activation: use dedicated fixed-point λ for SeLU (fixes #1287) #1296
Conversation
…arning#1287) * Replaced literal cast 1.050700987… → `res_T` with a `static const ap_fixed<16,6> lambda`, preserving range and ~1.5 × 10⁻² LSB precision regardless of user-chosen `res_T`. * Removed redundant datareg scope; now a single per-element branch: if (x ≥ 0) y = λ · x else y = selu_table[idx] (with index clamped to [0,N-1]). * Guard against negative-index underflow; behaviour for <0 inputs unchanged. * Keeps identical latency/area on Vivado 2023.1 & Vitis 2024.1, but eliminates silent overflow/rounding when `res_T` is narrow (e.g., ap_fixed<8,2>). Fixes fastmachinelearning#1287
|
Why use 6 integer bits to represent "1.0507"? |
Yeah, my bad. It was late and I must have confused with another project of mine. I'm updating it to use 2 integer bits. |
|
One could also use ufixed since the value is positive, to not use a sign bit. |
|
Can you also fix the pre-commit issue? (I think it's trailing whitespace) |
| @@ -717,9 +717,9 @@ template <class data_T, class res_T, typename CONFIG_T> void selu(data_T data[CO | |||
| initialized = true; | |||
| } | |||
|
|
|||
| typedef ap_fixed<16,2> selu_const_t; // ±512 range, ≈1.5e-2 LSB | |||
| typedef ap_ufixed<16,2> selu_const_t; | |||
There was a problem hiding this comment.
Would just 1 integer bit be enough?
There was a problem hiding this comment.
yeah, kept one more just to be sure, do you want me to bring it to 1?
There was a problem hiding this comment.
I think 1 makes sense, since the number is in the range.
|
For pre-commit, add the pre-commit package to your python setup and do |
|
guys it still finds trailing whitespaces, but I can't find them, any suggestions? |
valerioedu
changed the title
valerioedu
changed the title
|
guys I'm opening a new PR since this one is getting messy |
res_Twith astatic const ap_fixed<16,2> lambda, preserving range and ~1.5 × 10⁻² LSB precision regardless of user-chosenres_T.else y = selu_table[idx] (with index clamped to [0,N-1]).
eliminates silent overflow/rounding when
res_Tis narrow (e.g., ap_fixed<8,2>).Fixes #1287