_prod_TT fails with CL_OUT_OF_RESOURCE on Mali-T628 #60
Description
After checking that Caffe no longer requests zero-sized OpenCL buffers (#59), I still encounter the same result on SqueezeNet 1.1:
*** Aborted at 1486746813 (unix time) try "date -d @1486746813" if you are using GNU date ***
PC: @ 0xb5c8798a mcl_entrypoints_valid_event_list
*** SIGSEGV (@0x45ccc412) received by PID 29568 (TID 0xb01c6000) from PID 1171047442; stack trace: ***
@ 0xb5774270 (unknown)
(GoogleNet behaves similarly to what I report for SqueezeNet, just doesn't segfault. AlexNet works fine.)
The first call that goes bad is:
"call": "clEnqueueNDRangeKernel",
"name": "_prod_TT",
"queue": "0xd70c0",
"kernel": "0x70f100",
"gwo": [0, 0, 0],
"gws": [16, 3200, 1],
"lws": [8, 8, 1],
"event_wait_list": [],
"event": "0",
"timestamp": {
"start": "2017-02-10T16:55:05.482380",
"end": "2017-02-10T16:55:05.523968"
},
"output profiling_error": -5,
"profiling": {
"queued": 532575944823,
"submit": 532575944823,
"start": 13079693137503322112,
"end": 34359738377
},
"errcode": 0
}
The local work size (8x8) should work for any kernel. I suspect the problem might be in the way the arguments are set, but don't have any other clue at the moment.
Please find attached a trace from dividiti's OpenCL profiler up to the failing call:
naibaf7-caffe-60.dvdt-prof.txt. Please see #59 on how to reproduce.