[FEAT] CUDA integration

[lighter-zz]

run tests

install and compile zeknox properly. then

./test_gpu.sh

Benchmark

tl;dr: CPU 12 s, GPU 3.5 s; on AWS g6

CPU timing

[INFO  plonky2::util::timing] 12.0002s to prove
[DEBUG plonky2::util::timing] | 0.2251s to run 1000133 generators
[DEBUG plonky2::util::timing] | 0.0402s to compute full witness
[DEBUG plonky2::util::timing] | 0.0062s to compute wire polynomials
[DEBUG plonky2::util::timing] | 6.8379s to compute wires commitment
[DEBUG plonky2::util::timing] | | 0.0487s to IFFT
[DEBUG plonky2::util::timing] | | 1.7425s to FFT + blinding
[DEBUG plonky2::util::timing] | | 0.6274s to transpose LDEs
[DEBUG plonky2::util::timing] | | 4.4080s to build Merkle tree
[DEBUG plonky2::util::timing] | 0.0707s to compute partial products
[DEBUG plonky2::util::timing] | 1.5780s to commit to partial products, Z's and, if any, lookup polynomials
[DEBUG plonky2::util::timing] | | 0.0109s to IFFT
[DEBUG plonky2::util::timing] | | 0.2813s to FFT + blinding
[DEBUG plonky2::util::timing] | | 0.0214s to transpose LDEs
[DEBUG plonky2::util::timing] | | 1.2628s to build Merkle tree
[DEBUG plonky2::util::timing] | 1.0890s to compute quotient polys
[DEBUG plonky2::util::timing] | 0.0083s to split up quotient polys
[DEBUG plonky2::util::timing] | 1.6540s to commit to quotient polys
[DEBUG plonky2::util::timing] | | 0.3139s to FFT + blinding
[DEBUG plonky2::util::timing] | | 0.0236s to transpose LDEs
[DEBUG plonky2::util::timing] | | 1.3143s to build Merkle tree
[DEBUG plonky2::util::timing] | 0.0427s to construct the opening set, including lookups
[DEBUG plonky2::util::timing] | 0.4050s to compute opening proofs
[DEBUG plonky2::util::timing] | | 0.1607s to reduce batch of 255 polynomials
[DEBUG plonky2::util::timing] | | 0.0013s to reduce batch of 2 polynomials
[DEBUG plonky2::util::timing] | | 0.0489s to perform final FFT 524288
[DEBUG plonky2::util::timing] | | 0.1215s to fold codewords in the commitment phase
[DEBUG plonky2::util::timing] | | 0.0696s to find proof-of-work witness

GPU timing

[INFO  plonky2::util::timing] 3.5132s to prove
[DEBUG plonky2::util::timing] | 0.2128s to run 1000133 generators
[DEBUG plonky2::util::timing] | 0.0428s to compute full witness
[DEBUG plonky2::util::timing] | 0.0062s to compute wire polynomials
[DEBUG plonky2::util::timing] | 1.7011s to compute wires commitment
[DEBUG plonky2::util::timing] | | 0.0266s to IFFT
[DEBUG plonky2::util::timing] | | 0.3741s to FFT + blinding
[DEBUG plonky2::util::timing] | | 0.5789s to transpose LDEs
[DEBUG plonky2::util::timing] | | 0.7118s to build Merkle tree
[DEBUG plonky2::util::timing] | 0.0558s to compute partial products
[DEBUG plonky2::util::timing] | 0.2503s to commit to partial products, Z's and, if any, lookup polynomials
[DEBUG plonky2::util::timing] | | 0.0042s to IFFT
[DEBUG plonky2::util::timing] | | 0.0500s to FFT + blinding
[DEBUG plonky2::util::timing] | | 0.0210s to transpose LDEs
[DEBUG plonky2::util::timing] | | 0.1736s to build Merkle tree
[DEBUG plonky2::util::timing] | 0.7050s to compute quotient polys
[DEBUG plonky2::util::timing] | 0.0059s to split up quotient polys
[DEBUG plonky2::util::timing] | 0.2160s to commit to quotient polys
[DEBUG plonky2::util::timing] | | 0.0404s to FFT + blinding
[DEBUG plonky2::util::timing] | | 0.0167s to transpose LDEs
[DEBUG plonky2::util::timing] | | 0.1574s to build Merkle tree
[DEBUG plonky2::util::timing] | 0.0308s to construct the opening set, including lookups
[DEBUG plonky2::util::timing] | 0.2460s to compute opening proofs
[DEBUG plonky2::util::timing] | | 0.1483s to reduce batch of 255 polynomials
[DEBUG plonky2::util::timing] | | 0.0012s to reduce batch of 2 polynomials
[DEBUG plonky2::util::timing] | | 0.0441s to perform final FFT 524288
[DEBUG plonky2::util::timing] | | 0.0195s to fold codewords in the commitment phase
[DEBUG plonky2::util::timing] | | 0.0302s to find proof-of-work witness