| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-vcvt/sse4.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2022 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/intrinsics-polyfill.h" |
| 16 |
#include "xnnpack/vcvt.h" |
| 17 |
|
| 18 |
|
| 19 |
void xnn_qu8_vcvt_ukernel__sse41_u16( |
| 20 |
size_t batch, |
| 21 |
const uint8_t* input, |
| 22 |
uint8_t* output, |
| 23 |
const struct xnn_qu8_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 vinput_zero_point = _mm_set1_epi16(params->scalar.input_zero_point); |
| 31 |
const __m128i vmultiplier = _mm_set1_epi16(-params->scalar.multiplier); |
| 32 |
const __m128i voutput_zero_point = _mm_set1_epi16(params->scalar.output_zero_point); |
| 33 |
XNN_FORCE_REALIZATION(vinput_zero_point); |
| 34 |
XNN_FORCE_REALIZATION(vmultiplier); |
| 35 |
XNN_FORCE_REALIZATION(voutput_zero_point); |
| 36 |
for (; batch >= 16 * sizeof(uint8_t); batch -= 16 * sizeof(uint8_t)) { |
| 37 |
__m128i vacc0 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input)); |
| 38 |
__m128i vacc1 = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) (input + 8))); |
| 39 |
input += 16; |
| 40 |
|
| 41 |
vacc0 = _mm_sub_epi16(vinput_zero_point, vacc0); |
| 42 |
vacc1 = _mm_sub_epi16(vinput_zero_point, vacc1); |
| 43 |
|
| 44 |
vacc0 = _mm_slli_epi16(vacc0, 7); |
| 45 |
vacc1 = _mm_slli_epi16(vacc1, 7); |
| 46 |
|
| 47 |
vacc0 = _mm_mulhrs_epi16(vacc0, vmultiplier); |
| 48 |
vacc1 = _mm_mulhrs_epi16(vacc1, vmultiplier); |
| 49 |
|
| 50 |
vacc0 = _mm_adds_epi16(vacc0, voutput_zero_point); |
| 51 |
vacc1 = _mm_adds_epi16(vacc1, voutput_zero_point); |
| 52 |
|
| 53 |
const __m128i vy0 = _mm_packus_epi16(vacc0, vacc1); |
| 54 |
|
| 55 |
_mm_storeu_si128((__m128i*) output, vy0); |
| 56 |
output += 16; |
| 57 |
} |
| 58 |
for (; batch >= 8 * sizeof(uint8_t); batch -= 8 * sizeof(uint8_t)) { |
| 59 |
__m128i vacc = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input)); |
| 60 |
vacc = _mm_sub_epi16(vinput_zero_point, vacc); |
| 61 |
vacc = _mm_slli_epi16(vacc, 7); |
| 62 |
vacc = _mm_mulhrs_epi16(vacc, vmultiplier); |
| 63 |
vacc = _mm_adds_epi16(vacc, voutput_zero_point); |
| 64 |
input += 8; |
| 65 |
|
| 66 |
const __m128i vy = _mm_packus_epi16(vacc, vacc); |
| 67 |
_mm_storel_epi64((__m128i*) output, vy); |
| 68 |
output += 8; |
| 69 |
} |
| 70 |
if XNN_UNLIKELY(batch != 0) { |
| 71 |
assert(batch >= 1 * sizeof(uint8_t)); |
| 72 |
assert(batch <= 7 * sizeof(uint8_t)); |
| 73 |
|
| 74 |
__m128i vacc = _mm_cvtepu8_epi16(_mm_loadl_epi64((const __m128i*) input)); |
| 75 |
vacc = _mm_sub_epi16(vinput_zero_point, vacc); |
| 76 |
vacc = _mm_slli_epi16(vacc, 7); |
| 77 |
vacc = _mm_mulhrs_epi16(vacc, vmultiplier); |
| 78 |
vacc = _mm_adds_epi16(vacc, voutput_zero_point); |
| 79 |
|
| 80 |
__m128i vy = _mm_packus_epi16(vacc, vacc); |
| 81 |
if (batch & (4 * sizeof(uint8_t))) { |
| 82 |
_mm_storeu_si32(output, vy); |
| 83 |
vy = _mm_srli_epi64(vy, 32); |
| 84 |
output += 4; |
| 85 |
} |
| 86 |
if (batch & (2 * sizeof(uint8_t))) { |
| 87 |
_mm_storeu_si16(output, vy); |
| 88 |
vy = _mm_srli_epi32(vy, 16); |
| 89 |
output += 2; |
| 90 |
} |
| 91 |
if (batch & (1 * sizeof(uint8_t))) { |
| 92 |
*output = (uint8_t) _mm_extract_epi8(vy, 0); |
| 93 |
} |
| 94 |
} |
| 95 |
} |