#1870 Tune up for better experience on integrated intel with low memory (#1872)

* More deterministic vsize calculation.   Add control for choosing downscale method.
* Quick hack to make GLTF preview work again

6 files changed, 120 insertions, 36 deletions

diff --git a/indra/llrender/llgl.h b/indra/llrender/llgl.h
index 909dad2e85..be1c0a532a 100644
--- a/indra/llrender/llgl.h
+++ b/indra/llrender/llgl.h
@@ -102,6 +102,9 @@ public:
     bool mIsNVIDIA;
     bool mIsIntel;
 
+    // hints to the render pipe
+    U32 mDownScaleMethod = 0; // see settings.xml RenderDownScaleMethod
+
 #if LL_DARWIN
     // Needed to distinguish problem cards on older Macs that break with Materials
     bool mIsMobileGF;
diff --git a/indra/llrender/llimagegl.cpp b/indra/llrender/llimagegl.cpp
index 956bcef352..34200ef5cb 100644
--- a/indra/llrender/llimagegl.cpp
+++ b/indra/llrender/llimagegl.cpp
@@ -57,6 +57,9 @@ const F32 MIN_TEXTURE_LIFETIME = 10.f;
 U32 wpo2(U32 i);
 
 
+U32 LLImageGL::sFrameCount = 0;
+
+
 // texture memory accounting (for macOS)
 static LLMutex sTexMemMutex;
 static std::unordered_map<U32, U64> sTextureAllocs;
@@ -1003,7 +1006,7 @@ bool should_stagger_image_set(bool compressed)
 #else
     // glTexSubImage2D doesn't work with compressed textures on select tested Nvidia GPUs on Windows 10 -Cosmic,2023-03-08
     // Setting media textures off-thread seems faster when not using sub_image_lines (Nvidia/Windows 10) -Cosmic,2023-03-31
-    return !compressed && on_main_thread();
+    return !compressed && on_main_thread() && !gGLManager.mIsIntel;
 #endif
 }
 
@@ -1186,12 +1189,36 @@ void LLImageGL::generateTextures(S32 numTextures, U32 *textures)
 }
 
 // static
+void LLImageGL::updateClass()
+{
+    sFrameCount++;
+}
+
+// static
 void LLImageGL::deleteTextures(S32 numTextures, const U32 *textures)
 {
+    // wait a few frames before actually deleting the textures to avoid
+    // synchronization issues with the GPU
+    static std::vector<U32> sFreeList[4];
+
     if (gGLManager.mInited)
     {
-        free_tex_images(numTextures, textures);
-        glDeleteTextures(numTextures, textures);
+        LL_PROFILE_ZONE_SCOPED_CATEGORY_TEXTURE;
+        U32 idx = sFrameCount % 4;
+
+        for (S32 i = 0; i < numTextures; ++i)
+        {
+            sFreeList[idx].push_back(textures[i]);
+        }
+
+        idx = (sFrameCount + 3) % 4;
+
+        if (!sFreeList[idx].empty())
+        {
+            glDeleteTextures((GLsizei) sFreeList[idx].size(), sFreeList[idx].data());
+            free_tex_images((GLsizei) sFreeList[idx].size(), sFreeList[idx].data());
+            sFreeList[idx].resize(0);
+        }
     }
 }
 
@@ -2335,44 +2362,85 @@ bool LLImageGL::scaleDown(S32 desired_discard)
     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 (gGLManager.mDownScaleMethod == 0)
+    { // use an FBO to downscale the texture
+        // allocate new texture
+        U32 temp_texname = 0;
+        generateTextures(1, &temp_texname);
+        gGL.getTexUnit(0)->bindManual(LLTexUnit::TT_TEXTURE, temp_texname, true);
+        {
+            LL_PROFILE_ZONE_NAMED_CATEGORY_TEXTURE("scaleDown - glTexImage2D");
+            glTexImage2D(mTarget, 0, mFormatPrimary, desired_width, desired_height, 0, mFormatPrimary, mFormatType, NULL);
+        }
 
+        // account for new texture getting created
+        alloc_tex_image(desired_width, desired_height, mFormatPrimary);
 
-    if (sScratchPBO == 0)
-    {
-        glGenBuffers(1, &sScratchPBO);
-        sScratchPBOSize = 0;
-    }
+        // Use render-to-texture to scale down the texture
+        {
+            LL_PROFILE_ZONE_NAMED_CATEGORY_TEXTURE("scaleDown - glFramebufferTexture2D");
+            glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, mTarget, temp_texname, 0);
+        }
 
-    glBindBuffer(GL_PIXEL_PACK_BUFFER, sScratchPBO);
+        glViewport(0, 0, desired_width, desired_height);
 
-    if (size > sScratchPBOSize)
-    {
-        glBufferData(GL_PIXEL_PACK_BUFFER, size, NULL, GL_STREAM_COPY);
-        sScratchPBOSize = size;
+        // draw a full screen triangle
+        gGL.getTexUnit(0)->bind(this);
+        glDrawArrays(GL_TRIANGLES, 0, 3);
+        gGL.getTexUnit(0)->unbind(LLTexUnit::TT_TEXTURE);
+
+        // delete old texture and assign new texture name
+        deleteTextures(1, &mTexName);
+        mTexName = temp_texname;
+
+        if (mHasMipMaps)
+        { // generate mipmaps if needed
+            LL_PROFILE_ZONE_NAMED_CATEGORY_TEXTURE("scaleDown - glGenerateMipmap");
+            gGL.getTexUnit(0)->bind(this);
+            glGenerateMipmap(mTarget);
+            gGL.getTexUnit(0)->unbind(LLTexUnit::TT_TEXTURE);
+        }
     }
+    else
+    { // use a PBO to downscale the texture
+        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;
+        }
 
-    glGetTexImage(mTarget, mip, mFormatPrimary, mFormatType, nullptr);
+        glBindBuffer(GL_PIXEL_PACK_BUFFER, sScratchPBO);
 
-    free_tex_image(mTexName);
+        if (size > sScratchPBOSize)
+        {
+            glBufferData(GL_PIXEL_PACK_BUFFER, size, NULL, GL_STREAM_COPY);
+            sScratchPBOSize = size;
+        }
 
-    glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
+        glGetTexImage(mTarget, mip, mFormatPrimary, mFormatType, nullptr);
 
-    glBindBuffer(GL_PIXEL_UNPACK_BUFFER, sScratchPBO);
-    glTexImage2D(mTarget, 0, mFormatPrimary, desired_width, desired_height, 0, mFormatPrimary, mFormatType, nullptr);
-    glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
+        free_tex_image(mTexName);
 
-    alloc_tex_image(desired_width, desired_height, mFormatPrimary);
+        glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
 
-    if (mHasMipMaps)
-    {
-        LL_PROFILE_ZONE_NAMED_CATEGORY_TEXTURE("scaleDown - glGenerateMipmap");
-        glGenerateMipmap(mTarget);
-    }
+        glBindBuffer(GL_PIXEL_UNPACK_BUFFER, sScratchPBO);
+        glTexImage2D(mTarget, 0, mFormatPrimary, desired_width, desired_height, 0, mFormatPrimary, mFormatType, nullptr);
+        glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 
-    gGL.getTexUnit(0)->unbind(LLTexUnit::TT_TEXTURE);
+        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;
 
diff --git a/indra/llrender/llimagegl.h b/indra/llrender/llimagegl.h
index 3892e9c014..5073701c30 100644
--- a/indra/llrender/llimagegl.h
+++ b/indra/llrender/llimagegl.h
@@ -62,6 +62,9 @@ class LLImageGL : public LLRefCount
     friend class LLTexUnit;
 public:
 
+    // call once per frame
+    static void updateClass();
+
     // 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
@@ -276,7 +279,7 @@ protected:
 public:
     static std::unordered_set<LLImageGL*> sImageList;
     static S32 sCount;
-
+    static U32 sFrameCount;
     static F32 sLastFrameTime;
 
     // Global memory statistics
diff --git a/indra/llrender/llrender.h b/indra/llrender/llrender.h
index be9f3895e7..e6b4118beb 100644
--- a/indra/llrender/llrender.h
+++ b/indra/llrender/llrender.h
@@ -297,7 +297,11 @@ public:
         ALTERNATE_DIFFUSE_MAP = 1,
         NORMAL_MAP            = 1,
         SPECULAR_MAP          = 2,
-        NUM_TEXTURE_CHANNELS  = 3,
+        BASECOLOR_MAP         = 3,
+        METALLIC_ROUGHNESS_MAP = 4,
+        GLTF_NORMAL_MAP           = 5,
+        EMISSIVE_MAP          = 6,
+        NUM_TEXTURE_CHANNELS  = 7,
     };
 
     enum eVolumeTexIndex : U8
diff --git a/indra/llrender/llrendertarget.cpp b/indra/llrender/llrendertarget.cpp
index efdde77a32..60159a0497 100644
--- a/indra/llrender/llrendertarget.cpp
+++ b/indra/llrender/llrendertarget.cpp
@@ -426,14 +426,17 @@ void LLRenderTarget::bindTarget()
                             GL_COLOR_ATTACHMENT1,
                             GL_COLOR_ATTACHMENT2,
                             GL_COLOR_ATTACHMENT3};
-    glDrawBuffers(static_cast<GLsizei>(mTex.size()), drawbuffers);
 
     if (mTex.empty())
     { //no color buffer to draw to
         glDrawBuffer(GL_NONE);
         glReadBuffer(GL_NONE);
     }
-
+    else
+    {
+        glDrawBuffers(static_cast<GLsizei>(mTex.size()), drawbuffers);
+        glReadBuffer(GL_COLOR_ATTACHMENT0);
+    }
     check_framebuffer_status();
 
     glViewport(0, 0, mResX, mResY);
@@ -519,7 +522,8 @@ void LLRenderTarget::flush()
     llassert(sCurFBO == mFBO);
     llassert(sBoundTarget == this);
 
-    if (mGenerateMipMaps == LLTexUnit::TMG_AUTO) {
+    if (mGenerateMipMaps == LLTexUnit::TMG_AUTO)
+    {
         LL_PROFILE_GPU_ZONE("rt generate mipmaps");
         bindTexture(0, 0, LLTexUnit::TFO_TRILINEAR);
         glGenerateMipmap(GL_TEXTURE_2D);
@@ -540,6 +544,8 @@ void LLRenderTarget::flush()
         glViewport(gGLViewport[0], gGLViewport[1], gGLViewport[2], gGLViewport[3]);
         sCurResX = gGLViewport[2];
         sCurResY = gGLViewport[3];
+        glReadBuffer(GL_BACK);
+        glDrawBuffer(GL_BACK);
     }
 }
 
diff --git a/indra/llrender/llrendertarget.h b/indra/llrender/llrendertarget.h
index 340276a752..a1adf93fa1 100644
--- a/indra/llrender/llrendertarget.h
+++ b/indra/llrender/llrendertarget.h
@@ -172,6 +172,8 @@ public:
     // *HACK
     void swapFBORefs(LLRenderTarget& other);
 
+    static LLRenderTarget* sBoundTarget;
+
 protected:
     U32 mResX;
     U32 mResY;
@@ -186,8 +188,6 @@ protected:
     U32 mMipLevels;
 
     LLTexUnit::eTextureType mUsage;
-
-    static LLRenderTarget* sBoundTarget;
 };
 
 #endif