Zero-Copy Memory Access
Prior to the implementation of restricted USM, you had to access host’s data from the device using one of the following methods:
- Through SYCL buffers
- By copying data between the host and the device using explicit USM
Both of these methods resulted in data transfers between the host and the device memory on discrete cards such as the Intel® FPGA Programmable Acceleration Card (PAC) D5005 (previously known as
Intel® FPGA Programmable Acceleration Card (PAC) with Intel® Stratix® 10 SX FPGA
).
With host allocations on devices supporting USM host allocation, a kernel can directly access data over PCIe (no copying required). By using host allocations in designs that have infrequent random accesses to large pieces of data, you can improve throughput and latency of the design as a large piece of data no longer requires copying in full to the device. For detailed explanation of this concept, refer to the
Zero-copy Data Transfer tutorial on GitHub.
- SYCL buffers created with host allocations set ashostDatain the constructor still result in data copy from the host to the device memory.
- Shared allocation for thepac_s10_usmboard that is in theintel_s10sx_pacBSP does not yield any change in the behavior or performance over host allocation. Both its host and shared allocations reside in the host.