/* { dg-do run } */ /* { dg-options "-O2 -march=x86-64-v3 -mavx10.2" } */ /* { dg-require-effective-target avx10_2 } */ #ifndef AVX10_2 #define AVX10_2 #define AVX10_2_512 #define AVX10_512BIT #endif #include "avx10-helper.h" #include "fp8-helper.h" #define SIZE_SRC (AVX512F_LEN / 16) #define SIZE (AVX512F_LEN / 8) #include "avx512f-mask-type.h" void CALC (unsigned char *r, _Float16 *s1, _Float16 *s2) { _Float16 temp; Float16Union ut = {.f16 = temp}; int i, hf8_bf8, saturate; hf8_bf8 = 1; saturate = 1; for (i = 0; i < SIZE; i++) { r[i] = 0; if (i < SIZE_SRC) { Float16Union usrc2 = {.f16 = s2[i]}; ut.u16 = usrc2.u16; } else { Float16Union usrc1 = {.f16 = s1[i-SIZE_SRC]}; ut.u16 = usrc1.u16; } r[i] = convert_fp16_to_fp8(ut.f16, 0, hf8_bf8, saturate); } } void TEST (void) { int i,sign; UNION_TYPE (AVX512F_LEN, i_b) res1, res2, res3; UNION_TYPE (AVX512F_LEN, h) src1, src2; MASK_TYPE mask = MASK_VALUE; unsigned char res_ref[SIZE]; sign = 1; for (i = 0; i < SIZE_SRC; i++) { src1.a[i] = (_Float16)(sign * (1.5 * (1 << (i % 3)))); src2.a[i] = (_Float16)(-sign * (2.5 * (1 << (i % 3)))); sign = -sign; } for (i = 0; i < SIZE; i++) res2.a[i] = DEFAULT_VALUE; CALC(res_ref, src1.a, src2.a); res1.x = INTRINSIC (_cvts_2ph_bf8) (src1.x, src2.x); if (UNION_CHECK (AVX512F_LEN, i_b) (res1, res_ref)) abort (); res2.x = INTRINSIC (_mask_cvts_2ph_bf8) (res2.x, mask, src1.x, src2.x); MASK_MERGE (i_b) (res_ref, mask, SIZE); if (UNION_CHECK (AVX512F_LEN, i_b) (res2, res_ref)) abort (); res3.x = INTRINSIC (_maskz_cvts_2ph_bf8) (mask, src1.x, src2.x); MASK_ZERO (i_b) (res_ref, mask, SIZE); if (UNION_CHECK (AVX512F_LEN, i_b) (res3, res_ref)) abort (); }