Merge pull request #16 from Human9000-bit/fxhash

migrate internal hash maps to rustc's FxHasher

Author: enthropy7

Cargo.lock

@@ -43,6 +43,7 @@ keywords = ["computer-vision", "deep-learning", "image-processing", "onnx", "vid
 
 [workspace.dependencies]
 thiserror = "2"
+rustc-hash = "2.1.3"
 
 [workspace.lints.rust]
 unsafe_code = "deny"

Cargo.toml

@@ -46,6 +46,7 @@ yscv-kernels = { version = "0.1", path = "../../crates/yscv-kernels", default-fe
 serde = { version = "1.0", features = ["derive"] }
 serde_json = "1.0"
 thiserror.workspace = true
+rustc-hash.workspace = true
 
 [lints]
 workspace = true

apps/llm-bench/Cargo.toml

@@ -1,3 +1,4 @@
+use rustc_hash::FxHashMap;
 /// NCHWc layout coverage probe — Step 0 of NCHWc-everywhere plan.
 ///
 /// Loads a tracker ONNX model and classifies every Conv op as NCHWc-eligible
@@ -7,8 +8,7 @@
 /// Usage:
 ///   cargo run --release --no-default-features -p yscv-llm-bench --bin nchwc_coverage \
 ///     -- private/private/model.onnx
-use std::collections::{HashMap, HashSet};
-
+use rustc_hash::FxHashSet;
 use yscv_onnx::{OnnxAttribute, OnnxModel, OnnxNode, load_onnx_model_from_file};
 
 fn main() {
@@ -76,8 +76,8 @@ fn is_passthrough(op_type: &str) -> bool {
 /// Build a map: tensor_name → shape, tracing through Identity nodes to
 /// resolve weight-sharing aliases. In Siamese networks, branch1 weights
 /// are Identity(branch0_initializer) — treat them as static weights.
-fn build_effective_initializers(model: &OnnxModel) -> HashMap<String, Vec<usize>> {
-    let mut effective: HashMap<String, Vec<usize>> = model
+fn build_effective_initializers(model: &OnnxModel) -> FxHashMap<String, Vec<usize>> {
+    let mut effective: FxHashMap<String, Vec<usize>> = model
         .initializers
         .iter()
         .map(|(k, v)| (k.clone(), v.shape().to_vec()))
@@ -139,8 +139,8 @@ fn run_coverage_probe(model: &OnnxModel) {
     // Classify each conv op
     let mut eligible = 0usize;
     let mut rejected = 0usize;
-    let mut kind_counts: HashMap<String, usize> = HashMap::new();
-    let mut reject_reasons: HashMap<String, usize> = HashMap::new();
+    let mut kind_counts: FxHashMap<String, usize> = FxHashMap::default();
+    let mut reject_reasons: FxHashMap<String, usize> = FxHashMap::default();
 
     for (_, node) in &conv_ops {
         let weight_name = node.inputs.get(1).map(|s| s.as_str()).unwrap_or("");
@@ -331,7 +331,7 @@ fn run_coverage_probe(model: &OnnxModel) {
     println!();
     println!("QLinearConv ops in model: {}", qlinear_conv.len());
     // Show all non-Conv node op types and counts
-    let mut op_counts: HashMap<&str, usize> = HashMap::new();
+    let mut op_counts: FxHashMap<&str, usize> = FxHashMap::default();
     for node in &model.nodes {
         *op_counts.entry(node.op_type.as_str()).or_default() += 1;
     }
@@ -349,7 +349,7 @@ fn run_coverage_probe(model: &OnnxModel) {
     // Non-passthrough, non-Conv ops require NHWC → insert converter at boundary
 
     // Build tensor → producing node index
-    let mut tensor_producer: HashMap<&str, usize> = HashMap::new();
+    let mut tensor_producer: FxHashMap<&str, usize> = FxHashMap::default();
     for (i, node) in model.nodes.iter().enumerate() {
         for out in &node.outputs {
             if !out.is_empty() {
@@ -421,8 +421,8 @@ fn run_coverage_probe(model: &OnnxModel) {
     // And number of exit converters (nchwc tensors consumed by non-nchwc non-Conv nodes)
 
     // Use effective_weights keys as the "static" set (includes Identity-aliased weights)
-    let static_names: HashSet<&str> = effective_weights.keys().map(|s| s.as_str()).collect();
-    let initializer_names: HashSet<&str> = static_names.clone();
+    let static_names: FxHashSet<&str> = effective_weights.keys().map(|s| s.as_str()).collect();
+    let initializer_names: FxHashSet<&str> = static_names.clone();
     let dynamic_inputs: Vec<&str> = model
         .inputs
         .iter()
@@ -437,7 +437,7 @@ fn run_coverage_probe(model: &OnnxModel) {
     }
 
     // Mark which tensors are in nchwc domain (after potential converter)
-    let mut nchwc_tensors: HashSet<String> = HashSet::new();
+    let mut nchwc_tensors: FxHashSet<String> = FxHashSet::default();
     // Graph inputs start as NHWC; we will count converters needed for them
     // if any eligible node consumes them directly.
 
@@ -446,7 +446,7 @@ fn run_coverage_probe(model: &OnnxModel) {
     let mut internal_boundaries = 0usize;
 
     // Track which dynamic inputs get a converter
-    let mut converted_inputs: HashSet<String> = HashSet::new();
+    let mut converted_inputs: FxHashSet<String> = FxHashSet::default();
 
     for (i, node) in model.nodes.iter().enumerate() {
         if !node_eligible[i] && !matches!(node.op_type.as_str(), "Conv" | "Conv_Relu" | "Conv_SiLU")

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

@@ -35,6 +35,7 @@ use std::collections::HashMap;
 use std::path::{Path, PathBuf};
 use std::process::ExitCode;
 
+use rustc_hash::FxHashMap;
 use serde::Deserialize;
 use yscv_onnx::quantize::{
     CalibrationCollector, fold_constant_qdq_weights_for_yscv_fast, prune_unused_initializers,
@@ -385,8 +386,8 @@ struct TrackingDelta {
 /// the output maps by shape, then compare argmax cell + the box at the fp32
 /// peak cell. Returns `None` if no 1-channel score map is present.
 fn decoded_tracking_delta(
-    out_fp32: &HashMap<String, Tensor>,
-    out_qdq: &HashMap<String, Tensor>,
+    out_fp32: &FxHashMap<String, Tensor>,
+    out_qdq: &FxHashMap<String, Tensor>,
 ) -> Option<TrackingDelta> {
     let mut score: Option<(&Tensor, &Tensor)> = None;
     let mut bbox: Option<(&Tensor, &Tensor)> = None;

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

@@ -18,6 +18,7 @@ yscv-onnx = { version = "0.1", path = "../yscv-onnx", optional = true }
 yscv-tensor = { version = "0.1", path = "../yscv-tensor" }
 yscv-video = { version = "0.1", path = "../yscv-video" }
 thiserror.workspace = true
+rustc-hash.workspace = true
 
 [dev-dependencies]
 criterion = { version = "0.5", default-features = false, features = ["cargo_bench_support"] }

crates/yscv-detect/Cargo.toml

@@ -118,9 +118,9 @@ pub fn soft_nms(detections: &mut Vec<Detection>, sigma: f32, score_threshold: f3
 /// each group independently, then merges and returns results sorted by score
 /// descending.
 pub fn batched_nms(detections: &[Detection], iou_threshold: f32) -> Vec<Detection> {
-    use std::collections::HashMap;
+    use rustc_hash::FxHashMap;
 
-    let mut by_class: HashMap<usize, Vec<Detection>> = HashMap::new();
+    let mut by_class: FxHashMap<usize, Vec<Detection>> = FxHashMap::default();
     for det in detections {
         by_class.entry(det.class_id).or_default().push(*det);
     }

crates/yscv-detect/src/nms.rs

@@ -360,7 +360,7 @@ pub fn detect_yolov8_onnx(
     img_width: usize,
     config: &YoloConfig,
 ) -> Result<Vec<Detection>, crate::DetectError> {
-    use std::collections::HashMap;
+    use rustc_hash::FxHashMap;
 
     let input_name = model
         .inputs
@@ -373,7 +373,7 @@ pub fn detect_yolov8_onnx(
         image_data.to_vec(),
     )?;
 
-    let mut inputs = HashMap::new();
+    let mut inputs = FxHashMap::default();
     inputs.insert(input_name, tensor);
 
     let outputs = yscv_onnx::run_onnx_model(model, inputs)?;

crates/yscv-detect/src/yolo.rs

@@ -18,6 +18,7 @@ serde_json = "1.0"
 roxmltree = "0.20"
 csv = "1.3"
 thiserror.workspace = true
+rustc-hash.workspace = true
 
 [dev-dependencies]
 criterion = { version = "0.5", default-features = false, features = ["cargo_bench_support"] }

crates/yscv-eval/Cargo.toml

@@ -1,4 +1,4 @@
-use std::collections::HashMap;
+use rustc_hash::{FxBuildHasher, FxHashMap};
 
 use yscv_detect::{BoundingBox, Detection};
 
@@ -11,7 +11,8 @@ pub(crate) fn build_detection_dataset_from_coco(
     predictions: Vec<CocoPredictionWire>,
 ) -> Result<Vec<DetectionDatasetFrame>, EvalError> {
     let mut frames = Vec::with_capacity(ground_truth.images.len());
-    let mut image_index_by_id = HashMap::with_capacity(ground_truth.images.len());
+    let mut image_index_by_id =
+        FxHashMap::with_capacity_and_hasher(ground_truth.images.len(), FxBuildHasher);
 
     for (frame_idx, image) in ground_truth.images.into_iter().enumerate() {
         if image_index_by_id.insert(image.id, frame_idx).is_some() {