architecture/network/gpu/mocks

Reusable WebGPU mock helpers for owner-local GPU tests.

The returned devices are intentionally shallow: they record the calls the implementation is expected to make without pulling in a real WebGPU backend. Optionally, a mock device can emulate a CPU forward pass so that parity tests can compare GPU read-back values against the CPU source of truth.

architecture/network/gpu/mocks/gpu.mock.ts

computeMockForwardPass

computeMockForwardPass(
  entries: { binding: number; resource: { buffer: GPUBuffer; }; }[],
  nodeCount: number,
): Float32Array<ArrayBufferLike>

Compute the pre-activation values for every non-input node.

This mirrors the struct-packed gather-and-activate kernel: each node's pre-activation value is the sum of incoming weights times the current source activations (read from the bound node buffer) plus the node's bias. Input nodes are never processed by a dispatch, so they retain the values uploaded by the caller.

computeMockLevels

computeMockLevels(
  connections: MockConnection[],
  nodeCount: number,
): Uint32Array<ArrayBufferLike>

Compute a simple topological level for every node from the connection list.

Level 0 contains nodes with no incoming connections; every other node gets one level above its highest predecessor. This is sufficient for the mock because GPU-eligible networks are feed-forward and the real kernel already skips disabled connections via its flags check.

createMockGPUDevice

createMockGPUDevice(
  options: any,
): MockGPUDevice

Build a fake WebGPU device that records every call made by the inference path.

Parameters:

options - Optional limits, emulator network, or output generator.

createMockNavigatorGPU

createMockNavigatorGPU(
  options: { adapter?: any; device?: any; } | undefined,
): Navigator & { gpu: GPU; }

Build a fake navigator object that exposes a mock gpu property.

MockBuffer

Narrow a generic GPUBuffer to the mock's internal backing-store shape.

MockConnection

Connection struct mirrored from the WGSL Connection layout.

MockGenerateOutput

MockGenerateOutput(
  context: MockGenerateOutputContext,
): number[] | Float32Array<ArrayBufferLike>

Optional callback that can write deterministic output values into the mock.

MockGenerateOutputContext

Context supplied to the optional per-dispatch output generator.

MockGPUDevice

MockGPUDeviceOptions

MockGPURecordings

MockParams

Per-dispatch params mirrored from the WGSL Params layout.

readBoundConnections

readBoundConnections(
  entries: { binding: number; resource: { buffer: GPUBuffer; }; }[],
): MockConnection[]

Read the bound connection buffer as an array of connection structs.

readBoundFloatArray

readBoundFloatArray(
  entries: { binding: number; resource: { buffer: GPUBuffer; }; }[],
  binding: number,
): Float32Array<ArrayBufferLike>

Read a bound buffer as a typed array, returning an empty array if the buffer is not a mock-backed GPUBuffer.

readBoundParams

readBoundParams(
  entries: { binding: number; resource: { buffer: GPUBuffer; }; }[],
): MockParams | undefined

Read the bound params uniform as a typed struct.

architecture/network/gpu/__mocks__