SebastianBoblestETAS opened a new issue, #11394: URL: https://github.com/apache/tvm/issues/11394
I already reported this issue here: https://discuss.tvm.apache.org/t/problem-with-allocateconstnodes-in-cmsis-nn-code/12806 It seems that AllocateConstNodes can be overwritten in models that contain both CMSIS-NN functions and functions generated by TVM. Concretely, tvmgen_conv1d_ave_fused_cast_subtract_clip_cast in conv1d_ave_lib1.c uses constant_0 that is defined as static const int32_t __attribute__((section(".rodata.tvm"), aligned(16))) constant_0[1] = { -0x00000080 }; in conv1d_ave_lib1.c. At the same time tvmgen_conv1d_ave_cmsis_nn_main_0 is also called with constant_0 as "filter" argument, but the filter_dims are cmsis_nn_dims filter_dims = {8,1,2,3}; in conv1d_ave_lib2.c I use a local installation of TVM on Linux that is only a few hours old. I attached a zip file that contains the following: * A txt file with the tvmc command that I used, including the command line output from tvm * The tflite model * The corresponding keras model in h5 format * the python script that I used to quantize the keras model * The tar file that tvmc produced with the relevant files in codegen/host/src I hope I did not confuse anything here or make a mistake in the invocation of tvmc. Best, Sebastian @grant-arm @Mousius @manupa-arm @ashutosh-arm [cmsis_report.zip](https://github.com/apache/tvm/files/8741594/cmsis_report.zip) -- 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]
