| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-igemm/c4-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/gemm.h" |
| 17 |
#include "xnnpack/unaligned.h" |
| 18 |
|
| 19 |
|
| 20 |
void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_2x2c4__armsimd32( |
| 21 |
size_t mr, |
| 22 |
size_t nc, |
| 23 |
size_t kc, |
| 24 |
size_t ks, |
| 25 |
const int8_t** restrict a, |
| 26 |
const void* restrict w, |
| 27 |
int8_t* restrict c, |
| 28 |
size_t cm_stride, |
| 29 |
size_t cn_stride, |
| 30 |
size_t a_offset, |
| 31 |
const int8_t* zero, |
| 32 |
const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) |
| 33 |
{ |
| 34 |
assert(mr != 0); |
| 35 |
assert(mr <= 2); |
| 36 |
assert(nc != 0); |
| 37 |
assert(kc != 0); |
| 38 |
assert(ks != 0); |
| 39 |
assert(ks % (2 * sizeof(void*)) == 0); |
| 40 |
assert(a != NULL); |
| 41 |
assert(w != NULL); |
| 42 |
assert(c != NULL); |
| 43 |
|
| 44 |
kc = round_up_po2(kc, 4 * sizeof(int8_t)); |
| 45 |
int8_t* c0 = c; |
| 46 |
int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride); |
| 47 |
if XNN_UNPREDICTABLE(mr != 2) { |
| 48 |
c1 = c0; |
| 49 |
} |
| 50 |
|
| 51 |
const float vmagic_bias = params->fp32_armsimd32.magic_bias; |
| 52 |
do { |
| 53 |
int32_t vacc0x0 = ((const int32_t*) w)[0]; |
| 54 |
int32_t vacc0x1 = ((const int32_t*) w)[1]; |
| 55 |
int32_t vacc1x0 = vacc0x0; |
| 56 |
int32_t vacc1x1 = vacc0x1; |
| 57 |
w = (const void*) ((const int32_t*) w + 2); |
| 58 |
|
| 59 |
size_t p = ks; |
| 60 |
do { |
| 61 |
const int8_t* restrict a0 = a[0]; |
| 62 |
assert(a0 != NULL); |
| 63 |
if XNN_UNPREDICTABLE(a0 != zero) { |
| 64 |
a0 = (const int8_t*) ((uintptr_t) a0 + a_offset); |
| 65 |
} |
| 66 |
const int8_t* restrict a1 = a[1]; |
| 67 |
assert(a1 != NULL); |
| 68 |
if XNN_UNPREDICTABLE(a1 != zero) { |
| 69 |
a1 = (const int8_t*) ((uintptr_t) a1 + a_offset); |
| 70 |
} |
| 71 |
a += 2; |
| 72 |
|
| 73 |
size_t k = kc; |
| 74 |
do { |
| 75 |
const int8x4_t va0 = (int8x4_t) unaligned_load_s32(a0); a0 += 4; |
| 76 |
const int8x4_t va1 = (int8x4_t) unaligned_load_s32(a1); a1 += 4; |
| 77 |
|
| 78 |
const int16x2_t va0c02 = __sxtb16(va0); |
| 79 |
const int16x2_t va0c13 = __sxtb16(__ror(va0, 8)); |
| 80 |
const int16x2_t va1c02 = __sxtb16(va1); |
| 81 |
const int16x2_t va1c13 = __sxtb16(__ror(va1, 8)); |
| 82 |
|
| 83 |
const int8x4_t vb0 = *((const int8x4_t*) w); w = (const int8_t*) w + 4; |
| 84 |
const int16x2_t vb0c02 = __sxtb16(vb0); |
| 85 |
|
| 86 |
vacc0x0 = __smlad(va0c02, vb0c02, vacc0x0); |
| 87 |
vacc1x0 = __smlad(va1c02, vb0c02, vacc1x0); |
| 88 |
|
| 89 |
const int16x2_t vb0c13 = __sxtb16(__ror(vb0, 8)); |
| 90 |
vacc0x0 = __smlad(va0c13, vb0c13, vacc0x0); |
| 91 |
vacc1x0 = __smlad(va1c13, vb0c13, vacc1x0); |
| 92 |
const int8x4_t vb1 = *((const int8x4_t*) w); w = (const int8_t*) w + 4; |
| 93 |
const int16x2_t vb1c02 = __sxtb16(vb1); |
| 94 |
|
| 95 |
vacc0x1 = __smlad(va0c02, vb1c02, vacc0x1); |
| 96 |
vacc1x1 = __smlad(va1c02, vb1c02, vacc1x1); |
| 97 |
|
| 98 |
const int16x2_t vb1c13 = __sxtb16(__ror(vb1, 8)); |
| 99 |
vacc0x1 = __smlad(va0c13, vb1c13, vacc0x1); |
| 100 |
vacc1x1 = __smlad(va1c13, vb1c13, vacc1x1); |
| 101 |
|
| 102 |
k -= 4 * sizeof(int8_t); |
| 103 |
} while (k != 0); |
| 104 |
p -= 2 * sizeof(void*); |
| 105 |
} while (p != 0); |
| 106 |
|
| 107 |
float vfpacc0x0 = (float) vacc0x0; |
| 108 |
float vfpacc0x1 = (float) vacc0x1; |
| 109 |
float vfpacc1x0 = (float) vacc1x0; |
| 110 |
float vfpacc1x1 = (float) vacc1x1; |
| 111 |
|
| 112 |
const float vscale0 = ((const float*) w)[0]; |
| 113 |
vfpacc0x0 *= vscale0; |
| 114 |
vfpacc1x0 *= vscale0; |
| 115 |
const float vscale1 = ((const float*) w)[1]; |
| 116 |
vfpacc0x1 *= vscale1; |
| 117 |
vfpacc1x1 *= vscale1; |
| 118 |
w = (const void*) ((const float*) w + 2); |
| 119 |
|
| 120 |
vfpacc0x0 += vmagic_bias; |
| 121 |
vfpacc0x1 += vmagic_bias; |
| 122 |
vfpacc1x0 += vmagic_bias; |
| 123 |
vfpacc1x1 += vmagic_bias; |
| 124 |
|
| 125 |
int32_t vout0x0 = (int32_t) float_as_uint32(vfpacc0x0); |
| 126 |
int32_t vout0x1 = (int32_t) float_as_uint32(vfpacc0x1); |
| 127 |
int32_t vout1x0 = (int32_t) float_as_uint32(vfpacc1x0); |
| 128 |
int32_t vout1x1 = (int32_t) float_as_uint32(vfpacc1x1); |
| 129 |
|
| 130 |
const int32_t vmagic_bias_less_zero_point = params->fp32_armsimd32.magic_bias_less_zero_point; |
| 131 |
vout0x0 = __qsub(vout0x0, vmagic_bias_less_zero_point); |
| 132 |
vout0x1 = __qsub(vout0x1, vmagic_bias_less_zero_point); |
| 133 |
vout1x0 = __qsub(vout1x0, vmagic_bias_less_zero_point); |
| 134 |
vout1x1 = __qsub(vout1x1, vmagic_bias_less_zero_point); |
| 135 |
|
| 136 |
vout0x0 = __ssat(vout0x0, 8); |
| 137 |
vout0x1 = __ssat(vout0x1, 8); |
| 138 |
vout1x0 = __ssat(vout1x0, 8); |
| 139 |
vout1x1 = __ssat(vout1x1, 8); |
| 140 |
|
| 141 |
const uint32_t vout0 = (uint32_t) (uint8_t) vout0x0 | ((uint32_t) vout0x1 << 8); |
| 142 |
const uint32_t vout1 = (uint32_t) (uint8_t) vout1x0 | ((uint32_t) vout1x1 << 8); |
| 143 |
|
| 144 |
uint32_t vout = (uint32_t) (uint16_t) vout1 | (vout0 << 16); |
| 145 |
|
| 146 |
const int8x4_t voutput_min = (int8x4_t) params->fp32_armsimd32.output_min; |
| 147 |
__ssub8((int8x4_t) vout, voutput_min); |
| 148 |
vout = (uint32_t) __sel((uint8x4_t) vout, (uint8x4_t) voutput_min); |
| 149 |
|
| 150 |
const int8x4_t voutput_max = (int8x4_t) params->fp32_armsimd32.output_max; |
| 151 |
__ssub8((int8x4_t) vout, voutput_max); |
| 152 |
vout = (uint32_t) __sel((uint8x4_t) voutput_max, (uint8x4_t) vout); |
| 153 |
|
| 154 |
if XNN_LIKELY(nc >= 2) { |
| 155 |
unaligned_store_u16(c1, (uint16_t) vout); |
| 156 |
vout >>= 16; |
| 157 |
unaligned_store_u16(c0, (uint16_t) vout); |
| 158 |
|
| 159 |
c1 = (int8_t*) ((uintptr_t) c1 + cn_stride); |
| 160 |
c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); |
| 161 |
|
| 162 |
a = (const int8_t**restrict) ((uintptr_t) a - ks); |
| 163 |
nc -= 2; |
| 164 |
} else { |
| 165 |
*c1 = (int8_t) vout; |
| 166 |
vout >>= 16; |
| 167 |
*c0 = (int8_t) vout; |
| 168 |
|
| 169 |
nc = 0; |
| 170 |
} |
| 171 |
} while (nc != 0); |
| 172 |
} |