From c93691717a6e7ca67e32b4dedd233b8c63b6daf2 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Wed, 26 Sep 2018 11:25:40 +0100 Subject: COMPMID-1523: Fuse BN node with convolution. Change-Id: I146936c9e98b343496a4b61cdbadf0eaa38e885a Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/154008 Reviewed-by: Michele DiGiorgio Reviewed-by: Giuseppe Rossini Tested-by: bsgcomp --- arm_compute/core/CL/CLKernels.h | 1 + .../CL/kernels/CLFuseBatchNormalizationKernel.h | 101 +++++++++++++++++++++ arm_compute/runtime/CL/CLFunctions.h | 1 + .../CL/functions/CLFuseBatchNormalization.h | 93 +++++++++++++++++++ 4 files changed, 196 insertions(+) create mode 100644 arm_compute/core/CL/kernels/CLFuseBatchNormalizationKernel.h create mode 100644 arm_compute/runtime/CL/functions/CLFuseBatchNormalization.h (limited to 'arm_compute') diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h index f6759b9f1e..95b6f9b70f 100644 --- a/arm_compute/core/CL/CLKernels.h +++ b/arm_compute/core/CL/CLKernels.h @@ -65,6 +65,7 @@ #include "arm_compute/core/CL/kernels/CLFillBorderKernel.h" #include "arm_compute/core/CL/kernels/CLFlattenLayerKernel.h" #include "arm_compute/core/CL/kernels/CLFloorKernel.h" +#include "arm_compute/core/CL/kernels/CLFuseBatchNormalizationKernel.h" #include "arm_compute/core/CL/kernels/CLGEMMInterleave4x4Kernel.h" #include "arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyKernel.h" #include "arm_compute/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.h" diff --git a/arm_compute/core/CL/kernels/CLFuseBatchNormalizationKernel.h b/arm_compute/core/CL/kernels/CLFuseBatchNormalizationKernel.h new file mode 100644 index 0000000000..05a57c171e --- /dev/null +++ b/arm_compute/core/CL/kernels/CLFuseBatchNormalizationKernel.h @@ -0,0 +1,101 @@ +/* + * Copyright (c) 2018 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef __ARM_COMPUTE_CLFUSEBATCHNORMALIZATIONKERNEL_H__ +#define __ARM_COMPUTE_CLFUSEBATCHNORMALIZATIONKERNEL_H__ + +#include "arm_compute/core/CL/ICLKernel.h" + +namespace arm_compute +{ +// Forward declarations +class ICLTensor; + +/** OpenCL kernel to fuse the batch normalization node to a preceding convolution node */ +class CLFuseBatchNormalizationKernel : public ICLKernel +{ +public: + /** Default constructor */ + CLFuseBatchNormalizationKernel(); + /** Prevent instances of this class from being copied (As this class contains pointers) */ + CLFuseBatchNormalizationKernel(const CLFuseBatchNormalizationKernel &) = delete; + /** Prevent instances of this class from being copied (As this class contains pointers) */ + CLFuseBatchNormalizationKernel &operator=(const CLFuseBatchNormalizationKernel &) = delete; + /** Allow instances of this class to be moved */ + CLFuseBatchNormalizationKernel(CLFuseBatchNormalizationKernel &&) = default; + /** Allow instances of this class to be moved */ + CLFuseBatchNormalizationKernel &operator=(CLFuseBatchNormalizationKernel &&) = default; + /** Default destructor */ + ~CLFuseBatchNormalizationKernel() = default; + /** Set the source, destination of the kernel + * + * @param[in] conv_weights Convolution layer weights tensor. Data type supported: F16/F32 + * @param[in] bn_mean Batch normalization layer mean tensor. Same as @p conv_weights + * @param[in] bn_var Batch normalization layer variance tensor. Same as @p conv_weights + * @param[out] fused_weights Output fused weights tensor. Same as @p conv_weights + * @param[out] fused_bias Output fused bias tensor. Same as @p conv_weights + * @param[in] conv_bias (Optional) Convolution layer bias tensor. Same as @p conv_weights + * @param[in] bn_beta (Optional) Batch normalization layer beta tensor. Same as @p conv_weights + * @param[in] bn_gamma (Optional) Batch normalization layer gamma tensor. Same as @p conv_weights + * @param[in] epsilon (Optional) Batch normalization layer epsilon parameter. Defaults to 0.001f. + */ + void configure(const ICLTensor *conv_weights, const ICLTensor *bn_mean, const ICLTensor *bn_var, ICLTensor *fused_weights, ICLTensor *fused_bias, + const ICLTensor *conv_bias = nullptr, const ICLTensor *bn_beta = nullptr, const ICLTensor *bn_gamma = nullptr, + float epsilon = 0.001f); + /** Static function to check if given info will lead to a valid configuration of @ref CLFuseBatchNormalizationKernel + * + * @param[in] conv_weights Convolution layer weights tensor. Data type supported: F16/F32 + * @param[in] bn_mean Batch normalization layer mean tensor. Same as @p conv_weights + * @param[in] bn_var Batch normalization layer variance tensor. Same as @p conv_weights + * @param[in] fused_weights Output fused weights tensor. Same as @p conv_weights + * @param[in] fused_bias Output fused bias tensor. Same as @p conv_weights + * @param[in] conv_bias (Optional) Convolution layer bias tensor. Same as @p conv_weights + * @param[in] bn_beta (Optional) Batch normalization layer beta tensor. Same as @p conv_weights + * @param[in] bn_gamma (Optional) Batch normalization layer gamma tensor. Same as @p conv_weights + * @param[in] epsilon (Optional) Batch normalization layer epsilon parameter. Defaults to 0.001f. + * + * @return a status + */ + static Status validate(const ITensorInfo *conv_weights, const ITensorInfo *bn_mean, const ITensorInfo *bn_var, + const ITensorInfo *fused_weights, const ITensorInfo *fused_bias, + const ITensorInfo *conv_bias = nullptr, const ITensorInfo *bn_beta = nullptr, const ITensorInfo *bn_gamma = nullptr, + float epsilon = 0.001f); + + // Inherited methods overridden: + void run(const Window &window, cl::CommandQueue &queue) override; + +private: + const ICLTensor *_conv_weights; + const ICLTensor *_conv_bias; + const ICLTensor *_bn_mean; + const ICLTensor *_bn_var; + const ICLTensor *_bn_gamma; + const ICLTensor *_bn_beta; + ICLTensor *_fused_weights; + ICLTensor *_fused_bias; + float _epsilon; + bool _run_in_place_weights; + bool _run_in_place_bias; +}; +} // namespace arm_compute +#endif /*__ARM_COMPUTE_CLFUSEBATCHNORMALIZATIONKERNEL_H__ */ diff --git a/arm_compute/runtime/CL/CLFunctions.h b/arm_compute/runtime/CL/CLFunctions.h index 6a614f7704..014e25bbb3 100644 --- a/arm_compute/runtime/CL/CLFunctions.h +++ b/arm_compute/runtime/CL/CLFunctions.h @@ -65,6 +65,7 @@ #include "arm_compute/runtime/CL/functions/CLFlattenLayer.h" #include "arm_compute/runtime/CL/functions/CLFloor.h" #include "arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h" +#include "arm_compute/runtime/CL/functions/CLFuseBatchNormalization.h" #include "arm_compute/runtime/CL/functions/CLGEMM.h" #include "arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h" #include "arm_compute/runtime/CL/functions/CLGEMMInterleave4x4.h" diff --git a/arm_compute/runtime/CL/functions/CLFuseBatchNormalization.h b/arm_compute/runtime/CL/functions/CLFuseBatchNormalization.h new file mode 100644 index 0000000000..777a80f8b9 --- /dev/null +++ b/arm_compute/runtime/CL/functions/CLFuseBatchNormalization.h @@ -0,0 +1,93 @@ +/* + * Copyright (c) 2018 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef __ARM_COMPUTE_CLFUSEBATCHNORMALIZATION_H__ +#define __ARM_COMPUTE_CLFUSEBATCHNORMALIZATION_H__ + +#include "arm_compute/core/CL/kernels/CLFuseBatchNormalizationKernel.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/runtime/IFunction.h" + +namespace arm_compute +{ +// Forward declarations +class ICLTensor; + +/** Basic function to fuse the batch normalization node to a preceding convolution node */ +class CLFuseBatchNormalization : public IFunction +{ +public: + /** Default constructor */ + CLFuseBatchNormalization(); + /** Prevent instances of this class from being copied (As this class contains pointers) */ + CLFuseBatchNormalization(const CLFuseBatchNormalization &) = delete; + /** Prevent instances of this class from being copied (As this class contains pointers) */ + CLFuseBatchNormalization &operator=(const CLFuseBatchNormalization &) = delete; + /** Allow instances of this class to be moved */ + CLFuseBatchNormalization(CLFuseBatchNormalization &&) = default; + /** Allow instances of this class to be moved */ + CLFuseBatchNormalization &operator=(CLFuseBatchNormalization &&) = default; + /** Default destructor */ + ~CLFuseBatchNormalization() = default; + /** Set the input and output tensors. + * + * @param[in] conv_weights Convolution layer weights tensor. Data type supported: F16/F32 + * @param[in] bn_mean Batch normalization layer mean tensor. Same as @p conv_weights + * @param[in] bn_var Batch normalization layer variance tensor. Same as @p conv_weights + * @param[out] fused_weights Output fused weights tensor. Same as @p conv_weights + * @param[out] fused_bias Output fused bias tensor. Same as @p conv_weights + * @param[in] conv_bias (Optional) Convolution layer bias tensor. Same as @p conv_weights + * @param[in] bn_beta (Optional) Batch normalization layer beta tensor. Same as @p conv_weights + * @param[in] bn_gamma (Optional) Batch normalization layer gamma tensor. Same as @p conv_weights + * @param[in] epsilon (Optional) Batch normalization layer epsilon parameter. Defaults to 0.001f. + */ + void configure(const ICLTensor *conv_weights, const ICLTensor *bn_mean, const ICLTensor *bn_var, ICLTensor *fused_weights, ICLTensor *fused_bias, + const ICLTensor *conv_bias = nullptr, const ICLTensor *bn_beta = nullptr, const ICLTensor *bn_gamma = nullptr, + float epsilon = 0.001f); + /** Static function to check if given info will lead to a valid configuration of @ref CLFuseBatchNormalization + * + * @param[in] conv_weights Convolution layer weights tensor. Data type supported: F16/F32 + * @param[in] bn_mean Batch normalization layer mean tensor. Same as @p conv_weights + * @param[in] bn_var Batch normalization layer variance tensor. Same as @p conv_weights + * @param[in] fused_weights Output fused weights tensor. Same as @p conv_weights + * @param[in] fused_bias Output fused bias tensor. Same as @p conv_weights + * @param[in] conv_bias (Optional) Convolution layer bias tensor. Same as @p conv_weights + * @param[in] bn_beta (Optional) Batch normalization layer beta tensor. Same as @p conv_weights + * @param[in] bn_gamma (Optional) Batch normalization layer gamma tensor. Same as @p conv_weights + * @param[in] epsilon (Optional) Batch normalization layer epsilon parameter. Defaults to 0.001f. + * + * @return a status + */ + static Status validate(const ITensorInfo *conv_weights, const ITensorInfo *bn_mean, const ITensorInfo *bn_var, + const ITensorInfo *fused_weights, const ITensorInfo *fused_bias, + const ITensorInfo *conv_bias = nullptr, const ITensorInfo *bn_beta = nullptr, const ITensorInfo *bn_gamma = nullptr, + float epsilon = 0.001f); + + // Inherited methods overridden: + void run() override; + +private: + CLFuseBatchNormalizationKernel _fuse_bn_kernel; +}; +} // namespace arm_compute +#endif /*__ARM_COMPUTE_CLFUSEBATCHNORMALIZATION_H__ */ -- cgit v1.2.1