leavelet/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 28 Packages Projects Releases Wiki Activity

CUTLASS 2.6 (#298)

Browse Source 
CUTLASS 2.6
This commit is contained in:
Manish Gupta
2021-07-23 00:40:53 -04:00
committed by GitHub
parent 6c29fe20ba
commit e5d51840e8
308 changed files with 32408 additions and 4722 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/13_two_tensor_op_fusion/README.md
+4
View File
@@ -48,6 +48,10 @@ addition to its own input activation tile. Therefore the input activation warp t
2nd GEMM/Conv only depends on the output warp accumulator of the 1st GEMM/Conv in the
register file, and the operation can be fully register-file-resident.
When applying the above constraint to convolutions, it is required that the 2nd Convolution
kernel doesn't have halos such that data used by each threadblock doesn't depend on any other
threadblock. Typically this requires the 2nd Convolution uses 1x1 filter without any paddings.
# Copyright
Copyright (c) 2017-2021, NVIDIA CORPORATION. All rights reserved.