Improve the transfers of DIGIs and clusters in Raw2Cluster

[makortel]

Digis

After #62 the DIGIS+errors are still always transferred back to CPU. In the pixel tracking workflow they are not used, so it should be straightforward do make their transfer on-demand. Basically this requires moving the transfers

cmssw/EventFilter/SiPixelRawToDigi/plugins/SiPixelRawToClusterGPUKernel.cu

Lines 645 to 655 in cdcd677

	cudaCheck(cudaMemcpyAsync(pdigi_h, pdigi_d, wordCounter*sizeof(uint32_t), cudaMemcpyDefault, stream.id()));
	cudaCheck(cudaMemcpyAsync(rawIdArr_h, rawIdArr_d, wordCounter*sizeof(uint32_t), cudaMemcpyDefault, stream.id()));
	if (includeErrors) {
	cudaCheck(cudaMemcpyAsync(error_h, error_d, vsize, cudaMemcpyDefault, stream.id()));
	cudaStreamSynchronize(stream.id());
	error_h->set_data(data_h);
	int size = error_h->size();
	cudaCheck(cudaMemcpyAsync(data_h, data_d, size*esize, cudaMemcpyDefault, stream.id()));
	}
	// End of Raw2Digi and passing data for cluserisation

from the kernel to the "transfer function"

cmssw/EventFilter/SiPixelRawToDigi/plugins/SiPixelRawToClusterHeterogeneous.cc

Lines 517 to 518 in cdcd677

	void SiPixelRawToClusterHeterogeneous::convertGPUtoCPU(const SiPixelRawToClusterHeterogeneous::GPUProduct& gpu,
	SiPixelRawToClusterHeterogeneous::CPUProduct& cpu) const {

On a longer term (going beyond the quadruplet pixel tracks), the error information will likely become useful, so one could think of splitting the DIGIs and the errors to two separate HeterogeneousProducts (with the cost of them having different product labels...).

Clusters

The clusters, on the other hand, are needed in the CPU for various reasons (RecHits need to make edm::Refs, MC matching relies on clusters etc). After #62 the CPU clusters are first put as part of HeterogeneousProduct, and then copied to a standard edmNew::DetSetVector<SiPixelCluster> by SiPixelClusterHeterogeneousConverter. Since the standard collection is needed anyway, it would be more optimal to produce it directly already by SiPixelRawToClusterHeterogeneous. (this will then lead to a funny situation of module siPixelDigis taking siPixelCluters as an input, but ok)

[cmsbot]

A new Issue was created by @makortel Matti Kortelainen.

can you please review it and eventually sign/assign? Thanks.

cms-bot commands are listed here

[makortel]

@felicepantaleo @VinInn

[makortel]

With the approach of #100 this issue would get fixed automatically.