Scrypt mining is still broken in that it can not utilize high thread concurrencies (I need to use a thread concurrency of 24000 for my 7950s to mine effectively). The bug seems to be with the calculation for memory allocation. It can be easily duplicated by setting the thread concurrency above 8192, eg 12288. In reaper, any thread concurrency may be used as long as it is a multiple of 64 and produces a memory padding that fits within the card's memory space.
Using "set GPU_MAX_ALLOC_PERCENT=100" in Windows allows the program to run, but it doesn't hash at all and does not create a memory pad on the GPU.
I reported this several weeks ago and it still hasn't been addressed.
That's cause I have no idea what the problem is. Setting GPU_MAX_ALLOC_PERCENT does nothing on windows, only linux.Using "set GPU_MAX_ALLOC_PERCENT=100" in Windows allows the program to run, but it doesn't hash at all and does not create a memory pad on the GPU.
I reported this several weeks ago and it still hasn't been addressed.
Code:
[2012-10-05 16:01:03] Started cgminer 2.7.6
[2012-10-05 16:01:03] Started cgminer 2.7.6
[2012-10-05 16:01:04] Probing for an alive pool
[2012-10-05 16:01:04] Long-polling activated for http://ltc.kattare.com:9332/LP
[2012-10-05 16:01:06] LONGPOLL from pool 0 detected new block
[2012-10-05 16:01:15] Maximum buffer memory device 0 supports says 524288000, your scrypt settings come to 0
[2012-10-05 16:01:15] Error -61: clCreateBuffer (padbuffer8), decrease CT or increase LG
[2012-10-05 16:01:15] Failed to init GPU thread 0, disabling device 0
[2012-10-05 16:01:15] Restarting the GPU from the menu will not fix this.
[2012-10-05 16:01:15] Try restarting cgminer.
Press enter to continue:
There's the error code for you.
You should be able to easily replicate this just by setting --thread-concurrency 12288 (which works fine on reaper).
I'm pretty sure the problem has to do with these lines,
Code:
clState->padbufsize = bufsize;
clState->padbuffer8 = clCreateBuffer(clState->context, CL_MEM_READ_WRITE, bufsize, NULL, &status);
size_t bufsize = 128 * ipt * cgpu->thread_concurrency;
ipt is an unsigned integer, but thread_concurrency is a signed one. Maybe that's all there is to the overflow. I'll dick around with it before the next release.
EDIT: and by the way, I can't replicate any of this without windows...
