The invention relates to a device and a method for writing data elements from a coprocessor into a FIFO, a device and method for reading data elements from a FIFO memory into a coprocessor as well as a multiprocessing computer system.

New applications and services like High Definition TV (HDTV), Broadcast Satellite Services, Video-conferencing, Interactive Storage Media etc. require a huge amount of data processing in the video domain as well as in the audio domain. Thus, the data rates involved in those applications make computation very time consuming.

When coping with the computing requirements in signal processing, designers usually follow two architectural approaches, namely the dedicated and the programmable approach. The dedicated architecture aims to fully exploit the computational features of algorithms. VLSI implementations of dedicated architecture are optimised with regard to area, power and performance. The advantage of a dedicated architecture is that they provide a good performance, while they lack the flexibility to further extend an algorithm set. On the other hand the programmable approach can satisfy the requirement for a greater flexibility, since algorithms can be further extended simply by applying the necessary software modifications. Due to the fact, that a number of applications can run on the hardware, the cost per application hardware is reduced. The disadvantage of a programmable architecture is that the architecture as well as application models must be faster, since the computational properties of algorithms are not fully exploited. This problem can be solved by implementing multiprocessing strategies. It is therefore essential to provide an application model which has sufficient parallelism in order to fully exploit the computational resources of a multiprocessor architecture. One such application model which provides parallelism for signal processing applications is the Kahn process network model. An significant increase in processing speed can be achieved as soon as the application are executed in a multiprocessor architecture.

A typical Kahn process network is composed of a number of process networks, processes and FIFO memories (First In-First Out). In such a process network a process is implemented by a processor or a coprocessor and a FIFO represents a communication channel, i. e. the processors communicate with each other via FIFO buffers.

The FIFOs have a single reader and single writer. A process usually defines the functionality

of the application. It interacts with its environment through input and output ports. A process can be in one of three states, namely running, blocked or ready. For example, a process will be blocked if it tries to read from an empty FIFO or if it tries to write into a full FIFO.

In the Kahn process network the processes communicate with each other in a pipelined manner, i. e. the slowest process in the pipeline determines its throughput. If the computation is properly balanced among processes if will reduce the effective parallelism within the application model.

Usually, processes communicate with each other via FIFOs. Now when these processes are implemented as software processes, an implementation of a FIFO includes a buffer for storing the data, a read pointer, a write pointer as well as room and data semaphores for synchronising data and room in the buffer. Therefore, writing into a FIFO is implemented firstly by waiting for free room, i. e. space, in said FIFO. When there is room available in said FIFO, data is input into the FIFO and the write pointer is incremented.

Finally, it is signalled that valid data has been added into the FIFO. Reading from a FIFO is symmetrical to writing, i. e. the roles of room and data as well as the writer and reader pointer have changed.

In a multiprocessing environment there is usually a conflict for the available communication resources, like buses etc.. The communication expenses increase as the number of processors in the architecture increases. Therefore optimising the application model with regard to a reduced communication will result in an improved speed. A reduced communication is achievable by carefully choosing the token structure and controlling the number of tokens transferred over the communication channels.

It is therefore an object of the invention to improve the communication between a coprocessor and a controller in a multiprocessing environment.

This object is solved by a device for writing data elements from a coprocessor into a FIFO memory according to claim 1, a method for writing data elements from a coprocessor into a FIFO memory according to claim 7, a device for reading data elements from a FIFO memory into a coprocessor according to claim 12, a method for reading data elements from a FIFO memory into a coprocessor according to claim 18, and a multiprocessing computer system according to claim 23.

The invention is based on the idea to maintain two counters for an input or output port of a FIFO.

According to the invention, a device for writing data elements from a coprocessor into a FIFO memory is provided. Said device is embedded in a multiprocessing

environment comprising at least one coprocessor, a FIFO memory and a controller. Said device comprises a first counter for counting the available room in said FIFO memory, and a second counter for counting the number of data elements written into said FIFO memory.

Said device further comprises a control means for checking said first counter for available room in said FIFO memory, and for checking said second counter whether a predetermined number N of data elements have been written into said FIFO memory. Said control means decrements the count of said first counter and increments the count of said second counter, after a data element has been written into said FIFO memory. Said device finally comprises an output means for outputting data elements to said FIFO memory. Said control means issues a first message when the count of said second counter has reached said predetermined number N and issues a first call for available room in said FIFO memory to said controller.

Said output means forwards said first message and/or said first call to said controller.

The advantage of a writing device according to the invention is that by using said first and second counter for writing data into the FIFO memory there is no need for a coprocessor performing a write operation to request the controller for two semaphore operations (a P and a V operation) when writing to said FIFO, thereby reducing the communication traffic between coprocessor and controller.

According to an embodiment of the present invention said first message indicates that sufficient data elements have been written into said FIFO memory. Thus advantageously the communication traffic between coprocessor and controller is further reduced since said message is only transmitted after N data elements have been written into said FIFO.

According to a further embodiment of the present invention, said control means increments a write pointer, when data elements were output to said FIFO memory.

According to still a further embodiment of the present invention, said control means performs a wrap-around test after said write pointer was incremented.

According to an embodiment of the present invention, said control means resets said second counter after issuing said first message, i. e. the counter is reset after its count has reached N and said message has been issued so that the count can start again.

According to a further embodiment of the present invention, said control means issues said first call for available room in said FIFO memory to said controller before said count of said first counter becomes zero. This is advantageous since the latency of such a call is hidden.

The invention also concerns a method for writing data elements from a coprocessor to a FIFO. Said method corresponds to the device for writing data elements as described above.

Moreover, the invention also provides a device and method for reading data elements from a FIFO into a coprocessor. Said device and method are complementary to the writing device and writing method. Here, data is transferred from the FIFO to the coprocessor. Instead of the first message and first call, a second message indicating that sufficient data elements have been read from said FIFO, i. e. that room is available and a second call requesting available data elements in said FIFO, is issued.

According to the invention a multiprocessing computer system according to claim 23 is also provided. Furthermore, the invention also provides a computer program according to claim 24 for implementing said methods.

The invention will now be explained in more detail with reference to the drawing, in which: Fig. 1 shows a block diagram of a multiprocessing computer system according to an embodiment of the invention, Fig. 2 shows a block diagram of a writing device associated to coprocessors in Fig. 1, and Fig. 3 shows a block diagram of a reading device associated to coprocessors in Fig. 1.

Fig. 1 shows a block diagram of a multiprocessing computer system according to an embodiment of the invention. Said computer system comprises memory banks 1, preferably in form of FIFO memories, three CPU's 3, three coprocessors 2, a controller 4 and an interconnect, like a bus, for connecting said FIFO memories with the CPU's 3, the coprocessors 2 and the controller 4. Said controller 4 is preferably embodied as a programmable processor executing semaphore operations on request of the coprocessors 2 and preferably communicates with the coprocessors 2 by means of a dedicated ring network.

Part of the processes are implemented in software and run on a programmable processor, i. e. a CPU 3, while the other part of the processes are implemented in dedicated hardware and

run on coprocessors 2. The memory 1 is shared between the processors 3 and the coprocessors 2 and is usually a random access memory.

For more details on this so called CAKE architecture please refer to "Exploring Design Space of Parallel Realizations: MPEG-2 Decoder Case Study", by Basant et al. in Proc. of CODES 2001, Copenhagen, April 2001, pp. 92-97.

Fig. 2 shows a block diagram of a writing device associated to coprocessors 2 in Fig. 1. Said writing device comprises a first counter 10 and a second counter 20, a control means 30 and an output means 40, wherein said control means 30 is connected to said first and second counter 10,20 and said output means 40, respectively. Said control means 30 has a terminal 31 to which a coprocessor 2 can be connected. Said output means 40 has a first terminal 41 connectable to said FIFO 1 and a second terminal 42 connectable to the controller 4, wherein said connection is preferably implemented via a dedicated ring network.

Said first counter 10 is used to count the available room in said FIFO 1, i. e. the count of said first counter 10 is room available. It contains the number of free spaces in said FIFO 1, into which data can be written to before a call for free space or room need to be issued. Said second counter 20 is used to count the available data in said FIFO 1, i. e. the count of said second counter 20 is data_produced. It contains the number of data that has been written into the FIFO 1 but which have not been reported to the controller 4 yet.

First of all, the control means 30 checks the count of the first counter 10, i. e. the room-available counter 10. If the count of the first counter 10 is zero, the control means 30 issues a first call C1, i. e. a get-room call, to the controller 4 asking for room in the FIFO 1. If the controller 4 return a message that no room is currently available in the FIFO, the control means 30 repeats to issue a get-room call until the controller 4 returns that there is space available in the FIFO 1. When there is space available in said FIFO the output means 40 outputs data to the FIFO 1, where the data is stored. Next the control means 30 increments a write pointer and performs a pointer wrap-around test.

The wrap around test is performed by comparing the write/read pointer with a end-of-buffer address and if they equal each other, the write/read pointer is set to the begin- of-buffer address. Preferably the begin-of-buffer as well as the end-of-buffer addresses are stored in the coprocessor.

Thereafter, the control means 30 decrements the count of the first counter 10, i. e. the room-available counter, and increments the count of the second counter 20, i. e. the data-produced counter. The control means 30 determines whether the count of the second counter 20 has reached a predetermined number N, indicating that N data elements have been

written into the FIFO and are hence available to be read from said FIFO 1. As soon as the count of the second counter 20 has reached the predetermined number N, the control means 30 issues a first message Ml, i. e. put data, to the controller 4. Alternatively, the put data message may have an argument indicating how many data have been produced.

Fig. 3 shows a block diagram of a reading device associated to coprocessors 2 in Fig. 1. Said reading device comprises a third counter 50 and a fourth counter 60, a control means 70 and an input means 80, wherein said control means 70 is connected to said third and fourth counter 50,60 and said input means 80, respectively. Said control means 70 has a terminal 71 to which a coprocessor 2 can be connected. Said input means 80 has a first terminal 81 connectable to said FIFO 1 and a second terminal 82 connectable to the controller 4, wherein said connection is preferably implemented via a dedicated ring network.

Said third counter 50 is used to count the available data in said FIFO 1, i. e. the count of said first counter 10 is data available. It contains the number of free data elements in said FIFO 1 which can be read before a call for data need to be issued. Said fourth counter 60 is used to count the available free space or room in said FIFO 1, i. e. the count of said second counter 20 is room_produced. It contains the number of data that have been read from the FIFO buffer 1 but which have not been reported to the controller 4 yet.

First of all, the control means 70 checks the count of the third counter 50, i. e. the data-available counter. If the count of the third counter 50 is zero, the control means 70 issues a second call C2, i. e. a get-data call, to the controller 4 asking for data in the FIFO 1.

If the controller 4 return a message that no data is currently available in the FIFO 1, the control means 70 repeats to issue a get data call until the controller 4 returns that there is data available in the FIFO 1. When there is data available in said FIFO 1 the input means 80 inputs or reads data from the FIFO 1. Next the control means 70 increments a read pointer and performs a pointer wrap-around test.

Thereafter, the control means 70 decrements the count of the third counter 50, i. e. the data-available counter, and increments the count of the fourth counter 60, i. e. the room_produced counter. The control means 70 determines whether the count of the fourth counter 60 has reached a predetermined number N, indicating that N data elements have been read from the FIFO 1 and that hence there are spaces available into which data can be written to. As soon as the count of the fourth counter 60 has reached the predetermined number N, the control means 70 issues a second message M2, i. e. put room, to the controller 4.

Alternatively, the put room message may have an argument indicating how many data has been produced.

Accordingly, the controller 4 implemented as a programmable processor must be able to accept a get-room call and a put data message from the control means 30 of the writing device as well as a get-data call and a put room message from the control means 70 of the reading device. When the controller 4 receives a get room call, it resets the value of a corresponding room semaphore to zero and returns the previous value of the room semaphore. This value is then returned to the coprocessor 2 or its associated writing device which issued the get-room call. When the controller 4 receives the put data message from a writing device associated to a coprocessor 2, it increments the data semaphore by a specific value. This value may the predetermined value N or alternatively the argument of the put data message.

The implantation of the reaction of the controller 4 towards the get-data call and the put room message is similar to the implementations of the get-room call and the put data message. When the controller 4 receives a get-data call, it resets the value of a corresponding data semaphore to zero and returns the previous value of the data semaphore.

This value is then returned to the coprocessor 2 or its associated reading device which issued the get-data call. When the controller 4 receives the put-room message from a reading device associated to a coprocessor 2, it increments the room semaphore by a specific value.

This value may the predetermined value N or alternatively the argument of the put vroom message.

The above described communication scheme can be improved by issuing a get-data or get room call requesting data or room in the FIFO 1 before the first counter 10, i. e. data-available counter, or the third counter 50, i. e. the room-available counter, have become zero. This improvement will hide the latency of the getdata/getroom calls.

As controller 4 a low-end MIPS PR 1910 processor can be used.

As mentioned above processes may be implemented either on a programmable processor or on a dedicated coprocessor. The choice between those two alternatives is the classical hardware/software co-design question. Preferably, in the FIFO communication protocol buffer access, pointer increments and pointer wrap-around tests are implemented in hardware, i. e. dedicated coprocessors, while semaphore operations are implemented in software. It is desirable to perform semaphore operations in software, i. e. the processes are executed on the programmable processors, because of their complexity, e. g. waking up sleeping processes.