Reduce OpenCL start time by using a self made compiler cache.

- store compiled OpenCL binary
- load OpenCl binary if available

fix double definition of define `MEM_CHUNK`

guard error types those are only defined in OpenCL >1.1

- add new option `comp_mode` to the amd config
- disable `if guards` within opencl kernel if `comp_mode : false`

allow usage of Mesa OpenCl

- add new option for `strided_index`
- add additional option if `strided_index == 2` to controll the memory chunk with

reduce usage of registers: based on the suggestion of @enerc77

The AMD jconf.cpp would only accept an array. The config sample, and the nvidia and cpu config work with the value 'null', as they accept a 'kNullType'. This means at current, AMD GPUs could not be turned off, the config file wouldn't parse on load.

This change makes AMD consistent with the others, and can now be disabled.

fix that the GPU name is printed before the name is querried from OpenCL

- enable L1 cache for Nvidia Volta GPUs and newer
- remove explicit cache controll for Volta GPU and newer

This pull request increases the hash rate for Volta GPUs by ~5%

increase the intensity limit for AEON

The cpu miner backend uses the wrong ranges of nonces instead of using `[startNonce,startNonce + nonce_chunk)`
(startNonce,startNonce + nonce_chunk]` is used. This will results in an overlap with nonces used by the gpu back-ends.

revert the change that the OpenCl function `rotate` is used instead of `SKEIN_ROT`

- Eliminate modulus math (It runs slow inside microcode).
- Convert whatever the hell was going on into a rotate op. Like... Someone kinda reinvented the wheel in order to do a simple rotate.

reduce memory usage to 1GiB for NVIDIA devices with <=6 SMX

Ignore any affinity >=64 and throw a warning.

- if the intensity is zero than do not suggest a config
- remove the links to old issues

update VEGA names

bug was introduced with #67

- increase the L3 sanity check to 2GiB
- fix usage of byte instead of KB

Changed capitalization of "macOS"

Squashed the commit

Occassionally the auto adjust doesn't find enough memory and the
intensity is detected too low and aligned to 0 with the compute units.
This patch fixes this situation by issuing a warning with a suggestion
to set environment vars and then ignoring the alignment to 0

Per several issues:

Principally:
https://github.com/fireice-uk/xmr-stak/issues/81

Related:
https://github.com/fireice-uk/xmr-stak/issues/490
https://github.com/fireice-uk/xmr-stak/issues/472

remove branch differgences in AMD OpenCl code based on #454

a Please enter the commit message for your changes. Lines starting

Add messages


Missing include


1

- fix cuda9.1 compile (remove includ eof device_functions.hpp/ removed with cuda9.1)
- remove NVIDIA Volta gpus for MAC OSX

handle all error codes from the cuda api calls.

Makes casting more explicit.

When compiled with VS2017, the negative applied to the uint wait
time is ignored. Fixed by casting first.

- fix indention

Be more conservative with the auto suggestion.

- increase bfactor if `smx <= 6`
- limit memory for pascal < GTX1070 to 2GiB
- limt memory for pascal <= GTX1080 to 4GiB

- add macro `CUDA_CHECK_MSG_KERNEL` and `CUDA_CHECK_MSG`
- add suggestion of typicle errors can be solved