GPUTraceAggregation

Overview

GPUTraceAggregation computes dense counts and duration statistics over canonical trace columns. GPUTraceTimeBuckets computes interval occupancy and clipped active duration over trace time. Both are exported from @luma.gl/experimental/gpu-trace; their implementation reuses the generic GPUGroupAggregation and GPUHistogram operations.

GitHub

Scroll page · Ctrl/⌘ + scroll to interact

The live viewer uses a separate analysis command graph. It can profile the entire trace, the visible viewport, or a measured time interval; viewport updates are deferred until navigation settles. The graph keeps group counts, duration histograms, and clipped time buckets on the GPU and reads back only the small fixed-size chart result. Unchanged scope generations reuse cached output.

At a glance

Question	Answer
Problem	Compute trace-domain counts, duration statistics, and trace-time occupancy.
Reads / writes	Reads canonical trace columns, scope masks, and time intervals; writes dense groups, buckets, and summaries.
Ownership	Public inputs and outputs are caller-owned; scratch storage is graph-owned transient memory.
Output contract	Bounded group rows and time buckets with counts, clipped durations, and validation status.
Expected work	Linear selected-row aggregation plus bounded group or bucket finalization.
Chunks	Preserves canonical source partitions and aggregates into explicit result domains.
Conditions / budgets	Caches by scope generation and supports budgeted full-trace analysis.
Neighborhood	viewport/measurement/full-trace mask → GPUTraceAggregation → charts and cross-filtering.

CostScope cardinality, group count, and bucket count—not only final chart size.

Common mistakeDo not reuse a cached result across a different analysis scope or selection generation.

Reusable chart-output layout

GPUTraceAnalyticsOutputLayout packs named uint32 and float32 series into one compact result buffer. It calculates aligned offsets, creates typed graph views for contributors, and decodes the same named series after a small asynchronous readback. Applications therefore do not need to duplicate byte-offset arithmetic between their GPU graph and chart UI.

import {GPUTraceAnalyticsOutputLayout} from '@luma.gl/experimental/gpu-trace';

const outputLayout = new GPUTraceAnalyticsOutputLayout([
  {id: 'service-counts', format: 'uint32', length: serviceCount},
  {id: 'service-duration', format: 'float32', length: serviceCount},
  {id: 'time-bucket-counts', format: 'uint32', length: 64}
]);

const resultBuffer = device.createBuffer({
  byteLength: outputLayout.byteLength,
  usage: Buffer.STORAGE | Buffer.COPY_SRC
});

const resultHandle = graph.importBuffer({id: 'analytics-results'}, resultBuffer);
const serviceCounts = outputLayout.createUint32View(graph, resultHandle, 'service-counts');

// After an explicitly requested compact readback:
const counts = outputLayout.decodeUint32(bytes, 'service-counts');

The layout owns no buffers, submission policy, or charts. The same views can remain GPU-resident for a renderer; decoding is only for applications that choose to draw small DOM or canvas charts.

Usage

import {GPUTraceAggregation} from '@luma.gl/experimental/gpu-trace';

new GPUTraceAggregation({
  trace: traceView,
  dimension: 'process',
  metric: 'duration-mean',
  selection: interaction.visibleMask,
  output: meanDurationByProcess
}).addToGraph(graph);

Built-in dimensions are lane, group, process, thread, and classification. Applications may instead supply a dense unsigned key column, such as an operation dictionary ID. Inputs may be packed or interleaved and may preserve multiple source chunks.

The supported metrics are count, duration-sum, duration-min, duration-max, and duration-mean. Count outputs use uint32; duration outputs use float32. Output length defines the accepted dense key range, and keys outside that range are ignored.

Trace-time buckets

import {GPUTraceTimeBuckets} from '@luma.gl/experimental/gpu-trace';

new GPUTraceTimeBuckets({
  trace: traceView,
  domain: [traceStart, traceEnd],
  selection: interaction.visibleMask,
  countOutput: intersectingSpanCounts,
  durationOutput: clippedDurationSums
}).addToGraph(graph);

The output length defines an equal-width trace-time partition. A span contributes once to every bucket it intersects, and its duration contribution is clipped at the bucket boundaries. For traces that guarantee non-overlapping spans within each lane, dividing a bucket's duration sum by bucketDuration * laneCount yields utilization. Overlapping spans intentionally produce summed occupancy rather than an interval union.

Execution contract

• The contributor declares graph work only; it never compiles, submits, or maps buffers.
• The optional selection must contain one zero/nonzero row per canonical span and preserve input chunk boundaries when the trace uses multiple chunks.
• Outputs are cleared on every graph encoding and remain GPU-resident.
• Dataset-wide aggregations should be cached or encoded only when their input generation changes; do not attach a full-trace aggregation to every viewport-only frame.