For almost all the sponge constructions defined in the Poseidon paper,
the domain can be defined completely statically. Variable-length hashing
requires knowledge of the message length, but that can be provided to
the fixed padding function in a subsequent commit, and in any case we
can't use variable-length inputs in a circuit.
In order to make the changeover easier to review, we redefined
`CellValue<F>` to be `AssignedCell<F, F>`. Now we remove that type and
rename throughout the codebase.
As the underlying `Region` methods now return `AssignedCell` instead of
`Cell`, we can simplify all the places where we then constructed a
`CellValue` struct.
The witnessed points are cm_old, g_d_old, pk_d_old, ak.
g_d_new and pk_d_new are currently also witnessed as affine points,
which diverges from the spec.
Co-authored-by: Daira Hopwood <daira@jacaranda.org>
- Placing the Poseidon `state` columns after the `partial_sbox` column
instead of before it causes them to line up with vast stretch of free
space, enabling the pad-and-add region to be layed out there.
- Using the `Region::assign_advice_from_constant` API to initialise the
Poseidon state removes fixed-column contention between that region and
fixed-base scalar multiplication, enabling it to also be layed out
within the free space.
- If https://github.com/zcash/halo2/issues/334 were implemented then
this region would disappear.
- The overflow check in variable-base scalar mul is also moved into the
columns with free space.
Previously, fixed_y_q was a non-binary selector that both loaded
the y_Q value and toggled the y_Q gate.
Now, the gate is toggled by a q_s4 simple selector, while the value
is loaded into a separate fixed column.
The Action circuit only used standard PLONK in one place. Since it
used non-binary selectors, it cannot be optimised by the halo2
selector optimisations. We now replace it with a custom gate which
uses a binary selector.
The Sinsemilla chip witnesses message pieces in individual regions, and
then copies them into the `hash_piece` region to initialize the running
sum. Previously these occured in the same column, but we can reduce the
utilized rows of the Action circuit by moving these into a less-used
column.
If https://github.com/zcash/halo2/issues/334 is implemented, this change
would be unnecessary, as the witnessed message piece regions would never
be assigned into the circuit.
We were configuring multiple instances of this across all of the advice
columns, in order to spread their assignments. However, we are actually
more constrained by columns than rows, and we have comparatively few
rows of range check logic required for the Action circuit.
We now use a single LookupRangeCheckConfig for the entire circuit. The
reduction in lookup arguments and fixed columns cuts the proof size in
half (now at 6048 bytes when using `floor_planner::V1`).
Co-authored-by: therealyingtong <yingtong@z.cash>
- Move Poseidon into the right-hand advice columns. The Action circuit
has 33 Sinsemilla invocations with 510-bit inputs (the 32 Merkle path
hashes, and Commit^ivk). Poseidon fits within the row count of one of
these invocations, so we can run it in parallel with these.
- Share fixed columns between ECC and Poseidon chips. Poseidon requires
four advice columns, while ECC incomplete addition requires six, so we
could choose to configure them in parallel. However, we only use a
single Poseidon invocation, and we have the rows to accomodate it
serially with fixed-base scalar mul. Sharing the ECC chip's 8 Lagrange
coefficient fixed columns instead reduces the proof size.
- We position Poseidon in the right-most 6 fixed columns, anticipating
a further optimisation to Sinsemilla that will occupy the left-most
2 fixed columns.
therealyingtong
Jack Grigg
Deirdre Connolly
Deirdre Connolly
Sean Bowe
Daira Hopwood