miniav_tools_minigpu

GPU compute codec backend for miniav_tools, powered by minigpu (WGSL via Google Dawn).
Self-registers with the miniav_tools_platform_interface registry on import.

Importing package:miniav_tools_minigpu/miniav_tools_minigpu.dart is enough. Once registered, MiniAVTools.createEncoder(...) will use this backend for codecs it supports at priority 30 (below WebCodecs, above software fallback).

What it provides

• MJPEG encoder — DCT and VLC computed entirely in WGSL compute shaders. No libavcodec dependency; runs on any Dawn-capable GPU (Vulkan, D3D12, Metal).
• GpuCodecPipeline — composable pipeline abstraction for chaining compute passes (kFrameInputKey → … → kEncodedOutputKey).
• MinigpuMjpegPipeline — concrete MJPEG pipeline built on top of GpuCodecPipeline.

Planned (Phase B/C):

• VP8/VP9-style intra-only "raw GPU" codec for screen capture.
• Custom ML-inference-optimised video codec.

Usage

import 'package:miniav_tools/miniav_tools.dart';
import 'package:miniav_tools_minigpu/miniav_tools_minigpu.dart'; // self-registers

final encoder = await MiniAVTools.createEncoder(EncoderConfig(
  codec: VideoCodec.mjpeg,
  width: 1920, height: 1080,
));

Direct pipeline access

import 'package:miniav_tools_minigpu/miniav_tools_minigpu.dart';

final pipeline = MinigpuMjpegPipeline();
await pipeline.init(width: 1920, height: 1080);

final encoded = await pipeline.encode(frame);
await pipeline.dispose();

Supported codecs

Dependencies

• minigpu — GPU compute facade
• gpu_tensor — tensor buffers over Dawn
• gpu_pipeline — shader pipeline helpers
• miniav_tools_platform_interface

See also

• miniav_tools — user-facing facade
• miniav_tools_ffmpeg — FFmpeg backend (software + hardware)
• miniav_tools_web — WebCodecs backend (browser)
• Design doc