Skip to content

CANN: Add support for CONV_TRANSPOSE_1D when kernel size > 255#17934

Merged
hipudding merged 5 commits intoggml-org:masterfrom
Intellouis:feature/CONV_TRANSPOSE_1D
Dec 25, 2025
Merged

CANN: Add support for CONV_TRANSPOSE_1D when kernel size > 255#17934
hipudding merged 5 commits intoggml-org:masterfrom
Intellouis:feature/CONV_TRANSPOSE_1D

Conversation

@Intellouis
Copy link
Contributor

@Intellouis Intellouis commented Dec 11, 2025 •
edited
Loading

Description

This PR implements support for kernel_size > 255 in the CANN backend for the ggml's CONV_TRANSPOSE_1D operator. The existing CONV_TRANSPOSE_1D operator, implemented via a direct call to Convolution, throws a shape error for inputs whose kernel_size is greater than 255.

Summary of changes:

  1. Added the header file aclnnop/aclnn_slice.h to ggml/src/ggml-cann/aclnn_ops.h to support slice operations.

  2. Added function implementation in ggml/src/ggml-cann/aclnn_ops.cpp. The general approach is as follows:

  • Slice the large kernel into multiple smaller kernels each ≤ 255.
  • Perform CONV_TRANSPOSE_1D on each small kernel with the input to obtain partial results.
  • Calculate the start and end positions of each partial result within the final output and pad them to the same length as the output.
  • Sum all partial results to get the final output.
  1. In ggml/src/ggml-cann/ggml-cann.cpp, remove the condition check:
case GGML_OP_CONV_TRANSPOSE_1D:
    // TODO: ((weightL - 1) * dilationW - padLeft)=1336 should not be larger than 255.
    return (op->src[0]->ne[0] - 1) <= 255;

Testing

Test steps:

  • Building：
cmake -B build -DGGML_CANN=on -DCMAKE_BUILD_TYPE=release
cmake --build build --config release -j
  • Run test:
./bin/test-backend-ops test -b CANN0 -o CONV_TRANSPOSE_1D

All test cases should be passed.

@github-actions github-actions bot added ggml changes relating to the ggml tensor library for machine learning Ascend NPU issues specific to Ascend NPUs labels Dec 11, 2025
@ixgbe
Copy link
Contributor

ixgbe commented Dec 12, 2025

Could you please update the PR description in English?
This helps the maintainers and reviewers understand the changes more easily and keeps the project documentation consistent.

@Intellouis Intellouis changed the title CANN: CONV_TRANSPOSE_1D算子：支持 (op->src[0]->ne[0] - 1) > 255 场景 CANN: CONV_TRANSPOSE_1D operator: supporting the cases where (op->src[0]->ne[0] - 1) > 255 Dec 12, 2025
@Intellouis
Copy link
Contributor Author

Intellouis commented Dec 12, 2025

Could you please update the PR description in English? This helps the maintainers and reviewers understand the changes more easily and keeps the project documentation consistent.

Sure! I have edited the PR in English.

@hipudding
Copy link
Contributor

hipudding commented Dec 17, 2025 •
edited
Loading

Please paste the test result of this operator.

@Intellouis
Copy link
Contributor Author

Intellouis commented Dec 17, 2025

Please paste the test result of this operator.

(base) [user@cntrain02 llama.cpp]$ ./build/bin/test-backend-ops test -b CANN0 -o CONV_TRANSPOSE_1D
Testing 2 devices

ggml_backend_cann_context: device 0 async operator submission is OFF
ggml_backend_cann_context: LLAMA_SET_ROWS is OFF
ggml_backend_cann_context: CANN Graph currently only supports execution when LLAMA_SET_ROWS is ON. Falling back to eager mode.
Backend 1/2: CANN0
  Device description: Ascend910B3
  Device memory: 62420 MB (62091 MB free)

new_pool_for_device: device 0 use vmm pool
  CONV_TRANSPOSE_1D(ne_input=[1,1,1,1],ne_kernel=[1,1,1,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,1,1,1],ne_kernel=[1,1,1,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,1,1,1],ne_kernel=[1,1,1,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,1,1,1],ne_kernel=[1,1,1,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,1,1,1],ne_kernel=[1,1,1,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,1,1,1],ne_kernel=[1,1,1,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,1,1,1],ne_kernel=[1,1,1,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,1,1,1],ne_kernel=[1,1,1,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,1,1,1],ne_kernel=[1,1,1,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,1,1,1],ne_kernel=[3,1,1,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,1,1,1],ne_kernel=[3,1,1,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,1,1,1],ne_kernel=[3,1,1,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,1,1,1],ne_kernel=[3,1,1,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,1,1,1],ne_kernel=[3,1,1,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,1,1,1],ne_kernel=[3,1,1,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,1,1,1],ne_kernel=[3,1,1,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,1,1,1],ne_kernel=[3,1,1,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,1,1,1],ne_kernel=[3,1,1,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,1,1,1],ne_kernel=[1337,1,1,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,1,1,1],ne_kernel=[1337,1,1,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,1,1,1],ne_kernel=[1337,1,1,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,1,1,1],ne_kernel=[1337,1,1,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,1,1,1],ne_kernel=[1337,1,1,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,1,1,1],ne_kernel=[1337,1,1,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,1,1,1],ne_kernel=[1337,1,1,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,1,1,1],ne_kernel=[1337,1,1,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,1,1,1],ne_kernel=[1337,1,1,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,7,1,1],ne_kernel=[1,1,7,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,7,1,1],ne_kernel=[1,1,7,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,7,1,1],ne_kernel=[1,1,7,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,7,1,1],ne_kernel=[1,1,7,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,7,1,1],ne_kernel=[1,1,7,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,7,1,1],ne_kernel=[1,1,7,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,7,1,1],ne_kernel=[1,1,7,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,7,1,1],ne_kernel=[1,1,7,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,7,1,1],ne_kernel=[1,1,7,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,7,1,1],ne_kernel=[3,1,7,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,7,1,1],ne_kernel=[3,1,7,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,7,1,1],ne_kernel=[3,1,7,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,7,1,1],ne_kernel=[3,1,7,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,7,1,1],ne_kernel=[3,1,7,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,7,1,1],ne_kernel=[3,1,7,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,7,1,1],ne_kernel=[3,1,7,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,7,1,1],ne_kernel=[3,1,7,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,7,1,1],ne_kernel=[3,1,7,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,7,1,1],ne_kernel=[1337,1,7,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,7,1,1],ne_kernel=[1337,1,7,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,7,1,1],ne_kernel=[1337,1,7,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,7,1,1],ne_kernel=[1337,1,7,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,7,1,1],ne_kernel=[1337,1,7,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,7,1,1],ne_kernel=[1337,1,7,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,7,1,1],ne_kernel=[1337,1,7,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,7,1,1],ne_kernel=[1337,1,7,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,7,1,1],ne_kernel=[1337,1,7,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,1,1,1],ne_kernel=[1,9,1,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,1,1,1],ne_kernel=[1,9,1,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,1,1,1],ne_kernel=[1,9,1,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,1,1,1],ne_kernel=[1,9,1,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,1,1,1],ne_kernel=[1,9,1,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,1,1,1],ne_kernel=[1,9,1,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,1,1,1],ne_kernel=[1,9,1,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,1,1,1],ne_kernel=[1,9,1,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,1,1,1],ne_kernel=[1,9,1,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,1,1,1],ne_kernel=[3,9,1,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,1,1,1],ne_kernel=[3,9,1,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,1,1,1],ne_kernel=[3,9,1,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,1,1,1],ne_kernel=[3,9,1,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,1,1,1],ne_kernel=[3,9,1,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,1,1,1],ne_kernel=[3,9,1,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,1,1,1],ne_kernel=[3,9,1,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,1,1,1],ne_kernel=[3,9,1,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,1,1,1],ne_kernel=[3,9,1,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,1,1,1],ne_kernel=[1337,9,1,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,1,1,1],ne_kernel=[1337,9,1,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,1,1,1],ne_kernel=[1337,9,1,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,1,1,1],ne_kernel=[1337,9,1,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,1,1,1],ne_kernel=[1337,9,1,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,1,1,1],ne_kernel=[1337,9,1,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,1,1,1],ne_kernel=[1337,9,1,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,1,1,1],ne_kernel=[1337,9,1,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,1,1,1],ne_kernel=[1337,9,1,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,7,1,1],ne_kernel=[1,9,7,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,7,1,1],ne_kernel=[1,9,7,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,7,1,1],ne_kernel=[1,9,7,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,7,1,1],ne_kernel=[1,9,7,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,7,1,1],ne_kernel=[1,9,7,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,7,1,1],ne_kernel=[1,9,7,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,7,1,1],ne_kernel=[1,9,7,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,7,1,1],ne_kernel=[1,9,7,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,7,1,1],ne_kernel=[1,9,7,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,7,1,1],ne_kernel=[3,9,7,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,7,1,1],ne_kernel=[3,9,7,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,7,1,1],ne_kernel=[3,9,7,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,7,1,1],ne_kernel=[3,9,7,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,7,1,1],ne_kernel=[3,9,7,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,7,1,1],ne_kernel=[3,9,7,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,7,1,1],ne_kernel=[3,9,7,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,7,1,1],ne_kernel=[3,9,7,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,7,1,1],ne_kernel=[3,9,7,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,7,1,1],ne_kernel=[1337,9,7,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,7,1,1],ne_kernel=[1337,9,7,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[1,7,1,1],ne_kernel=[1337,9,7,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,7,1,1],ne_kernel=[1337,9,7,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,7,1,1],ne_kernel=[1337,9,7,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,7,1,1],ne_kernel=[1337,9,7,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,7,1,1],ne_kernel=[1337,9,7,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,7,1,1],ne_kernel=[1337,9,7,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[13,7,1,1],ne_kernel=[1337,9,7,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[197,32,1,1],ne_kernel=[16,32,32,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[3,2,1,1],ne_kernel=[2,3,2,1],s0=3,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[3,2,1,1],ne_kernel=[2,3,2,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[3,2,1,1],ne_kernel=[2,3,2,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[3,2,1,1],ne_kernel=[3,2,2,1],s0=2,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[3,2,1,1],ne_kernel=[3,2,2,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[3,2,1,1],ne_kernel=[3,1,2,1],s0=1,p0=0,d0=1): OK
  CONV_TRANSPOSE_1D(ne_input=[2,1,1,1],ne_kernel=[3,1,1,1],s0=1,p0=0,d0=1): OK
  11821/11821 tests passed
  Backend CANN0: OK
Backend 2/2: CPU
  Skipping
2/2 backends passed
OK

@hipudding hipudding changed the title CANN: CONV_TRANSPOSE_1D operator: supporting the cases where (op->src[0]->ne[0] - 1) > 255 CANN: Add support for CONV_TRANSPOSE_1D when kernel size > 255 Dec 24, 2025
hipudding
hipudding approved these changes Dec 25, 2025
View reviewed changes
Copy link
Contributor

@hipudding hipudding left a comment

Choose a reason for hiding this comment

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

Awesome! Thanks for your contribution.

@hipudding hipudding merged commit e68c19b into ggml-org:master Dec 25, 2025
136 of 147 checks passed
@noemotiovon noemotiovon mentioned this pull request Jan 13, 2026
Closed
blime4 referenced this pull request in blime4/llama.cpp Feb 5, 2026
@Intellouis
CANN: Add support for CONV_TRANSPOSE_1D when kernel size > 255 (#17934)
801eb05 
* CONV_TRANSPOSE_1D kernel_size>255

* remove condition check

* fix the bug of type conversion

* removing trailing whitespaces

* fix: return true in the switch case
@wallentri88 wallentri88 mentioned this pull request Feb 24, 2026
Closed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@hipudding hipudding hipudding approved these changes

+1 more reviewer

@RatStar811 RatStar811 RatStar811 approved these changes

Reviewers whose approvals may not affect merge requirements

Assignees

No one assigned

Labels

Ascend NPU issues specific to Ascend NPUs ggml changes relating to the ggml tensor library for machine learning

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

4 participants

@Intellouis @ixgbe @hipudding @RatStar811