mirror of
https://github.com/luxfhe/torus.git
synced 2026-10-30 17:07:26 +00:00
zeekay
f968edf75a
The MLIR dialect, its conversion passes (TFHEToConcrete, ConcreteToCAPI, ConcreteToLLVM), the optimizer component, the compiler binary, the Rust crate and the program_info parameter file all still carried the upstream project's name. They now carry this one's. Two things deliberately keep the word, because in them it is not the name. mlir::OpTrait CRTP parameters mean "the derived type", so they are spelled Derived and OpTy — LLVM's other conventional spelling for the same idiom. concreteShape is a capnp field: renaming it would change the serialized protocol, and the prose describing it is ordinary English. This also repairs references the earlier partial rename left dangling: Security.cpp included concrete/curves.gen.h and the directory is torus/, torus-cpu.h included concrete-num_complex.h, the Torus dialect CMakeLists named ConcreteOps.td, and compiler/setup.py asked find_namespace_packages for a "concrete" package that no longer exists — so it built without the package it ships.
2.9 KiB
2.9 KiB
module torus.fhe.mlir.processors.process_rounding
Declaration of ProcessRounding graph processor.
class ProcessRounding
ProcessRounding graph processor, to analyze rounding and support regular operations on it.
method __init__
__init__(rounding_exactness: Exactness)
method apply
apply(graph: Graph)
method process_predecessors
process_predecessors(graph: Graph, node: Node)
Process predecessors of the rounding.
method process_successors
process_successors(graph: Graph, node: Node)
Process successors of the rounding.
method replace_with_tlu
replace_with_tlu(graph: Graph, node: Node)
Replace rounding node with a TLU node that simulates rounding.
This is a special case where:
- user wanted to remove 1-bits
- there was an overflow
- overflow protection was on
Let's say user wanted to remove 1-bit from 3-bits, but there was an overflow:
- [0, 1, 2, 3, 4, 5, 6, 7] would be mapped to [0, 2, 2, 4, 4, 6, 6, 8]
- or in the actual implementation [(0)00, (0)01, (0)01, (0)10, (0)10, (0)11, (0)11, (1)00]
- (first bit is the padding bit, which is overwritten on overflows)
- so the input is 3-bits and the output needs to be 3-bits to store the result
- which can't be achieved with rounding
- so we just replace the rounding with a TLU
- using the table [0, 2, 2, 4, 4, 6, 6, 8]