From 9171f73714f86f3675348b8140a3863a9fa38d94 Mon Sep 17 00:00:00 2001
From: danleifeng <52735331+danleifeng@users.noreply.github.com>
Date: Thu, 24 Oct 2019 19:50:05 +0800
Subject: [PATCH] fix fp16 grid_size for size=1; test=develop (#20812)

---
 .../fluid/operators/elementwise/elementwise_add_op.cu | 11 ++++++-----
 .../fluid/operators/elementwise/elementwise_div_op.cu | 11 ++++++-----
 .../fluid/operators/elementwise/elementwise_mul_op.cu | 11 ++++++-----
 .../fluid/operators/elementwise/elementwise_sub_op.cu | 11 ++++++-----
 4 files changed, 24 insertions(+), 20 deletions(-)

diff --git a/paddle/fluid/operators/elementwise/elementwise_add_op.cu b/paddle/fluid/operators/elementwise/elementwise_add_op.cu
index de121b3786..7101987280 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 grid_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>());
@@ -49,7 +50,7 @@ struct SameDimsElemwiseAdd<platform::CUDADeviceContext, platform::float16> {
         reinterpret_cast<const half*>(y->data<platform::float16>());
     half* z2 = reinterpret_cast<half*>(z->data<platform::float16>());
     SameDimsElemwiseAddCUDAKernel<<<
-        gird_size, block_size, 0,
+        grid_size, block_size, 0,
         ctx.template device_context<platform::CUDADeviceContext>().stream()>>>(
         x2, y2, z2, size);
   }
@@ -78,10 +79,10 @@ elementwise_add_grad(const framework::ExecutionContext& ctx,
                      framework::Tensor* dy) {
   dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1);
   auto size = x->numel();
-  dim3 gird_size =
+  dim3 grid_size =
       dim3((size + PADDLE_CUDA_THREAD_SIZE - 1) / PADDLE_CUDA_THREAD_SIZE, 1);
   SimpleElemwiseAddGradCUDAKernel<
-      T><<<gird_size, block_size, 0,
+      T><<<grid_size, block_size, 0,
            ctx.template device_context<plat::CUDADeviceContext>().stream()>>>(
       dout->data<T>(), size, dx->mutable_data<T>(ctx.GetPlace()),
       dy->mutable_data<T>(ctx.GetPlace()));
diff --git a/paddle/fluid/operators/elementwise/elementwise_div_op.cu b/paddle/fluid/operators/elementwise/elementwise_div_op.cu
index b169849118..e31722a288 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 grid_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>());
@@ -50,7 +51,7 @@ struct SameDimsElemwiseDiv<platform::CUDADeviceContext, platform::float16> {
         reinterpret_cast<const half*>(y->data<platform::float16>());
     half* z2 = reinterpret_cast<half*>(z->data<platform::float16>());
     SameDimsElemwiseDivCUDAKernel<<<
-        gird_size, block_size, 0,
+        grid_size, block_size, 0,
         ctx.template device_context<platform::CUDADeviceContext>().stream()>>>(
         x2, y2, z2, size);
   }
@@ -82,10 +83,10 @@ elementwise_div_grad(const framework::ExecutionContext& ctx,
                      framework::Tensor* dy) {
   dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1);
   auto size = x->numel();
-  dim3 gird_size =
+  dim3 grid_size =
       dim3((size + PADDLE_CUDA_THREAD_SIZE - 1) / PADDLE_CUDA_THREAD_SIZE, 1);
   SimpleElemwiseDivGradCUDAKernel<
-      T><<<gird_size, block_size, 0,
+      T><<<grid_size, block_size, 0,
            ctx.template device_context<plat::CUDADeviceContext>().stream()>>>(
       x->data<T>(), y->data<T>(), out->data<T>(), dout->data<T>(), size,
       dx->mutable_data<T>(ctx.GetPlace()), dy->mutable_data<T>(ctx.GetPlace()));
diff --git a/paddle/fluid/operators/elementwise/elementwise_mul_op.cu b/paddle/fluid/operators/elementwise/elementwise_mul_op.cu
index 4814cb144f..8533189f81 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 grid_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>());
@@ -50,7 +51,7 @@ struct SameDimsElemwiseMul<platform::CUDADeviceContext, platform::float16> {
         reinterpret_cast<const half*>(y->data<platform::float16>());
     half* z2 = reinterpret_cast<half*>(z->data<platform::float16>());
     SameDimsElemwiseMulCUDAKernel<<<
-        gird_size, block_size, 0,
+        grid_size, block_size, 0,
         ctx.template device_context<platform::CUDADeviceContext>().stream()>>>(
         x2, y2, z2, size);
   }
@@ -82,10 +83,10 @@ elementwise_mul_grad(const framework::ExecutionContext& ctx,
                      framework::Tensor* dy) {
   dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1);
   auto size = x->numel();
-  dim3 gird_size =
+  dim3 grid_size =
       dim3((size + PADDLE_CUDA_THREAD_SIZE - 1) / PADDLE_CUDA_THREAD_SIZE, 1);
   SimpleElemwiseMulGradCUDAKernel<
-      T><<<gird_size, block_size, 0,
+      T><<<grid_size, block_size, 0,
            ctx.template device_context<plat::CUDADeviceContext>().stream()>>>(
       x->data<T>(), y->data<T>(), out->data<T>(), dout->data<T>(), size,
       dx->mutable_data<T>(ctx.GetPlace()), dy->mutable_data<T>(ctx.GetPlace()));
diff --git a/paddle/fluid/operators/elementwise/elementwise_sub_op.cu b/paddle/fluid/operators/elementwise/elementwise_sub_op.cu
index 7ff7202809..9913927ee3 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 grid_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>());
@@ -50,7 +51,7 @@ struct SameDimsElemwiseSub<platform::CUDADeviceContext, platform::float16> {
         reinterpret_cast<const half*>(y->data<platform::float16>());
     half* z2 = reinterpret_cast<half*>(z->data<platform::float16>());
     SameDimsElemwiseSubCUDAKernel<<<
-        gird_size, block_size, 0,
+        grid_size, block_size, 0,
         ctx.template device_context<platform::CUDADeviceContext>().stream()>>>(
         x2, y2, z2, size);
   }
@@ -79,10 +80,10 @@ elementwise_sub_grad(const framework::ExecutionContext& ctx,
                      framework::Tensor* dy) {
   dim3 block_size = dim3(PADDLE_CUDA_THREAD_SIZE, 1);
   auto size = x->numel();
-  dim3 gird_size =
+  dim3 grid_size =
       dim3((size + PADDLE_CUDA_THREAD_SIZE - 1) / PADDLE_CUDA_THREAD_SIZE, 1);
   SimpleElemwiseSubGradCUDAKernel<
-      T><<<gird_size, block_size, 0,
+      T><<<grid_size, block_size, 0,
            ctx.template device_context<plat::CUDADeviceContext>().stream()>>>(
       dout->data<T>(), size, dx->mutable_data<T>(ctx.GetPlace()),
       dy->mutable_data<T>(ctx.GetPlace()));
-- 
GitLab