Repository navigation
cpu: evaluate arrays past 2^31 elements and sum with compensation - #3731
Draft
melonakos wants to merge 1 commit into
Draft
melonakos wants to merge 1 commit into
melonakos wants to merge 1 commit into
Conversation
Copy link
Copy Markdown
Member
Sorry, something went wrong.
The CPU JIT evaluation loop counted elements in int, so an array of more than 2^31 elements was left unevaluated and every reduction over it returned 0; the linear index is now dim_t through Node::calc. The all-elements add reduction also lost the low-order bits in one pass (1e9 float ones summed to 2^24), so it now carries a Kahan correction, as #3687 does for the dimension-wise sum. Both halves of #3571 on the CPU backend.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Two CPU-backend halves of #3571. The JIT evaluation loop counted elements in int, so an array of more than 2^31 elements was never evaluated and every reduction over it returned 0; the linear index is now dim_t through Node::calc and the loop. Separately, the all-elements add reduction accumulated in one pass, so 1e9 float ones summed to 16777216 (2^24) and norm came out as 4096; it now carries a Kahan correction, the same thing #3687 does for the dimension-wise sum, which still plateaus at 2^24 until that PR lands.
Verified here with 1e9 and 2.2e9-element arrays (sum, count, max, norm, f32/f64/u8 all exact after the change), and the reduce and jit suites pass on CPU and on OpenCL (Intel Arc B580). Adds a 1e8-element exact-sum test; the 2^31 case is not in the suite because it needs 9 GB. The OpenCL and CUDA JIT kernels index with int too and are not touched here.