What is R_MULT used in forward_convolutional_layer_q

[LiangLeon]

Hi
I have learned how to get the weights multiplier and input multiplier.
but I have question of the following code which is in the forward_convolutional_layer_q function
float ALPHA1 = R_MULT / (l.input_quant_multipler * l.weights_quant_multipler);
for (i = 0; i < l.outputs; ++i) {
l.output[i] = output_q[i] * ALPHA1; // cuDNN: alpha1
}
R_MULT is a constant and the value is 32.
Could you give brief explanation about why we need R_MULT and how to set the value?
I really appreciate your time and great work.

[ArtyZe]

Hi
I have learned how to get the weights multiplier and input multiplier.
but I have question of the following code which is in the forward_convolutional_layer_q function
float ALPHA1 = R_MULT / (l.input_quant_multipler * l.weights_quant_multipler);
for (i = 0; i < l.outputs; ++i) {
l.output[i] = output_q[i] * ALPHA1; // cuDNN: alpha1
}
R_MULT is a constant and the value is 32.
Could you give brief explanation about why we need R_MULT and how to set the value?
I really appreciate your time and great work.

In gemm function:

maybe protect from overflow ?
so when dequant output from int to float value, need to mul 32 too