| --- |
| library_name: kernels |
| license: apache-2.0 |
| tags: |
| - kernel |
| - webgpu |
| - wgsl |
| --- |
| # ai.onnx.MatMulInteger |
|
|
| `ai.onnx` · standard ONNX operator · ONNX opset ≥ 10 |
|
|
| ## Description |
|
|
| Computes an integer matrix product with 8-bit inputs, `int32` accumulation, and independently optional zero points that default to 0. The package implements rank-1 dot products, rank-2 products, rank-2/rank-3 broadcasting, rank-3 products, and rank-4-by-rank-4 products. Scalar zero points are supported throughout; `b_zero_point` additionally supports `[N]` for rank-2 B and `[batch, 1, N]` for non-broadcast rank-3 B. Other standard ONNX matmul rank combinations and N-D per-row/per-column zero-point layouts are unsupported. |
|
|
| See the [ONNX `MatMulInteger` spec](https://onnx.ai/onnx/operators/onnx__MatMulInteger.html) for the reference semantics. |
|
|
| ## Inputs |
|
|
| | Name | Upstream name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `a` | `A` | `TA` | — | — | N-dimensional integer matrix A (int8 or uint8). | required | |
| | `b` | `B` | `TB` | — | — | N-dimensional integer matrix B (int8 or uint8). | required | |
| | `a_zero_point` | — | `TA` | — | — | Optional scalar zero point for A; defaults to 0. Standard N-D per-row layouts are unsupported. | optional | |
| | `b_zero_point` | — | `TB` | — | — | Optional zero point for B; defaults to 0. Supports a scalar, `[N]` for rank-2 B, or `[batch, 1, N]` for non-broadcast rank-3 B; other standard N-D per-column layouts are unsupported. | optional | |
|
|
| ## Outputs |
|
|
| | Name | Upstream name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `y` | `Y` | `TY` | derived | ONNX MatMul result of `a` and `b` | int32 matrix product result of A * B. | required | |
|
|
| ## Type constraints |
|
|
| | Variable | Allowed dtypes | |
| | --- | --- | |
| | `TA` | `uint8`, `int8` | |
| | `TB` | `uint8`, `int8` | |
| | `TY` | `int32` | |
|
|
| ## Implementation variants |
|
|
| One implementation is selected per call from the device capabilities, the request shapes and the dtypes; these notes say what each one covers. |
|
|
| - `sgmat_precast_a0_bnone_f16` — Prepare exact f16 representations of adjusted byte integers once, then load them directly into f32-accumulating subgroup matrices. Row groups follow matrix geometry, output rows, invocation limits and workgroup storage; bounded partial sums retain exact modular int32 arithmetic. |
| - `portable_precast_a0_bnone_f16` — Prepare lossless f16 representations of adjusted byte integers, then use vector workgroup GEMM with exact bounded f32 partials and modular int32 totals. Geometry follows output rows, invocation and storage limits; f32 storage requires enough row reuse to amortize preparation. |
| - `sgmat_precast_a0_bnone` — Prepare exact f32 representations of adjusted byte integers once, then load them directly into f32-accumulating subgroup matrices. Row groups follow matrix geometry, output rows, invocation limits and workgroup storage; bounded partial sums retain exact modular int32 arithmetic. |
| - `portable_precast_a0_bnone` — Prepare lossless f32 representations of adjusted byte integers, then use vector workgroup GEMM with exact bounded f32 partials and modular int32 totals. Geometry follows output rows, invocation and storage limits; f32 storage requires enough row reuse to amortize preparation. |
| - `sgmat_precast_a0_bscalar_f16` — Prepare exact f16 representations of adjusted byte integers once, then load them directly into f32-accumulating subgroup matrices. Row groups follow matrix geometry, output rows, invocation limits and workgroup storage; bounded partial sums retain exact modular int32 arithmetic. |
| - `portable_precast_a0_bscalar_f16` — Prepare lossless f16 representations of adjusted byte integers, then use vector workgroup GEMM with exact bounded f32 partials and modular int32 totals. Geometry follows output rows, invocation and storage limits; f32 storage requires enough row reuse to amortize preparation. |
| - `sgmat_precast_a0_bscalar` — Prepare exact f32 representations of adjusted byte integers once, then load them directly into f32-accumulating subgroup matrices. Row groups follow matrix geometry, output rows, invocation limits and workgroup storage; bounded partial sums retain exact modular int32 arithmetic. |
| - `portable_precast_a0_bscalar` — Prepare lossless f32 representations of adjusted byte integers, then use vector workgroup GEMM with exact bounded f32 partials and modular int32 totals. Geometry follows output rows, invocation and storage limits; f32 storage requires enough row reuse to amortize preparation. |
| - `sgmat_precast_a0_bcolumn_f16` — Prepare exact f16 representations of adjusted byte integers once, then load them directly into f32-accumulating subgroup matrices. Row groups follow matrix geometry, output rows, invocation limits and workgroup storage; bounded partial sums retain exact modular int32 arithmetic. |
| - `portable_precast_a0_bcolumn_f16` — Prepare lossless f16 representations of adjusted byte integers, then use vector workgroup GEMM with exact bounded f32 partials and modular int32 totals. Geometry follows output rows, invocation and storage limits; f32 storage requires enough row reuse to amortize preparation. |
| - `sgmat_precast_a0_bcolumn` — Prepare exact f32 representations of adjusted byte integers once, then load them directly into f32-accumulating subgroup matrices. Row groups follow matrix geometry, output rows, invocation limits and workgroup storage; bounded partial sums retain exact modular int32 arithmetic. |
| - `portable_precast_a0_bcolumn` — Prepare lossless f32 representations of adjusted byte integers, then use vector workgroup GEMM with exact bounded f32 partials and modular int32 totals. Geometry follows output rows, invocation and storage limits; f32 storage requires enough row reuse to amortize preparation. |
| - `sgmat_precast_a1_bnone_f16` — Prepare exact f16 representations of adjusted byte integers once, then load them directly into f32-accumulating subgroup matrices. Row groups follow matrix geometry, output rows, invocation limits and workgroup storage; bounded partial sums retain exact modular int32 arithmetic. |
| - `portable_precast_a1_bnone_f16` — Prepare lossless f16 representations of adjusted byte integers, then use vector workgroup GEMM with exact bounded f32 partials and modular int32 totals. Geometry follows output rows, invocation and storage limits; f32 storage requires enough row reuse to amortize preparation. |
| - `sgmat_precast_a1_bnone` — Prepare exact f32 representations of adjusted byte integers once, then load them directly into f32-accumulating subgroup matrices. Row groups follow matrix geometry, output rows, invocation limits and workgroup storage; bounded partial sums retain exact modular int32 arithmetic. |
| - `portable_precast_a1_bnone` — Prepare lossless f32 representations of adjusted byte integers, then use vector workgroup GEMM with exact bounded f32 partials and modular int32 totals. Geometry follows output rows, invocation and storage limits; f32 storage requires enough row reuse to amortize preparation. |
| - `sgmat_precast_a1_bscalar_f16` — Prepare exact f16 representations of adjusted byte integers once, then load them directly into f32-accumulating subgroup matrices. Row groups follow matrix geometry, output rows, invocation limits and workgroup storage; bounded partial sums retain exact modular int32 arithmetic. |
| - `portable_precast_a1_bscalar_f16` — Prepare lossless f16 representations of adjusted byte integers, then use vector workgroup GEMM with exact bounded f32 partials and modular int32 totals. Geometry follows output rows, invocation and storage limits; f32 storage requires enough row reuse to amortize preparation. |
| - `sgmat_precast_a1_bscalar` — Prepare exact f32 representations of adjusted byte integers once, then load them directly into f32-accumulating subgroup matrices. Row groups follow matrix geometry, output rows, invocation limits and workgroup storage; bounded partial sums retain exact modular int32 arithmetic. |
| - `portable_precast_a1_bscalar` — Prepare lossless f32 representations of adjusted byte integers, then use vector workgroup GEMM with exact bounded f32 partials and modular int32 totals. Geometry follows output rows, invocation and storage limits; f32 storage requires enough row reuse to amortize preparation. |
| - `sgmat_precast_a1_bcolumn_f16` — Prepare exact f16 representations of adjusted byte integers once, then load them directly into f32-accumulating subgroup matrices. Row groups follow matrix geometry, output rows, invocation limits and workgroup storage; bounded partial sums retain exact modular int32 arithmetic. |
| - `portable_precast_a1_bcolumn_f16` — Prepare lossless f16 representations of adjusted byte integers, then use vector workgroup GEMM with exact bounded f32 partials and modular int32 totals. Geometry follows output rows, invocation and storage limits; f32 storage requires enough row reuse to amortize preparation. |
| - `sgmat_precast_a1_bcolumn` — Prepare exact f32 representations of adjusted byte integers once, then load them directly into f32-accumulating subgroup matrices. Row groups follow matrix geometry, output rows, invocation limits and workgroup storage; bounded partial sums retain exact modular int32 arithmetic. |
| - `portable_precast_a1_bcolumn` — Prepare lossless f32 representations of adjusted byte integers, then use vector workgroup GEMM with exact bounded f32 partials and modular int32 totals. Geometry follows output rows, invocation and storage limits; f32 storage requires enough row reuse to amortize preparation. |
| - `sgmat_exact_a0_bnone` — Exact integer products through bounded f32 subgroup-matrix partials accumulated with modular int32 additions. Tile height follows workgroup memory and invocation limits. |
| - `sgmat_exact_a0_bscalar` — Exact integer products through bounded f32 subgroup-matrix partials accumulated with modular int32 additions. Tile height follows workgroup memory and invocation limits. |
| - `sgmat_exact_a0_bcolumn` — Exact integer products through bounded f32 subgroup-matrix partials accumulated with modular int32 additions. Tile height follows workgroup memory and invocation limits. |
| - `sgmat_exact_a1_bnone` — Exact integer products through bounded f32 subgroup-matrix partials accumulated with modular int32 additions. Tile height follows workgroup memory and invocation limits. |
| - `sgmat_exact_a1_bscalar` — Exact integer products through bounded f32 subgroup-matrix partials accumulated with modular int32 additions. Tile height follows workgroup memory and invocation limits. |
| - `sgmat_exact_a1_bcolumn` — Exact integer products through bounded f32 subgroup-matrix partials accumulated with modular int32 additions. Tile height follows workgroup memory and invocation limits. |
| - `dp4a_rank2_b_zero_point_per_column` — Packed-dot rank-2 matmul with per-column B zero points. Large prefill tiles use twice as many output-column lanes on variable 16–32-lane devices with sufficient workgroup capacity. |
| - `dp4a_rank2_a_zero_point_b_per_column` — Packed-dot rank-2 matmul with per-column B zero points. Large prefill tiles use twice as many output-column lanes on variable 16–32-lane devices with sufficient workgroup capacity. |
|
|
| ## Device requirements |
|
|
| Some implementation variants require `subgroup-matrix`, `shader-f16`, and `subgroups`. These are route-specific capabilities, not package-wide requirements; availability also depends on the request shape and dtype. |
|
|
| ## Files |
|
|
| - [`metadata.json`](build/webgpu/metadata.json) — kernel metadata (id, digests, per-variant templates, provenance) |
| - [`manifest.json`](build/webgpu/manifest.json) — the op contract (source of truth) |
| - [`test.json`](build/webgpu/test.json) — correctness cases |
| - [`bench.json`](build/webgpu/bench.json) — benchmark cases |
| - [`matmul-integer-batched.wgsl.jinja`](build/webgpu/matmul-integer-batched.wgsl.jinja) |
| - [`quant-dp4a-matmul.wgsl.jinja`](build/webgpu/quant-dp4a-matmul.wgsl.jinja) |
| - [`quant-exact-matrix.wgsl.jinja`](build/webgpu/quant-exact-matrix.wgsl.jinja) |
| - [`quant-exact-portable.wgsl.jinja`](build/webgpu/quant-exact-portable.wgsl.jinja) |
| - [`quant-exact-prepare.wgsl.jinja`](build/webgpu/quant-exact-prepare.wgsl.jinja) |
| - [`quant-matmul-accumulate-rank2.wgsl.jinja`](build/webgpu/quant-matmul-accumulate-rank2.wgsl.jinja) |
| - [`quant-matmul-accumulate-rank4.wgsl.jinja`](build/webgpu/quant-matmul-accumulate-rank4.wgsl.jinja) |
|
|
| ## Use with `@huggingface/kernels` |
|
|
| ```sh |
| npm install --save-exact @huggingface/kernels@0.0.1-preview.3 |
| ``` |
|
|
| Required output shapes and logical data types are inferred from the supplied inputs and attributes; result tensors are allocated automatically. |
|
|
| The `version: 1` option selects the published kernel contract; it is independent of any operator opset, contrib `since_version`, or model version. |
| It follows the `v1` branch as fixes land. To pin exact artifact bytes, pass a 40-character commit `revision` instead of `version`. |
|
|
| Replace each `*Data` placeholder with a typed array containing the corresponding input data. |
|
|
| ```js |
| import { getKernel } from "@huggingface/kernels"; |
| |
| const kernel = await getKernel("webgpu-kernels/ai.onnx.MatMulInteger", { version: 1 }); |
| const { y } = await kernel({ a: { data: aData, shape: [1, 1] }, b: { data: bData, shape: [1, 1] } }); |
| ``` |
|
|
