From 828fbd578436dbd528de9b9ea2e14da83bbac7b1 Mon Sep 17 00:00:00 2001
From: danleifeng <52735331+danleifeng@users.noreply.github.com>
Date: Thu, 24 Oct 2019 15:29:53 +0800
Subject: [PATCH] fix fp16 grid_size for size=1; test=release/1.6 (#20806)

---
 paddle/fluid/operators/elementwise/elementwise_add_op.cu | 5 +++--
 paddle/fluid/operators/elementwise/elementwise_div_op.cu | 5 +++--
 paddle/fluid/operators/elementwise/elementwise_mul_op.cu | 5 +++--
 paddle/fluid/operators/elementwise/elementwise_sub_op.cu | 5 +++--
 4 files changed, 12 insertions(+), 8 deletions(-)

diff --git a/paddle/fluid/operators/elementwise/elementwise_add_op.cu b/paddle/fluid/operators/elementwise/elementwise_add_op.cu
index de121b3786f..f83330fbe53 100644
--- a/paddle/fluid/operators/elementwise/elementwise_add_op.cu
+++ b/paddle/fluid/operators/elementwise/elementwise_add_op.cu
@@ -40,8 +40,9 @@ struct SameDimsElemwiseAdd<platform::CUDADeviceContext, platform::float16> {
                   const framework::Tensor* x, const framework::Tensor* y,
                   framework::Tensor* z) {
     auto size = x->numel();
-    dim3 gird_size = dim3(
-        (size / 2 + PADDLE_CUDA_THREAD_SIZE - 1) / PADDLE_CUDA_THREAD_SIZE, 1);
+    dim3 gird_size = dim3(((size + 1) / 2 + PADDLE_CUDA_THREAD_SIZE - 1) /
+                              PADDLE_CUDA_THREAD_SIZE,
+                          1);
     dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1);
     const half* x2 =
         reinterpret_cast<const half*>(x->data<platform::float16>());
diff --git a/paddle/fluid/operators/elementwise/elementwise_div_op.cu b/paddle/fluid/operators/elementwise/elementwise_div_op.cu
index b1698491180..787ff374228 100644
--- a/paddle/fluid/operators/elementwise/elementwise_div_op.cu
+++ b/paddle/fluid/operators/elementwise/elementwise_div_op.cu
@@ -41,8 +41,9 @@ struct SameDimsElemwiseDiv<platform::CUDADeviceContext, platform::float16> {
                   const framework::Tensor* x, const framework::Tensor* y,
                   framework::Tensor* z) {
     auto size = x->numel();
-    dim3 gird_size = dim3(
-        (size / 2 + PADDLE_CUDA_THREAD_SIZE - 1) / PADDLE_CUDA_THREAD_SIZE, 1);
+    dim3 gird_size = dim3(((size + 1) / 2 + PADDLE_CUDA_THREAD_SIZE - 1) /
+                              PADDLE_CUDA_THREAD_SIZE,
+                          1);
     dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1);
     const half* x2 =
         reinterpret_cast<const half*>(x->data<platform::float16>());
diff --git a/paddle/fluid/operators/elementwise/elementwise_mul_op.cu b/paddle/fluid/operators/elementwise/elementwise_mul_op.cu
index 4814cb144f0..efda1db2969 100644
--- a/paddle/fluid/operators/elementwise/elementwise_mul_op.cu
+++ b/paddle/fluid/operators/elementwise/elementwise_mul_op.cu
@@ -41,8 +41,9 @@ struct SameDimsElemwiseMul<platform::CUDADeviceContext, platform::float16> {
                   const framework::Tensor* x, const framework::Tensor* y,
                   framework::Tensor* z) {
     auto size = x->numel();
-    dim3 gird_size = dim3(
-        (size / 2 + PADDLE_CUDA_THREAD_SIZE - 1) / PADDLE_CUDA_THREAD_SIZE, 1);
+    dim3 gird_size = dim3(((size + 1) / 2 + PADDLE_CUDA_THREAD_SIZE - 1) /
+                              PADDLE_CUDA_THREAD_SIZE,
+                          1);
     dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1);
     const half* x2 =
         reinterpret_cast<const half*>(x->data<platform::float16>());
diff --git a/paddle/fluid/operators/elementwise/elementwise_sub_op.cu b/paddle/fluid/operators/elementwise/elementwise_sub_op.cu
index 7ff72028091..5df7abcf5a9 100644
--- a/paddle/fluid/operators/elementwise/elementwise_sub_op.cu
+++ b/paddle/fluid/operators/elementwise/elementwise_sub_op.cu
@@ -41,8 +41,9 @@ struct SameDimsElemwiseSub<platform::CUDADeviceContext, platform::float16> {
                   const framework::Tensor* x, const framework::Tensor* y,
                   framework::Tensor* z) {
     auto size = x->numel();
-    dim3 gird_size = dim3(
-        (size / 2 + PADDLE_CUDA_THREAD_SIZE - 1) / PADDLE_CUDA_THREAD_SIZE, 1);
+    dim3 gird_size = dim3(((size + 1) / 2 + PADDLE_CUDA_THREAD_SIZE - 1) /
+                              PADDLE_CUDA_THREAD_SIZE,
+                          1);
     dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1);
     const half* x2 =
         reinterpret_cast<const half*>(x->data<platform::float16>());
-- 
GitLab