| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/f32-vscaleexpminusmax/avx512f-p5-scalef.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/intrinsics-polyfill.h" |
| 15 |
#include "xnnpack/vscaleexpminusmax.h" |
| 16 |
|
| 17 |
|
| 18 |
void xnn_f32_vscaleexpminusmax_ukernel__avx512f_p5_scalef_u16( |
| 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 |
const __m512 vlog2e = _mm512_set1_ps(0x1.715476p+0f); |
| 31 |
const __m512 vminus_ln2_hi = _mm512_set1_ps(-0x1.62E43p-1f); |
| 32 |
const __m512 vminus_ln2_lo = _mm512_set1_ps(0x1.05C61p-29f); |
| 33 |
|
| 34 |
const __m512 vc0 = _mm512_set1_ps(1.0f); |
| 35 |
const __m512 vc1 = _mm512_set1_ps(0x1.FFFFF6p-1f); |
| 36 |
const __m512 vc2 = _mm512_set1_ps(0x1.FFFDC6p-2f); |
| 37 |
const __m512 vc3 = _mm512_set1_ps(0x1.555A80p-3f); |
| 38 |
const __m512 vc4 = _mm512_set1_ps(0x1.573A1Ap-5f); |
| 39 |
const __m512 vc5 = _mm512_set1_ps(0x1.0F9F9Cp-7f); |
| 40 |
|
| 41 |
const __m512 vscale = _mm512_set1_ps(scale); |
| 42 |
const __m512 vi_max = _mm512_set1_ps(max); |
| 43 |
|
| 44 |
for (; batch >= 16 * sizeof(float); batch -= 16 * sizeof(float)) { |
| 45 |
// Load 16 (1x16) inputs at a time. |
| 46 |
const __m512 vi0 = _mm512_loadu_ps(input); |
| 47 |
input += 16; |
| 48 |
|
| 49 |
// Subtract maximum input x := i - i_max. |
| 50 |
const __m512 vx0 = _mm512_sub_ps(vi0, vi_max); |
| 51 |
|
| 52 |
// Compute reduced argument batch := round(x / log(2)). |
| 53 |
__m512 vn0 = _mm512_roundscale_ps(_mm512_mul_ps(vx0, vlog2e), 0); |
| 54 |
|
| 55 |
// Compute reduced argument t := x - batch * log(2). |
| 56 |
// Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy. |
| 57 |
__m512 vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_hi, vx0); |
| 58 |
|
| 59 |
vt0 = _mm512_fmadd_ps(vn0, vminus_ln2_lo, vt0); |
| 60 |
|
| 61 |
// Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2]. |
| 62 |
__m512 vp0 = _mm512_fmadd_ps(vc5, vt0, vc4); |
| 63 |
|
| 64 |
vp0 = _mm512_fmadd_ps(vp0, vt0, vc3); |
| 65 |
|
| 66 |
vp0 = _mm512_fmadd_ps(vp0, vt0, vc2); |
| 67 |
|
| 68 |
vp0 = _mm512_fmadd_ps(vp0, vt0, vc1); |
| 69 |
|
| 70 |
vp0 = _mm512_fmadd_ps(vp0, vt0, vc0); |
| 71 |
|
| 72 |
// Reconstruct the final f value: |
| 73 |
// f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))) |
| 74 |
// = 2**batch * p |
| 75 |
__m512 vf0 = _mm512_scalef_ps(vp0, vn0); |
| 76 |
|
| 77 |
// Multiply by scale. |
| 78 |
vf0 = _mm512_mul_ps(vf0, vscale); |
| 79 |
|
| 80 |
// Store 16 (1x16) outputs at a time. |
| 81 |
_mm512_storeu_ps(output, vf0); |
| 82 |
_mm512_storeu_ps(output + 0, vf0); |
| 83 |
output += 16; |
| 84 |
} |
| 85 |
for (; batch >= 16 * sizeof(float); batch -= 16 * sizeof(float)) { |
| 86 |
// Load 16 inputs at a time. |
| 87 |
const __m512 vi = _mm512_loadu_ps(input); |
| 88 |
input += 16; |
| 89 |
|
| 90 |
// Subtract maximum input x := i - i_max. |
| 91 |
const __m512 vx = _mm512_sub_ps(vi, vi_max); |
| 92 |
|
| 93 |
// Compute reduced argument batch := round(x / log(2)). |
| 94 |
__m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0); |
| 95 |
|
| 96 |
// Compute reduced argument t := x - batch * log(2). |
| 97 |
// Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy. |
| 98 |
__m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx); |
| 99 |
vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt); |
| 100 |
|
| 101 |
// Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2]. |
| 102 |
__m512 vp = _mm512_fmadd_ps(vc5, vt, vc4); |
| 103 |
vp = _mm512_fmadd_ps(vp, vt, vc3); |
| 104 |
vp = _mm512_fmadd_ps(vp, vt, vc2); |
| 105 |
vp = _mm512_fmadd_ps(vp, vt, vc1); |
| 106 |
vp = _mm512_fmadd_ps(vp, vt, vc0); |
| 107 |
|
| 108 |
// Reconstruct the final f value: |
| 109 |
// f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))) |
| 110 |
// = 2**batch * p |
| 111 |
__m512 vf = _mm512_scalef_ps(vp, vn); |
| 112 |
|
| 113 |
// Multiply by scale. |
| 114 |
vf = _mm512_mul_ps(vf, vscale); |
| 115 |
|
| 116 |
// Store 16 outputs at a time. |
| 117 |
_mm512_storeu_ps(output, vf); |
| 118 |
output += 16; |
| 119 |
} |
| 120 |
if (batch != 0) { |
| 121 |
// Prepare mask for valid 32-bit batch (depends on batch). |
| 122 |
batch >>= XNN_LOG2_SIZEOF_FLOAT; |
| 123 |
const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << batch) - UINT32_C(1))); |
| 124 |
|
| 125 |
// Load up to 15 inputs at a time. |
| 126 |
const __m512 vi = _mm512_mask_loadu_ps(_mm512_undefined_ps(), vmask, input); |
| 127 |
|
| 128 |
// Subtract maximum input x := i - i_max. |
| 129 |
const __m512 vx = _mm512_sub_ps(vi, vi_max); |
| 130 |
|
| 131 |
// Compute reduced argument batch := round(x / log(2)). |
| 132 |
__m512 vn = _mm512_roundscale_ps(_mm512_mul_ps(vx, vlog2e), 0); |
| 133 |
|
| 134 |
// Compute reduced argument t := x - batch * log(2). |
| 135 |
// Use Cody-Waite range reduction method (note two constants to represent log(2)) to improve accuracy. |
| 136 |
__m512 vt = _mm512_fmadd_ps(vn, vminus_ln2_hi, vx); |
| 137 |
vt = _mm512_fmadd_ps(vn, vminus_ln2_lo, vt); |
| 138 |
|
| 139 |
// Compute degree-5 polynomial approximation for exp(t) on [-log(2)/2, log(2)/2]. |
| 140 |
__m512 vp = _mm512_fmadd_ps(vc5, vt, vc4); |
| 141 |
vp = _mm512_fmadd_ps(vp, vt, vc3); |
| 142 |
vp = _mm512_fmadd_ps(vp, vt, vc2); |
| 143 |
vp = _mm512_fmadd_ps(vp, vt, vc1); |
| 144 |
vp = _mm512_fmadd_ps(vp, vt, vc0); |
| 145 |
|
| 146 |
// Reconstruct the final f value: |
| 147 |
// f = 2**batch * (1 + t * (c1 + t * (c2 + t * (c3 + t * (c4 + t * c5))))) |
| 148 |
// = 2**batch * p |
| 149 |
__m512 vf = _mm512_scalef_ps(vp, vn); |
| 150 |
|
| 151 |
// Multiply by scale. |
| 152 |
vf = _mm512_mul_ps(vf, vscale); |
| 153 |
|
| 154 |
// Store up to 15 outputs at a time. |
| 155 |
_mm512_mask_storeu_ps(output, vmask, vf); |
| 156 |
} |
| 157 |
} |