Improved architecture with shared memory

Description of equivalent WO 03085524 A2

IMPROVED ARCHITECTURE WITH SHARED MEMORY

0002Field of the Invention

0003The present invention relates generally to integrated circuits (ICs). More particularly, the invention relates to an improved architecture with shared memory.

0004Background of the Invention

0005Fig. 1 shows a block diagram of a portion of a conventional System-on-Chip (SOC) 100, such as a digital signal processor (DSP) . As shown, the SOC includes a processor 110 coupled to a memory module 160 via a bus 180. The memory module stores a computer program comprising a sequence of instructions. During operation of the SOC, the processor retrieves and executes the computer instructions from memory to perform the desired function.

0006An SOC may be provided with multiple processors that execute, for example, the same program. Depending on the application, the processors can execute different programs or share the same program. Generally, each processor is associated with its own memory module to improve performance because a memory module can only be accessed by one processor during each clock cycle. Thus, with its own memory, a processor need not wait for memory to be free since it is the only processor that will be accessing its associated memory module. However, the improved performance is achieved at the sacrifice of chip size since duplicate memory modules are required for each processor. As evidenced from the above discussion, it is desirable to provide systems in which the processors can share a memory module to reduce chip size without incurring the performance penalty of conventional designs.

0007Summary of the Invention The invention relates, in one embodiment, to a method of sharing a memory module between a plurality of processors. The memory module is divided into n banks, where n = at least 2. Each bank can be accessed by one or more processors at any one time. The memory module is mapped to allocate sequential addresses to alternate banks of the memory, where sequential data are stored in alternate banks due to the mapping of the memory. In one embodiment, the memory banks are divided into x blocks, where x = at least 1, wherein each block can be accessed by one of the plurality of processors at any one time. In another embodiment, the method further includes synchronizing the processors to access different blocks at any one time. In another embodiment, first and second signal paths are provided between the memory module and a processor. The first signal path couples a cache to a processor and memory module for enabling the processor to fetch a plurality of data words from different banks simultaneously. This reduces memory latency caused by memory contention. The second signal path couples the memory module directly to the processor.

0008Brief Description of the Drawings

0009Fig. 1 shows a block diagram of conventional SOC; Fig. 2 shows a system in accordance with one embodiment of the invention;

0010Figs. 3-5 show a flow of FCU in accordance with different embodiments of the invention;

0011Fig. 6 shows a system in accordance with another embodiment of the invention; Figs. 7-8 show flow diagrams of an arbitration unit in accordance with various embodiments of the invention; and

0012Figs. 9-10 show memory modules in accordance with various embodiments of the invention.

0013Preferred Embodiments of the Invention

0014Fig. 2 shows a block diagram of a portion of a system 200 in accordance with one embodiment of the invention. The system comprises, for example, multiple digital signal processors (DSPs) for multi-port digital subscriber line (DSL) applications on a single chip. The system comprises m processors 210, where m is a whole number equal to or greater than 2. Illustratively, the system comprises first and second processors 210a-b (m = 2) . Providing more than two processors in the system is also useful. The processors are coupled to a memory module 260 via respective memory buses 218a and 218b. The memory bus, for example, is 16 bits wide. Other size buses can also be used, depending on the width of each data byte. Data bytes accessed by the processors are stored in the memory module. In one embodiment, the data bytes comprise program instructions, whereby the processors fetch instructions from the memory module for execution.

0015In accordance with one embodiment of the invention, the memory module is shared between the processors without noticeable performance degradation, eliminating the need to provide duplicate memory modules for each processor. Noticeable performance degradation is avoided by separating the memory module into n number of independently operable banks 265, where n is an integer greater than or equal to 2. Preferably, n = the number of processors in the system (i.e. n = m) . Since the memory banks operate independently, processors can simultaneously access different banks of the memory module during the same clock cycle . In another embodiment, a memory bank is subdivided into x number of independently accessible blocks 275a-p, where x is an integer greater than or equal to 1. In one embodiment, each bank is subdivided into 8 independently accessible blocks. Generally, the greater the number of blocks, the lower the probability of contention. The number of blocks, in one embodiment, is selected to optimize performance and reduce contention.

0016In one embodiment, each processor (210a or 210b) has a bus (218a or 218b) coupled to each bank. The blocks of the memory array, each have, for example control circuitry 278 to appropriately place data on the bus to the processors. The control circuitry comprises, for example, multiplexing circuitry or tri-state buffers to direct the data to the right processor. Each bank, for example, is subdivided into 8 blocks. By providing independent blocks within a bank, processors can advantageously access different blocks, irrespective of whether they are from the same bank or not. This further increases system performance by reducing potential conflicts between processors.

0017Furthermore, the memory is mapped so that contiguous memory addresses are rotated between the different memory banks. For example, in a two-bank memory module (e.g., bank 0 and bank 1) , one bank (bank 0) would be assigned the even addresses while odd addresses are assigned to the other bank (bank 1) . This would result in data bytes in sequential addresses being stored in alternate memory banks, such as data byte 1 in bank 0, data byte 2 in bank 1, data byte 3 in bank 0 and so forth. The data bytes, in one embodiment, comprise instructions in a program. Since program instructions are executed in sequence with the exception of jumps (e.g., branch and loop instructions), a processor would generally access different banks of the memory module after each cycle during program execution. By synchronizing or staggering the processors to execute the program so that the processors access different memory banks in the same cycle, multiple processors can execute the same program stored in memory module 260 simultaneously. A flow control unit (FCU) 245 synchronizes the processors to access different memory blocks to prevent memory conflicts or contentions. In the event of a memory conflict (e.g. two processors accessing the same block simultaneously), the FCU locks one of the processors (e.g. inserts a wait state or cycle) while allowing the other processor to access the memory. This should synchronize the processors to access different memory banks in the next clock cycle. Once synchronized, both processors can access the memory module during the same clock cycle until a memory conflict caused by, for example, a jump instruction, occurs. If both processors (210a and 210b) tries to access block 275a in the same cycle, a wait state is inserted in, for example, processor 210b for one cycle, such that processor 210a first accesses block 275a. In the next clock cycle, processor 210a accesses block 275b and processor 210b accesses block 275a. The processors 210a and 210b are hence synchronized to access different memory banks in the subsequent clock cycles.

0018Optionally, the processors can be provided with respective critical memory modules 215. The critical memory module, for example, is smaller than the main memory module 260 and is used for storing programs or subroutines which are accessed frequently by the processors (e.g., MIPS critical) . The use of critical memory modules enhances system performance by reducing memory conflicts without going to the extent of significantly increasing chip size. A control circuit 214 is provided. The control circuit is coupled to bus 217 and 218 to appropriately multiplex data from memory module 260 or critical memory module 215. In one embodiment, the control circuit comprises tri-state buffers to decouple and couple the appropriate bus to the processor .

0019In one embodiment, the FCU is implemented as a state machine. Fig. 3 shows a general process flow of a FCU state machine in accordance with one embodiment of the invention. As shown, the FCU controls accesses by the processors (e.g., A or B) . At step 310, the FCU is initialized. During operation, the processors issue respective memory addresses (A_Add or B_Ad) corresponding to the memory access in the next clock cycle. The FCU compares A_Add and B_Add at step 320 to determine whether there is a memory conflict or not (e.g., whether the processors are accessing the same or different memory blocks) . In one embodiment, the FCU checks the addresses to determine if any critical memory modules are accessed (not shown) . If either processor A or processor B is accessing its respective local critical memory, no conflict occurs. If no conflict exists, the processors access the memory module at step 340 in the same cycle. If a conflict exists, the FCU determines the priority of access by the processors at step 350. If processor A has a higher priority, the FCU allows processor A to access the memory while processor B executes a wait state at step 360. If processor B has a higher priority, processor B accesses the memory while processor A executes a wait state at step 370. After step 340, 360, or 370, the FCU returns to step 320 to compare the addresses for the next memory access by the processors. For example, if a conflict exists, such as at step 360, a wait state is inserted for processor B while processor A accesses the memory at address A_Add. Hence, both processors are synchronized to access different memory blocks in subsequent cycles.

0020Fig. 4 shows a process flow 401 of an FCU in accordance with another embodiment of the invention. In the case of a conflict, the FCU assigns access priority at step 460 by examining processor A to determine whether it has executed a jump or not. In one embodiment, if processor B has executed a jump, then processor B is locked (e.g. a wait state is executed) while processor A is granted access priority. Otherwise, processor A is locked and processor B is granted access priority.

0021In one embodiment, the FCU compares the addresses of processor A and processor B in step 440 to determine if the processors are accessing the same memory block. In the event that the processors are accessing different memory blocks (i.e., no conflict), the FCU allows both processors to access the memory simultaneously at step 430. If a conflict exists, the FCU compares, for example, the least significant bits of the current and previous addresses of processor A to determine access priority in step 460. If the least significant bits are not equal (i.e. the current and previous addresses are consecutive) , processor B may have caused the conflict by executing a jump. As such, the FCU proceeds to step 470, locking processor B while allowing processor A to access the memory. If the least significant bits are equal, processor A is locked and processor B accesses the memory at step 480.

0022Fig. 5 shows an FCU 501 in accordance to an alternative embodiment of the invention. Prior to operation, the FCU is initialized at step 510. At step 520, the FCU compares the addresses of processors to determine it they access different memory blocks. If the processors are accessing different memory blocks, both processors are allowed access at step 530. However, if the processors are accessing the same memory block, a conflict exists. During a conflict, the FCU determines which of the processors caused the conflict, e.g., performed a jump. In one embodiment, at steps 550 and 555, the least significant bits of the current and previous addresses of the processors are compared. If processor A caused the jump (e.g., least significant bits of previous and current address of processor A are equal while least significant bits of previous and current address of processor B are not), the FCU proceeds to step 570. At step 570, the FCU locks processor A and allows processor B to access the memory at step 570. If processor B caused the jump, the FCU locks processor B while allowing processor A to access the memory at step 560.

0023A situation may occur where both processors performed a jump. In such a case, the FCU proceeds to step 580 and examines a priority register which contains the information indicating which processor has priority. In one embodiment, the priority register is toggled to alternate the priority between the processors. As shown in Fig. 5, the FCU toggles the priority register at step 580 prior to determining which processor has priority. Alternatively, the priority register can be toggled after priority has been determined. In one embodiment, a 1 in the priority register indicates that processor A has priority (step 585) while a 0 indicates that processor B has priority (step

0024590) . Using a 1 to indicate that B has priority and a 0 to indicate that A has priority is also useful. The same process can also be performed in the event a conflict occurred in which neither processor performed a jump (e.g., least significant bits of the current and previous addresses of processor A or of processor B are not the same) .

0025In alternative embodiments, other types of arbitration schemes can be also be employed by the FCU to synchronize the processors. In one embodiment, the processors may be assigned a specific priority level vis-a-vis the other processor or processors.

0026Fig. 6 shows a block diagram of a portion of a system 600 in accordance with one embodiment of the invention. The system comprises, for example, multiple digital signal processors (DSPs) for multi-port digital subscriber line (DSL) applications on a single chip. The system comprises m processors 610, where m is a whole number equal to or greater than 2. Illustratively, the system comprises first and second processors 610a-b (m = 2) . Providing more than two processors is also useful.

0027A memory module 660 is provided for sharing among the processors. Data words accessed by the processors are stored in the memory module. A data word comprises a group of bits (e.g. 32 bits). In one embodiment, the data words comprise program instructions, which are accessed by the processors from the memory module via memory buses (e.g. 618a and 618b) for execution. The data words can also comprise application data. In accordance with one embodiment of the invention, the memory module is shared between the processors without noticeable performance degradation, eliminating the need to provide duplicate memory modules for each processor. Noticeable performance degradation is avoided by separating the memory module into n number of independently operable banks (e.g. 665a and 665b), where n is a number greater than or equal to 2. Preferably, n = the number of processors in the system (i.e. n = m) . Since the memory banks operate independently, the different banks can be simultaneously accessed during the same clock cycle. In another embodiment, the banks can be further subdivided into x number of independently accessible blocks 675a-p, where x is an integer greater than or equal to 1. A bank, for example, is subdivided into 8 independently accessible blocks. Generally, the greater the number of blocks, the lower the probability of contention. The number of blocks, in one embodiment, is selected to optimize performance and reduce contention.

0028The blocks of the memory array have, for example, control circuitry 668 to appropriately place data on the memory buses (e.g. 618a or 618b) to the processors (610a or 610b) . The control circuitry comprises, for example, multiplexing circuitry or tri-state buffers to direct the data to the respective processors. By providing independent blocks within a bank, the processors can advantageously access different blocks simultaneously, irrespective of whether they are from the same bank or not. This further increases system performance by reducing potential conflicts between processors. Furthermore, the memory is mapped so that contiguous memory addresses are rotated between the different memory banks. For example, in a two-bank memory module (e.g., bank 0 and bank 1) , one bank (bank 0) would be assigned the even addresses while odd addresses are assigned to the other bank (bank 1) . This would result in data words in sequential addresses being located in alternate memory banks, such as data word 1 in bank 0, data word 2 in bank 1, data word 3 in bank 0 and so forth. In one embodiment, the data words comprise program instructions. Since program instructions are executed in sequence with the exception of jumps (e.g., branch and loop instructions), a processor would generally access different banks of the memory module during program execution. By synchronizing or staggering the processors to execute the program so that the processors access different memory banks in the same cycle, multiple processors can execute the same program stored in memory module 660 simultaneously.

0029An arbitration control unit (ACU) 645 being coupled to the processor via the data bus and to the memory module via the memory bus is provided. The ACU controls access to the memory by the processors. In the event of a memory contention (e.g., two processors accessing the same bank simultaneously) , the ACU determines which processor has priority to access the memory module while the other processors are locked (e.g. by executing a wait state or cycle) . This generally synchronizes the processors to access different banks in the subsequent clock cycles.

0030In one embodiment, a priority register is provided to indicate which processor has priority. In the case of a system with two processors, the priority register may comprise one bit (P bit) . Additional bits may be included to accommodate additional number of processors. The priority register is updated after the occurrence of contention to rotate the priority between the processors. For example, a value of l' in the P bit indicates that the first processor has priority and a '0' indicates that the second processor has priority. During each cycle where a contention occurs, the P bit is toggled, switching the priority of the processors. Other types of arbitration schemes are also useful.

0031Optionally, the processors can be provided with respective critical memory modules 615. The critical memory module, for example, is smaller than the main memory module 660 and is used for storing programs or subroutines which are accessed frequently by the processors (e.g., MIPS critical) . The use of critical memory modules enhances system performance by reducing memory conflicts without going to the extent of significantly increasing chip size. The ACU 645 is coupled to n control logic units (CLUs) , one for each of the n processors. Illustratively, the ACU comprises first CLU 648a and second CLU 648b for first processor 610a and second processor 610b respectively. When a CLU is activated, its respective processor is allowed access to the memory module. In one embodiment, the CLU is coupled to a processor and to the n banks of memory module, enabling the processor to access the n memory banks simultaneously. Since the bandwidth of a processor is equal to the bandwidth of a memory bank, the CLU allows the processor to fetch from memory more words than needed. In one embodiment, the processor can potentially fetch twice the data words needed.

0032In one embodiment, the CLU comprises first (cache) and second (normal) signal paths. The cache signal path comprises, for example, a cache register (633a or 633b) and a multiplexer (636a or 636b). When the cache path is selected, the processor coupled to the CLU accesses the first and second memory banks (665a-b) . In one embodiment, the current address location (Addr) , as specified by the processor, and the next address (Addr + 1) are accessed. The multiplexer selects the word at (Addr + 1) and stores it in the cache while the word at the current address (Addr) is passed to the processor. The address of the word stored in the cache is stored in, for example, a cache address register (640a or 640b) . If the second path (normal) is selected, the processor accesses the current memory location. The CLU passes the data word at the current memory location to the processor via the second path. By providing a cache to store data in subsequent addresses, the probability of data access from memory is lowered, hence reducing memory latency caused by memory contention .

0033The processors can be provided with respective critical memory modules 615a and 615b. The critical memory module, for example, is smaller than the main memory module 660 and is used for storing data (e.g. programs or subroutines) which are accessed frequently by the processors (e.g., MIPS critical). The use of critical memory modules enhances system performance by reducing memory conflicts without going to the extent of significantly increasing chip size.

0034Fig. 7 shows a process flow of an ACU state machine in accordance with one embodiment of the invention. As shown, the ACU controls accesses by first and second processors (A or B) . The ACU system is initialized (710), for example, before system operation (e.g., system power up). Initialization includes, for example, setting the priority bit to indicate which processor has priority in the event of a memory contention. The priority register, for example, is set to give processor A priority.

0035During operation of the system, the processors issue respective memory addresses corresponding to the memory access in the next clock cycle (A_Addr and B_Addr representing the memory addresses currently issued by processor A and processor B) . The ACU determines whether there is a memory contention or not at steps 720 and 722, e.g., whether the processors are accessing the same memory range or not. The memory range coincides, in one embodiment, with a memory block. In another embodiment, the memory range coincides with memory blocks in different banks, the memory blocks comprising consecutive addresses. If no contention exists, processors A and B access respective banks of the memory module at step 750. In one embodiment, the CLUs of processors A and B are activated with the normal signal paths selected. Thus, each processor retrieves data words from respective memory banks at addresses A_{A r} and B_Addr.

0036If a contention occurs, the ACU evaluates the priority register to determine which processor has access priority at step 726. The processor P with access priority (e.g., processor A) is allowed access to the memory while the other processor P' with lower priority executes a wait state (e.g., processor B) at step 728. Hence, if the processors subsequently access data words in sequential locations in the next cycles, different banks will be accessed without executing wait-states. By synchronizing or staggering the processors to execute the program so that the processors access different memory banks in the same cycle, multiple processors can execute the same program stored in memory module 660 simultaneously without contention. In one embodiment, the CLU of processor P is activated with the cache signal path selected, at step 730. The data from the current address P_Adr and the next consecutive address P_Addr₊ι are fetched from the memory banks. The data in the current address P_{A d}r is passed to the processor P for access and data in the next address P_Addr+ι is stored in the cache register. The ACU updates the priority at step 332 for the next contention evaluation at step 722.

0037The ACU determines at step 734 if a new address P_Addr specified by the processor P in the next cycle matches the address of the cache data (i.e. cache hit). If a cache miss occurs, the process is repeated by evaluating the addresses specified by processors A and B for contention at step 720. In one embodiment, the data in the cache register associated with processor P is discarded.

0038A cache hit would allow processor P to continue execution by retrieving the data from the cache instead of memory, thus avoiding the insertion of a wait-state at step 736. In one embodiment, the CLU of processor P' is activated with the cache signal path selected at step 734. The data from the current address P'_Addr and the next address P'_Addr₊ι are fetched from the memory banks. The data in the current address P' d_dr is passed to the processor P' for access and the data in the next address P'_Addr+ι is stored in the cache register associated with P' . If there is a cache hit for processor P' in the next cycle, the cache data is accessed by the processor P' at step 740. The data in the current address P_Addr of processor P accessed by the processor and the data in the next address P_{Ad r+}i is stored in the cache register associated with P. There is no need to check for contention as only one processor is accessing the memory. The determination of a cache hit for processor P is repeated at step 734. If a cache miss for P' occurs at step 738, the ACU repeats the whole process at step.

0039In another embodiment shown in Fig. 8, the ACU comprises the cache signal path for each processor, the cache signal path allowing more data words to be fetched from memory than requested by the processor. The cache signal path comprises, for example, a cache register and a multiplexer. During the operation of the system, the addresses issued by processors A and B are evaluated for contention at steps 820 and 822. If contention exists, the priority is evaluated and wait states are inserted for the processor with lower priority as previously described in Fig. 7.

0040If no contention exists, the caches associated with both processors are evaluated for cache hit at step 852. If no cache hits are found, processors A and B access respective banks of the memory module at step 850 via the respective cache signal paths. In one embodiment, the CLUs of processors A and B are activated with the cache paths selected. The data in the current memory addresses (A_{Ad r} and B_Addr) s passed to the respective processors for access and data in the next consecutive addresses (A_Addr₊ι and B_Addr+ι) is stored in the respective cache registers. If cache hits are detected for both processors, the respective cache contents are accessed by the processors at step 862 and the process repeats at step 820. If a cache hit is found for only one of the processors, memory access may continue for the other processor without the need to test for contention since only one processor is accessing the memory. For example, if a cache hit is detected for processor A and a cache miss is detected for processor B, the contents of the cache associated with processor A is accessed while the data from the current memory address B_Addr is accessed by processor B at step 854. Data from the memory at the next location B_Addr₊i is stored in the cache associated with processor B. In the next cycle, the cache for processor B is monitored again for a cache hit. If a cache hit occurs, the cache contents for processor B is retrieved at step 856. The data from memory at address A_{A d}r will be fetched for processor A. A cache miss at step 858 will cause the process to be repeated from step 820. Figs. 9-10 illustrate the mapping of memory in accordance with different embodiments of the invention. Referring to Fig. 9, a memory module 260 with 2 banks (Bank 0 and Bank 1) each subdivided into 8 blocks (Blocks 0-7) is shown. Illustratively, assuming that the memory module comprises 512Kb of memory with a width of 16 bits, each block being allocated 2K addressable locations (2K x 16 bits x 16 blocks). In one embodiment, even addresses are allocated to bank 0 (i.e., 0, 2, 4 ...32K-2) and odd addresses to bank 1 (i.e., 1, 3, 5...32K-1). Block 0 of bank 0 would have addresses 0, 2, 4 ... 4K-2; block 1 of bank 1 would have addresses 1, 3, 5...4K-1.

0041Referring to Fig. 10, a memory module with 4 banks (Banks 0-3) each subdivided into 8 blocks (Blocks 0-7) is shown. Assuming that the memory module 512Kb of memory with a width of 16 bits, than each block is allocated IK addressable locations (IK x 16bits x 32 blocks) . In the case where the memory module comprises 4 banks, as shown in Fig. 10, the addresses would be allocated as follows: Bank 0: every fourth address from 0 (i.e., 0, 4, 8, etc. ) Bank 1: every fourth address from 1 (i.e., 1, 5, 9, etc. ) Bank 2: every fourth address from 2 (i.e., 2, 6, 10, etc.) Bank 3: every fourth address from 3 (i.e., 3, 7, 11, etc . ) The memory mapping can be generalized for n banks as follows : Bank 0: every n^h address beginning with 0 (i.e., 0, n, 2n, 3n, etc. ) Bank 1: every n^h address beginning with 1 (i.e., 1, 1+n, l+2n, l+3n, etc.)

0042Bank n-1: every n^h address beginning with n-1 (i.e., n-1, n-l+n, n-l+2n, etc.) While the invention has been particularly shown and described with reference to various embodiments, it will be recognized by those skilled in the art that modifications and changes may be made to the present invention without departing from the spirit and scope thereof. The scope of the invention should therefore be determined not with reference to the above description but with reference to the appended claims along with their full scope of equivalents .