Improve native array kernels and separate build layers

.config/nextest.toml

@@ -0,0 +1,19 @@
+# `cargo nextest run` runs each test in its own process and every binary at once.
+[profile.default]
+# A test running past 5 s is reported as slow. One still running at 30 s is stopped and fails.
+slow-timeout = { period = "5s", terminate-after = 6 }
+fail-fast = false
+
+# Timings run last, with nothing beside them: threaded correctness tests such as `linalg` would slow them.
+[[profile.default.overrides]]
+filter = 'binary(speed)'
+threads-required = 'num-cpus'
+priority = -100
+
+# Optional selection excluding timings, used by focused cross-repo diagnostics beside basedpl's untimed selection.
+[profile.untimed]
+default-filter = 'not binary(speed)'
+
+# Optional timing-only selection. An explicit cross-repo ('untimed', 'timed', 'speed') run uses this after basedpl timings. Plain run_tests uses default and includes everything.
+[profile.speed]
+default-filter = 'binary(speed)'

.github/workflows/ci.yml

@@ -49,10 +49,7 @@ jobs:
       - uses: dtolnay/rust-toolchain@stable
       - id: auth
         uses: rust-lang/crates-io-auth-action@v1
-      - run: cargo publish -p rustymath --no-verify
-        env:
-          CARGO_REGISTRY_TOKEN: ${{ steps.auth.outputs.token }}
-      - run: cargo publish -p rustymath-buffer --no-verify
+      - run: cargo publish --workspace --no-verify
         env:
           CARGO_REGISTRY_TOKEN: ${{ steps.auth.outputs.token }}
       - uses: actions/download-artifact@v8

.github/workflows/test.yml

@@ -15,6 +15,7 @@ jobs:
     steps:
       - uses: actions/checkout@v7
       - uses: dtolnay/rust-toolchain@stable
+      - uses: taiki-e/install-action@nextest
       # The repository has no `Cargo.lock`. Creating one first makes rust-cache's key change when a dependency publishes a new version.
       - run: cargo generate-lockfile
       - uses: Swatinem/rust-cache@v2
@@ -25,8 +26,8 @@ jobs:
           python-version: '3.12'
       - run: pip install -e '.[dev]'
       - run: cargo develop
-      - run: cargo test --release
-      - run: pytest -q
+      - run: cargo nextest run --cargo-profile release -E 'not binary(speed)'
+      - run: pytest -q --ignore=tests/test_speed.py
         if: ${{ !cancelled() }}
       - run: nbdev-test
         if: ${{ !cancelled() }}

Cargo.toml

@@ -1,16 +1,32 @@
 [workspace]
-members = ["buffer", "py"]
+members = ["core", "kernels", "timing", "buffer", "py"]
 default-members = [".", "buffer", "py"]
 
 [workspace.package]
-version = "0.1.56"
+version = "0.1.57"
 edition = "2024"
 
 [workspace.dependencies]
-rustymath = { path = ".", version = ">=0.1.55, <1", default-features = false }
+rustymath = { path = ".", version = ">=0.1.56, <1", default-features = false }
 rustymath-buffer = { path = "buffer", version = ">=0.1.39, <1" }
 pyo3 = ">=0.29"
 pyo3-build-config = ">=0.29"
+rustymath-core = { path = "core", version = ">=0.1.56, <1", default-features = false }
+rustymath-kernels = { path = "kernels", version = ">=0.1.56, <1" }
+rustymath-timing = { path = "timing", version = ">=0.1.56, <1" }
+half = { version = ">=2.7.1, <3", features = ["num-traits"] }
+hashbrown = ">=0.15, <1"
+foldhash = ">=0.2.0, <1"
+num-bigint = ">=0.4.6, <0.5"
+num-complex = ">=0.4.6, <1"
+num-integer = ">=0.1.46, <1"
+num-rational = ">=0.4.2, <1"
+num-traits = ">=0.2.19, <1"
+rand = { version = ">=0.10.2, <1", features = ["chacha"] }
+strength_reduce = ">=0.2.4, <1"
+rayon = ">=1.12.0, <2"
+web-time = ">=1.1.0, <2"
+libc = { version = ">=0.2.190, <1", features = ["extra_traits"] }
 
 [workspace.metadata.fastfmt]
 join_comments = true
@@ -29,40 +45,28 @@ name = "rustymath"
 test = false
 doctest = false
 
-# The timing suite: plain `cargo test` skips it; `cargo test --test speed` runs it.
-[[test]]
-name = "speed"
-test = false
-
 [features]
 default = ["native"]
 # Compiles rustymath's own linear algebra into macOS builds too, beside Accelerate, so `RUSTYMATH_NATIVE=1` can select it at run time.
 # Other targets always compile it. The Mac release builds turn it off with `--no-default-features`.
-native = []
+native = ["rustymath-core/native"]
 
 [dependencies]
-half = { version = ">=2.7.1, <3", features = ["num-traits"] }
-hashbrown = ">=0.15, <1"
-foldhash = ">=0.2.0, <1"
-num-bigint = ">=0.4.6, <0.5"
-num-complex = ">=0.4.6, <1"
-num-integer = ">=0.1.46, <1"
-num-rational = ">=0.4.2, <1"
-num-traits = ">=0.2.19, <1"
-rand = { version = ">=0.10.2, <1", default-features = false, features = ["alloc", "chacha"] }
-strength_reduce = ">=0.2.4, <1"
-
-# The browser build has no threads. There `parallel` runs its pieces in turn on the calling thread.
-[target.'cfg(not(target_arch = "wasm32"))'.dependencies]
-rayon = ">=1.12.0, <2"
+rustymath-core.workspace = true
+rustymath-kernels.workspace = true
+num-complex.workspace = true
+num-traits.workspace = true
 
-# The browser's clock, for the timing loop in `timing`, as basedpl's browser build uses it.
-[target.'cfg(all(target_arch = "wasm32", target_os = "unknown"))'.dependencies]
-web-time = ">=1.1.0, <2"
-
-[target.'cfg(unix)'.dependencies]
-# The memory guard calls `libc` to read physical memory and, on Linux, to ask for huge pages. On macOS, the timing lock also makes its QoS call through it, to favour the performance cores for a timing.
-libc = ">=0.2.190, <1"
+[dev-dependencies]
+rustymath-timing.workspace = true
+half.workspace = true
+hashbrown.workspace = true
+foldhash.workspace = true
+num-bigint.workspace = true
+num-integer.workspace = true
+num-rational.workspace = true
+rand.workspace = true
+strength_reduce.workspace = true
 
 [profile.test]
 inherits = "release"
@@ -77,6 +81,10 @@ incremental = true
 # Line tables change with every shifted line and recompile the code below it.
 debug = false
 
+[profile.quick]
+inherits = "release"
+lto = "off"
+
 [profile.dist]
 inherits = "release"
 lto = "thin"