Multi-path fused multiply-add with power control
AssigneeApple Inc.
InventorsLiang-Kai Wang, Ting Yu, Yu Sun
Life of this patent
Estimating the term…
Abstract
Techniques are disclosed relating to circuitry configured to perform floating-point operations such as fused multiply-addition (FMA) with multiple paths and power control. In some embodiments, an FMA unit includes a near path and multiple far paths and is configured to select a path based on a determined exponent difference. In some embodiments, the FMA unit is configured to operate portions of non-selected paths in a low power state.
Claims20
An apparatus, comprising:
- pipelined multiplication circuitry that includes multiple pipeline stages and is configured to generate fused multiply-addition results;
- classification circuitry configured to classify fused multiply-addition operation based on a difference between an exponent of an addend and a sum of exponents of multiplicands;
- wherein the multiplication circuitry includes multiple paths and is configured to select paths for performing operations based on the classification, wherein the paths include:
- near-path circuitry configured to perform operations where the difference is in a first range of values between a first threshold value and a second threshold value;
- first far-path circuitry configured to perform operations where the difference is in a second range of values that does not overlap with the first range of values;
- second far-path circuitry configured to perform operations where the difference is in a third range of values that does not overlap with the first or second ranges of values;
- and
- third far-path circuitry configured to perform operations where the difference is in a fourth range of values that does not overlap with the first, second, or third ranges of values;
- wherein the apparatus is configured to, in response to the apparatus assigning a fused multiply-addition operation to the first far-path circuitry using the classification circuitry, operate at least a portion of the near-path circuitry and at least a portion of the second far-path circuitry in a low-power state.
A non-transitory computer readable storage medium having stored thereon design information that specifies a design of at least a portion of a hardware integrated circuit in a format recognized by a semiconductor fabrication system that is configured to use the design information to produce the circuit according to the design, including:
- pipelined multiplication circuitry that includes multiple pipeline stages and is configured to generate fused multiply-addition results;
- classification circuitry configured to classify fused multiply-addition operation based on a difference between an exponent of an addend and a sum of exponents of multiplicands;
- wherein the multiplication circuitry includes multiple paths and is configured to select paths for performing operations based on the classification, wherein the paths include:
- near-path circuitry configured to perform operations where the difference is in a first range of values between a first threshold value and a second threshold value;
- first far-path circuitry configured to perform operations where the difference is in a second range of values that does not overlap with the first range of values;
- second far-path circuitry configured to perform operations where the difference is in a third range of values that does not overlap with the first or second ranges of values;
- and
- third far-path circuitry configured to perform operations where the difference is in a fourth range of values that does not overlap with the first, second, or third ranges of values
- wherein the circuit is configured to, in response to assigning a fused multiply-addition operation to the first far-path circuitry using the classification circuitry, operate at least a portion of the near-path circuitry and at least a portion of the first second far-path circuitry in a low-power state.
A method, comprising:
- classifying, by classification circuitry, ones of a plurality of fused multiply-addition operations based on a difference between an exponent of an addend and a sum of exponents of multiplicands;
- selecting, by control circuitry, ones of multiple paths to perform ones of the plurality of fused multiply-addition operations, including:
- selecting near-path circuitry in response to the difference being in a first range of values between a first threshold value and a second threshold value;
- selecting first far-path circuitry in response to the difference being in a second range of values that does not overlap with the first range of values;
- selecting second far-path circuitry in response to the difference being in a third range of values that does not overlap with the first or second ranges of values;
- and
- selecting third far-path circuitry in response to the difference being in a fourth range of values that does not overlap with the first, second, or third ranges of values;
- and
- in response to selecting the first far-path circuitry for an operation, operating at least a portion of the near-path circuitry and at least a portion of the first second far-path circuitry in a low-power state.
Dependencies are inferred from the claim text.