[Triton-MLIR] Replace triton.extract_slice with tensor.extract_slice and support more general tensor slicing (#837)

## Features

- Allow taking a block of tensor slice, as long as each dimension is
contiguous (unit stride).
- Fix some problems in `insert_slice_async`'s semantic.
- More general verification for ops that return shared layout encoding.

## Known Limitations

- `insert_slice_async` still uses the old semantic. May submit another
PR later to support similar semantic like `tensor.extract_slice`.
- No encoding verification for `tensor.extract_slice`.
- 3d tensor ops are broken.
- Strided accesses are not allowed.
- May cause a little performance slowdown since we are passing strides
as values but not constants (e.g., int).
It would be difficult to pass strides as attributes when we have control
flows. A block argument is possible to accept tensors with different
strides.
This commit is contained in:
Keren Zhou
2022-11-06 22:59:03 -08:00
committed by GitHub
parent a4ff0c362c
commit fdd59900f7
26 changed files with 507 additions and 339 deletions

View File

@@ -1,6 +1,7 @@
#include "triton/Analysis/Allocation.h"
#include "mlir/Analysis/Liveness.h"
#include "mlir/Analysis/SliceAnalysis.h"
#include "mlir/Dialect/Tensor/IR/Tensor.h"
#include "triton/Analysis/Alias.h"
#include "triton/Analysis/Utility.h"
#include "triton/Dialect/TritonGPU/IR/Dialect.h"
@@ -76,13 +77,13 @@ SmallVector<unsigned> getScratchConfigForReduce(triton::ReduceOp op) {
auto srcShape = srcTy.getShape();
auto axis = op.axis();
bool fast_reduce = axis == 1; // FIXME(Qingyi): The fastest-changing dimension
bool fastReduce = axis == 1; // FIXME(Qingyi): The fastest-changing dimension
SmallVector<unsigned> smemShape;
for (auto d : srcShape)
smemShape.push_back(d);
if (fast_reduce) {
if (fastReduce) {
unsigned sizeInterWarps = srcLayout.getWarpsPerCTA()[axis];
smemShape[axis] = sizeInterWarps;
} else {
@@ -123,7 +124,7 @@ private:
// For example: %a = scf.if -> yield
// %a must be allocated elsewhere by other operations.
// FIXME(Keren): extract and insert are always alias for now
if (!maybeSharedAllocationOp(op) || isa<triton::gpu::ExtractSliceOp>(op) ||
if (!maybeSharedAllocationOp(op) || isa<tensor::ExtractSliceOp>(op) ||
isa<triton::gpu::InsertSliceAsyncOp>(op)) {
return;
}