[mlir][Math] Add constant folder for Atan2Op.

[jacquesguan]

Summary

This patch adds constant folder for Atan2Op which only supports single and double precision floating-point.

[Mogball]

mlir/test/mlir-cpu-runner/math-polynomial-approx.mlir

455

This doesn't seem correct. I don't think the math atan2 is returning the correct results for undefined

[jacquesguan]

Fix undefined case.

[jacquesguan]

mlir/test/mlir-cpu-runner/math-polynomial-approx.mlir

455

Done.

mlir/include/mlir/Dialect/Math/IR/MathOps.td

let description = [{
 
     Example:
 
     ```mlir
     // Scalar variant.
     %a = math.atan2 %b, %c : f32
     ```
   }];
+  let hasFolder = 1;
 }
 
 //===----------------------------------------------------------------------===//
 // CeilOp
 //===----------------------------------------------------------------------===//
 
 def Math_CeilOp : Math_FloatUnaryOp<"ceil"> {
   let summary = "ceiling of the specified value";

mlir/lib/Dialect/Math/IR/MathOps.cpp

return constFoldUnaryOpConditional<FloatAttr>(
           return APFloat(atanf(a.convertToFloat()));
         default:
           return {};
         }
       });
 }
 
 //===----------------------------------------------------------------------===//
+// Atan2Op folder
+//===----------------------------------------------------------------------===//
+
+OpFoldResult math::Atan2Op::fold(ArrayRef<Attribute> operands) {
+  return constFoldBinaryOpConditional<FloatAttr>(
+      operands, [](const APFloat &a, const APFloat &b) -> Optional<APFloat> {
+        if (a.isZero() && b.isZero())
+          return llvm::APFloat::getNaN(a.getSemantics());
+
+        if (a.getSizeInBits(a.getSemantics()) == 64 &&
+            b.getSizeInBits(b.getSemantics()) == 64)
+          return APFloat(atan2(a.convertToDouble(), b.convertToDouble()));
+
+        if (a.getSizeInBits(a.getSemantics()) == 32 &&
+            b.getSizeInBits(b.getSemantics()) == 32)
+          return APFloat(atan2f(a.convertToFloat(), b.convertToFloat()));
+
+        return {};
+      });
+}
+
+//===----------------------------------------------------------------------===//
 // CeilOp folder
 //===----------------------------------------------------------------------===//
 
 OpFoldResult math::CeilOp::fold(ArrayRef<Attribute> operands) {
   return constFoldUnaryOp<FloatAttr>(operands, [](const APFloat &a) {
     APFloat result(a);
     result.roundToIntegral(llvm::RoundingMode::TowardPositive);
     return result;

mlir/test/Dialect/Math/canonicalize.mlir

 // CHECK-LABEL: @atan_fold_vec
 // CHECK-NEXT: %[[cst:.+]] = arith.constant dense<[0.000000e+00, 0.785398185, 0.000000e+00, 0.785398185]> : vector<4xf32>
 // CHECK-NEXT:   return %[[cst]]
 func.func @atan_fold_vec() -> (vector<4xf32>) {
   %v1 = arith.constant dense<[0.0, 1.0, 0.0, 1.0]> : vector<4xf32>
   %0 = math.atan %v1 : vector<4xf32>
   return %0 : vector<4xf32>
 }
+
+// CHECK-LABEL: @atan2_fold
+// CHECK-NEXT: %[[cst:.+]] = arith.constant 0.000000e+00 : f32
+// CHECK-NEXT:   return %[[cst]]
+func.func @atan2_fold() -> f32 {
+  %c1 = arith.constant 0.0 : f32
+  %c2 = arith.constant 1.0 : f32
+  %r = math.atan2 %c1, %c2 : f32
+  return %r : f32
+}
+
+// CHECK-LABEL: @atan2_fold_vec
+// CHECK-NEXT: %[[cst:.+]] = arith.constant dense<[0.000000e+00, 0.000000e+00, 0.463647604, 0.463647604]> : vector<4xf32>
+// CHECK-NEXT:   return %[[cst]]
+func.func @atan2_fold_vec() -> (vector<4xf32>) {
+  %v1 = arith.constant dense<[0.0, 0.0, 1.0, 1.0]> : vector<4xf32>
+  %v2 = arith.constant dense<[1.0, 1.0, 2.0, 2.0]> : vector<4xf32>
+  %0 = math.atan2 %v1, %v2 : vector<4xf32>
+  return %0 : vector<4xf32>
+}
+

mlir/test/mlir-cpu-runner/math-polynomial-approx.mlir

func.func @atan2() {
   // CHECK: 3.14159
   %atan2_2 = math.atan2 %zero, %neg_one : f32
   vector.print %atan2_2 : f32
 
   // CHECK: -1.5708
   %atan2_3 = math.atan2 %neg_one, %zero : f32
   vector.print %atan2_3 : f32
 
   // CHECK: nan
   %atan2_4 = math.atan2 %zero, %zero : f32
   vector.print %atan2_4 : f32
 
   // CHECK: 1.10715
   %atan2_5 = math.atan2 %two, %one : f32
   vector.print %atan2_5 : f32
 
   // CHECK: 2.03444
   %x6 = arith.constant -1.0 : f32
   %y6 = arith.constant 2.0 : f32
   %atan2_6 = math.atan2 %two, %neg_one : f32
   vector.print %atan2_6 : f32
 
   // CHECK: -2.03444
   %atan2_7 = math.atan2 %neg_two, %neg_one : f32
   vector.print %atan2_7 : f32
 
   // CHECK: -1.10715
   %atan2_8 = math.atan2 %neg_two, %one : f32
   vector.print %atan2_8 : f32
 
-  // CHECK: 0.463643
+  // CHECK: 0.463648
   %atan2_9 = math.atan2 %one, %two : f32
   vector.print %atan2_9 : f32
 
   // CHECK: 2.67795
   %x10 = arith.constant -2.0 : f32
   %y10 = arith.constant 1.0 : f32
   %atan2_10 = math.atan2 %one, %neg_two : f32
   vector.print %atan2_10 : f32
 
   // CHECK: -2.67795
   %x11 = arith.constant -2.0 : f32
   %y11 = arith.constant -1.0 : f32
   %atan2_11 = math.atan2 %neg_one, %neg_two : f32
   vector.print %atan2_11 : f32
 
-  // CHECK: -0.463643
+  // CHECK: -0.463648
   %atan2_12 = math.atan2 %neg_one, %two : f32
   vector.print %atan2_12 : f32
 
   return
 }
 
 
 func.func @main() {