Cooperative groups #2307
Cooperative groups #2307
Conversation
|
Looks good so far :-) One functionality that we will need is the possibility of querying the maximum number of blocks that can be used with a given kernel, so the user can store it and use it for launching the kernel. |
|
For some reason gridSync locks up, when compiled CUDA Clang, if numberOfBlocks > 2 * multiProcessorCount |
|
|
||
| //! Hello world kernel, utilizing grid synchronization. | ||
| //! Prints hello world from a thread, performs grid sync. | ||
| //! and prints the sum of indixes of this thread and the opposite thread (the sums have to be the same). |
There was a problem hiding this comment.
[nit] Could you explain what is the opposite thread here.
There was a problem hiding this comment.
Thread, that has the same distance from the end of the grid dimension, as this from the start. So, if the IDs range from 0 to 9, these are 0 and 9, 1 and 8, 2 and 7 and so on. Their sum is constant, so we can check, if grid sync was performed successfully
Could you add more details to the PR definition |
That depends on the desired scope of the PR, I've deliberately made it vague, so we can decide, when to merge it. |
This looks like an upstream issue, at least locally, when compiling with clang 17.0.6 and CUDA 12.1.1 and 12.5 |
Is it supposed to work? |
Yes, it is, maximum number reported is 16 * multiProcessorCount, and the kernel refuses to launch, on both nvcc and clang if this number is exceeded. |
MichaelVarvarin
force-pushed
the
cooperative-groups
branch
2 times, most recently
from
bb5ccaa
to
9614d9c
Compare
Nice 👍🏻 |
Doesn't actually work, unfortunately, I will investigate it. Would be funny, if I find a second compiler bug |
MichaelVarvarin
force-pushed
the
cooperative-groups
branch
from
26c4803
to
565c944
Compare
MichaelVarvarin
force-pushed
the
cooperative-groups
branch
2 times, most recently
from
36c2af3
to
334cea9
Compare
…nching the specified cooperative kernel
MichaelVarvarin
force-pushed
the
cooperative-groups
branch
from
334cea9
to
4ad8bae
Compare
| static inline Error_t launchCooperativeKernel( | ||
| void const* func, | ||
| dim3 gridDim, | ||
| dim3 blockDim, | ||
| void** args, | ||
| size_t sharedMem, | ||
| Stream_t stream) | ||
| { | ||
| return ::cudaLaunchCooperativeKernel(func, gridDim, blockDim, args, sharedMem, stream); | ||
| } | ||
|
|
There was a problem hiding this comment.
Could you change this to be templated on the func argument ?
| static inline Error_t launchCooperativeKernel( | |
| void const* func, | |
| dim3 gridDim, | |
| dim3 blockDim, | |
| void** args, | |
| size_t sharedMem, | |
| Stream_t stream) | |
| { | |
| return ::cudaLaunchCooperativeKernel(func, gridDim, blockDim, args, sharedMem, stream); | |
| } | |
| template <typename TFunc> | |
| static inline Error_t launchCooperativeKernel( | |
| TFunc func, | |
| dim3 gridDim, | |
| dim3 blockDim, | |
| void** args, | |
| size_t sharedMem, | |
| Stream_t stream) | |
| { | |
| return ::cudaLaunchCooperativeKernel(func, gridDim, blockDim, args, sharedMem, stream); | |
| } | |
Same for the HIP implementation.
| void const* kernelArgs[] = {&threadElemExtent, &task.m_kernelFnObj, &args...}; | ||
|
|
||
| TApi::launchCooperativeKernel( | ||
| reinterpret_cast<void*>(kernelName), |
There was a problem hiding this comment.
Can you check if the cast can be removed, after implementing https://github.com/alpaka-group/alpaka/pull/2307/files#r1714986927 ?
There was a problem hiding this comment.
No, you now have to do the same cast inside of launchCooperativeKernel
…td::threads accelerator
|
@MichaelVarvarin Sorry for removing the draft state. I thought I start the GitHub Action jobs. Not sure what was going wrong. |
Add support for cooperative groups and related functionality