diff options
Diffstat (limited to 'indra/llrender/llimagegl.cpp')
| -rw-r--r-- | indra/llrender/llimagegl.cpp | 274 |
1 files changed, 144 insertions, 130 deletions
diff --git a/indra/llrender/llimagegl.cpp b/indra/llrender/llimagegl.cpp index 9dc140b5b9..b3b79bd6c4 100644 --- a/indra/llrender/llimagegl.cpp +++ b/indra/llrender/llimagegl.cpp @@ -53,13 +53,75 @@ const F32 MIN_TEXTURE_LIFETIME = 10.f; //assumes i is a power of 2 > 0 U32 wpo2(U32 i); + +// texture memory accounting (for OS X) +static LLMutex sTexMemMutex; +static std::unordered_map<U32, U32> sTextureAllocs; +static U64 sTextureBytes = 0; + +// track a texture alloc on the currently bound texture. +// asserts that no currently tracked alloc exists +static void alloc_tex_image(U32 width, U32 height, U32 pixformat) +{ + U32 texUnit = gGL.getCurrentTexUnitIndex(); + U32 texName = gGL.getTexUnit(texUnit)->getCurrTexture(); + S32 size = LLImageGL::dataFormatBytes(pixformat, width, height); + + llassert(size >= 0); + + sTexMemMutex.lock(); + llassert(sTextureAllocs.find(texName) == sTextureAllocs.end()); + + sTextureAllocs[texName] = size; + sTextureBytes += size; + + sTexMemMutex.unlock(); +} + +// track texture free on given texName +static void free_tex_image(U32 texName) +{ + sTexMemMutex.lock(); + auto iter = sTextureAllocs.find(texName); + if (iter != sTextureAllocs.end()) + { + llassert(iter->second <= sTextureBytes); // sTextureBytes MUST NOT go below zero + + sTextureBytes -= iter->second; + + sTextureAllocs.erase(iter); + } + + sTexMemMutex.unlock(); +} + +// track texture free on given texNames +static void free_tex_images(U32 count, const U32* texNames) +{ + for (int i = 0; i < count; ++i) + { + free_tex_image(texNames[i]); + } +} + +// track texture free on currently bound texture +static void free_cur_tex_image() +{ + U32 texUnit = gGL.getCurrentTexUnitIndex(); + U32 texName = gGL.getTexUnit(texUnit)->getCurrTexture(); + free_tex_image(texName); +} + +// static +U64 LLImageGL::getTextureBytesAllocated() +{ + return sTextureBytes; +} + //statics U32 LLImageGL::sUniqueCount = 0; U32 LLImageGL::sBindCount = 0; -S32Bytes LLImageGL::sGlobalTextureMemory(0); -S32Bytes LLImageGL::sBoundTextureMemory(0); -S32Bytes LLImageGL::sCurBoundTextureMemory(0); S32 LLImageGL::sCount = 0; BOOL LLImageGL::sGlobalUseAnisotropic = FALSE; @@ -220,6 +282,7 @@ S32 LLImageGL::dataFormatBits(S32 dataformat) case GL_RGBA: return 32; case GL_SRGB_ALPHA: return 32; case GL_BGRA: return 32; // Used for QuickTime media textures on the Mac + case GL_DEPTH_COMPONENT: return 24; default: LL_ERRS() << "LLImageGL::Unknown format: " << dataformat << LL_ENDL; return 0; @@ -282,15 +345,6 @@ void LLImageGL::updateStats(F32 current_time) { LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE; sLastFrameTime = current_time; - sBoundTextureMemory = sCurBoundTextureMemory; - sCurBoundTextureMemory = S32Bytes(0); -} - -//static -S32 LLImageGL::updateBoundTexMem(const S32Bytes mem, const S32 ncomponents, S32 category) -{ - LLImageGL::sCurBoundTextureMemory += mem ; - return LLImageGL::sCurBoundTextureMemory.value(); } //---------------------------------------------------------------------------- @@ -298,7 +352,7 @@ S32 LLImageGL::updateBoundTexMem(const S32Bytes mem, const S32 ncomponents, S32 //static void LLImageGL::destroyGL(BOOL save_state) { - for (S32 stage = 0; stage < gGLManager.mNumTextureUnits; stage++) + for (S32 stage = 0; stage < gGLManager.mNumTextureImageUnits; stage++) { gGL.getTexUnit(stage)->unbind(LLTexUnit::TT_TEXTURE); } @@ -466,7 +520,6 @@ void LLImageGL::init(BOOL usemipmaps) mPickMaskHeight = 0; mUseMipMaps = usemipmaps; mHasExplicitFormat = FALSE; - mAutoGenMips = FALSE; mIsMask = FALSE; mNeedsAlphaAndPickMask = TRUE ; @@ -623,7 +676,7 @@ void LLImageGL::forceUpdateBindStats(void) const mLastBindTime = sLastFrameTime; } -BOOL LLImageGL::updateBindStats(S32Bytes tex_mem) const +BOOL LLImageGL::updateBindStats() const { if (mTexName != 0) { @@ -635,7 +688,6 @@ BOOL LLImageGL::updateBindStats(S32Bytes tex_mem) const { // we haven't accounted for this texture yet this frame sUniqueCount++; - updateBoundTexMem(tex_mem, mComponents, mCategory); mLastBindTime = sLastFrameTime; return TRUE ; @@ -741,7 +793,7 @@ BOOL LLImageGL::setImage(const U8* data_in, BOOL data_hasmips /* = FALSE */, S32 if (is_compressed) { S32 tex_size = dataFormatBytes(mFormatPrimary, w, h); - glCompressedTexImage2DARB(mTarget, gl_level, mFormatPrimary, w, h, 0, tex_size, (GLvoid *)data_in); + glCompressedTexImage2D(mTarget, gl_level, mFormatPrimary, w, h, 0, tex_size, (GLvoid *)data_in); stop_glerror(); } else @@ -812,6 +864,7 @@ BOOL LLImageGL::setImage(const U8* data_in, BOOL data_hasmips /* = FALSE */, S32 if (LLRender::sGLCoreProfile) { + LL_PROFILE_GPU_ZONE("generate mip map"); glGenerateMipmap(mTarget); } stop_glerror(); @@ -936,7 +989,7 @@ BOOL LLImageGL::setImage(const U8* data_in, BOOL data_hasmips /* = FALSE */, S32 if (is_compressed) { S32 tex_size = dataFormatBytes(mFormatPrimary, w, h); - glCompressedTexImage2DARB(mTarget, 0, mFormatPrimary, w, h, 0, tex_size, (GLvoid *)data_in); + glCompressedTexImage2D(mTarget, 0, mFormatPrimary, w, h, 0, tex_size, (GLvoid *)data_in); stop_glerror(); } else @@ -1015,30 +1068,12 @@ BOOL LLImageGL::preAddToAtlas(S32 discard_level, const LLImageRaw* raw_image) mFormatType = GL_UNSIGNED_BYTE; break; case 3: -#if USE_SRGB_DECODE - if (gGLManager.mHasTexturesRGBDecode) - { - mFormatInternal = GL_SRGB8; - } - else -#endif - { - mFormatInternal = GL_RGB8; - } + mFormatInternal = GL_RGB8; mFormatPrimary = GL_RGB; mFormatType = GL_UNSIGNED_BYTE; break; case 4: -#if USE_SRGB_DECODE - if (gGLManager.mHasTexturesRGBDecode) - { - mFormatInternal = GL_SRGB8_ALPHA8; - } - else -#endif - { - mFormatInternal = GL_RGBA8; - } + mFormatInternal = GL_RGBA8; mFormatPrimary = GL_RGBA; mFormatType = GL_UNSIGNED_BYTE; break; @@ -1231,6 +1266,7 @@ void LLImageGL::deleteTextures(S32 numTextures, const U32 *textures) { if (gGLManager.mInited) { + free_tex_images(numTextures, textures); glDeleteTextures(numTextures, textures); } } @@ -1368,7 +1404,55 @@ void LLImageGL::setManualImage(U32 target, S32 miplevel, S32 intformat, S32 widt stop_glerror(); { LL_PROFILE_ZONE_NAMED("glTexImage2D"); + LL_PROFILE_ZONE_NUM(width); + LL_PROFILE_ZONE_NUM(height); + + free_cur_tex_image(); +#if 0 glTexImage2D(target, miplevel, intformat, width, height, 0, pixformat, pixtype, use_scratch ? scratch : pixels); +#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); + } + + U8* src = (U8*)(use_scratch ? scratch : pixels); + if (src) + { + LL_PROFILE_ZONE_NAMED("glTexImage2D copy"); + U32 components = dataFormatComponents(pixformat); + U32 type_width = 0; + + switch (pixtype) + { + case GL_UNSIGNED_BYTE: + case GL_BYTE: + case GL_UNSIGNED_INT_8_8_8_8_REV: + type_width = 1; + break; + case GL_UNSIGNED_SHORT: + case GL_SHORT: + type_width = 2; + break; + case GL_UNSIGNED_INT: + case GL_INT: + case GL_FLOAT: + type_width = 4; + break; + default: + LL_ERRS() << "Unknown type: " << pixtype << LL_ENDL; + } + + U32 line_width = width * components * type_width; + for (U32 y = 0; y < height; ++y) + { + glTexSubImage2D(target, miplevel, 0, y, width, 1, pixformat, pixtype, src); + src += line_width; + } + } +#endif + alloc_tex_image(width, height, pixformat); } stop_glerror(); @@ -1482,30 +1566,12 @@ BOOL LLImageGL::createGLTexture(S32 discard_level, const LLImageRaw* imageraw, S mFormatType = GL_UNSIGNED_BYTE; break; case 3: - #if USE_SRGB_DECODE - if (gGLManager.mHasTexturesRGBDecode) - { - mFormatInternal = GL_SRGB8; - } - else - #endif - { - mFormatInternal = GL_RGB8; - } + mFormatInternal = GL_RGB8; mFormatPrimary = GL_RGB; mFormatType = GL_UNSIGNED_BYTE; break; case 4: - #if USE_SRGB_DECODE - if (gGLManager.mHasTexturesRGBDecode) - { - mFormatInternal = GL_SRGB8_ALPHA8; - } - else - #endif - { - mFormatInternal = GL_RGBA8; - } + mFormatInternal = GL_RGBA8; mFormatPrimary = GL_RGBA; mFormatType = GL_UNSIGNED_BYTE; break; @@ -1534,6 +1600,7 @@ BOOL LLImageGL::createGLTexture(S32 discard_level, const U8* data_in, BOOL data_ // Call with void data, vmem is allocated but unitialized { LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE; + LL_PROFILE_GPU_ZONE("createGLTexture"); checkActiveThread(); bool main_thread = on_main_thread(); @@ -1593,7 +1660,7 @@ BOOL LLImageGL::createGLTexture(S32 discard_level, const U8* data_in, BOOL data_ if (mUseMipMaps) { - mAutoGenMips = gGLManager.mHasMipMapGeneration; + mAutoGenMips = true; } mCurrentDiscardLevel = discard_level; @@ -1614,11 +1681,6 @@ BOOL LLImageGL::createGLTexture(S32 discard_level, const U8* data_in, BOOL data_ // things will break if we don't unbind after creation gGL.getTexUnit(0)->unbind(mBindTarget); - if (old_texname != 0) - { - sGlobalTextureMemory -= mTextureMemory; - } - //if we're on the image loading thread, be sure to delete old_texname and update mTexName on the main thread if (!defer_copy) { @@ -1639,7 +1701,6 @@ BOOL LLImageGL::createGLTexture(S32 discard_level, const U8* data_in, BOOL data_ mTextureMemory = (S32Bytes)getMipBytes(mCurrentDiscardLevel); - sGlobalTextureMemory += mTextureMemory; mTexelsInGLTexture = getWidth() * getHeight(); // mark this as bound at this point, so we don't throw it out immediately @@ -1649,51 +1710,6 @@ BOOL LLImageGL::createGLTexture(S32 discard_level, const U8* data_in, BOOL data_ return TRUE; } -void LLImageGLThread::updateClass() -{ - LL_PROFILE_ZONE_SCOPED; - - // update available vram one per second - static LLFrameTimer sTimer; - - if (sTimer.getElapsedSeconds() < 1.f) - { - return; - } - - sTimer.reset(); - - auto func = []() - { - if (gGLManager.mHasATIMemInfo) - { - S32 meminfo[4]; - glGetIntegerv(GL_TEXTURE_FREE_MEMORY_ATI, meminfo); - LLImageGLThread::sFreeVRAMMegabytes = meminfo[0]; - - } - else if (gGLManager.mHasNVXMemInfo) - { - S32 free_memory; - glGetIntegerv(GL_GPU_MEMORY_INFO_CURRENT_AVAILABLE_VIDMEM_NVX, &free_memory); - LLImageGLThread::sFreeVRAMMegabytes = free_memory / 1024; - } - }; - - - // post update to background thread if available, otherwise execute immediately - auto queue = LL::WorkQueue::getInstance("LLImageGL"); - if (sEnabled) - { - queue->post(func); - } - else - { - llassert(queue == nullptr); - func(); - } -} - void LLImageGL::syncToMainThread(LLGLuint new_tex_name) { LL_PROFILE_ZONE_SCOPED; @@ -1701,7 +1717,16 @@ void LLImageGL::syncToMainThread(LLGLuint new_tex_name) { LL_PROFILE_ZONE_NAMED("cglt - sync"); - if (gGLManager.mHasSync) + if (gGLManager.mIsNVIDIA) + { + // wait for texture upload to finish before notifying main thread + // upload is complete + auto sync = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0); + glFlush(); + glClientWaitSync(sync, 0, GL_TIMEOUT_IGNORED); + glDeleteSync(sync); + } + else { // post a sync to the main thread (will execute before tex name swap lambda below) // glFlush calls here are partly superstitious and partly backed by observation @@ -1724,10 +1749,6 @@ void LLImageGL::syncToMainThread(LLGLuint new_tex_name) } }); } - else - { - glFinish(); - } } ref(); @@ -1739,8 +1760,11 @@ void LLImageGL::syncToMainThread(LLGLuint new_tex_name) syncTexName(new_tex_name); unref(); }); + + LL_PROFILER_GPU_COLLECT; } + void LLImageGL::syncTexName(LLGLuint texname) { if (texname != 0) @@ -1830,7 +1854,7 @@ BOOL LLImageGL::readBackRaw(S32 discard_level, LLImageRaw* imageraw, bool compre return FALSE ; } - glGetCompressedTexImageARB(mTarget, gl_discard, (GLvoid*)(imageraw->getData())); + glGetCompressedTexImage(mTarget, gl_discard, (GLvoid*)(imageraw->getData())); //stop_glerror(); } else @@ -1872,7 +1896,6 @@ void LLImageGL::destroyGLTexture() { if(mTextureMemory != S32Bytes(0)) { - sGlobalTextureMemory -= mTextureMemory; mTextureMemory = (S32Bytes)0; } @@ -2424,13 +2447,9 @@ void LLImageGL::checkActiveThread() glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, nummips); */ -std::atomic<S32> LLImageGLThread::sFreeVRAMMegabytes(4096); //if free vram is unknown, default to 4GB - LLImageGLThread::LLImageGLThread(LLWindow* window) - // We want exactly one thread, but a very large capacity: we never want - // anyone, especially inner-loop render code, to have to block on post() - // because we're full. - : ThreadPool("LLImageGL", 1, 1024*1024) + // We want exactly one thread. + : LL::ThreadPool("LLImageGL", 1) , mWindow(window) { LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE; @@ -2438,7 +2457,7 @@ LLImageGLThread::LLImageGLThread(LLWindow* window) mFinished = false; mContext = mWindow->createSharedContext(); - ThreadPool::start(); + LL::ThreadPool::start(); } void LLImageGLThread::run() @@ -2448,13 +2467,8 @@ void LLImageGLThread::run() // WorkQueue, likewise cleanup afterwards. mWindow->makeContextCurrent(mContext); gGL.init(false); - ThreadPool::run(); + LL::ThreadPool::run(); gGL.shutdown(); mWindow->destroySharedContext(mContext); } -S32 LLImageGLThread::getFreeVRAMMegabytes() -{ - return sFreeVRAMMegabytes; -} - |