SIMD test in dedicated workspace

Cargo.toml

@@ -1,18 +1,21 @@
+[workspace]
+members = ["simd_test"]
+
 [package]
 name = "breakwater"
 version = "0.1.0"
 edition = "2021"
 
 [dependencies]
 chrono = "0.4"
-const_format = "0.2"
 clap = { version = "4.3", features = ["derive"] }
-rusttype = "0.9"
-number_prefix = "0.4"
+const_format = "0.2"
 env_logger = "0.10"
 lazy_static = "1.4"
 log = "0.4"
+number_prefix = "0.4"
 prometheus_exporter = "0.8"
+rusttype = "0.9"
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1.0"
 simple_moving_average = "0.1"
@@ -48,6 +51,6 @@ opt-level = 3
 
 [profile.release]
 opt-level = 3
-lto = "fat"
-codegen-units = 1
+# lto = "fat"
+# codegen-units = 1
 # panic = "abort" # You can enable this, but I prefer to get actual stack traces

simd_test/Cargo.toml

@@ -0,0 +1,37 @@
+[package]
+name = "simd_test"
+version = "0.1.0"
+edition = "2021"
+
+[dependencies]
+breakwater = { path = "../" }
+
+# chrono = "0.4"
+# clap = { version = "4.3", features = ["derive"] }
+# const_format = "0.2"
+# env_logger = "0.10"
+# lazy_static = "1.4"
+# log = "0.4"
+# number_prefix = "0.4"
+# prometheus_exporter = "0.8"
+# rusttype = "0.9"
+# serde = { version = "1.0", features = ["derive"] }
+# serde_json = "1.0"
+# simple_moving_average = "0.1"
+# thread-priority = "0.13"
+tokio = { version = "1.29", features = ["fs", "rt-multi-thread", "net", "io-util", "macros", "process", "signal", "sync", "time"] }
+# vncserver = { version ="0.2", optional = true}
+
+[dev-dependencies]
+criterion = {version = "0.5", features = ["async_tokio"]}
+pixelbomber = "0.4.1"
+# rstest = "0.18"
+# rand = "0.8"
+
+[lib]
+name = "simd_test"
+path = "src/lib.rs"
+
+[[bench]]
+name = "benchmarks"
+harness = false

simd_test/benches/benchmarks.rs

@@ -0,0 +1,162 @@
+#![feature(portable_simd)]
+
+use breakwater::{framebuffer::FrameBuffer, parser::ParserState, test::helpers::DevNullTcpStream};
+use criterion::{
+    BenchmarkId, Criterion, {criterion_group, criterion_main},
+};
+use pixelbomber::image_handler;
+use pixelbomber::image_handler::ImageConfig;
+use std::{sync::Arc, time::Duration};
+
+const FRAMEBUFFER_WIDTH: usize = 1920;
+const FRAMEBUFFER_HEIGHT: usize = 1080;
+
+async fn invoke_parse_pixelflut_commands_base(
+    input: &[u8],
+    fb: &Arc<FrameBuffer>,
+    parser_state: ParserState,
+) {
+    let mut stream = DevNullTcpStream::default();
+    breakwater::parser::parse_pixelflut_commands(input, fb, &mut stream, parser_state).await;
+}
+
+async fn invoke_parse_pixelflut_commands_simd(
+    input: &[u8],
+    fb: &Arc<FrameBuffer>,
+    parser_state: ParserState,
+) {
+    let mut stream = DevNullTcpStream::default();
+    simd_test::parser::parse_pixelflut_commands(input, fb, &mut stream, parser_state).await;
+}
+
+fn benchmark_base(c: &mut Criterion, name: &str, file: &str, config: ImageConfig) {
+    let mut commands = image_handler::load(vec![file], &config);
+    let command = commands.pop().unwrap();
+    c.bench_with_input(
+        BenchmarkId::new(name, format!("{FRAMEBUFFER_WIDTH} x {FRAMEBUFFER_HEIGHT}")),
+        &command,
+        |b, input| {
+            let fb = Arc::new(FrameBuffer::new(FRAMEBUFFER_WIDTH, FRAMEBUFFER_HEIGHT));
+            let parser_state = ParserState::default();
+            b.to_async(tokio::runtime::Runtime::new().unwrap())
+                .iter(|| invoke_parse_pixelflut_commands_base(input, &fb, parser_state.clone()));
+        },
+    );
+}
+
+fn benchmark_simd(c: &mut Criterion, name: &str, file: &str, config: ImageConfig) {
+    let mut commands = image_handler::load(vec![file], &config);
+    let command = commands.pop().unwrap();
+    c.bench_with_input(
+        BenchmarkId::new(name, format!("{FRAMEBUFFER_WIDTH} x {FRAMEBUFFER_HEIGHT}")),
+        &command,
+        |b, input| {
+            let fb = Arc::new(FrameBuffer::new(FRAMEBUFFER_WIDTH, FRAMEBUFFER_HEIGHT));
+            let parser_state = ParserState::default();
+            b.to_async(tokio::runtime::Runtime::new().unwrap())
+                .iter(|| invoke_parse_pixelflut_commands_simd(input, &fb, parser_state.clone()));
+        },
+    );
+}
+
+fn add_benches(c: &mut Criterion) {
+    // benchmark_base(
+    //     c,
+    //     "parse_draw_commands_ordered",
+    //     "benches/non-transparent.png",
+    //     image_handler::ImageConfigBuilder::new()
+    //         .width(FRAMEBUFFER_WIDTH as u32)
+    //         .height(FRAMEBUFFER_HEIGHT as u32)
+    //         .shuffle(false)
+    //         .build(),
+    // );
+
+    benchmark_base(
+        c,
+        "base_parse_draw_commands_shuffled",
+        "benches/non-transparent.png",
+        image_handler::ImageConfigBuilder::new()
+            .width(FRAMEBUFFER_WIDTH as u32)
+            .height(FRAMEBUFFER_HEIGHT as u32)
+            .shuffle(true)
+            .build(),
+    );
+
+    // benchmark_base(
+    //     c,
+    //     "parse_mixed_draw_commands",
+    //     "benches/mixed.png",
+    //     image_handler::ImageConfigBuilder::new()
+    //         .width(FRAMEBUFFER_WIDTH as u32)
+    //         .height(FRAMEBUFFER_HEIGHT as u32)
+    //         .shuffle(false)
+    //         .gray_usage(true)
+    //         .build(),
+    // );
+
+    // benchmark_base(
+    //     c,
+    //     "parse_draw_commands_with_offset",
+    //     "benches/non-transparent.png",
+    //     image_handler::ImageConfigBuilder::new()
+    //         .width(FRAMEBUFFER_WIDTH as u32)
+    //         .height(FRAMEBUFFER_HEIGHT as u32)
+    //         .shuffle(false)
+    //         .offset_usage(true)
+    //         .build(),
+    // );
+
+    // benchmark_simd(
+    //     c,
+    //     "parse_draw_commands_ordered",
+    //     "benches/non-transparent.png",
+    //     image_handler::ImageConfigBuilder::new()
+    //         .width(FRAMEBUFFER_WIDTH as u32)
+    //         .height(FRAMEBUFFER_HEIGHT as u32)
+    //         .shuffle(false)
+    //         .build(),
+    // );
+
+    benchmark_simd(
+        c,
+        "simd_parse_draw_commands_shuffled",
+        "benches/non-transparent.png",
+        image_handler::ImageConfigBuilder::new()
+            .width(FRAMEBUFFER_WIDTH as u32)
+            .height(FRAMEBUFFER_HEIGHT as u32)
+            .shuffle(true)
+            .build(),
+    );
+
+    // benchmark_simd(
+    //     c,
+    //     "parse_mixed_draw_commands",
+    //     "benches/mixed.png",
+    //     image_handler::ImageConfigBuilder::new()
+    //         .width(FRAMEBUFFER_WIDTH as u32)
+    //         .height(FRAMEBUFFER_HEIGHT as u32)
+    //         .shuffle(false)
+    //         .gray_usage(true)
+    //         .build(),
+    // );
+
+    // benchmark_simd(
+    //     c,
+    //     "parse_draw_commands_with_offset",
+    //     "benches/non-transparent.png",
+    //     image_handler::ImageConfigBuilder::new()
+    //         .width(FRAMEBUFFER_WIDTH as u32)
+    //         .height(FRAMEBUFFER_HEIGHT as u32)
+    //         .shuffle(false)
+    //         .offset_usage(true)
+    //         .build(),
+    // );
+}
+
+criterion_group!(
+    name = benches;
+    config = Criterion::default().warm_up_time(Duration::from_secs(2)).measurement_time(Duration::from_secs(3));
+    targets = add_benches,
+);
+
+criterion_main!(benches);

simd_test/src/lib.rs

@@ -0,0 +1,3 @@
+#![feature(portable_simd)]
+
+pub mod parser;