Before we tracked the size of the teams reduction buffer in order to
allocate it at runtime per kernel launch. This patch splits the number
into two parts, the size of the reduction data (=all reduction
variables) and the (maximal) length of the buffer. This will allow us to
allocate less if we need less, e.g., if we have less teams than the
maximal length. It also allows us to move code from clangs codegen into
the runtime as we now know how large the reduction data is.
We used to perform team reduction on global memory allocated in the
runtime and by clang. This was racy as multiple instances of a kernel,
or different kernels with team reductions, would use the same locations.
Since we now have the kernel launch environment, we can allocate dynamic
memory per-launch, allowing us to move all the state into a non-racy
place.
Fixes: https://github.com/llvm/llvm-project/issues/70249
The KernelEnvironment is for compile time information about a kernel. It
allows the compiler to feed information to the runtime. The
KernelLaunchEnvironment is for dynamic information *per* kernel launch.
It allows the rutime to feed information to the kernel that is not
shared with other invocations of the kernel. The first use case is to
replace the globals that synchronize teams reductions with per-launch
versions. This allows concurrent teams reductions. More uses cases will
follow, e.g., per launch memory pools.
Fixes: https://github.com/llvm/llvm-project/issues/70249
The patch contains a basic BumpAllocator for (AMD)GPUs to allow us to
run more tests. The allocator implements `malloc`, both internally and
externally, while we continue to default to the NVIDIA `malloc` when we
target NVIDIA GPUs. Once we have smarter or customizable allocators we
should consider this choice, for now, this allocator is better than
none. It traps if it is out of memory, making it easy to debug. Heap
size is configured via `LIBOMPTARGET_HEAP_SIZE` and defaults to 512MB.
It allows to track allocation statistics via
`LIBOMPTARGET_DEVICE_RTL_DEBUG=8` (together with
`-fopenmp-target-debug=8`). Two tests were added, and one was enabled.
This is the next step towards fixing
https://github.com/llvm/llvm-project/issues/66708
Summary:
The `omp_get_num_procs()` function should return the amount of
parallelism availible. On the GPU, this was not defined. We have elected
to define this function as the maximum amount of wavefronts / warps that
can be simultaneously resident on the device. For AMDGPU this is the
number of CUs multiplied byth CU's per wave. For NVPTX this is the
maximum threads per SM divided by the warp size and multiplied by the
number of SMs.
The changes in D157738 allowed for us to emit stub globals on the device
in the offloading entry section. These globals contain the addresses of
device functions and allow us to map host functions to their
corresponding device equivalent. This patch provides the initial support
required to build a table on the device to lookup the associated value.
This is done by finding these entries and creating a global table on the
device that can be searched with a simple binary search.
This requires an allocation, which supposedly should be automatically
freed at plugin shutdown. This includes a basic test which looks up device
pointers via a host pointer using the added function. This will need to be built
upon to provide full support for these calls in the runtime.
To support reverse offloading it would also be useful to provide a reverse table
that allows us to get host functions from device stubs.
Depends on D157738
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D157918
This feature was supposed to allow you to trace execution inside of
Libomptarget. However, this never really worked properly. The printing
was always reoganized, only worked for single threads, and pretty much
only told you a handful of things about a runtime library that's an
implementation detail to all users. Despite this, it contributed about
40% of the total filesize of the deviceRTL. This patch simply removes
this functionalit which I think was past due.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D157001
This patch introduces per kernel environment. Previously, flags such as execution mode are set through global variables with name like `__kernel_name_exec_mode`. They are accessible on the host by reading the corresponding global variable, but not from the device. Besides, some assumptions, such as no nested parallelism, are not per kernel basis, preventing us applying per kernel optimization in the device runtime.
This is a combination and refinement of patch series D116908, D116909, and D116910.
Depend on D155886.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D142569
This patch introduces per kernel environment. Previously, flags such as execution mode are set through global variables with name like `__kernel_name_exec_mode`. They are accessible on the host by reading the corresponding global variable, but not from the device. Besides, some assumptions, such as no nested parallelism, are not per kernel basis, preventing us applying per kernel optimization in the device runtime.
This is a combination and refinement of patch series D116908, D116909, and D116910.
Depend on D155886.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D142569
If we store a constant in an ICV it is easier for the optimizer to
propagate it. Since we often use the full block for the thread limit and
the parallel team size, we can instead replace that dynamic value with a
constant that otherwise cannot occur, here 0.
We ended up with `llvm.assume(icmp ne ptr as(4) null, as(4) @str)`
because the string in address space 4 was not known to be non-null.
There is no need to create these assumes.
AMDGPU provides a fixed frequency clock since some generations back.
However, the frequency is variable by card and must be looked up at
runtime. This patch adds a new device environment line for the clock
frequency so that we can use it in the same way as NVPTX. This is the
correct implementation and the version in ASO should be replaced.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D154456
With https://reviews.llvm.org/D137524, memory scope and ordering
attributes are being used to generate the required instructions for
atomic inc/dec on AMDGPU. This patch adds the memory scope attribute to
the atomic::inc API and uses the device scope in reduction. Without
the device scope in atomic_inc, the default system scope leads to
unnecessary L2 write-backs/invalidates.
Reviewed By: arsenm
Differential Revision: https://reviews.llvm.org/D154172
These files aren't fully formatted. I'm guessing this was a holdover
from when `clang-format` was totally broken for OpenMP offloading.
Format the files to be more consistent.
Reviewed By: tianshilei1992
Differential Revision: https://reviews.llvm.org/D151226
This reverts commit 35cfadfbe2.
It makes a couple of buildbots unhappy because of the following test failures:
- `Transforms/OpenMP/add_attributes.ll'`
- `mapping/declare_mapper_target_data.cpp` on AMDGPU
This patch introduces per kernel environment. Previously, flags such as execution mode are set through global variables with name like `__kernel_name_exec_mode`. They are accessible on the host by reading the corresponding global variable, but not from the device. Besides, some assumptions, such as no nested parallelism, are not per kernel basis, preventing us applying per kernel optimization in the device runtime.
This is a combination and refinement of patch series D116908, D116909, and D116910.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D142569
It turns out that the __builtin_amdgcn_s_barrier() alone does not emit
a fence. We somehow got away with this and assumed it would work as it
(hopefully) is correct on the NVIDIA path where we just emit a
__syncthreads. After talking to @arsenm we now (mostly) align with the
OpenCL barrier implementation [1] and emit explicit fences for AMDGPUs.
It seems this was the underlying cause for #59759, but I am not 100%
certain. There is a chance this simply hides the problem.
Fixes: https://github.com/llvm/llvm-project/issues/59759
[1] 07b347366e/opencl/src/workgroup/wgbarrier.cl (L21)
Code generation support for 'parallel masked' directive.
The `EmitOMPParallelMaskedDirective` was implemented.
In addition, the appropiate device functions were added.
Fix#59939.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D143527
It turns out that the `__builtin_amdgcn_s_barrier()` alone does not emit
a fence. We somehow got away with this and assumed it would work as it
(hopefully) is correct on the NVIDIA path where we just emit a
`__syncthreads`. After talking to @arsenm we now (mostly) align with the
OpenCL barrier implementation [1] and emit explicit fences for AMDGPUs.
It seems this was the underlying cause for #59759, but I am not 100%
certain. There is a chance this simply hides the problem.
Fixes: https://github.com/llvm/llvm-project/issues/59759
[1] 07b347366e/opencl/src/workgroup/wgbarrier.cl (L21)
Reviewed By: ye-luo
Differential Revision: https://reviews.llvm.org/D145290
This patch breaks the handling of `printf` in the OpenMP library. Usiing
`-ffreestanding` prevents clang from emitting LLVM builtins, which we
use for OpenMP printing support. Shelve this until we have functioning
`printf` in the GPU `libc` and we can remove that code.
This reverts commit a92eaa3ebe.
The `stdint.h` header provides the standard types. Previously we used
`-nostdinc` and defined these ourselves. This patch switches to a
freestanding version which should work properly. Without
`-ffreestanding` the `stdint.h` header will include other libraries. But
in a freestanding environment it should work given the primitives.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D145963
This patch makes preparation for a series that will enable per-kernel information
used in both host and device runtime. Some variables/enums, such as `OMPTgtExecModeFlags`,
have to be shared by both of them. A new header `OMPDeviceConstants.h` is added,
containing code that will be shared by them. We will introduce more variables soon.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D142320
This patch fixes the support for tasking on the device.
Note: AMDGPU doesn't support it yet because of no support for `malloc` and `free`.
Fix#59946.
```
➜ ./test_parallel_master_device
[OMPVV_RESULT: test_parallel_master_device.c] Test passed on the device.
```
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D141562
A previous patch merged the static and bitcode versions of the
deviceRTL. We previously used the static library's separate compilation
to set a special flag that prevented `IsSPMDMode` from being put in the
used list and preventing it from being optimized out. When they were
merged we could no longer do this separate compilation that allowed
users of LTO to get more optimal code.
This patch rearranges the code. The `IsSPMDMode` global is now
transitively used by its inclusion in the changed `__keep_alive`
function. This allows us to then manually delete the `__keep_alive`
function from the module when building the static library via
`llvm-extract`. The result is that the bitcode library correctly will
maintain the needed shared state, while the static library will be able
to internalize it and optimize it out.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D135280
If we have thread states, the program is going to be rather slow. If we
don't, we want to avoid wasting shared memory. This patch introduces a
slight penalty (malloc + indirection) for the slow path and reduces
resource usage for the fast path.
Differential Revision: https://reviews.llvm.org/D135037
We should use OpenMP atomics but they don't take variable orderings.
Maybe we should expose all of this in the header but that solves only
part of the problem anyway.
Differential Revision: https://reviews.llvm.org/D135036
Parallel regions are outlined as functions with capture variables explicitly generated as distinct parameters in the function's argument list. That complicates the fork_call interface in the OpenMP runtime: (1) the fork_call is variadic since there is a variable number of arguments to forward to the outlined function, (2) wrapping/unwrapping arguments happens in the OpenMP runtime, which is sub-optimal, has been a source of ABI bugs, and has a hardcoded limit (16) in the number of arguments, (3) forwarded arguments must cast to pointer types, which complicates debugging. This patch avoids those issues by aggregating captured arguments in a struct to pass to the fork_call.
Reviewed By: jdoerfert, jhuber6, ABataev
Differential Revision: https://reviews.llvm.org/D102107
The OpenMP device runtime needs to support the OpenMP standard. However
constructs like nested parallelism are very uncommon in real application
yet lead to complexity in the runtime that is sometimes difficult to
optimize out. As a stop-gap for performance we should supply an argument
that selectively disables this feature. This patch adds the
`-fopenmp-assume-no-nested-parallelism` argument which explicitly
disables the usee of nested parallelism in OpenMP.
Reviewed By: carlo.bertolli
Differential Revision: https://reviews.llvm.org/D132074
We will add some simple implementation of libc functions starting from
this patch, and the first one is `memcmp`, which is reported in #56929. Note that
`malloc` and `free` are not included in this patch because of the use of
`declare variant`. In the near future we will implement the two functions w/o
using any vendor provided function.
This fixes#56929.
Reviewed By: jhuber6
Differential Revision: https://reviews.llvm.org/D131182
The behaviour of this patch is not great, but it has some side-effects
that are required for OpenMPOpt to work. The problem is that when we use
`-mlink-builtin-bitcode` we only import used symbols from the runtime.
Then OpenMPOpt will insert calls to symbols that were not previously
included. This patch removed this implicit behaviour as these functions
were kept alive by the `noinline` simply because it kept calls to them
in the module. This caused regression in some tests that relied on some
OpenMPOpt passes without using LTO. Reverting for the LLVM15 release but
will try to fix it more correctly on main.
This reverts commit d61d72dae6.
Fixes#56752
We previously used the `noinline` attributes to specify some defintions
which should be kept alive in the runtime. These were then stripped
immediately in the OpenMPOpt module pass. However, Since the changes in
D130298, we not explicitly state which functions will have external
visiblity in the bitcode library. Additionally the OpenMPOpt module pass
should run before the inliner pass, so this shouldn't make a difference
in whether or not the functions will be alive for the initial pass of
OpenMPOpt. This should simplify the interface, and additionally save
time spend on scanning funciton names for noinline.
Reviewed By: jdoerfert
Differential Revision: https://reviews.llvm.org/D130368
We can help optimizations by making sure we use the team state whenever
it is clear there is no thread state. To this end we introduce a new
state flag (`state::HasThreadState`) and explicit control for the
`state::ValueRAII` helpers, including a dedicated "assert equal".
Differential Revision: https://reviews.llvm.org/D130113
