Skip to content

Reland upstream commit f9688ab #2517

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
whitneywhtsang merged 3 commits into from
Nov 4, 2024
Merged

Reland upstream commit f9688ab #2517

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
e1460b1
Revert "Revert "[BACKEND] Small fixes for dot operand properties (#48…
whitneywhtsang Oct 22, 2024
4b7942a
[intel] Small fixes for dot operand properties
whitneywhtsang Oct 22, 2024
c637c07
fix order for DPAS
anmyachev Oct 21, 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)>,

InterfaceMethod<"Return element sizes per thread for dot operands.", "SmallVector<unsigned>",
"getElemsPerThreadForOperands", (ins "ArrayRef<int64_t>":$tensorShape,
Expand Down Expand Up @@ -919,11 +921,11 @@ 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 @@ -1030,12 +1032,12 @@ 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 @@ -1235,18 +1237,18 @@ 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 @@ -1361,7 +1363,16 @@ 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");
if (getOpIdx() == 0)
contigPerThread[rank - 1] = kWidth;
else
contigPerThread[rank - 2] = kWidth;
return contigPerThread;
};
}];
}
Expand Down
Loading
Loading