Skip to content

P
Paddle

机器未来 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
未验证 提交 7f00716c 编写于 作者: K kexinzhao 提交者： GitHub
浏览文件
操作

Add context wait in type_transform (#8850)

上级 6f50dee4
隐藏空白更改
内联 并排
Showing with 33 addition and 25 deletion
paddle/fluid/framework/data_type_transform.cc
浏览文件 @ 7f00716c
......@@ -53,6 +53,7 @@ struct CastDataType {
auto* context = static_cast<const platform::CUDADeviceContext*>(ctx_);
trans(*context, in_begin, in_end, out_begin,
CastDataTypeFunctor<InType, OutType>());
context->Wait();
#endif
} else {
PADDLE_THROW("Unsupported place!");
......
paddle/fluid/framework/data_type_transform_test.cc
浏览文件 @ 7f00716c
......@@ -50,13 +50,13 @@ TEST(DataTypeTransform, CPUTransform) {
TransDataType(kernel_fp32, kernel_fp64, in, &out);
double* out_data_double = out.data<double>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_double[i], static_cast<double>(i / 3));
EXPECT_EQ(out_data_double[i], static_cast<double>(i / 3));
}
TransDataType(kernel_fp32, kernel_int32, in, &out);
int* out_data_int = out.data<int>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_int[i], static_cast<int>(i / 3));
EXPECT_EQ(out_data_int[i], static_cast<int>(i / 3));
}
}
......@@ -76,31 +76,31 @@ TEST(DataTypeTransform, CPUTransform) {
TransDataType(kernel_fp16, kernel_fp32, in, &out);
float* out_data_float = out.data<float>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_float[i], static_cast<float>(ptr[i]));
EXPECT_EQ(out_data_float[i], static_cast<float>(ptr[i]));
}
TransDataType(kernel_fp16, kernel_fp64, in, &out);
double* out_data_double = out.data<double>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_double[i], static_cast<double>(ptr[i]));
EXPECT_EQ(out_data_double[i], static_cast<double>(ptr[i]));
}
TransDataType(kernel_fp16, kernel_int32, in, &out);
int* out_data_int = out.data<int>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_int[i], static_cast<int>(ptr[i]));
EXPECT_EQ(out_data_int[i], static_cast<int>(ptr[i]));
}
TransDataType(kernel_fp16, kernel_int64, in, &out);
int64_t* out_data_int64 = out.data<int64_t>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_int64[i], static_cast<int64_t>(ptr[i]));
EXPECT_EQ(out_data_int64[i], static_cast<int64_t>(ptr[i]));
}
TransDataType(kernel_fp16, kernel_bool, in, &out);
bool* out_data_bool = out.data<bool>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_bool[i], static_cast<bool>(ptr[i]));
EXPECT_EQ(out_data_bool[i], static_cast<bool>(ptr[i]));
}
// transform float to float16
......@@ -112,7 +112,7 @@ TEST(DataTypeTransform, CPUTransform) {
TransDataType(kernel_fp32, kernel_fp16, in, &out);
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_float[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_float[i]).x);
}
// transform double to float16
......@@ -124,7 +124,7 @@ TEST(DataTypeTransform, CPUTransform) {
TransDataType(kernel_fp64, kernel_fp16, in, &out);
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_double[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_double[i]).x);
}
// transform int to float16
......@@ -136,7 +136,7 @@ TEST(DataTypeTransform, CPUTransform) {
TransDataType(kernel_int32, kernel_fp16, in, &out);
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_int[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_int[i]).x);
}
// transform int64 to float16
......@@ -148,7 +148,7 @@ TEST(DataTypeTransform, CPUTransform) {
TransDataType(kernel_int64, kernel_fp16, in, &out);
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_int64[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_int64[i]).x);
}
// transform bool to float16
......@@ -160,7 +160,7 @@ TEST(DataTypeTransform, CPUTransform) {
TransDataType(kernel_bool, kernel_fp16, in, &out);
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_bool[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_bool[i]).x);
}
}
}
paddle/fluid/framework/data_type_transform_test.cu
浏览文件 @ 7f00716c
......@@ -49,15 +49,16 @@ TEST(DataTypeTransform, GPUTransform) {
float arr[6] = {0, 1, 2, 3, 4, 5};
int data_number = sizeof(arr) / sizeof(arr[0]);
memcpy(in_ptr, arr, sizeof(arr));
TensorCopy(in, gpu_place, context, &in_gpu);
TensorCopy(in, gpu_place, context, &in_gpu);
context.Wait();
TransDataType(kernel_fp32, kernel_fp64, in_gpu, &out_gpu);
TensorCopy(out_gpu, cpu_place, context, &out);
context.Wait();
double* out_data_double = out.data<double>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_double[i], static_cast<double>(arr[i]));
EXPECT_EQ(out_data_double[i], static_cast<double>(arr[i]));
}
TransDataType(kernel_fp32, kernel_int32, in_gpu, &out_gpu);
......@@ -66,7 +67,7 @@ TEST(DataTypeTransform, GPUTransform) {
int* out_data_int = out.data<int>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_int[i], static_cast<int>(arr[i]));
EXPECT_EQ(out_data_int[i], static_cast<int>(arr[i]));
}
}
......@@ -83,6 +84,7 @@ TEST(DataTypeTransform, GPUTransform) {
int data_number = sizeof(arr) / sizeof(arr[0]);
memcpy(ptr, arr, sizeof(arr));
TensorCopy(in, gpu_place, context, &in_gpu);
context.Wait();
// transform from float16 to other data types
TransDataType(kernel_fp16, kernel_fp32, in_gpu, &out_gpu);
......@@ -91,7 +93,7 @@ TEST(DataTypeTransform, GPUTransform) {
float* out_data_float = out.data<float>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_float[i], static_cast<float>(ptr[i]));
EXPECT_EQ(out_data_float[i], static_cast<float>(ptr[i]));
}
TransDataType(kernel_fp16, kernel_fp64, in_gpu, &out_gpu);
......@@ -100,7 +102,7 @@ TEST(DataTypeTransform, GPUTransform) {
double* out_data_double = out.data<double>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_double[i], static_cast<double>(ptr[i]));
EXPECT_EQ(out_data_double[i], static_cast<double>(ptr[i]));
}
TransDataType(kernel_fp16, kernel_int32, in_gpu, &out_gpu);
......@@ -109,7 +111,7 @@ TEST(DataTypeTransform, GPUTransform) {
int* out_data_int = out.data<int>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_int[i], static_cast<int>(ptr[i]));
EXPECT_EQ(out_data_int[i], static_cast<int>(ptr[i]));
}
TransDataType(kernel_fp16, kernel_int64, in_gpu, &out_gpu);
......@@ -118,7 +120,7 @@ TEST(DataTypeTransform, GPUTransform) {
int64_t* out_data_int64 = out.data<int64_t>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_int64[i], static_cast<int64_t>(ptr[i]));
EXPECT_EQ(out_data_int64[i], static_cast<int64_t>(ptr[i]));
}
TransDataType(kernel_fp16, kernel_bool, in_gpu, &out_gpu);
......@@ -127,7 +129,7 @@ TEST(DataTypeTransform, GPUTransform) {
bool* out_data_bool = out.data<bool>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(out_data_bool[i], static_cast<bool>(ptr[i]));
EXPECT_EQ(out_data_bool[i], static_cast<bool>(ptr[i]));
}
// transform float to float16
......@@ -137,13 +139,14 @@ TEST(DataTypeTransform, GPUTransform) {
}
TensorCopy(in, gpu_place, context, &in_gpu);
context.Wait();
TransDataType(kernel_fp32, kernel_fp16, in_gpu, &out_gpu);
TensorCopy(out_gpu, cpu_place, context, &out);
context.Wait();
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_float[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_float[i]).x);
}
// transform double to float16
......@@ -154,13 +157,14 @@ TEST(DataTypeTransform, GPUTransform) {
}
TensorCopy(in, gpu_place, context, &in_gpu);
context.Wait();
TransDataType(kernel_fp64, kernel_fp16, in_gpu, &out_gpu);
TensorCopy(out_gpu, cpu_place, context, &out);
context.Wait();
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_double[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_double[i]).x);
}
// transform int to float16
......@@ -170,13 +174,14 @@ TEST(DataTypeTransform, GPUTransform) {
}
TensorCopy(in, gpu_place, context, &in_gpu);
context.Wait();
TransDataType(kernel_int32, kernel_fp16, in_gpu, &out_gpu);
TensorCopy(out_gpu, cpu_place, context, &out);
context.Wait();
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_int[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_int[i]).x);
}
// transform int64 to float16
......@@ -187,13 +192,14 @@ TEST(DataTypeTransform, GPUTransform) {
}
TensorCopy(in, gpu_place, context, &in_gpu);
context.Wait();
TransDataType(kernel_int64, kernel_fp16, in_gpu, &out_gpu);
TensorCopy(out_gpu, cpu_place, context, &out);
context.Wait();
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_int64[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_int64[i]).x);
}
// transform bool to float16
......@@ -203,13 +209,14 @@ TEST(DataTypeTransform, GPUTransform) {
}
TensorCopy(in, gpu_place, context, &in_gpu);
context.Wait();
TransDataType(kernel_bool, kernel_fp16, in_gpu, &out_gpu);
TensorCopy(out_gpu, cpu_place, context, &out);
context.Wait();
ptr = out.data<float16>();
for (int i = 0; i < data_number; ++i) {
ASSERT_EQ(ptr[i].x, static_cast<float16>(in_data_bool[i]).x);
EXPECT_EQ(ptr[i].x, static_cast<float16>(in_data_bool[i]).x);
}
}
}
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或