Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2018 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_ASSEMBLY_HELPER_H__ |
| 25 | #define __ARM_ASSEMBLY_HELPER_H__ |
| 26 | |
| 27 | #include "arm_compute/core/ITensor.h" |
| 28 | #include "support/ToolchainSupport.h" |
| 29 | |
| 30 | #include "arm_compute/core/Helpers.h" |
| 31 | #include "arm_compute/core/IAccessWindow.h" |
| 32 | #include "arm_compute/core/Log.h" |
| 33 | #include "arm_compute/core/NEON/kernels/assembly/NEGEMMAssemblyWrapper.h" |
| 34 | #include "arm_compute/core/NEON/kernels/assembly/arm_gemm.hpp" |
| 35 | #include "arm_compute/core/TensorInfo.h" |
| 36 | #include "arm_compute/core/Types.h" |
| 37 | #include "arm_compute/core/Validate.h" |
| 38 | #include "arm_compute/core/Window.h" |
| 39 | #include "arm_compute/runtime/NEON/NEScheduler.h" |
| 40 | |
| 41 | namespace arm_compute |
| 42 | { |
Alex Gilday | c357c47 | 2018-03-21 13:54:09 +0000 | [diff] [blame] | 43 | /** Assembly kernel glue */ |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 44 | template <typename TypeInput, typename TypeOutput> |
| 45 | class AssemblyKernelGlue final |
| 46 | { |
| 47 | public: |
Alex Gilday | c357c47 | 2018-03-21 13:54:09 +0000 | [diff] [blame] | 48 | /** Operator type */ |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 49 | using TypeOperator = TypeInput; |
Alex Gilday | c357c47 | 2018-03-21 13:54:09 +0000 | [diff] [blame] | 50 | /** Result type */ |
| 51 | using TypeResult = TypeOutput; |
| 52 | /** Default constructor. */ |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 53 | AssemblyKernelGlue() |
| 54 | : _gemm_kernel_asm(nullptr), _optimised_kernel(nullptr), _a(nullptr), _b(nullptr), _d(nullptr) |
| 55 | { |
| 56 | } |
Alex Gilday | c357c47 | 2018-03-21 13:54:09 +0000 | [diff] [blame] | 57 | /** Assembly Gemm */ |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 58 | using AssemblyGemm = arm_gemm::GemmCommon<TypeInput, TypeOutput>; |
| 59 | |
Alex Gilday | c357c47 | 2018-03-21 13:54:09 +0000 | [diff] [blame] | 60 | /** Prevent instances of this class from being copy constructed */ |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 61 | const AssemblyKernelGlue<TypeInput, TypeOutput> &operator=(const AssemblyKernelGlue<TypeInput, TypeOutput> &) = delete; |
Alex Gilday | c357c47 | 2018-03-21 13:54:09 +0000 | [diff] [blame] | 62 | /** Prevent instances of this class from being copied */ |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 63 | AssemblyKernelGlue(const AssemblyKernelGlue<TypeInput, TypeOutput> &) = delete; |
| 64 | |
Alex Gilday | c357c47 | 2018-03-21 13:54:09 +0000 | [diff] [blame] | 65 | /** Assembly Gemm kernel */ |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 66 | std::unique_ptr<AssemblyGemm> _gemm_kernel_asm; |
Alex Gilday | c357c47 | 2018-03-21 13:54:09 +0000 | [diff] [blame] | 67 | /** Optimised NEON kernel */ |
| 68 | std::unique_ptr<INEKernel> _optimised_kernel; |
| 69 | /** Input A */ |
| 70 | const ITensor *_a; |
| 71 | /** Input B */ |
| 72 | const ITensor *_b; |
| 73 | /** Output */ |
| 74 | ITensor *_d; |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 75 | |
| 76 | /** Configures the arrays pointers and strides in the assembly kernel and executes the assembly kernel. |
| 77 | * The call to set_arrays is needed to deal with the input sizes containing batches (dims > 2) |
| 78 | */ |
| 79 | inline void run() |
| 80 | { |
| 81 | const int lda = _a->info()->strides_in_bytes().y() / sizeof(TypeInput); |
| 82 | const int ldb = _b->info()->strides_in_bytes().y() / sizeof(TypeInput); |
| 83 | const int ldd = _d->info()->strides_in_bytes().y() / sizeof(TypeOutput); |
| 84 | |
Michalis Spyrou | e7e96e0 | 2018-04-13 13:44:10 +0100 | [diff] [blame] | 85 | const int batch_stride_a = _a->info()->strides_in_bytes().z() / sizeof(TypeInput); |
| 86 | const int batch_stride_d = _d->info()->strides_in_bytes().z() / sizeof(TypeOutput); |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 87 | |
Michalis Spyrou | e7e96e0 | 2018-04-13 13:44:10 +0100 | [diff] [blame] | 88 | const int multi_stride_a = _a->info()->strides_in_bytes()[3] / sizeof(TypeInput); |
| 89 | const int multi_stride_b = _b->info()->strides_in_bytes().z() / sizeof(TypeInput); |
| 90 | const int multi_stride_d = _d->info()->strides_in_bytes()[3] / sizeof(TypeOutput); |
| 91 | |
| 92 | const auto in0_ptr = reinterpret_cast<const TypeInput *>(_a->buffer()); |
| 93 | const auto in1_ptr = reinterpret_cast<const TypeInput *>(_b->buffer()); |
| 94 | auto out_ptr = reinterpret_cast<TypeOutput *>(_d->buffer()); |
| 95 | |
| 96 | _gemm_kernel_asm->set_arrays(in0_ptr, lda, batch_stride_a, multi_stride_a, in1_ptr, ldb, multi_stride_b, out_ptr, ldd, batch_stride_d, multi_stride_d); |
| 97 | NEScheduler::get().schedule(_optimised_kernel.get(), Window::DimX); |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 98 | } |
| 99 | }; |
| 100 | |
Alex Gilday | c357c47 | 2018-03-21 13:54:09 +0000 | [diff] [blame] | 101 | /** Float 32 assembly kernel glue */ |
| 102 | using AssemblyKernelGlueF32 = AssemblyKernelGlue<float, float>; |
| 103 | /** Uint 8 to Uint 32 kernel glue */ |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 104 | using AssemblyKernelGlueU8U32 = AssemblyKernelGlue<uint8_t, uint32_t>; |
Alex Gilday | c357c47 | 2018-03-21 13:54:09 +0000 | [diff] [blame] | 105 | /** Int 8 to Int 32 kernel glue */ |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 106 | using AssemblyKernelGlueS8S32 = AssemblyKernelGlue<int8_t, int32_t>; |
| 107 | |
Alex Gilday | c357c47 | 2018-03-21 13:54:09 +0000 | [diff] [blame] | 108 | /** Allocate a workspace tensor. |
| 109 | * |
| 110 | * @param[in] workspace_size Size to allocate. |
| 111 | * @param[out] workspace Tensor to allocate. |
| 112 | * @param[in] memory_group Tensor memory group. |
| 113 | * @param[in] alignment Workspace memory alignment. |
| 114 | * @param[in] num_threads Number of workspace threads. |
| 115 | */ |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 116 | inline void allocate_workspace(size_t workspace_size, Tensor &workspace, MemoryGroup &memory_group, size_t alignment, unsigned int num_threads) |
| 117 | { |
| 118 | ARM_COMPUTE_ERROR_ON_MSG(workspace_size == 0, "size cannot be 0"); |
| 119 | workspace.allocator()->init(TensorInfo(TensorShape{ (workspace_size + alignment - 1) * num_threads }, 1, DataType::S8)); |
| 120 | workspace.allocator()->allocate(); |
| 121 | } |
| 122 | |
Alex Gilday | c357c47 | 2018-03-21 13:54:09 +0000 | [diff] [blame] | 123 | /** Create a wrapper kernel. |
| 124 | * |
Alex Gilday | c357c47 | 2018-03-21 13:54:09 +0000 | [diff] [blame] | 125 | * @param[in] a Input tensor A. |
| 126 | * @param[in] b Input tensor B. |
Pablo Tello | 7fad9b1 | 2018-03-14 17:55:27 +0000 | [diff] [blame] | 127 | * @param[out] d Output tensor. |
Alex Gilday | c357c47 | 2018-03-21 13:54:09 +0000 | [diff] [blame] | 128 | * @param[in] alpha Alpha value. |
| 129 | * @param[in] beta Beta value. |
| 130 | * @param[out] workspace Workspace tensor |
| 131 | * @param[in] memory_group Tensor memory group. |
| 132 | * @param[out] asm_glue Assembly glue kernel. |
| 133 | * |
Pablo Tello | 7fad9b1 | 2018-03-14 17:55:27 +0000 | [diff] [blame] | 134 | * @return the wrapper kernel. |
Alex Gilday | c357c47 | 2018-03-21 13:54:09 +0000 | [diff] [blame] | 135 | */ |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 136 | template <typename T> |
Pablo Tello | 7fad9b1 | 2018-03-14 17:55:27 +0000 | [diff] [blame] | 137 | inline bool setup_assembly_kernel(const ITensor *a, const ITensor *b, ITensor *d, float alpha, float beta, |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 138 | Tensor &workspace, MemoryGroup &memory_group, T &asm_glue) |
| 139 | { |
Pablo Tello | 7fad9b1 | 2018-03-14 17:55:27 +0000 | [diff] [blame] | 140 | const CPUInfo &ci = NEScheduler::get().cpu_info(); |
| 141 | const int M = d->info()->tensor_shape().y(); |
| 142 | const int N = d->info()->tensor_shape().x(); |
| 143 | const int K = a->info()->tensor_shape().x(); |
Michalis Spyrou | e7e96e0 | 2018-04-13 13:44:10 +0100 | [diff] [blame] | 144 | const int batches = a->info()->tensor_shape().total_size_upper(2); |
| 145 | const int multis = b->info()->tensor_shape().z(); |
Pablo Tello | 7fad9b1 | 2018-03-14 17:55:27 +0000 | [diff] [blame] | 146 | unsigned int num_threads = NEScheduler::get().num_threads(); |
Michalis Spyrou | e7e96e0 | 2018-04-13 13:44:10 +0100 | [diff] [blame] | 147 | |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 148 | // unique_ptr to a Gemm object |
Pablo Tello | 7fad9b1 | 2018-03-14 17:55:27 +0000 | [diff] [blame] | 149 | std::unique_ptr<typename T::AssemblyGemm> |
Michalis Spyrou | e7e96e0 | 2018-04-13 13:44:10 +0100 | [diff] [blame] | 150 | asm_gemm(arm_gemm::gemm<typename T::TypeOperator, typename T::TypeResult>(ci, M, N, K, batches, multis, false, false, alpha, beta, num_threads, false)); |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 151 | // arm_compute wrapper for the Gemm object (see above) |
Pablo Tello | 7fad9b1 | 2018-03-14 17:55:27 +0000 | [diff] [blame] | 152 | std::unique_ptr<NEGEMMAssemblyWrapper<typename T::AssemblyGemm>> |
| 153 | acl_gemm_wrapper = support::cpp14::make_unique<NEGEMMAssemblyWrapper<typename T::AssemblyGemm>>(); |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 154 | if(acl_gemm_wrapper != nullptr && asm_gemm != nullptr) |
| 155 | { |
| 156 | acl_gemm_wrapper->configure(asm_gemm.get()); |
| 157 | const size_t workspace_size = asm_gemm->get_working_size(); |
| 158 | if(workspace_size) |
| 159 | { |
| 160 | // Allocate workspace |
Pablo Tello | 7fad9b1 | 2018-03-14 17:55:27 +0000 | [diff] [blame] | 161 | const unsigned int alignment = 4096; |
| 162 | allocate_workspace(workspace_size, workspace, memory_group, alignment, num_threads); |
| 163 | ARM_COMPUTE_ERROR_ON_NULLPTR(workspace.buffer()); |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 164 | asm_gemm->set_working_space(reinterpret_cast<typename T::TypeResult *>(workspace.buffer())); |
| 165 | } |
Pablo Tello | 7fad9b1 | 2018-03-14 17:55:27 +0000 | [diff] [blame] | 166 | |
| 167 | //if we disable this code below in brackets then ConvLayer deadlocks when threads > 1 and |
| 168 | //the shapes are In=1x1x1024 Weights=1x1x1024x1001 Biases=1001 Out=1x1x1001 |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 169 | { |
Pablo Tello | 7fad9b1 | 2018-03-14 17:55:27 +0000 | [diff] [blame] | 170 | const unsigned int window_size = asm_gemm->get_window_size(); |
| 171 | if(window_size < num_threads) |
| 172 | { |
| 173 | num_threads = window_size; |
| 174 | asm_gemm->set_nthreads(num_threads); |
| 175 | } |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 176 | } |
Pablo Tello | 7fad9b1 | 2018-03-14 17:55:27 +0000 | [diff] [blame] | 177 | |
Pablo Tello | eb82fd2 | 2018-02-23 13:43:50 +0000 | [diff] [blame] | 178 | asm_glue._gemm_kernel_asm = std::move(asm_gemm); |
| 179 | asm_glue._optimised_kernel = std::move(acl_gemm_wrapper); |
| 180 | // We need to setup the ptrs in the run() method |
| 181 | asm_glue._a = a; |
| 182 | asm_glue._b = b; |
| 183 | asm_glue._d = d; |
| 184 | return true; |
| 185 | } |
| 186 | return false; |
| 187 | } |
| 188 | } |
| 189 | #endif /* __ARM_ASSEMBLY_HELPER_H__ */ |