WGPU backend for cross-platform GPU compression (Metal, Vulkan, WebGPU)

[noahgift]

Summary

Add WGPU backend alongside cudarc to enable GPU-accelerated compression on all platforms.

Current State

trueno-zram GPU support:
├── cudarc (CUDA) → NVIDIA Linux only
└── No Mac, AMD, Intel, or browser support

Proposed State

trueno-zram GPU backends:
├── cudarc (CUDA) → NVIDIA Linux/Windows (existing)
├── wgpu (Metal) → macOS (Apple Silicon + Intel)
├── wgpu (Vulkan) → Linux AMD/Intel, Android
├── wgpu (DX12) → Windows AMD/Intel
└── wgpu (WebGPU) → Browsers (WASM)

Target Platforms

Platform	GPU API	Status
Linux + NVIDIA	CUDA	✅ Implemented
Linux + AMD/Intel	Vulkan via WGPU	🔲 Planned
macOS Intel	Metal via WGPU	🔲 Planned
macOS Apple Silicon	Metal via WGPU	🔲 Planned
Windows	DX12/Vulkan via WGPU	🔲 Planned
Browser/WASM	WebGPU via WGPU	🔲 Planned

Implementation

Dependencies

[dependencies]
wgpu = "23"

Backend Selection

pub enum GpuBackend {
    Cuda(CudaContext),      // Existing
    Wgpu(wgpu::Device),     // New
}

impl GpuBackend {
    pub fn auto_detect() -> Option<Self> {
        // 1. Try CUDA first (fastest on NVIDIA)
        if let Some(cuda) = CudaContext::new().ok() {
            return Some(GpuBackend::Cuda(cuda));
        }
        // 2. Fall back to WGPU (Metal, Vulkan, DX12)
        if let Some(wgpu) = WgpuContext::new().ok() {
            return Some(GpuBackend::Wgpu(wgpu));
        }
        None
    }
}

Compute Shader (WGSL)

@group(0) @binding(0) var<storage, read> input_pages: array<u32>;
@group(0) @binding(1) var<storage, read_write> output: array<u32>;

@compute @workgroup_size(256)
fn lz4_compress(@builtin(global_invocation_id) id: vec3<u32>) {
    let page_idx = id.x;
    // LZ4 compression kernel
}

Use Cases

1. Intel Mac via SSH - Metal GPU for coverage builds
2. Apple Silicon Macs - Native Metal acceleration
3. AMD GPU workstations - Vulkan backend
4. Browser-based tools - WebGPU for web IDEs
5. Edge devices - Portable WASM deployment

Acceptance Criteria

• Add wgpu feature flag
• Implement WgpuContext with device auto-detection
• Port LZ4 compress kernel to WGSL
• Port Zstd compress kernel to WGSL
• Benchmark Metal vs CUDA performance
• Test on Intel Mac (ssh mac)
• WASM compilation works

Performance Targets

Platform	Target Throughput
CUDA (RTX 4090)	60 GB/s
Metal (M1 Max)	30 GB/s
Vulkan (RX 7900)	40 GB/s
WebGPU (browser)	10 GB/s

Related

• Issue GPU batch compression for large workloads (5× PCIe rule) #1: GPU batch compression
• ZRAM-001 spec: GPU/CUDA subsystem
• wgpu docs: https://wgpu.rs

Labels

enhancement, gpu, cross-platform, wgpu

[noahgift]

Triage (2026-03-31): Feature request (WGPU backend for cross-platform GPU compression). Keeping open as backlog.

[noahgift]

Triage: P3 — feature request. WGPU backend for cross-platform GPU compression (Metal, Vulkan, WebGPU). Enhancement.