This PR adds a BF16 data-type, along with FP32 <-> BF16 conversion instructions in the LLVM codegen. Other kinds of ops on bfloat16 are not yet supported.
Improved codegen for the Ampere GPUs.
* Make the layout pass recognize the multistage pipelined pattern.
* Now the pipeline pass can automate the multistage pipelining transformation.
* Remove extra barriers (from the prefetch pass & WAR) on Ampere.
* Update the code generator (generator.cc) to make Triton generate n-buffered shared memory loads/stores.
Membar pass on top of master is buggy with asynchronous copy. For example, it doesn't wait for asynchronous copies to complete before recoalescing accumulator in GEMM, which leads to undefined behavior when the program doesn't enter the loop. This PR proposes
* Now using unordered instead of ordered float (fixes NaN issues)
* Bool -> int32 now converts to 1 rather than -1
* Reduce extend arguments to 32-bits if possible
* update membar pass when data is double buffered
* Add instruction prefetch_s
* prefetch tests pass (except the 1 warp case)
* Fix the 1-warp bug
* Add back prefetch files
* Disable prefetch on a100
* Always add war barrier on sm>=80
This massively simplifies implementation of `reassociate` and also fixes
a bunch of bug. The pass could still be improved, but can already be used
to generate constant pointer offsets in eg the matmul epilogue
This PR implements a major overhaul of the frontend for Triton, and replaces Triton-C by a pure Python API in which kernels are defined as @triton.jit decorated functions. The documentation and tutorials have also been updated to accommodate these changes.
See documentations for more information on the new API
This PR adds an automatic memory alignment mechanism in the Triton runtime. Specifically, the JIT compiler detects the alignment (in bytes) of each pointer argument as well as the largest power of two divisor (between 1 and 16) of each integer argument. Proper .aligned and .multipleof attributes are then added to the Triton-IR on-the-fly for all auto-tunable kernels. There is a cache that remembers all the kernels compiled for each possible configuration.
This PR also includes substantial cleaning of the Python API. This adds 2-3us overhead, mostly due to accessing integer #defines from the auto-tuned compilation options. The previous solution was slightly faster but hacky and potentially unsafe, so this is preferred for now.
Improved handling of asynchronous copy, scheduling and synchronization for A100. Now achieving CUTLASS-like performance on large square dense matrix multiplication tasks
- A100 support via mma.16816
- Thread swizzling for conflict-free shared memory accesses without
padding
- Complete overhaul of the LLVM code generation in
codegen/selection/generator.cc to remove overengineering
- Added debugging capabilities in the Python binding
- Compilation error for kernels that spill
torch-blocksparse package:
* Now using warp shuffle in reductions when possible
* Various bugfixes in layout inference
* Added INFINITY, exponential and select
* Better error messages for unimplemented constructs
