| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-f32-vcvt/sse2.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2021 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 <emmintrin.h> |
| 13 |
|
| 14 |
#include "xnnpack/common.h" |
| 15 |
#include "xnnpack/intrinsics-polyfill.h" |
| 16 |
#include "xnnpack/vcvt.h" |
| 17 |
|
| 18 |
|
| 19 |
void xnn_qu8_f32_vcvt_ukernel__sse2_u32( |
| 20 |
size_t batch, |
| 21 |
const uint8_t* input, |
| 22 |
float* output, |
| 23 |
const struct xnn_qu8_f32_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 24 |
{ |
| 25 |
assert(batch != 0); |
| 26 |
assert(batch % sizeof(uint8_t) == 0); |
| 27 |
assert(input != NULL); |
| 28 |
assert(output != NULL); |
| 29 |
|
| 30 |
const __m128i vmagic_exp = _mm_set1_epi16(UINT16_C(0x4B00)); |
| 31 |
const __m128 vmagic_bias = _mm_set1_ps((float) (INT32_C(0x00800000) + (int32_t) params->scalar.zero_point)); |
| 32 |
const __m128 vscale = _mm_set1_ps(params->scalar.scale); |
| 33 |
const __m128i vzero = _mm_setzero_si128(); |
| 34 |
XNN_FORCE_REALIZATION(vmagic_exp); |
| 35 |
XNN_FORCE_REALIZATION(vmagic_bias); |
| 36 |
XNN_FORCE_REALIZATION(vscale); |
| 37 |
for (; batch >= 32 * sizeof(uint8_t); batch -= 32 * sizeof(uint8_t)) { |
| 38 |
__m128i vx01234567 = _mm_loadl_epi64((const __m128i*) input); |
| 39 |
__m128i vx89ABCDEF = _mm_loadl_epi64((const __m128i*) (input + 8)); |
| 40 |
__m128i vxGHIJKLMN = _mm_loadl_epi64((const __m128i*) (input + 16)); |
| 41 |
__m128i vxOPQRSTUV = _mm_loadl_epi64((const __m128i*) (input + 24)); |
| 42 |
input += 32; |
| 43 |
|
| 44 |
|
| 45 |
vx01234567 = _mm_unpacklo_epi8(vx01234567, vzero); |
| 46 |
vx89ABCDEF = _mm_unpacklo_epi8(vx89ABCDEF, vzero); |
| 47 |
vxGHIJKLMN = _mm_unpacklo_epi8(vxGHIJKLMN, vzero); |
| 48 |
vxOPQRSTUV = _mm_unpacklo_epi8(vxOPQRSTUV, vzero); |
| 49 |
|
| 50 |
__m128 vy0123 = _mm_castsi128_ps(_mm_unpacklo_epi16(vx01234567, vmagic_exp)); |
| 51 |
__m128 vy4567 = _mm_castsi128_ps(_mm_unpackhi_epi16(vx01234567, vmagic_exp)); |
| 52 |
__m128 vy89AB = _mm_castsi128_ps(_mm_unpacklo_epi16(vx89ABCDEF, vmagic_exp)); |
| 53 |
__m128 vyCDEF = _mm_castsi128_ps(_mm_unpackhi_epi16(vx89ABCDEF, vmagic_exp)); |
| 54 |
__m128 vyGHIJ = _mm_castsi128_ps(_mm_unpacklo_epi16(vxGHIJKLMN, vmagic_exp)); |
| 55 |
__m128 vyKLMN = _mm_castsi128_ps(_mm_unpackhi_epi16(vxGHIJKLMN, vmagic_exp)); |
| 56 |
__m128 vyOPQR = _mm_castsi128_ps(_mm_unpacklo_epi16(vxOPQRSTUV, vmagic_exp)); |
| 57 |
__m128 vySTUV = _mm_castsi128_ps(_mm_unpackhi_epi16(vxOPQRSTUV, vmagic_exp)); |
| 58 |
|
| 59 |
vy0123 = _mm_sub_ps(vy0123, vmagic_bias); |
| 60 |
vy4567 = _mm_sub_ps(vy4567, vmagic_bias); |
| 61 |
vy89AB = _mm_sub_ps(vy89AB, vmagic_bias); |
| 62 |
vyCDEF = _mm_sub_ps(vyCDEF, vmagic_bias); |
| 63 |
vyGHIJ = _mm_sub_ps(vyGHIJ, vmagic_bias); |
| 64 |
vyKLMN = _mm_sub_ps(vyKLMN, vmagic_bias); |
| 65 |
vyOPQR = _mm_sub_ps(vyOPQR, vmagic_bias); |
| 66 |
vySTUV = _mm_sub_ps(vySTUV, vmagic_bias); |
| 67 |
|
| 68 |
vy0123 = _mm_mul_ps(vy0123, vscale); |
| 69 |
vy4567 = _mm_mul_ps(vy4567, vscale); |
| 70 |
vy89AB = _mm_mul_ps(vy89AB, vscale); |
| 71 |
vyCDEF = _mm_mul_ps(vyCDEF, vscale); |
| 72 |
vyGHIJ = _mm_mul_ps(vyGHIJ, vscale); |
| 73 |
vyKLMN = _mm_mul_ps(vyKLMN, vscale); |
| 74 |
vyOPQR = _mm_mul_ps(vyOPQR, vscale); |
| 75 |
vySTUV = _mm_mul_ps(vySTUV, vscale); |
| 76 |
|
| 77 |
_mm_storeu_ps(output, vy0123); |
| 78 |
_mm_storeu_ps(output + 4, vy4567); |
| 79 |
_mm_storeu_ps(output + 8, vy89AB); |
| 80 |
_mm_storeu_ps(output + 12, vyCDEF); |
| 81 |
_mm_storeu_ps(output + 16, vyGHIJ); |
| 82 |
_mm_storeu_ps(output + 20, vyKLMN); |
| 83 |
_mm_storeu_ps(output + 24, vyOPQR); |
| 84 |
_mm_storeu_ps(output + 28, vySTUV); |
| 85 |
output += 32; |
| 86 |
} |
| 87 |
for (; batch >= 8 * sizeof(uint8_t); batch -= 8 * sizeof(uint8_t)) { |
| 88 |
__m128i vx = _mm_loadl_epi64((const __m128i*) input); |
| 89 |
vx = _mm_unpacklo_epi8(vx, vzero); |
| 90 |
input += 8; |
| 91 |
|
| 92 |
__m128 vy_lo = _mm_castsi128_ps(_mm_unpacklo_epi16(vx, vmagic_exp)); |
| 93 |
__m128 vy_hi = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp)); |
| 94 |
|
| 95 |
vy_lo = _mm_sub_ps(vy_lo, vmagic_bias); |
| 96 |
vy_hi = _mm_sub_ps(vy_hi, vmagic_bias); |
| 97 |
|
| 98 |
vy_lo = _mm_mul_ps(vy_lo, vscale); |
| 99 |
vy_hi = _mm_mul_ps(vy_hi, vscale); |
| 100 |
|
| 101 |
_mm_storeu_ps(output, vy_lo); |
| 102 |
_mm_storeu_ps(output + 4, vy_hi); |
| 103 |
output += 8; |
| 104 |
} |
| 105 |
if XNN_UNLIKELY(batch != 0) { |
| 106 |
assert(batch >= 1 * sizeof(uint8_t)); |
| 107 |
assert(batch <= 7 * sizeof(uint8_t)); |
| 108 |
|
| 109 |
__m128i vx = _mm_loadl_epi64((const __m128i*) input); |
| 110 |
vx = _mm_unpacklo_epi8(vx, vzero); |
| 111 |
|
| 112 |
__m128 vy = _mm_castsi128_ps(_mm_unpacklo_epi16(vx, vmagic_exp)); |
| 113 |
vy = _mm_sub_ps(vy, vmagic_bias); |
| 114 |
vy = _mm_mul_ps(vy, vscale); |
| 115 |
|
| 116 |
if (batch & (4 * sizeof(uint8_t))) { |
| 117 |
_mm_storeu_ps(output, vy); |
| 118 |
vy = _mm_castsi128_ps(_mm_unpackhi_epi16(vx, vmagic_exp)); |
| 119 |
vy = _mm_sub_ps(vy, vmagic_bias); |
| 120 |
vy = _mm_mul_ps(vy, vscale); |
| 121 |
output += 4; |
| 122 |
} |
| 123 |
if (batch & (2 * sizeof(uint8_t))) { |
| 124 |
_mm_storel_pi((__m64*) output, vy); |
| 125 |
vy = _mm_movehl_ps(vy, vy); |
| 126 |
output += 2; |
| 127 |
} |
| 128 |
if (batch & (1 * sizeof(uint8_t))) { |
| 129 |
_mm_store_ss(output, vy); |
| 130 |
} |
| 131 |
} |
| 132 |
} |