From 1009e87127e08a58561373c3f83f22b5f0922641 Mon Sep 17 00:00:00 2001 From: Michalis Spyrou Date: Mon, 27 Jul 2020 12:48:34 +0100 Subject: COMPMID-3385: Async support to CLArithmetic* kernels/functions Pt.2 Signed-off-by: Michalis Spyrou Change-Id: Idc5ac2dd2ba5295c00c88b44a783645327a27e15 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/3617 Tested-by: Arm Jenkins Reviewed-by: Georgios Pinitas Comments-Addressed: Arm Jenkins --- .../CL/kernels/CLPixelWiseMultiplicationKernel.h | 48 +++--- arm_compute/core/utils/misc/InfoHelpers.h | 10 +- arm_compute/runtime/CL/functions/CLLSTMLayer.h | 174 ++++++++++----------- .../CL/functions/CLPixelWiseMultiplication.h | 157 ++++++++++++++++++- arm_compute/runtime/CL/functions/CLQLSTMLayer.h | 20 +-- arm_compute/runtime/common/LSTMParams.h | 36 ++--- 6 files changed, 297 insertions(+), 148 deletions(-) (limited to 'arm_compute') diff --git a/arm_compute/core/CL/kernels/CLPixelWiseMultiplicationKernel.h b/arm_compute/core/CL/kernels/CLPixelWiseMultiplicationKernel.h index bb98eb83bf..86159fc915 100644 --- a/arm_compute/core/CL/kernels/CLPixelWiseMultiplicationKernel.h +++ b/arm_compute/core/CL/kernels/CLPixelWiseMultiplicationKernel.h @@ -62,16 +62,16 @@ public: * - (QSYMM16,QSYMM16) -> QSYMM16 * - (QSYMM16,QSYMM16) -> S32 * - * @param[in] input1 An input tensor. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32. - * @param[in] input2 An input tensor. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32. - * @param[out] output The output tensor. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32. + * @param[in] input1 An input tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32. + * @param[in] input2 An input tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32. + * @param[out] output The output tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32. * @param[in] scale Scale to apply after multiplication. * Scale must be positive and its value must be either 1/255 or 1/2^n where n is between 0 and 15. * @param[in] overflow_policy Overflow policy. Supported overflow policies: Wrap, Saturate * @param[in] rounding_policy Rounding policy. Supported rounding modes: to zero, to nearest even. * @param[in] act_info (Optional) Activation layer information in case of a fused activation. */ - void configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float scale, + void configure(ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, float scale, ConvertPolicy overflow_policy, RoundingPolicy rounding_policy, const ActivationLayerInfo &act_info = ActivationLayerInfo()); /** Initialise the kernel's input, output and border mode. * @@ -90,16 +90,16 @@ public: * - (QSYMM16,QSYMM16) -> S32 * * @param[in] compile_context The compile context to be used. - * @param[in] input1 An input tensor. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32. - * @param[in] input2 An input tensor. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32. - * @param[out] output The output tensor. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32. + * @param[in] input1 An input tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32. + * @param[in] input2 An input tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32. + * @param[out] output The output tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32. * @param[in] scale Scale to apply after multiplication. * Scale must be positive and its value must be either 1/255 or 1/2^n where n is between 0 and 15. * @param[in] overflow_policy Overflow policy. Supported overflow policies: Wrap, Saturate * @param[in] rounding_policy Rounding policy. Supported rounding modes: to zero, to nearest even. * @param[in] act_info (Optional) Activation layer information in case of a fused activation. */ - void configure(const CLCompileContext &compile_context, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, float scale, + void configure(const CLCompileContext &compile_context, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, float scale, ConvertPolicy overflow_policy, RoundingPolicy rounding_policy, const ActivationLayerInfo &act_info = ActivationLayerInfo()); /** Static function to check if given info will lead to a valid configuration of @ref CLPixelWiseMultiplicationKernel * @@ -132,13 +132,13 @@ public: ConvertPolicy overflow_policy, RoundingPolicy rounding_policy, const ActivationLayerInfo &act_info = ActivationLayerInfo()); // Inherited methods overridden: - void run(const Window &window, cl::CommandQueue &queue) override; + void run_op(const InputTensorMap &inputs, const OutputTensorMap &outputs, const Window &window, cl::CommandQueue &queue) override; BorderSize border_size() const override; private: - const ICLTensor *_input1; - const ICLTensor *_input2; - ICLTensor *_output; + const ITensorInfo *_input1; + const ITensorInfo *_input2; + ITensorInfo *_output; }; /** Interface for the complex pixelwise multiplication kernel. */ @@ -157,21 +157,21 @@ public: CLComplexPixelWiseMultiplicationKernel &operator=(CLComplexPixelWiseMultiplicationKernel &&) = default; /** Initialise the kernel's input, output and border mode. * - * @param[in] input1 An input tensor. Data types supported: F32. Number of channels supported: 2. - * @param[in] input2 An input tensor. Data types supported: same as @p input1. Number of channels supported: same as @p input1. - * @param[out] output The output tensor, Data types supported: same as @p input1. Number of channels supported: same as @p input1. + * @param[in] input1 An input tensor info. Data types supported: F32. Number of channels supported: 2. + * @param[in] input2 An input tensor info. Data types supported: same as @p input1. Number of channels supported: same as @p input1. + * @param[out] output The output tensor info. Data types supported: same as @p input1. Number of channels supported: same as @p input1. * @param[in] act_info (Optional) Activation layer information in case of a fused activation. */ - void configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, const ActivationLayerInfo &act_info = ActivationLayerInfo()); + void configure(ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const ActivationLayerInfo &act_info = ActivationLayerInfo()); /** Initialise the kernel's input, output and border mode. * * @param[in] compile_context The compile context to be used. - * @param[in] input1 An input tensor. Data types supported: F32. Number of channels supported: 2. - * @param[in] input2 An input tensor. Data types supported: same as @p input1. Number of channels supported: same as @p input1. - * @param[out] output The output tensor, Data types supported: same as @p input1. Number of channels supported: same as @p input1. + * @param[in] input1 An input tensor info. Data types supported: F32. Number of channels supported: 2. + * @param[in] input2 An input tensor info. Data types supported: same as @p input1. Number of channels supported: same as @p input1. + * @param[out] output The output tensor info. Data types supported: same as @p input1. Number of channels supported: same as @p input1. * @param[in] act_info (Optional) Activation layer information in case of a fused activation. */ - void configure(const CLCompileContext &compile_context, const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output, const ActivationLayerInfo &act_info = ActivationLayerInfo()); + void configure(const CLCompileContext &compile_context, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const ActivationLayerInfo &act_info = ActivationLayerInfo()); /** Static function to check if given info will lead to a valid configuration of @ref CLComplexPixelWiseMultiplicationKernel * * @param[in] input1 An input tensor info. Data types supported: F32. Number of channels supported: 2. @@ -184,13 +184,13 @@ public: static Status validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ActivationLayerInfo &act_info = ActivationLayerInfo()); // Inherited methods overridden: - void run(const Window &window, cl::CommandQueue &queue) override; + void run_op(const InputTensorMap &inputs, const OutputTensorMap &outputs, const Window &window, cl::CommandQueue &queue) override; BorderSize border_size() const override; private: - const ICLTensor *_input1; - const ICLTensor *_input2; - ICLTensor *_output; + const ITensorInfo *_input1; + const ITensorInfo *_input2; + ITensorInfo *_output; }; } // namespace arm_compute #endif /*ARM_COMPUTE_CLPIXELWISEMULTIPLICATIONKERNEL_H */ diff --git a/arm_compute/core/utils/misc/InfoHelpers.h b/arm_compute/core/utils/misc/InfoHelpers.h index ffde82bce2..ced0d24b56 100644 --- a/arm_compute/core/utils/misc/InfoHelpers.h +++ b/arm_compute/core/utils/misc/InfoHelpers.h @@ -86,7 +86,7 @@ inline void build_lstm_params_tensor_info(const LSTMParams &lstm_params, { ARM_COMPUTE_ERROR_ON_NULLPTR(lstm_params.input_to_input_weights(), lstm_params.recurrent_to_input_weights(), lstm_params.input_gate_bias()); - const ITensorInfo *cell_to_input_weights_info = (lstm_params.has_peephole_opt()) ? lstm_params.cell_to_input_weights()->info() : nullptr; + ITensorInfo *cell_to_input_weights_info = (lstm_params.has_peephole_opt()) ? lstm_params.cell_to_input_weights()->info() : nullptr; lstm_params_info->set_cifg_params(lstm_params.input_to_input_weights()->info(), lstm_params.recurrent_to_input_weights()->info(), cell_to_input_weights_info, lstm_params.input_gate_bias()->info()); } @@ -100,10 +100,10 @@ inline void build_lstm_params_tensor_info(const LSTMParams &lstm_params, ARM_COMPUTE_ERROR_ON_NULLPTR(lstm_params.input_layer_norm_weights()); } - const ITensorInfo *forget_info = lstm_params.forget_layer_norm_weights()->info(); - const ITensorInfo *cell_info = lstm_params.cell_layer_norm_weights()->info(); - const ITensorInfo *output_info = lstm_params.output_layer_norm_weights()->info(); - const ITensorInfo *input_info = lstm_params.has_cifg_opt() ? nullptr : lstm_params.input_layer_norm_weights()->info(); + ITensorInfo *forget_info = lstm_params.forget_layer_norm_weights()->info(); + ITensorInfo *cell_info = lstm_params.cell_layer_norm_weights()->info(); + ITensorInfo *output_info = lstm_params.output_layer_norm_weights()->info(); + ITensorInfo *input_info = lstm_params.has_cifg_opt() ? nullptr : lstm_params.input_layer_norm_weights()->info(); lstm_params_info->set_layer_normalization_params(input_info, forget_info, cell_info, output_info); } diff --git a/arm_compute/runtime/CL/functions/CLLSTMLayer.h b/arm_compute/runtime/CL/functions/CLLSTMLayer.h index abfcc3a62f..1a8b33463d 100644 --- a/arm_compute/runtime/CL/functions/CLLSTMLayer.h +++ b/arm_compute/runtime/CL/functions/CLLSTMLayer.h @@ -28,7 +28,6 @@ #include "arm_compute/core/CL/kernels/CLCopyKernel.h" #include "arm_compute/core/CL/kernels/CLMemsetKernel.h" -#include "arm_compute/core/CL/kernels/CLPixelWiseMultiplicationKernel.h" #include "arm_compute/core/Types.h" #include "arm_compute/runtime/CL/CLTensor.h" #include "arm_compute/runtime/CL/functions/CLActivationLayer.h" @@ -37,6 +36,7 @@ #include "arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h" #include "arm_compute/runtime/CL/functions/CLGEMM.h" #include "arm_compute/runtime/CL/functions/CLMeanStdDevNormalizationLayer.h" +#include "arm_compute/runtime/CL/functions/CLPixelWiseMultiplication.h" #include "arm_compute/runtime/IMemoryManager.h" #include "arm_compute/runtime/MemoryGroup.h" #include "arm_compute/runtime/common/LSTMParams.h" @@ -97,7 +97,7 @@ public: const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights, const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights, const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias, - const ICLTensor *output_state_in, const ICLTensor *cell_state_in, + const ICLTensor *output_state_in, ICLTensor *cell_state_in, ICLTensor *scratch_buffer, ICLTensor *output_state_out, ICLTensor *cell_state_out, ICLTensor *output, const LSTMParams &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold = 0.f, float projection_threshold = 0.f); /** Initialize function's tensors. @@ -143,7 +143,7 @@ public: const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights, const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights, const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias, - const ICLTensor *output_state_in, const ICLTensor *cell_state_in, + const ICLTensor *output_state_in, ICLTensor *cell_state_in, ICLTensor *scratch_buffer, ICLTensor *output_state_out, ICLTensor *cell_state_out, ICLTensor *output, const LSTMParams &lstm_params, const ActivationLayerInfo &activation_info, float cell_threshold = 0.f, float projection_threshold = 0.f); @@ -200,90 +200,90 @@ public: void prepare() override; private: - MemoryGroup _memory_group; - CLFullyConnectedLayer _fully_connected_input_gate; - CLArithmeticAddition _accum_input_gate1; - CLArithmeticSubtraction _subtract_input_gate; - CLPixelWiseMultiplicationKernel _pixelwise_mul_input_gate; - CLActivationLayer _activation_input_gate; - CLFullyConnectedLayer _fully_connected_forget_gate; - CLArithmeticAddition _accum_forget_gate1; - CLPixelWiseMultiplicationKernel _pixelwise_mul_forget_gate; - CLActivationLayer _activation_forget_gate; - CLFullyConnectedLayer _fully_connected_cell_state; - CLGEMM _gemm_cell_state1; - CLTransposeKernel _transpose_cell_state; - CLArithmeticAddition _accum_cell_state1; - CLArithmeticAddition _accum_cell_state2; - CLPixelWiseMultiplicationKernel _pixelwise_mul_cell_state1; - CLActivationLayer _activation_cell_state; - CLActivationLayer _cell_clip; - CLPixelWiseMultiplicationKernel _pixelwise_mul_cell_state2; - CLFullyConnectedLayer _fully_connected_output; - CLPixelWiseMultiplicationKernel _pixelwise_mul_output_state1; - CLArithmeticAddition _accum_output1; - CLActivationLayer _activation_output; - CLActivationLayer _activation_output_state; - CLPixelWiseMultiplicationKernel _pixelwise_mul_output_state2; - CLFullyConnectedLayer _fully_connected_output_state; - CLActivationLayer _projection_clip; - CLCopyKernel _copy_cell_state; - CLCopyKernel _copy_output; - CLConcatenateLayer _concat_scratch_buffer; - CLConcatenateLayer _concat_inputs_forget_gate; - CLConcatenateLayer _concat_weights_forget_gate; - CLConcatenateLayer _concat_weights_input_gate; - CLConcatenateLayer _concat_weights_output; - CLMemsetKernel _ones_memset_kernel; - CLMeanStdDevNormalizationLayer _mean_std_norm_input_gate; - CLPixelWiseMultiplicationKernel _pixelwise_mul_input_gate_coeff; - CLArithmeticAddition _accum_input_gate_bias; - CLMeanStdDevNormalizationLayer _mean_std_norm_forget_gate; - CLPixelWiseMultiplicationKernel _pixelwise_mul_forget_gate_coeff; - CLArithmeticAddition _accum_forget_gate_bias; - CLMeanStdDevNormalizationLayer _mean_std_norm_cell_gate; - CLPixelWiseMultiplicationKernel _pixelwise_mul_cell_gate_coeff; - CLArithmeticAddition _accum_cell_gate_bias; - CLMeanStdDevNormalizationLayer _mean_std_norm_output_gate; - CLPixelWiseMultiplicationKernel _pixelwise_mul_output_gate_coeff; - CLArithmeticAddition _accum_output_gate_bias; - CLTensor _input_gate_out1; - CLTensor _input_gate_out2; - CLTensor _input_gate_out3; - CLTensor _input_gate_out4; - CLTensor _forget_gate_out1; - CLTensor _forget_gate_out2; - CLTensor _forget_gate_out3; - CLTensor _forget_gate_out4; - CLTensor _forget_gate_out5; - CLTensor _forget_gate_out6; - CLTensor _cell_state_out1; - CLTensor _cell_state_out2; - CLTensor _cell_state_out3; - CLTensor _cell_state_out4; - CLTensor _cell_state_out5; - CLTensor _output1; - CLTensor _output2; - CLTensor _output3; - CLTensor _output4; - CLTensor _cell_state_activation; - CLTensor _output_state1; - CLTensor _ones; - CLTensor _input_layer_norm_out1; - CLTensor _input_layer_norm_out2; - CLTensor _forget_layer_norm_out1; - CLTensor _forget_layer_norm_out2; - CLTensor _cell_layer_norm_out1; - CLTensor _cell_layer_norm_out2; - CLTensor _output_layer_norm_out1; - CLTensor _output_layer_norm_out2; - bool _run_peephole_opt; - bool _run_cifg_opt; - bool _perform_cell_clipping; - bool _has_projection_weights; - bool _perform_projection_clipping; - bool _is_prepared; - bool _is_layer_norm_lstm; + MemoryGroup _memory_group; + CLFullyConnectedLayer _fully_connected_input_gate; + CLArithmeticAddition _accum_input_gate1; + CLArithmeticSubtraction _subtract_input_gate; + CLPixelWiseMultiplication _pixelwise_mul_input_gate; + CLActivationLayer _activation_input_gate; + CLFullyConnectedLayer _fully_connected_forget_gate; + CLArithmeticAddition _accum_forget_gate1; + CLPixelWiseMultiplication _pixelwise_mul_forget_gate; + CLActivationLayer _activation_forget_gate; + CLFullyConnectedLayer _fully_connected_cell_state; + CLGEMM _gemm_cell_state1; + CLTransposeKernel _transpose_cell_state; + CLArithmeticAddition _accum_cell_state1; + CLArithmeticAddition _accum_cell_state2; + CLPixelWiseMultiplication _pixelwise_mul_cell_state1; + CLActivationLayer _activation_cell_state; + CLActivationLayer _cell_clip; + CLPixelWiseMultiplication _pixelwise_mul_cell_state2; + CLFullyConnectedLayer _fully_connected_output; + CLPixelWiseMultiplication _pixelwise_mul_output_state1; + CLArithmeticAddition _accum_output1; + CLActivationLayer _activation_output; + CLActivationLayer _activation_output_state; + CLPixelWiseMultiplication _pixelwise_mul_output_state2; + CLFullyConnectedLayer _fully_connected_output_state; + CLActivationLayer _projection_clip; + CLCopyKernel _copy_cell_state; + CLCopyKernel _copy_output; + CLConcatenateLayer _concat_scratch_buffer; + CLConcatenateLayer _concat_inputs_forget_gate; + CLConcatenateLayer _concat_weights_forget_gate; + CLConcatenateLayer _concat_weights_input_gate; + CLConcatenateLayer _concat_weights_output; + CLMemsetKernel _ones_memset_kernel; + CLMeanStdDevNormalizationLayer _mean_std_norm_input_gate; + CLPixelWiseMultiplication _pixelwise_mul_input_gate_coeff; + CLArithmeticAddition _accum_input_gate_bias; + CLMeanStdDevNormalizationLayer _mean_std_norm_forget_gate; + CLPixelWiseMultiplication _pixelwise_mul_forget_gate_coeff; + CLArithmeticAddition _accum_forget_gate_bias; + CLMeanStdDevNormalizationLayer _mean_std_norm_cell_gate; + CLPixelWiseMultiplication _pixelwise_mul_cell_gate_coeff; + CLArithmeticAddition _accum_cell_gate_bias; + CLMeanStdDevNormalizationLayer _mean_std_norm_output_gate; + CLPixelWiseMultiplication _pixelwise_mul_output_gate_coeff; + CLArithmeticAddition _accum_output_gate_bias; + CLTensor _input_gate_out1; + CLTensor _input_gate_out2; + CLTensor _input_gate_out3; + CLTensor _input_gate_out4; + CLTensor _forget_gate_out1; + CLTensor _forget_gate_out2; + CLTensor _forget_gate_out3; + CLTensor _forget_gate_out4; + CLTensor _forget_gate_out5; + CLTensor _forget_gate_out6; + CLTensor _cell_state_out1; + CLTensor _cell_state_out2; + CLTensor _cell_state_out3; + CLTensor _cell_state_out4; + CLTensor _cell_state_out5; + CLTensor _output1; + CLTensor _output2; + CLTensor _output3; + CLTensor _output4; + CLTensor _cell_state_activation; + CLTensor _output_state1; + CLTensor _ones; + CLTensor _input_layer_norm_out1; + CLTensor _input_layer_norm_out2; + CLTensor _forget_layer_norm_out1; + CLTensor _forget_layer_norm_out2; + CLTensor _cell_layer_norm_out1; + CLTensor _cell_layer_norm_out2; + CLTensor _output_layer_norm_out1; + CLTensor _output_layer_norm_out2; + bool _run_peephole_opt; + bool _run_cifg_opt; + bool _perform_cell_clipping; + bool _has_projection_weights; + bool _perform_projection_clipping; + bool _is_prepared; + bool _is_layer_norm_lstm; }; } // namespace arm_compute #endif /* ARM_COMPUTE_CLLSTMLAYER_H */ diff --git a/arm_compute/runtime/CL/functions/CLPixelWiseMultiplication.h b/arm_compute/runtime/CL/functions/CLPixelWiseMultiplication.h index b87daba1f8..ca8d77e6b7 100644 --- a/arm_compute/runtime/CL/functions/CLPixelWiseMultiplication.h +++ b/arm_compute/runtime/CL/functions/CLPixelWiseMultiplication.h @@ -24,18 +24,141 @@ #ifndef ARM_COMPUTE_CLPIXELWISEMULTIPLICATION_H #define ARM_COMPUTE_CLPIXELWISEMULTIPLICATION_H -#include "arm_compute/core/Types.h" -#include "arm_compute/runtime/CL/ICLSimpleFunction.h" +#include "arm_compute/core/CL/kernels/CLFillBorderKernel.h" +#include "arm_compute/runtime/CL/ICLOperator.h" +#include "arm_compute/runtime/IFunction.h" namespace arm_compute { // Forward declaration class ICLTensor; +namespace experimental +{ +/** Basic function to run @ref CLPixelWiseMultiplicationKernel. */ +class CLPixelWiseMultiplication : public ICLOperator +{ +public: + /** Default Constructor */ + CLPixelWiseMultiplication(); + /** Initialise the kernel's inputs, output and convertion policy. + * + * Valid configurations (Input1,Input2) -> Output : + * + * - (U8,U8) -> U8 + * - (U8,U8) -> S16 + * - (U8,S16) -> S16 + * - (S16,U8) -> S16 + * - (S16,S16) -> S16 + * - (F16,F16) -> F16 + * - (F32,F32) -> F32 + * - (QASYMM8,QASYMM8) -> QASYMM8 + * - (QASYMM8_SIGNED,QASYMM8_SIGNED) -> QASYMM8_SIGNED + * - (QSYMM16,QSYMM16) -> QSYMM16 + * - (QSYMM16,QSYMM16) -> S32 + * + * @param[in] compile_context The compile context to be used. + * @param[in, out] input1 An input tensor. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32. + * The input tensor is [in, out] because its TensorInfo might be modified inside the kernel in case of broadcasting of dimension 0. + * @param[in, out] input2 An input tensor. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32. + * The input tensor is [in, out] because its TensorInfo might be modified inside the kernel in case of broadcasting of dimension 0. + * @param[out] output The output tensor. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32. + * @param[in] scale Scale to apply after multiplication. + * Scale must be positive and its value must be either 1/255 or 1/2^n where n is between 0 and 15. + * @param[in] overflow_policy Overflow policy. Supported overflow policies: Wrap, Saturate + * @param[in] rounding_policy Rounding policy. Supported rounding modes: to zero, to nearest even. + * @param[in] act_info (Optional) Activation layer information in case of a fused activation. + */ + void configure(const CLCompileContext &compile_context, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, float scale, + ConvertPolicy overflow_policy, RoundingPolicy rounding_policy, const ActivationLayerInfo &act_info = ActivationLayerInfo()); + /** Static function to check if given info will lead to a valid configuration of @ref CLPixelWiseMultiplication + * + * Valid configurations (Input1,Input2) -> Output : + * + * - (U8,U8) -> U8 + * - (U8,U8) -> S16 + * - (U8,S16) -> S16 + * - (S16,U8) -> S16 + * - (S16,S16) -> S16 + * - (F16,F16) -> F16 + * - (F32,F32) -> F32 + * - (QASYMM8,QASYMM8) -> QASYMM8 + * - (QASYMM8_SIGNED,QASYMM8_SIGNED) -> QASYMM8_SIGNED + * - (QSYMM16,QSYMM16) -> QSYMM16 + * - (QSYMM16,QSYMM16) -> S32 + * + * + * @param[in] input1 An input tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32. + * @param[in] input2 An input tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32. + * @param[in] output The output tensor info. Data types supported: U8/QASYMM8/QASYMM8_SIGNED/S16/QSYMM16/F16/F32. + * @param[in] scale Scale to apply after multiplication. + * Scale must be positive and its value must be either 1/255 or 1/2^n where n is between 0 and 15. + * @param[in] overflow_policy Overflow policy. Supported overflow policies: Wrap, Saturate + * @param[in] rounding_policy Rounding policy. Supported rounding modes: to zero, to nearest even. + * @param[in] act_info (Optional) Activation layer information in case of a fused activation. + * + * @return a status + */ + static Status validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float scale, + ConvertPolicy overflow_policy, RoundingPolicy rounding_policy, const ActivationLayerInfo &act_info = ActivationLayerInfo()); + + // Inherited methods overridden: + void run(InputTensorMap inputs, OutputTensorMap outputs, OperatorTensorMap workspace) override; + +private: + CLFillBorderKernel _border_handler; +}; + +/** Basic function to run @ref CLComplexPixelWiseMultiplicationKernel. */ +class CLComplexPixelWiseMultiplication : public ICLOperator +{ +public: + /** Default Constructor */ + CLComplexPixelWiseMultiplication(); + /** Initialise the kernel's inputs, output. + * + * @param[in] compile_context The compile context to be used. + * @param[in, out] input1 An input tensor. Data types supported: F32. Number of channels supported: 2. + * The input tensor is [in, out] because its TensorInfo might be modified inside the kernel in case of broadcasting of dimension 0. + * @param[in, out] input2 An input tensor. Data types supported: same as @p input1. Number of channels supported: same as @p input1. + * The input tensor is [in, out] because its TensorInfo might be modified inside the kernel in case of broadcasting of dimension 0. + * @param[out] output The output tensor, Data types supported: same as @p input1. Number of channels supported: same as @p input1. + * @param[in] act_info (Optional) Activation layer information in case of a fused activation. + */ + void configure(const CLCompileContext &compile_context, ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output, const ActivationLayerInfo &act_info = ActivationLayerInfo()); + /** Static function to check if given info will lead to a valid configuration of @ref CLComplexPixelWiseMultiplication + * + * @param[in] input1 An input tensor info. Data types supported: F32. Number of channels supported: 2. + * @param[in] input2 An input tensor info. Data types supported: same as @p input1. Number of channels supported: same as @p input1. + * @param[in] output The output tensor info, Data types supported: same as @p input1. Number of channels supported: same as @p input1. + * @param[in] act_info (Optional) Activation layer information in case of a fused activation. + */ + static Status validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ActivationLayerInfo &act_info = ActivationLayerInfo()); + + // Inherited methods overridden: + void run(InputTensorMap inputs, OutputTensorMap outputs, OperatorTensorMap workspace) override; + +private: + CLFillBorderKernel _border_handler; +}; +} // namespace experimental + /** Basic function to run @ref CLPixelWiseMultiplicationKernel. */ -class CLPixelWiseMultiplication : public ICLSimpleFunction +class CLPixelWiseMultiplication : public IFunction { public: + /** Default Constructor */ + CLPixelWiseMultiplication(); + /** Default Destructor */ + ~CLPixelWiseMultiplication(); + /** Prevent instances of this class from being copied (As this class contains pointers) */ + CLPixelWiseMultiplication(const CLPixelWiseMultiplication &) = delete; + /** Default move constructor */ + CLPixelWiseMultiplication(CLPixelWiseMultiplication &&); + /** Prevent instances of this class from being copied (As this class contains pointers) */ + CLPixelWiseMultiplication &operator=(const CLPixelWiseMultiplication &) = delete; + /** Default move assignment operator */ + CLPixelWiseMultiplication &operator=(CLPixelWiseMultiplication &&); /** Initialise the kernel's inputs, output and convertion policy. * * Valid configurations (Input1,Input2) -> Output : @@ -125,12 +248,31 @@ public: */ static Status validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, float scale, ConvertPolicy overflow_policy, RoundingPolicy rounding_policy, const ActivationLayerInfo &act_info = ActivationLayerInfo()); + + // Inherited methods overridden: + void run() override; + +private: + struct Impl; + std::unique_ptr _impl; }; /** Basic function to run @ref CLComplexPixelWiseMultiplicationKernel. */ -class CLComplexPixelWiseMultiplication : public ICLSimpleFunction +class CLComplexPixelWiseMultiplication : public IFunction { public: + /** Default Constructor */ + CLComplexPixelWiseMultiplication(); + /** Default Destructor */ + ~CLComplexPixelWiseMultiplication(); + /** Prevent instances of this class from being copied (As this class contains pointers) */ + CLComplexPixelWiseMultiplication(const CLComplexPixelWiseMultiplication &) = delete; + /** Default move constructor */ + CLComplexPixelWiseMultiplication(CLComplexPixelWiseMultiplication &&); + /** Prevent instances of this class from being copied (As this class contains pointers) */ + CLComplexPixelWiseMultiplication &operator=(const CLComplexPixelWiseMultiplication &) = delete; + /** Default move assignment operator */ + CLComplexPixelWiseMultiplication &operator=(CLComplexPixelWiseMultiplication &&); /** Initialise the kernel's inputs, output. * * @param[in, out] input1 An input tensor. Data types supported: F32. Number of channels supported: 2. @@ -160,6 +302,13 @@ public: * @param[in] act_info (Optional) Activation layer information in case of a fused activation. */ static Status validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output, const ActivationLayerInfo &act_info = ActivationLayerInfo()); + + // Inherited methods overridden: + void run() override; + +private: + struct Impl; + std::unique_ptr _impl; }; } // namespace arm_compute #endif /*ARM_COMPUTE_CLPIXELWISEMULTIPLICATION_H */ diff --git a/arm_compute/runtime/CL/functions/CLQLSTMLayer.h b/arm_compute/runtime/CL/functions/CLQLSTMLayer.h index 0aea91ae8e..53f337bc61 100644 --- a/arm_compute/runtime/CL/functions/CLQLSTMLayer.h +++ b/arm_compute/runtime/CL/functions/CLQLSTMLayer.h @@ -26,13 +26,13 @@ #include "arm_compute/core/CL/kernels/CLCopyKernel.h" #include "arm_compute/core/CL/kernels/CLGEMMLowpReductionKernel.h" -#include "arm_compute/core/CL/kernels/CLPixelWiseMultiplicationKernel.h" #include "arm_compute/core/CL/kernels/CLQLSTMLayerNormalizationKernel.h" #include "arm_compute/core/Types.h" #include "arm_compute/runtime/CL/functions/CLActivationLayer.h" #include "arm_compute/runtime/CL/functions/CLElementwiseOperations.h" #include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h" #include "arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h" +#include "arm_compute/runtime/CL/functions/CLPixelWiseMultiplication.h" #include "arm_compute/runtime/CL/functions/CLTranspose.h" #include "arm_compute/runtime/common/LSTMParams.h" @@ -52,7 +52,7 @@ class ICLTensor; * -# @ref CLGEMMLowpMatrixMultiplyCore Quantized matrix multiplication core. Accumulators are 32-bit integers * -# @ref CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint Convert 32-bit integers into QSYMM16 * -# @ref CLGEMMLowpMatrixAReductionKernel For precomputing effective biases to use - * -# @ref CLPixelWiseMultiplicationKernel Elementwise multiplication + * -# @ref CLPixelWiseMultiplication Elementwise multiplication * -# @ref CLTranspose Transpose function for reshaping the weights * */ class CLQLSTMLayer : public IFunction @@ -113,7 +113,7 @@ public: const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights, const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights, const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias, - const ICLTensor *cell_state_in, const ICLTensor *output_state_in, + ICLTensor *cell_state_in, const ICLTensor *output_state_in, ICLTensor *cell_state_out, ICLTensor *output_state_out, ICLTensor *output, const LSTMParams &lstm_params); @@ -163,7 +163,7 @@ public: const ICLTensor *input_to_forget_weights, const ICLTensor *input_to_cell_weights, const ICLTensor *input_to_output_weights, const ICLTensor *recurrent_to_forget_weights, const ICLTensor *recurrent_to_cell_weights, const ICLTensor *recurrent_to_output_weights, const ICLTensor *forget_gate_bias, const ICLTensor *cell_bias, const ICLTensor *output_gate_bias, - const ICLTensor *cell_state_in, const ICLTensor *output_state_in, + ICLTensor *cell_state_in, const ICLTensor *output_state_in, ICLTensor *cell_state_out, ICLTensor *output_state_out, ICLTensor *output, const LSTMParams &lstm_params); @@ -306,7 +306,7 @@ private: CLArithmeticAddition _projection_bias_add{}; CLGEMMLowpMatrixMultiplyCore _mm_input_to_forget{}; CLGEMMLowpMatrixMultiplyCore _mm_recurrent_to_forget{}; - CLPixelWiseMultiplicationKernel _pixelwise_mul_cell_to_forget{}; + CLPixelWiseMultiplication _pixelwise_mul_cell_to_forget{}; CLGEMMLowpOutputStage _input_to_forget_outstage{}; CLGEMMLowpOutputStage _recurrent_to_forget_outstage{}; CLGEMMLowpOutputStage _cell_to_forget_outstage{}; @@ -325,12 +325,12 @@ private: CLGEMMLowpMatrixMultiplyCore _mm_recurrent_to_input{}; CLGEMMLowpOutputStage _recurrent_to_input_outstage{}; CLArithmeticAddition _accumulate_input_recurrent_input{}; - CLPixelWiseMultiplicationKernel _pixelwise_mul_cell_to_input{}; + CLPixelWiseMultiplication _pixelwise_mul_cell_to_input{}; CLGEMMLowpOutputStage _cell_to_input_outstage{}; CLArithmeticAddition _accumulate_cell_input{}; CLActivationLayer _input_gate_sigmoid{}; - CLPixelWiseMultiplicationKernel _pixelwise_mul_forget_cell{}; - CLPixelWiseMultiplicationKernel _pixelwise_mul_input_cell{}; + CLPixelWiseMultiplication _pixelwise_mul_forget_cell{}; + CLPixelWiseMultiplication _pixelwise_mul_input_cell{}; CLArithmeticAddition _add_forget_cell{}; CLActivationLayer _cell_clip{}; CLGEMMLowpMatrixMultiplyCore _mm_input_to_output{}; @@ -338,12 +338,12 @@ private: CLGEMMLowpMatrixMultiplyCore _mm_recurrent_to_output{}; CLGEMMLowpOutputStage _recurrent_to_output_outstage{}; CLArithmeticAddition _accumulate_input_recurrent_output{}; - CLPixelWiseMultiplicationKernel _pixelwise_mul_cell_to_output{}; + CLPixelWiseMultiplication _pixelwise_mul_cell_to_output{}; CLGEMMLowpOutputStage _cell_to_output_outstage{}; CLArithmeticAddition _accumulate_cell_to_output{}; CLActivationLayer _output_gate_sigmoid{}; CLActivationLayer _hidden_tanh{}; - CLPixelWiseMultiplicationKernel _pixelwise_mul_hidden{}; + CLPixelWiseMultiplication _pixelwise_mul_hidden{}; CLGEMMLowpOutputStage _hidden_outstage{}; CLGEMMLowpMatrixMultiplyCore _mm_projection{}; CLGEMMLowpOutputStage _projection_outstage{}; diff --git a/arm_compute/runtime/common/LSTMParams.h b/arm_compute/runtime/common/LSTMParams.h index 82fca7e7a2..ffb4ddd9d3 100644 --- a/arm_compute/runtime/common/LSTMParams.h +++ b/arm_compute/runtime/common/LSTMParams.h @@ -81,7 +81,7 @@ public: * * @return Reference to this LSTMParams object */ - LSTMParams &set_cifg_params(const T *input_to_input_weights, const T *recurrent_to_input_weights, const T *cell_to_input_weights, const T *input_gate_bias) + LSTMParams &set_cifg_params(const T *input_to_input_weights, const T *recurrent_to_input_weights, T *cell_to_input_weights, const T *input_gate_bias) { _input_to_input_weights = input_to_input_weights; _recurrent_to_input_weights = recurrent_to_input_weights; @@ -111,7 +111,7 @@ public: * * @return Reference to this LSTMParams object */ - LSTMParams &set_peephole_params(const T *cell_to_forget_weights, const T *cell_to_output_weights) + LSTMParams &set_peephole_params(T *cell_to_forget_weights, T *cell_to_output_weights) { _cell_to_forget_weights = cell_to_forget_weights; _cell_to_output_weights = cell_to_output_weights; @@ -127,8 +127,8 @@ public: * * @return Reference to this LSTMParams object */ - LSTMParams &set_layer_normalization_params(const T *input_layer_norm_weights, const T *forget_layer_norm_weights, - const T *cell_layer_norm_weights, const T *output_layer_norm_weights) + LSTMParams &set_layer_normalization_params(T *input_layer_norm_weights, T *forget_layer_norm_weights, + T *cell_layer_norm_weights, T *output_layer_norm_weights) { _input_layer_norm_weights = input_layer_norm_weights; _forget_layer_norm_weights = forget_layer_norm_weights; @@ -204,7 +204,7 @@ public: return _recurrent_to_input_weights; } - const T *cell_to_input_weights() const + T *cell_to_input_weights() const { return _cell_to_input_weights; } @@ -214,12 +214,12 @@ public: return _input_gate_bias; } - const T *cell_to_forget_weights() const + T *cell_to_forget_weights() const { return _cell_to_forget_weights; } - const T *cell_to_output_weights() const + T *cell_to_output_weights() const { return _cell_to_output_weights; } @@ -234,22 +234,22 @@ public: return _projection_bias; } - const T *input_layer_norm_weights() const + T *input_layer_norm_weights() const { return _input_layer_norm_weights; } - const T *forget_layer_norm_weights() const + T *forget_layer_norm_weights() const { return _forget_layer_norm_weights; } - const T *cell_layer_norm_weights() const + T *cell_layer_norm_weights() const { return _cell_layer_norm_weights; } - const T *output_layer_norm_weights() const + T *output_layer_norm_weights() const { return _output_layer_norm_weights; } @@ -317,16 +317,16 @@ public: private: const T *_input_to_input_weights; const T *_recurrent_to_input_weights; - const T *_cell_to_input_weights; + T *_cell_to_input_weights; const T *_input_gate_bias; - const T *_cell_to_forget_weights; - const T *_cell_to_output_weights; + T *_cell_to_forget_weights; + T *_cell_to_output_weights; const T *_projection_weights; const T *_projection_bias; - const T *_input_layer_norm_weights; - const T *_forget_layer_norm_weights; - const T *_cell_layer_norm_weights; - const T *_output_layer_norm_weights; + T *_input_layer_norm_weights; + T *_forget_layer_norm_weights; + T *_cell_layer_norm_weights; + T *_output_layer_norm_weights; float _cell_clip; float _projection_clip; float _input_intermediate_scale; -- cgit v1.2.1