US RE36183 E
A data shuffler of the pipeline type receives successive trains of n sequential data words and rearranges data words in each train according to a predetermined order. It comprises p (p≦n) elementary processing units arranged in series. Each unit comprises an input, an output, a one-word storage register (20), a steering means (21) in order, in response to a binary control, to connect the input to the output either directly or through the register; and means for periodically supplying to each of the p steering means a sequence of n control bits determined as a function of said predetermined order.
1. A pipeline data shuffler receiving successive trains of sequential data words, and rearranging said sequential data words in each said train according to a predetermined order comprising:
a plurality of elementary processing units arranged and connected in series, each said elementary processing unit comprising:
a one-word storage register for temporarily storing one of said data words receiving from said input;
steering means for controlling the elementary processing unit to connect the input to the output through the register for temporarily storing the data word in response to a first level of a control signal and for connecting the input directly to the output thereby bypassing the register in response to a second level of the control signal; and
means, coupled to each said elementary processing unit, for periodically and concurrently supplying to each of the steering means of each said elementary processing unit a sequence of said control signals based upon a number of sequential data words in said each train and determined as a function of said predetermined order to rearrange the sequential data words.
2. A data shuffler according to claim 1, wherein, when successive elementary units are controlled by identical sequences, the steering means which separate them are suppressed.
3. A data shuffler according to claim 1, wherein a number of the elementary processing units is not more than a number of the sequential data words in each of the successive trains.
4. A data shuffler according to claim 1, wherein a number of the control signals is equal to a number of the sequential data words in each of the successive trains. .Iadd.
5. A method of rearranging sequential data from a predetermined initial order in which a first datum appears before a second datum to a predetermined arrival order in which the second datum appears before the first datum, comprising the steps of:
passing the data through a series of registers each having an input and an output, the registers connected in series; and
selectively bypassing the second datum from the input of a register of the series to the output of the register of the series while the first datum is held in the register, so the second datum appears in the arrival order before the first datum. .Iaddend..Iadd.6. The method of claim 5, wherein the steps of passing and selectively bypassing are repeated until the arrival order is achieved. .Iaddend..Iadd.7. A data shuffler for rearranging sequential data from a predetermined initial order to a predetermined arrival order, comprising:
a plurality of registers, each of a size to hold one datum, and each having an input and an output;
at least one selector having two inputs and one output, one of the inputs connected to the output of one of the registers to receive a datum stored therein, another of the inputs connected to the input of the one of the registers to receive a datum not stored in the one of the registers and the output connected to the input of another of the registers; and
a selector control including a control store representing a mapping of the predetermined initial order to the predetermined arrival order as a plurality of binary commands to control the selector so as to determine when the output of the selector represents the datum present at the one input and when the output of the selector represents the datum present at
the other input. .Iaddend..Iadd.8. The method of claim 5, wherein the predetermined initial order and the predetermined arrival order each include a predetermined number of data words, the method further comprising the steps of:
storing for each register of the series of registers a sequence of commands equal in number to the predetermined number of data words;
determining for each data word whether the data word is selectively bypassed based on a value of corresponding command. .Iaddend..Iadd.9. A method of rearranging sequential data from a predetermined initial order in which a first datum appears before a second datum to a predetermined arrival order in which the second datum appears before the first datum, comprising the steps of:
passing the data through a register having an input and an output; and
selectively bypassing the second datum from the input of the register to the output of the register while the first datum is held in the register, so the second datum appears in the arrival order before the first datum. .Iaddend..Iadd.10. The method of claim 9, wherein the steps of passing and selectively bypassing are repeated until the arrival order is achieved.
.Iaddend..Iadd.11. The method of claim 9, wherein the predetermined initial order and the predetermined arrival order each include a predetermined number of data words, the method further comprising the steps of:
storing for the register a sequence of commands equal in number to the predetermined number of data words;
determining for each data word whether the data word is selectively bypassed based on a value of a command corresponding to a position of the data word within the predetermined initial order. .Iaddend..Iadd.12. The method of claim 11, the predetermined initial order and the predetermined arrival order each including a predetermined number of data, wherein the sequence of commands is a number of bits equal to the predetermined number of data, each bit designating which input of the selector is represented at the output upon arrival at the selector of datum whose position within the predetermined initial order corresponds to each bit. .Iaddend..Iadd.13. A method of rearranging a serial input sequence of a predetermined number of data having a predetermined initial order, producing a serial output sequence of the predetermined number of data having a predetermined sorted order, the method comprising the step of:
precalculating a sequence of commands equal in number to the predetermined number of data, the precalculated commands defining rearrangements of pairs of data such that when in the initial order a first datum is in a position before a second datum and in the sorted order the second datum is in a sorted position before the first datum a command in the sequence of commands corresponding to the first datum defines delaying output of the
first datum while the second datum is output. .Iaddend..Iadd.14. The method of claim 13, further comprising the steps of:
applying the serial input sequence to an input of a selectable delay element;
selecting to pass data of the serial input sequence applied to the input of the selectable delay element directly to an output thereof in accordance with a command of the sequence commands corresponding to the position of each datum in the serial input sequence. .Iaddend..Iadd.15. The method of claim 13, wherein the step of applying further comprises the steps of:
providing a predetermined number of selectable delay elements connected in series, the predetermined number equal to a greatest number of delays needed to rearrange data of the serial input sequence from positions of the data in the initial order to sorted positions of the data in the sorted order;
applying the serial input sequence to an input of one of the selectable delay elements; and
selecting to pass data of the serial input sequence applied to the input of the selectable delay element directly to an output thereof in accordance with commands of the sequence of commands corresponding to positions of the data in the serial input sequence. .Iaddend.
The present invention relates to a data shuffler, that is, a circuit receiving a train of data and supplying at the output the same data arranged in a different and predetermined order.
Numerous applications, especially in the field of picture processing, require the use of operators or circuits capable of changing the order of a series of data according to a predetermined sequence.
An example of such an application is the calculation of the two-dimensional transform, such as discrete cosine transform, of an image block in a TV image where this shuffling problem occurs several times:
data shuffling for calculating the monodimensional transform,
conversion of the coefficient matrix scanning into a zigzag scanning.
Another application is the ciphering of a video signal through pixel mixing.
Hereinafter the state of the art and the invention will be described in the particular case it is desired to carry out zigzag scanning of a data block. However, it will be noted that it is only a specific case of the invention which is described in detail to make the invention better understood.
FIGS. 1A and 1B illustrate a zigzag scanning process. Considering a series of data, for example 16 data such as shown in FIG. 1A in the form of a 4×4 image block, zigzag scanning consists in reading those data according to successive diagonals, that is, as illustrated in FIG. 1B, in the following order:
1, 2, 5, 9, 6, 3, 4, 7, 10, 13, 14, 11, 8, 12, 15, 16.
The most conventional method for carrying out such a data shuffling consists in using a circuit of the type illustrated in FIG. 2. This circuit comprises two random access memories (RAM) M1 and M2 each of which having the size of the train or block of data that it is desired to shuffle. Each of those memories is addressable either in the natural order of FIG. 1A through a counter C or in the order corresponding to a stored sequence in a ROM 10, the content of which corresponds to the drawing of FIG. 1B for a zigzag scanning.
Each train or block of data arriving on an input 11 is written in one of the memories according to natural order while the other memory is read according to zigzag order towards an output 12. During the next step, the data are entered in the memory which has just been read while the memory in which the data have just been written is read.
This method requires a memory capacity of two data trains or blocks in the RAMs M1 and M2 and a sequence of address words in ROM 10.
By way of example, considering 4×4 blocks and 12bits words, the memory capacity of the RAMs will have to be 32 words of 12 bits (384 bits) and the memory capacity of ROM 10 will be 16 words of 4 bits (64 bits), that is, a total memory capacity of 448 bits.
In case of 64-words blocks (8×8), it is necessary to provide for a RAM of 128 (2×64) words of 12 bits (1536 bits) and a ROM having a capacity of 64 words of 6 bits (384 bits), that is, a total memory capacity of 1920 bits.
In order to reduce the required memory capacity, it has been devised to use the simplified diagram illustrated in FIG. 3 comprising one RAM memory MO, the memory capacity of which is equal to the size of the incoming data block, that it is desired to transform, receiving the data trains on an input 11 and supplying them in a determined order on an output 12. This memory MO is addressed by a ROM 13 controlled by a counter C.
Initially, memory MO is filled with the first block of data normally arranged according to the order from 1 to 16. For the next data block, the ROM determines the addressing according to a determined order. Each time a datum is read at a specified address, a datum from the incoming train is simultaneously written at the place that has just been read. This architecture permits to divide by two the RAM memory capacity and to simplify the address circuits since the reading/writing operations are sequentially carried out in the same case without readdressing. However, it is necessary to provide for an increased capacity of ROM 13 for supplying the sequential addresses required for processing the successive data blocks. This structure has been used when the data shuffling corresponds to a matrix transposition (symmetry with respect to a diagonal). Indeed, this operation is involutive, that is, after two transpositions the initial order occurs again. Therefore, it is only necessary to store one address sequence in ROM 13.
But, the operation is more complex in case of zigzag scanning as illustrated in FIGS. 4A-4F for 4×4 data blocks. The first address sequence, corresponding to natural order, is illustrated in FIG. 4A. The next address sequence is illustrated in FIG. 4B where it can be appreciated that the third address corresponds to case 5 and the fourth one to case 9. The table of FIG. 4B is obtained by following the arrows represented in FIG. 4A. For the next data block, following the same zigzag sequence as the one which permits to pass from FIG. 4A to FIG. 4B and by applying this sequence to FIG. 4B, it can be seen that the successive order will be 1, 2, 6, 10 . . . (FIG. 4C). One similarly passes from each figure to the next one. Considering FIG. 4F, it will be seen that by applying a zigzag scanning to the data that are stored therein, the natural sequence of FIG. 4A is obtained again. Therefore, it is necessary to provide for six address cycles in the ROM to be able to convert again, according to a zigzag order, the successive data trains.
Considering the same digital data as above, for 4×4 blocks of 12-bits data, it will therefore be necessary to provide for a 16-words RAM capacity of 12 bits (192 bits) and a 4-bits ROM capacity of 6×16 address words (384 bits), that is, a total memory capacity of 576 bits.
In case of a zigzag scanning of 8×8 blocks, it will be necessary to carry out one after the other 136 different sequences before obtaining again the initial sequence which requires a storage ROM of 8704 (136×64) 6-bits words, that is, 51 kbits, to which are to be added the 64 words of 12 bits of the RAM. Therefore, it is clear, that in this case the initial two-RAMs approach such as illustrated in FIG. 2 is far more economical than the one-RAM approach such as illustrated in FIG. 3.
Moreover, in the very schematic FIGS. 2 and 3, various decoders, which are necessarily associated with the storage ROM and are not represented, also necessitate a non negligible silicon surface area.
Another drawback of the prior art data shufflers is that the manufacture of memory structures is unavoidably delicate and involves to provide for sorting and testing steps to check that none of the memory points is faulty.
A further drawbacks of the prior art structures is that the process for using the described shufflers requires initial filling of a full memory block. Therefore, the latency time is equal to the duration of the initial filling of a complete memory, that is, to the duration for introducing all the words of a data train.
An object of the invention is to provide for a data shuffler palliating the drawbacks of the two above prior art circuits and more specifically:
decreasing the necessary silicon surface,
increasing the reliability of the system,
reducing latency time.
To achieve those objects and others, the invention provides for a pipeline-type data shuffler receiving successive trains of n sequential data words and rearranging in each train the data words according to a predetermined order, comprising:
p (p≦n) elementary processing units arranged in series, each unit comprising an input, an output, a one-word storage register, and a steering means in order, in response to a binary control, to connect the input to the output either directly or through the register; and
means for periodically supplying to each of the p steering means a sequence of n control bits determined as a function of said predetermined order.
Thus, the invention completely avoids the use of RAM memories and uses only individual one-word data registers which are a priori much more reliable devices than memories. In addition, it will be noted that the circuit according to the invention operates in pipeline mode, that is, reduces latency time as much as possible.
The foregoing and other objects, features and advantages of the invention will be apparent from the following detailed description of preferred embodiments as illustrated in the accompanying figures wherein:
FIGS. 1A and 1B are tables illustrating a zigzag-type scanning;
FIG. 2 shows a first data shuffler according to the prior art;
FIG. 3 shows a second data shuffler according to the prior art;
FIGS. 4A, 4B, 4C, 4D, 4E and 4F show successive addressing steps associated with the circuit of FIG. 3;
FIG. 5 illustrates an elementary processing unit according to the invention;
FIG. 6 is a time diagram illustrating the operation of the elementary unit of FIG. 5 for a specific sequence of selection control bits;
FIG. 7 shows a shuffler according to the invention;
FIG. 8 illustrates a variant of the elementary unit according to the invention; and
FIG. 9 shows a data shuffler according to the invention using the variant of FIG. 8.
The invention provides for a circuit permitting to sequentially rearrange a data train from a determined initial order to a determined arrival order through successive iterations carried out in successive elementary units.
More specifically, considering zigzag scanning, the initial order is the natural order going from 1 to 16 and the arrival order is the above-indicated order 1, 2, 5, 9, 6 . . . 15, 16. If E is the input sequence and S the output sequence, it will be possible to use, to pass from E to S, a so-called bubble sorting algorithm, that is, for each datum in a first series, the arrival series is observed to see if this datum is before or after the next datum. If the datum is before, it remains in place; if it is after, it is shifted until a datum is arranged after in the arrival series. Thus, to pass from series E to series S, the intermediate series X1-X4 illustrated in the following table are passed.
__________________________________________________________________________E 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16X1 1 2 3 5 6 4 7 9 10 11 8 13 14 12 15 16X2 1 2 5 6 3 4 9 7 10 11 13 14 8 12 15 16X3 1 2 5 6 3 9 4 7 10 13 14 11 8 12 15 16X4 1 2 5 6 9 3 4 7 10 13 14 11 8 12 15 16S 1 2 5 9 6 3 4 7 10 13 14 11 8 12 15 16__________________________________________________________________________
More specifically, series X1 is obtained from the initial series E as follows.
Elements 1 and 2 of the initial series are in this order in the arrival series. Therefore, they are not shifted. Similarly, element 3 is in series S before element 4; therefore, it remains in place. But, element 4 is in series S after element 5; therefore it is interchanged with element 5. In the same way, it is after element 6 and is therefore interchanged with element 6. But this element 4 is before element 7; therefore it remains in place, followed by element 7, and so on. The next element is element 7. Element 8 comes after 9, 10, 11 in the arrival order but before 12; the following elements are therefore 9, 10, 11, 8. Element 12 comes after 13 and 14 in the arrival order but before 15; the following elements therefore are 13, 14, 12. Lastly, come 15 and 16 which are in this order in the arrival series.
Therefore, through successive iterations it is possible to pass from series E to series S. In the specific case considered, four intermediate steps X1, X2, X3 and X4 are to be provided for with the described algorithm.
FIG. 5 represents an elementary processing unit permitting to implement bubble sorting algorithm as above described for passing from a series to the next series. E designates the input of this unit and X1 its output. This unit comprises a register 20 actuated at the rate of a clock CK corresponding to the introduction rate of the data and a selection circuit 21 actuated through a signal SEL1.
FIG. 6 illustrates a time diagram designed to explain the operation of the circuit of FIG. 5. The signal SEL1 is a binary signal, corresponding for example to an element of a succession of zeros and ones stored in a ROM. When the signal SEL1 is at a low level, the register is actuated and its output is connected to output X1. When the signal SEL1 is at a high level, the input E is directly connected to the output X1, no shifting being applied to the register. The last two lines of FIG. 6 represent signals E and X1, signal X1 being shifted with respect to the signal E by a clock interval. In signal X1, the above described series X1 is effectively obtained.
In order to pass from E to S, it is possible to use, as illustrated in FIG. 7, a sequence of units I-V, each of which corresponds to a unit of FIG. 5. Those units are controlled by series of zeros and ones stored in the columns of a ROM 22 comprising as many columns as units and as many lines as data words in a train or block to be shuffled. Since there are five units in the case described, the output will be shifted by five clock intervals with respect to the input. That is, the latency time of the circuit according to the invention is then 5 whereas it was equal to the number of elements of the train of data to be shuffled, that is, 16, for the prior art circuits.
On the other hand, it happens that in practice there often are several series of identical control orders for successive units (the last two ones in case of FIG. 7). Therefore, it is possible to use simplified elementary units such as the one represented in FIG. 8, the registers of those units being controlled by the same clock as the registers of identical units of the same subgroup. Therefore, FIG. 9 illustrates a variant of FIG. 7 wherein a unit of the type shown in FIG. 8 is substituted for unit IV. The two inputs of this unit are interconnected and its two outputs are directed towards the two non-interconnected inputs of a unit of the type shown in FIG. 5.
Therefore, it can be appreciated that in practice, the ROM memory designed to store the selection control bits has a more reduced size than the above represented size.
Latency time remains 5, as previously since there are 5 units but the critical transit time of data is reduced since a selection circuit has been suppressed.
With a circuit according to the invention, for 16-data blocks, one will use only 5 one-word registers and one ROM comprising four times 16 bits. For 8×8 blocks, it is necessary to provide for 27 registers and it is shown that due to optimization using simplified circuits such as the one of FIG. 8, one uses 15 selectors only, that is, it is necessary to store only 15 times 64 bits (64=8×8 being the number of words in the series that it is desired to shuffle). The ROM storage capacity is therefore 15 times 64=960 bits to which have to be added 27 registers of a 12-bits word each. Those values are to be compared with the 1920 memory bits of the prior art using two RAMs and with the 51 kbits of the prior art using one RAM.
Therefore, the invention permits to obtain a shuffling substantially reducing the number of necessary memory cases, but in addition it uses registers which are particularly reliable components and a low capacity ROM.
The invention has been described only in the specific case of zigzag shuffling. More generally, it can be applied to any chosen data shuffling. It is only for very simple shuffling, such as simple matrix transpositions that the devices of the prior art may exhibit an equivalent simplicity.
A further advantage of the invention is that, for a given shuffling operation, if it is desired to carry out an inverse shuffling operation (for example to pass from the zigzag scanning to a linear scanning), the same optimized structure such as the one described in FIG. 9 permits with a different content of the control ROMs to make the inverse operation. Moreover, the new ROM content is the same as the first one, with only line interchanges. Thus, instead of providing for a second ROM, it will be possible, by modifying the pointer progression of a single ROM in a determined way, to use the same control bit memory.
Moreover, while the control bits stored in a ROM have been described above, it is possible to use any other means for sending a predetermined series of control bits, for example programmable logic arrays (PLA).
Citations de brevets