[Mlir-commits] [mlir] [mlir][vector] Add tests for `populateSinkVectorBroadcastPatterns` (1/n) (PR #102286)

Andrzej WarzyƄski llvmlistbot at llvm.org
Wed Aug 7 01:55:26 PDT 2024


https://github.com/banach-space updated https://github.com/llvm/llvm-project/pull/102286

>From 1739ecd78cfd85341eccb1dfe667d041a86518be Mon Sep 17 00:00:00 2001
From: Andrzej Warzynski <andrzej.warzynski at arm.com>
Date: Wed, 7 Aug 2024 09:40:02 +0100
Subject: [PATCH] [mlir][vector] Add tests for
 `populateSinkVectorBroadcastPatterns` (1/n)

Adds tests for scalable vectors in:
  * sink-vector-broadcast.mlir

This test file excercises patterns grouped under
`populateSinkVectorBroadcastPatterns`, which includes:
  * `ReorderElementwiseOpsOnBroadcast`,
  * `ReorderCastOpsOnBroadcast`.

Right now there are only tests for the former. However, I've noticed
that "vector-reduce-to-contract.mlir" contains tests for the latter and
I've left a few TODOs to group these tests back together in one file.

Additionally, added some helpful `notifyMatchFailure` messages in
`ReorderElementwiseOpsOnBroadcast`.
---
 .../Vector/Transforms/VectorTransforms.cpp    |  17 ++-
 .../Dialect/Vector/sink-vector-broadcast.mlir | 120 +++++++++++++++---
 .../Vector/vector-reduce-to-contract.mlir     |  10 ++
 3 files changed, 121 insertions(+), 26 deletions(-)

diff --git a/mlir/lib/Dialect/Vector/Transforms/VectorTransforms.cpp b/mlir/lib/Dialect/Vector/Transforms/VectorTransforms.cpp
index 6777e589795c8..3ba14d7f900ae 100644
--- a/mlir/lib/Dialect/Vector/Transforms/VectorTransforms.cpp
+++ b/mlir/lib/Dialect/Vector/Transforms/VectorTransforms.cpp
@@ -979,15 +979,18 @@ struct ReorderElementwiseOpsOnBroadcast final
     if (!llvm::isa<ShapedType>(op->getResults()[0].getType()))
       return failure();
     if (!OpTrait::hasElementwiseMappableTraits(op))
+      return rewriter.notifyMatchFailure(
+          op, "Op doesn't have ElementwiseMappableTraits");
+    if (op->getNumOperands() == 0)
       return failure();
-    if (op->getNumOperands() == 0 ||
-        op->getResults()[0].getType() != op->getOperand(0).getType()) {
-      return failure();
-    }
-    // Avoid operations that only accept vector types, since broadcast
-    // source might be scalar types.
+    if (op->getResults()[0].getType() != op->getOperand(0).getType())
+      return rewriter.notifyMatchFailure(op,
+                                         "result and operand type mismatch");
     if (isa<vector::FMAOp>(op)) {
-      return failure();
+      return rewriter.notifyMatchFailure(
+          op,
+          "Op only accepts vector types - not supported as broadcast source "
+          "might be a scalar");
     }
 
     // Get the type of the lhs operand
diff --git a/mlir/test/Dialect/Vector/sink-vector-broadcast.mlir b/mlir/test/Dialect/Vector/sink-vector-broadcast.mlir
index e7863a9e8b7b7..ae55b6696c676 100644
--- a/mlir/test/Dialect/Vector/sink-vector-broadcast.mlir
+++ b/mlir/test/Dialect/Vector/sink-vector-broadcast.mlir
@@ -1,16 +1,32 @@
 // RUN: mlir-opt %s -test-sink-vector-broadcast -split-input-file | FileCheck %s
 
+//-----------------------------------------------------------------------------
+// [Pattern: ReorderElementwiseOpsOnBroadcast]
+//-----------------------------------------------------------------------------
+
 // CHECK-LABEL:   func.func @broadcast_scalar_with_bcast(
 // CHECK-SAME:     %[[ARG_0:.*]]: index, %[[ARG_1:.*]]: index) -> vector<1x4xindex> {
 // CHECK:           %[[ADD:.*]] = arith.addi %[[ARG_0]], %[[ARG_1]] : index
 // CHECK:           %[[BCAST:.*]] = vector.broadcast %[[ADD]] : index to vector<1x4xindex>
 // CHECK:           return %[[BCAST]] : vector<1x4xindex>
 
-func.func @broadcast_scalar_with_bcast( %arg1: index, %arg2: index) -> vector<1x4xindex> {
+func.func @broadcast_scalar_with_bcast(%arg1: index, %arg2: index) -> vector<1x4xindex> {
   %0 = vector.broadcast %arg1 : index to vector<1x4xindex>
   %1 = vector.broadcast %arg2 : index to vector<1x4xindex>
-  %2 = arith.addi %0, %1 : vector<1x4xindex>
-  return %2 : vector<1x4xindex>
+  %2 = arith.addi %0, %1 : vector<1x4xindex> return %2 : vector<1x4xindex>
+}
+
+// CHECK-LABEL:   func.func @broadcast_scalar_with_bcast_scalable(
+// CHECK-SAME:     %[[ARG_0:.*]]: index, %[[ARG_1:.*]]: index) -> vector<1x[4]xindex> {
+// CHECK:           %[[ADD:.*]] = arith.addi %[[ARG_0]], %[[ARG_1]] : index
+// CHECK:           %[[BCAST:.*]] = vector.broadcast %[[ADD]] : index to vector<1x[4]xindex>
+// CHECK:           return %[[BCAST]] : vector<1x[4]xindex>
+
+func.func @broadcast_scalar_with_bcast_scalable(%arg1: index, %arg2: index) -> vector<1x[4]xindex> {
+  %0 = vector.broadcast %arg1 : index to vector<1x[4]xindex>
+  %1 = vector.broadcast %arg2 : index to vector<1x[4]xindex>
+  %2 = arith.addi %0, %1 : vector<1x[4]xindex>
+  return %2 : vector<1x[4]xindex>
 }
 
 // -----
@@ -21,13 +37,26 @@ func.func @broadcast_scalar_with_bcast( %arg1: index, %arg2: index) -> vector<1x
 // CHECK:           %[[ADD:.*]] = arith.addi %[[ARG1]], %[[ARG2]] : index
 // CHECK:           %[[BCAST:.*]] = vector.broadcast %[[ADD]] : index to vector<1x4xindex>
 // CHECK:           return %[[BCAST]] : vector<1x4xindex>
-func.func @broadcast_scalar_with_bcast_and_splat( %arg1: index, %arg2: index) -> vector<1x4xindex> {
+func.func @broadcast_scalar_with_bcast_and_splat(%arg1: index, %arg2: index) -> vector<1x4xindex> {
   %0 = vector.splat %arg1 : vector<1x4xindex>
   %1 = vector.broadcast %arg2 : index to vector<1x4xindex>
   %2 = arith.addi %0, %1 : vector<1x4xindex>
   return %2 : vector<1x4xindex>
 }
 
+// CHECK-LABEL:   func.func @broadcast_scalar_with_bcast_and_splat_scalable(
+// CHECK-SAME:      %[[ARG1:.*]]: index,
+// CHECK-SAME:      %[[ARG2:.*]]: index) -> vector<1x[4]xindex> {
+// CHECK:           %[[ADD:.*]] = arith.addi %[[ARG1]], %[[ARG2]] : index
+// CHECK:           %[[BCAST:.*]] = vector.broadcast %[[ADD]] : index to vector<1x[4]xindex>
+// CHECK:           return %[[BCAST]] : vector<1x[4]xindex>
+func.func @broadcast_scalar_with_bcast_and_splat_scalable(%arg1: index, %arg2: index) -> vector<1x[4]xindex> {
+  %0 = vector.splat %arg1 : vector<1x[4]xindex>
+  %1 = vector.broadcast %arg2 : index to vector<1x[4]xindex>
+  %2 = arith.addi %0, %1 : vector<1x[4]xindex>
+  return %2 : vector<1x[4]xindex>
+}
+
 // -----
 
 // CHECK-LABEL:   func.func @broadcast_vector(
@@ -37,13 +66,27 @@ func.func @broadcast_scalar_with_bcast_and_splat( %arg1: index, %arg2: index) ->
 // CHECK:           %[[BCAST:.*]] = vector.broadcast %[[ADDF]] : vector<4xf32> to vector<3x4xf32>
 // CHECK:           return %[[BCAST]] : vector<3x4xf32>
 
-func.func @broadcast_vector( %arg1: vector<4xf32>, %arg2: vector<4xf32>) -> vector<3x4xf32> {
+func.func @broadcast_vector(%arg1: vector<4xf32>, %arg2: vector<4xf32>) -> vector<3x4xf32> {
   %arg1_bcast = vector.broadcast %arg1 : vector<4xf32> to vector<3x4xf32>
   %arg2_bcast = vector.broadcast %arg2 : vector<4xf32> to vector<3x4xf32>
   %2 = arith.addf %arg1_bcast, %arg2_bcast : vector<3x4xf32>
   return %2 : vector<3x4xf32>
 }
 
+// CHECK-LABEL:   func.func @broadcast_vector_scalable(
+// CHECK-SAME:      %[[ARG_0:.*]]: vector<[4]xf32>,
+// CHECK-SAME:      %[[ARG_1:.*]]: vector<[4]xf32>) -> vector<3x[4]xf32> {
+// CHECK:           %[[ADDF:.*]] = arith.addf %[[ARG_0]], %[[ARG_1]] : vector<[4]xf32>
+// CHECK:           %[[BCAST:.*]] = vector.broadcast %[[ADDF]] : vector<[4]xf32> to vector<3x[4]xf32>
+// CHECK:           return %[[BCAST]] : vector<3x[4]xf32>
+
+func.func @broadcast_vector_scalable(%arg1: vector<[4]xf32>, %arg2: vector<[4]xf32>) -> vector<3x[4]xf32> {
+  %arg1_bcast = vector.broadcast %arg1 : vector<[4]xf32> to vector<3x[4]xf32>
+  %arg2_bcast = vector.broadcast %arg2 : vector<[4]xf32> to vector<3x[4]xf32>
+  %2 = arith.addf %arg1_bcast, %arg2_bcast : vector<3x[4]xf32>
+  return %2 : vector<3x[4]xf32>
+}
+
 // -----
 
 // CHECK-LABEL:   func.func @broadcast_scalar_and_vec(
@@ -53,13 +96,27 @@ func.func @broadcast_vector( %arg1: vector<4xf32>, %arg2: vector<4xf32>) -> vect
 // CHECK:            %[[BCAST:.*]] = vector.broadcast %[[ARG2]] : vector<4xindex> to vector<1x4xindex>
 // CHECK:            %[[ADD:.*]] = arith.addi %[[SPLAT]], %[[BCAST]] : vector<1x4xindex>
 // CHECK:            return %[[ADD]] : vector<1x4xindex>
-func.func @broadcast_scalar_and_vec( %arg1: index, %arg2: vector<4xindex>) -> vector<1x4xindex> {
+func.func @broadcast_scalar_and_vec(%arg1: index, %arg2: vector<4xindex>) -> vector<1x4xindex> {
   %0 = vector.splat %arg1 : vector<1x4xindex>
   %1 = vector.broadcast %arg2 : vector<4xindex> to vector<1x4xindex>
   %2 = arith.addi %0, %1 : vector<1x4xindex>
   return %2 : vector<1x4xindex>
 }
 
+// CHECK-LABEL:   func.func @broadcast_scalar_and_vec_scalable(
+// CHECK-SAME:       %[[ARG1:.*]]: index,
+// CHECK-SAME:       %[[ARG2:.*]]: vector<[4]xindex>) -> vector<1x[4]xindex> {
+// CHECK:            %[[SPLAT:.*]] = vector.splat %[[ARG1]] : vector<1x[4]xindex>
+// CHECK:            %[[BCAST:.*]] = vector.broadcast %[[ARG2]] : vector<[4]xindex> to vector<1x[4]xindex>
+// CHECK:            %[[ADD:.*]] = arith.addi %[[SPLAT]], %[[BCAST]] : vector<1x[4]xindex>
+// CHECK:            return %[[ADD]] : vector<1x[4]xindex>
+func.func @broadcast_scalar_and_vec_scalable(%arg1: index, %arg2: vector<[4]xindex>) -> vector<1x[4]xindex> {
+  %0 = vector.splat %arg1 : vector<1x[4]xindex>
+  %1 = vector.broadcast %arg2 : vector<[4]xindex> to vector<1x[4]xindex>
+  %2 = arith.addi %0, %1 : vector<1x[4]xindex>
+  return %2 : vector<1x[4]xindex>
+}
+
 // -----
 
 // CHECK-LABEL:   func.func @broadcast_vector_and_scalar(
@@ -69,12 +126,25 @@ func.func @broadcast_scalar_and_vec( %arg1: index, %arg2: vector<4xindex>) -> ve
 // CHECK:           %[[ADD:.*]] = arith.addi %[[BCAST]], %[[ARG_1]] : vector<4xi32>
 // CHECK:           return %[[ADD]] : vector<4xi32>
 
-func.func @broadcast_vector_and_scalar( %arg1: i32, %arg2: vector<4xi32>) -> vector<4xi32> {
+func.func @broadcast_vector_and_scalar(%arg1: i32, %arg2: vector<4xi32>) -> vector<4xi32> {
   %arg1_bcast = vector.broadcast %arg1 : i32 to vector<4xi32>
   %2 = arith.addi %arg1_bcast, %arg2 : vector<4xi32>
   return %2 : vector<4xi32>
 }
 
+// CHECK-LABEL:   func.func @broadcast_vector_and_scalar_scalable(
+// CHECK-SAME:      %[[ARG_0:.*]]: i32,
+// CHECK-SAME:      %[[ARG_1:.*]]: vector<[4]xi32>) -> vector<[4]xi32> {
+// CHECK:           %[[BCAST:.*]] = vector.broadcast %[[ARG_0]] : i32 to vector<[4]xi32>
+// CHECK:           %[[ADD:.*]] = arith.addi %[[BCAST]], %[[ARG_1]] : vector<[4]xi32>
+// CHECK:           return %[[ADD]] : vector<[4]xi32>
+
+func.func @broadcast_vector_and_scalar_scalable(%arg1: i32, %arg2: vector<[4]xi32>) -> vector<[4]xi32> {
+  %arg1_bcast = vector.broadcast %arg1 : i32 to vector<[4]xi32>
+  %2 = arith.addi %arg1_bcast, %arg2 : vector<[4]xi32>
+  return %2 : vector<[4]xi32>
+}
+
 // -----
 
 #matmat_accesses = [
@@ -87,12 +157,12 @@ func.func @broadcast_vector_and_scalar( %arg1: i32, %arg2: vector<4xi32>) -> vec
   iterator_types = ["parallel", "parallel", "reduction"]
 }
 
-// CHECK-LABEL:   func.func @broadcast_not_elementwise() -> vector<2x2xf32> {
-// CHECK-DAG:       %[[VAL_0:.*]] = arith.constant dense<1.000000e+00> : vector<2x2xf32>
-// CHECK-DAG:       %[[VAL_1:.*]] = arith.constant dense<2.000000e+00> : vector<2x2xf32>
-// CHECK-DAG:       %[[VAL_2:.*]] = arith.constant dense<3.000000e+00> : vector<2x2xf32>
-// CHECK:           %[[VAL_3:.*]] = vector.contract {indexing_maps = [#map, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %[[VAL_0]], %[[VAL_1]], %[[VAL_2]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
-func.func @broadcast_not_elementwise() -> vector<2x2xf32> {
+// CHECK-LABEL:   func.func @negative_not_elementwise
+// CHECK-DAG:       %[[F1:.*]] = arith.constant dense<1.000000e+00> : vector<2x2xf32>
+// CHECK-DAG:       %[[F2:.*]] = arith.constant dense<2.000000e+00> : vector<2x2xf32>
+// CHECK-DAG:       %[[F3:.*]] = arith.constant dense<3.000000e+00> : vector<2x2xf32>
+// CHECK:           %[[RES:.*]] = vector.contract {indexing_maps = [#map, #map1, #map2], iterator_types = ["parallel", "parallel", "reduction"], kind = #vector.kind<add>} %[[F1]], %[[F2]], %[[F3]] : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
+func.func @negative_not_elementwise() -> vector<2x2xf32> {
   %f1 = arith.constant 1.0: f32
   %f2 = arith.constant 2.0: f32
   %f3 = arith.constant 3.0: f32
@@ -100,27 +170,39 @@ func.func @broadcast_not_elementwise() -> vector<2x2xf32> {
   %A = vector.broadcast %f1 : f32 to vector<2x2xf32>
   %B = vector.broadcast %f2 : f32 to vector<2x2xf32>
   %C = vector.broadcast %f3 : f32 to vector<2x2xf32>
-  %mm1 = vector.contract #matmat_trait %A, %B, %C
+  %res = vector.contract #matmat_trait %A, %B, %C
     : vector<2x2xf32>, vector<2x2xf32> into vector<2x2xf32>
 
-  return %mm1 : vector<2x2xf32>
+  return %res : vector<2x2xf32>
 }
 
-// CHECK-LABEL: func.func @dont_sink_cmp(
+// -----
+
+// The source and the result for arith.cmp have different types - not supported
+
+// CHECK-LABEL: func.func @negative_source_and_result_mismatch
 //       CHECK:   %[[BROADCAST:.+]] = vector.broadcast
 //       CHECK:   %[[RETURN:.+]] = arith.cmpf uno, %[[BROADCAST]], %[[BROADCAST]]
 //       CHECK:   return %[[RETURN]]
-func.func @dont_sink_cmp(%arg0 : f32, %arg1 : vector<1xf32>) -> vector<1xi1> {
+func.func @negative_source_and_result_mismatch(%arg0 : f32, %arg1 : vector<1xf32>) -> vector<1xi1> {
   %0 = vector.broadcast %arg0 : f32 to vector<1xf32>
   %1 = arith.cmpf uno, %0, %0 : vector<1xf32>
   return %1 : vector<1xi1>
 }
 
-// CHECK-LABEL: func.func @dont_sink_fma(
+// -----
+
+// vector.fma only supports vectors - currently it's not possible to replace this with e.g.:
+//    %scalar_res = vector.fma %scalar_1, %scalar2
+//    %vec_res = vector.broadcast %scalar_res
+//
+// TODO: It should be possible to support this case
+
+// CHECK-LABEL: func.func @negative_op_only_supports_vectors
   //     CHECK:   %[[BROADCAST:.+]] = vector.broadcast
   //     CHECK:   %[[RESULT:.+]] = vector.fma %[[BROADCAST]]
   //     CHECK:   return %[[RESULT]]
-func.func @dont_sink_fma(%arg0 : f32) -> vector<1xf32> {
+func.func @negative_op_only_supports_vectors(%arg0 : f32) -> vector<1xf32> {
   %0 = vector.broadcast %arg0 : f32 to vector<1xf32>
   %1 = vector.fma %0, %0, %0 : vector<1xf32>
   return %1 : vector<1xf32>
diff --git a/mlir/test/Dialect/Vector/vector-reduce-to-contract.mlir b/mlir/test/Dialect/Vector/vector-reduce-to-contract.mlir
index 23a44b7c03f8f..c0dbea81df892 100644
--- a/mlir/test/Dialect/Vector/vector-reduce-to-contract.mlir
+++ b/mlir/test/Dialect/Vector/vector-reduce-to-contract.mlir
@@ -246,8 +246,12 @@ func.func @contract_broadcast_would_have_no_reduction_dim_pair(%arg0 : vector<1x
 
 
 //===----------------------------------------------------------------------===//
+// [Pattern: ReorderCastOpsOnBroadcast]
+//
 // Reorder casting ops and vector ops. The casting ops have almost identical
 // pattern, so only arith.extsi op is tested.
+//
+// TODO: Potential duplication with sink-vector-broadcast.mlir
 //===----------------------------------------------------------------------===//
 
 // -----
@@ -272,6 +276,11 @@ func.func @broadcast_scalar_extsi(%a : i8) -> vector<2x4xi32> {
 
 // -----
 
+//===----------------------------------------------------------------------===//
+// [Pattern: ReorderElementwiseOpsOnTranspose]
+//
+// TODO: Potential duplication with sink-vector-broadcast.mlir
+//===----------------------------------------------------------------------===//
 func.func @transpose_extsi(%a : vector<4x2xi8>) -> vector<2x4xi32> {
   // CHECK: %[[EXT:.+]] = arith.extsi %{{.+}} : vector<4x2xi8> to vector<4x2xi32>
   // CHECK: vector.transpose %[[EXT]], [1, 0] : vector<4x2xi32> to vector<2x4xi32>
@@ -282,6 +291,7 @@ func.func @transpose_extsi(%a : vector<4x2xi8>) -> vector<2x4xi32> {
 
 //===----------------------------------------------------------------------===//
 // Reorder elementwise ops and vector ops.
+// TODO: Potential duplication with sink-vector-broadcast.mlir
 //===----------------------------------------------------------------------===//
 
 // -----



More information about the Mlir-commits mailing list