From 6acc6add8412c6d3841a49684610fc5a6526312e Mon Sep 17 00:00:00 2001 From: Isabella Gottardi Date: Fri, 2 Feb 2018 17:19:18 +0000 Subject: COMPMID-846: Create a ConvolutionLayer for NEON Change-Id: I98bbef40bfac5b05134be4ef9fb54d14c0c9e8e8 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/118806 Tested-by: Jenkins Reviewed-by: Anthony Barbier --- arm_compute/runtime/NEON/NEFunctions.h | 1 + .../runtime/NEON/functions/NEConvolutionLayer.h | 132 +---- .../NEON/functions/NEGEMMConvolutionLayer.h | 184 ++++++ .../runtime/NEON/functions/NEWinogradLayer.h | 16 + src/runtime/NEON/functions/NEConvolutionLayer.cpp | 636 ++------------------ .../NEON/functions/NEGEMMConvolutionLayer.cpp | 652 +++++++++++++++++++++ src/runtime/NEON/functions/NEWinogradLayer.cpp | 54 +- tests/benchmark/NEON/ConvolutionLayer.cpp | 26 +- tests/validation/NEON/ConvolutionLayer.cpp | 109 ++-- 9 files changed, 1061 insertions(+), 749 deletions(-) create mode 100644 arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h create mode 100644 src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp diff --git a/arm_compute/runtime/NEON/NEFunctions.h b/arm_compute/runtime/NEON/NEFunctions.h index 077cf577e7..1531377e2e 100644 --- a/arm_compute/runtime/NEON/NEFunctions.h +++ b/arm_compute/runtime/NEON/NEFunctions.h @@ -60,6 +60,7 @@ #include "arm_compute/runtime/NEON/functions/NEFloor.h" #include "arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h" #include "arm_compute/runtime/NEON/functions/NEGEMM.h" +#include "arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h" #include "arm_compute/runtime/NEON/functions/NEGEMMInterleave4x4.h" #include "arm_compute/runtime/NEON/functions/NEGEMMLowpAssemblyMatrixMultiplyCore.h" #include "arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h" diff --git a/arm_compute/runtime/NEON/functions/NEConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEConvolutionLayer.h index f80f67d944..6ab1350b25 100644 --- a/arm_compute/runtime/NEON/functions/NEConvolutionLayer.h +++ b/arm_compute/runtime/NEON/functions/NEConvolutionLayer.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -26,79 +26,27 @@ #include "arm_compute/runtime/IFunction.h" -#include "arm_compute/core/NEON/kernels/NECol2ImKernel.h" -#include "arm_compute/core/NEON/kernels/NEFillBorderKernel.h" -#include "arm_compute/core/NEON/kernels/NEGEMMAssemblyBaseKernel.h" -#include "arm_compute/core/NEON/kernels/NEGEMMInterleave4x4Kernel.h" -#include "arm_compute/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.h" -#include "arm_compute/core/NEON/kernels/NEGEMMTranspose1xWKernel.h" -#include "arm_compute/core/NEON/kernels/NEIm2ColKernel.h" -#include "arm_compute/core/NEON/kernels/NEWeightsReshapeKernel.h" #include "arm_compute/core/Types.h" #include "arm_compute/runtime/MemoryGroup.h" -#include "arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h" -#include "arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h" -#include "arm_compute/runtime/Tensor.h" +#include "arm_compute/runtime/NEON/functions/NEDirectConvolutionLayer.h" +#include "arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h" +#include "arm_compute/runtime/NEON/functions/NEWinogradLayer.h" #include namespace arm_compute { class ITensor; -/** Function to reshape and perform 1xW transposition on the weights. This function calls the following kernels: - * -# @ref NEWeightsReshapeKernel - * -# @ref NEGEMMTranspose1xWKernel (executed in case GEMM is required for the operation) - */ -class NEConvolutionLayerReshapeWeights : public IFunction -{ -public: - /** Constructor */ - NEConvolutionLayerReshapeWeights(std::shared_ptr 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/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 ITensor *weights, const ITensor *biases, ITensor *output, bool transpose1xW); - /** Static function to check if given info will lead to a valid configuration of @ref NEConvolutionLayerReshapeWeights - * - * @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[in] 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. - * - * @return an error status - */ - static Status validate(const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose1xW); - - // Inherited methods overridden: - void run() override; - -private: - MemoryGroup _memory_group; - NEWeightsReshapeKernel _weights_reshape_kernel; - NEGEMMTranspose1xWKernel _weights_transposed_kernel; - Tensor _weights_reshaped; - bool _transpose1xW; -}; - -/** Basic function to simulate a convolution layer. This function calls the following NEON kernels: - * -# @ref NEWeightsReshapeKernel (executed only once for each configuration) - * -# @ref NEIm2ColKernel - * -# @ref NEGEMMInterleave4x4Kernel (executed only in case GEMM is required for the operation) - * -# @ref NEGEMMMatrixMultiplyKernel or @ref NEGEMMLowpMatrixMultiplyCore (if quantized asymmetric) - * -# @ref NEGEMMLowpQuantizeDownInt32ToUint8Scale (if quantized asymmetric) - * -# @ref NECol2ImKernel +/** Basic function to simulate a convolution layer. This function calls one of the following NEON functions: + * -# @ref NEGEMMConvolutionLayer (executed only in case GEMM is required for the operation) + * -# @ref NEWinogradLayer (executed only in case Winograd is required for the operation) + * -# @ref NEDirectConvolutionLayer (executed only in case Direct Convolution is required for the operation) */ class NEConvolutionLayer : public IFunction { public: /** Constructor */ - NEConvolutionLayer(const std::shared_ptr &memory_manager = nullptr); + NEConvolutionLayer(std::shared_ptr memory_manager = nullptr); /** Set the input and output tensors. * @@ -114,7 +62,7 @@ public: * @param[in] weights_info Specifies if the weights tensor has been reshaped with NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights * tensor has also been transposed with NEGEMMTranspose1xWKernel. Data type supported: Same as @p input. */ - void configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info = WeightsInfo()); + void configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *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 NEConvolutionLayer * * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], @@ -133,51 +81,31 @@ public: */ static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info = WeightsInfo()); - - // Inherited methods overridden: - void run() override; - -private: - /** Configures the appropriate matrix multiply routine + /** Static function to check if given info will return the convolution called by @ref NEConvolutionLayer * - * @param[in] input Input tensor. Data types supported: QS8/QASYMM8/QS16/F16/F32. - * @param[in] weights Weights tensor. Data type supported: Same as @p input. - * @param[out] output Output tensor. Data types supported: Same as @p input, - * except for input of QASYMM8 type where output should be of S32 type. - */ - void configure_mm(const ITensor *input, const ITensor *weights, ITensor *output); - /** Prepare the appropriate assembly optimized kernel + * @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[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 NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights + * tensor has also been transposed with NEGEMMTranspose1xWKernel. Data type supported: Same as @p input. * - * @param[in] ci CPU information - * @param[in] M M parameter of matrix multiplication - * @param[in] N N parameter of matrix multiplication - * @param[in] K K parameter of matrix multiplication + * @return the Convolution Method Hint */ - void configure_asm_mm(const struct CPUInfo &ci, int M, int N, int K); + 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 = WeightsInfo()); -private: - MemoryGroup _memory_group; - NEIm2ColKernel _input_im2col_kernel; - NEGEMMInterleave4x4Kernel _input_interleave_kernel; - NEConvolutionLayerReshapeWeights _reshape_weights; - NEGEMMMatrixMultiplyKernel _mm_kernel; - std::unique_ptr _mm_optimised_kernel; - NEGEMMLowpMatrixMultiplyCore _mm_gemmlowp; - NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint _gemmlowp_output_stage; - NECol2ImKernel _output_col2im_kernel; - - Tensor _input_im2col_reshaped; - Tensor _input_interleaved_reshaped; - Tensor _weights_reshaped; - Tensor _gemm_output; - Tensor _tmp_output; - Tensor _workspace; + // Inherited methods overridden: + void run() override; - bool _append_bias; - bool _is_fully_connected_convolution; - bool _are_weights_reshaped; - bool _is_quantized; - bool _is_interleaved_transposed; +private: + std::shared_ptr _memory_manager; + std::unique_ptr _function; /**< Function to run */ }; } #endif /* __ARM_COMPUTE_NECONVOLUTIONLAYER_H__ */ diff --git a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h new file mode 100644 index 0000000000..c3c7f825a9 --- /dev/null +++ b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h @@ -0,0 +1,184 @@ +/* + * 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_NEGEMMCONVOLUTIONLAYER_H__ +#define __ARM_COMPUTE_NEGEMMCONVOLUTIONLAYER_H__ + +#include "arm_compute/runtime/IFunction.h" + +#include "arm_compute/core/NEON/kernels/NECol2ImKernel.h" +#include "arm_compute/core/NEON/kernels/NEFillBorderKernel.h" +#include "arm_compute/core/NEON/kernels/NEGEMMAssemblyBaseKernel.h" +#include "arm_compute/core/NEON/kernels/NEGEMMInterleave4x4Kernel.h" +#include "arm_compute/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.h" +#include "arm_compute/core/NEON/kernels/NEGEMMTranspose1xWKernel.h" +#include "arm_compute/core/NEON/kernels/NEIm2ColKernel.h" +#include "arm_compute/core/NEON/kernels/NEWeightsReshapeKernel.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/runtime/MemoryGroup.h" +#include "arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h" +#include "arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h" +#include "arm_compute/runtime/Tensor.h" + +#include + +namespace arm_compute +{ +class ITensor; + +/** Function to reshape and perform 1xW transposition on the weights. This function calls the following kernels: + * -# @ref NEWeightsReshapeKernel + * -# @ref NEGEMMTranspose1xWKernel (executed in case GEMM is required for the operation) + */ +class NEConvolutionLayerReshapeWeights : public IFunction +{ +public: + /** Constructor */ + NEConvolutionLayerReshapeWeights(std::shared_ptr 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/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 ITensor *weights, const ITensor *biases, ITensor *output, bool transpose1xW); + /** Static function to check if given info will lead to a valid configuration of @ref NEConvolutionLayerReshapeWeights + * + * @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[in] 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. + * + * @return an error status + */ + static Status validate(const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose1xW); + + // Inherited methods overridden: + void run() override; + +private: + MemoryGroup _memory_group; + NEWeightsReshapeKernel _weights_reshape_kernel; + NEGEMMTranspose1xWKernel _weights_transposed_kernel; + Tensor _weights_reshaped; + bool _transpose1xW; +}; + +/** Basic function to simulate a convolution layer. This function calls the following NEON kernels: + * -# @ref NEWeightsReshapeKernel (executed only once for each configuration) + * -# @ref NEIm2ColKernel + * -# @ref NEGEMMInterleave4x4Kernel (executed only in case GEMM is required for the operation) + * -# @ref NEGEMMMatrixMultiplyKernel or @ref NEGEMMLowpMatrixMultiplyCore (if quantized asymmetric) + * -# @ref NEGEMMLowpQuantizeDownInt32ToUint8Scale (if quantized asymmetric) + * -# @ref NECol2ImKernel + */ +class NEGEMMConvolutionLayer : public IFunction +{ +public: + /** Constructor */ + NEGEMMConvolutionLayer(const std::shared_ptr &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/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 NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights + * tensor has also been transposed with NEGEMMTranspose1xWKernel. Data type supported: Same as @p input. + */ + void configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *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 NEGEMMConvolutionLayer + * + * @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[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 NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights + * tensor has also been transposed with NEGEMMTranspose1xWKernel. 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()); + + // Inherited methods overridden: + void run() override; + +private: + /** Configures the appropriate matrix multiply routine + * + * @param[in] input Input tensor. Data types supported: QS8/QASYMM8/QS16/F16/F32. + * @param[in] weights Weights tensor. Data type supported: Same as @p input. + * @param[out] output Output tensor. Data types supported: Same as @p input, + * except for input of QASYMM8 type where output should be of S32 type. + */ + void configure_mm(const ITensor *input, const ITensor *weights, ITensor *output); + /** Prepare the appropriate assembly optimized kernel + * + * @param[in] ci CPU information + * @param[in] M M parameter of matrix multiplication + * @param[in] N N parameter of matrix multiplication + * @param[in] K K parameter of matrix multiplication + */ + void configure_asm_mm(const struct CPUInfo &ci, int M, int N, int K); + +private: + MemoryGroup _memory_group; + NEIm2ColKernel _input_im2col_kernel; + NEGEMMInterleave4x4Kernel _input_interleave_kernel; + NEConvolutionLayerReshapeWeights _reshape_weights; + NEGEMMMatrixMultiplyKernel _mm_kernel; + std::unique_ptr _mm_optimised_kernel; + NEGEMMLowpMatrixMultiplyCore _mm_gemmlowp; + NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint _gemmlowp_output_stage; + NECol2ImKernel _output_col2im_kernel; + + Tensor _input_im2col_reshaped; + Tensor _input_interleaved_reshaped; + Tensor _weights_reshaped; + Tensor _gemm_output; + Tensor _tmp_output; + Tensor _workspace; + + bool _append_bias; + bool _is_fully_connected_convolution; + bool _are_weights_reshaped; + bool _is_quantized; + bool _is_interleaved_transposed; +}; +} +#endif /* __ARM_COMPUTE_NECONVOLUTIONGEMMLAYER_H__ */ diff --git a/arm_compute/runtime/NEON/functions/NEWinogradLayer.h b/arm_compute/runtime/NEON/functions/NEWinogradLayer.h index f57be697b5..a939f82854 100644 --- a/arm_compute/runtime/NEON/functions/NEWinogradLayer.h +++ b/arm_compute/runtime/NEON/functions/NEWinogradLayer.h @@ -67,6 +67,22 @@ public: // Inherited methods overridden: void run() override; + /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMConvolutionLayer + * + * @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: 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. + * Currently only 3x3 kernels are supported. + * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p weights. + * @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. Currently only unit strides are supported. + * + * @return a status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info); + /** Prevent instances of this class from being copied (As this class contains pointers) */ NEWinogradLayer(const NEWinogradLayer &) = delete; /** Prevent instances of this class from being copied (As this class contains pointers) */ diff --git a/src/runtime/NEON/functions/NEConvolutionLayer.cpp b/src/runtime/NEON/functions/NEConvolutionLayer.cpp index 335267522b..0a491589ff 100644 --- a/src/runtime/NEON/functions/NEConvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEConvolutionLayer.cpp @@ -23,630 +23,96 @@ */ #include "arm_compute/runtime/NEON/functions/NEConvolutionLayer.h" -#include "arm_compute/core/NEON/kernels/arm32/NEGEMMAArch32Kernel.h" -#include "arm_compute/core/NEON/kernels/arm64/NEGEMMAArch64Kernel.h" -#include "arm_compute/core/NEON/kernels/arm64/NEGEMMAArch64NativeKernel.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/NEON/NEScheduler.h" #include "support/ToolchainSupport.h" -namespace arm_compute -{ -#include "arm_compute/core/NEON/kernels/assembly/gemm_interleaved.hpp" -#include "arm_compute/core/NEON/kernels/assembly/kernels/a32_sgemm_8x6.hpp" -#include "arm_compute/core/NEON/kernels/assembly/kernels/a64_sgemm_12x8.hpp" -} // namespace arm_compute - #include #include namespace arm_compute { -namespace -{ -TensorShape get_reshaped_weights_shape(const ITensorInfo *weights, bool append_bias) -{ - const unsigned int mat_weights_cols = weights->dimension(3); - const unsigned int mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + (append_bias ? 1 : 0); - return TensorShape(mat_weights_cols, mat_weights_rows); -} -} // namespace - -NEConvolutionLayerReshapeWeights::NEConvolutionLayerReshapeWeights(std::shared_ptr memory_manager) - : _memory_group(std::move(memory_manager)), _weights_reshape_kernel(), _weights_transposed_kernel(), _weights_reshaped(), _transpose1xW(false) -{ -} - -void NEConvolutionLayerReshapeWeights::configure(const ITensor *weights, const ITensor *biases, ITensor *output, bool transpose1xW) -{ - // Perform validation step - ARM_COMPUTE_ERROR_ON_NULLPTR(weights, output); - ARM_COMPUTE_ERROR_THROW_ON(NEConvolutionLayerReshapeWeights::validate(weights->info(), - (biases != nullptr) ? biases->info() : nullptr, - output->info(), - transpose1xW)); - - // Check if bias are present, if yes they will be embedded to the weights matrix - const bool append_biases = (biases != nullptr) && !is_data_type_quantized_asymmetric(weights->info()->data_type()); - //const unsigned bias_element = (append_biases) ? 1 : 0; - const ITensor *biases_to_use = (append_biases) ? biases : nullptr; - - _transpose1xW = transpose1xW; - - if(transpose1xW) - { - // Create tensor to store the reshaped weights - TensorInfo info_wr = weights->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(get_reshaped_weights_shape(weights->info(), append_biases)); - - _weights_reshaped.allocator()->init(info_wr); - _memory_group.manage(&_weights_reshaped); - - _weights_reshape_kernel.configure(weights, biases, &_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()); -} - -Status NEConvolutionLayerReshapeWeights::validate(const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose1xW) -{ - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4); - if(!is_data_type_quantized_asymmetric(weights->data_type())) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(weights, output); - } - // Check if bias are present, if yes they will be embedded to the weights matrix - const bool append_bias = (biases != nullptr); - - if(append_bias) - { - ARM_COMPUTE_ERROR_ON(is_data_type_quantized_asymmetric(weights->data_type())); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(weights, biases); - ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(3)); - ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); - } - - // Checks performed when biases are present - if(append_bias) - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases); - ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(3)); - ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); - } - - if(transpose1xW) - { - TensorInfo weights_reshaped = weights->clone()->set_tensor_shape(get_reshaped_weights_shape(weights, append_bias)); - ARM_COMPUTE_RETURN_ON_ERROR(NEWeightsReshapeKernel::validate(weights, biases, &weights_reshaped)); - ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMTranspose1xWKernel::validate(&weights_reshaped, output)); - } - else - { - ARM_COMPUTE_RETURN_ON_ERROR(NEWeightsReshapeKernel::validate(weights, biases, output)); - } - - return Status{}; -} - -void NEConvolutionLayerReshapeWeights::run() -{ - _memory_group.acquire(); - - NEScheduler::get().schedule(&_weights_reshape_kernel, 3); - - if(_transpose1xW) - { - NEScheduler::get().schedule(&_weights_transposed_kernel, Window::DimY); - } - - _memory_group.release(); -} - -namespace -{ -TensorShape get_reshaped_weights_shape_conv(const ITensorInfo *weights, bool append_bias, bool is_fully_connected_convolution) +NEConvolutionLayer::NEConvolutionLayer(std::shared_ptr memory_manager) + : _memory_manager(std::move(memory_manager)), _function() { - unsigned int mat_weights_cols = weights->dimension(3); - unsigned int mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + (append_bias ? 1 : 0); - - if(is_fully_connected_convolution) - { - // Create tensor to store the reshaped weights - return TensorShape(mat_weights_cols, mat_weights_rows); - } - else - { - // Create tensor to store transposed weights - const float transpose_width = 16.0f / weights->element_size(); - return TensorShape(mat_weights_rows * static_cast(transpose_width), static_cast(std::ceil(mat_weights_cols / transpose_width))); - } } -Status validate_and_initialize_values(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const PadStrideInfo &conv_info, const WeightsInfo &weights_info, DataType &dt, - bool &append_bias, - bool &are_weights_reshaped, unsigned int &kernel_width, unsigned int &kernel_height, - bool &is_fully_connected_convolution, bool &is_interleaved_transposed, bool &is_quantized, - unsigned int &mat_weights_cols, unsigned int &mat_weights_rows, - unsigned int &conv_w, unsigned int &conv_h) -{ - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, weights); - ARM_COMPUTE_RETURN_ERROR_ON(!weights_info.are_reshaped() && weights->dimension(2) != input->dimension(2)); - ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4); - ARM_COMPUTE_RETURN_ERROR_ON(weights_info.are_reshaped() && is_data_type_quantized_asymmetric(input->data_type())); - - dt = input->data_type(); - is_quantized = is_data_type_quantized_asymmetric(dt); - - if(biases != nullptr) - { - if(is_quantized) - { - ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32); - } - else - { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); - } - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases); - ARM_COMPUTE_RETURN_ERROR_ON(!weights_info.are_reshaped() && biases->dimension(0) != weights->dimension(3)); - ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); - } - - append_bias = (biases != nullptr) && (!is_quantized); - are_weights_reshaped = weights_info.are_reshaped(); - kernel_width = (are_weights_reshaped) ? weights_info.kernel_size().first : weights->dimension(0); - kernel_height = (are_weights_reshaped) ? weights_info.kernel_size().second : weights->dimension(1); - mat_weights_cols = weights->dimension(3); - mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + (append_bias ? 1 : 0); - - std::tie(conv_w, conv_h) = scaled_dimensions(input->dimension(0), input->dimension(1), kernel_width, kernel_height, - conv_info); - - // Check if its a "fully connected" convolution - is_fully_connected_convolution = ((conv_w == 1) && (conv_h == 1)); - is_interleaved_transposed = (!is_fully_connected_convolution && !is_quantized); - - return Status{}; -} -} // namespace - -NEConvolutionLayer::NEConvolutionLayer(const std::shared_ptr &memory_manager) - : _memory_group(memory_manager), _input_im2col_kernel(), _input_interleave_kernel(), _reshape_weights(), _mm_kernel(), _mm_optimised_kernel(nullptr), _mm_gemmlowp(memory_manager), - _gemmlowp_output_stage(), _output_col2im_kernel(), _input_im2col_reshaped(), _input_interleaved_reshaped(), _weights_reshaped(), _gemm_output(), _tmp_output(), _workspace(), _append_bias(false), - _is_fully_connected_convolution(false), _are_weights_reshaped(false), _is_quantized(false), _is_interleaved_transposed(false) -{ -} - -void NEConvolutionLayer::configure_mm(const ITensor *input, const ITensor *weights, ITensor *output) -{ - if(_is_quantized) - { - // 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 - { - _mm_kernel.configure(input, weights, output, 1.f); - } -} - -void NEConvolutionLayer::configure_asm_mm(const struct CPUInfo &ci, int M, int N, int K) -{ - ARM_COMPUTE_UNUSED(ci); - ARM_COMPUTE_UNUSED(M); - ARM_COMPUTE_UNUSED(N); - ARM_COMPUTE_UNUSED(K); -#if defined(__arm__) || defined(__aarch64__) -#if defined(__arm__) - GemmInterleaved gemm(&ci, M, N, K, false, false); -#elif defined(__aarch64__) - GemmInterleaved gemm(&ci, M, N, K, false, false); -#endif /* defined(__arm__) || defined(__aarch64__) */ - - constexpr size_t alignment = 4096; - _workspace.allocator()->init(TensorInfo(TensorShape{ (gemm.get_working_size() + alignment - 1) * NEScheduler::get().num_threads() }, 1, DataType::U8)); - _memory_group.manage(&_workspace); -#endif /* defined(__arm__) || defined(__aarch64__) */ -} - -void NEConvolutionLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info) +void NEConvolutionLayer::configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info) { // Perform validate step ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); + ARM_COMPUTE_ERROR_THROW_ON(NEConvolutionLayer::validate(input->info(), weights->info(), ((biases != nullptr) ? biases->info() : nullptr), output->info(), conv_info, weights_info)); - DataType dt{}; - unsigned int kernel_width = 0; - unsigned int kernel_height = 0; - unsigned int mat_weights_cols = 0; - unsigned int mat_weights_rows = 0; - unsigned int conv_w = 0; - unsigned int conv_h = 0; - - Status status = validate_and_initialize_values(input->info(), weights->info(), (biases == nullptr) ? nullptr : biases->info(), conv_info, weights_info, dt, _append_bias, _are_weights_reshaped, - kernel_width, kernel_height, - _is_fully_connected_convolution, _is_interleaved_transposed, _is_quantized, - mat_weights_cols, mat_weights_rows, conv_w, conv_h); - - ARM_COMPUTE_ERROR_THROW_ON(status); - - const unsigned int fixed_point_position = input->info()->fixed_point_position(); - const ITensor *biases_to_use = (_append_bias) ? biases : nullptr; - -#if defined(__arm__) - if(NEScheduler::get().cpu_info().CPU == CPUTarget::ARMV7 && dt == DataType::F32) - { - _mm_optimised_kernel = support::cpp14::make_unique(); - } -#elif defined(__aarch64__) - if(NEScheduler::get().cpu_info().CPU >= CPUTarget::ARMV8 && dt == DataType::F32) - { - _mm_optimised_kernel = support::cpp14::make_unique(); - } -#endif /* defined(__arm__) || defined(__aarch64__) */ - - // Reshape weights if needed - if(_mm_optimised_kernel != nullptr) + switch(NEConvolutionLayer::get_convolution_method(input->info(), weights->info(), ((biases != nullptr) ? biases->info() : nullptr), output->info(), conv_info, + weights_info)) { - if(_are_weights_reshaped) + case ConvolutionMethod::WINOGRAD: { - mat_weights_cols = weights_info.num_kernels(); - mat_weights_rows = weights->info()->dimension(1); + auto f = arm_compute::support::cpp14::make_unique(_memory_manager); + f->configure(input, weights, biases, output, conv_info); + _function = std::move(f); + break; } - else + case ConvolutionMethod::GEMM: { - TensorShape reshaped_weights_shape{ mat_weights_cols, mat_weights_rows }; - - // Create tensor to store the reshaped weights - _weights_reshaped.allocator()->init(TensorInfo(reshaped_weights_shape, 1, dt, fixed_point_position)); - _reshape_weights.configure(weights, biases, &_weights_reshaped, false /* 1xW transpose */); - weights = &_weights_reshaped; + auto f = arm_compute::support::cpp14::make_unique(_memory_manager); + f->configure(input, weights, biases, output, conv_info, weights_info); + _function = std::move(f); + break; } - } - else - { - if(_are_weights_reshaped) + case ConvolutionMethod::DIRECT: { - if(_is_fully_connected_convolution || _is_quantized) - { - mat_weights_cols = weights_info.num_kernels(); - mat_weights_rows = weights->info()->dimension(1); - } - else - { - const unsigned int transpose_width = 16 / input->info()->element_size(); - mat_weights_cols = weights_info.num_kernels(); - mat_weights_rows = weights->info()->dimension(0) / transpose_width + (_append_bias ? 1 : 0); - } + auto f = arm_compute::support::cpp14::make_unique(_memory_manager); + f->configure(input, weights, biases, output, conv_info); + _function = std::move(f); + break; } - else - { - TensorShape reshaped_weights_shape; - - if(_is_fully_connected_convolution || _is_quantized) - { - reshaped_weights_shape = TensorShape{ mat_weights_cols, mat_weights_rows }; - } - else - { - // Create tensor to store transposed weights - const float transpose_width = 16.0f / input->info()->element_size(); - reshaped_weights_shape = TensorShape{ mat_weights_rows *static_cast(transpose_width), - static_cast(std::ceil(mat_weights_cols / transpose_width)) }; - } - - // Create tensor to store the reshaped weights - _weights_reshaped.allocator()->init(TensorInfo(reshaped_weights_shape, 1, dt, fixed_point_position)); - _reshape_weights.configure(weights, biases_to_use, &_weights_reshaped, _is_interleaved_transposed /* 1xW transpose */); - 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); - _input_im2col_reshaped.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col)); - _memory_group.manage(&_input_im2col_reshaped); - - // Create tensor (interleave) to prepare input tensor for GEMM - if(!_is_fully_connected_convolution && _mm_optimised_kernel == nullptr) - { - TensorShape shape_interleaved(shape_im2col); - shape_interleaved.set(0, shape_interleaved.x() * 4); - shape_interleaved.set(1, std::ceil(shape_interleaved.y() / 4.f)); - _input_interleaved_reshaped.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_interleaved)); - _memory_group.manage(&_input_interleaved_reshaped); - } - - // Create GEMM output tensor - TensorShape shape_gemm(_input_im2col_reshaped.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. - 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 kernels - // Configure im2col - _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _append_bias); - - // Configure matrix multiply - if(_mm_optimised_kernel != nullptr) - { - struct CPUInfo ci = NEScheduler::get().cpu_info(); - - const int M = _gemm_output.info()->tensor_shape().y(); - const int N = _gemm_output.info()->tensor_shape().x(); - const int K = _input_im2col_reshaped.info()->tensor_shape().x(); - -#if defined(__aarch64__) - if((N <= 128) && (K <= 128)) - { - _mm_optimised_kernel = support::cpp14::make_unique(); - } - else -#endif /* defined(__aarch64__) */ - { - configure_asm_mm(ci, M, N, K); - } - - // Configure matrix multiplication kernel - _mm_optimised_kernel->configure(&_input_im2col_reshaped, weights, &_gemm_output, &_workspace); - - _workspace.allocator()->allocate(); - } - else - { - if(_is_interleaved_transposed) - { - // Configure GEMMInterleave4x4. _input_interleaved_reshaped will be auto configured in the kernel - _input_interleave_kernel.configure(&_input_im2col_reshaped, &_input_interleaved_reshaped); - - // Configure GEMM - configure_mm(&_input_interleaved_reshaped, weights, &_gemm_output); - _input_interleaved_reshaped.allocator()->allocate(); - } - else - { - configure_mm(&_input_im2col_reshaped, weights, &_gemm_output); - } - } - - _input_im2col_reshaped.allocator()->allocate(); - - // Configure output stage for quantized case - if(_is_quantized) - { - const QuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? input->info()->quantization_info() : output->info()->quantization_info(); - - float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output_quant_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_quant_info.offset); - } - - // Configure Col2Im - _output_col2im_kernel.configure(_is_quantized ? &_tmp_output : &_gemm_output, output, Size2D(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(); + default: + ARM_COMPUTE_ERROR("Not supported."); + break; } } Status NEConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info) { - DataType dt{}; - bool append_bias{}; - bool are_weights_reshaped{}; - bool is_fully_connected_convolution{}; - bool is_interleaved_transposed{}; - bool is_quantized{}; - unsigned int kernel_width = 0; - unsigned int kernel_height = 0; - unsigned int mat_weights_cols = 0; - unsigned int mat_weights_rows = 0; - unsigned int conv_w = 0; - unsigned int conv_h = 0; - - Status status = validate_and_initialize_values(input, weights, biases, conv_info, weights_info, dt, append_bias, are_weights_reshaped, kernel_width, kernel_height, - is_fully_connected_convolution, is_interleaved_transposed, is_quantized, mat_weights_cols, mat_weights_rows, - conv_w, conv_h); - - ARM_COMPUTE_RETURN_ON_ERROR(status); - - std::unique_ptr reshaped_weights = weights->clone(); - bool optimised_kernel = false; - -#if defined(__arm__) - if(NEScheduler::get().cpu_info().CPU == CPUTarget::ARMV7 && dt == DataType::F32) + switch(NEConvolutionLayer::get_convolution_method(input, weights, biases, output, conv_info, weights_info)) { - optimised_kernel = true; - } -#elif defined(__aarch64__) - if(NEScheduler::get().cpu_info().CPU >= CPUTarget::ARMV8 && dt == DataType::F32) - { - optimised_kernel = true; - } -#endif /* defined(__arm__) || defined(__aarch64__) */ - - // Reshape weights if needed - if(optimised_kernel) - { - if(are_weights_reshaped) - { - mat_weights_cols = weights_info.num_kernels(); - mat_weights_rows = weights->dimension(1); - } - else - { - TensorShape reshaped_weights_shape{ mat_weights_cols, mat_weights_rows }; - - // Create tensor to store the reshaped weights - reshaped_weights->set_tensor_shape(get_reshaped_weights_shape_conv(weights, append_bias, is_fully_connected_convolution)); - ARM_COMPUTE_RETURN_ON_ERROR(NEConvolutionLayerReshapeWeights::validate(weights, biases, reshaped_weights.get(), !is_fully_connected_convolution /* 1xW transpose */)); - weights = reshaped_weights.get(); - } - } - else - { - if(are_weights_reshaped) - { - const unsigned int transpose_width = 16 / input->element_size(); - mat_weights_cols = weights_info.num_kernels(); - mat_weights_rows = weights->dimension(0) / transpose_width + (append_bias ? 1 : 0); - } - else - { - TensorShape reshaped_weights_shape; - - if(is_fully_connected_convolution || is_quantized) - { - reshaped_weights_shape = TensorShape{ mat_weights_cols, mat_weights_rows }; - } - else - { - // Create tensor to store transposed weights - const float transpose_width = 16.0f / input->element_size(); - reshaped_weights_shape = TensorShape{ mat_weights_rows *static_cast(transpose_width), - static_cast(std::ceil(mat_weights_cols / transpose_width)) }; - } - - // Create tensor to store the reshaped weights - reshaped_weights->set_tensor_shape(get_reshaped_weights_shape_conv(weights, append_bias, is_fully_connected_convolution)); - ARM_COMPUTE_RETURN_ON_ERROR(NEConvolutionLayerReshapeWeights::validate(weights, biases, reshaped_weights.get(), !is_fully_connected_convolution /* 1xW transpose */)); - weights = reshaped_weights.get(); - } - } - - // Validate im2col - const unsigned int mat_input_cols = mat_weights_rows; - const unsigned int mat_input_rows = conv_w * conv_h; - TensorShape shape_im2col = input->tensor_shape(); - shape_im2col.set(0, mat_input_cols); - shape_im2col.set(1, mat_input_rows); - shape_im2col.set(2, 1); - TensorInfo im2_col_info = input->clone()->set_tensor_shape(shape_im2col); - ARM_COMPUTE_RETURN_ON_ERROR(NEIm2ColKernel::validate(input, &im2_col_info, Size2D(weights->dimension(0), weights->dimension(1)), conv_info, append_bias)); - - // Create GEMM output tensor - TensorShape shape_gemm(im2_col_info.tensor_shape()); - shape_gemm.set(0, mat_weights_cols); - shape_gemm.set(1, mat_input_rows); - TensorInfo gemm_output_info = input->clone()->set_tensor_shape(shape_gemm); - - // Validate GEMM interleave and multiply - if(is_interleaved_transposed) - { - TensorShape shape_interleaved = shape_im2col; - shape_interleaved.set(0, shape_interleaved.x() * 4); - shape_interleaved.set(1, std::ceil(shape_interleaved.y() / 4.f)); - TensorInfo input_interleaved_info = input->clone()->set_tensor_shape(shape_interleaved); - ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMInterleave4x4Kernel::validate(&im2_col_info, &input_interleaved_info)); - ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixMultiplyKernel::validate(&input_interleaved_info, weights, &gemm_output_info)); - } - else - { - ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixMultiplyKernel::validate(&im2_col_info, weights, &gemm_output_info)); + case ConvolutionMethod::WINOGRAD: + //Validate Winograd + NEWinogradLayer::validate(input, weights, biases, output, conv_info); + break; + case ConvolutionMethod::GEMM: + //Validate Gemm-based Convolution + NEGEMMConvolutionLayer::validate(input, weights, biases, output, conv_info, weights_info); + break; + case ConvolutionMethod::DIRECT: + //Validate Gemm-based Convolution + NEDirectConvolutionLayer::validate(input, weights, biases, output, conv_info); + default: + ARM_COMPUTE_ERROR("Not supported."); + break; } - ARM_COMPUTE_RETURN_ON_ERROR(NECol2ImKernel::validate(&gemm_output_info, output, Size2D(conv_w, conv_h))); - - ARM_COMPUTE_RETURN_ERROR_ON_MSG((output->dimension(0) != conv_w) || (output->dimension(1) != conv_h), "Output shape does not match the expected one"); - return Status{}; } -void NEConvolutionLayer::run() +ConvolutionMethod NEConvolutionLayer::get_convolution_method(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, + const WeightsInfo &weights_info) { - // Run weights reshaping (Runs once for every configure) - if(!_are_weights_reshaped) - { - _are_weights_reshaped = true; - _reshape_weights.run(); - } - - _memory_group.acquire(); - - // Run input reshaping - NEScheduler::get().schedule(&_input_im2col_kernel, Window::DimY); - - // Runs matrix multiply on reshaped matrices - if(_mm_optimised_kernel != nullptr) - { - NEScheduler::get().schedule(_mm_optimised_kernel.get(), Window::DimY); - } - else - { - if(_is_interleaved_transposed) - { - // Run interleave - NEScheduler::get().schedule(&_input_interleave_kernel, Window::DimY); - } - - // Runs matrix multiply on reshaped matrices - if(_is_quantized) - { - _mm_gemmlowp.run(); - } - else - { - NEScheduler::get().schedule(&_mm_kernel, Window::DimY); - } - } - - // Run output stage for quantized case - if(_is_quantized) + ARM_COMPUTE_UNUSED(output); + ARM_COMPUTE_UNUSED(weights_info); + if((input->data_type() == DataType::F32) && (weights->dimension(0) == 3) && (weights->dimension(1) == 3) && (weights->num_dimensions() <= 4) && (conv_info.stride().first == 1) + && (conv_info.stride().second == 1) && (biases != nullptr)) { - _gemmlowp_output_stage.run(); + return ConvolutionMethod::WINOGRAD; } + return ConvolutionMethod::GEMM; +} - // Reshape output matrix - NEScheduler::get().schedule(&_output_col2im_kernel, Window::DimY); - - _memory_group.release(); +void NEConvolutionLayer::run() +{ + _function->run(); } } // namespace arm_compute diff --git a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp new file mode 100644 index 0000000000..d0a16ef40d --- /dev/null +++ b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp @@ -0,0 +1,652 @@ +/* + * 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/NEON/functions/NEGEMMConvolutionLayer.h" + +#include "arm_compute/core/NEON/kernels/arm32/NEGEMMAArch32Kernel.h" +#include "arm_compute/core/NEON/kernels/arm64/NEGEMMAArch64Kernel.h" +#include "arm_compute/core/NEON/kernels/arm64/NEGEMMAArch64NativeKernel.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/NEON/NEScheduler.h" +#include "support/ToolchainSupport.h" + +namespace arm_compute +{ +#include "arm_compute/core/NEON/kernels/assembly/gemm_interleaved.hpp" +#include "arm_compute/core/NEON/kernels/assembly/kernels/a32_sgemm_8x6.hpp" +#include "arm_compute/core/NEON/kernels/assembly/kernels/a64_sgemm_12x8.hpp" +} // namespace arm_compute + +#include +#include + +namespace +{ +arm_compute::TensorShape get_reshaped_weights_shape(const arm_compute::ITensorInfo *weights, bool append_bias) +{ + const unsigned int mat_weights_cols = weights->dimension(3); + const unsigned int mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + (append_bias ? 1 : 0); + return arm_compute::TensorShape(mat_weights_cols, mat_weights_rows); +} +} // namespace + +namespace arm_compute +{ +NEConvolutionLayerReshapeWeights::NEConvolutionLayerReshapeWeights(std::shared_ptr memory_manager) + : _memory_group(std::move(memory_manager)), _weights_reshape_kernel(), _weights_transposed_kernel(), _weights_reshaped(), _transpose1xW(false) +{ +} + +void NEConvolutionLayerReshapeWeights::configure(const ITensor *weights, const ITensor *biases, ITensor *output, bool transpose1xW) +{ + // Perform validation step + ARM_COMPUTE_ERROR_ON_NULLPTR(weights, output); + ARM_COMPUTE_ERROR_THROW_ON(NEConvolutionLayerReshapeWeights::validate(weights->info(), + (biases != nullptr) ? biases->info() : nullptr, + output->info(), + transpose1xW)); + + // Check if bias are present, if yes they will be embedded to the weights matrix + const bool append_biases = (biases != nullptr) && !is_data_type_quantized_asymmetric(weights->info()->data_type()); + //const unsigned bias_element = (append_biases) ? 1 : 0; + const ITensor *biases_to_use = (append_biases) ? biases : nullptr; + + _transpose1xW = transpose1xW; + + if(transpose1xW) + { + // Create tensor to store the reshaped weights + TensorInfo info_wr = weights->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(get_reshaped_weights_shape(weights->info(), append_biases)); + + _weights_reshaped.allocator()->init(info_wr); + _memory_group.manage(&_weights_reshaped); + + _weights_reshape_kernel.configure(weights, biases, &_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()); +} + +Status NEConvolutionLayerReshapeWeights::validate(const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose1xW) +{ + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4); + if(!is_data_type_quantized_asymmetric(weights->data_type())) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, output); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(weights, output); + } + // Check if bias are present, if yes they will be embedded to the weights matrix + const bool append_bias = (biases != nullptr); + + if(append_bias) + { + ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(weights->data_type())); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(weights, biases); + ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(3)); + ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); + } + + // Checks performed when biases are present + if(append_bias) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases); + ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(3)); + ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); + } + + if(transpose1xW) + { + TensorInfo weights_reshaped = weights->clone()->set_tensor_shape(get_reshaped_weights_shape(weights, append_bias)); + ARM_COMPUTE_RETURN_ON_ERROR(NEWeightsReshapeKernel::validate(weights, biases, &weights_reshaped)); + ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMTranspose1xWKernel::validate(&weights_reshaped, output)); + } + else + { + ARM_COMPUTE_RETURN_ON_ERROR(NEWeightsReshapeKernel::validate(weights, biases, output)); + } + + return Status{}; +} + +void NEConvolutionLayerReshapeWeights::run() +{ + _memory_group.acquire(); + + NEScheduler::get().schedule(&_weights_reshape_kernel, 3); + + if(_transpose1xW) + { + NEScheduler::get().schedule(&_weights_transposed_kernel, Window::DimY); + } + + _memory_group.release(); +} + +namespace +{ +TensorShape get_reshaped_weights_shape_conv(const ITensorInfo *weights, bool append_bias, bool is_fully_connected_convolution) +{ + unsigned int mat_weights_cols = weights->dimension(3); + unsigned int mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + (append_bias ? 1 : 0); + + if(is_fully_connected_convolution) + { + // Create tensor to store the reshaped weights + return TensorShape(mat_weights_cols, mat_weights_rows); + } + else + { + // Create tensor to store transposed weights + const float transpose_width = 16.0f / weights->element_size(); + return TensorShape(mat_weights_rows * static_cast(transpose_width), static_cast(std::ceil(mat_weights_cols / transpose_width))); + } +} + +Status validate_and_initialize_values(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const PadStrideInfo &conv_info, const WeightsInfo &weights_info, DataType &dt, + bool &append_bias, + bool &are_weights_reshaped, unsigned int &kernel_width, unsigned int &kernel_height, + bool &is_fully_connected_convolution, bool &is_interleaved_transposed, bool &is_quantized, + unsigned int &mat_weights_cols, unsigned int &mat_weights_rows, + unsigned int &conv_w, unsigned int &conv_h) +{ + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, weights); + ARM_COMPUTE_RETURN_ERROR_ON(!weights_info.are_reshaped() && weights->dimension(2) != input->dimension(2)); + ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4); + ARM_COMPUTE_RETURN_ERROR_ON(weights_info.are_reshaped() && is_data_type_quantized_asymmetric(input->data_type())); + + dt = input->data_type(); + is_quantized = is_data_type_quantized_asymmetric(dt); + + if(biases != nullptr) + { + if(is_quantized) + { + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32); + } + else + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); + } + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases); + ARM_COMPUTE_RETURN_ERROR_ON(!weights_info.are_reshaped() && biases->dimension(0) != weights->dimension(3)); + ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); + } + + append_bias = (biases != nullptr) && (!is_quantized); + are_weights_reshaped = weights_info.are_reshaped(); + kernel_width = (are_weights_reshaped) ? weights_info.kernel_size().first : weights->dimension(0); + kernel_height = (are_weights_reshaped) ? weights_info.kernel_size().second : weights->dimension(1); + mat_weights_cols = weights->dimension(3); + mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + (append_bias ? 1 : 0); + + std::tie(conv_w, conv_h) = scaled_dimensions(input->dimension(0), input->dimension(1), kernel_width, kernel_height, + conv_info); + + // Check if its a "fully connected" convolution + is_fully_connected_convolution = ((conv_w == 1) && (conv_h == 1)); + is_interleaved_transposed = (!is_fully_connected_convolution && !is_quantized); + + return Status{}; +} +} // namespace + +NEGEMMConvolutionLayer::NEGEMMConvolutionLayer(const std::shared_ptr &memory_manager) + : _memory_group(memory_manager), _input_im2col_kernel(), _input_interleave_kernel(), _reshape_weights(), _mm_kernel(), _mm_optimised_kernel(nullptr), _mm_gemmlowp(memory_manager), + _gemmlowp_output_stage(), _output_col2im_kernel(), _input_im2col_reshaped(), _input_interleaved_reshaped(), _weights_reshaped(), _gemm_output(), _tmp_output(), _workspace(), _append_bias(false), + _is_fully_connected_convolution(false), _are_weights_reshaped(false), _is_quantized(false), _is_interleaved_transposed(false) +{ +} + +void NEGEMMConvolutionLayer::configure_mm(const ITensor *input, const ITensor *weights, ITensor *output) +{ + if(_is_quantized) + { + // 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 + { + _mm_kernel.configure(input, weights, output, 1.f); + } +} + +void NEGEMMConvolutionLayer::configure_asm_mm(const struct CPUInfo &ci, int M, int N, int K) +{ + ARM_COMPUTE_UNUSED(ci); + ARM_COMPUTE_UNUSED(M); + ARM_COMPUTE_UNUSED(N); + ARM_COMPUTE_UNUSED(K); +#if defined(__arm__) || defined(__aarch64__) +#if defined(__arm__) + GemmInterleaved gemm(&ci, M, N, K, false, false); +#elif defined(__aarch64__) + GemmInterleaved gemm(&ci, M, N, K, false, false); +#endif /* defined(__arm__) || defined(__aarch64__) */ + + constexpr size_t alignment = 4096; + _workspace.allocator()->init(TensorInfo(TensorShape{ (gemm.get_working_size() + alignment - 1) * NEScheduler::get().num_threads() }, 1, DataType::U8)); + _memory_group.manage(&_workspace); +#endif /* defined(__arm__) || defined(__aarch64__) */ +} + +void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info) +{ + // Perform validate step + ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output); + + DataType dt{}; + unsigned int kernel_width = 0; + unsigned int kernel_height = 0; + unsigned int mat_weights_cols = 0; + unsigned int mat_weights_rows = 0; + unsigned int conv_w = 0; + unsigned int conv_h = 0; + + Status status = validate_and_initialize_values(input->info(), weights->info(), (biases == nullptr) ? nullptr : biases->info(), conv_info, weights_info, dt, _append_bias, _are_weights_reshaped, + kernel_width, kernel_height, + _is_fully_connected_convolution, _is_interleaved_transposed, _is_quantized, + mat_weights_cols, mat_weights_rows, conv_w, conv_h); + + ARM_COMPUTE_ERROR_THROW_ON(status); + + const unsigned int fixed_point_position = input->info()->fixed_point_position(); + const ITensor *biases_to_use = (_append_bias) ? biases : nullptr; + +#if defined(__arm__) + if(NEScheduler::get().cpu_info().CPU == CPUTarget::ARMV7 && dt == DataType::F32) + { + _mm_optimised_kernel = support::cpp14::make_unique(); + } +#elif defined(__aarch64__) + if(NEScheduler::get().cpu_info().CPU >= CPUTarget::ARMV8 && dt == DataType::F32) + { + _mm_optimised_kernel = support::cpp14::make_unique(); + } +#endif /* defined(__arm__) || defined(__aarch64__) */ + + // Reshape weights if needed + if(_mm_optimised_kernel != nullptr) + { + if(_are_weights_reshaped) + { + mat_weights_cols = weights_info.num_kernels(); + mat_weights_rows = weights->info()->dimension(1); + } + else + { + TensorShape reshaped_weights_shape{ mat_weights_cols, mat_weights_rows }; + + // Create tensor to store the reshaped weights + _weights_reshaped.allocator()->init(TensorInfo(reshaped_weights_shape, 1, dt, fixed_point_position)); + _reshape_weights.configure(weights, biases, &_weights_reshaped, false /* 1xW transpose */); + weights = &_weights_reshaped; + } + } + else + { + if(_are_weights_reshaped) + { + if(_is_fully_connected_convolution || _is_quantized) + { + mat_weights_cols = weights_info.num_kernels(); + mat_weights_rows = weights->info()->dimension(1); + } + else + { + const unsigned int transpose_width = 16 / input->info()->element_size(); + mat_weights_cols = weights_info.num_kernels(); + mat_weights_rows = weights->info()->dimension(0) / transpose_width + (_append_bias ? 1 : 0); + } + } + else + { + TensorShape reshaped_weights_shape; + + if(_is_fully_connected_convolution || _is_quantized) + { + reshaped_weights_shape = TensorShape{ mat_weights_cols, mat_weights_rows }; + } + else + { + // Create tensor to store transposed weights + const float transpose_width = 16.0f / input->info()->element_size(); + reshaped_weights_shape = TensorShape{ mat_weights_rows *static_cast(transpose_width), + static_cast(std::ceil(mat_weights_cols / transpose_width)) }; + } + + // Create tensor to store the reshaped weights + _weights_reshaped.allocator()->init(TensorInfo(reshaped_weights_shape, 1, dt, fixed_point_position)); + _reshape_weights.configure(weights, biases_to_use, &_weights_reshaped, _is_interleaved_transposed /* 1xW transpose */); + 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); + _input_im2col_reshaped.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col)); + _memory_group.manage(&_input_im2col_reshaped); + + // Create tensor (interleave) to prepare input tensor for GEMM + if(!_is_fully_connected_convolution && _mm_optimised_kernel == nullptr) + { + TensorShape shape_interleaved(shape_im2col); + shape_interleaved.set(0, shape_interleaved.x() * 4); + shape_interleaved.set(1, std::ceil(shape_interleaved.y() / 4.f)); + _input_interleaved_reshaped.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_interleaved)); + _memory_group.manage(&_input_interleaved_reshaped); + } + + // Create GEMM output tensor + TensorShape shape_gemm(_input_im2col_reshaped.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. + 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 kernels + // Configure im2col + _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _append_bias); + + // Configure matrix multiply + if(_mm_optimised_kernel != nullptr) + { + struct CPUInfo ci = NEScheduler::get().cpu_info(); + + const int M = _gemm_output.info()->tensor_shape().y(); + const int N = _gemm_output.info()->tensor_shape().x(); + const int K = _input_im2col_reshaped.info()->tensor_shape().x(); + +#if defined(__aarch64__) + if((N <= 128) && (K <= 128)) + { + _mm_optimised_kernel = support::cpp14::make_unique(); + } + else +#endif /* defined(__aarch64__) */ + { + configure_asm_mm(ci, M, N, K); + } + + // Configure matrix multiplication kernel + _mm_optimised_kernel->configure(&_input_im2col_reshaped, weights, &_gemm_output, &_workspace); + + _workspace.allocator()->allocate(); + } + else + { + if(_is_interleaved_transposed) + { + // Configure GEMMInterleave4x4. _input_interleaved_reshaped will be auto configured in the kernel + _input_interleave_kernel.configure(&_input_im2col_reshaped, &_input_interleaved_reshaped); + + // Configure GEMM + configure_mm(&_input_interleaved_reshaped, weights, &_gemm_output); + _input_interleaved_reshaped.allocator()->allocate(); + } + else + { + configure_mm(&_input_im2col_reshaped, weights, &_gemm_output); + } + } + + _input_im2col_reshaped.allocator()->allocate(); + + // Configure output stage for quantized case + if(_is_quantized) + { + const QuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? input->info()->quantization_info() : output->info()->quantization_info(); + + float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output_quant_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_quant_info.offset); + } + + // Configure Col2Im + _output_col2im_kernel.configure(_is_quantized ? &_tmp_output : &_gemm_output, output, Size2D(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(); + } +} + +Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, + const WeightsInfo &weights_info) +{ + DataType dt{}; + bool append_bias{}; + bool are_weights_reshaped{}; + bool is_fully_connected_convolution{}; + bool is_interleaved_transposed{}; + bool is_quantized{}; + unsigned int kernel_width = 0; + unsigned int kernel_height = 0; + unsigned int mat_weights_cols = 0; + unsigned int mat_weights_rows = 0; + unsigned int conv_w = 0; + unsigned int conv_h = 0; + + Status status = validate_and_initialize_values(input, weights, biases, conv_info, weights_info, dt, append_bias, are_weights_reshaped, kernel_width, kernel_height, + is_fully_connected_convolution, is_interleaved_transposed, is_quantized, mat_weights_cols, mat_weights_rows, + conv_w, conv_h); + + ARM_COMPUTE_RETURN_ON_ERROR(status); + + std::unique_ptr reshaped_weights = weights->clone(); + bool optimised_kernel = false; + +#if defined(__arm__) + if(NEScheduler::get().cpu_info().CPU == CPUTarget::ARMV7 && dt == DataType::F32) + { + optimised_kernel = true; + } +#elif defined(__aarch64__) + if(NEScheduler::get().cpu_info().CPU >= CPUTarget::ARMV8 && dt == DataType::F32) + { + optimised_kernel = true; + } +#endif /* defined(__arm__) || defined(__aarch64__) */ + + // Reshape weights if needed + if(optimised_kernel) + { + if(are_weights_reshaped) + { + mat_weights_cols = weights_info.num_kernels(); + mat_weights_rows = weights->dimension(1); + } + else + { + TensorShape reshaped_weights_shape{ mat_weights_cols, mat_weights_rows }; + + // Create tensor to store the reshaped weights + reshaped_weights->set_tensor_shape(get_reshaped_weights_shape_conv(weights, append_bias, is_fully_connected_convolution)); + ARM_COMPUTE_RETURN_ON_ERROR(NEConvolutionLayerReshapeWeights::validate(weights, biases, reshaped_weights.get(), !is_fully_connected_convolution /* 1xW transpose */)); + weights = reshaped_weights.get(); + } + } + else + { + if(are_weights_reshaped) + { + const unsigned int transpose_width = 16 / input->element_size(); + mat_weights_cols = weights_info.num_kernels(); + mat_weights_rows = weights->dimension(0) / transpose_width + (append_bias ? 1 : 0); + } + else + { + TensorShape reshaped_weights_shape; + + if(is_fully_connected_convolution || is_quantized) + { + reshaped_weights_shape = TensorShape{ mat_weights_cols, mat_weights_rows }; + } + else + { + // Create tensor to store transposed weights + const float transpose_width = 16.0f / input->element_size(); + reshaped_weights_shape = TensorShape{ mat_weights_rows *static_cast(transpose_width), + static_cast(std::ceil(mat_weights_cols / transpose_width)) }; + } + + // Create tensor to store the reshaped weights + reshaped_weights->set_tensor_shape(get_reshaped_weights_shape_conv(weights, append_bias, is_fully_connected_convolution)); + ARM_COMPUTE_RETURN_ON_ERROR(NEConvolutionLayerReshapeWeights::validate(weights, biases, reshaped_weights.get(), !is_fully_connected_convolution /* 1xW transpose */)); + weights = reshaped_weights.get(); + } + } + + // Validate im2col + const unsigned int mat_input_cols = mat_weights_rows; + const unsigned int mat_input_rows = conv_w * conv_h; + TensorShape shape_im2col = input->tensor_shape(); + shape_im2col.set(0, mat_input_cols); + shape_im2col.set(1, mat_input_rows); + shape_im2col.set(2, 1); + TensorInfo im2_col_info = input->clone()->set_tensor_shape(shape_im2col); + ARM_COMPUTE_RETURN_ON_ERROR(NEIm2ColKernel::validate(input, &im2_col_info, Size2D(weights->dimension(0), weights->dimension(1)), conv_info, append_bias)); + + // Create GEMM output tensor + TensorShape shape_gemm(im2_col_info.tensor_shape()); + shape_gemm.set(0, mat_weights_cols); + shape_gemm.set(1, mat_input_rows); + TensorInfo gemm_output_info = input->clone()->set_tensor_shape(shape_gemm); + + // Validate GEMM interleave and multiply + if(is_interleaved_transposed) + { + TensorShape shape_interleaved = shape_im2col; + shape_interleaved.set(0, shape_interleaved.x() * 4); + shape_interleaved.set(1, std::ceil(shape_interleaved.y() / 4.f)); + TensorInfo input_interleaved_info = input->clone()->set_tensor_shape(shape_interleaved); + ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMInterleave4x4Kernel::validate(&im2_col_info, &input_interleaved_info)); + ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixMultiplyKernel::validate(&input_interleaved_info, weights, &gemm_output_info)); + } + else + { + ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixMultiplyKernel::validate(&im2_col_info, weights, &gemm_output_info)); + } + + ARM_COMPUTE_RETURN_ON_ERROR(NECol2ImKernel::validate(&gemm_output_info, output, Size2D(conv_w, conv_h))); + + ARM_COMPUTE_RETURN_ERROR_ON_MSG((output->dimension(0) != conv_w) || (output->dimension(1) != conv_h), "Output shape does not match the expected one"); + + return Status{}; +} + +void NEGEMMConvolutionLayer::run() +{ + // Run weights reshaping (Runs once for every configure) + if(!_are_weights_reshaped) + { + _are_weights_reshaped = true; + _reshape_weights.run(); + } + + _memory_group.acquire(); + + // Run input reshaping + NEScheduler::get().schedule(&_input_im2col_kernel, Window::DimY); + + // Runs matrix multiply on reshaped matrices + if(_mm_optimised_kernel != nullptr) + { + NEScheduler::get().schedule(_mm_optimised_kernel.get(), Window::DimY); + } + else + { + if(_is_interleaved_transposed) + { + // Run interleave + NEScheduler::get().schedule(&_input_interleave_kernel, Window::DimY); + } + + // Runs matrix multiply on reshaped matrices + if(_is_quantized) + { + _mm_gemmlowp.run(); + } + else + { + NEScheduler::get().schedule(&_mm_kernel, Window::DimY); + } + } + + // Run output stage for quantized case + if(_is_quantized) + { + _gemmlowp_output_stage.run(); + } + + // Reshape output matrix + NEScheduler::get().schedule(&_output_col2im_kernel, Window::DimY); + + _memory_group.release(); +} +} // namespace arm_compute diff --git a/src/runtime/NEON/functions/NEWinogradLayer.cpp b/src/runtime/NEON/functions/NEWinogradLayer.cpp index e343583b36..0ac6d0966d 100644 --- a/src/runtime/NEON/functions/NEWinogradLayer.cpp +++ b/src/runtime/NEON/functions/NEWinogradLayer.cpp @@ -23,6 +23,7 @@ */ #include "arm_compute/runtime/NEON/functions/NEWinogradLayer.h" +#include "arm_compute/core/Error.h" #include "arm_compute/core/Utils.h" #include "arm_compute/core/Validate.h" #include "arm_compute/runtime/NEON/NEScheduler.h" @@ -46,6 +47,33 @@ inline Tensor4DShape internal_get_input_shape(const arm_compute::ITensor *input) namespace arm_compute { +namespace +{ +Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info) +{ + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, biases); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(0) != 3 && weights->dimension(0) != 5, "Only 3 and 5 kernels are supported"); + ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4); + + if(biases != nullptr) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); + ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); + } + + // Get parameters from conv_info + unsigned int stride_x = 0; + unsigned int stride_y = 0; + std::tie(stride_x, stride_y) = conv_info.stride(); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(stride_y != 1 || stride_x != 1, "Winograd layer only supports unit strides."); + + ARM_COMPUTE_UNUSED(output); + + return Status{}; +} +} //namespace + NEWinogradLayer::NEWinogradLayer(std::shared_ptr memory_manager) : _memory_group(std::move(memory_manager)), _batched_gemm_kernel(nullptr), _transform_input_kernel(nullptr), _transform_output_kernel(nullptr), _transform_weights_kernel(nullptr), _permute_input(), _permute_weights(), _permute_output(), _input_workspace(), _output_workspace(), _kernel_storage(), _input_nhwc(), _output_nhwc(), _weights_hwio(), _input(), _weights(), _output(), @@ -55,16 +83,9 @@ NEWinogradLayer::NEWinogradLayer(std::shared_ptr memory_manager) void NEWinogradLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, biases); - ARM_COMPUTE_ERROR_ON_MSG(weights->info()->dimension(0) != 3 && weights->info()->dimension(0) != 5, "Only 3 and 5 kernels are supported"); - ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4); - - if(biases != nullptr) - { - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); - ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1); - } + ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, biases, output); + ARM_COMPUTE_UNUSED(conv_info); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), biases->info(), output->info(), conv_info)); _weights = weights; _input = input; @@ -119,19 +140,15 @@ void NEWinogradLayer::configure(const ITensor *input, const ITensor *weights, co constexpr size_t storage_alignment = 64; const size_t kernel_storage_size = transform_weights_kernel->get_weight_storage_size(out_channels, in_channels) * data_type_size; _kernel_storage.allocator()->init(TensorInfo(TensorShape{ (kernel_storage_size + storage_alignment - 1) }, 1, DataType::U8)); - _memory_group.manage(&_kernel_storage); - _memory_group.manage(&_input_nhwc); _kernel_storage.allocator()->allocate(); // Input storage const size_t input_storage_size = transform_input_kernel->get_input_storage_size(in_shape.n_batches, in_shape.n_channels, in_shape.n_rows, in_shape.n_cols, use_same_padding) * data_type_size; _input_workspace.allocator()->init(TensorInfo(TensorShape{ (input_storage_size + storage_alignment - 1) }, 1, DataType::U8)); - _memory_group.manage(&_input_workspace); _input_workspace.allocator()->allocate(); // Output storage const size_t output_storage_size = transform_output_kernel->get_output_storage_size(in_shape.n_batches, in_shape.n_rows, in_shape.n_cols, out_channels, use_same_padding) * data_type_size; _output_workspace.allocator()->init(TensorInfo(TensorShape{ (output_storage_size + storage_alignment - 1) }, 1, DataType::U8)); - _memory_group.manage(&_output_workspace); _output_workspace.allocator()->allocate(); // configure and allocate dst tensor to be used to convert from winograd domain to spatial domain when calling to reshape_output() @@ -227,4 +244,13 @@ void NEWinogradLayer::run() _permute_output.run(); _memory_group.release(); } + +Status NEWinogradLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, biases, output); + ARM_COMPUTE_RETURN_ERROR_ON(validate_arguments(input, weights, biases, output, conv_info)); + + return Status{}; +} + } // namespace arm_compute diff --git a/tests/benchmark/NEON/ConvolutionLayer.cpp b/tests/benchmark/NEON/ConvolutionLayer.cpp index d871a6958c..1be95a50c1 100644 --- a/tests/benchmark/NEON/ConvolutionLayer.cpp +++ b/tests/benchmark/NEON/ConvolutionLayer.cpp @@ -54,7 +54,7 @@ const auto data_types = framework::dataset::make("DataType", { DataType::F32, Da #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ } // namespace -using NEConvolutionLayerFixture = ConvolutionLayerFixture; +using NEGEMMConvolutionLayerFixture = ConvolutionLayerFixture; TEST_SUITE(NEON) #if defined(__aarch64__) @@ -77,53 +77,53 @@ REGISTER_FIXTURE_DATA_TEST_CASE(SqueezeNetWinogradLayer, NEWinogradLayerFixture, framework::dataset::make("Batches", 1))); #endif /* __aarch64__ */ -REGISTER_FIXTURE_DATA_TEST_CASE(AlexNetConvolutionLayer, NEConvolutionLayerFixture, framework::DatasetMode::ALL, +REGISTER_FIXTURE_DATA_TEST_CASE(AlexNetConvolutionLayer, NEGEMMConvolutionLayerFixture, framework::DatasetMode::ALL, framework::dataset::combine(framework::dataset::combine(datasets::AlexNetConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", 1))); -REGISTER_FIXTURE_DATA_TEST_CASE(LeNet5ConvolutionLayer, NEConvolutionLayerFixture, framework::DatasetMode::ALL, +REGISTER_FIXTURE_DATA_TEST_CASE(LeNet5ConvolutionLayer, NEGEMMConvolutionLayerFixture, framework::DatasetMode::ALL, framework::dataset::combine(framework::dataset::combine(datasets::LeNet5ConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", 1))); -REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV1ConvolutionLayer, NEConvolutionLayerFixture, framework::DatasetMode::ALL, +REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV1ConvolutionLayer, NEGEMMConvolutionLayerFixture, framework::DatasetMode::ALL, framework::dataset::combine(framework::dataset::combine(datasets::GoogLeNetInceptionV1ConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", 1))); -REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV4ConvolutionLayer, NEConvolutionLayerFixture, framework::DatasetMode::ALL, +REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV4ConvolutionLayer, NEGEMMConvolutionLayerFixture, framework::DatasetMode::ALL, framework::dataset::combine(framework::dataset::combine(datasets::GoogLeNetInceptionV4ConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", 1))); -REGISTER_FIXTURE_DATA_TEST_CASE(SqueezeNetConvolutionLayer, NEConvolutionLayerFixture, framework::DatasetMode::ALL, +REGISTER_FIXTURE_DATA_TEST_CASE(SqueezeNetConvolutionLayer, NEGEMMConvolutionLayerFixture, framework::DatasetMode::ALL, framework::dataset::combine(framework::dataset::combine(datasets::SqueezeNetConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", 1))); TEST_SUITE(NIGHTLY) -REGISTER_FIXTURE_DATA_TEST_CASE(AlexNetConvolutionLayer, NEConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, +REGISTER_FIXTURE_DATA_TEST_CASE(AlexNetConvolutionLayer, NEGEMMConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, framework::dataset::combine(framework::dataset::combine(datasets::AlexNetConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", { 4, 8 }))); -REGISTER_FIXTURE_DATA_TEST_CASE(LeNet5ConvolutionLayer, NEConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, +REGISTER_FIXTURE_DATA_TEST_CASE(LeNet5ConvolutionLayer, NEGEMMConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, framework::dataset::combine(framework::dataset::combine(datasets::LeNet5ConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", { 4, 8 }))); -REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV1ConvolutionLayer, NEConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, +REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV1ConvolutionLayer, NEGEMMConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, framework::dataset::combine(framework::dataset::combine(datasets::GoogLeNetInceptionV1ConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", { 4, 8 }))); -REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV4ConvolutionLayer, NEConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, +REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV4ConvolutionLayer, NEGEMMConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, framework::dataset::combine(framework::dataset::combine(datasets::GoogLeNetInceptionV4ConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", { 4, 8 }))); -REGISTER_FIXTURE_DATA_TEST_CASE(SqueezeNetConvolutionLayer, NEConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, +REGISTER_FIXTURE_DATA_TEST_CASE(SqueezeNetConvolutionLayer, NEGEMMConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, framework::dataset::combine(framework::dataset::combine(datasets::SqueezeNetConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", { 4, 8 }))); // 8 batches use about 2GB of memory which is too much for most devices! -REGISTER_FIXTURE_DATA_TEST_CASE(VGG16ConvolutionLayer, NEConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, +REGISTER_FIXTURE_DATA_TEST_CASE(VGG16ConvolutionLayer, NEGEMMConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, framework::dataset::combine(framework::dataset::combine(datasets::VGG16ConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", { 1, 2 }))); -REGISTER_FIXTURE_DATA_TEST_CASE(YOLOV2ConvolutionLayer, NEConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, +REGISTER_FIXTURE_DATA_TEST_CASE(YOLOV2ConvolutionLayer, NEGEMMConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, framework::dataset::combine(framework::dataset::combine(datasets::YOLOV2ConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", { 1, 4, 8 }))); diff --git a/tests/validation/NEON/ConvolutionLayer.cpp b/tests/validation/NEON/ConvolutionLayer.cpp index ed4f160a37..59db279ac7 100644 --- a/tests/validation/NEON/ConvolutionLayer.cpp +++ b/tests/validation/NEON/ConvolutionLayer.cpp @@ -23,6 +23,7 @@ */ #include "arm_compute/core/Types.h" #include "arm_compute/runtime/NEON/functions/NEConvolutionLayer.h" +#include "arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h" #include "arm_compute/runtime/NEON/functions/NEWinogradLayer.h" #include "arm_compute/runtime/Tensor.h" #include "arm_compute/runtime/TensorAllocator.h" @@ -68,6 +69,44 @@ const auto CNNDataTypes = framework::dataset::make("DataType", TEST_SUITE(NEON) +TEST_SUITE(ConvolutionLayer) +DATA_TEST_CASE(ValidateConvolutionMethod, framework::DatasetMode::ALL, zip(zip(zip(zip(zip( + framework::dataset::make("InputInfo", { TensorInfo(TensorShape(8U, 8U, 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(3U, 3U, 5U, 21U), 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(1U), 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(6U, 6U, 1U), 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, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(2, 1, 0, 0), + PadStrideInfo(3, 2, 1, 0) + })), + framework::dataset::make("Expected", { ConvolutionMethod::WINOGRAD, ConvolutionMethod::WINOGRAD, ConvolutionMethod::GEMM, ConvolutionMethod::GEMM })), + input_info, weights_info, biases_info, output_info, conv_info, expected) +{ + ConvolutionMethod is_valid = NEConvolutionLayer::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); + ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS); +} +TEST_SUITE_END() + TEST_SUITE(WinogradLayer) template using NEWinogradLayerFixture = WinogradLayerValidationFixture; @@ -82,7 +121,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEWinogradLayerFixture, framework::Datas TEST_SUITE_END() TEST_SUITE_END() -TEST_SUITE(ConvolutionLayer) +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) @@ -107,7 +146,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::da const QuantizationInfo weights_quantization_info = weights.info()->quantization_info(); // Create and configure function - NEConvolutionLayer conv; + NEGEMMConvolutionLayer conv; conv.configure(&src, &weights, &bias, &dst, info); // Validate valid region @@ -130,21 +169,21 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::da } template -using NEConvolutionLayerFixture = ConvolutionValidationFixture; +using NEGEMMConvolutionLayerFixture = ConvolutionValidationFixture; TEST_SUITE(Float) #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionLayerFixture, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), - framework::dataset::make("DataType", DataType::F16))) +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerFixture, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true, false })), + framework::dataset::make("DataType", DataType::F16))) { // Validate output validate(Accessor(_target), _reference, tolerance_f16); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), - framework::dataset::make("DataType", DataType::F16))) +FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true, false })), + framework::dataset::make("DataType", DataType::F16))) { // Validate output validate(Accessor(_target), _reference, tolerance_f16); @@ -153,16 +192,16 @@ TEST_SUITE_END() #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionLayerFixture, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), - framework::dataset::make("DataType", DataType::F32))) +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerFixture, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true, false })), + framework::dataset::make("DataType", DataType::F32))) { // Validate output validate(Accessor(_target), _reference, tolerance_f32); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), - framework::dataset::make("DataType", DataType::F32))) +FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true, false })), + framework::dataset::make("DataType", DataType::F32))) { // Validate output validate(Accessor(_target), _reference, tolerance_f32); @@ -171,23 +210,23 @@ TEST_SUITE_END() TEST_SUITE_END() template -using NEConvolutionLayerFixedPointFixture = ConvolutionValidationFixedPointFixture; +using NEGEMMConvolutionLayerFixedPointFixture = ConvolutionValidationFixedPointFixture; TEST_SUITE(FixedPoint) TEST_SUITE(QS8) // We test for fixed point precision [4,6] -FIXTURE_DATA_TEST_CASE(RunTiny, NEConvolutionLayerFixedPointFixture, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), - framework::dataset::make("DataType", DataType::QS8)), - framework::dataset::make("FractionalBits", 4, 7))) +FIXTURE_DATA_TEST_CASE(RunTiny, NEGEMMConvolutionLayerFixedPointFixture, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true, false })), + framework::dataset::make("DataType", DataType::QS8)), + framework::dataset::make("FractionalBits", 4, 7))) { // Validate output validate(Accessor(_target), _reference, tolerance_q); } -FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionLayerFixedPointFixture, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), - framework::dataset::make("DataType", DataType::QS8)), - framework::dataset::make("FractionalBits", 4, 7))) +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerFixedPointFixture, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true, false })), + framework::dataset::make("DataType", DataType::QS8)), + framework::dataset::make("FractionalBits", 4, 7))) { // Validate output validate(Accessor(_target), _reference, tolerance_q); @@ -196,7 +235,7 @@ TEST_SUITE_END() TEST_SUITE(QS16) // Testing for fixed point position [1,14) -FIXTURE_DATA_TEST_CASE(RunTiny, NEConvolutionLayerFixedPointFixture, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyConvolutionLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunTiny, NEGEMMConvolutionLayerFixedPointFixture, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyConvolutionLayerDataset(), framework::dataset::make("ReshapeWeights", { true, false })), framework::dataset::make("DataType", DataType::QS16)), framework::dataset::make("FractionalBits", 1, 14))) @@ -204,10 +243,10 @@ FIXTURE_DATA_TEST_CASE(RunTiny, NEConvolutionLayerFixedPointFixture, fr // Validate output validate(Accessor(_target), _reference, tolerance_q); } -FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionLayerFixedPointFixture, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), - framework::dataset::make("DataType", DataType::QS16)), - framework::dataset::make("FractionalBits", 1, 14))) +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerFixedPointFixture, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true, false })), + framework::dataset::make("DataType", DataType::QS16)), + framework::dataset::make("FractionalBits", 1, 14))) { // Validate output validate(Accessor(_target), _reference, tolerance_q); @@ -216,11 +255,11 @@ TEST_SUITE_END() TEST_SUITE_END() template -using NEConvolutionLayerQuantizedFixture = ConvolutionValidationQuantizedFixture; +using NEGEMMConvolutionLayerQuantizedFixture = ConvolutionValidationQuantizedFixture; TEST_SUITE(Quantized) TEST_SUITE(QASYMM8) -FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionLayerQuantizedFixture, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerQuantizedFixture, 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) }))) @@ -228,10 +267,10 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionLayerQuantizedFixture, fr // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEConvolutionLayerQuantizedFixture, 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, 10) }))) +FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMConvolutionLayerQuantizedFixture, 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, 10) }))) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); -- cgit v1.2.1