Debian Codegraph @

gcc/testsuite/gcc.target/i386/amxtranspose-tcmmrlfp16ps-2.c

raw
1 /* { dg-do run { target { ! ia32 } } } */
2 /* { dg-require-effective-target amx_transpose } */
3 /* { dg-require-effective-target amx_complex } */
4 /* { dg-require-effective-target avx512fp16 } */
5 /* { dg-options "-O2 -mamx-transpose -mamx-complex -mavx512fp16" } */
6 #define AMX_TRANSPOSE
7 #define AMX_COMPLEX
8 #define DO_TEST test_amx_transpose_tcmmrlfp16ps
9 void test_amx_transpose_tcmmrlfp16ps ();
10 #include "amx-helper.h"
11
12 void calc_matrix_tcmmrlfp16ps (__tile *dst, __tile *src1, __tile *src2)
13 {
14 uint16_t *src1_buf = (uint16_t *) src1->buf;
15 uint16_t *src2_buf = (uint16_t *) src2->buf;
16 float *dst_buf = (float *) dst->buf;
17
18 int K = src1->rows;
19 int M = src1->colsb / 4;
20 int N = src2->colsb / 4;
21 int m, k, n, t;
22
23 for (m = 0; m < M; m++)
24 for (k = 0; k < K; k++)
25 for (n = 0; n < N; n++)
26 for (t = 0; t < 2; t+=2)
27 dst_buf[m * N + n] +=
28 (make_fp16_f32(src1_buf[k * 2 * M + 2 * m + t]) *
29 make_fp16_f32(src2_buf[k * 2 * N + 2 * n + t])) -
30 (make_fp16_f32(src1_buf[k * 2 * M + 2 * m + t + 1]) *
31 make_fp16_f32(src2_buf[k * 2 * N + 2 * n + t + 1]));
32 }
33
34 void test_amx_transpose_tcmmrlfp16ps ()
35 {
36 __tilecfg_u cfg;
37 __tile dst, dst_ref, src1, src2;
38 uint8_t tmp_dst_buf[1024], tmp_dst_zero_buf[1024];
39
40 init_fp16_max_tile_buffer (tmp_dst_buf);
41 init_fp16_max_tile_zero_buffer (tmp_dst_zero_buf);
42
43 init_tile_config (&cfg);
44 init_tile_reg_and_src_with_buffer (1, dst, tmp_dst_zero_buf);
45 init_tile_reg_and_src_with_buffer (2, src1, tmp_dst_buf);
46 init_tile_reg_and_src_with_buffer (3, src2, tmp_dst_buf);
47
48 calc_matrix_tcmmrlfp16ps (&dst, &src1, &src2);
49
50 _tile_tcmmrlfp16ps (1, 2, 3);
51 _tile_stored (1, dst_ref.buf, _STRIDE);
52
53 if (!check_tile_register (&dst_ref, &dst))
54 abort ();
55 }