PatentDe  


Dokumentenidentifikation EP1212677 29.09.2005
EP-Veröffentlichungsnummer 0001212677
Titel REGISTERSPEICHER ZUR VERARBEITUNG VON 2-D MATRIX
Anmelder Intel Corp., Santa Clara, Calif., US
Erfinder CHEN, K., George, Portland, US
Vertreter derzeit kein Vertreter bestellt
DE-Aktenzeichen 60022206
Vertragsstaaten DE, FR, GB
Sprache des Dokument EN
EP-Anmeldetag 26.06.2000
EP-Aktenzeichen 009417429
WO-Anmeldetag 26.06.2000
PCT-Aktenzeichen PCT/US00/17630
WO-Veröffentlichungsnummer 0001008005
WO-Veröffentlichungsdatum 01.02.2001
EP-Offenlegungsdatum 12.06.2002
EP date of grant 24.08.2005
Veröffentlichungstag im Patentblatt 29.09.2005
IPC-Hauptklasse G06F 9/30
IPC-Nebenklasse G06F 9/302   G06F 9/318   

Beschreibung[en]
BACKGROUND 1. FIELD

The present invention relates generally to computer systems and, more specifically, to processor architectures.

2. DESCRIPTION

Some processors are designed to provide extensions to their instruction set architecture (ISA) for multimedia operations. For example, the MMX™ instructions supported by Pentium® II, Pentium® III, and Celeron™ processors, commercially available from Intel Corporation of Santa Clara, California, implement various functions useful for multimedia applications, such as digital signal processing, and audio and video processing. These instructions support "single instruction multiple data" (SIMD) operations on multimedia and communications data types. Although the use of these instructions provide an improvement over combinations of pre-existing instructions to perform a given function, and individual MMX™ instructions are efficient for some types of processing, various impediments to faster multimedia processing still remain. For example, many implementations of block-based image and video processing algorithms (such as joint photographic experts group (JPEG) and moving picture expert group (MPEG) schemes) result in the data, stored in a set of registers accessible as operands for the MMX™ instructions, being transposed during matrix mathematical operations.

EP-A-0743594 discloses a method of effecting a matrix transpose operation in a computer. The method uses a computer instructions which restructures a data string by retaining first and last sub-strings of the data string in unchanged positions and interchanges the position of at least two intermediate sub-strings. The data string is formed from sub-strings each representing one or more data value in a matrix. The computer instruction can be effected in a single register store having a predetermined bit capacity addressable by a single address, or in a pair of such register stores.

The transposition of data among registers incurs significant overhead, however, thereby slowing overall processor throughput for multimedia processing. Therefore, any techniques for avoiding or minimizing these delays would be a valuable advance in the processor art.

SUMMARY

According to a first aspect of this invention there is provided a processor as claimed in claim 1 herein.

According to a second aspect of this invention there is provided a method as claimed in claim 7 herein.

According to a third aspect of this invention there is provided a system as claimed in claim 9 herein.

Features of the invention are defined in the dependent claims herein.

BRIEF DESCRIPTION OF THE DRAWINGS

The features and advantages of the present invention will become apparent from the following detailed description of the present invention in which:

  • Figure 1 is a diagram of a set of MMX™ registers according to the prior art;
  • Figure 2 is a diagram of the set of MMX™ registers storing an 8 pixel by 8 pixel block of image data;
  • Figure 3 is a diagram of the set of MMX™ registers storing the transposed 8 pixel by 8 pixel block of image data;
  • Figure 4 is a diagram of a virtual MMX™ register set linked to the MMX™ register set according to an embodiment of the present invention;
  • Figure 5 is a diagram of the set of virtual MMX™ registers storing the transposed 8 pixel by 8 pixel block of image data; and
  • Figure 6 is a diagram of a system having a processor with a MMX™ register set and a virtual MMX™ register set according to an embodiment of the present invention.

DETAILED DESCRIPTION

An embodiment of the present invention comprises a method and apparatus for extending MMX™ registers to be more efficient for two dimensional (2-D) matrix operations.

Reference in the specification to "one embodiment" or "an embodiment" of the present invention means that a particular feature, structure or characteristic described in connection with the embodiment is included in at least one embodiment of the present invention. Thus, the appearances of the phrase "in one embodiment" appearing in various places throughout the specification are not necessarily all referring to the same embodiment.

When executing an instruction, a processor typically references one or more register operands. For MMX™ instructions, the register operands may be one or more of a special set of registers called MMX™ registers. Figure 1 is a diagram of a set of MMX™ registers according to the prior art. In the register set 10 shown in Figure 1, there are eight MMX™ registers labeled mm0 12 through mm7 14. In other embodiments, the number of registers may be higher or lower than eight. Each register includes a plurality of units of data, ordered from a low unit 16 to a high unit 18 as shown. In one embodiment, a unit comprises a byte. In other embodiments, a unit may comprise a word, a double word, or other storage unit. In at least one known system, the number of units (e.g., bytes) per MMX™ register is eight, although in other systems other numbers of units may be employed. To efficiently implement SIMD multimedia processing using the MMX™ registers, data to be processed should be aligned in such a way that multiple related data items are arranged in a single MMX™ register. For example, suppose an 8 pixel by 8 pixel block of image data is laid out in the MMX™ registers as shown in Figure 2, with each pixel value P(i, j) being represented in one unit and the overall set of registers representing a matrix. The 8 pixel by 8 pixel block may be a portion of a larger image. In this example, the first row of image data of the block is stored in the first MMX™ register, mm0 12, with the first column of the first row being stored in the low unit of mm0 and the last column of the first row being stored in the high unit of mm0, the second row of image data is stored in the second MMX™ register, mm1 20, with the first column of the second row being stored in the low unit of mm1 and the last column of the second row being stored in the high unit of mm1, and so on.

Once the data is stored in the MMX™ registers as shown, the processor may execute instructions to efficiently operate on the 8x8 matrix one row at a time. This type of processing is commonly used in block-based imaging applications, for example, as well as other applications. For example, all of the data of row 0 may be added to the data of row 3 using a single MMX™ instruction as shown below.

PADDB MMO, MM3 ; add row 0 to row 3 and store results in row 0.

However, to operate on the 8x8 matrix one column at a time presents problems because each column's data is spread out among the eight MMX™ registers. For example, the first column's data is distributed among the low units of mm0 12 through mm7 14, and the last column's data is distributed among the high units of mm0 through mm7, respectively. To continue to gain the benefits of using MMX™ SIMD processing, it becomes necessary to transpose the 8x8 matrix as shown in Figure 3. Matrix transposition is well known in mathematics. After transposition, the first MMX™ register mm0 12 stores the original first column of data, with the first row of the original first column being stored in the low unit 16 and the last row of the original first column being stored in the high unit 18 as shown. Similarly, the other MMX™ registers store columns of the 8x8 matrix as shown.

Although an 8x8 matrix transposition may be performed using known pack and unpack instructions, this processing is very inefficient and incurs significant processing overhead. Generally, transposition processing for an 8x8 matrix may be implemented by the performance of four 4x4 matrix transpositions, with each transposition requiring at least 12 processing cycles on a Pentium®III processor. Thus, at least 64 cycles are used merely to reposition the data so that an MMX™ instruction may be used to operate on the eight elements of a given column.

Embodiments of the present invention overcome the need to transpose the data in the MMX™ registers through a series of pack and unpack instructions by providing an additional set of registers in the processor to assist in multimedia processing. In embodiments of the present invention, an equivalent "mirror" set of the MMX™ registers is designed into the processor architecture. This register set may be called the virtual MMX register set, or VMX set. Figure 4 is a diagram of the virtual MMX register set linked to the MMX™ register set according to an embodiment of the present invention. In one embodiment, the number of MMX™ registers in MMX™ register set 10 is the same as the number of VMX registers provided in VMX register set 22. In addition, in one embodiment, the number of VMX registers is greater than or equal to the number of units in each MMX™ register (with a unit being a byte, word, double word, or other storage unit, depending on the implementation). VMX register set 22 stores the transposed matrix data from the MMX™ register set and may be automatically updated by register update logic 23 whenever any unit of any register in the MMX™ set is modified. Hence, a load of one row of MMX™ register set 10 automatically results in a load of one column of VMX register set 22. For example, the data from the first MMX™ register mm0 12 may be automatically stored in the low units of the VMX registers VMO through VM7, the data from the second MMX™ register mm1 20 may be automatically stored in the second lowest units of the VMX registers, and so on.

Referring back to Figure 2, if the MMX™ registers are loaded with an 8x8 matrix designated P0,0 to P7,7 as shown, then the VMX registers may be automatically loaded by register update logic with the transposed matrix as shown in Figure 5. To operate on row elements of the matrix, a program may simply reference one or more of MMX™ registers mm0 12 through mm7 14. However, to operate on column elements of the matrix, a program references the VMX registers vm0 24 through vm7 26, instead. Since the MMX™ registers are mirrored in the processor hardware with the VMX registers, there are no coherency issues. All MMX™ instructions may be operated with either of the register sets by using references to the appropriate registers as needed. No changes to the instruction set of the processor are needed, only operand references may be changed in a program.

One advantage of embodiments of the present invention over existing processor architectures is that the invention provides greater parallelism for matrix operations. This is achieved by avoiding costly transposes for column operations through the execution of multiple pack and unpack instructions.

One example of how the present invention may be employed is in the Discrete Cosine Transform (DCT) algorithm used for processing 8 pixel by 8 pixel blocks in many video processing schemes. Currently, to perform an 8x8 DCT, processing includes performing a 1x8 column transform first, transposing the 8x8 matrix, performing another 1x8 column transform, and then transposing the results again to get the DCT coefficients. The most optimized DCT currently running on Pentium® class processors takes about 300 cycles of processing. Of this amount, about 100 cycles are used to perform the transpose operations for an 8x8 matrix. Thus, implementation of the present invention yields an approximately 30% improvement for the DCT processing. Similar performance gains may be achieved for inverse DCTs. Although the example of a DCT is discussed herein, embodiments of the present invention may be useful for any matrix operations, including those used in various image and video compressions algorithms.

Although the present invention has been discussed above in the context of a two dimensional (2-D) matrix, the concept may be adapted to three or more dimensions. For example, a third register set may be included in the design of the processor to store other transpositions of the matrix data.

Figure 6 is a diagram of a system having a processor with a MMX™ register set and a virtual MMX™ register set according to an embodiment of the present invention. System 50 may include processor 52 coupled to memory 54. Processor 52 comprises various components well-known in the art, many of which have been omitted from Figure 6 for clarity. instruction memory 56 stores instructions that may reference one or more of MMX™ registers 10 or one or more of VMX registers 22. Register update logic 23 coordinates the automatic updating of the VMX registers 22 when the MMX™ registers change. Multiplexor (MUX) 58 selects data from either MMX™ registers or VMX registers for input to arithmetic logic unit (ALU) 60. The ALU produces data for data memory 52.

The present invention allows MMX™ registers to be manipulated in a more intuitive manner. With the addition of the mirror register set, implementation of any block-based algorithms may be simplified and their performance increased. Some examples of applications which may benefit from the present invention include discrete cosine transforms (DCTs) used in video compression algorithms, matrix transforms in three dimensional (3-D) graphics algorithms, and others.

While this invention has been described with reference to illustrative embodiments, this description is not intended to be construed in a limiting sense. Various modifications of the illustrative embodiments, as well as other embodiments of the invention, which are apparent to persons skilled in the art to which the inventions pertains are deemed to lie within the scope of the invention.


Anspruch[de]
  1. Prozessor, der folgendes umfasst:
    • eine erste Gruppe von Registern (10), welche eine Datenmatrix speichern;
    • eine zweite Gruppe von Registern (22), die eine Transposition der Datenmatrix speichern; gekennzeichnet durch
    • eine Logik (23), die automatisch die zweite Gruppe von Registern modifiziert, so dass das Transpositionsverhältnis zwischen der in der ersten Gruppe von Registern gespeicherten Datenmatrix und der in der zweiten Gruppe von Registern gespeicherten Transposition der Datenmatrix aufrechterhalten wird, als Reaktion auf eine Modifikation der ersten Gruppe von Registern.
  2. Prozessor nach Anspruch 1, wobei eine Modifikation einer Zeile der Matrix in der ersten Gruppe von Registern automatisch zu einer entsprechenden Modifikation einer Spalte der transponierten Kopie der Matrix in der zweiten Gruppe von Registern führt.
  3. Prozessor nach Anspruch 1, wobei die erste Gruppe von Registern eine erste Anzahl von Registern umfasst, wobei jedes Register eine erste Anzahl von Speichereinheiten umfasst, wobei die zweite Gruppe von Registern eine zweite Anzahl von Registern umfasst, wobei jedes Register eine zweite Anzahl von Speichereinheiten umfasst, und wobei die zweite Anzahl von Speichereinheiten größer oder gleich der ersten Anzahl von Registern ist.
  4. Prozessor nach Anspruch 3, wobei die erste Anzahl von Registern der zweiten Anzahl von Registern entspricht.
  5. Prozessor nach Anspruch 4, wobei die erste Gruppe von Registern MMX™ Register umfasst, wobei die erste Anzahl von Registern gleich acht ist, und wobei die Datenmatrix Bilddaten umfasst.
  6. Prozessor nach Anspruch 1, wobei der Prozessor einen Befehl ausführt, der sich auf ein Register der ersten Gruppe von Registern bezieht, so dass eine Zeile der Datematrix bearbeitet wird, und wobei ein Befehl ausgeführt wird, der sich auf ein Register der zweiten Gruppe von Registern bezieht, so dass eine Spalte der Datenmatrix bearbeitet wird.
  7. Verfahren der Verwendung von zwei Gruppen von Registern zur Matrixverarbeitung durch einen Prozessor, wobei das Verfahren die folgenden Schritte umfasst:
    • das Speichern einer Datenmatrix in einer ersten Gruppe von Registern (10), wobei die erste Gruppe von Registern eine erste Anzahl von Registern umfasst, wobei jedes Register eine erste Anzahl von Speichereinheiten umfasst, wobei jede Speichereinheit ein Element der Matrix speichert;
    • das Transponieren der Datenmatrix in eine zweite Gruppe von Registern (22), wobei die zweite Gruppe von Registern eine zweite Anzahl von Registern umfasst, wobei jedes Register eine zweite Anzahl von Speichereinheiten umfasst;
    • das automatische Modifizieren (23) der zweiten Gruppe von Registern, so dass das Transpositionsverhältnis zwischen der in der ersten Gruppe von Registern gespeicherten Datenmatrix und der Transposition der in der zweiten Gruppe von Registern gespeicherten Datenmatrix aufrechterhalten wird, als Reaktion auf eine Modifikation der ersten Gruppe von Registern; und
    • die Bezugnahme auf eines der Register der ersten Gruppe von Registern, so dass eine Zeile der Datenmatrix bearbeitet wird; und die Bezugnahme auf eines der Register der zweiten Gruppe von Registern, so dass eine Spalte der Datenmatrix bearbeitet wird.
  8. Verfahren nach Anspruch 7, wobei dieses ferner den Einsatz der beiden Gruppen von Registern für eine DCT-Verarbeitung (Discrete Cosine Transform) einer Matrix von Bilddaten durch einen Prozessor zumindest teilweise durch Bezugnahme auf eines der Register der zweiten Gruppe von Registern umfasst, so dass eine Spalte der Matrix bearbeitet wird.
  9. System (50), das folgendes umfasst:
    • einen Speicher (54);
    • einen mit dem Speicher gekoppelten Prozessor (52), wobei der Prozessor folgendes umfasst:
      • eine erste Gruppe von Registern (10), welche eine Datenmatrix speichern;
      • eine zweite Gruppe von Registern (22), welche eine Transposition der Datenmatrix speichern; gekennzeichnet durch eine Logik (23), die automatisch die zweite Gruppe von Registern modifiziert, so dass das Transpositionsverhältnis zwischen der in der ersten Gruppe von Registern gespeicherten Datenmatrix und der in der zweiten Gruppe von Registern gespeicherten Transposition der Datenmatrix aufrechterhalten wird, als Reaktion auf eine Modifikation der ersten Gruppe von Registern.
Anspruch[en]
  1. A processor comprising:
    • a first set of registers (10) storing a matrix of data;
    • a second set of registers (22) storing a transpose of the matrix of data; characterised by logic (23) that automatically modifies the second set of registers, so as to maintain the transposition relationship between the matrix of data stored in the first set of registers and the transpose of the matrix of data stored in the second set of registers, in response to a modification of the first set of registers.
  2. The processor of claim 1, wherein a modification of a row of the matrix in the first set of registers automatically results in a corresponding modification of a column of the transposed copy of the matrix in the second set of registers.
  3. The processor of claim 1, wherein the first set of registers comprises a first number of registers, each register comprising a first number of storage units, the second set of registers composes a second number of registers, each register comprising a second number of storage units, and the second number of storage units is greater than or equal to the first number of registers.
  4. The processor of claim 3, wherein the first number of registers is equal to the second number of registers.
  5. The processor of claim 4, wherein the first set of registers comprise MMX™ registers, the first number of registers is eight, and the matrix of data comprises image data.
  6. The processor of claim 1, wherein the processor executes an instruction referencing one of the first set of registers to operate on a row of the matrix of data and executes an instruction referencing one of the second set of registers to operate on a column of the matrix of data.
  7. A method of using two sets of registers for matrix processing by a processor comprising:
    • storing a matrix of data into a first set of registers (10), me first set of registers comprising a first number of registers, each register comprising a first number of storage units, each storage unit storing an element of the matrix;
    • transposing the matrix of data into a second set of registers (22), the second set of registers comprising a second number of registers, each register comprising a second number of storage units;
    • automatically modifying (23) the second set of registers, so as to maintain the transposition relationship between the matrix of data stored in the first set of registers and the transpose of the matrix of data stored in the second set of registers, in response to a modification of the first set of registers;
    • referencing one of the first set of registers to operate on a row of the matrix of data; and referencing one of the second set of registers to operate on a column of the matrix of data.
  8. The method of claim 7 further comprising using the two sets of registers for discrete cosine transform (DCT) processing of a matrix of image data by a processor at least in part by referencing one of the second set of registers to operate on a column of the matrix.
  9. A system (50) comprising:
    • a memory (54);
    • a processor (52) coupled to the memory, the processor comprising:
      • a first set of registers (10) storing a matrix of data;
      • a second set of registers (22) storing a transpose of the matrix of data; characterised by logic (23) that automatically modifies the second set of registers, so as to maintain the transposition relationship between the matrix of data stored in the first set of registers and the transpose of the matrix of data stored in the second set of registers, in response to a modification of the first set of registers.
Anspruch[fr]
  1. Processeur comprenant :
    • un premier jeu de registres (10) stockant une matrice de données ;
    • un deuxième jeu de registres (22) stockant une transposée de la matrice de données ; caractérisé par
    • une logique (23) qui modifie automatiquement le deuxième jeu de registres, afin de maintenir la relation de transposée entre la matrice de données stockée dans le premier jeu de registres et la transposition de la matrice de données stockée dans le deuxième jeu de registres, en réponse à une modification du premier jeu de registres.
  2. Processeur selon la revendication 1, dans lequel une modification d'une rangée de la matrice dans le premier jeu de registres résulte automatiquement en une modification correspondante d'une colonne de la copie transposée de la matrice dans le deuxième jeu de registres.
  3. Processeur selon la revendication 1, dans lequel le premier jeu de registres comprend un premier nombre de registres, chaque registre comprenant un premier nombre d'unités de stockage, le deuxième jeu de registres comprend un deuxième nombre de registres, chaque registre comprenant un deuxième nombre d'unités de stockage, et le deuxième nombre d'unités de stockage est supérieur ou égal au premier nombre de registres.
  4. Processeur selon la revendication 3, dans lequel le premier nombre de registres est égal au deuxième nombre de registres.
  5. Processeur selon la revendication 4, dans lequel le premier jeu de registres comprend des registres MMX™, le premier nombre de registres est huit, et la matrice de données comprend des données d'image.
  6. Processeur selon la revendication 1, dans lequel le processeur exécute une instruction référençant l'un du premier jeu de registres pour opérer sur une rangée de la matrice de données et exécute une instruction référençant l'un du deuxième jeu de registres pour opérer sur une colonne de la matrice de données.
  7. Procédé d'utilisation de deux jeux de registres pour le traitement matriciel par un processeur comprenant les étapes consistant à :
    • stocker une matrice de données dans un premier jeu de registres (10), le premier jeu de registres comprenant un premier nombre de registres, chaque registre comprenant un premier nombre d'unités de stockage, chaque unité de stockage stockant un élément de la matrice ;
    • transposer la matrice de données dans un deuxième jeu de registres (22), le deuxième jeu de registres comprenant un deuxième nombre de registres, chaque registre comprenant un deuxième nombre d'unités de stockage ;
    • modifier automatiquement (33) le deuxième jeu de registres, afin de maintenir la relation de transposition entre la matrice de données stockée dans le premier jeu de registres et la transposée de la matrice de données stockée dans le deuxième jeu de registres, en réponse à une modification du premier jeu de registres ;
    • référencer l'un du premier jeu de registres pour opérer sur une rangée de la matrice de données ; et
    • référencer l'un du deuxième jeu de registres pour opérer sur une colonne de la matrice de données.
  8. Procédé selon la revendication 7, comprenant en outre l'étape consistant à utiliser les deux jeux de registres pour un traitement par transformée en cosinus discrète (en anglais « Discrete Cosine Transform » - DCT) d'une matrice de données d'image par un processeur au moins en partie en référençant l'un du deuxième jeu de registres pour opérer sur une colonne de la matrice.
  9. Système (50) comprenant :
    • une mémoire (54) ;
    • un processeur (52) couplé à la mémoire, le processeur comprenant :
      • un premier jeu de registres (10) stockant une matrice de données ;
      • un deuxième jeu de registres (22) stockant une transposée de la matrice de données; caractérisé par
      • une logique (23) qui modifie automatiquement le deuxième jeu de registres, afin de maintenir la relation de transposition entre la matrice de données stockée dans le premier jeu de registres et la transposée de la matrice de données stockée dans le deuxième jeu de registres, en réponse à une modification du premier jeu de registres.






IPC
A Täglicher Lebensbedarf
B Arbeitsverfahren; Transportieren
C Chemie; Hüttenwesen
D Textilien; Papier
E Bauwesen; Erdbohren; Bergbau
F Maschinenbau; Beleuchtung; Heizung; Waffen; Sprengen
G Physik
H Elektrotechnik

Anmelder
Datum

Patentrecherche

Patent Zeichnungen (PDF)

Copyright © 2008 Patent-De Alle Rechte vorbehalten. eMail: info@patent-de.com