Hi I am trying to port the project to the ROCm platform(AMD GPU), and found there is i

And in defien_auto_correlate.h, there are two functions defined <code class="notransla

Try to port the project to ROCm about oskar HOT 3 CLOSED

coversb commented on August 19, 2024

Try to port the project to ROCm

from oskar.

Comments (3)

coversb commented on August 19, 2024

Hello, now I almost finished the porting. But in ctest part, there are some FAILED items.
I saw the log, and I think it's issue about 'Single precision GPU test'. Do you have any suggestion to debug this?
Thanks!

[ RUN      ] auto_correlate.matrix_singleGPU_doubleGPU
/public/home/chenhao/workspace/oskar/oskar/oskar/correlate/test/Test_auto_correlate.cpp:27: Failure
Expected: (max_rel_error) < (tol), actual: 0.959102 vs 0.002
RELATIVE ERROR MIN: 0 MAX: 0.9591 AVG: 0.07807 STD: 0.19071
/public/home/chenhao/workspace/oskar/oskar/oskar/correlate/test/Test_auto_correlate.cpp:33: Failure
Expected: (avg_rel_error) < (tol), actual: 0.0780705 vs 1e-05
RELATIVE ERROR MIN: 0 MAX: 0.9591 AVG: 0.07807 STD: 0.19071
   > Matrix.
     Single precision GPU: 2.27 ms, Double precision GPU: 0.04 ms
[  FAILED  ] auto_correlate.matrix_singleGPU_doubleGPU (775 ms)

from oskar.

coversb commented on August 19, 2024

And in defien_auto_correlate.h, there are two functions defined OSKAR_ACORR_GPU and OSKAR_ACORR_CPU, I found in the code, it uses OSKAR_ACORR_CPU, does they all work fine ? Maybe just the speed to run is different?
Thanks~

from oskar.

coversb commented on August 19, 2024

Almost finished the ROCm porting, and pass all unit test, close it now.