I used gpu on TX1 for Sift(a feature matching algorithm).After sevel minutes’ executing it crashed with the error warning:
gk20a gpu.0:gk20a_fifo_handle_mmu_fault:mmu fault on engine 0,engine subid 0(gpc),client 10(t1 3),addr 0x00000000:0x22400000,type 2(pte),info 0x00000a82,inst_ptr 0x148b56000
gk20a gpu.0:gk20a_fifo_set_ctx_mmu_error_tsg:TSG 0 generated a mmu fault
gk20a gpu.0:gk20a_error_notifier:error notifier set to 31 for ch 502
safeCall()Runtime API error in file<>,line 247:unspecified launch failure.[224.191590] gk20a gpu.0 :gk20a_set_error_notifier:error notifier set to 31 for ch 501
gk20a gpu.0: gk20a_set_error_notifier :error notifier set t 31 for ch 500
gk20a gpu.0: gk20a_set_error_notifier :error notifier set t 31 for ch 499
gk20a gpu.0: fifo_error_isr:channel reset initiated from fifo_error_isr;intr=0x100000000

I don’t know how to locate the error ,and the file<>,line 247: safeCall(cudaThreadSynchronize());
Will you help me how to deal with this bug.Thanks.


Could you share the source for us checking?

One possibility is compiling with the wrong GPU architecture.
TX1 GPU architecture is sm_53. Please remember to build the kernel with sm_53.

By the way, we have FAST implementation under OpenVX standard.
If you are interested in this, please install VisionWorks via JetPack.