__attribute__((noinline)) float dot512(float *x1, float *x2, size_t len) {
assert(len % 16 == 0);
__m512 sum = _mm512_setzero_ps();
if (len > 15) {
size_t limit = len - 15;
for (size_t i = 0; i < limit; i += 16) {
__m512 v1 = _mm512_loadu_ps(x1 + i);
__m512 v2 = _mm512_loadu_ps(x2 + i);
sum = _mm512_add_ps(sum, _mm512_mul_ps(v1, v2));
}
}
float buffer[16];
_mm512_storeu_ps(buffer, sum);
return buffer[0] + buffer[1] + buffer[2] + buffer[3] + buffer[4] + buffer[5] +
buffer[6] + buffer[7] + buffer[8] + buffer[9] + buffer[10] +
buffer[11] + buffer[12] + buffer[13] + buffer[14] + buffer[15];
}
void extern
avx512f_test (void)
{
x = _mm512_loadu_ps (p);
x = _mm512_mask_loadu_ps (x, m, p);
x = _mm512_maskz_loadu_ps (m, p);
_mm512_storeu_ps (p, x);
_mm512_mask_storeu_ps (p, m, x);
}
__attribute__((noinline)) float dot512fma2(float *x1, float *x2, size_t len) {
assert(len % 32 == 0);
__m512 sum = _mm512_setzero_ps();
if (len > 31) {
size_t limit = len - 31;
for (size_t i = 0; i < limit; i += 32) {
__m512 v11 = _mm512_loadu_ps(x1 + i);
__m512 v21 = _mm512_loadu_ps(x2 + i);
__m512 v12 = _mm512_loadu_ps(x1 + i + 16);
__m512 v22 = _mm512_loadu_ps(x2 + i + 16);
sum = _mm512_fmadd_ps(v11, v21, sum);
sum = _mm512_fmadd_ps(v12, v22, sum);
}
}
float buffer[16];
_mm512_storeu_ps(buffer, sum);
return buffer[0] + buffer[1] + buffer[2] + buffer[3] + buffer[4] + buffer[5] +
buffer[6] + buffer[7] + buffer[8] + buffer[9] + buffer[10] +
buffer[11] + buffer[12] + buffer[13] + buffer[14] + buffer[15];
}
void convert_f32_bf16(float* in, libxsmm_bfloat16* out, int len)
{
int i;
#ifdef _OPENMP
#pragma omp parallel for private(i)
#endif
for ( i = 0; i < len; i+=16 ) {
__m512 vfp32 = gxm_fp32_to_bfp16_rne_adjustment_avx512f( _mm512_loadu_ps( in+i ) );
__m256i vbfp16 = gxm_fp32_to_bfp16_truncate_avx512f( vfp32 );
_mm256_storeu_si256( (__m256i*)(out+i), vbfp16 );
}
}
int main(int, char**argv)
{
/* AVX512 Foundation */
__m512i i;
__m512d d;
__m512 f;
__mmask16 m = ~1;
i = _mm512_maskz_loadu_epi32(0, argv);
d = _mm512_loadu_pd((double *)argv + 64);
f = _mm512_loadu_ps((float *)argv + 128);
#ifdef __AVX512ER__
/* AVX512 Exponential and Reciprocal */
f = _mm512_exp2a23_round_ps(f, 8);
#endif
#ifdef __AVX512CD__
/* AVX512 Conflict Detection */
i = _mm512_maskz_conflict_epi32(m, i);
#endif
#ifdef __AVX512PF__
/* AVX512 Prefetch */
_mm512_mask_prefetch_i64scatter_pd(argv, 0xf, i, 2, 2);
#endif
#ifdef __AVX512DQ__
/* AVX512 Doubleword and Quadword support */
m = _mm512_movepi32_mask(i);
#endif
#ifdef __AVX512BW__
/* AVX512 Byte and Word support */
i = _mm512_mask_loadu_epi8(i, m, argv - 8);
#endif
#ifdef __AVX512VL__
/* AVX512 Vector Length */
__m256i i2 = _mm256_maskz_loadu_epi32(0, argv);
_mm256_mask_storeu_epi32(argv + 1, m, i2);
#endif
#ifdef __AVX512IFMA__
/* AVX512 Integer Fused Multiply-Add */
i = _mm512_madd52lo_epu64(i, i, i);
#endif
#ifdef __AVX512VBMI__
/* AVX512 Vector Byte Manipulation Instructions */
i = _mm512_permutexvar_epi8(i, i);
#endif
_mm512_mask_storeu_epi64(argv, m, i);
_mm512_mask_storeu_ps(argv + 64, m, f);
_mm512_mask_storeu_pd(argv + 128, m, d);
return 0;
}
KFR_INTRINSIC f32avx512 read(cunaligned_t, csize_t<16>, const f32* ptr) { return _mm512_loadu_ps(ptr); }
inline
void load(const float *data)
{
val1 = _mm512_loadu_ps(data + 0);
val2 = _mm512_loadu_ps(data + 16);
}
