function depth_sep_conv_layer{}; +}; +} // namespace test +} // namespace arm_compute +#endif /* ARM_COMPUTE_TEST_DEPTHWISESEPARABLECONVOLUTIONLAYERFIXTURE */ -- cgit v1.2.1