diff options
author | Isabella Gottardi <isabella.gottardi@arm.com> | 2018-02-06 14:52:43 +0000 |
---|---|---|
committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:47:18 +0000 |
commit | f07d28d9ee8ae73a93fe433f72855b6dcf58ad90 (patch) | |
tree | 6ad19c89540f36e1ba5c6af7ff061bee773c43d6 | |
parent | 21f67d6763c82d78278f6bca6c6f9e42bb5ee1b9 (diff) | |
download | ComputeLibrary-f07d28d9ee8ae73a93fe433f72855b6dcf58ad90.tar.gz |
COMPMID-845: Create a ConvolutionLayer for CL
Change-Id: Ifcc406d2d0a99c911d6b6c875657b0e0028255d5
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/119148
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
-rw-r--r-- | arm_compute/core/Types.h | 8 | ||||
-rw-r--r-- | arm_compute/runtime/CL/CLFunctions.h | 1 | ||||
-rw-r--r-- | arm_compute/runtime/CL/functions/CLConvolutionLayer.h | 132 | ||||
-rw-r--r-- | arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h | 153 | ||||
-rw-r--r-- | src/runtime/CL/functions/CLConvolutionLayer.cpp | 332 | ||||
-rw-r--r-- | src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp | 353 | ||||
-rw-r--r-- | tests/validation/CL/ConvolutionLayer.cpp | 134 | ||||
-rw-r--r-- | utils/TypePrinter.h | 66 |
8 files changed, 766 insertions, 413 deletions
diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h index 417369cd9b..24c73ca7c1 100644 --- a/arm_compute/core/Types.h +++ b/arm_compute/core/Types.h @@ -1058,5 +1058,13 @@ struct IOFormatInfo std::string row_delim; bool align_columns; }; + +/** Available ConvolutionMethod*/ +enum class ConvolutionMethod +{ + GEMM, /**< Convolution using GEMM */ + DIRECT, /**< Direct convolution */ + WINOGRAD /**< Convolution using Winograd */ +}; } #endif /* __ARM_COMPUTE_TYPES_H__ */ diff --git a/arm_compute/runtime/CL/CLFunctions.h b/arm_compute/runtime/CL/CLFunctions.h index 630b9535d9..a5bbc41a17 100644 --- a/arm_compute/runtime/CL/CLFunctions.h +++ b/arm_compute/runtime/CL/CLFunctions.h @@ -60,6 +60,7 @@ #include "arm_compute/runtime/CL/functions/CLFloor.h" #include "arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h" #include "arm_compute/runtime/CL/functions/CLGEMM.h" +#include "arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h" #include "arm_compute/runtime/CL/functions/CLGEMMInterleave4x4.h" #include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h" #include "arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h" diff --git a/arm_compute/runtime/CL/functions/CLConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLConvolutionLayer.h index f6672cef1d..53d59c3176 100644 --- a/arm_compute/runtime/CL/functions/CLConvolutionLayer.h +++ b/arm_compute/runtime/CL/functions/CLConvolutionLayer.h @@ -26,71 +26,18 @@ #include "arm_compute/runtime/IFunction.h" -#include "arm_compute/core/CL/kernels/CLCol2ImKernel.h" -#include "arm_compute/core/CL/kernels/CLFillBorderKernel.h" -#include "arm_compute/core/CL/kernels/CLGEMMInterleave4x4Kernel.h" -#include "arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h" -#include "arm_compute/core/CL/kernels/CLGEMMTranspose1xWKernel.h" -#include "arm_compute/core/CL/kernels/CLIm2ColKernel.h" -#include "arm_compute/core/CL/kernels/CLWeightsReshapeKernel.h" -#include "arm_compute/core/Types.h" -#include "arm_compute/runtime/CL/CLMemoryGroup.h" -#include "arm_compute/runtime/CL/CLTensor.h" -#include "arm_compute/runtime/CL/functions/CLGEMM.h" -#include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h" -#include "arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h" +#include "arm_compute/runtime/CL/functions/CLDirectConvolutionLayer.h" +#include "arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h" #include "arm_compute/runtime/IMemoryManager.h" #include <memory> namespace arm_compute { -class ICLTensor; - -/** Function to reshape and transpose the weights. This function calls the following kernels: - * -# @ref CLWeightsReshapeKernel - * -# @ref CLGEMMTranspose1xWKernel - */ -class CLConvolutionLayerReshapeWeights : public IFunction -{ -public: - /** Constructor */ - CLConvolutionLayerReshapeWeights(std::shared_ptr<IMemoryManager> memory_manager = nullptr); - /** Set the input and output tensors. - * - * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. - * Data type supported: QS8/QASYMM8/QS16/F16/F32. - * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p weights. - * @param[out] output Destination tensor. Data types supported: Same as @p weights. - * @param[in] transpose1xW True if the weights are to undergo a 1xW transposition after reshaping (in case of GEMM operation), false otherwise. - * Data types supported: Same as @p weights. - */ - void configure(const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, bool transpose1xW); - // Inherited methods overridden: - void run() override; - -private: - CLMemoryGroup _memory_group; - CLWeightsReshapeKernel _weights_reshape_kernel; - CLGEMMTranspose1xWKernel _weights_transposed_kernel; - CLTensor _weights_reshaped; - bool _transpose1xW; -}; - /** Basic function to compute the convolution layer. This function calls the following OpenCL kernels/functions: * - * Note: weights already reshaped for quantized asymmetric is not supported - * - * -# @ref CLIm2ColKernel - * -# @ref CLGEMMLowpMatrixMultiplyCore (if quantized asymmetric) - * -# @ref CLGEMMLowpQuantizeDownInt32ToUint8Scale (if quantized asymmetric) - * -# @ref CLCol2ImKernel - * - * if the weights are already reshaped: - * -# @ref CLGEMMInterleave4x4Kernel - * -# @ref CLGEMMMatrixMultiplyKernel - * else - * -# @ref CLGEMM + * -# @ref CLGEMMConvolutionLayer + * -# @ref CLDirectConvolutionLayer */ class CLConvolutionLayer : public IFunction { @@ -108,46 +55,49 @@ public: * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. * Data types supported: Same as @p input. * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. - * @param[in] weights_info Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel. If this is not part of the fully connected layer the weights - * tensor has also been transposed with CLGEMMTranspose1xWKernel. Data type supported: Same as @p input. + * @param[in] weights_info Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel. Data type supported: Same as @p input. + */ + void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info = WeightsInfo()); + /** Static function to check if given info will lead to a valid configuration of @ref CLConvolutionLayer + * + * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], + * while every optional dimension from 4 and above represent a batch of inputs. + * Data types supported: QS8/QASYMM8/QS16/F16/F32. + * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input. + * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported:Same as @p input. + * @param[in] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. + * Data types supported: Same as @p input. + * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. + * @param[in] weights_info Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel. Data type supported: Same as @p input. + * + * @return a status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, + const WeightsInfo &weights_info = WeightsInfo()); + /** Static function to check if given info will return the convolution called by @ref CLConvolutionLayer + * + * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], + * while every optional dimension from 4 and above represent a batch of inputs. + * Data types supported: QS8/QASYMM8/QS16/F16/F32. + * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported:Same as @p input. + * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported:Same as @p input. + * @param[in] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. + * Data types supported: Same as @p input. + * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. + * @param[in] weights_info Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel. Data type supported: Same as @p input. + * @param[in] gpu_target Specifies the @p GPUTarget. + * + * @return a status */ - void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info = WeightsInfo()); + static ConvolutionMethod get_convolution_method(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, + const WeightsInfo &weights_info, const GPUTarget gpu_target); // Inherited methods overridden: void run() override; private: - /** Configures the appropriate matrix multiply routine - * - * @param input Input tensor. Data types supported: QS8/QASYMM8/QS16/F16/F32. - * @param weights Weights tensor. Data type supported: Same as @p input. - * @param output Output tensor. Data types supported: Same as @p input, - * except for input of QASYMM8 type where output should be of S32 type. - * @param is_interleaved_transposed Flag that signals if matrix is interleaved transposed - */ - void configure_mm(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output, bool is_interleaved_transposed, bool are_weights_reshaped); - -private: - CLMemoryGroup _memory_group; - CLConvolutionLayerReshapeWeights _reshape_weights; - CLIm2ColKernel _im2col_kernel; - CLGEMMInterleave4x4Kernel _interleave_kernel; - CLGEMMMatrixMultiplyKernel _mm_kernel; - CLGEMM _mm_gemm; - CLGEMMLowpMatrixMultiplyCore _mm_gemmlowp; - CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint _gemmlowp_output_stage; - CLCol2ImKernel _col2im_kernel; - - CLTensor _im2col_output; - CLTensor _interleave_output; - CLTensor _weights_reshaped; - CLTensor _weights_transposed; - CLTensor _gemm_output; - CLTensor _tmp_output; - - bool _are_weights_reshaped; - bool _is_quantized; - bool _is_interleaved_transposed; + std::shared_ptr<IMemoryManager> _memory_manager; + std::unique_ptr<IFunction> _function; /**< Function to run */ }; } #endif /* __ARM_COMPUTE_CLCONVOLUTIONLAYER_H__ */ diff --git a/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h b/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h new file mode 100644 index 0000000000..7126688f8b --- /dev/null +++ b/arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h @@ -0,0 +1,153 @@ +/* + * Copyright (c) 2017-2018 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. + */ +#ifndef __ARM_COMPUTE_CLGEMMCONVOLUTIONLAYER_H__ +#define __ARM_COMPUTE_CLGEMMCONVOLUTIONLAYER_H__ + +#include "arm_compute/runtime/IFunction.h" + +#include "arm_compute/core/CL/kernels/CLCol2ImKernel.h" +#include "arm_compute/core/CL/kernels/CLFillBorderKernel.h" +#include "arm_compute/core/CL/kernels/CLGEMMInterleave4x4Kernel.h" +#include "arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h" +#include "arm_compute/core/CL/kernels/CLGEMMTranspose1xWKernel.h" +#include "arm_compute/core/CL/kernels/CLIm2ColKernel.h" +#include "arm_compute/core/CL/kernels/CLWeightsReshapeKernel.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/runtime/CL/CLMemoryGroup.h" +#include "arm_compute/runtime/CL/CLTensor.h" +#include "arm_compute/runtime/CL/functions/CLGEMM.h" +#include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h" +#include "arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h" +#include "arm_compute/runtime/IMemoryManager.h" + +#include <memory> + +namespace arm_compute +{ +class ICLTensor; + +/** Function to reshape and transpose the weights. This function calls the following kernels: + * -# @ref CLWeightsReshapeKernel + * -# @ref CLGEMMTranspose1xWKernel + */ +class CLConvolutionLayerReshapeWeights : public IFunction +{ +public: + /** Constructor */ + CLConvolutionLayerReshapeWeights(std::shared_ptr<IMemoryManager> memory_manager = nullptr); + /** Set the input and output tensors. + * + * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. + * Data type supported: QS8/QASYMM8/QS16/F16/F32. + * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p weights. + * @param[out] output Destination tensor. Data types supported: Same as @p weights. + * @param[in] transpose1xW True if the weights are to undergo a 1xW transposition after reshaping (in case of GEMM operation), false otherwise. + * Data types supported: Same as @p weights. + */ + void configure(const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, bool transpose1xW); + // Inherited methods overridden: + void run() override; + +private: + CLMemoryGroup _memory_group; + CLWeightsReshapeKernel _weights_reshape_kernel; + CLGEMMTranspose1xWKernel _weights_transposed_kernel; + CLTensor _weights_reshaped; + bool _transpose1xW; +}; + +/** Basic function to compute the convolution layer. This function calls the following OpenCL kernels/functions: + * + * Note: weights already reshaped for quantized asymmetric is not supported + * + * -# @ref CLIm2ColKernel + * -# @ref CLGEMMLowpMatrixMultiplyCore (if quantized asymmetric) + * -# @ref CLGEMMLowpQuantizeDownInt32ToUint8Scale (if quantized asymmetric) + * -# @ref CLCol2ImKernel + * + * if the weights are already reshaped: + * -# @ref CLGEMMInterleave4x4Kernel + * -# @ref CLGEMMMatrixMultiplyKernel + * else + * -# @ref CLGEMM + */ +class CLGEMMConvolutionLayer : public IFunction +{ +public: + /** Default constructor */ + CLGEMMConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr); + /** Set the input and output tensors. + * + * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], + * while every optional dimension from 4 and above represent a batch of inputs. + * Data types supported: QS8/QASYMM8/QS16/F16/F32. + * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. Data type supported: Same as @p input. + * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. + * Data type supported: Should match @p input data type, except for input of QASYMM8 type where biases should be of S32 type. + * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. + * Data types supported: Same as @p input. + * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. + * @param[in] weights_info Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel. If this is not part of the fully connected layer the weights + * tensor has also been transposed with CLGEMMTranspose1xWKernel. Data type supported: Same as @p input. + */ + void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info = WeightsInfo()); + + // Inherited methods overridden: + void run() override; + +private: + /** Configures the appropriate matrix multiply routine + * + * @param input Input tensor. Data types supported: QS8/QASYMM8/QS16/F16/F32. + * @param weights Weights tensor. Data type supported: Same as @p input. + * @param output Output tensor. Data types supported: Same as @p input, + * except for input of QASYMM8 type where output should be of S32 type. + * @param is_interleaved_transposed Flag that signals if matrix is interleaved transposed + */ + void configure_mm(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output, bool is_interleaved_transposed, bool are_weights_reshaped); + +private: + CLMemoryGroup _memory_group; + CLConvolutionLayerReshapeWeights _reshape_weights; + CLIm2ColKernel _im2col_kernel; + CLGEMMInterleave4x4Kernel _interleave_kernel; + CLGEMMMatrixMultiplyKernel _mm_kernel; + CLGEMM _mm_gemm; + CLGEMMLowpMatrixMultiplyCore _mm_gemmlowp; + CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint _gemmlowp_output_stage; + CLCol2ImKernel _col2im_kernel; + + CLTensor _im2col_output; + CLTensor _interleave_output; + CLTensor _weights_reshaped; + CLTensor _weights_transposed; + CLTensor _gemm_output; + CLTensor _tmp_output; + + bool _are_weights_reshaped; + bool _is_quantized; + bool _is_interleaved_transposed; +}; +} +#endif /* __ARM_COMPUTE_CLGEMMCONVOLUTIONLAYER_H__ */ diff --git a/src/runtime/CL/functions/CLConvolutionLayer.cpp b/src/runtime/CL/functions/CLConvolutionLayer.cpp index d1533b6f24..c430174fe7 100644 --- a/src/runtime/CL/functions/CLConvolutionLayer.cpp +++ b/src/runtime/CL/functions/CLConvolutionLayer.cpp @@ -24,10 +24,8 @@ #include "arm_compute/runtime/CL/functions/CLConvolutionLayer.h" #include "arm_compute/core/PixelValue.h" -#include "arm_compute/core/Size2D.h" #include "arm_compute/core/Utils.h" #include "arm_compute/core/Validate.h" -#include "arm_compute/core/utils/quantization/AsymmHelpers.h" #include "arm_compute/runtime/CL/CLScheduler.h" #include <cmath> @@ -36,315 +34,87 @@ using namespace arm_compute; -CLConvolutionLayerReshapeWeights::CLConvolutionLayerReshapeWeights(std::shared_ptr<IMemoryManager> memory_manager) - : _memory_group(std::move(memory_manager)), _weights_reshape_kernel(), _weights_transposed_kernel(), _weights_reshaped(), _transpose1xW(false) -{ -} - -void CLConvolutionLayerReshapeWeights::configure(const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, bool transpose1xW) -{ - ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4); - - if(biases != nullptr) - { - ARM_COMPUTE_ERROR_ON(is_data_type_quantized_asymmetric(weights->info()->data_type())); - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases); - ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != weights->info()->dimension(3)); - ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1); - } - - const bool append_biases = (biases != nullptr) && !is_data_type_quantized_asymmetric(weights->info()->data_type()); - const unsigned bias_element = (append_biases) ? 1 : 0; - const ICLTensor *biases_to_use = (append_biases) ? biases : nullptr; - - _transpose1xW = transpose1xW; - - if(transpose1xW) - { - // Create tensor to store the reshaped weights - const unsigned int mat_weights_cols = weights->info()->dimension(3); - const unsigned int mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + bias_element; - TensorShape shape_wr(mat_weights_cols, mat_weights_rows); - const DataType dt = weights->info()->data_type(); - const int fixed_point_position = weights->info()->fixed_point_position(); - TensorInfo info_wr(shape_wr, 1, dt, fixed_point_position); - - _weights_reshaped.allocator()->init(info_wr); - _memory_group.manage(&_weights_reshaped); - _weights_reshape_kernel.configure(weights, biases_to_use, &_weights_reshaped); - _weights_transposed_kernel.configure(&_weights_reshaped, output); - _weights_reshaped.allocator()->allocate(); - } - else - { - _weights_reshape_kernel.configure(weights, biases_to_use, output); - } - - output->info()->set_quantization_info(weights->info()->quantization_info()); -} - -void CLConvolutionLayerReshapeWeights::run() -{ - _memory_group.acquire(); - - CLScheduler::get().enqueue(_weights_reshape_kernel); - if(_transpose1xW) - { - CLScheduler::get().enqueue(_weights_transposed_kernel); - } - - _memory_group.release(); -} - CLConvolutionLayer::CLConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager) - : _memory_group(memory_manager), _reshape_weights(), _im2col_kernel(), _interleave_kernel(), _mm_kernel(), _mm_gemm(memory_manager), _mm_gemmlowp(memory_manager), _gemmlowp_output_stage(), - _col2im_kernel(), _im2col_output(), _interleave_output(), _weights_reshaped(), _weights_transposed(), _gemm_output(), _tmp_output(), _are_weights_reshaped(false), _is_quantized(false), - _is_interleaved_transposed(false) + : _memory_manager(std::move(memory_manager)), _function() { } -void CLConvolutionLayer::configure_mm(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output, bool is_interleaved_transposed, bool are_weights_reshaped) +void CLConvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info) { - if(_is_quantized) - { - if(are_weights_reshaped) - { - ARM_COMPUTE_ERROR("Weights already reshaped are not suppported with gemmlowp"); - } - else - { - // Since we need negative offsets for computing convolution, we need to change QuantizationInfo() - // Extract and negate input and weights offset - const QuantizationInfo input_quantization_info = input->info()->quantization_info(); - const QuantizationInfo weights_quantization_info = weights->info()->quantization_info(); - - input->info()->set_quantization_info(QuantizationInfo(input_quantization_info.scale, -input_quantization_info.offset)); - weights->info()->set_quantization_info(QuantizationInfo(weights_quantization_info.scale, -weights_quantization_info.offset)); - - _mm_gemmlowp.configure(input, weights, output, GEMMInfo(false, false, true /* Reshape weights only for the first run*/)); + ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); + ARM_COMPUTE_ERROR_THROW_ON(CLConvolutionLayer::validate(input->info(), weights->info(), ((biases != nullptr) ? biases->info() : nullptr), output->info(), conv_info, weights_info)); - // Revert back QuantizatioInfo as input and weights could be used in other convolution layers - input->info()->set_quantization_info(input_quantization_info); - weights->info()->set_quantization_info(weights_quantization_info); - } - } - else + switch(CLConvolutionLayer::get_convolution_method(input->info(), weights->info(), ((biases != nullptr) ? biases->info() : nullptr), output->info(), conv_info, + weights_info, CLScheduler::get().target())) { - if(are_weights_reshaped) + case ConvolutionMethod::DIRECT: { - // Configure matrix multiply kernel - _mm_kernel.configure(input, weights, output, 1.f, is_interleaved_transposed); + auto f = arm_compute::support::cpp14::make_unique<CLDirectConvolutionLayer>(); + f->configure(input, weights, biases, output, conv_info); + _function = std::move(f); + break; } - else + case ConvolutionMethod::GEMM: { - // Configure matrix multiply function - _mm_gemm.configure(input, weights, nullptr, output, 1.0f, 0.0f, GEMMInfo(false, false, true /* Reshape weights only for the first run*/)); + auto f = arm_compute::support::cpp14::make_unique<CLGEMMConvolutionLayer>(_memory_manager); + f->configure(input, weights, biases, output, conv_info, weights_info); + _function = std::move(f); + break; } + default: + ARM_COMPUTE_ERROR("Not supported."); + break; } } -void CLConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info) +Status CLConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, + const WeightsInfo &weights_info) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32); - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, weights); - ARM_COMPUTE_ERROR_ON(weights_info.are_reshaped() && CLScheduler::get().target() == GPUTarget::BIFROST); - ARM_COMPUTE_ERROR_ON(!weights_info.are_reshaped() && weights->info()->dimension(2) != input->info()->dimension(2)); - ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4); - ARM_COMPUTE_ERROR_ON(weights_info.are_reshaped() && is_data_type_quantized_asymmetric(input->info()->data_type())); + ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); - _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type()); + //Configure if the parameters match the direct convolution or the gemm-based + const GPUTarget gpu_target = CLScheduler::get().target(); - if(biases != nullptr) + switch(CLConvolutionLayer::get_convolution_method(input, weights, biases, output, conv_info, weights_info, gpu_target)) { - if(_is_quantized) + case ConvolutionMethod::DIRECT: { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32); + // Validate direct convolution layer + CLDirectConvolutionLayerKernel::validate(input, weights, biases, output, conv_info, gpu_target); + break; } - else + case ConvolutionMethod::GEMM: { - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); + // Validate gemm-based convolution layer + /* TODO COMPMID-754: Add validation methods for CLGEMMConvolutionLayer + CLGEMMConvolutionLayerKernel::validate(input, weights, biases, output, conv_info, weights_info); */ + break; } - ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases); - ARM_COMPUTE_ERROR_ON(!weights_info.are_reshaped() && biases->info()->dimension(0) != weights->info()->dimension(3)); - ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1); + default: + ARM_COMPUTE_ERROR("Not supported."); + break; } - const DataType dt = input->info()->data_type(); - - // Set the GPU target for matrix multiply and im2col and col2im - _mm_kernel.set_target(CLScheduler::get().target()); - _im2col_kernel.set_target(CLScheduler::get().target()); - _col2im_kernel.set_target(CLScheduler::get().target()); - - const bool append_bias = (biases != nullptr) && (!_is_quantized); - _are_weights_reshaped = weights_info.are_reshaped(); - - const unsigned bias_element = (append_bias) ? 1 : 0; - const ICLTensor *biases_to_use = (append_bias) ? biases : nullptr; - - // Get parameters from conv_info - unsigned int stride_x = 0; - unsigned int stride_y = 0; - std::tie(stride_x, stride_y) = conv_info.stride(); - - // Get convolved dimensions - unsigned int conv_w = 0; - unsigned int conv_h = 0; - - const unsigned int kernel_width = (_are_weights_reshaped) ? weights_info.kernel_size().first : weights->info()->dimension(0); - const unsigned int kernel_height = (_are_weights_reshaped) ? weights_info.kernel_size().second : weights->info()->dimension(1); - std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), kernel_width, kernel_height, - conv_info); - - // Check if its a "fully connected" convolution - const bool is_fully_connected_convolution = ((conv_w == 1) && (conv_h == 1)); - _is_interleaved_transposed = (!is_fully_connected_convolution) && (!_is_quantized) && (_are_weights_reshaped); - - unsigned int mat_weights_cols = weights->info()->dimension(3); - unsigned int mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + bias_element; - - // Reshape weights if needed - if(_are_weights_reshaped) - { - if(is_fully_connected_convolution || _is_quantized) - { - mat_weights_cols = weights->info()->dimension(0); - mat_weights_rows = weights->info()->dimension(1); - } - else - { - mat_weights_cols = weights_info.num_kernels(); - const unsigned int quarter_reshaped_cols = weights->info()->dimension(0) / 4; - mat_weights_rows = quarter_reshaped_cols + bias_element; - } - } - else - { - // _weights_reshaped will be auto configured in the kernel. - // Just append biases and do not transpose 1xW as it will be reshaped in CLGEMM - _reshape_weights.configure(weights, biases_to_use, &_weights_reshaped, false); - - weights = &_weights_reshaped; - } - - // Create tensor to store im2col reshaped inputs - const unsigned int mat_input_cols = mat_weights_rows; - const unsigned int mat_input_rows = conv_w * conv_h; - TensorShape shape_im2col = input->info()->tensor_shape(); - shape_im2col.set(0, mat_input_cols); - shape_im2col.set(1, mat_input_rows); - shape_im2col.set(2, 1); - // FIXME: input->clone() doesn't work with subtensors for grouped convolutions. - TensorInfo im2col_reshaped_info(shape_im2col, 1, dt, input->info()->fixed_point_position()); - im2col_reshaped_info.set_quantization_info(input->info()->quantization_info()); - _im2col_output.allocator()->init(im2col_reshaped_info); - _memory_group.manage(&_im2col_output); - - // Create GEMM output tensor - TensorShape shape_gemm = _im2col_output.info()->tensor_shape(); - shape_gemm.set(0, mat_weights_cols); - shape_gemm.set(1, mat_input_rows); - const DataType gemm_data_type = _is_quantized ? DataType::S32 : dt; - // GEMM output should be S32 for acquiring raw integer accumulator without quantized postprocessing for quantized asymmetric input. - // FIXME: input->clone() doesn't work with subtensors for grouped convolutions. - TensorInfo info_gemm(shape_gemm, 1, gemm_data_type, input->info()->fixed_point_position()); - info_gemm.set_quantization_info(output->info()->quantization_info()); - _gemm_output.allocator()->init(info_gemm); - _memory_group.manage(&_gemm_output); - - // Configure im2col - _im2col_kernel.configure(input, &_im2col_output, Size2D(kernel_width, kernel_height), conv_info, append_bias); - - // Configure matrix multiply - if(_is_interleaved_transposed) - { - // Configure GEMMInterleave4x4. _input_interleaved_reshaped will be auto configured in the kernel - _memory_group.manage(&_interleave_output); - _interleave_kernel.configure(&_im2col_output, &_interleave_output); - - // Configure GEMM - configure_mm(&_interleave_output, weights, &_gemm_output, true, _are_weights_reshaped); - _interleave_output.allocator()->allocate(); - } - else - { - configure_mm(&_im2col_output, weights, &_gemm_output, false, _are_weights_reshaped); - } - _im2col_output.allocator()->allocate(); - - // Configure output stage for quantized case - if(_is_quantized) - { - float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output->info()->quantization_info().scale; - int output_multiplier, output_shift; - quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift); - _memory_group.manage(&_tmp_output); - _gemmlowp_output_stage.configure(&_gemm_output, biases, &_tmp_output, output_multiplier, output_shift, output->info()->quantization_info().offset); - } - - // Configure Col2Im - _col2im_kernel.configure(_is_quantized ? &_tmp_output : &_gemm_output, output, std::make_pair(conv_w, conv_h)); - if(_is_quantized) - { - _tmp_output.allocator()->allocate(); - } - _gemm_output.allocator()->allocate(); + return Status{}; +} - ARM_COMPUTE_ERROR_ON_MSG((output->info()->dimension(0) != conv_w) || (output->info()->dimension(1) != conv_h), "Output shape does not match the expected one"); +ConvolutionMethod CLConvolutionLayer::get_convolution_method(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, + const WeightsInfo &weights_info, const GPUTarget gpu_target) +{ + ARM_COMPUTE_UNUSED(input); + ARM_COMPUTE_UNUSED(biases); + ARM_COMPUTE_UNUSED(output); + ARM_COMPUTE_UNUSED(conv_info); + ARM_COMPUTE_UNUSED(weights_info); - // Allocate intermediate tensor - if(!_are_weights_reshaped) + if((gpu_target == GPUTarget::BIFROST) && (weights->dimension(0) == 5) && (weights->dimension(1) == 5)) { - _weights_reshaped.allocator()->allocate(); + return ConvolutionMethod::DIRECT; } + return ConvolutionMethod::GEMM; } void CLConvolutionLayer::run() { - // Run weights reshaping (Runs once for every configure) - if(!_are_weights_reshaped) - { - _are_weights_reshaped = true; - _reshape_weights.run(); - } - - _memory_group.acquire(); - - // Run im2col - CLScheduler::get().enqueue(_im2col_kernel); - - // Note: _is_interleaved_transposed is true only if the weights passed to the function have been passed already reshaped - // and if we do not have QASYMM8 data type. If this flag is true, we need to run the - // gemm kernel instead of gemm function - if(_is_interleaved_transposed) - { - // Run interleave4x4 kernel - CLScheduler::get().enqueue(_interleave_kernel); - - // Run matrix multiply kernel - CLScheduler::get().enqueue(_mm_kernel); - } - else - { - // Runs CLGEMM or CLGEMMLowpMatrixMultiplyCore functions - if(_is_quantized) - { - // Run gemmlowp - _mm_gemmlowp.run(); - - // Run output stage - _gemmlowp_output_stage.run(); - } - else - { - // Run gemm - _mm_gemm.run(); - } - } - - // Reshape output matrix - CLScheduler::get().enqueue(_col2im_kernel, false); - - _memory_group.release(); + _function->run(); } diff --git a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp new file mode 100644 index 0000000000..c4cfe1e24c --- /dev/null +++ b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp @@ -0,0 +1,353 @@ +/* + * Copyright (c) 2017-2018 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/runtime/CL/functions/CLGEMMConvolutionLayer.h" + +#include "arm_compute/core/PixelValue.h" +#include "arm_compute/core/Size2D.h" +#include "arm_compute/core/Utils.h" +#include "arm_compute/core/Validate.h" +#include "arm_compute/core/utils/quantization/AsymmHelpers.h" +#include "arm_compute/runtime/CL/CLScheduler.h" + +#include <cmath> +#include <memory> +#include <tuple> + +using namespace arm_compute; + +CLConvolutionLayerReshapeWeights::CLConvolutionLayerReshapeWeights(std::shared_ptr<IMemoryManager> memory_manager) + : _memory_group(std::move(memory_manager)), _weights_reshape_kernel(), _weights_transposed_kernel(), _weights_reshaped(), _transpose1xW(false) +{ +} + +void CLConvolutionLayerReshapeWeights::configure(const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, bool transpose1xW) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(weights, output); + ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4); + + if(biases != nullptr) + { + ARM_COMPUTE_ERROR_ON(is_data_type_quantized_asymmetric(weights->info()->data_type())); + ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases); + ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != weights->info()->dimension(3)); + ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1); + } + + const bool append_biases = (biases != nullptr) && !is_data_type_quantized_asymmetric(weights->info()->data_type()); + const unsigned bias_element = (append_biases) ? 1 : 0; + const ICLTensor *biases_to_use = (append_biases) ? biases : nullptr; + + _transpose1xW = transpose1xW; + + if(transpose1xW) + { + // Create tensor to store the reshaped weights + const unsigned int mat_weights_cols = weights->info()->dimension(3); + const unsigned int mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + bias_element; + TensorShape shape_wr(mat_weights_cols, mat_weights_rows); + const DataType dt = weights->info()->data_type(); + const int fixed_point_position = weights->info()->fixed_point_position(); + TensorInfo info_wr(shape_wr, 1, dt, fixed_point_position); + + _weights_reshaped.allocator()->init(info_wr); + _memory_group.manage(&_weights_reshaped); + _weights_reshape_kernel.configure(weights, biases_to_use, &_weights_reshaped); + _weights_transposed_kernel.configure(&_weights_reshaped, output); + _weights_reshaped.allocator()->allocate(); + } + else + { + _weights_reshape_kernel.configure(weights, biases_to_use, output); + } + + output->info()->set_quantization_info(weights->info()->quantization_info()); +} + +void CLConvolutionLayerReshapeWeights::run() +{ + _memory_group.acquire(); + + CLScheduler::get().enqueue(_weights_reshape_kernel); + if(_transpose1xW) + { + CLScheduler::get().enqueue(_weights_transposed_kernel); + } + + _memory_group.release(); +} + +CLGEMMConvolutionLayer::CLGEMMConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager) + : _memory_group(memory_manager), _reshape_weights(), _im2col_kernel(), _interleave_kernel(), _mm_kernel(), _mm_gemm(memory_manager), _mm_gemmlowp(memory_manager), _gemmlowp_output_stage(), + _col2im_kernel(), _im2col_output(), _interleave_output(), _weights_reshaped(), _weights_transposed(), _gemm_output(), _tmp_output(), _are_weights_reshaped(false), _is_quantized(false), + _is_interleaved_transposed(false) +{ +} + +void CLGEMMConvolutionLayer::configure_mm(const ICLTensor *input, const ICLTensor *weights, ICLTensor *output, bool is_interleaved_transposed, bool are_weights_reshaped) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights); + if(_is_quantized) + { + if(are_weights_reshaped) + { + ARM_COMPUTE_ERROR("Weights already reshaped are not suppported with gemmlowp"); + } + else + { + // Since we need negative offsets for computing convolution, we need to change QuantizationInfo() + // Extract and negate input and weights offset + const QuantizationInfo input_quantization_info = input->info()->quantization_info(); + const QuantizationInfo weights_quantization_info = weights->info()->quantization_info(); + + input->info()->set_quantization_info(QuantizationInfo(input_quantization_info.scale, -input_quantization_info.offset)); + weights->info()->set_quantization_info(QuantizationInfo(weights_quantization_info.scale, -weights_quantization_info.offset)); + + _mm_gemmlowp.configure(input, weights, output, GEMMInfo(false, false, true /* Reshape weights only for the first run*/)); + + // Revert back QuantizatioInfo as input and weights could be used in other convolution layers + input->info()->set_quantization_info(input_quantization_info); + weights->info()->set_quantization_info(weights_quantization_info); + } + } + else + { + if(are_weights_reshaped) + { + // Configure matrix multiply kernel + _mm_kernel.configure(input, weights, output, 1.f, is_interleaved_transposed); + } + else + { + // Configure matrix multiply function + _mm_gemm.configure(input, weights, nullptr, output, 1.0f, 0.0f, GEMMInfo(false, false, true /* Reshape weights only for the first run*/)); + } + } +} + +void CLGEMMConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); + ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, weights); + ARM_COMPUTE_ERROR_ON(weights_info.are_reshaped() && CLScheduler::get().target() == GPUTarget::BIFROST); + ARM_COMPUTE_ERROR_ON(!weights_info.are_reshaped() && weights->info()->dimension(2) != input->info()->dimension(2)); + ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4); + ARM_COMPUTE_ERROR_ON(weights_info.are_reshaped() && is_data_type_quantized_asymmetric(input->info()->data_type())); + + _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type()); + + if(biases != nullptr) + { + if(_is_quantized) + { + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32); + } + else + { + ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); + } + ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases); + ARM_COMPUTE_ERROR_ON(!weights_info.are_reshaped() && biases->info()->dimension(0) != weights->info()->dimension(3)); + ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1); + } + + const DataType dt = input->info()->data_type(); + + // Set the GPU target for matrix multiply and im2col and col2im + _mm_kernel.set_target(CLScheduler::get().target()); + _im2col_kernel.set_target(CLScheduler::get().target()); + _col2im_kernel.set_target(CLScheduler::get().target()); + + const bool append_bias = (biases != nullptr) && (!_is_quantized); + _are_weights_reshaped = weights_info.are_reshaped(); + + const unsigned bias_element = (append_bias) ? 1 : 0; + const ICLTensor *biases_to_use = (append_bias) ? biases : nullptr; + + // Get parameters from conv_info + unsigned int stride_x = 0; + unsigned int stride_y = 0; + std::tie(stride_x, stride_y) = conv_info.stride(); + + // Get convolved dimensions + unsigned int conv_w = 0; + unsigned int conv_h = 0; + + const unsigned int kernel_width = (_are_weights_reshaped) ? weights_info.kernel_size().first : weights->info()->dimension(0); + const unsigned int kernel_height = (_are_weights_reshaped) ? weights_info.kernel_size().second : weights->info()->dimension(1); + std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), kernel_width, kernel_height, + conv_info); + + // Check if its a "fully connected" convolution + const bool is_fully_connected_convolution = ((conv_w == 1) && (conv_h == 1)); + _is_interleaved_transposed = (!is_fully_connected_convolution) && (!_is_quantized) && (_are_weights_reshaped); + + unsigned int mat_weights_cols = weights->info()->dimension(3); + unsigned int mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + bias_element; + + // Reshape weights if needed + if(_are_weights_reshaped) + { + if(is_fully_connected_convolution || _is_quantized) + { + mat_weights_cols = weights->info()->dimension(0); + mat_weights_rows = weights->info()->dimension(1); + } + else + { + mat_weights_cols = weights_info.num_kernels(); + const unsigned int quarter_reshaped_cols = weights->info()->dimension(0) / 4; + mat_weights_rows = quarter_reshaped_cols + bias_element; + } + } + else + { + // _weights_reshaped will be auto configured in the kernel. + // Just append biases and do not transpose 1xW as it will be reshaped in CLGEMM + _reshape_weights.configure(weights, biases_to_use, &_weights_reshaped, false); + + weights = &_weights_reshaped; + } + + // Create tensor to store im2col reshaped inputs + const unsigned int mat_input_cols = mat_weights_rows; + const unsigned int mat_input_rows = conv_w * conv_h; + TensorShape shape_im2col = input->info()->tensor_shape(); + shape_im2col.set(0, mat_input_cols); + shape_im2col.set(1, mat_input_rows); + shape_im2col.set(2, 1); + // FIXME: input->clone() doesn't work with subtensors for grouped convolutions. + TensorInfo im2col_reshaped_info(shape_im2col, 1, dt, input->info()->fixed_point_position()); + im2col_reshaped_info.set_quantization_info(input->info()->quantization_info()); + _im2col_output.allocator()->init(im2col_reshaped_info); + _memory_group.manage(&_im2col_output); + + // Create GEMM output tensor + TensorShape shape_gemm = _im2col_output.info()->tensor_shape(); + shape_gemm.set(0, mat_weights_cols); + shape_gemm.set(1, mat_input_rows); + const DataType gemm_data_type = _is_quantized ? DataType::S32 : dt; + // GEMM output should be S32 for acquiring raw integer accumulator without quantized postprocessing for quantized asymmetric input. + // FIXME: input->clone() doesn't work with subtensors for grouped convolutions. + TensorInfo info_gemm(shape_gemm, 1, gemm_data_type, input->info()->fixed_point_position()); + info_gemm.set_quantization_info(output->info()->quantization_info()); + _gemm_output.allocator()->init(info_gemm); + _memory_group.manage(&_gemm_output); + + // Configure im2col + _im2col_kernel.configure(input, &_im2col_output, Size2D(kernel_width, kernel_height), conv_info, append_bias); + + // Configure matrix multiply + if(_is_interleaved_transposed) + { + // Configure GEMMInterleave4x4. _input_interleaved_reshaped will be auto configured in the kernel + _memory_group.manage(&_interleave_output); + _interleave_kernel.configure(&_im2col_output, &_interleave_output); + + // Configure GEMM + configure_mm(&_interleave_output, weights, &_gemm_output, true, _are_weights_reshaped); + _interleave_output.allocator()->allocate(); + } + else + { + configure_mm(&_im2col_output, weights, &_gemm_output, false, _are_weights_reshaped); + } + _im2col_output.allocator()->allocate(); + + // Configure output stage for quantized case + if(_is_quantized) + { + float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output->info()->quantization_info().scale; + int output_multiplier, output_shift; + quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift); + _memory_group.manage(&_tmp_output); + _gemmlowp_output_stage.configure(&_gemm_output, biases, &_tmp_output, output_multiplier, output_shift, output->info()->quantization_info().offset); + } + + // Configure Col2Im + _col2im_kernel.configure(_is_quantized ? &_tmp_output : &_gemm_output, output, std::make_pair(conv_w, conv_h)); + if(_is_quantized) + { + _tmp_output.allocator()->allocate(); + } + _gemm_output.allocator()->allocate(); + + ARM_COMPUTE_ERROR_ON_MSG((output->info()->dimension(0) != conv_w) || (output->info()->dimension(1) != conv_h), "Output shape does not match the expected one"); + + // Allocate intermediate tensor + if(!_are_weights_reshaped) + { + _weights_reshaped.allocator()->allocate(); + } +} + +void CLGEMMConvolutionLayer::run() +{ + // Run weights reshaping (Runs once for every configure) + if(!_are_weights_reshaped) + { + _are_weights_reshaped = true; + _reshape_weights.run(); + } + + _memory_group.acquire(); + + // Run im2col + CLScheduler::get().enqueue(_im2col_kernel); + + // Note: _is_interleaved_transposed is true only if the weights passed to the function have been passed already reshaped + // and if we do not have QASYMM8 data type. If this flag is true, we need to run the + // gemm kernel instead of gemm function + if(_is_interleaved_transposed) + { + // Run interleave4x4 kernel + CLScheduler::get().enqueue(_interleave_kernel); + + // Run matrix multiply kernel + CLScheduler::get().enqueue(_mm_kernel); + } + else + { + // Runs CLGEMM or CLGEMMLowpMatrixMultiplyCore functions + if(_is_quantized) + { + // Run gemmlowp + _mm_gemmlowp.run(); + + // Run output stage + _gemmlowp_output_stage.run(); + } + else + { + // Run gemm + _mm_gemm.run(); + } + } + + // Reshape output matrix + CLScheduler::get().enqueue(_col2im_kernel, false); + + _memory_group.release(); +} diff --git a/tests/validation/CL/ConvolutionLayer.cpp b/tests/validation/CL/ConvolutionLayer.cpp index 46cb097986..b7f9241c88 100644 --- a/tests/validation/CL/ConvolutionLayer.cpp +++ b/tests/validation/CL/ConvolutionLayer.cpp @@ -25,6 +25,7 @@ #include "arm_compute/runtime/CL/CLTensor.h" #include "arm_compute/runtime/CL/CLTensorAllocator.h" #include "arm_compute/runtime/CL/functions/CLConvolutionLayer.h" +#include "arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h" #include "tests/CL/CLAccessor.h" #include "tests/PaddingCalculator.h" #include "tests/datasets/LargeConvolutionLayerDataset.h" @@ -64,6 +65,57 @@ const auto CNNDataTypes = framework::dataset::make("DataType", TEST_SUITE(CL) TEST_SUITE(ConvolutionLayer) +DATA_TEST_CASE(ValidateConvolutionMethod, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip( + framework::dataset::make("InputInfo", { TensorInfo(TensorShape(17U, 31U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(17U, 31U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(23U, 27U, 5U, 4U), 1, DataType::F32, 0), + TensorInfo(TensorShape(3U, 3U, 2U, 1U), 1, DataType::F32, 0), + TensorInfo(TensorShape(33U, 27U, 7U, 4U), 1, DataType::F32, 0) + }), + framework::dataset::make("WeightsInfo", { TensorInfo(TensorShape(5U, 5U, 2U, 19U), 1, DataType::F32, 0), + TensorInfo(TensorShape(5U, 5U, 2U, 19U), 1, DataType::F32, 0), + TensorInfo(TensorShape(3U, 3U, 5U, 21U), 1, DataType::F32, 0), + TensorInfo(TensorShape(3U, 3U, 5U, 21U), 1, DataType::F32, 0), + TensorInfo(TensorShape(5U, 5U, 7U, 16U), 1, DataType::F16, 0) + })), + framework::dataset::make("BiasesInfo", { TensorInfo(TensorShape(19U), 1, DataType::F32, 0), + TensorInfo(TensorShape(19U), 1, DataType::F32, 0), + TensorInfo(TensorShape(21U), 1, DataType::F32, 0), + TensorInfo(TensorShape(21U), 1, DataType::F32, 0), + TensorInfo(TensorShape(16U), 1, DataType::F32, 0) + })), + framework::dataset::make("OutputInfo", { TensorInfo(TensorShape(15U, 15U, 19U), 1, DataType::F32, 0), + TensorInfo(TensorShape(15U, 15U, 19U), 1, DataType::F32, 0), + TensorInfo(TensorShape(21U, 25U, 21U, 4U), 1, DataType::F32, 0), + TensorInfo(TensorShape(11U, 25U, 21U), 1, DataType::F32, 0), + TensorInfo(TensorShape(11U, 12U, 16U, 4U), 1, DataType::F32, 0) + })), + framework::dataset::make("ConvInfo", { PadStrideInfo(1, 2, 1, 1), + PadStrideInfo(1, 2, 1, 1), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(2, 1, 0, 0), + PadStrideInfo(3, 2, 1, 0) + })), + framework::dataset::make("GpuTarget", { GPUTarget::BIFROST, + GPUTarget::MIDGARD, + GPUTarget::G70, + GPUTarget::MIDGARD, + GPUTarget::BIFROST + })), + + framework::dataset::make("Expected", { ConvolutionMethod::DIRECT, ConvolutionMethod::GEMM, ConvolutionMethod::GEMM, ConvolutionMethod::GEMM, ConvolutionMethod::DIRECT })), + input_info, weights_info, biases_info, output_info, conv_info, gpu_target, expected) +{ + ConvolutionMethod is_valid = CLConvolutionLayer::get_convolution_method(&input_info.clone()->set_is_resizable(false), + &weights_info.clone()->set_is_resizable(false), + &biases_info.clone()->set_is_resizable(false), + &output_info.clone()->set_is_resizable(false), conv_info, WeightsInfo(), gpu_target); + ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS); +} +TEST_SUITE_END() + +TEST_SUITE(GEMMConvolutionLayer) + DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::SmallConvolutionLayerDataset(), datasets::LargeConvolutionLayerDataset()), CNNDataTypes), input_shape, weights_shape, bias_shape, output_shape, info, data_type) { @@ -87,7 +139,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::da const QuantizationInfo weights_quantization_info = weights.info()->quantization_info(); // Create and configure function - CLConvolutionLayer conv; + CLGEMMConvolutionLayer conv; conv.configure(&src, &weights, &bias, &dst, info); // Validate valid region @@ -110,22 +162,22 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::da } template <typename T> -using CLConvolutionLayerFixture = ConvolutionValidationFixture<CLTensor, CLAccessor, CLConvolutionLayer, T>; +using CLGEMMConvolutionLayerFixture = ConvolutionValidationFixture<CLTensor, CLAccessor, CLGEMMConvolutionLayer, T>; TEST_SUITE(Float) TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", - DataType::F16))) +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), + framework::dataset::make("DataType", + DataType::F16))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", - DataType::F16))) +FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), + framework::dataset::make("DataType", + DataType::F16))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); @@ -133,18 +185,18 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionLayerFixture<half>, framework::Dat TEST_SUITE_END() TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", - DataType::F32))) +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), + framework::dataset::make("DataType", + DataType::F32))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_f32); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", - DataType::F32))) +FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), + framework::dataset::make("DataType", + DataType::F32))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_f32); @@ -153,25 +205,25 @@ TEST_SUITE_END() TEST_SUITE_END() template <typename T> -using CLConvolutionLayerFixedPointFixture = ConvolutionValidationFixedPointFixture<CLTensor, CLAccessor, CLConvolutionLayer, T>; +using CLGEMMConvolutionLayerFixedPointFixture = ConvolutionValidationFixedPointFixture<CLTensor, CLAccessor, CLGEMMConvolutionLayer, T>; TEST_SUITE(FixedPoint) TEST_SUITE(QS8) // We test for fixed point precision [4,6] -FIXTURE_DATA_TEST_CASE(RunTiny, CLConvolutionLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", - DataType::QS8)), - framework::dataset::make("FractionalBits", 4, 7))) +FIXTURE_DATA_TEST_CASE(RunTiny, CLGEMMConvolutionLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), + framework::dataset::make("DataType", + DataType::QS8)), + framework::dataset::make("FractionalBits", 4, 7))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fixed); } -FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", - DataType::QS8)), - framework::dataset::make("FractionalBits", 4, 7))) +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMConvolutionLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), + framework::dataset::make("DataType", + DataType::QS8)), + framework::dataset::make("FractionalBits", 4, 7))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fixed); @@ -180,7 +232,7 @@ TEST_SUITE_END() TEST_SUITE(QS16) // Testing for fixed point position [1,14) -FIXTURE_DATA_TEST_CASE(RunTiny, CLConvolutionLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyConvolutionLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunTiny, CLGEMMConvolutionLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyConvolutionLayerDataset(), framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QS16)), @@ -189,11 +241,11 @@ FIXTURE_DATA_TEST_CASE(RunTiny, CLConvolutionLayerFixedPointFixture<int16_t>, fr // Validate output validate(CLAccessor(_target), _reference, tolerance_fixed); } -FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", - DataType::QS16)), - framework::dataset::make("FractionalBits", 1, 14))) +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMConvolutionLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), + framework::dataset::make("DataType", + DataType::QS16)), + framework::dataset::make("FractionalBits", 1, 14))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fixed); @@ -202,11 +254,11 @@ TEST_SUITE_END() TEST_SUITE_END() template <typename T> -using CLConvolutionLayerQuantizedFixture = ConvolutionValidationQuantizedFixture<CLTensor, CLAccessor, CLConvolutionLayer, T>; +using CLGEMMConvolutionLayerQuantizedFixture = ConvolutionValidationQuantizedFixture<CLTensor, CLAccessor, CLGEMMConvolutionLayer, T>; TEST_SUITE(Quantized) TEST_SUITE(QASYMM8) -FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) }))) @@ -214,10 +266,10 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionLayerQuantizedFixture<uint8_t>, fr // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", DataType::QASYMM8)), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 0) }))) +FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), + framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 0) }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); diff --git a/utils/TypePrinter.h b/utils/TypePrinter.h index 52699b67de..63fba35052 100644 --- a/utils/TypePrinter.h +++ b/utils/TypePrinter.h @@ -24,6 +24,7 @@ #ifndef __ARM_COMPUTE_TEST_TYPE_PRINTER_H__ #define __ARM_COMPUTE_TEST_TYPE_PRINTER_H__ +#include "arm_compute/core/CL/CLTypes.h" #include "arm_compute/core/Dimensions.h" #include "arm_compute/core/Error.h" #include "arm_compute/core/HOGInfo.h" @@ -932,5 +933,70 @@ inline std::string to_string(const HOGInfo &type) return str.str(); } +inline ::std::ostream &operator<<(::std::ostream &os, const ConvolutionMethod &conv_method) +{ + switch(conv_method) + { + case ConvolutionMethod::GEMM: + os << "GEMM"; + break; + case ConvolutionMethod::DIRECT: + os << "DIRECT"; + break; + case ConvolutionMethod::WINOGRAD: + os << "WINOGRAD"; + break; + default: + ARM_COMPUTE_ERROR("NOT_SUPPORTED!"); + } + + return os; +} + +inline std::string to_string(const ConvolutionMethod &conv_method) +{ + std::stringstream str; + str << conv_method; + return str.str(); +} + +inline ::std::ostream &operator<<(::std::ostream &os, const GPUTarget &gpu_target) +{ + switch(gpu_target) + { + case GPUTarget::GPU_ARCH_MASK: + os << "GPU_ARCH_MASK"; + break; + case GPUTarget::MIDGARD: + os << "MIDGARD"; + break; + case GPUTarget::BIFROST: + os << "BIFROST"; + break; + case GPUTarget::T600: + os << "T600"; + break; + case GPUTarget::T700: + os << "T700"; + break; + case GPUTarget::T800: + os << "T800"; + break; + case GPUTarget::G70: + os << "G70"; + break; + default: + ARM_COMPUTE_ERROR("NOT_SUPPORTED!"); + } + + return os; +} + +inline std::string to_string(const GPUTarget &gpu_target) +{ + std::stringstream str; + str << gpu_target; + return str.str(); +} } // namespace arm_compute #endif /* __ARM_COMPUTE_TEST_TYPE_PRINTER_H__ */ |