i use 20201208_0508315 release.
my code is
ncnn::Net aNet;
aNet.load_param(param_path.c_str());
aNet.load_model(model_path.c_str());
arm64 device(iPhone 7) work fine, but armv7 device(iPhone 4S) crash at load_model().
call stack is
thread #6, queue = 'com.testApp, stop reason = EXC_BAD_ACCESS (code=EXC_ARM_DA_ALIGN, address=0x169aca2c)
* frame #0: 0x00e4c04c testApp`__kmp_hyper_barrier_release(barrier_type, kmp_info*, int, int, int) + 460
frame #1: 0x00e4d7b4 testApp`__kmp_fork_barrier(int, int) + 210
frame #2: 0x00e3e3dc testApp`__kmp_fork_call + 4784
frame #3: 0x00e39b9e testApp`__kmpc_fork_call + 62
frame #4: 0x00c9efbe testApp`ncnn::conv3x3s1_winograd64_transform_kernel_pack4_neon(ncnn::Mat const&, ncnn::Mat&, int, int) + 194
frame #5: 0x00c9cc78 testApp`ncnn::Convolution_arm::create_pipeline(ncnn::Option const&) + 5184
frame #6: 0x00c8eda4 testApp`ncnn::Net::load_model(ncnn::DataReader const&) + 294
frame #7: 0x00c8eeb2 testApp`ncnn::Net::load_model(char const*) + 48
thread #7
frame #0: 0x00e7578e testApp`__kmp_suspend_64 + 58
frame #1: 0x00e6eb64 testApp`kmp_flag_64::wait(kmp_info*, int) + 704
frame #2: 0x00e6cecc testApp`__kmp_hyper_barrier_release(barrier_type, kmp_info*, int, int, int) + 76
frame #3: 0x00e6e7b4 testApp`__kmp_fork_barrier(int, int) + 210
frame #4: 0x00e60938 testApp`__kmp_launch_thread + 146
frame #5: 0x00e74a18 testApp`__kmp_launch_worker(void*) + 66
frame #6: 0x37c08872 libsystem_pthread.dylib`<redacted> + 138
frame #7: 0x37c087e6 libsystem_pthread.dylib`_pthread_start + 110
frame #8: 0x37c06740 libsystem_pthread.dylib`thread_start + 8
hi, please provide your model param file for reproducing the issue
I also reproduced with example param and model files.
I tested it with previous releases. (Imported ncnn.framework and openmp.framework for each package.)
20210124_5e4ea0b => crash
I suspect there is some regression bug during the openmp upgrading from 7.0.0 to 11.0.0
Try linking with openmp 20200226 with the latest ncnn and see if it makes any difference
Most helpful comment
hi, please provide your model param file for reproducing the issue