From 3f366fee3312a7fc011f1110c117a718e6b47ca2 Mon Sep 17 00:00:00 2001
From: Qi Li <qili93@qq.com>
Date: Wed, 2 Jun 2021 19:21:00 +0800
Subject: [PATCH] [ROCM] fix fused_fc_elementwise_layernorm, test=develop
 (#33281)

---
 paddle/fluid/platform/cuda_device_function.h | 12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/paddle/fluid/platform/cuda_device_function.h b/paddle/fluid/platform/cuda_device_function.h
index 4095720f71e..35214330238 100644
--- a/paddle/fluid/platform/cuda_device_function.h
+++ b/paddle/fluid/platform/cuda_device_function.h
@@ -31,6 +31,7 @@ namespace platform {
 #endif
 
 inline static int RoundToPowerOfTwo(int dim) {
+#ifdef PADDLE_WITH_CUDA
   if (dim > 512) {
     return 1024;
   } else if (dim > 256) {
@@ -44,6 +45,17 @@ inline static int RoundToPowerOfTwo(int dim) {
   } else {
     return 32;
   }
+#else  // HIP results in error or nan if > 256
+  if (dim > 128) {
+    return 256;
+  } else if (dim > 64) {
+    return 128;
+  } else if (dim > 32) {
+    return 64;
+  } else {
+    return 32;
+  }
+#endif
 }
 
 #define CUDA_LAUNCH_KERNEL_BASE(dim, ...)  \
-- 
GitLab