From ef6120f32f41250984545c74b0417209aebcf349 Mon Sep 17 00:00:00 2001
From: Qi Li <qili93@qq.com>
Date: Thu, 3 Jun 2021 13:58:49 +0800
Subject: [PATCH] [ROCM] fix fused_fc_elementwise_layernorm, test=develop
 (#33281) (#33299)

---
 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 dde9531e591..5a86bb46e6a 100644
--- a/paddle/fluid/platform/cuda_device_function.h
+++ b/paddle/fluid/platform/cuda_device_function.h
@@ -32,6 +32,7 @@ namespace platform {
 #endif
 
 inline static int RoundToPowerOfTwo(int dim) {
+#ifdef PADDLE_WITH_CUDA
   if (dim > 512) {
     return 1024;
   } else if (dim > 256) {
@@ -45,6 +46,17 @@ 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