From 60d6eab63d70d6892eea86543f2c84227ec3ac9a Mon Sep 17 00:00:00 2001
From: Jean-Marc Valin <jmvalin@amazon.com>
Date: Tue, 6 Jul 2021 17:33:33 -0400
Subject: [PATCH] Doing a bit of unrolling to speed things up

---
 dnn/vec_avx.h | 35 ++++++++++++++++++++++++++++++++++-
 1 file changed, 34 insertions(+), 1 deletion(-)

diff --git a/dnn/vec_avx.h b/dnn/vec_avx.h
index f18c771a..e4ccdb41 100644
--- a/dnn/vec_avx.h
+++ b/dnn/vec_avx.h
@@ -502,7 +502,40 @@ static inline void sparse_sgemv_accum8x4(float *_out, const qweight *w, int rows
       colblocks = *idx++;
       y = &out[i];
       vy0 = _mm256_loadu_si256((const __m256i *)&y[0]);
-      for (j=0;j<colblocks;j++)
+      j=0;
+#if 1 /* Unrolling by 4 gives some gain, comment out if it does not. */
+      for (;j<colblocks-3;j+=4)
+      {
+         __m256i tmp;
+         __m256i vxj;
+         __m256i vw;
+         vxj = _mm256_set1_epi32(*(int*)&x[*idx++]);
+         vw = _mm256_loadu_si256((const __m256i *)w); //_mm256_lddqu_si256?
+         tmp = _mm256_maddubs_epi16(vxj, vw); //swap?
+         tmp = _mm256_madd_epi16(tmp, ones);
+         vy0 = _mm256_add_epi32(vy0, tmp);
+         w += 32;
+         vxj = _mm256_set1_epi32(*(int*)&x[*idx++]);
+         vw = _mm256_loadu_si256((const __m256i *)w); //_mm256_lddqu_si256?
+         tmp = _mm256_maddubs_epi16(vxj, vw); //swap?
+         tmp = _mm256_madd_epi16(tmp, ones);
+         vy0 = _mm256_add_epi32(vy0, tmp);
+         w += 32;
+         vxj = _mm256_set1_epi32(*(int*)&x[*idx++]);
+         vw = _mm256_loadu_si256((const __m256i *)w); //_mm256_lddqu_si256?
+         tmp = _mm256_maddubs_epi16(vxj, vw); //swap?
+         tmp = _mm256_madd_epi16(tmp, ones);
+         vy0 = _mm256_add_epi32(vy0, tmp);
+         w += 32;
+         vxj = _mm256_set1_epi32(*(int*)&x[*idx++]);
+         vw = _mm256_loadu_si256((const __m256i *)w); //_mm256_lddqu_si256?
+         tmp = _mm256_maddubs_epi16(vxj, vw); //swap?
+         tmp = _mm256_madd_epi16(tmp, ones);
+         vy0 = _mm256_add_epi32(vy0, tmp);
+         w += 32;
+      }
+#endif
+      for (;j<colblocks;j++)
       {
          __m256i tmp;
          __m256i vxj;