OpenCL Kernel Optimization Tips
5 likes1,991 views
The document provides optimization tips for OpenCL kernels that focus on finding a balance between the device, toolchain, and problem being solved. It discusses considering the device type and memory characteristics, using profiling tools suited for the target platform, ensuring the problem is data parallel, and manually optimizing aspects like work group size and memory access patterns rather than relying on automatic features. Optimization requires understanding tradeoffs between these elements rather than taking a single-minded approach.
OpenCL Kernel Optimization Tips
- 1. OpenCL Kernel Optimization Tips Champ Yen (champ.yen@gmail.com) http://champyen.blogspot.com ver.20140820
- 2. Optimization - a form of balance Device/Platform Features Runtime Toolchain Problem Algorithm Optimization Optimization is not only greedy searching in single direction. It is more like to find a good balance point between device, toolchain and the problem.
- 3. Device - Computation ● device type ○ cpu - powerful single thread performance ○ gpu - many threads, great total throughput ● ISA design ○ scalar-based ○ vector-based ● # of compute unit/processing elements ● estimate impact of using divergence & barrier ● capability of asynchronous data transfer
- 4. Device - Memory ● get basic memory characteristics: ○ size ○ latency ○ throughput ○ coalescing effect ○ addressing mode ● global memory - unified or not ● local memory - real or not ● penalty of oversize
- 5. Toolchain/Runtime ● document/tutorial/guide for debugging, profiling and optimization. ● there is no perfect runtime/toolchain ● profiling/debugging tools. ● it is not always a good idea to debug/optimization on different platforms. ● automatic optimization MAY NOT HELP the thinking of optimization ● tricky forms of computation/memory operations. ○ MAD operations ○ memory access mode
- 6. Problem/Algorithms ● DATA PARALLEL! ● multi-stages is not always bad. ○ doing all things together uses more memory resource in one workitem. ● vectorized is not always a good idea ● use appropriate work group size ○ bad memory access pattern, less coalescing ○ may cause lower cache hit rate ○ less local memory for each workitem ○ may be less private memory for each workitem. ● different form of implementation ● do optimization things manually. ○ DO NOT relies on automatic features.
- 7. Q & A