@@ -748,36 +748,74 @@ static const std::string gpu_pipeline =
748
748
" func.func(convert-parallel-loops-to-gpu),"
749
749
// insert-gpu-allocs pass can have client-api = opencl or vulkan args
750
750
" func.func(insert-gpu-allocs{in-regions=1}),"
751
- // ** imex GPU passes
752
- // "drop-regions,"
753
- // "canonicalize,"
754
- // // "normalize-memrefs,"
755
- // // "gpu-decompose-memrefs,"
756
- // "func.func(lower-affine),"
757
- // "gpu-kernel-outlining,"
758
- // "canonicalize,"
759
- // "cse,"
760
- // // The following set-spirv-* passes can have client-api = opencl or
761
- // vulkan
762
- // // args
763
- // "set-spirv-capabilities{client-api=opencl},"
764
- // "gpu.module(set-spirv-abi-attrs{client-api=opencl}),"
765
- // "canonicalize,"
766
- // "fold-memref-alias-ops,"
767
- // "imex-convert-gpu-to-spirv{enable-vc-intrinsic=1},"
768
- // "spirv.module(spirv-lower-abi-attrs),"
769
- // "spirv.module(spirv-update-vce),"
770
- // // "func.func(llvm-request-c-wrappers),"
771
- // "serialize-spirv,"
772
- // "expand-strided-metadata,"
773
- // "lower-affine,"
774
- // "convert-gpu-to-gpux,"
775
- // "convert-func-to-llvm,"
776
- // "convert-math-to-llvm,"
777
- // "convert-gpux-to-llvm,"
778
- // "finalize-memref-to-llvm,"
779
- // "reconcile-unrealized-casts";
780
- // ** nv GPU passes
751
+ " drop-regions,"
752
+ " canonicalize,"
753
+ // "normalize-memrefs,"
754
+ // "gpu-decompose-memrefs,"
755
+ " func.func(lower-affine),"
756
+ " gpu-kernel-outlining,"
757
+ " canonicalize,"
758
+ " cse,"
759
+ // The following set-spirv-* passes can have client-api = opencl or vulkan
760
+ // args
761
+ " set-spirv-capabilities{client-api=opencl},"
762
+ " gpu.module(set-spirv-abi-attrs{client-api=opencl}),"
763
+ " canonicalize,"
764
+ " fold-memref-alias-ops,"
765
+ " imex-convert-gpu-to-spirv{enable-vc-intrinsic=1},"
766
+ " spirv.module(spirv-lower-abi-attrs),"
767
+ " spirv.module(spirv-update-vce),"
768
+ // "func.func(llvm-request-c-wrappers),"
769
+ " serialize-spirv,"
770
+ " expand-strided-metadata,"
771
+ " lower-affine,"
772
+ " convert-gpu-to-gpux,"
773
+ " convert-func-to-llvm,"
774
+ " convert-math-to-llvm,"
775
+ " convert-gpux-to-llvm,"
776
+ " finalize-memref-to-llvm,"
777
+ " reconcile-unrealized-casts" ;
778
+
779
+ static const std::string cuda_pipeline =
780
+ " add-gpu-regions,"
781
+ " canonicalize,"
782
+ " ndarray-dist,"
783
+ " func.func(dist-coalesce),"
784
+ " func.func(dist-infer-elementwise-cores),"
785
+ " convert-dist-to-standard,"
786
+ " canonicalize,"
787
+ " overlap-comm-and-compute,"
788
+ " add-comm-cache-keys,"
789
+ " lower-distruntime-to-idtr,"
790
+ " convert-ndarray-to-linalg,"
791
+ " canonicalize,"
792
+ " func.func(tosa-make-broadcastable),"
793
+ " func.func(tosa-to-linalg),"
794
+ " func.func(tosa-to-tensor),"
795
+ " canonicalize,"
796
+ " linalg-fuse-elementwise-ops,"
797
+ " arith-expand,"
798
+ " memref-expand,"
799
+ " arith-bufferize,"
800
+ " func-bufferize,"
801
+ " func.func(empty-tensor-to-alloc-tensor),"
802
+ " func.func(scf-bufferize),"
803
+ " func.func(tensor-bufferize),"
804
+ " func.func(bufferization-bufferize),"
805
+ " func.func(linalg-bufferize),"
806
+ " func.func(linalg-detensorize),"
807
+ " func.func(tensor-bufferize),"
808
+ " region-bufferize,"
809
+ " canonicalize,"
810
+ " func.func(finalizing-bufferize),"
811
+ " imex-remove-temporaries,"
812
+ " func.func(convert-linalg-to-parallel-loops),"
813
+ " func.func(scf-parallel-loop-fusion),"
814
+ // is add-outer-parallel-loop needed?
815
+ " func.func(imex-add-outer-parallel-loop),"
816
+ " func.func(gpu-map-parallel-loops),"
817
+ " func.func(convert-parallel-loops-to-gpu),"
818
+ " func.func(insert-gpu-allocs{in-regions=1}),"
781
819
" func.func(insert-gpu-copy),"
782
820
" drop-regions,"
783
821
" canonicalize,"
@@ -799,7 +837,9 @@ static const std::string gpu_pipeline =
799
837
800
838
const std::string _passes (get_text_env (" SHARPY_PASSES" ));
801
839
static const std::string &pass_pipeline =
802
- _passes != " " ? _passes : (useGPU () ? gpu_pipeline : cpu_pipeline);
840
+ _passes != " " ? _passes
841
+ : (useGPU () ? (useCUDA () ? cuda_pipeline : gpu_pipeline)
842
+ : cpu_pipeline);
803
843
804
844
JIT::JIT (const std::string &libidtr)
805
845
: _context (::mlir::MLIRContext::Threading::DISABLED), _pm (&_context),
@@ -851,23 +891,24 @@ JIT::JIT(const std::string &libidtr)
851
891
_crunnerlib = mlirRoot + " /lib/libmlir_c_runner_utils.so" ;
852
892
_runnerlib = mlirRoot + " /lib/libmlir_runner_utils.so" ;
853
893
if (!std::ifstream (_crunnerlib)) {
854
- throw std::runtime_error (" Cannot find libmlir_c_runner_utils.so " );
894
+ throw std::runtime_error (" Cannot find lib: " + _crunnerlib );
855
895
}
856
896
if (!std::ifstream (_runnerlib)) {
857
- throw std::runtime_error (" Cannot find libmlir_runner_utils.so " );
897
+ throw std::runtime_error (" Cannot find lib: " + _runnerlib );
858
898
}
859
899
860
900
if (useGPU ()) {
861
901
auto gpuxlibstr = get_text_env (" SHARPY_GPUX_SO" );
862
902
if (!gpuxlibstr.empty ()) {
863
903
_gpulib = std::string (gpuxlibstr);
864
904
} else {
865
- // auto imexRoot = get_text_env("IMEXROOT");
866
- // imexRoot = !imexRoot.empty() ? imexRoot : std::string(CMAKE_IMEX_ROOT);
867
- // _gpulib = imexRoot + "/lib/liblevel-zero-runtime.so";
868
- // _gpulib = imexRoot + "/lib/liblevel-zero-runtime.so";
869
- // for nv gpu
870
- _gpulib = mlirRoot + " /lib/libmlir_cuda_runtime.so" ;
905
+ if (useCUDA ()) {
906
+ _gpulib = mlirRoot + " /lib/libmlir_cuda_runtime.so" ;
907
+ } else {
908
+ auto imexRoot = get_text_env (" IMEXROOT" );
909
+ imexRoot = !imexRoot.empty () ? imexRoot : std::string (CMAKE_IMEX_ROOT);
910
+ _gpulib = imexRoot + " /lib/liblevel-zero-runtime.so" ;
911
+ }
871
912
if (!std::ifstream (_gpulib)) {
872
913
throw std::runtime_error (" Cannot find lib: " + _gpulib);
873
914
}
0 commit comments