Closed esseivaju closed 2 years ago
Changes listed in the original comment have been implemented. The tests below are still failing:
HistoContainer_t
OneToManyAssoc_t
prefixScan_t
radixSort_t
Fixing those will require updating many Kernels so it should probably be done in a future PR as this one is getting fairly large. Some of these tests (e.g. prefixScan_t
/ prefixScan.h
) are still pretty much identical to the original cudauvm
version, the only change being update to #define
directives to compile the correct code with nvc++
so it's suggesting some issues with nvc++
it should probably be done in a future PR as this one is getting fairly large.
I agree, this starts to be a good chunk to integrate. Visually looks good, could you just confirm that you agree with merging now?
Yes we can merge now I will fix the remaining issues with HistoContainer, prefixScan and radixSort in the next pull request.
As discussed in #368 this PR contains an updated to cudart api utilization to get closer to standard c++ and no cuda dependencies.
Changes
CUDACore
which are no longer necessaryEDProducer
instead ofEDProducerExternalWork
Current issues to be fixed
cudaMalloc
/cudaMemcpy
, this should be updated to usestd::unique_ptr
SiPixelRawToClusterCUDA
module). I updated thegpuClustering_t
test to work withstd::unique_ptr
and it fails with the following output: