Self-timed Fused Multiply-add Unit Performance Improvement

The paper presents the results of researching and developing methods accelerating a self-timed unit that performs a fused multiply-add-subtract operation under three operands following the IEEE754 standard. The paper proposes a normalization order of the sum and difference speculative estimation when using ternary self-timed coding. This estimation uses a difference of the 24 most significant bits of the product and the third aligned operand for counting leading zeroes. This technique provides a 20% reduction in complexity when implementing the shifter in the normalization block. The analysis shows that the developed methods accelerate the normalization stage by an average of 32% on the statistical set of the input operands.