Skip to content

[BACKEND] Minor Bugfixes for SharedToDotOperand MMAv3 #5030

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
lezcano merged 4 commits into from
Nov 4, 2024
Merged

[BACKEND] Minor Bugfixes for SharedToDotOperand MMAv3 #5030

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
1d738e1
Fix order for Hopper mmaLayout
ggengnv Oct 31, 2024
d874589
Fix bad rebase: use mmaBitwidth to compute numRep
ggengnv Nov 1, 2024
c2821a8
Revert getOrder and use getWarpOrder in SharedToDot
ggengnv Nov 1, 2024
246dac8
Get mmaLayout's order instead of dotOp's
ggengnv Nov 1, 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
6 changes: 3 additions & 3 deletions ...ty/nvidia/lib/TritonNVIDIAGPUToLLVM/ConvertLayoutOpToLLVM/SharedToDotOperandMMAv2OrV3.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -659,15 +659,15 @@ Value loadArg(ConversionPatternRewriter &rewriter, Location loc,

int kWidth = encoding.getKWidth();
auto numRep = mmaLayout.getMMAv2OrV3RepForOperand(
shapePerCTA, bitwidth, kWidth, encoding.getOpIdx());
shapePerCTA, mmaBitwidth, kWidth, encoding.getOpIdx());

auto warpsPerCTA = mmaLayout.getWarpsPerCTA();
auto order = triton::gpu::getOrder(mmaLayout);
auto warpOrder = mmaLayout.getWarpOrder();
Comment on lines 664 to +665
Copy link
Copy Markdown
Contributor

@lezcano lezcano Nov 2, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What I don't quite understand is why you need the warp order and warps of the mma layout rather than reasoning about the layout from the dot operand, which is the one that you are moving around.

At any rate, this code will be removed in a months time, so probably not worth the time investment. Might be better to think how to fix the mmav3 pipeliner when one of the operands comes from a register.

Value warp = udiv(thread, i32_val(32));
Value lane = urem(thread, i32_val(32));

SmallVector<Value> multiDimWarpId =
delinearize(rewriter, loc, warp, warpsPerCTA, order);
delinearize(rewriter, loc, warp, warpsPerCTA, warpOrder);
Value warpB = urem(multiDimWarpId[0], i32_val(shapePerCTA[0]));
int warpsPerTile;
Value warpM = urem(multiDimWarpId[1], i32_val(shapePerCTA[1] / 16));
Expand Down