[alpaka] Refactor prefixScan implementation

[antoniopetre]

The prefixScan algorithm is implemented in Alpaka using two kernels, while a single kernel is used for Native CUDA.

I refactored the prefixScan implementation in order to use a single kernel (similar with the Native CUDA implementation).

[fwyzard]

Fixed conflicts and applied code formatting.

[fwyzard]

Rebased and fixed conflicts.

[fwyzard]

Rebased and fixed conflicts.

[makortel]

In general looks ok.

[makortel]

On Cori (with CUDA 11.2) I got the following failure when running ./alpaka --cuda

Processing 1000 events, of which 1 concurrently, with 1 threads.
terminate called after throwing an instance of 'std::runtime_error'
  what():  .../pixeltrack-standalone/external/alpaka/include/alpaka/mem/buf/BufUniformCudaHipRt.hpp(101) 'cudaFree(reinterpret_cast<void*>(memPtr))' returned error  : 'cudaErrorIllegalAddress': 'an illegal memory access was encountered'!

I'm really puzzled what BufUniformCudaHipRt is doing here (ok, maybe it is something that works with both CUDA and HIP). The master version runs fine.

[makortel]

Here is a stack trace of the exception

#0  __cxxabiv1::__cxa_throw (obj=obj@entry=0xb42e930, tinfo=0x2aaaac5909d0 <typeinfo for std::runtime_error>, dest=0x2aaaac2c3b90 <std::runtime_error::~runtime_error()>) at ../../.././libstdc++-v3/libsupc++/eh_throw.cc:80
#1  0x00002aaaab8f1ad8 in alpaka::uniform_cuda_hip::detail::rtCheck (line=<optimized out>, file=<optimized out>, desc=<optimized out>, error=<optimized out>) at .../pixeltrack-standalone/external/alpaka/include/alpaka/core/UniformCudaHip.hpp:67
#2  alpaka::uniform_cuda_hip::detail::rtCheckIgnore<>(cudaError const&, char const*, char const*, int const&) (error=<optimized out>, cmd=<optimized out>, file=<optimized out>, line=<optimized out>) at .../pixeltrack-standalone/external/alpaka/include/alpaka/core/UniformCudaHip.hpp:88
#3  0x00002aaab6cf3324 in alpaka::traits::CurrentThreadWaitFor<alpaka::uniform_cuda_hip::detail::QueueUniformCudaHipRtBase, void>::currentThreadWaitFor (queue=...) at /global/common/cori_cle7/software/sles15_cgpu/gcc/8.3.0/include/c++/8.3.0/bits/shared_ptr_base.h:1018
#4  alpaka::wait<alpaka::QueueUniformCudaHipRtNonBlocking> (awaited=...) at .../pixeltrack-standalone/external/alpaka/include/alpaka/wait/Traits.hpp:38
#5  alpaka_cuda_async::gpuVertexFinder::Producer::makeAsync (this=this@entry=0xbd5db8, tksoa=tksoa@entry=0x2aaae6000000, ptMin=<optimized out>, queue=...) at .../pixeltrack-standalone/src/alpaka/plugin-PixelVertexFinding/alpaka/gpuVertexFinder.cc:179
#6  0x00002aaab6cf952a in alpaka_cuda_async::PixelVertexProducerAlpaka::produce (this=0xbd5da8, iEvent=..., iSetup=...) at .../pixeltrack-standalone/src/alpaka/plugin-PixelVertexFinding/alpaka/PixelVertexProducerAlpaka.cc:53
#7  0x00002aaab6cfa1d4 in edm::EDProducer::doProduce (eventSetup=..., event=..., this=<optimized out>) at .../pixeltrack-standalone/src/alpaka/Framework/EDProducer.h:19
#8  edm::WorkerT<alpaka_cuda_async::PixelVertexProducerAlpaka>::doWorkAsync(edm::Event&, edm::EventSetup const&, edm::WaitingTask*)::{lambda(std::__exception_ptr::exception_ptr const*)#1}::operator()(std::__exception_ptr::exception_ptr const*) (iPtr=<optimized out>, this=<optimized out>)
    at .../pixeltrack-standalone/src/alpaka/Framework/Worker.h:69
#9  edm::FunctorWaitingTask<edm::WorkerT<alpaka_cuda_async::PixelVertexProducerAlpaka>::doWorkAsync(edm::Event&, edm::EventSetup const&, edm::WaitingTask*)::{lambda(std::__exception_ptr::exception_ptr const*)#1}>::execute() (this=0x2aaab7f3fd40) at .../pixeltrack-standalone/src/alpaka/Framework/WaitingTask.h:78
#10 0x00002aaaabd6d07d in tbb::internal::custom_scheduler<tbb::internal::IntelSchedulerTraits>::process_bypass_loop (this=this@entry=0x2aaab7f93e00, context_guard=..., t=t@entry=0x2aaab7f3fd40, isolation=isolation@entry=0) at ../../include/tbb/task.h:992
#11 0x00002aaaabd6d375 in tbb::internal::custom_scheduler<tbb::internal::IntelSchedulerTraits>::local_wait_for_all (this=0x2aaab7f93e00, parent=..., child=<optimized out>) at ../../include/tbb/task.h:992
#12 0x000000000041bf5b in tbb::task::wait_for_all (this=0x2aaab7f97d40) at .../pixeltrack-standalone/external/tbb/include/tbb/task.h:992
#13 edm::EventProcessor::runToCompletion (this=this@entry=0x7fffffff5960) at .../pixeltrack-standalone/src/alpaka/bin/EventProcessor.cc:37
#14 0x00000000004112ce in main (argc=<optimized out>, argv=<optimized out>) at .../pixeltrack-standalone/src/alpaka/bin/main.cc:176

[fwyzard]

Fixed conflicts, rebased, etc.

[fwyzard]

While the validation is good, now I see a small but systematic loss in performance.

Before:

$ CUDA_VISIBLE_DEVICES=0 numactl -N 0 ./alpaka --cuda --numberOfThreads 8 --numberOfStreams 16 --validation --maxEvents 10000; echo; for N in 1 2 3 4; do CUDA_VISIBLE_DEVICES=0 numactl -N 0 ./alpaka --cuda --numberOfThreads 8 --numberOfStreams 16 --maxEvents 10000; done
Processing 10000 events, of which 16 concurrently, with 8 threads.
CountValidator: all 10000 events passed validation
 Average relative track difference 0.000880287 (all within tolerance)
 Average absolute vertex difference 0.0007 (all within tolerance)
Processed 10000 events in 4.353466e+01 seconds, throughput 229.702 events/s.

Processing 10000 events, of which 16 concurrently, with 8 threads.
Processed 10000 events in 4.096583e+01 seconds, throughput 244.106 events/s.
Processed 10000 events in 4.049791e+01 seconds, throughput 246.926 events/s.
Processed 10000 events in 4.007989e+01 seconds, throughput 249.502 events/s.
Processed 10000 events in 4.102423e+01 seconds, throughput 243.758 events/s.

After:

$ CUDA_VISIBLE_DEVICES=0 numactl -N 0 ./alpaka --cuda --numberOfThreads 8 --numberOfStreams 16 --validation --maxEvents 10000; echo; for N in 1 2 3 4; do CUDA_VISIBLE_DEVICES=0 numactl -N 0 ./alpaka --cuda --numberOfThreads 8 --numberOfStreams 16 --maxEvents 10000; done
Processing 10000 events, of which 16 concurrently, with 8 threads.
CountValidator: all 10000 events passed validation
 Average relative track difference 0.00088813 (all within tolerance)
 Average absolute vertex difference 0.0004 (all within tolerance)
Processed 10000 events in 4.477171e+01 seconds, throughput 223.355 events/s.

Processing 10000 events, of which 16 concurrently, with 8 threads.
Processed 10000 events in 4.160250e+01 seconds, throughput 240.37 events/s.
Processed 10000 events in 4.151133e+01 seconds, throughput 240.898 events/s.
Processed 10000 events in 4.216559e+01 seconds, throughput 237.16 events/s.
Processed 10000 events in 4.186934e+01 seconds, throughput 238.838 events/s.

So 2-3% slower.