Patent Grant
7555513
The present invention relates generally to processors, and more particularly to single instruction multiple data (SIMD) processors including arithmetic logic units (ALUs) having operand store interconnects facilitating the execution of certain instructions.
SIMD processors are best suited to performing the same operation on multiple pieces of data simultaneously. Typically, parallel processing portions of a single arithmetic logic unit (often viewed as individual parallel ALUs) operate on portions of operands simultaneously.
Certain calculations used in digital signal processing, on the other hand, repeatedly calculate the outcome of an operation applied to data that is shifted in time.
For example, shift invariant convolutions are often used in digital signal processing, to apply a frequency domain filter to a time domain representation of a signal. Specifically, the Nth order shift-invariant convolution of two discrete time signals h(n) and q(n) can be expressed mathematically as
z(n)=h(0)*q(n)+h(1)*q(n−1)+ . . . +h(N−1)*q(n−N+1)
where N is the order of the convolution, q(n), n=0, 1, 2, . . . is the sequence of samples input to the filter and h(i), i=0, 1, . . . , N−1 is the impulse response of the filter and z(n), n=0, 1, 2, . . . is the output from the filter. In general the length of the sequence of input samples need not be bounded, if output samples are produced as input samples are received. For example given an impulse response h(i), i=0 . . . N−1 and the subsequence of N input samples starting from discrete time t0−N+1:
q(t0−N+1), . . . , q(t0−1), q(t0)
the value of the convolution between h(i) and q(n) can be computed at discrete time t0. Further, given an impulse response h(i), i=0 . . . N−1 and the longer subsequence of N+k−1 input samples starting from discrete time t0−N+1, for some value of k>1:
q(t0−N+1), . . . , q(t0−1), q(t0), . . . , q(t0+k−1)
the value of the convolution between h(i) and q(n) can be computed at k discrete times t0 . . . t0+k−1.
For convenience, and without loss of generality, we define a new representation of the impulse response as a(i), i=0 . . . P−1 where
a(i)=h(N−1−i) for i<N
a(i)=not defined, N≦i<P
and N≦P.
Specifically a(0)=h(N−1), a(1)=h(N−2), . . . , a(N−1)=h(0)
Similarly without loss of generality we define a representation of the subsequence of N+k−1 input samples starting from discrete time t0−N+1 as x(n), n=0 . . . P−1 where
x(n)=q(n+t0−N+1) for n<N+k−1
x(n)=not defined, N+k−1≦n<P
In a similar vein we lastly define a representation of the subsequence of k output samples starting from discrete time t0 as y(n), n=0 . . . k−1 where
y(n)=z(t0+n), 0≦n<k
With these definitions in place an alternate representation of the convolution can be expressed as:
y(n)=a(0)*x(n)+a(1)*x(n+1)+ . . . +a(N−1)*x(n+N−1), n=0 . . . k−1
Conventional processors calculate shift-invariant convolutions, y(n) by executing a sequence of basic arithmetic operations such as multiply, add and multiply-and-accumulate.
As a further example, pattern matching techniques often require a processor to assess a best match of a series of target values (referred to as a target) and reference samples sequential in time or spatial position. Expressed mathematically, it is often desirable to assess j, for which the dissimilarity between the target a(0)a(1) . . . a(N−1) and reference samples x(j)x(j+1) . . . x(j+N−1) is minimized. It is often similarly desirable to find a pattern match of the target to interpolated samples of the reference. Again, conventional processors perform such pattern matching operations by executing a sequence of basic arithmetic operations.
Clearly, a SIMD processor capable of calculating shift invariant convolutions for multiple values of n concurrently would be beneficial, particularly in computationally intensive applications that benefit from high speeds. Similarly, a processor capable of performing several pattern matching operations concurrently would provide benefits.
Accordingly, an enhanced SIMD processor includes an ALU having data interconnects facilitating the concurrent processing of overlapping data portions of at least one operand store. Such interconnects may facilitate the calculation of shift-invariant convolutions, and pattern matches.
In accordance with an aspect of the present invention, a SIMD processor includes an ALU having a plurality of logic blocks, each of the logic blocks includes a first set of inputs and a second set of inputs. A plurality of first interconnects comprising Nxm1 data lines, interconnects a plurality of N data portions of a first operand to the first inputs of each of the logic blocks thereby providing the same data to the first set of inputs of each of the logic blocks. A plurality of groups of second interconnects, with each of the groups comprising Nxm2 data lines, interconnects a different subset of N data portions of a second operand to each of the second set of inputs of the plurality of logic blocks. Each of the subsets of N data portions of the second operand has data units in at least one other of the subsets of N data portions of the second operand.
In accordance with an aspect of the present invention, a method of calculating the Nth order convolution of sub-vectors of entries of vectors A=a(0)a(1)a(2) . . . a(P−1) and X=x(0)x(1)x(2) . . . x(P−1) at a single instruction, multiple data (SIMD) processor, includes concurrently calculating y(i)=a(j)*x(i)+a(j+1)*x(i+1) . . . +a(j+N−1)*x(i+N−1), for i=0 to i=k−1 wherein each y(i) is calculated by a logic block of the SIMD processor.
In accordance with yet a further aspect of the present invention, a method of calculating the Nth order sum of absolute differences between a target of the form a(0)a(1)a(2) . . . a(N−1) and reference samples of the form x(j)x(j+1) . . . x(j+N−1) at a single instruction, multiple data (SIMD) processor, includes concurrently calculating, y(i)=|a(0)−x(i)|+|a(1)−x(i+1)|+ . . . +|a(N−1)−x(i+N−1)|, for i=0 to i=k−1 wherein each y(i) is calculated by a logic block of the SIMD processor.
In accordance with yet another aspect of the invention, the Nth order sum of absolute sum and differences between a target of the form a(0)a(1)a(2) . . . a(N−1) and sub-samples of reference samples of the form x(j)x(j+1) . . . x(j+N−1) at a single instruction, multiple data (SIMD) processor may be calculated. A method includes interpolating adjacent data units x(i) of x(j)x(j+1) . . . x(j+N−1) to form interpolated data units l(j)l(j+1) . . . ; concurrently calculating, y(i)=|a(0)−l(i)|+|a(1)−l(i+1)|+ . . . +|a(N−1)−l(i+N−1)|, for i=0 to i=k−1. Each y(i) is calculated by a logic block of the SIMD processor.
In accordance with yet another aspect of the present invention, a SIMD processor includes an arithmetic logic unit (ALU) comprising first and second logic blocks. Each of the logic blocks includes a first set of inputs and a second set of inputs; a plurality of first interconnects, interconnecting a plurality of data portions of a first operand to the first inputs of each of the first and second logic blocks, thereby providing the same data to the first set of inputs of the first and second logic blocks; and first and second groups of second interconnects, with each of the first and second groups interconnecting a different subset of data portions of a second operand to the second set of inputs of one of the logic blocks. At least some of the interconnects of the first and second groups provide the same data storage portions from the second operand store to the second set of inputs of the first and second logic blocks.
Other aspects and features of the present invention will become apparent to those of ordinary skill in the art upon review of the following description of specific embodiments of the invention in conjunction with the accompanying figures.
In the figures which illustrate by way of example only, embodiments of the present invention,
Processor readable instructions are stored within instruction memory 30. Resulting calculations performed by processor 10 may be stored in operand storage 24 in communication with instruction decoder logic 18. Operand storage 24 may be any combination of addressable storage elements including processor registers and memory, addressable by individual locations. In the depicted embodiment, operand storage 24 forms part of processor 10. A person of ordinary skill will readily recognize that storage 24 may be any suitable combination of memory and registers local to processor or computer readable memory external to processor 10.
Addressing logic 18, based on instructions as decoded by instruction decoder 16 controls which elements within operand storage processed by processor 10 is accordingly adapted to address the various combinations of addressable elements within operand storage 24.
SIMD architecture is generally well known, and described in John L. Hennessy, David A. Patterson, David Goldberg, Computer Architecture: A Quantitative Approach (Morgan Kaufmann, 2002)—ISBN: 1558605967, the contents of which are hereby incorporated herein by reference.
As will be appreciated, processor 10 fetches instructions to be executed from instruction memory 30, at addresses specified by program counter 20. Instructions are decoded by instruction decoder 16. Branch unit 22 adjusts program counter 20 in accordance with decoded instructions. Arithmetic or logical instructions within decoded processor instructions are provided to ALU 12 which, in turn, operates on data provided by one or more operand stores in accordance with a decoded instruction to generate a result. Results are provided to a result store. Typically operand stores and result store are contained within operand storage 24. Generally, operand read data addresses and write data addresses may form part of a provided instruction and may be decoded by instruction decoder 16.
Notably, operand storage 24 includes two read ports, and thus may concurrently provide operand data A and X from two read addresses at data lines RD1 and RD2, as specified by read data address lines RA1 and RA2. Data at write data lines WD is written to operand storage 24 at a location specified by WA.
Conveniently ALU 12 includes multiple parallel logic blocks for performing the same arithmetic or logical instruction on multiple data elements. As a result SIMD processors, generally, (and in particular SIMD processor 10) are particularly well suited for certain classes of arithmetic calculations.
A simplified schematic block diagram of portions of an ALU 12 is illustrated in
Results calculated by logic blocks 100-0 and 100-1 may be combined into results data 112. Results data 112 may be written back into operand storage 24 (as for example specified at write data address lines WA). Results data 112 is accessible in data portions having bit size of m3.
Interconnect fabric 120 interconnects to provide operands 108 and 110 to logic blocks 100-0, 100-1 to facilitate certain arithmetic or logic calculations as detailed below. In its simplest form, interconnect fabric 120 interconnects N of the P1 data units of operand 108 to the first logic block 100-1 (with N≦P1), and the same N of the P1 data units to the second logic block 100-2. Interconnect fabric 120 further provides non-identical subsets of the P2 data units of operand 110 to each of the first and second logic blocks. In the depicted embodiment interconnect provides N-element subsets of the P2 data units of operand 110 (with N<P2). At least one of the N subsets is provided to each of the two logic blocks. In the example of
Now, two identical logic blocks 100-0, 100-2, allow for the concurrent calculation of
op(A,X1) and op(A,X2), where
A is an Nxm1 bit vector, X1 and X2 are Nxm2 bit subsets of vector X, and op represents the arithmetic or logical operation performed by logic blocks 100-0, 100-1.
Now, combination of two or more logic blocks 100-0 and 100-1 and a suitable interconnect fabric 120, results in an ALU 12 that is particularly well suited to calculate shift-invariant convolutions of two operands. As noted, shift invariant convolutions are often used in digital signal processing, to apply a frequency domain filter to a time domain representation of a signal. Specifically, using the terminology developed above, the Nth order shift-invariant convolution of two discrete time signals can be expressed mathematically as
where N is the order of the convolution, and A=[a(0) . . . a(N−1)]′, an Nx1 column vector of coefficients, and X(n)=[x(n) . . . x(n+N−1)]′, an Nx1 column vector.
Typically, the vector a(0) . . . a(N−1) represents an ordering of filter coefficients, and x(n) . . . x(n+N−1) represents N samples of an input signal. Shift invariant convolutions are further detailed in A. V. Oppenheim and R. W. Schafer. Digital Signal Processing. Prentice-Hall, Inc.: Englewood Cliffs, N.J., 1975, the contents of which are hereby incorporated by reference. Each coefficient a(i) and each x(i) sample of the input signal may be represented as an m bit integer.
Then y(n)=A′*X(n), where A′ is the vector transpose of A, and the * represents standard matrix multiplication (in this case the inner product between two vectors).
Now, it may be observed that y(n)=A′*X(n), y(n+1)=A′*X(n+1), y(n+k−1)=A′*X(n+k−1).
As may now be apparent, provided logic blocks 100-0, 100-1 are capable of calculating an inner product, with an appropriate interconnect fabric 120, logic blocks 100-0 and 100-1 as depicted in
To better appreciate this, a portion of ALU 12 of processor 10 is again schematically depicted in
As illustrated ALU 12 includes k logic blocks 100-0, 100-1 . . . 100-(k-1) (individually and collectively logic blocks 100) each operable to calculate the N point inner product of its inputs. Each logic block 100 has first and second sets of Nxm bit inputs and calculates the N point inner product of its inputs.
Interconnect fabric 120 includes a plurality of first interconnects 102, interconnecting a plurality of (N) m-bit data storage portions of operand 108 (in this case N groups of m adjacent bits of data within operand 108) to one set of inputs of each of operation blocks 100. Thus, in the depicted embodiment, each of first interconnects 102 is formed as Nxm data lines. Fabric 120 further includes interconnects 102 interconnecting the first set of inputs of the various logic blocks 100 so that each of the logic blocks 100 receives the same portion (i.e. N, m-bit data portions) of operand 108. That is, each logic block 100 receives at one of its inputs a(j)a(j+1) . . . a(j+N−1), with N+j<P, where each a(i) represents a data storage unit of m adjacent bits of operand 108. For simplicity, in the embodiment of
Further, each logic block 100 is provided with a different subset of N, m bit data storage portions of operand 110 by way of a group of second interconnects 104-0, 104-1 . . . 104-(k-1) (individually and collectively second interconnects 104). Each group of second interconnects 104 provides a subset of N, m bit data portions of operand 110, each subset offset from another group by m bits (i.e. one data unit) so that one group of interconnects 104-0 provides x(0), x(1), . . . , x(N−1) to the first logic block, the second group of second interconnects 104-1 provides bits x(1), x(2), . . . , x(N) to logic block 100-1. The third group of second interconnects 104-2 similarly provides bit groups x(2), x(3), . . . , x(N+1) to logic block 100-2. The ith interconnect 104-i provides the group consisting of x(i), x(i+1), . . . , x(i+N−1) to the ith logic block 100-i. Again x(i) represents a group of m adjacent bits within operand X.
Put another way, each group of interconnects 104 provides a N element subset of entries of vector X to a logic block 100. Each subset consists of N adjacent entries of vector X. The multiple subsets are each offset by one entry. As such, each group of second interconnects 104 is formed as Nxm data lines, interconnecting an Nxm bit subset of the operand X.
Conveniently, then, first and second interconnects 102 and 104 each provide Nxm bits of data to the ith logic block 100-i.
Each identical logic block 100, in turn, includes combinational logic to calculate y(n) (as defined above, for j=0), the inner product between A and X(n). So the ith logic block 100-i, calculates
y(i)=a(0)*x(i)+a(1)*x(i+1)+ . . . +a(N−1)*x(i+N−1)
k blocks 100 concurrently calculate y(0),y(1) . . . y(k−1), the shift invariant Nth order convolution of h(n) and q(n). Each y(i) has m3 bits. All y(i)s may be concatenated into results data 112 and stored in operand store 24. Interconnects 106-0, 106-1 . . . 106-(k-1) are groups of m3 bit data lines each providing result y(i) to results data 112.
As noted, first and second operands 108, 110 and results data 112 are read from or written to operand storage 24. As such, fabric 120 may be formed of interconnects 102, 104 and 106 in the form of direct or indirect electrical interconnects, in communication with operand storage 24. First and second operands may, for example, be read from memory or registers of processor 10.
Conveniently then, SIMD processor 10 may concurrently calculate the Nth order convolution for k values (i.e. y(0)y(1) . . . y(k−1)) with a single instruction, in the form
For example, ALU 12 may perform 4th order shift invariant convolutions in a SIMD processor whose operand stores store 8 samples of data, each sample having 8 bits (i.e. m=8). In this case the processor may fetch 4×8 bits of the coefficient vector A from operand 108, and 4 subsets of data vector X, each subset having 4×8 bits, from operand 110, and compute an output Y of length 4×8 bits. The results data 112 may be written back to the operand store 24 with appropriate padding, or can be added to a vector accumulator. Note that not all data from operand 110 need be used in generating the reduced set of outputs y(3) . . . y(0). In this case the m bits representing x(7) from operand 110 would be ignored.
Interconnect fabric 120 of logic blocks 100 depicted in
As noted the each logic block 100 is provided with N data units of target A, and a unique N entry subset of reference X having N data units. That is, the ith logic block 100-i is provided with entries a(0+j)a(1+j) . . . a(N−1+j), and x(i)x(i+1) . . . x(i+N−1). Again, for simplicity of illustration, j=0.
As such, each logic block 100 may further calculate the N-point sum of absolute differences (SAD) at its inputs to determine a difference metric between a target in A, and N data units of the reference X.
Thus, the ith logic block 100-i may calculate either the N point inner product, or N point sum of absolute differences.
Specifically, using the same values at its inputs, the ith logic block 100-i may calculate,
y(i)=a(0)*x(i)+a(1)*x(i+1)+ . . . +a(N−1)*x(i+N−1), or
SAD(i)=|a(0)−x(i)|+|a(1)−x(i+1)|+ . . . +|a(N−1)−x(i+N−1)|
k logic blocks 100 can thus calculate SAD(0), SAD(1) . . . SAD(k−1). Again, all sum of absolute differences may be concatenated and stored in operand storage 24 by way of interconnects 116. As well, an optimal pattern matching j may be assessed by determining the smallest of the calculated sum of absolute differences, reflecting the value of j for which the dissimilarity between a(0)a(1) . . . a(N−1) and x(0+j)x(1+j) . . . x(j+N−1) is minimized.
As will be appreciated, although exemplified ALU 12 incorporates logic blocks 100 that calculate N point inner products and N point sum of absolute differences, logic blocks 100 could easily be adapted to calculate other metrics, such as the P point sum of absolute differences, P point inner products, P point Euclidean distances between portions of the match target A and the reference X, and the like. Each calculation could be performed in response to a different processor instruction that may take the form of the CONVOLUTION_N instruction set out above. The number of logic circuits 118 incorporated in each logic block 100 will be limited by the number of interconnects 116 and conventional fabrication constraints.
In an alternate embodiment, illustrated in
Now, as will be appreciated, a SIMD processor and ALU exemplary of embodiments of the present invention may include logic blocks equivalent to logic blocks 100 and suitable interconnects. Results of calculations of logic blocks 100 may be stored in a result store as dictated by the instruction invoking use of logic blocks 100. Similarly, although exemplified logic blocks operate on Nxm sized operand and result stores, and calculate Nth order convolutions and SADs, varied data unit bit sizes and order operations are possible, by modifying the number of interconnects, and logic blocks 100 in manners understood by those of ordinary skill.
SIMD processors and ALUs exemplary of embodiments of the present invention may be formed using conventional manufacturing techniques, including very large scale integration techniques.
Of course, the above described embodiments are intended to be illustrative only and in no way limiting. The described embodiments of carrying out the invention are susceptible to many modifications of form, arrangement of parts, details and order of operation. The invention, rather, is intended to encompass all such modification within its scope, as defined by the claims.
| Number | Name | Date | Kind |
|---|---|---|---|
| 5530661 | Garbe et al. | Jun 1996 | A |
| 5659780 | Wu | Aug 1997 | A |
| 5742529 | Mennemeier et al. | Apr 1998 | A |
| 6282556 | Chehrazi et al. | Aug 2001 | B1 |
| 6292814 | Sazzad | Sep 2001 | B1 |
| 6526430 | Hung et al. | Feb 2003 | B1 |
| 6529930 | Sazzad et al. | Mar 2003 | B1 |
| 7054895 | Koba et al. | May 2006 | B2 |
| 7085795 | Debes et al. | Aug 2006 | B2 |
| 7126991 | Mimar | Oct 2006 | B1 |
| Number | Date | Country | |
|---|---|---|---|
| 20060095712 A1 | May 2006 | US |
