Data bus bandwidth allocation apparatus and method

[0002] The invention relates generally to data bus bandwidth allocation apparatus and methods and more particularly to data bus bandwidth allocation apparatus and methods for allocating bandwidth for memory reads.

[0003] Data processing systems, such as computers, telecommunication systems and other devices, may use common memory that is shared by a plurality of processors or software applications. For example, with computers, a processor, such as a host processor, may generate data for other processing devices or processing engines and may need to communicate the processed data over a bus. In systems that also employ video processing or graphics processing devices, such systems may employ memory such as frame buffers which store data that is used by memory data requesters such as 3D drawing engines, display engines, graphic interface providing engines, and other suitable engines. Memory controllers use request arbiters to arbitrate among client's memory read and write requests to decide which client will receive data from system memory or from the frame buffer. A memory controller typically processes the memory requests to obtain the data from the frame buffer during memory reads or from an input data buffer (FIFO) associated with the host processor.

[0004] A problem arises when the amount of data delivered by the parallel access to system and local memory exceeds the data throughput capacity of the bus(es) providing the data transport from the memory controller to the clients. As such, memory controllers can receive more data from memory than they may be able to output to memory read backbones (e.g., buses) to the plurality of requesters requesting data from the frame buffer or from the host processor. As a result, collisions can occur creating an efficiency problem and potential data throughput bottlenecks. One solution is to add an additional bus to the memory read backbone for peak demand periods when, for example, all requesters are requesting data and their demands can be fulfilled by concurrent read activity to both frame buffer and system memory. However, such a system can be prohibitively costly due to layout complexity on an integrated circuit and may also require three (or more) port data return buffers that can accept read data on all ports every clock cycle. Some known processing systems use dual read buses to facilitate higher throughput from memory reads but would require a third bus to handle the peak bandwidth requirements.

[0005] The problem can be compounded when the host system's memory controller returns data in an unregulated fashion, namely, whenever the host processor does not make requests and the full bandwidth of the system memory becomes available. Typically, the memory controller can control the rate at which data is read from the frame buffer but has no control over when and how much data is available from the host processor. As such, there is no control over one data source but such systems typically have the ability to control the amount and frequency of data the memory controller obtains from the frame buffer memory. Such systems use FIFO data buffers to help reduce overflow problems, but even with deep buffers overflowing can not be avoided if the ratio between memory bandwidth and transport bandwidth is high. However, with real time requesters, such as audio and video processors, data can be lost if not processed when made available. Also, known systems, such as video and graphics processing systems, may include memory request sequencers which obtain the appropriate amount of data per request from a frame buffer over one or more channels. In addition, such systems may have a multiplexing scheme which multiplexes data from the host processor with data from the frame buffer so that it is passed to the memory read backbone to the requisite memory requesters. However, known systems typically encounter data collision conditions through such multiplexing schemes when a plurality of requesters are requesting data from the frame buffer and data from another source such as a host processor.

[0006] Consequently, there exists a need for a data bus bandwidth allocation apparatus that facilitates suitable bandwidth allocation in a system that has an unregulated bus, such as a bus from the host processor or other source, and a regulated bus such as a memory bus between a memory controller and a frame buffer memory or other requested bus.

[0007] The disclosed invention will be more readily understood in view of the below-identified figures, wherein

[0008] FIG. 1 is a block diagram illustrating one example of a data bus bandwidth allocation apparatus in accordance with one embodiment of the invention;

[0009] FIG. 2 is a block diagram illustrating one example of a sequencer with an adjustable delay in accordance with one embodiment of the invention;

[0010] FIG. 3 is a block diagram illustrating an example of a data issue rate regulator in accordance with one embodiment of the invention; and

[0011] FIG. 4 is a flow chart illustrating an example of one method of operation of the apparatus shown in FIG. 1.

[0012] Briefly, a data bus bandwidth allocation apparatus and method uses buffer entry feedback data from a buffer, such as an overflow buffer, that receives requested data over an unregulated bus. The data bus bandwidth allocation method and apparatus generates data issue delay data based on the buffer entry feedback data to adjust data commands to a data source, such as frame buffer memory used to feed a regulated bus. In one embodiment, the data issue rate regulator utilizes a programmable threshold corresponding to a threshold of data entries in a FIFO buffer. The overflow FIFO has feedback indicating, for example, the number of free entries or the number of full entries. The feedback data is used by the data issue rate regulator. The data issue rate regulator provides data rate regulation information to an adjustable delay sequencer. The adjustable delay sequencer selectively throttles data reads from the frame buffer memory so that data collisions do not occur over the memory read backbone. As such, the data that is provided over the unregulated bus is allowed to freely be sent over the memory read backbone. The rate at which data is obtained from the frame buffer is regulated under control of the adjustable delay sequencer which adjusts the amount of delay between consecutive memory reads to facilitate an efficient bandwidth allocation of the memory read backbone.

[0013] FIG. 1 illustrates an apparatus 10 incorporating one example of a data bus bandwidth allocation apparatus that includes a buffer 12, such as an overflow FIFO, operatively coupled to receive requested data from an unregulated bus, a data issue rate regulator 16 operatively coupled to receive buffer entry feedback data 18 from the buffer 12, and an adjustable delay sequencer 20 operative to adjust data read commands to a data source, such as frame buffer 22 that feeds the regulated bus 26aand 26b. In this example, the regulated bus 26aand 26bis a memory command bus.

[0014] The sequencer 20 with the adjustable delay control receives data issue delay data 24 from the data issue rate regulator 16. The data issue delay data 24 indicates, for example, the amount of delay that the sequencer 20 needs to provide for adjusting data read commands over regulated channels 26aand 26bfrom the frame buffer 22 to allow all of the data from the unregulated bus to be transferred over the memory read backbone 25. As such, the adjustable delay sequencer 20 adjusts data command flow over the regulated bus 26aand 26bin response to the data issue delay data to avoid data collision between data returned from frame buffer memory 22 and data returned from FIFO buffer 12.

[0015] The buffer entry feedback data 18 may represent, for example, the number of empty entries left in the buffer 12, also referred to herein as the uncontrollable request channel FIFO. The entry feedback data 18 may also be the number of full entries or any other suitable feedback data. The data issue rate regulator 16 utilizes a programmable threshold 28 to control the amount of rate regulation. As such, if the threshold 28 is adjusted, the rate at which the sequencer issues read commands for the frame buffer will vary accordingly. The entry feedback data 18 and stored threshold value 28 may be provided by a state machine or may take any suitable form.

[0016] As shown, the system may also, if desired, include a plurality of request FIFOs 30a-30n, which store read requests from a plurality of memory requesters that generate memory requests as known in the art. For example, memory requesters may include video processing engines, 3D graphics drawing engines, GUI engines or any other suitable data processing mechanism which requests data to be read from memory. Also as known in the art, a typical request may include, for example, the address to be read from memory, the operation code such as whether the request is a read or a write request, the size or amount of data to be read, and any other suitable data. The request FIFOs 30a-30nreceive read requests 32a-32nand store them in a first in, first out fashion. A request arbiter 34 receives the request and distinguishes between a request from a requester requesting data over the unregulated bus 14 and a requester requesting data over a regulated bus 26aand 26b. The memory command bus 26aand 26bbetween the frame buffer and the sequencer is considered to be regulated since the rate at which data may be obtained over the bus is controlled by the control sequencer or other suitable device. In contrast, the unregulated bus 14 provides data to the FIFO 12 at a rate that is uncontrollable by the sequencer or other suitable device . The arbiter routs the request for the unregulated channel to a host processor interface 40 which notifies the host that data is requested.

[0017] In addition to controlling the data read commands, the sequencer 20 also generates a data select signal 46 to control selection of data that is communicated over the memory read backbone 25. A multiplexing circuit 48 is responsive to the data select signal to select data either from the buffer 12 that receives data from the unregulated bus or from the frame buffer 22. Accordingly, the FIFO 12 is coupled to provide data 50 received over the unregulated bus 14 to the multiplexing circuit 48. In addition, the frame buffer is coupled to the multiplexer 48 through one or more read backbone buses 52aand 52b. These datapaths may also include FIFO buffers to avoid data loss during unavoidable collisions.

[0018] The unregulated bus 14 may be any suitable bus such as an AGP and/or PCI type bus as known in the art or any other suitable bus(es). In this embodiment, the unregulated bus has a bandwidth of approximately 1 gigabyte per second and the regulated bus has a bandwidth of approximately 4.5 gigabytes per second (two 64 bit channels) so that the bandwidth of the unregulated bus is less than the bandwidth of the regulated bus. However, it will be recognized by one of ordinary skill in the art that any suitable bandwidths may also benefit from the disclosed invention.

[0019] FIG. 2 illustrates one embodiment of the sequencer 20 with the adjustable delay to variably control the rate at which data is obtained for a memory request. The sequencer 20 includes an input FIFO 200 that receives the memory command 202 included in the selected request 42. The sequencer also includes a cycle issue controller 204 and a cycle sequencer 206.

[0020] The memory command input FIFO 200 stores received memory commands on a first in first out basis. The data in the memory command is the data that is typically included in memory requests. As such, read cycle parameters 208 are determined from the memory command and used by the cycle sequencer 206. Read cycle parameters may include, for example, the read address, the size of the data or the amount of data to be obtained, and any other suitable information. As such, the input FIFO 200 serves as a memory request input command FIFO that stores memory request commands for obtaining data over the regulated bus.

[0021] The cycle issue controller 204 receives the data issue delay data 24 from the issue rate regulator and receives cycle request data 208. The issue delay data 24 represents an additional time delay before starting a memory cycle, this results in a delay of the returned data by the same amount of time. The cycle issue controller 204 generates start cycle data 210 indicating when to start the read cycle for reading memory from the frame buffer. The cycle issue controller is a finite state machine that traverses a sequence of operational states according to rules determined by the type of memory, the requested memory cycle and a set of timing rules. The timing rules, as known in the art, include the sequence timing of signals to a RAM interface.

[0022] The cycle sequencer 206 receives the start cycle data 210 along with the cycle parameters 209 and generates the requisite control signals over the regulated bus 26aand 26bto control the read operation from the frame buffer. The control signals are indicated as 212. In addition, the cycle sequencer 206 generates the data select signal 46 to also control when the data that is read from the buffer 12 is passed to the memory read backbone 25. The cycle issue controller 204 and cycle sequencer 206 may be any suitable hardware or software or combination thereof.

[0023] In operation, the sequencer 20 stores the memory request commands for at least one controlled channel in the memory request command FIFO. It then generates the start cycle data in response to the data issue delay data and in response to the cycle request data that is stored in the memory request input command FIFO. The sequencer produces a frame buffer read control signal 212 and a data select signal 46 in response to the start cycle data and in response to the cycle parameters to select data from the frame buffer or data from the buffer 12 based on the data issue delay data.

[0024] FIG. 3 illustrates one embodiment of the data issue rate regulator 16 which receives the buffer entry feedback data 18 and the threshold 28 to determine whether a delay is necessary to allow the data on the unregulated bus to be transferred over the memory read backbone 25 to avoid data collision. In this example, the data issue rate regulator 16 includes a comparator circuit 300, a gain factor table 302, and a gating circuit 304. In operation, the comparator 300 compares the number of entries that have been used in the buffer 12 to the threshold to determine whether or not the number of entries that are filled is greater than the threshold. If the number of entries that are filled is greater than the threshold, and a gain factor associated with that number of entries indicates that a delay is required, and the issue delay data 24 is generated indicating the amount of delay required. The gain factor table may be, for example, a table that is accessed to determine the amount of delay necessary for a given number of entries that exceed the threshold. For example, a larger delay may be necessary if the number of entries that has exceeded the threshold is large whereas a smaller delay may be required if the number of entries exceeds the threshold by a lower amount.

[0025] FIG. 4 illustrates one example of the operation of the apparatus in FIG. 1 wherein the apparatus determines whether it has received a memory request as shown in block 400. If a memory request has been received from the arbiter, a determination is made as to whether the request target is the system memory or the frame buffer. If the request is directed towards the frame buffer, a request to the sequencer 20 is made.

[0026] If the sequencer is ready to accept the request it acknowledges the cycle using the RTR/RTS handshake 25. A counter is then set equal to the delay as shown in block 406. The sequencer then determines whether the delay has expired as shown in block 408. If the delay has not expired, a counter is decremented as shown in block 410 and the sequencer holds the command data for the frame buffer until the delay period has expired, as shown in block 412. If the delay period has expired, the read cycle is started as shown in block 414. The process repeats each time the feedback data and the threshold information indicate that a delay is necessary. It will be recognized that only data generating commands or data reads to memory need to be delayed. Refresh commands or other commands that do not deliver data back to a requester need not be delayed.

[0027] It should be understood that the implementation of other variations and modifications of the invention in its various aspects will be apparent to those of ordinary skill in the art, and that the invention is not limited by the specific embodiments described. It is therefore contemplated to cover by the present invention, any and all modifications, variations, or equivalents that fall within the spirit and scope of the basic underlying principles disclosed and claimed herein.