/* { dg-do run { target { ! ia32 } } } */ /* { dg-require-effective-target amx_tile } */ /* { dg-require-effective-target amx_bf16 } */ /* { dg-options "-O2 -mamx-tile -mamx-bf16" } */ #include #define AMX_BF16 #define DO_TEST test_amx_bf16_dpbf16ps void test_amx_bf16_dpbf16ps (); #include "amx-check.h" /* Transformation functions between bf16/float */ static uint16_t make_bf16 (float f) { uint32_t u = (uint32_t)f; u = (u >> 16) & 0xffff; return (uint16_t)u; } static float make_f32 (uint16_t bf) { uint32_t u = (uint32_t)(bf << 16); return (float)u; } /* 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 = 16.1f * i + 3.4f * j; ptr[i * 32 + j] = make_bf16(f); } } void calc_matrix_dpbf16ps (__tile *dst, __tile *src1, __tile *src2) { uint16_t *src1_buf = (uint16_t *)src1->buf; uint16_t *src2_buf = (uint16_t *)src2->buf; float *dst_buf = (float *)dst->buf; int M = src1->rows; int N = src1->colsb / 4; int K = src2->colsb / 4; int i, j, k, t; for (i = 0; i < M; i++) for (j = 0; j < N; j++) for (k = 0; k < K; k++) for (t = 0; t < 2; t+=2) { dst_buf[i * N + k] += (make_f32(src1_buf[i * 4 * N + 4 * j + t]) * make_f32(src2_buf[j * 4 * K + 4 * k + t])) + (make_f32(src1_buf[i * 4 * N + 4 * j + t + 1]) * make_f32(src2_buf[j * 4 * K + 4 * k + t + 1])); } } void test_amx_bf16_dpbf16ps () { __tilecfg_u cfg; __tile dst, dst_ref, src1, src2; uint8_t tmp_dst_buf[1024]; init_bf16_max_tile_buffer (tmp_dst_buf); init_tile_config (&cfg); init_tile_reg_and_src_with_buffer (1, dst, tmp_dst_buf); init_tile_reg_and_src_with_buffer (2, src1, tmp_dst_buf); init_tile_reg_and_src_with_buffer (3, src2, tmp_dst_buf); calc_matrix_dpbf16ps (&dst, &src1, &src2); _tile_dpbf16ps (1, 2, 3); _tile_stored (1, dst_ref.buf, _STRIDE); if (!check_tile_register (&dst_ref, &dst)) abort(); }