Allocation free synthesis

[winston-h-zhang]

Addresses #137.

This PR introduces WitnessViewCS, a slightly modified version of WitnessCS that does not own the witness buffers. Instead, we have RecursiveSNARK own the witness buffers, and WitnessViewCS takes mutable handles into those buffers.

By allowing RecursiveSNARK to directly manage its own witness memory, we remove the extra allocations WitnessCS demands.

Benchmark results from downstream lurk-rs confirm that this improves performance: argumentcomputer/lurk-beta#922

Pasted Benchmark

Click to view benchmark

Test	Base	PR	%
LEM Fibonacci Prove - rc = 100/fib/num-100	6.0±0.06s	5.6±0.04s	-6.67%
LEM Fibonacci Prove - rc = 100/fib/num-200	12.4±0.07s	12.0±0.04s	-3.23%
LEM Fibonacci Prove - rc = 600/fib/num-100	5.3±0.06s	N/A	N/A
LEM Fibonacci Prove - rc = 600/fib/num-200	12.1±0.04s	N/A	N/A

[huitseeker]

I like the direction of this, but it adds a ConstraintSystem that :

• for what it's worth, isn't implementing the trait correctly (e.g. new),
• adds 100 lines there for a ConstraintSystem that could live in bellpepper,
• seems to operate the same changes as Pre-allocate in SatisfyingAssignment<G> #124 + refactor: Refactor R1CS for in-place operations and Vec inputs #121, which collectively are a change that's about 10x smaller than here. without needing an additional ConstraintSystem

[winston-h-zhang]

Benchmarks

Table of Contents

• Benchmark Results
  • LEM Fibonacci Prove Fresh - rc = 100
  • LEM Fibonacci Prove Fresh - rc = 600
  • LEM Fibonacci Prove Load - rc = 100
  • LEM Fibonacci Prove Load - rc = 600

Benchmark Results

LEM Fibonacci Prove Fresh - rc = 100

	main/fib	witness-view-cs/fib	new-witness-cs/fib
num-100	6.39 s (1.00x)	7.14 s (1.11x)	6.91 s (1.00x)
num-200	13.59 s (1.00x)	14.26 s (1.05x)	14.76 s (1.00x)
adthrpt	222.2 (1.00x)	224.7 (1.01x)	203.8 (0.92x)

LEM Fibonacci Prove Fresh - rc = 600

	main/fib	witness-view-cs/fib	new-witness-cs/fib
num-100	4.54 s (1.00x)	4.46 s (0.98x)	4.45 s (1.00x)
num-200	10.58 s (1.00x)	10.38 s (0.98x)	10.40 s (1.00x)
adthrpt	298.0 (1.00x)	304.0 (1.02x)	302.5 (1.01x)

LEM Fibonacci Prove Load - rc = 100

	main/fib	witness-view-cs/fib	new-witness-cs/fib
num-100	6.39 s (1.00x)	6.30 s (0.98x)	6.26 s (1.00x)
num-200	13.05 s (1.00x)	12.68 s (0.97x)	12.66 s (1.00x)
adthrpt	240.2 (1.00x)	250.8 (1.04x)	250.0 (1.04x)

LEM Fibonacci Prove Load - rc = 600

	main/fib	witness-view-cs/fib	new-witness-cs/fib
num-100	5.78 s (1.00x)	5.74 s (1.00x)	5.77 s (1.00x)
num-200	15.25 s (1.00x)	13.33 s (0.87x)	13.59 s (1.00x)
adthrpt	190.1 (1.00x)	237.2 (1.25x)	230.2 (1.21x)

---

Made with criterion-table

[winston-h-zhang]

I remain skeptical of using the WitnessCS API, since it still allocates large vectors on a per-prove-step basis. After running benchmarks though, it seems the difference is not significant. I've posted the results for transparency.

I'll hold off closing this for now, but a new PR with the WitnessCS API will be up, and we can merge that one.