Render-cache controller for multithreading, multi-core graphics processor

An embodiment of the present invention relates generally to computer graphics architecture, and more particularly, to a method and apparatus for rendering graphics. Other embodiments are also described.

In the field of computer graphics, rendering refers to the process of adding realism to computer graphics by adding three-dimensional (3D) qualities, such as shadowing, color variation, and shade, to objects displayed on a two-dimensional display. Modern 3D graphics processors are commonly used to render 3D graphical images for display. Rendering is typically accomplished by breaking the objects up into a series of primitives such as polygons, typically, triangles. For each pixel that makes up the polygon, attribute values are assigned for attributes such as color, specular color, alpha (transparency), fog, surface texture, and depth. The attribute values are then combined, or otherwise processed, by the graphics processor to achieve a precise pixel value for each pixel that makes up the object. To render a realistic image, often several processing operations, involving a variety of attributes, must be performed for each individual pixel that makes up the object or image.

A graphics processor is generally limited by the clock speed with which it can process the individual pixels of a 3D computer image. One way in which system designers have improved the efficiency of graphics processors is by designing processors to perform multiple pixel processing operations at the same time. For example, to increase the overall efficiency of graphics processors, system designers have developed multithreading, multi-core graphics processors. As the name suggests, a multithreading, multi-core graphics processor has multiple cores, or pixel processing units, that operate on pixels in parallel. Each core of the engine is directed to process pixels by a stream of instructions referred to as a thread. One of the advantages of a multithreading, multi-core approach to pixel processing is that the graphics processor can switch between threads if, for example, one thread is required to wait for pixel data to be fetched from main memory.

To further improve efficiency, many graphics processors have been designed with an on-chip cache to store pixel data that is being processed. In particular, if the on-chip cache contains the pixel data that is required for processing, the processing occurs more quickly because the graphics processor need not wait for the pixel data to be fetched from main memory every time it is required for a pixel processing operation. The reduced number of main memory accesses improves the overall efficiency of the graphics processor.

However, implementing a caching technique with a multithreading, multi-core graphics processor poses several challenges. Due to thread-switching in a multithreading, multi-core graphics engine, the order in which pixel processing operations are completed may be different than the order in which the threads are dispatched to the multi-core engine. For example, if two threads are dispatched to the multi-core engine to perform processing on the same pixel (e.g., each thread specifies the same x,y coordinates, representing the same pixel), due to thread-switching, the pixel processing operations are not guaranteed to be performed in the order in which the threads are dispatched to the multi-core engine. Consequently, if the pixel processing operations are performed out of order, the resulting pixel data may be incorrect.

Another problem is maintaining pixel data coherency for the cache and main memory. For example, after a pixel processing operation is performed and the resulting pixel value is written to the cache, the value for that particular pixel, as stored in the cache, will be different than the corresponding value stored in main memory. Consequently, a main memory read operation results in an improper pixel value being read.

One solution to these problems is to implement two separate cache controllers, for example, one cache controller in the graphics engine to maintain data coherency between the render-cache and main memory, and a separate controller to maintain the order in which threads are dispatched to the core. However, a problem with this prior art solution is the overall size of the logic required for implementing the solutions. For example, because each cache-controller has its own content addressable memory, the overall size of the logic (e.g., number of gates required to implement the solution) is larger than desired.

Reference in the specification to “an embodiment” or “one embodiment” means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the invention. The appearance of the phrase “for one embodiment” or “in one embodiment” in various places in the specification does not necessarily refer to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments. Embodiments of the present invention are illustrated by way of example, and not by way of limitation, in the figures of the accompanying drawings, in which like references indicate similar elements, and in which:

FIG. 1 is a block diagram illustrating a computer graphics system including a multithreading, multi-core graphics processor;

FIG. 2 is a combination block/data flow diagram illustrating, for one embodiment of the present invention, the flow of pixel data through a multithreading, multi-core graphics engine;

FIG. 3 is a block diagram illustrating a render-cache controller for one embodiment of the present invention; and

FIG. 4 is a flow diagram illustrating a method for pre-allocating pixel data to a streaming render-cache for processing by a multithreading, multi-core graphics engine.

An embodiment of the invention is directed to a method and apparatus for rendering three-dimensional (3D) graphics using a streaming render-cache with a multi-threaded, multi-core graphics processor. In the following description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of the present invention. It will be evident, however, to one skilled in the art that the present invention may be practiced without these specific details.

Referring to FIG. 1 , a block diagram of a computer graphics system 10 including a multithreading, multi-core graphics processor 18 for rendering 3D graphical images is illustrated. The computer graphics system 10 includes a central processing unit (CPU) 12 connected with a main memory 14 via core logic 16 , as well as a graphics processor 18 . The graphics processor 18 includes a multithreading, multi-core graphics engine 20 , a render-cache controller 22 , and an associated streaming render-cache 24 . The core logic 14 , sometimes referred to as a bridge or bridge logic, controls the exchange of data between the main memory 14 , the CPU 12 and the graphics processor 18 , in addition to handling input and output functions for the system. Generally, the graphics processor 18 determines the graphic information to be sent to a display, based on instructions and data received from the CPU 12 and the main memory 14 .

For one embodiment of the present invention, the CPU 12 executes a series of instructions directing the graphics processor 18 to generate one or more images for display. Accordingly, the CPU 12 communicates instructions to the graphics processor 18 identifying the location in the main memory 14 of the data from which to generate the graphic image or images. The data may include geometric shapes defined by a fixed set of vertices, each vertex being assigned attribute values for a variety of attributes. For example, the attributes may include, but are not limited to: color, specular color, alpha (transparency), fog, surface texture, and depth. Based on the attribute values of each vertex, the graphics processor 18 traverses the pixels of the polygon and assigns attribute values for corresponding attributes to each pixel comprising the polygon to be rendered. Theses attribute values, generically referred to herein as pixel data, are the operands, or inputs, for pixel processing operations. The multithreading, multi-core graphics engine may be capable of a wide variety of pixel processing operations.

FIG. 2 is a combination block/data flow diagram illustrating, for one embodiment of the present invention, the flow of pixel data through a multithreading, multi-core graphics engine 20 . The graphics processor 18 illustrated in FIG. 2 includes raster logic 30 that generates threads (e.g., a sequence of pixel processing instructions) that are dispatched by a thread dispatcher 34 to the multithreading, multi-core graphics engine 20 . The graphics engine 20 includes multiple cores, and each core may operate in parallel and may be capable of thread-switching. While the graphics engine 20 shown in FIG. 2 includes six cores, it will be appreciated that in practice the actual number of cores may vary.

Each core within the multi-core graphics engine 20 performs pixel processing operations on pixel data based on instructions received via threads from the thread dispatcher 34 . Each thread identifies the location of the pixel data to be processed, by indicating one or more cache-line addresses of the render-cache 24 where pixel data is stored. Each time a thread becomes active within a core, the graphics engine 20 performs a pixel data access operation 36 to fetch pixel data from the render-cache 24 . Furthermore, each core of the graphics engine 20 is capable of thread-switching. For example, in FIG. 2 , the references “T 0 ” and “T 1 ” within each core represent different threads. While each core can only process one thread at any given moment in time, each core has the ability to switch amongst the threads if, for example, the pixel data required for a pixel processing operation associated with a particular thread is not yet stored in the render-cache 24 . In such a case, a core may switch from one thread to another, for example, from thread T 0 to thread T 1 .

Once the graphics engine 20 has completed a pixel processing operation, it may write the processed pixel data to the same location in the render-cache 24 from which it was originally read. Consequently, as new threads are dispatched, if a cache hit occurs, the processed pixel data may be used in a subsequent pixel processing operation without having to be fetched from main memory 14 or some other slower graphics memory (not shown). However, if the processed pixel data is not required in a subsequent pixel processing operation, it may eventually be written to main memory, or a graphics memory (not shown) so that it can be output to a display if necessary.

As briefly described above, the graphics processor 18 includes raster logic 30 . For one embodiment of the present invention, the raster logic 30 analyzes data representing an object to be rendered by traversing, or walking, a primitive and generating pixel data (e.g., attribute values) for each pixel that is part of the primitive.

As the raster logic 30 traverses the primitive, it generates threads, which when executed within the multithreading, multi-core graphics engine 18 , cause the individual cores to perform pixel processing operations (e.g., pixel shading operations) using sub-spans that are, for example, 2 pixels by 2 pixels. However, before each thread is dispatched to the multi-core engine 20 , the raster logic 30 performs a pre-allocation and in-flight check routine 32 to allow the pixel data for each pixel of the sub-span to be in the render-cache 24 and readily accessible to the multi-core engine 20 when the thread is received by the multi-core engine 20 . By pre-allocating the pixel data to the render-cache, the latency of the main memory 14 access is hidden from the multi-core graphics engine 20 .

The routine 32 may involve two basic operations. First, the render-cache controller 22 is checked to determine whether the pixel data for each pixel in the sub-span has been previously allocated to the render-cache 24 . If the render-cache controller 22 indicates that the pixel data, corresponding to a particular pixel to be processed in connection with the thread, has not been previously allocated (e.g., a cache miss), then the pixel data corresponding to that particular pixel is allocated to the render-cache 24 . However, if the render-cache controller 22 indicates that the pixel data, corresponding to a particular pixel to be processed by the thread, has already been allocated to the render-cache 24 (e.g., a cache hit), then the render-cache controller 22 determines the state of the previously allocated pixel data.

In particular, the render-cache controller 22 determines: (1) whether the previously allocated pixel data is waiting to be processed by the graphics engine (e.g., not yet read by the graphics engine), (2) whether the previously allocated pixel data has been read by the graphics engine, but the result of the pixel processing operation not yet written back to the render-cache 24 , or alternatively, (3) whether the previously allocated pixel data has already been read, processed by the graphics engine 20 and written back to the render-cache 24 . For example, the possibility exists that the pixel data in the render-cache 24 was allocated in connection with a previously dispatched thread that has yet to be processed by the graphics engine 20 . This situation is referred to as a pixel “in-flight”, or not yet “retired” from the core. For example, a pixel is said to be “in-flight” when it is in a transitive state, meaning that it has been read from the render-cache 24 by the graphics engine 20 , but not yet processed and written back to the render-cache 24 . For one embodiment of the present invention, the render-cache controller 22 prevents a thread from being dispatched if any pixel data corresponding to pixels in the sub-span being processed by the particular thread are in-flight. This ensures that pixels are processed in the proper order, particularly when multiple threads require access to pixel data associated with pixels having the same X and Y coordinates and the result is dependent on the order of processing.

FIG. 3 is a block diagram illustrating a render-cache controller 22 for one embodiment of the present invention. The render-cache controller 22 includes a cache-line status array 40 , a pixel mask array 42 , and a content addressable memory (CAM) 44 . Together, these three components maintain the data coherency of the render-cache and ensure that threads are dispatched to the graphics engine 20 in the proper order.

The CAM 44 maps X and Y pixel coordinates to corresponding cache-line addresses in the render-cache 24 . For example, the CAM 44 receives as input data representing the pixel coordinates of a pixel. If the CAM 44 has a matching entry, the CAM 44 outputs an address representing the location in the render-cache where the pixel data associated with the pixel coordinates is stored. For one embodiment of the invention, the CAM 44 is as wide as 30 bits and has a depth of 128 entries. In order to operate at a high frequency, the look-up function, or comparator function, of the CAM 44 can be done in parallel and can be finalized over more than one clock cycle in a pipelined manner. For one embodiment of the present invention, the CAM 44 is fully associative, meaning that any main memory address has the full freedom to be replicated at any address in the render-cache 24 .

The pixel mask array 42 indicates whether pixel data associated with a particular pixel stored in the render-cache 24 is in-flight. For example, before dispatching a thread to perform a pixel processing operation on a particular pixel stored in the render-cache 24 , the pixel mask array 42 is checked to determine whether the particular pixel in render-cache 24 is waiting to be processed by the graphics engine 20 in connection with a previously dispatched thread. For one embodiment of the present invention, the pixel mask array 42 has the same depth, or number of entries, as the CAM 44 . Furthermore, for one embodiment of the invention, each entry comprises a single bit corresponding to a cache-line in the render-cache 24 . Whether the bit is set or not determines whether the pixel data stored in the corresponding cache-line is in-flight. For example, for one embodiment of the invention, a bit in the pixel mask array 42 that is set indicates that the pixel data stored in the cache-line associated with the bit in the pixel mask array 42 is in-flight. Consequently, any thread instructing the graphics engine 20 to perform a pixel processing operation on pixel data stored in that particular cache-line of the render-cache 24 will be blocked from being dispatched until the graphics engine 20 has completed processing the pixel data and written the processed pixel data back to the cache-line of the render-cache 24 .

For one embodiment of the present invention, the pixel mask array 42 is updated, or reset, when the graphics engine 20 writes the resulting processed pixel data to the render-cache 24 . For example, for one embodiment of the invention, when the graphics engine 20 accesses the render-cache 24 during a write operation, the cache-line address that is being written to is communicated to the render-cache controller 22 via a pipeline 38 . Consequently, the render-cache controller 22 determines that the cache-line contains processed pixel data and accordingly, the bit in the pixel mask array 42 corresponding to the cache-line address is cleared, or reset. Similarly, when the graphics engine 20 reads a particular cache-line, the address of the cache-line is pipelined to the render-cache controller 22 and the bit in the pixel mask array 42 corresponding to the cache-line is set, indicating that the pixel data associated with the cache-line address is in-flight.

For one embodiment of the present invention, the cache-line status array 40 is a single bit array of the same size of the CAM 44 . Like the pixel mask array 42 , each bit in the cache-line status array 40 corresponds with a cache-line in the render-cache 24 and indicates whether the graphics engine 20 has accessed the pixel data stored at the address of the cache-line. A bit in the cache-line status array 40 is set when pixel data at the cache-line address corresponding to the bit is in-flight, and reset when not in-flight. The cache-line status array 40 is used to determine available cache-lines during the pre-allocation routine. For example, after a cache miss occurs, the cache-line status array 40 is checked to determine a cache-line address of the render cache 24 that is available to allocate new pixel data. The entries in the cache-line status array 40 indicate which cache-line addresses are currently in use, and therefore unavailable to the allocation routine. An available cache-line address may be selected based on one of many well-known cache-aging algorithms, such as the least recently used algorithm. The cache-line that is selected must be scheduled for eviction. For example, the pixel data that is stored at the selected cache-line address must be written to main memory 14 , or a graphics memory (not shown) before new pixel data is written to the cache-line address.

For one embodiment of the present invention, the size of the render-cache 22 and associated CAM 44 can be selected based on the maximum number of sub-spans being processed at any given time within the graphics engine 20 . The maximum number of sub-spans being processed by the graphics engine 20 at any given time can be used as an upper bound to size the render-cache 24 and the associated CAM 44 . For example, if the raster logic 30 generates sub-spans that are 2 pixels by 2 pixels, four entries are required in the render-cache 24 and the CAM 44 for every active thread in the engine.

FIG. 4 is a flow diagram illustrating a method 50 for pre-allocating pixel data to a streaming render-cache for processing by a multithreading, multi-core graphics engine. At operation 52 , the method begins with a cache tag comparison. For each pixel being processed, based on the pixel's X and Y coordinates, a look-up operation is performed to determine whether the particular pixel has been previously allocated to the render-cache.

If the cache-tag comparison results in a cache miss, then at operation 60 , an available cache-line address is selected according to a cache-scheduling or cache-aging policy. At operation 62 , the pixel data stored in the selected cache-line is evicted, or written to main memory 14 . At operation 64 , new pixel data is read from main memory and written to the selected, available cache-line address of the render-cache. Next, at operation 64 , the CAM is updated with the render-cache address where the particular pixel is stored. In addition, at operation 56 , the pixel mask array is updated to indicate that pixel data associated with the particular pixel is now waiting to be processed. Finally, at operation 58 , a thread is dispatched to the multi-core engine. The thread includes the cache-line address indicating the location in the render-cache where the graphics engine can access the pixel data associated with the particular pixel.

If, however, the cache-tag comparison results in a hit, then at operation 54 , the pixel mask array is checked to determine whether the previously allocated pixel data is in flight. For example the pixel mask array is checked to determine whether the graphics engine has already processed the previously allocated pixel data, or whether the pixel data is waiting to be processed. If the previously allocated pixel data is still waiting to be processed, then the render-cache controller 22 blocks the thread from being dispatched. Only after the pixel mask array indicates that previously allocated pixel has been retired from the core is the thread dispatcher allowed to dispatch a thread including the cache-line address of the particular pixel. For example, if at operation 54 , the pixel mask array indicates that the previously allocated pixel data has already been processed, then at operation 56 , the pixel mask array is updated to indicate that the pixel data is now waiting to be processed. Accordingly, at operation 58 , a thread is dispatched to the graphics engine including the render-cache address where the pixel data is located.

The embodiments of the invention described above provide several advantages over prior art multithreading, multi-core graphics processors. One advantage is that the need for having two separate CAMs is eliminated. For example, there is no need to have one CAM, at the thread dispatch point, to control the order in which threads are dispatched, and a second CAM, at the multi-core engine, to maintain the data coherency of the render-cache. A second advantage of the render-cache controller described herein is that it provides control for a fully associative render-cache that requires only one lookup operation, while maintaining data coherency by allowing only one cache-line in-flight. In addition, the render-cache controller is advantageous because it can be sized based on the working set of the multi-core engine and it hides the latency of the main memory by pre-allocating pixel data to the render-cache, thereby taking advantage of the core's thread-switching and compute cycles. Finally, the render-cache controller takes advantage of the spatial locality of pixel data in main memory. For example, because the render-cache controller maintains the order in which threads are dispatched, cache hits are likely to occur. If, however, a cache miss does occur, the main memory access should not require a page swap, because the required pixel data will likely be found within the page that is in main memory.

Thus, a method and apparatus for rendering graphics using a render-cache with a multi-threaded, multi-core graphics processor is provided with reference to specific exemplary embodiments. It will be evident that various modifications and changes may be made to theses embodiments without departing from the broader spirit and scope of the invention. Accordingly, the specification and drawings are to be regarded in an illustrative rather than a restrictive sense.