Version 5 (11 Nov 2016)

Optimization: major 2x speedup (eXtremal) by storing 8 atomic counters in 1 uint, and by reducing branch divergence when iterating over and XORing Xi's. Note that as a result of these optimizations, sa-solver compiled with NR_ROWS_LOG=20 now only finds 182 out of 196 existing solutions ("make test" verification data was adjusted accordingly)
Defaulting OPTIM_SIMPLIFY_ROUND to 1; GPU memory usage down to 0.8 GB per instance
Optimization: significantly reduce CPU usage and PCIe bandwidth (before: ~100 MB/s/GPU, after: 0.5 MB/s/GPU), accomplished by filtering invalid solutions on-device
Optimization: reduce size of collisions[] array; +7% speed increase measured on RX 480 and R9 Nano using AMDGPU-PRO 16.40
Implement stratum method client.reconnect
Avoid segfault when encountering an out-of-range input
For simplicity -i <header> now only accepts 140-byte headers
Update README.md with Nvidia performance numbers
Fix mining on Xeon Phi and CPUs (fix OpenCL warnings)
Fix compilation warnings and 32-bit platforms

Version 4 (08 Nov 2016)

Add Nvidia GPU support (fix more unaligned memory accesses)
Add nerdralph's optimization (OPTIM_SIMPLIFY_ROUND) for potential +30% speedup, especially useful on Nvidia GPUs
Drop the Python 3.5 dependency; now requires only Python 3.3 or above (lhl)
Drop the libsodium dependency; instead use our own SHA256 implementation
Add nicehash compatibility (stratum servers fixing 17 bytes of the nonce)
Only apply set_target to next mining job
Do not abandon previous mining jobs if clean_jobs is false
Fix KeyError's when displaying stats
Be more robust about different types of network errors during connection
Remove bytes.hex() which was only supported on Python 3.5+.

SILENTARMY is now a full miner, not just a solver; the solver binary was renamed "sa-solver" and the miner is the script "silentarmy"
Multi-GPU support
Stratum support for pool mining
Reduce GPU memory usage to 671 MB (NR_ROWS_LOG=19) or 1208 MB (NR_ROWS_LOG=20, default, ~10% faster than 19) per Equihash instance
Rename --list-gpu to --list and list all OpenCL devices (not just GPUs)
Add support for multiple OpenCL platforms: --list now scans all available platforms, numbering devices using globally unique IDs
Improve correctness: find ~0.09% more solutions

Support GCN 1.0 / remove unaligned memory accesses (because of this bug, previously SILENTARMY always reported 0 solutions on GCN 1.0 hardware)
Minor performance improvement (~1%)
Get rid of "kernel.cl" and move the OpenCL code to a C string embedded in the binary during compilation
Update README with instructions for installing Radeon Software Crimson Edition (fglrx.ko) in addition to AMDGPU-PRO (amdgpu.ko)