Commit Graph

50 Commits

Author SHA1 Message Date
Philippe Tillet
c847cc6320 [DRIVER] Added options for developers to cache PTX file so that ti can
be manually modified
2021-02-09 00:09:10 -05:00
Philippe Tillet
0b23f95b20 [RUNTIME] Added option to print LLVM-IR
Also includes appropriate driver code change for that
2021-01-31 01:01:32 -05:00
Philippe Tillet
e11077eab9 [RUNTIME] Disable error on spills 2021-01-14 20:33:34 -05:00
Philippe Tillet
af080740f2 [GENERAL] Merged v1.0alpha into master. Added features are:
- A100 support via mma.16816
- Thread swizzling for conflict-free shared memory accesses without
padding
- Complete overhaul of the LLVM code generation in
codegen/selection/generator.cc to remove overengineering
- Added debugging capabilities in the Python binding
- Compilation error for kernels that spill
2021-01-11 19:23:24 -05:00
Philippe Tillet
ec5c2ad571 [DRIVER] Better exception handling of invalid ptx 2020-11-27 17:00:11 -05:00
Philippe Tillet
7710e048f4 [DRIVER] Simplified Driver API by substantially removing reliance on driver::context 2020-11-26 00:38:25 -05:00
Philippe Tillet
947ed0d46c [DRIVER] Added (slow) support for CUDA11 and Ampere 2020-11-25 13:52:03 -05:00
Philippe Tillet
6c5284ed3b [GENERAL] Various bugfixes 2020-11-11 14:44:56 -05:00
Philippe Tillet
a2d54b5ad3 [General] LLVM-9 -> LLVM-10 2020-11-07 22:46:18 -05:00
Philippe Tillet
9be1d5afc2 [GENERAL] Various improvements:
* Sparse einsum in triton.ops.einsum
* Hacky support for fixed-tile-size atomic-add
* Various bugfixes in parser
2020-10-25 12:16:40 -07:00
Philippe Tillet
db7a72bfe3 [DRIVER] Removed OpenCL support
There is no plan to support OpenCL anytime soon (Vulkan would be preferred). Removing the adequate portion of the driver code
2020-10-13 20:57:32 -07:00
Philippe Tillet
0cbee3ec56 [CODEGEN] More work on the CPU backend 2020-09-14 10:59:40 -04:00
Philippe Tillet
cfcda09679 [DRIVER] Now always using PTXv6.4 2020-07-29 13:50:01 -04:00
Philippe Tillet
52eca7676b [TRITON] Fixed misaligned address issue 2020-06-05 12:00:42 -04:00
Philippe Tillet
cd21151b98 [GENERAL] Fixed some undefined behavior with GCC-9 2020-05-11 11:07:21 -04:00
Philippe Tillet
7c09ff80eb [CORE] Fixed several issues that arose in the development of the
torch-blocksparse package:

* Now using warp shuffle in reductions when possible
* Various bugfixes in layout inference
* Added INFINITY, exponential and select
* Better error messages for unimplemented constructs
2020-03-31 18:57:28 -04:00
Philippe Tillet
a27b96cad1 [CORE][DRIVER] Now only using PTX6.4 if CUDA10.1+ is detected 2020-03-05 21:29:22 -05:00
Philippe Tillet
f2daff85d2 [GENERAL] Improved caching mechanism:
* Now computing hash in libtriton
* Now only compiling a single pytorch hook per function signature
2020-02-24 16:36:50 -05:00
Philippe Tillet
f278d9741a [GENERAL] Merged einsum feature branch. Various feature, performance
improvements and bugfixes:

* Added preliminary support for extended Einstein summation in PyTriton
* Significant performance improvement on FP32 kernels containing matrix
multiplication
* Added re-coalescing pass for FP16 kernels containing matrix
multiplication
* Various bugfixes
2020-01-20 12:42:48 -05:00
Philippe Tillet
f4bbbbe5e4 [PYTHON][OPS] Bugfix in conv fprop 2019-11-01 00:43:02 -04:00
Philippe Tillet
96cba9036a [tests] [unit] added 1D and 3D reduction test 2019-10-20 17:48:19 -04:00
Philippe Tillet
b43454c9b7 [codegen] [membar] view do not write to shared memory 2019-10-17 22:38:41 -04:00
Philippe Tillet
e787ce0cab [codegen] more cleaning 2019-10-13 02:26:30 -04:00
Philippe Tillet
ee3803b577 more cleaning 2019-10-11 19:29:24 -04:00
Philippe Tillet
323c90e431 ugh 2019-10-11 19:05:54 -04:00
Philippe Tillet
a1e0512703 [codegen] more progress 2019-10-03 14:11:50 -04:00
Philippe Tillet
86a3e5d897 [codegen] now matrix-multiplication is bank-conflict free for all
layouts
2019-10-01 16:57:59 -04:00
Philippe Tillet
ed1b2bc563 more work on padding 2019-09-27 22:15:30 -04:00
Philippe Tillet
a3bf3a1804 [codegen] more hmma row-major handling 2019-09-24 19:35:46 -04:00
Philippe Tillet
c24d55db23 [codegen] more work on hmma coalescing 2019-09-23 20:38:27 -04:00
Philippe Tillet
f0013f8bf1 [codegen] [allocation] fixed issues in HMMA 2019-09-23 17:54:42 -04:00
Philippe Tillet
b95ac15d48 [codegen] [selection] fixed synchronization issue with double-buffering 2019-09-23 13:56:46 -04:00
Philippe Tillet
001973630e [codegen] cleaned up shared memory and double-buffering logic 2019-09-21 22:21:40 -04:00
Philippe Tillet
e35be1ddcf [ir][instruction] added identifier for each instruction 2019-09-19 16:25:36 -04:00
Philippe Tillet
1fd9be27ee [tests][bench] now benchmarking all variants of copy 2019-09-17 22:17:58 -04:00
Philippe Tillet
e01e623333 [codegen][auto-coalesce] more debugging 2019-09-16 20:34:08 -04:00
Philippe Tillet
579a662e60 [codegen][coalesce] more bugfixes 2019-09-13 14:17:21 -04:00
Philippe Tillet
11ff27d638 [codegen][coalesce] some bugfix for phi-nodes 2019-09-12 22:44:07 -04:00
Philippe Tillet
32234c2612 ugh 2019-09-08 17:35:24 -04:00
Philippe Tillet
18848cbb71 [driver] now passing std::unique_ptr<> instead of cloning LLVM module
when compiling it
2019-09-05 17:25:58 -04:00
Philippe Tillet
b747959a57 trying to work around tensorflow limitations 2019-09-04 01:54:43 -04:00
Philippe Tillet
2ccc915011 [python][examples] added template for blocksparse 2019-09-03 20:44:27 -04:00
Philippe Tillet
97fdb5b6be [tests] added missing files 2019-09-03 12:44:35 -04:00
Philippe Tillet
a842d337c5 [general] various cleaning and bugfix:
* added copy1d and copy2d benchmark
* fixed issue in reassociation pass
2019-09-02 23:00:49 -04:00
Philippe Tillet
7e0af2118c [codegen] worked around bug seemingly from nvptx/ptxas by simplifying multiplications by 1:
- Generated LLVM-IR looked correct
- Illegal addressing disappeared when running cuda-memcheck
- Illegal addressing disappeared when using nvptx-short-pointer
2019-08-30 16:45:14 -07:00
Philippe Tillet
d457482539 [codegen] fixed issue in double buffering pointer update 2019-08-28 17:50:45 -07:00
Philippe Tillet
37cbcfabd0 [examples] back to 96 TFLOPS on V100 2019-08-26 22:49:14 -07:00
Philippe Tillet
b4ae06a714 tracking down performance regression 2019-08-26 20:38:39 -07:00
Philippe Tillet
cb04ec0b3b some more cleaning 2019-08-23 19:22:38 -07:00
Philippe Tillet
732156b942 [general] rename *.cpp -> *.cc 2019-08-23 19:06:39 -07:00