Fixed-point Softmax() calls exp_on_negative_values() twice

[cousteaulecommandant]

Fixed-point Softmax() calls exp_on_negative_values() twice per element: once to calculate the value of the exp of each element

tflite-micro/tensorflow/lite/kernels/internal/reference/softmax.h

Line 131 in ef64591

FixedPoint0 exp_in_0 = exp_on_negative_values(scaled_diff_f8);

and another one a few lines earlier to calculate the sum of all exps

tflite-micro/tensorflow/lite/kernels/internal/reference/softmax.h

Lines 109 to 110 in ef64591

	sum_of_exps = sum_of_exps + gemmlowp::Rescale<kAccumulationIntegerBits>(
	exp_on_negative_values(scaled_diff_f8));

The exp_on_negative_values() function is rather slow, taking most of the time of Softmax(), so calling it twice per element makes the whole Softmax() function be about twice as slow as necessary.

Could this be optimized so that it is only called once per element? Store the exp results in a temporary array (for example, reusing output_data), computing the sum as they're stored, and then applying the scaling to the precomputed exp results rather than computing them again.