[Mlir-commits] [mlir] andrzej/update collapse inner 2 (PR #94604)
llvmlistbot at llvm.org
llvmlistbot at llvm.org
Thu Jun 6 04:11:53 PDT 2024
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-mlir
Author: Andrzej Warzyński (banach-space)
<details>
<summary>Changes</summary>
- **[mlir][vector] Update tests for collapse 1/n (nfc)**
- **[mlir][vector] Update tests for collapse 2/n (nfc)**
---
Full diff: https://github.com/llvm/llvm-project/pull/94604.diff
1 Files Affected:
- (modified) mlir/test/Dialect/Vector/vector-transfer-collapse-inner-most-dims.mlir (+144-41)
``````````diff
diff --git a/mlir/test/Dialect/Vector/vector-transfer-collapse-inner-most-dims.mlir b/mlir/test/Dialect/Vector/vector-transfer-collapse-inner-most-dims.mlir
index b4cb640108bae..9fb648b39ffc5 100644
--- a/mlir/test/Dialect/Vector/vector-transfer-collapse-inner-most-dims.mlir
+++ b/mlir/test/Dialect/Vector/vector-transfer-collapse-inner-most-dims.mlir
@@ -1,12 +1,17 @@
// RUN: mlir-opt %s -test-vector-transfer-collapse-inner-most-dims -split-input-file | FileCheck %s
-func.func @contiguous_inner_most_view(%in: memref<1x1x8x1xf32, strided<[3072, 8, 1, 1], offset: ?>>) -> vector<1x8x1xf32>{
+//-----------------------------------------------------------------------------
+// 1. vector.transfer_read
+//-----------------------------------------------------------------------------
+
+func.func @contiguous_inner_most(%in: memref<1x1x8x1xf32, strided<[3072, 8, 1, 1], offset: ?>>) -> vector<1x8x1xf32>{
%c0 = arith.constant 0 : index
%cst = arith.constant 0.0 : f32
%0 = vector.transfer_read %in[%c0, %c0, %c0, %c0], %cst {in_bounds = [true, true, true]} : memref<1x1x8x1xf32, strided<[3072, 8, 1, 1], offset: ?>>, vector<1x8x1xf32>
return %0 : vector<1x8x1xf32>
}
-// CHECK: func @contiguous_inner_most_view(%[[SRC:.+]]: memref<1x1x8x1xf32, strided<[3072, 8, 1, 1], offset: ?>>
+
+// CHECK: func @contiguous_inner_most(%[[SRC:.+]]: memref<1x1x8x1xf32, strided<[3072, 8, 1, 1], offset: ?>>
// CHECK: %[[SRC_0:.+]] = memref.subview %[[SRC]]
// CHECK-SAME: memref<1x1x8x1xf32, strided<[3072, 8, 1, 1], offset: ?>> to memref<1x1x8xf32, strided<[3072, 8, 1], offset: ?>>
// CHECK: %[[VEC:.+]] = vector.transfer_read %[[SRC_0]]
@@ -14,15 +19,61 @@ func.func @contiguous_inner_most_view(%in: memref<1x1x8x1xf32, strided<[3072, 8,
// CHECK: %[[RESULT:.+]] = vector.shape_cast %[[VEC]]
// CHECK: return %[[RESULT]]
+// Same as the top example within this split, but with the inner vector
+// dim scalable. Note that this example only makes sense when "8 = [8]" (i.e.
+// vscale = 1). This is assumed (implicitly) via the `in_bounds` attribute.
+
+func.func @contiguous_inner_most_scalable_inner_dim(%in: memref<1x1x8x1xf32, strided<[3072, 8, 1, 1], offset: ?>>) -> vector<1x[8]x1xf32>{
+ %c0 = arith.constant 0 : index
+ %cst = arith.constant 0.0 : f32
+ %0 = vector.transfer_read %in[%c0, %c0, %c0, %c0], %cst {in_bounds = [true, true, true]} : memref<1x1x8x1xf32, strided<[3072, 8, 1, 1], offset: ?>>, vector<1x[8]x1xf32>
+ return %0 : vector<1x[8]x1xf32>
+}
+
+// CHECK: func @contiguous_inner_most_scalable_inner_dim(%[[SRC:.+]]: memref<1x1x8x1xf32, strided<[3072, 8, 1, 1], offset: ?>>
+// CHECK: %[[SRC_0:.+]] = memref.subview %[[SRC]]
+// CHECK-SAME: memref<1x1x8x1xf32, strided<[3072, 8, 1, 1], offset: ?>> to memref<1x1x8xf32, strided<[3072, 8, 1], offset: ?>>
+// CHECK: %[[VEC:.+]] = vector.transfer_read %[[SRC_0]]
+// CHECK-SAME: memref<1x1x8xf32, strided<[3072, 8, 1], offset: ?>>, vector<1x[8]xf32>
+// CHECK: %[[RESULT:.+]] = vector.shape_cast %[[VEC]]
+// CHECK: return %[[RESULT]]
+
+// Same as the top example within this split, but the trailing unit dim was
+// replaced with a dyn dim - not supported
+
+func.func @non_unit_trailing_dim(%in: memref<1x1x8x?xf32, strided<[3072, 8, 1, 1], offset: ?>>) -> vector<1x8x1xf32>{
+ %c0 = arith.constant 0 : index
+ %cst = arith.constant 0.0 : f32
+ %0 = vector.transfer_read %in[%c0, %c0, %c0, %c0], %cst {in_bounds = [true, true, true]} : memref<1x1x8x?xf32, strided<[3072, 8, 1, 1], offset: ?>>, vector<1x8x1xf32>
+ return %0 : vector<1x8x1xf32>
+}
+
+// CHECK-LABEL: func @non_unit_trailing_dim
+// CHECK-NOT: memref.subview
+// CHECK-NOT: vector.shape_cast
+
+// Same as the top example within this split, but with a scalable unit dim in
+// the output vector - not supported
+
+func.func @negative_scalable_unit_dim(%in: memref<1x1x8x1xf32, strided<[3072, 8, 1, 1], offset: ?>>) -> vector<1x8x[1]xf32>{
+ %c0 = arith.constant 0 : index
+ %cst = arith.constant 0.0 : f32
+ %0 = vector.transfer_read %in[%c0, %c0, %c0, %c0], %cst {in_bounds = [true, true, true]} : memref<1x1x8x1xf32, strided<[3072, 8, 1, 1], offset: ?>>, vector<1x8x[1]xf32>
+ return %0 : vector<1x8x[1]xf32>
+}
+// CHECK-LABEL: func @scalable_unit_dim
+// CHECK-NOT: memref.subview
+// CHECK-NOT: vector.shape_cast
+
// -----
-func.func @contiguous_outer_dyn_inner_most_view(%a: index, %b: index, %memref: memref<?x?x8x1xf32>) -> vector<8x1xf32> {
+func.func @contiguous_inner_most_dynamic_outer(%a: index, %b: index, %memref: memref<?x?x8x1xf32>) -> vector<8x1xf32> {
%c0 = arith.constant 0 : index
%pad = arith.constant 0.0 : f32
%v = vector.transfer_read %memref[%a, %b, %c0, %c0], %pad {in_bounds = [true, true]} : memref<?x?x8x1xf32>, vector<8x1xf32>
return %v : vector<8x1xf32>
}
-// CHECK: func.func @contiguous_outer_dyn_inner_most_view(
+// CHECK: func.func @contiguous_inner_most_dynamic_outer
// CHECK-SAME: %[[IDX0:[a-zA-Z0-9]+]]
// CHECK-SAME: %[[IDX1:[a-zA-Z0-9]+]]
// CHECK-SAME: %[[SRC:[a-zA-Z0-9]+]]
@@ -38,56 +89,132 @@ func.func @contiguous_outer_dyn_inner_most_view(%a: index, %b: index, %memref: m
// CHECK: %[[RESULT:.+]] = vector.shape_cast %[[VEC]]
// CHECK: return %[[RESULT]]
+// Same as the top example within this split, but with the inner vector
+// dim scalable. Note that this example only makes sense when "8 = [8]" (i.e.
+// vscale = 1). This is assumed (implicitly) via the `in_bounds` attribute.
+
+func.func @contiguous_inner_most_outer_dim_dyn_scalable_inner_dim(%a: index, %b: index, %memref: memref<?x?x8x1xf32>) -> vector<[8]x1xf32> {
+ %c0 = arith.constant 0 : index
+ %pad = arith.constant 0.0 : f32
+ %v = vector.transfer_read %memref[%a, %b, %c0, %c0], %pad {in_bounds = [true, true]} : memref<?x?x8x1xf32>, vector<[8]x1xf32>
+ return %v : vector<[8]x1xf32>
+}
+// CHECK-LABEL: func @contiguous_inner_most_outer_dim_dyn_scalable_inner_dim
+// CHECK-SAME: %[[IDX0:[a-zA-Z0-9]+]]
+// CHECK-SAME: %[[IDX1:[a-zA-Z0-9]+]]
+// CHECK-SAME: %[[SRC:[a-zA-Z0-9]+]]
+// CHECK: %[[VIEW:.+]] = memref.subview %[[SRC]]{{.*}} memref<?x?x8x1xf32> to memref<?x?x8xf32, strided<[?, 8, 1], offset: ?>>
+// CHECK: %[[VEC_READ:.+]] = vector.transfer_read %[[VIEW]]
+// CHECK-SAME: {in_bounds = [true]}
+// CHECK-SAME: memref<?x?x8xf32, strided<[?, 8, 1], offset: ?>>, vector<[8]xf32>
+// CHECK: vector.shape_cast %[[VEC_READ]]
+
// -----
-func.func @contiguous_inner_most_dim(%A: memref<16x1xf32>, %i:index, %j:index) -> (vector<8x1xf32>) {
+func.func @contiguous_inner_most_dim_non_zero_idxs(%A: memref<16x1xf32>, %i:index, %j:index) -> (vector<8x1xf32>) {
%c0 = arith.constant 0 : index
%f0 = arith.constant 0.0 : f32
%1 = vector.transfer_read %A[%i, %j], %f0 : memref<16x1xf32>, vector<8x1xf32>
return %1 : vector<8x1xf32>
}
-// CHECK: func @contiguous_inner_most_dim(%[[SRC:.+]]: memref<16x1xf32>, %[[I:.+]]: index, %[[J:.+]]: index) -> vector<8x1xf32>
+// CHECK: func @contiguous_inner_most_dim_non_zero_idxs(%[[SRC:.+]]: memref<16x1xf32>, %[[I:.+]]: index, %[[J:.+]]: index) -> vector<8x1xf32>
// CHECK: %[[SRC_0:.+]] = memref.subview %[[SRC]]
// CHECK-SAME: memref<16x1xf32> to memref<16xf32, strided<[1]>>
// CHECK: %[[V:.+]] = vector.transfer_read %[[SRC_0]]
-// CHECK: %[[RESULT]] = vector.shape_cast %[[V]] : vector<8xf32> to vector<8x1xf32>
+// CHECK: %[[RESULT:.+]] = vector.shape_cast %[[V]] : vector<8xf32> to vector<8x1xf32>
// CHECK: return %[[RESULT]]
+// Same as the top example within this split, but with the inner vector
+// dim scalable. Note that this example only makes sense when "8 = [8]" (i.e.
+// vscale = 1). This is assumed (implicitly) via the `in_bounds` attribute.
+
+func.func @contiguous_inner_most_dim_non_zero_idxs_scalable_inner_dim(%A: memref<16x1xf32>, %i:index, %j:index) -> (vector<[8]x1xf32>) {
+ %c0 = arith.constant 0 : index
+ %f0 = arith.constant 0.0 : f32
+ %1 = vector.transfer_read %A[%i, %j], %f0 : memref<16x1xf32>, vector<[8]x1xf32>
+ return %1 : vector<[8]x1xf32>
+}
+// CHECK-LABEL: func @contiguous_inner_most_dim_non_zero_idxs_scalable_inner_dim(
+// CHECK-SAME: %[[SRC:.+]]: memref<16x1xf32>, %[[I:.+]]: index, %[[J:.+]]: index) -> vector<[8]x1xf32>
+// CHECK: %[[SRC_0:.+]] = memref.subview %[[SRC]]
+// CHECK-SAME: memref<16x1xf32> to memref<16xf32, strided<[1]>>
+// CHECK: %[[V:.+]] = vector.transfer_read %[[SRC_0]]
+// CHECK: %[[RESULT:.+]] = vector.shape_cast %[[V]] : vector<[8]xf32> to vector<[8]x1xf32>
+// CHECK: return %[[RESULT]]
+
// -----
-func.func @contiguous_inner_most_dim_bounds(%A: memref<1000x1xf32>, %i:index, %ii:index) -> (vector<4x1xf32>) {
+func.func @contiguous_inner_most_dim_with_subview(%A: memref<1000x1xf32>, %i:index, %ii:index) -> (vector<4x1xf32>) {
%c0 = arith.constant 0 : index
%cst = arith.constant 0.0 : f32
%0 = memref.subview %A[%i, 0] [40, 1] [1, 1] : memref<1000x1xf32> to memref<40x1xf32, strided<[1, 1], offset: ?>>
%1 = vector.transfer_read %0[%ii, %c0], %cst {in_bounds = [true, true]} : memref<40x1xf32, strided<[1, 1], offset: ?>>, vector<4x1xf32>
return %1 : vector<4x1xf32>
}
-// CHECK: func @contiguous_inner_most_dim_bounds(%[[SRC:.+]]: memref<1000x1xf32>, %[[II:.+]]: index, %[[J:.+]]: index) -> vector<4x1xf32>
+// CHECK: func @contiguous_inner_most_dim_with_subview(%[[SRC:.+]]: memref<1000x1xf32>, %[[II:.+]]: index, %[[J:.+]]: index) -> vector<4x1xf32>
// CHECK: %[[SRC_0:.+]] = memref.subview %[[SRC]]
// CHECK: %[[SRC_1:.+]] = memref.subview %[[SRC_0]]
// CHECK: %[[V:.+]] = vector.transfer_read %[[SRC_1]]
// CHECK-SAME: {in_bounds = [true]}
// CHECK-SAME: vector<4xf32>
+// Same as the top example within this split, but with the inner vector
+// dim scalable. Note that this example only makes sense when "8 = [8]" (i.e.
+// vscale = 1). This is assumed (implicitly) via the `in_bounds` attribute.
+
+func.func @contiguous_inner_most_dim_with_subview_scalable_inner_dim(%A: memref<1000x?xf32>, %i:index, %ii:index, %j:index) -> (vector<[4]x1xf32>) {
+ %c0 = arith.constant 0 : index
+ %cst = arith.constant 0.0 : f32
+ %0 = memref.subview %A[%i, 0] [40, %j] [1, 1] : memref<1000x?xf32> to memref<40x?xf32, strided<[?, 1], offset: ?>>
+ %1 = vector.transfer_read %0[%ii, %c0], %cst {in_bounds = [true, true]} : memref<40x?xf32, strided<[?, 1], offset: ?>>, vector<[4]x1xf32>
+ return %1 : vector<[4]x1xf32>
+}
+// CHECK-LABEL: func @contiguous_inner_most_dim_with_bounds_trailing_dim_dyn_scalable_vec
+// CHECK-SAME: %[[SRC:.+]]: memref<1000x?xf32>
+// CHECK: %[[SRC_0:.+]] = memref.subview %[[SRC]]
+// CHECK: %[[V:.+]] = vector.transfer_read %[[SRC_1]]
+// CHECK-SAME: {in_bounds = [true]}
+// CHECK-SAME: vector<[4]xf32>
+
// -----
-func.func @contiguous_inner_most_dim_bounds_2d(%A: memref<1000x1x1xf32>, %i:index, %ii:index) -> (vector<4x1x1xf32>) {
+func.func @contiguous_inner_most_dim_2d(%A: memref<1000x1x1xf32>, %i:index, %ii:index) -> (vector<4x1x1xf32>) {
%c0 = arith.constant 0 : index
%cst = arith.constant 0.0 : f32
%0 = memref.subview %A[%i, 0, 0] [40, 1, 1] [1, 1, 1] : memref<1000x1x1xf32> to memref<40x1x1xf32, strided<[1, 1, 1], offset: ?>>
%1 = vector.transfer_read %0[%ii, %c0, %c0], %cst {in_bounds = [true, true, true]} : memref<40x1x1xf32, strided<[1, 1, 1], offset: ?>>, vector<4x1x1xf32>
return %1 : vector<4x1x1xf32>
}
-// CHECK: func @contiguous_inner_most_dim_bounds_2d(%[[SRC:.+]]: memref<1000x1x1xf32>, %[[II:.+]]: index, %[[J:.+]]: index) -> vector<4x1x1xf32>
+// CHECK: func @contiguous_inner_most_dim_2d(%[[SRC:.+]]: memref<1000x1x1xf32>, %[[II:.+]]: index, %[[J:.+]]: index) -> vector<4x1x1xf32>
// CHECK: %[[SRC_0:.+]] = memref.subview %[[SRC]]
// CHECK: %[[SRC_1:.+]] = memref.subview %[[SRC_0]]
// CHECK: %[[V:.+]] = vector.transfer_read %[[SRC_1]]
// CHECK-SAME: {in_bounds = [true]}
// CHECK-SAME: vector<4xf32>
+// Same as the top example within this split, but with the inner vector
+// dim scalable. Note that this example only makes sense when "8 = [8]" (i.e.
+// vscale = 1). This is assumed (implicitly) via the `in_bounds` attribute.
+
+func.func @contiguous_inner_most_dim_2d_scalable_inner_dim(%A: memref<1000x1x1xf32>, %i:index, %ii:index) -> (vector<[4]x1x1xf32>) {
+ %c0 = arith.constant 0 : index
+ %cst = arith.constant 0.0 : f32
+ %0 = memref.subview %A[%i, 0, 0] [40, 1, 1] [1, 1, 1] : memref<1000x1x1xf32> to memref<40x1x1xf32, strided<[1, 1, 1], offset: ?>>
+ %1 = vector.transfer_read %0[%ii, %c0, %c0], %cst {in_bounds = [true, true, true]} : memref<40x1x1xf32, strided<[1, 1, 1], offset: ?>>, vector<[4]x1x1xf32>
+ return %1 : vector<[4]x1x1xf32>
+}
+// CHECK-LABEL: func @contiguous_inner_most_dim_2d_scalable_inner_dim(
+// CHECK-SAME: %[[SRC:.+]]: memref<1000x1x1xf32>, %[[II:.+]]: index, %[[J:.+]]: index) -> vector<[4]x1x1xf32>
+// CHECK: %[[SRC_0:.+]] = memref.subview %[[SRC]]
+// CHECK: %[[SRC_1:.+]] = memref.subview %[[SRC_0]]
+// CHECK: %[[V:.+]] = vector.transfer_read %[[SRC_1]]
+// CHECK-SAME: {in_bounds = [true]}
+// CHECK-SAME: vector<[4]xf32>
+// CHECK: vector.shape_cast %[[V]]
+
// -----
-func.func @contiguous_inner_most_dim_out_of_bounds_2d(%arg0: memref<1x1xf32>) -> vector<4x8xf32> {
+func.func @negative_out_of_bounds(%arg0: memref<1x1xf32>) -> vector<4x8xf32> {
%c0 = arith.constant 0 : index
%cst = arith.constant 0.000000e+00 : f32
%0 = vector.transfer_read %arg0[%c0, %c0], %cst : memref<1x1xf32>, vector<4x8xf32>
@@ -95,7 +222,7 @@ func.func @contiguous_inner_most_dim_out_of_bounds_2d(%arg0: memref<1x1xf32>) ->
}
// The inner most unit dim can not be dropped. In this context, we do not
// generate rank-reduced memref.subview ops.
-// CHECK: func.func @contiguous_inner_most_dim_out_of_bounds_2d
+// CHECK: func.func @negative_out_of_bounds
// CHECK-SAME: %[[SRC:[a-zA-Z0-9]+]]
// CHECK-NOT: memref.subview
// CHECK: %[[READ:.+]] = vector.transfer_read %[[SRC]]
@@ -103,6 +230,10 @@ func.func @contiguous_inner_most_dim_out_of_bounds_2d(%arg0: memref<1x1xf32>) ->
// -----
+//-----------------------------------------------------------------------------
+// 2. vector.transfer_write
+//-----------------------------------------------------------------------------
+
func.func @drop_two_inner_most_dim_for_transfer_write(%arg0: memref<1x512x16x1x1xf32>, %arg1: vector<1x16x16x1x1xf32>, %arg2: index) {
%c0 = arith.constant 0 : index
vector.transfer_write %arg1, %arg0[%c0, %arg2, %c0, %c0, %c0]
@@ -177,21 +308,6 @@ func.func @non_unit_strides(%arg0: memref<512x16x1xf32, strided<[8192, 16, 4], o
// -----
-func.func @leading_scalable_dimension_transfer_read(%dest : memref<24x1xf32>) -> vector<[4]x1xf32> {
- %c0 = arith.constant 0 : index
- %pad = arith.constant 0.0 : f32
- %0 = vector.transfer_read %dest[%c0, %c0], %pad {in_bounds = [true, true]} : memref<24x1xf32>, vector<[4]x1xf32>
- return %0 : vector<[4]x1xf32>
-}
-// CHECK: func.func @leading_scalable_dimension_transfer_read
-// CHECK-SAME: %[[DEST:[a-zA-Z0-9]+]]
-// CHECK: %[[SUBVIEW:.+]] = memref.subview %[[DEST]][0, 0] [24, 1] [1, 1] : memref<24x1xf32> to memref<24xf32, strided<[1]>>
-// CHECK: %[[READ:.+]] = vector.transfer_read %[[SUBVIEW]]{{.*}} {in_bounds = [true]} : memref<24xf32, strided<[1]>>, vector<[4]xf32>
-// CHECK: %[[CAST:.+]] = vector.shape_cast %[[READ]] : vector<[4]xf32> to vector<[4]x1xf32>
-// CHECK: return %[[CAST]]
-
-// -----
-
// Negative test: [1] (scalable 1) is _not_ a unit dimension.
func.func @trailing_scalable_one_dim_transfer_read(%dest : memref<24x1xf32>) -> vector<4x[1]xf32> {
%c0 = arith.constant 0 : index
@@ -217,16 +333,3 @@ func.func @leading_scalable_dimension_transfer_write(%dest : memref<24x1xf32>, %
// CHECK: %[[SUBVIEW:.+]] = memref.subview %[[DEST]][0, 0] [24, 1] [1, 1] : memref<24x1xf32> to memref<24xf32, strided<[1]>>
// CHECK: %[[CAST:.+]] = vector.shape_cast %[[VEC]] : vector<[4]x1xf32> to vector<[4]xf32>
// CHECK: vector.transfer_write %[[CAST]], %[[SUBVIEW]]{{.*}} {in_bounds = [true]} : vector<[4]xf32>, memref<24xf32, strided<[1]>>
-
-// -----
-
-// Negative test: [1] (scalable 1) is _not_ a unit dimension.
-func.func @trailing_scalable_one_dim_transfer_write(%dest : memref<24x1xf32>, %vec: vector<4x[1]xf32>, %index: index) {
- %c0 = arith.constant 0 : index
- vector.transfer_write %vec, %dest[%index, %c0] {in_bounds = [true, true]} : vector<4x[1]xf32>, memref<24x1xf32>
- return
-}
-// CHECK: func.func @trailing_scalable_one_dim_transfer_write
-// CHECK-NOT: vector.shape_cast
-// CHECK: vector.transfer_write {{.*}} : vector<4x[1]xf32>, memref<24x1xf32>
-// CHECK-NOT: vector.shape_cast
``````````
</details>
https://github.com/llvm/llvm-project/pull/94604
More information about the Mlir-commits
mailing list