I’m working with a Jetson TX1 and I’m learning NVVP Visual Profiler and its extended analysis and instruction level hints like global memory access (un-coalesced read/write).
I’m using the -lineinfo option and I’m able to collect timeline with kernel timings, perform latency/compute/memory analysis, and I can also see some information about execution count on a line-per-line basis.
One of the main things I do not understand is how to collect additional information that sometimes in a random way the profiler gives me, e.g. a warning on a line of my kernel where I’m writing a single byte to memory not exploiting coalescing.
I tried also collecting all metrics and events together by enabling them and the run again the profiler, and sometimes I was able to get the pie-chart with stall information, but nothing more (apart from GPU Details, but I’m looking for guided analysis hints).
More in detail, doing compute analysis from guided analysis and selecting instruction level profiling, it replays my executable and shows me another tab with my source code and a column of green/red bars. In the top menu bar I can see “hot spot - execution count”, but any other option in the tear-down menu is disabled (e.g. Global Access, Registers, …) even if sometimes happened to me to see register pressure and global access.
Can someone give me some hints on how to activate/switch-on additional instruction-level analysis or point me to a tutorial with a kernel that surely must produce some hints in the profiler and shows how to see warnings in the profiler?
Any help would be appreciated