ggml : support AVX512VNNI (llama/6280)

This change causes some quants (e.g. Q4_0, Q8_0) to go faster on some
architectures (e.g. AMD Zen 4).

diff --git a/src/ggml-quants.c b/src/ggml-quants.c
index 2eaca0593fd8b111f79cfc6dd46470702a7e88ea..f26798accdb9a5f4564a007d3873edb03b2bbed6 100644 (file)
--- a/src/ggml-quants.c
+++ b/src/ggml-quants.c
@@ -132,7 +132,7 @@ static inline __m256 sum_i16_pairs_float(const __m256i x) {
 }
 
 static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
-#if __AVXVNNI__
+#if defined(__AVXVNNI__) || defined(__AVX512VNNI__)
     const __m256i zero = _mm256_setzero_si256();
     const __m256i summed_pairs = _mm256_dpbusd_epi32(zero, ax, sy);
     return _mm256_cvtepi32_ps(summed_pairs);