From 04aaae1d79482cad2564412f3b32e70298ac7789 Mon Sep 17 00:00:00 2001
From: Yann Follet <131855179+YannFollet@users.noreply.github.com>
Date: Fri, 28 Apr 2023 19:59:48 +0800
Subject: [PATCH] add avx2 for dot_q8_0_q8_0, 2x faster than scalar (#1211)

---
 ggml.c | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)

diff --git a/ggml.c b/ggml.c
index 3422a9448..1fbf2955d 100644
--- a/ggml.c
+++ b/ggml.c
@@ -3626,6 +3626,24 @@ static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void *
     }
 
     *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
+#elif defined(__AVX2__)
+    // Initialize accumulator with zeros
+    __m256 acc = _mm256_setzero_ps();
+
+    // Main loop
+    for (int i = 0; i < nb; ++i) {
+        // Compute combined scale for the block
+        const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) );
+        __m256i bx = _mm256_loadu_si256((const __m256i *)x[i].qs);
+        __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
+
+        const __m256 q = mul_sum_i8_pairs_float(bx, by);
+
+        // Multiply q with scale and accumulate
+        acc = _mm256_fmadd_ps( d, q, acc );
+    }
+
+    *s = hsum_float_8(acc);
 #else
     // scalar
     float sumf = 0.0;