Image rendering

[00002] This invention relates to image rendering.

[00003] One way to accelerate the speed by which a three-dimensional image can be rendered from stored image data is to divide the image into subportions and distribute the subportions to different devices for rendering in parallel. One scheme, called scan line interleaving (SLI) generates the image line by line, with every other line rendered by one device and the intervening lines rendered by a second device. A different scheme divides the image in half and renders the two halves with different devices. A third scheme divides the image into identically-sized rectangular chunks and assigns the chunks to different devices for rendering.

[00004] In general, in one aspect; the invention features rendering an image by dividing the image into chunks, rendering the chunks in one of at least two devices, and determining which one of the at least two devices renders each one of at least some of the chunks based on the progress of at least one device of the at least two devices in the rendering of other chunks.

[00005] In another aspect, the invention features rendering an image by dividing the image into chunks, creating a command buffer for each of the chunks (each command buffer containing instructions necessary to render the chunk), placing each command buffer in a ring buffer, and processing each of the command buffers with one of at least two devices. One device processes at least one of the command buffers starting at a head of the ring buffer and the other device processes at least one of the command buffers starting at a tail of the ring buffer.

[00006] In another aspect, the invention features a computer system including a processor configured to divide an image into chunks and render at least some of the chunks, a system bus to carry the chunks to memory, and a graphics device configured to render at least some of the chunks, the chunks being determined by the progress of the processor in rendering chunks.

[00007] Other advantages and features will become apparent from the following description and from the claims.

[00013] In a specific example of the invention shown in FIGS. 1 and 2, an image 10 to be rendered is stored in an, I/O unit 50, e.g., a disk drive. A peripheral component interconnect (PCI) bus 52 carries the image 10 from the I/O unit 50 through a bridge 54 and a system bus 56 to a system memory 58. A chipset 64 controls a graphics controller 66 in a graphics card 70 (a graphics device). The chipset 64 allows the graphics controller 66 to fetch the image 10 from system memory 58 via the system bus 56 and an accelerated graphics port (AGP) 62.

[00014] Two devices (e.g., central processing unit (CPU) 72 and graphics card 70) render the image 10 from the stored image data. After the image 10 is fully rendered, a RAMDAC (random access memory digital-analog converter) 68 converts the digital image to an analog image and sends it for display on the computer's monitor.

[00015] Once the graphics controller 66 fetches the image 10, driver software in the graphics controller 66 divides the image 10 into chunks 12, e.g., identically-sized rectangles, to facilitate rendering. The software in the graphics controller 66 then creates a command buffer 14 for each chunk 12 in system memory 58. Each command buffer 14 contains all of the commands required to render one chunk 12.

[00016] As seen in FIG. 3, the software in the graphics controller 66 organizes all command buffers 14 into an ordered list 16. An optimization algorithm built into the software in the graphics controller 66 may create the ordered list 16. One possible optimization algorithm compares the estimated time needed to render the chunks 12 stored in each command buffer 14 to the relative performance capabilities of devices 18 and 20, the devices which will do the rendering, described further below. For example, the more detailed a chunk 12 is (e.g., ten colors v. two colors; middle of image 10 v. edge of image 10), the longer it takes to render. The optimization algorithm could place the more complex chunks 12 at a point in the ordered list 16 where the quicker device 18 or 20 will more likely first access it. The probability of first access is assessed given that the devices 18 and 20 may vary in speed and that device 18 starts rendering chunks 12 at the top of the ordered list 16 with a head command buffer 14a while device 20 starts at the bottom with a tail command buffer 14n. Ordering the least complex chunks 12 in the middle of the ordered list 16 tends to maximize the time device 18 and 20 render in parallel.

[00017] At the top of the ordered list 16, the software places two additional command buffers 22 and 24 to form a ring buffer 26 that is stored in system memory 58. The command buffers 22 and 24 contain the commands (described further below) necessary for the master device 18 to inform the slave device 20 of the address of the tail command buffer 14n in the ordered list 16 of command buffers 14.

[00018] The ring buffer 26 is a temporary storage area for a series of elements (here, command buffers 14), placed in a list having a head (first or top) and a tail (last or bottom). Two devices process elements in the list, one working from the head down and the other from the tail up. Here, the master 18 starts processing the ring buffer 26 at the head (command buffer 22) and the slave 20 starts at the tail (tail command buffer 14n). Pointers associated with each device direct the devices from element to element. Once both pointers point to the same element, the list is empty and processing of the list ceases. So here, master 18 and slave 20 process the ring buffer 26 simultaneously until their pointers 29 and 31 both point to the same command buffer 14.

[00019] The software identifies the master device 18 and the slave device 20. Ideally, one device is a discrete device, preferably the master 18, while the other device is an integrated device, preferably the slave 20. A discrete device is a separate and distinct device, e.g., CPU, whereas an integrated device is a collection of distinct elements, e.g., graphics card. In this example, the integrated device (slave 20) attaches to the discrete device (master 18). Often an integrated slave device is slower than a discrete master device.

[00020] After identifying the master 18 and slave 20, the software loads the address of the first command buffer 22 (the head of ring buffer 26) to a master pointer register 28. The master 18 reads this address to get pointer 29. The master 18 processes command buffers starting at this point, in this case, command buffer 22. The master 18 continues processing command buffers in the ring buffer 26 one by one as its pointer 29 moves down from the head of ring buffer 26.

[00021] Also referring to FIG. 4, master 18 first processes command buffer 22, enabling it to inform the slave 20 where to start processing-command buffers 14. The master 18 loads the address of the tail command buffer 14n in the ring buffer 26 to a slave pointer register 30. The slave 20 reads this address to get pointer 31. The slave 20 starts its processing at the tail command buffer 14n and works its way up the ring buffer 26, directed by pointer 31.

[00022] Now with the master 18 pointing to the head command buffer 14a in the ordered list 16 and the slave 20 pointing to the tail command buffer 14n, they can process command buffers 14 in parallel. A hardware mechanism in the graphics controller 66, implemented in both devices 18 and 20, distributes the command buffers 14 between the devices 18 and 20 in a synchronized and load balanced fashion.

[00023] The devices 18 and 20 each process the command buffers 14 in parallel at their own synchronous rates, in their respective directions. Rendering with a discrete device and an integrated device accommodates a performance level difference between thee two devices. The more powerful device, probably the discrete device, works at its performance level while obtaining whatever performance boost the other device provides.

[00024] Also referring to the flowchart of FIG. 5, each time device 18 or 20 points to the next command buffer 14, instructions 32 in a conditional batch buffer (CBB) 34 determine which device 0.18 or 20 renders the chunk 12. When master pointer 29 or slave pointer 31 first points to a command buffer 14, an instruction 40 asks whether the slave pointer 31 points to a command buffer 14 below the master pointer 29. If so, that means two or more command buffers 14 remain in the ring buffer 26. Then, in an action 42, the device 18 or 20 pointing to that command buffer 14 processes it, rendering a chunk 12. Local memory 60 (FIG. 1) stores the rendered chunk 12 in a batch buffer 36. In action 44, that rendering device's pointer 29 or 31 moves to point to the next command buffer 14. At this stage, the instructions 32 repeat, beginning with instruction 40.

[00025] If the slave pointer 31 does not point to a command buffer 14 below the master pointer 29, then only one, final command buffer 14 remains in the ring buffer 26. Instruction 46 asks which device 18 or 20 triggered this set of instructions 32 (i.e. which pointer 29 or 31 just pointed to the final command buffer 14). If it was the master 18, the master 18 waits in action 48 for the slave 20 to finish rendering the final chunk in action 74 and trigger instruction 76. Otherwise, the slave 20 processes the final command buffer 14 in action 74, rendering the final chunk.

[00026] Once the slave 20 finishes rendering the final chunk in action 74, one of the devices 18 or 20 stops functioning while the other assembles the final, rendered image in local memory 60, from the batch buffers 36 stored there, where the final image can be accessed by the RAMDAC 68 and displayed on the monitor. Instruction 76 asks which device 18 or 20 should assemble the final image in local memory 60. The driver software determines which device 18 or 20 should assemble the final image given the specific device configuration. If the software chooses the master 18, the master 18 in action 78 assembles the final image from all of the rendered chunks and the slave 20 terminates. After assembling the final image, the master 18 terminates in action 80. If the software chooses the slave 20, the same events occur, with the devices 18 and 20 reversed, in actions 82 and 84. Thus, having rendered a complete image, devices 18 and 20 are terminated until called to render another image.

[00027] Other embodiments are within the scope of the following claims.