Add GPU compute capability check for float16 math function test #8946
Conversation
| return count; | ||
| } | ||
|
|
||
| int GetCUDAComputeCapability(int id) { |
There was a problem hiding this comment.
Since this is the cudaDeviceProp of a GPU card, could we get this in init.cc file? Thus, we can cache this version number other than call it in every float16 method.
There was a problem hiding this comment.
Good point! Done.
| using namespace paddle::platform; | ||
|
|
||
| // fp16 GEMM in cublas requires GPU compute capability >= 53 | ||
| if (GetCUDAComputeCapability(0) < 53) { |
There was a problem hiding this comment.
I am not sure if this GetCUDAComputeCapability(0) < 53 code should be moved into math_function.cu?
There was a problem hiding this comment.
Adding processing code for compute capability < 53 case could be a future TODO item.
Before that is done, I think we need code like this to skip the fp16 test if necessary, otherwise one of the CI machine with old GPUs will fail the test.
I have also added PADDLE_ENFORCE_GE(compute_cap, 53) in math_function.cu to give a more specific error message for users.
kexinzhao
left a comment
kexinzhao
left a comment
There was a problem hiding this comment.
Thank you for the review!
| return count; | ||
| } | ||
|
|
||
| int GetCUDAComputeCapability(int id) { |
There was a problem hiding this comment.
Good point! Done.
| using namespace paddle::platform; | ||
|
|
||
| // fp16 GEMM in cublas requires GPU compute capability >= 53 | ||
| if (GetCUDAComputeCapability(0) < 53) { |
There was a problem hiding this comment.
Adding processing code for compute capability < 53 case could be a future TODO item.
Before that is done, I think we need code like this to skip the fp16 test if necessary, otherwise one of the CI machine with old GPUs will fail the test.
I have also added PADDLE_ENFORCE_GE(compute_cap, 53) in math_function.cu to give a more specific error message for users.
2 participants
fix #8947