[Mlir-commits] [mlir] [mlir][AMDGPU] Add support for AMD f16 math library calls (PR #108809)

[Daniel Hernandez-Juarez]

================
@@ -63,59 +63,61 @@ void mlir::populateMathToROCDLConversionPatterns(LLVMTypeConverter &converter,
   // Handled by mathToLLVM: math::SqrtOp
   // Handled by mathToLLVM: math::TruncOp
   populateOpPatterns<math::AcosOp>(converter, patterns, "__ocml_acos_f32",
-                                   "__ocml_acos_f64");
+                                   "__ocml_acos_f64", "__ocml_acos_f16");
   populateOpPatterns<math::AcoshOp>(converter, patterns, "__ocml_acosh_f32",
-                                    "__ocml_acosh_f64");
+                                    "__ocml_acosh_f64", "__ocml_acosh_f16");
   populateOpPatterns<math::AsinOp>(converter, patterns, "__ocml_asin_f32",
-                                   "__ocml_asin_f64");
+                                   "__ocml_asin_f64", "__ocml_asin_f16");
   populateOpPatterns<math::AsinhOp>(converter, patterns, "__ocml_asinh_f32",
-                                    "__ocml_asinh_f64");
+                                    "__ocml_asinh_f64", "__ocml_asinh_f16");
   populateOpPatterns<math::AtanOp>(converter, patterns, "__ocml_atan_f32",
-                                   "__ocml_atan_f64");
+                                   "__ocml_atan_f64", "__ocml_atan_f16");
   populateOpPatterns<math::AtanhOp>(converter, patterns, "__ocml_atanh_f32",
-                                    "__ocml_atanh_f64");
+                                    "__ocml_atanh_f64", "__ocml_atanh_f16");
   populateOpPatterns<math::Atan2Op>(converter, patterns, "__ocml_atan2_f32",
-                                    "__ocml_atan2_f64");
+                                    "__ocml_atan2_f64", "__ocml_atan2_f16");
   populateOpPatterns<math::CbrtOp>(converter, patterns, "__ocml_cbrt_f32",
-                                   "__ocml_cbrt_f64");
+                                   "__ocml_cbrt_f64", "__ocml_cbrt_f16");
   populateOpPatterns<math::CeilOp>(converter, patterns, "__ocml_ceil_f32",
-                                   "__ocml_ceil_f64");
+                                   "__ocml_ceil_f64", "__ocml_ceil_f16");
   populateOpPatterns<math::CosOp>(converter, patterns, "__ocml_cos_f32",
-                                  "__ocml_cos_f64");
+                                  "__ocml_cos_f64", "__ocml_cos_f16");
   populateOpPatterns<math::CoshOp>(converter, patterns, "__ocml_cosh_f32",
-                                   "__ocml_cosh_f64");
+                                   "__ocml_cosh_f64", "__ocml_cosh_f16");
   populateOpPatterns<math::SinhOp>(converter, patterns, "__ocml_sinh_f32",
-                                   "__ocml_sinh_f64");
-  populateOpPatterns<math::ExpOp>(converter, patterns, "", "__ocml_exp_f64");
+                                   "__ocml_sinh_f64", "__ocml_sinh_f16");
+  populateOpPatterns<math::ExpOp>(converter, patterns, "", "__ocml_exp_f64",
+                                  "__ocml_exp_f16");
----------------
dhernandez0 wrote:

Same thing as @krzysz00 comment regarding `__ocml_log_f16`, it seems `__ocml_exp_f16` is not just calling the intrinsic (while `__ocml_exp_f32` just calls its intrinsic). So, we might want to keep it.

https://github.com/llvm/llvm-project/pull/108809