Crate taconite_gaic 

GAIC (Grid Anchor based Image Cropping, VGG16 backbone) on an AMD XDNA NPU, replaying the bundle iron/applications/gaic/export_gaic.py writes, through taconite.

Gaic::features runs the backbone once per image — the 13 VGG16 convs, each one flm.GEMM over an im2col view (see iron/applications/gaic/gaic_npu.py) — and reduces it to the 32-channel 1/16-scale map every crop is scored from; Gaic::score scores any number of candidate boxes against it (RoI + RoD align on the host, the 5184 -> 768 FC on the NPU, 768 -> 128 -> 1 on the host). anchors makes GAIC’s candidate sets and preprocess the network input, both matching GAIC-Pytorch’s demo exactly.

Host glue per conv, in one threaded pass: the GEMM’s output (bf16, pixel-major [H][W + 2][OC], two junk columns a row) -> + bias, ReLU, [2x2 max-pool] -> the next conv’s A source written straight into the shared input buffer. The activation buffers are host_only (uncached) BOs, so rows move through them with whole-row copies.

One Gaic per process: its kernels stay resident as 9 of the NPU’s 16 hardware contexts. Send, not Sync.

Modules

align

RoIAlignAvg / RoDAlignAvg: ports of GAIC-Pytorch’s CPU kernels (untils/{roi,rod}_align/src/*.cpp, forward) followed by the 2x2 stride-1 average pool the *Avg modules apply. f32 throughout, like the C code.

anchors

Candidate crops, as GAIC-Pytorch’s dataset/candidate_generation.py makes them (same f64 arithmetic, same truncating int()), in the network input’s pixels: [x1, y1, x2, y2].

bundle

The bundle iron/applications/gaic/export_gaic.py writes, in the format every IRON bundle shares (read with taconite_bundle): manifest.txt naming the compiled kernels and the layer table, the packed weights, host-side parameters and the self-check’s references in the tensor store. See that script’s docstring for every record and tensor.

preprocess

The network input, as GAIC-Pytorch’s demo.py makes it: resize (short side 256, each side rounded to a multiple of 32) with PIL’s LANCZOS filter, then ToTensor + ImageNet normalisation.

Structs

Features

The reduced feature map of one image: [reddim][h][w] f32 at 1/16 of the network input (input_w x input_h), what every box is scored from.

Gaic

Timing

Wall time per stage, accumulated since the last Gaic::reset_timing.

Enums

Error