blst_evm384

Assembly implementation of EVM384 precompiles using blst

Includes translation of assembly to Go assembly for current x86 processors

Assembly code performs generic 384-bit constant time modular operations (add, sub, mul)

Includes a version of constant time C code for 64-bit operating systems

See EVM384 post for more details on the proposal

Status

Warning - this library is a work in progress

For EVM384 performance demonstration purposes only

C

Build

./build.sh

Test

Random testing comparing C code output with assembly

./test_evm384

Re-run benchmark

./bench_evm384

Note the benchmark expects a stable frequency

Stablize CPU Operation

In order to get consistent results run to run and to compare against other platforms, a true operation cycle count is collected. This requires the CPU frequency to be stable during the run.

Disable turbo and frequency scaling (requires root access)

for i in /sys/devices/system/cpu/cpu\*/cpufreq/scaling_governor
do
  echo "performance" | sudo tee /sys/devices/system/cpu/cpu$i/cpufreq/scaling_governor
done

Intel

echo "1" | sudo tee /sys/devices/system/cpu/intel_pstate/no_turbo

AMD

echo "0" | sudo tee /sys/devices/system/cpu/cpufreq/boost

Another option is to disable hyper threading if results are inconsistent

Go

go test -bench=.

Performance

Example of results from 3970X AMD running at stable 3.7 GHz

Assembly

Function	cycles/op	ns/op at 3.7 GHz
Add	18.6	5.0
Sub	18.8	5.1
Mul	119.7	32.3

Use the cycles/op to determine performance on your machine. Divide the cycles/op by your frequency (in GHz) to get nsecs/op.

For example from above:

18.6 / 3.7 = 5.0 ns/op

Go Assembly

Function	cycles/op	ns/op at 3.7 GHz
Add	18.8	5.1
Mul	128.0	34.6

Note the functions are generic 384-bit, therefore performance should be field size independent for all curves 384-bit and under.