[Triton-MLIR][BACKEND] insert_slice_async on GPUs < sm80 (#908)

`insert_slice_async` is decomposed into `load + insert_slice` in the
backend.

Not sure if V100 perf can match the master branch though in this way.
Maybe the performance can be improved if instructions are arranged in
the following form:

```
%0 = load
%1 = load 
%2 = load 
...
insert_slice %0
insert_slice %1
insert_slice %2
```

Tested on A100 when manually enabling this decomposition.
Tests on V100 haven't been integrated yet, we can divide the tests into
two phases:
1. Test only load, insert_slice, and insert_slice_async, given TritonGPU
IRs in `test_backend.py`.
2. End to end gemm tests on V100.
This commit is contained in:
Keren Zhou
2022-11-24 14:05:54 -08:00
committed by GitHub
parent f98aed1258
commit 153aecb339
16 changed files with 351 additions and 137 deletions

View File

@@ -26,13 +26,14 @@ ChangeResult SharedMemoryAliasAnalysis::visitOperation(
// These ops may allocate a new shared memory buffer.
auto result = op->getResult(0);
// FIXME(Keren): extract and insert are always alias for now
if (auto extractSliceOp = dyn_cast<tensor::ExtractSliceOp>(op)) {
if (isa<tensor::ExtractSliceOp>(op)) {
// extract_slice %src
aliasInfo = AliasInfo(operands[0]->getValue());
pessimistic = false;
} else if (auto insertSliceOp =
dyn_cast<triton::gpu::InsertSliceAsyncOp>(op)) {
} else if (isa<tensor::InsertSliceOp>(op) ||
isa<triton::gpu::InsertSliceAsyncOp>(op)) {
// insert_slice_async %src, %dst, %index
// insert_slice %src into %dst[%offsets]
aliasInfo = AliasInfo(operands[1]->getValue());
pessimistic = false;
} else if (isSharedEncoding(result)) {