gaspersic/triton
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
448 Commits 14 Branches 9 Tags
073be1d2ee38353a6e71c535b3460bc160fd3822
Commit Graph

448 Commits

This Branch
This Branch
All Branches
Author SHA1 Message Date
Philippe Tillet
073be1d2ee [FRONTEND] check that tensors have power-of-two number of elements (#499) 2022-04-14 19:30:02 -07:00
Philippe Tillet
5c7122004c [TUTORIALS] Tutorial shouldn't expose clock. Just removed it. 2022-04-14 17:33:44 -07:00
Philippe Tillet
dc4d40faec [FRONTEND] now mangle constexpr float containing "e-" 2022-04-14 10:26:48 -07:00
Philippe Tillet
25f6689508 [FRONTEND] rename current stream monkey patch (#495) 2022-04-13 11:45:55 -07:00
Philippe Tillet
76bfac9f15 [FRONTEND] Improved constexpr handling (#493) 2022-04-12 00:02:54 -07:00
Philippe Tillet
14b0fd4cfb [FRONTEND] Added possibility for users to customize current stream query (#492) 2022-04-07 12:11:32 -07:00
Philippe Tillet
6424771f55 [CI] Documentation fixup 2022-04-07 09:42:35 -07:00
Philippe Tillet
9f08ecd684 [FRONTEND] Semantic analysis refactor (#491) 
Moved dispatch.cc to semantic.py (@ptillet)
Integer signedness analysis was moved from C++ to python (@daadaada)
Cleaner frontend types (@daadaada)
Moved SSA construction to a separate object (@ptillet)


Co-authored-by: Yan Da <dyanab@connect.ust.hk>
 2022-04-06 16:13:53 -07:00
Philippe Tillet
2bed6fc850 [LANG] Added support for device functions (#484) 2022-04-03 20:58:16 -07:00
apd10
e85c7a7fc7 Bugfix in ptxas path. (#487) 
Bug: "ret" value is destroyed when a failing "ptxas --version" is run
overwriting the previous valid "ret" value.

Fix: keep rets only for those runs which are successful. Pick the first
one
 2022-03-30 20:45:41 -07:00
Philippe Tillet
bace26143d [TUTORIALS] Removed leftover print 2022-03-28 16:53:23 -07:00
Philippe Tillet
e0cc488055 [FRONTEND] Added tl.clock and tl.globaltimer (#485) 2022-03-28 16:15:43 -07:00
Philippe Tillet
76a9ee50a8 Revert "[FRONTEND] Semantic analysis refactor (#473)" (#483) 
This reverts commit 539961072c.
 2022-03-24 17:16:50 -07:00
Philippe Tillet
ea6d1f1b85 [DRIVER] LLVM driver fixup (#482) 
Current way of doing things is probably not super thread safe. init is shared between threads and some threads my not call the LLVMInitialize* function.
 2022-03-23 00:24:45 -07:00
Keren Zhou
a4f68165cd [FRONTEND] Hot fix for lineno (#481) 
Override __reduce__ to make CompilationError pickable and print out error messages
 2022-03-22 22:09:49 -07:00
daadaada
539961072c [FRONTEND] Semantic analysis refactor (#473) 
Moved dispatch.cc to semantic.py
Integer signedness now moved from C++ to python
Cleaner frontend type

Co-authored-by: Phil Tillet <phil@openai.com>
 2022-03-16 21:25:30 -07:00
Yongjik Kim
0dd2ec2e3a [FRONTEND] Add an assert in case we get a CPU tensor. (#478) 2022-03-16 14:38:56 -07:00
Philippe Tillet
d4d8eaf6c0 [FRONTEND] improved caching mechanism (#474) 
Co-authored-by: Greg Brockman <gdb@gregbrockman.com>
Co-authored-by: Christopher Hesse <christopherhesse@users.noreply.github.com>
 2022-03-15 12:20:51 -07:00
Doğukan Tuna
21f8a0646d [DOCS] Minor README.md (#470) 
Added binary distribution for quick installation
 2022-03-05 00:50:37 -08:00
Philippe Tillet
a50a47a85b [CODEGEN] Reverted some changes from previous PR; fixed vectorization characteristics of mma layout (#469) 2022-03-04 01:53:31 -08:00
Philippe Tillet
bb5765df5c [CODEGEN] Now padding shared memory for layout conversion (#468) 2022-03-03 22:19:05 -08:00
daadaada
d9dd97492f Use unique_ptr in ir::context_impl (#462) 
Co-authored-by: Philippe Tillet <Phil.Tillet@gmail.com>
 2022-02-24 16:07:10 -08:00
Philippe Tillet
98ed7db8c1 [CODEGEN] Improvements and bugfixes (#463) 2022-02-24 14:56:24 -08:00
daadaada
a9dfdcaaa9 [FRONTEND] Make the performance model work for int8, tf32, and fp32 (#456) 2022-02-11 22:34:42 -08:00
Philippe Tillet
9b100302d3 [FRONTEND] Now using pybind11 to release GIL (#458) 2022-02-10 01:57:39 -08:00
Philippe Tillet
40093a9878 [DOCS] Multiple versions are now supported (#457) 2022-02-09 01:32:41 -08:00
Philippe Tillet
4941bc7001 [DOCS] Some more fixes (#455) 2022-02-08 16:53:56 -08:00
Philippe Tillet
2fdf0a4fe8 [DOCS] changed build command 2022-02-08 11:45:21 -08:00
Philippe Tillet
077d6c8ff0 [DOCS] re-activated tutorials 2022-02-08 11:42:39 -08:00
Philippe Tillet
822ddcd14b [DOCS] Added versioning (#453) 2022-02-08 11:28:18 -08:00
Philippe Tillet
7b48340ffd [CI] Some fixes for the build (#451) 2022-02-06 19:11:33 -08:00
Philippe Tillet
5a8a544d10 [OPS][BLOCKSPARSE] Improved robustness, clarity and performance (#450) 
* dds layout now internally re-uses dsd code path for increased code 
* at_mask and kp_mask related things are now dropped from the softmax API. I couldn't think of any case where it was needed beyond is_causal. And if there is any, we should probably find a way to get it implemented statically so that users don't have to materialize masks.
 * fixed bug in blocksparse matmul that caused troubles when layout had a full row/col of zeros
 * blocksparse softmax now no longer modifies any data in-place
 * blocksparse softmax now takes an is_dense arguments that provides better performance. Passing is_dense=True, is_causal=True is the best way to achieve triangular attention.
  * unit tests now test backward pass
 2022-02-06 18:00:45 -08:00
Philippe Tillet
69ff52ea1f [CODEGEN] removed buggy (and mostly useless) optimization in peephole pass (#449) 2022-02-05 21:37:23 -08:00
TC
137bb67fad [LANG] Add fp16 to fp8 conversion (#444) 2022-02-02 20:42:09 -08:00
Philippe Tillet
3b20170fa3 Merge pull request #448 from openai/v2.0 
`v2.0` is now merged into `master`
 2022-01-30 20:49:08 -08:00
Philippe Tillet
b0d6e2f322 [STYLE] run autopep 2022-01-30 20:27:44 -08:00
Philippe Tillet
2922dc141c Merge branch 'master' into v2.0 2022-01-30 20:25:01 -08:00
Philippe Tillet
807d8a1945 [ALL] Merge master (#447) 2022-01-30 20:21:20 -08:00
Philippe Tillet
bef76b142a [BACKEND] float division is now approximate by default (#446) 2022-01-29 18:29:29 -08:00
Philippe Tillet
bd52e530a0 [OPS][BLOCKSPARSE] Fix padding issue in DSD LUT (#445) 2022-01-28 21:40:30 -08:00
daadaada
e68d6a7776 [BACKEND] Making the warp-level tile "more square" to increase data-reuse for tl.dot. (#442) 
* Increase smem data-reuse for some layouts

* tweak

* Keep the original tiling logic for sm < 80

Co-authored-by: Philippe Tillet <phil@openai.com>
 2022-01-27 09:59:54 -08:00
daadaada
59d371c6eb [BACKEND] Added Int8 mma (#440) 2022-01-27 09:12:44 -08:00
Benjamin Lefaudeux
3a23c1dd33 [BACKEND] minor, hotfix for gcc compilation (#439) 2022-01-23 14:24:02 -08:00
Philippe Tillet
ccf9abe0ba [FRONTEND][RANDOM] Improved backward compatibility of RNG (#438) 
The unsigned int PR definitely improved our RNG. However, it requires
different floating point arithmetics which, means the results are not
bit-wise identical to how they were before. This commit revives backward
compatibility, but we should change it back to the "right" way later.
 2022-01-21 18:05:55 -08:00
Philippe Tillet
4c97d1ecd7 [FRONTEND] Bunch of fixes here and there (#436) 2022-01-20 10:55:59 -08:00
Philippe Tillet
e0c5709cc8 [FRONTEND] Fixed semantics bug on ptr to bool conversions (#432) 2022-01-17 18:00:03 -08:00
daadaada
2a944ded53 [TESTS] Added bfloat16 tests (#430) 2022-01-13 23:38:32 -08:00
Philippe Tillet
4c94359199 [FRONTEND] Alignment fix-up (#428) 2022-01-11 23:11:58 -08:00
Philippe Tillet
bbc78f6516 [FRONTEND][RANDOM] Make sure offset dtype is always uint32 before calling uint32_to_uniform_float (#427) 2022-01-11 11:08:49 -08:00
Botao Yu
bf32205edc [OPS][BLOCKSPARSE] Remove unnecessary loop and add cuda bool layout support (#425) 2022-01-11 11:07:16 -08:00