mirror of
https://github.com/pytorch/pytorch.git
synced 2025-10-20 12:54:11 +08:00
ROCm Sparsity through HipSparseLT (#150578)
TLDR: - This pull request introduces support for hipSPARSELt in ROCm, current usage would be semi-structure sparsity. - Require **ROCm 6.4** && **gfx942/gfx950**. - The average performance uplift (compare to dense operation) is ~ 20% in ROCm 6.4 but expect further performance lift along the way. ### Dense vs. Sparse Performance Comparison #### **NT (Row-major)** **Average Uplift**: `1.20` | M | N | K | hipsparselt-bench (us) | hipblaslt-bench get all (us) | Uplift | |-------|--------|--------|-------------------------|-------------------------------|--------| | 14336 | 8 | 4096 | 20.05 | 25.3 | 1.26 | | 4096 | 8 | 14336 | 21.07 | 25.28 | 1.20 | | 3072 | 3072 | 10240 | 299.05 | 351.82 | 1.18 | | 3072 | 1536 | 768 | 18.56 | 20.05 | 1.08 | | 3072 | 17664 | 768 | 163.13 | 173.91 | 1.07 | | 3072 | 196608 | 768 | 1717.30 | 1949.63 | 1.14 | | 3072 | 24576 | 768 | 206.84 | 242.98 | 1.17 | | 3072 | 6144 | 768 | 53.90 | 56.88 | 1.06 | | 3072 | 98304 | 768 | 833.77 | 962.28 | 1.15 | | 768 | 1536 | 768 | 8.53 | 19.65 | 2.30 | | 768 | 17664 | 768 | 46.02 | 46.84 | 1.02 | | 768 | 196608 | 768 | 463.15 | 540.46 | 1.17 | | 768 | 24576 | 768 | 54.32 | 59.55 | 1.10 | | 768 | 6144 | 768 | 19.47 | 20.15 | 1.03 | | 768 | 98304 | 768 | 231.88 | 258.73 | 1.12 | --- #### **NN (Row-major)** **Average Uplift**: `1.13` | M | N | K | hipsparselt-bench (us) | hipblaslt-bench get all (us) | Uplift | |-----|--------|-------|-------------------------|-------------------------------|--------| | 768 | 1536 | 3072 | 27.50 | 28.78 | 1.05 | | 768 | 17664 | 3072 | 125.06 | 158.94 | 1.27 | | 768 | 196608 | 3072 | 1568.38 | 1767.12 | 1.13 | | 768 | 24576 | 3072 | 171.05 | 203.49 | 1.19 | | 768 | 6144 | 3072 | 58.72 | 60.39 | 1.03 | | 768 | 98304 | 3072 | 787.15 | 887.60 | 1.13 | ------------------------- This pull request introduces support for hipSPARSELt in ROCm, alongside various updates and improvements to the codebase and test suite. The changes primarily involve adding configuration flags, updating conditional checks, and ensuring compatibility with hipSPARSELt. ### ROCm and hipSPARSELt Support: * [`BUILD.bazel`](diffhunk://#diff-7fc57714ef13c3325ce2a1130202edced92fcccc0c6db34a72f7b57f60d552a3R292): Added `@AT_HIPSPARSELT_ENABLED@` substitution to enable hipSPARSELt support. * [`aten/CMakeLists.txt`](diffhunk://#diff-0604597797bb21d7c39150f9429d6b2ace10b79ab308514ad03f76153ae8249bR104-R110): Introduced a conditional flag to enable hipSPARSELt support based on ROCm version. * [`aten/src/ATen/CMakeLists.txt`](diffhunk://#diff-ce80f3115ab2f6be5142f0678a1fc92c6b2d7727766ce44f48726c99e720f777R37): Added `AT_HIPSPARSELT_ENABLED` configuration. * [`aten/src/ATen/cuda/CUDAConfig.h.in`](diffhunk://#diff-8bb82da825ca87c28233abacffa1b0566c73a54990b7a77f3f5108d3718fea15R11): Defined `AT_HIPSPARSELT_ENABLED` macro. * `caffe2/CMakeLists.txt`, `cmake/Dependencies.cmake`, `cmake/public/LoadHIP.cmake`: Included hipSPARSELt in the ROCm dependencies. [[1]](diffhunk://#diff-c5ee05f1e918772792ff6f2a3f579fc2f182e57b1709fd786ef6dc711fd68b27R1380) [[2]](diffhunk://#diff-12e8125164bbfc7556b1781a8ed516e333cc0bf058acb7197f7415be44606c72L1084-R1084) [[3]](diffhunk://#diff-b98e27b9a5f196a6965a99ee5a7bb15b3fc633d6375b767635b1b04ccb2fd3d5R153) ### Codebase Updates: * [`aten/src/ATen/native/sparse/cuda/cuSPARSELtOps.cpp`](diffhunk://#diff-ae921dd1584ab98fdd9c25a3521047795de702223f5b65fdaa45a5bd92b4d1f3R1-R6): Added hipSPARSELt support checks and initialization functions. Updated various methods to conditionally handle hipSPARSELt. [[1]](diffhunk://#diff-ae921dd1584ab98fdd9c25a3521047795de702223f5b65fdaa45a5bd92b4d1f3R1-R6) [[2]](diffhunk://#diff-ae921dd1584ab98fdd9c25a3521047795de702223f5b65fdaa45a5bd92b4d1f3R22-R67) [[3]](diffhunk://#diff-ae921dd1584ab98fdd9c25a3521047795de702223f5b65fdaa45a5bd92b4d1f3R78-R85) [[4]](diffhunk://#diff-ae921dd1584ab98fdd9c25a3521047795de702223f5b65fdaa45a5bd92b4d1f3R97-R109) [[5]](diffhunk://#diff-ae921dd1584ab98fdd9c25a3521047795de702223f5b65fdaa45a5bd92b4d1f3R183-R188) [[6]](diffhunk://#diff-ae921dd1584ab98fdd9c25a3521047795de702223f5b65fdaa45a5bd92b4d1f3L134-R200) [[7]](diffhunk://#diff-ae921dd1584ab98fdd9c25a3521047795de702223f5b65fdaa45a5bd92b4d1f3R213-R222) [[8]](diffhunk://#diff-ae921dd1584ab98fdd9c25a3521047795de702223f5b65fdaa45a5bd92b4d1f3L217-R285) ### Test Suite Updates: * [`test/test_sparse_semi_structured.py`](diffhunk://#diff-b7b57bc1e34145ef89c7929751d5d26aeecc8edfb37da9c60e9d3f0a1335133cR50-R65): Added checks for hipSPARSELt availability and updated test conditions to skip tests not supported on ROCm. [[1]](diffhunk://#diff-b7b57bc1e34145ef89c7929751d5d26aeecc8edfb37da9c60e9d3f0a1335133cR50-R65) [[2]](diffhunk://#diff-b7b57bc1e34145ef89c7929751d5d26aeecc8edfb37da9c60e9d3f0a1335133cR228) [[3]](diffhunk://#diff-b7b57bc1e34145ef89c7929751d5d26aeecc8edfb37da9c60e9d3f0a1335133cR239) [[4]](diffhunk://#diff-b7b57bc1e34145ef89c7929751d5d26aeecc8edfb37da9c60e9d3f0a1335133cR250) [[5]](diffhunk://#diff-b7b57bc1e34145ef89c7929751d5d26aeecc8edfb37da9c60e9d3f0a1335133cR579) [[6]](diffhunk://#diff-b7b57bc1e34145ef89c7929751d5d26aeecc8edfb37da9c60e9d3f0a1335133cR624) [[7]](diffhunk://#diff-b7b57bc1e34145ef89c7929751d5d26aeecc8edfb37da9c60e9d3f0a1335133cR661) [[8]](diffhunk://#diff-b7b57bc1e34145ef89c7929751d5d26aeecc8edfb37da9c60e9d3f0a1335133cR695) [[9]](diffhunk://#diff-b7b57bc1e34145ef89c7929751d5d26aeecc8edfb37da9c60e9d3f0a1335133cR730) [[10]](diffhunk://#diff-b7b57bc1e34145ef89c7929751d5d26aeecc8edfb37da9c60e9d3f0a1335133cR755) [[11]](diffhunk://#diff-b7b57bc1e34145ef89c7929751d5d26aeecc8edfb37da9c60e9d3f0a1335133cR771) [[12]](diffhunk://#diff-b7b57bc1e34145ef89c7929751d5d26aeecc8edfb37da9c60e9d3f0a1335133cR809) [[13]](diffhunk://#diff-b7b57bc1e34145ef89c7929751d5d26aeecc8edfb37da9c60e9d3f0a1335133cR844) [[14]](diffhunk://#diff-b7b57bc1e34145ef89c7929751d5d26aeecc8edfb37da9c60e9d3f0a1335133cL840-R854) [[15]](diffhunk://#diff-b7b57bc1e34145ef89c7929751d5d26aeecc8edfb37da9c60e9d3f0a1335133cR1005) Pull Request resolved: https://github.com/pytorch/pytorch/pull/150578 Approved by: https://github.com/jeffdaily
This commit is contained in:
committed by
PyTorch MergeBot
parent
ad26ec6abe
commit
43390d8b13
@ -290,6 +290,7 @@ header_template_rule(
|
||||
substitutions = {
|
||||
"@AT_CUDNN_ENABLED@": "1",
|
||||
"@AT_CUSPARSELT_ENABLED@": "0",
|
||||
"@AT_HIPSPARSELT_ENABLED@": "0",
|
||||
"@AT_ROCM_ENABLED@": "0",
|
||||
"@AT_MAGMA_ENABLED@": "0",
|
||||
"@NVCC_FLAGS_EXTRA@": "",
|
||||
|
Reference in New Issue
Block a user