[WIP] ECAL Reco on GPU v2 #366
Conversation
| char const* msg = cudaGetErrorString(result); | ||
| printf("\n%s, line %d:\ncudaCheckFromDevice(%s);\n%s: %s\n", | ||
| file, line, cmd, error, msg); | ||
| assert(0); |
There was a problem hiding this comment.
Just to note that we have experienced an assert() in kernel code (or in a function called by a kernel) leading to a device-side synchronization between CUDA streams, see #186 and #183. On the other hand I see I had noted in #183 (comment) that in case the condition can be checked at compile time the assert would not lead to syncrhonization.
There was a problem hiding this comment.
thx a lot Mattti! I did not know about this at all! just to note, i added this to provide at least some check on launching kernels from device side. Feel free to recommend a better check in here... this was copied from what you guys have in cudaCheck...
There was a problem hiding this comment.
I don't really have better ideas. But check with the profiler whether the kernels get run in parallel or not , the effect should be rather visible (see discussion and profile timelines in #183).
PR description:
Next iteration of ecal multifit. This is Work in Progress. The idea is
PR validation:
using the ecal validation exe