It is currently necessary for optimal performance in quantized workloads to add a special-purpose instruction in the IR. Backward compatibility with this instruction is *NOT* guaranteed.
This PR completely rewrites the runtime of Triton to be more lean and
clearly separate the compilation step from the just-in-time caching logic.
This should substantially reduce launch overhead.
This PR adds several optimization capabilities in the compiler backend:
- Now using inline PTX for `tl.store`, making it possible to use things like evict_last
- For A100, mma layout can be directly converted to shared memory
- For A100, an additional "transpose" argument in `dot` allows tensors to be loaded once and used both row- and col- major.
- Fixed liveness analysis; this was broken.
- Now can load/store directly mma layout without converting. Useful for when tl.dot accumulator is initialized with DRAM data inside of an inner loop.
- `tl.dot` can now take LHS inputs in registers when it comes from a previous `tl.dot` instruction. Useful for e.g. fused attention.
This is a more stable commit that produce bitwise identical code to earlier
versions. Using commits after this one may lead to slightly different numerics
Moved dispatch.cc to semantic.py (@ptillet)
Integer signedness analysis was moved from C++ to python (@daadaada)
Cleaner frontend types (@daadaada)
Moved SSA construction to a separate object (@ptillet)
Co-authored-by: Yan Da <dyanab@connect.ust.hk>
A forthcoming PR will update the RNG to use these types.
Also:
- Add tests for the `//`, `<<`, and `>>` operators.
- Change `TensorWrapper` to unwrap objects when the resulting object would be simpler.
- Clean up `throw_unreachable`, since it was triggering compiler warnings.
- `BF16TyID` was missing a repr implementation.
- Throw a better exception on impossible casts.
- Add a few assertions. Tested with a debug build.
- Add `pointer_dtype.__str__` to aid kernel debugging.
Significantly improves the performance of `triton.ops.matmul` in memory-bound settings via the use of many more block configs coupled with a performance model to drive the auto-tuning process.
- Don't override `self.args` in `CompilationError`, and show the line number and column in error messages. This causes it to generate an easier-to-read backtrace.
- Better `__str__` on `TensorWrapper`, `dtype`, and `block`.
- Removed driver module -- accelerator runtime is handled by pytorch
- Added basic support for ROCM based on @micmelesse 's PR -- now can execute empty kernel on AMD devices without any compile-time changes
- Now only using PREFER_SHARED for kernels when the size of shared memory is greater than 49k. Otherwise there can be poor L1 performance for broadcast tensors
