lupesko commented on a change in pull request #11780: fix flaky test
test_operator_gpu.test_countsketch
URL: https://github.com/apache/incubator-mxnet/pull/11780#discussion_r203629802
##########
File path: src/operator/contrib/count_sketch.cu
##########
@@ -165,7 +165,8 @@ inline void CountSketchBackward(const Tensor<gpu, 2,
DType> &in_grad,
nthreads, in_grad_ptr+bstart*in_dim, h_ptr,
s_ptr, out_grad_ptr+bstart*out_dim, batchlen,
in_dim, out_dim);
- MSHADOW_CUDA_POST_KERNEL_CHECK(sketch_backward_kernel);
+ cudaError_t err = cudaDeviceSynchronize();
Review comment:
DRY is not only about lines of code. Consider a case where there is an issue
with the flow in line 168, and the developer decides to change it from
cudaDeviceSynchronize to another API. Without DRY, the developer will need to
remember to look for other places in the code that call the API. Sometimes,
this will be missed. With DRY, different parts of the code are funneled to the
one function or whatever that has this logic, and by updating in one location
all code paths are updated. Makes sense?
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
