tezos-opencl: Occasional CL_INVALID_VALUE on passwords transfer to device during auto-tune

I previously only saw this on Xeon Phi, but today it also occurred on the old Titan Kepler (just once, not reproducible):

binary size 3450038
LWS=7 GWS=49 (7 blocks) PASS,
Test mask: ?a?a?l?u?d?d?s
Calculating best GWS for LWS=32; max. 100 ms single kernel invocation.
Raw speed figures including buffer transfers:
Tuning for password length 7
xfer: 2.528 us, init: 268.160 us, crypt: 4x35.609 ms, final: 81.717 ms, res xfer: 2.112 us
gws:       256     1140 c/s     2334720 rounds/s  224.429 ms per crypt_all()!
xfer: 4.256 us, init: 265.888 us, crypt: 4x114.488 ms (exceeds 100 ms)
xfer: 1.664 us, init: 276.224 us, crypt: 4x119.263 ms (exceeds 100 ms)
Calculating best LWS for GWS=256
Testing LWS=32 GWS=256 ... 229.469 ms+
Testing LWS=64 GWS=256 ... 229.605 ms
Testing LWS=128 GWS=256 ... 225.517 ms+
Testing LWS=256 GWS=256 ... 224.184 ms+
Calculating best GWS for LWS=256; max. 200 ms single kernel invocation.
Raw speed figures including buffer transfers:
xfer: 50.336 us, init: 79.685 ms, crypt: 4x224.813 ms (exceeds 200 ms)
0: OpenCL CL_INVALID_VALUE (-30) error in opencl_tezos_fmt_plug.c:312 - Copy data to gpu
 (error occurred)LWS=256 GWS=5376 (21 blocks) DONE
Raw:    5144 c/s real, 8885 c/s virtual, Dev#6 util: 100%

(The device was in other concurrent use.)

These errors look benign in all cases I've seen so far, but I am puzzled by them anyway. I suppose we might end up closing this issue with no action, but I just want to have this recorded in here and maybe someone will see the problem.

Inspect opencl_tezos_fmt_plug.c at line 312, where the logged CL_INVALID_VALUE occurs during password transfer to the GPU. Reproduce the tezos-opencl auto-tune on a concurrently used device and determine whether the transfer error has a specific cause or is benign; done means identifying a fix or documenting why no action is needed.