Georgios Pinitas | 48b3ef8 | 2019-10-14 19:03:09 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2017-2018 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 | |
| 25 | /* As some of the merges need these headers, but are all included in the |
| 26 | * arm_gemm namespace, put these headers here. */ |
| 27 | #include <algorithm> |
| 28 | |
| 29 | #include <arm_neon.h> |
| 30 | |
| 31 | #include "arm_gemm.hpp" |
| 32 | #include "asmlib.hpp" |
| 33 | #include "utils.hpp" |
| 34 | |
| 35 | namespace arm_gemm { |
| 36 | |
| 37 | template<unsigned int twidth, unsigned int height, bool sve=false, typename Tin, typename Tout> |
| 38 | void MergeResults(Tout * out, const Tin * in, int ldc, int y0, int ymax, int x0, int xmax, const Tout *bias, Activation act, bool append) { |
| 39 | // For SVE cases, multiply the width up by the vector length. |
| 40 | // Use the *input* type to determine this, since this will be what the kernel operated on. |
| 41 | const int width = twidth * (sve ? get_vector_length<Tin>() : 1); |
| 42 | |
| 43 | const int full_y_blocks = (ymax - y0) / height; |
| 44 | const int y_remainder = (ymax - y0) % height; |
| 45 | const int y_blocks = full_y_blocks + (y_remainder ? 1 : 0); |
| 46 | |
| 47 | const int full_x_blocks = (xmax - x0) / width; |
| 48 | const int x_remainder = (xmax - x0) % width; |
| 49 | const int x_blocks = full_x_blocks + (x_remainder ? 1 : 0); |
| 50 | |
| 51 | for (int y_block = 0; y_block < y_blocks; y_block++) { |
| 52 | int ybase = y0 + (y_block * height); |
| 53 | |
| 54 | int fill_rows = (y_block < full_y_blocks) ? height : y_remainder; |
| 55 | |
| 56 | for (int x_block = 0; x_block < x_blocks; x_block++) { |
| 57 | int xbase = x0 + (x_block * width); |
| 58 | |
| 59 | int fill_cols = (x_block < full_x_blocks) ? width : x_remainder; |
| 60 | |
| 61 | for (int row=0; row < fill_rows; row++) { |
| 62 | for (int col=0; col < fill_cols; col++) { |
| 63 | Tout &r = out[(ybase + row) * ldc + xbase + col]; |
| 64 | Tout v = in[row * width + col]; |
| 65 | |
| 66 | if (append) { |
| 67 | v += r; |
| 68 | } |
| 69 | |
| 70 | if (bias) { |
| 71 | v += bias[xbase + col]; |
| 72 | } |
| 73 | |
| 74 | switch(act.type) { |
| 75 | default: |
| 76 | case Activation::Type::None: |
| 77 | break; |
| 78 | |
| 79 | case Activation::Type::ReLU: |
| 80 | v = std::max(v, static_cast<Tout>(0)); |
| 81 | break; |
| 82 | |
| 83 | case Activation::Type::BoundedReLU: |
| 84 | v = std::max(std::min(v, static_cast<Tout>(act.param1)), static_cast<Tout>(0)); |
| 85 | break; |
| 86 | } |
| 87 | |
| 88 | r = v; |
| 89 | } |
| 90 | } |
| 91 | |
| 92 | in += (width * height); |
| 93 | } |
| 94 | } |
| 95 | } |
| 96 | |
| 97 | #include "merges/list.hpp" |
| 98 | |
| 99 | #if defined(__aarch64__) && defined(__ARM_FP16_ARGS) |
| 100 | template void MergeResults<12u, 8u, false, float, __fp16>(__fp16*, float const*, int, int, int, int, int, __fp16 const*, Activation, bool); |
| 101 | #endif |
| 102 | |
| 103 | #if defined(__arm__) && defined(__ARM_FP16_ARGS) |
| 104 | template void MergeResults<8u, 6u, false, float, __fp16>(__fp16*, float const*, int, int, int, int, int, __fp16 const*, Activation, bool); |
| 105 | #endif |
| 106 | |
| 107 | } // namespace arm_gemm |