Scatter-gather intelligent memory architecture for unstructured streaming data on multiprocessor systems
A scatter/gather technique optimizes unstructured streaming memory accesses, providing off-chip bandwidth efficiency by accessing only useful data at a fine granularity, and off-loading memory access overhead by supporting address calculation, data shuffling, and format conversion.
1. A processor comprising:
a core including:
a computation processor; and
a scatter/gather engine coupled to the computation processor, the scatter/gather engine to generate sub-cache line sized non-sequential data accesses to a memory based on an access pattern communicated to the scatter/gather engine from and defined by an application, to communicate sub-cache line sized data with the memory, wherein the scatter/gather engine includes an access processor, an access pattern generator, and a cache interface, wherein the access pattern generator is to generate the sub-cache line sized non-sequential data accesses to the memory based on an indirect access pattern; and
a cache coupled to the core, wherein data is to be transferred between the cache and the memory using full-cache line sized transfers.
2. The processor of claim 1 , wherein the access processor is to compute memory addresses for the sub-cache line sized non-sequential data accesses and perform data format conversion.
3. The processor of claim 1 , wherein the access processor is to generate the sub-cache line sized non-sequential data accesses to the memory based on the indirect access pattern.
4. The processor of claim 1 , wherein the access pattern generator is to generate the sub-cache line sized non-sequential data accesses to the memory based on a stride-based access pattern.
5. The processor of claim 1 , wherein the scatter/gather engine further comprises:
a stream port coupled to the computation processor, the stream port including a buffer capable of storing ordered data accessible by the access processor and the computation processor.
6. The processor of claim 1 , wherein the cache interface in conjunction with the cache is to provide data coherence when a same data is accessed through both the cache and the scatter/gather engine.
7. The processor of claim 1 , further comprising:
a memory controller coupled to the scatter/gather engine and the memory, the memory controller to support both cache line and sub-cache line sized data accesses to the memory.
8. A machine-readable non-transitory medium having stored thereon instructions, which if performed by a machine cause the machine to perform a method comprising:
transferring full-cache line sized data between a cache of a multicore processor and an off-chip memory;
generating, by a scatter/gather engine of the multicore processor, sub-cache line sized non-sequential data accesses to the off-chip memory based on an indirect access pattern defined and communicated by an application to transfer sub-cache line sized data between the multicore processor and the off-chip memory, the sub-cache line sized data having fewer bits than the full-cache line sized data transfers; and
computing memory addresses for the sub-cache line sized non-sequential data accesses, performing data format conversion, and generating the sub-cache line sized non-sequential data access addresses for the off-chip memory based on the indirect access pattern.
9. The machine-readable non-transitory medium of claim 8 , wherein the method further comprises:
allocating a stream port in the scatter/gather engine to handle the sub-cache line sized data; and
directing access to the off-chip memory through the allocated stream port.
10. The machine-readable non-transitory medium of claim 8 , wherein the method further comprises:
allocating a stream port in the scatter/gather engine to a thread in the multicore processor; and
in response to a thread context switch, releasing the stream port after write data stored in the stream port has been written to the off-chip memory.
11. The machine-readable non-transitory medium of claim 8 , wherein the method further comprises enforcing data coherency when a same data is accessed through both the cache and the scatter/gather engine.
12. The machine-readable non-transitory medium of claim 8 , wherein the method further comprises enforcing data coherency via a mutual exclusion of the data in a buffer in the scatter/gather engine or in the cache.
13. The machine-readable non-transitory medium of claim 12 , wherein the method further comprises enforcing the data coherency via an address range check in a directory.