Add support for cooperative groups and grid synchronization #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
3 times, most recently
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
Add cooperative groups and gridSync to SYCL
MichaelVarvarin
force-pushed
the
cooperative-groups
branch
from
b3f7021
to
fab1f24
Compare
MichaelVarvarin
changed the title
This reverts commit 07aed98.
Add 4 new functions
createTaskCooperativeKernel(TWorkDiv const&, TKernelFnObj const& kernelFnObj, TArgs&&... args)that creates a task to later launch cooperative kernel, a new type of kernel with support for synchronizing the whole grid (look into CUDA documentation on cooperative kernels),execCooperative(TQueue& queue, TWorkDiv const& workDiv, TKernelFnObj const& kernelFnObj, TArgs&&... args)that immediately executes cooperative kernel,getMaxActiveBlocks(TDev const& device, TKernelFnObj const& kernelFnObj, alpaka::Vec<TDim, TIdx> const& blockThreadExtent, alpaka::Vec<TDim, TIdx> const& threadElemExtent, TArgs const&... args) -> intthat queries the maximum number of blocks supported for cooperative launch for the specified kernel andsyncGridThreads(acc)a barrier for all threads on the grid, similar tosyncBlockThreads(acc). Also adds a new device prop//m_cooperativeLaunchthat specifies, whether device support launch of cooperative kernels. Is queried at runtime for CUDA and HIP (for CUDA any GPU launched in the last 10 years should support it), is hard-coded to false for CpuSerial and to true for all else.