squall/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity
171 Commits 1 Branch 0 Tags 45 MiB
6fc5008803
Commit Graph

171 Commits

This Branch
This Branch
All Branches
Author SHA1 Message Date
Artem Belevich
fb8b3a98b7 Addressed code review comments. 2019-05-10 10:24:52 -07:00
gthomascollignon
d9d357877f Added missing file (#48) 2019-05-09 14:07:52 -07:00
Artem Belevich
e18292db46 Make CUTLASS compileable with Clang. 
Requires a recent clang build (r359248 or newer).

Enable compilation with clang with these options:
cmake -DCUDA_COMPILER=clang -DCMAKE_CXX_COMPILER=/path/to/clang++
 2019-05-02 11:00:22 -07:00
Timmy
fe3438a3c1 cutlass 1.3.1 (#46) 
CUTLASS 1.3.1 patch resolves failing text with NVRTC.
 2019-04-19 16:54:52 -07:00
Andrew Kerr
877bdcace6
Cutlass 1.3 Release (#42) 
CUTLASS 1.3 Release
- Efficient GEMM kernel targeting Volta Tensor Cores via mma.sync instruction added in CUDA 10.1.
 2019-03-20 10:49:17 -07:00
Andrew Kerr
19a9d64e3c
Removed patch version from README. 
Removed patch version from README.
 2018-12-19 15:20:43 -08:00
Andrew Kerr
80e6f7c860
Merge pull request #38 from NVIDIA/resolve_maxwell 
Resolved issue for incorrect SGEMM on Maxwell architecture.
 2018-12-19 15:17:41 -08:00
akerr
822b0952cd Resolved issue for incorrect SGEMM on Maxwell architecture. 2018-12-19 15:07:16 -08:00
Andrew Kerr
ed2ed4d667
Merge pull request #33 from NVIDIA/cutlass_1.2 
CUTLASS 1.2
 2018-10-26 14:59:50 -07:00
Andrew Kerr
4db423c40f
Minor edit to CHANGELOG. 2018-10-26 14:58:31 -07:00
Andrew Kerr
b2bc0d3b79 Updating Doxygen docs 2018-10-26 14:54:58 -07:00
akerr
74df0331f2 CUTLASS 1.2 2018-10-26 14:38:46 -07:00
Andrew Kerr
2332df492e
Merge pull request #30 from NVIDIA/fix_utilities_example 
Fixed cutlass_utilities example.
 2018-09-29 15:09:18 -07:00
akerr
cfe4b933ef CUDA 9 lacks host-side conversions from float=>half. Instead, we must reinterpret_cast<> from cutlass::half_t => half. 2018-09-29 15:04:20 -07:00
Andrew Kerr
6877595a5e
Merge pull request #28 from NVIDIA/cutlass_1.1 
Fixed typeo
 2018-09-28 12:59:49 -07:00
Andrew Kerr
69e3709da4
Fixed typeo 
Fixed typeo
 2018-09-28 12:59:20 -07:00
Andrew Kerr
d419094c28
Merge pull request #26 from NVIDIA/cutlass_1.1 
Clarification to README
 2018-09-21 11:44:47 -07:00
akerr
1a7ac522f8 Clarification to README 2018-09-20 11:04:03 -07:00
Andrew Kerr
bf6eec53eb
Merge pull request #25 from NVIDIA/cutlass_1.1 
Updated CUTLASS.md
 2018-09-19 21:33:04 -07:00
akerr
206e38dac5 Updated copyright of CUTLASS.md 2018-09-19 21:31:12 -07:00
Andrew Kerr
d85f6a1cec
Merge pull request #24 from NVIDIA/cutlass_1.1 
Cutlass 1.1
 2018-09-19 21:16:53 -07:00
akerr
0826572c4c Reduced range of random values to avoid bit-level inconsistencies for large matrices. 2018-09-19 21:11:48 -07:00
akerr
77d1e0ca81 Updated README and CHANGELOG. 2018-09-19 20:42:51 -07:00
akerr
d7137f9c0a Updated doxygen 2018-09-19 14:02:08 -07:00
akerr
461f417b9d Checkpointing CUTLASS 1.1 release. 2018-09-18 16:58:03 -07:00
Andrew Kerr
cf0301e00f
Merge pull request #15 from NVIDIA/release_1.0.1_edits 
Minor edits to README and changelog pursuant CUTLASS 1.0.1 patch.
 2018-06-26 13:59:01 -07:00
akerr
b9bb0d1a49 Edits to README and changelog pursuant CUTLASS 1.0.1 patch. 2018-06-26 13:57:39 -07:00
Andrew Kerr
e1c4ba501b
Merge pull request #13 from NVIDIA/cutlass_v1.0.1 
Cutlass v1.0.1
 2018-06-12 08:25:56 -07:00
akerr
c566e83e6d Updated changelog. 2018-06-11 14:54:07 -07:00
akerr
374882be53 Replaced GoogleTest copy with submodule. Added updates to support intra-threadblock reductions. Added tests for same. 2018-06-11 11:47:15 -07:00
akerr
2c496c3e9e Replaced GoogleTest copy with Git submodule. 2018-06-11 11:32:41 -07:00
Andrew Kerr
9fd55460c6
Merge pull request #10 from NVIDIA/cutlass_v1.0_rel 
Minor updates to usage and README.
 2018-05-18 12:27:31 -07:00
akerr
480732c2e8 Minor updates to usage and readme. 2018-05-17 15:10:55 -07:00
Andrew Kerr
68aaee8773
Merge pull request #9 from NVIDIA/cutlass_v1.0_rel 
Updated URL to Doxygen and modified usage statement
 2018-05-17 11:12:37 -07:00
akerr
acb90e962a Updated url to Doxygen and modified usage statement in performance test program. 2018-05-17 11:11:05 -07:00
Andrew Kerr
96bc3f227f
Merge pull request #8 from NVIDIA/cutlass_v1.0_rel 
Configured Github Pages
 2018-05-16 15:26:55 -07:00
akerr
25ff282403 Moved Doxygen documents. 2018-05-16 15:25:24 -07:00
Andrew Kerr
9d5726a568 Set theme jekyll-theme-minimal 2018-05-16 13:49:06 -07:00
Andrew Kerr
6f0d271d8d
CUTLASS v1.0 
CUTLASS v1.0 released.
 2018-05-16 13:47:13 -07:00
akerr
923dfb42ce Updated README.md 2018-05-16 12:50:10 -07:00
akerr
6f6f269a0a Updated README.md 2018-05-16 12:47:07 -07:00
akerr
2028ebe120 CUTLASS v1.0 release 2018-05-16 11:44:56 -07:00
Andrew Kerr
84377249a1
Merge pull request #2 from Artem-B/clang-fixes 
Merging "Clang fixes" into master.
 2018-01-04 15:52:53 -08:00
akerr
901287175f Merge branch 'Artem-B-clang-fixes' 2018-01-04 15:46:08 -08:00
Artem Belevich
1c9b54df16 Whitespace fix. 2018-01-03 16:42:51 -08:00
Artem Belevich
39616514d0 Reworked CUDA_LOG macro to print location&the message with one printf. 
This replies on the fact that clang allows using device-side features
from __host__/__device__ functions from __host__ ones as long as we
don't have to generate code for that. Wrapping thread/blockIdx in
__host__ __device__ function allows using CUDA_LOG everywhere during
host and device compilation.
 2018-01-03 16:36:50 -08:00
Artem Belevich
df4b4e4bb6 Added _cuda_ to the name of the executable to indicate that it's not clang's version. 2017-12-11 16:34:10 -08:00
Artem Belevich
81957b3a3d Force inlining of few functions that rely on that for performance. 
Clang is less agressive than nvccnvcc, so number of functions did not getn
inlined into the kernel by default. That prevented SROA from eliminating
loads/stores to temporary buffers and resulted in abysmal performance.

Replaced inline with __forceinline__ to ensure that we do inline the
functions necessary for optimal performance.
 2017-12-11 14:52:30 -08:00
Artem Belevich
ce2b3f695d Fixed debug macros for clang. 
Unlike nvcc, clang always sees both host and device-side code during
compilation. CUDA_LOG macro is used in both host and device code, so when it
expanded to contain device-only code, that resulted in errors when it was used
from the host-side functions.

In order to make CUDA_LOG work with clang it was split into two parts -- a pair
of target-attribute-based overloaded functions that perform host or device
specific parts of logging, and a printf which works on both sides.
 2017-12-11 14:52:30 -08:00
Artem Belevich
e9e7cd4d44 Make cutlass compilable with clang. 
E.g:
PATH=/nvcc/path/bin:/clang/path/bin:$PATH make sm=35,60 compiler=clang all
 2017-12-11 14:52:30 -08:00