| # Native dependency patches |
|
|
| The runtime builds immutable upstream XNNPACK and pthreadpool revisions, then |
| applies the patches in this directory from CMake. These patches are intended |
| to include only functionality required to enable YOLO26n inference on |
| Raspberry Pi 5 using Linux AArch64 and four worker threads. They are not |
| intended to remain useful to arbitrary XNNPACK or pthreadpool consumers. |
|
|
| ## Patch responsibilities |
|
|
| | Patch | Responsibility | |
| | --- | --- | |
| | [`xnnpack-graph-api-pruning.patch`](xnnpack-graph-api-pruning.patch) | Removes graph entry points and implementations that the program cannot reach. It contains no new behavior. | |
| | [`xnnpack-yolo26-graph-extensions.patch`](xnnpack-yolo26-graph-extensions.patch) | Adds SiLU and flags for the three explicit requests through which our fixed graph selects the new implementations. This is the only functional change to XNNPACK's graph layer. | |
| | [`xnnpack-operator-pruning.patch`](xnnpack-operator-pruning.patch) | Removes operator dispatch and microparameter implementations that the Pi 5 runtime cannot reach. It contains no new behavior. | |
| | [`xnnpack-yolo26-operator-extensions.patch`](xnnpack-yolo26-operator-extensions.patch) | Implements the tagged concatenate, registers standalone SiLU, and connects tagged convolution operators to the fused GEMM or indirect-GEMM paths. It selects implementations but does not contain arithmetic kernels. | |
| | [`xnnpack-yolo26-kernels.patch`](xnnpack-yolo26-kernels.patch) | Adds the Arm64 FP16 convolution-plus-SiLU assembly and standalone FP16/FP32 SiLU kernels. | |
| | [`xnnpack-rpi5-bti-diagnostic.patch`](xnnpack-rpi5-bti-diagnostic.patch) | Downgrades XNNPACK's GCC BTI warning because the fixed Raspberry Pi 5 Cortex-A76 target cannot enforce BTI. It does not change compilation. | |
| | [`pthreadpool-spin-wait.patch`](pthreadpool-spin-wait.patch) | Extends the worker spin-wait period used by the fixed four-thread runtime. This reduces wake-up latency between short operators. | |
|
|
| The six XNNPACK patches apply in the table's order, the exact application commands and dependency revisions live in [`../CMakeLists.txt`](../CMakeLists.txt). |
|
|
| ## Specialized concatenate |
|
|
| `BuildP2ToP3` explicitly marks the model.2 concatenate with |
| `XNN_FLAG_YOLO26_FUSED_C2F_CONCATENATE`. XNNPACK accepts that request only for |
| three batch-one NHWC FP16 inputs with 16 channels each and a batch-one NHWC |
| FP16 output with 48 channels. It validates matching spatial dimensions and |
| fails runtime construction instead of silently selecting the specialization |
| for a different contract. |
|
|
| The operator parallelizes over spatial rows and performs the three channel |
| copies together. Its spatial extents are derived during reshape, so the same |
| tagged operation is used at 320, 480, 640, 800, 960, and 1280. Every untagged |
| concatenate retains XNNPACK's ordinary per-input copy path. |
|
|
| ## Custom kernels |
|
|
| The added `f16-vsilu` and `f32-vsilu` files are deliberate |
| hand-maintained kernel derivatives. There is no project-supported |
| generator command for them. Do not run an upstream generator and assume the |
| result retains the fused SiLU behavior; changes must be made in the patch and |
| reviewed as arithmetic changes. |
|
|
| The algorithm, constants, registers, stack frame, stock and fused entry paths, |
| and tail handling are documented beside the assembly that implements them in |
| `xnnpack-yolo26-kernels.patch`. |
|
|
| ## Maintenance rules |
|
|
| - Keep each change with the layer that owns it. Do not put assembly in the |
| operator-extension patch or functional changes in either pruning patch. |
| - Keep implementation contracts beside the changed code in the patch. This |
| README is not a substitute for source comments. |
| - Do not restore general-purpose operators or architectures without a concrete |
| runtime requirement. This build deliberately supports only its fixed target. |
| - Update the pinned dependency revision and its patches together. A patch that |
| no longer applies is a configuration failure requiring review, not something |
| CMake should silently bypass. |
| - Treat dispatch, kernel, and pthreadpool changes as performance-sensitive. |
| Record new Raspberry Pi 5 performance evidence and review it against the |
| relevant reference result. |
| - Treat graph/API changes as topology-sensitive. Re-run the native and public |
| runtime behavior tests at every supported resolution. Use the Ultralytics |
| harness when arithmetic or execution changes require renewed accuracy |
| evidence. |
|
|
| Upstream dependency revisions are pinned in [`../CMakeLists.txt`](../CMakeLists.txt), |
| and their deployment notices are retained under [`../licenses/`](../licenses/). |
|
|
