Fix F16 -> F32 upcasting to support Pascal GPUs

Based on code from #2780 provided by @wkpark
triton-lang · May 10, 2024 · dd5e1d1 · dd5e1d1
1 parent 84b8e5f
commit dd5e1d1
Showing 1 changed file with 9 additions and 0 deletions.
diff --git a/third_party/nvidia/lib/TritonNVIDIAGPUToLLVM/ElementwiseOpToLLVM.cpp b/third_party/nvidia/lib/TritonNVIDIAGPUToLLVM/ElementwiseOpToLLVM.cpp
@@ -476,6 +476,15 @@ struct FpToFpOpConversion
       return outVals;
     }
 
+   if (srcElementType.isF16() && dstElementType.isF32()) {
+      SmallVector<Value> outVals;
+      for (Value v : operands[0]) {
+        outVals.push_back(
+            convertFp16ToFp32(loc, rewriter, v));
+      }
+      return outVals;
+    }
+
     if (srcElementType.isF32() && dstElementType.isBF16()) {
       assert(roundingMode.has_value() &&
              "rounding mode must be specified for fp32->bf16 conversion");