HaikuInsights

Native developer test instrument for Haiku OS

HaikuInsights is a hard fork of HaikuBench. It preserves the useful native benchmark foundation while rebuilding its measurement, result, worker and reporting boundaries into a trustworthy developer instrument.

The project asks five questions:

1. What capability is present?
2. Which Haiku, driver or API path exposes it?
3. Was that path actually exercised?
4. Was the result correct?
5. How well did the confirmed path perform?

Current status

0.0.0-dev is the inherited-code sanitation series. Existing benchmark values, scores, capability claims and exported reports are not yet treated as calibrated HaikuInsights evidence. See docs/roadmap.md and docs/prueffeld/phase-0000-haikuinsights-identity-and-provenance.md.

Non-goals for the sanitation series

• no recreation of commercial profiling products
• no new Intel, NVIDIA or Radeon-specific inspection stack
• no public leaderboard
• no claim that an API renderer string alone proves hardware execution

Screenshots

Quick Start

# required build dependencies
pkgman install mesa_devel glu_devel vulkan_devel

# optional Vulkan runtime path and GLSL compiler
pkgman install vulkan mesa_lavapipe shaderc

# build, smoke-test and run
make
make smoke
./HaikuInsights

Benchmarks

System Bench — 20 tests, adaptive sampling

Every test warms up the CPU for 200 ms, then samples adaptively: at least 3 runs, up to 10, stopping as soon as the coefficient of variation drops below 2%. The UI and exports report the trimmed mean (robust to a single outlier via the MAD rule) along with stddev, p50, p95 and the actual number of samples taken. JSON export also carries min/max, CoV and trimmed mean for later report validation.

Section	Tests	What it measures
CPU (3)	Integer, Float, Multi-thread	Arithmetic throughput, transcendentals, HT/multi-core scaling
Memory (4)	Seq Read, Seq Write, Copy, Latency	DDR bandwidth and random-access latency (pointer-chasing)
Cache (3)	L1, L2, L3	Per-level bandwidth with 8 independent accumulators
Kernel (8)	Semaphores, Threads, Ports, Areas, Atomics, Syscalls	Haiku kernel primitive throughput and overhead
Messaging (2)	BMessage Flatten, BLooper	Haiku IPC serialization and looper round-trip

2D Benchmark — 24 tests, 6 levels

All rendering through BView, no GPU required. Detects whether 2D hardware acceleration is active by identifying the accelerant driver via BScreen::GetDeviceInfo() and ioctl B_GET_ACCELERANT_SIGNATURE. Reports device name, chipset, VRAM and individual hook availability (FillRect, ScreenBlit, InvertRect).

Level	Theme	Tests
1	Basic Fills	FillRect, StrokeRect, FillRoundRect, StrokeLine
2	Geometry	FillTriangle, FillEllipse, FillPolygon, DrawString
3	Compositing	BShape curves, gradients, alpha blending, bitmap scaling
4	Advanced	24-point star, clipped regions, conic gradients, text storm
5	Extreme	Nebula particles, fractal trees, plasma renderer, everything combined
6	Software 3D	Teapot wireframe, flat-shaded, x4 rotating, x16 army

GPU Benchmark — 6 tests, OpenGL

Test	Description
Fill Rate	200 full-screen quads per frame
Geometry	10 000 small triangles per frame
Texture	256x256 re-upload + 50 textured quads
Alpha Blending	300 semi-transparent overlapping quads
Stencil Buffer	100 stencil masks + 150 masked draws
Combined Stress	25 rotating lit objects (teapots, spheres, tori)

Includes a live rotating teapot preview during idle.

Vulkan Benchmark — 4 tests

Test	Description
Memory Bandwidth	Host-visible 64 MB buffer write throughput
Compute Shader	1M invocations, 256 FMA ops each (GFLOPS)
Buffer Copy	Device-to-device 32 MB vkCmdCopyBuffer
Buffer Fill	64 MB vkCmdFillBuffer

Vulkan is loaded dynamically at runtime. If the Vulkan loader or a driver (e.g. mesa_lavapipe) is not installed, the benchmark shows an error message instead of crashing. The compute shader is compiled from GLSL at runtime using glslc (shaderc).

Teapot 3D — interactive OpenGL demo

Renders 1 to 64 Utah Teapots with Phong lighting via BGLView. Use the +/- buttons to add or remove teapots and watch the FPS counter.

2D Hardware Acceleration Detection

The 2D benchmark panel reports the video card's acceleration status:

• Driver: active accelerant name (e.g. intel_extreme.accelerant)
• Device: GPU name, chipset and VRAM from BScreen::GetDeviceInfo()
• 2D Acceleration: ACTIVE (hardware) or NONE (software rendering)
• Hooks: FillRect, ScreenBlit, InvertRect availability

Detection uses ioctl B_GET_ACCELERANT_SIGNATURE on /dev/graphics/ to identify the active driver, then classifies it against the known set of Haiku accelerant drivers:

Driver	2D Acceleration
intel_extreme, radeon_hd, radeon, nvidia, ati, matrox, via, intel_810	Hardware accelerated
vesa, framebuffer, virtio_gpu	Software only

Export

Clicking Export reports generates two files directly on the Desktop:

• HaikuInsights_YYYY-MM-DD_HH-MM-SS.md — human-readable report
• HaikuInsights_YYYY-MM-DD_HH-MM-SS.json — machine-readable data

Both files include:

• App version, date, Machine ID
• Full system info (CPU, cores, RAM, OS version)
• All benchmark results with mean, stddev, p50, p95, CoV
• Overall composite score broken down by subsystem

JSON structure

{
  "format_version": 1,
  "app_version": "1.1.0",
  "machine_id": "HB-A3F7B21C",
  "hardware": "Intel Core i3 M 370 | 4 cores | 3757 MB RAM",
  "os": "Haiku hrev59722",
  "system_bench": {
    "valid": true,
    "runs": 5,
    "results": {
      "cpu_integer_mops": {
        "mean": 59.10, "stddev": 0.30,
        "median": 59.05, "p95": 59.50,
        "min": 58.70, "max": 59.60,
        "cov": 0.005, "trimmed_mean": 59.08,
        "n_samples": 3, "n_dropped": 0
      },
      ...
    }
  },
  "graphics": {
    "teapot": "Teapot 3D: 25 FPS (4 teapots)",
    "bench_2d": "2D Bench: 18795 fill/s | nebula 5215.3/s | extreme 14.3/s",
    "gpu_opengl": "GPU Bench: 24.1 FPS (...)",
    "vulkan": null
  }
}

Temperature Monitoring

ACPI thermal zones are read in real-time and displayed:

• In the main window (System panel)
• As an overlay on all OpenGL benchmark windows
• In the 2D and GPU benchmark result windows

Colors: green (< 65 C), yellow (65-80 C), red (> 80 C).

Running the test suite

make smoke

The root smoke target builds the application, verifies that its Haiku resource container is readable, and runs all inherited Phase A unit tests (BenchStats, AdaptiveRunner, BenchWarmup) plus the integration smoke test (114 checks total). ./tests/run_all.sh remains the direct test entry point.

Architecture

HaikuInsights/
├── HaikuInsights           binary (with HVIF icon)
├── icon.hvif               application icon source
├── HaikuInsights.rdef       application signature, version and icon resources
├── Makefile
├── LICENSE                 MIT
├── README.md
├── CHANGELOG.md
├── haikubench-1.1.0.recipe  HaikuPorts recipe
├── WattHaiku.cpp           BApplication entry point
├── MainWindow.h/cpp        Main window, export, score
├── SysBenchmark.h/cpp      20 system tests
├── AdaptiveRunner.h/cpp    Convergence-driven multi-run sampler
├── BenchStats.h/cpp        Statistics (mean, median, p95, CoV, trimmed mean)
├── BenchWarmup.h/cpp       CPU warm-up and buffer touch
├── CpuDatabase.h/cpp       CPUID brand string detection
├── TeapotWindow.h/cpp      OpenGL teapot demo
├── Bench2DWindow.h/cpp     24 BView rendering tests + 2D accel detection
├── GpuBenchWindow.h/cpp    6 OpenGL tests + teapot preview
├── VulkanBenchWindow.h/cpp 4 Vulkan compute tests
├── TempOverlay.h/cpp       ACPI temperature overlay
├── Version.h               Version constant
└── tests/                  Unit and integration tests

Requirements

• Haiku OS (x86_64), R1/beta5 or newer
• mesa_devel, glu_devel, vulkan_devel (required build dependencies)
• vulkan (optional Vulkan ICD loader at runtime)
• mesa_lavapipe (optional Vulkan software-device provider)
• shaderc (optional runtime provider of glslc)
• shaderc_devel is not consumed by the current source or link boundary

Provenance

HaikuInsights started from HaikuBench commit 611f819e0a8871b55397ef486cf070fbdab07f35 by Andrea Bernardi. The fork keeps the original MIT license and copyright notices. See docs/provenance.md for the maintained boundary.

License

MIT License. See LICENSE.