]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
ggml : fix conv2d_dw SVE path (ggml/1380)
authorGeorgi Gerganov <redacted>
Tue, 4 Nov 2025 18:40:52 +0000 (20:40 +0200)
committerGeorgi Gerganov <redacted>
Wed, 5 Nov 2025 08:41:51 +0000 (10:41 +0200)
commit852ce5180acd4e3641f6b71783a343647eee058d
tree7069cdb97601e4804c23b9c2f0d32e33fe470d9a
parent9aa63374f2101f4eaef425888699907b238ed2c0
ggml : fix conv2d_dw SVE path (ggml/1380)

* Fix test-conv2d-dw failure on ARM SVE by using runtime vector length

The ggml_compute_forward_conv_2d_dw_cwhn function was using a hardcoded GGML_F32_EPR (8) for SIMD vectorization, but on ARM SVE the actual vector length varies by hardware. This caused incorrect computation when processing CWHN layout tensors on ARM machines.

Fix by using svcntw() to get the runtime SVE vector length instead of the compile-time constant.

Co-authored-by: ggerganov <redacted>
* ci : reduce sam score threshold

* ci : update bbox checks for sam test

---------

Co-authored-by: copilot-swe-agent[bot] <redacted>
Co-authored-by: ggerganov <redacted>
ggml/src/ggml-cpu/ops.cpp