DSP block for implementing large multiplier on a programmable integrated circuit device

Description

BACKGROUND OF THE INVENTION

This invention relates to programmable integrated circuit devices—e.g., programmable logic devices (PLDs), and, more particularly, to the use of specialized processing blocks which may be included in such devices to perform large multiplications.

As applications for which PLDs are used increase in complexity, it has become more common to design PLDs to include specialized processing blocks in addition to blocks of generic programmable logic resources. Such specialized processing blocks may include a concentration of circuitry on a PLD that has been partly or fully hardwired to perform one or more specific tasks, such as a logical or a mathematical operation. A specialized processing block may also contain one or more specialized structures, such as an array of configurable memory elements. Examples of structures that are commonly implemented in such specialized processing blocks include: multipliers, arithmetic logic units (ALUs), barrel-shifters, various memory elements (such as FIFO/LIFO/SIPO/RAM/ROM/CAM blocks and register files), AND/NAND/OR/NOR arrays, etc., or combinations thereof.

One particularly useful type of specialized processing block that has been provided on PLDs is a digital signal processing (DSP) block, which may be used to process, e.g., audio signals. Such blocks are frequently also referred to as multiply-accumulate (“MAC”) blocks, because they include structures to perform multiplication operations, and sums and/or accumulations of multiplication operations.

For example, the STRATIX® III PLD sold by Altera Corporation, of San Jose, Calif., includes DSP blocks, each of which includes the equivalent of four 18-bit-by-18-bit multipliers. Each of those DSP blocks also includes adders and registers, as well as programmable connectors (e.g., multiplexers) that allow the various components to be configured in different ways. In each such block, the multipliers can be configured as two operations each involving two 18-bit-by-18-bit multipliers, or as one larger (36-bit-by 36-bit) multiplier. In addition, one 18-bit-by-18-bit complex multiplication (which decomposes into two 18-bit-by-18-bit multiplication operations for each of the real and imaginary parts) can be performed. Also, up to two individual 18-bit-by-18-bit multiplications can be performed. Moreover, the two members of any pair of multipliers can be configured for multiplications smaller than 18-bit-by-18-bit.

Although such a DSP block may be configured as a multiplier as large as 36-bit-by-36-bit, a user may want to create a larger multiplier. For example, while a 36-bit-by-36-bit multiplier will support 25-bit-by-25-bit single-precision multiplication under the IEEE 754-1985 standard, it is too small for double-precision multiplication, or for 36-bit-by-36-bit complex multiplication. While the multipliers from several DSP blocks can be used together to implement double-precision multiplication, or larger complex multiplication, the logic needed to interconnect the multipliers has heretofore been programmed by the user in the general-purpose programmable logic outside the DSP block, making it slow and less efficient, and consuming general-purpose resources that might be put to other uses. Moreover, such architectures have relied on two or more carry-propagate operations to arrive at the final product, and a carry-propagate adder is a relatively slow adder configuration.

SUMMARY OF THE INVENTION

The present invention relates to specialized processing blocks for programmable integrated circuit devices such as PLDs, which are provided with links to adjacent blocks, allowing multiple blocks to be combined for larger operations than can be performed within any single specialized processing block, reducing or eliminating reliance on general-purpose programmable resources of the device.

In accordance with this invention, large multipliers may be implemented by connecting or chaining multiple fine-grained DSP blocks—i.e., DSP blocks that individually perform relatively small multiplications, but not smaller than might be useful by itself. Preferably, the DSP blocks to be connected are physically adjacent to one another to minimize propagation delay. Within the DSP blocks, most multiplication and addition operations are performed with carry-save adders. The connections between adjacent DSP blocks preferably also make use of the carry-save adder signal bus. A carry-propagate adder, which is slower, may be used for generating the final output of the chain of DSP blocks.

Thus, in accordance with the present invention, there is provided a specialized processing block for a programmable integrated circuit device having a plurality of specialized processing blocks. The specialized processing block includes multiplier circuitry that performs at least one multiplication and provides partial sum/carry signals for each of the at least one multiplication, a chain output for propagating partial sum/carry signals to another specialized processing block, a chain input for receiving partial sum/carry signals propagated from another specialized processing block, and combining circuitry that combines the partial sum/carry signals for each multiplication and any partial sum/carry signals propagated from another specialized processing block, for propagation to the output.

A programmable logic device incorporating such specialized processing blocks also is provided.

BRIEF DESCRIPTION OF THE DRAWINGS

Further features of the invention, its nature and various advantages, will be apparent upon consideration of the following detailed description, taken in conjunction with the accompanying drawings, in which like reference characters refer to like parts throughout, and in which:

FIG. 1 is schematic representation of a DSP block according to a preferred embodiment of the present invention;

FIG. 2 is a representation of the decomposition of a 54-bit-by-54-bit multiplication into 18-bit-by-18-bit multiplications;

FIGS. 3a and 3b, hereinafter referred to collectively as FIG. 3, are a schematic representation of the concatenation of a plurality of DSP blocks in accordance with an embodiment of the present invention to perform the decomposed multiplications of FIG. 2;

FIG. 4 is a representation of the decomposition of a 36-bit-by-18-bit multiplication into 18-bit-by-18-bit multiplications;

FIG. 5 is a representation of the decomposition of the addition of two 36-bit-by-36-bit multiplications into 18-bit-by-18-bit multiplications; and

FIG. 6 is a simplified block diagram of an illustrative system employing a programmable logic device incorporating the present invention.

DETAILED DESCRIPTION OF THE INVENTION

A DSP block in accordance with the present invention may include one or more N-bit-by-N-bit multipliers. Such a DSP block may include the following functional blocks:

• • 1. Partial product generation, preferably based on Booth recoding;
  • 2. A respective carry-save adder for partial products from each multiplier;
  • 3. A further carry-save adder for partial sum/carry from all multipliers within the DSP block, as well as the chained-in partial sum/carry from any adjacent DSP block; and
  • 4. A carry-propagate adder for the final output.

Such a DSP architecture provides the logic to generate partial sum/carry from multipliers of the aforementioned relatively small operand sizes, and to add the partial sum/carry from the current DSP block and an adjacent DSP block. Because the chain-in from any adjacent DSP block is based on partial sum/carry signals from a carry-save adder, the propagation delay to add the chain-in results is that contributed by one 4-to-2 compressor (or about 1.5 times of the delay of a full adder). In contrast, previous blocks relied on output signals from a carry-propagate adder, or possibly a carry-lookahead adder, for the chain-in results. In the case of a carry-propagate adder, the delay includes the combined delay of a full adder in every block in the chain, and while the delay for a carry-lookahead adder is smaller, it is still much larger than the delay of a full adder.

The trade-off for the reduction in delay using a carry-save adder is that the overall area of a DSP block based on carry-save adders for chaining is about slightly larger than previous DSP blocks relying on carry-propagate adders for chaining. For example, in the case of the STRATIX® family of programmable logic devices available from Altera Corporation, of San Jose, Calif., the increase in area is about 1.5%.

The fast interconnection based on the partial sum/carry signals allows the creation of very large multipliers with high performance from multiple DSP blocks. In accordance with the invention, a 54-bit-by-54-bit multiplier implemented using 18-bit-by-18-bit multiplier modules may be significantly faster than a comparable multiplier created from DSP blocks based on carry-propagate adders. Moreover, compared with other previously-known architectures in which the final adder is implemented in programmable logic of the PLD core, the present invention may provide a substantial f_max improvement, in addition to saving substantial programmable logic resources and interconnection resources in the PLD core.

Performing multiplication operations without utilizing the PLD core fabric can save a significant amount of power. In one example of a 54-bit-by-54-bit multiplier, assuming that power consumption is proportional to block area for the same power density, and even assuming the power density of a DSP block is four times that of programmable logic in the PLD core, and that the DSP block area according to the present invention increases by about 10% over the area of the previously-known DSP blocks (including the area added by the carry-save adders as well as additional area added by other circuitry such as shifters and multiplexers), then the power consumption of a 54-bit-by-54-bit multiplier configured according to the present invention may nevertheless be reduced by about 50% by the elimination of the use of programmable logic.

In a fine-grained DSP architecture according to the invention, each DSP block may be small enough to fit into a single row of logic blocks in the PLD floorplan. That enables row-redundancy support of more of the DSP blocks than in devices with multi-row DSP blocks, enabling an improvement in yield. In addition, bypass multiplexers used to turn off row redundancy for rows where there are multi-row DSP blocks can be eliminated, reducing delay in a critical path of DSP system. This can enhance f_max of the DSP system.

A fine-grained DSP architecture better reflects that not all PLD users require large multiplying operations. According to some estimates, 90% of existing DSP designs utilize operands of 18-bit-by-18-bit or smaller, while 27% utilize operands of 9-bit-by-9-bit or smaller. That means that in a large DSP module with a 36-bit-by-36-bit multiplier, the DSP utilization rate is only about 29%. According to the present invention, assuming a fine-grained multiplier size of 18-bit-by 18, the DSP utilization rate can be close to 80% while still supporting 36-bit-by-36-bit and even larger multiplications. In terms of die area, implementing 18-bit-by-18-bit multipliers in a coarse-grained 36-bit-by-36-bit architecture may consume about four times as much area as implementing the same multipliers in a fine-grained 18-bit-by-18-bit architecture.

A fine-grained architecture also allows an increase in the DSP density by eliminating odd remaining rows that cannot fit a large DSP block. For example, in some large PLD devices, using single-row DSP blocks can improve the DSP block count by up to 9% (depending on the total number of rows) as compared to using DSP blocks that span four rows.

Because no core logic circuits or interconnection routings are needed in the construction of large multipliers using the present invention, more PLD fabric resources are available for other uses, and congestion may be reduced leading to better fitting of user logic designs. The architecture according to the invention is fully scalable and therefore applicable to PLDs regardless of density. In fact, higher-density devices would have more adjacent DSP modules, and therefore would support even larger multipliers.

There is no theoretical limit on chaining, and therefore the only physical limit is the number of adjacent DSP blocks in one column. For example, in the largest model of the PLD product sold by Altera Corporation, of San Jose, Calif., under the name STRATIX® IV, which has 138 rows of logic blocks, there are 138 adjacent DSP blocks in a column. In accordance with the present invention, such a device, which has 276 adjacent 18-bit-by-18-bit multipliers per DSP block column, could support a multiplier as large as 272×18×18, broken down to (17×18)-by-(16×18) or 306-bits-by-288-bits.

A DSP block according to the present invention includes at least one N-bit-by-N-bit multiplier. A preferred embodiment of a DSP block 100, shown in FIG. 1, includes two N-bit-by-N-bit multipliers, each implemented using a respective carry-save adder 101, 102. The choice of N should be large enough to be useful but small enough to allow DSP block 100 to fit into one row of PLD logic blocks. If, as estimated above, about 90% of DSP designs use multiplications of size 18-bit-by-18-bit or smaller, and about 27% are 9-bit-by-9-bit or smaller, then 18-bit-by-18-bit multipliers, which can be configured as smaller multipliers as well (in addition to being combined for larger multiplications) are a reasonable selection, and therefore in this example N=18.

The operands 103 can be entered asynchronously, or can be registered in registers 104, under control of multiplexers 105. If each multiplier is to be used independently, the respective output of respective CSA 101, 102 can be selected by a respective one of multiplexers 111, 112. Compressor 106 is provided to allow the two multiplications to be combined. Compressor 106 may include shifting, multiplexing and inverting inputs 116, as shown. If compressor 106 is used, it may be selected by multiplexers 111, 112. An additional compressor 107 is provided in accordance with the invention to allow chain-in input 108 from an adjacent one of DSP blocks 100 to be combined with the output of the current operations as processed through compressor 106. Again, if compressor 107 is used, it may be selected by multiplexers 111, 112. The outputs of compressors 106, 107 also may be selected by multiplexer 121 as chain-out output 122.

Carry-propagate adders 109, 110 are provided to combine the current carry-save output with carry-in input 113 from a previous one (if any) of DSP blocks 100, to provide partial final output 114, and, unless this is the final DSP block 100 in the chain, a carry-out output 115 which will serve as a carry-in input 113 to a subsequent one of DSP blocks 100. This allows the carry-propagate adders from the various DSP blocks 100 in the chain to be used as a single larger carry-propagate adder while propagating only carry-out 115, rather than both carry-out 15 and sum 114, to subsequent DSP blocks 100.

Both chain-in input 108 and carry-in input 113 may bypass DSP block 100 completely, and be output directly to chain-out output 122 and carry-out output 115, respectively, under control of respective multiplexers 123, 124. This allows for redundancy in the event that DSP block 100 is defective and must be completely bypassed, allowing the chain and carry signals to propagate to the next one of DSP blocks 100.

Registered output 130 may routed back to compressor 107, where it may be selected by multiplexer 131 to enable an accumulator function. Signal 130 also may be routed to multiplexer 121 to enable pipeline operation in large multipliers. This allows the output register 132 to serve as a pipeline register when pipelining is used, rather than adding an additional register for that purpose downstream of multiplexer 121 (whether inside or outside of DSP block 100).

In this configuration, the two N-bit-by-N-bit multipliers may be implemented in one DSP block, and they can be programmed as two independent multipliers. They also can work together to maximize the performance of two-multiplier operations, and provide a high degree of flexibility in building larger multipliers. The partial sum and partial carry vectors from the compressor blocks may be chosen for the chain-out signals to minimize the use of the carry-propagate adders and the associated impacts on delay and power. Output registers may be provided after the compressors to improve performance for large multipliers, and to assist in the construction of filters, such as a finite impulse response (FIR) filter. The user can trade output latency with higher clock frequency by using the registers to create pipeline stages.

For a multiplication operation, SxT, where one or both of the operands (S and/or T) are wider than N bits, the operands can be extended at the most significant bit position to a multiple of N, and divided in segments of N bits each. The product can be written as the sum of outputs of a plurality of N-bit-by-N-bit multipliers. For example, where S and T are both 54 bits wide (as in the case of double-precision floating point operations), each of S and T can be broken into three respective 18-bit segments A3|A2|A1 and B3|B2|B1.

The multiplication operation can be decomposed as shown in FIG. 2 into nine partial products A1×B1, A2×B1, etc., which can be performed by nine 18-bit-by-18-bit multipliers and combined, after appropriate left-shifting by appropriate multiples of 18 bits (representing corresponding powers of 2¹⁸), to form 108-bit product 200. The operation can be written as:

P ⁡ [ 107 : 0 ] = ⁢ { A ⁢ ⁢ 3 | A ⁢ ⁢ 2 | A ⁢ ⁢ 1 } × { B ⁢ ⁢ 3 ❘ B ⁢ ⁢ 2 ❘ B ⁢ ⁢ 1 } = ⁢ A ⁢ ⁢ 3 × B ⁢ ⁢ 3 × 2 72 + ( A ⁢ ⁢ 3 × B ⁢ ⁢ 2 + A ⁢ ⁢ 2 × B ⁢ ⁢ 3 ) × 2 54 + ( A ⁢ ⁢ 3 × B ⁢ ⁢ 1 + ⁢ A ⁢ ⁢ 2 × B2 + A ⁢ ⁢ 1 × B ⁢ ⁢ 3 ) × 2 36 + ( A ⁢ ⁢ 2 × B ⁢ ⁢ 1 + A ⁢ ⁢ 1 × B ⁢ ⁢ 2 ) × 2 18 + ⁢ A ⁢ ⁢ 1 × B ⁢ ⁢ 1

Using DSP block 100 of FIG. 1 with N=18, which includes two 18-bit-by-18-bit multipliers, the nine multiplications can be performed using five DSP blocks 100 (with one multiplier left over) as shown schematically in FIG. 3. FIG. 2 is annotated to show which of the DSP blocks 100 in FIG. 3 (identified as DSP #1, DSP #2, etc.) is used for which partial products. In each of DSP #1 through DSP #4, the least significant 18 bits go through the carry-propagate adder stage 300 to generate part of the final product while the rest of the partial sum and partial carry vectors are sent to the next DSP block 100 through the chain-out/chain-in bus, as follows:

Throughout the operation of the foregoing 54-bit-by-54-bit multiplier, carry-save signal buses are used, and there is only one carry-propagate operation for the final outputs, through 18 bits in each of DSP #1, #2, #3 and #4, and 36 bits in DSP #5. This does not need logic or routing resources in PLD logic core.

It should be noted that other operations also may be performed. For example, 36-bit-by 18, 36-bit-by 36, 54-bit-by-18-bit and 54-bit-by-36-bit multiplications, as well as a sum of four 18-bit-by-18-bit multiplications or two 36-bit-by-36-bit multiplications, all can be performed in a single column of chained DSP blocks 100.

Thus it is seen that the present invention provides a generalized way to build large multipliers from a basic DSP block, with only one carry-propagate-adder operation is required for any size of operands. In the embodiment shown, all operations involving two multipliers may be performed efficiently within a single DSP block, such as addition or subtraction of two N-bit-by-N-bit multiplications, or a 2N-bit-by-N-bit multiplication. The latter is diagrammed in FIG. 4, which shows the decomposition of such an operation into two N-bit-by-N-bit multiplications. Other mixtures of multiplications and additions may be performed when multiple DSP blocks are concatenated together through the chain and carry connections, such as addition or subtraction of three or more N-bit-by-N-bit multiplications, addition or subtraction of two 2N-bit-by-2N-bit multiplications. FIG. 5 shows how the addition of two 2N-bit-by-2N-bit multiplications may be decomposed into eight N-bit-by-N-bit multiplications, which can be performed in four DSP blocks.

The invention eliminates the need to use logic circuitry or routing in the PLD logic core in the construction of large multipliers. This reduces power dissipation in the PLD, avoids routing congestion in the logic core, and maintains the maximum speed allowed within the DSP block. The modular construction of large multipliers also is more amenable to row-based redundancy protection in the PLD, and avoids the overhead of providing dedicated multipliers for large operand width in an PLD, as the width of operands is limited only by the number of adjacent DSP blocks with chain and carry connections.

Thus it is seen that a large multiplication that requires more than one specialized processing block of a PLD can be performed using fewer or no general-purpose programmable resources of the PLD.

A PLD 280 incorporating such circuitry according to the present invention may be used in many kinds of electronic devices. One possible use is in a data processing system 900 shown in FIG. 6. Data processing system 900 may include one or more of the following components: a processor 281; memory 282; I/O circuitry 283; and peripheral devices 284. These components are coupled together by a system bus 285 and are populated on a circuit board 286 which is contained in an end-user system 287.

System 900 can be used in a wide variety of applications, such as computer networking, data networking, instrumentation, video processing, digital signal processing, or any other application where the advantage of using programmable or reprogrammable logic is desirable. PLD 280 can be used to perform a variety of different logic functions. For example, PLD 280 can be configured as a processor or controller that works in cooperation with processor 281. PLD 280 may also be used as an arbiter for arbitrating access to a shared resources in system 900. In yet another example, PLD 280 can be configured as an interface between processor 281 and one of the other components in system 900. It should be noted that system 900 is only exemplary, and that the true scope and spirit of the invention should be indicated by the following claims.

Various technologies can be used to implement PLDs 280 as described above and incorporating this invention.

It will be understood that the foregoing is only illustrative of the principles of the invention, and that various modifications can be made by those skilled in the art without departing from the scope and spirit of the invention. For example, the various elements of this invention can be provided on a PLD in any desired number and/or arrangement. One skilled in the art will appreciate that the present invention can be practiced by other than the described embodiments, which are presented for purposes of illustration and not of limitation, and the present invention is limited only by the claims that follow.