[SYCL-MLIR] Support _Float16 type (#7267)

Emit warning when using this type in an unsupported target, as promotion is not yet implemented.

Signed-off-by: Victor Perez <[email protected]>

Author: victor-eds

polygeist/tools/cgeist/Lib/CodeGenTypes.cc

@@ -1586,6 +1586,14 @@ mlir::Type CodeGenTypes::getMLIRType(clang::QualType QT, bool *implicitRef,
       return builder.getF80Type();
     if (T->isFP128Ty())
       return builder.getF128Type();
+    if (T->is16bitFPTy()) {
+      if (CGM.getTarget().shouldEmitFloat16WithExcessPrecision()) {
+        llvm::WithColor::warning()
+            << "Experimental usage of _Float16. Code generated will be illegal "
+               "for this target. Use with caution.\n";
+      }
+      return builder.getF16Type();
+    }
 
     if (auto IT = dyn_cast<llvm::IntegerType>(T)) {
       return builder.getIntegerType(IT->getBitWidth());

polygeist/tools/cgeist/Test/Verification/float16.c

@@ -0,0 +1,63 @@
+// RUN: cgeist %s --function=* -S -march=x86-64         2>&1 | FileCheck %s --check-prefix=CHECK-EXTEND
+// RUN: cgeist %s --function=* -S -march=sapphirerapids 2>&1 | FileCheck %s --check-prefix=CHECK-NATIVE
+
+// COM: sapphirerapids supports _Float16 natively
+
+// CHECK-EXTEND: warning: Experimental usage of _Float16.
+// CHECK-NATIVE-NOT: warning: Experimental usage of _Float16.
+
+// CHECK-EXTEND-LABEL:  func.func @type(%arg0: f16) -> f16
+// CHECK-EXTEND-NEXT:     return %arg0 : f16
+// CHECK-EXTEND-NEXT:   }
+
+// CHECK-NATIVE-LABEL:  func.func @type(%arg0: f16) -> f16
+// CHECK-NATIVE-NEXT:     return %arg0 : f16
+// CHECK-NATIVE-NEXT:   }
+
+_Float16 type(_Float16 arg) {
+  return arg;
+}
+
+// CHECK-EXTEND-LABEL:  func.func @arith(%arg0: f16, %arg1: f16, %arg2: f16, %arg3: f16, %arg4: f16) -> f16
+// CHECK-EXTEND-NEXT:     %[[ADD:.*]] = arith.addf %arg0, %arg1 : f16
+// CHECK-EXTEND-NEXT:     %[[NEG:.*]] = arith.negf %arg3 : f16
+// CHECK-EXTEND-NEXT:     %[[MUL:.*]] = arith.mulf %arg2, %[[NEG]] : f16
+// CHECK-EXTEND-NEXT:     %[[DIV:.*]] = arith.divf %[[MUL]], %arg4 : f16
+// CHECK-EXTEND-NEXT:     %[[SUB:.*]] = arith.subf %[[ADD]], %[[DIV]] : f16
+// CHECK-EXTEND-NEXT:     return %[[SUB]] : f16
+// CHECK-EXTEND-NEXT:   }
+
+// CHECK-NATIVE-LABEL:  func.func @arith(%arg0: f16, %arg1: f16, %arg2: f16, %arg3: f16, %arg4: f16) -> f16
+// CHECK-NATIVE-NEXT:     %[[ADD:.*]] = arith.addf %arg0, %arg1 : f16
+// CHECK-NATIVE-NEXT:     %[[NEG:.*]] = arith.negf %arg3 : f16
+// CHECK-NATIVE-NEXT:     %[[MUL:.*]] = arith.mulf %arg2, %[[NEG]] : f16
+// CHECK-NATIVE-NEXT:     %[[DIV:.*]] = arith.divf %[[MUL]], %arg4 : f16
+// CHECK-NATIVE-NEXT:     %[[SUB:.*]] = arith.subf %[[ADD]], %[[DIV]] : f16
+// CHECK-NATIVE-NEXT:     return %[[SUB]] : f16
+// CHECK-NATIVE-NEXT:   }
+
+_Float16 arith(_Float16 a,
+	       _Float16 b,
+	       _Float16 c,
+	       _Float16 d,
+	       _Float16 e) {
+  return (+a) + b - c * (-d) / e;
+}
+
+// CHECK-EXTEND-LABEL:  func.func @compound_assign(%arg0: memref<?xf16>, %arg1: f16)
+// CHECK-EXTEND-NEXT:     %[[ORIG:.*]] = affine.load %arg0[0] : memref<?xf16>
+// CHECK-EXTEND-NEXT:     %[[ADD:.*]] = arith.addf %[[ORIG]], %arg1 : f16
+// CHECK-EXTEND-NEXT:     affine.store %[[ADD]], %arg0[0] : memref<?xf16>
+// CHECK-EXTEND-NEXT:     return
+// CHECK-EXTEND-NEXT:   }
+
+// CHECK-NATIVE-LABEL:  func.func @compound_assign(%arg0: memref<?xf16>, %arg1: f16)
+// CHECK-NATIVE-NEXT:     %[[ORIG:.*]] = affine.load %arg0[0] : memref<?xf16>
+// CHECK-NATIVE-NEXT:     %[[ADD:.*]] = arith.addf %[[ORIG]], %arg1 : f16
+// CHECK-NATIVE-NEXT:     affine.store %[[ADD]], %arg0[0] : memref<?xf16>
+// CHECK-NATIVE-NEXT:     return
+// CHECK-NATIVE-NEXT:   }
+
+void compound_assign(_Float16 *a, _Float16 b) {
+  *a += b;
+}