[mlir] Add NVVM lowering for std.pow

[tpopp]

Add test case

[frgossen]

Thanks!
Can you also add this for ROCM?
Can be a separate CL ofc.

mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp

 
 void mlir::configureGpuToNVVMConversionLegality(ConversionTarget &target) {
   target.addIllegalOp<FuncOp>();
   target.addLegalDialect<::mlir::LLVM::LLVMDialect>();
   target.addLegalDialect<::mlir::NVVM::NVVMDialect>();
   target.addIllegalDialect<gpu::GPUDialect>();
   target.addIllegalOp<LLVM::CosOp, LLVM::ExpOp, LLVM::FAbsOp, LLVM::FCeilOp,
                       LLVM::FFloorOp, LLVM::LogOp, LLVM::Log10Op, LLVM::Log2Op,
-                      LLVM::SinOp, LLVM::SqrtOp>();
+                      LLVM::PowOp, LLVM::SinOp, LLVM::SqrtOp>();
 
   // TODO: Remove once we support replacing non-root ops.
   target.addLegalOp<gpu::YieldOp, gpu::GPUModuleOp, gpu::ModuleEndOp>();
 }
 
 void mlir::populateGpuToNVVMConversionPatterns(
     LLVMTypeConverter &converter, OwningRewritePatternList &patterns) {
   populateWithGenerated(converter.getDialect()->getContext(), patterns);
void mlir::populateGpuToNVVMConversionPatterns(
   patterns.insert<OpToFuncCallLowering<FloorFOp>>(converter, "__nv_floorf",
                                                   "__nv_floor");
   patterns.insert<OpToFuncCallLowering<LogOp>>(converter, "__nv_logf",
                                                "__nv_log");
   patterns.insert<OpToFuncCallLowering<Log10Op>>(converter, "__nv_log10f",
                                                  "__nv_log10");
   patterns.insert<OpToFuncCallLowering<Log2Op>>(converter, "__nv_log2f",
                                                 "__nv_log2");
+  patterns.insert<OpToFuncCallLowering<PowFOp>>(converter, "__nv_powf",
+                                                "__nv_pow");
   patterns.insert<OpToFuncCallLowering<RsqrtOp>>(converter, "__nv_rsqrtf",
                                                  "__nv_rsqrt");
   patterns.insert<OpToFuncCallLowering<SinOp>>(converter, "__nv_sinf",
                                                "__nv_sin");
   patterns.insert<OpToFuncCallLowering<SqrtOp>>(converter, "__nv_sqrtf",
                                                 "__nv_sqrt");
   patterns.insert<OpToFuncCallLowering<TanhOp>>(converter, "__nv_tanhf",
                                                 "__nv_tanh");
 }
 
 std::unique_ptr<OperationPass<gpu::GPUModuleOp>>
 mlir::createLowerGpuOpsToNVVMOpsPass(unsigned indexBitwidth) {
   return std::make_unique<LowerGpuOpsToNVVMOpsPass>(indexBitwidth);
 }

mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir

func @gpu_exp(%arg_f32 : f32, %arg_f64 : f64) -> (f32, f64) {
       %result64 = std.exp %arg_f64 : f64
       // CHECK: llvm.call @__nv_exp(%{{.*}}) : (!llvm.double) -> !llvm.double
       std.return %result32, %result64 : f32, f64
     }
     "test.finish" () : () -> ()
   }) : () -> ()
 }
 
+// -----
+
+gpu.module @test_module {
+  // CHECK: llvm.func @__nv_powf(!llvm.float) -> !llvm.float
+  // CHECK: llvm.func @__nv_pow(!llvm.double) -> !llvm.double
+  // CHECK-LABEL: func @gpu_pow
+  func @gpu_pow(%arg_f32 : f32, %arg_f64 : f64) -> (f32, f64) {
+    %result32 = std.pow %arg_f32, %arg_f32 : f32
+    // CHECK: llvm.call @__nv_powf(%{{.*}}, %{{.*}}) : (!llvm.float) -> !llvm.float
+    %result64 = std.pow %arg_f64, %arg_f64 : f64
+    // CHECK: llvm.call @__nv_pow(%{{.*}}, %{{.*}}) : (!llvm.double) -> !llvm.double
+    std.return %result32, %result64 : f32, f64
+  }
+}