Perf optimization. I-cache aware?

[daminakaTT]

Ticket

N/A

Problem description

erisc has performance issue
Current performance is 95% 6.35 B/cycle, 5% 6.83 B/cycle

What's changed

• 2 element unrolling of checksum calculation
  95% 6.35 B/cycle, 6.93 B/cycle
• Remove unrolling from copy loop
  100% 7.05 B/cycle
• Both
  100% 7.06 B/cycle

Originally two different performances might come from cache-memory access difference.
For accumulation, 2 elements unroll is the best.
For simple copy, no unrolling is the best.
I'm assuming that unroll for copy makes instruction to be big so instruction copy happens many time.
2 elements unroll might be thanks to 16 bytes operation?

Checklist

• All post commit CI passes
• Blackhole Post commit CI passes (if applicable)
• Model regression CI passes (if applicable)
• Device performance regression CI passes (if applicable)
• (For models and ops writers) Full new models tests CI passes (if applicable)
• New/Existing tests provide coverage for changes