VK: scale auto shader compiler workers on wide CPUs
This commit is contained in:
@@ -220,9 +220,20 @@ namespace vk
|
||||
{
|
||||
if (num_worker_threads == 0)
|
||||
{
|
||||
// Select optimal number of compiler threads
|
||||
// Select a conservative but modern default for async pipeline compilation.
|
||||
// Older heuristics topped out too early on high-core CPUs and left large
|
||||
// shader bursts queued longer than necessary.
|
||||
const auto hw_threads = utils::get_thread_count();
|
||||
if (hw_threads > 12)
|
||||
|
||||
if (hw_threads >= 24)
|
||||
{
|
||||
num_worker_threads = 12;
|
||||
}
|
||||
else if (hw_threads >= 16)
|
||||
{
|
||||
num_worker_threads = 8;
|
||||
}
|
||||
else if (hw_threads > 12)
|
||||
{
|
||||
num_worker_threads = 6;
|
||||
}
|
||||
@@ -238,6 +249,9 @@ namespace vk
|
||||
{
|
||||
num_worker_threads = 1;
|
||||
}
|
||||
|
||||
rsx_log.notice("Async pipeline compiler auto-selected %d worker(s) for %u host thread(s).",
|
||||
num_worker_threads, hw_threads);
|
||||
}
|
||||
|
||||
ensure(num_worker_threads >= 1);
|
||||
|
||||
Reference in New Issue
Block a user