diff --git a/indra/llrender/llrender.cpp b/indra/llrender/llrender.cpp
index 8ab8d3151b749db13203959be5bc16d0a9542fb2..7c974934aec142a1cac356c613d3a15243e969ec 100644
--- a/indra/llrender/llrender.cpp
+++ b/indra/llrender/llrender.cpp
@@ -35,6 +35,7 @@
#include "llrendertarget.h"
#include "lltexture.h"
#include "llshadermgr.h"
+#include "llmd5.h"
#if LL_WINDOWS
extern void APIENTRY gl_debug_callback(GLenum source,
@@ -66,6 +67,14 @@ LLVector2 LLRender::sUIGLScaleFactor = LLVector2(1.f, 1.f);
static const U32 LL_NUM_TEXTURE_LAYERS = 32;
static const U32 LL_NUM_LIGHT_UNITS = 8;
+struct LLVBCache
+{
+ LLPointer<LLVertexBuffer> vb;
+ std::chrono::steady_clock::time_point touched;
+};
+
+static std::unordered_map<std::size_t, LLVBCache> sVBCache;
+
static const GLenum sGLTextureType[] =
{
GL_TEXTURE_2D,
@@ -1635,24 +1644,105 @@ void LLRender::flush()
if (mBuffer)
{
- if (mBuffer->useVBOs() && !mBuffer->isLocked())
- { //hack to only flush the part of the buffer that was updated (relies on stream draw using buffersubdata)
- mBuffer->getVertexStrider(mVerticesp, 0, count);
- mBuffer->getTexCoord0Strider(mTexcoordsp, 0, count);
- mBuffer->getColorStrider(mColorsp, 0, count);
+
+ LLMD5 hash;
+ U32 attribute_mask = LLGLSLShader::sCurBoundShaderPtr->mAttributeMask;
+
+ {
+ LL_PROFILE_ZONE_NAMED_CATEGORY_VERTEX("vb cache hash");
+
+ hash.update((U8*)mVerticesp.get(), count * sizeof(LLVector4a));
+ if (attribute_mask & LLVertexBuffer::MAP_TEXCOORD0)
+ {
+ hash.update((U8*)mTexcoordsp.get(), count * sizeof(LLVector2));
+ }
+
+ if (attribute_mask & LLVertexBuffer::MAP_COLOR)
+ {
+ hash.update((U8*)mColorsp.get(), count * sizeof(LLColor4U));
+ }
+
+ hash.finalize();
+ }
+
+ size_t vhash[2];
+ hash.raw_digest((unsigned char*) vhash);
+
+ // check the VB cache before making a new vertex buffer
+ // This is a giant hack to deal with (mostly) our terrible UI rendering code
+ // that was built on top of OpenGL immediate mode. Huge performance wins
+ // can be had by not uploading geometry to VRAM unless absolutely necessary.
+ // Most of our usage of the "immediate mode" style draw calls is actually
+ // sending the same geometry over and over again.
+ // To leverage this, we maintain a running hash of the vertex stream being
+ // built up before a flush, and then check that hash against a VB
+ // cache just before creating a vertex buffer in VRAM
+ auto& cache = sVBCache.find(vhash[0]);
+
+ LLPointer<LLVertexBuffer> vb;
+
+ if (cache != sVBCache.end())
+ {
+ LL_PROFILE_ZONE_NAMED_CATEGORY_VERTEX("vb cache hit");
+ // cache hit, just use the cached buffer
+ vb = cache->second.vb;
+ cache->second.touched = std::chrono::steady_clock::now();
+ }
+ else
+ {
+ LL_PROFILE_ZONE_NAMED_CATEGORY_VERTEX("vb cache miss");
+ vb = new LLVertexBuffer(attribute_mask, GL_STATIC_DRAW);
+ vb->allocateBuffer(count, 0, true);
+ vb->setPositionData((LLVector4a*) mVerticesp.get());
+
+ if (attribute_mask & LLVertexBuffer::MAP_TEXCOORD0)
+ {
+ vb->setTexCoordData(mTexcoordsp.get());
+ }
+
+ if (attribute_mask & LLVertexBuffer::MAP_COLOR)
+ {
+ vb->setColorData(mColorsp.get());
+ }
+
+ vb->unbind();
+
+ sVBCache[vhash[0]] = { vb , std::chrono::steady_clock::now() };
+
+ static U32 miss_count = 0;
+ miss_count++;
+ if (miss_count > 1024)
+ {
+ LL_PROFILE_ZONE_NAMED_CATEGORY_VERTEX("vb cache clean");
+ miss_count = 0;
+ auto now = std::chrono::steady_clock::now();
+
+ using namespace std::chrono_literals;
+ // every 1024 misses, clean the cache of any VBs that haven't been touched in the last second
+ for (auto& iter = sVBCache.begin(); iter != sVBCache.end(); )
+ {
+ if (now - iter->second.touched > 1s)
+ {
+ iter = sVBCache.erase(iter);
+ }
+ else
+ {
+ ++iter;
+ }
+ }
+ }
}
- mBuffer->flush();
- mBuffer->setBuffer(immediate_mask);
+ vb->setBuffer(immediate_mask);
if (mMode == LLRender::QUADS && sGLCoreProfile)
{
- mBuffer->drawArrays(LLRender::TRIANGLES, 0, count);
+ vb->drawArrays(LLRender::TRIANGLES, 0, count);
mQuadCycle = 1;
}
else
{
- mBuffer->drawArrays(mMode, 0, count);
+ vb->drawArrays(mMode, 0, count);
}
}
else
diff --git a/indra/llrender/llvertexbuffer.cpp b/indra/llrender/llvertexbuffer.cpp
index 20261dcb8a2f3622e04ed6712ab2c5f051ba1ba0..57be21cf6e62b1e435f9a98597fc43ea3179c689 100644
--- a/indra/llrender/llvertexbuffer.cpp
+++ b/indra/llrender/llvertexbuffer.cpp
@@ -1180,7 +1180,6 @@ bool expand_region(LLVertexBuffer::MappedRegion& region, S32 start, S32 end)
U8* LLVertexBuffer::mapVertexBuffer(S32 type, S32 index, S32 count, bool map_range)
{
LL_PROFILE_ZONE_SCOPED_CATEGORY_VERTEX;
- bindGLBuffer(true);
if (mFinal)
{
LL_ERRS() << "LLVertexBuffer::mapVeretxBuffer() called on a finalized buffer." << LL_ENDL;
@@ -1256,7 +1255,6 @@ U8* LLVertexBuffer::mapVertexBuffer(S32 type, S32 index, S32 count, bool map_ran
U8* LLVertexBuffer::mapIndexBuffer(S32 index, S32 count, bool map_range)
{
LL_PROFILE_ZONE_SCOPED_CATEGORY_VERTEX;
- bindGLIndices(true);
if (mFinal)
{
LL_ERRS() << "LLVertexBuffer::mapIndexBuffer() called on a finalized buffer." << LL_ENDL;
@@ -2070,6 +2068,24 @@ void LLVertexBuffer::setupVertexBufferFast(U32 data_mask)
void* ptr = (void*)(base + mOffsets[TYPE_VERTEX]);
glVertexAttribPointer(loc, 3, GL_FLOAT, GL_FALSE, LLVertexBuffer::sTypeSize[TYPE_VERTEX], ptr);
}
- }
+}
+
+void LLVertexBuffer::setPositionData(const LLVector4a* data)
+{
+ bindGLBuffer();
+ flush_vbo(GL_ARRAY_BUFFER, 0, sizeof(LLVector4a) * getNumVerts(), (U8*) data);
+}
+
+void LLVertexBuffer::setTexCoordData(const LLVector2* data)
+{
+ bindGLBuffer();
+ flush_vbo(GL_ARRAY_BUFFER, mOffsets[TYPE_TEXCOORD0], mOffsets[TYPE_TEXCOORD0] + sTypeSize[TYPE_TEXCOORD0] * getNumVerts(), (U8*)data);
+}
+
+void LLVertexBuffer::setColorData(const LLColor4U* data)
+{
+ bindGLBuffer();
+ flush_vbo(GL_ARRAY_BUFFER, mOffsets[TYPE_COLOR], mOffsets[TYPE_COLOR] + sTypeSize[TYPE_COLOR] * getNumVerts(), (U8*) data);
+}
diff --git a/indra/llrender/llvertexbuffer.h b/indra/llrender/llvertexbuffer.h
index 74b951884ddad7e8d360960d899f6b8b7932b4ec..926d37b0523278eff7b2d38b5f8f379c6952ddd2 100644
--- a/indra/llrender/llvertexbuffer.h
+++ b/indra/llrender/llvertexbuffer.h
@@ -212,6 +212,10 @@ class LLVertexBuffer : public LLRefCount
bool getMetallicRoughnessTexcoordStrider(LLStrider<LLVector2>& strider, S32 index=0, S32 count = -1, bool map_range = false);
bool getEmissiveTexcoordStrider(LLStrider<LLVector2>& strider, S32 index=0, S32 count = -1, bool map_range = false);
+ void setPositionData(const LLVector4a* data);
+ void setTexCoordData(const LLVector2* data);
+ void setColorData(const LLColor4U* data);
+
bool useVBOs() const;
bool isEmpty() const { return mEmpty; }