WasmGPU.compute.kernels.gemmF32¶
Summary¶
gemmF32 computes row-major out = alpha * A * B + beta * out for f32 matrices.
Syntax¶
WasmGPU.compute.kernels.gemmF32(a: StorageBuffer, b: StorageBuffer, m: number, n: number, k: number, opts?: GemmF32Options): StorageBuffer
Parameters¶
| Name | Type | Required | Description |
|---|---|---|---|
a |
StorageBuffer |
Yes | Row-major f32 matrix with shape (m, k). |
b |
StorageBuffer |
Yes | Row-major f32 matrix with shape (k, n). |
m |
number |
Yes | Non-negative output row count. |
n |
number |
Yes | Non-negative output column count. |
k |
number |
Yes | Non-negative shared inner dimension. |
opts |
GemmF32Options |
No | Output, scale factors, encoder, label, and limit-validation controls. |
Returns¶
StorageBuffer - Row-major f32 matrix with shape (m, n). A new output enables COPY_SRC; opts.out is returned when supplied.
Type Details¶
type GemmF32Options = {
out?: StorageBuffer;
alpha?: number; // default: 1
beta?: number; // default: 0
encoder?: GPUCommandEncoder;
label?: string;
validateLimits?: boolean;
};
The operation computes out = alpha * A * B + beta * out. Each dimension must fit in u32, and each derived element count must be a safe integer that fits in u32. A supplied output needs m * n * 4 bytes and must differ from both inputs. When opts.encoder is supplied, commands are recorded but not submitted by this call.
Example¶
const wgpu = await WasmGPU.create(document.querySelector("canvas"));
const a = wgpu.compute.createStorageBuffer({ data: new Float32Array([1, 2, 3, 4]) });
const b = wgpu.compute.createStorageBuffer({ data: new Float32Array([5, 6, 7, 8]) });
const out = wgpu.compute.kernels.gemmF32(a, b, 2, 2, 2);
console.log(Array.from(await wgpu.compute.readback.readF32(out)));