Skip to content

Latest commit

 

History

History
42 lines (23 loc) · 3.09 KB

memory.md

File metadata and controls

42 lines (23 loc) · 3.09 KB
Raw

Memory Management

In this section, we demonstrate how to manage the memory in host side with L3 level APIs provided by ThunderGP.

Currently, Xilinx Multiple-SLRs FPGAs have many independent DRAM banks, and many graph algorithms manipulate at least five types of data. Managing the data and the partitions among the banks using the traditional OpenCL APIs need a lot of redundant code for configuration, which makes the code hard to maintaining. Therefore we developed an unique ID based memory management module(he_mem), and basing on this module, a hierarchy of data structures is constructed:

drawing

All the compuations and operations are heirarchically based on these object, which makes it easy to understand the relationship among the objects. There are two classes of objects in ThunderGP, we call them kernel descriptor and partition descriptor.

Partition Descriptor

Partition descriptor maintains all the partitioned data among the available DRAM banks, As illustrated in the above figure:

  • partitionDescriptor: the purple blocks are partitionDescriptor . There are M partitionDescriptor during the processing, where the M is determined by the size of graph dataset dynamically. Following figure shows how does ThunderGP perform the partitioning on the graph, each partitions hold edges in a fixed range of destination.

drawing

  • subPartitionDescriptor: partitionDescriptor consist of N subPartitionDescriptor (the grey blocks), the N is determined by how many CUs are available in the FPGA side. The subPartitionDescriptor structure consists of three types of partitioned data of edge: the destination of edges, the source of edges and the property of edges. Each type of data is contained in he_mem elements (the yellow blocks). Following figure shows how does ThunderGP perform the sub-partitioning on the partitions, the edges of each partition is equally divided into N parts, here the N is 4.

drawing

Kernel Descriptor

There are two types of kernel descriptor: the gatherScatterDescriptor and applyDescriptor. Showing in the heirarchy figure, the blue blocks are the kernel descriptor, both of these two types of descriptors have an instance of OpenCL kernel (the green blocks). The property of vertices are also included in gatherScatterDescriptor.

Appendix

API list for he_mem

Function Description
int he_mem_init() initialize the he_mem structure
cl_mem* get_cl_mem_pointer() get the OpenCL cl_mem instance in the specific he_mem
void* get_host_mem_pointer() get the host memory pointer in the specific he_mem
he_mem_t* get_he_mem() get the instance of he_mem by the unique ID
int transfer_data_from_pl() transfer the data from FPGA to CPU
int transfer_data_to_pl() transfer the data from CPU to FPGA