From 1ada7d4d6f838dc0842fc89159747755c516ce24 Mon Sep 17 00:00:00 2001
From: Jean-Marc Valin <jmvalin@amazon.com>
Date: Fri, 3 Nov 2023 02:46:38 -0400
Subject: [PATCH] Vectorizing sgemv for multiples of 4 with SSE
---
dnn/vec_avx.h | 17 +++++++++++++++++
1 file changed, 17 insertions(+)
diff --git a/dnn/vec_avx.h b/dnn/vec_avx.h
index 767d7e193..a1d6cad27 100644
--- a/dnn/vec_avx.h
+++ b/dnn/vec_avx.h
@@ -709,6 +709,23 @@ static inline void sgemv(float *out, const float *weights, int rows, int cols, i
}
_mm256_storeu_ps (&y[0], vy0);
}
+ for (;i<rows-3;i+=4)
+ {
+ float *y;
+ __m128 vy0;
+ y = &out[i];
+ vy0 = _mm_setzero_ps();
+ for (j=0;j<cols;j++)
+ {
+ __m128 vxj;
+ __m128 vw;
+ vxj = _mm_broadcast_ss(&x[j]);
+
+ vw = _mm_loadu_ps(&weights[j*col_stride + i]);
+ vy0 = _mm_fmadd_ps(vw, vxj, vy0);
+ }
+ _mm_storeu_ps (&y[0], vy0);
+ }
for (;i<rows;i++)
{
out[i] = 0;
--
GitLab