9.5 Reducing the effect of serial computations
You can reduce the impact of serial components in your application by reducing and optimizing the computations.
- Use memory mapping instead of memory copies to transfer data.
- Optimize communication code.
- To reduce latency, optimize the communication code that sends and receives data.
- Keep messages small.
- Reduce communication overhead by sending only the data that is required.
- Use power of two sized memory blocks for communication.
- Ensure the sizes of memory blocks used for communication are a power of two. This makes
the data more cacheable.
- Send more data in a smaller number of transfers.
- Compute values instead of reading them from memory.
- A simple computation is likely to be faster than reading from memory.
- Do serial computations on the application processors.
- Application processors are optimized for low latency tasks.
- Use clEnqueueFillBuffer() to fill buffers.
- The Mali™ OpenCL driver contains an optimized implementation of
clEnqueueFillBuffer(). Use in place of manually implementing a buffer fill in your application.
- Use clEnqueueFillImage() to fill images.
- The Mali OpenCL driver contains an optimized implementation of
clEnqueueFillImage(). Use this in place of manually implementing an image fill in your application.
PDF version