WIP: CFAVML-NONE: Add ndarray extension library

Cargo.toml

@@ -3,7 +3,11 @@ resolver = "2"
 members = [
     "cfavml",
     "cfavml-gemm",
+    "cfavml-ndarray",
     "cfavml-utils",
     # Testing and profiling
     "cfavml/asm-view"
 ]
+
+[profile.test]
+overflow-checks = false

cfavml-gemm/Cargo.toml

@@ -7,7 +7,7 @@ description = "BLAS-like general matrix multiplication extension for `cfavml`."
 [dependencies]
 num_cpus = "1.16.0"
 
-cfavml = { version = "0.3", path = "../cfavml" }
+cfavml = { version = "0.4", path = "../cfavml" }
 cfavml-utils = { version = "0.1", path = "../cfavml-utils" }
 
 [dev-dependencies]

cfavml-ndarray/Cargo.toml

@@ -0,0 +1,32 @@
+[package]
+name = "cfavml-ndarray"
+version = "0.1.0"
+edition = "2021"
+rust-version = "1.75"
+authors = ["Harrison Burt <[email protected]>"]
+description = "Accelerate your ndarray operations with CFAVML's SIMD acceleration."
+keywords = ["linear-algebra", "vector", "simd"]
+categories = ["concurrency"]
+readme = "README.md"
+repository = "https://github.com/ChillFish8/cfavml"
+license = "MIT OR Apache-2.0"
+
+[dependencies]
+ndarray = "0.16"
+
+cfavml = { version = "0.4.0", path = "../cfavml" }
+
+[features]
+# Enables cfavml's nightly only optimizations, which includes AVX512 support.
+nightly = ["cfavml/nightly"]
+
+[dev-dependencies]
+ndarray-rand = "0.15.0"
+paste = "1.0.15"
+divan = "0.1.14"
+
+mimalloc = { version = "0.1.43", default-features = false }
+
+[[bench]]
+name = "bench_arithmetic_ops"
+harness = false

cfavml-ndarray/README.md

@@ -0,0 +1,11 @@
+# cfavml-ndarray
+
+Accelerate your ndarray operations using CFAVML's SIMD routines.
+
+This library acts as an extension system to your existing ndarray workloads, allowing 
+you to incrementally switch over to the CFAVML optimized routines brining SIMD acceleration
+across all primitive integer and float types across x86 and ARM hardware.
+
+### NOTE
+
+Currently, this library is a WIP and requires alloc and std.

cfavml-ndarray/benches/bench_arithmetic_ops.rs

@@ -0,0 +1,174 @@
+use std::hint::black_box;
+
+use divan::Bencher;
+use divan::counter::ItemsCount;
+use ndarray::Array3;
+use ndarray_rand::rand_distr::Uniform;
+use ndarray_rand::RandomExt;
+
+#[global_allocator]
+static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc;
+
+fn main() {
+    divan::main();
+}
+
+
+#[divan::bench(
+    sample_count = 32, 
+    sample_size = 25,
+    counters = [ItemsCount::new(2usize * 10 * 256 * 256)],
+)]
+fn bench_cfavml_two_array_no_broadcast_no_time_alloc(bencher: Bencher) {
+    let a = Array3::random((10, 256, 256), Uniform::new(1.0, 10.0));
+    let b = Array3::random((10, 256, 256), Uniform::new(1.0, 10.0));
+
+    bencher
+        .with_inputs(|| {
+            a.clone()
+        })
+        .bench_local_values(|a| {
+             cfavml_ndarray::ops::mul(a, black_box(&b))
+        });
+}
+
+#[divan::bench(
+    sample_count = 32, 
+    sample_size = 25,
+    counters = [ItemsCount::new(2usize * 10 * 256 * 256)],
+)]
+fn bench_cfavml_two_array_no_broadcast_time_alloc(bencher: Bencher) {
+    let a = Array3::random((10, 256, 256), Uniform::new(1.0, 10.0));
+    let b = Array3::random((10, 256, 256), Uniform::new(1.0, 10.0));
+
+    bencher.bench_local(|| {
+        let a = black_box(&a).clone();
+        cfavml_ndarray::ops::mul(a, black_box(&b))
+    });
+}
+
+#[divan::bench(
+    sample_count = 32, 
+    sample_size = 25,
+    counters = [ItemsCount::new(2usize * 10 * 256 * 256)],
+)]
+fn bench_cfavml_one_array_broadcast_value_no_time_alloc(bencher: Bencher) {
+    let a = Array3::random((10, 256, 256), Uniform::new(1.0, 10.0));
+
+    bencher
+        .with_inputs(|| {
+            a.clone()
+        })
+        .bench_local_values(|a| {
+            cfavml_ndarray::ops::mul(a, black_box(3.0))
+        });
+}
+
+#[divan::bench(
+    sample_count = 32, 
+    sample_size = 25,
+    counters = [ItemsCount::new(2usize * 10 * 256 * 256)],
+)]
+fn bench_cfavml_one_array_broadcast_value_time_alloc(bencher: Bencher) {
+    let a = Array3::random((10, 256, 256), Uniform::new(1.0, 10.0));
+
+    bencher.bench_local(|| {
+        let a = black_box(&a).clone();
+        cfavml_ndarray::ops::mul(a, black_box(3.0))
+    });
+}
+
+#[divan::bench(
+    sample_count = 32, 
+    sample_size = 25,
+    counters = [ItemsCount::new(2usize * 10 * 256 * 256)],
+)]
+fn bench_default_two_array_no_broadcast_no_time_alloc(bencher: Bencher) {
+    let a = Array3::random((10, 256, 256), Uniform::new(1.0, 10.0));
+    let b = Array3::random((10, 256, 256), Uniform::new(1.0, 10.0));
+
+    bencher
+        .with_inputs(|| {
+            a.clone()
+        })
+        .bench_local_values(|a| {
+            a * black_box(&b)
+        });
+}
+
+#[divan::bench(
+    sample_count = 32, 
+    sample_size = 25,
+    counters = [ItemsCount::new(2usize * 10 * 256 * 256)],
+)]
+fn bench_default_two_array_no_broadcast_time_alloc(bencher: Bencher) {
+    let a = Array3::random((10, 256, 256), Uniform::new(1.0, 10.0));
+    let b = Array3::random((10, 256, 256), Uniform::new(1.0, 10.0));
+
+    bencher.bench_local(|| {
+        let a = black_box(&a).clone();
+        a * black_box(&b)
+    });
+}
+
+#[divan::bench(
+    sample_count = 32, 
+    sample_size = 25,
+    counters = [ItemsCount::new(2usize * 10 * 256 * 256)],
+)]
+fn bench_default_two_array_no_broadcast_two_views(bencher: Bencher) {
+    let a = Array3::random((10, 256, 256), Uniform::new(1.0, 10.0));
+    let b = Array3::random((10, 256, 256), Uniform::new(1.0, 10.0));
+
+    bencher.bench_local(|| {
+        let a = black_box(&a);
+        let b = black_box(&b);
+        a * b
+    });
+}
+
+#[divan::bench(
+    sample_count = 32, 
+    sample_size = 25,
+    counters = [ItemsCount::new(2usize * 10 * 256 * 256)],
+)]
+fn bench_default_one_array_broadcast_value_no_time_alloc(bencher: Bencher) {
+    let a = Array3::random((10, 256, 256), Uniform::new(1.0, 10.0));
+
+    bencher
+        .with_inputs(|| {
+            a.clone()
+        })
+        .bench_local_values(|a| {
+            a * black_box(3.0)
+        });
+}
+
+#[divan::bench(
+    sample_count = 32, 
+    sample_size = 25,
+    counters = [ItemsCount::new(2usize * 10 * 256 * 256)],
+)]
+fn bench_default_one_array_broadcast_value_time_alloc(bencher: Bencher) {
+    let a = Array3::random((10, 256, 256), Uniform::new(1.0, 10.0));
+
+    bencher.bench_local(|| {
+        let a = black_box(&a).clone();
+        a * black_box(3.0)
+    });
+}
+
+#[divan::bench(
+    sample_count = 32, 
+    sample_size = 25,
+    counters = [ItemsCount::new(2usize * 10 * 256 * 256)],
+)]
+fn bench_default_one_array_broadcast_value_with_view(bencher: Bencher) {
+    let a = Array3::random((10, 256, 256), Uniform::new(1.0, 10.0));
+
+    bencher.bench_local(|| {
+        let a = black_box(&a);
+        let b = black_box(3.0);
+        a * b
+    });
+}