Summary:
This patch just changes the interface to make starting the thread
multiple times permissable since it will only be done the first time.
Note that this does not refcount it or anything, so it's onto the user
to make sure that they don't shut down the thread before everyone is
done using it. That is the case today because the shutDown portion is
run by a single thread in the destructor phase.
Another question is if we should make this thread truly global state,
because currently it will be private to each plugin instance, so if you
have an AMD and NVIDIA image there will be two, similarly if you have
those inside of a shared library.
Summary:
Pretty dumb mistake of me, forgot that this is run per-device and
per-plugin, which fell through the cracks with my testing because I have
two GPUs that use different plugins.
Summary:
Handling the RPC server requires running through list of jobs that the
device has requested to be done. Currently this is handled by the thread
that does the waiting for the kernel to finish. However, this is not
sound on NVIDIA architectures and only works for async launches in the
OpenMP model that uses helper threads.
However, we also don't want to have this thread doing work
unnnecessarily. For this reason we track the execution of kernels and
cause the thread to sleep via a condition variable (usually backed by
some kind of futex or other intelligent sleeping mechanism) so that the
thread will be idle while no kernels are running.
Summary:
This patch is an NFC renaming to make using the offloading entry type
more portable between other targets. Right now this is just moving its
definition to LLVM so others can use it. Future work will rework the
struct layout.
Summary:
This patch adds an RPC interface that lives directly in the OpenMP
device runtime. This allows OpenMP to implement custom opcodes.
Currently this is only providing the host call interface, which is the
raw version of reverse offloading. Previously this lived in `libc/` as
an extension which is not the correct place.
The interface here uses a weak symbol for the RPC client by the same
name that the `libc` interface uses. This means that it will defer to
the libc one if both are present so we don't need to set up multiple
instances.
The presense of this symbol is what controls whether or not we set up
the RPC server. Because this is an external symbol it normally won't be
optimized out, so there's a special pass in OpenMPOpt that deletes this
symbol if it is unused during linking. That means at `O0` the RPC server
will always be present now, but will be removed trivially if it's not
used at O1 and higher.
We had three `utils::` namespaces, all with different "meaning" (host,
device, hsa_utils). We should, when we can, keep "include/Shared"
accessible from host and device, thus RefCountTy has been moved to a
separate header. `hsa_utils` was introduced to make `utils::` less
overloaded. And common functionality was de-duplicated, e.g.,
`utils::advance` and `utils::advanceVoidPtr` -> `utils:advancePtr`. Type
punning now checks for the size of the result to make sure it matches
the source type.
No functional change was intended.
This pull request is a revised version of #76587. This pull request
fixes some build issues that were present in the previous version of
this change.
> This pull request is the first part of an ongoing effort to extends
PGO instrumentation to GPU device code. This PR makes the following
changes:
>
> - Adds blank registration functions to device RTL
> - Gives PGO globals protected visibility when targeting a supported
GPU
> - Handles any addrspace casts for PGO calls
> - Implements PGO global extraction in GPU plugins (currently only
dumps info)
>
> These changes can be tested by supplying `-fprofile-instrument=clang`
while targeting a GPU.
Through the new `-foffload-via-llvm` flag, CUDA kernels can now be
lowered to the LLVM/Offload API. On the Clang side, this is simply done
by using the OpenMP offload toolchain and emitting calls to `llvm*`
functions to orchestrate the kernel launch rather than `cuda*`
functions. These `llvm*` functions are implemented on top of the
existing LLVM/Offload API.
As we are about to redefine the Offload API, this wil help us in the
design process as a second offload language.
We do not support any CUDA APIs yet, however, we could:
https://www.osti.gov/servlets/purl/1892137
For proper host execution we need to resurrect/rebase
https://tianshilei.me/wp-content/uploads/2021/12/llpp-2021.pdf
(which was designed for debugging).
```
❯❯❯ cat test.cu
extern "C" {
void *llvm_omp_target_alloc_shared(size_t Size, int DeviceNum);
void llvm_omp_target_free_shared(void *DevicePtr, int DeviceNum);
}
__global__ void square(int *A) { *A = 42; }
int main(int argc, char **argv) {
int DevNo = 0;
int *Ptr = reinterpret_cast<int *>(llvm_omp_target_alloc_shared(4, DevNo));
*Ptr = 7;
printf("Ptr %p, *Ptr %i\n", Ptr, *Ptr);
square<<<1, 1>>>(Ptr);
printf("Ptr %p, *Ptr %i\n", Ptr, *Ptr);
llvm_omp_target_free_shared(Ptr, DevNo);
}
❯❯❯ clang++ test.cu -O3 -o test123 -foffload-via-llvm --offload-arch=native
❯❯❯ llvm-objdump --offloading test123
test123: file format elf64-x86-64
OFFLOADING IMAGE [0]:
kind elf
arch gfx90a
triple amdgcn-amd-amdhsa
producer openmp
❯❯❯ LIBOMPTARGET_INFO=16 ./test123
Ptr 0x155448ac8000, *Ptr 7
Ptr 0x155448ac8000, *Ptr 42
```
Similar to (de)allocation traces, we can record kernel launch stack
traces and display them in case of an error. However, the AMD GPU plugin
signal handler, which is invoked on memroy faults, cannot pinpoint the
offending kernel. Insteade print `<NUM>`, set via
`OFFLOAD_TRACK_NUM_KERNEL_LAUNCH_TRACES=<NUM>`, many traces. The
recoding/record uses a ring buffer of fixed size (for now 8).
For `trap` errors, we print the actual kernel name, and trace if
recorded.
As a first step towards a GPU sanitizer we now can track allocations and
deallocations in order to report double frees, and other problems during
deallocation.
This pull request is the first part of an ongoing effort to extends PGO
instrumentation to GPU device code. This PR makes the following changes:
- Adds blank registration functions to device RTL
- Gives PGO globals protected visibility when targeting a supported GPU
- Handles any addrspace casts for PGO calls
- Implements PGO global extraction in GPU plugins (currently only dumps
info)
These changes can be tested by supplying `-fprofile-instrument=clang`
while targeting a GPU.
Sometimes it might be beneficial to spawn more thread blocks instead of
reusing existing for multiple loop iterations.
**Alternatives considered:**
Make `DefaultNumBlocks` settable via an environment variable.
---------
Co-authored-by: Joseph Huber <huberjn@outlook.com>
We already used a flat array of kernel launch parameters for the AMD GPU
launch but now we also use this scheme for the NVIDIA GPU launch. The
only remaining/required use of the indirection is the host plugin (due
ot ffi). This allows to us simplify the use for non-OpenMP kernel
launch.
Summary:
Currently, we register images into a linear table according to the
logical OpenMP device identifier. We then initialize all of these images
as one block. This logic requires that images are compatible with *all*
devices instead of just the one that it can run on. This prevents us
from running on systems with heterogeneous devices (i.e. image 1 runs on
device 0 image 0 runs on device 1).
This patch reworks the logic by instead making the compatibility check a
per-device query. We then scan every device to see if it's compatible
and do it as they come.
Summary:
Initializing the plugins requires initializing the runtime like CUDA or
HSA. This has a considerable overhead on most platforms, so we should
only actually initialize a plugin if it is needed by any image that is
loaded.
Summary:
Previously, the R&R support was global state initialized by a global
constructor. This is bad because it prevents us from adequately
constraining the lifetime of the library. Additionally, we want to
minimize the amount of global state floating around.
This patch moves the R&R support into a plugin member like everything
else. This means there will be multiple copies of the R&R implementation
floating around, but this was already the case given the fact that we
currently handle everything with dynamic libraries.
Summary:
Currently this is only used for the zero-copy handling. However, this
can easily be moved into `libomptarget` so that we do not need to bother
setting the requires flags in the plugin. The advantage here is that we
no longer need to do this for every device redundently. Additionally,
these requires flags are specifically OpenMP related, so they should
live in `libomptarget`.
This patch overhauls the `libomptarget` and plugin interface. Currently,
we define a C API and compile each plugin as a separate shared library.
Then, `libomptarget` loads these API functions and forwards its internal
calls to them. This was originally designed to allow multiple
implementations of a library to be live. However, since then no one has
used this functionality and it prevents us from using much nicer
interfaces. If the old behavior is desired it should instead be
implemented as a separate plugin.
This patch replaces the `PluginAdaptorTy` interface with the
`GenericPluginTy` that is used by the plugins. Each plugin exports a
`createPlugin_<name>` function that is used to get the specific
implementation. This code is now shared with `libomptarget`.
There are some notable improvements to this.
1. Massively improved lifetimes of life runtime objects
2. The plugins can use a C++ interface
3. Global state does not need to be duplicated for each plugin +
libomptarget
4. Easier to use and add features and improve error handling
5. Less function call overhead / Improved LTO performance.
Additional changes in this plugin are related to contending with the
fact that state is now shared. Initialization and deinitialization is
now handled correctly and in phase with the underlying runtime, allowing
us to actually know when something is getting deallocated.
Depends on https://github.com/llvm/llvm-project/pull/86971https://github.com/llvm/llvm-project/pull/86875https://github.com/llvm/llvm-project/pull/86868
This patch overhauls the `libomptarget` and plugin interface. Currently,
we define a C API and compile each plugin as a separate shared library.
Then, `libomptarget` loads these API functions and forwards its internal
calls to them. This was originally designed to allow multiple
implementations of a library to be live. However, since then no one has
used this functionality and it prevents us from using much nicer
interfaces. If the old behavior is desired it should instead be
implemented as a separate plugin.
This patch replaces the `PluginAdaptorTy` interface with the
`GenericPluginTy` that is used by the plugins. Each plugin exports a
`createPlugin_<name>` function that is used to get the specific
implementation. This code is now shared with `libomptarget`.
There are some notable improvements to this.
1. Massively improved lifetimes of life runtime objects
2. The plugins can use a C++ interface
3. Global state does not need to be duplicated for each plugin +
libomptarget
4. Easier to use and add features and improve error handling
5. Less function call overhead / Improved LTO performance.
Additional changes in this plugin are related to contending with the
fact that state is now shared. Initialization and deinitialization is
now handled correctly and in phase with the underlying runtime, allowing
us to actually know when something is getting deallocated.
Depends on https://github.com/llvm/llvm-project/pull/86971https://github.com/llvm/llvm-project/pull/86875https://github.com/llvm/llvm-project/pull/86868
Summary:
The `GenericDeviceTy::dataDelete` method doesn't verify the
`TargetAllocTy` of the of the device pointer. Because of this, it can
use the `MemoryManager` to free the ptr. However, the
`TARGET_ALLOC_HOST` and `TARGET_ALLOC_SHARED` types are not allocated
using the `MemoryManager` in the `GenericDeviceTy::dataAlloc` method.
Since the `MemoryManager` uses the `DeviceAllocatorTy::free` operation
without specifying the type of the ptr, some plugins may use incorrect
operations to free ptrs of certain types. In particular, this bug causes
the CUDA plugin to use the `cuMemFree` operation on ptrs of type
`TARGET_ALLOC_HOST`, resulting in an unchecked error, as shown in the
output snippet of the test
`offload/test/api/omp_host_pinned_memory_alloc.c`:
```
omptarget --> Notifying about an unmapping: HstPtr=0x00007c6114200000
omptarget --> Call to llvm_omp_target_free_host for device 0 and address 0x00007c6114200000
omptarget --> Call to omp_get_num_devices returning 1
omptarget --> Call to omp_get_initial_device returning 1
PluginInterface --> MemoryManagerTy:🆓 target memory 0x00007c6114200000.
PluginInterface --> Cannot find its node. Delete it on device directly.
TARGET CUDA RTL --> Failure to free memory: Error in cuMemFree[Host]: invalid argument
omptarget --> omp_target_free deallocated device ptr
```
This patch fixes this by adding the check of the device pointer type
before calling the appropriate operation for each type.
In a nutshell, this moves our libomptarget code to populate the offload
subproject.
With this commit, users need to enable the new LLVM/Offload subproject
as a runtime in their cmake configuration.
No further changes are expected for downstream code.
Tests and other components still depend on OpenMP and have also not been
renamed. The results below are for a build in which OpenMP and Offload
are enabled runtimes. In addition to the pure `git mv`, we needed to
adjust some CMake files. Nothing is intended to change semantics.
```
ninja check-offload
```
Works with the X86 and AMDGPU offload tests
```
ninja check-openmp
```
Still works but doesn't build offload tests anymore.
```
ls install/lib
```
Shows all expected libraries, incl.
- `libomptarget.devicertl.a`
- `libomptarget-nvptx-sm_90.bc`
- `libomptarget.rtl.amdgpu.so` -> `libomptarget.rtl.amdgpu.so.18git`
- `libomptarget.so` -> `libomptarget.so.18git`
Fixes: https://github.com/llvm/llvm-project/issues/75124
---------
Co-authored-by: Saiyedul Islam <Saiyedul.Islam@amd.com>
