CUDA : How to detect shared memory bank conflict on device with compute capabiliy >= 7.2?

Question

On device with compute capability <= 7.2 , I always use

nvprof --events shared_st_bank_conflict

but when i run it on RTX2080ti with CUDA10 , it returns

Warning: Skipping profiling on device 0 since profiling is not supported on devices with compute capability greater than 7.2

So how can i detect whether there's share memory bank conflict on this devices ?

I've installed Nvidia Nsight Systems and Nsight Compute , find no such profiling report...

thks

you should use nsight compute – Robert Crovella Mar 22 '19 at 20:24 — Robert Crovella, Mar 22 '19 at 20:24

score 2 · Answer 1 · edited Mar 10 '21 at 11:03

2

You can use --metrics:

Either

nv-nsight-cu-cli --metrics l1tex__data_bank_conflicts_pipe_lsu_mem_shared_op_ld.sum

for conflicts when reading (load'ing) from shared memory, or

nv-nsight-cu-cli --metrics l1tex__data_bank_conflicts_pipe_lsu_mem_shared_op_st.sum

for conflicting when writing (store'ing) to shared memory.

edited Mar 10 '21 at 11:03

einpoklum

118,144
57
340
684

answered Feb 16 '20 at 22:09

Deepak C U

21
4

score 1 · Answer 2 · answered Mar 22 '19 at 18:53

1

It seems this is a problem, and is addressed in this post to the NVIDIA forums. Apparently it should be supported using one of the Nsight tools (either the CLI or UI).

answered Mar 22 '19 at 18:53

CUDA : How to detect shared memory bank conflict on device with compute capabiliy >= 7.2?

2 Answers2