enable operator gpu unittest #3050
Conversation
QiJune
force-pushed
the
op_gpu_test
branch
2 times, most recently
from
2aae619
to
4a1f7bd
Compare
QiJune
requested review from
reyoung,
wangkuiyi,
gangliao,
hedaoyuan and
jacquesqiao
| @@ -1,3 +1,4 @@ | |||
| #define EIGEN_USE_GPU | |||
There was a problem hiding this comment.
Have to add this macro, otherwise, will cause segment fault
| std::memcpy(dst, array.data(), sizeof(T) * array.size()); | ||
| } | ||
|
|
||
| #ifndef PADDLE_ONLY_CPU | ||
| template <typename T> | ||
| void PyCUDATensorSetFromArray( |
There was a problem hiding this comment.
partial specialization of function template is not allowed. So just define another function here
| places = [] | ||
| places.append(core.CPUPlace()) | ||
| if core.is_compile_gpu(): | ||
| places.append(core.GPUPlace(0)) |
There was a problem hiding this comment.
Run CPU OpKernel first, and then GPU OpKernel
| @@ -8,8 +8,8 @@ class TestAddOp(unittest.TestCase): | |||
|
|
|||
| def setUp(self): | |||
| self.type = "add_two" | |||
| self.X = numpy.random.random((342, 345)).astype("float32") | |||
| self.Y = numpy.random.random((342, 345)).astype("float32") | |||
| self.X = numpy.random.random((102, 105)).astype("float32") | |||
There was a problem hiding this comment.
Sometimes, GPU memory is not enough for unit-test, so reduce the size
| # TODO(qijun) gcc 4.9 or later versions raise SEGV due to the optimization problem. | ||
| # Use Debug mode instead for now. | ||
| if(CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 4.9 OR CMAKE_CXX_COMPILER_VERSION VERSION_EQUAL 4.9) | ||
| set(CMAKE_BUILD_TYPE "Debug" CACHE STRING "" FORCE) |
There was a problem hiding this comment.
If gcc 4.8 and nvcc 8.0, both debug and release will be fine.
If gcc 5.4 and nvcc 8.0, debug is fine, but release will cause segment fault.
There was a problem hiding this comment.
How about we constraint GCC version in Dockerfile, other than checking here?
I know checking in CMake provides a guarantee, but I am afraid adding too many such constraints would complicate our building system too much.
There was a problem hiding this comment.
Yes, I will constraint GCC version in Dockerfile.
But some others may compile paddle in gcc5.4 environment, so, the unittest test_framework will fall.
| @@ -62,9 +61,11 @@ inline T* Tensor::mutable_data(platform::Place place) { | |||
| if (platform::is_cpu_place(place)) { | |||
| holder_.reset(new PlaceholderImpl<T, platform::CPUPlace>( | |||
| boost::get<platform::CPUPlace>(place), size)); | |||
| } else if (platform::is_gpu_place(place)) { | |||
| #ifdef PADDLE_ONLY_CPU | |||
| PADDLE_THROW("'GPUPlace' is not supported in CPU only device."); | |||
| } | |||
There was a problem hiding this comment.
{ is outside of macro, so } is needed inside macro
No description provided.