Auto witgen

[chriseth]

No description provided.

[chriseth]

This can now generate the following code for the binary machine:

//Known inputs: main_binary::operation_id, main_binary::A, main_binary::B
let main_binary::sel[1]_u3 = 0
let main_binary::sel[1]_u2 = 0
let main_binary::sel[1]_u1 = 0
let main_binary::sel[1]_d0 = 0
let main_binary::sel[1]_d1 = 0
let main_binary::sel[1]_d2 = 0
let main_binary::sel[1]_d3 = 0
let main_binary::sel[2]_u3 = 0
let main_binary::sel[2]_u2 = 0
let main_binary::sel[2]_u1 = 0
let main_binary::sel[2]_d0 = 0
let main_binary::sel[2]_d1 = 0
let main_binary::sel[2]_d2 = 0
let main_binary::sel[2]_d3 = 0
let main_binary::operation_id_u1 = main_binary::operation_id_d0
let main_binary::A_byte_u1 = (main_binary::A_d0 & 0xff000000) / 16777216;
let main_binary::A_u1 = main_binary::A_d0 & 0xffffff;
let main_binary::B_byte_u1 = (main_binary::B_d0 & 0xff000000) / 16777216;
let main_binary::B_u1 = main_binary::B_d0 & 0xffffff;
let mut lookup_20_3: T = 0.into();
fixed_lookup_machine.process_lookup_direct((20, vec![LookupCell::Input(&(main_binary::operation_id_d0)), LookupCell::Input(&(main_binary::A_byte_u1)), LookupCell::Input(&(main_binary::B_byte_u1)), LookupCell::Output(&mut lookup_20_3)]))
let main_binary::C_byte_u1 = lookup_20_3
let main_binary::sel[0]_d0 = 1
let main_binary::operation_id_u2 = main_binary::operation_id_u1
let main_binary::A_byte_u2 = (main_binary::A_u1 & 0xff0000) / 65536;
let main_binary::A_u2 = main_binary::A_u1 & 0xffff;
let main_binary::B_byte_u2 = (main_binary::B_u1 & 0xff0000) / 65536;
let main_binary::B_u2 = main_binary::B_u1 & 0xffff;
let mut lookup_20_3: T = 0.into();
fixed_lookup_machine.process_lookup_direct((20, vec![LookupCell::Input(&(main_binary::operation_id_u1)), LookupCell::Input(&(main_binary::A_byte_u2)), LookupCell::Input(&(main_binary::B_byte_u2)), LookupCell::Output(&mut lookup_20_3)]))
let main_binary::C_byte_u2 = lookup_20_3
let main_binary::operation_id_u3 = main_binary::operation_id_u2
let main_binary::A_byte_u3 = (main_binary::A_u2 & 0xff00) / 256;
let main_binary::A_u3 = main_binary::A_u2 & 0xff;
let main_binary::B_byte_u3 = (main_binary::B_u2 & 0xff00) / 256;
let main_binary::B_u3 = main_binary::B_u2 & 0xff;
let mut lookup_20_3: T = 0.into();
fixed_lookup_machine.process_lookup_direct((20, vec![LookupCell::Input(&(main_binary::operation_id_u2)), LookupCell::Input(&(main_binary::A_byte_u3)), LookupCell::Input(&(main_binary::B_byte_u3)), LookupCell::Output(&mut lookup_20_3)]))
let main_binary::C_byte_u3 = lookup_20_3
let main_binary::A_byte_u4 = main_binary::A_u3
let main_binary::B_byte_u4 = main_binary::B_u3
let mut lookup_20_3: T = 0.into();
fixed_lookup_machine.process_lookup_direct((20, vec![LookupCell::Input(&(main_binary::operation_id_u3)), LookupCell::Input(&(main_binary::A_byte_u4)), LookupCell::Input(&(main_binary::B_byte_u4)), LookupCell::Output(&mut lookup_20_3)]))
let main_binary::C_byte_u4 = lookup_20_3
let main_binary::sel[0]_u3 = 0
let main_binary::sel[0]_u2 = 0
let main_binary::sel[0]_u1 = 0
let main_binary::sel[0]_d0 = 0
let main_binary::sel[0]_d1 = 0
let main_binary::sel[0]_d2 = 0
let main_binary::sel[0]_d3 = 0

[chriseth]

Ran an extended version of binary_large_test.asm that just loops infinitely and extended the degree of the main machine to 2**20. The binary machine is jit-computed. The compilation time is included in the statistics below.

 == Witgen profile (12469602 events)
   43.7% (   12.4s): Main Machine
   25.9% (    7.3s): Secondary machine 0: main_binary (BlockMachine)
   25.0% (    7.1s): FixedLookup
    5.5% (    1.5s): witgen (outer code)
    0.0% ( 470.0ns): range constraint multiplicity witgen
  ---------------------------
    ==> Total: 28.293288704s

Used rows in binary: 8161892
Used rows in main: 4194304

Witgen takes 800 ns for one lookup to the binary machine (this includes the time spent in the fixed lookup machine)

The same .asm file run on current main:

 == Witgen profile (36955286 events)
   61.4% (   39.0s): Secondary machine 0: main_binary (BlockMachine)
   20.3% (   12.9s): Main Machine
   15.9% (   10.1s): FixedLookup
    2.5% (    1.6s): witgen (outer code)
    0.0% ( 412.0ns): range constraint multiplicity witgen
  ---------------------------
    ==> Total: 63.642381077s
    ```

[chriseth]

Make this extern "C"?

[chriseth]

Change LookupCell to use "C" repr

[chriseth]

Now I got it down to roughly 600. I think we can save another 100ms by improving how the known bits are stored (padded to a u32 for each new row).

[chriseth]

Yep, now we have 540ns per block

[chriseth]

We should change process_lookup_direct not to take a Vec, because then we need to allocate. Instead, we should take a mutable ref to a slice and create it on the stack.