Floating-Point Adder Normalization Circuit for Denormal Operand Handling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing processor architectures face performance overhead and increased power consumption when handling denormal floating-point numbers, as they often require additional circuits for detection and normalization, which can increase silicon area and latency, especially since denormal numbers are rarely encountered.
Innovation Solution
The processor speculatively converts floating-point multiply instructions into multiply-add instructions, utilizing the normalization circuit of a floating-point adder to normalize or denormalize products by forcing an addend of zero, allowing the existing hardware to handle unnormal or denormal outputs without additional latency or overhead.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If additional denormal detection and normalization circuits are added to handle denormal numbers, then support for denormal floating-point numbers is improved, but silicon area and circuit complexity increase
Solution Approach 1:
The floating-point adder unit is designed to perform multiple functions: it can execute both floating-point addition operations and floating-point multiplication operations. By making the adder unit universal, the patent eliminates the need for separate dedicated denormal handling circuits, as the same adder unit that handles normalization for addition can also handle normalization for multiplication results, thereby supporting denormal numbers without increasing overall circuit complexity
Solution Approach 2:
The patent merges the denormal handling functionality into the existing floating-point adder unit. Instead of adding separate detection and normalization circuits for multiplication, the design combines these functions with the adder unit's existing normalization capability. This merging approach allows denormal support to be achieved using shared hardware resources, reducing the need for additional dedicated circuits
2Reliability
If additional denormal detection and normalization circuits are added, then support for denormal floating-point numbers is improved, but latency and throughput delay increase
Solution Approach 1:
The patent performs preliminary detection of denormal conditions by examining the exponent fields of the input operands before the actual multiplication computation. By detecting potential denormal results in advance (when both operands have exponents indicating denormal values), the system can prepare the adder unit accordingly, allowing the normalization process to begin immediately after multiplication without adding extra latency during the critical computation path
Solution Approach 2:
The floating-point adder unit serves dual purposes: it performs both floating-point addition and floating-point multiplication operations, including the normalization of multiplication results. This multi-functionality eliminates the need for separate dedicated normalization circuits that would add latency, as the same hardware unit handles normalization for both addition and multiplication operations
3Reliability
If additional denormal handling circuits are added, then support for denormal floating-point numbers is improved, but power consumption increases
Solution Approach 1:
The floating-point adder unit is designed to perform multiple functions including floating-point addition, floating-point multiplication, and normalization of multiplication results. By making the adder unit universal, the patent eliminates the need for separate dedicated denormal handling circuits, thereby reducing overall power consumption while maintaining support for denormal floating-point numbers
Solution Approach 2:
The patent merges denormal handling functionality into the existing floating-point adder unit. Instead of adding separate power-consuming circuits for denormal detection and normalization, the design combines these functions with the adder unit's existing capabilities, allowing denormal support to be achieved using shared hardware resources with minimal additional power overhead
4Device complexity
If denormal numbers are represented as zero, then power consumption and circuit complexity are reduced, but accuracy is lost
Solution Approach 1:
The patent extracts the denormal handling functionality from separate dedicated circuits and integrates it into the existing floating-point adder unit. By taking out the need for separate detection and normalization circuits and incorporating these functions into the universal adder, the design maintains accuracy for denormal numbers while reducing overall circuit complexity
Solution Approach 2:
The floating-point adder unit is designed to perform multiple functions including handling denormal multiplication results. By making the adder unit universal and capable of detecting and normalizing denormal values, the patent maintains measurement precision for denormal numbers without requiring additional dedicated circuits that would increase complexity
Data Source
AI summary
In a denormal support mode, the normalization circuit of a floating-point adder is used to normalize or denormalized the output of a floating-point multiplier. Each floating-point multiply instruction is speculatively converted to a multiply-add instruction, with the addend forced to zero. This preserves the value of the product, while normalizing or denormalizing the product using the floating-point adder's normalization circuit. When the operands to the multiply operation are available, they are inspected. If the operands will not generate an unnormal intermediate product or a denormal final product, the add operation is suppressed, such as by operand-forwarding. Additionally, each non-fused floating-point multiply-add instruction is replaced with a multiply-add instruction having a zero addend, and a floating-point add instruction having the addend of the original multiply-add instruction is inserted into the instruction stream. Upon inspection of the operands, if an unnormal intermediate result or a denormal final result will not occur, the addend may be restored to the multiply-add instruction and the add instruction converted to a NOP.