Why shared memory bank conflict number is not equal?

Developer Tools Nsight Compute
1

I use nv-nsight-compute-cli to get shared memory bank conflict number, but i find its value is not equal when first run and the second run. like below

屏幕截图
屏幕截图1865×596 97.2 KB

why??

2

erfeng.jef,

Can you please provide the GPU, Nsight Compute version and specify whether this is the CUDA sample matrixMul or a different matrixMul.

The two counters listed can be collected in a single pass on Volta and Turing. If matrixMul is deterministic then the expectation is the counters should be equal.

3

matrixMul source code is here
matrixMul_7.txt (120.1 KB)

GPU is GTX 2080Ti
nsight compute cli version is

root@5e8d51ebe68d:/media/jefby/nvidia_gpu_inst_parser/cuda_template# nv-nsight-cu-cli --version
NVIDIA (R) Nsight Compute Command Line Profiler
Copyright (c) 2012-2019 NVIDIA Corporation
Version 2019.5.0 (Build 27346997)