[Mlir-commits] [mlir] [mlir][linalg] Refine tensor.extract vectorization (PR #99299)

Andrzej WarzyƄski llvmlistbot at llvm.org
Tue Jul 30 11:50:31 PDT 2024


https://github.com/banach-space updated https://github.com/llvm/llvm-project/pull/99299

>From 20b886a94bfb44004600b6bf6fdbffa50a6370d8 Mon Sep 17 00:00:00 2001
From: Andrzej Warzynski <andrzej.warzynski at arm.com>
Date: Wed, 17 Jul 2024 11:08:12 +0100
Subject: [PATCH] [mlir][linalg] Relax tensor.extract vectorization

Simplifies the vectorization of tensor.extract so that:
* all cases that read into a genuinely multi-dim vector (*) are
  considered a gather load,
* all other cases are considered as potential contiguous loads.

This change means that the following extraction from a "column" tensor
is correctly identified as a scalar load followed by a broadcast (rather
than a gather load).

```mlir
func.func @vectorize_scalar_broadcast_column_tensor(%in: tensor<1x1x4xi32>) -> tensor<1x1x4xi32> {
  %c4 = arith.constant 4 : index
  %c0 = arith.constant 0 : index
  %cst = arith.constant dense<[...]> : tensor<15x1xi32>

  %out = linalg.generic {
    indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
    iterator_types = ["parallel", "parallel", "parallel"]}
    outs(%in : tensor<1x1x4xi32>) {

  ^bb0(%out: i32):
    %8 = linalg.index 0 : index
    %idx_0 = linalg.index 0 : index
    %extracted = tensor.extract %cst[%idx_0, %c0] : tensor<15x1xi32>
    linalg.yield %extracted : i32
  } -> tensor<1x1x4xi32>

  return %out:tensor<1x1x4xi32>
}
```

(*) `vector<1x4x1xf32>` is considered as 1D vector in this context.
---
 .../Linalg/Transforms/Vectorization.cpp       | 35 +++++-------
 .../Linalg/vectorize-tensor-extract.mlir      | 56 +++++++++++++++++++
 2 files changed, 71 insertions(+), 20 deletions(-)

diff --git a/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp b/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
index 9185663799e52..34d9c4247b916 100644
--- a/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
+++ b/mlir/lib/Dialect/Linalg/Transforms/Vectorization.cpp
@@ -944,27 +944,22 @@ getTensorExtractMemoryAccessPattern(tensor::ExtractOp extractOp,
   if (linalgOp.hasDynamicShape())
     return VectorMemoryAccessKind::Gather;
 
-  // 1. Assume that it's a gather load when reading _into_:
-  //    * an n-D "vector", like `tensor<1x2x4xi32` or `tensor<2x1x4xi32>`, or
-  //    * a 1-D "vector" with the trailing dim equal 1, e.g. `tensor<1x4x1xi32`.
-  // TODO: Relax these conditions.
-  // FIXME: This condition assumes non-dynamic sizes.
-  if ((llvm::count_if(targetShape,
-                      [](int64_t dimSize) { return dimSize > 1; }) != 1) ||
-      targetShape.back() == 1)
-    return VectorMemoryAccessKind::Gather;
-
-  // 2. Assume that it's a gather load when reading _from_ a tensor for which
-  // the trailing dimension is 1, e.g. `tensor<1x4x1xi32>`.
-  // TODO: Relax this condition.
-  if (inputShape.getShape().back() == 1)
+  // True for vectors that are effectively 1D, e.g. `vector<1x4x1xi32>`, false
+  // otherwise.
+  bool isOutput1DVector = (llvm::count_if(targetShape, [](int64_t dimSize) {
+                             return dimSize > 1;
+                           }) == 1);
+
+  // 1. Assume that it's a gather load when reading non-1D vector.
+  if (!isOutput1DVector)
     return VectorMemoryAccessKind::Gather;
 
   bool leadingIdxsLoopInvariant = true;
 
-  // 3. Analyze the leading indices of `extractOp`.
+  // 2. Analyze the leading indices of `extractOp`.
   // Look at the way each index is calculated and decide whether it is suitable
-  // for a contiguous load, i.e. whether it's loop invariant.
+  // for a contiguous load, i.e. whether it's loop invariant. If not, it's a
+  // gather load.
   auto indices = extractOp.getIndices();
   auto leadIndices = indices.drop_back(1);
 
@@ -980,13 +975,13 @@ getTensorExtractMemoryAccessPattern(tensor::ExtractOp extractOp,
     return VectorMemoryAccessKind::Gather;
   }
 
-  // 4. Analyze the trailing index for `extractOp`.
+  // 3. Analyze the trailing index for `extractOp`.
   // At this point we know that the leading indices are loop invariant. This
   // means that is potentially a scalar or a contiguous load. We can decide
   // based on the trailing idx.
   auto extractOpTrailingIdx = indices.back();
 
-  // 4a. Scalar broadcast load
+  // 3a. Scalar broadcast load
   // If the trailing index is loop invariant then this is a scalar load.
   if (leadingIdxsLoopInvariant &&
       isLoopInvariantIdx(linalgOp, extractOpTrailingIdx)) {
@@ -995,7 +990,7 @@ getTensorExtractMemoryAccessPattern(tensor::ExtractOp extractOp,
     return VectorMemoryAccessKind::ScalarBroadcast;
   }
 
-  // 4b. Contiguous loads
+  // 3b. Contiguous loads
   // The trailing `extractOp` index should increment with every loop iteration.
   // This effectively means that it must be based on the trailing loop index.
   // This is what the following bool captures.
@@ -1009,7 +1004,7 @@ getTensorExtractMemoryAccessPattern(tensor::ExtractOp extractOp,
     return VectorMemoryAccessKind::Contiguous;
   }
 
-  // 5. Fallback case - gather load.
+  // 4. Fallback case - gather load.
   LDBG("Found gather load: " << extractOp);
   return VectorMemoryAccessKind::Gather;
 }
diff --git a/mlir/test/Dialect/Linalg/vectorize-tensor-extract.mlir b/mlir/test/Dialect/Linalg/vectorize-tensor-extract.mlir
index 85e1c56dd45a0..ac75a19cbeb28 100644
--- a/mlir/test/Dialect/Linalg/vectorize-tensor-extract.mlir
+++ b/mlir/test/Dialect/Linalg/vectorize-tensor-extract.mlir
@@ -595,3 +595,59 @@ module attributes {transform.with_named_sequence} {
      transform.yield
    }
 }
+
+
+// -----
+
+func.func @vectorize_scalar_broadcast_column_tensor(%in: tensor<1x1x4xi32>) -> tensor<1x1x4xi32> {
+  %c4 = arith.constant 4 : index
+  %c0 = arith.constant 0 : index
+  %cst = arith.constant dense<[[0], [1], [2], [3], [4], [5], [6], [7], [8], [9], [10], [11], [12], [13], [14]]> : tensor<15x1xi32>
+
+  %out = linalg.generic {indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>], iterator_types = ["parallel", "parallel", "parallel"]} outs(%in : tensor<1x1x4xi32>) {
+  ^bb0(%out: i32):
+    %8 = linalg.index 0 : index
+    %idx_0 = linalg.index 0 : index
+    %extracted = tensor.extract %cst[%idx_0, %c0] : tensor<15x1xi32>
+    linalg.yield %extracted : i32
+  } -> tensor<1x1x4xi32>
+
+  return %out:tensor<1x1x4xi32>
+}
+
+// CHECK: #[[$ATTR_1:.+]] = affine_map<(d0, d1) -> (0, 0, 0)>
+// CHECK-LABEL:   func.func @vectorize_scalar_broadcast_column_tensor(
+// CHECK-SAME:      %[[VAL_0:.*]]: tensor<1x1x4xi32>) -> tensor<1x1x4xi32> {
+// CHECK:           %[[VAL_1:.*]] = arith.constant 4 : index
+// CHECK:           %[[VAL_2:.*]] = arith.constant 0 : index
+// CHECK:           %[[VAL_3:.*]] = arith.constant dense<{{\[\[}}0], [1], [2], [3], [4], [5], [6], [7], [8], [9], [10], [11], [12], [13], [14]]> : tensor<15x1xi32>
+// CHECK:           %[[VAL_4:.*]] = arith.constant 1 : index
+// CHECK:           %[[VAL_5:.*]] = arith.constant 1 : index
+// CHECK:           %[[VAL_6:.*]] = arith.constant 4 : index
+// CHECK:           %[[VAL_7:.*]] = arith.constant 0 : index
+// CHECK:           %[[VAL_8:.*]] = arith.constant 0 : i32
+// CHECK:           %[[VAL_9:.*]] = vector.transfer_read %[[VAL_0]]{{\[}}%[[VAL_7]], %[[VAL_7]], %[[VAL_7]]], %[[VAL_8]] : tensor<1x1x4xi32>, vector<1x1x4xi32>
+// CHECK:           %[[VAL_10:.*]] = vector.step : vector<1xindex>
+// CHECK:           %[[VAL_11:.*]] = vector.broadcast %[[VAL_10]] : vector<1xindex> to vector<4x1x1xindex>
+// CHECK:           %[[VAL_12:.*]] = vector.transpose %[[VAL_11]], [2, 1, 0] : vector<4x1x1xindex> to vector<1x1x4xindex>
+// CHECK:           %[[VAL_13:.*]] = vector.step : vector<1xindex>
+// CHECK:           %[[VAL_14:.*]] = vector.broadcast %[[VAL_13]] : vector<1xindex> to vector<4x1x1xindex>
+// CHECK:           %[[VAL_15:.*]] = vector.transpose %[[VAL_14]], [2, 1, 0] : vector<4x1x1xindex> to vector<1x1x4xindex>
+// CHECK:           %[[VAL_16:.*]] = arith.constant dense<true> : vector<1x1x4xi1>
+// CHECK:           %[[VAL_17:.*]] = arith.constant dense<0> : vector<1x1x4xi32>
+// CHECK:           %[[VAL_18:.*]] = arith.constant 0 : index
+// CHECK:           %[[VAL_19:.*]] = arith.constant 0 : i32
+// CHECK:           %[[VAL_20:.*]] = vector.shape_cast %[[VAL_15]] : vector<1x1x4xindex> to vector<4xindex>
+// CHECK:           %[[VAL_21:.*]] = vector.extractelement %[[VAL_20]]{{\[}}%[[VAL_19]] : i32] : vector<4xindex>
+// CHECK:           %[[VAL_22:.*]] = arith.constant 0 : i32
+// CHECK:           %[[VAL_23:.*]] = vector.transfer_read %[[VAL_3]]{{\[}}%[[VAL_21]], %[[VAL_2]]], %[[VAL_22]] {in_bounds = [true, true, true], permutation_map = #[[$ATTR_1]]} : tensor<15x1xi32>, vector<1x1x4xi32>
+// CHECK:           %[[VAL_24:.*]] = arith.constant 0 : index
+// CHECK:           %[[VAL_25:.*]] = vector.transfer_write %[[VAL_23]], %[[VAL_0]]{{\[}}%[[VAL_24]], %[[VAL_24]], %[[VAL_24]]] : vector<1x1x4xi32>, tensor<1x1x4xi32>
+
+module attributes {transform.with_named_sequence} {
+  transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
+    %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
+    transform.structured.vectorize %0  vector_sizes [1, 1, 4]{ vectorize_nd_extract } : !transform.any_op
+    transform.yield
+  }
+}



More information about the Mlir-commits mailing list