armsx/ARMSX3
0
0
Fork 0
mirror of https://github.com/ARMSX2/ARMSX3.git synced 2026-08-24 16:58:52 -07:00
Code Issues Packages Projects Releases Wiki Activity

Revert "VK: make the fence poll mechanism selectable at runtime"

Browse Source 
This reverts commit b91c6551ed.
This commit is contained in:
jpolo1224
2026-08-21 01:23:10 -04:00
parent b91c6551ed
commit 0262053a91
2 changed files with 4 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files
rpcs3/Emu/RSX/VK/VKGSRender.cpp
+1 -3
View File
@@ -1256,9 +1256,7 @@ void VKGSRender::check_present_status()
while (!m_queued_frames.empty())
{
auto ctx = m_queued_frames.front();
// Speculative: this walks the queue looking for frames that have already finished and
// gives up on the first that has not, so it must not be the thing that waits for one.
if (!ctx->swap_command_buffer->poke(false))
if (!ctx->swap_command_buffer->poke())
{
return;
}
rpcs3/Emu/RSX/VK/VKGSRenderTypes.hpp
+3 -45
View File
@@ -77,48 +77,8 @@ namespace vk
vk::command_buffer::reset();
}
// How to ask "is this submission done", selectable at runtime via driver_env.txt.
//
// wait0 (default) vkWaitForFences with a zero timeout, today's behaviour
// status vkGetFenceStatus, what this used before
// defer speculative callers do not ask the driver at all
//
// A toggle rather than a change because this has already been "fixed" once on one
// device: vkGetFenceStatus measured 19.7ms per call and never once returned
// VK_NOT_READY, and the zero-timeout wait that replaced it is measurably better but has
// the same signature -- 1.9ms per call at 0.0% not ready on Adreno 740, roughly 5ms a
// frame in Minecraft. The driver appears to treat a zero timeout as "wait" either way,
// which is a driver behaviour and not something to assume holds everywhere.
//
// The profiler already reports "fence polls N/frame, T ns each, P% not ready", so each
// mode can be judged directly on whatever device is testing.
static int fence_poll_mode()
bool poke()
{
static const int mode = []
{
const char* const env = std::getenv("ARMSX3_FENCE_POLL");
if (!env) return 0;
if (std::string_view(env) == "status") return 1;
if (std::string_view(env) == "defer") return 2;
return 0;
}();
return mode;
}
// allow_block=false marks a caller that only wants to reclaim early and is happy to be
// told "not yet": check_present_status returns immediately on a false, and poke_all
// ignores the result entirely. The mandatory callers -- next() and reset() -- keep the
// default, so the ring can never be starved by this.
bool poke(bool allow_block = true)
{
if (!allow_block && fence_poll_mode() == 2 && is_pending)
{
// Not asked, so not known. The work is reclaimed by a later frame, or by the
// blocking path in next()/reset() if the ring gets that far.
return false;
}
reader_lock lock(guard_mutex);
if (!is_pending)
@@ -142,9 +102,7 @@ namespace vk
// vkWaitForFences with a zero timeout is specified to return VK_TIMEOUT
// without waiting, and is the same question with an answer that arrives.
RSX_PROF_SCOPE(fence_poll);
fence_status = fence_poll_mode() == 1
? vkGetFenceStatus(pool->get_owner(), m_submit_fence->handle)
: vkWaitForFences(pool->get_owner(), 1, &m_submit_fence->handle, VK_TRUE, 0);
fence_status = vkWaitForFences(pool->get_owner(), 1, &m_submit_fence->handle, VK_TRUE, 0);
}
if (rsx::prof::enabled()) [[unlikely]]
@@ -377,7 +335,7 @@ namespace vk
{
for (auto& cb : m_cb_list)
{
cb.poke(false);
cb.poke();
}
}