From dd5e1d1cf290f946eaaa99043474cf4087aad27f Mon Sep 17 00:00:00 2001
From: sasha0552 <admin@sasha0552.org>
Date: Fri, 10 May 2024 06:02:20 +0000
Subject: [PATCH] Fix F16 -> F32 upcasting to support Pascal GPUs

Based on code from openai/triton#2780 provided by @wkpark
---
 .../lib/TritonNVIDIAGPUToLLVM/ElementwiseOpToLLVM.cpp    | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/third_party/nvidia/lib/TritonNVIDIAGPUToLLVM/ElementwiseOpToLLVM.cpp b/third_party/nvidia/lib/TritonNVIDIAGPUToLLVM/ElementwiseOpToLLVM.cpp
index a87fc936d47b..e4ad66ceaca9 100644
--- 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");