| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-vadd/sse-mul32-ld32.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2020 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/unaligned.h" |
| 17 |
#include "xnnpack/vbinary.h" |
| 18 |
|
| 19 |
|
| 20 |
void xnn_qu8_vadd_minmax_ukernel__sse41_mul32_ld32_u16( |
| 21 |
size_t batch, |
| 22 |
const uint8_t* input_a, |
| 23 |
const uint8_t* input_b, |
| 24 |
uint8_t* output, |
| 25 |
const struct xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 26 |
{ |
| 27 |
assert(batch != 0); |
| 28 |
assert(batch % sizeof(uint8_t) == 0); |
| 29 |
assert(input_a != NULL); |
| 30 |
assert(input_b != NULL); |
| 31 |
assert(output != NULL); |
| 32 |
|
| 33 |
const __m128i vbias = _mm_set1_epi32(params->scalar.bias); |
| 34 |
const __m128i va_multiplier = _mm_set1_epi32(params->scalar.a_multiplier); |
| 35 |
const __m128i vb_multiplier = _mm_set1_epi32(params->scalar.b_multiplier); |
| 36 |
const __m128i vshift = _mm_cvtsi32_si128((int) params->scalar.shift); |
| 37 |
const __m128i voutput_zero_point = _mm_set1_epi16(params->scalar.output_zero_point); |
| 38 |
const __m128i voutput_min = _mm_set1_epi8(params->scalar.output_min); |
| 39 |
const __m128i voutput_max = _mm_set1_epi8(params->scalar.output_max); |
| 40 |
|
| 41 |
XNN_FORCE_REALIZATION(vbias); |
| 42 |
XNN_FORCE_REALIZATION(va_multiplier); |
| 43 |
XNN_FORCE_REALIZATION(vb_multiplier); |
| 44 |
XNN_FORCE_REALIZATION(vshift); |
| 45 |
XNN_FORCE_REALIZATION(voutput_zero_point); |
| 46 |
XNN_FORCE_REALIZATION(voutput_min); |
| 47 |
XNN_FORCE_REALIZATION(voutput_max); |
| 48 |
|
| 49 |
for (; batch >= 16 * sizeof(uint8_t); batch -= 16 * sizeof(uint8_t)) { |
| 50 |
const __m128i va0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a))); |
| 51 |
const __m128i vb0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b))); |
| 52 |
const __m128i va4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 4))); |
| 53 |
const __m128i vb4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 4))); |
| 54 |
const __m128i va89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 8))); |
| 55 |
const __m128i vb89AB = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 8))); |
| 56 |
const __m128i vaCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 12))); |
| 57 |
const __m128i vbCDEF = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 12))); |
| 58 |
input_a += 16; |
| 59 |
input_b += 16; |
| 60 |
|
| 61 |
__m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier)); |
| 62 |
__m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier)); |
| 63 |
__m128i vacc89AB = _mm_add_epi32(vbias, _mm_mullo_epi32(va89AB, va_multiplier)); |
| 64 |
__m128i vaccCDEF = _mm_add_epi32(vbias, _mm_mullo_epi32(vaCDEF, va_multiplier)); |
| 65 |
|
| 66 |
vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vb0123, vb_multiplier)); |
| 67 |
vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vb4567, vb_multiplier)); |
| 68 |
vacc89AB = _mm_add_epi32(vacc89AB, _mm_mullo_epi32(vb89AB, vb_multiplier)); |
| 69 |
vaccCDEF = _mm_add_epi32(vaccCDEF, _mm_mullo_epi32(vbCDEF, vb_multiplier)); |
| 70 |
|
| 71 |
vacc0123 = _mm_sra_epi32(vacc0123, vshift); |
| 72 |
vacc4567 = _mm_sra_epi32(vacc4567, vshift); |
| 73 |
vacc89AB = _mm_sra_epi32(vacc89AB, vshift); |
| 74 |
vaccCDEF = _mm_sra_epi32(vaccCDEF, vshift); |
| 75 |
|
| 76 |
const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); |
| 77 |
const __m128i vout89ABCDEF = _mm_adds_epi16(_mm_packs_epi32(vacc89AB, vaccCDEF), voutput_zero_point); |
| 78 |
|
| 79 |
__m128i vout0123456789ABCDEF = _mm_packus_epi16(vout01234567, vout89ABCDEF); |
| 80 |
|
| 81 |
vout0123456789ABCDEF = _mm_max_epu8(vout0123456789ABCDEF, voutput_min); |
| 82 |
|
| 83 |
vout0123456789ABCDEF = _mm_min_epu8(vout0123456789ABCDEF, voutput_max); |
| 84 |
|
| 85 |
_mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF); |
| 86 |
output += 16; |
| 87 |
} |
| 88 |
if XNN_UNLIKELY(batch != 0) { |
| 89 |
do { |
| 90 |
const __m128i va0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a))); |
| 91 |
const __m128i vb0123 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b))); |
| 92 |
const __m128i va4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_a + 4))); |
| 93 |
const __m128i vb4567 = _mm_cvtepu8_epi32(_mm_cvtsi32_si128((int) unaligned_load_s32(input_b + 4))); |
| 94 |
input_a += 8; |
| 95 |
input_b += 8; |
| 96 |
|
| 97 |
__m128i vacc0123 = _mm_add_epi32(vbias, _mm_mullo_epi32(va0123, va_multiplier)); |
| 98 |
__m128i vacc4567 = _mm_add_epi32(vbias, _mm_mullo_epi32(va4567, va_multiplier)); |
| 99 |
|
| 100 |
vacc0123 = _mm_add_epi32(vacc0123, _mm_mullo_epi32(vb0123, vb_multiplier)); |
| 101 |
vacc4567 = _mm_add_epi32(vacc4567, _mm_mullo_epi32(vb4567, vb_multiplier)); |
| 102 |
|
| 103 |
vacc0123 = _mm_sra_epi32(vacc0123, vshift); |
| 104 |
vacc4567 = _mm_sra_epi32(vacc4567, vshift); |
| 105 |
|
| 106 |
const __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(vacc0123, vacc4567), voutput_zero_point); |
| 107 |
|
| 108 |
__m128i vout0123456701234567 = _mm_packus_epi16(vout01234567, vout01234567); |
| 109 |
vout0123456701234567 = _mm_max_epu8(vout0123456701234567, voutput_min); |
| 110 |
vout0123456701234567 = _mm_min_epu8(vout0123456701234567, voutput_max); |
| 111 |
|
| 112 |
if XNN_LIKELY(batch >= (8 * sizeof(uint8_t))) { |
| 113 |
_mm_storel_epi64((__m128i*) output, vout0123456701234567); |
| 114 |
output += 8; |
| 115 |
batch -= 8 * sizeof(uint8_t); |
| 116 |
} else { |
| 117 |
if (batch & (4 * sizeof(uint8_t))) { |
| 118 |
unaligned_store_u32(output, (uint32_t) _mm_cvtsi128_si32(vout0123456701234567)); |
| 119 |
vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32); |
| 120 |
output += 4; |
| 121 |
} |
| 122 |
if (batch & (2 * sizeof(uint8_t))) { |
| 123 |
unaligned_store_u16(output, (uint16_t) _mm_extract_epi16(vout0123456701234567, 0)); |
| 124 |
vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16); |
| 125 |
output += 2; |
| 126 |
} |
| 127 |
if (batch & (1 * sizeof(uint8_t))) { |
| 128 |
*output = (uint8_t) _mm_extract_epi8(vout0123456701234567, 0); |
| 129 |
} |
| 130 |
batch = 0; |
| 131 |
} |
| 132 |
} while (batch != 0); |
| 133 |
} |
| 134 |
} |