feat: implement 128bit fft

Cargo.toml

@@ -13,13 +13,16 @@ keywords = ["fft"]
 [dependencies]
 num-complex = "0.4"
 dyn-stack = { version = "0.8", default-features = false }
+pulp = "0.10"
+bytemuck = "1.13"
 aligned-vec = { version = "0.5", default-features = false }
 serde = { version = "1.0", optional = true, default-features = false }
 
 [features]
-default = ["std"]
-nightly = []
+default = ["std", "fft128"]
+nightly = ["pulp/nightly", "bytemuck/nightly_stdsimd"]
 std = []
+fft128 = []
 serde = ["dep:serde", "num-complex/serde"]
 
 [dev-dependencies]
@@ -28,9 +31,13 @@ rustfft = "6.0"
 fftw-sys = { version = "0.6", default-features = false, features = ["system"] }
 rand = "0.8"
 bincode = "1.3"
+more-asserts = "0.3.1"
+
+[target.'cfg(target_os = "linux")'.dev-dependencies]
+rug = "1.19.0"
 
 [[bench]]
-name = "fft"
+name = "bench"
 harness = false
 
 [package.metadata.docs.rs]

README.md

@@ -3,33 +3,38 @@ that processes vectors of sizes that are powers of two. It was made to be used
 as a backend in Zama's `concrete` library.
 
 This library provides two FFT modules:
- - The ordered module FFT applies a forward/inverse FFT that takes its input in standard
- order, and outputs the result in standard order. For more detail on what the FFT
- computes, check the ordered module-level documentation.
- - The unordered module FFT applies a forward FFT that takes its input in standard order,
- and outputs the result in a certain permuted order that may depend on the FFT plan. On the
- other hand, the inverse FFT takes its input in that same permuted order and outputs its result
- in standard order. This is useful for cases where the order of the coefficients in the
- Fourier domain is not important. An example is using the Fourier transform for vector
- convolution. The only operations that are performed in the Fourier domain are elementwise, and
- so the order of the coefficients does not affect the results.
+
+- The ordered module FFT applies a forward/inverse FFT that takes its input in standard
+  order, and outputs the result in standard order. For more detail on what the FFT
+  computes, check the ordered module-level documentation.
+- The unordered module FFT applies a forward FFT that takes its input in standard order,
+  and outputs the result in a certain permuted order that may depend on the FFT plan. On the
+  other hand, the inverse FFT takes its input in that same permuted order and outputs its result
+  in standard order. This is useful for cases where the order of the coefficients in the
+  Fourier domain is not important. An example is using the Fourier transform for vector
+  convolution. The only operations that are performed in the Fourier domain are elementwise, and
+  so the order of the coefficients does not affect the results.
+
+Additionally, an optional 128-bit negacyclic FFT module is provided.
 
 ## Features
 
- - `std` (default): This enables runtime arch detection for accelerated SIMD
-   instructions, and an FFT plan that measures the various implementations to
-   choose the fastest one at runtime.
- - `nightly`: This enables unstable Rust features to further speed up the FFT,
-   by enabling AVX512F instructions on CPUs that support them. This feature
-   requires a nightly Rust
- toolchain.
- - `serde`: This enables serialization and deserialization functions for the
-   unordered plan. These allow for data in the Fourier domain to be serialized
-   from the permuted order to the standard order, and deserialized from the
-   standard order to the permuted order. This is needed since the inverse
-   transform must be used with the same plan that computed/deserialized the
-   forward transform (or more specifically, a plan with the same internal base
-   FFT size).
+- `std` (default): This enables runtime arch detection for accelerated SIMD
+  instructions, and an FFT plan that measures the various implementations to
+  choose the fastest one at runtime.
+- `fft128` (default): This flag provides access to the 128-bit FFT, which is accessible in the
+  `concrete_fft::fft128` module.
+- `nightly`: This enables unstable Rust features to further speed up the FFT,
+  by enabling AVX512F instructions on CPUs that support them. This feature
+  requires a nightly Rust
+  toolchain.
+- `serde`: This enables serialization and deserialization functions for the
+  unordered plan. These allow for data in the Fourier domain to be serialized
+  from the permuted order to the standard order, and deserialized from the
+  standard order to the permuted order. This is needed since the inverse
+  transform must be used with the same plan that computed/deserialized the
+  forward transform (or more specifically, a plan with the same internal base
+  FFT size).
 
 ## Example
 
@@ -65,8 +70,8 @@ for (actual, expected) in transformed_inv.iter().map(|z| z / N as f64).zip(data)
 
 ## Links
 
- - [Zama](https://www.zama.ai/)
- - [Concrete](https://github.com/zama-ai/concrete)
+- [Zama](https://www.zama.ai/)
+- [Concrete](https://github.com/zama-ai/concrete)
 
 ## License