mlemanczyk/even-perfect-numbers-scanner
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Make by-divisor GPU kernels reuse read-only operands

Browse Source
This commit is contained in:
Marek Lemańczyk committed 2025-10-08 18:47:16 +02:00
1 parent cd1debe9d6
commit 6918bd28ff
8 files changed
+594 -250
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

todo.txt
+3
View File
@@ -1,2 +1,5 @@
- implement full routing for CPU or GPU execution based on the (...)-device=cpu|gpu switches. The current --mersenne=bydivisor GPU flow appears to run every stage on the GPU and ignore switches such as --primes-device for the primality check.
- investigate why the --mersenne=bydivisor GPU path runs out of host memory. Exceptions must remain visible.
- Reimplement the GPU cycle heuristics for MersenneNumberIncrementalGpuTester and re-enable Scan_handles_various_prime_exponents once the device path is fixed.
- Extend the GPU numeric type probe to cover proposed custom types such as GpuUInt256 and GpuUInt512 once their definitions are available so we can plan the wide-order heuristics on the device.
- document the GPU by-divisor regressions introduced in commit cd1debe ("Various micro- and medium- optimizations"): the refactor dropped divisor-cycle caching, remainder stepper reuse, and honoring the cached cycle parameter so we can plan which pieces to restore.