Skip to content

[FXML-5890] Order tiling worklist #532

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
cferry-AMD merged 16 commits into from
May 6, 2025
Merged

[FXML-5890] Order tiling worklist #532

Changes from 1 commit
Commits
Show all changes
16 commits
Select commit Hold shift + click to select a range
1c64247
WIP: worklist traversal
cferry-AMD Apr 25, 2025
7dcfd68
Merge remote-tracking branch 'origin/feature/fused-ops' into corentin…
cferry-AMD Apr 28, 2025
d20bb09
Add a test
cferry-AMD Apr 29, 2025
998575f
Make insert_worklist function generic
cferry-AMD Apr 29, 2025
5bf0b8b
Change wording: tiling -> fusion
cferry-AMD Apr 29, 2025
a6fd34e
Remove remark, replace by attribute -> more compact test
cferry-AMD Apr 30, 2025
9af785f
Use llvm::dbgs() + FileCheck
cferry-AMD Apr 30, 2025
5402418
Apply suggestions from code review
cferry-AMD Apr 30, 2025
9a5f9c6
Merge remote-tracking branch 'origin/feature/fused-ops' into corentin…
cferry-AMD May 5, 2025
f1b7413
Remove reverse worklist test
cferry-AMD May 5, 2025
37a8962
Apply suggestions from code review
cferry-AMD May 5, 2025
efe04e4
Add op with tiling_priority
cferry-AMD May 5, 2025
01b0345
Add test swapping powf operands
cferry-AMD May 5, 2025
fd7c6b9
Description did not save
cferry-AMD May 5, 2025
e064275
revise test comments
cferry-AMD May 5, 2025
0ccb35d
Typo in test
cferry-AMD May 6, 2025
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
21 changes: 8 additions & 13 deletions mlir/test/Dialect/Linalg/tile-sort.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,17 +1,16 @@
// RUN: mlir-opt %s -transform-interpreter -split-input-file -debug-only=tile-using-interface 2>&1 | FileCheck %s

func.func @tile_order_ceil_then_negf(%arg: tensor<256xf32>) -> tensor<256xf32> {
// Worklist:
// * Tiling linalg.powf -> {linalg.ceil (priority = 0), linalg.negf (priority = 1)}
// * Tiling linalg.ceil -> {linalg.negf (priority = 1)}
// * Tiling linalg.negf -> {linalg.ceil (priority = 0)}
// Ops are tiled by lower priority: linalg.powf, linalg.ceil (1st operand of powf, priority = 0),
// linalg.negf (2nd oprand of powf, priority = 1), linalg.ceil (operand of negf, priority = 0)
%empty = tensor.empty() : tensor<256xf32>
%0 = linalg.ceil {tiling_priority = 0 : i64} ins(%arg: tensor<256xf32>) outs(%empty: tensor<256xf32>) -> tensor<256xf32>
%empty1 = tensor.empty() : tensor<256xf32>
%1 = linalg.negf {tiling_priority = 1 : i64} ins(%0 : tensor<256xf32>) outs(%empty1: tensor<256xf32>) -> tensor<256xf32>
%empty2 = tensor.empty() : tensor<256xf32>
%2 = linalg.powf {tile} ins(%0, %1: tensor<256xf32>, tensor<256xf32>) outs(%empty2: tensor<256xf32>) -> tensor<256xf32>

// The order of these checks is the order in which the ops are actually tiled.
// CHECK: worklist: producer is %{{.*}} = linalg.ceil
// CHECK: worklist: producer is %{{.*}} = linalg.negf
// CHECK: worklist: producer is %{{.*}} = linalg.ceil
Expand All @@ -30,10 +29,8 @@ module attributes {transform.with_named_sequence} {
// -----

func.func @tile_negf_then_ceil(%arg: tensor<256xf32>) -> tensor<256xf32> {
// Worklist:
// * Tiling linalg.powf -> {linalg.negf (priority = 0), linalg.ceil (priority = 1)}
// * Tiling linalg.negf -> {linalg.ceil (priority = 1), linalg.ceil (priority = 1)}
// * Tiling linalg.ceil -> {linalg.ceil (priority = 1)}
// Ops are tiled by lower priority: linalg.powf, linalg.negf (2nd operand of powf, priority = 0),
// linalg.ceil (1st oprand of powf, priority = 1), linalg.ceil (operand of negf, priority = 1)
%empty = tensor.empty() : tensor<256xf32>
%0 = linalg.ceil {tiling_priority = 1 : i64} ins(%arg: tensor<256xf32>) outs(%empty: tensor<256xf32>) -> tensor<256xf32>
%empty1 = tensor.empty() : tensor<256xf32>
Expand All @@ -59,11 +56,9 @@ module attributes {transform.with_named_sequence} {
// -----

func.func @tile_negf_then_ceil_swap_in_powf(%arg: tensor<256xf32>) -> tensor<256xf32> {
// This gives the same tiling order regardless of the operand order in the powf
// Worklist:
// * Tiling linalg.powf -> {linalg.negf (priority = 0), linalg.ceil (priority = 1)}
// * Tiling linalg.negf -> {linalg.ceil (priority = 1), linalg.ceil (priority = 1)}
// * Tiling linalg.ceil -> {linalg.ceil (priority = 1)}
// This gives the same tiling order as above regardless of the operand order in the powf
// linalg.powf, linalg.negf (1st operand of powf, priority = 0),
// linalg.ceil (2nd oprand of powf, priority = 1), linalg.ceil (operand of negf, priority = 1)
%empty = tensor.empty() : tensor<256xf32>
%0 = linalg.ceil {tiling_priority = 1 : i64} ins(%arg: tensor<256xf32>) outs(%empty: tensor<256xf32>) -> tensor<256xf32>
%empty1 = tensor.empty() : tensor<256xf32>
Expand Down