1836 dont store texture in system memory unless absolutely necessary (#1843)

* #1836 Texture memory usage overhaul.  Much decrufting
- don't keep a copy of textures in system memory
- use GPU to downrez textures instead of reloading from cache
- use GPU to generate brightness/darkness bumpmaps

4 files changed, 109 insertions, 213 deletions

diff --git a/indra/llrender/llgltexture.cpp b/indra/llrender/llgltexture.cpp
index e614f45986..4dcca5a726 100644
--- a/indra/llrender/llgltexture.cpp
+++ b/indra/llrender/llgltexture.cpp
@@ -351,20 +351,6 @@ void LLGLTexture::forceUpdateBindStats(void) const
     return mGLTexturep->forceUpdateBindStats() ;
 }
 
-U32 LLGLTexture::getTexelsInAtlas() const
-{
-    llassert(mGLTexturep.notNull()) ;
-
-    return mGLTexturep->getTexelsInAtlas() ;
-}
-
-U32 LLGLTexture::getTexelsInGLTexture() const
-{
-    llassert(mGLTexturep.notNull()) ;
-
-    return mGLTexturep->getTexelsInGLTexture() ;
-}
-
 bool LLGLTexture::isGLTextureCreated() const
 {
     llassert(mGLTexturep.notNull()) ;
@@ -372,13 +358,6 @@ bool LLGLTexture::isGLTextureCreated() const
     return mGLTexturep->isGLTextureCreated() ;
 }
 
-S32  LLGLTexture::getDiscardLevelInAtlas() const
-{
-    llassert(mGLTexturep.notNull()) ;
-
-    return mGLTexturep->getDiscardLevelInAtlas() ;
-}
-
 void LLGLTexture::destroyGLTexture()
 {
     if(mGLTexturep.notNull() && mGLTexturep->getHasGLTexture())
diff --git a/indra/llrender/llgltexture.h b/indra/llrender/llgltexture.h
index 0901243f8f..a7de20dc5c 100644
--- a/indra/llrender/llgltexture.h
+++ b/indra/llrender/llgltexture.h
@@ -51,10 +51,10 @@ public:
         BOOST_NONE          = 0,
         BOOST_AVATAR        ,
         BOOST_AVATAR_BAKED  ,
-        BOOST_SCULPTED      ,
         BOOST_TERRAIN       , // Needed for minimap generation for now. Lower than BOOST_HIGH so the texture stats don't get forced, i.e. texture stats are manually managed by minimap/terrain instead.
 
         BOOST_HIGH          = 10,
+        BOOST_SCULPTED      ,
         BOOST_BUMP          ,
         BOOST_UNUSED_1      , // Placeholder to avoid disrupting habits around texture debug
         BOOST_SELECTED      ,
@@ -75,7 +75,6 @@ public:
         AVATAR_SCRATCH_TEX,
         DYNAMIC_TEX,
         MEDIA,
-        ATLAS,
         OTHER,
         MAX_GL_IMAGE_CATEGORY
     };
@@ -156,10 +155,7 @@ public:
     bool       isJustBound()const ;
     void       forceUpdateBindStats(void) const;
 
-    U32        getTexelsInAtlas() const ;
-    U32        getTexelsInGLTexture() const ;
     bool       isGLTextureCreated() const ;
-    S32        getDiscardLevelInAtlas() const ;
     LLGLTextureState getTextureState() const { return mTextureState; }
 
     //---------------------------------------------------------------------------------------------
diff --git a/indra/llrender/llimagegl.cpp b/indra/llrender/llimagegl.cpp
index 7e5cd628c1..956bcef352 100644
--- a/indra/llrender/llimagegl.cpp
+++ b/indra/llrender/llimagegl.cpp
@@ -41,6 +41,7 @@
 #include "llrender.h"
 #include "llwindow.h"
 #include "llframetimer.h"
+#include <unordered_set>
 
 extern LL_COMMON_API bool on_main_thread();
 
@@ -130,10 +131,9 @@ S32 LLImageGL::sCount                   = 0;
 
 bool LLImageGL::sGlobalUseAnisotropic   = false;
 F32 LLImageGL::sLastFrameTime           = 0.f;
-bool LLImageGL::sAllowReadBackRaw       = false ;
 LLImageGL* LLImageGL::sDefaultGLTexture = NULL ;
 bool LLImageGL::sCompressTextures = false;
-std::set<LLImageGL*> LLImageGL::sImageList;
+std::unordered_set<LLImageGL*> LLImageGL::sImageList;
 
 
 bool LLImageGLThread::sEnabledTextures = false;
@@ -150,6 +150,9 @@ S32 LLImageGL::sMaxCategories = 1 ;
 
 //optimization for when we don't need to calculate mIsMask
 bool LLImageGL::sSkipAnalyzeAlpha;
+U32  LLImageGL::sScratchPBO = 0;
+U32  LLImageGL::sScratchPBOSize = 0;
+
 
 //------------------------
 //****************************************************************************************************
@@ -159,20 +162,6 @@ bool LLImageGL::sSkipAnalyzeAlpha;
 //**************************************************************************************
 //below are functions for debug use
 //do not delete them even though they are not currently being used.
-void check_all_images()
-{
-    for (std::set<LLImageGL*>::iterator iter = LLImageGL::sImageList.begin();
-         iter != LLImageGL::sImageList.end(); iter++)
-    {
-        LLImageGL* glimage = *iter;
-        if (glimage->getTexName() && glimage->isGLTextureCreated())
-        {
-            gGL.getTexUnit(0)->bind(glimage) ;
-            glimage->checkTexSize() ;
-            gGL.getTexUnit(0)->unbind(glimage->getTarget()) ;
-        }
-    }
-}
 
 void LLImageGL::checkTexSize(bool forced) const
 {
@@ -252,6 +241,11 @@ void LLImageGL::initClass(LLWindow* window, S32 num_catagories, bool skip_analyz
     LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE;
     sSkipAnalyzeAlpha = skip_analyze_alpha;
 
+    if (sScratchPBO == 0)
+    {
+        glGenBuffers(1, &sScratchPBO);
+    }
+
     if (thread_texture_loads || thread_media_updates)
     {
         LLImageGLThread::createInstance(window);
@@ -265,6 +259,12 @@ void LLImageGL::cleanupClass()
 {
     LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE;
     LLImageGLThread::deleteSingleton();
+    if (sScratchPBO != 0)
+    {
+        glDeleteBuffers(1, &sScratchPBO);
+        sScratchPBO = 0;
+        sScratchPBOSize = 0;
+    }
 }
 
 
@@ -360,66 +360,19 @@ void LLImageGL::updateStats(F32 current_time)
 //----------------------------------------------------------------------------
 
 //static
-void LLImageGL::destroyGL(bool save_state)
+void LLImageGL::destroyGL()
 {
     for (S32 stage = 0; stage < gGLManager.mNumTextureImageUnits; stage++)
     {
         gGL.getTexUnit(stage)->unbind(LLTexUnit::TT_TEXTURE);
     }
-
-    sAllowReadBackRaw = true ;
-    for (std::set<LLImageGL*>::iterator iter = sImageList.begin();
-         iter != sImageList.end(); iter++)
-    {
-        LLImageGL* glimage = *iter;
-        if (glimage->mTexName)
-        {
-            if (save_state && glimage->isGLTextureCreated() && glimage->mComponents)
-            {
-                glimage->mSaveData = new LLImageRaw;
-                if(!glimage->readBackRaw(glimage->mCurrentDiscardLevel, glimage->mSaveData, false)) //necessary, keep it.
-                {
-                    glimage->mSaveData = NULL ;
-                }
-            }
-
-            glimage->destroyGLTexture();
-            stop_glerror();
-        }
-    }
-    sAllowReadBackRaw = false ;
-}
-
-//static
-void LLImageGL::restoreGL()
-{
-    for (std::set<LLImageGL*>::iterator iter = sImageList.begin();
-         iter != sImageList.end(); iter++)
-    {
-        LLImageGL* glimage = *iter;
-        if(glimage->getTexName())
-        {
-            LL_ERRS() << "tex name is not 0." << LL_ENDL ;
-        }
-        if (glimage->mSaveData.notNull())
-        {
-            if (glimage->getComponents() && glimage->mSaveData->getComponents())
-            {
-                glimage->createGLTexture(glimage->mCurrentDiscardLevel, glimage->mSaveData, 0, true, glimage->getCategory());
-                stop_glerror();
-            }
-            glimage->mSaveData = NULL; // deletes data
-        }
-    }
 }
 
 //static
 void LLImageGL::dirtyTexOptions()
 {
-    for (std::set<LLImageGL*>::iterator iter = sImageList.begin();
-         iter != sImageList.end(); iter++)
+    for (auto& glimage : sImageList)
     {
-        LLImageGL* glimage = *iter;
         glimage->mTexOptionsDirty = true;
         stop_glerror();
     }
@@ -542,10 +495,6 @@ void LLImageGL::init(bool usemipmaps)
     mHeight = 0;
     mCurrentDiscardLevel = -1;
 
-    mDiscardLevelInAtlas = -1 ;
-    mTexelsInAtlas = 0 ;
-    mTexelsInGLTexture = 0 ;
-
     mAllowCompression = true;
 
     mTarget = GL_TEXTURE_2D;
@@ -622,9 +571,6 @@ bool LLImageGL::setSize(S32 width, S32 height, S32 ncomponents, S32 discard_leve
             return false;
         }
 
-        // pickmask validity depends on old image size, delete it
-        freePickMask();
-
         mWidth = width;
         mHeight = height;
         mComponents = ncomponents;
@@ -1025,98 +971,6 @@ bool LLImageGL::setImage(const U8* data_in, bool data_hasmips /* = false */, S32
     return true;
 }
 
-bool LLImageGL::preAddToAtlas(S32 discard_level, const LLImageRaw* raw_image)
-{
-    //not compatible with core GL profile
-    llassert(!LLRender::sGLCoreProfile);
-
-    if (gGLManager.mIsDisabled)
-    {
-        LL_WARNS() << "Trying to create a texture while GL is disabled!" << LL_ENDL;
-        return false;
-    }
-    llassert(gGLManager.mInited);
-    stop_glerror();
-
-    if (discard_level < 0)
-    {
-        llassert(mCurrentDiscardLevel >= 0);
-        discard_level = mCurrentDiscardLevel;
-    }
-
-    // Actual image width/height = raw image width/height * 2^discard_level
-    S32 w = raw_image->getWidth() << discard_level;
-    S32 h = raw_image->getHeight() << discard_level;
-
-    // setSize may call destroyGLTexture if the size does not match
-    if (!setSize(w, h, raw_image->getComponents(), discard_level))
-    {
-        LL_WARNS() << "Trying to create a texture with incorrect dimensions!" << LL_ENDL;
-        return false;
-    }
-
-    if (!mHasExplicitFormat)
-    {
-        switch (mComponents)
-        {
-            case 1:
-                // Use luminance alpha (for fonts)
-                mFormatInternal = GL_LUMINANCE8;
-                mFormatPrimary  = GL_LUMINANCE;
-                mFormatType     = GL_UNSIGNED_BYTE;
-                break;
-            case 2:
-                // Use luminance alpha (for fonts)
-                mFormatInternal = GL_LUMINANCE8_ALPHA8;
-                mFormatPrimary  = GL_LUMINANCE_ALPHA;
-                mFormatType     = GL_UNSIGNED_BYTE;
-                break;
-            case 3:
-                mFormatInternal = GL_RGB8;
-                mFormatPrimary = GL_RGB;
-                mFormatType    = GL_UNSIGNED_BYTE;
-                break;
-            case 4:
-                mFormatInternal = GL_RGBA8;
-                mFormatPrimary = GL_RGBA;
-                mFormatType    = GL_UNSIGNED_BYTE;
-                break;
-            default:
-                LL_ERRS() << "Bad number of components for texture: " << (U32) getComponents() << LL_ENDL;
-        }
-    }
-
-    mCurrentDiscardLevel = discard_level;
-    mDiscardLevelInAtlas = discard_level;
-    mTexelsInAtlas = raw_image->getWidth() * raw_image->getHeight() ;
-    mLastBindTime = sLastFrameTime;
-    mGLTextureCreated = false ;
-
-    glPixelStorei(GL_UNPACK_ROW_LENGTH, raw_image->getWidth());
-    stop_glerror();
-
-    if(mFormatSwapBytes)
-    {
-        glPixelStorei(GL_UNPACK_SWAP_BYTES, 1);
-        stop_glerror();
-    }
-
-    return true ;
-}
-
-void LLImageGL::postAddToAtlas()
-{
-    if(mFormatSwapBytes)
-    {
-        glPixelStorei(GL_UNPACK_SWAP_BYTES, 0);
-        stop_glerror();
-    }
-
-    glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
-    gGL.getTexUnit(0)->setTextureFilteringOption(mFilterOption);
-    stop_glerror();
-}
-
 U32 type_width_from_pixtype(U32 pixtype)
 {
     U32 type_width = 0;
@@ -1752,7 +1606,6 @@ bool LLImageGL::createGLTexture(S32 discard_level, const U8* data_in, bool data_
 
 
     mTextureMemory = (S64Bytes)getMipBytes(mCurrentDiscardLevel);
-    mTexelsInGLTexture = getWidth() * getHeight();
 
     // mark this as bound at this point, so we don't throw it out immediately
     mLastBindTime = sLastFrameTime;
@@ -1830,8 +1683,7 @@ void LLImageGL::syncTexName(LLGLuint texname)
 
 bool LLImageGL::readBackRaw(S32 discard_level, LLImageRaw* imageraw, bool compressed_ok) const
 {
-    llassert_always(sAllowReadBackRaw) ;
-    //LL_ERRS() << "should not call this function!" << LL_ENDL ;
+    LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE;
 
     if (discard_level < 0)
     {
@@ -2297,6 +2149,8 @@ void LLImageGL::analyzeAlpha(const void* data_in, U32 w, U32 h)
 //----------------------------------------------------------------------------
 U32 LLImageGL::createPickMask(S32 pWidth, S32 pHeight)
 {
+    LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE;
+    freePickMask();
     U32 pick_width = pWidth/2 + 1;
     U32 pick_height = pHeight/2 + 1;
 
@@ -2314,7 +2168,6 @@ U32 LLImageGL::createPickMask(S32 pWidth, S32 pHeight)
 //----------------------------------------------------------------------------
 void LLImageGL::freePickMask()
 {
-    // pickmask validity depends on old image size, delete it
     if (mPickMask != NULL)
     {
         delete [] mPickMask;
@@ -2352,16 +2205,16 @@ void LLImageGL::updatePickMask(S32 width, S32 height, const U8* data_in)
         return ;
     }
 
-    freePickMask();
-
     if (mFormatType != GL_UNSIGNED_BYTE ||
         ((mFormatPrimary != GL_RGBA)
       && (mFormatPrimary != GL_SRGB_ALPHA)))
     {
         //cannot generate a pick mask for this texture
+        freePickMask();
         return;
     }
 
+
 #ifdef SHOW_ASSERT
     const U32 pickSize = createPickMask(width, height);
 #else // SHOW_ASSERT
@@ -2460,6 +2313,73 @@ void LLImageGL::resetCurTexSizebar()
     sCurTexSizeBar = -1 ;
     sCurTexPickSize = -1 ;
 }
+
+bool LLImageGL::scaleDown(S32 desired_discard)
+{
+    LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE;
+
+    if (mTarget != GL_TEXTURE_2D)
+    {
+        return false;
+    }
+
+    desired_discard = llmin(desired_discard, mMaxDiscardLevel);
+
+    if (desired_discard <= mCurrentDiscardLevel)
+    {
+        return false;
+    }
+
+    S32 mip = desired_discard - mCurrentDiscardLevel;
+
+    S32 desired_width = getWidth(desired_discard);
+    S32 desired_height = getHeight(desired_discard);
+
+    U64 size = getBytes(desired_discard);
+    llassert(size <= 2048*2048*4); // we shouldn't be using this method to downscale huge textures, but it'll work
+    gGL.getTexUnit(0)->bind(this);
+
+
+    if (sScratchPBO == 0)
+    {
+        glGenBuffers(1, &sScratchPBO);
+        sScratchPBOSize = 0;
+    }
+
+    glBindBuffer(GL_PIXEL_PACK_BUFFER, sScratchPBO);
+
+    if (size > sScratchPBOSize)
+    {
+        glBufferData(GL_PIXEL_PACK_BUFFER, size, NULL, GL_STREAM_COPY);
+        sScratchPBOSize = size;
+    }
+
+    glGetTexImage(mTarget, mip, mFormatPrimary, mFormatType, nullptr);
+
+    free_tex_image(mTexName);
+
+    glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
+
+    glBindBuffer(GL_PIXEL_UNPACK_BUFFER, sScratchPBO);
+    glTexImage2D(mTarget, 0, mFormatPrimary, desired_width, desired_height, 0, mFormatPrimary, mFormatType, nullptr);
+    glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
+
+    alloc_tex_image(desired_width, desired_height, mFormatPrimary);
+
+    if (mHasMipMaps)
+    {
+        LL_PROFILE_ZONE_NAMED_CATEGORY_TEXTURE("scaleDown - glGenerateMipmap");
+        glGenerateMipmap(mTarget);
+    }
+
+    gGL.getTexUnit(0)->unbind(LLTexUnit::TT_TEXTURE);
+
+    mCurrentDiscardLevel = desired_discard;
+
+    return true;
+}
+
+
 //----------------------------------------------------------------------------
 #if LL_IMAGEGL_THREAD_CHECK
 void LLImageGL::checkActiveThread()
diff --git a/indra/llrender/llimagegl.h b/indra/llrender/llimagegl.h
index 5c7a5ce821..3892e9c014 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 <unordered_set>
 
 #define LL_IMAGEGL_THREAD_CHECK 0 //set to 1 to enable thread debugging for ImageGL
 
@@ -83,9 +84,8 @@ public:
     // needs to be called every frame
     static void updateStats(F32 current_time);
 
-    // Save off / restore GL textures
-    static void destroyGL(bool save_state = true);
-    static void restoreGL();
+    // cleanup GL state
+    static void destroyGL();
     static void dirtyTexOptions();
 
     static bool checkSize(S32 width, S32 height);
@@ -148,6 +148,10 @@ public:
     S32  getDiscardLevel() const        { return mCurrentDiscardLevel; }
     S32  getMaxDiscardLevel() const     { return mMaxDiscardLevel; }
 
+    // override the current discard level
+    // should only be used for local textures where you know exactly what you're doing
+    void setDiscardLevel(S32 level) { mCurrentDiscardLevel = level; }
+
     S32  getCurrentWidth() const { return mWidth ;}
     S32  getCurrentHeight() const { return mHeight ;}
     S32  getWidth(S32 discard_level = -1) const;
@@ -194,26 +198,26 @@ public:
     void setFilteringOption(LLTexUnit::eTextureFilterOptions option);
     LLTexUnit::eTextureFilterOptions getFilteringOption(void) const { return mFilterOption; }
 
-    LLGLenum getTexTarget()const { return mTarget ;}
-    S8       getDiscardLevelInAtlas()const {return mDiscardLevelInAtlas;}
-    U32      getTexelsInAtlas()const { return mTexelsInAtlas ;}
-    U32      getTexelsInGLTexture()const {return mTexelsInGLTexture;}
-
+    LLGLenum getTexTarget()const { return mTarget; }
 
     void init(bool usemipmaps);
     virtual void cleanup(); // Clean up the LLImageGL so it can be reinitialized.  Be careful when using this in derived class destructors
 
     void setNeedsAlphaAndPickMask(bool need_mask);
 
-    bool preAddToAtlas(S32 discard_level, const LLImageRaw* raw_image);
-    void postAddToAtlas() ;
-
 #if LL_IMAGEGL_THREAD_CHECK
     // thread debugging
     std::thread::id mActiveThread;
     void checkActiveThread();
 #endif
 
+    // scale down to the desired discard level using GPU
+    // returns true if texture was scaled down
+    // desired discard will be clamped to max discard
+    // if desired discard is less than or equal to current discard, no scaling will occur
+    // only works for GL_TEXTURE_2D target
+    bool scaleDown(S32 desired_discard);
+
 public:
     // Various GL/Rendering options
     S64Bytes mTextureMemory;
@@ -240,15 +244,10 @@ private:
 
     bool     mGLTextureCreated ;
     LLGLuint mTexName;
-    //LLGLuint mNewTexName = 0; // tex name set by background thread to be applied in main thread
     U16      mWidth;
     U16      mHeight;
     S8       mCurrentDiscardLevel;
 
-    S8       mDiscardLevelInAtlas;
-    U32      mTexelsInAtlas ;
-    U32      mTexelsInGLTexture;
-
     bool mAllowCompression;
 
 protected:
@@ -275,7 +274,7 @@ protected:
 
     // STATICS
 public:
-    static std::set<LLImageGL*> sImageList;
+    static std::unordered_set<LLImageGL*> sImageList;
     static S32 sCount;
 
     static F32 sLastFrameTime;
@@ -301,6 +300,8 @@ public:
 private:
     static S32 sMaxCategories;
     static bool sSkipAnalyzeAlpha;
+    static U32 sScratchPBO;
+    static U32 sScratchPBOSize;
 
     //the flag to allow to call readBackRaw(...).
     //can be removed if we do not use that function at all.