| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-f32-vcvt/sse4.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 <immintrin.h> |
| 13 |
|
| 14 |
#include "xnnpack/common.h" |
| 15 |
#include "xnnpack/unaligned.h" |
| 16 |
#include "xnnpack/vcvt.h" |
| 17 |
|
| 18 |
|
| 19 |
void xnn_qu8_f32_vcvt_ukernel__sse41_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 vzero_point = _mm_set1_epi32(params->scalar.zero_point); |
| 31 |
const __m128 vscale = _mm_set1_ps(params->scalar.scale); |
| 32 |
XNN_FORCE_REALIZATION(vzero_point); |
| 33 |
XNN_FORCE_REALIZATION(vscale); |
| 34 |
for (; batch >= 32 * sizeof(uint8_t); batch -= 32 * sizeof(uint8_t)) { |
| 35 |
__m128i vx0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input))); |
| 36 |
__m128i vx4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input + 4))); |
| 37 |
__m128i vx89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input + 8))); |
| 38 |
__m128i vxCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input + 12))); |
| 39 |
__m128i vxGHIJ = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input + 16))); |
| 40 |
__m128i vxKLMN = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input + 20))); |
| 41 |
__m128i vxOPQR = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input + 24))); |
| 42 |
__m128i vxSTUV = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input + 28))); |
| 43 |
input += 32; |
| 44 |
|
| 45 |
vx0123 = _mm_sub_epi32(vx0123, vzero_point); |
| 46 |
vx4567 = _mm_sub_epi32(vx4567, vzero_point); |
| 47 |
vx89AB = _mm_sub_epi32(vx89AB, vzero_point); |
| 48 |
vxCDEF = _mm_sub_epi32(vxCDEF, vzero_point); |
| 49 |
vxGHIJ = _mm_sub_epi32(vxGHIJ, vzero_point); |
| 50 |
vxKLMN = _mm_sub_epi32(vxKLMN, vzero_point); |
| 51 |
vxOPQR = _mm_sub_epi32(vxOPQR, vzero_point); |
| 52 |
vxSTUV = _mm_sub_epi32(vxSTUV, vzero_point); |
| 53 |
|
| 54 |
__m128 vy0123 = _mm_cvtepi32_ps(vx0123); |
| 55 |
__m128 vy4567 = _mm_cvtepi32_ps(vx4567); |
| 56 |
__m128 vy89AB = _mm_cvtepi32_ps(vx89AB); |
| 57 |
__m128 vyCDEF = _mm_cvtepi32_ps(vxCDEF); |
| 58 |
__m128 vyGHIJ = _mm_cvtepi32_ps(vxGHIJ); |
| 59 |
__m128 vyKLMN = _mm_cvtepi32_ps(vxKLMN); |
| 60 |
__m128 vyOPQR = _mm_cvtepi32_ps(vxOPQR); |
| 61 |
__m128 vySTUV = _mm_cvtepi32_ps(vxSTUV); |
| 62 |
|
| 63 |
vy0123 = _mm_mul_ps(vy0123, vscale); |
| 64 |
vy4567 = _mm_mul_ps(vy4567, vscale); |
| 65 |
vy89AB = _mm_mul_ps(vy89AB, vscale); |
| 66 |
vyCDEF = _mm_mul_ps(vyCDEF, vscale); |
| 67 |
vyGHIJ = _mm_mul_ps(vyGHIJ, vscale); |
| 68 |
vyKLMN = _mm_mul_ps(vyKLMN, vscale); |
| 69 |
vyOPQR = _mm_mul_ps(vyOPQR, vscale); |
| 70 |
vySTUV = _mm_mul_ps(vySTUV, vscale); |
| 71 |
|
| 72 |
_mm_storeu_ps(output, vy0123); |
| 73 |
_mm_storeu_ps(output + 4, vy4567); |
| 74 |
_mm_storeu_ps(output + 8, vy89AB); |
| 75 |
_mm_storeu_ps(output + 12, vyCDEF); |
| 76 |
_mm_storeu_ps(output + 16, vyGHIJ); |
| 77 |
_mm_storeu_ps(output + 20, vyKLMN); |
| 78 |
_mm_storeu_ps(output + 24, vyOPQR); |
| 79 |
_mm_storeu_ps(output + 28, vySTUV); |
| 80 |
output += 32; |
| 81 |
} |
| 82 |
for (; batch >= 4 * sizeof(uint8_t); batch -= 4 * sizeof(uint8_t)) { |
| 83 |
__m128i vx = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input))); |
| 84 |
vx = _mm_sub_epi32(vx, vzero_point); |
| 85 |
input += 4; |
| 86 |
|
| 87 |
__m128 vy = _mm_cvtepi32_ps(vx); |
| 88 |
vy = _mm_mul_ps(vy, vscale); |
| 89 |
|
| 90 |
_mm_storeu_ps(output, vy); |
| 91 |
output += 4; |
| 92 |
} |
| 93 |
if XNN_UNLIKELY(batch != 0) { |
| 94 |
assert(batch >= 1 * sizeof(uint8_t)); |
| 95 |
assert(batch <= 3 * sizeof(uint8_t)); |
| 96 |
|
| 97 |
__m128i vx = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input))); |
| 98 |
vx = _mm_sub_epi32(vx, vzero_point); |
| 99 |
|
| 100 |
__m128 vy = _mm_cvtepi32_ps(vx); |
| 101 |
vy = _mm_mul_ps(vy, vscale); |
| 102 |
|
| 103 |
if (batch & (2 * sizeof(uint8_t))) { |
| 104 |
_mm_storel_pi((__m64*) output, vy); |
| 105 |
vy = _mm_movehl_ps(vy, vy); |
| 106 |
output += 2; |
| 107 |
} |
| 108 |
if (batch & (1 * sizeof(uint8_t))) { |
| 109 |
_mm_store_ss(output, vy); |
| 110 |
} |
| 111 |
} |
| 112 |
} |