Add sqrt lowering from standard to NVVM

Differential Revision: https://reviews.llvm.org/D92850

diff --git a/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp b/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
index 4b7bb6e..467c622 100644 (file)
--- a/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
+++ b/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
@@ -138,7 +138,7 @@ struct LowerGpuOpsToNVVMOpsPass
     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::Log2Op, LLVM::SinOp, LLVM::SqrtOp>();
     target.addIllegalOp<FuncOp>();
     target.addLegalDialect<NVVM::NVVMDialect>();
     // TODO: Remove once we support replacing non-root ops.
@@ -187,6 +187,8 @@ void mlir::populateGpuToNVVMConversionPatterns(
                                                  "__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");
 }

diff --git a/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir b/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir
index 347cc48..1c0257d 100644 (file)
--- a/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir
+++ b/mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir
@@ -315,6 +315,26 @@ gpu.module @test_module {
 
 // -----
 
+gpu.module @test_module {
+  // CHECK: llvm.func @__nv_sqrtf(!llvm.float) -> !llvm.float
+  // CHECK: llvm.func @__nv_sqrt(!llvm.double) -> !llvm.double
+  // CHECK-LABEL: func @gpu_sqrt
+  func @gpu_sqrt(%arg_f16 : f16, %arg_f32 : f32, %arg_f64 : f64)
+      -> (f16, f32, f64) {
+    %result16 = std.sqrt %arg_f16 : f16
+    // CHECK: llvm.fpext %{{.*}} : !llvm.half to !llvm.float
+    // CHECK-NEXT: llvm.call @__nv_sqrtf(%{{.*}}) : (!llvm.float) -> !llvm.float
+    // CHECK-NEXT: llvm.fptrunc %{{.*}} : !llvm.float to !llvm.half
+    %result32 = std.sqrt %arg_f32 : f32
+    // CHECK: llvm.call @__nv_sqrtf(%{{.*}}) : (!llvm.float) -> !llvm.float
+    %result64 = std.sqrt %arg_f64 : f64
+    // CHECK: llvm.call @__nv_sqrt(%{{.*}}) : (!llvm.double) -> !llvm.double
+    std.return %result16, %result32, %result64 : f16, f32, f64
+  }
+}
+
+// -----
+
 // Test that we handled properly operation with SymbolTable other than module op
 gpu.module @test_module {
   "test.symbol_scope"() ({