[mlir][linalg] Allow pack consumer fusion if the tile size is greater than dimension size. by hanhanW · Pull Request #149438 · llvm/llvm-project

hanhanW · 2025-07-18T02:07:23Z

This happens only when you use larger tile size, which is greater than or equal to the dimension size. In this case, it is a full slice, so it is fusible.

The IR can be generated during the TileAndFuse process. It is hard to fix in such driver, so we enable the naive fusion for the case.

llvmbot · 2025-07-18T02:07:52Z

@llvm/pr-subscribers-mlir

Author: Han-Chung Wang (hanhanW)

Changes

This happens only when you use larger tile size, which is greater than or equal to the dimension size. In this case, it is a full slice, so it is fusible.

The IR can be generated during the TileAndFuse process. It is hard to fix in such driver, so we enable the naive fusion for the case.

Full diff: https://github.com/llvm/llvm-project/pull/149438.diff

2 Files Affected:

(modified) mlir/lib/Dialect/Linalg/Transforms/TilingInterfaceImpl.cpp (+1-1)
(modified) mlir/test/Interfaces/TilingInterface/tile-and-fuse-consumer.mlir (+51)

diff --git a/mlir/lib/Dialect/Linalg/Transforms/TilingInterfaceImpl.cpp b/mlir/lib/Dialect/Linalg/Transforms/TilingInterfaceImpl.cpp
index 5a10883a6043c..eed431570452a 100644
--- a/mlir/lib/Dialect/Linalg/Transforms/TilingInterfaceImpl.cpp
+++ b/mlir/lib/Dialect/Linalg/Transforms/TilingInterfaceImpl.cpp
@@ -911,7 +911,7 @@ struct PackOpTiling
         int64_t destDimSize = packOp.getDestType().getDimSize(dim);
         bool isTiled = failed(cstTileSize) ||
                        ShapedType::isDynamic(srcDimSize) ||
-                       cstTileSize.value() != srcDimSize;
+                       cstTileSize.value() < srcDimSize;
         if (!isTiled) {
           outerDimOffsets.push_back(offsets[dim]);
           if (ShapedType::isStatic(destDimSize)) {
diff --git a/mlir/test/Interfaces/TilingInterface/tile-and-fuse-consumer.mlir b/mlir/test/Interfaces/TilingInterface/tile-and-fuse-consumer.mlir
index 7b0a8494a8acb..e9465baac7509 100644
--- a/mlir/test/Interfaces/TilingInterface/tile-and-fuse-consumer.mlir
+++ b/mlir/test/Interfaces/TilingInterface/tile-and-fuse-consumer.mlir
@@ -451,6 +451,57 @@ module attributes {transform.with_named_sequence} {
 
 // -----
 
+#map = affine_map<(d0) -> (-d0 + 4, 16)>
+func.func @fuse_pack_consumer_if_single_iteration(%arg0: tensor<4x4xf32>) -> tensor<1x4x16x1xf32> {
+  %0 = tensor.empty() : tensor<1x4x16x1xf32>
+  %1 = tensor.empty() : tensor<4x4xf32>
+  %2 = scf.forall (%arg1) = (0) to (4) step (16) shared_outs(%arg2 = %1) -> (tensor<4x4xf32>) {
+    %3 = affine.min #map(%arg1)
+    %extracted_slice = tensor.extract_slice %arg0[%arg1, 0] [%3, 4] [1, 1] : tensor<4x4xf32> to tensor<?x4xf32>
+    %extracted_slice_0 = tensor.extract_slice %arg2[%arg1, 0] [%3, 4] [1, 1] : tensor<4x4xf32> to tensor<?x4xf32>
+    %4 = linalg.exp ins(%extracted_slice : tensor<?x4xf32>) outs(%extracted_slice_0 : tensor<?x4xf32>) -> tensor<?x4xf32>
+    scf.forall.in_parallel {
+      tensor.parallel_insert_slice %4 into %arg2[%arg1, 0] [%3, 4] [1, 1] : tensor<?x4xf32> into tensor<4x4xf32>
+    }
+  }
+  %cst = arith.constant 0.000000e+00 : f32
+  %pack = linalg.pack %2 padding_value(%cst : f32) outer_dims_perm = [0, 1] inner_dims_pos = [0, 1] inner_tiles = [16, 1] into %0 : tensor<4x4xf32> -> tensor<1x4x16x1xf32>
+  return %pack : tensor<1x4x16x1xf32>
+}
+
+module attributes {transform.with_named_sequence} {
+  transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
+    %0 = transform.structured.match ops{["tensor.parallel_insert_slice"]} in %arg0 : (!transform.any_op) -> !transform.any_op
+    %1 = transform.structured.match ops{["scf.forall"]} in %arg0 : (!transform.any_op) -> !transform.any_op
+    %consumer, %fused_consumer = transform.test.fuse_consumer %0 in(%1) : (!transform.any_op, !transform.any_op) -> (!transform.any_op, !transform.any_op)
+    transform.yield
+  }
+}
+//      CHECK: #[[MAP:.*]] = affine_map<(d0) -> (-d0 + 4, 16)>
+//      CHECK: func.func @fuse_pack_consumer_if_single_iteration(
+// CHECK-SAME:     %[[ARG0:[a-zA-Z0-9]+]]
+//  CHECK-DAG:   %[[PACK_INIT:.*]] = tensor.empty() : tensor<1x4x16x1xf32>
+//  CHECK-DAG:   %[[ELEM_INIT:.*]] = tensor.empty() : tensor<4x4xf32>
+//  CHECK-DAG:   %[[PAD_VAL:.*]] = arith.constant 0.000000e+00 : f32
+//      CHECK:   %{{.*}}:2 = scf.forall (%[[IV:.*]]) = (0) to (4) step (16)
+// CHECK-SAME:      shared_outs(%[[ELEM_OUT_ARG:.*]] = %[[ELEM_INIT]], %[[PACK_OUT_ARG:.*]] = %[[PACK_INIT]])
+//  CHECK-DAG:      %[[SIZE:.+]] = affine.min #[[MAP]](%[[IV]])
+//  CHECK-DAG:      %[[ELEM_SRC:.*]] = tensor.extract_slice %[[ARG0]][%[[IV]], 0] [%[[SIZE]], 4] [1, 1]
+//  CHECK-DAG:      %[[ELEM_DEST:.*]] = tensor.extract_slice %[[ELEM_OUT_ARG]][%[[IV]], 0] [%[[SIZE]], 4] [1, 1]
+//      CHECK:      %[[ELEM:.*]] = linalg.exp
+// CHECK-SAME:        ins(%[[ELEM_SRC]]
+// CHECK-SAME:        outs(%[[ELEM_DEST]]
+//  CHECK-DAG:      %[[TILED_PACK_DEST:.*]] = tensor.extract_slice %[[PACK_OUT_ARG]][%[[IV]], 0, 0, 0] [1, 4, 16, 1] [1, 1, 1, 1]
+//      CHECK:      %[[PACK:.*]] = linalg.pack %[[ELEM]]
+// CHECK-SAME:        padding_value(%[[PAD_VAL]] : f32)
+// CHECK-SAME:        outer_dims_perm = [0, 1] inner_dims_pos = [0, 1] inner_tiles = [16, 1]
+// CHECK-SAME:        into %[[TILED_PACK_DEST]]
+//      CHECK:      scf.forall.in_parallel {
+//      CHECK:          tensor.parallel_insert_slice %[[ELEM]] into %[[ELEM_OUT_ARG]][%[[IV]], 0] [%[[SIZE]], 4] [1, 1]
+//      CHECK:          tensor.parallel_insert_slice %[[PACK]] into %[[PACK_OUT_ARG]][%[[IV]], 0, 0, 0] [1, 4, 16, 1] [1, 1, 1, 1]
+
+// -----
+
 // It is valid to fuse the pack op in perfect tiling scenario when the dimension
 // is dynamic and padding is not needed.

llvmbot · 2025-07-18T02:07:54Z

@llvm/pr-subscribers-mlir-linalg

Author: Han-Chung Wang (hanhanW)

Changes

This happens only when you use larger tile size, which is greater than or equal to the dimension size. In this case, it is a full slice, so it is fusible.

The IR can be generated during the TileAndFuse process. It is hard to fix in such driver, so we enable the naive fusion for the case.

Full diff: https://github.com/llvm/llvm-project/pull/149438.diff

2 Files Affected:

(modified) mlir/lib/Dialect/Linalg/Transforms/TilingInterfaceImpl.cpp (+1-1)
(modified) mlir/test/Interfaces/TilingInterface/tile-and-fuse-consumer.mlir (+51)

diff --git a/mlir/lib/Dialect/Linalg/Transforms/TilingInterfaceImpl.cpp b/mlir/lib/Dialect/Linalg/Transforms/TilingInterfaceImpl.cpp
index 5a10883a6043c..eed431570452a 100644
--- a/mlir/lib/Dialect/Linalg/Transforms/TilingInterfaceImpl.cpp
+++ b/mlir/lib/Dialect/Linalg/Transforms/TilingInterfaceImpl.cpp
@@ -911,7 +911,7 @@ struct PackOpTiling
         int64_t destDimSize = packOp.getDestType().getDimSize(dim);
         bool isTiled = failed(cstTileSize) ||
                        ShapedType::isDynamic(srcDimSize) ||
-                       cstTileSize.value() != srcDimSize;
+                       cstTileSize.value() < srcDimSize;
         if (!isTiled) {
           outerDimOffsets.push_back(offsets[dim]);
           if (ShapedType::isStatic(destDimSize)) {
diff --git a/mlir/test/Interfaces/TilingInterface/tile-and-fuse-consumer.mlir b/mlir/test/Interfaces/TilingInterface/tile-and-fuse-consumer.mlir
index 7b0a8494a8acb..e9465baac7509 100644
--- a/mlir/test/Interfaces/TilingInterface/tile-and-fuse-consumer.mlir
+++ b/mlir/test/Interfaces/TilingInterface/tile-and-fuse-consumer.mlir
@@ -451,6 +451,57 @@ module attributes {transform.with_named_sequence} {
 
 // -----
 
+#map = affine_map<(d0) -> (-d0 + 4, 16)>
+func.func @fuse_pack_consumer_if_single_iteration(%arg0: tensor<4x4xf32>) -> tensor<1x4x16x1xf32> {
+  %0 = tensor.empty() : tensor<1x4x16x1xf32>
+  %1 = tensor.empty() : tensor<4x4xf32>
+  %2 = scf.forall (%arg1) = (0) to (4) step (16) shared_outs(%arg2 = %1) -> (tensor<4x4xf32>) {
+    %3 = affine.min #map(%arg1)
+    %extracted_slice = tensor.extract_slice %arg0[%arg1, 0] [%3, 4] [1, 1] : tensor<4x4xf32> to tensor<?x4xf32>
+    %extracted_slice_0 = tensor.extract_slice %arg2[%arg1, 0] [%3, 4] [1, 1] : tensor<4x4xf32> to tensor<?x4xf32>
+    %4 = linalg.exp ins(%extracted_slice : tensor<?x4xf32>) outs(%extracted_slice_0 : tensor<?x4xf32>) -> tensor<?x4xf32>
+    scf.forall.in_parallel {
+      tensor.parallel_insert_slice %4 into %arg2[%arg1, 0] [%3, 4] [1, 1] : tensor<?x4xf32> into tensor<4x4xf32>
+    }
+  }
+  %cst = arith.constant 0.000000e+00 : f32
+  %pack = linalg.pack %2 padding_value(%cst : f32) outer_dims_perm = [0, 1] inner_dims_pos = [0, 1] inner_tiles = [16, 1] into %0 : tensor<4x4xf32> -> tensor<1x4x16x1xf32>
+  return %pack : tensor<1x4x16x1xf32>
+}
+
+module attributes {transform.with_named_sequence} {
+  transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
+    %0 = transform.structured.match ops{["tensor.parallel_insert_slice"]} in %arg0 : (!transform.any_op) -> !transform.any_op
+    %1 = transform.structured.match ops{["scf.forall"]} in %arg0 : (!transform.any_op) -> !transform.any_op
+    %consumer, %fused_consumer = transform.test.fuse_consumer %0 in(%1) : (!transform.any_op, !transform.any_op) -> (!transform.any_op, !transform.any_op)
+    transform.yield
+  }
+}
+//      CHECK: #[[MAP:.*]] = affine_map<(d0) -> (-d0 + 4, 16)>
+//      CHECK: func.func @fuse_pack_consumer_if_single_iteration(
+// CHECK-SAME:     %[[ARG0:[a-zA-Z0-9]+]]
+//  CHECK-DAG:   %[[PACK_INIT:.*]] = tensor.empty() : tensor<1x4x16x1xf32>
+//  CHECK-DAG:   %[[ELEM_INIT:.*]] = tensor.empty() : tensor<4x4xf32>
+//  CHECK-DAG:   %[[PAD_VAL:.*]] = arith.constant 0.000000e+00 : f32
+//      CHECK:   %{{.*}}:2 = scf.forall (%[[IV:.*]]) = (0) to (4) step (16)
+// CHECK-SAME:      shared_outs(%[[ELEM_OUT_ARG:.*]] = %[[ELEM_INIT]], %[[PACK_OUT_ARG:.*]] = %[[PACK_INIT]])
+//  CHECK-DAG:      %[[SIZE:.+]] = affine.min #[[MAP]](%[[IV]])
+//  CHECK-DAG:      %[[ELEM_SRC:.*]] = tensor.extract_slice %[[ARG0]][%[[IV]], 0] [%[[SIZE]], 4] [1, 1]
+//  CHECK-DAG:      %[[ELEM_DEST:.*]] = tensor.extract_slice %[[ELEM_OUT_ARG]][%[[IV]], 0] [%[[SIZE]], 4] [1, 1]
+//      CHECK:      %[[ELEM:.*]] = linalg.exp
+// CHECK-SAME:        ins(%[[ELEM_SRC]]
+// CHECK-SAME:        outs(%[[ELEM_DEST]]
+//  CHECK-DAG:      %[[TILED_PACK_DEST:.*]] = tensor.extract_slice %[[PACK_OUT_ARG]][%[[IV]], 0, 0, 0] [1, 4, 16, 1] [1, 1, 1, 1]
+//      CHECK:      %[[PACK:.*]] = linalg.pack %[[ELEM]]
+// CHECK-SAME:        padding_value(%[[PAD_VAL]] : f32)
+// CHECK-SAME:        outer_dims_perm = [0, 1] inner_dims_pos = [0, 1] inner_tiles = [16, 1]
+// CHECK-SAME:        into %[[TILED_PACK_DEST]]
+//      CHECK:      scf.forall.in_parallel {
+//      CHECK:          tensor.parallel_insert_slice %[[ELEM]] into %[[ELEM_OUT_ARG]][%[[IV]], 0] [%[[SIZE]], 4] [1, 1]
+//      CHECK:          tensor.parallel_insert_slice %[[PACK]] into %[[PACK_OUT_ARG]][%[[IV]], 0, 0, 0] [1, 4, 16, 1] [1, 1, 1, 1]
+
+// -----
+
 // It is valid to fuse the pack op in perfect tiling scenario when the dimension
 // is dynamic and padding is not needed.

egebeysel

LGTM, just one comment :)

egebeysel · 2025-07-18T05:53:43Z

A comment here explaining the < would be nice, something like what you have in the description.

egebeysel · 2025-07-18T05:58:30Z

Q: wouldn't/shouldn't this (and in general one-iteration loops) be folded away? Probably it should happen at a different point/place, but still just wondering :)

If this is scf.for, we can fold them away. What I'm not sure is the case that distribution mapping attributes are present. I think the semantic is that you'll need to distribute the computation using the core_id/thread_id. It may be correct to still fold it away though.

adam-smnk

Looks good

… than dimension size. This only happens when you use tile size which is greater than or equal to the dimension size. In this case, it is a full slice, so it is fusiable. The IR can be generated during the TileAndFuse process. It is hard to fix in such driver, so we enable the naive fusion for the case. Signed-off-by: hanhanW <hanhan0912@gmail.com>

Signed-off-by: hanhanW <hanhan0912@gmail.com>

… than dimension size. (llvm#149438) This happens only when you use larger tile size, which is greater than or equal to the dimension size. In this case, it is a full slice, so it is fusible. The IR can be generated during the TileAndFuse process. It is hard to fix in such driver, so we enable the naive fusion for the case. --------- Signed-off-by: hanhanW <hanhan0912@gmail.com>

hanhanW requested review from MaheshRavishankar, adam-smnk and egebeysel July 18, 2025 02:07

hanhanW requested review from dcaballe, nicolasvasilache and rengolin as code owners July 18, 2025 02:07

llvmbot added mlir:linalg mlir labels Jul 18, 2025

egebeysel approved these changes Jul 18, 2025

View reviewed changes

adam-smnk approved these changes Jul 18, 2025

View reviewed changes

adam-smnk reviewed Jul 18, 2025

View reviewed changes

Comment thread mlir/lib/Dialect/Linalg/Transforms/TilingInterfaceImpl.cpp Outdated

hanhanW added 3 commits July 18, 2025 10:00

add a comment

4c992de

Signed-off-by: hanhanW <hanhan0912@gmail.com>

correct the comment

6c51a22

Signed-off-by: hanhanW <hanhan0912@gmail.com>

hanhanW force-pushed the allow-fake-tiling branch from cd4089d to 6c51a22 Compare July 18, 2025 17:00

hanhanW merged commit 3ea6da5 into llvm:main Jul 18, 2025
9 checks passed

hanhanW deleted the allow-fake-tiling branch July 18, 2025 17:42

hanhanW mentioned this pull request Jul 18, 2025

[CPU] Switch CPUDoubleTilingExpert pipeline to use IREE::CPU::LoweringConfigAttr. iree-org/iree#21354

Merged

This was referenced Jul 23, 2025

test abhinavgaba/llvm-project#2

Closed

Add dataFence plugin interface abhinavgaba/llvm-project#3

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[mlir][linalg] Allow pack consumer fusion if the tile size is greater than dimension size.#149438

[mlir][linalg] Allow pack consumer fusion if the tile size is greater than dimension size.#149438
hanhanW merged 3 commits into
llvm:mainfrom
hanhanW:allow-fake-tiling

hanhanW commented Jul 18, 2025

Uh oh!

llvmbot commented Jul 18, 2025

Uh oh!

llvmbot commented Jul 18, 2025

Uh oh!

egebeysel left a comment

Uh oh!

egebeysel Jul 18, 2025

Uh oh!

egebeysel Jul 18, 2025

Uh oh!

hanhanW Jul 18, 2025

Uh oh!

adam-smnk left a comment

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

hanhanW commented Jul 18, 2025

Uh oh!

llvmbot commented Jul 18, 2025

Uh oh!

llvmbot commented Jul 18, 2025

Uh oh!

egebeysel left a comment

Choose a reason for hiding this comment

Uh oh!

egebeysel Jul 18, 2025

Choose a reason for hiding this comment

Uh oh!

egebeysel Jul 18, 2025

Choose a reason for hiding this comment

Uh oh!

hanhanW Jul 18, 2025

Choose a reason for hiding this comment

Uh oh!

adam-smnk left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants