diff options
Diffstat (limited to 'indra/llrender')
| -rw-r--r-- | indra/llrender/llimagegl.cpp | 56 | ||||
| -rw-r--r-- | indra/llrender/llimagegl.h | 32 | ||||
| -rw-r--r-- | indra/llrender/llrender.cpp | 7 |
3 files changed, 82 insertions, 13 deletions
diff --git a/indra/llrender/llimagegl.cpp b/indra/llrender/llimagegl.cpp index 6bcc34938c..95575c009b 100644 --- a/indra/llrender/llimagegl.cpp +++ b/indra/llrender/llimagegl.cpp @@ -166,6 +166,8 @@ U64 LLImageGL::getTextureBytesAllocated() //statics U32 LLImageGL::sUniqueCount = 0; +std::atomic<U32> LLImageGL::sOOMErrorCount(0); +thread_local bool LLImageGL::sStampBindFrame = true; U32 LLImageGL::sBindCount = 0; S32 LLImageGL::sCount = 0; @@ -777,6 +779,7 @@ void LLImageGL::setImage(const LLImageRaw* imageraw) bool LLImageGL::setImage(const U8* data_in, bool data_hasmips /* = false */, S32 usename /* = 0 */) { LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE; + LLImageGLStampBypass no_stamp; // upload binds are not visibility const bool is_compressed = isCompressed(); @@ -1502,6 +1505,12 @@ void LLImageGL::setManualImage(U32 target, S32 miplevel, S32 intformat, S32 widt { free_cur_tex_image(); } + + // Drain stale GL errors so an OOM detected below belongs to this alloc. + // Otherwise a failed glTexImage2D is swallowed in release while + // alloc_tex_image still counts the bytes, inflating the used-VRAM figure. + while (glGetError() != GL_NO_ERROR) {} + const bool use_sub_image = should_stagger_image_set(compress); if (!use_sub_image) { @@ -1511,19 +1520,30 @@ void LLImageGL::setManualImage(U32 target, S32 miplevel, S32 intformat, S32 widt else { // break up calls to a manageable size for the GL command buffer - { - LL_PROFILE_ZONE_NAMED("glTexImage2D alloc"); - glTexImage2D(target, miplevel, intformat, width, height, 0, pixformat, pixtype, nullptr); - } + LL_PROFILE_ZONE_NAMED("glTexImage2D alloc"); + glTexImage2D(target, miplevel, intformat, width, height, 0, pixformat, pixtype, nullptr); + } - U8* src = (U8*)(pixels); - if (src) + if (glGetError() == GL_OUT_OF_MEMORY) + { + ++sOOMErrorCount; + LL_WARNS_ONCE("Texture") << "glTexImage2D failed with GL_OUT_OF_MEMORY (" + << width << "x" << height << " mip " << miplevel + << ") - not counting bytes" << LL_ENDL; + } + else + { + if (use_sub_image) { - LL_PROFILE_ZONE_NAMED("glTexImage2D copy"); - sub_image_lines(target, miplevel, 0, 0, width, height, pixformat, pixtype, src, width); + U8* src = (U8*)(pixels); + if (src) + { + LL_PROFILE_ZONE_NAMED("glTexImage2D copy"); + sub_image_lines(target, miplevel, 0, 0, width, height, pixformat, pixtype, src, width); + } } + alloc_tex_image(width, height, intformat, 1); } - alloc_tex_image(width, height, intformat, 1); } stop_glerror(); } @@ -1668,6 +1688,7 @@ bool LLImageGL::createGLTexture(S32 discard_level, const U8* data_in, bool data_ LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE; LL_PROFILE_GPU_ZONE("createGLTexture"); checkActiveThread(); + LLImageGLStampBypass no_stamp; // creation binds are not visibility bool main_thread = on_main_thread(); @@ -2090,12 +2111,21 @@ S32 LLImageGL::dimDerivedMaxDiscard(S32 width, S32 height) void LLImageGL::stampBound() const { - // Skip the store on same-frame re-binds - bindFast is per-draw and - // would dirty this cache line per bind per texture otherwise. + // Both stamps skip same-frame re-binds (bindFast runs per draw). They dedupe + // separately, so a non-camera pass touching the time stamp first doesn't stop + // a real camera bind from setting the frame stamp later the same frame. if (mLastBindTime != sLastFrameTime) { mLastBindTime = sLastFrameTime; } + if (sStampBindFrame) + { + const U32 frame = LLFrameTimer::getFrameCount(); + if (mLastBindFrame != frame) + { + mLastBindFrame = frame; + } + } } S64 LLImageGL::getBytes(S32 discard_level) const @@ -2520,6 +2550,10 @@ bool LLImageGL::scaleDown(S32 desired_discard) { LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE; + // Don't let eviction re-arm visibility: the glGenerateMipmap re-bind below + // would otherwise stamp mLastBindFrame and keep the texture fetch-eligible. + LLImageGLStampBypass no_stamp; + if (mTarget != GL_TEXTURE_2D || mFormatInternal == -1 // not initialized ) diff --git a/indra/llrender/llimagegl.h b/indra/llrender/llimagegl.h index 0869ae54fe..57ca79b3dd 100644 --- a/indra/llrender/llimagegl.h +++ b/indra/llrender/llimagegl.h @@ -39,6 +39,7 @@ #include "llrender.h" #include "threadpool.h" #include "workqueue.h" +#include <atomic> #include <unordered_set> #define LL_IMAGEGL_THREAD_CHECK 0 //set to 1 to enable thread debugging for ImageGL @@ -238,8 +239,17 @@ public: public: // Various GL/Rendering options S64Bytes mTextureMemory; - mutable F32 mLastBindTime = 0.f; // wall-clock time at last stampBound; drives the streaming cooldown - F32 mGLCreateTime = 0.f; // wall-clock time the GL texture was created; cooldown fallback for never-bound textures + mutable F32 mLastBindTime = 0.f; // wall-clock time at last stampBound (bind or bind-attempt) + mutable U32 mLastBindFrame = 0; // frame index (LLFrameTimer::getFrameCount) at last CAMERA-pass + // stampBound; 0 = never. Drives visibility GC + fetch gating. + F32 mGLCreateTime = 0.f; // wall-clock time the GL texture was created + + // When false, stampBound skips the mLastBindFrame stamp (mLastBindTime still + // updates). Set false around non-camera passes (probes, shadows, impostors) + // and administrative binds (upload, scaleDown - via LLImageGLStampBypass) so + // those binds don't count as camera visibility. Thread-local so a GL upload + // thread can't flip it on the render thread mid-frame. + static thread_local bool sStampBindFrame; private: U32 createPickMask(S32 pWidth, S32 pHeight); @@ -300,6 +310,11 @@ public: // Global memory statistics static U32 sBindCount; // Tracks number of texture binds for current frame static U32 sUniqueCount; // Tracks number of unique texture binds for current frame + // glTexImage2D GL_OUT_OF_MEMORY failures detected (bytes NOT counted for + // these). Written from whichever thread runs texture creation; read by + // the streaming 1Hz pressure log. Nonzero = the driver is refusing + // allocations and the VRAM budget is unreliable. + static std::atomic<U32> sOOMErrorCount; static bool sGlobalUseAnisotropic; static LLImageGL* sDefaultGLTexture ; static bool sAutomatedTest; @@ -355,6 +370,19 @@ public: }; +// RAII: suppress the mLastBindFrame stamp for the current scope. Use around +// administrative binds (upload, create, scaleDown) so they don't count as +// camera visibility - otherwise the GC's own scaleDown re-stamps what it just +// aged out and oscillates. Saves/restores, so it nests correctly. +class LLImageGLStampBypass +{ +public: + LLImageGLStampBypass() : mPrev(LLImageGL::sStampBindFrame) { LLImageGL::sStampBindFrame = false; } + ~LLImageGLStampBypass() { LLImageGL::sStampBindFrame = mPrev; } +private: + bool mPrev; +}; + class LLImageGLThread : public LLSimpleton<LLImageGLThread>, LL::ThreadPool { public: diff --git a/indra/llrender/llrender.cpp b/indra/llrender/llrender.cpp index 5e845fbcce..f0a1c44507 100644 --- a/indra/llrender/llrender.cpp +++ b/indra/llrender/llrender.cpp @@ -245,6 +245,11 @@ bool LLTexUnit::bind(LLTexture* texture, bool for_rendering, bool forceBind) texture->setActive() ; texture->updateBindStatsForTester() ; } + // updateBindStats only stamps time; the GC and fetch gate use + // the frame stamp, so stamp it here too or bind()-drawn faces + // (bump/material/media) oscillate. Admin/non-camera binds are + // already suppressed via LLImageGLStampBypass / sStampBindFrame. + gl_tex->stampBound(); mHasMipMaps = gl_tex->mHasMipMaps; if (gl_tex->mTexOptionsDirty) { @@ -325,6 +330,8 @@ bool LLTexUnit::bind(LLImageGL* texture, bool for_rendering, bool forceBind, S32 glBindTexture(sGLTextureType[texture->getTarget()], mCurrTexture); stop_glerror(); texture->updateBindStats(); + // Frame-stamp fresh binds too - see bind(LLTexture*) above. + texture->stampBound(); mHasMipMaps = texture->mHasMipMaps; if (texture->mTexOptionsDirty) { |