Mimalloc in every bench

Cargo.toml

@@ -44,6 +44,11 @@ keywords = ["computer-vision", "deep-learning", "image-processing", "onnx", "vid
 [workspace.dependencies]
 thiserror = "2"
 rustc-hash = "2.1.3"
+# Every benchmark and latency harness in the workspace installs this as its
+# global allocator, so measurements are allocator-comparable across targets.
+# Library crates deliberately do not — a library must leave that choice to the
+# binary at the top of the dependency graph.
+mimalloc = { version = "0.1", default-features = false }
 
 [workspace.lints.rust]
 unsafe_code = "deny"

README.md

@@ -214,10 +214,20 @@ All hot paths have hand-tuned SIMD for three architectures with runtime CPU dete
 | **Threading** | GCD dispatch_apply (~0.3µs) | std::thread::scope (~1µs) | std::thread::scope (~1µs) |
 | **GPU inference** | MPSGraph | wgpu/Vulkan | wgpu/Vulkan |
 | **Softmax** | Fused NEON | Fused AVX/SSE | Fused NEON |
-| **Allocator** | mimalloc | mimalloc | mimalloc |
+| **Bench-harness allocator** | mimalloc | mimalloc | mimalloc |
 
 SIMD dispatch is automatic at runtime — no need for `-C target-cpu` flags (though they help: `-C target-cpu=apple-m1` or `-C target-cpu=native` for best codegen). The framework detects CPU features once through `yscv-cpu` and routes kernels through cached `host_cpu().features` gates. `yscv_kernels::runtime_dispatch_report()` exposes the typed CPU/kernel selection snapshot, while `runtime_config_report()` records active `YSCV_*` A/B overrides for reproducible benchmark logs. 570 `#[target_feature]`-gated functions total, all with scalar fallback for WASM/RISC-V/Miri.
 
+The allocator row applies to the benchmark and latency harnesses, which all
+install mimalloc so their numbers stay comparable. The library crates set no
+global allocator — that is a whole-program choice, so it belongs to your
+binary. If you want mimalloc in your own build, add the two lines yourself:
+
+```rust
+#[global_allocator]
+static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc;
+```
+
 ### Recommended release profile
 
 For best performance, the workspace already includes an optimized release profile in `Cargo.toml`:

apps/bench/Cargo.toml

@@ -20,7 +20,7 @@ yscv-video = { path = "../../crates/yscv-video" }
 yscv-detect = { path = "../../crates/yscv-detect" }
 yscv-track = { path = "../../crates/yscv-track" }
 yscv-optim = { path = "../../crates/yscv-optim" }
-mimalloc = { version = "0.1", default-features = false }
+mimalloc.workspace = true
 
 [[bin]]
 name = "yscv-bench"

apps/bench/src/fpv_latency.rs

@@ -13,6 +13,9 @@
 //! can model a real board's timing distribution without requiring that
 //! HW to be physically attached.
 
+#[global_allocator]
+static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc;
+
 use std::sync::Arc;
 use std::sync::atomic::{AtomicUsize, Ordering};
 use std::time::{Duration, Instant};

apps/camera-face-tool/Cargo.toml

@@ -22,6 +22,7 @@ thiserror.workspace = true
 
 [dev-dependencies]
 criterion = { version = "0.5", default-features = false, features = ["cargo_bench_support"] }
+mimalloc.workspace = true
 
 [[bench]]
 name = "camera_face_runtime_ops"

apps/camera-face-tool/benches/camera_face_runtime_ops.rs

@@ -1,3 +1,6 @@
+#[global_allocator]
+static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc;
+
 use criterion::{Criterion, black_box, criterion_group, criterion_main};
 use serde_json::{Value, json};
 use yscv_detect::{

apps/llm-bench/Cargo.toml

@@ -47,6 +47,7 @@ serde = { version = "1.0", features = ["derive"] }
 serde_json = "1.0"
 thiserror.workspace = true
 rustc-hash.workspace = true
+mimalloc.workspace = true
 
 [lints]
 workspace = true

apps/llm-bench/src/bin/bench_tracker.rs

@@ -1,5 +1,8 @@
 //! Minimal yscv tracker benchmark that uses the current workspace crates.
 
+#[global_allocator]
+static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc;
+
 use std::path::Path;
 use std::process::ExitCode;
 use std::time::Instant;

apps/llm-bench/src/bin/calib_accuracy.rs

@@ -23,6 +23,9 @@
 //! cargo run --release --bin calib_accuracy -p yscv-llm-bench
 //! ```
 
+#[global_allocator]
+static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc;
+
 use yscv_onnx::quantize::{
     CalibrationCollector,
     derive::{QuantTarget, derive_mse_optimal, derive_percentile, derive_symmetric},

apps/llm-bench/src/bin/compute_gap.rs

@@ -1,3 +1,6 @@
+#[global_allocator]
+static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc;
+
 use std::time::Instant;
 
 use yscv_kernels::{

apps/llm-bench/src/bin/inspect.rs

@@ -2,6 +2,9 @@
 //! of an ONNX model. Used to detect cached-decoder protocols
 //! (past_key_values.{i}.{key,value}) before wiring up the bench harness.
 
+#[global_allocator]
+static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc;
+
 use yscv_onnx::load_onnx_model_from_file;
 
 fn main() {

apps/llm-bench/src/bin/kernel_bench.rs

@@ -11,6 +11,9 @@
 //! cargo run --release --bin kernel_bench -p yscv-llm-bench
 //! ```
 
+#[global_allocator]
+static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc;
+
 use std::time::Instant;
 
 use yscv_kernels::{

apps/llm-bench/src/bin/nchwc_coverage.rs

@@ -1,3 +1,6 @@
+#[global_allocator]
+static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc;
+
 use rustc_hash::FxHashMap;
 /// NCHWc layout coverage probe — Step 0 of NCHWc-everywhere plan.
 ///

apps/llm-bench/src/bin/quantize_tracker.rs

@@ -31,6 +31,9 @@
 //! `calib_accuracy --model`. Without `--output` the QDQ save+reload step
 //! is skipped — useful when iterating on the calibration count alone.
 
+#[global_allocator]
+static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc;
+
 use std::path::{Path, PathBuf};
 use std::process::ExitCode;
 

apps/llm-bench/src/main.rs

@@ -31,6 +31,9 @@
 //! See `apps/llm-bench/README.md` for end-to-end download + quantize +
 //! bench instructions for TinyLlama-1.1B.
 
+#[global_allocator]
+static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc;
+
 use std::path::PathBuf;
 use std::process::ExitCode;
 use std::time::Instant;

benchmarks/onnx-models/benches/onnx_models.rs

@@ -1,3 +1,6 @@
+#[global_allocator]
+static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc;
+
 use criterion::{BenchmarkId, Criterion, black_box, criterion_group, criterion_main};
 use yscv_onnx::{OnnxModel, OnnxRunner, load_onnx_model_from_file, optimize_onnx_graph};
 use yscv_onnx_model_bench::{asset_dir, download_assets, make_inputs, model_cases};

benchmarks/onnx-models/examples/bench_yolo.rs

@@ -3,6 +3,9 @@
 //! Set BENCH_COOLDOWN=<secs> to insert a cooldown pause between benchmarks.
 //! The default is 0; callers that need thermal control must opt in.
 
+#[global_allocator]
+static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc;
+
 use rustc_hash::FxHashMap;
 use yscv_onnx::load_onnx_model_from_file;
 use yscv_onnx_model_bench::{asset_dir, download_assets, make_inputs, model_cases};

benchmarks/rust-competitors/Cargo.toml

@@ -6,6 +6,9 @@ publish = false
 
 [dependencies]
 tract-onnx = { version = "0.21", optional = true }
+# Matches the allocator the yscv harnesses use, so competitor numbers are not
+# a yscv-on-mimalloc vs tract-on-system-malloc comparison.
+mimalloc = { version = "0.1", default-features = false }
 
 [features]
 default = ["tract"]

benchmarks/rust-competitors/src/main.rs

@@ -1,6 +1,9 @@
 //! Benchmark competitor ONNX runtimes (tract, candle) against yscv.
 //! Outputs JSON timing results for automated comparison.
 
+#[global_allocator]
+static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc;
+
 use std::time::Instant;
 
 fn main() {

crates/yscv-autograd/Cargo.toml

@@ -17,6 +17,7 @@ thiserror.workspace = true
 
 [dev-dependencies]
 criterion = { version = "0.5", default-features = false, features = ["cargo_bench_support"] }
+mimalloc.workspace = true
 
 [[bench]]
 name = "autograd_graph_ops"

crates/yscv-autograd/benches/autograd_graph_ops.rs

@@ -1,3 +1,6 @@
+#[global_allocator]
+static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc;
+
 use criterion::{Criterion, black_box, criterion_group, criterion_main};
 use yscv_autograd::Graph;
 use yscv_tensor::Tensor;

crates/yscv-cli/Cargo.toml

@@ -26,6 +26,7 @@ thiserror.workspace = true
 
 [dev-dependencies]
 criterion = { version = "0.5", default-features = false, features = ["cargo_bench_support"] }
+mimalloc.workspace = true
 
 [[bench]]
 name = "cli_runtime_ops"

crates/yscv-cli/benches/cli_runtime_ops.rs

@@ -1,3 +1,6 @@
+#[global_allocator]
+static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc;
+
 use criterion::{Criterion, black_box, criterion_group, criterion_main};
 use yscv_detect::{
     BoundingBox, CLASS_ID_FACE, CLASS_ID_PERSON, Rgb8FaceDetectScratch, Rgb8PeopleDetectScratch,

crates/yscv-detect/Cargo.toml

@@ -22,6 +22,7 @@ rustc-hash.workspace = true
 
 [dev-dependencies]
 criterion = { version = "0.5", default-features = false, features = ["cargo_bench_support"] }
+mimalloc.workspace = true
 proptest = "1"
 
 [[bench]]

crates/yscv-detect/benches/detect_heatmap_scratch.rs

@@ -1,3 +1,6 @@
+#[global_allocator]
+static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc;
+
 use criterion::{Criterion, black_box, criterion_group, criterion_main};
 use yscv_detect::{
     FrameFaceDetectScratch, FramePeopleDetectScratch, HeatmapDetectScratch,

crates/yscv-eval/Cargo.toml

@@ -22,6 +22,7 @@ rustc-hash.workspace = true
 
 [dev-dependencies]
 criterion = { version = "0.5", default-features = false, features = ["cargo_bench_support"] }
+mimalloc.workspace = true
 
 [[bench]]
 name = "eval_metrics_ops"

crates/yscv-eval/benches/eval_metrics_ops.rs

@@ -1,3 +1,6 @@
+#[global_allocator]
+static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc;
+
 use std::time::Duration;
 
 use criterion::{Criterion, black_box, criterion_group, criterion_main};