| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-vcvt/armsimd32.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 <arm_acle.h> |
| 13 |
|
| 14 |
#include "xnnpack/intrinsics-polyfill.h" |
| 15 |
#include "xnnpack/math.h" |
| 16 |
#include "xnnpack/unaligned.h" |
| 17 |
#include "xnnpack/vcvt.h" |
| 18 |
|
| 19 |
|
| 20 |
void xnn_qu8_vcvt_ukernel__armsimd32_u8( |
| 21 |
size_t batch, |
| 22 |
const uint8_t* input, |
| 23 |
uint8_t* output, |
| 24 |
const struct xnn_qu8_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 25 |
{ |
| 26 |
assert(batch != 0); |
| 27 |
assert(batch % sizeof(uint8_t) == 0); |
| 28 |
assert(input != NULL); |
| 29 |
assert(output != NULL); |
| 30 |
|
| 31 |
const uint16x2_t vminus_input_zero_point = (uint16x2_t) broadcast2x_uint16(-params->scalar.input_zero_point); |
| 32 |
const int32_t vbias = (int32_t) ((uint32_t) (int32_t) params->scalar.output_zero_point << 1) + INT32_C(1); |
| 33 |
const int32_t vmultiplier = (int32_t) params->scalar.multiplier << 9; |
| 34 |
for (; batch >= 8 * sizeof(uint8_t); batch -= 8 * sizeof(uint8_t)) { |
| 35 |
const uint8x4_t vx0123 = (uint8x4_t) unaligned_indexed_load_u32(input, 0); |
| 36 |
const uint8x4_t vx4567 = (uint8x4_t) unaligned_indexed_load_u32(input, 1); |
| 37 |
input += 8; |
| 38 |
|
| 39 |
const uint16x2_t vx02 = __uxtab16(vminus_input_zero_point, vx0123); |
| 40 |
const uint16x2_t vx13 = __uxtab16(vminus_input_zero_point, __ror(vx0123, 8)); |
| 41 |
const uint16x2_t vx46 = __uxtab16(vminus_input_zero_point, vx4567); |
| 42 |
const uint16x2_t vx57 = __uxtab16(vminus_input_zero_point, __ror(vx4567, 8)); |
| 43 |
|
| 44 |
int32_t vacc0 = __smlawb(vmultiplier, vx02, vbias); |
| 45 |
int32_t vacc1 = __smlawb(vmultiplier, vx13, vbias); |
| 46 |
int32_t vacc2 = __smlawt(vmultiplier, vx02, vbias); |
| 47 |
int32_t vacc3 = __smlawt(vmultiplier, vx13, vbias); |
| 48 |
int32_t vacc4 = __smlawb(vmultiplier, vx46, vbias); |
| 49 |
int32_t vacc5 = __smlawb(vmultiplier, vx57, vbias); |
| 50 |
int32_t vacc6 = __smlawt(vmultiplier, vx46, vbias); |
| 51 |
int32_t vacc7 = __smlawt(vmultiplier, vx57, vbias); |
| 52 |
|
| 53 |
vacc0 = __usat(math_asr_s32(vacc0, 1), 8); |
| 54 |
vacc1 = __usat(math_asr_s32(vacc1, 1), 8); |
| 55 |
vacc2 = __usat(math_asr_s32(vacc2, 1), 8); |
| 56 |
vacc3 = __usat(math_asr_s32(vacc3, 1), 8); |
| 57 |
vacc4 = __usat(math_asr_s32(vacc4, 1), 8); |
| 58 |
vacc5 = __usat(math_asr_s32(vacc5, 1), 8); |
| 59 |
vacc6 = __usat(math_asr_s32(vacc6, 1), 8); |
| 60 |
vacc7 = __usat(math_asr_s32(vacc7, 1), 8); |
| 61 |
|
| 62 |
output[0] = (uint8_t) vacc0; |
| 63 |
output[1] = (uint8_t) vacc1; |
| 64 |
output[2] = (uint8_t) vacc2; |
| 65 |
output[3] = (uint8_t) vacc3; |
| 66 |
output[4] = (uint8_t) vacc4; |
| 67 |
output[5] = (uint8_t) vacc5; |
| 68 |
output[6] = (uint8_t) vacc6; |
| 69 |
output[7] = (uint8_t) vacc7; |
| 70 |
output += 8; |
| 71 |
} |
| 72 |
for (; batch >= 4 * sizeof(uint8_t); batch -= 4 * sizeof(uint8_t)) { |
| 73 |
const uint8x4_t vx0123 = (uint8x4_t) unaligned_load_u32(input); |
| 74 |
input += 4; |
| 75 |
|
| 76 |
const uint16x2_t vx02 = __uxtab16(vminus_input_zero_point, vx0123); |
| 77 |
const uint16x2_t vx13 = __uxtab16(vminus_input_zero_point, __ror(vx0123, 8)); |
| 78 |
|
| 79 |
int32_t vacc0 = __smlawb(vmultiplier, vx02, vbias); |
| 80 |
int32_t vacc1 = __smlawb(vmultiplier, vx13, vbias); |
| 81 |
int32_t vacc2 = __smlawt(vmultiplier, vx02, vbias); |
| 82 |
int32_t vacc3 = __smlawt(vmultiplier, vx13, vbias); |
| 83 |
|
| 84 |
vacc0 = __usat(math_asr_s32(vacc0, 1), 8); |
| 85 |
vacc1 = __usat(math_asr_s32(vacc1, 1), 8); |
| 86 |
vacc2 = __usat(math_asr_s32(vacc2, 1), 8); |
| 87 |
vacc3 = __usat(math_asr_s32(vacc3, 1), 8); |
| 88 |
|
| 89 |
output[0] = (uint8_t) vacc0; |
| 90 |
output[1] = (uint8_t) vacc1; |
| 91 |
output[2] = (uint8_t) vacc2; |
| 92 |
output[3] = (uint8_t) vacc3; |
| 93 |
output += 4; |
| 94 |
} |
| 95 |
if XNN_UNLIKELY(batch != 0) { |
| 96 |
const uint8x4_t vx0123 = (uint8x4_t) unaligned_load_u32(input); |
| 97 |
|
| 98 |
const uint16x2_t vx02 = __uxtab16(vminus_input_zero_point, vx0123); |
| 99 |
const uint16x2_t vx13 = __uxtab16(vminus_input_zero_point, __ror(vx0123, 8)); |
| 100 |
|
| 101 |
int32_t vacc0 = __smlawb(vmultiplier, vx02, vbias); |
| 102 |
int32_t vacc1 = __smlawb(vmultiplier, vx13, vbias); |
| 103 |
const int32_t vacc2 = __smlawt(vmultiplier, vx02, vbias); |
| 104 |
|
| 105 |
vacc0 = __usat(math_asr_s32(vacc0, 1), 8); |
| 106 |
vacc1 = __usat(math_asr_s32(vacc1, 1), 8); |
| 107 |
|
| 108 |
if (batch & (2 * sizeof(uint8_t))) { |
| 109 |
output[0] = (uint8_t) vacc0; |
| 110 |
output[1] = (uint8_t) vacc1; |
| 111 |
vacc0 = __usat(math_asr_s32(vacc2, 1), 8); |
| 112 |
output += 2; |
| 113 |
} |
| 114 |
if (batch & (1 * sizeof(uint8_t))) { |
| 115 |
output[0] = (uint8_t) vacc0; |
| 116 |
} |
| 117 |
} |
| 118 |
} |