Digital signal processor configuration including multiplying units coupled to plural accumlators for enhanced parallel mac processing

[0002] Not applicable.

[0003] The present invention relates generally to digital signal processors. In particular, the present invention relates to the association or coupling between accumulators and multiplying units in a programming model for a high performance digital signal processor.

[0004] A variety of types of operations are needed for a digital signal processor (“DSP”) to accomplish desired tasks. These operations are performed on data elements, operands, and the like and typically include mathematical operations, logic operations, shifting operations, and other data manipulation operations. For example, these operations might include adding, subtracting, multiplying, dividing, selecting, combining, arithmetic shifting, logic shifting, and the like. These operations may be identified in program instructions and may be executed by functional units, execution units, processing elements, or the like.

[0005] One such functional unit that is often utilized by DSPs is a multiply-accumulate (“MAC”) unit. MAC units multiply two or more operands together and add the product to a value already stored in an accumulator. This value may be a fraction or integer, real or complex, and may be positive, negative or zero. The accumulator is a register that is at least wide enough to hold the largest product produced by the multiplier. It can be used as a source or a destination register for operations.

[0006] In trying to achieve faster processing while minimizing the physical size of the processor and the power requirements, a problem exists in whether to have accumulator registers dedicated to each of the MAC units or whether to have several accumulator registers available for random use by all of a processor's MAC units. In addition, it is important that the configuration between accumulator registers and MAC units facilitate different types of operations, such as scalar and vector operations, which are typically required by different types of DSP programming models.

[0007] More specifically, there are several types of programming models for digital signal processors which use MAC units. A first type of programming model is an instruction parallelism model which is defined by its ability to simultaneously execute different instructions. This model uses a horizontal approach to parallelism where several instructions are included in a long instruction word that is fetched and executed every cycle. This model may be embodied in a very long instruction word (“VLIW”) model or a super-scalar model, among others. Instruction parallelism models are very effective in telecommunication applications.

[0008] Data parallelism models are a second type of model and may be able to simultaneously execute multiple operations of a single instruction, where each operation can be performed with different data. Data parallelism models utilize vector operations and are embodied in a single instruction multiple data (“SIMD”) model. Data parallelism models are very efficient in block based applications such as image processing, filtering applications and multimedia applications.

[0009] In super-scalar and SIMD processors, typically the accumulator registers are either dedicated to specific MAC units or else a very small number of accumulator registers are available for all MAC units. These configurations may increase the processing time, consume too much physical chip space, or increase power consumption.

[0010] A more efficient configuration between MAC units and accumulator registers is needed, especially for processors that include both horizontal parallelism and vertical parallelism.

[0011] Accordingly, it would be extremely advantageous to develop a configuration between MAC units and accumulator registers which permits the execution of both horizontal operations and vertical operations and balances the need for faster processing with the need to minimize the physical size of the processor and minimize power consumption.

[0012] The above and other advantages are carried out in one form by a digital signal processor which includes a multiply element that multiplies at least two data elements or operands together, resulting in a product. In addition, two or more accumulating elements are operatively associated with or coupled to each multiplying element. Each accumulating element contains a stored value which is replaced by the sum of the stored value and the product from the multiply element.

[0013] A more complete understanding of the present invention may be derived by referring to the detailed description and claims when considered in connection with the Figures, wherein like reference numbers refer to like items throughout the Figures, and:

[0014] FIG. 1 shows a block diagram of an exemplary configuration of a digital signal processor; and

[0015] FIG. 2 shows a block diagram of an exemplary configuration between functional units.

[0016] The present invention relates to the coupling and configuration of multiply-accumulate (“MAC”) units for use within digital signal processors (“DSPs”). In general, a number of accumulators are dedicated to each MAC unit in a manner that enables both scalar and vector operations.

[0017] The present invention may be described herein in terms of functional block components and various processing steps. It should be appreciated that such functional blocks may be realized by any number of components configured to perform specified functions. For example, the present invention may employ various integrated circuit components, which may carry out a variety of functions. In addition, those skilled in the art will appreciate that the present invention may be practiced in any number of data processing contexts and that the digital signal processing configuration described herein is merely one exemplary application for the invention. Further, it should be noted that the present invention may employ any number of conventional techniques for data storage, transport, manipulation, processing, and the like. Such general techniques that may be known to those skilled in the art are not described in detail herein.

[0018] The following U.S. Patent Applications are related to the present invention. More specifically, these U.S. Patent Applications detail, inter alia, exemplary instruction set architectures for permitting a plurality of types of instructions to be executed by processing elements and exemplary arrangements of other processing elements such as functional units and register files for achieving efficient execution of instructions while balancing time, cost and spacial concerns. Therefore, the following U.S. Patent Applications are hereby incorporated by reference: Reconfigurable Functional Units for Implementing A Hybrid VLIW-SIMD Programming Model, by inventor Moataz A Mohamed, Ser. No. 09/172,315, filed Oct. 14, ,1998 and Efficient Forwarding Paths and Operand Sharing in a Digital Signal Processor, by inventors Moataz A. Mohamed, John R. Spence and Kenneth W. Malich, Ser. No. 09/173,257, filed Oct. 14, 1998.

[0019] Referring to FIG. 1, an exemplary DSP 100 includes an instruction packet 102, a scheduler 104, a plurality of MAC units 106, a plurality of accumulators 108, and data registers 110 and 112. Instruction packet 102 typically comprises a number of individual instructions 102a-102gand a header or template field 103. Instruction packet 102 is preferably 256 bits wide, although other bit widths may be used. Instructions 102a-102gand header 103 are typically 16 or 32 bits wide, although other bit widths maybe used. Each individual instruction 102a-102gtypically identifies an operation to be performed on certain data elements or operands. The data elements may be stored in data registers 110 and 112 which may be scalar or vector.

[0020] As identified above, these operations might include mathematical operations, logic operations, shifting operations, and other data manipulation operations. For example, these operations might include adding, subtracting, multiplying, dividing, selecting, combining, arithmetic shifting, logic shifting, and the like. The operation might also be a MAC operation which is common in DSP systems.

[0021] A MAC operation is one in which two or more operands are multiplied and the product is added to a value in an accumulator. An accumulator is a register having a stored value which may be increased by adding a new value to the stored value resulting in a sum which is then stored in the accumulator. MAC units are preferably 32 bits×32 bits so that they can execute one 32×32 bit operation, one 32×16 bit operation, or two 16×16 bit operations, although other dimensions may also be suitable. Accumulators are preferably 10×40 or 10×48 bit registers, but other register widths may also be suitable.

[0022] An exemplary instruction 102amight be the following MAC instruction:

[0023] MAC ACC1 SR1 SR2

[0024] Instruction 102ais requesting that the values stored in scalar register one 110band scalar register two 110cbe multiplied by MAC unit one 106band the product be added to a value stored in accumulator one 108b. This sum will replace the value previously stored in accumulator one 108b.

[0025] More particularly, DSP 100, in accordance with a preferred embodiment of the present invention, may execute instruction 102aas follows. Instructions 102a-102gof instruction packet 102 are sent to a scheduler 104. Scheduler 104 typically contains information on the data path structure, i.e., a mapping of how the registers and functional units are associated, coupled, related or connected. Scheduler 104 also contains information on the operating status of each functional unit, such as whether a given functional unit is currently executing an instruction or whether it is in an idle state and available to execute an instruction. Scheduler 104 receives instructions 102a-102gand determines which functional units are available to execute instructions 102a-102g. For example, as described above, instruction 102amay identify that accumulator one 108bis to be used. In accordance with a preferred embodiment of the present invention, accumulator one 108bis operatively associated with or coupled to MAC unit one 106b. Scheduler 104 will therefore send instruction 102ato MAC unit one 106bto be executed. MAC unit one 106bmay obtain data operands from scalar register one 110band scalar register two 110cand multiply these values to generate a result, total, product or the like. MAC unit one 106bthen sends the product to accumulator one 108b. Accumulator one 108badds the product to its existing stored value and stores the sum as its new value.

[0026] As exemplified with this illustration, instruction 102aneed not specify what MAC unit should be used in execution. Scheduler 104 selects which MAC unit should be used based upon the existing coupling configuration among the MAC units and the accumulators and the accumulators identified in instruction 102a.

[0027] The following configuration illustrates a preferable relationship between MAC units and accumulators. If there are n MAC units, the MAC units are preferably numbered with consecutive integers from zero to n, where n is an integer. Accumulators that are operatively associated with or coupled with each MAC unit are numbered I, I+1*n, I+2*n, I+3*n, etc. where I is an integer in the set of 0 to n, i.e., I is a number of a given MAC unit. Any number of MAC units can be used in a given DSP and any number of accumulators can be coupled with a given MAC unit.

[0028] Referring now to FIG. 2, in accordance with a preferred embodiment of the present invention, a preferred configuration 200 includes a plurality of first functional units 202a-202e, with each first functional unit 202a-202eunit being operatively associated or coupled with a plurality of second functional units 204a-204j. Exemplifying the above configuration, an exemplary processor has five first functional units 202a-202enumbered zero through four. (In the above example, n=5.) Each first functional unit 202a-202eis operatively associated with two second functional units 204a-204jas follows: first functional unit zero (I=0) 202ais operatively associated with second functional units zero (I) 204aand five (I+n) 204f, first functional unit one 202bis operatively associated with second functional units one 204band six 204g, first functional unit two 202cis operatively associated with second functional units two 204cand seven 204h, first functional unit three 202dis operatively associated with second functional units three 204dand eight 204i, and first functional unit four 202eis operatively associated with second fictional units four 204eand nine 204j.

[0029] Again with reference to FIG. 2, in accordance with an alternate preferred embodiment of the present invention, first and second functional units may be any type of execution units or processing elements, such as MAC units and accumulators. Thus, a preferred configuration 200 includes a plurality of MAC units 202a-202e, with each MAC unit being operatively associated with a plurality of accumulators 204a-204j. Exemplifying the above configuration, a preferred configuration has five MAC units numbered zero through four. (In the above example, n=5.) Each MAC unit is coupled to two accumulators as follows: MAC unit zero (I=0) 202ais coupled to accumulators zero (I) 204aand five (I+n) 204f, MAC unit one 202bis coupled to accumulators one 204band six 204g, MAC unit two 202cis coupled to accumulators two 204cand seven 204h, MAC unit three 202dis coupled to accumulators three 204dand eight 204i, and MAC unit four 202eis coupled to accumulators four 204eand nine 204j.

[0030] If an exemplary programming model operatively associates three accumulators for each MAC unit, again assuming five MAC units total (n=5), the numbering scheme would be as follows: MAC unit zero (I=0) 202ais operatively associated with accumulators zero (I) 204a, five (I+n) 204fand ten (I+2*n) (not shown), MAC unit one 202bis operatively associated with accumulators one 204b, six 204gand eleven (not shown), etc.

[0031] Dedicated multiple accumulators being operatively associated with each MAC unit permits data to be stored in an efficient and accessible manner. More specifically, with reference back to FIG. 1, assume a programmer desires to sequentially execute the following two instructions:

[0032] Instruction 1: MAC ACC0 SR1 SR2

[0033] Instruction 2: MAC ACC3 SR5 SR7

[0034] Both instructions utilize the same MAC unit, that is, MAC unit zero 106a. Because more than one accumulator is coupled to MAC unit zero 106a, the value in accumulator zero 108acan be stored in an easily accessible place after the execution of Instruction 1, while not preventing MAC unit zero 106afrom executing additional instructions, such as Instruction 2, with the use of accumulator three 108d. If accumulator zero 108awas the only accumulator coupled to MAC unit zero 106a, the value in accumulator zero 108awould either have to be returned to memory, which consumes power and time, or replaced by the next instruction that requires the use of MAC unit zero 106a.

[0035] This configuration, which associates non-sequential accumulators to each MAC unit, is optimal for both vertical and horizontal parallelism. In the case of vertical parallelism, a group of consecutive MAC units may be treated as a vector unit. For example, referring to FIG. 2, MAC unit zero 202aand MAC unit one 202bcould be combined to form VMAC 01 and could execute a vector MAC instruction. Assuming each MAC unit 202 is 32 bits×32 bits, VMAC 01 could execute a two-element vector operation (where each element is a 32-bit word) or a four-element vector operation (where each element is a 16-bit word). When consecutive MAC units are treated as a vector unit, the accumulators will automatically be grouped to form a vector accumulator based on the existing coupling configuration described above. Thus, accumulator zero 204a, accumulator one 204b, accumulator five 204fand accumulator six 204gmay all be grouped together to execute the vector MAC operation. This grouping ensures the correct vector semantics for the vector operation.

[0036] In the case of horizontal parallelism, if a programmer uses consecutive accumulators, the configuration ensures that each instruction will be scheduled to different MAC units. This enables the instructions to be executed simultaneously because no two consecutive instructions will be waiting for a single MAC unit to complete the two instructions. For example, an instruction packet contains two instructions as follows:

[0037] MAC ACC1 SR1 SR2

[0038] MAC ACC2 SR3 SR4

[0039] The first instruction will be executed by MAC unit one 202band the second instruction will be executed by MAC unit two 202c. This is because, although the specific MAC units are not identified in the instructions, in the configuration, accumulator one 204bis coupled to MAC unit one 202band accumulator two 204cis coupled to MAC unit two 202c. Thus, both instructions can be executed simultaneously, without serialization, i.e., incurring a delay waiting for a MAC unit to complete a first execution, so that it may then complete a second execution.

[0040] In the event that one instruction packet contains two instructions utilizing non-sequential accumulators, where the accumulators are both coupled to the same MAC unit, serialization can occur. For example, an instruction packet contains two instructions as follows:

[0041] MAC ACC0 SR1 SR2

[0042] MAC ACC5 SR3 SR4

[0043] In this case, both instructions are sent from the scheduler to MAC unit zero 202abecause accumulator zero 204aand accumulator five 204fare both coupled to MAC unit zero 202a. But, of course, both instructions cannot be executed simultaneously. Serialization in this case can be avoided by instructions or logic in the scheduler. When the scheduler identifies that two instructions require the same MAC unit, the scheduler could then generate an accumulator copy instruction which copies data from a specified accumulator into an accumulator that is coupled to a MAC unit that is currently idle or available to execute an instruction. The scheduler then sends the instruction to the idle or available MAC unit for execution.

[0044] It will be understood that the foregoing description is of preferred exemplary embodiments of this invention and that this invention is not so limited. Various modifications may be made in the design, arrangement, and implementation of this method and apparatus without departing from the spirit and scope of the subject invention, as set forth in the claims below.