Skip to content
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

Allow Layouts to propogate to local_load #5219

Merged
merged 4 commits into from
Nov 21, 2024
Merged

Allow Layouts to propogate to local_load #5219

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
15a76c4
Allow Layouts to propogate to local_load
mbrookhart Nov 21, 2024
f2abd1c
Improve lit test
mbrookhart Nov 21, 2024
d703e4b
Update test/TritonGPU/combine.mlir
mbrookhart Nov 21, 2024
708b65f
Update test/TritonGPU/combine.mlir
mbrookhart Nov 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
3 changes: 2 additions & 1 deletion lib/Dialect/TritonGPU/Transforms/Utility.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -563,7 +563,8 @@ bool canFoldIntoConversion(Operation *op, Attribute targetEncoding) {
}
return isa<triton::gpu::ConvertLayoutOp, arith::ConstantOp,
triton::MakeRangeOp, triton::SplatOp, triton::HistogramOp,
triton::gpu::LocalAllocOp, triton::gpu::LocalStoreOp>(op);
triton::gpu::LocalAllocOp, triton::gpu::LocalLoadOp,
triton::gpu::LocalStoreOp>(op);
}

scf::ForOp replaceForOpWithNewSignature(
Expand Down
18 changes: 18 additions & 0 deletions test/TritonGPU/combine.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2685,3 +2685,21 @@ module attributes {"triton_gpu.num-warps" = 1 : i32, "triton_gpu.threads-per-war
tt.return
}
}

// -----

#blocked = #triton_gpu.blocked<{sizePerThread = [1, 1, 1, 1, 4], threadsPerWarp = [2, 1, 16, 1, 1], warpsPerCTA = [1, 1, 2, 2, 1], order = [4, 0, 1, 2, 3]}>
#blocked2 = #triton_gpu.blocked<{sizePerThread = [1, 1, 1, 1, 4], threadsPerWarp = [1, 1, 32, 1, 1], warpsPerCTA = [1, 1, 1, 1, 4], order = [4, 3, 2, 1, 0]}>
#blocked1 = #triton_gpu.blocked<{sizePerThread = [1, 1, 1, 1, 4], threadsPerWarp = [2, 1, 16, 1, 1], warpsPerCTA = [1, 2, 2, 1, 1], order = [4, 0, 3, 2, 1]}>
#shared = #triton_gpu.shared<{vec = 1, perPhase = 1, maxPhase = 1, order = [4, 0, 1, 2, 3], hasLeadingOffset = false}>
module attributes {"triton_gpu.num-ctas" = 1 : i32, "triton_gpu.num-warps" = 4 : i32, triton_gpu.target = "cuda:100", "triton_gpu.threads-per-warp" = 32 : i32} {
// CHECK-LABEL: lift_convert_to_local_load
// CHECK-NOT: convert_layout
mbrookhart marked this conversation as resolved.
Show resolved Hide resolved
// CHECK: tt.return
tt.func public @lift_convert_to_local_load(%arg0 : !triton_gpu.memdesc<2x1x32x4x4xi8, #shared, #triton_gpu.shared_memory, mutable>) -> tensor<2x4x32x1x4xi8, #blocked2> {
%1 = triton_gpu.local_load %arg0 : !triton_gpu.memdesc<2x1x32x4x4xi8, #shared, #triton_gpu.shared_memory, mutable> -> tensor<2x1x32x4x4xi8, #blocked>
%2 = tt.trans %1 {order = array<i32: 0, 3, 2, 1, 4>} : tensor<2x1x32x4x4xi8, #blocked> -> tensor<2x4x32x1x4xi8, #blocked1>
%3 = triton_gpu.convert_layout %2 : tensor<2x4x32x1x4xi8, #blocked1> -> tensor<2x4x32x1x4xi8, #blocked2>
tt.return %3 : tensor<2x4x32x1x4xi8, #blocked2>
}
}
Loading