From e94e24509c8b7868c22b9007712ba0ed88ca1c93 Mon Sep 17 00:00:00 2001
From: Yida Wang <yidawa@gmail.com>
Date: Thu, 1 Mar 2018 12:14:44 -0800
Subject: [PATCH] remove the pragma primitives for better performance when the
threads are binded (#949)
---
topi/python/topi/x86/conv2d_avx_1x1.py | 9 ---------
topi/python/topi/x86/conv2d_avx_common.py | 9 ---------
2 files changed, 18 deletions(-)
diff --git a/topi/python/topi/x86/conv2d_avx_1x1.py b/topi/python/topi/x86/conv2d_avx_1x1.py
index cc264d04a..afd0be2e2 100644
--- a/topi/python/topi/x86/conv2d_avx_1x1.py
+++ b/topi/python/topi/x86/conv2d_avx_1x1.py
@@ -77,9 +77,6 @@ def _schedule_conv(s, data, data_pad, data_vec, kernel, kernel_vec, conv_out, ou
batch, ic_chunk, ih, ic_block, iw = s[A1].op.axis
parallel_axis = s[A1].fuse(ic_chunk, ih)
s[A1].parallel(parallel_axis)
- s[A1].pragma(batch, "parallel_launch_point")
- s[A1].pragma(parallel_axis, "parallel_stride_pattern")
- s[A1].pragma(batch, "parallel_barrier_when_finish")
# schedule kernel pack
oc_chunk, ic_chunk, oh, ow, ic_block, oc_block = s[W].op.axis
@@ -88,9 +85,6 @@ def _schedule_conv(s, data, data_pad, data_vec, kernel, kernel_vec, conv_out, ou
s[W].vectorize(oc_block)
parallel_axis = s[W].fuse(oc_chunk, oh)
s[W].parallel(parallel_axis)
- s[W].pragma(parallel_axis, "parallel_launch_point")
- s[W].pragma(parallel_axis, "parallel_stride_pattern")
- s[W].pragma(parallel_axis, "parallel_barrier_when_finish")
C, O0, O = conv_out, output, last
CC = s.cache_write(C, 'global')
@@ -128,8 +122,5 @@ def _schedule_conv(s, data, data_pad, data_vec, kernel, kernel_vec, conv_out, ou
s[O].vectorize(oc_block)
s[O].parallel(parallel_axis)
- s[O].pragma(batch, "parallel_launch_point")
- s[O].pragma(parallel_axis, "parallel_stride_pattern")
- s[O].pragma(batch, "parallel_barrier_when_finish")
return s
diff --git a/topi/python/topi/x86/conv2d_avx_common.py b/topi/python/topi/x86/conv2d_avx_common.py
index 4f5be019f..f4c0e453e 100644
--- a/topi/python/topi/x86/conv2d_avx_common.py
+++ b/topi/python/topi/x86/conv2d_avx_common.py
@@ -90,9 +90,6 @@ def _schedule_conv(s, data, data_pad, data_vec, kernel, kernel_vec, conv_out, ou
batch, ic_chunk, ih, ic_block, iw = s[A1].op.axis
parallel_axis = s[A1].fuse(ic_chunk, ih)
s[A1].parallel(parallel_axis)
- s[A1].pragma(batch, "parallel_launch_point")
- s[A1].pragma(parallel_axis, "parallel_stride_pattern")
- s[A1].pragma(batch, "parallel_barrier_when_finish")
# schedule kernel pack
oc_chunk, ic_chunk, oh, ow, ic_block, oc_block = s[W].op.axis
@@ -101,9 +98,6 @@ def _schedule_conv(s, data, data_pad, data_vec, kernel, kernel_vec, conv_out, ou
s[W].vectorize(oc_block)
parallel_axis = s[W].fuse(oc_chunk, oh)
s[W].parallel(parallel_axis)
- s[W].pragma(parallel_axis, "parallel_launch_point")
- s[W].pragma(parallel_axis, "parallel_stride_pattern")
- s[W].pragma(parallel_axis, "parallel_barrier_when_finish")
# schedule conv
C, O0, O = conv_out, output, last
@@ -144,8 +138,5 @@ def _schedule_conv(s, data, data_pad, data_vec, kernel, kernel_vec, conv_out, ou
s[O].vectorize(oc_block)
s[O].parallel(parallel_axis)
- s[O].pragma(batch, "parallel_launch_point")
- s[O].pragma(parallel_axis, "parallel_stride_pattern")
- s[O].pragma(batch, "parallel_barrier_when_finish")
return s
--
GitLab