US 20060221087 A1
Systems and methods for balancing a load among multiple graphics processors that render different portions of a frame. A display area is partitioned into portions for each of two (or more) graphics processors. The graphics processors render their respective portions of a frame and return feedback data indicating completion of the rendering. Based on the feedback data, an imbalance can be detected between respective loads of two of the graphics processors. In the event that an imbalance exists, the display area is re-partitioned to increase a size of the portion assigned to the less heavily loaded processor and to decrease a size of the portion assigned to the more heavily loaded processor.
28. A method for load balancing in a graphics processing system, the method comprising:
assigning a portion of a rendering process for a frame to be performed by each of a plurality of graphics processors in the graphics processing system;
instructing the graphics processors to perform the rendering process, wherein each graphics processor performs the portion of the rendering process assigned thereto;
instructing the graphics processors to provide feedback data reflecting respective rendering times for each graphics processor; and
for at least a first pair of graphics processors selected from the plurality of graphics processors:
determining, based on the feedback data provided by the first pair of graphics processors, whether an imbalance exists between respective loads of the first pair of graphics processors; and
in the event that an imbalance exists, shifting a subset of the portion of the rendering process assigned to a more heavily loaded one of the first pair of graphics processors to the less heavily loaded one of the first pair of graphics processors.
29. The method of
30. The method of
assigning a first processor identifier to one of the first pair of graphics processors; and
assigning a second processor identifier to the other of the first pair of graphics processors,
wherein the feedback data received from the graphics processors in the first pair of graphics processors includes the processor identifiers.
31. The method of
32. The method of
33. The method of
determining, based on the feedback data provided by the second pair of graphics processors, whether an imbalance exists between respective loads of the second pair of graphics processors; and
in the event that an imbalance exists, shifting a subset of the portion of the rendering process assigned to a more heavily loaded one of the second pair of graphics processors to the less heavily loaded one of the second pair of graphics processors.
34. The method of
35. The method of
36. The method of
instructing one graphics processor of the first pair of graphics processors to transfer result data from the portion of the rendering process assigned thereto to another one of the plurality of graphics processors.
37. The method of
38. The method of
39. The method of
defining a plurality of storage locations, each storage location associated with a different one of a plurality of frames,
wherein the act of instructing the graphics processors to provide feedback data includes:
instructing a first one of the first pair of graphics processors to store, after completion of the portion of the rendering process assigned thereto for one of the plurality of frames, a first processor identifier in the one of the storage locations associated with the one of the frames; and
instructing a second one of the first pair of graphics processors to store, after completion of the portion of the rendering process assigned thereto for the one of the frames, a second processor identifier in the one of the storage locations associated with the one of the frames,
wherein the processor identifier written by the first one of the first pair of graphics processors that was last to finish the portion of the rendering process assigned thereto overwrites the processor identifier of the second one of the first pair of graphics processors.
This application is a continuation of U.S. patent application Ser. No. 10/642,905, filed Aug. 18, 2003, which disclosure is incorporated herein by reference for all purposes.
The present disclosure is related to the following commonly assigned co pending U.S. Patent Applications: application Ser. No. 10/643,072, filed on the same date as the present application, entitled “Private Addressing in a Multi Processor Graphics Processing System” and application Ser. No. 10/639,893, filed Aug. 12, 2003, entitled “Programming Multiple Chips from a Command Buffer,” the respective disclosures of which are incorporated herein by reference for all purposes.
The present invention relates generally to graphics processing subsystems with multiple processors and in particular to adaptive load balancing for such graphics processing subsystems.
Graphics processing subsystems are designed to render realistic animated images in real time, e.g., at 30 or more frames per second. These subsystems are most often implemented on expansion cards that can be inserted into appropriately configured slots on a motherboard of a computer system and generally include one or more dedicated graphics processing units (GPUs) and dedicated graphics memory. The typical GPU is a highly complex integrated circuit device optimized to perform graphics computations (e.g., matrix transformations, scan-conversion and/or other rasterization techniques, texture blending, etc.) and write the results to the graphics memory. The GPU is a “slave” processor that operates in response to commands received from a driver program executing on a “master” processor, generally the central processing unit (CPU) of the system.
To meet the demands for realism and speed, some GPUs include more transistors than typical CPUs. In addition, graphics memories have become quite large in order to improve speed by reducing traffic on the system bus; some graphics cards now include as much as 256 MB of memory. But despite these advances, a demand for even greater realism and faster rendering persists.
As one approach to meeting this demand, some manufacturers have begun to develop “multi-chip” graphics processing subsystems in which two or more GPUs, usually on the same card, operate in parallel. Parallel operation substantially increases the number of rendering operations that can be carried out per second without requiring significant advances in GPU design. To minimize resource conflicts between the GPUs, each GPU is generally provided with its own dedicated memory area, including a display buffer to which the GPU writes pixel data it renders.
In a multi-chip system, the processing burden may be divided among the GPUs in various ways. For example, each GPU may be instructed to render pixel data for a different portion of the displayable image, such as a number of lines of a raster-based display. The image is displayed by reading out the pixel data from each GPU's display buffer in an appropriate sequence. As a more concrete example, a graphics processing subsystem may use two GPUs to generate a displayable image consisting of M rows of pixel data; the first GPU can be instructed to render rows 1 through P, while the second GPU is instructed to render rows P+1 through M. To preserve internal consistency of the displayed image (“frame coherence”), each GPU is prevented from rendering a subsequent frame until the other GPU has also finished the current frame so that both portions of the displayed image are updated in the same scanout pass.
Ideally, the display area (or screen) is partitioned in such a way that each GPU requires an equal amount of time to render its portion of the image. If the rendering times are unequal, a GPU that finishes its portion of the frame first will be idle, wasting valuable computational resources. In general, simply partitioning the display area equally among the GPUs is not an optimal solution because the rendering complexity of different parts of an image can vary widely. For example, in a typical scene from a video game, the foreground characters and/or vehicles—which are often complex objects rendered from a large number of primitives—tend to appear near the bottom of the image, while the top portion of the image is often occupied by a relatively static background that can be rendered from relatively few primitives and texture maps. When such an image is split into top and bottom halves, the GPU that renders the top half will generally complete its portion of the image, then wait for the other GPU to finish. To avoid this idle time, it would be desirable to divide the display area unequally, with the top portion being larger than the bottom portion. In general, the optimal division depends on the particular scene being rendered and may vary over time even within a single video game or other graphics application.
It would, therefore, be desirable to provide a mechanism whereby the processing load on each GPU can be monitored and the division of the display area among the GPUs can be dynamically adjusted to balance the loads.
The present invention provides systems and methods for balancing a load among multiple graphics processors that render different portions of a frame.
According to one aspect of the invention, a method is provided for load balancing for graphics processors configured to operate in parallel. A display area is partitioned into at least a first portion to be rendered by a first one of the graphics processors and a second portion to be rendered by a second one of the graphics processors. The graphics processors are instructed to render a frame, wherein the first and second graphics processors perform rendering for the first and second portions of the display area, respectively. Feedback data for the frame is received from the first and second graphics processors, the feedback data reflecting respective rendering times for the first and second graphics processors. Based on the feedback data, it is determined whether an imbalance exists between respective loads of the first and second graphics processors. In the event that an imbalance exists, based on the feedback data, the one of the first and second graphics processors that is more heavily loaded is identified; the display area is re-partitioned to increase a size of the one of the first and second portions of the display area that is rendered by the more heavily loaded one of the first and second graphics processors and to decrease a size of the other of the first and second portions of the display area.
According to another aspect of the invention, a method is provided for load balancing for graphics processors configured to operate in parallel. A display area is partitioned into at least a first portion to be rendered by a first graphics processor and a second portion to be rendered by a second graphics processor. The graphics processors are instructed to render a number of frames, wherein the first and second graphics processors perform rendering for the first and second portions of the display area, respectively. Feedback data for each of the frames is received from the first and second graphics processors, the feedback data for each frame indicating which of the first and second graphics processors was last to finish rendering the frame. Based on the feedback data, it is determined whether an imbalance exists between respective loads of the first and second graphics processors. In the event that an imbalance exists, based on the feedback data, the one of the first and second graphics processors that is more heavily loaded is identified; the display area is re-partitioned to increase a size of the one of the first and second portions of the display area that is rendered by the more heavily loaded one of the first and second graphics processors and to decrease a size of the other of the first and second portions of the display area.
In some embodiments, a storage location is associated with each one of the frames, and receiving the feedback data for each of the frames includes instructing the first graphics processor to store a first processor identifier in the associated one of the storage locations for each of the frames after rendering the first portion of the display area for that frame; and instructing the second graphics processor to store a second processor identifier different from the first processor identifier in the associated one of the storage locations for each of the frames after rendering the second portion of the display area for that frame. Each of the first and second identifiers may have a different numeric value and determination of whether an imbalance exists may include computing a load coefficient from the numeric values stored in the storage locations. The load coefficient may be, e.g., an average of the recorded numeric values that can be compared to an arithmetic mean of the numeric values of the processor identifiers in order to determine whether an imbalance exists.
In some embodiments, during the act of re-partitioning, an amount by which the size of the first portion of the display area is reduced depends at least in part on a magnitude of the difference between the load coefficient and the arithmetic mean.
In some embodiments, the plurality of graphics processors further includes a third graphics processor. During the act of partitioning, the display area may be partitioned into at least three bands including a first band that corresponds to the first portion of the display area, a second band that corresponds to the second portion of the display area, and a third band that corresponds to a third portion of the display area to be rendered by the third graphics processor, wherein the first band is adjacent to the second band and the second band is adjacent to the third band. Additional feedback data may be received for each of the frames, the additional feedback data indicating which of the second and third graphics processors was last to finish rendering the frame. Based on the feedback data, it may be determined whether an imbalance exists between respective loads of the second and third graphics processors. In the event that an imbalance exists, it may be determined which of the second and third graphics processors is more heavily loaded, and the display area may be re-partitioned to increase a size of the one of the second and third portions of the display area that is rendered by the more heavily loaded one of the second and third graphics processors and to decrease a size of the other of the second and third portions of the display area.
According to yet another aspect of the invention, a driver for a graphics processing subsystem having multiple graphics processors includes a command stream generator, an imbalance detecting module, and a partitioning module. The command stream generator is configured to generate a command stream for the graphics processors, the command stream including a set of rendering commands for a frame and an instruction to each of a first one and a second one of the graphics processors to transmit feedback data indicating that the respective processor has executed the set of rendering commands. The imbalance detecting module is configured to receive the feedback data transmitted by the first and second graphics processors and to determine from the feedback data whether an imbalance exists between respective loads of the first and second graphics processors. The partitioning module is configured to partition a display area into a plurality of portions, each portion to be rendered by a different one of the graphics processors, the plurality of portions including a first portion to be rendered by the first graphics processor and a second portion to be rendered by the second graphics processor. The partitioning module is further configured such that, in response to a determination by the imbalance detecting module that an imbalance exists, the partitioning module increases a size of the one of the first and second portions of the display area that is rendered by the more heavily loaded one of the first and second graphics processors and decreases a size of the other of the first and second portions of the display area.
The following detailed description together with the accompanying drawings will provide a better understanding of the nature and advantages of the present invention.
The present invention provides systems and methods for balancing a load among multiple graphics processors that render different portions of a frame. In some embodiments, load balancing is performed by determining whether one of two graphics processors finishes rendering a frame last more often than the other. If one of the processors finishes last more often, a portion of the processing burden (e.g., a number of lines of pixels to render) is shifted from that processor to the other processor. The comparison can be repeated and the load adjusted as often as desired. The technique of pairwise load comparisons and balancing can be extended to systems with any number of graphics processors.
Graphics processing subsystem 112 is advantageously implemented using a printed circuit card adapted to be connected to an appropriate bus slot (e.g., PCI or AGP) on a motherboard of system 100. In this embodiment, graphics processing subsystem 112 includes two (or more) graphics processing units (GPUs) 14 a, 114 b, each of which is advantageously implemented as a separate integrated circuit device (e.g., programmable processor or application-specific integrated circuit (ASIC)). GPUs 114 a, 114 b are configured to perform various rendering functions in response to instructions (commands) received via system bus 106. In some embodiments, the rendering functions correspond to various steps in a graphics processing pipeline by which geometry data describing a scene is transformed to pixel data for displaying on display device 110. These functions can include, for example, lighting transformations, coordinate transformations, scan-conversion of geometric primitives to rasterized data, shading computations, shadow rendering, texture blending, and so on. Numerous implementations of rendering functions are known in the art and may be implemented in GPUs 114 a, 114 b. GPUs 114 a, 114 b are advantageously configured identically so that any graphics processing instruction can be executed by either GPU with substantially identical results.
Each GPU 114 a, 114 b has an associated graphics memory 116 a, 116 b, which may be implemented using one or more integrated-circuit memory devices of generally conventional design. Graphics memories 116 a, 116 b may contain various physical or logical subdivisions, such as display buffers 122 a, 122 b and command buffers 124 a, 124 b. Display buffers 122 a, 122 b store pixel data for an image (or for a part of an image) that is read by scanout control logic 120 and transmitted to display device 110 for display. This pixel data may be generated from scene data provided to GPUs 114 a, 114 b via system bus 106 or generated by various processes executing on CPU 102 and provided to display buffers 122 a, 122 b via system bus 106. In some embodiments, display buffers 122 a, 122 b can be double buffered so that while data for a first image is being read for display from a “front” buffer, data for a second image can be written to a “back” buffer without affecting the currently displayed image. Command buffers 124 a, 124 b are used to queue commands received via system bus 106 for execution by respective GPUs 114 a, 114 b, as described below. Other portions of graphics memories 116 a, 116 b may be used to store data required by respective GPUs 114 a, 114 b (such as texture data, color lookup tables, etc.), executable program code for GPUs 114 a, 114 b, and so on.
For each graphics memory 116 a, 116 b, a memory interface 123 a, 123 b is also provided for controlling access to the respective graphics memory. Memory interfaces 123 a, 123 b can be integrated with respective GPUs 114 a, 114 b or with respective memories 116 a, 116 b, or they can be implemented as separate integrated circuit devices. In one embodiment, all memory access requests originating from GPU 114 a are sent to memory interface 123 a. If the target address of the request corresponds to a location in memory 116 a, memory interface 123 a accesses the appropriate location; if not, then memory interface 123 a forwards the request to a bridge unit 130, which is described below. Memory interface 123 a also receives all memory access requests targeting locations in memory 116 a; these requests may originate from scanout control logic 120, CPU 102, or other system components, as well as from GPU 114 a or 114 b. Similarly, memory interface 123 b receives all memory access requests that originate from GPU 114 b or that target locations in memory 116 b.
Bridge unit 130 is configured to manage communication between components of graphics processing subsystem 112 (including memory interfaces 123 a, 123 b) and other components of system 100. For example, bridge unit 130 may receive all incoming data transfer requests from system bus 106 and distribute (or broadcast) the requests to one or more of memory interfaces 123 a, 123 b. Bridge unit 130 may also receive data transfer requests originating from components of graphics processing subsystem 112 (such as GPUs 114 a, 114 b) that reference memory locations external to graphics processing subsystem 112 and transmit these requests via system bus 106. In addition, in some embodiments, bridge unit 130 facilitates access by either of GPUs 114 a, 114 b to the memory 116 b, 116 a associated with the other of GPUs 114 a, 114 b. Examples of implementations of bridge unit 130 are described in detail in the above-referenced co-pending application Ser. No. 10/643,072; a detailed description is omitted herein as not being critical to understanding the present invention.
In operation, a graphics driver program (or other program) executing on CPU 102 delivers rendering commands and associated data for processing by GPUs 114 a, 114 b. In some embodiments, CPU 102 communicates asynchronously with each of GPUs 114 a, 114 b using a command buffer, which may be implemented in any memory accessible to both the CPU 102 and the GPUs 114 a, 114 b. In one embodiment, the command buffer is stored in system memory 104 and is accessible to GPUs 114 a, 114 b via direct memory access (DMA) transfers. In another embodiment, each GPU 114 a, 114 b has a respective command buffer 124 a, 124 b in its memory 116 a, 116 b; these command buffers are accessible to CPU 102 via DMA transfers. The command buffer stores a number of rendering commands and sets of rendering data. In one embodiment, a rendering command may be associated with rendering data, with the rendering command defining a set of rendering operations to be performed by the GPU on the associated rendering data. In some embodiments, the rendering data is stored in the command buffer adjacent to the associated rendering command.
CPU 102 writes a command stream including rendering commands and data sets to the command buffer for each GPU 114 a, 114 b (e.g., command buffers 124 a, 124 b). In some embodiments, the same rendering commands and data are written to each GPU's command buffer (e.g., using a broadcast mode of bridge chip 130); in other embodiments, CPU 102 writes to each GPU's command buffer separately. Where the same command stream is provided to both GPUs 114 a, 114 b, the command stream may include tags or other parameters to indicate which of the GPUs should process a particular command.
Each command buffer 124 a, 124 b is advantageously implemented as a first-in, first-out buffer (FIFO) that is written by CPU 102 and read by the respective one of GPUs 114 a, 114 b; reading and writing can occur asynchronously. In one embodiment, CPU 102 periodically writes new commands and data to each command buffer at a location determined by a “put” pointer, which CPU 102 increments after each write. Asynchronously, each of GPUs 114 a, 114 b continuously reads and processes commands and data sets previously stored in its command buffer 124 a, 124 b; each GPU 114 a, 114 b maintains a “get” pointer to identify the read location in its command buffer 124 a, 124 b, and the get pointer is incremented after each read. Provided that CPU 102 stays sufficiently far ahead of GPUs 114 a, 114 b, the GPUs are able to render images without incurring idle time waiting for CPU 102. In some embodiments, depending on the size of the command buffer and the complexity of a scene, CPU 102 may write commands and data sets for frames several frames ahead of a frame being rendered by GPUs 114 a, 114 b.
The command buffer may be of fixed size (e.g., 5 megabytes) and may be written and read in a wraparound fashion (e.g., after writing to the last location, CPU 102 may reset the “put” pointer to the first location). A more detailed description of embodiments of command buffers and techniques for writing commands and data to command buffers in a multi-chip graphics processing system is provided in the above-referenced co-pending application Ser. No. 10/639,893.
Scanout control logic 120 reads pixel data for an image from frame buffers 122 a, 122 b and transfers the data to display device 110 to be displayed. Scanout can occur at a constant refresh rate (e.g., 80 Hz); the refresh rate can be a user selectable parameter and need not correspond to the rate at which new frames of image data are written to display buffers 122 a, 122 b. Scanout control logic 120 may also perform other operations such as adjustment of color values, generating composite screen images by combining the pixel data in either of the display buffers 122 a, 122 b with data for a video or cursor overlay image or the like obtained from either of graphics memories 116 a, 116 b or another data source (not shown), digital to analog conversion, and so on.
GPUs 114 a, 114 b are advantageously operated in parallel to increase the rate at which new frames of image data can be rendered. In one embodiment, referred to herein as “spatial parallelism,” each GPU 114 a, 114 b generates pixel data for a different portion (e.g., a horizontal or vertical band) of each frame; scanout control logic 120 reads a first portion (e.g., the top portion) of the pixel data for a frame from display buffer 122 a and a second portion (e.g., the bottom portion) from display buffer 122 b. For spatial parallelism, rendering commands and accompanying data may be written in parallel to both command buffers 124 a, 124 b (e.g., using a broadcast mode of bridge unit 130), but commands and/or data can also be selectively written to one or more of the command buffers (e.g., different parameters for a command that defines the viewable area might be written to the different command buffers so that each GPU renders the correct portion of the image).
An example of spatial parallelism is shown in
In accordance with an embodiment of the present invention, each GPU provides feedback data to the graphics driver program (or another program executing on CPU 102). The feedback data provides information about the time taken by a particular GPU to render its portion of the image. The graphics driver program uses this feedback to dynamically balance the load among the GPUs by modifying the clip rectangle from time to time, e.g., by changing the dividing line to a different line P′, based on the relative loads on the two GPUs.
An example of a command stream 300 that may be written to either (or both) of command buffers 124 a, 124 b is shown in
The clip rectangle command is followed by one or more rendering commands 304 and associated rendering data for a frame F0. These commands and data may include, for instance, definitions of primitives and/or objects making up the scene, coordinate transformations, lighting transformations, shading commands, texture commands, and any other type of rendering commands and/or data, typically culminating in the writing of pixel data to display buffers 122 a, 122 b (and reading of that data by scanout control logic 120).
Following the last rendering command 304 for frame F0 is a “write notifier” (WN) command 306. The write notifier command instructs the GPU to write feedback data to system memory indicating that it has finished the frame F0. This feedback data can be read by the graphics driver program and used to balance the load among the GPUs. Specific embodiments of feedback data are described below.
Write notifier command 306 is followed by rendering commands 308 and associated rendering data for the next frame F1, which in turn are followed by another write notifier command 310, and so on. After some number (Q) of frames, there is a write notifier command 322 followed by a new clip rectangle command 324. At this point, the clip rectangles for each GPU may be modified by the graphics driver program based on the feedback data received in response to the various write notifier commands (e.g., commands 306, 310). For example, where the display area is divided as shown in
It will be appreciated that the system described herein is illustrative and that variations and modifications are possible. For instance, while two GPUs, with respective memories, are shown, any number of GPUs can be used, and multiple GPUs might share a memory. The memory interfaces described herein may be integrated with a GPU and/or a memory in a single integrated circuit device (chip) or implemented as separate chips. The bridge unit may be integrated with any of the memory interface and/or GPU chips, or may be implemented on a separate chip. The various memories can be implemented using one or more integrated circuit devices. Graphics processing subsystems can be implemented using various expansion card formats, including PCI, PCIX (PCI Express), AGP (Accelerated Graphics Port), and so on. Some or all of the components of a graphics processing subsystem may be mounted directly on a motherboard; for instance, one of the GPUs can be a motherboard-mounted graphics co-processor. Computer systems suitable for practicing the present invention may also include various other components, such as high-speed DMA (direct memory access) chips, and a single system may implement multiple bus protocols (e.g., PCI and AGP buses may both be present) with appropriate components provided for interconnecting the buses. One or more command buffers may be implemented in the main system memory rather than graphics subsystem memory, and commands may include an additional parameter indicating which GPU(s) is (are) to receive or process the command. While the present description may refer to asynchronous operation, those skilled in the art will recognize that the invention may also be implemented in systems where the CPU communicates synchronously with the GPUs.
Embodiments of feedback data and load balancing techniques based on the feedback data will now be described. In one embodiment, each GPU 114 a, 114 b is assigned an identifier that it stores in a designated location in its local memory 116 a, 116 b; the identifier may also be stored in an on-chip register of each GPU 114 a, 114 b. For example, GPU 114 a can be assigned an identifier “0” while GPU 114 b is assigned an identifier “1.” These identifiers, which advantageously have numerical values, may be assigned, e.g., at system startup or application startup. As described below, the identifier may be used as feedback data for purposes of load balancing.
It should be noted that in this embodiment each GPU is instructed to write to the same location in system memory; as a result, the second GPU to finish frame k overwrites the identifier of the first GPU in array element feedback[k]. Thus, after both GPUs have finished a particular frame k, the value stored in feedback[k] indicates which GPU was last to finish the frame k.
At step 410, the frame counter is incremented to the next frame, modulo B. This causes the feedback array to be overwritten in a circular fashion every B frames, so that the contents of the array generally reflect the last B frames that have been rendered. In one embodiment, the frame counter value for each frame is provided with the write notification command to each GPU; in another embodiment, each GPU maintains its own frame counter and updates the frame counter after writing the identifier to the appropriate location in system memory in response to the write notifier command.
The information in the feedback array can be used by a graphics driver program (or another program executing on CPU 102) for load balancing, as illustrated in
At step 501, a clip rectangle command is issued (e.g., placed in the command stream) for each GPU. This initial clip rectangle command may partition the display area equally between the GPUs (e.g., using P=M/2) or unequally. For example, a developer of an application program may empirically determine a value of P that approximately balances that load and provide that value to the graphics driver program via an appropriate command. The initial size of the portion of the display area allocated to each GPU is not critical, as the sizes will typically be changed from time to time to balance the load.
At step 502, the graphics driver determines whether it is time to balance the load between the GPUs. Various criteria may be used in this determination; for example, the graphics driver may balance the load after some number (Q) of frames, where Q might be, e.g., 1, 2, 5, 10, 20, etc. Q advantageously does not exceed the number of entries B in the feedback array, but Q need not be equal to B. Alternatively, load balancing may be performed at regular time intervals (e.g., once per second) or according to any other criteria. If it is not time to balance the load, process 500 waits (step 504), then checks the load balancing criteria again at step 502.
When it is time to balance the load, the graphics driver averages Q values from the feedback array at step 506, thereby computing a load coefficient. In one embodiment Q is equal to B (the length of the feedback array), but other values may be chosen. It should be noted that the graphics driver and the GPUs may operate asynchronously with the CPU as described above, so that the graphics driver might not know whether the GPUs have finished a particular frame and the GPUs may be rendering a frame that is several frames earlier in the command stream than a current frame in the graphics driver. Where the feedback array is written in a circular fashion, as in process 400 described above, selecting Q to be equal to B provides an average over the B most recently rendered frames. In some embodiments, a weighted average may be used, e.g., giving a larger weight to more recently-rendered frames.
The load coefficient is used to determine whether an adjustment to the clip rectangles for the GPUs needs to be made. If the GPUs are equally loaded, the likelihood of either GPU finishing a frame first is about 50%, and the average value over a suitable number of frames (e.g., 20) will be about 0.5 if identifier values of 0 and 1 are used. An average value in excess of 0.5 indicates that GPU-1 (which renders the bottom portion of the image) is more heavily loaded than GPU-0, and an average value below 0.5 indicates that GPU-0 (which renders the top portion of the image) is more heavily loaded than GPU-1.
Accordingly, at step 510 it is determined whether the load coefficient exceeds a “high” threshold. The high threshold is preselected and may be exactly 0.5 or a somewhat higher value (e.g., 0.55 or 0.6). If the load coefficient exceeds the high threshold, then the loads are adjusted at step 512 by moving the boundary line P in
After the new boundary line P is determined, a new clip rectangle command is issued for each GPU (step 522) and the process returns to step 504 to wait until it is time to balance the load again. In an alternative embodiment, a new clip rectangle command is issued at step 522 only if the boundary line changes. In conjunction with the new clip rectangle command, a message may be sent to the scanout control logic so that the appropriate display buffer is selected to provide each line of pixel data (e.g., by modifying one or more scanout parameters related to selection of display buffers). Changes in the parameters of the scanout control logic are advantageously synchronized with rendering of the frame in which the new clip rectangle takes effect; accordingly, in some embodiments, the clip rectangle command may also update the scanout parameters in order to display the next rendered frame correctly.
In some embodiments, when the boundary line is shifted to balance the load, it may be useful to transfer data from one display buffer to another. For example, in
It will be appreciated that the processes described herein are illustrative and that variations and modifications are possible. Steps described as sequential may be executed in parallel, order of steps may be varied, and steps may be modified or combined. Optimal selection of the number of frames to average (Q) and/or the frequency of balancing generally depends on various tradeoffs. For instance, a small value of Q provides faster reactions to changes in the scene being rendered, while a larger value of Q will tend to produce more stable results (by minimizing the effect of fluctuations) as well as reducing any effect of an entry in the feedback array for a frame that only one GPU has finished (such an entry would not accurately reflect the last GPU to finish that frame). More frequent balancing may reduce GPU idle time, while less frequent balancing tends to reduce any overhead (such as data transfers between the memories of different GPUs) associated with changing clip rectangles. In one embodiment, checking the balance every 20 frames with Q=B=20 is effective, but in general, optimal values depend on various implementation details. It should be noted that checking the balance can occur quite frequently; e.g., if 30 frames are rendered per second and checking occurs every 20 frames, then the balance may change about every 0.67 seconds.
The identifiers for different GPUs may have any value. Correspondingly, the high threshold and low threshold may have any values, and the two threshold values may be equal (e.g., both equal to 0.5), so long as the high threshold is not less than the low threshold. Both thresholds are advantageously set to values near or equal to the arithmetic mean of the two identifiers; an optimal selection of thresholds in a particular system may be affected by considerations such as the frequency of load rebalancing and any overhead associated with changing the clip rectangles assigned to each GPU. The threshold comparison is advantageously defined such that there is some condition for which the load is considered balanced (e.g., if the average is exactly equal to the arithmetic mean).
Prior to rendering images or writing any feedback data, the feedback array may be initialized, e.g., by randomly selecting either of the GPU identifiers for each entry or by filling alternating entries with different identifiers. Such initialization reduces the likelihood of a spurious imbalance being detected in the event that checking the load balance occurs before the GPUs have written values to all of the entries that are being used to determine the load coefficient.
In one alternative embodiment, the amount by which the partition changes (e.g., the number of lines by which the boundary line P is shifted) may depend on the magnitude of the difference between the load coefficient and the arithmetic mean. For example, if the load coefficient is greater than 0.5 but less than 0.6, a downward shift of four lines might be used, while for a load coefficient greater than 0.6, a shift of eight lines might be used; similar shifts in the opposite direction can be implemented for load coefficients below the arithmetic mean. In some embodiments, the difference in size of the two clip rectangles is limited to ensure that each GPU is always rendering at least a minimum portion (e.g., 10% or 25%) of the display area.
Instead of averaging, a load coefficient may be defined in other ways. For instance, the sum of the recorded identifier values may be used as the load coefficient. In the embodiment described above, with Q=20, the stored identifier values (0 or 1) would sum to 10 if the load is balanced; high and low thresholds may be set accordingly. Other arithmetic operations that may be substituted for those described herein will also be apparent to those of ordinary skill in the art and are within the scope of the present invention.
In another alternative embodiment, different feedback data may be used instead of or in addition to the GPU identifiers described above. For example, instead of providing one feedback array in system memory, with both GPUs writing feedback data to the same location for a given frame, each GPU may write to a corresponding entry of a different feedback array, and the feedback data may include timing information, e.g., a timestamp indicating when each GPU finished a particular frame. In this embodiment, the graphics driver is configured to use the timing information to determine whether one GPU is consistently using more time per frame than another and adjust the clip rectangles accordingly to balance the load. It should be noted that, in some system implementations, timestamps might not accurately reflect the performance of the GPUs; in addition, determining relative loads from sequences of timestamps for each GPU generally requires more computational steps than simply computing a load coefficient as described above. Nevertheless, it is to be understood that embodiments of the invention may include timing information in the feedback data instead of or in addition to GPU identifiers.
Multi-processor graphics processing systems may include more than two GPUs, and processes 400 and 500 may be adapted for use in such systems. For example, one embodiment of the present invention provides three GPUs, with each GPU being assigned a different horizontal band of the display area, as shown in
More specifically, in one embodiment, the command stream for each GPU is similar to that of
To balance the loads, the graphics driver adjusts the value of K based on a load coefficient determined from the feedback01 array, e.g., in accordance with process 500 of
It will be appreciated that this load-balancing technique may be further extended to systems with any number of GPUs. For instance, the display area can be divided into any number of horizontal bands, with each band being assigned to a different GPU. In such embodiments, the number of feedback arrays is generally one less than the number of GPUs. Alternatively, vertical bands may be used.
It should also be noted that the identifier of a particular GPU need not be unique across all GPUs, as long as the two GPUs that write to each feedback array have identifiers that are different from each other. For example, in the embodiment shown in
In another alternative embodiment, a combination of horizontal and vertical partitions of the display area may be used to assign portions of the display area to GPUs. For example,
In yet another alternative embodiment, the vertical boundary line J might also be adjustable. For instance, GPU-0 and GPU-1 could each be assigned a secondary (column) identifier value of “0” while GPU-2 and GPU-3 are each assigned a secondary identifier with a value of “1.” A third feedback array feedbackC may be provided, with each GPU writing its secondary identifier to the feedbackC array in addition to writing its primary identifier to the appropriate one of the feedback01 and feedback23 arrays. The vertical boundary line J can then be adjusted based on the average value of entries in the feedbackC array. Alternatively, the primary identifier (which has values 0-3) can be associated with the vertical division while the secondary identifier (which has values 0 and 1) is associated with the horizontal division.
The techniques described herein may also be employed in a “multi-card” graphics processing subsystem in which different GPUs reside on different expansion cards connected by a high-speed bus, such as a PCIX (64-bit PCI Express) bus or a 3GIO (third-generation input/output) bus presently being developed. An example of a multi-card system 900 is shown in
In this arrangement, spatial parallelism can be implemented, with each GPU 914 a, 914 b rendering a portion of each frame to its display buffer 922 a, 922 b. In order to display the frame, pixel data from display buffer 922 b is transferred (e.g., using a conventional block transfer, or Blit, operation) via bus 908 to display buffer 922 a, from which it is read by scanout control logic 920.
Load balancing as described above can be implemented in this system and advantageously takes into consideration time consumed by the data transfers. For example,
In this embodiment, pixel data from display buffer 922 b is transferred to display buffer 922 a prior to scanout. Accordingly, for GPU 914 b, the rendering commands 1004 b are followed by a Blit command 1006 that instructs GPU 914 b to transfer pixel data from local display buffer 922 b to display buffer 922 a on card 912 a so that it can be scanned out. Since GPU 914 a writes pixel data directly to display buffer 922 a, a Blit command is not required in command stream 1000 a, so the rendering commands 1004 a for GPU 914 a are followed by a “no-op” 1005. The no-op may be, e.g., a command that simply delays execution of a following command (such commands are known in the art), no command, or a command instructing GPU 914 a to ignore a Blit command that appears in its command stream.
A write notifier command 1008 a for frame F0 follows the no-op command 1005 in command stream 1000 a, and a corresponding write notifier command 1008 b follows Blit command 1006. The write notifier commands 1008 a, 1008 b may be implemented similarly to the write notifier commands described above with reference to process 400 of
It should be noted that the time required for the Blit operations is accounted for in the load balancing process because the write notifier command 1008 b for a frame F0 is not executed by GPU 914 b until after the Blit operation for the frame F0 is executed. Thus, the rendering time for GPU 914 a is balanced against the rendering time plus the Blit time for GPU 914 b.
In some multi-card embodiments used to render scenes in which foreground regions (most often but not always at the bottom of the display area) are consistently more complex than background regions, a performance advantage can be gained by assigning GPU 914 a to process the background region of the scene and assigning GPU 914 b to process the foreground region. For example, in
Those of ordinary skill in the art will recognize that a similar implementation might also be used in embodiments of a single-card multi-processor system in which pixel data from all GPUs is transferred to a single display buffer prior to scanout. For example, in system 112 of
While the invention has been described with respect to specific embodiments, one skilled in the art will recognize that numerous modifications are possible. For instance, in a multi-processor graphics processing system, any number of GPUs may be included on a graphics card, and any number of cards may be provided; e.g., a four-GPU subsystem might be implemented using two cards with two GPUs each, or a three-GPU subsystem might include a first card with one GPU and a second card with two GPUs. One or more of the GPUs may be a motherboard-mounted graphics co-processor.
Rendering of a display frame may be divided among the GPUs in horizontal bands and/or vertical bands. Those of skill in the art will recognize that use of vertical bands may result in more uniform sizes of the regions rendered by different GPUs (since image complexity usually varies less from left to right than from top to bottom), while use of horizontal bands may simplify the scanout operation in a horizontal row-oriented display device (since only one GPU's display buffer would be accessed to read a particular row of pixels). In addition, a frame may be partitioned among the GPUs along both horizontal and vertical boundaries, and load balancing may be performed along either or both boundaries as described above.
Embodiments of the invention may be implemented using special-purpose hardware, software executing on general-purpose or special-purpose processors, or any combination thereof. The embodiments have been described in terms of functional blocks that might or might not correspond to separate integrated circuit devices in a particular implementation. Although the present disclosure may refer to a general-purpose computing system, those of ordinary skill in the art with access to the present disclosure will recognize that the invention may be employed in a variety of other embodiments, including special-purpose computing systems such as video game consoles or any other computing system that provides graphics processing capability with multiple graphics processors.
Computer programs embodying various features of the present invention may be encoded on computer-readable media for storage and/or transmission; suitable media include magnetic disk or tape, optical storage media such as compact disk (CD) or DVD (digital video disk), flash memory, and carrier signals for transmission via wired, optical, and/or wireless networks conforming to a variety of protocols, including the Internet. Computer-readable media encoded with the program code may be packaged with a compatible device such as a multi-processor graphics card or provided separately from other devices (e.g., via Internet download).
Thus, although the invention has been described with respect to specific embodiments, it will be appreciated that the invention is intended to cover all modifications and equivalents within the scope of the following claims.