Remove `static constexpr` within the global kernel. This is not
supported by all CUDA versions.

optimize the algorithm for cryptonight_gpu autosuggestion

psychocrypt committed 9 minutes ago
 - use precomuted indicies within the loop
 - `cn_explode_gpu` use all threads to load the state

- fix race condition during shared memory access
- optimize memory access

Co-authored-by: psychocrypt <psychocryptHPC@gmail.com>
Co-authored-by: fireice-uk <fireice-uk@users.noreply.github.com>

- add helper method `GetAllAlgorithms()` to get all active POW
algorithms
- select max scratchpad memory size based on the dev pool and user
algorithms

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>

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`

- 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

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.

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

Read memory in bigger chunks per thread to increase the used memory bandwith.
Use for Kepla and Fermi GPUs the old autosuggestion instead of the new settings for cryptonight_v8.

I disabled a few algorithms for fatser compile and missed to re-enable them.

`uint` is unknown in windows, therefore switch to the better type `uint32_t`

- introduce a new schema where two threads work together on one hash
- update autoadjustment
- remove an mistake where shared memory was shrinked for gpus < sm_70

Avoid branche differegence

- fix that shared memory for fast div is always used even if an algorithm is not using it
- optimize fast div algo
- store `division_result` (64_bit) per thread instead of shuffle around and store it as 32bit

- use optimzed div and sqrt
- reduce memory footprint

Add fast version for div and sqrt for the cuda backend

- use shared memory to exchange

implement `cryptonight_v8`

5% more hash per second

- add cryptonight_heavy derivate cryptonight_bittube2
- add coin bittube
- remove coin ipbc because this coin is now called bittube

- update pools.txt
- add new algorithm `cryptonight_haven`
- update all backends

- rename cryptonight_fast to cryptonight_masari
- set dev pool to cryptonight_monero

If CUDA 9.X is used and the miner is compiled for `sm_70` and used with Volta GPUs than the miner
deadlocks if `threads` is not a multiple of `32`.

- use `__activemask()` to get all active lanes

solve #1494

- add algorithm `cryptonight_v7_stellite` (internal named: `cryptonight_stellite`)

- add algorithm `cryptonight_lite_v7_xor`
- update documentation