diff options
Diffstat (limited to 'arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h')
-rw-r--r-- | arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h | 205 |
1 files changed, 90 insertions, 115 deletions
diff --git a/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h index 811dc82843..6ad5aa7bfa 100644 --- a/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h +++ b/arm_compute/runtime/NEON/functions/NEDepthwiseConvolutionLayer.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 ARM Limited. + * Copyright (c) 2017-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -24,18 +24,18 @@ #ifndef ARM_COMPUTE_NEDEPTHWISECONVOLUTION_H #define ARM_COMPUTE_NEDEPTHWISECONVOLUTION_H -#include "arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayer3x3Kernel.h" -#include "arm_compute/core/NEON/kernels/NEDepthwiseConvolutionLayerNativeKernel.h" -#include "arm_compute/core/NEON/kernels/NEDirectConvolutionLayerOutputStageKernel.h" -#include "arm_compute/core/NEON/kernels/NEFillBorderKernel.h" +#include "arm_compute/runtime/IMemoryManager.h" +#include "arm_compute/runtime/MemoryGroup.h" #include "arm_compute/runtime/NEON/functions/NEActivationLayer.h" #include "arm_compute/runtime/NEON/functions/NEPermute.h" -#include "arm_compute/runtime/NEON/functions/assembly/NEDepthwiseConvolutionAssemblyDispatch.h" + +#include <memory> namespace arm_compute { // Forward declarations class ITensor; +class NEDepthwiseConvolutionLayerNativeKernel; /** Function to execute a depthwise convolution. */ @@ -52,8 +52,24 @@ public: NEDepthwiseConvolutionLayer &operator=(const NEDepthwiseConvolutionLayer &) = delete; /** Default move assignment operator */ NEDepthwiseConvolutionLayer &operator=(NEDepthwiseConvolutionLayer &&) = default; + /** Default destructor */ + ~NEDepthwiseConvolutionLayer(); /** Initialize the function's source, destination, weights and convolution information. * + * Valid data layouts: + * - NHWC + * - NCHW + * + * Valid data type configurations: + * |src0 |src1 |src2 |dst | + * |:--------------|:------------------|:------|:--------------| + * |F16 |F16 |F16 |F16 | + * |F32 |F32 |F32 |F32 | + * |QASYMM8 |QASYMM8 |S32 |QASYMM8 | + * |QASYMM8 |QSYMM8_PER_CHANNEL |S32 |QASYMM8 | + * |QASYMM8_SIGNED |QASYMM8_SIGNED |S32 |QASYMM8_SIGNED | + * |QASYMM8_SIGNED |QSYMM8_PER_CHANNEL |S32 |QASYMM8_SIGNED | + * * @param[in, out] input Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32 * @param[out] output Destination tensor. Data type supported: same as @p input. * @param[in] weights Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. @@ -65,8 +81,14 @@ public: * @param[in] act_info (Optional) Activation layer information in case of a fused activation. * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). */ - void configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, - unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U)); + void configure(ITensor *input, + const ITensor *weights, + const ITensor *biases, + ITensor *output, + const PadStrideInfo &conv_info, + unsigned int depth_multiplier = 1, + const ActivationLayerInfo &act_info = ActivationLayerInfo(), + const Size2D &dilation = Size2D(1U, 1U)); /** Static function to check if given info will lead to a valid configuration of @ref NEDepthwiseConvolutionLayer * @@ -83,40 +105,27 @@ public: * * @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 = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U)); + static Status validate(const ITensorInfo *input, + const ITensorInfo *weights, + const ITensorInfo *biases, + const ITensorInfo *output, + const PadStrideInfo &conv_info, + unsigned int depth_multiplier = 1, + const ActivationLayerInfo &act_info = ActivationLayerInfo(), + const Size2D &dilation = Size2D(1U, 1U)); // Inherited methods overriden: void run() override; void prepare() override; private: - /** Static function to choose the best depthwise convolution function for @ref NEDepthwiseConvolutionLayer - * - * @param[in] input Source tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32 - * @param[in] weights Weights tensor info. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. - * Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8/QASYMM8_SIGNED. - * @param[in] biases Biases tensor info. A 1D tensor with shape [IFM]. Must be nullptr if not needed. - * Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED. - * @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] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU for 3x3 quantized are supported. - * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). - * - * @return a Depthwise Convolution Function - */ - static DepthwiseConvolutionFunction get_depthwiseconvolution_function(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, - const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1, - ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U)); - - /** Basic function to execute optimized depthwise convolution routines. This function calls the following NEON kernels: + /** Basic function to execute optimized depthwise convolution routines. This function calls the following kernels: * * @note At the moment 3x3 and 5x5 convolution of stride 1, 2 are supported * * -# @ref NEFillBorderKernel (if pad_x or pad_y > 0) and no assembly kernel implementation is present * -# @ref NEDepthwiseConvolutionLayer3x3Kernel if 3x3 and no assembly kernel implementation is present - * -# @ref NEDepthwiseConvolutionAssemblyDispatch if assembly kernel implementation is present + * -# @ref cpu::CpuDepthwiseConvolutionAssemblyDispatch if assembly kernel implementation is present * -# @ref NEDirectConvolutionLayerOutputStageKernel if re-quantization of output is required * -# @ref NEActivationLayer if fused activation is required * @@ -131,9 +140,13 @@ private: /** Default move constructor */ NEDepthwiseConvolutionLayerOptimizedInternal(NEDepthwiseConvolutionLayerOptimizedInternal &&) = default; /** Prevent instances of this class from being copied (As this class contains pointers) */ - NEDepthwiseConvolutionLayerOptimizedInternal &operator=(const NEDepthwiseConvolutionLayerOptimizedInternal &) = delete; + NEDepthwiseConvolutionLayerOptimizedInternal & + operator=(const NEDepthwiseConvolutionLayerOptimizedInternal &) = delete; /** Default move assignment operator */ - NEDepthwiseConvolutionLayerOptimizedInternal &operator=(NEDepthwiseConvolutionLayerOptimizedInternal &&) = default; + NEDepthwiseConvolutionLayerOptimizedInternal & + operator=(NEDepthwiseConvolutionLayerOptimizedInternal &&) = default; + /** Default destructor */ + ~NEDepthwiseConvolutionLayerOptimizedInternal() = default; /** Initialize the function's source, destination, kernels and border_size. * * @param[in, out] input Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32. (Written to only for border filling). @@ -146,8 +159,14 @@ private: * @param[in] act_info (Optional) Activation layer information in case of a fused activation. * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). */ - void configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, - unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U)); + void configure(ITensor *input, + const ITensor *weights, + const ITensor *biases, + ITensor *output, + const PadStrideInfo &conv_info, + unsigned int depth_multiplier = 1, + const ActivationLayerInfo &act_info = ActivationLayerInfo(), + const Size2D &dilation = Size2D(1U, 1U)); /** Static function to check if given info will lead to a valid configuration of @ref NEDepthwiseConvolutionLayer3x3 * @@ -163,71 +182,26 @@ private: * * @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 = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U)); + static Status validate(const ITensorInfo *input, + const ITensorInfo *weights, + const ITensorInfo *biases, + const ITensorInfo *output, + const PadStrideInfo &conv_info, + unsigned int depth_multiplier = 1, + const ActivationLayerInfo &act_info = ActivationLayerInfo(), + const Size2D &dilation = Size2D(1U, 1U)); // Inherited methods overriden: void run() override; void prepare() override; private: - /** Configure the kernels/functions for the generic pipeline. - * - * @param[in, out] input Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32. (Written to only for border filling). - * @param[in] weights Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input. - * @param[in] biases Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed. - * Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED. - * @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 Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. - * @param[in] act_info Activation layer information in case of a fused activation. - * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). - * - */ - void configure_generic(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, - unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation = Size2D(1U, 1U)); - /** Configure the kernels/functions for the optimized pipeline. - * - * @param[in] input Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32. (Written to only for border filling). - * @param[in] weights Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. Data type supported: Same as @p input. - * @param[in] biases Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed. - * Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED. - * @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 Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. - * @param[in] act_info Activation layer information in case of a fused activation. - */ - void configure_optimized(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, - unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation = Size2D(1U, 1U)); - /** Run generic kernel */ - void run_generic(); - /** Run optimized function */ - void run_optimized(); - - MemoryGroup _memory_group; - NEDepthwiseConvolutionLayer3x3Kernel _dwc_kernel; - NEDepthwiseConvolutionAssemblyDispatch _dwc_optimized_func; - NEDirectConvolutionLayerOutputStageKernel _output_stage_kernel; - NEFillBorderKernel _border_handler; - NEPermute _permute_input; - NEPermute _permute_weights; - NEPermute _permute_output; - NEActivationLayer _activationlayer_function; - Tensor _accumulator; - Tensor _permuted_input; - Tensor _permuted_weights; - Tensor _permuted_output; - const ITensor *_original_weights; - bool _has_bias; - bool _is_quantized; - bool _is_optimized; - bool _is_nchw; - bool _permute; - bool _is_activationlayer_enabled; - bool _is_prepared; + MemoryGroup _memory_group; + struct Impl; + std::unique_ptr<Impl> _impl; }; - /** Basic function to execute a generic depthwise convolution. This function calls the following NEON kernel: + /** Basic function to execute a generic depthwise convolution. This function calls the following kernel: * * -# @ref NEDepthwiseConvolutionLayerNativeKernel * @@ -245,6 +219,8 @@ private: NEDepthwiseConvolutionLayerGeneric &operator=(const NEDepthwiseConvolutionLayerGeneric &) = delete; /** Default move assignment operator */ NEDepthwiseConvolutionLayerGeneric &operator=(NEDepthwiseConvolutionLayerGeneric &&) = default; + /** Default destructor */ + ~NEDepthwiseConvolutionLayerGeneric() = default; /** Initialize the function's source, destination, weights and convolution information. * * @param[in, out] input Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32. (Written to only for border filling). @@ -258,8 +234,14 @@ private: * @param[in] act_info (Optional) Activation layer information in case of a fused activation. * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). */ - void configure(ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, - unsigned int depth_multiplier = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U)); + void configure(ITensor *input, + const ITensor *weights, + const ITensor *biases, + ITensor *output, + const PadStrideInfo &conv_info, + unsigned int depth_multiplier = 1, + const ActivationLayerInfo &act_info = ActivationLayerInfo(), + const Size2D &dilation = Size2D(1U, 1U)); /** Static function to check if given info will lead to a valid configuration of @ref NEDepthwiseConvolutionLayerGeneric * @@ -276,32 +258,25 @@ private: * * @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 = 1, const ActivationLayerInfo &act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U)); + static Status validate(const ITensorInfo *input, + const ITensorInfo *weights, + const ITensorInfo *biases, + const ITensorInfo *output, + const PadStrideInfo &conv_info, + unsigned int depth_multiplier = 1, + const ActivationLayerInfo &act_info = ActivationLayerInfo(), + const Size2D &dilation = Size2D(1U, 1U)); // Inherited methods overriden: void run() override; - void prepare() override; private: - NEDepthwiseConvolutionLayerNativeKernel _depthwise_conv_kernel; - NEFillBorderKernel _fill_border; - NEPermute _permute_input; - NEPermute _permute_weights; - NEPermute _permute_output; - NEActivationLayer _activationlayer_function; - Tensor _permuted_input; - Tensor _permuted_weights; - Tensor _permuted_output; - bool _is_prepared; - bool _is_nchw; - bool _is_activationlayer_enabled; - const ITensor *_original_weights; + struct Impl; + std::unique_ptr<Impl> _impl; }; - - DepthwiseConvolutionFunction _depth_conv_func; - NEDepthwiseConvolutionLayerOptimizedInternal _func_optimized; - NEDepthwiseConvolutionLayerGeneric _func_generic; + MemoryGroup _memory_group; + struct Impl; + std::unique_ptr<Impl> _impl; }; } // namespace arm_compute -#endif /* ARM_COMPUTE_NEDEPTHWISECONVOLUTION_H */
\ No newline at end of file +#endif /* ARM_COMPUTE_NEDEPTHWISECONVOLUTION_H */ |