Lines Matching refs:quantized

11 of quantized 8bit values and matrices of real numbers. The choice of a
20 output pipeline, allowing the user to implement their own preferred quantized
36 quantized uint8 values, the mapping between real and quantized uint8 values must
50 mean that the result of the quantized calculations do no longer readily provide
64 In order for such zero-padding to be feasible in a quantized implementation of
66 quantized form, i.e. that it correspond exactly to some quantized value, which
70 quantized value for padding, that does not exactly correspond to the real value
130 *   `zero_point` is integral, more specifically is one of the possible quantized
139 to quantized values (typically uint8), we turn to applying this knowledge to
141 multiplication of matrices of quantized values.
147 Now suppose that we have already quantized these two matrices according to the
150 are quantized as
172 Now our goal is to represent this result itself as a quantized matrix, i.e.
181 Here we need to keep in mind that our goal is to specify what the quantized
191 result_real_value in terms of the quantized operands, and we obtain:
203 "quantized matrix multiplication" should actually compute, in order to be able
206 ## Implementation of quantized matrix multiplication
208 Having obtained the mathematical form (5) of quantized matrix multiplication, we
222 cost goes. Luckily, it only involves integers: the quantized operands matrix
257 as we are typically mapping the `[0..255]` quantized uint8 value range to an
344 Example code showing how to perfom a quantized matrix multiplication in the