 mlemanczyk
|
c3254e3fda
|
Temporary restored delegating all work to CPU in CPU mode in preparation for a branch-less split.
|
2025-11-30 15:40:03 +01:00 |
|
|
mlemanczyk
|
6a1198c417
|
CPU vs GPU work delegation adjusted to better utilize CPU.
|
2025-11-29 12:09:33 +01:00 |
|
|
mlemanczyk
|
c311b27dcc
|
Prime testing using divisor tables depending on the last digit separated to its own class. PrimeTester switched back to standard endings.
|
2025-11-27 16:45:02 +01:00 |
|
|
mlemanczyk
|
a296e62d4d
|
--min-k command-line argument added. Partial factoring switched back to CPU.
|
2025-11-26 22:03:52 +01:00 |
|
|
mlemanczyk
|
13498dc460
|
Most device buffers integrated into PrimeOrderCalculatorAccelerator. PrimeOrderCalculator switched to partial factoring on GPU.
|
2025-11-26 03:16:39 +01:00 |
|
|
mlemanczyk
|
b9ca95bf8c
|
PrimeOrderCalculatorAccelerator.Rent is now created outside of the loops and re-used for all calculations.
|
2025-11-24 22:51:53 +01:00 |
|
|
mlemanczyk
|
070e2a14fd
|
MontgomeryDivisorData converted to a pooled class to limit GC pressure. MontgomeryDivisorDataGpu added for GPU operations. PrimeOrderCalculator kernels precompiled during warmup.
|
2025-11-24 15:14:31 +01:00 |
|
|
mlemanczyk
|
6d1ff1372e
|
Factoring switched to PrimeTester.IsPrime for the best performance. PrimeOrderCalculator.PartialFactor switched to GPU factorization. Every so often it will send work to GPU instead of CPU. BatchedResources renamed to MersenneNumberDivisorByDivisorAccelerator for consistency and moved to its own file. It now uses array instead of dictionary for lookup free performance.
|
2025-11-23 00:07:03 +01:00 |
|
|
mlemanczyk
|
54fdff0a4e
|
GPU calculations frequency adjusted
|
2025-11-19 21:38:07 +01:00 |
|
|
mlemanczyk
|
5cac51479e
|
GPU memory allocations protected with lock to prevent shared memory access issues. GPU usage limited for the best performance.
|
2025-11-19 16:49:01 +01:00 |
|
|
mlemanczyk
|
02d16bf05c
|
Tiny cleanup
|
2025-11-18 06:02:43 +01:00 |
|
|
mlemanczyk
|
801b1be3e3
|
Cleanup
|
2025-11-18 04:22:39 +01:00 |
|
|
mlemanczyk
|
342eff2726
|
GpuKernelPool.GetOrAddKernels removed in --mersenne=bydivisor mode. PartialFactorResult.WithAdditionalPrime will now modify the current instance instead of creating new copy for better performance. SmallPrimeFactorViews renamed to SmallPrimeFactorGpuTables for consistency with other static tables.
|
2025-11-17 23:33:59 +01:00 |
|
|
mlemanczyk
|
f630bfb598
|
HeuristicCombinedPrimeTesterAccelerator integrated with PrimeOrderCalculatorAccelerator.
|
2025-11-17 17:38:23 +01:00 |
|
|
mlemanczyk
|
183b147135
|
Heuristic tables split into 2 groups
|
2025-11-17 16:04:27 +01:00 |
|
|
mlemanczyk
|
0540411cad
|
Accelerator index added for dictionary-less use-cases
|
2025-11-15 02:21:45 +01:00 |
|
|
mlemanczyk
|
42447c4e43
|
BinaryGcd switched to the best performing variant according to benchmarks.
|
2025-11-14 11:43:16 +01:00 |
|
|
mlemanczyk
|
0d2a88c45f
|
GpuPrimeWorkLimiter.Acquire() / .Release() removed to save time on the hot-path with billions of calls. We'll rely on AcceleratorStreamPool limits and synchronization.
|
2025-11-13 17:52:13 +01:00 |
|
|
mlemanczyk
|
c53aa517f7
|
GpuScratchBuffer / -Pool naively integrated with PriemOrderCalculatorAccelerator (formerly Pow2MontgomeryAccelerator). FactorEntry replaced with 2 arrays - unused field removed. Small cleanup.
|
2025-11-13 14:41:27 +01:00 |
|
|
mlemanczyk
|
43863d03fb
|
Performance improvements
|
2025-11-13 00:51:32 +01:00 |
|
|
mlemanczyk
|
5bc65464d0
|
Initial PollardRho implementation for GPU. Added pool for accelerator streams shared across the threads. GPU processing performance and stabilization improvements. GpuUInt128.BinaryGcd switched to the version which outperformed other in benchmarks.
|
2025-11-12 21:18:30 +01:00 |
|
|
mlemanczyk
|
fdc2da7dae
|
Accelerator leak fixed.
|
2025-11-11 16:40:44 +01:00 |
|
|
mlemanczyk
|
cf5f78c2ef
|
PrimeTesterByLastDigitAccelerator integrated into Pow2MontgomeryGpuCalculator for better performance.
|
2025-11-11 13:36:21 +01:00 |
|
|
mlemanczyk
|
fa89958736
|
Stabilization improvements. Stream will be released when releasing GPU lease to prevent application crashes.
|
2025-11-10 14:30:30 +01:00 |
|
|
mlemanczyk
|
d0b1639a5d
|
GPU stabilization improvements to prevent calculation crashes
|
2025-11-09 21:52:16 +01:00 |
|
|
Marek Lemańczyk
|
4c4d4243fb
|
Merge remote-tracking branch 'github/codex/optimize-evenperfectbitscanner-for-static-arrays-khd9w8' into work
# Conflicts:
# PerfectNumbers.Core/Gpu/GpuContextPool.cs
# PerfectNumbers.Core/Gpu/GpuKernelLease.cs
# PerfectNumbers.Core/Gpu/GpuKernelPool.cs
# PerfectNumbers.Core/Gpu/GpuStaticTableInitializer.cs
# PerfectNumbers.Core/Gpu/Kernels/PrimeOrderSpecialMaxKernels.cs
# PerfectNumbers.Core/PrimeOrderCalculator.Cpu.cs
# PerfectNumbers.Core/PrimeOrderCalculator.Gpu.cs
# PerfectNumbers.Core/PrimeTester.cs
|
2025-11-03 23:59:12 +01:00 |
|
|
Marek Lemańczyk
|
76dfe108a9
|
Address feedback on GPU pooling and heuristic factoring
|
2025-11-02 11:22:49 +01:00 |
|
|
Marek Lemańczyk
|
5339c3c304
|
Use span growth instead of dictionaries in partial factor
|
2025-11-02 11:22:38 +01:00 |
|
|
mlemanczyk
|
2281d9535e
|
Revert "Merge pull request #251 from mlemanczyk/codex/refactor-prime-checking-on-gpu-n5t2pm"
This reverts commit 936ce239db, reversing
changes made to e14b77e504.
|
2025-11-02 11:22:14 +01:00 |
|
|
Marek Lemańczyk
|
8c662c5c85
|
Address feedback on GPU pooling and heuristic factoring
|
2025-10-31 21:47:42 +01:00 |
|
|
Marek Lemańczyk
|
567b92792c
|
Use span growth instead of dictionaries in partial factor
|
2025-10-31 18:52:00 +01:00 |
|
|
mlemanczyk
|
fc5700f82c
|
IsPrime with cancellation token removed.
|
2025-10-29 22:31:28 +01:00 |
|
|
mlemanczyk
|
0850227881
|
Switched to heuristic GPU for a better performance than the standard one.
|
2025-10-28 22:15:33 +01:00 |
|
|
mlemanczyk
|
ab045a62c4
|
Revert "Temporarily enabling heuristic GPU order checks."
This reverts commit 4db06cfe80.
|
2025-10-28 21:04:09 +01:00 |
|
|
mlemanczyk
|
4db06cfe80
|
Temporarily enabling heuristic GPU order checks.
|
2025-10-28 20:49:41 +01:00 |
|
|
Marek Lemańczyk
|
bd0727131f
|
Merge pull request #238 from mlemanczyk/codex/optimize-evenperfectbitscanner-loops-kit1so
Hoist prime order thresholds in CPU heuristics
|
2025-10-27 21:31:23 +01:00 |
|
|
Marek Lemańczyk
|
94a4ed8328
|
Hoist prime order thresholds in CPU heuristics
|
2025-10-27 21:24:43 +01:00 |
|
|
mlemanczyk
|
d381abe0cc
|
We'll now clear the dictionary after renting, to avoid unnecessary clearing if it's not going to be used anymore.
|
2025-10-27 16:54:49 +01:00 |
|
|
mlemanczyk
|
50d66a5186
|
Switched to GPU version in some places.
|
2025-10-27 15:30:10 +01:00 |
|
|
mlemanczyk
|
f2ce40ade9
|
Additional logging added.
|
2025-10-26 20:00:32 +01:00 |
|
|
Marek Lemańczyk
|
3758524450
|
Extract divisor validator and rename legacy CPU prime helper
|
2025-10-26 18:59:07 +01:00 |
|
|
Marek Lemańczyk
|
1e5eadac34
|
Refine heuristic prime tester integration
|
2025-10-26 18:15:27 +01:00 |
|
|
Marek Lemańczyk
|
43c7bafd5d
|
Restore CPU PrimeTester usage and rename fallback helpers
|
2025-10-26 15:19:49 +01:00 |
|
|
mlemanczyk
|
1ed5afe689
|
Switched to .IsPrime which performed best in benchmarks.
|
2025-10-26 04:05:00 +01:00 |
|
|
Marek Lemańczyk
|
15c229665d
|
Add heuristic prime sieves for cached divisor sequences
|
2025-10-26 03:21:29 +01:00 |
|
|
Marek Lemańczyk
|
fab9ba0413
|
Restore reviewed divisor heuristics
|
2025-10-24 19:50:32 +02:00 |
|
|
Marek Lemańczyk
|
bbe0c02069
|
Use bit masks for divisor heuristics
|
2025-10-24 19:50:27 +02:00 |
|
|
Marek Lemańczyk
|
1e5a9128cb
|
Split CPU and GPU exponent steppers
|
2025-10-24 08:49:53 +02:00 |
|
|
Marek Lemańczyk
|
7feeb08943
|
Clarify Pollard-Rho parity guard and seed CPU batches
|
2025-10-23 19:50:39 +02:00 |
|
|
mlemanczyk
|
865a88fac0
|
Switched to GPU for prime testing
|
2025-10-22 00:40:30 +02:00 |
|