NVIDIA GPU Optimisation Tool CUDA Memcheck

RidgeRun CUDA Optimisation Guide
GPU Architecture
Execution process Memory hierarchy Communication & Concurrency
Optimisation Workflow
Tools CUDA Memcheck CUDA Profiler Computational Budget Tool
Optimisation Recipes
Types of optimisations Coarse optimisations Workload offloading Problem size Communication overlapping Correct memory access patterns Inter-thread communication Memory management Fine optimisations Increase arithmetic intensity Function approximation Condition and loops replacement Inlining
Common pitfalls when optimising
Examples
Empirical Experiments
Simple bounding test Multi-threaded bounding test
Contact Us

CUDA Memcheck

For debugging most of the issues, you can use cudamemcheck. You can refer to the CUDA-MEMCHECK NVIDIA documentation. The tool allows the developer to identify:

• memory access errors and memory leaks (memcheck)
• shared memory data access hazards (racecheck): remember that shared memory is "multi-threaded"
• uninitialised device global memory access (initcheck)
• thread synchronisation hazards (synccheck)

The usage is pretty simple:

cuda-memcheck [memcheck_options] app_name [app_options]

where some of the memcheck_options are:

• --show-backtrace
• --track-unused-memory
• --tool: memcheck (default), racecheck, initcheck, synccheck
• --report-api-errors
• --check-api-memory-access
• --print-level
• --save

It is better to read them from --help option.

So, what are they for in our optimisation?

Simple: it is better to have identified issues previously you touch the code. Thus, you won't doubt yourself.