Michele Di Giorgio | 93b75e0 | 2021-06-21 12:00:43 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2016-2021 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 | */ |
Georgios Pinitas | 7891a73 | 2021-08-20 21:39:25 +0100 | [diff] [blame^] | 24 | #include "src/cpu/kernels/CpuGemmMatrixAdditionKernel.h" |
Michele Di Giorgio | 93b75e0 | 2021-06-21 12:00:43 +0100 | [diff] [blame] | 25 | |
| 26 | #include "arm_compute/core/Helpers.h" |
| 27 | #include "arm_compute/core/Types.h" |
| 28 | #include "arm_compute/core/Validate.h" |
| 29 | #include "src/core/CPP/Validate.h" |
| 30 | #include "src/core/NEON/NEFixedPoint.h" |
| 31 | #include "src/core/helpers/AutoConfiguration.h" |
| 32 | #include "src/core/helpers/WindowHelpers.h" |
| 33 | |
| 34 | #include <arm_neon.h> |
| 35 | |
| 36 | namespace arm_compute |
| 37 | { |
| 38 | namespace cpu |
| 39 | { |
| 40 | namespace kernels |
| 41 | { |
| 42 | namespace |
| 43 | { |
| 44 | void matrix_addition_f32(const ITensor *src, ITensor *dst, const Window &window, float beta) |
| 45 | { |
| 46 | ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); |
| 47 | const float32x4_t beta_f32 = vdupq_n_f32(beta); |
| 48 | |
| 49 | constexpr int window_step_x = 16; |
| 50 | const auto window_start_x = static_cast<int>(window.x().start()); |
| 51 | const auto window_end_x = static_cast<int>(window.x().end()); |
| 52 | |
| 53 | Window win = window.collapse_if_possible(window, Window::DimZ); |
| 54 | win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 55 | |
| 56 | Iterator in(src, win); |
| 57 | Iterator out(dst, win); |
| 58 | |
| 59 | execute_window_loop(win, [&](const Coordinates &) |
| 60 | { |
| 61 | const auto in_ptr = reinterpret_cast<const float *>(in.ptr()); |
| 62 | const auto out_ptr = reinterpret_cast<float *>(out.ptr()); |
| 63 | |
| 64 | int x = window_start_x; |
| 65 | for(; x < (window_end_x - window_step_x); x += window_step_x) |
| 66 | { |
| 67 | float32x4x4_t alpha_ab = vld4q_f32(out_ptr + x); |
| 68 | const float32x4x4_t c = vld4q_f32(in_ptr + x); |
| 69 | |
| 70 | // Multiply matrix C by its weight and accumulate |
| 71 | alpha_ab.val[0] = vmlaq_f32(alpha_ab.val[0], c.val[0], beta_f32); |
| 72 | alpha_ab.val[1] = vmlaq_f32(alpha_ab.val[1], c.val[1], beta_f32); |
| 73 | alpha_ab.val[2] = vmlaq_f32(alpha_ab.val[2], c.val[2], beta_f32); |
| 74 | alpha_ab.val[3] = vmlaq_f32(alpha_ab.val[3], c.val[3], beta_f32); |
| 75 | |
| 76 | vst4q_f32(out_ptr + x, alpha_ab); |
| 77 | } |
| 78 | |
| 79 | // Left-over loop |
| 80 | for(; x < window_end_x; ++x) |
| 81 | { |
| 82 | *(out_ptr + x) += *(in_ptr + x) * beta; |
| 83 | } |
| 84 | }, |
| 85 | in, out); |
| 86 | } |
| 87 | |
| 88 | #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
| 89 | void matrix_addition_f16(const ITensor *src, ITensor *dst, const Window &window, float beta) |
| 90 | { |
| 91 | ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); |
| 92 | const float16x8_t beta_f16 = vdupq_n_f16(beta); |
| 93 | |
| 94 | constexpr int window_step_x = 16; |
| 95 | const auto window_start_x = static_cast<int>(window.x().start()); |
| 96 | const auto window_end_x = static_cast<int>(window.x().end()); |
| 97 | |
| 98 | Window win = window.collapse_if_possible(window, Window::DimZ); |
| 99 | win.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 100 | |
| 101 | Iterator in(src, win); |
| 102 | Iterator out(dst, win); |
| 103 | |
Freddie Liardet | bb5d42c | 2021-07-28 14:24:02 +0100 | [diff] [blame] | 104 | execute_window_loop(win, [&](const Coordinates &) |
Michele Di Giorgio | 93b75e0 | 2021-06-21 12:00:43 +0100 | [diff] [blame] | 105 | { |
| 106 | const auto in_ptr = reinterpret_cast<const float16_t *>(in.ptr()); |
| 107 | const auto out_ptr = reinterpret_cast<float16_t *>(out.ptr()); |
| 108 | |
| 109 | int x = window_start_x; |
| 110 | for(; x < (window_end_x - window_step_x); x += window_step_x) |
| 111 | { |
| 112 | float16x8x2_t alpha_ab = vld2q_f16(out_ptr + x); |
| 113 | const float16x8x2_t c = vld2q_f16(in_ptr + x); |
| 114 | // Multiply matrix C by its weight and accumulate |
| 115 | alpha_ab.val[0] = vaddq_f16(alpha_ab.val[0], vmulq_f16(c.val[0], beta_f16)); |
| 116 | alpha_ab.val[1] = vaddq_f16(alpha_ab.val[1], vmulq_f16(c.val[1], beta_f16)); |
| 117 | |
| 118 | vst2q_f16(out_ptr + x, alpha_ab); |
| 119 | } |
| 120 | |
| 121 | // Left-over loop |
| 122 | for(; x < window_end_x; ++x) |
| 123 | { |
| 124 | *(out_ptr + x) += *(in_ptr + x) * static_cast<float16_t>(beta); |
| 125 | } |
| 126 | }, |
| 127 | in, out); |
| 128 | } |
| 129 | #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |
| 130 | |
| 131 | } // namespace |
| 132 | |
| 133 | void CpuGemmMatrixAdditionKernel::configure(const ITensorInfo *src, ITensorInfo *dst, float beta) |
| 134 | { |
| 135 | ARM_COMPUTE_UNUSED(dst); |
| 136 | ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst); |
| 137 | |
| 138 | // Perform validation step |
| 139 | ARM_COMPUTE_ERROR_THROW_ON(CpuGemmMatrixAdditionKernel::validate(src, dst, beta)); |
| 140 | |
| 141 | _beta = beta; |
| 142 | switch(src->data_type()) |
| 143 | { |
| 144 | case DataType::F32: |
| 145 | _func = &matrix_addition_f32; |
| 146 | break; |
| 147 | case DataType::F16: |
| 148 | #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC |
| 149 | _func = &matrix_addition_f16; |
| 150 | break; |
| 151 | #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ |
| 152 | default: |
| 153 | ARM_COMPUTE_ERROR("Data type not supported"); |
| 154 | break; |
| 155 | } |
| 156 | |
| 157 | // Configure kernel window |
| 158 | Window win = calculate_max_window(*src, Steps()); |
| 159 | ICPPKernel::configure(win); |
| 160 | } |
| 161 | |
| 162 | Status CpuGemmMatrixAdditionKernel::validate(const ITensorInfo *src, const ITensorInfo *dst, float beta) |
| 163 | { |
| 164 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); |
| 165 | ARM_COMPUTE_UNUSED(beta); |
| 166 | |
| 167 | ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(src); |
| 168 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32); |
| 169 | |
| 170 | if(dst->total_size() > 0) |
| 171 | { |
| 172 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, dst); |
| 173 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, dst); |
| 174 | } |
| 175 | return Status{}; |
| 176 | } |
| 177 | |
| 178 | void CpuGemmMatrixAdditionKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) |
| 179 | { |
| 180 | ARM_COMPUTE_UNUSED(info); |
| 181 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 182 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window); |
| 183 | ARM_COMPUTE_ERROR_ON(tensors.empty()); |
| 184 | |
| 185 | const ITensor *src = tensors.get_const_tensor(TensorType::ACL_SRC); |
| 186 | ITensor *dst = tensors.get_tensor(TensorType::ACL_DST); |
| 187 | |
| 188 | if(_beta != 0.0f) |
| 189 | { |
| 190 | (*_func)(src, dst, window, _beta); |
| 191 | } |
| 192 | } |
| 193 | |
| 194 | const char *CpuGemmMatrixAdditionKernel::name() const |
| 195 | { |
| 196 | return "CpuGemmMatrixAdditionKernel"; |
| 197 | } |
| 198 | } // namespace kernels |
| 199 | } // namespace cpu |
| 200 | } // namespace arm_compute |