diff options
author | Dave Parks <davep@lindenlab.com> | 2022-08-26 10:51:42 -0500 |
---|---|---|
committer | Dave Parks <davep@lindenlab.com> | 2022-08-26 10:51:42 -0500 |
commit | 9bee2a92d2eb1f48f8ca0061e546c5e9df38fe79 (patch) | |
tree | e05c4291ba21b26b7cb4d4d0e46423a17663385e /indra | |
parent | c58fd13c63048d96bd0cbf7a2ac4aa286e3e3b96 (diff) |
SL-17997 Follow up from beta breakers results.
Diffstat (limited to 'indra')
-rw-r--r-- | indra/llrender/llimagegl.cpp | 17 | ||||
-rw-r--r-- | indra/llrender/llimagegl.h | 5 | ||||
-rw-r--r-- | indra/llwindow/llwindowwin32.cpp | 55 |
3 files changed, 44 insertions, 33 deletions
diff --git a/indra/llrender/llimagegl.cpp b/indra/llrender/llimagegl.cpp index ae10142e7a..7ba0cc2fde 100644 --- a/indra/llrender/llimagegl.cpp +++ b/indra/llrender/llimagegl.cpp @@ -54,7 +54,6 @@ const F32 MIN_TEXTURE_LIFETIME = 10.f; U32 wpo2(U32 i); -#if LL_DARWIN // texture memory accounting (for OS X) static LLMutex sTexMemMutex; static std::unordered_map<U32, U32> sTextureAllocs; @@ -119,22 +118,6 @@ U64 LLImageGL::getTextureBytesAllocated() return sTextureBytes; } -#else - -#define alloc_tex_image(width, height, pixformat) (void) width; (void) height; (void) pixformat; -#define free_tex_image(texName) (void) texName; -#define free_tex_images(count, texNames) (void) count; (void) texNames; -#define free_cur_tex_image() - -// static -U64 LLImageGL::getTextureBytesAllocated() -{ - // UNIMPLEMENTED OUTSIDE OF OS X, DO NOT CALL - llassert(false); - return 0; -} -#endif - //statics U32 LLImageGL::sUniqueCount = 0; diff --git a/indra/llrender/llimagegl.h b/indra/llrender/llimagegl.h index 619421e68a..1bcdd040e0 100644 --- a/indra/llrender/llimagegl.h +++ b/indra/llrender/llimagegl.h @@ -53,7 +53,10 @@ class LLImageGL : public LLRefCount friend class LLTexUnit; public: - // For OS X use only -- get an estimate of how many bytes have been allocated in vram for textures + // Get an estimate of how many bytes have been allocated in vram for textures. + // Does not include mipmaps. + // NOTE: multiplying this number by two gives a good estimate for total + // video memory usage based on testing in lagland against an NVIDIA GPU. static U64 getTextureBytesAllocated(); // These 2 functions replace glGenTextures() and glDeleteTextures() 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<U32> 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() |