Skip to content

[mlir][SVE] Add e2e for 1D depthwise WC convolution #85225

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
Mar 22, 2024
Merged

[mlir][SVE] Add e2e for 1D depthwise WC convolution #85225

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
a5f7480
[mlir][SVE] Add e2e for 1D depthwise WC convolution
banach-space Mar 14, 2024
41de1a7
fixup! [BOLT] Add BB index to BAT (#86044)
banach-space Mar 22, 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
60 changes: 60 additions & 0 deletions mlir/test/Integration/Dialect/Linalg/CPU/ArmSVE/1d-depthwise-conv.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,60 @@
// DEFINE: %{compile} = mlir-opt %s \
// DEFINE: -transform-interpreter -test-transform-dialect-erase-schedule \
// DEFINE: -one-shot-bufferize="bufferize-function-boundaries" -lower-vector-mask -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} = conv
// DEFINE: %{run} = %mcr_aarch64_cmd %t -e %{entry_point} -entry-point-result=void --march=aarch64 --mattr="+sve"\
// DEFINE: -shared-libs=%mlir_runner_utils,%mlir_c_runner_utils

// RUN: %{compile} | %{run} | FileCheck %s
Comment on lines +5 to +9
Copy link
Member

@MacDue MacDue Mar 23, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Here it's both using input + output files (via -o), but also piping things. This seems to work when %mcr_aarch64_cmd is an emulator, but if you replace it with mlir-cpu-runner (on SVE hardware) the test fails on the first run (but works on the second run).

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P.s. -convert-vector-to-scf -arm-sve-legalize-vector-storage -convert-vector-to-llvm="enable-arm-sve" can be replaced with -convert-vector-to-scf=full-unroll.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks, good catch!

Yeah, the RUN should be something like this: // RUN: rm -f %t && %{compile} && %{run} | FileCheck %s instead. However, the current RUN line, when fails, produces:

# .---command stderr------------
# | cannot open input file '/llvm-project/build/release_vls/tools/mlir/test/Integration/Dialect/Linalg/CPU/ArmSVE/Output/1d-depthwise-conv.mlir.tmp': No such file or directory
# | could not parse the input IR
# `-----------------------------

rather than (from https://lab.llvm.org/buildbot/#/builders/176/builds/9331):

# .---command stderr------------
# | Error: entry point not found
# `-----------------------------

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

A guess: It may depend on the shell/system if the result is an empty file or no file (Error: entry point not found seems to indicate the input was empty).

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

That could be the case 🤔

There's https://github.com/llvm/llvm-project/blob/main/mlir/test/Integration/Dialect/Linalg/CPU/ArmSVE/matmul.mlir that works fine with this configuration - we probably just missed that it failed first time it was run. Also a guess :)

If I don't hear back from Omair with other suggestion, I'll merge the updated version around Monday and just monitor the bots to see what happens. Thanks Ben!

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

No worries, sounds good to me :)

In the matmul test there's a lone // RUN: %{compile}, with the run steps separate, which would not have this issue.


func.func @conv() {
// Define input/output tensors
%input_init = tensor.empty() : tensor<1x8x6xi32>
%output_init = tensor.empty() : tensor<1x7x6xi32>

%five = arith.constant 5 : i32
%zero = arith.constant 0 : i32
%input = linalg.fill ins(%five : i32) outs(%input_init : tensor<1x8x6xi32>) -> tensor<1x8x6xi32>
%output = linalg.fill ins(%zero : i32) outs(%output_init : tensor<1x7x6xi32>) -> tensor<1x7x6xi32>

// Define the filter tensor
%filter = arith.constant dense<[
[ 1, 2, 3, 4, 5, 6],
[ 11, 12, 13, 14, 15, 16]
]> : tensor<2x6xi32>

// static sizes -> dynamic sizes
%input_dyn = tensor.cast %input_init : tensor<1x8x6xi32> to tensor<1x8x?xi32>
%output_dyn = tensor.cast %output : tensor<1x7x6xi32> to tensor<1x7x?xi32>
%filter_dyn = tensor.cast %filter : tensor<2x6xi32> to tensor<2x?xi32>

// Run the convolution
%res = linalg.depthwise_conv_1d_nwc_wc
ins(%input_dyn, %filter_dyn : tensor<1x8x?xi32>, tensor<2x?xi32>)
outs(%output_dyn : tensor<1x7x?xi32>) -> tensor<1x7x?xi32>

// Print the results
// CHECK: SVE: START OF TEST OUTPUT
vector.print str "SVE: START OF TEST OUTPUT\n"

// CHECK-NEXT: Unranked Memref base@ = {{.*}} rank = 3 offset = 0 sizes = [1, 7, 6] strides = [42, 6, 1] data =
// CHECK-COUNT-7: [60, 70, 80, 90, 100, 110]
%xf = tensor.cast %res : tensor<1x7x?xi32> to tensor<*xi32>
call @printMemrefI32(%xf) : (tensor<*xi32>) -> ()

// CHECK-NEXT: SVE: END OF TEST OUTPUT
vector.print str "SVE: END OF TEST OUTPUT\n"

return
}

module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
%0 = transform.structured.match ops{["linalg.depthwise_conv_1d_nwc_wc"]} in %arg0 : (!transform.any_op) -> !transform.any_op
transform.structured.vectorize %0 vector_sizes [1, 7, [8], 2] : !transform.any_op
transform.yield
}
}

func.func private @printMemrefI32(%ptr : tensor<*xi32>) attributes { llvm.emit_c_interface }