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