This patch adds `olGetEventElapsedTime` to the new LLVM Offload API, as
requested in
[#185728](https://github.com/llvm/llvm-project/issues/185728), and adds
the corresponding support in `plugins-nextgen`.
A main motivation for this change is to make it possible to measure the
elapsed time of work submitted to a queue, especially kernel launches.
This is relevant to the intended use of the new Offload API for
microbenchmarking GPU libc math functions.
### Summary
The new API returns the elapsed time, in milliseconds, between two
events on the same device.
To support the common pattern `create start event → enqueue kernel →
create end event → sync end event → get elapsed time`, `olCreateEvent`
now always creates and records a backend event through the device
interface. For backends that materialize real event state, this gives
the event concrete backend state that can be used for elapsed-time
measurement. For backends that do not materialize backend event state,
`EventInfo` may still remain null and existing event operations continue
to treat such events as trivially complete.
Previously, an event created on an empty queue could be represented only
as a logical event. That representation was sufficient for sync and
completion queries, but it was not suitable for elapsed-time measurement
because there was no backend event state to timestamp. The new behavior
preserves the meaning of completion of prior work while also allowing
backends with timing support to attach real event state.
### Changes in `plugins-nextgen`
#### Common interface
Add elapsed-time support to the common device and plugin interfaces:
* `GenericPluginTy::get_event_elapsed_time`
* `GenericDeviceTy::getEventElapsedTime`
* `GenericDeviceTy::getEventElapsedTimeImpl`
#### AMDGPU
* Add the required ROCr declarations and wrappers.
* Enable queue profiling at queue creation time.
* Record events by enqueuing a real barrier marker packet on the stream.
* Retain the timing signal needed to query the recorded marker later.
* Implement `getEventElapsedTimeImpl` using
`hsa_amd_profiling_get_dispatch_time`, converting the result to
milliseconds with `HSA_SYSTEM_INFO_TIMESTAMP_FREQUENCY`.
This follows the ROCm/HIP approach of enabling queue profiling at HSA
queue creation time, while keeping the AMDGPU queue path simpler than
the lazy-enable alternative discussed during review.
#### CUDA
* Add the required CUDA driver declarations and wrappers.
* Implement `getEventElapsedTimeImpl` with `cuEventElapsedTime`.
#### Host
* Add `getEventElapsedTimeImpl` that stores `0.0f` in the output
pointer, when present, and returns success.
Reason: the host plugin does not materialize backend event state and
already treats event operations as trivially successful. Returning
`0.0f` preserves that model without introducing a new failure mode.
#### Level Zero
* Add `getEventElapsedTimeImpl`, but leave it unimplemented.
Reason: the Level Zero plugin currently does not provide standalone
backend event support for this event model. For example, `waitEventImpl`
/ `syncEventImpl` are still unimplemented there.
---------
Signed-off-by: Leandro Augusto Lacerda Campos <leandrolcampos@yahoo.com.br>
Signed-off-by: Leandro A. Lacerda Campos <leandrolcampos@yahoo.com.br>
This PR adds extends liboffload olMemRegister API to handle a case when
a memory block may have been mapped before calling olMemRegister to
support some use cases in libomptarget
Summary:
This PR adds a pointer argument to the initialization routine to be used
for global options. Right now this is used to allow the user to
constrain which backends they wish to use.
If a null argument is passed, the same behavior as before is observed.
This is epxected to be extensible by forcing the user to encode the size
of the struct. So, old executables will encode which fields they have
access to.
We use a macro helper to get this struct rather than a runtime call so
that the current state of the size is baked into the executable rather
than something looked up by the runtime. Otherwise it would just return
the size that the (potentially newer) runtime would see
The purpose of this PR is to add support of host as an offloading device
to liboffload. Both OpenMP and sycl support offloading to a host as
their normal workflow and therefore would require such capability from
liboffload library.
Summary:
We provide an RPC server to manage calls initiated by the device to run
on the host. This is very useful for the built-in handling we have,
however there are cases where we would want to extend this
functionality.
Cases like Fortran or MPI would be useful, but we cannot put references
to these in the core offloading runtime. This way, we can provide this
as a library interface that registers custom handlers for whatever code
people want.
## Summary
- Fix uninitialized output parameter in `olQueryQueue_impl` when
`Queue->AsyncInfo->Queue` is null
- Set `IsQueueWorkCompleted` to `true` when no underlying queue exists
(no pending work)
- Resolves test failure on AMDGPU for
`olQueryQueueTest.SuccessEmptyAsyncQueueCheckResult`
Fixes#178462.
## Test plan
- [x] Fixed
`OffloadAPI/queue.unittests/olQueryQueueTest/SuccessEmptyAsyncQueueCheckResult/AMDGPU_AMD_Radeon_RX_7700_XT_0`
test
- [ ] CI tests pass
---------
Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
Co-authored-by: Joseph Huber <huberjn@outlook.com>
Per documentation the call to dataExchange API (move memory block
between different devices) is permitted only if isDataExchangable() call
returned true. While almost all platforms support memory transfer
between different devices, in the case when the transfer is attempted
between devices belonging to different platforms if they are present on
the same machine which can lead to unexpected results. This PR adds a
check if dataExchange can be called and if not uses a workaround by
initiating memory transfer through host.
Add liboffload asynchronous queue query API for libomptarget migration
This PR adds liboffload asynchronous queue query API that needed to make
libomptarget to use liboffload
Add liboffload memory data locking API for libomptarget migration
This PR adds liboffload memory data locking API that needed to make
libomptarget to use liboffload
The code to recognize the level_zero plugin as a liboffload backend was
split from #158900. This PR adds the support back.
---------
Co-authored-by: Alexey Sachkov <alexey.sachkov@intel.com>
Co-authored-by: Nick Sarnie <nick.sarnie@intel.com>
Co-authored-by: Joseph Huber <huberjn@outlook.com>
Introduced in OpenMP 6.0, the device UID shall be a unique identifier of
a device on a given system. (Not necessarily a UUID.) Since it is not
guaranteed that the (U)UIDs defined by the device vendor libraries, such
as HSA, do not overlap with those of other vendors, the device UIDs in
offload are always combined with the offload plugin name. In case the
vendor library does not specify any device UID for a given device, we
fall back to the offload-internal device ID.
The device UID can be retrieved using the `llvm-offload-device-info`
tool.
Summary:
The Offloading library wraps around the underlying plugins. The problem
is that we currently initialize all plugins we find, even if they are
not needed for the program. This is very expensive for trivial uses, as
fully heterogenous usage is quite rare. In practice this means that you
will always pay a 200 ms penalty for having CUDA installed.
This patch changes the behavior to provide accessors into the plugins
and devices that allows them to be initialized lazily. We use a
once_flag, this should properly take a fast-path check while still
blocking on concurrent use.
Making full use of this will require a way to filter platforms more
specifically. I'm thinking of what this would look like as an API.
I'm thinking that we either have an extra iterate function that takes a
callback on the platform, or we just provide a helper to find all the
devices that can run a given image. Maybe both?
Fixes: https://github.com/llvm/llvm-project/issues/159636
Summary:
Need to verify this actually has a device. We really need to rework this
to point to a real impolementation, or streamline it to handle this
automatically.
Currently, devices store a raw pointer to back to their owning Platform.
Platforms are stored directly inside of a vector. Modifying this vector
risks invalidating all the platform pointers stored in devices.
This patch allocates platforms individually, and changes devices to
store a reference to its platform instead of a pointer. This is safe,
because platforms are guaranteed to outlive the devices they contain.
Summary:
The host plugin is basically OpenMP specific and doesn't work very well.
Previously we were skipping over it in the list instead of just not
adding it at all.
If olMemAlloc happens to allocate memory that was already allocated
elsewhere (possibly by another device on another platform), it is now
thrown away and a new allocation generated.
A new `AllocBases` vector is now available, which is an ordered list
of allocation start addresses.
Summary:
This exposes the 'isDeviceCompatible' routine for checking if a binary
*can* be loaded. This is useful if people don't want to consume errors
everywhere when figuring out which image to put to what device.
I don't know if this is a good name, I was thining like `olIsCompatible`
or whatever. Let me know what you think.
Long term I'd like to be able to do something similar to what OpenMP
does where we can conditionally only initialize devices if we need them.
That's going to be support needed if we want this to be more
generic.
Summary:
Currently we have this `__tgt_device_image` indirection which just takes
a reference to some pointers. This was all find and good when the only
usage of this was from a section of GPU code that came from an ELF
constant section. However, we have expanded beyond that and now need to
worry about managing lifetimes. We have code that references the image
even after it was loaded internally. This patch changes the
implementation to instaed copy the memory buffer and manage it locally.
This PR reworks the JIT and other image handling to directly manage its
own memory. We now don't need to duplicate this behavior externally at
the Offload API level. Also we actually free these if the user unloads
them.
Upside, less likely to crash and burn. Downside, more latency when
loading an image.
Previously, `olDestroyQueue` would not actually destroy the queue,
instead leaving it for the device to clean up when it was destroyed.
Now, the queue is either released immediately if it is complete or put
into a list of "pending" queues if it is not. Whenever we create a new
queue, we check this list to see if any are now completed. If there are
any we release their resources and use them instead of pulling from
the pool.
This prevents long running programs that create and drop many queues
without syncing them from leaking memory all over the place.
The order entries in the tablegen API files are iterated is not the
order
they appear in the file. To avoid any issues with the order changing
in future, we now generate all definitions of a certain class before
class that can use them.
This is a NFC; the definitions don't actually change, just the order
they exist in in the OffloadAPI.h header.
Both these functions update an `AllocInfoMap` structure in the context,
however they did not use any locks, causing random failures in threaded
code. Now they use a mutex.
This is equivalent to `cuOccupancyMaxPotentialBlockSize`. It is
currently
only implemented on Cuda; AMDGPU and Host return unsupported.
---------
Co-authored-by: Callum Fare <callum@codeplay.com>
Add the following properties in Offload device info:
* VENDOR_ID
* NUM_COMPUTE_UNITS
* [SINGLE|DOUBLE|HALF]_FP_CONFIG
* NATIVE_VECTOR_WIDTH_[CHAR|SHORT|INT|LONG|FLOAT|DOUBLE|HALF]
* MAX_CLOCK_FREQUENCY
* MEMORY_CLOCK_RATE
* ADDRESS_BITS
* MAX_MEM_ALLOC_SIZE
* GLOBAL_MEM_SIZE
Add a bitfield option to enumerators, allowing the values to be
bit-shifted instead of incremented. Generate the per-type enums using
`foreach` to reduce code duplication.
Use macros in unit test definitions to reduce code duplication.
This sprinkles a few mutexes around the plugin interface so that the
olLaunchKernel CTS test now passes when ran on multiple threads.
Part of this also involved changing the interface for device synchronise
so that it can optionally not free the underlying queue (which
introduced a race condition in liboffload).
Add a device function to check if a device queue is empty. If liboffload
tries to create an event for an empty queue, we create an "empty" event
that is already complete.
This allows `olCreateEvent`, `olSyncEvent` and `olWaitEvent` to run
quickly for empty queues.
`MAX_WORK_GROUP_SIZE` now represents the maximum total number of work
groups the device can allocate, rather than the maximum per dimension.
`MAX_WORK_GROUP_SIZE_PER_DIMENSION` has been added, which has the old
behaviour.
Rather than having every "enqueue"-type function have an output pointer
specifically for an output event, just provide an `olCreateEvent`
entrypoint which pushes an event to the queue.
For example, replace:
```cpp
olMemcpy(Queue, ..., EventOut);
```
with
```cpp
olMemcpy(Queue, ...);
olCreateEvent(Queue, EventOut);
```
This is a hotfix for #148615 - it fixes the issue for me locally.
I think a broader issue is that in the test environment we're calling
olShutDown from a global destructor in the test binaries. We should do
something more controlled, either calling olInit/olShutDown in every
test, or move those to a GTest global environment. I didn't do that
originally because it looked like it needed changes to LLVM's GTest
wrapper.
