Skip to content

[BACKEND] Small fixes for dot operand properties #4895

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Jokeren merged 21 commits into from
Oct 15, 2024
Merged

[BACKEND] Small fixes for dot operand properties #4895

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
21 commits
Select commit Hold shift + click to select a range
3bf9218
Update
Jokeren Oct 11, 2024
c7ee428
Update
Jokeren Oct 11, 2024
5428fb2
Update
Jokeren Oct 11, 2024
55bae07
Update
Jokeren Oct 11, 2024
5a8a9ec
Update
Jokeren Oct 11, 2024
dc9588f
Update Dialect.cpp
Jokeren Oct 13, 2024
a3ef441
Update Dialect.cpp
Jokeren Oct 13, 2024
f9159e9
Update
Jokeren Oct 15, 2024
7789dde
Update
Jokeren Oct 15, 2024
78061cb
Update
Jokeren Oct 15, 2024
539e815
Update
Jokeren Oct 15, 2024
e01a3ca
Update
Jokeren Oct 15, 2024
8f4f760
Update
Jokeren Oct 15, 2024
d5b26f2
Update
Jokeren Oct 15, 2024
dfa69e4
Update
Jokeren Oct 15, 2024
7e90ac8
Update
Jokeren Oct 15, 2024
5706933
Update comments
Jokeren Oct 15, 2024
b4856ae
Update
Jokeren Oct 15, 2024
17378ed
Update
Jokeren Oct 15, 2024
d7ea5d3
Update
Jokeren Oct 15, 2024
e264941
Merge branch 'main' into keren/dot-op-fix
Jokeren Oct 15, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
57 changes: 34 additions & 23 deletions include/triton/Dialect/TritonGPU/IR/TritonGPUAttrDefs.td
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -781,22 +781,24 @@ def MmaEncodingTrait : AttrInterface<"MmaEncodingTrait"> {

InterfaceMethod<"Return shape per CTA.",
"SmallVector<unsigned>",
"getShapePerCTATileForDotOperands",
"getShapePerCTATileForOperand",
(ins "ArrayRef<int64_t>":$tensorShape,
"unsigned":$opIdx)>,
"int":$kWidth,
"int":$opIdx)>,

InterfaceMethod<"Return total element size per thread for dot operands.",
"unsigned",
"getTotalElemsPerThreadForOperands",
"getTotalElemsPerThreadForOperand",
(ins "ArrayRef<int64_t>":$tensorShape,
"Type":$eltTy,
"unsigned":$kWidth,
"unsigned":$opIdx)>,
"int":$kWidth,
"int":$opIdx)>,

InterfaceMethod<"Return size per thread for dot operands.",
"SmallVector<unsigned>",
"getSizePerThreadForOperands",
(ins "unsigned":$opIdx)>,
"getSizePerThreadForOperand",
(ins "int":$opIdx,
"int":$kWidth)>,
];
}

Expand Down Expand Up @@ -914,11 +916,11 @@ V [ 0,4,8...60 1,5...61 2,6...62 3,7...63 ] [ 128,132...188 129,
bool supportReduction() const {
return true;
}
SmallVector<unsigned> getSizePerThreadForOperands(unsigned opIdx) const;
SmallVector<unsigned> getShapePerCTATileForDotOperands(ArrayRef<int64_t> shape, int opIdx) const;
unsigned getTotalElemsPerThreadForOperands(ArrayRef<int64_t> shape, Type eltTy, int kWidth, int opIdx) const;
SmallVector<int64_t> getMFMAInstrShapeForOperands(int kWidth, int opIdx) const;
SmallVector<int64_t> getMFMARepForOperands(ArrayRef<int64_t> operandShape, int kWidth, int opIdx) const;
SmallVector<unsigned> getSizePerThreadForOperand(int kWidth, int opIdx) const;
SmallVector<unsigned> getShapePerCTATileForOperand(ArrayRef<int64_t> shape, int kWidth, int opIdx) const;
unsigned getTotalElemsPerThreadForOperand(ArrayRef<int64_t> shape, Type eltTy, int kWidth, int opIdx) const;
SmallVector<int64_t> getInstrShapeForOperand(int kWidth, int opIdx) const;
SmallVector<int64_t> getRepForOperand(ArrayRef<int64_t> operandShape, int kWidth, int opIdx) const;

SmallVector<unsigned> getContigPerThread() {
auto rank = getWarpsPerCTA().size();
Expand Down Expand Up @@ -1021,12 +1023,12 @@ Row | warp 0 warp 2
bool supportReduction() const {
return true;
}
SmallVector<unsigned> getSizePerThreadForOperands(unsigned opIdx) const;
SmallVector<unsigned> getShapePerCTATileForDotOperands(ArrayRef<int64_t> shape, int opIdx) const;
unsigned getTotalElemsPerThreadForOperands(ArrayRef<int64_t> shape, Type eltTy, int kWidth, int opIdx) const;
SmallVector<unsigned> getSizePerThreadForOperand(int kWidth, int opIdx) const;
SmallVector<unsigned> getShapePerCTATileForOperand(ArrayRef<int64_t> shape, int kWidth, int opIdx) const;
unsigned getTotalElemsPerThreadForOperand(ArrayRef<int64_t> shape, Type eltTy, int kWidth, int opIdx) const;
SmallVector<int64_t> getElemsPerInstrForOperands() const;
SmallVector<int64_t> getRepForOperands(ArrayRef<int64_t> operandShape,
Type elemType, int kWidth, int opIdx) const;
SmallVector<int64_t> getRepForOperand(ArrayRef<int64_t> operandShape,
Type elemType, int kWidth, int opIdx) const;
static SmallVector<unsigned> getMNKDimPerInstr();

SmallVector<unsigned> getContigPerThread() {
Expand Down Expand Up @@ -1222,18 +1224,18 @@ For example, the matrix L corresponding to blockTileSize=[32,16] is:
SmallVector<int> getMMAv1Rep(int opIdx) const;
SmallVector<int> getMMAv1ShapePerWarp(int opIdx) const;
int getMMAv1Vec(int opIdx) const;
SmallVector<int64_t> getMMAv2Rep(ArrayRef<int64_t> shape,
int bitwidth, int opIdx) const;
SmallVector<int64_t> getMMAv2RepForOperand(ArrayRef<int64_t> shape,
int bitwidth, int kWidth, int opIdx) const;

bool supportReduction() const {
if (isAmpere() || isHopper()) {
return true;
}
return false;
};
SmallVector<unsigned> getSizePerThreadForOperands(unsigned opIdx) const;
SmallVector<unsigned> getShapePerCTATileForDotOperands(ArrayRef<int64_t> shape, int opIdx) const;
unsigned getTotalElemsPerThreadForOperands(ArrayRef<int64_t> shape, Type eltTy, int kWidth, int opIdx) const;
SmallVector<unsigned> getSizePerThreadForOperand(int kWidth, int opIdx) const;
SmallVector<unsigned> getShapePerCTATileForOperand(ArrayRef<int64_t> shape, int kWidth, int opIdx) const;
unsigned getTotalElemsPerThreadForOperand(ArrayRef<int64_t> shape, Type eltTy, int kWidth, int opIdx) const;

SmallVector<unsigned> getContigPerThread() {
assert(isVolta() || isAmpere() || isHopper());
Expand Down Expand Up @@ -1344,7 +1346,16 @@ elements along the K dim, or they use all elements of the tensor along the K dim
let genVerifyDecl = 1;
let extraClassDeclaration = extraDistributedDeclaration # [{
SmallVector<unsigned> getContigPerThread() {
return getSizePerThread();
auto rank = getWarpsPerCTA().size();
assert(rank == 2 || rank == 3);
SmallVector<unsigned> contigPerThread(rank, 1);
auto kWidth = getKWidth();
assert(kWidth != 0 && "Do not support kWidth=0");
Comment thread
lezcano marked this conversation as resolved.
if (getOpIdx() == 0)
contigPerThread[rank - 1] = kWidth;
else
contigPerThread[rank - 2] = kWidth;
return contigPerThread;
};
}];
}
Expand Down
Loading