If your kernel is taking too much time, the track you should follow is trying to lower the kernel execution time. It will basically give you the first hint about what is taking too much time to complete.
Note: For peak performance, please refer to the matrixMulCUBLAS example. Performance= 35.35 GFlop/s, Time= 3.708 msec, Size= 131072000 Ops, WorkgroupSize= 1024 threads/blockĬhecking computed result for correctness: OK GPU Device 0: "GeForce GT 640M LE" with compute capability 3.0 =27694= NVPROF is profiling process 27694, command: matrixMul It will, by default, throw information about the API calls and how much the kernels consume.
#NVPROF CUDALAUNCH HOW TO#
You can learn how to use Nsight in the proper way and update this wiki later :D Provided that most of RidgeRun's work is on Tegra, we can focus on nvprof and, then, using Nsight.
#NVPROF CUDALAUNCH MANUAL#
You can find more information in this User manual for NVIDIA profiling tools for optimizing the performance of CUDA applications. Actually, Nsight is very recommended by NVIDIA for performing profiling. Please Contact RidgeRun OR email to if you have any questions.ĭepending on your setup, Nsight may be so useful since it integrates a user interface and guides the developer through the analysis process. Please come back soon to read the completed information on Ridgerun's support for this technology. RidgeRun CUDA Optimisation Guide RidgeRun documentation is currently under development.
0 kommentar(er)
