Default MaximumFrameLatency increased

[Spodi]

Defaults jitter fix (MaximumFrameLatency) to be on (2), instead of off (1), if nothing is set by a port. Keeping it at 1 has a lower latency, but I think it's more sensible to get a smoother experience by default.

[briaguya-ai]

Digging into this a bit https://github.com/search?q=repo%3AKenix3%2Flibultraship%20maximum_frame_latency&type=code

"apply"

libultraship/src/graphic/Fast3D/gfx_dxgi.cpp

Lines 111 to 129 in b704048

	static void apply_maximum_frame_latency(bool first) {
	DXGI_SWAP_CHAIN_DESC swap_desc = {};
	dxgi.swap_chain->GetDesc(&swap_desc);
	ComPtr<IDXGISwapChain2> swap_chain2;
	if ((swap_desc.Flags & DXGI_SWAP_CHAIN_FLAG_FRAME_LATENCY_WAITABLE_OBJECT) &&
	dxgi.swap_chain->QueryInterface(__uuidof(IDXGISwapChain2), &swap_chain2) == S_OK) {
	ThrowIfFailed(swap_chain2->SetMaximumFrameLatency(dxgi.maximum_frame_latency));
	if (first) {
	dxgi.waitable_object = swap_chain2->GetFrameLatencyWaitableObject();
	WaitForSingleObject(dxgi.waitable_object, INFINITE);
	}
	} else {
	ComPtr<IDXGIDevice1> device1;
	ThrowIfFailed(dxgi.swap_chain->GetDevice(__uuidof(IDXGIDevice1), &device1));
	ThrowIfFailed(device1->SetMaximumFrameLatency(dxgi.maximum_frame_latency));
	}
	dxgi.applied_maximum_frame_latency = dxgi.maximum_frame_latency;
	}

the line being changed in this PR

libultraship/src/graphic/Fast3D/gfx_dxgi.cpp

Line 465 in b704048

dxgi.maximum_frame_latency = 1;

a bunch of logic in swap buffers end

libultraship/src/graphic/Fast3D/gfx_dxgi.cpp

Lines 866 to 909 in b704048

	static void gfx_dxgi_swap_buffers_end() {
	LARGE_INTEGER t0, t1, t2;
	QueryPerformanceCounter(&t0);
	QueryPerformanceCounter(&t1);
	if (dxgi.applied_maximum_frame_latency > dxgi.maximum_frame_latency) {
	// If latency is decreased, you have to wait the same amout of times as the old latency was set to
	int times_to_wait = dxgi.applied_maximum_frame_latency;
	int latency = dxgi.maximum_frame_latency;
	dxgi.maximum_frame_latency = 1;
	apply_maximum_frame_latency(false);
	if (dxgi.waitable_object != nullptr) {
	while (times_to_wait > 0) {
	WaitForSingleObject(dxgi.waitable_object, INFINITE);
	times_to_wait--;
	}
	}
	dxgi.maximum_frame_latency = latency;
	apply_maximum_frame_latency(false);
	return; // Make sure we don't wait a second time on the waitable object, since that would hang the program
	} else if (dxgi.applied_maximum_frame_latency != dxgi.maximum_frame_latency) {
	apply_maximum_frame_latency(false);
	}
	if (!dxgi.dropped_frame) {
	if (dxgi.waitable_object != nullptr) {
	WaitForSingleObject(dxgi.waitable_object, INFINITE);
	}
	// else TODO: maybe sleep until some estimated time the frame will be shown to reduce lag
	}
	DXGI_FRAME_STATISTICS stats;
	dxgi.swap_chain->GetFrameStatistics(&stats);
	QueryPerformanceCounter(&t2);
	dxgi.zero_latency = dxgi.pending_frame_stats.rbegin()->first == stats.PresentCount;
	// printf(L"done %I64u gpu:%d wait:%d freed:%I64u frame:%u %u monitor:%u t:%I64u\n", (unsigned long
	// long)(t0.QuadPart - dxgi.qpc_init), (int)(t1.QuadPart - t0.QuadPart), (int)(t2.QuadPart - t0.QuadPart), (unsigned
	// long long)(t2.QuadPart - dxgi.qpc_init), dxgi.pending_frame_stats.rbegin()->first, stats.PresentCount,
	// stats.SyncRefreshCount, (unsigned long long)(stats.SyncQPCTime.QuadPart - dxgi.qpc_init));
	}

what is called when ports set this using Fast3dWindow::SetMaximumFrameLatency

libultraship/src/graphic/Fast3D/gfx_dxgi.cpp

Lines 925 to 927 in b704048

	static void gfx_dxgi_set_maximum_frame_latency(int latency) {
	dxgi.maximum_frame_latency = latency;
	}

a call to apply when creating swapchain

libultraship/src/graphic/Fast3D/gfx_dxgi.cpp

Lines 966 to 996 in b704048

	void gfx_dxgi_create_swap_chain(IUnknown* device, std::function<void()>&& before_destroy_fn) {
	bool win8 = IsWindows8OrGreater(); // DXGI_SCALING_NONE is only supported on Win8 and beyond
	bool dxgi_13 = dxgi.CreateDXGIFactory2 != nullptr; // DXGI 1.3 introduced waitable object
	DXGI_SWAP_CHAIN_DESC1 swap_chain_desc = {};
	swap_chain_desc.BufferCount = 3;
	swap_chain_desc.Width = 0;
	swap_chain_desc.Height = 0;
	swap_chain_desc.Format = DXGI_FORMAT_R8G8B8A8_UNORM;
	swap_chain_desc.BufferUsage = DXGI_USAGE_RENDER_TARGET_OUTPUT;
	swap_chain_desc.Scaling = win8 ? DXGI_SCALING_NONE : DXGI_SCALING_STRETCH;
	swap_chain_desc.SwapEffect =
	dxgi.dxgi1_4 ? DXGI_SWAP_EFFECT_FLIP_DISCARD : // Introduced in DXGI 1.4 and Windows 10
	DXGI_SWAP_EFFECT_FLIP_SEQUENTIAL; // Apparently flip sequential was also backported to Win 7 Platform Update
	swap_chain_desc.Flags = dxgi_13 ? DXGI_SWAP_CHAIN_FLAG_FRAME_LATENCY_WAITABLE_OBJECT : 0;
	if (dxgi.tearing_support) {
	swap_chain_desc.Flags |= DXGI_SWAP_CHAIN_FLAG_ALLOW_TEARING; // Now we can use DXGI_PRESENT_ALLOW_TEARING
	}
	swap_chain_desc.SampleDesc.Count = 1;
	ThrowIfFailed(
	dxgi.factory->CreateSwapChainForHwnd(device, dxgi.h_wnd, &swap_chain_desc, nullptr, nullptr, &dxgi.swap_chain));
	ThrowIfFailed(dxgi.factory->MakeWindowAssociation(dxgi.h_wnd, DXGI_MWA_NO_ALT_ENTER));
	apply_maximum_frame_latency(true);
	ThrowIfFailed(dxgi.swap_chain->GetDesc1(&swap_chain_desc));
	dxgi.swap_chain_device = device;
	dxgi.before_destroy_swap_chain_fn = std::move(before_destroy_fn);
	}

there seems to be a lot here - if nothing else

libultraship/src/graphic/Fast3D/gfx_dxgi.cpp

Line 875 in b704048

dxgi.maximum_frame_latency = 1;

probably needs to be updated to 2 as well, but any explanation of why so much logic is needed for what seems like something we should just set and forget would be really helpful

[briaguya-ai]

after rereading this

libultraship/src/graphic/Fast3D/gfx_dxgi.cpp

Line 875 in b704048

dxgi.maximum_frame_latency = 1;

probably needs to be updated to 2 as well, but any explanation of why so much logic is needed for what seems like something we should just set and forget would be really helpful

it seems that spot can stay 1

breaking down

libultraship/src/graphic/Fast3D/gfx_dxgi.cpp

Lines 871 to 887 in b704048

	if (dxgi.applied_maximum_frame_latency > dxgi.maximum_frame_latency) {
	// If latency is decreased, you have to wait the same amout of times as the old latency was set to
	int times_to_wait = dxgi.applied_maximum_frame_latency;
	int latency = dxgi.maximum_frame_latency;
	dxgi.maximum_frame_latency = 1;
	apply_maximum_frame_latency(false);
	if (dxgi.waitable_object != nullptr) {
	while (times_to_wait > 0) {
	WaitForSingleObject(dxgi.waitable_object, INFINITE);
	times_to_wait--;
	}
	}
	dxgi.maximum_frame_latency = latency;
	apply_maximum_frame_latency(false);
	return; // Make sure we don't wait a second time on the waitable object, since that would hang the program
	} else if (dxgi.applied_maximum_frame_latency != dxgi.maximum_frame_latency) {

current latency value is stored in local var

libultraship/src/graphic/Fast3D/gfx_dxgi.cpp

Lines 871 to 874 in b704048

	if (dxgi.applied_maximum_frame_latency > dxgi.maximum_frame_latency) {
	// If latency is decreased, you have to wait the same amout of times as the old latency was set to
	int times_to_wait = dxgi.applied_maximum_frame_latency;
	int latency = dxgi.maximum_frame_latency;

it's set to 1 to "run out" the existing latency frames or something?

libultraship/src/graphic/Fast3D/gfx_dxgi.cpp

Lines 875 to 882 in b704048

	dxgi.maximum_frame_latency = 1;
	apply_maximum_frame_latency(false);
	if (dxgi.waitable_object != nullptr) {
	while (times_to_wait > 0) {
	WaitForSingleObject(dxgi.waitable_object, INFINITE);
	times_to_wait--;
	}
	}

it is set back to the new applied latency value (which was stored in the local var)

libultraship/src/graphic/Fast3D/gfx_dxgi.cpp

Lines 883 to 886 in b704048

	dxgi.maximum_frame_latency = latency;
	apply_maximum_frame_latency(false);
	return; // Make sure we don't wait a second time on the waitable object, since that would hang the program

[briaguya-ai]

i still feel the ability to change this at runtime is adding a ton of complexity we likely don't need, but from a "minimal change maximum impact" perspective this seems like a good one

[Spodi]

libultraship/src/graphic/Fast3D/gfx_dxgi.cpp

Lines 871 to 874 in b704048

	if (dxgi.applied_maximum_frame_latency > dxgi.maximum_frame_latency) {
	// If latency is decreased, you have to wait the same amout of times as the old latency was set to
	int times_to_wait = dxgi.applied_maximum_frame_latency;
	int latency = dxgi.maximum_frame_latency;

it's set to 1 to "run out" the existing latency frames or something?

libultraship/src/graphic/Fast3D/gfx_dxgi.cpp

Lines 875 to 882 in b704048

	dxgi.maximum_frame_latency = 1;
	apply_maximum_frame_latency(false);
	if (dxgi.waitable_object != nullptr) {
	while (times_to_wait > 0) {
	WaitForSingleObject(dxgi.waitable_object, INFINITE);
	times_to_wait--;
	}
	}

it is set back to the new applied latency value (which was stored in the local var)

libultraship/src/graphic/Fast3D/gfx_dxgi.cpp

Lines 883 to 886 in b704048

	dxgi.maximum_frame_latency = latency;
	apply_maximum_frame_latency(false);
	return; // Make sure we don't wait a second time on the waitable object, since that would hang the program

Yes, setting it to 1 and than waiting in a loop is to get the latency down again. This is the "undocumented stuff" I talked about. It's never mentioned in the official docs, that you can't just set it at runtime and latency will stay up instead.
In fact the previous aproach here was to create a completely new swapchain. Even the one that wrote the old approach was seemingly supprised, that latency won't go down automatically (according to the comments).
Creating a new swapchain had the problem of confusing ReShade, SK and whatever else is out there. And sometimes also did strange things on its own, when the jitter fix option was changed too fast and therefore multiple swapchains where created in a quick succesion.