XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1 | // Auto-generated file. Do not edit! |
| 2 | // Template: src/f32-gemm/scalar.c.in |
| 3 | // Generator: tools/xngen |
| 4 | // |
| 5 | // Copyright 2019 Google LLC |
| 6 | // |
| 7 | // This source code is licensed under the BSD-style license found in the |
| 8 | // LICENSE file in the root directory of this source tree. |
| 9 | |
| 10 | #include <assert.h> |
| 11 | |
| 12 | #include <xnnpack/gemm.h> |
| 13 | #include <xnnpack/math.h> |
| 14 | |
| 15 | |
| 16 | void xnn_f32_gemminc_ukernel_1x4__scalar( |
| 17 | size_t mr, |
| 18 | size_t nc, |
| 19 | size_t kc, |
| 20 | const float* restrict a, |
| 21 | size_t a_stride, |
| 22 | const float* restrict w, |
| 23 | float* restrict c, |
| 24 | size_t cm_stride, |
| 25 | size_t cn_stride, |
| 26 | const float*restrict acc, |
Marat Dukhan | eb09a6b | 2020-04-08 17:34:32 -0700 | [diff] [blame] | 27 | const union xnn_f32_minmax_params params[restrict static 1]) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 28 | { |
| 29 | assert(mr != 0); |
| 30 | assert(mr <= 1); |
| 31 | assert(nc != 0); |
| 32 | assert(kc != 0); |
| 33 | assert(kc % sizeof(float) == 0); |
| 34 | assert(a != NULL); |
| 35 | assert(w != NULL); |
| 36 | assert(c != NULL); |
| 37 | assert(acc != NULL); |
| 38 | |
| 39 | const float* a0 = a; |
| 40 | float* c0 = c; |
| 41 | |
| 42 | do { |
| 43 | float vacc00 = acc[0]; |
| 44 | float vacc01 = acc[1]; |
| 45 | float vacc02 = acc[2]; |
| 46 | float vacc03 = acc[3]; |
| 47 | acc += 4; |
| 48 | |
| 49 | size_t k = kc; |
| 50 | do { |
| 51 | const float va0 = *a0++; |
| 52 | |
| 53 | const float vb0 = w[0]; |
| 54 | const float vb1 = w[1]; |
| 55 | const float vb2 = w[2]; |
| 56 | const float vb3 = w[3]; |
| 57 | w += 4; |
| 58 | |
| 59 | vacc00 += va0 * vb0; |
| 60 | vacc01 += va0 * vb1; |
| 61 | vacc02 += va0 * vb2; |
| 62 | vacc03 += va0 * vb3; |
| 63 | |
| 64 | k -= sizeof(float); |
| 65 | } while (k != 0); |
| 66 | |
| 67 | const float vmin = params->scalar.min; |
| 68 | vacc00 = math_max_f32(vacc00, vmin); |
| 69 | vacc01 = math_max_f32(vacc01, vmin); |
| 70 | vacc02 = math_max_f32(vacc02, vmin); |
| 71 | vacc03 = math_max_f32(vacc03, vmin); |
| 72 | |
| 73 | const float vmax = params->scalar.max; |
| 74 | vacc00 = math_min_f32(vacc00, vmax); |
| 75 | vacc01 = math_min_f32(vacc01, vmax); |
| 76 | vacc02 = math_min_f32(vacc02, vmax); |
| 77 | vacc03 = math_min_f32(vacc03, vmax); |
| 78 | |
| 79 | if XNN_LIKELY(nc >= 4) { |
| 80 | c0[0] = vacc00; |
| 81 | c0[1] = vacc01; |
| 82 | c0[2] = vacc02; |
| 83 | c0[3] = vacc03; |
| 84 | c0 = (float*) ((uintptr_t) c0 + cn_stride); |
| 85 | |
| 86 | a0 = (const void*) ((uintptr_t) a0 - kc); |
| 87 | |
| 88 | nc -= 4; |
| 89 | } else { |
| 90 | if (nc & 2) { |
| 91 | c0[0] = vacc00; |
| 92 | c0[1] = vacc01; |
| 93 | vacc00 = vacc02; |
| 94 | c0 += 2; |
| 95 | } |
| 96 | if (nc & 1) { |
| 97 | c0[0] = vacc00; |
| 98 | } |
| 99 | |
| 100 | nc = 0; |
| 101 | } |
| 102 | } while (nc != 0); |
| 103 | } |