| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/f32-vscaleexpminusmax/avx2-p5.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 <immintrin.h> |
| 13 |
|
| 14 |
#include "xnnpack/common.h" |
| 15 |
#include "xnnpack/vscaleexpminusmax.h" |
| 16 |
|
| 17 |
|
| 18 |
void xnn_f32_vscaleexpminusmax_ukernel__avx2_p5_u80( |
| 19 |
size_t batch, |
| 20 |
const float* input, |
| 21 |
float* output, |
| 22 |
float scale, |
| 23 |
float max) |
| 24 |
{ |
| 25 |
assert(batch != 0); |
| 26 |
assert(batch % sizeof(float) == 0); |
| 27 |
assert(input != NULL); |
| 28 |
assert(output != NULL); |
| 29 |
|
| 30 |
static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0}; |
| 31 |
|
| 32 |
const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f); |
| 33 |
// The smallest x for which expf(x) is normalized. |
| 34 |
const __m256 vdenorm_cutoff = _mm256_set1_ps(-0x1.5D589Ep6f); |
| 35 |
const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f); |
| 36 |
const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E43p-1f); |
| 37 |
const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.05C61p-29f); |
| 38 |
|
| 39 |
const __m256 vc1 = _mm256_set1_ps(0x1.FFFFF6p-1f); |
| 40 |
const __m256 vc2 = _mm256_set1_ps(0x1.FFFDC6p-2f); |
| 41 |
const __m256 vc3 = _mm256_set1_ps(0x1.555A80p-3f); |
| 42 |
const __m256 vc4 = _mm256_set1_ps(0x1.573A1Ap-5f); |
| 43 |
const __m256 vc5 = _mm256_set1_ps(0x1.0F9F9Cp-7f); |
| 44 |
|
| 45 |
const __m256 vscale = _mm256_set1_ps(scale); |
| 46 |
const __m256 vi_max = _mm256_set1_ps(max); |
| 47 |
|
| 48 |
for (; batch >= 80 * sizeof(float); batch -= 80 * sizeof(float)) { |
| 49 |
// Load 80 (10x8) inputs at a time. |
| 50 |
const __m256 vi0 = _mm256_loadu_ps(input); |
| 51 |
const __m256 vi1 = _mm256_loadu_ps(input + 8); |
| 52 |
const __m256 vi2 = _mm256_loadu_ps(input + 16); |
| 53 |
const __m256 vi3 = _mm256_loadu_ps(input + 24); |
| 54 |
const __m256 vi4 = _mm256_loadu_ps(input + 32); |
| 55 |
const __m256 vi5 = _mm256_loadu_ps(input + 40); |
| 56 |
const __m256 vi6 = _mm256_loadu_ps(input + 48); |
| 57 |
const __m256 vi7 = _mm256_loadu_ps(input + 56); |
| 58 |
const __m256 vi8 = _mm256_loadu_ps(input + 64); |
| 59 |
const __m256 vi9 = _mm256_loadu_ps(input + 72); |
| 60 |
input += 80; |
| 61 |
|
| 62 |
// Subtract maximum input x := i - i_max. This implies x <= 0. |
| 63 |
const __m256 vx0 = _mm256_sub_ps(vi0, vi_max); |
| 64 |
const __m256 vx1 = _mm256_sub_ps(vi1, vi_max); |
| 65 |
const __m256 vx2 = _mm256_sub_ps(vi2, vi_max); |
| 66 |
const __m256 vx3 = _mm256_sub_ps(vi3, vi_max); |
| 67 |
const __m256 vx4 = _mm256_sub_ps(vi4, vi_max); |
| 68 |
const __m256 vx5 = _mm256_sub_ps(vi5, vi_max); |
| 69 |
const __m256 vx6 = _mm256_sub_ps(vi6, vi_max); |
| 70 |
const __m256 vx7 = _mm256_sub_ps(vi7, vi_max); |
| 71 |
const __m256 vx8 = _mm256_sub_ps(vi8, vi_max); |
| 72 |
const __m256 vx9 = _mm256_sub_ps(vi9, vi_max); |
| 73 |
|
| 74 |
// Compute reduced argument batch := round(x / log(2)). |
| 75 |
__m256 vn0 = _mm256_fmadd_ps(vx0, vlog2e, vmagic_bias); |
| 76 |
__m256 vn1 = _mm256_fmadd_ps(vx1, vlog2e, vmagic_bias); |
| 77 |
__m256 vn2 = _mm256_fmadd_ps(vx2, vlog2e, vmagic_bias); |
| 78 |
__m256 vn3 = _mm256_fmadd_ps(vx3, vlog2e, vmagic_bias); |
| 79 |
__m256 vn4 = _mm256_fmadd_ps(vx4, vlog2e, vmagic_bias); |
| 80 |
__m256 vn5 = _mm256_fmadd_ps(vx5, vlog2e, vmagic_bias); |
| 81 |
__m256 vn6 = _mm256_fmadd_ps(vx6, vlog2e, vmagic_bias); |
| 82 |
__m256 vn7 = _mm256_fmadd_ps(vx7, vlog2e, vmagic_bias); |
| 83 |
__m256 vn8 = _mm256_fmadd_ps(vx8, vlog2e, vmagic_bias); |
| 84 |
__m256 vn9 = _mm256_fmadd_ps(vx9, vlog2e, vmagic_bias); |
| 85 |
|
| 86 |
// Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e. |
| 87 |
// -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly. |
| 88 |
const __m256 vs0 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn0), 23)); |
| 89 |
const __m256 vs1 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn1), 23)); |
| 90 |
const __m256 vs2 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn2), 23)); |
| 91 |
const __m256 vs3 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn3), 23)); |
| 92 |
const __m256 vs4 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn4), 23)); |
| 93 |
const __m256 vs5 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn5), 23)); |
| 94 |
const __m256 vs6 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn6), 23)); |
| 95 |
const __m256 vs7 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn7), 23)); |
| 96 |
const __m256 vs8 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn8), 23)); |
| 97 |
const __m256 vs9 = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn9), 23)); |
| 98 |
|
| 99 |
// Subtract the large number back to get final batch := round(x / log(2)). |
| 100 |
vn0 = _mm256_sub_ps(vn0, vmagic_bias); |
| 101 |
vn1 = _mm256_sub_ps(vn1, vmagic_bias); |
| 102 |
vn2 = _mm256_sub_ps(vn2, vmagic_bias); |
| 103 |
vn3 = _mm256_sub_ps(vn3, vmagic_bias); |
| 104 |
vn4 = _mm256_sub_ps(vn4, vmagic_bias); |
| 105 |
vn5 = _mm256_sub_ps(vn5, vmagic_bias); |
| 106 |
vn6 = _mm256_sub_ps(vn6, vmagic_bias); |
| 107 |
vn7 = _mm256_sub_ps(vn7, vmagic_bias); |
| 108 |
vn8 = _mm256_sub_ps(vn8, vmagic_bias); |
| 109 |
vn9 = _mm256_sub_ps(vn9, vmagic_bias); |
| 110 |
|
| 111 |
// Compute reduced argument t := x - batch * log(2). |
| 112 |
// Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy. |
| 113 |
__m256 vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_hi, vx0); |
| 114 |
__m256 vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_hi, vx1); |
| 115 |
__m256 vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_hi, vx2); |
| 116 |
__m256 vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_hi, vx3); |
| 117 |
__m256 vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_hi, vx4); |
| 118 |
__m256 vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_hi, vx5); |
| 119 |
__m256 vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_hi, vx6); |
| 120 |
__m256 vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_hi, vx7); |
| 121 |
__m256 vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_hi, vx8); |
| 122 |
__m256 vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_hi, vx9); |
| 123 |
|
| 124 |
vt0 = _mm256_fmadd_ps(vn0, vminus_ln2_lo, vt0); |
| 125 |
vt1 = _mm256_fmadd_ps(vn1, vminus_ln2_lo, vt1); |
| 126 |
vt2 = _mm256_fmadd_ps(vn2, vminus_ln2_lo, vt2); |
| 127 |
vt3 = _mm256_fmadd_ps(vn3, vminus_ln2_lo, vt3); |
| 128 |
vt4 = _mm256_fmadd_ps(vn4, vminus_ln2_lo, vt4); |
| 129 |
vt5 = _mm256_fmadd_ps(vn5, vminus_ln2_lo, vt5); |
| 130 |
vt6 = _mm256_fmadd_ps(vn6, vminus_ln2_lo, vt6); |
| 131 |
vt7 = _mm256_fmadd_ps(vn7, vminus_ln2_lo, vt7); |
| 132 |
vt8 = _mm256_fmadd_ps(vn8, vminus_ln2_lo, vt8); |
| 133 |
vt9 = _mm256_fmadd_ps(vn9, vminus_ln2_lo, vt9); |
| 134 |
|
| 135 |
// Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2]. |
| 136 |
__m256 vp0 = _mm256_fmadd_ps(vc5, vt0, vc4); |
| 137 |
__m256 vp1 = _mm256_fmadd_ps(vc5, vt1, vc4); |
| 138 |
__m256 vp2 = _mm256_fmadd_ps(vc5, vt2, vc4); |
| 139 |
__m256 vp3 = _mm256_fmadd_ps(vc5, vt3, vc4); |
| 140 |
__m256 vp4 = _mm256_fmadd_ps(vc5, vt4, vc4); |
| 141 |
__m256 vp5 = _mm256_fmadd_ps(vc5, vt5, vc4); |
| 142 |
__m256 vp6 = _mm256_fmadd_ps(vc5, vt6, vc4); |
| 143 |
__m256 vp7 = _mm256_fmadd_ps(vc5, vt7, vc4); |
| 144 |
__m256 vp8 = _mm256_fmadd_ps(vc5, vt8, vc4); |
| 145 |
__m256 vp9 = _mm256_fmadd_ps(vc5, vt9, vc4); |
| 146 |
|
| 147 |
vp0 = _mm256_fmadd_ps(vp0, vt0, vc3); |
| 148 |
vp1 = _mm256_fmadd_ps(vp1, vt1, vc3); |
| 149 |
vp2 = _mm256_fmadd_ps(vp2, vt2, vc3); |
| 150 |
vp3 = _mm256_fmadd_ps(vp3, vt3, vc3); |
| 151 |
vp4 = _mm256_fmadd_ps(vp4, vt4, vc3); |
| 152 |
vp5 = _mm256_fmadd_ps(vp5, vt5, vc3); |
| 153 |
vp6 = _mm256_fmadd_ps(vp6, vt6, vc3); |
| 154 |
vp7 = _mm256_fmadd_ps(vp7, vt7, vc3); |
| 155 |
vp8 = _mm256_fmadd_ps(vp8, vt8, vc3); |
| 156 |
vp9 = _mm256_fmadd_ps(vp9, vt9, vc3); |
| 157 |
|
| 158 |
vp0 = _mm256_fmadd_ps(vp0, vt0, vc2); |
| 159 |
vp1 = _mm256_fmadd_ps(vp1, vt1, vc2); |
| 160 |
vp2 = _mm256_fmadd_ps(vp2, vt2, vc2); |
| 161 |
vp3 = _mm256_fmadd_ps(vp3, vt3, vc2); |
| 162 |
vp4 = _mm256_fmadd_ps(vp4, vt4, vc2); |
| 163 |
vp5 = _mm256_fmadd_ps(vp5, vt5, vc2); |
| 164 |
vp6 = _mm256_fmadd_ps(vp6, vt6, vc2); |
| 165 |
vp7 = _mm256_fmadd_ps(vp7, vt7, vc2); |
| 166 |
vp8 = _mm256_fmadd_ps(vp8, vt8, vc2); |
| 167 |
vp9 = _mm256_fmadd_ps(vp9, vt9, vc2); |
| 168 |
|
| 169 |
vp0 = _mm256_fmadd_ps(vp0, vt0, vc1); |
| 170 |
vp1 = _mm256_fmadd_ps(vp1, vt1, vc1); |
| 171 |
vp2 = _mm256_fmadd_ps(vp2, vt2, vc1); |
| 172 |
vp3 = _mm256_fmadd_ps(vp3, vt3, vc1); |
| 173 |
vp4 = _mm256_fmadd_ps(vp4, vt4, vc1); |
| 174 |
vp5 = _mm256_fmadd_ps(vp5, vt5, vc1); |
| 175 |
vp6 = _mm256_fmadd_ps(vp6, vt6, vc1); |
| 176 |
vp7 = _mm256_fmadd_ps(vp7, vt7, vc1); |
| 177 |
vp8 = _mm256_fmadd_ps(vp8, vt8, vc1); |
| 178 |
vp9 = _mm256_fmadd_ps(vp9, vt9, vc1); |
| 179 |
|
| 180 |
// Reconstruct the final f value: |
| 181 |
// f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))) |
| 182 |
// = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))) |
| 183 |
// = s + (t * s) * p |
| 184 |
vt0 = _mm256_mul_ps(vt0, vs0); |
| 185 |
vt1 = _mm256_mul_ps(vt1, vs1); |
| 186 |
vt2 = _mm256_mul_ps(vt2, vs2); |
| 187 |
vt3 = _mm256_mul_ps(vt3, vs3); |
| 188 |
vt4 = _mm256_mul_ps(vt4, vs4); |
| 189 |
vt5 = _mm256_mul_ps(vt5, vs5); |
| 190 |
vt6 = _mm256_mul_ps(vt6, vs6); |
| 191 |
vt7 = _mm256_mul_ps(vt7, vs7); |
| 192 |
vt8 = _mm256_mul_ps(vt8, vs8); |
| 193 |
vt9 = _mm256_mul_ps(vt9, vs9); |
| 194 |
|
| 195 |
__m256 vf0 = _mm256_fmadd_ps(vt0, vp0, vs0); |
| 196 |
__m256 vf1 = _mm256_fmadd_ps(vt1, vp1, vs1); |
| 197 |
__m256 vf2 = _mm256_fmadd_ps(vt2, vp2, vs2); |
| 198 |
__m256 vf3 = _mm256_fmadd_ps(vt3, vp3, vs3); |
| 199 |
__m256 vf4 = _mm256_fmadd_ps(vt4, vp4, vs4); |
| 200 |
__m256 vf5 = _mm256_fmadd_ps(vt5, vp5, vs5); |
| 201 |
__m256 vf6 = _mm256_fmadd_ps(vt6, vp6, vs6); |
| 202 |
__m256 vf7 = _mm256_fmadd_ps(vt7, vp7, vs7); |
| 203 |
__m256 vf8 = _mm256_fmadd_ps(vt8, vp8, vs8); |
| 204 |
__m256 vf9 = _mm256_fmadd_ps(vt9, vp9, vs9); |
| 205 |
|
| 206 |
// For inputs below zero cutoff, replace output with +0.0f. |
| 207 |
// Note that for NaN inputs, comparison result is false, and outputs are left unchanged. |
| 208 |
vf0 = _mm256_andnot_ps(_mm256_cmp_ps(vx0, vdenorm_cutoff, _CMP_LT_OS), vf0); |
| 209 |
vf1 = _mm256_andnot_ps(_mm256_cmp_ps(vx1, vdenorm_cutoff, _CMP_LT_OS), vf1); |
| 210 |
vf2 = _mm256_andnot_ps(_mm256_cmp_ps(vx2, vdenorm_cutoff, _CMP_LT_OS), vf2); |
| 211 |
vf3 = _mm256_andnot_ps(_mm256_cmp_ps(vx3, vdenorm_cutoff, _CMP_LT_OS), vf3); |
| 212 |
vf4 = _mm256_andnot_ps(_mm256_cmp_ps(vx4, vdenorm_cutoff, _CMP_LT_OS), vf4); |
| 213 |
vf5 = _mm256_andnot_ps(_mm256_cmp_ps(vx5, vdenorm_cutoff, _CMP_LT_OS), vf5); |
| 214 |
vf6 = _mm256_andnot_ps(_mm256_cmp_ps(vx6, vdenorm_cutoff, _CMP_LT_OS), vf6); |
| 215 |
vf7 = _mm256_andnot_ps(_mm256_cmp_ps(vx7, vdenorm_cutoff, _CMP_LT_OS), vf7); |
| 216 |
vf8 = _mm256_andnot_ps(_mm256_cmp_ps(vx8, vdenorm_cutoff, _CMP_LT_OS), vf8); |
| 217 |
vf9 = _mm256_andnot_ps(_mm256_cmp_ps(vx9, vdenorm_cutoff, _CMP_LT_OS), vf9); |
| 218 |
|
| 219 |
// Multiply by scale. |
| 220 |
vf0 = _mm256_mul_ps(vf0, vscale); |
| 221 |
vf1 = _mm256_mul_ps(vf1, vscale); |
| 222 |
vf2 = _mm256_mul_ps(vf2, vscale); |
| 223 |
vf3 = _mm256_mul_ps(vf3, vscale); |
| 224 |
vf4 = _mm256_mul_ps(vf4, vscale); |
| 225 |
vf5 = _mm256_mul_ps(vf5, vscale); |
| 226 |
vf6 = _mm256_mul_ps(vf6, vscale); |
| 227 |
vf7 = _mm256_mul_ps(vf7, vscale); |
| 228 |
vf8 = _mm256_mul_ps(vf8, vscale); |
| 229 |
vf9 = _mm256_mul_ps(vf9, vscale); |
| 230 |
|
| 231 |
// Store 80 (10x8) outputs at a time. |
| 232 |
_mm256_storeu_ps(output, vf0); |
| 233 |
_mm256_storeu_ps(output + 8, vf1); |
| 234 |
_mm256_storeu_ps(output + 16, vf2); |
| 235 |
_mm256_storeu_ps(output + 24, vf3); |
| 236 |
_mm256_storeu_ps(output + 32, vf4); |
| 237 |
_mm256_storeu_ps(output + 40, vf5); |
| 238 |
_mm256_storeu_ps(output + 48, vf6); |
| 239 |
_mm256_storeu_ps(output + 56, vf7); |
| 240 |
_mm256_storeu_ps(output + 64, vf8); |
| 241 |
_mm256_storeu_ps(output + 72, vf9); |
| 242 |
output += 80; |
| 243 |
} |
| 244 |
for (; batch >= 8 * sizeof(float); batch -= 8 * sizeof(float)) { |
| 245 |
// Load 8 inputs at a time. |
| 246 |
const __m256 vi = _mm256_loadu_ps(input); |
| 247 |
input += 8; |
| 248 |
|
| 249 |
// Subtract maximum input x := i - i_max. This implies x <= 0. |
| 250 |
const __m256 vx = _mm256_sub_ps(vi, vi_max); |
| 251 |
|
| 252 |
// Compute reduced argument batch := round(x / log(2)). |
| 253 |
__m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias); |
| 254 |
|
| 255 |
// Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e. |
| 256 |
// -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly. |
| 257 |
const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23)); |
| 258 |
|
| 259 |
// Subtract the large number back to get final batch := round(x / log(2)). |
| 260 |
vn = _mm256_sub_ps(vn, vmagic_bias); |
| 261 |
|
| 262 |
// Compute reduced argument t := x - batch * log(2). |
| 263 |
// Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy. |
| 264 |
__m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx); |
| 265 |
vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt); |
| 266 |
|
| 267 |
// Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2]. |
| 268 |
__m256 vp = _mm256_fmadd_ps(vc5, vt, vc4); |
| 269 |
vp = _mm256_fmadd_ps(vp, vt, vc3); |
| 270 |
vp = _mm256_fmadd_ps(vp, vt, vc2); |
| 271 |
vp = _mm256_fmadd_ps(vp, vt, vc1); |
| 272 |
|
| 273 |
// Reconstruct the final f value: |
| 274 |
// f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))) |
| 275 |
// = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))) |
| 276 |
// = s + (t * s) * p |
| 277 |
vt = _mm256_mul_ps(vt, vs); |
| 278 |
__m256 vf = _mm256_fmadd_ps(vt, vp, vs); |
| 279 |
|
| 280 |
// For inputs below zero cutoff, replace output with +0.0f. |
| 281 |
// Note that for NaN inputs, comparison result is false, and outputs are left unchanged. |
| 282 |
vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf); |
| 283 |
|
| 284 |
// Multiply by scale. |
| 285 |
vf = _mm256_mul_ps(vf, vscale); |
| 286 |
|
| 287 |
// Store 64 (8x8) outputs at a time. |
| 288 |
_mm256_storeu_ps(output, vf); |
| 289 |
output += 8; |
| 290 |
} |
| 291 |
if (batch != 0) { |
| 292 |
assert(batch >= 1 * sizeof(float)); |
| 293 |
assert(batch <= 7 * sizeof(float)); |
| 294 |
const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - batch)); |
| 295 |
|
| 296 |
// Load up to 7 inputs at a time. |
| 297 |
const __m256 vi = _mm256_maskload_ps(input, vmask); |
| 298 |
|
| 299 |
// Subtract maximum input x := i - i_max. This implies x <= 0. |
| 300 |
const __m256 vx = _mm256_sub_ps(vi, vi_max); |
| 301 |
|
| 302 |
// Compute reduced argument batch := round(x / log(2)). |
| 303 |
__m256 vn = _mm256_fmadd_ps(vx, vlog2e, vmagic_bias); |
| 304 |
|
| 305 |
// Create a floating-point number s (scale) such that s == 2**batch for inputs which don't cause underflow, i.e. |
| 306 |
// -87.33642 <= x <= 0.0, and -126 <= batch <= 0 accordingly. |
| 307 |
const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23)); |
| 308 |
|
| 309 |
// Subtract the large number back to get final batch := round(x / log(2)). |
| 310 |
vn = _mm256_sub_ps(vn, vmagic_bias); |
| 311 |
|
| 312 |
// Compute reduced argument t := x - batch * log(2). |
| 313 |
// Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy. |
| 314 |
__m256 vt = _mm256_fmadd_ps(vn, vminus_ln2_hi, vx); |
| 315 |
vt = _mm256_fmadd_ps(vn, vminus_ln2_lo, vt); |
| 316 |
|
| 317 |
// Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2]. |
| 318 |
__m256 vp = _mm256_fmadd_ps(vc5, vt, vc4); |
| 319 |
vp = _mm256_fmadd_ps(vp, vt, vc3); |
| 320 |
vp = _mm256_fmadd_ps(vp, vt, vc2); |
| 321 |
vp = _mm256_fmadd_ps(vp, vt, vc1); |
| 322 |
|
| 323 |
// Reconstruct the final f value: |
| 324 |
// f = s * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))) |
| 325 |
// = s + (t * s) * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5)))) |
| 326 |
// = s + (t * s) * p |
| 327 |
vt = _mm256_mul_ps(vt, vs); |
| 328 |
__m256 vf = _mm256_fmadd_ps(vt, vp, vs); |
| 329 |
|
| 330 |
// For inputs below zero cutoff, replace output with +0.0f. |
| 331 |
// Note that for NaN inputs, comparison result is false, and outputs are left unchanged. |
| 332 |
vf = _mm256_andnot_ps(_mm256_cmp_ps(vx, vdenorm_cutoff, _CMP_LT_OS), vf); |
| 333 |
|
| 334 |
// Multiply by scale. |
| 335 |
vf = _mm256_mul_ps(vf, vscale); |
| 336 |
|
| 337 |
// Store up to 7 outputs at a time. |
| 338 |
_mm256_maskstore_ps(output, vmask, vf); |
| 339 |
} |
| 340 |
} |