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 ff96319
Browse filesBrowse the repository at this point in the historyBrowse files
[Espressif] Fuse binary ReLU and HardTanh patterns
Quantize single-consumer add, subtract and multiply followed by ReLU or
HardTanh, including ReLU6, at their pattern boundaries so the reference has
no intermediate rounding. Fold the activation into Fused Quant output
qparams and clamp bounds before ESP replacement. Reuse existing shape and
arithmetic eligibility checks with the final output scale; unsupported
producers retain portable execution. Forked and separately quantized
activations keep their original boundaries. Clamp bounds use the producer's
floating dtype to preserve BF16 and FP16 portable fallback results.
Read the quantizer and FoldActivations first, then the binary lowering
and schema/wrapper changes. Shared dialect and S3 implementation cases cover
forks, broadcasts, saturation, fallback and narrower clamp grids, including
asymmetric, positive and negative HardTanh bounds and rounding at endpoints.
Authored with OpenAI Codex.
Copy file name to clipboardExpand all lines: backends/espressif/ops/operators.yaml
+3-3Lines changed: 3 additions & 3 deletions
Original file line number
Diff line number
Diff line change
@@ -10,19 +10,19 @@
10
10
- arg_meta: null
11
11
kernel_name: espressif::quantized_lut_int8_out
12
12
13
-
- func: esp_nn::quantized_mul_int8.out(Tensor input1, Tensor input2, int input1_offset, int input2_offset, int output_offset, int output_multiplier, int output_shift, *, Tensor(a!) out) -> Tensor(a!)
13
+
- func: esp_nn::quantized_mul_int8.out(Tensor input1, Tensor input2, int input1_offset, int input2_offset, int output_offset, int output_multiplier, int output_shift, int activation_min, int activation_max, *, Tensor(a!) out) -> Tensor(a!)
14
14
variants: function
15
15
kernels:
16
16
- arg_meta: null
17
17
kernel_name: espressif::quantized_mul_int8_out
18
18
19
-
- func: esp_nn::quantized_mul_broadcast_int8.out(Tensor input1, Tensor input2, int input1_offset, int input2_offset, int output_offset, int output_multiplier, int output_shift, *, Tensor(a!) out) -> Tensor(a!)
19
+
- func: esp_nn::quantized_mul_broadcast_int8.out(Tensor input1, Tensor input2, int input1_offset, int input2_offset, int output_offset, int output_multiplier, int output_shift, int activation_min, int activation_max, *, Tensor(a!) out) -> Tensor(a!)
- func: esp_nn::quantized_add_int8.out(Tensor input1, Tensor input2, int input1_offset, int input2_offset, int output_offset, int input1_multiplier, int input1_shift, int input2_multiplier, int input2_shift, int left_shift, int output_multiplier, int output_shift, *, Tensor(a!) out) -> Tensor(a!)
25
+
- func: esp_nn::quantized_add_int8.out(Tensor input1, Tensor input2, int input1_offset, int input2_offset, int output_offset, int input1_multiplier, int input1_shift, int input2_multiplier, int input2_shift, int left_shift, int output_multiplier, int output_shift, int activation_min, int activation_max, *, Tensor(a!) out) -> Tensor(a!)
0 commit comments