Merge pull request #5829 from secondlife/geenz/texture-quality

Texture streaming rework

3 files changed, 125 insertions, 17 deletions

diff --git a/indra/llrender/llimagegl.cpp b/indra/llrender/llimagegl.cpp
index 4a3d32c7ff..6bcc34938c 100644
--- a/indra/llrender/llimagegl.cpp
+++ b/indra/llrender/llimagegl.cpp
@@ -66,8 +66,8 @@ static LLMutex sTexMemMutex;
 static std::unordered_map<U32, U64> sTextureAllocs;
 static U64 sTextureBytes = 0;
 
-// track a texture alloc on the currently bound texture.
-// asserts that no currently tracked alloc exists
+// Per-mip upload paths call this once per level; only free_tex_image
+// removes a texture's accounting entirely.
 void LLImageGLMemory::alloc_tex_image(U32 width, U32 height, U32 intformat, U32 count)
 {
     U32 texUnit = gGL.getCurrentTexUnitIndex();
@@ -80,15 +80,46 @@ void LLImageGLMemory::alloc_tex_image(U32 width, U32 height, U32 intformat, U32
 
     sTexMemMutex.lock();
 
-    // it is a precondition that no existing allocation exists for this texture
-    llassert(sTextureAllocs.find(texName) == sTextureAllocs.end());
-
-    sTextureAllocs[texName] = size;
+    auto iter = sTextureAllocs.find(texName);
+    if (iter != sTextureAllocs.end())
+    {
+        iter->second += size;
+    }
+    else
+    {
+        sTextureAllocs[texName] = size;
+    }
     sTextureBytes += size;
 
     sTexMemMutex.unlock();
 }
 
+// Add mip 1..N bytes to existing accounting. Use after glGenerateMipmap.
+void LLImageGLMemory::account_extra_mip_bytes(U32 base_width, U32 base_height, U32 intformat)
+{
+    U64 extra = 0;
+    U32 w = base_width;
+    U32 h = base_height;
+    while (w > 1 || h > 1)
+    {
+        w = w > 1 ? w >> 1 : 1;
+        h = h > 1 ? h >> 1 : 1;
+        extra += LLImageGL::dataFormatBytes(intformat, w, h);
+    }
+
+    U32 texUnit = gGL.getCurrentTexUnitIndex();
+    U32 texName = gGL.getTexUnit(texUnit)->getCurrTexture();
+
+    sTexMemMutex.lock();
+    auto iter = sTextureAllocs.find(texName);
+    if (iter != sTextureAllocs.end())
+    {
+        iter->second += extra;
+        sTextureBytes += extra;
+    }
+    sTexMemMutex.unlock();
+}
+
 // track texture free on given texName
 void LLImageGLMemory::free_tex_image(U32 texName)
 {
@@ -684,7 +715,10 @@ void LLImageGL::dump()
 //----------------------------------------------------------------------------
 void LLImageGL::forceUpdateBindStats(void) const
 {
-    mLastBindTime = sLastFrameTime;
+    // Intentionally a no-op: mLastBindTime is written only by real bind
+    // paths so the staleness signal reflects actual GPU use. Callers that
+    // still invoke this (avatar "keep alive" sites, deleted-texture
+    // fallback) no longer falsely refresh staleness.
 }
 
 bool LLImageGL::updateBindStats() const
@@ -838,7 +872,7 @@ bool LLImageGL::setImage(const U8* data_in, bool data_hasmips /* = false */, S32
                     mMipLevels = wpo2(llmax(w, h));
 
                     //use legacy mipmap generation mode (note: making this condional can cause rendering issues)
-                    // -- but making it not conditional triggers deprecation warnings when core profile is enabled
+                    // - but making it not conditional triggers deprecation warnings when core profile is enabled
                     //      (some rendering issues while core profile is enabled are acceptable at this point in time)
                     if (!LLRender::sGLCoreProfile)
                     {
@@ -864,6 +898,7 @@ bool LLImageGL::setImage(const U8* data_in, bool data_hasmips /* = false */, S32
                     {
                         LL_PROFILE_GPU_ZONE("generate mip map");
                         glGenerateMipmap(mTarget);
+                        account_extra_mip_bytes(w, h, mFormatInternal);
                     }
                     stop_glerror();
                 }
@@ -1461,7 +1496,12 @@ void LLImageGL::setManualImage(U32 target, S32 miplevel, S32 intformat, S32 widt
         LL_PROFILE_ZONE_NUM(width);
         LL_PROFILE_ZONE_NUM(height);
 
-        free_cur_tex_image();
+        // Release prior accounting only on the base mip; per-mip iteration
+        // accumulates the rest via the additive alloc_tex_image.
+        if (miplevel == 0)
+        {
+            free_cur_tex_image();
+        }
         const bool use_sub_image = should_stagger_image_set(compress);
         if (!use_sub_image)
         {
@@ -1613,7 +1653,6 @@ bool LLImageGL::createGLTexture(S32 discard_level, const LLImageRaw* imageraw, S
     {
         destroyGLTexture();
         mCurrentDiscardLevel = discard_level;
-        mLastBindTime = sLastFrameTime;
         mGLTextureCreated = false;
         return true ;
     }
@@ -1729,9 +1768,7 @@ bool LLImageGL::createGLTexture(S32 discard_level, const U8* data_in, bool data_
 
 
     mTextureMemory = (S64Bytes)getMipBytes(mCurrentDiscardLevel);
-
-    // mark this as bound at this point, so we don't throw it out immediately
-    mLastBindTime = sLastFrameTime;
+    mGLCreateTime = sLastFrameTime;
 
     checkActiveThread();
     return true;
@@ -1858,7 +1895,7 @@ bool LLImageGL::readBackRaw(S32 discard_level, LLImageRaw* imageraw, bool compre
     LLGLint is_compressed = 0;
     if (compressed_ok)
     {
-        glGetTexLevelParameteriv(mTarget, is_compressed, GL_TEXTURE_COMPRESSED, (GLint*)&is_compressed);
+        glGetTexLevelParameteriv(mTarget, gl_discard, GL_TEXTURE_COMPRESSED, (GLint*)&is_compressed);
     }
 
     //-----------------------------------------------------------------------------------------------
@@ -2039,6 +2076,28 @@ S32 LLImageGL::getWidth(S32 discard_level) const
     return width;
 }
 
+// static
+S32 LLImageGL::dimDerivedMaxDiscard(S32 width, S32 height)
+{
+    if (width <= 0 || height <= 0)
+    {
+        return 0;
+    }
+    // max(w,h) - min() caps short on rectangular textures
+    // (1024x512 reaches 1x1 at discard 10, not 9).
+    return (S32)floorf(log2f((F32)llmax(width, 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.
+    if (mLastBindTime != sLastFrameTime)
+    {
+        mLastBindTime = sLastFrameTime;
+    }
+}
+
 S64 LLImageGL::getBytes(S32 discard_level) const
 {
     if (discard_level < 0)
@@ -2468,7 +2527,12 @@ bool LLImageGL::scaleDown(S32 desired_discard)
         return false;
     }
 
-    desired_discard = llmin(desired_discard, mMaxDiscardLevel);
+    // GL pyramid reaches 1x1 regardless of codec levels;
+    // mMaxDiscardLevel is hardcapped at MAX_DISCARD_LEVEL.
+    S32 dim_max_discard = (mWidth > 0 && mHeight > 0)
+        ? dimDerivedMaxDiscard(mWidth, mHeight)
+        : (S32)mMaxDiscardLevel;
+    desired_discard = llmin(desired_discard, dim_max_discard);
 
     if (desired_discard <= mCurrentDiscardLevel)
     {
@@ -2501,6 +2565,7 @@ bool LLImageGL::scaleDown(S32 desired_discard)
                 LL_PROFILE_ZONE_NAMED_CATEGORY_TEXTURE("scaleDown - glGenerateMipmap");
                 gGL.getTexUnit(0)->bind(this);
                 glGenerateMipmap(mTarget);
+                account_extra_mip_bytes(desired_width, desired_height, mFormatInternal);
                 gGL.getTexUnit(0)->unbind(LLTexUnit::TT_TEXTURE);
             }
         }
@@ -2546,6 +2611,7 @@ bool LLImageGL::scaleDown(S32 desired_discard)
         {
             LL_PROFILE_ZONE_NAMED_CATEGORY_TEXTURE("scaleDown - glGenerateMipmap");
             glGenerateMipmap(mTarget);
+            account_extra_mip_bytes(desired_width, desired_height, mFormatInternal);
         }
 
         gGL.getTexUnit(0)->unbind(LLTexUnit::TT_TEXTURE);
diff --git a/indra/llrender/llimagegl.h b/indra/llrender/llimagegl.h
index 6b4492c09e..0c85446b84 100644
--- a/indra/llrender/llimagegl.h
+++ b/indra/llrender/llimagegl.h
@@ -51,6 +51,11 @@ class LLWindow;
 namespace LLImageGLMemory
 {
     void alloc_tex_image(U32 width, U32 height, U32 intformat, U32 count);
+
+    // Add mip 1..N bytes to existing accounting. Call after glGenerateMipmap
+    // when only the base mip was accounted; without this the bytes counter
+    // undercounts mipmap-generated textures by ~25%.
+    void account_extra_mip_bytes(U32 base_width, U32 base_height, U32 intformat);
     void free_tex_image(U32 texName);
     void free_tex_images(U32 count, const U32* texNames);
     void free_cur_tex_image();
@@ -151,6 +156,15 @@ public:
     S32  getDiscardLevel() const        { return mCurrentDiscardLevel; }
     S32  getMaxDiscardLevel() const     { return mMaxDiscardLevel; }
 
+    // floor(log2(max(w, h))) - deepest GL pyramid level (down to 1x1).
+    // Returns 0 for non-positive inputs.
+    static S32 dimDerivedMaxDiscard(S32 width, S32 height);
+
+    // Record the wall-clock bind time - every bind path that touches a
+    // streaming-managed texture must call this, or the staleness signal
+    // sees the texture as never-bound and ramps it toward eviction.
+    void stampBound() const;
+
     // 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; }
@@ -224,7 +238,8 @@ public:
 public:
     // Various GL/Rendering options
     S64Bytes mTextureMemory;
-    mutable F32  mLastBindTime; // last time this was bound, by discard level
+    mutable F32  mLastBindTime = 0.f; // wall-clock time at last stampBound; drives streaming staleness
+    F32          mGLCreateTime = 0.f; // wall-clock time the GL texture was created; staleness fallback for never-bound textures
 
 private:
     U32 createPickMask(S32 pWidth, S32 pHeight);
diff --git a/indra/llrender/llrender.cpp b/indra/llrender/llrender.cpp
index 57be8570af..5e845fbcce 100644
--- a/indra/llrender/llrender.cpp
+++ b/indra/llrender/llrender.cpp
@@ -197,6 +197,10 @@ void LLTexUnit::bindFast(LLTexture* texture)
     glActiveTexture(GL_TEXTURE0 + mIndex);
     gGL.mCurrTextureUnitIndex = mIndex;
     mCurrTexture = gl_tex->getTexName();
+    mCurrTexType = gl_tex->getTarget();
+    // bindFast bypasses updateBindStats(); stamp directly so the staleness
+    // signal sees per-frame use of batched textures.
+    gl_tex->stampBound();
     if (!mCurrTexture)
     {
         LL_PROFILE_ZONE_NAMED("MISSING TEXTURE");
@@ -249,11 +253,17 @@ bool LLTexUnit::bind(LLTexture* texture, bool for_rendering, bool forceBind)
                         setTextureFilteringOption(gl_tex->mFilterOption);
                     }
                 }
+                else
+                {
+                    // Already current - still being used, keep it fresh.
+                    gl_tex->stampBound();
+                }
             }
             else
             {
                 //if deleted, will re-generate it immediately
                 texture->forceImmediateUpdate() ;
+                gl_tex->stampBound();
 
                 gl_tex->forceUpdateBindStats() ;
                 return texture->bindDefaultImage(mIndex);
@@ -325,6 +335,11 @@ bool LLTexUnit::bind(LLImageGL* texture, bool for_rendering, bool forceBind, S32
             stop_glerror();
         }
     }
+    else
+    {
+        // Already current - still being used, keep it fresh.
+        texture->stampBound();
+    }
 
     stop_glerror();
 
@@ -1718,7 +1733,16 @@ LLVertexBuffer* LLRender::genBuffer(U32 attribute_mask, S32 count)
     LLVertexBuffer * vb = new LLVertexBuffer(attribute_mask);
     vb->allocateBuffer(count, 0);
 
-    vb->setBuffer();
+    // Non-Apple path uses glBufferSubData inside setXxxData, so the VBO
+    // must already be bound. On Apple, the VBO is lazily created in
+    // _unmapBuffer (LLAppleVBOPool); calling setBuffer() here would bind
+    // mGLBuffer == 0 and then setupVertexBuffer would issue
+    // glVertexAttribIPointer with a non-null offset against no bound
+    // GL_ARRAY_BUFFER -> GL_INVALID_OPERATION in core profile.
+    if (!gGLManager.mIsApple)
+    {
+        vb->setBuffer();
+    }
 
     vb->setPositionData(mVerticesp.get());
 
@@ -1733,6 +1757,9 @@ LLVertexBuffer* LLRender::genBuffer(U32 attribute_mask, S32 count)
     }
 
 #if LL_DARWIN
+    // unmapBuffer creates the GL buffer, uploads, and leaves it bound,
+    // drawBuffer's later setBuffer() then runs setupVertexBuffer against
+    // a valid VBO.
     vb->unmapBuffer();
 #endif
     vb->unbind();