The present invention relates to an information processing device and a data access method which read out a rectangular area inside of image data from a DRAM, the DRAM having a burst-mode which burst-transfers data of successive column addresses; the data access method processes the rectangular area of image data.
In recent years, multiple functions are being implemented in system LSIs; from the standpoint of system cost reduction, architecture such as unified memory, such as DRAM, are often used. It has also become common for diverse DRAM access requests to be made to a single DRAM.
In addition, since system LSIs implemented with multiple functions, such as media processing, demand a high bandwidth, speed enhancement in DRAM is becoming increasingly necessary. Accordingly, memory manufacturers are competing fiercely in DRAM speed enhancement.
However, since the operation frequency of a DRAM memory cell itself has not changed from before, the minimum access size to DRAM is rapidly increasing from the perspective of system LSI. As a result, there are no problems transferring long data that is equivalent to a burst length, however there is the problem that when transmitting short data, the transference load of unnecessary data increases and the effective bandwidth will decrease. For example, there is the problem that, for media processing, the effective bandwidth for a motion compensation process necessary for video decoding will drop; this problem has conventionally been avoided by utilizing high cost DRAM, which tolerates this drop in effective bandwidth.
For increasing the effective bandwidth, the image processing device disclosed in Patent Document 1, for example, aims to improve effective bandwidth by including a cache memory for caching single slice data from a frame memory for storing frame data.
[Patent Document 1] Japanese Laid-Open Patent Application No. 2000-175201 Publication
Since in the past, access to reference images in motion compensation processing has been access on a line basis, in contrast to a rectangular area basis i.e. short data-length lateral access, there is the problem that the transfer load of unnecessary data increases and the effective bandwidth drops. In other words, since the minimal access size for DRAM is rapidly increasing, as mentioned above, when data shorter than the burst length on the line direction is transferred, there is the problem that the transfer load of unnecessary data increases, and that the effective bandwidth drops.
In the conventional address mapping between a physical address of the DRAM and a logical address of image data, in order to earn a burst length for the line access direction, mapping is performed so that the column addresses are successive with the line direction. As a result, in access to data that is shorter than the burst length on the line direction, such as motion compensation processing, there is the problem that the transfer load of unnecessary data increases and the effective bandwidth drops.
Further, even when using a cache memory which caches data on a slice basis, there is the problem that the effective bandwidth cannot be raised for access to rectangular areas which exceed a boundary of a slice, as in motion compensation processing.
The present invention is conceived to solve the aforementioned problems and has as an object to provide an information processing device and data access method which reduce the transfer load of unnecessary data and improve effective bandwidth, even when data is accessed on a rectangular area basis, as in motion compensation processing.
In order to achieve the above object, the information processing device in the present invention includes a DRAM having a burst mode which burst-transfers data at successive column addresses; one or more data processing units which issue an access request; and an address conversion unit which converts access addresses which are included in the access request issued from the one or more data processing units. At least one of the one or more data processing units accesses an M×N rectangular area, where M and N are integers, and the address conversion unit converts access addresses so that a column address of data at the (K+m)th column, where K and m are integers and m≦M, of an Lth line, and a column address of data at a Kth column of an (L+n)th line, where L and n are integers and n≦N, become successive.
According to this configuration, there is the effect that the invalid transfer load can be reduced and the effective bandwidth can be improved in a short burst-length access method like rectangular access.
Here, some or all areas of the DRAM is a frame memory which stores image data, the rectangular area is N pixels×N lines in the image data, where M and N are integers, and the data processing unit may have a structure that performs one of motion compensation and motion estimation.
Here, n may be equal to 2n′ (n=2n′).
According to this configuration, there is the effect that unnecessary transfer loads can be reduced and the effective bandwidth can be improved in field estimation access.
Here, another one of the data processing units may access the image data on a line basis, and may successively read out all data of 2n lines.
According to this configuration, there is the effect that invalid transfer loads can be reduced and the effective bandwidth can be improved in DRAM which is mapped more than the address conversion unit above.
Here, the data processing unit may decode an inputted stream on a basis of two or more macroblocks, by motion compensation, the DRAM may store the image data decoded by the data processing unit, the information processing device may further include: a memory featuring a smaller storage capacity and faster access speed than the DRAM; a data transfer unit which transfers data from the DRAM to the memory; and the data processing unit accesses the image data stored in the DRAM as reference data.
Here, the image data stored in the DRAM may be split into split regions larger in size than the rectangular area, and the data transfer unit may transfer data on a split region basis from the DRAM to the memory, based on the access request from the data processing unit.
Here, the data transfer unit may have a structure that includes a register which holds a size of the split region.
Since plural rectangular areas can be transferred simultaneously, there is the effect that invalid transfer loads can be reduced and the effective bandwidth can be improved.
The data transfer unit may transfer data from the DRAM to the memory when a predetermined number n of read-out requests are outputted from the data processing unit.
Here, the data transfer unit may have a structure that includes a register which holds the size of the split region and the number n.
The data transfer unit may transfer the split region which includes all rectangular areas, from the DRAM to the memory when the access request from the data processing unit requests the rectangular areas which are adjacent or overlapping.
The data processing unit includes: a motion vector estimation unit which estimates plural motion vectors according to plural macroblocks from the inputted stream; and a decoding unit which decodes the inputted stream on a macroblock basis, and stores the decoding result in the DRAM.
A decoding sequence of the macroblocks may be changed based on the plural motion vectors so that addresses for accessing the DRAM vectors become successive.
A data access method in the present invention is a data access method for accessing a rectangular area made up of M pixels×N lines in image data from a DRAM, the DRAM having a burst mode which burst-transfers data of continuous column addresses, and storing the image data, and has an input step of inputting an access request for the rectangular area; and an address changing step of changing an access addresses included in an access request issued in the access step. In the address converting step, addresses are converted so that a column address of data at the (K+m)th column, where K and m are integers and m≦M, of the Lth line, and a column address of the data at the Kth column of the (L+n) line, where L and n are integers and n≦N, become successive.
As above and according to the information processing device and the data access method in the present invention, there is the effect that invalid transfer loads can be reduced and the effective bandwidth improved in comparison to data access which is shorter than the burst length of a rectangular area.
Below, an information processing device according to the embodiments of the present invention is described in detail based on diagrams.
In the figure, the memory 1 is a DRAM which stores image data, having a burst mode for burst-transferring data stored in successive column addresses.
The plural masters 13, 14 and 15 are connected to a common command bus 20 and a common data bus 21, output a command as a memory access request to the command processing unit 11, and input and output data into and from the data processing unit 12. Commands from the plural masters 13, 14 and 15 are serialized in priority rank and sent to the memory 1 through the memory interface 10. For writing, write data is sent to the memory 1 in synchronization with the sending of the command through the memory interface 10. For read data, after a command is accepted in the memory 1 and an appropriate time has elapsed, the read data is returned to data processing unit 12 through the memory interface 10. Afterwards, the read data is returned to one of the masters 13, 14 and 15 through data bus 21.
The command processing unit 11 has an address conversion unit 100, as shown in
In
Although the master 14 intends to access, for example, the sequence of the circled numbers 1, 3, . . . , 15, 2, 4, 6, . . . , 16 (see
The decoding device 201 can transfer data from either the memory 1 or the memory 3. The capacity of the memory 3 is smaller than that of the memory 1. For example, the memory 1 is a DRAM and the memory 3 is a faster SRAM. The decoding unit 203 decodes the stream by the motion compensation process. In the motion compensation process, the decoding unit 203 must take the decoded image data already in the memory 1 as a reference image and read out a rectangular area from the reference image. When so this processing is performed, a data transfer unit 4 transfers the necessary data among the reference images stored in the memory 1 to the memory 3, not on a rectangular area basis but on a split region basis.
The present invention is applicable to an information processing device which includes a DRAM that stores image data and has a burst mode, which burst-transfers successive data at successive column addresses, and applies to an information processing device, for example a network terminal which reproduces a compression-encoded stream, a digital camera, a cellular telephone with camera, a DVD recorder/player, a PDA, a personal computer and so on which are capable of recording and playing back DTV (digital television) movies.
Number | Date | Country | Kind |
---|---|---|---|
2004/120740 | Apr 2004 | JP | national |
Filing Document | Filing Date | Country | Kind | 371c Date |
---|---|---|---|---|
PCT/JP2005/004966 | 3/18/2005 | WO | 00 | 10/11/2006 |
Publishing Document | Publishing Date | Country | Kind |
---|---|---|---|
WO2005/109205 | 11/17/2005 | WO | A |
Number | Name | Date | Kind |
---|---|---|---|
5239512 | Fautier et al. | Aug 1993 | A |
5850483 | Takabatake et al. | Dec 1998 | A |
6075889 | Hamilton, Jr. et al. | Jun 2000 | A |
6075899 | Yoshioka et al. | Jun 2000 | A |
6205181 | Hu et al. | Mar 2001 | B1 |
6212231 | Takano | Apr 2001 | B1 |
6353438 | Van Hook et al. | Mar 2002 | B1 |
6745320 | Mitsuishi | Jun 2004 | B1 |
6750909 | Tsai | Jun 2004 | B1 |
6791625 | Kohashi et al. | Sep 2004 | B1 |
6807311 | Callway et al. | Oct 2004 | B1 |
6819326 | Jaspers | Nov 2004 | B2 |
7110663 | Tsukagoshi et al. | Sep 2006 | B1 |
20020196260 | Candler et al. | Dec 2002 | A1 |
20030113026 | Srinivasan et al. | Jun 2003 | A1 |
20030151609 | Champion | Aug 2003 | A1 |
20040258399 | Ishii et al. | Dec 2004 | A1 |
20050013368 | Gallant et al. | Jan 2005 | A1 |
20060263069 | Tsukagoshi et al. | Nov 2006 | A1 |
Number | Date | Country |
---|---|---|
0859524 | Aug 1998 | EP |
1089567 | Apr 2001 | EP |
2289199 | Nov 1995 | GB |
7-220059 | Aug 1994 | JP |
10-191236 | Jul 1998 | JP |
2000-175201 | Jun 2000 | JP |
2000-348168 | Dec 2000 | JP |
2004518343 | Jun 2004 | JP |
9809444 | Mar 1998 | WO |
0036841 | Jun 2000 | WO |
0062550 | Oct 2000 | WO |
02056600 | Jul 2002 | WO |
Number | Date | Country | |
---|---|---|---|
20070208919 A1 | Sep 2007 | US |