Page MenuHomePhabricator

Please use GitHub pull requests for new patches. Phabricator shutdown timeline

Differential D91252 Diff 304490 mlir/lib/Conversion/GPUToROCDL/LowerGpuOpsToROCDLOps.cpp

Changeset View

Standalone View

View Options

mlir/lib/Conversion/GPUToROCDL/LowerGpuOpsToROCDLOps.cpp

Show First 20 Lines • Show All 66 Lines • ▼ Show 20 Lines void runOnOperation() override {
populateVectorToLLVMConversionPatterns(converter, llvmPatterns); populateVectorToLLVMConversionPatterns(converter, llvmPatterns);
populateVectorToROCDLConversionPatterns(converter, llvmPatterns); populateVectorToROCDLConversionPatterns(converter, llvmPatterns);
populateStdToLLVMConversionPatterns(converter, llvmPatterns); populateStdToLLVMConversionPatterns(converter, llvmPatterns);
populateGpuToROCDLConversionPatterns(converter, llvmPatterns); populateGpuToROCDLConversionPatterns(converter, llvmPatterns);
LLVMConversionTarget target(getContext()); LLVMConversionTarget target(getContext());
target.addIllegalDialect<gpu::GPUDialect>(); target.addIllegalDialect<gpu::GPUDialect>();
target.addIllegalOp<LLVM::CosOp, LLVM::ExpOp, LLVM::FAbsOp, LLVM::FCeilOp, target.addIllegalOp<LLVM::CosOp, LLVM::ExpOp, LLVM::FAbsOp, LLVM::FCeilOp,
LLVM::FFloorOp, LLVM::LogOp, LLVM::Log10Op, LLVM::FFloorOp, LLVM::LogOp, LLVM::Log10Op,
LLVM::Log2Op>(); LLVM::Log2Op, LLVM::SinOp, LLVM::TanhOp>();
Lint: Pre-merge checks
Inline Actions

clang-tidy: error: no member named 'TanhOp' in namespace 'mlir::LLVM' [clang-diagnostic-error]
not useful
clang-tidy: error: expected unqualified-id [clang-diagnostic-error]
not useful

Lint: Pre-merge checks: clang-tidy: error: no member named 'TanhOp' in namespace 'mlir::LLVM' [clang-diagnostic-error]…
herhutUnsubmitted
Not Done
ReplyInline Actions

Just wondering, why is there no TanOp here? Could you add this, as well?

herhut: Just wondering, why is there no `TanOp` here? Could you add this, as well?
akuegelAuthorUnsubmitted
Done
ReplyInline Actions

Yes, looks like this was an oversight. I have added it now as well.

akuegel: Yes, looks like this was an oversight. I have added it now as well.
target.addIllegalOp<FuncOp>(); target.addIllegalOp<FuncOp>();
target.addLegalDialect<ROCDL::ROCDLDialect>(); target.addLegalDialect<ROCDL::ROCDLDialect>();
// TODO: Remove once we support replacing non-root ops. // TODO: Remove once we support replacing non-root ops.
target.addLegalOp<gpu::YieldOp, gpu::GPUModuleOp, gpu::ModuleEndOp>(); target.addLegalOp<gpu::YieldOp, gpu::GPUModuleOp, gpu::ModuleEndOp>();
if (failed(applyPartialConversion(m, target, std::move(llvmPatterns)))) if (failed(applyPartialConversion(m, target, std::move(llvmPatterns))))
signalPassFailure(); signalPassFailure();
} }
}; };
Show All 24 Linesvoid mlir::populateGpuToROCDLConversionPatterns(
patterns.insert<OpToFuncCallLowering<FloorFOp>>(converter, "__ocml_floor_f32", patterns.insert<OpToFuncCallLowering<FloorFOp>>(converter, "__ocml_floor_f32",
"__ocml_floor_f64"); "__ocml_floor_f64");
patterns.insert<OpToFuncCallLowering<LogOp>>(converter, "__ocml_log_f32", patterns.insert<OpToFuncCallLowering<LogOp>>(converter, "__ocml_log_f32",
"__ocml_log_f64"); "__ocml_log_f64");
patterns.insert<OpToFuncCallLowering<Log10Op>>(converter, "__ocml_log10_f32", patterns.insert<OpToFuncCallLowering<Log10Op>>(converter, "__ocml_log10_f32",
"__ocml_log10_f64"); "__ocml_log10_f64");
patterns.insert<OpToFuncCallLowering<Log2Op>>(converter, "__ocml_log2_f32", patterns.insert<OpToFuncCallLowering<Log2Op>>(converter, "__ocml_log2_f32",
"__ocml_log2_f64"); "__ocml_log2_f64");
patterns.insert<OpToFuncCallLowering<SinOp>>(converter, "__ocml_sin_f32",
"__ocml_sin_f64");
patterns.insert<OpToFuncCallLowering<TanhOp>>(converter, "__ocml_tanh_f32", patterns.insert<OpToFuncCallLowering<TanhOp>>(converter, "__ocml_tanh_f32",
"__ocml_tanh_f64"); "__ocml_tanh_f64");
} }
herhutUnsubmitted
Done
ReplyInline Actions

Can you keep these sorted?

herhut: Can you keep these sorted?
std::unique_ptr<OperationPass<gpu::GPUModuleOp>> std::unique_ptr<OperationPass<gpu::GPUModuleOp>>
mlir::createLowerGpuOpsToROCDLOpsPass(unsigned indexBitwidth) { mlir::createLowerGpuOpsToROCDLOpsPass(unsigned indexBitwidth) {
return std::make_unique<LowerGpuOpsToROCDLOpsPass>(indexBitwidth); return std::make_unique<LowerGpuOpsToROCDLOpsPass>(indexBitwidth);
} }