| 1 |
/* { dg-do run { target { ! ia32 } } } */ |
| 2 |
/* { dg-require-effective-target amx_transpose } */ |
| 3 |
/* { dg-require-effective-target amx_tf32 } */ |
| 4 |
/* { dg-options "-O2 -mamx-transpose -mamx-tf32" } */ |
| 5 |
#define AMX_TRANSPOSE |
| 6 |
#define AMX_TF32 |
| 7 |
#define DO_TEST test_amx_transpose_tmmultf32ps |
| 8 |
void test_amx_transpose_tmmultf32ps(); |
| 9 |
#include "amx-helper.h" |
| 10 |
|
| 11 |
void calc_matrix_tmmultf32ps(__tile *dst, __tile *src1, __tile *src2) |
| 12 |
{ |
| 13 |
float *src1_buf = (float *) src1->buf; |
| 14 |
float *src2_buf = (float *) src2->buf; |
| 15 |
float *dst_buf = (float *) dst->buf; |
| 16 |
|
| 17 |
int K = src1->rows; |
| 18 |
int M = src1->colsb / 4; |
| 19 |
int N = src2->colsb / 4; |
| 20 |
int m, n, k; |
| 21 |
|
| 22 |
for (m = 0; m < M; m++) |
| 23 |
for (k = 0; k < K; k++) |
| 24 |
for (n = 0; n < N; n++) |
| 25 |
dst_buf[m * N + n] += |
| 26 |
zero_lower_mantissa_bits_fp32 (silence_snan_fp32 (src1_buf[k * M + m])) * |
| 27 |
zero_lower_mantissa_bits_fp32 (silence_snan_fp32 (src2_buf[k * N + n])); |
| 28 |
|
| 29 |
} |
| 30 |
|
| 31 |
void test_amx_transpose_tmmultf32ps () |
| 32 |
{ |
| 33 |
__tilecfg_u cfg; |
| 34 |
__tile dst, dst_ref, src1, src2; |
| 35 |
uint8_t tmp_dst_buf[1024]; |
| 36 |
|
| 37 |
init_fp32_max_tile_buffer (tmp_dst_buf); |
| 38 |
|
| 39 |
init_tile_config (&cfg); |
| 40 |
init_tile_reg_and_src_with_buffer (1, dst, tmp_dst_buf); |
| 41 |
init_tile_reg_and_src_with_buffer (2, src1, tmp_dst_buf); |
| 42 |
init_tile_reg_and_src_with_buffer (3, src2, tmp_dst_buf); |
| 43 |
|
| 44 |
calc_matrix_tmmultf32ps (&dst, &src1, &src2); |
| 45 |
|
| 46 |
_tile_tmmultf32ps (1, 2, 3); |
| 47 |
_tile_stored (1, dst_ref.buf, _STRIDE); |
| 48 |
|
| 49 |
if (!check_tile_register (&dst_ref, &dst)) |
| 50 |
abort (); |
| 51 |
} |