commit 87b6cbb6fdec3789a409b02665e731ba67b4377c
author Jongsoo Park <jongsoo@fb.com> Mon Mar 18 16:28:47 2019 -0700
committer Facebook Github Bot <facebook-github-bot@users.noreply.github.com> Mon Mar 18 16:31:42 2019 -0700
tree a4b1556fc136b91412303016117364157aaabd3a
parent 0a8efce51ea38abdaee643ee6758516853171c1e

fix bug in pool_dnnlowp_op_avx2.cc (#18141)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18141

VLEN should've been 32

Reviewed By: jianyuh

Differential Revision: D14510780

fbshipit-source-id: ddf12746e1c69677a268432432ddb088cc210084

caffe2/quantization/server/pool_dnnlowp_op_avx2.cc

diff --git a/caffe2/quantization/server/pool_dnnlowp_op_avx2.cc b/caffe2/quantization/server/pool_dnnlowp_op_avx2.cc
index 92d0816..8c35ebd 100644
--- a/caffe2/quantization/server/pool_dnnlowp_op_avx2.cc
+++ b/caffe2/quantization/server/pool_dnnlowp_op_avx2.cc
@@ -34,7 +34,7 @@
       wstart = wstart > 0 ? wstart : 0;
 
       uint8_t* Yh = Ydata_temp + (ph * pooled_width + pw) * channels;
-      constexpr int VLEN = 8;
+      constexpr int VLEN = 32;
       // vectorized loop
       for (int c = 0; c < channels / VLEN * VLEN; c += VLEN) {
         __m256i Y_v = _mm256_setzero_si256();