@@ -728,43 +728,46 @@ static const std::string gpu_pipeline =
728
728
" linalg-fuse-elementwise-ops,"
729
729
" arith-expand,"
730
730
" memref-expand,"
731
- " arith-bufferize,"
732
- " func-bufferize,"
733
731
" func.func(empty-tensor-to-alloc-tensor),"
734
- " func.func(scf-bufferize),"
735
- " func.func(tensor-bufferize),"
736
- " func.func(bufferization-bufferize),"
737
- " func.func(linalg-bufferize),"
738
- " func.func(linalg-detensorize),"
739
- " func.func(tensor-bufferize),"
732
+ " func.func(tile-for-gpu{tile-sizes=32 in-regions}),"
733
+ " func.func(tile-for-gpu{tile-sizes=1 in-regions}),"
740
734
" region-bufferize,"
741
735
" canonicalize,"
742
- " func.func(finalizing-bufferize),"
736
+ " one-shot-bufferize,"
737
+ " cse,"
738
+ " canonicalize,"
739
+ " scf-forall-to-parallel,"
740
+ " cse,"
741
+ " canonicalize,"
743
742
" imex-remove-temporaries,"
744
- " func.func(convert-linalg-to-parallel-loops),"
745
- " func.func(scf-parallel-loop-fusion),"
746
- // is add-outer-parallel-loop needed?
747
- " func.func(imex-add-outer-parallel-loop),"
743
+ " buffer-deallocation-pipeline,"
744
+ " func.func(convert-linalg-to-loops),"
748
745
" func.func(gpu-map-parallel-loops),"
749
- " func.func(convert-parallel-loops-to-gpu),"
746
+ " convert-parallel-loops-to-gpu,"
747
+ " canonicalize,"
748
+ " cse,"
750
749
" func.func(insert-gpu-allocs{in-regions=1}),"
751
750
" func.func(insert-gpu-copy),"
752
751
" drop-regions,"
753
752
" canonicalize,"
754
- " func.func(lower-affine),"
755
753
" gpu-kernel-outlining,"
754
+ " convert-scf-to-cf,"
755
+ " convert-cf-to-llvm,"
756
756
" canonicalize,"
757
757
" cse,"
758
- " gpu.module(strip-debuginfo,convert-gpu-to-nvvm),"
759
- " nvvm-attach-target,"
758
+ " gpu.module(strip-debuginfo,"
759
+ " convert-gpu-to-nvvm),"
760
+ " nvvm-attach-target{chip=sm_80 O=3},"
760
761
" func.func(gpu-async-region),"
761
762
" expand-strided-metadata,"
762
763
" lower-affine,"
763
764
" gpu-to-llvm,"
764
- " gpu-module-to-binary{format=fatbin},"
765
765
" convert-func-to-llvm,"
766
766
" convert-math-to-llvm,"
767
767
" finalize-memref-to-llvm,"
768
+ " canonicalize,"
769
+ " cse,"
770
+ " gpu-module-to-binary{format=fatbin},"
768
771
" reconcile-unrealized-casts" ;
769
772
770
773
const std::string _passes (get_text_env (" SHARPY_PASSES" ));
0 commit comments