1 /* 2 * Copyright (c) 2017-2021 Arm Limited. 3 * 4 * SPDX-License-Identifier: MIT 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a copy 7 * of this software and associated documentation files (the "Software"), to 8 * deal in the Software without restriction, including without limitation the 9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or 10 * sell copies of the Software, and to permit persons to whom the Software is 11 * furnished to do so, subject to the following conditions: 12 * 13 * The above copyright notice and this permission notice shall be included in all 14 * copies or substantial portions of the Software. 15 * 16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 22 * SOFTWARE. 23 */ 24 #ifndef ARM_COMPUTE_CL_SOFTMAX_KERNEL_H 25 #define ARM_COMPUTE_CL_SOFTMAX_KERNEL_H 26 27 #include "arm_compute/core/Error.h" 28 #include "arm_compute/core/KernelDescriptors.h" 29 #include "src/core/common/Macros.h" 30 #include "src/gpu/cl/ClCompileContext.h" 31 #include "src/gpu/cl/IClKernel.h" 32 33 namespace arm_compute 34 { 35 namespace opencl 36 { 37 namespace kernels 38 { 39 /** Interface for max, shifting, exponentiating and summing the logits */ 40 class ClLogits1DMaxShiftExpSumKernel : public IClKernel 41 { 42 /**< Grid size (obtained through auto-tuning) */ 43 static const unsigned int _grid_size; 44 /**< Vector size in the serial case (obtained through auto-tuning) */ 45 static const unsigned int _serial_vector_size; 46 /**< Vector size in the parallel case (obtained through auto-tuning, enables the best memory access pattern for Bifrost) .*/ 47 static const unsigned int _parallel_vector_size; 48 49 public: 50 /** Info for whether a parallel reduction will be run and the vector size of the execution. */ 51 using ParallelReductionInfo = std::tuple<bool, unsigned int>; 52 53 ClLogits1DMaxShiftExpSumKernel(); 54 ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClLogits1DMaxShiftExpSumKernel); 55 /** Configure the kernel using the given information about tensors 56 * 57 * @param[in] compile_context The compile context to be used. 58 * @param[in] src Source tensor. Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32 59 * @param[in,out] max Max values tensor. Data types supported: same as @p src 60 * @param[out] dst Destination tensor. Data types supported: same as @p src 61 * @param[out] sum Sum of 1D logits tensor. Data types supported: same as @p src 62 * @param[in] info Contains information consumed by kernels for softmax described in @ref SoftmaxKernelInfo. 63 */ 64 void configure(const CLCompileContext &compile_context, const ITensorInfo &src, ITensorInfo &max, ITensorInfo &dst, ITensorInfo &sum, const SoftmaxKernelInfo &info); 65 /** Static function to check if given info will lead to a valid configuration 66 * 67 * Similar to @ref ClLogits1DMaxShiftExpSumKernel::configure() 68 * 69 * @return a status 70 */ 71 static Status validate(const ITensorInfo &src, const ITensorInfo &max, const ITensorInfo &dst, const ITensorInfo &sum); 72 /** Checks if the given size is eligible for parallel reduction 73 * 74 * @note Serial reduction is launched for width < (_grid_size * _serial_vector_size). 75 * @note Parallel reduction is launched for width >= (_grid_size * _serial_vector_size) and vector_size is forced to 4. 76 * 77 * @param[in] size Size to check 78 * 79 * @return A two-element tuple where the first element is a boolean specifying if a parallel reduction will be run, 80 * while the second element is the vector size of the execution. 81 */ 82 static ParallelReductionInfo is_parallel_reduction(size_t size); 83 84 // Inherited methods overridden: 85 void run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue) override; 86 }; 87 88 /** Interface for calculating the final step of the Softmax Layer where each logit value is multiplied by the inverse of the sum of the logits. */ 89 class ClLogits1DNormKernel : public IClKernel 90 { 91 public: 92 ClLogits1DNormKernel(); 93 ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(ClLogits1DNormKernel); 94 95 /** Set the input and output tensors. 96 * 97 * @param[in] compile_context The compile context to be used. 98 * @param[in] src Source tensor. Data types supported: S32/F16/F32. If this kernel is used for log softmax, only F32/F16 is supported. 99 * @param[in] sum Sum tensor. Dimensions should be dim(input)-1. Data types supported: same as @p input 100 * @param[out] dst Destination tensor. Data types supported: QASYMM8/QASYMM8_SIGNED for S32 @p input, or same as @p input 101 * @param[in] info Contains information consumed by kernels for softmax described in @ref SoftmaxKernelInfo. 102 */ 103 void configure(const CLCompileContext &compile_context, const ITensorInfo &src, const ITensorInfo &sum, ITensorInfo &dst, const SoftmaxKernelInfo &info); 104 /** Static function to check if given info will lead to a valid configuration 105 * 106 * Similar to @ref ClLogits1DNormKernel::configure() 107 * 108 * @return a status 109 */ 110 static Status validate(const ITensorInfo &src, const ITensorInfo &sum, const ITensorInfo &dst, const SoftmaxKernelInfo &info); 111 112 // Inherited methods overridden: 113 void run_op(ITensorPack &tensors, const Window &window, ::cl::CommandQueue &queue) override; 114 }; 115 } // namespace kernels 116 } // namespace opencl 117 } // namespace arm_compute 118 #endif /* ARM_COMPUTE_CL_SOFTMAX_KERNEL_H */ 119