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 85163e8
Browse filesBrowse the repository at this point in the historyBrowse files
Pin the delegate, operator-library, and packaged ESP-IDF components to
ESP-NN 1.5.0 for the S3 AvgPool row-stride fix, faster multiply, and
convolution/pooling alignment fixes. Remove the temporary rectangular
AvgPool checks from lowering and the kernel wrapper. Exercise wide and
tall outputs through the shared operator cases and four delegate cases.
All 76 operator-library cases and 20 delegate programs in each of
optimized and ANSI modes pass on the S3 emulator. Generated component
manifest validation and scoped lintrunner pass. Compared with 1.4.2,
matching test firmware grows by 2896 bytes for the operator library and
4512 bytes for the optimized delegate; ANSI shrinks by 224 bytes.
The new S3 multiply frames require 216 additional stack bytes.
Authored with OpenAI Codex.
Copy file name to clipboardExpand all lines: backends/espressif/README.md
+2-5Lines changed: 2 additions & 5 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -24,17 +24,14 @@ Fused Quant frontend convention. The two flows have distinct quantizers:
24
24
|`EspNNQuantizer`| Existing delegate: PT2E Q/DQ graphs partitioned by `EspNNPartitioner` into `EspNNBackend` delegates |
25
25
|`get_esp_fused_quant_quantizer`| Operator library: shared Fused Quant annotation and fusion, followed by `EspPassManager` lowering to direct ESP-NN operators or portable fallback |
26
26
27
-
Both flows use ESP-NN kernels. The operator library currently accelerates
27
+
Both flows use ESP-NN 1.5.0 kernels. The operator library currently accelerates
28
28
int8 tensor add with equal input shapes and multiply with equal shapes or a
29
29
contiguous trailing block broadcast across the full tensor, including
30
30
one-element tensors. It also accelerates values-only 4-D int8 MaxPool2d with matching input
31
31
and output qparams, unit dilation, floor mode, and channel counts divisible by
32
32
four. AvgPool2d uses the same shape and channel restrictions, with shared
33
33
input/output qparams, floor mode, and no divisor override. Padding is supported
34
-
when `count_include_pad=False`. AvgPool2d acceleration currently requires square
35
-
outputs or a single output row because ESP-NN 1.4.1's optimized S3 kernel has
36
-
an output-row-stride bug; other output shapes retain portable execution.
37
-
The upstream fix is available in ESP-NN 1.4.2; this backend still pins 1.4.1.
34
+
when `count_include_pad=False`.
38
35
Its integer kernel rounds ties away from zero; the FQ reference rounds ties to
39
36
even, so results can differ by one int8 step.
40
37
The shared Fused Quant layout pass introduces contiguous NHWC tensors for
0 commit comments