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
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>
MergeResults(Tout * out,const Tin * in,int ldc,int y0,int ymax,int x0,int xmax,const Tout * bias,Activation act,bool append)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 // NOTE: The following code is disabled to avoid calling get_vector_length(), so templated MergeResults will not
40 // be correct for SVE cases. This is OK as we have specialisations for all needed SVE cases anyway.
41 //
42 // For SVE cases, multiply the width up by the vector length.
43 // Use the *input* type to determine this, since this will be what the kernel operated on.
44 // const int width = twidth * (sve ? get_vector_length<Tin>() : 1);
45 const int width = twidth;
46
47 const int full_y_blocks = (ymax - y0) / height;
48 const int y_remainder = (ymax - y0) % height;
49 const int y_blocks = full_y_blocks + (y_remainder ? 1 : 0);
50
51 const int full_x_blocks = (xmax - x0) / width;
52 const int x_remainder = (xmax - x0) % width;
53 const int x_blocks = full_x_blocks + (x_remainder ? 1 : 0);
54
55 for (int y_block = 0; y_block < y_blocks; y_block++) {
56 int ybase = y0 + (y_block * height);
57
58 int fill_rows = (y_block < full_y_blocks) ? height : y_remainder;
59
60 for (int x_block = 0; x_block < x_blocks; x_block++) {
61 int xbase = x0 + (x_block * width);
62
63 int fill_cols = (x_block < full_x_blocks) ? width : x_remainder;
64
65 for (int row=0; row < fill_rows; row++) {
66 for (int col=0; col < fill_cols; col++) {
67 Tout &r = out[(ybase + row) * ldc + xbase + col];
68 Tout v = in[row * width + col];
69
70 if (append) {
71 v += r;
72 }
73
74 if (bias) {
75 v += bias[xbase + col];
76 }
77
78 switch(act.type) {
79 default:
80 case Activation::Type::None:
81 break;
82
83 case Activation::Type::ReLU:
84 v = std::max(v, static_cast<Tout>(0));
85 break;
86
87 case Activation::Type::BoundedReLU:
88 v = std::max(std::min(v, static_cast<Tout>(act.param1)), static_cast<Tout>(0));
89 break;
90 }
91
92 r = v;
93 }
94 }
95
96 in += (width * height);
97 }
98 }
99 }
100
101 #include "merges/list.hpp"
102
103 /* Cortex-A53 8x6 SGEMM kernel uses a templated merge as the optimized merge
104 * generator cannot cope with the width (6) not being a multiple of VL (4). */
105 #ifdef __aarch64__
106 template void MergeResults<6u, 8u, false, float, float>(float *, float const*, int, int, int, int, int, float const *, Activation, bool);
107 #endif
108
109 #if defined(__aarch64__) && defined(__ARM_FP16_ARGS)
110 template void MergeResults<12u, 8u, false, float, __fp16>(__fp16*, float const*, int, int, int, int, int, __fp16 const*, Activation, bool);
111 #endif
112
113 #if defined(__arm__) && defined(__ARM_FP16_ARGS)
114 template void MergeResults<8u, 6u, false, float, __fp16>(__fp16*, float const*, int, int, int, int, int, __fp16 const*, Activation, bool);
115 #endif
116
117 } // namespace arm_gemm
118