Make AVX512IFMA opt-in backend (#695)

* Make AVX512IFMA opt-in backend * Updated README to have backend info * Added entry to changelog --------- Co-authored-by: Michael Rosenberg <[email protected]>
dalek-cryptography · Sep 8, 2024 · b636fb8 · b636fb8
1 parent 0964f80
commit b636fb8
Show file tree

Hide file tree

Showing 11 changed files with 87 additions and 45 deletions.
diff --git a/curve25519-dalek/CHANGELOG.md b/curve25519-dalek/CHANGELOG.md
@@ -3,6 +3,10 @@
 Entries are listed in reverse chronological order per undeprecated
 major series.
 
+## Unreleased
+
+* Move AVX-512 backend selection logic to a separate CFG flag that requires nightly
+
 ## 4.x series
 
 ### 4.1.3

diff --git a/curve25519-dalek/Cargo.toml b/curve25519-dalek/Cargo.toml
@@ -76,7 +76,7 @@ curve25519-dalek-derive = { version = "0.1", path = "../curve25519-dalek-derive"
 level = "warn"
 check-cfg = [
  'cfg(allow_unused_unsafe)',
- 'cfg(curve25519_dalek_backend, values("fiat", "serial", "simd"))',
+ 'cfg(curve25519_dalek_backend, values("fiat", "serial", "simd", "unstable_avx512"))',
  'cfg(curve25519_dalek_diagnostics, values("build"))',
  'cfg(curve25519_dalek_bits, values("32", "64"))',
  'cfg(nightly)',

diff --git a/curve25519-dalek/README.md b/curve25519-dalek/README.md
@@ -90,11 +90,12 @@ This release also does a lot of dependency updates and relaxations to unblock up
 
 Curve arithmetic is implemented and used by one of the following backends:
 
-| Backend | Selection | Implementation | Bits / Word sizes |
-| :--- | :--- | :--- | :--- |
-| `serial` | Automatic | An optimized, non-parllel implementation | `32` and `64` |
-| `fiat` | Manual | Formally verified field arithmetic from [fiat-crypto] | `32` and `64` |
-| `simd` | Automatic | Intel AVX2 / AVX512 IFMA accelerated backend | `64` only |
+| Backend | Selection | Implementation | Bits / Word sizes |
+| :--- | :--- | :--- | :--- |
+| `serial` | Automatic | An optimized, non-parllel implementation | `32` and `64` |
+| `fiat` | Manual | Formally verified field arithmetic from [fiat-crypto] | `32` and `64` |
+| `simd` | Automatic | Intel AVX2 accelerated backend | `64` only |
+| `unstable_avx512` | Manual | Intel AVX512 IFMA accelerated backend (requires nightly) | `64` only |
 
 At runtime, `curve25519-dalek` selects an arithmetic backend from the set of backends it was compiled to support. For Intel x86-64 targets, unless otherwise specified, it will build itself with `simd` support, and default to `serial` at runtime if the appropriate CPU features aren't detected. See [SIMD backend] for more details.
 
@@ -148,16 +149,16 @@ $ cargo build --target i686-unknown-linux-gnu
 
 ## SIMD backend
 
-The specific SIMD backend (AVX512 / AVX2 / `serial` default) is selected automatically at runtime, depending on the currently available CPU features, and whether Rust nightly is being used for compilation. The precise conditions are specified below.
+When the `simd` backend is selected, the AVX2 or `serial` implementation is selected automatically at runtime, depending on the currently available CPU features. Similarly, when the `unstable_avx512` backend is selected, the AVX512 implementation is selected automatically at runtime if available, or else selection falls through to the aforementioned `simd` backend logic.
 
 For a given CPU feature, you can also specify an appropriate `-C target_feature` to build a binary which assumes the required SIMD instructions are always available. Don't do this if you don't have a good reason.
 
 | Backend | `RUSTFLAGS` | Requires nightly? |
 | :--- | :--- | :--- |
-| avx2 | `-C target_feature=+avx2` | no |
-| avx512 | `-C target_feature=+avx512ifma,+avx512vl` | yes |
+| AVX2 | `-C target_feature=+avx2` | no |
+| AVX512 | `-C target_feature=+avx512ifma,+avx512vl` | yes |
 
-If compiled on a non-nightly compiler, `curve25519-dalek` will not include AVX512 code, and therefore will never select it at runtime.
+To reiterate, the `simd` backend will NOT use AVX512 code under any circumstance. The only way to enable AVX512 currently is to select the `unstable_avx512` backend and use a nightly compiler.
 
 # Documentation
 

diff --git a/curve25519-dalek/build.rs b/curve25519-dalek/build.rs
@@ -35,13 +35,16 @@ fn main() {
 
  println!("cargo:rustc-cfg=curve25519_dalek_bits=\"{curve25519_dalek_bits}\"");
 
- if rustc_version::version_meta()
+ let nightly = if rustc_version::version_meta()
  .expect("failed to detect rustc version")
  .channel
  == rustc_version::Channel::Nightly
  {
  println!("cargo:rustc-cfg=nightly");
- }
+ true
+ } else {
+ false
+ };
 
  let rustc_version = rustc_version::version().expect("failed to detect rustc version");
  if rustc_version.major == 1 && rustc_version.minor <= 64 {
@@ -51,25 +54,44 @@ fn main() {
  }
 
  // Backend overrides / defaults
- let curve25519_dalek_backend =
- match std::env::var("CARGO_CFG_CURVE25519_DALEK_BACKEND").as_deref() {
- Ok("fiat") => "fiat",
- Ok("serial") => "serial",
- Ok("simd") => {
- // simd can only be enabled on x86_64 & 64bit target_pointer_width
- match is_capable_simd(&target_arch, curve25519_dalek_bits) {
- true => "simd",
- // If override is not possible this must result to compile error
- // See: issues/532
- false => panic!("Could not override curve25519_dalek_backend to simd"),
+ let curve25519_dalek_backend = match std::env::var("CARGO_CFG_CURVE25519_DALEK_BACKEND")
+ .as_deref()
+ {
+ Ok("fiat") => "fiat",
+ Ok("serial") => "serial",
+ Ok("simd") => {
+ // simd can only be enabled on x86_64 & 64bit target_pointer_width
+ match is_capable_simd(&target_arch, curve25519_dalek_bits) {
+ true => "simd",
+ // If override is not possible this must result to compile error
+ // See: issues/532
+ false => panic!("Could not override curve25519_dalek_backend to simd"),
+ }
+ }
+ Ok("unstable_avx512") if nightly => {
+ // simd can only be enabled on x86_64 & 64bit target_pointer_width
+ match is_capable_simd(&target_arch, curve25519_dalek_bits) {
+ true => {
+ // In addition enable Avx2 fallback through simd stable backend
+ // NOTE: Compiler permits duplicate / multi value on the same key
+ println!("cargo:rustc-cfg=curve25519_dalek_backend=\"simd\"");
+
+ "unstable_avx512"
  }
+ // If override is not possible this must result to compile error
+ // See: issues/532
+ false => panic!("Could not override curve25519_dalek_backend to unstable_avx512"),
  }
- // default between serial / simd (if potentially capable)
- _ => match is_capable_simd(&target_arch, curve25519_dalek_bits) {
- true => "simd",
- false => "serial",
- },
- };
+ }
+ Ok("unstable_avx512") if !nightly => {
+ panic!("Could not override curve25519_dalek_backend to unstable_avx512, as this is nigthly only");
+ }
+ // default between serial / simd (if potentially capable)
+ _ => match is_capable_simd(&target_arch, curve25519_dalek_bits) {
+ true => "simd",
+ false => "serial",
+ },
+ };
  println!("cargo:rustc-cfg=curve25519_dalek_backend=\"{curve25519_dalek_backend}\"");
 }
 

diff --git a/curve25519-dalek/src/backend/mod.rs b/curve25519-dalek/src/backend/mod.rs
@@ -46,14 +46,14 @@ pub mod vector;
 enum BackendKind {
  #[cfg(curve25519_dalek_backend = "simd")]
  Avx2,
- #[cfg(all(curve25519_dalek_backend = "simd", nightly))]
+ #[cfg(all(curve25519_dalek_backend = "unstable_avx512", nightly))]
  Avx512,
  Serial,
 }
 
 #[inline]
 fn get_selected_backend() -> BackendKind {
- #[cfg(all(curve25519_dalek_backend = "simd", nightly))]
+ #[cfg(all(curve25519_dalek_backend = "unstable_avx512", nightly))]
  {
  cpufeatures::new!(cpuid_avx512, "avx512ifma", "avx512vl");
  let token_avx512: cpuid_avx512::InitToken = cpuid_avx512::init();
@@ -88,7 +88,7 @@ where
  #[cfg(curve25519_dalek_backend = "simd")]
  BackendKind::Avx2 =>
  vector::scalar_mul::pippenger::spec_avx2::Pippenger::optional_multiscalar_mul::<I, J>(scalars, points),
- #[cfg(all(curve25519_dalek_backend = "simd", nightly))]
+ #[cfg(all(curve25519_dalek_backend = "unstable_avx512", nightly))]
  BackendKind::Avx512 =>
  vector::scalar_mul::pippenger::spec_avx512ifma_avx512vl::Pippenger::optional_multiscalar_mul::<I, J>(scalars, points),
  BackendKind::Serial =>
@@ -100,7 +100,7 @@ where
 pub(crate) enum VartimePrecomputedStraus {
  #[cfg(curve25519_dalek_backend = "simd")]
  Avx2(vector::scalar_mul::precomputed_straus::spec_avx2::VartimePrecomputedStraus),
- #[cfg(all(curve25519_dalek_backend = "simd", nightly))]
+ #[cfg(all(curve25519_dalek_backend = "unstable_avx512", nightly))]
  Avx512ifma(
  vector::scalar_mul::precomputed_straus::spec_avx512ifma_avx512vl::VartimePrecomputedStraus,
  ),
@@ -120,7 +120,7 @@ impl VartimePrecomputedStraus {
  #[cfg(curve25519_dalek_backend = "simd")]
  BackendKind::Avx2 =>
  VartimePrecomputedStraus::Avx2(vector::scalar_mul::precomputed_straus::spec_avx2::VartimePrecomputedStraus::new(static_points)),
- #[cfg(all(curve25519_dalek_backend = "simd", nightly))]
+ #[cfg(all(curve25519_dalek_backend = "unstable_avx512", nightly))]
  BackendKind::Avx512 =>
  VartimePrecomputedStraus::Avx512ifma(vector::scalar_mul::precomputed_straus::spec_avx512ifma_avx512vl::VartimePrecomputedStraus::new(static_points)),
  BackendKind::Serial =>
@@ -150,7 +150,7 @@ impl VartimePrecomputedStraus {
  dynamic_scalars,
  dynamic_points,
  ),
- #[cfg(all(curve25519_dalek_backend = "simd", nightly))]
+ #[cfg(all(curve25519_dalek_backend = "unstable_avx512", nightly))]
  VartimePrecomputedStraus::Avx512ifma(inner) => inner.optional_mixed_multiscalar_mul(
  static_scalars,
  dynamic_scalars,
@@ -181,7 +181,7 @@ where
  BackendKind::Avx2 => {
  vector::scalar_mul::straus::spec_avx2::Straus::multiscalar_mul::<I, J>(scalars, points)
  }
- #[cfg(all(curve25519_dalek_backend = "simd", nightly))]
+ #[cfg(all(curve25519_dalek_backend = "unstable_avx512", nightly))]
  BackendKind::Avx512 => {
  vector::scalar_mul::straus::spec_avx512ifma_avx512vl::Straus::multiscalar_mul::<I, J>(
  scalars, points,
@@ -210,7 +210,7 @@ where
  scalars, points,
  )
  }
- #[cfg(all(curve25519_dalek_backend = "simd", nightly))]
+ #[cfg(all(curve25519_dalek_backend = "unstable_avx512", nightly))]
  BackendKind::Avx512 => {
  vector::scalar_mul::straus::spec_avx512ifma_avx512vl::Straus::optional_multiscalar_mul::<
  I,
@@ -228,7 +228,7 @@ pub fn variable_base_mul(point: &EdwardsPoint, scalar: &Scalar) -> EdwardsPoint
  match get_selected_backend() {
  #[cfg(curve25519_dalek_backend = "simd")]
  BackendKind::Avx2 => vector::scalar_mul::variable_base::spec_avx2::mul(point, scalar),
- #[cfg(all(curve25519_dalek_backend = "simd", nightly))]
+ #[cfg(all(curve25519_dalek_backend = "unstable_avx512", nightly))]
  BackendKind::Avx512 => {
  vector::scalar_mul::variable_base::spec_avx512ifma_avx512vl::mul(point, scalar)
  }
@@ -242,7 +242,7 @@ pub fn vartime_double_base_mul(a: &Scalar, A: &EdwardsPoint, b: &Scalar) -> Edwa
  match get_selected_backend() {
  #[cfg(curve25519_dalek_backend = "simd")]
  BackendKind::Avx2 => vector::scalar_mul::vartime_double_base::spec_avx2::mul(a, A, b),
- #[cfg(all(curve25519_dalek_backend = "simd", nightly))]
+ #[cfg(all(curve25519_dalek_backend = "unstable_avx512", nightly))]
  BackendKind::Avx512 => {
  vector::scalar_mul::vartime_double_base::spec_avx512ifma_avx512vl::mul(a, A, b)
  }

diff --git a/curve25519-dalek/src/backend/vector/mod.rs b/curve25519-dalek/src/backend/vector/mod.rs
@@ -16,7 +16,7 @@ pub mod packed_simd;
 
 pub mod avx2;
 
-#[cfg(nightly)]
+#[cfg(all(curve25519_dalek_backend = "unstable_avx512", nightly))]
 pub mod ifma;
 
 pub mod scalar_mul;
diff --git a/curve25519-dalek/src/backend/vector/scalar_mul/pippenger.rs b/curve25519-dalek/src/backend/vector/scalar_mul/pippenger.rs
@@ -11,7 +11,10 @@
 
 #[curve25519_dalek_derive::unsafe_target_feature_specialize(
  "avx2",
- conditional("avx512ifma,avx512vl", nightly)
+ conditional(
+ "avx512ifma,avx512vl",
+ all(curve25519_dalek_backend = "unstable_avx512", nightly)
+ )
 )]
 pub mod spec {
 

diff --git a/curve25519-dalek/src/backend/vector/scalar_mul/precomputed_straus.rs b/curve25519-dalek/src/backend/vector/scalar_mul/precomputed_straus.rs
@@ -13,7 +13,10 @@
 
 #[curve25519_dalek_derive::unsafe_target_feature_specialize(
  "avx2",
- conditional("avx512ifma,avx512vl", nightly)
+ conditional(
+ "avx512ifma,avx512vl",
+ all(curve25519_dalek_backend = "unstable_avx512", nightly)
+ )
 )]
 pub mod spec {
 

diff --git a/curve25519-dalek/src/backend/vector/scalar_mul/straus.rs b/curve25519-dalek/src/backend/vector/scalar_mul/straus.rs
@@ -13,7 +13,10 @@
 
 #[curve25519_dalek_derive::unsafe_target_feature_specialize(
  "avx2",
- conditional("avx512ifma,avx512vl", nightly)
+ conditional(
+ "avx512ifma,avx512vl",
+ all(curve25519_dalek_backend = "unstable_avx512", nightly)
+ )
 )]
 pub mod spec {
 

diff --git a/curve25519-dalek/src/backend/vector/scalar_mul/variable_base.rs b/curve25519-dalek/src/backend/vector/scalar_mul/variable_base.rs
@@ -2,7 +2,10 @@
 
 #[curve25519_dalek_derive::unsafe_target_feature_specialize(
  "avx2",
- conditional("avx512ifma,avx512vl", nightly)
+ conditional(
+ "avx512ifma,avx512vl",
+ all(curve25519_dalek_backend = "unstable_avx512", nightly)
+ )
 )]
 pub mod spec {
 

diff --git a/curve25519-dalek/src/backend/vector/scalar_mul/vartime_double_base.rs b/curve25519-dalek/src/backend/vector/scalar_mul/vartime_double_base.rs
@@ -13,7 +13,10 @@
 
 #[curve25519_dalek_derive::unsafe_target_feature_specialize(
  "avx2",
- conditional("avx512ifma,avx512vl", nightly)
+ conditional(
+ "avx512ifma,avx512vl",
+ all(curve25519_dalek_backend = "unstable_avx512", nightly)
+ )
 )]
 pub mod spec {