NVIDIA is using clang as device compiler so the reciprocal optimizations was disabled with #2104.

- re-enable optimized reciprocal calculation

- fix broken compile: change used `ULL` to `UL` because `UL` is defined as 64bit
- fix memory copy to shared memory via vload8 (somehow it create wrong access)

The auto config generates for AMD devices now by default two threads per GPU.

- remove the savety 128MiB memory now only from the max available GPU memory not from the avaialble memory for one alloc call
- extend the memory documentation in amd.txt

Due to a wrong implementation clamp was not working.

use for non clang (Rocm) OpenCL a optimized reciprocal calculation without lookup table.

Co-authored-by: SChernykh <sergey.v.chernykh@gmail.com>

Disable compatibility mode if intensity is a multiple of worksize. In that case enabled compaibility mode will only slow down the miner.

If two threads are using the same GPU device the start time of each hash round is optimized based on the average time needed to calculate a bunch of hashes.

This way to optimize the hash rate was first introduced by @SChernykh. This implementation based on the implementation in xmrig but differen in the details.

- introduce a new config option `interleave`
- implement thread interleaving

Use `mul24` to speedup the scratchpad index calculation.

Co-authored-by: SChernykh <sergey.v.chernykh@gmail.com>

Add new striding index where the memory is chunked by the size of the work group (worksize).

Co-authored-by: SChernykh <sergey.v.chernykh@gmail.com>

small optimization for non cryptonight_v8 algorithms

- optimize division

optimize cryptonight_heavy diff

- change a few 64bit variables into 32bit.
- provide defines type quallified

- remove useless `clFinish`
- avoid download num threads for skein&co and start always as much threads as in all other kernel (terminate useless threads)

Reduce local memory foot print to increase the occupancy.

Co-authored-by: SChernykh <sergey.v.chernykh@gmail.com>

port optimizations from OpenCL.

Co-authored-by: SChernykh <sergey.v.chernykh@gmail.com>

Use only the half AES matrix and compute the other half in place.
This PR increases the possible occupancy.

port OpenCl optimized division to CUDA

Co-authored-by: SChernykh <sergey.v.chernykh@gmail.com>

If CUDA is loaded before AMD but no CUDA is available it can be happen that the embadded OpenCL code is empty.
This is only an issue if the binary is builded static on a different system.

define shared memory in the outer scope

x-ref: https://github.com/xmrig/xmrig-amd/pull/192

- optimize kernel cn0 and cn2
- optimize vast int math
- use more 32bit variables

Co-authored-by: psychocrypt <psychocryptHPC@gmail.com>

optimize the devision in cryptonight_heavy and cryptonight_haven

import of https://github.com/xmrig/xmrig-amd/pull/185/commits/5d9b9334654df25cea7707f667990fd1577ed290

In the cuda backend for monero we start always twice as much threads as needed.
Those threads are than removed after the AES matrix is copied to the shared memory.
Never the less it is the result of an copy past bug.

- start correct number of threads for `monero`

Fix the fix from #1945. The initial fix produces invalid results.

The AMD compiler for OpenCL shipped with the driver 14XX is broken
and can not compile xmr-stak since the monero v8 changes are introduced.

- workaround a simple compare.
- add new device define `OPENCL_DRIVER_MAJOR`

Allow to ship the miner with multiple cuda backends those depends on different driver versions.
This will allow to support Turing/Volta and old Fermi GPU within one release.

- add support to search for the first working CUDA backend
- add some more messages to support better debugging (if a user has some issues)

- remove helper array to perform division
- tweak `get_reciprocal`

The name `comp_mode` for a memoy load pattern if a bad choosen name.
Therefore I changed it to `mem_mode` which also gives use the possibility
to add new mode later if needed.

- rename `comp_mode` to `mem_mode`
- fix documentation

In the current implementation the bit align is using signed integer which results in pulling in
ones in the case the sign bit is set.

- cast to unsigned integer before using bitshift

If `comp_mode` is false the results on a windows platform will be invalid.
The reason for that is that `ulong4` is in windows 16byte and in linux 32byte.

thx @xmrig for finding and solving the issue

fix #1873

Add a suggestion to an common line which is shown in the event of an crash under windows.

Add compatibility mode for CUDA to avoid invalid shares.

Use an array  instead of a if cascade to select the hasing function for CUDA.

- add more descriptive messages if memory allocation fails
- add gnu compiler flags: `noexecstack` to support systemd
- handle cases where memroy allocation fails

Co-authored-by: Tony Butler <spudz76@gmail.com>

Use volatile pointer to be sure that the compiler is not caching the values.

Fix wrong warning about unknown ASM type