[Mlir-commits] [mlir] [mlir][vector] Update tests/patterns for vector.transpose (PR #91359)
Andrzej Warzyński
llvmlistbot at llvm.org
Tue May 7 09:46:58 PDT 2024
https://github.com/banach-space created https://github.com/llvm/llvm-project/pull/91359
Pretty much all logic that we have today for lowering vector.transpose
assumes fixed length vectors (it's done via vector.shuffle that don't
support scalable vectors). This patch updates related tests and patterns
to capture and document that limitation more explicitly.
Note that `vector.transpose` is a valid operation in the context of
scalable vectors, but we are yet to implement the missing lowerings.
Most changes are implemented in the test file. Here's a summary:
* @transpose_nx8x2xf32 is renamed as @transpose_scalable and move near
other test using `lowering_strategy = "shuffle_1d"` (to avoid
duplicating TD sequences)
* tests specific to X86 (`avx2_lowering_strategy = true`) are moved to
a dedicated file (to seperate generic tests from target-specific
tests)
* `@transpose10_nx4xnx1xf32` duplicated `@transpose10_4xnx1xf32` and was
deleted (the latter is renamed as `@transpose10_4x1xf32_scalable` to
match its fixed-width counterpart: `@transpose10_4x1xf32`)
* The changes in LowerVectorTranspose.cpp are NFCs - they just make sure
that "scalable" vectors are filtered out at the very beginning
>From 952f78a47d7d3012f9a263348e2accbc037b41eb Mon Sep 17 00:00:00 2001
From: Andrzej Warzynski <andrzej.warzynski at arm.com>
Date: Sat, 4 May 2024 14:59:13 +0000
Subject: [PATCH] [mlir][vector] Update tests/patterns for vector.transpose
Pretty much all logic that we have today for lowering vector.transpose
assumes fixed length vectors (it's done via vector.shuffle that don't
support scalable vectors). This patch updates related tests and patterns
to capture and document that limitation more explicitly.
Note that `vector.transpose` is a valid operation in the context of
scalable vectors, but we are yet to implement the missing lowerings.
Most changes are implemented in the test file. Here's a summary:
* @transpose_nx8x2xf32 is renamed as @transpose_scalable and move near
other test using `lowering_strategy = "shuffle_1d"` (to avoid
duplicating TD sequences)
* tests specific to X86 (`avx2_lowering_strategy = true`) are moved to
a dedicated file (to seperate generic tests from target-specific
tests)
* `@transpose10_nx4xnx1xf32` duplicated `@transpose10_4xnx1xf32` and was
deleted (the latter is renamed as `@transpose10_4x1xf32_scalable` to
match its fixed-width counterpart: `@transpose10_4x1xf32`)
* The changes in LowerVectorTranspose.cpp are NFCs - they just make sure
that "scalable" vectors are filtered out at the very beginning
---
.../Transforms/LowerVectorTranspose.cpp | 7 +-
.../Vector/vector-transpose-lowering.mlir | 561 ++----------------
2 files changed, 43 insertions(+), 525 deletions(-)
diff --git a/mlir/lib/Dialect/Vector/Transforms/LowerVectorTranspose.cpp b/mlir/lib/Dialect/Vector/Transforms/LowerVectorTranspose.cpp
index 792550dcfaf2..ebb289831453 100644
--- a/mlir/lib/Dialect/Vector/Transforms/LowerVectorTranspose.cpp
+++ b/mlir/lib/Dialect/Vector/Transforms/LowerVectorTranspose.cpp
@@ -320,6 +320,10 @@ class TransposeOpLowering : public OpRewritePattern<vector::TransposeOp> {
LogicalResult matchAndRewrite(vector::TransposeOp op,
PatternRewriter &rewriter) const override {
+ if (op.getSourceVectorType().isScalable())
+ return rewriter.notifyMatchFailure(
+ op, "scalable vectors are not supported by this pattern");
+
auto loc = op.getLoc();
Value input = op.getVector();
@@ -352,9 +356,6 @@ class TransposeOpLowering : public OpRewritePattern<vector::TransposeOp> {
return success();
}
- if (inputType.isScalable())
- return failure();
-
// Handle a true 2-D matrix transpose differently when requested.
if (vectorTransformOptions.vectorTransposeLowering ==
vector::VectorTransposeLowering::Flat &&
diff --git a/mlir/test/Dialect/Vector/vector-transpose-lowering.mlir b/mlir/test/Dialect/Vector/vector-transpose-lowering.mlir
index 628a8ce50959..219a72df52a1 100644
--- a/mlir/test/Dialect/Vector/vector-transpose-lowering.mlir
+++ b/mlir/test/Dialect/Vector/vector-transpose-lowering.mlir
@@ -110,6 +110,17 @@ func.func @transpose(%arg0: vector<2x4xf32>) -> vector<4x2xf32> {
return %0 : vector<4x2xf32>
}
+/// Scalable vectors are not supported
+
+// CHECK-LABEL: func @transpose_scalable
+// CHECK-NOT: vector.shuffle
+// CHECK-NOT: vector.shape_cast
+// CHECK: vector.transpose
+func.func @transpose_scalable(%arg0: vector<2x[4]xf32>) -> vector<[4]x2xf32> {
+ %0 = vector.transpose %arg0, [1, 0] : vector<2x[4]xf32> to vector<[4]x2xf32>
+ return %0 : vector<[4]x2xf32>
+}
+
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%root : !transform.any_op {transform.readonly}) {
@@ -132,502 +143,22 @@ func.func @transpose(%arg0: vector<2x4xf32>) -> vector<4x2xf32> {
return %0 : vector<4x2xf32>
}
+/// Scalable vectors are not supported
-module attributes {transform.with_named_sequence} {
- transform.named_sequence @__transform_main(%root : !transform.any_op {transform.readonly}) {
- %func_op = transform.structured.match ops{["func.func"]} in %root : (!transform.any_op) -> !transform.op<"func.func">
- transform.apply_patterns to %func_op {
- transform.apply_patterns.vector.lower_transpose lowering_strategy = "flat_transpose"
- } : !transform.op<"func.func">
- transform.yield
- }
-}
-
-// -----
-
-// CHECK-LABEL: func @transpose4x8
-func.func @transpose4x8xf32(%arg0: vector<4x8xf32>) -> vector<8x4xf32> {
- // CHECK: vector.extract {{.*}}[0]
- // CHECK-NEXT: vector.extract {{.*}}[1]
- // CHECK-NEXT: vector.extract {{.*}}[2]
- // CHECK-NEXT: vector.extract {{.*}}[3]
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 1, 8, 9, 4, 5, 12, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 3, 10, 11, 6, 7, 14, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 1, 8, 9, 4, 5, 12, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 3, 10, 11, 6, 7, 14, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 1, 2, 3, 8, 9, 10, 11] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 1, 2, 3, 8, 9, 10, 11] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [4, 5, 6, 7, 12, 13, 14, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [4, 5, 6, 7, 12, 13, 14, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.insert {{.*}}[0]
- // CHECK-NEXT: vector.insert {{.*}}[1]
- // CHECK-NEXT: vector.insert {{.*}}[2]
- // CHECK-NEXT: vector.insert {{.*}}[3]
- // CHECK-NEXT: vector.shape_cast {{.*}} vector<4x8xf32> to vector<32xf32>
- // CHECK-NEXT: vector.shape_cast {{.*}} vector<32xf32> to vector<8x4xf32>
- %0 = vector.transpose %arg0, [1, 0] : vector<4x8xf32> to vector<8x4xf32>
- return %0 : vector<8x4xf32>
-}
-
-// CHECK-LABEL: func @transpose021_1x4x8
-func.func @transpose021_1x4x8xf32(%arg0: vector<1x4x8xf32>) -> vector<1x8x4xf32> {
- // CHECK: vector.extract {{.*}}[0, 0]
- // CHECK-NEXT: vector.extract {{.*}}[0, 1]
- // CHECK-NEXT: vector.extract {{.*}}[0, 2]
- // CHECK-NEXT: vector.extract {{.*}}[0, 3]
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 1, 8, 9, 4, 5, 12, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 3, 10, 11, 6, 7, 14, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 1, 8, 9, 4, 5, 12, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 3, 10, 11, 6, 7, 14, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 1, 2, 3, 8, 9, 10, 11] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 1, 2, 3, 8, 9, 10, 11] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [4, 5, 6, 7, 12, 13, 14, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [4, 5, 6, 7, 12, 13, 14, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.insert {{.*}}[0]
- // CHECK-NEXT: vector.insert {{.*}}[1]
- // CHECK-NEXT: vector.insert {{.*}}[2]
- // CHECK-NEXT: vector.insert {{.*}}[3]
- // CHECK-NEXT: vector.shape_cast {{.*}} vector<4x8xf32> to vector<32xf32>
- // CHECK-NEXT: vector.shape_cast {{.*}} vector<32xf32> to vector<1x8x4xf32>
- %0 = vector.transpose %arg0, [0, 2, 1] : vector<1x4x8xf32> to vector<1x8x4xf32>
- return %0 : vector<1x8x4xf32>
-}
-
-// CHECK-LABEL: func @transpose8x8
-func.func @transpose8x8xf32(%arg0: vector<8x8xf32>) -> vector<8x8xf32> {
- // CHECK: vector.extract {{.*}}[0]
- // CHECK-NEXT: vector.extract {{.*}}[1]
- // CHECK-NEXT: vector.extract {{.*}}[2]
- // CHECK-NEXT: vector.extract {{.*}}[3]
- // CHECK-NEXT: vector.extract {{.*}}[4]
- // CHECK-NEXT: vector.extract {{.*}}[5]
- // CHECK-NEXT: vector.extract {{.*}}[6]
- // CHECK-NEXT: vector.extract {{.*}}[7]
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [2, 3, 8, 9, 6, 7, 12, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [0, 1, 2, 3, 8, 9, 10, 11] : vector<8xf32>, vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [4, 5, 6, 7, 12, 13, 14, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.insert {{.*}}[0]
- // CHECK-NEXT: vector.insert {{.*}}[1]
- // CHECK-NEXT: vector.insert {{.*}}[2]
- // CHECK-NEXT: vector.insert {{.*}}[3]
- // CHECK-NEXT: vector.insert {{.*}}[4]
- // CHECK-NEXT: vector.insert {{.*}}[5]
- // CHECK-NEXT: vector.insert {{.*}}[6]
- // CHECK-NEXT: vector.insert {{.*}}[7]
- %0 = vector.transpose %arg0, [1, 0] : vector<8x8xf32> to vector<8x8xf32>
- return %0 : vector<8x8xf32>
-}
-
-// CHECK-LABEL: func @transpose021_1x8x8
-func.func @transpose021_1x8x8xf32(%arg0: vector<1x8x8xf32>) -> vector<1x8x8xf32> {
- // CHECK: vector.extract {{.*}}[0, 0]
- // CHECK-NEXT: vector.extract {{.*}}[0, 1]
- // CHECK-NEXT: vector.extract {{.*}}[0, 2]
- // CHECK-NEXT: vector.extract {{.*}}[0, 3]
- // CHECK-NEXT: vector.extract {{.*}}[0, 4]
- // CHECK-NEXT: vector.extract {{.*}}[0, 5]
- // CHECK-NEXT: vector.extract {{.*}}[0, 6]
- // CHECK-NEXT: vector.extract {{.*}}[0, 7]
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [2, 3, 8, 9, 6, 7, 12, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [0, 1, 2, 3, 8, 9, 10, 11] : vector<8xf32>, vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [4, 5, 6, 7, 12, 13, 14, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.insert {{.*}}[0]
- // CHECK-NEXT: vector.insert {{.*}}[1]
- // CHECK-NEXT: vector.insert {{.*}}[2]
- // CHECK-NEXT: vector.insert {{.*}}[3]
- // CHECK-NEXT: vector.insert {{.*}}[4]
- // CHECK-NEXT: vector.insert {{.*}}[5]
- // CHECK-NEXT: vector.insert {{.*}}[6]
- // CHECK-NEXT: vector.insert {{.*}}[7]
- // CHECK-NEXT: vector.shape_cast %{{.*}} : vector<8x8xf32> to vector<1x8x8xf32>
- %0 = vector.transpose %arg0, [0, 2, 1] : vector<1x8x8xf32> to vector<1x8x8xf32>
- return %0 : vector<1x8x8xf32>
-}
-
-// CHECK-LABEL: func @transpose120_8x1x8
-func.func @transpose120_8x1x8xf32(%arg0: vector<8x1x8xf32>) -> vector<1x8x8xf32> {
- // CHECK: vector.extract {{.*}}[0, 0]
- // CHECK-NEXT: vector.extract {{.*}}[1, 0]
- // CHECK-NEXT: vector.extract {{.*}}[2, 0]
- // CHECK-NEXT: vector.extract {{.*}}[3, 0]
- // CHECK-NEXT: vector.extract {{.*}}[4, 0]
- // CHECK-NEXT: vector.extract {{.*}}[5, 0]
- // CHECK-NEXT: vector.extract {{.*}}[6, 0]
- // CHECK-NEXT: vector.extract {{.*}}[7, 0]
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [2, 3, 8, 9, 6, 7, 12, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [0, 1, 2, 3, 8, 9, 10, 11] : vector<8xf32>, vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [4, 5, 6, 7, 12, 13, 14, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.insert {{.*}}[0]
- // CHECK-NEXT: vector.insert {{.*}}[1]
- // CHECK-NEXT: vector.insert {{.*}}[2]
- // CHECK-NEXT: vector.insert {{.*}}[3]
- // CHECK-NEXT: vector.insert {{.*}}[4]
- // CHECK-NEXT: vector.insert {{.*}}[5]
- // CHECK-NEXT: vector.insert {{.*}}[6]
- // CHECK-NEXT: vector.insert {{.*}}[7]
- // CHECK-NEXT: vector.shape_cast %{{.*}} : vector<8x8xf32> to vector<1x8x8xf32>
- %0 = vector.transpose %arg0, [1, 2, 0] : vector<8x1x8xf32> to vector<1x8x8xf32>
- return %0 : vector<1x8x8xf32>
-}
-
-// CHECK-LABEL: func @transpose120_8x8x1
-func.func @transpose120_8x8x1xf32(%arg0: vector<8x8x1xf32>) -> vector<8x1x8xf32> {
- // CHECK: vector.shape_cast %{{.*}} : vector<8x8x1xf32> to vector<8x8xf32>
- // CHECK-NEXT: vector.extract {{.*}}[0]
- // CHECK-NEXT: vector.extract {{.*}}[1]
- // CHECK-NEXT: vector.extract {{.*}}[2]
- // CHECK-NEXT: vector.extract {{.*}}[3]
- // CHECK-NEXT: vector.extract {{.*}}[4]
- // CHECK-NEXT: vector.extract {{.*}}[5]
- // CHECK-NEXT: vector.extract {{.*}}[6]
- // CHECK-NEXT: vector.extract {{.*}}[7]
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [2, 3, 8, 9, 6, 7, 12, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [0, 1, 2, 3, 8, 9, 10, 11] : vector<8xf32>, vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [4, 5, 6, 7, 12, 13, 14, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.insert {{.*}}[0]
- // CHECK-NEXT: vector.insert {{.*}}[1]
- // CHECK-NEXT: vector.insert {{.*}}[2]
- // CHECK-NEXT: vector.insert {{.*}}[3]
- // CHECK-NEXT: vector.insert {{.*}}[4]
- // CHECK-NEXT: vector.insert {{.*}}[5]
- // CHECK-NEXT: vector.insert {{.*}}[6]
- // CHECK-NEXT: vector.insert {{.*}}[7]
- // CHECK-NEXT: vector.shape_cast %{{.*}} : vector<8x8xf32> to vector<8x1x8xf32>
- %0 = vector.transpose %arg0, [1, 2, 0] : vector<8x8x1xf32> to vector<8x1x8xf32>
- return %0 : vector<8x1x8xf32>
-}
-
-// CHECK-LABEL: func @transpose102_8x8x1
-func.func @transpose102_8x8x1xf32(%arg0: vector<8x8x1xf32>) -> vector<8x8x1xf32> {
- // CHECK: vector.shape_cast %{{.*}} : vector<8x8x1xf32> to vector<8x8xf32>
- // CHECK-NEXT: vector.extract {{.*}}[0]
- // CHECK-NEXT: vector.extract {{.*}}[1]
- // CHECK-NEXT: vector.extract {{.*}}[2]
- // CHECK-NEXT: vector.extract {{.*}}[3]
- // CHECK-NEXT: vector.extract {{.*}}[4]
- // CHECK-NEXT: vector.extract {{.*}}[5]
- // CHECK-NEXT: vector.extract {{.*}}[6]
- // CHECK-NEXT: vector.extract {{.*}}[7]
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [2, 3, 8, 9, 6, 7, 12, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [0, 1, 2, 3, 8, 9, 10, 11] : vector<8xf32>, vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [4, 5, 6, 7, 12, 13, 14, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.insert {{.*}}[0]
- // CHECK-NEXT: vector.insert {{.*}}[1]
- // CHECK-NEXT: vector.insert {{.*}}[2]
- // CHECK-NEXT: vector.insert {{.*}}[3]
- // CHECK-NEXT: vector.insert {{.*}}[4]
- // CHECK-NEXT: vector.insert {{.*}}[5]
- // CHECK-NEXT: vector.insert {{.*}}[6]
- // CHECK-NEXT: vector.insert {{.*}}[7]
- // CHECK-NEXT: vector.shape_cast %{{.*}} : vector<8x8xf32> to vector<8x8x1xf32>
- %0 = vector.transpose %arg0, [1, 0, 2] : vector<8x8x1xf32> to vector<8x8x1xf32>
- return %0 : vector<8x8x1xf32>
-}
-
-// CHECK-LABEL: func @transpose201_8x1x8
-func.func @transpose201_8x1x8xf32(%arg0: vector<8x1x8xf32>) -> vector<8x8x1xf32> {
- // CHECK: vector.extract {{.*}}[0, 0]
- // CHECK-NEXT: vector.extract {{.*}}[1, 0]
- // CHECK-NEXT: vector.extract {{.*}}[2, 0]
- // CHECK-NEXT: vector.extract {{.*}}[3, 0]
- // CHECK-NEXT: vector.extract {{.*}}[4, 0]
- // CHECK-NEXT: vector.extract {{.*}}[5, 0]
- // CHECK-NEXT: vector.extract {{.*}}[6, 0]
- // CHECK-NEXT: vector.extract {{.*}}[7, 0]
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [2, 3, 8, 9, 6, 7, 12, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [0, 1, 2, 3, 8, 9, 10, 11] : vector<8xf32>, vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [4, 5, 6, 7, 12, 13, 14, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.insert {{.*}}[0]
- // CHECK-NEXT: vector.insert {{.*}}[1]
- // CHECK-NEXT: vector.insert {{.*}}[2]
- // CHECK-NEXT: vector.insert {{.*}}[3]
- // CHECK-NEXT: vector.insert {{.*}}[4]
- // CHECK-NEXT: vector.insert {{.*}}[5]
- // CHECK-NEXT: vector.insert {{.*}}[6]
- // CHECK-NEXT: vector.insert {{.*}}[7]
- // CHECK-NEXT: vector.shape_cast %{{.*}} : vector<8x8xf32> to vector<8x8x1xf32>
- %0 = vector.transpose %arg0, [2, 0, 1] : vector<8x1x8xf32> to vector<8x8x1xf32>
- return %0 : vector<8x8x1xf32>
-}
-
-// CHECK-LABEL: func @transpose201_1x8x8
-func.func @transpose201_1x8x8xf32(%arg0: vector<1x8x8xf32>) -> vector<8x1x8xf32> {
- // CHECK: vector.extract {{.*}}[0, 0]
- // CHECK-NEXT: vector.extract {{.*}}[0, 1]
- // CHECK-NEXT: vector.extract {{.*}}[0, 2]
- // CHECK-NEXT: vector.extract {{.*}}[0, 3]
- // CHECK-NEXT: vector.extract {{.*}}[0, 4]
- // CHECK-NEXT: vector.extract {{.*}}[0, 5]
- // CHECK-NEXT: vector.extract {{.*}}[0, 6]
- // CHECK-NEXT: vector.extract {{.*}}[0, 7]
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [2, 3, 8, 9, 6, 7, 12, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [0, 1, 2, 3, 8, 9, 10, 11] : vector<8xf32>, vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [4, 5, 6, 7, 12, 13, 14, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.insert {{.*}}[0]
- // CHECK-NEXT: vector.insert {{.*}}[1]
- // CHECK-NEXT: vector.insert {{.*}}[2]
- // CHECK-NEXT: vector.insert {{.*}}[3]
- // CHECK-NEXT: vector.insert {{.*}}[4]
- // CHECK-NEXT: vector.insert {{.*}}[5]
- // CHECK-NEXT: vector.insert {{.*}}[6]
- // CHECK-NEXT: vector.insert {{.*}}[7]
- // CHECK-NEXT: vector.shape_cast %{{.*}} : vector<8x8xf32> to vector<8x1x8xf32>
- %0 = vector.transpose %arg0, [2, 0, 1] : vector<1x8x8xf32> to vector<8x1x8xf32>
- return %0 : vector<8x1x8xf32>
-}
-
-// CHECK-LABEL: func @transpose210_8x1x8
-func.func @transpose210_8x1x8xf32(%arg0: vector<8x1x8xf32>) -> vector<8x1x8xf32> {
- // CHECK: vector.extract {{.*}}[0, 0]
- // CHECK-NEXT: vector.extract {{.*}}[1, 0]
- // CHECK-NEXT: vector.extract {{.*}}[2, 0]
- // CHECK-NEXT: vector.extract {{.*}}[3, 0]
- // CHECK-NEXT: vector.extract {{.*}}[4, 0]
- // CHECK-NEXT: vector.extract {{.*}}[5, 0]
- // CHECK-NEXT: vector.extract {{.*}}[6, 0]
- // CHECK-NEXT: vector.extract {{.*}}[7, 0]
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [2, 3, 8, 9, 6, 7, 12, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [0, 1, 2, 3, 8, 9, 10, 11] : vector<8xf32>, vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [4, 5, 6, 7, 12, 13, 14, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.insert {{.*}}[0]
- // CHECK-NEXT: vector.insert {{.*}}[1]
- // CHECK-NEXT: vector.insert {{.*}}[2]
- // CHECK-NEXT: vector.insert {{.*}}[3]
- // CHECK-NEXT: vector.insert {{.*}}[4]
- // CHECK-NEXT: vector.insert {{.*}}[5]
- // CHECK-NEXT: vector.insert {{.*}}[6]
- // CHECK-NEXT: vector.insert {{.*}}[7]
- // CHECK-NEXT: vector.shape_cast %{{.*}} : vector<8x8xf32> to vector<8x1x8xf32>
- %0 = vector.transpose %arg0, [2, 1, 0] : vector<8x1x8xf32> to vector<8x1x8xf32>
- return %0 : vector<8x1x8xf32>
-}
-
-// CHECK-LABEL: func @transpose210_8x8x1
-func.func @transpose210_8x8x1xf32(%arg0: vector<8x8x1xf32>) -> vector<1x8x8xf32> {
- // CHECK: vector.shape_cast %{{.*}} : vector<8x8x1xf32> to vector<8x8xf32>
- // CHECK-NEXT: vector.extract {{.*}}[0]
- // CHECK-NEXT: vector.extract {{.*}}[1]
- // CHECK-NEXT: vector.extract {{.*}}[2]
- // CHECK-NEXT: vector.extract {{.*}}[3]
- // CHECK-NEXT: vector.extract {{.*}}[4]
- // CHECK-NEXT: vector.extract {{.*}}[5]
- // CHECK-NEXT: vector.extract {{.*}}[6]
- // CHECK-NEXT: vector.extract {{.*}}[7]
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [2, 3, 8, 9, 6, 7, 12, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [0, 1, 2, 3, 8, 9, 10, 11] : vector<8xf32>, vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [4, 5, 6, 7, 12, 13, 14, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.insert {{.*}}[0]
- // CHECK-NEXT: vector.insert {{.*}}[1]
- // CHECK-NEXT: vector.insert {{.*}}[2]
- // CHECK-NEXT: vector.insert {{.*}}[3]
- // CHECK-NEXT: vector.insert {{.*}}[4]
- // CHECK-NEXT: vector.insert {{.*}}[5]
- // CHECK-NEXT: vector.insert {{.*}}[6]
- // CHECK-NEXT: vector.insert {{.*}}[7]
- // CHECK-NEXT: vector.shape_cast %{{.*}} : vector<8x8xf32> to vector<1x8x8xf32>
- %0 = vector.transpose %arg0, [2, 1, 0] : vector<8x8x1xf32> to vector<1x8x8xf32>
- return %0 : vector<1x8x8xf32>
-}
-
-// CHECK-LABEL: func @transpose210_1x8x8
-func.func @transpose210_1x8x8xf32(%arg0: vector<1x8x8xf32>) -> vector<8x8x1xf32> {
- // CHECK: vector.extract {{.*}}[0, 0]
- // CHECK-NEXT: vector.extract {{.*}}[0, 1]
- // CHECK-NEXT: vector.extract {{.*}}[0, 2]
- // CHECK-NEXT: vector.extract {{.*}}[0, 3]
- // CHECK-NEXT: vector.extract {{.*}}[0, 4]
- // CHECK-NEXT: vector.extract {{.*}}[0, 5]
- // CHECK-NEXT: vector.extract {{.*}}[0, 6]
- // CHECK-NEXT: vector.extract {{.*}}[0, 7]
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [0, 8, 1, 9, 4, 12, 5, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.shuffle {{.*}} [2, 10, 3, 11, 6, 14, 7, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [2, 3, 8, 9, 6, 7, 12, 13] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0xcc", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-NEXT: llvm.inline_asm asm_dialect = intel "vblendps $0, $1, $2, 0x33", "=x,x,x" {{.*}} : (vector<8xf32>, vector<8xf32>) -> vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [0, 1, 2, 3, 8, 9, 10, 11] : vector<8xf32>, vector<8xf32>
- // CHECK-COUNT-4: vector.shuffle {{.*}} [4, 5, 6, 7, 12, 13, 14, 15] : vector<8xf32>, vector<8xf32>
- // CHECK-NEXT: vector.insert {{.*}}[0]
- // CHECK-NEXT: vector.insert {{.*}}[1]
- // CHECK-NEXT: vector.insert {{.*}}[2]
- // CHECK-NEXT: vector.insert {{.*}}[3]
- // CHECK-NEXT: vector.insert {{.*}}[4]
- // CHECK-NEXT: vector.insert {{.*}}[5]
- // CHECK-NEXT: vector.insert {{.*}}[6]
- // CHECK-NEXT: vector.insert {{.*}}[7]
- // CHECK-NEXT: vector.shape_cast %{{.*}} : vector<8x8xf32> to vector<8x8x1xf32>
- %0 = vector.transpose %arg0, [2, 1, 0] : vector<1x8x8xf32> to vector<8x8x1xf32>
- return %0 : vector<8x8x1xf32>
+// CHECK-LABEL: func @transpose_scalable(
+func.func @transpose_scalable(%arg0: vector<2x[4]xf32>) -> vector<[4]x2xf32> {
+ // CHECK-NOT: vector.shape_cast
+ // CHECK-NOT: vector.flat_transpose
+ // CHECK: vector.transpose
+ %0 = vector.transpose %arg0, [1, 0] : vector<2x[4]xf32> to vector<[4]x2xf32>
+ return %0 : vector<[4]x2xf32>
}
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%root : !transform.any_op {transform.readonly}) {
%func_op = transform.structured.match ops{["func.func"]} in %root : (!transform.any_op) -> !transform.op<"func.func">
transform.apply_patterns to %func_op {
- transform.apply_patterns.vector.lower_transpose avx2_lowering_strategy = true
+ transform.apply_patterns.vector.lower_transpose lowering_strategy = "flat_transpose"
} : !transform.op<"func.func">
transform.yield
}
@@ -635,6 +166,7 @@ module attributes {transform.with_named_sequence} {
// -----
+// CHECK-LABEL: @transpose_shuffle16x16xf32(
func.func @transpose_shuffle16x16xf32(%arg0: vector<16x16xf32>) -> vector<16x16xf32> {
// CHECK: vector.shuffle {{.*}} [0, 16, 1, 17, 4, 20, 5, 21, 8, 24, 9, 25, 12, 28, 13, 29] : vector<16xf32>, vector<16xf32>
// CHECK: vector.shuffle {{.*}} [2, 18, 3, 19, 6, 22, 7, 23, 10, 26, 11, 27, 14, 30, 15, 31] : vector<16xf32>, vector<16xf32>
@@ -704,6 +236,14 @@ func.func @transpose_shuffle16x16xf32(%arg0: vector<16x16xf32>) -> vector<16x16x
return %0 : vector<16x16xf32>
}
+// CHECK-LABEL: @transpose_shuffle16x16xf32_scalable(
+func.func @transpose_shuffle16x16xf32_scalable(%arg0: vector<16x[16]xf32>) -> vector<[16]x16xf32> {
+ // CHECK-NOT: vector.shuffle
+ // CHECK: vector.transpose
+ %0 = vector.transpose %arg0, [1, 0] : vector<16x[16]xf32> to vector<[16]x16xf32>
+ return %0 : vector<[16]x16xf32>
+}
+
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%root : !transform.any_op {transform.readonly}) {
%func_op = transform.structured.match ops{["func.func"]} in %root : (!transform.any_op) -> !transform.op<"func.func">
@@ -786,6 +326,14 @@ func.func @transpose021_shuffle16x16xf32(%arg0: vector<1x16x16xf32>) -> vector<1
return %0 : vector<1x16x16xf32>
}
+// CHECK-LABEL: func @transpose021_shuffle16x16xf32_scalable
+func.func @transpose021_shuffle16x16xf32_scalable(%arg0: vector<1x16x[16]xf32>) -> vector<1x[16]x16xf32> {
+ // CHECK-NOT: vector.shuffle
+ // CHECK: vector.transpose
+ %0 = vector.transpose %arg0, [0, 2, 1] : vector<1x16x[16]xf32> to vector<1x[16]x16xf32>
+ return %0 : vector<1x[16]x16xf32>
+}
+
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%root : !transform.any_op {transform.readonly}) {
%func_op = transform.structured.match ops{["func.func"]} in %root : (!transform.any_op) -> !transform.op<"func.func">
@@ -830,23 +378,14 @@ func.func @transpose10_1xnx4xf32(%arg0: vector<1x[4]xf32>) -> vector<[4]x1xf32>
/// Scalable unit dim should not be lowered to shape_cast.
-// CHECK-LABEL: func @transpose10_4xnx1xf32
-func.func @transpose10_4xnx1xf32(%arg0: vector<4x[1]xf32>) -> vector<[1]x4xf32> {
+// CHECK-LABEL: func @transpose10_4x1xf32_scalable
+func.func @transpose10_4x1xf32_scalable(%arg0: vector<4x[1]xf32>) -> vector<[1]x4xf32> {
// CHECK-NOT: vector.shape_cast
// CHECK: vector.transpose %{{.*}} : vector<4x[1]xf32> to vector<[1]x4xf32>
%0 = vector.transpose %arg0, [1, 0] : vector<4x[1]xf32> to vector<[1]x4xf32>
return %0 : vector<[1]x4xf32>
}
-// CHECK-LABEL: func @transpose10_nx4xnx1xf32
-func.func @transpose10_nx4xnx1xf32(%arg0: vector<4x[1]xf32>) -> vector<[1]x4xf32> {
- // CHECK-NOT: vector.shape_cast
- // CHECK: vector.transpose %{{.*}} : vector<4x[1]xf32> to vector<[1]x4xf32>
- %0 = vector.transpose %arg0, [1, 0] : vector<4x[1]xf32> to vector<[1]x4xf32>
-
- return %0 : vector<[1]x4xf32>
-}
-
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%root : !transform.any_op {transform.readonly}) {
%func_op = transform.structured.match ops{["func.func"]} in %root : (!transform.any_op) -> !transform.op<"func.func">
@@ -856,25 +395,3 @@ module attributes {transform.with_named_sequence} {
transform.yield
}
}
-
-// -----
-
-// Scalable transposes should not be lowered to vector.shuffle.
-
-// CHECK-LABEL: func @transpose_nx8x2xf32
-func.func @transpose_nx8x2xf32(%arg0: vector<[8]x2xf32>) -> vector<2x[8]xf32> {
- // CHECK-NOT: vector.shuffle
- // CHECK: vector.transpose %{{.*}} : vector<[8]x2xf32> to vector<2x[8]xf32>
- %0 = vector.transpose %arg0, [1, 0] : vector<[8]x2xf32> to vector<2x[8]xf32>
- return %0 : vector<2x[8]xf32>
-}
-
-module attributes {transform.with_named_sequence} {
- transform.named_sequence @__transform_main(%root : !transform.any_op {transform.readonly}) {
- %func_op = transform.structured.match ops{["func.func"]} in %root : (!transform.any_op) -> !transform.op<"func.func">
- transform.apply_patterns to %func_op {
- transform.apply_patterns.vector.lower_transpose lowering_strategy = "shuffle_1d"
- } : !transform.op<"func.func">
- transform.yield
- }
-}
More information about the Mlir-commits
mailing list