Lampshade

Lampshade provides fast, composable GPU histograms, reduction, predicate masks, prefix scan, stream compaction, and unsigned integer radix sort for Rust applications using wgpu and WGSL.

Benchmarks

Resident GPU-buffer benchmarks include command recording, submission, execution, and reusable workspace management. They exclude host upload and validation readback. Reduction comparisons include the required four-byte scalar readback for both libraries. Inputs are deterministic, outputs are validated, and reported comparisons are medians of independent process medians.

The published-release regression harness runs identical resident workloads against crates.io 0.7 and the current checkout, writes raw runs and process medians to JSON, and enforces a 2% regression budget. The 0.8 typed-pipeline stabilization report records the final fixed-path gate and targeted rechecks. The Lampshade migration report verifies the renamed final source against the published 0.7 predecessor. The GPU-resident count report separates isolated scheduling cost from full compaction-to-sort/reduction results on RTX, Intel, and two Jetsons, plus fixed-path regression controls.

Against Massively 0.96

On an RTX 4070 Ti SUPER using Vulkan, Lampshade was faster in every overlapping 100-million-item workload:

The same comparison at 10 million items also favored Lampshade on two Jetson Orin Nano systems:

See the Massively harness and wgpu 30 report for the method, exact revisions, complete matrices, and machine-readable results.

Intel Vulkan

On Intel Alder Lake-N integrated graphics at 10 million items, Lampshade led Massively in every workload. Sort uses the capability-gated 4-bit radix path; reduction uses the portable kernel:

All 74 release tests passed. At 100M, reduction measured 21.983 ms versus 22.836 ms for Massively, a 1.04x lead. The Intel wide-radix report includes 1M-100M results, stage profiles, and measured regression controls. At 100M, the same four speedups are 9.78x, 4.79x, 2.44x, and 2.52x respectively.

Apple Metal

Upgrading from wgpu 28 to wgpu 30 removed the previous host-returning reduction deficit on an M3 Pro. These are final-candidate medians of three independent process medians:

Massively 0.96 could not initialize these Metal pipelines: its generated layouts requested 42 or 47 storage buffers against the adapter limit of 29. The harness records this as an unsupported comparison, not an artificial speedup. Reduction does run in both libraries and uses the same end-to-host scalar boundary. All 74 release GPU tests and every 100M benchmark validator pass on the M3 Pro. See the wgpu 30 report, the earlier Apple report, and the upstream issue.

Against wgpu_sort

On the tested NVIDIA Vulkan system at 100 million key/value pairs:

The wgpu_sort report documents the pinned baseline and reproduction harness.

Features

• Portable 1-256-bin u32 histograms with workgroup-private counters.
• Wrapping sum, minimum, and maximum reduction for u32 values.
• Inclusive and exclusive u32 prefix scan.
• Reusable comparison predicates that produce compaction-ready masks.
• Stable compaction of u32 values and KeyValue records.
• Stable radix sort for u32 values and (u32 key, u32 value) pairs.
• Explicit key-width bounds that skip unnecessary radix passes.
• Slice APIs for simple upload/execute/readback workflows.
• GPU-buffer APIs for composing work in one command encoder.
• Capacity-bounded sort and reduction driven by GPU-resident item counts.
• Typed buffer views and an ordered pipeline recorder that prepares shared GPU-count metadata automatically.
• Reusable scratch storage and no unsafe blocks in library code.

The measurements above were collected under the former wgpu-primitives package name. The 0.8 rebrand changes package/import names but not kernels or timing boundaries.

Installation

Lampshade 0.8 uses wgpu 30. Tokio is listed because the executable quick start below uses #[tokio::main]; library development dependencies do not propagate to applications.

[dependencies]

lampshade = "0.8"

tokio = { version = "1", features = ["macros", "rt-multi-thread"] }

The predecessor was published as wgpu-primitives = "0.7". Existing users can move to lampshade = "0.8" and change Rust imports from wgpu_primitives to lampshade. Because wgpu types appear in the public GPU-buffer APIs, upgrading from versions before 0.6 also requires wgpu 30.

Quick start

use lampshade::{
    Compactor, Context, MaskGenerator, Reducer, Scanner, Sorter, U32Predicate,
};

#[tokio::main]
async fn main() -> Result<(), lampshade::Error> {
    let context = Context::init().await?;
    let generator = MaskGenerator::from_context(&context);
    let mut reducer = Reducer::from_context(&context);
    let mut scanner = Scanner::from_context(&context);
    let mut compactor = Compactor::from_context(&context);
    let mut sorter = Sorter::from_context(&context);

    let input = [4, 17, 9, 22, 11, 3];
    let mask = generator
        .mask(&input, U32Predicate::GreaterThanOrEqual(10))
        .await?;

    assert_eq!(mask, [0, 1, 0, 1, 1, 0]);
    assert_eq!(reducer.sum(&input).await?, 66);
    assert_eq!(scanner.scan_exclusive(&[3, 1, 4, 1]).await?, [0, 3, 4, 8]);
    assert_eq!(compactor.compact(&input, &mask).await?, [17, 22, 11]);
    assert_eq!(sorter.sort(&input).await?, [3, 4, 9, 11, 17, 22]);
    Ok(())
}

See examples/ for standalone primitives and composed resident pipelines. The particle pipeline filters, stably compacts, and depth-sorts key/entity records with one submission and one final readback.

GPU-resident composition

The pipeline API below is part of Lampshade 0.8 and is not present in the published wgpu-primitives 0.7 predecessor.

Applications that already own a wgpu device should reuse it and record multiple primitives before submitting once. The stable pipeline API carries buffer ranges, capacities, and fixed or GPU-resident extents between operations:

let mut primitives = pipeline::Primitives::new(&device, &queue);
let input_view = pipeline::GpuSlice::from_range(&input_buffer, 0..item_count)?;
let mask_output = pipeline::GpuSliceMut::from_range(&mask_buffer, 0..item_count)?;
let compacted = pipeline::GpuSliceMut::from_range(&compacted_buffer, 0..item_count)?;
let sorted = pipeline::GpuSliceMut::from_range(&sorted_buffer, 0..item_count)?;
let sum = pipeline::GpuSliceMut::from_range(&sum_buffer, 0..1)?;
let count = pipeline::GpuCount::new(&output_count)?;

primitives.reserve_workspace(
    pipeline::WorkspaceRequirements::new(item_count)
        .predicate()
        .compact()
        .counted_sort()
        .counted_reduce(),
)?;
primitives.reserve_count(count, item_count)?;

let mut recorder = primitives.record(&mut encoder);
let mask = recorder.mask(
    input_view,
    mask_output,
    U32Predicate::GreaterThanOrEqual(10),
)?;
let compacted = recorder.compact(input_view, mask, compacted, count)?;
let sorted = recorder.sort(compacted, sorted, pipeline::SortOptions::default())?;
recorder.reduce(sorted, sum, U32Reduction::Sum)?;
drop(recorder);
queue.submit(Some(encoder.finish()));

resident_pipeline.rs composes u32 predicate, compaction, sort, and reduction. The particle example proves the same typed flow for KeyValue records: predicate, stable compaction, and stable sort by key. Compaction writes the selected count and later primitives consume it without a CPU synchronization point. The recorder caches a GpuCountPlan internally and schedules its preparation once after the count producer. Existing raw-buffer and explicit-plan APIs remain available.

GpuSlice ranges use element indices and may start at aligned nonzero offsets. Different read/write roles in one primitive must still use distinct underlying buffer handles: WebGPU treats writable storage use as exclusive even for disjoint static binding ranges. reserve_workspace prepares only the requested pipelines and grows only their capacity-dependent workspaces; bind groups and small uniform buffers may still be created while commands are recorded.

Plans default to CountedSortDispatch::Indirect, which sizes radix reduce/scatter launches to the GPU-selected prefix and is the portable choice for unknown or sparse counts. Its histogram scan remains capacity-sized. CountedSortDispatch::Capacity trades inactive workgroups for lower dispatch overhead and should be selected only with workload-specific benchmark evidence.

The command encoder preserves GPU execution order. Rust borrows the encoder and buffers only while recording; no input is cloned or read back. Use KeyValueSorter::new_for_adapter when adapter metadata is available so compatible fast paths can be selected.

The resident methods validate sizes, ranges, alignment, and usages, but do not inspect GPU data. Masks must contain only 0 or 1; declared key-width bounds must contain every key. Primitive participants that read and write must use distinct buffer handles. Full usage requirements are documented on each API at docs.rs.

Applications that own the adapter as well as the device should construct the facade with Primitives::new_for_adapter(&device, &queue, &adapter_info) so measured hardware-specific paths remain available. The repository-only standalone consumer validates this public API boundary and records typed-versus-raw overhead on discrete NVIDIA and integrated Intel GPUs.

See the architecture guide for the public convenience, resident composition, and private kernel/runtime layers. The typed-pipeline guide records the API contract and stabilization evidence.

How it works

• Histogram: each workgroup accumulates up to 256 counters in shared memory, then merges at most one count per bin into the global output. Values outside the requested range are ignored.
• Reduction: each workgroup combines a coalesced input range into one partial value; later passes repeat over the partials until one value remains. A count plan builds the hierarchy and indirect dispatch arguments from a GPU-resident length.
• Predicate mask: one thread evaluates each value or KeyValue field and writes a 0 or 1.
• Scan: workgroups scan local ranges, recursively scan block totals, then add those totals to produce global prefixes. Supported devices use subgroup operations; others use the portable shared-memory path.
• Compaction: an exclusive mask scan gives stable destination indices. Scatter combines block-local offsets with scanned block totals without materializing another full-size prefix pass.
• Radix sort: stable least-significant-digit passes ping-pong between buffers. Known key-width bounds reduce the pass count. Compatible NVIDIA Vulkan devices use 8-bit or 4-bit paths, capable Intel Vulkan devices use a 4-bit path, and other adapters retain the portable 2-bit path. GPU-counted sorting uses the portable kernel with either count-proportional indirect dispatch or explicit capacity dispatch while preserving the same stable ordering contract.

Profiling

GPU timestamp spans are available for every primitive when the adapter supports timestamp queries. Dispatches also carry stable labels for tools such as NVIDIA Nsight Graphics.

$env:WGPU_BACKEND = 'vulkan' # or 'dx12'
$env:WGPU_PRIMITIVES_PROFILE_CASES = 'compact_50'
$env:WGPU_PRIMITIVES_PROFILE_VALIDATE = '1'
cargo run --release --example profile_primitives

Cases include histogram, reduction, scan, sort, predicate, value compaction, and key/value compaction at selectable sizes and selectivities.

Roadmap

1. Validate AMD, more Intel and Apple GPUs, and additional driver versions.
2. Grow the CUB-like private kernel/workspace engine behind the existing safe, Thrust-like Rust APIs; split crates only when usage evidence justifies it.
3. Improve portable key-width detection for GPU-resident inputs.
4. Add derived primitives only when real workloads justify their API and cost.
5. Revisit full-width scatter when hardware counters or a new algorithm provide evidence for at least a 5% gain.

New primitives require a resident-buffer API, deterministic boundary tests, CPU-reference validation, and reproducible benchmarks.

Development

cargo fmt --all --check

cargo clippy --all-targets --all-features -- -D warnings

cargo test --release --all-targets

cargo check --examples --benches

cargo package

Criterion benches cover each primitive plus counted_pipeline and the raw-versus-typed particle_pipeline. GPU integration tests skip when no compatible adapter is available; CI uses Mesa's Vulkan software adapter.

License

MIT