From c045702e51e04006f81f6f473ab5326eb24aace6 Mon Sep 17 00:00:00 2001
From: Jean-Marc Valin <jmvalin@jmvalin.ca>
Date: Thu, 24 Dec 2020 02:46:19 -0500
Subject: [PATCH] Add non-dot-product AVX code
---
dnn/vec_avx.h | 42 ++++++++++++++++++++++++++++++++++++++++++
1 file changed, 42 insertions(+)
diff --git a/dnn/vec_avx.h b/dnn/vec_avx.h
index 80b022909..e11fae04e 100644
--- a/dnn/vec_avx.h
+++ b/dnn/vec_avx.h
@@ -218,3 +218,45 @@ static void sparse_sgemv_accum16(float *out, const float *weights, int rows, con
}
}
+#ifdef DOT_PROD
+#else
+static void sparse_sgemv_accum8x4(float *out, const float *weights, int rows, const int *idx, const float *x)
+{
+ int i, j;
+ for (i=0;i<rows;i+=8)
+ {
+ float * restrict y;
+ int cols;
+ __m256 vy0;
+ y = &out[i];
+ vy0 = _mm256_loadu_ps(&y[0]);
+ cols = *idx++;
+ for (j=0;j<cols;j++)
+ {
+ int id;
+ __m256 vxj;
+ __m256 vw;
+ id = *idx++;
+ vxj = _mm256_broadcast_ss(&x[4*id]);
+ vw = _mm256_loadu_ps(&weights[0]);
+ vy0 = _mm256_fmadd_ps(vw, vxj, vy0);
+
+ vxj = _mm256_broadcast_ss(&x[4*id+1]);
+ vw = _mm256_loadu_ps(&weights[8]);
+ vy0 = _mm256_fmadd_ps(vw, vxj, vy0);
+
+ vxj = _mm256_broadcast_ss(&x[4*id+2]);
+ vw = _mm256_loadu_ps(&weights[16]);
+ vy0 = _mm256_fmadd_ps(vw, vxj, vy0);
+
+ vxj = _mm256_broadcast_ss(&x[4*id+3]);
+ vw = _mm256_loadu_ps(&weights[24]);
+ vy0 = _mm256_fmadd_ps(vw, vxj, vy0);
+
+ weights += 32;
+ }
+ _mm256_storeu_ps (&y[0], vy0);
+ }
+}
+#endif
+
--
GitLab