Published s3/p4 .espdl models have (N/16)HWC16-packed Conv filters that the C/C++ Conv kernels cannot consume #320

Description

The published per-chip .espdl model files (e.g. models/human_face_detect/models/s3/*.espdl, .../p4/*.espdl) ship their Conv filters pre-packed in the chip's SIMD layout — each weight tensor's doc_string carries the tag layout ==> N16HWC16 / N16HWC16_UNALIGNED (esp-ppq reset_conv_filter_layout: aligned rows as [N/16][H][W][C][16], tail rows in plain order). The generic C/C++ Conv kernels (dl_base_conv2d.cpp, comment "filter in sequence [N, H, W, C]"; depthwise [H][W][C]) read the tensor as a plain array and nothing in the runtime unpacks it.

Consequently any build that takes the C path — IDF_TARGET=linux, or a chip without the PIE/TIE kernels — runs the shipped s3/p4 models with scrambled weights. Inference completes without error but the output is garbage: with v3.2.3 on the Linux target, human_face_detect_msr produces the same saturated score map for every input and a black frame "detects a face" at score 0.531. (Two further C-path defects contributed to that symptom; one is filed separately, one is already fixed on master.)

Present at v3.2.3 (7089b94) and on current master (12c0616).

Expected

Either (a) the runtime unpacks packed filter layouts when the C kernels are selected (keyed on the layout tag, which is already in the file), or (b) the packed-layout assumption is documented and the models directory also provides C-layout (ESPDL_C_* target) exports for the published models, or at least the C path refuses packed tensors loudly instead of computing the wrong function silently.

Reproduction

1. Build for IDF_TARGET=linux (C kernels).
2. Load human_face_detect_msr_s8_v1.espdl (s3), feed an all-zero input and a random input — the outputs are statistically identical; score0/score1 are pinned at ±127.
3. Unpack each 4-D Conv filter in the loader to [N][H][W][C] (group 1) / [H][W][C][N] (depthwise), using the esp-ppq packing rule and the node's group attribute. Every Conv node in MSR, MNP and MFN then matches an independent int8 reference bit-for-bit, and the s3 and p4 variants unpack to byte-identical filters for 233 of 234 parameters.

The unpack we used (~100 lines, in our host loader) is essentially the inverse of reset_conv_filter_layout in esp_ppq/parser/espdl/export_patterns.py; we can contribute it as a PR if you'd like it in fbs_loader/dl_module_conv.

Disclosure

This was found and the report written with Claude (Anthropic's AI assistant) working under the direction of Dr Liz James, as part of an assurance project that runs ESP-DL's published face models off-device. Claims above were verified by running the code.