Add rsqrt lowering from standard to ROCDL.

[akuegel]

Summary

Add a lowering for rsqrt from standard dialect to ROCDL.

[akuegel]

Friendly ping?

[frgossen]

Thanks!

mlir/lib/Conversion/GPUToROCDL/LowerGpuOpsToROCDLOps.cpp

void mlir::populateGpuToROCDLConversionPatterns(
   patterns.insert<OpToFuncCallLowering<FloorFOp>>(converter, "__ocml_floor_f32",
                                                   "__ocml_floor_f64");
   patterns.insert<OpToFuncCallLowering<LogOp>>(converter, "__ocml_log_f32",
                                                "__ocml_log_f64");
   patterns.insert<OpToFuncCallLowering<Log10Op>>(converter, "__ocml_log10_f32",
                                                  "__ocml_log10_f64");
   patterns.insert<OpToFuncCallLowering<Log2Op>>(converter, "__ocml_log2_f32",
                                                 "__ocml_log2_f64");
+  patterns.insert<OpToFuncCallLowering<RsqrtOp>>(converter, "__ocml_rsqrt_f32",
+                                                 "__ocml_rsqrt_f64");
   patterns.insert<OpToFuncCallLowering<SinOp>>(converter, "__ocml_sin_f32",
                                                "__ocml_sin_f64");
   patterns.insert<OpToFuncCallLowering<SqrtOp>>(converter, "__ocml_sqrt_f32",
                                                 "__ocml_sqrt_f64");
   patterns.insert<OpToFuncCallLowering<TanhOp>>(converter, "__ocml_tanh_f32",
                                                 "__ocml_tanh_f64");
 }
 
 std::unique_ptr<OperationPass<gpu::GPUModuleOp>>
 mlir::createLowerGpuOpsToROCDLOpsPass(unsigned indexBitwidth) {
   return std::make_unique<LowerGpuOpsToROCDLOpsPass>(indexBitwidth);
 }

mlir/test/Conversion/GPUToROCDL/gpu-to-rocdl.mlir

func @gpu_log2(%arg_f32 : f32, %arg_f64 : f64) -> (f32, f64) {
     // CHECK: llvm.call @__ocml_log2_f64(%{{.*}}) : (!llvm.double) -> !llvm.double
     std.return %result32, %result64 : f32, f64
   }
 }
 
 // -----
 
 gpu.module @test_module {
+  // CHECK: llvm.func @__ocml_rsqrt_f32(!llvm.float) -> !llvm.float
+  // CHECK: llvm.func @__ocml_rsqrt_f64(!llvm.double) -> !llvm.double
+  // CHECK-LABEL: func @gpu_rsqrt
+  func @gpu_rsqrt(%arg_f16 : f16, %arg_f32 : f32, %arg_f64 : f64)
+      -> (f16, f32, f64) {
+    %result16 = std.rsqrt %arg_f16 : f16
+    // CHECK: llvm.fpext %{{.*}} : !llvm.half to !llvm.float
+    // CHECK-NEXT: llvm.call @__ocml_rsqrt_f32(%{{.*}}) : (!llvm.float) -> !llvm.float
+    // CHECK-NEXT: llvm.fptrunc %{{.*}} : !llvm.float to !llvm.half
+    %result32 = std.rsqrt %arg_f32 : f32
+    // CHECK: llvm.call @__ocml_rsqrt_f32(%{{.*}}) : (!llvm.float) -> !llvm.float
+    %result64 = std.rsqrt %arg_f64 : f64
+    // CHECK: llvm.call @__ocml_rsqrt_f64(%{{.*}}) : (!llvm.double) -> !llvm.double
+    std.return %result16, %result32, %result64 : f16, f32, f64
+  }
+}
+
+// -----
+
+gpu.module @test_module {
   // CHECK: llvm.func @__ocml_sqrt_f32(!llvm.float) -> !llvm.float
   // CHECK: llvm.func @__ocml_sqrt_f64(!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 @__ocml_sqrt_f32(%{{.*}}) : (!llvm.float) -> !llvm.float