From e60024f0afecb6f92e383221084f1a7af02f8542 Mon Sep 17 00:00:00 2001 From: Dave Parks Date: Tue, 29 Mar 2022 15:41:00 -0500 Subject: SL-17005 WIP - Use D3D/DXGI to query for available VRAM on Windows --- indra/llwindow/llwindowwin32.cpp | 136 +++++++++++++++++++++++++++++++++++++++ 1 file changed, 136 insertions(+) (limited to 'indra/llwindow/llwindowwin32.cpp') diff --git a/indra/llwindow/llwindowwin32.cpp b/indra/llwindow/llwindowwin32.cpp index 9feccd7874..3afb78d1d2 100644 --- a/indra/llwindow/llwindowwin32.cpp +++ b/indra/llwindow/llwindowwin32.cpp @@ -47,6 +47,7 @@ #include "llglslshader.h" #include "llthreadsafequeue.h" #include "stringize.h" +#include "llframetimer.h" // System includes #include @@ -61,6 +62,10 @@ #include #include // std::pair +#include +#include + + // Require DirectInput version 8 #define DIRECTINPUT_VERSION 0x0800 @@ -347,6 +352,20 @@ struct LLWindowWin32::LLWindowWin32Thread : public LL::ThreadPool void run() override; + // initialzie DXGI adapter (for querying available VRAM) + void initDX(); + + // initialize D3D (if DXGI cannot be used) + void initD3D(); + + // call periodically to update available VRAM + void updateVRAMUsage(); + + U32 getAvailableVRAMMegabytes() + { + return mAvailableVRAM; + } + /// called by main thread to post work to this window thread template void post(CALLABLE&& func) @@ -395,6 +414,12 @@ struct LLWindowWin32::LLWindowWin32Thread : public LL::ThreadPool void gatherInput(); HWND mWindowHandle = NULL; HDC mhDC = 0; + + std::atomic mAvailableVRAM; + + IDXGIAdapter3* mDXGIAdapter = nullptr; + LPDIRECT3D9 mD3D = nullptr; + LPDIRECT3DDEVICE9 mD3DDevice = nullptr; }; @@ -4507,6 +4532,10 @@ std::vector LLWindowWin32::getDynamicFallbackFontList() return std::vector(); } +U32 LLWindowWin32::getAvailableVRAMMegabytes() +{ + return mWindowThread ? mWindowThread->getAvailableVRAMMegabytes() : 0; +} #endif // LL_WINDOWS @@ -4562,17 +4591,97 @@ private: std::string mPrev; }; +void LLWindowWin32::LLWindowWin32Thread::initDX() +{ + if (mDXGIAdapter == NULL) + { + IDXGIFactory4* pFactory = nullptr; + + HRESULT res = CreateDXGIFactory1(__uuidof(IDXGIFactory4), (void**)&pFactory); + + if (FAILED(res)) + { + LL_WARNS() << "CreateDXGIFactory1 failed: 0x" << std::hex << res << LL_ENDL; + } + else + { + res = pFactory->EnumAdapters(0, reinterpret_cast(&mDXGIAdapter)); + if (FAILED(res)) + { + LL_WARNS() << "EnumAdapters failed: 0x" << std::hex << res << LL_ENDL; + } + } + + pFactory->Release(); + } +} + +void LLWindowWin32::LLWindowWin32Thread::initD3D() +{ + if (mDXGIAdapter == NULL && mD3DDevice == NULL && mWindowHandle != 0) + { + mD3D = Direct3DCreate9(D3D_SDK_VERSION); + + D3DPRESENT_PARAMETERS d3dpp; + + ZeroMemory(&d3dpp, sizeof(d3dpp)); + d3dpp.Windowed = TRUE; + d3dpp.SwapEffect = D3DSWAPEFFECT_DISCARD; + + mD3D->CreateDevice(D3DADAPTER_DEFAULT, D3DDEVTYPE_HAL, mWindowHandle, D3DCREATE_SOFTWARE_VERTEXPROCESSING, &d3dpp, &mD3DDevice); + } +} + +void LLWindowWin32::LLWindowWin32Thread::updateVRAMUsage() +{ + LL_PROFILE_ZONE_SCOPED; + if (mDXGIAdapter != nullptr) + { + DXGI_QUERY_VIDEO_MEMORY_INFO info; + mDXGIAdapter->QueryVideoMemoryInfo(0, DXGI_MEMORY_SEGMENT_GROUP_LOCAL, &info); + + // try to use no more than the available reserve minus 10% + U32 target = info.AvailableForReservation / 1024 / 1024; + target -= target / 10; + + U32 used_vram = info.CurrentUsage / 1024 / 1024; + + mAvailableVRAM = used_vram < target ? target - used_vram : 0; + + /*LL_INFOS() << "\nLocal\nAFR: " << info.AvailableForReservation / 1024 / 1024 + << "\nBudget: " << info.Budget / 1024 / 1024 + << "\nCR: " << info.CurrentReservation / 1024 / 1024 + << "\nCU: " << info.CurrentUsage / 1024 / 1024 << LL_ENDL; + + mDXGIAdapter->QueryVideoMemoryInfo(0, DXGI_MEMORY_SEGMENT_GROUP_NON_LOCAL, &info); + LL_INFOS() << "\nNon-Local\nAFR: " << info.AvailableForReservation / 1024 / 1024 + << "\nBudget: " << info.Budget / 1024 / 1024 + << "\nCR: " << info.CurrentReservation / 1024 / 1024 + << "\nCU: " << info.CurrentUsage / 1024 / 1024 << LL_ENDL;*/ + } + else if (mD3DDevice != NULL) + { // fallback to D3D9 + mAvailableVRAM = mD3DDevice->GetAvailableTextureMem() / 1024 / 1024; + } + +} + void LLWindowWin32::LLWindowWin32Thread::run() { sWindowThreadId = std::this_thread::get_id(); LogChange logger("Window"); + initDX(); + while (! getQueue().done()) { LL_PROFILE_ZONE_SCOPED_CATEGORY_WIN32; if (mWindowHandle != 0) { + // lazily call initD3D inside this loop to catch when mWindowHandle has been set + initD3D(); + MSG msg; BOOL status; if (mhDC == 0) @@ -4605,6 +4714,13 @@ void LLWindowWin32::LLWindowWin32Thread::run() getQueue().runPending(); } + // update available vram once every 3 seconds + static LLFrameTimer vramTimer; + if (vramTimer.getElapsedTimeF32() > 3.f) + { + updateVRAMUsage(); + vramTimer.reset(); + } #if 0 { LL_PROFILE_ZONE_NAMED_CATEGORY_WIN32("w32t - Sleep"); @@ -4613,6 +4729,26 @@ void LLWindowWin32::LLWindowWin32Thread::run() } #endif } + + //clean up DXGI/D3D resources + if (mDXGIAdapter) + { + mDXGIAdapter->Release(); + mDXGIAdapter = nullptr; + } + + if (mD3DDevice) + { + mD3DDevice->Release(); + mD3DDevice = nullptr; + } + + if (mD3D) + { + mD3D->Release(); + mD3D = nullptr; + } + } void LLWindowWin32::post(const std::function& func) -- cgit v1.2.3 From da069765b1974e60df5ebf9e973c7fd00c52e329 Mon Sep 17 00:00:00 2001 From: Dave Parks Date: Wed, 8 Jun 2022 16:24:03 -0500 Subject: SL-17005 Tweak target VRAM usage and add some log spam to help zero in on good targets. --- indra/llwindow/llwindowwin32.cpp | 22 ++++++++++++++++------ 1 file changed, 16 insertions(+), 6 deletions(-) (limited to 'indra/llwindow/llwindowwin32.cpp') diff --git a/indra/llwindow/llwindowwin32.cpp b/indra/llwindow/llwindowwin32.cpp index da79660239..20443988ab 100644 --- a/indra/llwindow/llwindowwin32.cpp +++ b/indra/llwindow/llwindowwin32.cpp @@ -65,7 +65,6 @@ #include #include - // Require DirectInput version 8 #define DIRECTINPUT_VERSION 0x0800 @@ -4649,23 +4648,34 @@ void LLWindowWin32::LLWindowWin32Thread::updateVRAMUsage() mDXGIAdapter->QueryVideoMemoryInfo(0, DXGI_MEMORY_SEGMENT_GROUP_LOCAL, &info); // try to use no more than the available reserve minus 10% - U32 target = info.AvailableForReservation / 1024 / 1024; - target -= target / 10; + U32 target = info.Budget / 1024 / 1024; + + // EXPERIMENTAL + // Trying to zero in on a good target usage, code here should be tuned against observed behavior + // of various hardware. + if (target > 4096) // if 4GB are installed, try to leave 2GB free + { + target -= 2048; + } + else // if less than 4GB are installed, try not to use more than half of it + { + target /= 2; + } U32 used_vram = info.CurrentUsage / 1024 / 1024; mAvailableVRAM = used_vram < target ? target - used_vram : 0; - /*LL_INFOS() << "\nLocal\nAFR: " << info.AvailableForReservation / 1024 / 1024 + LL_INFOS("Window") << "\nLocal\nAFR: " << info.AvailableForReservation / 1024 / 1024 << "\nBudget: " << info.Budget / 1024 / 1024 << "\nCR: " << info.CurrentReservation / 1024 / 1024 << "\nCU: " << info.CurrentUsage / 1024 / 1024 << LL_ENDL; mDXGIAdapter->QueryVideoMemoryInfo(0, DXGI_MEMORY_SEGMENT_GROUP_NON_LOCAL, &info); - LL_INFOS() << "\nNon-Local\nAFR: " << info.AvailableForReservation / 1024 / 1024 + LL_INFOS("Window") << "\nNon-Local\nAFR: " << info.AvailableForReservation / 1024 / 1024 << "\nBudget: " << info.Budget / 1024 / 1024 << "\nCR: " << info.CurrentReservation / 1024 / 1024 - << "\nCU: " << info.CurrentUsage / 1024 / 1024 << LL_ENDL;*/ + << "\nCU: " << info.CurrentUsage / 1024 / 1024 << LL_ENDL; } else if (mD3DDevice != NULL) { // fallback to D3D9 -- cgit v1.2.3 From 47ee4aff475e5086bc532acf1b79a86adcb70d73 Mon Sep 17 00:00:00 2001 From: Cosmic Linden Date: Tue, 9 Aug 2022 11:34:09 -0700 Subject: SL-17005: Fix crash in initDX when no factory --- indra/llwindow/llwindowwin32.cpp | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) (limited to 'indra/llwindow/llwindowwin32.cpp') diff --git a/indra/llwindow/llwindowwin32.cpp b/indra/llwindow/llwindowwin32.cpp index 20443988ab..6e5c1e504a 100644 --- a/indra/llwindow/llwindowwin32.cpp +++ b/indra/llwindow/llwindowwin32.cpp @@ -4619,7 +4619,10 @@ void LLWindowWin32::LLWindowWin32Thread::initDX() } } - pFactory->Release(); + if (pFactory) + { + pFactory->Release(); + } } } -- cgit v1.2.3 From eac4540fc359b1390c8828f92c5ba75d44fd1308 Mon Sep 17 00:00:00 2001 From: Cosmic Linden Date: Tue, 9 Aug 2022 11:48:51 -0700 Subject: SL-17005: (WIP) (DEBUG) Add logging for graphics adapters --- indra/llwindow/llwindowwin32.cpp | 67 ++++++++++++++++++++++++++++++++++++++++ 1 file changed, 67 insertions(+) (limited to 'indra/llwindow/llwindowwin32.cpp') diff --git a/indra/llwindow/llwindowwin32.cpp b/indra/llwindow/llwindowwin32.cpp index 6e5c1e504a..9e7ea15cd7 100644 --- a/indra/llwindow/llwindowwin32.cpp +++ b/indra/llwindow/llwindowwin32.cpp @@ -4598,10 +4598,77 @@ private: std::string mPrev; }; +// Print hardware debug info about available graphics adapters in ordinal order +void debugEnumerateGraphicsAdapters() +{ + LL_INFOS("Window") << "Enumerating graphics adapters..." << LL_ENDL; + + IDXGIFactory1* factory; + HRESULT res = CreateDXGIFactory1(__uuidof(IDXGIFactory1), (void**)&factory); + if (FAILED(res) || !factory) + { + LL_WARNS() << "CreateDXGIFactory1 failed: 0x" << std::hex << res << LL_ENDL; + } + else + { + UINT graphics_adapter_index = 0; + IDXGIAdapter3* dxgi_adapter; + while (true) + { + res = factory->EnumAdapters(graphics_adapter_index, reinterpret_cast(&dxgi_adapter)); + if (FAILED(res)) + { + if (graphics_adapter_index == 0) + { + LL_WARNS() << "EnumAdapters failed: 0x" << std::hex << res << LL_ENDL; + } + else + { + LL_INFOS("Window") << "Done enumerating graphics adapters" << LL_ENDL; + } + } + else + { + DXGI_ADAPTER_DESC desc; + dxgi_adapter->GetDesc(&desc); + std::wstring description_w((wchar_t*)desc.Description); + std::string description(description_w.begin(), description_w.end()); + LL_INFOS("Window") << "Graphics adapter index: " << graphics_adapter_index << ", " + << "Description: " << description << ", " + << "DeviceId: " << desc.DeviceId << ", " + << "SubSysId: " << desc.SubSysId << ", " + << "AdapterLuid: " << desc.AdapterLuid.HighPart << "_" << desc.AdapterLuid.LowPart << ", " + << "DedicatedVideoMemory: " << desc.DedicatedVideoMemory / 1024 / 1024 << ", " + << "DedicatedSystemMemory: " << desc.DedicatedSystemMemory / 1024 / 1024 << ", " + << "SharedSystemMemory: " << desc.SharedSystemMemory / 1024 / 1024 << LL_ENDL; + } + + if (dxgi_adapter) + { + dxgi_adapter->Release(); + dxgi_adapter = NULL; + } + else + { + break; + } + + graphics_adapter_index++; + } + } + + if (factory) + { + factory->Release(); + } +} + void LLWindowWin32::LLWindowWin32Thread::initDX() { if (mDXGIAdapter == NULL) { + debugEnumerateGraphicsAdapters(); + IDXGIFactory4* pFactory = nullptr; HRESULT res = CreateDXGIFactory1(__uuidof(IDXGIFactory4), (void**)&pFactory); -- cgit v1.2.3 From c790bf7f49328b479f1bcbfe0268ca0bec6e2376 Mon Sep 17 00:00:00 2001 From: Cosmic Linden Date: Thu, 11 Aug 2022 11:10:25 -0700 Subject: SL-17005: (WIP) (Debug) Add logging and debug settings to assist in understanding Windows 8 failure modes. Most of this should be removed later --- indra/llwindow/llwindowwin32.cpp | 32 +++++++++++++++++++++++++++++--- 1 file changed, 29 insertions(+), 3 deletions(-) (limited to 'indra/llwindow/llwindowwin32.cpp') diff --git a/indra/llwindow/llwindowwin32.cpp b/indra/llwindow/llwindowwin32.cpp index 9e7ea15cd7..9e2e07102e 100644 --- a/indra/llwindow/llwindowwin32.cpp +++ b/indra/llwindow/llwindowwin32.cpp @@ -48,6 +48,8 @@ #include "llthreadsafequeue.h" #include "stringize.h" #include "llframetimer.h" +#include "commoncontrol.h" // TODO: Remove after testing +#include "llsd.h" // TODO: Remove after testing // System includes #include @@ -416,7 +418,9 @@ struct LLWindowWin32::LLWindowWin32Thread : public LL::ThreadPool std::atomic mAvailableVRAM; + bool mTryUseDXGIAdapter; // TODO: Remove after testing IDXGIAdapter3* mDXGIAdapter = nullptr; + bool mTryUseD3DDevice; // TODO: Remove after testing LPDIRECT3D9 mD3D = nullptr; LPDIRECT3DDEVICE9 mD3DDevice = nullptr; }; @@ -4549,6 +4553,14 @@ U32 LLWindowWin32::getAvailableVRAMMegabytes() inline LLWindowWin32::LLWindowWin32Thread::LLWindowWin32Thread() : ThreadPool("Window Thread", 1, MAX_QUEUE_SIZE) { + const LLSD skipDXGI{ LL::CommonControl::get("Global", "DisablePrimaryGraphicsMemoryAccounting") }; // TODO: Remove after testing + LL_WARNS() << "DisablePrimaryGraphicsMemoryAccounting: " << skipDXGI << ", as boolean: " << skipDXGI.asBoolean() << LL_ENDL; + mTryUseDXGIAdapter = !skipDXGI.asBoolean(); + LL_WARNS() << "mTryUseDXGIAdapter: " << mTryUseDXGIAdapter << LL_ENDL; + const LLSD skipD3D{ LL::CommonControl::get("Global", "DisableSecondaryGraphicsMemoryAccounting") }; // TODO: Remove after testing + LL_WARNS() << "DisableSecondaryGraphicsMemoryAccounting: " << skipD3D << ", as boolean: " << skipD3D.asBoolean() << LL_ENDL; + mTryUseD3DDevice = !skipD3D.asBoolean(); + LL_WARNS() << "mTryUseD3DDevice: " << mTryUseD3DDevice << LL_ENDL; ThreadPool::start(); } @@ -4665,7 +4677,7 @@ void debugEnumerateGraphicsAdapters() void LLWindowWin32::LLWindowWin32Thread::initDX() { - if (mDXGIAdapter == NULL) + if (mDXGIAdapter == NULL && mTryUseDXGIAdapter) { debugEnumerateGraphicsAdapters(); @@ -4684,6 +4696,10 @@ void LLWindowWin32::LLWindowWin32Thread::initDX() { LL_WARNS() << "EnumAdapters failed: 0x" << std::hex << res << LL_ENDL; } + else + { + LL_INFOS() << "EnumAdapters success" << LL_ENDL; + } } if (pFactory) @@ -4695,7 +4711,7 @@ void LLWindowWin32::LLWindowWin32Thread::initDX() void LLWindowWin32::LLWindowWin32Thread::initD3D() { - if (mDXGIAdapter == NULL && mD3DDevice == NULL && mWindowHandle != 0) + if (mDXGIAdapter == NULL && mD3DDevice == NULL && mTryUseD3DDevice && mWindowHandle != 0) { mD3D = Direct3DCreate9(D3D_SDK_VERSION); @@ -4705,7 +4721,16 @@ void LLWindowWin32::LLWindowWin32Thread::initD3D() d3dpp.Windowed = TRUE; d3dpp.SwapEffect = D3DSWAPEFFECT_DISCARD; - mD3D->CreateDevice(D3DADAPTER_DEFAULT, D3DDEVTYPE_HAL, mWindowHandle, D3DCREATE_SOFTWARE_VERTEXPROCESSING, &d3dpp, &mD3DDevice); + HRESULT res = mD3D->CreateDevice(D3DADAPTER_DEFAULT, D3DDEVTYPE_HAL, mWindowHandle, D3DCREATE_SOFTWARE_VERTEXPROCESSING, &d3dpp, &mD3DDevice); + + if (FAILED(res)) + { + LL_WARNS() << "(fallback) CreateDevice failed: 0x" << std::hex << res << LL_ENDL; + } + else + { + LL_INFOS() << "(fallback) CreateDevice success" << LL_ENDL; + } } } @@ -4768,6 +4793,7 @@ void LLWindowWin32::LLWindowWin32Thread::run() if (mWindowHandle != 0) { // lazily call initD3D inside this loop to catch when mWindowHandle has been set + // *TODO: Shutdown if this fails when mWindowHandle exists initD3D(); MSG msg; -- cgit v1.2.3 From 9bee2a92d2eb1f48f8ca0061e546c5e9df38fe79 Mon Sep 17 00:00:00 2001 From: Dave Parks Date: Fri, 26 Aug 2022 10:51:42 -0500 Subject: SL-17997 Follow up from beta breakers results. --- indra/llwindow/llwindowwin32.cpp | 55 +++++++++++++++++++++++++++++----------- 1 file changed, 40 insertions(+), 15 deletions(-) (limited to 'indra/llwindow/llwindowwin32.cpp') diff --git a/indra/llwindow/llwindowwin32.cpp b/indra/llwindow/llwindowwin32.cpp index 20443988ab..656439e71e 100644 --- a/indra/llwindow/llwindowwin32.cpp +++ b/indra/llwindow/llwindowwin32.cpp @@ -414,6 +414,7 @@ struct LLWindowWin32::LLWindowWin32Thread : public LL::ThreadPool HWND mWindowHandle = NULL; HDC mhDC = 0; + // best guess at available video memory in MB std::atomic mAvailableVRAM; IDXGIAdapter3* mDXGIAdapter = nullptr; @@ -4644,44 +4645,68 @@ void LLWindowWin32::LLWindowWin32Thread::updateVRAMUsage() LL_PROFILE_ZONE_SCOPED; if (mDXGIAdapter != nullptr) { + // NOTE: what lies below is hand wavy math based on compatibility testing and observation against a variety of hardware + // It doesn't make sense, but please don't refactor it to make sense. -- davep + DXGI_QUERY_VIDEO_MEMORY_INFO info; mDXGIAdapter->QueryVideoMemoryInfo(0, DXGI_MEMORY_SEGMENT_GROUP_LOCAL, &info); - // try to use no more than the available reserve minus 10% - U32 target = info.Budget / 1024 / 1024; +#if 0 // debug 0 budget and 0 CU + info.Budget = 0; + info.CurrentUsage = 0; +#endif + + U32 budget_mb = info.Budget / 1024 / 1024; + U32 afr_mb = info.AvailableForReservation / 1024 / 1024; + // correct for systems that misreport budget + if (budget_mb == 0) + { + // fall back to available for reservation clamped between 512MB and 2GB + budget_mb = llclamp(afr_mb, (U32) 512, (U32) 2048); + } + + U32 cu_mb = info.CurrentUsage / 1024 / 1024; + + // get an estimated usage based on texture bytes allocated + U32 eu_mb = LLImageGL::getTextureBytesAllocated() * 2 / 1024 / 1024; - // EXPERIMENTAL - // Trying to zero in on a good target usage, code here should be tuned against observed behavior - // of various hardware. - if (target > 4096) // if 4GB are installed, try to leave 2GB free + if (cu_mb == 0) + { // current usage is sometimes unreliable on Intel GPUs, fall back to estimated usage + cu_mb = llmax((U32)1, eu_mb); + } + F32 eu_error = (F32)((S32)eu_mb - (S32)cu_mb) / (F32)cu_mb; + + U32 target_mb = info.Budget / 1024 / 1024; + + if (target_mb > 4096) // if 4GB are installed, try to leave 2GB free { - target -= 2048; + target_mb -= 2048; } else // if less than 4GB are installed, try not to use more than half of it { - target /= 2; + target_mb /= 2; } - U32 used_vram = info.CurrentUsage / 1024 / 1024; - - mAvailableVRAM = used_vram < target ? target - used_vram : 0; + mAvailableVRAM = cu_mb < target_mb ? target_mb - cu_mb : 0; LL_INFOS("Window") << "\nLocal\nAFR: " << info.AvailableForReservation / 1024 / 1024 << "\nBudget: " << info.Budget / 1024 / 1024 << "\nCR: " << info.CurrentReservation / 1024 / 1024 - << "\nCU: " << info.CurrentUsage / 1024 / 1024 << LL_ENDL; + << "\nCU: " << info.CurrentUsage / 1024 / 1024 + << "\nEU: " << eu_mb << llformat(" (%.2f)", eu_error) + << "\nTU: " << target_mb + << "\nAM: " << mAvailableVRAM << LL_ENDL; - mDXGIAdapter->QueryVideoMemoryInfo(0, DXGI_MEMORY_SEGMENT_GROUP_NON_LOCAL, &info); + /*mDXGIAdapter->QueryVideoMemoryInfo(0, DXGI_MEMORY_SEGMENT_GROUP_NON_LOCAL, &info); LL_INFOS("Window") << "\nNon-Local\nAFR: " << info.AvailableForReservation / 1024 / 1024 << "\nBudget: " << info.Budget / 1024 / 1024 << "\nCR: " << info.CurrentReservation / 1024 / 1024 - << "\nCU: " << info.CurrentUsage / 1024 / 1024 << LL_ENDL; + << "\nCU: " << info.CurrentUsage / 1024 / 1024 << LL_ENDL;*/ } else if (mD3DDevice != NULL) { // fallback to D3D9 mAvailableVRAM = mD3DDevice->GetAvailableTextureMem() / 1024 / 1024; } - } void LLWindowWin32::LLWindowWin32Thread::run() -- cgit v1.2.3