NVIDIA/cutlass
1
0
Fork 0
mirror of https://github.com/NVIDIA/cutlass.git synced 2026-05-11 17:00:05 +00:00
Code Issues Packages Projects Releases Wiki Activity

Update generator.py

Browse Source 
stop generating analytical conv kernels to reduce kernel number
This commit is contained in:
Haicheng Wu
2022-05-08 21:47:15 -04:00
committed by GitHub
parent 6023038bae
commit 1604ebaf10
1 changed files with 5 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
tools/library/scripts/generator.py
View File

@@ -344,8 +344,9 @@ def CreateConv2dOperator(manifest, layout, tile_descriptions, data_type, alignme
# one exceptional case
# iterator algorithm (analytic and optimized)
iterator_algorithms = [IteratorAlgorithm.Analytic, IteratorAlgorithm.Optimized]
#iterator_algorithms = [IteratorAlgorithm.Analytic, IteratorAlgorithm.Optimized]
iterator_algorithms = [IteratorAlgorithm.Optimized]
# by default, only generate the largest tile size, largest alignment, and optimized iterator
if manifest.kernel_filter == '':
tile_descriptions = [tile_descriptions[0],]
@@ -526,7 +527,8 @@ def CreateConv3dOperator(manifest, layout, tile_descriptions, data_type, alignme
alignment_c = min(8, alignment)
# iterator algorithm (analytic and optimized)
iterator_algorithms = [IteratorAlgorithm.Analytic, IteratorAlgorithm.Optimized]
#iterator_algorithms = [IteratorAlgorithm.Analytic, IteratorAlgorithm.Optimized]
iterator_algorithms = [IteratorAlgorithm.Optimized]
# by default, only generate the largest tile size and optimized iterators
if manifest.kernel_filter == '':