From 6fa452174affe3e0c741679e6120bdce3daf9802 Mon Sep 17 00:00:00 2001
From: Alejandro Acosta <alejandro.acosta@codeplay.com>
Date: Fri, 29 Nov 2024 14:08:27 +0000
Subject: [PATCH] Fix __dp4a call

---
 include/cutlass/numeric_conversion.h | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/include/cutlass/numeric_conversion.h b/include/cutlass/numeric_conversion.h
index 664945d35..5b9fcf57b 100644
--- a/include/cutlass/numeric_conversion.h
+++ b/include/cutlass/numeric_conversion.h
@@ -3664,7 +3664,11 @@ struct NumericArrayConverter<float, int8_t, N, Round> {
 
     CUTLASS_PRAGMA_UNROLL
     for (int ii = 0; ii < PackedResultType::kElements; ++ii) {
+#if defined(__CUDA_ARCH__)
       t[ii] = __dp4a(x, mask[ii], 0);
+#else
+      t[ii] = x * mask[ii];
+#endif
       r[ii] = static_cast<float>(t[ii]);
     }
   #endif