| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/f32-ppmm/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/math.h" |
| 13 |
#include "xnnpack/ppmm.h" |
| 14 |
|
| 15 |
|
| 16 |
void xnn_f32_ppmm_minmax_ukernel_4x4__scalar( |
| 17 |
size_t mr, |
| 18 |
size_t nc, |
| 19 |
size_t kc, |
| 20 |
const float* restrict a, |
| 21 |
const float* restrict w, |
| 22 |
float* restrict c, |
| 23 |
size_t cm_stride, |
| 24 |
size_t cn_stride, |
| 25 |
const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) |
| 26 |
{ |
| 27 |
assert(mr != 0); |
| 28 |
assert(mr <= 4); |
| 29 |
assert(nc != 0); |
| 30 |
assert(kc != 0); |
| 31 |
assert(kc % sizeof(float) == 0); |
| 32 |
|
| 33 |
float* c0 = c; |
| 34 |
float* c1 = (float*) ((uintptr_t) c0 + cm_stride); |
| 35 |
if XNN_UNPREDICTABLE(mr < 2) { |
| 36 |
c1 = c0; |
| 37 |
} |
| 38 |
float* c2 = (float*) ((uintptr_t) c1 + cm_stride); |
| 39 |
if XNN_UNPREDICTABLE(mr <= 2) { |
| 40 |
c2 = c1; |
| 41 |
} |
| 42 |
float* c3 = (float*) ((uintptr_t) c2 + cm_stride); |
| 43 |
if XNN_UNPREDICTABLE(mr != 4) { |
| 44 |
c3 = c2; |
| 45 |
} |
| 46 |
|
| 47 |
do { |
| 48 |
float vacc0x0 = w[0]; |
| 49 |
float vacc0x1 = w[1]; |
| 50 |
float vacc0x2 = w[2]; |
| 51 |
float vacc0x3 = w[3]; |
| 52 |
float vacc1x0 = vacc0x0; |
| 53 |
float vacc1x1 = vacc0x1; |
| 54 |
float vacc1x2 = vacc0x2; |
| 55 |
float vacc1x3 = vacc0x3; |
| 56 |
float vacc2x0 = vacc0x0; |
| 57 |
float vacc2x1 = vacc0x1; |
| 58 |
float vacc2x2 = vacc0x2; |
| 59 |
float vacc2x3 = vacc0x3; |
| 60 |
float vacc3x0 = vacc0x0; |
| 61 |
float vacc3x1 = vacc0x1; |
| 62 |
float vacc3x2 = vacc0x2; |
| 63 |
float vacc3x3 = vacc0x3; |
| 64 |
w += 4; |
| 65 |
|
| 66 |
size_t k = kc; |
| 67 |
do { |
| 68 |
const float va0 = a[0]; |
| 69 |
const float va1 = a[1]; |
| 70 |
const float va2 = a[2]; |
| 71 |
const float va3 = a[3]; |
| 72 |
a += 4; |
| 73 |
|
| 74 |
const float vb0 = w[0]; |
| 75 |
const float vb1 = w[1]; |
| 76 |
const float vb2 = w[2]; |
| 77 |
const float vb3 = w[3]; |
| 78 |
w += 4; |
| 79 |
|
| 80 |
vacc0x0 += va0 * vb0; |
| 81 |
vacc1x0 += va1 * vb0; |
| 82 |
vacc2x0 += va2 * vb0; |
| 83 |
vacc3x0 += va3 * vb0; |
| 84 |
vacc0x1 += va0 * vb1; |
| 85 |
vacc1x1 += va1 * vb1; |
| 86 |
vacc2x1 += va2 * vb1; |
| 87 |
vacc3x1 += va3 * vb1; |
| 88 |
vacc0x2 += va0 * vb2; |
| 89 |
vacc1x2 += va1 * vb2; |
| 90 |
vacc2x2 += va2 * vb2; |
| 91 |
vacc3x2 += va3 * vb2; |
| 92 |
vacc0x3 += va0 * vb3; |
| 93 |
vacc1x3 += va1 * vb3; |
| 94 |
vacc2x3 += va2 * vb3; |
| 95 |
vacc3x3 += va3 * vb3; |
| 96 |
|
| 97 |
k -= sizeof(float); |
| 98 |
} while (k != 0); |
| 99 |
|
| 100 |
const float vmax = params->scalar.max; |
| 101 |
vacc0x0 = math_min_f32(vacc0x0, vmax); |
| 102 |
vacc1x0 = math_min_f32(vacc1x0, vmax); |
| 103 |
vacc2x0 = math_min_f32(vacc2x0, vmax); |
| 104 |
vacc3x0 = math_min_f32(vacc3x0, vmax); |
| 105 |
vacc0x1 = math_min_f32(vacc0x1, vmax); |
| 106 |
vacc1x1 = math_min_f32(vacc1x1, vmax); |
| 107 |
vacc2x1 = math_min_f32(vacc2x1, vmax); |
| 108 |
vacc3x1 = math_min_f32(vacc3x1, vmax); |
| 109 |
vacc0x2 = math_min_f32(vacc0x2, vmax); |
| 110 |
vacc1x2 = math_min_f32(vacc1x2, vmax); |
| 111 |
vacc2x2 = math_min_f32(vacc2x2, vmax); |
| 112 |
vacc3x2 = math_min_f32(vacc3x2, vmax); |
| 113 |
vacc0x3 = math_min_f32(vacc0x3, vmax); |
| 114 |
vacc1x3 = math_min_f32(vacc1x3, vmax); |
| 115 |
vacc2x3 = math_min_f32(vacc2x3, vmax); |
| 116 |
vacc3x3 = math_min_f32(vacc3x3, vmax); |
| 117 |
|
| 118 |
const float vmin = params->scalar.min; |
| 119 |
vacc0x0 = math_max_f32(vacc0x0, vmin); |
| 120 |
vacc1x0 = math_max_f32(vacc1x0, vmin); |
| 121 |
vacc2x0 = math_max_f32(vacc2x0, vmin); |
| 122 |
vacc3x0 = math_max_f32(vacc3x0, vmin); |
| 123 |
vacc0x1 = math_max_f32(vacc0x1, vmin); |
| 124 |
vacc1x1 = math_max_f32(vacc1x1, vmin); |
| 125 |
vacc2x1 = math_max_f32(vacc2x1, vmin); |
| 126 |
vacc3x1 = math_max_f32(vacc3x1, vmin); |
| 127 |
vacc0x2 = math_max_f32(vacc0x2, vmin); |
| 128 |
vacc1x2 = math_max_f32(vacc1x2, vmin); |
| 129 |
vacc2x2 = math_max_f32(vacc2x2, vmin); |
| 130 |
vacc3x2 = math_max_f32(vacc3x2, vmin); |
| 131 |
vacc0x3 = math_max_f32(vacc0x3, vmin); |
| 132 |
vacc1x3 = math_max_f32(vacc1x3, vmin); |
| 133 |
vacc2x3 = math_max_f32(vacc2x3, vmin); |
| 134 |
vacc3x3 = math_max_f32(vacc3x3, vmin); |
| 135 |
|
| 136 |
if XNN_LIKELY(nc >= 4) { |
| 137 |
c3[0] = vacc3x0; |
| 138 |
c3[1] = vacc3x1; |
| 139 |
c3[2] = vacc3x2; |
| 140 |
c3[3] = vacc3x3; |
| 141 |
c2[0] = vacc2x0; |
| 142 |
c2[1] = vacc2x1; |
| 143 |
c2[2] = vacc2x2; |
| 144 |
c2[3] = vacc2x3; |
| 145 |
c1[0] = vacc1x0; |
| 146 |
c1[1] = vacc1x1; |
| 147 |
c1[2] = vacc1x2; |
| 148 |
c1[3] = vacc1x3; |
| 149 |
c0[0] = vacc0x0; |
| 150 |
c0[1] = vacc0x1; |
| 151 |
c0[2] = vacc0x2; |
| 152 |
c0[3] = vacc0x3; |
| 153 |
|
| 154 |
a = (const float*) ((uintptr_t) a - kc * 4); |
| 155 |
|
| 156 |
c3 = (float*) ((uintptr_t) c3 + cn_stride); |
| 157 |
c2 = (float*) ((uintptr_t) c2 + cn_stride); |
| 158 |
c1 = (float*) ((uintptr_t) c1 + cn_stride); |
| 159 |
c0 = (float*) ((uintptr_t) c0 + cn_stride); |
| 160 |
|
| 161 |
nc -= 4; |
| 162 |
} else { |
| 163 |
if (nc & 2) { |
| 164 |
c3[0] = vacc3x0; |
| 165 |
c3[1] = vacc3x1; |
| 166 |
c2[0] = vacc2x0; |
| 167 |
c2[1] = vacc2x1; |
| 168 |
c1[0] = vacc1x0; |
| 169 |
c1[1] = vacc1x1; |
| 170 |
c0[0] = vacc0x0; |
| 171 |
c0[1] = vacc0x1; |
| 172 |
|
| 173 |
vacc3x0 = vacc3x2; |
| 174 |
vacc2x0 = vacc2x2; |
| 175 |
vacc1x0 = vacc1x2; |
| 176 |
vacc0x0 = vacc0x2; |
| 177 |
|
| 178 |
c3 += 2; |
| 179 |
c2 += 2; |
| 180 |
c1 += 2; |
| 181 |
c0 += 2; |
| 182 |
} |
| 183 |
if (nc & 1) { |
| 184 |
*c3 = vacc3x0; |
| 185 |
*c2 = vacc2x0; |
| 186 |
*c1 = vacc1x0; |
| 187 |
*c0 = vacc0x0; |
| 188 |
} |
| 189 |
|
| 190 |
nc = 0; |
| 191 |
} |
| 192 |
} while (nc != 0); |
| 193 |
} |