1. A system for balancing crossbar bandwidth in a multiprocessing system, comprising:
a crossbar switch having a plurality of links to processors in the multiprocessing system;
a plurality of synchronizers provided to receive micropackets from the links;
at least one wait register to hold the micropackets; and
an arbitrator operatively associated with the at least one wait register, the arbitrator determining which micropackets are output from the crossbar if a threshold is satisfied.
2. The system of claim 1 further comprising a routing agent operatively associated with the at least one wait register, the routing agent routing micropackets around the wait register if the wait register for the link is empty and the link wins arbitration.
3. The system of claim 1 further comprising a routing agent operatively associated with the at least one wait register, the routing agent writing only valid micropackets to the wait register.
4. The system of claim 3 wherein both valid and invalid micropackets are routed past the wait register.
5. The system of claim 1 wherein the threshold is different for different flow control classes.
6. The system of claim 1 further comprising an analysis module to evaluate the threshold during operation of the crossbar switch and change the threshold based on performance of the crossbar switch.
7. The system of claim 1 wherein the threshold is satisfied if all valid micropackets for a data packet are received at the crossbar switch.
8. The system of claim 1 wherein the threshold is satisfied if the wait register is full.
9. The system of claim 1 wherein the at least one wait register is encoded for a maximum expected number of valid micropackets in a data packet received from the link.
10. The system of claim 1 wherein the at least one wait register is a firmware writable i_wait-state register.
11. A method of balancing crossbar bandwidth in a multiprocessing system comprising:
receiving a data packet as a plurality of micropackets at a crossbar switch;
writing valid micropackets to a wait register until a threshold is satisfied; and
arbitrating for output from the crossbar switch if the threshold is satisfied.
12. The method of claim 11 further comprising suppressing invalid micropackets received at the crossbar switch.
13. The method of claim 11 further comprising satisfying the threshold if all valid micropackets for the data packet are received at the crossbar switch.
14. The method of claim 11 further comprising satisfying the threshold if the wait register is full.
15. The method of claim 11 further comprising changing the threshold based on performance of the crossbar switch.
16. The method of claim 11 further comprising bypassing the wait register if the wait register is empty and the input queue has won arbitration.
17. The method of claim 16 wherein both valid and invalid micropackets bypass the wait register.
18. A system for balancing crossbar bandwidth in a multiprocessing system, comprising:
means for receiving micropackets at a crossbar switch;
means for suppressing invalid micropackets received at the crossbar switch; and
means for outputting micropackets from the crossbar switch when a threshold is satisfied.
19. The system of claim 18 wherein the threshold is satisfied after a predetermined wait for the micropackets.
20. The system of claim 18 further comprising means for routing micropackets around the means for suppressing invalid micropackets if the means for suppressing invalid micropackets is empty.
The claims below are in addition to those above.
All refrences to claim(s) which appear below refer to the numbering after this setence.
1. An integrated circuit device, comprising:
a processing component;
a cache, which is arranged to store data for use by the processing component responsively to an addressing scheme based on memory addresses having an address length of m1 bits; and
first and second buses coupled between the processing component and the cache, the buses having bus widths of n1 and n2 bits, respectively, such that n1<m1,
the processing component and the cache each comprising a respective address bus expander coupled to the first bus in order to compact at least some of the memory addresses for transmission over the first bus so that each of the at least some memory addresses is transmitted over the first bus in one cycle of the first bus.
2. The device according to claim 1, wherein the data comprise data words having a word length of m2 bits stored at each address, such that n2<m2, and
wherein each of the processing component and the cache further comprises a respective second bus expander coupled to the second bus in order to compact at least some of the data words for transmission over the second bus so that each of the at least some of the data words is transmitted over the second bus in one cycle of the second bus.
3. The device according to claim 2, wherein the data words comprise data values for processing by the device, and wherein the processing component is arranged to load the compacted data words via the second bus from the cache for processing and to store the compacted data words via the second bus to the cache.
4. The device according to claim 2, wherein the data words comprise instructions for execution by the device, wherein the compacted words comprise compacted instructions, and wherein the processing component is arranged to fetch the compacted instructions via the second bus.
5. The device according to claim 2, wherein the address bus expander and the second bus expander are arranged to compact the memory addresses and the data words simultaneously, so as to transmit a compacted memory address and a compacted data word for storage at the memory address together in one cycle of the first and second buses.
6. The device according to claim 2, wherein the address bus expander and the second bus expander are arranged to compact the memory addresses and the data words by transmitting indices to values in respective tables held by the bus expanders, and wherein the cache is arranged to store at least some of the indices together with the data.
7. The device according to claim 1, wherein the address bus expander is arranged to compact each of the at least some of the memory addresses by dividing each of the memory addresses into at least first and second fields, storing values of the second field in a respective table such that the values in respective tables held by the address bus expander in the processing component and the address bus expander in the cache are identical, and if the second field of a memory address matches a value in the table, transmitting an index corresponding to the value over the first bus along with the first field in the one cycle of the bus.
8. The device according to claim 7, wherein the first field comprises a set of least significant bits (LSB) of the memory address, while the second field comprises a set of most significant bits (MSB) of the memory address.
9. The device according to claim 7, wherein the at least first and second fields comprise a third field, and wherein the address bus expander is arranged to compact each of the at least some of the memory addresses by transmitting first and second indices corresponding to the values of the first and third fields, respectively, over the first bus along with the first field.
10. The device according to claim 7, wherein the address bus expander in the processing component is arranged, when the second field of the memory address does not match any of the values in the table, to transmit both of the first and second fields over multiple cycles of the bus, and to cause the respective table of the bus expander to be updated in both the processing component and the cache.
11. The device according to claim 7, wherein the cache comprises lines of the data that are indexed according to the first field, each line containing a corresponding value of the second field, and
wherein the address bus expander in the cache is arranged, upon receiving the index over the first bus, to retrieve the value of the second field from the table responsively to the index, and
wherein the cache is arranged to determine whether a cache hit has occurred by checking the retrieved value against the corresponding value of the second field in the line that is indexed by the first field.
12. The device according to claim 11, wherein the address bus expander is arranged to retrieve the value of the second field from the table simultaneously with retrieval of the data from the line in the cache that is indexed by the first field for transmission of the data over the second bus to the processing component.
13. The device according to claim 1, wherein the cache that is coupled to the processing component by the first and second buses is a Level 1 (L1 ) cache, and further comprising a Level 2 (L2 ) cache, and third and fourth buses coupling the L2 cache to the L1 cache, the L1 cache and the L2 cache comprising further bus expanders coupled to at least one of the third and fourth buses.
14. A method for coupling a processing component to a cache in an integrated circuit device, the method comprising:
configuring the cache to store data for use by the processing component responsively to an addressing scheme based on memory addresses having an address length of ml bits;
coupling first and second buses coupled between the processing component and the cache, the buses having bus widths of n1 and n2 bits, respectively, such that n1<m1;
compacting at least some of the memory addresses for transmission over the first bus so that each of the at least some memory addresses can be transmitted over the first bus in one cycle of the first bus;
transmitting at least the compacted memory addresses over the first bus; and
conveying the data over the second bus responsively to the compacted memory addresses.
15. The method according to claim 14, wherein the data comprise data words having a word length of m2 bits stored at each address, such that n2<m2, and wherein conveying the data comprises compacting at least some of the data words for transmission over the second bus so that each of the at least some of the data words is transmitted over the second bus in one cycle of the second bus.
16. The method according to claim 15, wherein the data words comprise data values for processing by the integrated circuit device, and wherein conveying the data comprises loading the compacted data words via the second bus from the cache for processing by the processing component, and storing the compacted data words from the processing component via the second bus to the cache.
17. The method according to claim 15, wherein the data words comprise instructions for execution by the device, wherein the compacted words comprise compacted instructions, and wherein conveying the data comprises fetching the compacted instructions from the cache via the second bus to the processing component.
18. The method according to claim 15, wherein compacting the at least some of the memory addresses and compacting at least some of the data words comprise compacting at least a portion of the memory addresses and the data words simultaneously, so as to transmit a compacted memory address and a compacted data word for storage at the memory address together in one cycle of the first and second buses.
19. The method according to claim 15, wherein compacting the at least some of the memory addresses and compacting at least some of the data words comprise compacting the memory addresses and the data words by transmitting indices to values in respective tables held in the cache and the processing component, and storing at least some of the indices in the cache together with the data.
20. The method according to claim 14, wherein compacting the at least some of the memory addresses comprises:
dividing each of the at least some of the memory addresses into at least first and second fields; and
storing values of the second field in a respective table such that the values in respective tables held in the processing component and in the cache are identical, and
wherein transmitting at least the compacted memory addresses comprises, if the second field of a memory address matches a value in the table, transmitting an index corresponding to the value over the first bus along with the first field in the one cycle of the bus.
21. The method according to claim 20, wherein the first field comprises a set of least significant bits (LSB) of the memory address, while the second field comprises a set of most significant bits (MSB) of the memory address.
22. The method according to claim 20, wherein the at least first and second fields comprise a third field, and wherein transmitting the index comprises transmitting first and second indices corresponding to the values of the first and third fields, respectively, over the first bus along with the first field.
23. The method according to claim 20, wherein transmitting at least the compacted memory addresses comprises, when the second field of the memory address does not match any of the values in the table, transmitting both of the first and second fields over multiple cycles of the bus, and updating the respective table in both the processing component and the cache.
24. The method according to claim 20, wherein configuring the cache comprises indexing lines of the data that according to the first field, each line containing a corresponding value of the second field, and wherein conveying the data comprises:
receiving the index in one of the compacted memory addresses over the first bus; and
determining whether a cache hit has occurred by retrieving the value of the second field from the table responsively to the index, and checking the retrieved value against the corresponding value of the second field in the line that is indexed by the first field.
25. The method according to claim 24, wherein determining-whether the cache hit has occurred comprises simultaneously retrieving the value of the second field from the table and retrieving the data from the line that is indexed by the first field for transmission over the second bus to the processing component.
26. The method according to claim 14, wherein the cache that is coupled to the processing component by the first and second buses is a Level 1 (L1 ) cache, and further comprising coupling a Level 2 (L2 ) cache via third and fourth buses to the L1 cache, and compacting at least some of the memory addresses or the data for transmission over at least one of the third and fourth buses.