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 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];
}
__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];
}
KFR_INTRINSIC void write(cunaligned_t, f32* ptr, const f32avx512& x) { _mm512_storeu_ps(ptr, x.v); }
inline
void store(float *data) const
{
_mm512_storeu_ps(data + 0, val1);
_mm512_storeu_ps(data + 16, val2);
}
