ogl_beamforming

Commit: beda44ae53194d3ff72eae4d43397b11cf89c36c
Parent: 25eaa2baec4e1090349f5da57bdf2968411b6859
Author: Randy Palamar
Date:   Sun,  4 Jan 2026 10:27:18 -0700

util: avx512 memory copy

If the CPU being targeted supports AVX512BW then we might as well
use 64 byte wide instructions. masked load/store make it trivial
to clean up the remainder.

Diffstat:
M
util.c
 | 
9
+++++++++

1 file changed, 9 insertions(+), 0 deletions(-)
diff --git a/util.c b/util.c
@@ -12,7 +12,16 @@ function void
 mem_copy(void *restrict dest, void *restrict src, uz n)
 {
 	u8 *s = src, *d = dest;
+	#ifdef __AVX512BW__
+	for (; n >= 64; n -= 64, s += 64, d += 64)
+		_mm512_storeu_epi8(d, _mm512_loadu_epi8(s));
+	if (n > 0) {
+		__mmask64 k = _cvtu64_mask64(_bzhi_u64(-1, n));
+		_mm512_mask_storeu_epi8(d, k, _mm512_maskz_loadu_epi8(k, s));
+	}
+	#else
 	for (; n; n--) *d++ = *s++;
+	#endif
 }
 
 function void