| 1 |
#ifndef AMX_HELPER_H_INCLUDED |
| 2 |
#define AMX_HELPER_H_INCLUDED |
| 3 |
#include <immintrin.h> |
| 4 |
#include <xmmintrin.h> |
| 5 |
#include "amx-check.h" |
| 6 |
|
| 7 |
typedef union |
| 8 |
{ |
| 9 |
_Float16 f16; |
| 10 |
uint16_t u; |
| 11 |
} union16f_uw; |
| 12 |
|
| 13 |
typedef union |
| 14 |
{ |
| 15 |
__bf16 bf16; |
| 16 |
uint16_t u; |
| 17 |
} union16bh_uw; |
| 18 |
|
| 19 |
typedef union |
| 20 |
{ |
| 21 |
float f; |
| 22 |
uint32_t u; |
| 23 |
} union32f_ud; |
| 24 |
|
| 25 |
typedef union |
| 26 |
{ |
| 27 |
__m512 m; |
| 28 |
uint8_t u[64]; |
| 29 |
} union512_ub; |
| 30 |
|
| 31 |
#if defined(AMX_FP16) || defined(AMX_COMPLEX) || defined (AMX_AVX512) |
| 32 |
/* Transformation functions between fp16/float */ |
| 33 |
static uint16_t make_f32_fp16 (float f) |
| 34 |
{ |
| 35 |
union16f_uw tmp; |
| 36 |
__m128 b = _mm_set_ss (f); |
| 37 |
__m128h a; |
| 38 |
tmp.f16 = _mm_cvtsh_h (_mm_cvtss_sh (a, b)); |
| 39 |
return tmp.u; |
| 40 |
} |
| 41 |
|
| 42 |
static float make_fp16_f32 (uint16_t fp) |
| 43 |
{ |
| 44 |
union16f_uw tmp; |
| 45 |
tmp.u = fp; |
| 46 |
__m128h b = _mm_set_sh (tmp.f16); |
| 47 |
__m128 a; |
| 48 |
return _mm_cvtss_f32 (_mm_cvtsh_ss (a, b)); |
| 49 |
} |
| 50 |
|
| 51 |
/* Init tile buffer with fp16 pairs */ |
| 52 |
void init_fp16_max_tile_buffer (uint8_t* buf) |
| 53 |
{ |
| 54 |
int i, j; |
| 55 |
uint16_t* ptr = (uint16_t *) buf; |
| 56 |
|
| 57 |
for (i = 0; i < 16; i++) |
| 58 |
for (j = 0; j < 32; j++) |
| 59 |
{ |
| 60 |
float f = 2.5f * i + 1.25f * j; |
| 61 |
ptr[i * 32 + j] = make_f32_fp16 (f); |
| 62 |
} |
| 63 |
} |
| 64 |
|
| 65 |
/* Init tile fp16 pair buffer with zero */ |
| 66 |
void init_fp16_max_tile_zero_buffer (uint8_t* buf) |
| 67 |
{ |
| 68 |
int i, j; |
| 69 |
uint16_t* ptr = (uint16_t *) buf; |
| 70 |
|
| 71 |
for (i = 0; i < 16; i++) |
| 72 |
for (j = 0; j < 32; j++) |
| 73 |
ptr[i * 32 + j] = make_f32_fp16 (0.0f); |
| 74 |
} |
| 75 |
#endif |
| 76 |
|
| 77 |
#if defined (AMX_AVX512) || defined (AMX_BF16) |
| 78 |
/* Transformation functions between bf16/float */ |
| 79 |
static uint16_t make_f32_bf16 (float f) |
| 80 |
{ |
| 81 |
union16bh_uw tmp; |
| 82 |
tmp.bf16 = (__bf16) f; |
| 83 |
return tmp.u; |
| 84 |
} |
| 85 |
|
| 86 |
static float make_bf16_f32 (uint16_t bf) |
| 87 |
{ |
| 88 |
union16bh_uw tmp; |
| 89 |
tmp.u = bf; |
| 90 |
return _mm_cvtsbh_ss (tmp.bf16); |
| 91 |
} |
| 92 |
|
| 93 |
/* Init tile buffer with bf16 pairs */ |
| 94 |
void init_bf16_max_tile_buffer (uint8_t *buf) |
| 95 |
{ |
| 96 |
int i, j; |
| 97 |
uint16_t* ptr = (uint16_t *) buf; |
| 98 |
|
| 99 |
for (i = 0; i < 16; i++) |
| 100 |
for (j = 0; j < 32; j++) |
| 101 |
{ |
| 102 |
float f = 2.5f * i + 1.25f * j; |
| 103 |
ptr[i * 32 + j] = make_f32_bf16 (f); |
| 104 |
} |
| 105 |
} |
| 106 |
#endif |
| 107 |
|
| 108 |
/* Init tile buffer with fp32 */ |
| 109 |
void init_fp32_max_tile_buffer (uint8_t *buf) |
| 110 |
{ |
| 111 |
int i, j; |
| 112 |
float* ptr = (float *) buf; |
| 113 |
|
| 114 |
for (i = 0; i < 16; i++) |
| 115 |
for (j = 0; j < 16; j++) |
| 116 |
ptr[i * 16 + j] = 2.5f * i + 1.25f * j; |
| 117 |
} |
| 118 |
|
| 119 |
/* Init tile fp32 buffer with zero */ |
| 120 |
void init_fp32_max_tile_zero_buffer (uint8_t *buf) |
| 121 |
{ |
| 122 |
int i, j; |
| 123 |
float* ptr = (float *) buf; |
| 124 |
|
| 125 |
for (i = 0; i < 16; i++) |
| 126 |
for (j = 0; j < 16; j++) |
| 127 |
ptr[i * 16 + j] = 0.0f; |
| 128 |
} |
| 129 |
|
| 130 |
/* Init tile buffer with int32 */ |
| 131 |
void init_int32_max_tile_buffer (uint8_t *buf) |
| 132 |
{ |
| 133 |
int i, j; |
| 134 |
uint32_t *ptr = (uint32_t *)buf; |
| 135 |
|
| 136 |
for (i = 0; i < 16; i++) |
| 137 |
for (j = 0; j < 16; j++) |
| 138 |
ptr[i * 16 + j] = (uint32_t) (3 * j - 16 * i); |
| 139 |
} |
| 140 |
|
| 141 |
void |
| 142 |
init_fp8_max_tile_buffer (uint8_t *buf) |
| 143 |
{ |
| 144 |
int i, j; |
| 145 |
|
| 146 |
for (i = 0; i < 16; i++) |
| 147 |
for (j = 0; j < 64; j++) |
| 148 |
{ |
| 149 |
int idx = i * 64 + j; |
| 150 |
|
| 151 |
/* Positive Infinity (S11111.00) */ |
| 152 |
if (idx % 128 == 0) |
| 153 |
buf[idx] = 0x7C; |
| 154 |
|
| 155 |
/* Negative Infinity (S11111.00 with sign bit set) */ |
| 156 |
else if (idx % 128 == 1) |
| 157 |
buf[idx] = 0xFC; |
| 158 |
|
| 159 |
/* Positive NaN (S11111.01) */ |
| 160 |
else if (idx % 128 == 2) |
| 161 |
buf[idx] = 0x7D; |
| 162 |
|
| 163 |
/* Negative NaN (S11111.01 with sign bit set) */ |
| 164 |
else if (idx % 128 == 3) |
| 165 |
buf[idx] = 0xFD; |
| 166 |
|
| 167 |
/* insert Positive NaN (S11111.10) */ |
| 168 |
else if (idx % 128 == 4) |
| 169 |
buf[idx] = 0x7E; |
| 170 |
|
| 171 |
/* Negative NaN (S11111.10 with sign bit set) */ |
| 172 |
else if (idx % 128 == 5) |
| 173 |
buf[idx] = 0xFE; |
| 174 |
|
| 175 |
/* Positive NaN (S11111.11) */ |
| 176 |
else if (idx % 128 == 6) |
| 177 |
buf[idx] = 0x7F; |
| 178 |
|
| 179 |
/* Negative NaN (S11111.11 with sign bit set) */ |
| 180 |
else if (idx % 128 == 7) |
| 181 |
buf[idx] = 0xFF; |
| 182 |
|
| 183 |
else |
| 184 |
buf[idx] = (uint8_t) ((idx * 251) & 0xFF); |
| 185 |
} |
| 186 |
} |
| 187 |
|
| 188 |
#define COMPARE_ZMM(A, B) \ |
| 189 |
for (int j = 0; j < 16; j++) \ |
| 190 |
{ \ |
| 191 |
union32f_ud fu1, fu2; \ |
| 192 |
fu1.f = A[j]; \ |
| 193 |
fu2.f = B[j]; \ |
| 194 |
if (fu1.u != fu2.u) \ |
| 195 |
abort (); \ |
| 196 |
} |
| 197 |
|
| 198 |
#define COMPARE_ZMM_BF16(A, B) \ |
| 199 |
for (int j = 0; j < 32; j++) \ |
| 200 |
{ \ |
| 201 |
union16bh_uw fu1, fu2; \ |
| 202 |
fu1.bf16 = A[j]; \ |
| 203 |
fu2.bf16 = B[j]; \ |
| 204 |
if (fu1.u != fu2.u) \ |
| 205 |
abort(); \ |
| 206 |
} |
| 207 |
|
| 208 |
#define COMPARE_ZMM_FP16(A, B) \ |
| 209 |
for (int j = 0; j < 32; j++) \ |
| 210 |
{ \ |
| 211 |
union16f_uw fu1, fu2; \ |
| 212 |
fu1.f16 = A[j]; \ |
| 213 |
fu2.f16 = B[j]; \ |
| 214 |
if (fu1.u != fu2.u) \ |
| 215 |
abort(); \ |
| 216 |
} |
| 217 |
|
| 218 |
/* Mask low 13bits to zero */ |
| 219 |
static float zero_lower_mantissa_bits_fp32 (float x) |
| 220 |
{ |
| 221 |
union32f_ud tmp; |
| 222 |
tmp.f = x; |
| 223 |
tmp.u = tmp.u & 0xffffe000; |
| 224 |
return tmp.f; |
| 225 |
} |
| 226 |
|
| 227 |
/* Handle SNAN */ |
| 228 |
static float silence_snan_fp32 (float x) |
| 229 |
{ |
| 230 |
union32f_ud tmp; |
| 231 |
tmp.f = x; |
| 232 |
if ((((tmp.u & 0x7f800000) >> 23) == 0xff) && |
| 233 |
((tmp.u & 0x007fffff) != 0) && |
| 234 |
((tmp.u & 0x00400000) == 0)) |
| 235 |
tmp.u = tmp.u | 0x00400000; |
| 236 |
return tmp.f; |
| 237 |
} |
| 238 |
|
| 239 |
void init_pair_tile_src (int tmm_num, __tilepair *src, uint8_t *_buffer, int z) |
| 240 |
{ |
| 241 |
int rows, colsb, start, i, j, t, elements[2]; |
| 242 |
uint16_t *buffer = (uint16_t *) _buffer; |
| 243 |
uint16_t *ptr = (uint16_t *) src->buf; |
| 244 |
__tilecfg_u tmp; |
| 245 |
|
| 246 |
_tile_storeconfig (tmp.a); |
| 247 |
|
| 248 |
tmm_num &= ~1; |
| 249 |
|
| 250 |
rows = tmp.s.rows[tmm_num]; |
| 251 |
colsb = tmp.s.colsb[tmm_num]; |
| 252 |
start = tmp.s.start_row; |
| 253 |
|
| 254 |
zero_pair_tile_src (src); |
| 255 |
|
| 256 |
for (t = 0; t < 2; t++) |
| 257 |
elements[t] = tmp.s.colsb[tmm_num + t] / 4; |
| 258 |
|
| 259 |
src->colsb = (tmp.s.colsb[tmm_num] + tmp.s.colsb[tmm_num + 1]) / 2; |
| 260 |
src->rows = rows; |
| 261 |
|
| 262 |
while (start < 2 * rows) |
| 263 |
{ |
| 264 |
int r = start / 2; |
| 265 |
int w = start % 2; |
| 266 |
|
| 267 |
if (start < 2 * rows - z) |
| 268 |
for (t = 0; t < 2; t++) |
| 269 |
if (tmp.s.colsb[tmm_num + t] > 0) |
| 270 |
for (i = 0; i < elements[t]; i++) |
| 271 |
ptr[t * rows * colsb / 2 + r * elements[t] * 2 + 2 * i + w] = |
| 272 |
buffer[start * colsb / 2 + t * elements[0] + i]; |
| 273 |
start++; |
| 274 |
} |
| 275 |
} |
| 276 |
|
| 277 |
#endif |