From e73686ac797be2d19cd9bed26d690e1431e3d848 Mon Sep 17 00:00:00 2001 From: Usama Arif Date: Mon, 8 Apr 2019 17:30:48 +0100 Subject: COMPMID-2047: Add support for dilation in CLDepthwiseConvolution. Change-Id: I3106aa34bd168985a56791613d95072756be6e9b Signed-off-by: Usama Arif Reviewed-on: https://review.mlplatform.org/c/958 Comments-Addressed: Arm Jenkins Reviewed-by: Pablo Marquez Tested-by: Arm Jenkins --- .../CLDepthwiseConvolutionLayer3x3NCHWKernel.h | 18 +++++----- .../CLDepthwiseConvolutionLayer3x3NHWCKernel.h | 10 +++--- .../core/CL/kernels/CLDepthwiseIm2ColKernel.h | 16 +++++---- .../ICLDepthwiseConvolutionLayer3x3Kernel.h | 7 ++-- arm_compute/core/utils/misc/ShapeCalculator.h | 42 +++++----------------- 5 files changed, 38 insertions(+), 55 deletions(-) (limited to 'arm_compute/core') diff --git a/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h b/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h index b1c730d9a7..3b7fc7b7dc 100644 --- a/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h +++ b/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 ARM Limited. + * Copyright (c) 2018-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -41,31 +41,33 @@ public: * * @param[in] input Source tensor. DataType supported: QASYMM8/F16/F32. * @param[in] weights Weights tensor. A 3D tensor with dimensions [3, 3, IFM]. Data type supported: Same as @p input. - * @param[in] biases (Optional) Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed. + * @param[in] biases Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed. * Data type supported: Same as @p input. * @param[out] output Destination tensor. Data type supported: Same as @p input. * @param[in] conv_info Padding and stride information to use for the convolution. * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for QASYMM8 supported. + * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). */ void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier, - ActivationLayerInfo act_info) override; + ActivationLayerInfo act_info, const Size2D &dilation) override; /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer3x3NCHWKernel * - * @param[in] input Source tensor. DataType supported: F16/F32/QASYMM8. - * @param[in] weights Weights tensor. A 3D tensor with dimensions [3, 3, IFM]. Data type supported: Same as @p input. - * @param[in] biases Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed. + * @param[in] input Source tensor info. DataType supported: F16/F32/QASYMM8. + * @param[in] weights Weights tensor info. A 3D tensor with dimensions [3, 3, IFM]. Data type supported: Same as @p input. + * @param[in] biases Biases tensor info. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed. * Data type supported: Same as @p input. * @param[in] output Destination tensor. Data type supported: Same as @p input. * @param[in] conv_info Padding and stride information to use for the convolution. - * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. + * @param[in] depth_multiplier Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU are supported. * @param[in] gpu_target (Optional) GPU target to validate the kernel for. Defaults to midgard. + * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). * * @return a status */ static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier, - ActivationLayerInfo act_info = ActivationLayerInfo(), GPUTarget gpu_target = GPUTarget::MIDGARD); + ActivationLayerInfo act_info = ActivationLayerInfo(), GPUTarget gpu_target = GPUTarget::MIDGARD, const Size2D &dilation = Size2D(1U, 1U)); void run(const Window &window, cl::CommandQueue &queue) override; BorderSize border_size() const override; diff --git a/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h b/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h index 2fc9780a2f..7d0ecec13e 100644 --- a/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h +++ b/arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h @@ -42,30 +42,32 @@ public: * * @param[in] input Source tensor. DataType supported: QASYMM8. * @param[in] weights Weights tensor. A 3D tensor with dimensions [IFM, 3, 3]. Data type supported: Same as @p input. - * @param[in] biases (Optional) Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed. + * @param[in] biases Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed. * Data type supported: Same as @p input. * @param[out] output Destination tensor. Data type supported: Same as @p input. * @param[in] conv_info Padding and stride information to use for the convolution. * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU are supported. + * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). */ void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier, - ActivationLayerInfo act_info) override; + ActivationLayerInfo act_info, const Size2D &dilation) override; /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer3x3NHWCKernel * * @param[in] input Source tensor info. DataType supported: QASYMM8. * @param[in] weights Weights tensor info. A 3D tensor with dimensions [IFM, 3, 3]. Data type supported: Same as @p input. - * @param[in] biases (Optional) Biases tensor info. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed. + * @param[in] biases Biases tensor info. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed. * Data type supported: Same as @p input. * @param[in] output Destination tensor info. Data type supported: Same as @p input. * @param[in] conv_info Padding and stride information to use for the convolution. * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU are supported. + * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). * * @return a status */ static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier, - ActivationLayerInfo act_info = ActivationLayerInfo()); + ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U)); // Inherited methods overridden: void run(const Window &window, cl::CommandQueue &queue) override; diff --git a/arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h b/arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h index 00d9cb64e1..15798471a8 100644 --- a/arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h +++ b/arm_compute/core/CL/kernels/CLDepthwiseIm2ColKernel.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -58,22 +58,26 @@ public: * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. * @param[in] has_bias Boolean that specifies if the depthwise convolution has bias. * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. + * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). */ - void configure(const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias = false, unsigned int depth_multiplier = 1); + void configure(const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias = false, unsigned int depth_multiplier = 1, + const Size2D &dilation = Size2D(1U, 1U)); /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseIm2ColKernel * - * @param[in] input The input tensor to convert. 3 lower dimensions represent a single input [width, height, IFM], + * @param[in] input The input tensor info to convert. 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: QASYMM8/F32 - * @param[in] output The output tensor. First 3 lower dimensions represent a transform of each 3D input, + * @param[in] output The output tensor info. First 3 lower dimensions represent a transform of each 3D input, * while every dimension above 3 represents a batch. Data types supported: Same as @p input * @param[in] kernel_dims The kernel dimensions (width and height). * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. * @param[in] has_bias Boolean that specifies if the depthwise convolution has bias. - * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. + * @param[in] depth_multiplier Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. + * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). * * @return a status */ - static Status validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier); + static Status validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int depth_multiplier, + const Size2D &dilation = Size2D(1U, 1U)); // Inherited methods overridden: void run(const Window &window, cl::CommandQueue &queue) override; diff --git a/arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h b/arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h index 3396de2e46..92eca89fd8 100644 --- a/arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h +++ b/arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -52,15 +52,16 @@ public: * * @param[in] input Source tensor. DataType supported: QASYMM8/F16/F32. * @param[in] weights Weights tensor. A 3D tensor with dimensions [3, 3, IFM]. Data type supported: Same as @p input. - * @param[in] biases (Optional) Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed. + * @param[in] biases Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed. * Data type supported: Same as @p input. * @param[out] output Destination tensor. Data type supported: Same as @p input. * @param[in] conv_info Padding and stride information to use for the convolution. * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for QASYMM8 supported. + * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). */ virtual void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1, - ActivationLayerInfo act_info = ActivationLayerInfo()) = 0; + ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U)) = 0; protected: BorderSize _border_size; diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h index 26b337d5c5..b46b1b2535 100644 --- a/arm_compute/core/utils/misc/ShapeCalculator.h +++ b/arm_compute/core/utils/misc/ShapeCalculator.h @@ -402,10 +402,12 @@ inline TensorShape compute_transposed_shape(const ITensorInfo &input) * @param[in] weights Weights tensor info * @param[in] conv_info Padding and stride information to use for the convolution. * @param[in] depth_multiplier Multiplier to apply to the input's depth in order to retrieve the output's depth. + * @param[in] dilation Dilation, in elements, across x and y. Defaults to (1, 1). * * @return the calculated shape */ -inline TensorShape compute_depthwise_convolution_shape(const ITensorInfo &input, const ITensorInfo &weights, PadStrideInfo conv_info, unsigned int depth_multiplier) +inline TensorShape compute_depthwise_convolution_shape(const ITensorInfo &input, const ITensorInfo &weights, PadStrideInfo conv_info, unsigned int depth_multiplier, const Size2D &dilation = Size2D(1U, + 1U)) { const TensorShape input_shape{ input.tensor_shape() }; const TensorShape weights_shape{ weights.tensor_shape() }; @@ -415,43 +417,15 @@ inline TensorShape compute_depthwise_convolution_shape(const ITensorInfo &input, const int height_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); const int channel_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL); - unsigned int output_width = 0; - unsigned int output_height = 0; - std::tie(output_width, output_height) = scaled_dimensions(input_shape[width_idx], input_shape[height_idx], - weights_shape[width_idx], weights_shape[height_idx], - conv_info); - - TensorShape output_shape{ input_shape }; - output_shape.set(width_idx, output_width); - output_shape.set(height_idx, output_height); - output_shape.set(channel_idx, input_shape[channel_idx] * depth_multiplier); - - return output_shape; -} - -/** Calculate the depthwise convolution output shape of a tensor - * - * @param[in] input Input tensor info - * @param[in] weights_width Weights width - * @param[in] weights_height Weights height - * @param[in] conv_info Padding and stride information to use for the convolution. - * @param[in] depth_multiplier Multiplier to apply to the input's depth in order to retrieve the output's depth. - * - * @return the calculated shape - */ -inline TensorShape compute_depthwise_convolution_shape(const ITensorInfo &input, int weights_width, int weights_height, PadStrideInfo conv_info, unsigned int depth_multiplier) -{ - const TensorShape input_shape{ input.tensor_shape() }; - - const DataLayout data_layout = input.data_layout(); - const int width_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); - const int height_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); - const int channel_idx = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL); + const DataLayout weights_data_layout = weights.data_layout(); + const int weights_width_idx = get_data_layout_dimension_index(weights_data_layout, DataLayoutDimension::WIDTH); + const int weights_height_idx = get_data_layout_dimension_index(weights_data_layout, DataLayoutDimension::HEIGHT); unsigned int output_width = 0; unsigned int output_height = 0; std::tie(output_width, output_height) = scaled_dimensions(input_shape[width_idx], input_shape[height_idx], - weights_width, weights_width, conv_info); + weights_shape[weights_width_idx], weights_shape[weights_height_idx], + conv_info, dilation); TensorShape output_shape{ input_shape }; output_shape.set(width_idx, output_width); -- cgit v1.2.1