Plumb through tensor.pack e2e execution for llvm-cpu backend. (#11875)
All the tensor.pack ops with static inner_tile_sizes are vectorized, which are all covered by e2e tests.
diff --git a/compiler/src/iree/compiler/Codegen/Interfaces/PartitionableLoopsInterface.cpp b/compiler/src/iree/compiler/Codegen/Interfaces/PartitionableLoopsInterface.cpp
index 02abcb1..b6e6251 100644
--- a/compiler/src/iree/compiler/Codegen/Interfaces/PartitionableLoopsInterface.cpp
+++ b/compiler/src/iree/compiler/Codegen/Interfaces/PartitionableLoopsInterface.cpp
@@ -94,7 +94,7 @@
// loops, but that needs the interface to return the static sizes of the
// loops.
SmallVector<unsigned> partitionableLoops;
- auto interfaceOp = cast<OpTy>(op);
+ auto interfaceOp = cast<TilingInterface>(op);
for (auto [index, iteratorType] :
llvm::enumerate(interfaceOp.getLoopIteratorTypes())) {
if (iteratorType != utils::IteratorType::parallel) {
@@ -241,6 +241,10 @@
IREE::LinalgExt::AttentionOp::attachInterface<
AllParallelAsPartitionableLoops<IREE::LinalgExt::AttentionOp>>(*ctx);
});
+ registry.addExtension(+[](MLIRContext *ctx, tensor::TensorDialect *dialect) {
+ tensor::PackOp::attachInterface<
+ OuterParallelAsPartitionableLoops<tensor::PackOp>>(*ctx);
+ });
}
} // namespace iree_compiler
diff --git a/compiler/src/iree/compiler/Codegen/LLVMCPU/KernelDispatch.cpp b/compiler/src/iree/compiler/Codegen/LLVMCPU/KernelDispatch.cpp
index e865d2d..9f3d7eb 100644
--- a/compiler/src/iree/compiler/Codegen/LLVMCPU/KernelDispatch.cpp
+++ b/compiler/src/iree/compiler/Codegen/LLVMCPU/KernelDispatch.cpp
@@ -1068,10 +1068,15 @@
return workgroupTileSizes;
}
-static LogicalResult setRootConfig(func::FuncOp entryPointFn,
- IREE::LinalgExt::PackOp op) {
- SmallVector<int64_t> tileSizes =
- getLinalgExtDefaultWorkgroupTileSizes(op, defaultWorkgroupTileSize);
+template <typename OpTy>
+static LogicalResult setPackOpRootConfig(func::FuncOp entryPointFn, OpTy op) {
+ // TODO(hanchung): Retire IREE::LinalgExt::PackOp. This is for having
+ // consistent configurations for pack ops.
+ static_assert(
+ llvm::is_one_of<OpTy, IREE::LinalgExt::PackOp, tensor::PackOp>::value,
+ "applies to only pack operations");
+ SmallVector<int64_t> tileSizes = getLinalgExtDefaultWorkgroupTileSizes(
+ cast<TilingInterface>(op.getOperation()), defaultWorkgroupTileSize);
// The default function aims to returns the number of workload per workgroup,
// but it does not know that it is working on packed domain. We need to take
@@ -1702,14 +1707,16 @@
return setRootConfig(entryPointFn, op, LinalgOpInfo(op),
targetMLTransInfo);
})
- .Case<IREE::LinalgExt::FftOp, IREE::LinalgExt::PackOp,
- IREE::LinalgExt::UnPackOp, linalg::Mmt4DOp,
- linalg::Conv2DNhwcHwcfOp, linalg::Conv2DNchwFchwOp,
- linalg::PoolingNhwcSumOp, linalg::PoolingNhwcMaxOp,
- linalg::PoolingNhwcMaxUnsignedOp, linalg::PoolingNhwcMinOp,
- linalg::PoolingNhwcMinUnsignedOp, linalg::PoolingNchwSumOp,
- linalg::PoolingNchwMaxOp, linalg::DepthwiseConv2DNhwcHwcOp>(
+ .Case<IREE::LinalgExt::FftOp, IREE::LinalgExt::UnPackOp,
+ linalg::Mmt4DOp, linalg::Conv2DNhwcHwcfOp,
+ linalg::Conv2DNchwFchwOp, linalg::PoolingNhwcSumOp,
+ linalg::PoolingNhwcMaxOp, linalg::PoolingNhwcMaxUnsignedOp,
+ linalg::PoolingNhwcMinOp, linalg::PoolingNhwcMinUnsignedOp,
+ linalg::PoolingNchwSumOp, linalg::PoolingNchwMaxOp,
+ linalg::DepthwiseConv2DNhwcHwcOp>(
[&](auto op) { return setRootConfig(entryPointFn, op); })
+ .Case<IREE::LinalgExt::PackOp, tensor::PackOp>(
+ [&](auto op) { return setPackOpRootConfig(entryPointFn, op); })
.Case<linalg::ContractionOpInterface>(
[&](auto op) { return setRootConfig(entryPointFn, op); })
.Case<linalg::LinalgOp>(
diff --git a/compiler/src/iree/compiler/Codegen/LLVMCPU/Passes.cpp b/compiler/src/iree/compiler/Codegen/LLVMCPU/Passes.cpp
index e009447..c2903e2 100644
--- a/compiler/src/iree/compiler/Codegen/LLVMCPU/Passes.cpp
+++ b/compiler/src/iree/compiler/Codegen/LLVMCPU/Passes.cpp
@@ -665,6 +665,8 @@
OpPassManager &nestedModulePM = passManager.nest<ModuleOp>();
nestedModulePM.addNestedPass<func::FuncOp>(
IREE::LinalgExt::createLinalgExtVectorizationPass());
+ nestedModulePM.addNestedPass<func::FuncOp>(
+ createVectorizePackUnPackOpsPass());
addBufferizePasses(nestedModulePM);
nestedModulePM.addNestedPass<func::FuncOp>(
createSplitFullPartialTransferPass("linalg-copy"));
diff --git a/compiler/src/iree/compiler/Dialect/Flow/Transforms/FormDispatchRegions.cpp b/compiler/src/iree/compiler/Dialect/Flow/Transforms/FormDispatchRegions.cpp
index 6258e0d..6036f2c 100644
--- a/compiler/src/iree/compiler/Dialect/Flow/Transforms/FormDispatchRegions.cpp
+++ b/compiler/src/iree/compiler/Dialect/Flow/Transforms/FormDispatchRegions.cpp
@@ -196,7 +196,9 @@
}
return !isa<linalg::FillOp>(op);
}
- return isa<TilingInterface>(op) ||
+ // tensor::PadOp fusion is not ready. Explicitly marking it not a root op for
+ // now.
+ return (isa<TilingInterface>(op) && !isa<tensor::PadOp>(op)) ||
isa<LinalgExt::SetEncodingOp, LinalgExt::UnsetEncodingOp>(op);
}
@@ -676,7 +678,8 @@
// Only look for Linalg ops here. Avoid moving `linalg.fill` that aren't
// fused with anything else into their own dispatches since it is better
// to convert them to splats.
- if (!isa<linalg::LinalgOp>(op) || isa<linalg::FillOp>(op)) continue;
+ if (!isa<linalg::LinalgOp, tensor::PackOp>(op) || isa<linalg::FillOp>(op))
+ continue;
unsigned newGroup = numRootOps++;
setRootAttribute(context, &op, newGroup);
diff --git a/compiler/src/iree/compiler/Dialect/Flow/Transforms/InitializeEmptyTensors.cpp b/compiler/src/iree/compiler/Dialect/Flow/Transforms/InitializeEmptyTensors.cpp
index 93a05ed..acb6af2 100644
--- a/compiler/src/iree/compiler/Dialect/Flow/Transforms/InitializeEmptyTensors.cpp
+++ b/compiler/src/iree/compiler/Dialect/Flow/Transforms/InitializeEmptyTensors.cpp
@@ -39,7 +39,8 @@
LogicalResult matchAndRewrite(tensor::EmptyOp emptyTensorOp,
PatternRewriter &rewriter) const override {
if (llvm::all_of(emptyTensorOp->getUsers(), [](Operation *user) -> bool {
- return isa<linalg::LinalgOp, LinalgExt::LinalgExtOp>(user);
+ return isa<linalg::LinalgOp, LinalgExt::LinalgExtOp, tensor::PackOp>(
+ user);
})) {
return failure();
}
@@ -66,7 +67,8 @@
LogicalResult matchAndRewrite(tensor::EmptyOp emptyTensorOp,
PatternRewriter &rewriter) const override {
if (llvm::all_of(emptyTensorOp->getUsers(), [](Operation *user) -> bool {
- return isa<linalg::LinalgOp, LinalgExt::LinalgExtOp>(user);
+ return isa<linalg::LinalgOp, LinalgExt::LinalgExtOp, tensor::PackOp>(
+ user);
})) {
return failure();
}
diff --git a/tests/e2e/tensor_ops/BUILD b/tests/e2e/tensor_ops/BUILD
index 5d5b679..d2234b2 100644
--- a/tests/e2e/tensor_ops/BUILD
+++ b/tests/e2e/tensor_ops/BUILD
@@ -38,6 +38,7 @@
# keep sorted
[
"extract_slice.mlir",
+ "pack.mlir",
"tensor_insert_slice.mlir",
],
include = ["*.mlir"],
@@ -59,6 +60,7 @@
],
include = ["*.mlir"],
exclude = [
+ "pack.mlir",
"tensor_cast.mlir",
],
),
@@ -83,6 +85,7 @@
],
include = ["*.mlir"],
exclude = [
+ "pack.mlir",
"tensor_cast.mlir",
],
),
diff --git a/tests/e2e/tensor_ops/CMakeLists.txt b/tests/e2e/tensor_ops/CMakeLists.txt
index b3defcd..22b49c6 100644
--- a/tests/e2e/tensor_ops/CMakeLists.txt
+++ b/tests/e2e/tensor_ops/CMakeLists.txt
@@ -31,6 +31,7 @@
check_llvm-cpu_local-task
SRCS
"extract_slice.mlir"
+ "pack.mlir"
"tensor_insert_slice.mlir"
TARGET_BACKEND
"llvm-cpu"
diff --git a/tests/e2e/tensor_ops/pack.mlir b/tests/e2e/tensor_ops/pack.mlir
new file mode 100644
index 0000000..6569471
--- /dev/null
+++ b/tests/e2e/tensor_ops/pack.mlir
@@ -0,0 +1,498 @@
+func.func private @generate_2D_source(%height : index, %width : index) -> tensor<?x?xi32> {
+ %init_source = tensor.empty(%height, %width) : tensor<?x?xi32>
+ %source = linalg.generic {
+ indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>],
+ iterator_types = ["parallel", "parallel"]}
+ outs(%init_source : tensor<?x?xi32>) {
+ ^bb0(%b0 : i32):
+ %outer = linalg.index 0 : index
+ %inner = linalg.index 1 : index
+ %strided = arith.muli %outer, %width : index
+ %linearized = arith.addi %inner, %strided : index
+ %linearized_i32 = arith.index_cast %linearized : index to i32
+ linalg.yield %linearized_i32 : i32
+ } -> tensor<?x?xi32>
+ return %source : tensor<?x?xi32>
+}
+
+func.func @static_pack_simple() {
+ %iree_input = util.unfoldable_constant dense<[[0, 1, 2, 3], [4, 5, 6, 7], [8, 9, 10, 11], [12, 13, 14, 15]]> : tensor<4x4xi32>
+ %init = tensor.empty() : tensor<2x2x2x2xi32>
+ %pack = tensor.pack %iree_input inner_dims_pos = [0, 1] inner_tiles = [2, 2] into %init
+ : tensor<4x4xi32> -> tensor<2x2x2x2xi32>
+ check.expect_eq_const(%pack, dense<[[[[0, 1], [4, 5]], [[2, 3], [6, 7]]], [[[8, 9], [12, 13]], [[10 ,11], [14, 15]]]]> : tensor<2x2x2x2xi32>) : tensor<2x2x2x2xi32>
+ return
+}
+
+func.func @dynamic_pack_simple() {
+ %iree_input = flow.tensor.constant dense<[
+ [0, 1, 2, 3],
+ [4, 5, 6, 7],
+ [8, 9, 10, 11],
+ [12, 13, 14, 15]]> : tensor<4x4xi32> -> tensor<?x?xi32>
+ %c0 = arith.constant 0 : index
+ %c1 = arith.constant 1 : index
+ %c2 = arith.constant 2 : index
+ %in_d0 = tensor.dim %iree_input, %c0 : tensor<?x?xi32>
+ %in_d1 = tensor.dim %iree_input, %c1 : tensor<?x?xi32>
+ %out_d0 = arith.ceildivui %in_d0, %c2 : index
+ %out_d1 = arith.ceildivui %in_d1, %c2 : index
+ %init = tensor.empty(%out_d0, %out_d1) : tensor<?x?x2x2xi32>
+ %pack = tensor.pack %iree_input inner_dims_pos = [0, 1] inner_tiles = [2, 2] into %init
+ : tensor<?x?xi32> -> tensor<?x?x2x2xi32>
+ %cast = tensor.cast %pack : tensor<?x?x2x2xi32> to tensor<2x2x2x2xi32>
+ check.expect_eq_const(%cast, dense<[[[[0, 1], [4, 5]], [[2, 3], [6, 7]]], [[[8, 9], [12, 13]], [[10 ,11], [14, 15]]]]> : tensor<2x2x2x2xi32>) : tensor<2x2x2x2xi32>
+ return
+}
+
+func.func @static_pack_simple_pad_mode() {
+ %iree_input = util.unfoldable_constant dense<[[0, 1, 2, 3], [4, 5, 6, 7], [8, 9, 10, 11], [12, 13, 14, 15]]> : tensor<4x4xi32>
+ %pad = arith.constant 0 : i32
+ %init = tensor.empty() : tensor<2x2x3x3xi32>
+ %pack = tensor.pack %iree_input padding_value(%pad : i32) inner_dims_pos = [0, 1] inner_tiles = [3, 3] into %init
+ : tensor<4x4xi32> -> tensor<2x2x3x3xi32>
+ // After padding, the input is
+ // 0, 1, 2, 3, 0, 0
+ // 4, 5, 6, 7, 0, 0
+ // 8, 9, 10, 11, 0, 0
+ // 12, 13, 14, 15, 0, 0
+ // 0, 0, 0, 0, 0, 0
+ // 0, 0, 0, 0, 0, 0
+ check.expect_eq_const(%pack, dense<[[[[0, 1, 2], [4, 5, 6], [8, 9, 10]],
+ [[3, 0, 0], [7, 0, 0], [11, 0, 0]]],
+ [[[12, 13, 14], [0, 0, 0], [0, 0, 0]],
+ [[15, 0, 0], [0, 0, 0], [0, 0, 0]]]]> : tensor<2x2x3x3xi32>) : tensor<2x2x3x3xi32>
+ return
+}
+
+func.func @dynamic_pack_simple_pad_mode() {
+ %iree_input = flow.tensor.constant dense<[
+ [0, 1, 2, 3],
+ [4, 5, 6, 7],
+ [8, 9, 10, 11],
+ [12, 13, 14, 15]]> : tensor<4x4xi32> -> tensor<?x?xi32>
+ %pad = arith.constant 0 : i32
+ %c0 = arith.constant 0 : index
+ %c1 = arith.constant 1 : index
+ %c3 = arith.constant 3 : index
+ %in_d0 = tensor.dim %iree_input, %c0 : tensor<?x?xi32>
+ %in_d1 = tensor.dim %iree_input, %c1 : tensor<?x?xi32>
+ %out_d0 = arith.ceildivui %in_d0, %c3 : index
+ %out_d1 = arith.ceildivui %in_d1, %c3 : index
+ %init = tensor.empty(%out_d0, %out_d1) : tensor<?x?x3x3xi32>
+ %pack = tensor.pack %iree_input padding_value(%pad : i32) inner_dims_pos = [0, 1] inner_tiles = [3, 3] into %init
+ : tensor<?x?xi32> -> tensor<?x?x3x3xi32>
+ %cast = tensor.cast %pack : tensor<?x?x3x3xi32> to tensor<2x2x3x3xi32>
+ check.expect_eq_const(%cast, dense<[[[[0, 1, 2], [4, 5, 6], [8, 9, 10]],
+ [[3, 0, 0], [7, 0, 0], [11, 0, 0]]],
+ [[[12, 13, 14], [0, 0, 0], [0, 0, 0]],
+ [[15, 0, 0], [0, 0, 0], [0, 0, 0]]]]> : tensor<2x2x3x3xi32>) : tensor<2x2x3x3xi32>
+ return
+}
+
+func.func @static_pack_large() {
+ %height = arith.constant 128 : index
+ %width = arith.constant 256 : index
+ %0 = call @generate_2D_source(%height, %width) : (index, index) -> tensor<?x?xi32>
+ %source = tensor.cast %0 : tensor<?x?xi32> to tensor<128x256xi32>
+
+ %init_pack = tensor.empty() : tensor<4x16x32x16xi32>
+ %pack = tensor.pack %source inner_dims_pos = [0, 1] inner_tiles = [32, 16] into %init_pack
+ : tensor<128x256xi32> -> tensor<4x16x32x16xi32>
+
+ // Pack without padding is just a reshape followed by a transpose.
+ %reshape = tensor.expand_shape %source [[0, 1], [2, 3]] : tensor<128x256xi32> into tensor<4x32x16x16xi32>
+ %init_transpose = tensor.empty() : tensor<4x16x32x16xi32>
+ %transpose = linalg.transpose
+ ins(%reshape : tensor<4x32x16x16xi32>)
+ outs(%init_transpose : tensor<4x16x32x16xi32>)
+ permutation = [0, 2, 1, 3]
+ check.expect_eq(%pack, %transpose) : tensor<4x16x32x16xi32>
+ return
+}
+
+func.func @dynamic_pack_large() {
+ %d0 = util.unfoldable_constant 128 : index
+ %d1 = util.unfoldable_constant 256 : index
+ %source = call @generate_2D_source(%d0, %d1) : (index, index) -> tensor<?x?xi32>
+
+ %c32 = arith.constant 32 : index
+ %c16 = arith.constant 16 : index
+ %tiled_d0 = arith.ceildivui %d0, %c32 : index
+ %tiled_d1 = arith.ceildivui %d1, %c16 : index
+ %dyn_init_pack = tensor.empty(%tiled_d0, %tiled_d1) : tensor<?x?x32x16xi32>
+ %pack = tensor.pack %source inner_dims_pos = [0, 1] inner_tiles = [32, 16] into %dyn_init_pack
+ : tensor<?x?xi32> -> tensor<?x?x32x16xi32>
+ %cast_pack = tensor.cast %pack : tensor<?x?x32x16xi32> to tensor<4x16x32x16xi32>
+
+ %c128 = arith.constant 128 : index
+ %c256 = arith.constant 256 : index
+ %source2 = call @generate_2D_source(%c128, %c256) : (index, index) -> tensor<?x?xi32>
+ %static_source = tensor.cast %source2 : tensor<?x?xi32> to tensor<128x256xi32>
+ %reshape = tensor.expand_shape %static_source [[0, 1], [2, 3]] : tensor<128x256xi32> into tensor<4x32x16x16xi32>
+ %init_transpose = tensor.empty() : tensor<4x16x32x16xi32>
+ %transpose = linalg.transpose
+ ins(%reshape : tensor<4x32x16x16xi32>)
+ outs(%init_transpose : tensor<4x16x32x16xi32>)
+ permutation = [0, 2, 1, 3]
+ check.expect_eq(%cast_pack, %transpose) : tensor<4x16x32x16xi32>
+ return
+}
+
+func.func @static_pack_transpose_inner_dims_large() {
+ %height = arith.constant 128 : index
+ %width = arith.constant 256 : index
+ %0 = call @generate_2D_source(%height, %width) : (index, index) -> tensor<?x?xi32>
+ %source = tensor.cast %0 : tensor<?x?xi32> to tensor<128x256xi32>
+
+ %init_pack = tensor.empty() : tensor<4x16x16x32xi32>
+ %pack = tensor.pack %source inner_dims_pos = [1, 0] inner_tiles = [16, 32] into %init_pack
+ : tensor<128x256xi32> -> tensor<4x16x16x32xi32>
+ %reshape = tensor.expand_shape %source [[0, 1], [2, 3]] : tensor<128x256xi32> into tensor<4x32x16x16xi32>
+ %init_transpose = tensor.empty() : tensor<4x16x16x32xi32>
+ %transpose = linalg.transpose
+ ins(%reshape : tensor<4x32x16x16xi32>)
+ outs(%init_transpose : tensor<4x16x16x32xi32>)
+ permutation = [0, 2, 3, 1]
+
+ check.expect_eq(%pack, %transpose) : tensor<4x16x16x32xi32>
+ return
+}
+
+func.func @dynamic_pack_transpose_inner_dims_large() {
+ %d0 = util.unfoldable_constant 128 : index
+ %d1 = util.unfoldable_constant 256 : index
+ %source = call @generate_2D_source(%d0, %d1) : (index, index) -> tensor<?x?xi32>
+
+ %c32 = arith.constant 32 : index
+ %c16 = arith.constant 16 : index
+ %tiled_d0 = arith.ceildivui %d0, %c32 : index
+ %tiled_d1 = arith.ceildivui %d1, %c16 : index
+ %dyn_init_pack = tensor.empty(%tiled_d0, %tiled_d1) : tensor<?x?x16x32xi32>
+ %pack = tensor.pack %source inner_dims_pos = [1, 0] inner_tiles = [16, 32] into %dyn_init_pack
+ : tensor<?x?xi32> -> tensor<?x?x16x32xi32>
+ %cast_pack = tensor.cast %pack : tensor<?x?x16x32xi32> to tensor<4x16x16x32xi32>
+
+ %c128 = arith.constant 128 : index
+ %c256 = arith.constant 256 : index
+ %source2 = call @generate_2D_source(%c128, %c256) : (index, index) -> tensor<?x?xi32>
+ %static_source = tensor.cast %source2 : tensor<?x?xi32> to tensor<128x256xi32>
+ %reshape = tensor.expand_shape %static_source [[0, 1], [2, 3]] : tensor<128x256xi32> into tensor<4x32x16x16xi32>
+ %init_transpose = tensor.empty() : tensor<4x16x16x32xi32>
+ %transpose = linalg.transpose
+ ins(%reshape : tensor<4x32x16x16xi32>)
+ outs(%init_transpose : tensor<4x16x16x32xi32>)
+ permutation = [0, 2, 3, 1]
+
+ check.expect_eq(%cast_pack, %transpose) : tensor<4x16x16x32xi32>
+ return
+}
+
+func.func @static_pack_pad_large() {
+ %height = arith.constant 100 : index
+ %width = arith.constant 250 : index
+ %0 = call @generate_2D_source(%height, %width) : (index, index) -> tensor<?x?xi32>
+ %source = tensor.cast %0 : tensor<?x?xi32> to tensor<100x250xi32>
+ %padding_value = arith.constant 42 : i32
+
+ %init_pack = tensor.empty() : tensor<4x16x32x16xi32>
+ %pack = tensor.pack %source padding_value(%padding_value : i32)
+ inner_dims_pos = [0, 1] inner_tiles = [32, 16] into %init_pack
+ : tensor<100x250xi32> -> tensor<4x16x32x16xi32>
+
+ %pad = tensor.pad %source low[0, 0] high[28, 6] {
+ ^bb0(%b0 : index, %b1 : index):
+ tensor.yield %padding_value : i32
+ } : tensor<100x250xi32> to tensor<128x256xi32>
+ %reshape = tensor.expand_shape %pad [[0, 1], [2, 3]] : tensor<128x256xi32> into tensor<4x32x16x16xi32>
+ %init_transpose = tensor.empty() : tensor<4x16x32x16xi32>
+ %transpose = linalg.transpose
+ ins(%reshape : tensor<4x32x16x16xi32>)
+ outs(%init_transpose : tensor<4x16x32x16xi32>)
+ permutation = [0, 2, 1, 3]
+
+ check.expect_eq(%pack, %transpose) : tensor<4x16x32x16xi32>
+ return
+}
+
+func.func @static_pack_pad_transpose_outer_dims_large() {
+ %height = arith.constant 100 : index
+ %width = arith.constant 250 : index
+ %0 = call @generate_2D_source(%height, %width) : (index, index) -> tensor<?x?xi32>
+ %source = tensor.cast %0 : tensor<?x?xi32> to tensor<100x250xi32>
+ %padding_value = arith.constant 42 : i32
+
+ %init_pack = tensor.empty() : tensor<16x4x32x16xi32>
+ %pack = tensor.pack %source padding_value(%padding_value : i32)
+ outer_dims_perm = [1, 0] inner_dims_pos = [0, 1] inner_tiles = [32, 16] into %init_pack
+ : tensor<100x250xi32> -> tensor<16x4x32x16xi32>
+
+ %pad = tensor.pad %source low[0, 0] high[28, 6] {
+ ^bb0(%b0 : index, %b1 : index):
+ tensor.yield %padding_value : i32
+ } : tensor<100x250xi32> to tensor<128x256xi32>
+ %reshape = tensor.expand_shape %pad [[0, 1], [2, 3]] : tensor<128x256xi32> into tensor<4x32x16x16xi32>
+ %init_transpose = tensor.empty() : tensor<16x4x32x16xi32>
+ %transpose = linalg.transpose
+ ins(%reshape : tensor<4x32x16x16xi32>)
+ outs(%init_transpose : tensor<16x4x32x16xi32>)
+ permutation = [2, 0, 1, 3]
+
+ check.expect_eq(%pack, %transpose) : tensor<16x4x32x16xi32>
+ return
+}
+
+func.func @dynamic_pack_pad_large() {
+ %d0 = util.unfoldable_constant 100 : index
+ %d1 = util.unfoldable_constant 250 : index
+ %source = call @generate_2D_source(%d0, %d1) : (index, index) -> tensor<?x?xi32>
+ %padding_value = arith.constant 42 : i32
+
+ %c32 = arith.constant 32 : index
+ %c16 = arith.constant 16 : index
+ %tiled_d0 = arith.ceildivui %d0, %c32 : index
+ %tiled_d1 = arith.ceildivui %d1, %c16 : index
+ %dyn_init_pack = tensor.empty(%tiled_d0, %tiled_d1) : tensor<?x?x32x16xi32>
+ %pack = tensor.pack %source padding_value(%padding_value : i32)
+ inner_dims_pos = [0, 1] inner_tiles = [32, 16] into %dyn_init_pack
+ : tensor<?x?xi32> -> tensor<?x?x32x16xi32>
+ %cast_pack = tensor.cast %pack : tensor<?x?x32x16xi32> to tensor<4x16x32x16xi32>
+
+ %c100 = arith.constant 100 : index
+ %c250 = arith.constant 250 : index
+ %source2 = call @generate_2D_source(%c100, %c250) : (index, index) -> tensor<?x?xi32>
+ %static_source = tensor.cast %source2 : tensor<?x?xi32> to tensor<100x250xi32>
+ %pad = tensor.pad %static_source low[0, 0] high[28, 6] {
+ ^bb0(%b0 : index, %b1 : index):
+ tensor.yield %padding_value : i32
+ } : tensor<100x250xi32> to tensor<128x256xi32>
+ %reshape = tensor.expand_shape %pad [[0, 1], [2, 3]] : tensor<128x256xi32> into tensor<4x32x16x16xi32>
+ %init_transpose = tensor.empty() : tensor<4x16x32x16xi32>
+ %transpose = linalg.transpose
+ ins(%reshape : tensor<4x32x16x16xi32>)
+ outs(%init_transpose : tensor<4x16x32x16xi32>)
+ permutation = [0, 2, 1, 3]
+
+ check.expect_eq(%cast_pack, %transpose) : tensor<4x16x32x16xi32>
+ return
+}
+
+func.func @dynamic_pack_pad_transpose_outer_dims_large() {
+ %d0 = util.unfoldable_constant 100 : index
+ %d1 = util.unfoldable_constant 250 : index
+ %source = call @generate_2D_source(%d0, %d1) : (index, index) -> tensor<?x?xi32>
+ %padding_value = arith.constant 42 : i32
+
+ %c32 = arith.constant 32 : index
+ %c16 = arith.constant 16 : index
+ %tiled_d0 = arith.ceildivui %d0, %c32 : index
+ %tiled_d1 = arith.ceildivui %d1, %c16 : index
+ %dyn_init_pack = tensor.empty(%tiled_d1, %tiled_d0) : tensor<?x?x32x16xi32>
+ %pack = tensor.pack %source padding_value(%padding_value : i32)
+ outer_dims_perm = [1, 0] inner_dims_pos = [0, 1] inner_tiles = [32, 16] into %dyn_init_pack
+ : tensor<?x?xi32> -> tensor<?x?x32x16xi32>
+ %cast_pack = tensor.cast %pack : tensor<?x?x32x16xi32> to tensor<16x4x32x16xi32>
+
+ %c100 = arith.constant 100 : index
+ %c250 = arith.constant 250 : index
+ %source2 = call @generate_2D_source(%c100, %c250) : (index, index) -> tensor<?x?xi32>
+ %static_source = tensor.cast %source2 : tensor<?x?xi32> to tensor<100x250xi32>
+ %pad = tensor.pad %static_source low[0, 0] high[28, 6] {
+ ^bb0(%b0 : index, %b1 : index):
+ tensor.yield %padding_value : i32
+ } : tensor<100x250xi32> to tensor<128x256xi32>
+ %reshape = tensor.expand_shape %pad [[0, 1], [2, 3]] : tensor<128x256xi32> into tensor<4x32x16x16xi32>
+ %init_transpose = tensor.empty() : tensor<16x4x32x16xi32>
+ %transpose = linalg.transpose
+ ins(%reshape : tensor<4x32x16x16xi32>)
+ outs(%init_transpose : tensor<16x4x32x16xi32>)
+ permutation = [2, 0, 1, 3]
+
+ check.expect_eq(%cast_pack, %transpose) : tensor<16x4x32x16xi32>
+ return
+}
+
+func.func @static_pack_pad_transpose_inner_dims_large() {
+ %height = arith.constant 100 : index
+ %width = arith.constant 250 : index
+ %0 = call @generate_2D_source(%height, %width) : (index, index) -> tensor<?x?xi32>
+ %source = tensor.cast %0 : tensor<?x?xi32> to tensor<100x250xi32>
+ %padding_value = arith.constant 42 : i32
+
+ %init_pack = tensor.empty() : tensor<4x16x16x32xi32>
+ %pack = tensor.pack %source padding_value(%padding_value : i32)
+ inner_dims_pos = [1, 0] inner_tiles = [16, 32] into %init_pack
+ : tensor<100x250xi32> -> tensor<4x16x16x32xi32>
+
+ %pad = tensor.pad %source low[0, 0] high[28, 6] {
+ ^bb0(%b0 : index, %b1 : index):
+ tensor.yield %padding_value : i32
+ } : tensor<100x250xi32> to tensor<128x256xi32>
+ %reshape = tensor.expand_shape %pad [[0, 1], [2, 3]] : tensor<128x256xi32> into tensor<4x32x16x16xi32>
+ %init_transpose = tensor.empty() : tensor<4x16x16x32xi32>
+ %transpose = linalg.transpose
+ ins(%reshape : tensor<4x32x16x16xi32>)
+ outs(%init_transpose : tensor<4x16x16x32xi32>)
+ permutation = [0, 2, 3, 1]
+
+ check.expect_eq(%pack, %transpose) : tensor<4x16x16x32xi32>
+ return
+}
+
+func.func @dynamic_pack_pad_transpose_inner_dims_large() {
+ %d0 = util.unfoldable_constant 100 : index
+ %d1 = util.unfoldable_constant 250 : index
+ %source = call @generate_2D_source(%d0, %d1) : (index, index) -> tensor<?x?xi32>
+ %padding_value = arith.constant 42 : i32
+
+ %c16 = arith.constant 16 : index
+ %c32 = arith.constant 32 : index
+ %tiled_d0 = arith.ceildivui %d0, %c32 : index
+ %tiled_d1 = arith.ceildivui %d1, %c16 : index
+ %init_pack = tensor.empty(%tiled_d0, %tiled_d1) : tensor<?x?x16x32xi32>
+ %pack = tensor.pack %source padding_value(%padding_value : i32)
+ inner_dims_pos = [1, 0] inner_tiles = [16, 32] into %init_pack
+ : tensor<?x?xi32> -> tensor<?x?x16x32xi32>
+ %cast_pack = tensor.cast %pack : tensor<?x?x16x32xi32> to tensor<4x16x16x32xi32>
+
+ %c100 = arith.constant 100 : index
+ %c250 = arith.constant 250 : index
+ %source2 = call @generate_2D_source(%c100, %c250) : (index, index) -> tensor<?x?xi32>
+ %static_source = tensor.cast %source2 : tensor<?x?xi32> to tensor<100x250xi32>
+
+ %pad = tensor.pad %static_source low[0, 0] high[28, 6] {
+ ^bb0(%b0 : index, %b1 : index):
+ tensor.yield %padding_value : i32
+ } : tensor<100x250xi32> to tensor<128x256xi32>
+ %reshape = tensor.expand_shape %pad [[0, 1], [2, 3]] : tensor<128x256xi32> into tensor<4x32x16x16xi32>
+ %init_transpose = tensor.empty() : tensor<4x16x16x32xi32>
+ %transpose = linalg.transpose
+ ins(%reshape : tensor<4x32x16x16xi32>)
+ outs(%init_transpose : tensor<4x16x16x32xi32>)
+ permutation = [0, 2, 3, 1]
+
+ check.expect_eq(%cast_pack, %transpose) : tensor<4x16x16x32xi32>
+ return
+}
+
+func.func @static_pack_pad_transpose_inner_and_outer_dims_large() {
+ %height = arith.constant 100 : index
+ %width = arith.constant 250 : index
+ %0 = call @generate_2D_source(%height, %width) : (index, index) -> tensor<?x?xi32>
+ %source = tensor.cast %0 : tensor<?x?xi32> to tensor<100x250xi32>
+ %padding_value = arith.constant 42 : i32
+
+ %init_pack = tensor.empty() : tensor<16x4x16x32xi32>
+ %pack = tensor.pack %source padding_value(%padding_value : i32)
+ outer_dims_perm = [1, 0] inner_dims_pos = [1, 0] inner_tiles = [16, 32] into %init_pack
+ : tensor<100x250xi32> -> tensor<16x4x16x32xi32>
+
+ %pad = tensor.pad %source low[0, 0] high[28, 6] {
+ ^bb0(%b0 : index, %b1 : index):
+ tensor.yield %padding_value : i32
+ } : tensor<100x250xi32> to tensor<128x256xi32>
+ %reshape = tensor.expand_shape %pad [[0, 1], [2, 3]] : tensor<128x256xi32> into tensor<4x32x16x16xi32>
+ %init_transpose = tensor.empty() : tensor<16x4x16x32xi32>
+ %transpose = linalg.transpose
+ ins(%reshape : tensor<4x32x16x16xi32>)
+ outs(%init_transpose : tensor<16x4x16x32xi32>)
+ permutation = [2, 0, 3, 1]
+
+ check.expect_eq(%pack, %transpose) : tensor<16x4x16x32xi32>
+ return
+}
+
+func.func @dynamic_pack_pad_transpose_inner_and_outer_dims_large() {
+ %d0 = util.unfoldable_constant 100 : index
+ %d1 = util.unfoldable_constant 250 : index
+ %source = call @generate_2D_source(%d0, %d1) : (index, index) -> tensor<?x?xi32>
+ %padding_value = arith.constant 42 : i32
+
+ %c16 = arith.constant 16 : index
+ %c32 = arith.constant 32 : index
+ %tiled_d0 = arith.ceildivui %d0, %c32 : index
+ %tiled_d1 = arith.ceildivui %d1, %c16 : index
+ %init_pack = tensor.empty(%tiled_d1, %tiled_d0) : tensor<?x?x16x32xi32>
+ %pack = tensor.pack %source padding_value(%padding_value : i32)
+ outer_dims_perm = [1, 0] inner_dims_pos = [1, 0] inner_tiles = [16, 32] into %init_pack
+ : tensor<?x?xi32> -> tensor<?x?x16x32xi32>
+ %cast_pack = tensor.cast %pack : tensor<?x?x16x32xi32> to tensor<16x4x16x32xi32>
+
+ %c100 = arith.constant 100 : index
+ %c250 = arith.constant 250 : index
+ %source2 = call @generate_2D_source(%c100, %c250) : (index, index) -> tensor<?x?xi32>
+ %static_source = tensor.cast %source2 : tensor<?x?xi32> to tensor<100x250xi32>
+
+ %pad = tensor.pad %static_source low[0, 0] high[28, 6] {
+ ^bb0(%b0 : index, %b1 : index):
+ tensor.yield %padding_value : i32
+ } : tensor<100x250xi32> to tensor<128x256xi32>
+ %reshape = tensor.expand_shape %pad [[0, 1], [2, 3]] : tensor<128x256xi32> into tensor<4x32x16x16xi32>
+ %init_transpose = tensor.empty() : tensor<16x4x16x32xi32>
+ %transpose = linalg.transpose
+ ins(%reshape : tensor<4x32x16x16xi32>)
+ outs(%init_transpose : tensor<16x4x16x32xi32>)
+ permutation = [2, 0, 3, 1]
+
+ check.expect_eq(%cast_pack, %transpose) : tensor<16x4x16x32xi32>
+ return
+}
+
+// TODO(hanchung): Enable the test once we have better ideas about supporting
+// dynamic inner tiling sizes. We are not able to vectorized this case today.
+// func.func @fully_dynamic_pack_simple() {
+// %iree_input = flow.tensor.constant dense<[
+// [0, 1, 2, 3],
+// [4, 5, 6, 7],
+// [8, 9, 10, 11],
+// [12, 13, 14, 15]]> : tensor<4x4xi32> -> tensor<?x?xi32>
+// %c0 = arith.constant 0 : index
+// %c1 = arith.constant 1 : index
+// %c2 = util.unfoldable_constant 2 : index
+// %in_d0 = tensor.dim %iree_input, %c0 : tensor<?x?xi32>
+// %in_d1 = tensor.dim %iree_input, %c1 : tensor<?x?xi32>
+// %out_d0 = arith.ceildivui %in_d0, %c2 : index
+// %out_d1 = arith.ceildivui %in_d1, %c2 : index
+// %init = tensor.empty(%out_d0, %out_d1, %c2, %c2) : tensor<?x?x?x?xi32>
+// %pack = tensor.pack %iree_input inner_dims_pos = [0, 1] inner_tiles = [%c2, %c2] into %init
+// : tensor<?x?xi32> -> tensor<?x?x?x?xi32>
+// %cast = tensor.cast %pack : tensor<?x?x?x?xi32> to tensor<2x2x2x2xi32>
+// check.expect_eq_const(%cast, dense<[[[[0, 1], [4, 5]], [[2, 3], [6, 7]]], [[[8, 9], [12, 13]], [[10 ,11], [14, 15]]]]> : tensor<2x2x2x2xi32>) : tensor<2x2x2x2xi32>
+// return
+// }
+//
+// func.func @fully_dynamic_pack_pad_transpose_inner_and_outer_dims_large() {
+// %d0 = util.unfoldable_constant 100 : index
+// %d1 = util.unfoldable_constant 250 : index
+// %source = call @generate_2D_source(%d0, %d1) : (index, index) -> tensor<?x?xi32>
+// %padding_value = arith.constant 42 : i32
+//
+// %c16 = util.unfoldable_constant 16 : index
+// %c32 = util.unfoldable_constant 32 : index
+// %tiled_d0 = arith.ceildivui %d0, %c32 : index
+// %tiled_d1 = arith.ceildivui %d1, %c16 : index
+// %init_pack = tensor.empty(%tiled_d1, %tiled_d0, %c16, %c32) : tensor<?x?x?x?xi32>
+// %pack = tensor.pack %source padding_value(%padding_value : i32)
+// outer_dims_perm = [1, 0] inner_dims_pos = [1, 0] inner_tiles = [%c16, %c32] into %init_pack
+// : tensor<?x?xi32> -> tensor<?x?x?x?xi32>
+// %cast_pack = tensor.cast %pack : tensor<?x?x?x?xi32> to tensor<16x4x16x32xi32>
+//
+// %c100 = arith.constant 100 : index
+// %c250 = arith.constant 250 : index
+// %source2 = call @generate_2D_source(%c100, %c250) : (index, index) -> tensor<?x?xi32>
+// %static_source = tensor.cast %source2 : tensor<?x?xi32> to tensor<100x250xi32>
+//
+// %pad = tensor.pad %static_source low[0, 0] high[28, 6] {
+// ^bb0(%b0 : index, %b1 : index):
+// tensor.yield %padding_value : i32
+// } : tensor<100x250xi32> to tensor<128x256xi32>
+// %reshape = tensor.expand_shape %pad [[0, 1], [2, 3]] : tensor<128x256xi32> into tensor<4x32x16x16xi32>
+// %init_transpose = tensor.empty() : tensor<16x4x16x32xi32>
+// %transpose = linalg.transpose
+// ins(%reshape : tensor<4x32x16x16xi32>)
+// outs(%init_transpose : tensor<16x4x16x32xi32>)
+// permutation = [2, 0, 3, 1]
+//
+// check.expect_eq(%cast_pack, %transpose) : tensor<16x4x16x32xi32>
+// return
+// }
