From 1b14c75c0d591c4abe4d2d41b7e4e165fbf58382 Mon Sep 17 00:00:00 2001 From: Sheri Zhang Date: Mon, 9 Mar 2020 14:29:52 +0000 Subject: COMPMID-2968: Add support for QASYMM8_SIGNED in CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel Signed-off-by: Sheri Zhang Change-Id: I37e6e76dbd5546c0eaedfacd01ea905c37148e8a Signed-off-by: Sheri Zhang Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/2861 Comments-Addressed: Arm Jenkins Reviewed-by: Michele Di Giorgio Tested-by: Arm Jenkins --- arm_compute/core/CL/CLKernels.h | 2 +- ...CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel.h | 91 +++++++++++++++++++++ ...owpQuantizeDownInt32ToUint8ScaleByFloatKernel.h | 95 ---------------------- arm_compute/core/Types.h | 1 + 4 files changed, 93 insertions(+), 96 deletions(-) create mode 100644 arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel.h delete mode 100644 arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.h (limited to 'arm_compute/core') diff --git a/arm_compute/core/CL/CLKernels.h b/arm_compute/core/CL/CLKernels.h index f2e16ca139..b265aa2fe7 100644 --- a/arm_compute/core/CL/CLKernels.h +++ b/arm_compute/core/CL/CLKernels.h @@ -79,11 +79,11 @@ #include "arm_compute/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.h" #include "arm_compute/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.h" #include "arm_compute/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.h" +#include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel.h" #include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ScaleKernel.h" #include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.h" #include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h" #include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.h" -#include "arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.h" #include "arm_compute/core/CL/kernels/CLGEMMLowpReductionKernel.h" #include "arm_compute/core/CL/kernels/CLGEMMMatrixAccumulateBiasesKernel.h" #include "arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h" diff --git a/arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel.h b/arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel.h new file mode 100644 index 0000000000..439f569d07 --- /dev/null +++ b/arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel.h @@ -0,0 +1,91 @@ +/* + * Copyright (c) 2020 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_CLGEMMLOWPQUANTIZEDOWNINT32SCALEBYFLOATKERNEL_H +#define ARM_COMPUTE_CLGEMMLOWPQUANTIZEDOWNINT32SCALEBYFLOATKERNEL_H + +#include "arm_compute/core/CL/ICLKernel.h" + +namespace arm_compute +{ +// Forward declarations +class ICLTensor; + +/** OpenCL kernel used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8/QASYMM8_SIGNED + * + * This kernel takes a final int32 accumulator value (the output of @ref CLGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QASYMM8/QASYMM8_SIGNED value. + * The following computations will be performed by the kernel: + * + * -# Compute fixed point multiplication between each entry of input by result_fixedpoint_multiplier + * -# Add bias to final result if bias tensor is not a nullptr + * -# Requantize + * -# Add offset to each result + * -# Clamp the value between the specified min and max bounds + * -# Clamp the resulting int32 values to + * - to the [0..255] range and cast to QASYMM8. + * - to the [-128..127] range and cast to QASYMM8_SIGNED. + */ +class CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel : public ICLKernel +{ +public: + /** Constructor */ + CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel(); + /** Prevent instances of this class from being copied (As this class contains pointers)*/ + CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel(const CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel &) = delete; + /** Prevent instances of this class from being copied (As this class contains pointers)*/ + CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel &operator=(const CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel &) = delete; + /** Allow instances of this class to be moved */ + CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel(CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel &&) = default; + /** Allow instances of this class to be moved */ + CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel &operator=(CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel &&) = default; + /** Initialise the kernel's input and output. + * + * @param[in] input Input tensor. Data type supported: S32 + * @param[in] bias Biases tensor. Only shared biases supported and it can be a nullptr if the biases addition is not required. + * Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input. + * @param[out] output Output tensor. Data type supported: Data type supported: QASYMM8/QASYMM8_SIGNED + * @param[in] info Output stage info. Used to pass the quantized output data type + */ + void configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, const GEMMLowpOutputStageInfo *info); + /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMLowpQuantizeDownInt32ScaleByFloatKernel + * + * @param[in] input Input tensor. Data type supported: S32 + * @param[in] bias Biases tensor. Only shared biases supported and it can be a nullptr if the biases addition is not required. + * Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input. + * @param[in] output Output tensor. Data type supported: Data type supported: QASYMM8/QASYMM8_SIGNED + * @param[in] info Output stage info. Used to pass the quantized output data type + * + * @return a status + */ + static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const GEMMLowpOutputStageInfo *info); + + // Inherited methods overridden: + void run(const Window &window, cl::CommandQueue &queue) override; + +private: + const ICLTensor *_input; + const ICLTensor *_bias; + ICLTensor *_output; +}; +} // namespace arm_compute +#endif /* ARM_COMPUTE_CLGEMMLOWPQUANTIZEDOWNINT32SCALEBYFLOATKERNEL_H */ diff --git a/arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.h b/arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.h deleted file mode 100644 index 900a8c3b5d..0000000000 --- a/arm_compute/core/CL/kernels/CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel.h +++ /dev/null @@ -1,95 +0,0 @@ -/* - * Copyright (c) 2018-2019 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_CLGEMMLOWPQUANTIZEDOWNINT32TOUINT8SCALEBYFLOATKERNEL_H -#define ARM_COMPUTE_CLGEMMLOWPQUANTIZEDOWNINT32TOUINT8SCALEBYFLOATKERNEL_H - -#include "arm_compute/core/CL/ICLKernel.h" - -namespace arm_compute -{ -// Forward declarations -class ICLTensor; - -/** OpenCL kernel used to quantize down the int32 accumulator values of GEMMLowp to QASYMM8 - * - * This kernel takes a final int32 accumulator value (the output of @ref CLGEMMLowpMatrixMultiplyKernel), and processes it to obtain the final QASYMM8 value. - * The following computations will be performed by the kernel: - * - * -# Compute fixed point multiplication between each entry of input by result_fixedpoint_multiplier - * -# Add bias to final result if bias tensor is not a nullptr - * -# Requantize - * -# Add offset to each result - * -# Clamp the value between the specified min and max bounds - * -# Clamp the resulting int32 values to the [0..255] range and cast to QASYMM8. - */ -class CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel : public ICLKernel -{ -public: - /** Constructor */ - CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel(); - /** Prevent instances of this class from being copied (As this class contains pointers)*/ - CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel(const CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel &) = delete; - /** Prevent instances of this class from being copied (As this class contains pointers)*/ - CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel &operator=(const CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel &) = delete; - /** Allow instances of this class to be moved */ - CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel(CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel &&) = default; - /** Allow instances of this class to be moved */ - CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel &operator=(CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFloatKernel &&) = default; - /** Initialise the kernel's input and output. - * - * @param[in] input Input tensor. Data type supported: S32 - * @param[in] bias Biases tensor. Only shared biases supported and it can be a nullptr if the biases addition is not required. - * Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input. - * @param[out] output Output tensor. Data type supported: Data type supported: QASYMM8 - * @param[in] multiplier Float multiplier to be multiplied to each element of the input matrix - * @param[in] offset Offset to be applied to result before converting it back to QASYMM8 - * @param[in] min (Optional) Min value used to saturate down the output result before converting back to QASYMM8 - * @param[in] max (Optional) Max value used to saturate up the output result before converting back to QASYMM8, - * Along with @p min, this value can be used to implement "rectified linear unit" activation functions - */ - void configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, float multiplier, int offset, int min = 0, int max = 0); - /** Static function to check if given info will lead to a valid configuration of @ref CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel - * - * @param[in] input Input tensor. Data type supported: S32 - * @param[in] bias Biases tensor. Only shared biases supported and it can be a nullptr if the biases addition is not required. - * Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input. - * @param[in] output Output tensor. Data type supported: Data type supported: QASYMM8 - * @param[in] min (Optional) Min value used to saturate down the output result before converting back to QASYMM8 - * @param[in] max (Optional) Max value used to saturate up the output result before converting back to QASYMM8, - * Along with @p min, this value can be used to implement "rectified linear unit" activation functions - * - * @return a status - */ - static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min = 0, int max = 0); - - // Inherited methods overridden: - void run(const Window &window, cl::CommandQueue &queue) override; - -private: - const ICLTensor *_input; - const ICLTensor *_bias; - ICLTensor *_output; -}; -} // namespace arm_compute -#endif /* ARM_COMPUTE_CLGEMMLOWPQUANTIZEDOWNINT32TOUINT8SCALEBYFLOATKERNEL_H */ diff --git a/arm_compute/core/Types.h b/arm_compute/core/Types.h index 711b68f236..37a9679a21 100644 --- a/arm_compute/core/Types.h +++ b/arm_compute/core/Types.h @@ -1956,6 +1956,7 @@ struct GEMMLowpOutputStageInfo int32_t gemmlowp_max_bound{ std::numeric_limits::max() }; /**< GEMMLowp max value used to saturate down the output result before converting back to QASYMM8 */ std::vector gemmlowp_multipliers{}; /**< GEMMLowp output stage multiplier used for quantizing to QASYMM8 */ std::vector gemmlowp_shifts{}; /**< GEMMLowp output stage multiplier used for quantizing to QASYMM8 */ + float gemmlowp_real_multiplier{ 0 }; /**< GEMMLowp output stage real multiplier used for quantizing to QASYMM8 */ bool is_quantized_per_channel{ false }; /**< GEMMLowp quantized per-channel flag */ DataType output_data_type{ DataType::UNKNOWN }; /**< Output tensor data type to use if the output is not initialized */ }; -- cgit v1.2.1