Open zenogantner opened 5 years ago
-O1 and -O2 also give an FPE.
Here is the backtrace:
(gdb) bt
#0 0x00005555555e84b6 in fxdiv_init_uint64_t (d=<optimized out>) at /home/zgantner/src/NNPACK-darknet/deps/fxdiv/include/fxdiv.h:261
#1 0x00005555555e84b6 in fxdiv_init_size_t (d=<optimized out>) at /home/zgantner/src/NNPACK-darknet/deps/fxdiv/include/fxdiv.h:313
#2 0x00005555555e84b6 in compute_gemm_convolution_inference (input_channels=input_channels@entry=1024, output_channels=output_channels@entry=256, input_size=..., kernel_size=..., output_size=..., output_subsampling=..., input=0x7fffc5329010, kernel=0x7fffc502a010, bias=0x0, output=0x7fffc4efe010, workspace_buffer=0x0, workspace_size=0x0, activation=nnp_activation_identity, threadpool=0x5555566da040, profile=0x0, input_padding=...) at /home/zgantner/src/NNPACK-darknet/src/convolution-inference.c:798
#3 0x00005555555ea8c5 in nnp_convolution_inference (algorithm=<optimized out>, transform_strategy=nnp_convolution_transform_strategy_compute, input_channels=1024, output_channels=256, input_size=..., input_padding=..., kernel_size=Python Exception <class 'gdb.MemoryError'> Cannot access memory at address 0x8042:
#4 0x000055555557d583 in forward_convolutional_layer_nnpack ()
#5 0x00005555555c07cc in forward_network ()
#6 0x00005555555c10db in network_predict ()
#7 0x0000555555573e36 in test_detector ()
#8 0x00005555555746ad in run_detector ()
#9 0x0000555555559b9f in main ()
How to reproduce:
Install darknet-nnpack following the instructions in the README file.
I do not really understand what the issue is there.
If I compile the library with -O0 instead of -O3, there is no FPE.
I see the same issue with full-size YOLOv3.