diff options
author | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2018-08-13 11:20:41 +0100 |
---|---|---|
committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:54:54 +0000 |
commit | 916d1bcee42051721a82cfb46b52855c2fe56646 (patch) | |
tree | e3e38a8deddc558cabeda6fb7d14b2d45c8db2c4 /arm_compute/core | |
parent | 61de78aba1b405663c6620be15418873a2ee914a (diff) | |
download | ComputeLibrary-916d1bcee42051721a82cfb46b52855c2fe56646.tar.gz |
COMPMID-1498 - Enable grouping in CLGEMMConvolutionLayer
Change-Id: I15c7df21773145b03f42b6f78bd7ad2e5b8a5219
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/144126
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Diffstat (limited to 'arm_compute/core')
-rw-r--r-- | arm_compute/core/CL/kernels/CLCol2ImKernel.h | 5 | ||||
-rw-r--r-- | arm_compute/core/CL/kernels/CLIm2ColKernel.h | 4 | ||||
-rw-r--r-- | arm_compute/core/CL/kernels/CLWeightsReshapeKernel.h | 8 | ||||
-rw-r--r-- | arm_compute/core/NEON/kernels/NEIm2ColKernel.h | 4 | ||||
-rw-r--r-- | arm_compute/core/utils/misc/ShapeCalculator.h | 9 |
5 files changed, 15 insertions, 15 deletions
diff --git a/arm_compute/core/CL/kernels/CLCol2ImKernel.h b/arm_compute/core/CL/kernels/CLCol2ImKernel.h index 5c047ca091..2a18ae08c8 100644 --- a/arm_compute/core/CL/kernels/CLCol2ImKernel.h +++ b/arm_compute/core/CL/kernels/CLCol2ImKernel.h @@ -63,12 +63,11 @@ public: CLCol2ImKernel &operator=(CLCol2ImKernel &&) = default; /** Default destructor */ ~CLCol2ImKernel() = default; - /** Set the input and output of the kernel. * * @param[in] input The input tensor to convert. Data types supported: QASYMM8/F16/F32 * @param[out] output The output 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 + * while the rest represent batch of outputs. Data types supported: Same as @p input. Data layout: NCHW * @param[in] convolved_dims Output convolved dimensions. * @param[in] num_groups (Optional) Number of groups when performing a grouped convolution */ @@ -77,7 +76,7 @@ public: * * @param[in] input The input tensor to convert. Data types supported: QASYMM8/F16/F32 * @param[in] output The output 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 + * while the rest represent batch of outputs. Data types supported: Same as @p input. Data layout: NCHW * @param[in] convolved_dims Output convolved dimensions. * @param[in] num_groups (Optional) Number of groups when performing a grouped convolution * diff --git a/arm_compute/core/CL/kernels/CLIm2ColKernel.h b/arm_compute/core/CL/kernels/CLIm2ColKernel.h index c678f277cb..0647f5dcec 100644 --- a/arm_compute/core/CL/kernels/CLIm2ColKernel.h +++ b/arm_compute/core/CL/kernels/CLIm2ColKernel.h @@ -76,7 +76,7 @@ public: * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. * @param[in] has_bias In case biases are provided expands the matrix with 1. * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). - * @param[in] num_groups (Optional) Number of groups when performing a grouped convolution + * @param[in] num_groups (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is only supported for NCHW data layout */ void configure(const ICLTensor *input, ICLTensor *output, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const Size2D &dilation = Size2D(1U, 1U), unsigned int num_groups = 1); @@ -90,7 +90,7 @@ public: * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. * @param[in] has_bias In case biases are provided expands the matrix with 1. * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). - * @param[in] num_groups (Optional) Number of groups when performing a grouped convolution + * @param[in] num_groups (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is only supported for NCHW data layout * * @return a status */ diff --git a/arm_compute/core/CL/kernels/CLWeightsReshapeKernel.h b/arm_compute/core/CL/kernels/CLWeightsReshapeKernel.h index d3bbbaf470..bdc5792641 100644 --- a/arm_compute/core/CL/kernels/CLWeightsReshapeKernel.h +++ b/arm_compute/core/CL/kernels/CLWeightsReshapeKernel.h @@ -75,10 +75,10 @@ public: * @warning Appending biases to weights reshaped matrix is not supported for quantized asymmetric types. * @param[out] output The output tensor. Should be a 2D Tensor if there are no groups and the weights are not shared; a 3D Tensor otherwise. * Data types supported: Same as @p input - * @param[in] num_groups (Optional) Number of groups when performing a grouped convolution. + * @param[in] num_groups (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is only supported for NCHW data layout * Number of groups greater than one are only supported for NCHW data layout, and the number of weights must be a multiple of it. */ - void configure(const ICLTensor *input, const ICLTensor *biases, ICLTensor *output, const unsigned int num_groups = 1); + void configure(const ICLTensor *input, const ICLTensor *biases, ICLTensor *output, unsigned int num_groups = 1); /** Static function to check if given info will lead to a valid configuration of @ref CLWeightsReshapeKernel * * @param[in] input The input tensor to convert. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM] if shared, @@ -88,12 +88,12 @@ public: * @warning Appending biases to weights reshaped matrix is not supported for quantized asymmetric types. * @param[in] output The output tensor. Should be a 2D Tensor if there are no groups and the weights are not shared; a 3D Tensor otherwise. * Data types supported: Same as @p input - * @param[in] num_groups (Optional) Number of groups when performing a grouped convolution. + * @param[in] num_groups (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is only supported for NCHW data layout * Number of groups greater than one are only supported for NCHW data layout, and the number of weights must be a multiple of it. * * @return a status */ - static Status validate(const ITensorInfo *input, const ITensorInfo *biases, const ITensorInfo *output, const unsigned int num_groups = 1); + static Status validate(const ITensorInfo *input, const ITensorInfo *biases, const ITensorInfo *output, unsigned int num_groups = 1); // Inherited methods overridden: void run(const Window &window, cl::CommandQueue &queue) override; diff --git a/arm_compute/core/NEON/kernels/NEIm2ColKernel.h b/arm_compute/core/NEON/kernels/NEIm2ColKernel.h index 37145a38c1..38bdff0e98 100644 --- a/arm_compute/core/NEON/kernels/NEIm2ColKernel.h +++ b/arm_compute/core/NEON/kernels/NEIm2ColKernel.h @@ -84,7 +84,7 @@ public: * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. * @param[in] has_bias In case biases are provided expands the matrix with 1. * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). - * @param[in] num_groups (Optional) Number of groups when performing a grouped convolution + * @param[in] num_groups (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is not supported * @param[in] is_fully_connected (Optional) Determines whether this kernel will be called by @ref NEFullyConnectedLayer in order to validate the arguments * @param[in] is_flatten (Optional) Determines whether this kernel will be called by @ref NEFlattenLayer in order to validate the arguments */ @@ -100,7 +100,7 @@ public: * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. * @param[in] has_bias In case biases are provided expands the matrix with 1. * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). - * @param[in] num_groups (Optional) Number of groups when performing a grouped convolution + * @param[in] num_groups (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is not supported * @param[in] is_fully_connected (Optional)Determines whether this kernel will be called by @ref NEFullyConnectedLayer in order to validate the arguments * @param[in] is_flatten (Optional) Determines whether this kernel will be called by @ref NEFlattenLayer in order to validate the arguments * diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h index 8a00c22306..c40e7119b2 100644 --- a/arm_compute/core/utils/misc/ShapeCalculator.h +++ b/arm_compute/core/utils/misc/ShapeCalculator.h @@ -55,19 +55,20 @@ inline TensorShape compute_permutation_output_shape(const ITensorInfo &input, co permute(output_shape, perm); return output_shape; } -inline TensorShape compute_weights_reshaped_shape(const ITensorInfo &weights, bool has_bias = false, const unsigned int num_groups = 1) +inline TensorShape compute_weights_reshaped_shape(const ITensorInfo &weights, bool has_bias = false, unsigned int num_groups = 1) { // Number of groups greater than one are only supported for NCHW data layout, and the number of weights must be a multiple of it. - ARM_COMPUTE_ERROR_ON(num_groups == 0); - ARM_COMPUTE_ERROR_ON((weights.dimension(3) % num_groups) != 0); ARM_COMPUTE_ERROR_ON(weights.data_layout() == DataLayout::NHWC && num_groups > 1); + ARM_COMPUTE_ERROR_ON((weights.dimension(3) % num_groups) != 0); // Calculate output shape TensorShape weights_reshaped{ weights.tensor_shape() }; + weights_reshaped.set(3, weights_reshaped[3] / num_groups); + weights_reshaped.collapse(3); const size_t tmp_dim = weights_reshaped[0]; - weights_reshaped.set(0, weights_reshaped[1] / num_groups); + weights_reshaped.set(0, weights_reshaped[1]); weights_reshaped.set(1, tmp_dim + (has_bias ? 1 : 0)); if(weights.num_dimensions() < 5) { |