You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 45a0b7a
Browse filesBrowse the repository at this point in the historyBrowse files
Share AvgPool2d input/output observers in get_esp_fused_quant_quantizer and lower supported
static int8 pooling to the optimized ESP-NN kernel through the existing NHWC
layout flow. Register the operator, provide its integer reference, and cover
floor-mode pooling, padding without included pad values, batches, rounding,
large windows, and portable fallback through shared dialect and device cases.
Keep non-square outputs with multiple rows portable to avoid ESP-NN 1.4.1's
output-row-stride bug. Enforce the restriction in the wrapper as well, and
cover wide and tall outputs in both optimized channel paths. The upstream fix
is released in 1.4.2; dependency adoption remains separate.
Honor zero comparison tolerances in the ESP-IDF runner instead of silently
using its defaults. Allow float32 epsilon only for pooling tests with portable
sin branches. Validation passed 325 Python tests, 76 optimized ESP32-S3 emulator
implementation cases across the pooling and binary suites, and scoped lint.
Authored with OpenAI Codex.
0 commit comments