<git-pr-chain>
#### Commits in this PR
1. Fix segfault in assertion test.
The issue here is that we were not checking the return values of the
CUDA API
calls we were making. We call one function and then use the data it
returns as
input to another call. Obviously this doesn't work if the first call
returns
an error and doesn't actually return meaningful data.
I don't know why this was passing in CI, but it failed consistently for
me.
#### [PR chain](https://github.com/jlebar/git-pr-chain)
1. 👉#2520👈 **YOU ARE HERE**
</git-pr-chain>
<git-pr-chain>
#### Commits in this PR
1. Make kernel_static_print test work when called twice.
This test is checking that a message is printed when the kernel is
compiled.
But the test had nothing to force the kernel to be compiled every time
you ran
the test. So after you ran it once, the test would fail every time until
you
cleared the cache.
#### [PR chain](https://github.com/jlebar/git-pr-chain)
1. 👉#2518👈 **YOU ARE HERE**
1. #2520
</git-pr-chain>
There's no guarantee that `/tmp/triton/*/*.json` existing means
that the corresponding `/tmp/triton/*/*.cubin` file also exists because the tmp directory doesn't guarantee file stability.
When propagating layout we were generating a view op with mismatching
total number of element per threads. Lowering such op would require
exchanging data across threads.
This change prevents the optimizer from generating such cases. This may
require further optimizations in the future.
I noticed that Triton is using the `ptxas` version as part of the
version hash even for non-CUDA targets. This is an attempt at fixing
this. Moving the version calculation to the back-end makes sense to me
from an architectural standpoint, so that's my approach here. I'm not as
confident in the implementation, so please if folks have any feedback
let me know.
Without this change, a constexpr assignment (ie. `A = B & C`, where `B`
and `C` are both constexpr) is getting assigned to a triton tensor,
which becomes an issue when `A` is used as the condition of an If
statement.
Note: I had to add `not isinstance(node.value, ast.Constant)` to the
condition because if we are assigning `x = 0` then the assigned value is
also a constexpr, but in this case we do want to assign a triton tensor
to `x` so that we can do `x.to(tl.int64)` for example, which cannot be
done on a constexpr.
---------
Co-authored-by: Philippe Tillet <phil@openai.com>
* Add waves_per_eu in the tuning space
* Do not allocate tensor on device during kernel compilation step
* Add breakdown elapsed time
* Parallelize the post-processing step
* Parallelize the profile step with --ngpus
* Better timing info printout
By default, ptxas will enable fusion of mul/add to fma instructions. The
backend was also being configured unconditionally to enable this on
conversion from LLVM IR to PTX. This commit adds an option which can be
used to disable the FP fusion behavior in both locations.
In current implementation, warpsPerCTA is always set to [numWarps, 1]
for 2 tt.dot fusion scenario. But, it is not optimal for cases such that
tt.dot doesn't have enough parallelism on row dimension but on column
dimension.
### Problem
The previous change still didn't link libstdc++fs into libtriton.so,
which caused the runtime error: undefined symbol
_ZNKSt10filesystem7__cxx114path11parent_pathEv
`link_libraries(stdc++fs)` should be placed before `add_library`.
### What this PR does
This PR Makes the link_libraries(stdc++fs) global for all targets in the
CMake project. By doing so, we ensure that the stdc++fs library is
consistently linked to all targets, addressing potential build issues on
old linux OS like debian10 which uses GCC8.3.0 as the default C/C++
compiler.
The pipeliner was overallocating shared memory for the inputs
for current schedule. This reduces the shared memory usage to only
what is needed.
Note that improving membar analysis could allow taking advantage of
allocating extra buffers to remove barriers.
* Attempt to absorb upstream's changes to improve causal=True
* Add autotuner
* Optimize for AMD MI250
- add pre_load_v as a tuning parameter
- do not define N_CTX as constexpr
- perform the second dot before sum
- remove qk_scale out of the inner loop
- add more configs in the autotuner
Note that bwd kernel is disabled for now. This is because we enabled
autotuning and grid becomes a function. So ctx.grid[0] no longer works.
* Enable bwd kernel
* enforce cc=None on ROCm
* Comment
* Update approach to ignore integer cc values
Co-authored-by: Alexander Efimov <efimov.alexander@gmail.com>
---------
Co-authored-by: Alexander Efimov <efimov.alexander@gmail.com>
* Restructure ROCM Library Search
Currently there are a handful of ROCM dependant files which are required for
triton to run. The linker(ld.lld), the include files, and multiple hip/hsa
shared objects.
This change will provide three search areas to find these files. All in
the same order.
1. third_party/rocm. This location is within the python/triton directory
and is carried over when triton is built. IF all necessary files
are in this location there will be no need to have ROCM installed at
all on the system.
2. $ROCM_PATH environmental variable. If this exists it will override
all other locations to find ROCM necessary files
3. /opt/rocm. The default location for ROCm installations. Finding one
here will notify triton that ROCM is installed in this environment
To ease with step 3. A new script scripts/amd/setup_rocm_libs.sh
has been added to the repo. Executing this script will cause all necessary
ROCM files to be downloaded from their respective packages on repo.radeon.com
and installed in third_party/rocm. Allowing for triton to run without installing
the full ROCM stack. setup_rocm_libs.sh takes a env_var ROCM_VERSION if a user
wishes to install a ROCM version other than the default (currently 5.4.2)
When triton whls are built to support Pytorch, method 3 will be used to stay in
sync with PyTorch's approach of bringing along any libraries needed and not
requiring ROCM to be installed.
(cherry picked from commit e6aea90fb3e8218cb562e5d990719112d8282702)
* Fix default rocm path
Running into `fatal error: hip/hip_runtime.h: No such file or directory` with latest wheel due to incorrect directory for ROCm libs
(cherry picked from commit 292bae625b113eb65c66cfe4442da7a6456c988a)
* setup_rocm_libs.sh manylinux refactor
(cherry picked from commit f995f314ada4606cb78dc6233cd9c8effc356191)
* Set setup_rocm_libs.sh to be executable
(cherry picked from commit 05d67b9418cacda0d356c2102d7c1a887948b013)
* Revert to using numbered so files to fix upstream
(cherry picked from commit 34f8189eae57a23cc15b4b4f032fe25757e0db8e)
* Remove drm script
---------
Co-authored-by: Jeff Daily <jeff.daily@amd.com>
* [GEMM] Tuning script v2
* Extend tuning space to include BLOCK_SIZE = 256
Check LDS in a more smart way
* Added README
* Add git branch and commit to the default tuning result filename
This seems to have worked fine in opt mode (although it may be producing
undefined behavior), but in debug mode on a newer version of llvm, it
segfaults without this PR as the iterators get invalidated.
This is also consistent with other places it is done in this file.
Fix dependencies in wgmma_wait op to prevent the scheduler from moving
it past the uses of wgmma accumulator. We need to explicitly represent
the dependency between the wait and the accumulator uses otherwise LLVM
is free to re-order those.
This allows us to remove a workaround to prevent the re-ordering. We can
also remove the wait op added in the loop during pipelining.
Also fix the descritpor calcuation for wgmma, we should calculate the
same descriptor for the whole warpgroup.
Added a workaround for a bug that was exposed by different timing due to
those changes. We shouldn't insert operations between the loop and
async_wait or we may have race conditions.
