[MLIR][Linalg] Scalable Vectorization of Reduction on the Trailing Dimension #97788
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,134 @@ | ||
| // DEFINE: %{compile} = mlir-opt %s \ | ||
| // DEFINE: -transform-interpreter -test-transform-dialect-erase-schedule \ | ||
| // DEFINE: -one-shot-bufferize="bufferize-function-boundaries" -buffer-deallocation-pipeline -cse -canonicalize -convert-vector-to-scf -arm-sve-legalize-vector-storage \ | ||
| // DEFINE: -convert-vector-to-llvm="enable-arm-sve" -test-lower-to-llvm -o %t | ||
| // DEFINE: %{entry_point} = reduce_1d_f32 | ||
| // DEFINE: %{run} = %mcr_aarch64_cmd %t -e %{entry_point} -entry-point-result=void --march=aarch64 --mattr="+sve"\ | ||
| // DEFINE: -shared-libs=%mlir_native_utils_lib_dir/libmlir_runner_utils%shlibext,%mlir_native_utils_lib_dir/libmlir_c_runner_utils%shlibext | ||
|
|
||
| // RUN: %{compile} | ||
|
|
||
| // RUN: %{run} | FileCheck %s --check-prefix=REDUCE | ||
|
|
||
| // REDEFINE: %{entry_point} = generic_reduce_1d_f32 | ||
| // RUN: %{run} | FileCheck %s --check-prefix=GENERIC | ||
|
|
||
| func.func @reduce_1d_f32() { | ||
|
There was a problem hiding this comment. Would you mind adding a test for |
||
| // 1-D Tensor | ||
| %N = arith.constant 1000 : index | ||
| %c0_f32 = arith.constant 0.0 : f32 | ||
|
|
||
| // Allocate the input and output tensors | ||
| %A_alloc = bufferization.alloc_tensor(%N) : tensor<?xf32> | ||
| %C_alloc = bufferization.alloc_tensor() : tensor<f32> | ||
|
|
||
| // Initialise the tensors | ||
| %pi = arith.constant 3.1416 : f32 | ||
| %A_in = linalg.fill ins(%pi : f32) outs(%A_alloc : tensor<?xf32>) -> tensor<?xf32> | ||
| %C_in = tensor.insert %c0_f32 into %C_alloc[] : tensor<f32> | ||
|
|
||
| // Reduce | ||
| %C_out = linalg.reduce ins(%A_in : tensor<?xf32>) outs(%C_in: tensor<f32>) dimensions = [0] | ||
| (%in: f32, %init: f32) { | ||
| %0 = arith.addf %in, %init : f32 | ||
| linalg.yield %0 : f32 | ||
| } | ||
|
|
||
| // Print and verify the output | ||
| // REDUCE-LABEL: SVE: START OF TEST OUTPUT | ||
| vector.print str "SVE: START OF TEST OUTPUT\n" | ||
|
|
||
| // REDUCE-NEXT: Unranked Memref {{.*}} rank = 0 offset = 0 sizes = [] strides = [] data = | ||
| // REDUCE-NEXT: [3141.6] | ||
|
|
||
| %xf = tensor.cast %C_out : tensor<f32> to tensor<*xf32> | ||
| call @printMemrefF32(%xf) : (tensor<*xf32>) -> () | ||
|
|
||
| // REDUCE-NEXT: SVE: END OF TEST OUTPUT | ||
| vector.print str "SVE: END OF TEST OUTPUT\n" | ||
|
|
||
| return | ||
| } | ||
|
|
||
| func.func @generic_reduce_1d_f32() { | ||
| // 1-D Tensor | ||
| %N = arith.constant 1000 : index | ||
| %c0_f32 = arith.constant 0.0 : f32 | ||
|
|
||
| // Allocate the input and output tensors | ||
| %A_alloc = bufferization.alloc_tensor(%N) : tensor<?xf32> | ||
| %C_alloc = bufferization.alloc_tensor() : tensor<f32> | ||
|
|
||
| // Initialise the tensors | ||
| %pi = arith.constant 3.1416 : f32 | ||
| %A_in = linalg.fill ins(%pi : f32) outs(%A_alloc : tensor<?xf32>) -> tensor<?xf32> | ||
| %C_in = tensor.insert %c0_f32 into %C_alloc[] : tensor<f32> | ||
|
|
||
| // Reduce | ||
| %C_out = linalg.generic { indexing_maps = [affine_map<(d0) -> (d0)>, | ||
| affine_map<(d0) -> ()>], | ||
| iterator_types = ["reduction"] } | ||
| ins(%A_in : tensor<?xf32>) | ||
| outs(%C_in : tensor<f32>) { | ||
| ^bb(%in: f32, %out: f32) : | ||
| %0 = arith.addf %in, %out : f32 | ||
| linalg.yield %0 : f32 | ||
| } -> tensor<f32> | ||
|
|
||
| // Print and verify the output | ||
| // GENERIC-LABEL: SVE: START OF TEST OUTPUT | ||
| vector.print str "SVE: START OF TEST OUTPUT\n" | ||
|
|
||
| // GENERIC-NEXT: Unranked Memref {{.*}} rank = 0 offset = 0 sizes = [] strides = [] data = | ||
| // GENERIC-NEXT: [3141.6] | ||
|
|
||
| %xf = tensor.cast %C_out : tensor<f32> to tensor<*xf32> | ||
| call @printMemrefF32(%xf) : (tensor<*xf32>) -> () | ||
|
|
||
| // GENERIC-NEXT: SVE: END OF TEST OUTPUT | ||
| vector.print str "SVE: END OF TEST OUTPUT\n" | ||
|
|
||
| return | ||
| } | ||
|
|
||
| module attributes {transform.with_named_sequence} { | ||
| // A sequence that will tile and vectorise a Reduce Op | ||
| transform.named_sequence @tile_and_vectorize_reduce(%func | ||
| : !transform.op<"func.func"> {transform.readonly}) { | ||
|
|
||
| // Step 0: Get a handle to the reduce Op | ||
| %reduce = transform.structured.match ops{["linalg.reduce", "linalg.generic"]} in %func | ||
| : (!transform.op<"func.func">) -> !transform.any_op | ||
|
|
||
| // Step 1: Tile | ||
| %tiled_reduce, %loops:1 = transform.structured.tile_using_for %reduce tile_sizes [[4]] | ||
| : (!transform.any_op) -> (!transform.any_op, !transform.any_op) | ||
|
|
||
| // Step 2: Vectorize | ||
| transform.structured.vectorize %tiled_reduce vector_sizes [[4]] : !transform.any_op | ||
|
|
||
| // Step 3: Lower vector.multi_reduction | ||
| transform.apply_patterns to %func { | ||
| transform.apply_patterns.vector.lower_masked_transfers | ||
| transform.apply_patterns.vector.lower_multi_reduction lowering_strategy = "innerreduction" | ||
| } : !transform.op<"func.func"> | ||
|
|
||
| transform.yield | ||
| } | ||
|
|
||
| // A sequence that goes over all functions in tis module and applies | ||
| // "tile_and_vectorize_reduce" | ||
| transform.named_sequence @__transform_main(%module: !transform.any_op {transform.readonly}) { | ||
| %funcs = transform.structured.match ops{["func.func"]} in %module | ||
| : (!transform.any_op) -> !transform.op<"func.func"> | ||
|
|
||
| transform.foreach %funcs : !transform.op<"func.func"> { | ||
| ^bb2(%func : !transform.op<"func.func">): | ||
| transform.include @tile_and_vectorize_reduce failures(propagate) | ||
| (%func) : (!transform.op<"func.func">) -> () | ||
| } | ||
| transform.yield | ||
| } | ||
| } | ||
|
|
||
| func.func private @printMemrefF32(%ptr : tensor<*xf32>) | ||
Oops, something went wrong.
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Please document. Also, why
hasLinalgReductionraher thanisaLinalgReduction?