Switching crossbar for graphics pipeline
A shared local memory data crossbar may be implemented in multiple stages. With this approach, the number of multiplexer cells can be reduced by fifty percent (50%) or more in some embodiments.
1. A graphics processor comprising:
a register file to store information;
a plurality of arithmetic logic units coupled to the register file;
a plurality of texture units coupled to the register file; and
a crossbar switch array coupled between a plurality of input ports and a plurality of output ports, wherein the crossbar switch array comprises:
a first crossbar level having n×m crossbar switches coupled to the plurality of input ports;
a second crossbar level having r×p crossbar switches, wherein each of the r×p crossbar switches has only one connection to each of the n×m crossbar switches of the first crossbar level; and
a third crossbar level having m×k crossbar switches, wherein each of the m×k crossbar switches has only one connection to each of the r×p crossbar switches of the second crossbar level, wherein k, m, n, p and r are positive integers.
2. The graphics processor of claim 1 , wherein r is the number of crossbar switches in the first crossbar level, m is the number of crossbar switches in the second crossbar level and p is the number of crossbar switches in the third crossbar level.
3. The graphics processor of claim 1 , wherein the crossbar switch array comprises a fully non-blocking crossbar switch array.
4. The graphics processor of claim 3 , wherein the fully non-blocking crossbar switch array includes only 144 multiplexer cells.
5. The graphics processor of claim 3 , wherein the fully non-blocking crossbar switch array includes only three levels.
6. The graphics processor of claim 1 , wherein each of the n×m crossbar switches, the r×p crossbar switches and the m×k crossbar switches is implemented as a fully connected crossbar switch.
7. The graphics processor of claim 1 , further comprising a multiplexer selection circuit to route any one of the plurality of input ports to any one of the plurality of output ports.
8. The graphics processor of claim 1 , wherein each of the plurality of input ports is assigned to one of the r×p crossbar switches of the second crossbar level.
9. The graphics processor of claim 1 , further comprising a shared local memory coupled to the plurality of input ports.
10. The graphics processor of claim 9 , wherein the shared local memory includes slots and b banks, wherein s is equal to r×n and b is equal to p×k and m is equal to n+k−1.
11. The graphics processor of claim 1 , wherein n, m, r and k are equal to one another.
12. The graphics processor of claim 1 , further comprising at least one single instruction multiple data processor coupled to the plurality of output ports.
13. A processor comprising:
a shared local memory having a plurality of memory banks;
a crossbar switch coupled between the shared local memory and a plurality of single instruction multiple data processors, the crossbar switch comprising:
a first crossbar level having n×m crossbar switches coupled to the shared local memory;
a second crossbar level having r×p crossbar switches, wherein each of the r×p crossbar switches has only one connection to each of the n×m crossbar switches of the first crossbar level; and
a third crossbar level having m×k crossbar switches, wherein each of the m×k crossbar switches has only one connection to each of the r×p crossbar switches of the second crossbar level; and
the plurality of single instruction multiple data processors coupled to the crossbar switch, wherein k, m, n, p and r are positive integers.
14. The processor of claim 13 , wherein r is the number of crossbar switches in the first crossbar level, m is the number of crossbar switches in the second crossbar level and p is the number of crossbar switches in the third crossbar level.
15. The processor of claim 13 , wherein each of the n×m crossbar switches, the r×p crossbar switches and the m×k crossbar switches is implemented as a fully connected crossbar switch.
16. The processor of claim 13 , wherein n, m, r and k are equal to one another.
17. An apparatus comprising:
a plurality of shader cores; and
a cache having a plurality of ports, wherein the plurality of shader cores is coupled to the cache via a crossbar switch comprising: a first crossbar level having n×m crossbar switches coupled to the plurality of shader cores;
a second crossbar level having r×p crossbar switches, wherein each of the r×p crossbar switches has only one connection to each of the n×m crossbar switches of the first crossbar level; and
a third crossbar level having m×k crossbar switches coupled to the cache, wherein each of the m×k crossbar switches has only one connection to each of the r×p crossbar switches of the second crossbar level, wherein k, m, n, p and r are positive integers.
18. The apparatus of claim 17 , wherein r is the number of crossbar switches in the first crossbar level, m is the number of crossbar switches in the second crossbar level and p is the number of crossbar switches in the third crossbar level.
19. The apparatus of claim 17 , wherein each of the n×m crossbar switches, the r×p crossbar switches and the m×k crossbar switches is implemented as a fully connected crossbar switch.
20. The apparatus of claim 17 , wherein n, m, r and k are equal to one another.