Hierarchy from top to bottom:
- Host: the entire system
- Device group: multiple devices, e.g. one GPU and one CPU.
- Compute device
- Compute unit
- Processing element
- Work group
- Work item
GPU vs CPU hardware level.
https://youtu.be/e-2bTxKuS2U?list=PLTfYiv7-a3l7mYEdjk35wfY-KQj5yVXO2&t=319 mentions GPU has no cache.
TODO what is a platform?
http://stackoverflow.com/questions/3444664/does-any-opencl-host-have-more-than-one-platform
One CPU, one GPU, etc.
TODO vs core?
Can be obtained with: clGetDeviceInfo(CL_DEVICE_MAX_COMPUTE_UNITS)
TODO
Contains many work items.
Is basically a completely independent chunk of work.
Work items inside the same work group can share local memory, and can synchronize.
So ideally, we would like to have a single work group, with infinitely many work items inside it.
However, the laws of physics are cruel, and the following limits exist:
CL_DEVICE_MAX_WORK_GROUP_SIZE: maximum number of work items inside the work groupCL_DEVICE_MAX_WORK_ITEM_SIZES[CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS]: maximum number of work items on each dimension
This is due to memory localization on GPUs is important enough that OpenCL exposes this extra level.
We can however has as many work groups as we want.
Synchronization only works inside a single work groups: http://stackoverflow.com/questions/5895001/opencl-synchronization-between-work-groups
Size of the work group.
On CPU: always 1. TODO why?
On GPU; must divide Global size.
Work groups with different sizes.
Application: take care of edge cases of the data, e.g. image edges: https://software.intel.com/en-us/articles/opencl-20-non-uniform-work-groups
Each work item runs your kernel code in parallel to the other ones.
An work item can be seen as a thread.
Contains private memory, which no other work item can see.
TODO: can a single work item be run in parallel on the GPU?
- http://stackoverflow.com/questions/9885880/effect-of-private-memory-in-opencl
- http://stackoverflow.com/questions/9885880/is-private-memory-slower-than-local-memory
TODO: can private memory be slower than local memory? http://stackoverflow.com/questions/21872810/whats-the-advantage-of-the-local-memory-in-opencl
Might be faster, and global memory is limited.
HandsOnOpencl Example 8 shows how matrix multiplication becomes 10x faster with some local memory usage. Looks like memory access was the bottleneck.
It also shows how we must make an explicit copy to use private memory.
- http://stackoverflow.com/questions/8888718/how-to-declare-local-memory-in-opencl
- http://stackoverflow.com/questions/2541929/how-do-i-use-local-memory-in-opencl
- http://stackoverflow.com/questions/17574570/create-local-array-dynamic-inside-opencl-kernel
TODO.