Skip to content

[mlir][vector] Update tests for collapse 5/n (nfc) #96227

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
merged 2 commits into from
Jul 15, 2024
Merged

[mlir][vector] Update tests for collapse 5/n (nfc) #96227

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
871a7ee
[mlir][vector] Update tests for collapse 5/n (nfc)
banach-space Jun 20, 2024
1f5c0aa
fixup! [mlir][vector] Update tests for collapse 5/n (nfc)
banach-space Jul 15, 2024
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
122 changes: 122 additions & 0 deletions mlir/test/Dialect/Vector/vector-transfer-collapse-inner-most-dims.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -438,6 +438,24 @@ func.func @contiguous_inner_most_dim_non_zero_idx_in_bounds(%arg0: memref<16x1xf
// CHECK: %[[SC:.*]] = vector.shape_cast %[[VEC]] : vector<8x1xf32> to vector<8xf32>
// CHECK: vector.transfer_write %[[SC]], %[[SV]]{{\[}}%[[IDX]]] {in_bounds = [true]} : vector<8xf32>, memref<16xf32, strided<[1]>>

// Same as the top example within this split, but with the outer vector
// dim scalable. Note that this example only makes sense when "8 = [8]" (i.e.
// vscale = 1). This is assumed via the `in_bounds` attribute.

// TODO: Add a similar test for xfer_write

func.func @contiguous_inner_most_non_zero_idx_in_bounds_scalable(%arg0: memref<16x1xf32>, %arg1: vector<[8]x1xf32>, %i: index) {
vector.transfer_write %arg1, %arg0[%i, %i] {in_bounds = [true, true]} : vector<[8]x1xf32>, memref<16x1xf32>
return
}
// CHECK-LABEL: func.func @contiguous_inner_most_non_zero_idx_in_bounds_scalable(
// CHECK-SAME: %[[MEM:.*]]: memref<16x1xf32>,
// CHECK-SAME: %[[VEC:.*]]: vector<[8]x1xf32>
// CHECK-SAME: %[[IDX:.*]]: index) {
// CHECK: %[[SV:.*]] = memref.subview %[[MEM]][0, 0] [16, 1] [1, 1] : memref<16x1xf32> to memref<16xf32, strided<[1]>>
// CHECK: %[[SC:.*]] = vector.shape_cast %[[VEC]] : vector<[8]x1xf32> to vector<[8]xf32>
// CHECK: vector.transfer_write %[[SC]], %[[SV]]{{\[}}%[[IDX]]] {in_bounds = [true]} : vector<[8]xf32>, memref<16xf32, strided<[1]>>

// The index to be dropped is unknown and "out of bounds" - not safe to
// collapse.
func.func @negative_contiguous_inner_most_dim_non_zero_idx_out_of_bounds(%arg0: memref<16x1xf32>, %arg1: vector<8x1xf32>, %i: index) {
Expand All @@ -451,6 +469,86 @@ func.func @negative_contiguous_inner_most_dim_non_zero_idx_out_of_bounds(%arg0:

// -----

// Verify that the transformation does work even when the input is a "subview"

func.func @contiguous_inner_most_dim_with_subview(%A: memref<1000x1xf32>, %i:index, %ii:index, %vec: 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: ?>>
vector.transfer_write %vec, %0[%ii, %c0] {in_bounds = [true, true]} : vector<4x1xf32>, memref<40x1xf32, strided<[1, 1], offset: ?>>
return
}

// CHECK-LABEL: func.func @contiguous_inner_most_dim_with_subview(
// CHECK-SAME: %[[MEM:.*]]: memref<1000x1xf32>,
// CHECK-SAME: %[[IDX_1:.*]]: index, %[[IDX_2:.*]]: index,
// CHECK-SAME: %[[VEC:.*]]: vector<4x1xf32>) {
// CHECK: %[[SV_1:.*]] = memref.subview %[[MEM]]{{\[}}%[[IDX_1]], 0] [40, 1] [1, 1] : memref<1000x1xf32> to memref<40x1xf32, strided<[1, 1], offset: ?>>
// CHECK: %[[SV_2:.*]] = memref.subview %[[SV_1]][0, 0] [40, 1] [1, 1] : memref<40x1xf32, strided<[1, 1], offset: ?>> to memref<40xf32, strided<[1], offset: ?>>
// CHECK: %[[SC:.*]] = vector.shape_cast %[[VEC]] : vector<4x1xf32> to vector<4xf32>
// CHECK: vector.transfer_write %[[SC]], %[[SV_2]]{{\[}}%[[IDX_2]]] {in_bounds = [true]} : vector<4xf32>, memref<40xf32, strided<[1], offset: ?>>

// Same as the top example within this split, but with the outer vector
// dim scalable. Note that this example only makes sense when "4 = [4]" (i.e.
// vscale = 1). This is assumed via the `in_bounds` attribute.

func.func @contiguous_inner_most_dim_with_subview_scalable_inner_dim(%A: memref<1000x1xf32>, %i:index, %ii:index, %vec: vector<[4]x1xf32>) {
%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: ?>>
vector.transfer_write %vec, %0[%ii, %c0] {in_bounds = [true, true]} : vector<[4]x1xf32>, memref<40x1xf32, strided<[1, 1], offset: ?>>
return
}

// CHECK-LABEL: func.func @contiguous_inner_most_dim_with_subview_scalable_inner_dim
// CHECK-SAME: %[[MEM:.*]]: memref<1000x1xf32>,
// CHECK-SAME: %[[IDX_1:.*]]: index, %[[IDX_2:.*]]: index,
// CHECK-SAME: %[[VEC:.*]]: vector<[4]x1xf32>) {
// CHECK: %[[SV_1:.*]] = memref.subview %[[MEM]]{{\[}}%[[IDX_1]], 0] [40, 1] [1, 1] : memref<1000x1xf32> to memref<40x1xf32, strided<[1, 1], offset: ?>>
// CHECK: %[[SV_2:.*]] = memref.subview %[[SV_1]][0, 0] [40, 1] [1, 1] : memref<40x1xf32, strided<[1, 1], offset: ?>> to memref<40xf32, strided<[1], offset: ?>>
// CHECK: %[[SC:.*]] = vector.shape_cast %[[VEC]] : vector<[4]x1xf32> to vector<[4]xf32>
// CHECK: vector.transfer_write %[[SC]], %[[SV_2]]{{\[}}%[[IDX_2]]] {in_bounds = [true]} : vector<[4]xf32>, memref<40xf32, strided<[1], offset: ?>>

// -----

func.func @contiguous_inner_most_dim_with_subview_2d(%A: memref<1000x1x1xf32>, %i:index, %ii:index, %vec: 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: ?>>
vector.transfer_write %vec, %0[%ii, %c0, %c0] {in_bounds = [true, true, true]} : vector<4x1x1xf32>, memref<40x1x1xf32, strided<[1, 1, 1], offset: ?>>
return
}
// CHECK-LABEL: func.func @contiguous_inner_most_dim_with_subview_2d(
// CHECK-SAME: %[[MEM:.*]]: memref<1000x1x1xf32>,
// CHECK-SAME: %[[IDX_1:.*]]: index, %[[IDX_2:.*]]: index,
// CHECK-SAME: %[[VEC:.*]]: vector<4x1x1xf32>) {
// CHECK: %[[SV_1:.*]] = memref.subview %[[MEM]]{{\[}}%[[IDX_1]], 0, 0] [40, 1, 1] [1, 1, 1] : memref<1000x1x1xf32> to memref<40x1x1xf32, strided<[1, 1, 1], offset: ?>>
// CHECK: %[[SV_2:.*]] = memref.subview %[[SV_1]][0, 0, 0] [40, 1, 1] [1, 1, 1] : memref<40x1x1xf32, strided<[1, 1, 1], offset: ?>> to memref<40xf32, strided<[1], offset: ?>>
// CHECK: %[[SC:.*]] = vector.shape_cast %[[VEC]] : vector<4x1x1xf32> to vector<4xf32>
// CHECK: vector.transfer_write %[[SC]], %[[SV_2]]{{\[}}%[[IDX_2]]] {in_bounds = [true]} : vector<4xf32>, memref<40xf32, strided<[1], offset: ?>>

// Same as the top example within this split, but with the outer vector
// dim scalable. Note that this example only makes sense when "4 = [4]" (i.e.
// vscale = 1). This is assumed (implicitly) via the `in_bounds` attribute.

func.func @contiguous_inner_most_dim_with_subview_2d_scalable(%A: memref<1000x1x1xf32>, %i:index, %ii:index, %vec: 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: ?>>
vector.transfer_write %vec, %0[%ii, %c0, %c0] {in_bounds = [true, true, true]} : vector<[4]x1x1xf32>, memref<40x1x1xf32, strided<[1, 1, 1], offset: ?>>
return
}
// CHECK-LABEL: func.func @contiguous_inner_most_dim_with_subview_2d_scalable
// CHECK-SAME: %[[MEM:.*]]: memref<1000x1x1xf32>,
// CHECK-SAME: %[[IDX_1:.*]]: index, %[[IDX_2:.*]]: index,
// CHECK-SAME: %[[VEC:.*]]: vector<[4]x1x1xf32>) {
// CHECK: %[[SV_1:.*]] = memref.subview %[[MEM]]{{\[}}%[[IDX_1]], 0, 0] [40, 1, 1] [1, 1, 1] : memref<1000x1x1xf32> to memref<40x1x1xf32, strided<[1, 1, 1], offset: ?>>
// CHECK: %[[SV_2:.*]] = memref.subview %[[SV_1]][0, 0, 0] [40, 1, 1] [1, 1, 1] : memref<40x1x1xf32, strided<[1, 1, 1], offset: ?>> to memref<40xf32, strided<[1], offset: ?>>
// CHECK: %[[SC:.*]] = vector.shape_cast %[[VEC]] : vector<[4]x1x1xf32> to vector<[4]xf32>
// CHECK: vector.transfer_write %[[SC]], %[[SV_2]]{{\[}}%[[IDX_2]]] {in_bounds = [true]} : vector<[4]xf32>, memref<40xf32, strided<[1], offset: ?>>

// -----

func.func @drop_inner_most_dim(%arg0: memref<1x512x16x1xf32, strided<[8192, 16, 1, 1], offset: ?>>, %arg1: vector<1x16x16x1xf32>, %arg2: index) {
%c0 = arith.constant 0 : index
vector.transfer_write %arg1, %arg0[%c0, %arg2, %c0, %c0]
Expand All @@ -471,6 +569,30 @@ func.func @drop_inner_most_dim(%arg0: memref<1x512x16x1xf32, strided<[8192, 16,

// -----

// NOTE: This is an out-of-bounds access.

func.func @negative_non_unit_inner_vec_dim(%arg0: memref<4x1xf32>, %vec: vector<4x8xf32>) {
%c0 = arith.constant 0 : index
vector.transfer_write %vec, %arg0[%c0, %c0] : vector<4x8xf32>, memref<4x1xf32>
return
}
// CHECK: func.func @negative_non_unit_inner_vec_dim
// CHECK-NOT: memref.subview
// CHECK: vector.transfer_write

// -----

func.func @negative_non_unit_inner_memref_dim(%arg0: memref<4x8xf32>, %vec: vector<4x1xf32>) {
%c0 = arith.constant 0 : index
vector.transfer_write %vec, %arg0[%c0, %c0] : vector<4x1xf32>, memref<4x8xf32>
return
}
// CHECK: func.func @negative_non_unit_inner_memref_dim
// CHECK-NOT: memref.subview
// CHECK: vector.transfer_write

// -----

func.func @non_unit_strides(%arg0: memref<512x16x1xf32, strided<[8192, 16, 4], offset: ?>>, %arg1: vector<16x16x1xf32>, %arg2: index) {
%c0 = arith.constant 0 : index
vector.transfer_write %arg1, %arg0[%arg2, %c0, %c0]
Expand Down
Loading