cfRod edited a comment on issue #20766: URL: https://github.com/apache/incubator-mxnet/issues/20766#issuecomment-988759914
>ACL not supporting training specific operators makes sense, but it should ideally default to OneDNN's reference kernels It should fallback to reference as per this implementation list https://github.com/oneapi-src/oneDNN/blob/b2cd3a8e50a715f9326a35f4c503bd11e60235a5/src/cpu/cpu_deconvolution_list.cpp#L48. Since there is no deconvolution operator implemented in ACL here. For convolution, you can see the implementations here https://github.com/oneapi-src/oneDNN/blob/b2cd3a8e50a715f9326a35f4c503bd11e60235a5/src/cpu/cpu_convolution_list.cpp#L105 guarded by CPU_INSTANCE_AARCH64_ACL. and within the convolution we have checks to make sure FWD phases are enabled https://github.com/oneapi-src/oneDNN/blob/b2cd3a8e50a715f9326a35f4c503bd11e60235a5/src/cpu/aarch64/acl_convolution_utils.cpp#L47 Could you please check if MxNET implements deconvolution via oneDNN's deconvolution operator? from the logs it seems it is calling convolution ops from ACL ``` ) const [with T = dnnl::impl::cpu::aarch64::acl_indirect_gemm_resource_t; dnnl::impl::resource_mapper_t::key_t = const "...) at assert.c:92 ``` -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
