Target memory manager is introduced in this patch which aims to manage target
memory such that they will not be freed immediately when they are not used
because the overhead of memory allocation and free is very large. For CUDA
device, cuMemFree even blocks the context switch on device which affects
concurrent kernel execution.
The memory manager can be taken as a memory pool. It divides the pool into
multiple buckets according to the size such that memory allocation/free
distributed to different buckets will not affect each other.
In this version, we use the exact-equality policy to find a free buffer. This
is an open question: will best-fit work better here? IMO, best-fit is not good
for target memory management because computation on GPU usually requires GBs of
data. Best-fit might lead to a serious waste. For example, there is a free
buffer of size 1960MB, and now we need a buffer of size 1200MB. If best-fit,
the free buffer will be returned, leading to a 760MB waste.
The allocation will happen when there is no free memory left, and the memory
free on device will take place in the following two cases:
1. The program ends. Obviously. However, there is a little problem that plugin
library is destroyed before the memory manager is destroyed, leading to a fact
that the call to target plugin will not succeed.
2. Device is out of memory when we request a new memory. The manager will walk
through all free buffers from the bucket with largest base size, pick up one
buffer, free it, and try to allocate immediately. If it succeeds, it will
return right away rather than freeing all buffers in free list.
Update:
A threshold (8KB by default) is set such that users could control what size of memory
will be managed by the manager. It can also be configured by an environment variable
`LIBOMPTARGET_MEMORY_MANAGER_THRESHOLD`.
Reviewed By: jdoerfert, ye-luo, JonChesterfield
Differential Revision: https://reviews.llvm.org/D81054
Summary:
In current implementation, D2D memcpy is first to copy data back to host and then
copy from host to device. This is very efficient if the device supports D2D
memcpy, like CUDA.
In this patch, D2D memcpy will first try to use native supported driver API. If
it fails, fall back to original way. It is worth noting that D2D memcpy in this
scenerio contains two ideas:
- Same devices: this is the D2D memcpy in the CUDA context.
- Different devices: this is the PeerToPeer memcpy in the CUDA context.
My implementation merges this two parts. It chooses the best API according to
the source device and destination device.
Reviewers: jdoerfert, AndreyChurbanov, grokos
Reviewed By: jdoerfert
Subscribers: yaxunl, guansong, sstefan1, openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D80649
Summary:
Instead of using global variables with unpredicted time of
deinitialization, use dynamically allocated variables with functions
explicitly marked as global constructor/destructor and priority. This
allows to prevent the crash because of the incorrect order of dynamic
libraries deinitialization.
Reviewers: grokos, hfinkel
Subscribers: caomhin, kkwli0, openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D74837
Summary:
If the dynamically loaded module has been compiled with -fopenmp-targets
and has no target regions, it has empty target descriptor. It leads to a
crash at the runtime if another module has at least one target region
and at least one entry in its descriptor. The runtime library is unable
to load the empty binary descriptor and terminates the execution.
Caused by a clang-offload-wrapper.
Reviewers: grokos, jdoerfert
Subscribers: caomhin, kkwli0, openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D72472
Summary:
This patch adds support for using unified memory in the case of regular maps that happen when a target region is offloaded to the device.
For cases where only a single version of the data is required then the host address can be used. When variables need to be privatized in any way or globalized, then the copy to the device is still required for correctness.
Reviewers: ABataev, jdoerfert, Hahnfeld, AlexEichenberger, caomhin, grokos
Reviewed By: Hahnfeld
Subscribers: mgorny, guansong, openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D65001
llvm-svn: 368192
Remove loopTripCnt from threaded device stack after consuming it.
Added a libomptarget DP message to aid in future debugging and to
validate the added testcase, which only runs in Debug build.
Differential Revision: https://reviews.llvm.org/D64808
llvm-svn: 366349
Summary:
We used to call __kmpc_omp_taskwait function with global threadid set to
0. It may crash the application at the runtime if the thread executing
target region is not a master thread.
Reviewers: grokos, kkwli0
Subscribers: guansong, jdoerfert, caomhin, openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D64571
llvm-svn: 366220
Summary:
Target link variables are currently implemented by creating a copy of the variables on the device side and unified memory never gets exploited.
When the prgram uses the:
```
#pragma omp requires unified_shared_memory
```
directive in conjunction with a declare target link, the linked variable is no longer allocated on the device and the host version is used instead.
This behavior is overridden by performing an explicit mapping.
A Clang side patch is required.
Reviewers: ABataev, AlexEichenberger, grokos, Hahnfeld
Reviewed By: AlexEichenberger, grokos, Hahnfeld
Subscribers: Hahnfeld, jfb, guansong, jdoerfert, openmp-commits
Tags: #openmp
Differential Revision: https://reviews.llvm.org/D60223
llvm-svn: 361294
This adds AArch64 support to recently added part of the runtime responsible for offloading to target. This piece of code allows offloading-to-self on AArch64 machines.
Differential Revision: https://reviews.llvm.org/D30644
llvm-svn: 297070
Shilei Tian