[mlir][nvvm]: Add math::Exp2Op lowering to NVVM.

[akuegel]

Summary

There exists already a lowering to LLVM::Exp2Op.
On GPU we prefer to lower to an intrinsic instead.

[herhut]

Please use a [mlir][nvvm] tag in the commit message.

mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp

 
 } // anonymous namespace
 
 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::PowOp, LLVM::SinOp, LLVM::SqrtOp>();
+  target.addIllegalOp<LLVM::CosOp, LLVM::ExpOp, LLVM::Exp2Op, LLVM::FAbsOp,
+                      LLVM::FCeilOp, LLVM::FFloorOp, LLVM::LogOp, LLVM::Log10Op,
+                      LLVM::Log2Op, 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,
                                                RewritePatternSet &patterns) {
   populateWithGenerated(patterns);
void mlir::populateGpuToNVVMConversionPatterns(LLVMTypeConverter &converter,
   patterns.add<OpToFuncCallLowering<math::Atan2Op>>(converter, "__nv_atan2f",
                                                     "__nv_atan2");
   patterns.add<OpToFuncCallLowering<CeilFOp>>(converter, "__nv_ceilf",
                                               "__nv_ceil");
   patterns.add<OpToFuncCallLowering<math::CosOp>>(converter, "__nv_cosf",
                                                   "__nv_cos");
   patterns.add<OpToFuncCallLowering<math::ExpOp>>(converter, "__nv_expf",
                                                   "__nv_exp");
+  patterns.add<OpToFuncCallLowering<math::Exp2Op>>(converter, "__nv_exp2f",
+                                                   "__nv_exp2");
   patterns.add<OpToFuncCallLowering<math::ExpM1Op>>(converter, "__nv_expm1f",
                                                     "__nv_expm1");
   patterns.add<OpToFuncCallLowering<FloorFOp>>(converter, "__nv_floorf",
                                                "__nv_floor");
   patterns.add<OpToFuncCallLowering<math::LogOp>>(converter, "__nv_logf",
                                                   "__nv_log");
   patterns.add<OpToFuncCallLowering<math::Log1pOp>>(converter, "__nv_log1pf",
                                                     "__nv_log1p");

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

func @gpu_exp(%arg_f32 : f32, %arg_f64 : f64) -> (f32, f64) {
     // CHECK: llvm.call @__nv_expf(%{{.*}}) : (f32) -> f32
     %result64 = math.exp %arg_f64 : f64
     // CHECK: llvm.call @__nv_exp(%{{.*}}) : (f64) -> f64
     std.return %result32, %result64 : f32, f64
   }
 }
 
 // -----
+gpu.module @test_module {
+  // CHECK: llvm.func @__nv_exp2f(f32) -> f32
+  // CHECK: llvm.func @__nv_exp2(f64) -> f64
+  // CHECK-LABEL: func @gpu_exp2
+  func @gpu_exp2(%arg_f32 : f32, %arg_f64 : f64) -> (f32, f64) {
+    %result32 = math.exp2 %arg_f32 : f32
+    // CHECK: llvm.call @__nv_exp2f(%{{.*}}) : (f32) -> f32
+    %result64 = math.exp2 %arg_f64 : f64
+    // CHECK: llvm.call @__nv_exp2(%{{.*}}) : (f64) -> f64
+    std.return %result32, %result64 : f32, f64
+  }
+}
+
+// -----
 
 gpu.module @test_module {
   // CHECK: llvm.func @__nv_logf(f32) -> f32
   // CHECK: llvm.func @__nv_log(f64) -> f64
   // CHECK-LABEL: func @gpu_log
   func @gpu_log(%arg_f32 : f32, %arg_f64 : f64) -> (f32, f64) {
     %result32 = math.log %arg_f32 : f32
     // CHECK: llvm.call @__nv_logf(%{{.*}}) : (f32) -> f32