I replaced score calculation by cuda graph with one kernel execution.
The updated score calculation is now twice as fast as previous one.
Additionally, this PR addresses a bug that occurred when the souce points size exeeded approximately 5000 or when VGF was disabled.
I replaced score calculation by cuda graph with one kernel execution. The updated score calculation is now twice as fast as previous one. Additionally, this PR addresses a bug that occurred when the souce points size exeeded approximately 5000 or when VGF was disabled.
(Block size was fixed after this PR.)