From df4cf57c7394265b27d051cb1cf0152c53659126 Mon Sep 17 00:00:00 2001 From: Michele Di Giorgio Date: Wed, 9 Oct 2019 15:32:39 +0100 Subject: COMPMID-2306: CLDepthwiseConvolution: support for QUANT8_PER_CHANNEL_SYMM Change-Id: I18c886400daa2dcba0b91011bc4e503d807a4732 Signed-off-by: Michele Di Giorgio Reviewed-on: https://review.mlplatform.org/c/2143 Comments-Addressed: Arm Jenkins Reviewed-by: Giorgio Arena Tested-by: Arm Jenkins --- .../ICLDepthwiseConvolutionLayer3x3Kernel.h | 33 ++++++++++++++-------- 1 file changed, 21 insertions(+), 12 deletions(-) (limited to 'arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h') diff --git a/arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h b/arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h index 92eca89fd8..a6b4510115 100644 --- a/arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h +++ b/arm_compute/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h @@ -37,7 +37,7 @@ class ICLDepthwiseConvolutionLayer3x3Kernel : public ICLKernel public: /** Default constructor */ ICLDepthwiseConvolutionLayer3x3Kernel() - : _border_size(0), _input(), _output(), _weights(), _biases(), _conv_stride_y(1) + : _border_size(0), _input(), _output(), _weights(), _biases(), _conv_stride_y(1), _output_multipliers(), _output_shifts(), _is_quantized(false) { } /** Prevent instances of this class from being copied (As this class contains pointers) */ @@ -50,18 +50,24 @@ public: ICLDepthwiseConvolutionLayer3x3Kernel &operator=(ICLDepthwiseConvolutionLayer3x3Kernel &&) = default; /** Initialize the function's source, destination, conv and border_size. * - * @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 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). + * @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, QASYMM8/QSYMM8_PER_CHANNEL when input is QASYMM8. + * @param[in] biases Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed. + * Data type supported: Same as @p input, S32 when input is QASYMM8. + * @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 for QASYMM8. + * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). + * @param[in] output_multipliers (Optional) Output multipliers tensor for quantized computations. In case of per-channel quantization, + * the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32 + * @param[in] output_shifts (Optional) Output shifts tensor for quantized computations. In case of per-channel quantization, + * the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32 */ - 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(), const Size2D &dilation = Size2D(1U, 1U)) = 0; + 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(), const Size2D &dilation = Size2D(1U, 1U), + const ICLTensor *output_multipliers = nullptr, const ICLTensor *output_shifts = nullptr) = 0; protected: BorderSize _border_size; @@ -70,6 +76,9 @@ protected: const ICLTensor *_weights; const ICLTensor *_biases; unsigned int _conv_stride_y; + const ICLTensor *_output_multipliers; + const ICLTensor *_output_shifts; + bool _is_quantized; }; } // namespace arm_compute #endif /*__ARM_COMPUTE_ICLDEPTHWISECONVOLUTIONKERNEL3x3_H__ */ -- cgit v1.2.1