This is primarily a bug-fix release. Some buggy areas that were cleaned up include:
- 32bit vs. 64bit architecture issues
- automatic memory sizing and size detection
- per timing interval initialization/cleanup
- lat_mem_rd strided and random memory access patterns
- lat_mem_rd memory size now grows exponentially instead of linearly
- lat_ops / par_ops are now more accurate on more architectures
- stream used the wrong kernel in one instance
- automatic configuration of random number generators
- potential infinite loop in timing harness
The most significant non-bugfix change is in lat_mem_rd. It now has two modes: sequential strided access, and random access. Each access mode is common in real applications, and each is important. The old lat_mem_rd access pattern (reverse strided access) was a hybrid of the two patterns that was predictable by some hardware and not others and did not really reflect any common memory access pattern. We feel the new lat_mem_rd is more accurate and better matches common system usage.
Log in to post a comment.