Optimized curl transform formulas

[jserv]

powsrv.io team optimized the curl transform formulas, claiming 7% speedup. dcurl should benefit from the changes proposed in entangled PR #803.

Reference: Curl improvement in IRI.

[wusyong]

This uses Quine-McCluskey algo. to optimize transform function. While BUILD_SSE has significant improvement by ~10%, BUILD_AVX only gets ~1%. The benchmark is taken on node0 with commit 4afc057.

SSE

make check BUILD_STAT=1:

• Original:

[dcurl] Implementation CPU (Intel SSE) is initialized successfully
PoW execution times: 100 times.
Hash rate average value: 7997.251 kH/sec,
with the range +- 109.604 kH/sec including 95% of the hash rate values.

• Optimized:

[dcurl] Implementation CPU (Intel SSE) is initialized successfully
PoW execution times: 100 times.
Hash rate average value: 9075.665 kH/sec,
with the range +- 112.197 kH/sec including 95% of the hash rate values.

AVX

make check BUILD_AVX=1 BUILD_STAT=1:

• Original:
  Hash rate are around 94XX kH/s with multiple tests

[dcurl] Implementation CPU (Intel AVX) is initialized successfully
PoW execution times: 100 times.
Hash rate average value: 9452.758 kH/sec,
with the range +- 394.444 kH/sec including 95% of the hash rate values.

• Optimized:
  Hash rate are around 95XX kH/s with multiple tests

[dcurl] Implementation CPU (Intel AVX) is initialized successfully
PoW execution times: 100 times.
Hash rate average value: 9578.751 kH/sec,
with the range +- 341.911 kH/sec including 95% of the hash rate values.

[jserv]

@wusyong, Can you manually disable AVX2-specific execution paths and compare again?

[wusyong]

@jserv Results remain same. And after testing AVX2 on devorg the results have no change and sometimes worse. I even just removed operator for ngamma and alpha then it will be worse.

AVX2

Original:
Hash rate is around 25xxx kH/sec

[dcurl] Implementation CPU (Intel AVX) is initialized successfully
PoW execution times: 100 times.
Hash rate average value: 25524.854 kH/sec,
with the range +- 6373.506 kH/sec including 95% of the hash rate values.

Optimized:
Hash rate is around 25xxx~24XXX kH/sec

[dcurl] Implementation CPU (Intel AVX) is initialized successfully
PoW execution times: 100 times.
Hash rate average value: 24271.330 kH/sec,
with the range +- 7454.857 kH/sec including 95% of the hash rate values.

[jserv]

Extra parameters -march=native -mtune=native might be set to CFLAGS if we would like to ask GCC to operate normally without causing #GP(0) on any byte-granularity alignment (unlike SSE instructions).

IMHO, applying Quine-McCluskey algorithm for micro-optimization is reasonable, and we can rework AVX backend later. Please send pull request(s) along with appropriate explanation.

Reference: https://software.intel.com/sites/default/files/managed/b4/3a/319433-024.pdf

[muXxer]

You are welcome ;)