Techniques for minimizing memory bandwidth used for motion compensation

Abstract

In a motion compensation engine, a number of blocks are provided for re-ordering motion vector (MV) reference positions prior to fetch. An MV Sort & Group block outputs MVs one at a time to a Decomposer block. The Decomposer block takes each MV and decomposes it into a series of DRAM read commands consisting of DRAM addresses. This rectangular region is divided into pixel words, which correspond to addressable DRAM words. The addresses are then sent to an Overlap Remover block, which comprises a bitmap corresponding to the DRAM addresses sent to it from the Decomposer block. Before a group is received, the bitmap is cleared by setting all coordinates to “0”. Each address received causes the Overlap Remover to set a bit to “1” in the bitmap which corresponds to a relative (x,y) coordinate within a small bounded rectangular region. Addresses received within a group, which are the same as previous addresses, are overlapping addresses and the corresponding bit will simply remain set to “1”.

Description

FIELD OF THE INVENTION

The present invention relates to motion compensation in video stream decoding. In particular, the present invention is directed toward a technique for minimizing memory bandwidth used for motion compensation, particularly in decoding MPEG data streams.

BACKGROUND OF THE INVENTION

Motion compensation is a key ingredient of many video compression standards. The older MPEG-2 standard uses two motion vectors in the formation of the prediction of its B-type macroblock. In MPEG-2 encoded data streams, B-type frames may be comprised of a number of macroblocks. Each macroblock comprise six blocks of 8×8 pixels, where each pixel may comprise Y, Cb or Cr data. The first four blocks form a larger 16×16 matrix of Y pixel data. The next block forms an 8×8 matrix of Cb pixel data and the last block forms an 8×8 matrix of Cr pixel data.

FIG. 2 is a simplified diagram illustrating the relationship of a macroblock from an I-type or P-type frame and the motion vector. I-type or P-type frame 210 includes a macroblock 230 located at a particular position in the frame. B-type frame 220 uses motion vector 240 to locate macroblock 230 in a previously decoded I-type or P-type frame 210 and uses it to predict the pixel values for the macroblock being reconstructed at the current macroblock position in the B-type frame 220. In addition to the motion vector 240, residual difference data is included in the B-type frame 220, which is added to the predicted pixels to produce the final reconstructed pixels at the current macroblock position. Thus, a B-type frame may include motion vector data, which locate reference macroblocks used as predictors. Along with motion vector data, residual difference data is included to correct the predicted pixels to form the final pixels, which make up the B-type frame image. When a motion vector is encoded, a residual difference is also encoded for every pixel in the rectangular region the motion vector is assigned to. This difference data is transformed using the Discrete Cosine Transform (DCT), which produces a rectangular region of numbers that represent the spectral components of the original rectangular region of numbers. By quantizing and entropy encoding these spectral components, the final representation can be encoded in significantly fewer bits. This is how pixel data is compressed using the MPEG-2 standard.

Much of the compression of the pixel data can be attributed to the motion vectors of the P or B-type frames. Motion vectors and residual difference data have been discussed thus far as components of B-type frames, but they may also be components of P-type frames. A motion vector represents movement of a given macroblock from a previously decoded I-type or P-type frame. The motion vector can actually point to a new location between pixels ( 1/2 pixel granularity). In such cases, neighboring pixels would be interpolated to form prediction pixels.

Recently emerging standards such as MPEG4-2, DivX, WMV9 and H.264 (MPEG4-10) require support for a higher density of motion vectors, and their implementations require a heavier use of memory. WMV9 has a 4MV mode, which defines 4 motion vectors per reference frame for the same 16×16 macroblock region—one for each 8×8 sub-region. MPEG4-10 requires support for as many as 16 independent motion vectors per reference, per macroblock. MPEG4-2 has a Global Motion Compensation (GMC) mode, which defines a separate motion vector for each luminance pixel of the 16×16 macroblock, which adds up to 256 motion vectors per macroblock.

A higher motion vector density translates to a larger number of memory fetches where each fetch comprises fewer reference pixels. In the case of MPEG4-10, each row fetched may consist of as few as 4 pixels (32 bits). In the case of the MPEG4-2 GMC mode, each row might only be 2 pixels wide (16 bits).

These more recent compression standards also have higher memory bandwidth requirements. From the IC and system point of view, higher bandwidth requirements translate to a wider memory data path. Synchronous DRAM is widely used today including a variant called DDR (double-data-rate) DRAM. The smallest randomly readable 32-bit DDR DRAM word is 64-bits wide since one address reads data on both edges of the clock, rising and falling. Likewise, a 64-bit DDR DRAM has a 128-bit effective word size.

The combination of smaller reference block sizes and a larger memory data path results in more wasted motion compensation related memory bandwidth. Non-aligned reads fetch unused pixels to the left and to the right of the required pixels. Also, typically a great deal of overlap exists between motion vectors since they are more likely to point to the same memory words. Certain memory words are therefore fetched multiple times as they are used for separate motion compensation operations based on separate motion vectors.

Thus, it remains a requirement in the art to provide a technique for minimizing memory bandwidth used for motion compensation in video compression decoding.

The best solution for this dichotomy is a technique which sorts and groups motion vectors by proximity and then removes any overlap between them before generating memory fetch commands. The idea of reordering motion vectors before fetching is novel, along with the specific implementation disclosed herein.

In the motion compensation engine of the present invention, a number of blocks are provided for re-ordering motion vector reference positions prior to fetch. These blocks may be implemented in software or hardware since they do not perform time-critical operations. The term “reference position” is not a generally used term and is used with special meaning in the present invention. A reference position is similar to a motion vector, but it is more closely related to a physical memory address than a simple displacement. It results from combining the motion vector with the currently decoded block position, the reference field or frame it refers to, and the color component currently being processed.

Motion vector reference positions (MVs) are received one at a time. They are sorted in numerical order and then grouped by an MV Sort & Group block. The block contains a small local memory or buffer, which holds MVs as they are sent in. The MVs are sorted and stored as they arrive. After they are sorted, they are guaranteed to be in raster-scan proximity order, such that the closest ones to the right are first, followed by the closest ones vertically below.

Grouping simply involves a second pass over the sorted MVs, which calculates the smallest rectangle, which can hold a sub-group of them. The first MV is read out of the buffer and sent along to the Decomposer block, followed by the second MV. An ever-increasing rectangle, which bounds the MVs is calculated as each MV is sent. Once the rectangle is about to exceed a predetermined size, an “end of buffer” (EOB) signal marks the last MV of the group as that MV is sent to the Decomposer. All of the MVs between EOBs are part of the same group.

At this point, the MV Sort & Group block outputs MVs one at a time to the Decomposer block. The EOB is sent as a sideband signal to mark the last MV of the group.

The Decomposer block takes each MV and decomposes it into a series of DRAM read commands consisting of DRAM addresses. An MV represents a rectangular region of pixels from a reference frame. The width and height of that region are determined by the compression standard (MPEG2, MPEG4, WMV9, MPEG4-10), the mode of operation (GMC, 1MV, 4MV, 4×4, etc) and which sub-pixel the original motion vector points to. This rectangular region is broken down into pixel groups, which correspond to addressable DRAM words. If the DRAM word size were 64 bits, then the rectangle would be broken down into 8 pixel groups, each readable from DRAM in one clock cycle.

The addresses corresponding to each 8-pixel group are then sent along to the Overlap Remover block in raster-scan order (left-to-right, top-to-bottom). The EOB signal is also passed along, but modified to mark the last address of the last MV of the group.

The Overlap Remover block comprises of a local memory or buffer, which stores a bitmap corresponding to DRAM addresses passed to it from the Decomposer block. Before a group is received, the bitmap is cleared, meaning a “0” is written to every (x,y) coordinate. Each address received causes the Overlap Remover to set a bit to “1” in the bitmap at the location which corresponds to a relative (x,y) coordinate within a small bounded rectangular region. Addresses received within a group, which are the same as previous addresses, are overlapping addresses and the corresponding bit will simply remain set to “1”. Once the EOB signal is detected with the final address received for a group, the Overlap Remover reads the bitmap and translates the “1” bits back into addresses which it uses to generate DRAM read commands. The “0” bits are ignored and skipped over. The Overlap Remover sends each read command one at a time to the memory controller, which retrieves corresponding data words and returns them to the motion compensation engine.

Three FIFO blocks act to queue simple motion compensation commands. These commands are, in turn, used to process the buffers that are returned by the DRAM controller. These buffers contain pixels from small rectangular regions of reference memory. Each buffer is the result of a group of DRAM read commands sent to the memory controller by the Overlap Remover block. Fully processed macroblock prediction blocks are output and later combined with IDCT or reverse transform error terms to produce reconstructed macroblocks.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 illustrates a block diagram for an exemplary implementation of a motion compensation engine according to the present invention.

FIG. 2 is a diagram illustrating how a macroblock may be rendered using a motion vector in a B-type frame according to the prior art.

DETAILED DESCRIPTION OF THE INVENTION

FIG. 1 illustrates a block diagram for an exemplary implementation of a motion compensation engine according to the present invention. Blocks 110, 120, and 130 in the upper-left-hand corner implement the technique of re-ordering motion vector reference positions prior to fetch. These three blocks may be implemented in software or hardware since they do not perform time critical operations.

Block 105 represents a “reference position”, a term not generally used in the art, but is used with special meaning with regard to the present invention. A reference position is defined herein as similar to a motion vector, but it is more closely related to a physical memory address than a simple displacement. The reference position results from combining the motion vector with the currently decoded block position, the reference field or frame it refers to, and the color component currently being processed.

The general input to the block diagram in FIG. 1 is a bus, which sends in motion vector reference positions (MVs) 105 one at a time. Motion vector reference positions 105 are sorted in numerical order and then grouped by the MV Sort & Group block 110. MV Sort & Group block 110 contains a small local memory or buffer, which holds MVs as they are sent in.

MVs 105 are sorted and stored as they arrive in block 110. After they are sorted, they are guaranteed to be in raster-scan proximity order, such that the closest ones to the right are first in order, followed by the closest ones vertically below.

Grouping in block 110 simply involves a second pass over the sorted MVs, which calculates the smallest rectangle, which can hold a sub-group of them. The first MV is read out of the buffer and sent along to Decomposer block 120, followed by the second. An ever-increasing rectangle, which bounds the MVs may then be calculated as each MV is sent. Once the rectangle is about to exceed a predetermined size, an “end of buffer” (EOB) signal marks the last MV of the group as that MV is sent to Decomposer 120. All of the MVs between EOBs are part of the same group.

At this point in the diagram though, MV Sort & Group block 110 outputs MVs one at a time to Decomposer block 120. The EOB is sent as a sideband signal to mark the last MV of the group.

Decomposer block 120 takes each MV and decomposes it into a series of DRAM read commands consisting of DRAM addresses. An MV represents a rectangular region of pixel images from a reference frame. The width and height of that region are determined by the compression standard (MPEG2, MPEG4, WMV9, MPEG4-10), the mode of operation (GMC, 1MV, 4MV, 4×4, etc) and which sub-pixel image the original motion vector points to. This rectangular region is broken down into pixel groups, which correspond to addressable DRAM words. If the DRAM word size were 64 bits, then the rectangle would be divided into 8-pixel groups, each readable from DRAM in one clock cycle.

The addresses corresponding to each 8-pixel group are then sent along to Overlap Remover block 130 in raster-scan order (left-to-right, top-to-bottom). The EOB signal is also passed along, but modified to mark the last address of the last MV of the group.

Overlap Remover block 130 comprises a local memory or buffer, which stores a bitmap corresponding to DRAM addresses passed to it from Decomposer block 120. Before a group is received, the bitmap is cleared by setting each (x,y) coordinate to “0”. Each address received causes Overlap Remover 130 to set a bit in the bitmap to “1”, which corresponds to a relative (x,y) coordinate within a small bounded rectangular region. This bitmap acts in a manner similar to a memory mask. Addresses received within a group, which are the same as previous addresses, are overlapping addresses, and the corresponding bit will simply remain set to “1”. Thus, operations need only be performed once for a particular pixel image.

Once the EOB signal is detected with the final address received for a group, Overlap Remover 130 reads the bitmap and translates the “1” bits back into addresses which it uses to generate DRAM read commands. The “0” bits are ignored and skipped over. If a particular bit in the bitmap is set to “1” more than once, only one DRAM read command will still be generated. Thus, unlike the Prior Art, where individual DRAM read commands are generated based upon the Motion Vectors, the present invention, utilizing the bitmap mask, eliminates duplicative DRAM read commands, thus minimizing the memory bandwidth required for motion compensation operations.

Overlap remover 130 sends each read command 510 one at a time to the memory controller, which retrieves the corresponding data words 520, MUXed in MUX 450 with address data from Motion Compensation (MC) Pixel filter 420 and Motion Compensation (MC) Weight and Add 430, and returns them to the motion compensation engine 610 shown in the lower half of FIG. 1.

The three FIFO blocks 150, 160, and 170 at the center of FIG. 1 act to queue simple motion compensation commands, which in turn may be processed by Motion Compensation (MC) pixel filter controller 340, Motion Compensation (MC) weight and add controller 350, and final prediction out controller 360. MUXes 310, 320, and 330 may be used to selectively transmit data between FIFO 150 and motion compensation blocks 340, 350, and 360. The data from motion compensation blocks 340, 350, and 360 may then be combined in MUXes 370, 380, and 390 and in turn used by the blocks 420, 430 at the bottom of FIG. 1 to process the buffers 410 that are returned by the DRAM controller. These buffers 410 contain pixels from small rectangular regions of reference memory. Each buffer is the result of a group of DRAM read commands 510 sent to the memory controller by Overlap Remover block 130.

Fully processed macroblock prediction blocks 440 are output at the bottom of FIG. 1, which are later combined with IDCT or reverse transform error terms to produce reconstructed macroblocks.

As illustrated in FIG. 1, the use of the Motion Vector sort and group 110, Decomposer 120, and overlap Remover 130 results in a more efficient use of memory bandwidth. By using the sort and overlap removal process, duplicate calls to memory addresses are eliminated by setting bits in the memory bit map such that each address is accessed only once in the Motion Compensation process.

While the preferred embodiment and various alternative embodiments of the invention have been disclosed and described in detail herein, it may be apparent to those skilled in the art that various changes in form and detail may be made therein without departing from the spirit and scope thereof.

Claims

1. A method of optimizing memory access when performing motion compensation, comprising: receiving a plurality of motion vector reference positions, sorting the motion vector reference positions in numerical order, and grouping, in a buffer, the motion vector reference positions in raster-scan proximity order.

2. The method of claim 1, wherein grouping comprises: calculating a smallest representative rectangle which can hold a group of motion vector reference positions, and calculating another representative rectangle bounding the motion vector reference positions as each motion vector reference position is read from the buffer.

3. The method of claim 2, wherein once the another representative rectangle is about to exceed a predetermined size, an end of buffer (EOB) signal marks the last motion vector reference position of the group, such that all motion vector reference positions between EOBs are part of a same group.

4. The method of claim 3, further comprising: sorting a region of reference positions and then separating them into groups where each group contains reference positions which fall within a bounding rectangle constrained to be less than a specified size.

5. The method of claim 4, further comprising: decomposing, in a decomposer, each motion vector reference position into a series of memory read commands comprising memory addresses.

6. The method of claim 5, further comprising: dividing the decomposed addresses into address groups, each of which correspond to single addressable memory words.

7. The method of claim 6, wherein the memory word size is 64 bits and the rectangular region is broken down into 8-pixel words, each word read from memory one clock cycle.

8. The method of claim 7, wherein memory addresses corresponding to each 8-pixel word are sent to an overlap remover in raster-scan order, along with the EOB signal modified to mark a last address of a last motion vector reference position of the group.

9. The method of claim 6, further comprising: storing in the overlap remover, a bitmap corresponding to memory addresses sent from the decomposer, wherein each address received causes the overlap remover to set a bit to “1” in the bitmap corresponding to a relative coordinate within the rectangular region.

10. The method of claim 9, wherein addresses received within a group, that are the same as previous addresses, are overlapping addresses and the corresponding bit in the bitmap remains set to “1”.

11. The method of claim 10, wherein once the EOB signal is detected with the final address received for a group, the overlap remover reads the bitmap and translates the set bits back into addresses used to generate memory read commands.

12. The method of claim 11, further comprising: sending from the overlap remover, each read command one at a time to a memory controller which retrieves corresponding data words and returns them to the motion compensation engine without having to read the same memory address twice.

13. The method of claim 1, wherein the motion vector reference position represents a rectangular region of pixels from a reference frame, in which width and height of the rectangular region are determined by a compression standard, mode of operation, and to which sub-pixel an original motion vector points.

14. A method of optimizing memory access for motion compensation, comprising: sorting a plurality of motion vector reference positions into raster scan order for a representative rectangular area; converting the motion vector reference positions to memory addresses; generating a bitmap representing memory addresses and setting each bit to “1” in the bitmap once for each memory address converted from motion vector reference positions; and using the bitmap to generate a list of memory addresses for memory access such that duplicate access of the same memory address is avoided over the larger group of reference positions.

15. A motion compensation system of optimizing memory access when performing motion compensation, comprising: an input for receiving a plurality of motion vector reference positions; a sorter for sorting the motion vector reference positions in numerical order; and a buffer for storing the motion vector reference positions in raster-scan proximity order.

16. The motion compensation system of claim 15, wherein the sorter further comprises: a calculator for calculating a smallest representative rectangle which can hold a group of motion vector reference positions; and a calculator for calculating another representative rectangle bounding the motion vector reference positions as each motion vector reference position is read from the buffer.

17. The motion compensation system of claim 16, wherein once the another representative rectangle is about to exceed a predetermined size, an end of buffer (EOB) signal marks the last motion vector reference position of the group, such that all motion vector reference positions between EOBs are part of a same group.

18. The motion compensation system of claim 17, further comprising: another sorter for sorting a region of reference pixels in a rectangular buffer corresponding to the group

19. The motion compensation system of claim 18, further comprising: a decomposer for decomposing each motion vector reference position into a series of memory read commands comprising memory addresses.

20. The motion compensation system of claim 19, wherein the decomposer divides the decomposed addresses into pixel words, each word of which corresponds to addressable memory words.

21. The motion compensation system of claim 20, wherein the memory word size is 64 bits and the rectangular region is divided by the decomposer into 8-pixel words, each read from memory one clock cycle.

22. The motion compensation system of claim 21, further comprising: an overlap remover for receiving memory addresses corresponding to each 8-pixel word in raster-scan order, along with the EOB signal modified to mark a last address of a last motion vector reference position of the group.

23. The motion compensation system of claim 19, wherein the overlap remover stores a bitmap corresponding to memory addresses passed from the decomposer, wherein each address received causes the overlap remover to set a bit to “1” in the bitmap corresponding to a relative coordinate within the rectangular region.

24. The motion compensation system of claim 23, wherein addresses received within a group in the overlap remover that are the same as previous addresses are overlapping addresses and the corresponding bit in the bitmap remains set to “1”.

25. The motion compensation system of claim 24, wherein once the EOB signal is detected by the overlap remover with the final address received for a group, the overlap remover reads the bitmap and translates the “1” bits back into addresses used to generate memory read commands.

26. The motion compensation system of claim 25, wherein the overlap remover sends each read command one at a time to a memory controller which retrieves corresponding data words and returns them to the motion compensation engine without having to read the same memory address twice within a group of reference positions.

27. The motion compensation system of claim 15, wherein the motion vector reference position represents a rectangular region of pixels from a reference frame, in which width and height of the rectangular region are determined by a compression standard, mode of operation, and to which sub-pixel an original motion vector points.

28. A motion compensation system of optimizing memory access for motion compensation, comprising: a sorting block for sorting a plurality of motion vector reference positions into raster scan order for a rectangular area; a decomposer block for converting the motion vector reference positions to memory addresses; an overlap remover for generating a bitmap representing memory addresses and setting each bit in the bitmap once for each memory address converted from motion vector reference positions, and using the bitmap to generate a list of memory addresses for memory access such that duplicate access of the same memory address is avoided.