In this case it doesn't, since the function is only used for config
parsing (I hope). Add the check to prevent a memory leak, just in
case this changes in the future.
TODO: Option parsing functions should really be moved elsewhere and
renamed appropriately, like set_opt_kernel() or config_set_kernel(),
if config parsing is to remain as-is. It will probably change
drastically, so leaving as-is for now.
There are no other kernels, and differentiating kernels in the future
should be done in a universal way, preferably not by choosing a function.
Note how queue_kernel() is used (passed as function reference).
WIP!
Use a string instead of a state-machine-ey kernel selection mechanism
where kernel names have to be predefined. This should allow just dropping
new kernels into dir `kernel` without bloating the code in three other
places.
Is in dire need of a cleanup, function parameter check, edge case check -
all the usual testing.
In particular, checking these definitions/keywords:
* enum cl_kernels
* kname
* [c]gpu[s]->kernel (and similar)
* memory cleanup after strdup()?..
* chosen_kernel
* queue_scrypt_kernel
* strbuf
* initCl
The config file is written poorly: first off, not all available options
are written (they have to be written manually, since there's no struct
enumeration); then, default values are not commented out.
This setting allows to set the GPU intensity value directly without any modifiers, it does not
get any more raw than this! Look at the xintensity description raw for examples of regular
intensity values. You can also set this value through the ncurses interface by pressing:
G -> A -> select device id -> enter value.
Minor xintensity code cleanup as well.
Conflicts:
driver-opencl.c
miner.h
sgminer.c
Changing `intensity` and to it from `xintensity` works fine.
Changing `xintensity` sometimes fails to enqueue kernel.
For example, starting with --xintensity=128 (on a 5850) and
then changing to 64, 42, or 100 is reliable. However, changing to 127 is
not, and produces
[04:13:01] Error -54: Enqueueing kernel onto command queue. (clEnqueueNDRangeKernel)
[04:13:01] GPU 0 failure, disabling!
Manually enabling the disabled GPU is successful, but the GPU no longer
submits shares.
This might be a hardware limitation.
All of this is credited to ArGee of RGMiner, he did the initial ground work for this setting.
This new setting allows for a much finer grained intensity setting and also opens up for dual gpu threads on devices not previously able to. Note: make sure to use lower thread-concurrency values when you increase cpu threads.
Intensity is currently used to spawn GPU threads as a simple 2^value setting.
I:13 = 8192 threads
I:15 = 32768 threads
I:17 = 131072 threads
I:18 = 262144 threads
I:19 = 524288 threads
I:20 = 1048576 threads
Notice how the higher settings increase thread count tremendously.
Now enter the xintensity setting (Yes, I am a genius with my naming convention!).
It is simply a shader multiplier, obviously based on the amount of shaders you got on a card, this should allow the same value to scale with different card models.
6970 with 1536 shaders: xI:64 = 98304 threads
R9 280X with 2048 shaders: xI:64 = 131072 threads
R9 290 with 2560 shaders: xI:64 = 180224 threads
R9 290X with 2816 shaders: xI:64 = 163840 threads
6970 with 1536 shaders: xI:300 = 460800 threads
R9 280X with 2048 shaders: xI:300 = 614400 threads
R9 290 with 2560 shaders: xI:300 = 768000 threads
R9 290X with 2816 shaders: xI:300 = 844800 threads
It's now much easier to control thread intensity and it potentially allows for a uniform way of setting the intensity on your system. I'm very interested in constructive feedback, as I do not have access to a lot of different card models.
This change has been tested on 6970, R9 290, R9 290X - all with equal or a little better speeds than regular intensity setting after a little tuning, but your mileage may vary. Don't fret it, if this doesn't work for you, the regular intensity setting is still available.
Conflicts:
driver-opencl.c
sgminer.c