From 0a617904f98ab5960379099822e4891a08137e68 Mon Sep 17 00:00:00 2001 From: Dave Parks Date: Tue, 17 Sep 2024 18:14:22 -0500 Subject: #2590 Fix for horrible FPS on Intel Mac (#2591) * Work around for GHA mac runners not playing nice with Tracy * Delay VBO deletion for a few frames * Enable multithreaded GL driver and multithreaded media textures on Apple silicon --- indra/llrender/llgl.cpp | 5 + indra/llrender/llgl.h | 1 + indra/llrender/llvertexbuffer.cpp | 374 +++++++++++++++++++++----------------- 3 files changed, 217 insertions(+), 163 deletions(-) (limited to 'indra/llrender') diff --git a/indra/llrender/llgl.cpp b/indra/llrender/llgl.cpp index c62cacdce6..c5c9d50dee 100644 --- a/indra/llrender/llgl.cpp +++ b/indra/llrender/llgl.cpp @@ -1170,6 +1170,11 @@ bool LLGLManager::initGL() mGLVendorShort = "INTEL"; mIsIntel = true; } + else if (mGLVendor.find("APPLE") != std::string::npos) + { + mGLVendorShort = "APPLE"; + mIsApple = true; + } else { mGLVendorShort = "MISC"; diff --git a/indra/llrender/llgl.h b/indra/llrender/llgl.h index 17f825bd71..f5b1e8d786 100644 --- a/indra/llrender/llgl.h +++ b/indra/llrender/llgl.h @@ -102,6 +102,7 @@ public: bool mIsAMD; bool mIsNVIDIA; bool mIsIntel; + bool mIsApple = false; // hints to the render pipe U32 mDownScaleMethod = 0; // see settings.xml RenderDownScaleMethod diff --git a/indra/llrender/llvertexbuffer.cpp b/indra/llrender/llvertexbuffer.cpp index e9fa369b0c..0be799db9d 100644 --- a/indra/llrender/llvertexbuffer.cpp +++ b/indra/llrender/llvertexbuffer.cpp @@ -289,22 +289,58 @@ static GLuint gen_buffer() return ret; } -#define ANALYZE_VBO_POOL 0 +static void delete_buffers(S32 count, GLuint* buffers) +{ + LL_PROFILE_ZONE_SCOPED_CATEGORY_VERTEX; + // wait a few frames before actually deleting the buffers to avoid + // synchronization issues with the GPU + static std::vector sFreeList[4]; + + if (gGLManager.mInited) + { + U32 idx = LLImageGL::sFrameCount % 4; + + for (S32 i = 0; i < count; ++i) + { + sFreeList[idx].push_back(buffers[i]); + } + + idx = (LLImageGL::sFrameCount + 3) % 4; + + if (!sFreeList[idx].empty()) + { + glDeleteBuffers((GLsizei)sFreeList[idx].size(), sFreeList[idx].data()); + sFreeList[idx].resize(0); + } + } +} -#if LL_DARWIN -// experimental -- disable VBO pooling on OS X and use glMapBuffer +#define ANALYZE_VBO_POOL 0 + +// VBO Pool interface class LLVBOPool +{ + public: + virtual ~LLVBOPool() = default; + virtual void allocate(GLenum type, U32 size, GLuint& name, U8*& data) = 0; + virtual void free(GLenum type, U32 size, GLuint name, U8* data) = 0; + virtual U64 getVramBytesUsed() = 0; +}; + +// VBO Pool for Apple GPUs (as in M1/M2 etc, not Intel macs) +// Effectively disables VBO pooling +class LLAppleVBOPool final: public LLVBOPool { public: U64 mAllocated = 0; - U64 getVramBytesUsed() + U64 getVramBytesUsed() override { return mAllocated; } - void allocate(GLenum type, U32 size, GLuint& name, U8*& data) + void allocate(GLenum type, U32 size, GLuint& name, U8*& data) override { LL_PROFILE_ZONE_SCOPED_CATEGORY_VERTEX; STOP_GLERROR; @@ -324,7 +360,7 @@ public: } } - void free(GLenum type, U32 size, GLuint name, U8* data) + void free(GLenum type, U32 size, GLuint name, U8* data) override { LL_PROFILE_ZONE_SCOPED_CATEGORY_VERTEX; llassert(type == GL_ARRAY_BUFFER || type == GL_ELEMENT_ARRAY_BUFFER); @@ -339,19 +375,17 @@ public: STOP_GLERROR; if (name) { - glDeleteBuffers(1, &name); + delete_buffers(1, &name); } STOP_GLERROR; } }; -#else - -class LLVBOPool +// VBO Pool for GPUs that benefit from VBO pooling +class LLDefaultVBOPool final : public LLVBOPool { public: typedef std::chrono::steady_clock::time_point Time; - struct Entry { U8* mData; @@ -359,7 +393,7 @@ public: Time mAge; }; - ~LLVBOPool() + ~LLDefaultVBOPool() override { clear(); } @@ -377,7 +411,7 @@ public: U32 mMisses = 0; U32 mHits = 0; - U64 getVramBytesUsed() + U64 getVramBytesUsed() override { return mAllocated + mReserved; } @@ -393,7 +427,7 @@ public: size += block_size - (size % block_size); } - void allocate(GLenum type, U32 size, GLuint& name, U8*& data) + void allocate(GLenum type, U32 size, GLuint& name, U8*& data) override { LL_PROFILE_ZONE_SCOPED_CATEGORY_VERTEX; llassert(type == GL_ARRAY_BUFFER || type == GL_ELEMENT_ARRAY_BUFFER); @@ -449,7 +483,7 @@ public: clean(); } - void free(GLenum type, U32 size, GLuint name, U8* data) + void free(GLenum type, U32 size, GLuint name, U8* data) override { LL_PROFILE_ZONE_SCOPED_CATEGORY_VERTEX; llassert(type == GL_ARRAY_BUFFER || type == GL_ELEMENT_ARRAY_BUFFER); @@ -512,7 +546,7 @@ public: LL_PROFILE_ZONE_NAMED_CATEGORY_VERTEX("vbo cache timeout"); auto& entry = entries.back(); ll_aligned_free_16(entry.mData); - glDeleteBuffers(1, &entry.mGLName); + delete_buffers(1, &entry.mGLName); llassert(mReserved >= iter->first); mReserved -= iter->first; entries.pop_back(); @@ -548,7 +582,7 @@ public: for (auto& entry : entries.second) { ll_aligned_free_16(entry.mData); - glDeleteBuffers(1, &entry.mGLName); + delete_buffers(1, &entry.mGLName); } } @@ -557,7 +591,7 @@ public: for (auto& entry : entries.second) { ll_aligned_free_16(entry.mData); - glDeleteBuffers(1, &entry.mGLName); + delete_buffers(1, &entry.mGLName); } } @@ -567,7 +601,6 @@ public: mVBOPool.clear(); } }; -#endif static LLVBOPool* sVBOPool = nullptr; @@ -896,7 +929,16 @@ void LLVertexBuffer::drawArrays(U32 mode, U32 first, U32 count) const void LLVertexBuffer::initClass(LLWindow* window) { llassert(sVBOPool == nullptr); - sVBOPool = new LLVBOPool(); + if (gGLManager.mIsApple) + { + LL_INFOS() << "VBO Pooling Disabled" << LL_ENDL; + sVBOPool = new LLAppleVBOPool(); + } + else + { + LL_INFOS() << "VBO Pooling Enabled" << LL_ENDL; + sVBOPool = new LLDefaultVBOPool(); + } #if ENABLE_GL_WORK_QUEUE sQueue = new GLWorkQueue(); @@ -964,7 +1006,6 @@ void LLVertexBuffer::flushBuffers() { LL_PROFILE_ZONE_SCOPED_CATEGORY_VERTEX; // must only be called from main thread - llassert(LLCoros::on_main_thread_main_coro()); for (auto& buffer : sMappedBuffers) { buffer->_unmapBuffer(); @@ -1231,28 +1272,29 @@ U8* LLVertexBuffer::mapVertexBuffer(LLVertexBuffer::AttributeType type, U32 inde count = mNumVerts - index; } -#if !LL_DARWIN - U32 start = mOffsets[type] + sTypeSize[type] * index; - U32 end = start + sTypeSize[type] * count-1; - - bool flagged = false; - // flag region as mapped - for (U32 i = 0; i < mMappedVertexRegions.size(); ++i) + if (!gGLManager.mIsApple) { - MappedRegion& region = mMappedVertexRegions[i]; - if (expand_region(region, start, end)) + U32 start = mOffsets[type] + sTypeSize[type] * index; + U32 end = start + sTypeSize[type] * count-1; + + bool flagged = false; + // flag region as mapped + for (U32 i = 0; i < mMappedVertexRegions.size(); ++i) { - flagged = true; - break; + MappedRegion& region = mMappedVertexRegions[i]; + if (expand_region(region, start, end)) + { + flagged = true; + break; + } } - } - if (!flagged) - { - //didn't expand an existing region, make a new one - mMappedVertexRegions.push_back({ start, end }); + if (!flagged) + { + //didn't expand an existing region, make a new one + mMappedVertexRegions.push_back({ start, end }); + } } -#endif return mMappedData+mOffsets[type]+sTypeSize[type]*index; } @@ -1267,28 +1309,29 @@ U8* LLVertexBuffer::mapIndexBuffer(U32 index, S32 count) count = mNumIndices-index; } -#if !LL_DARWIN - U32 start = sizeof(U16) * index; - U32 end = start + sizeof(U16) * count-1; - - bool flagged = false; - // flag region as mapped - for (U32 i = 0; i < mMappedIndexRegions.size(); ++i) + if (!gGLManager.mIsApple) { - MappedRegion& region = mMappedIndexRegions[i]; - if (expand_region(region, start, end)) + U32 start = sizeof(U16) * index; + U32 end = start + sizeof(U16) * count-1; + + bool flagged = false; + // flag region as mapped + for (U32 i = 0; i < mMappedIndexRegions.size(); ++i) { - flagged = true; - break; + MappedRegion& region = mMappedIndexRegions[i]; + if (expand_region(region, start, end)) + { + flagged = true; + break; + } } - } - if (!flagged) - { - //didn't expand an existing region, make a new one - mMappedIndexRegions.push_back({ start, end }); + if (!flagged) + { + //didn't expand an existing region, make a new one + mMappedIndexRegions.push_back({ start, end }); + } } -#endif return mMappedIndexData + sizeof(U16)*index; } @@ -1301,37 +1344,40 @@ U8* LLVertexBuffer::mapIndexBuffer(U32 index, S32 count) // dst -- mMappedData or mMappedIndexData void LLVertexBuffer::flush_vbo(GLenum target, U32 start, U32 end, void* data, U8* dst) { -#if LL_DARWIN - // on OS X, flush_vbo doesn't actually write to the GL buffer, so be sure to call - // _mapBuffer to tag the buffer for flushing to GL - _mapBuffer(); - LL_PROFILE_ZONE_NAMED_CATEGORY_VERTEX("vb memcpy"); - STOP_GLERROR; - // copy into mapped buffer - memcpy(dst+start, data, end-start+1); -#else - llassert(target == GL_ARRAY_BUFFER ? sGLRenderBuffer == mGLBuffer : sGLRenderIndices == mGLIndices); - - // skip mapped data and stream to GPU via glBufferSubData - if (end != 0) + if (gGLManager.mIsApple) { - LL_PROFILE_ZONE_NAMED_CATEGORY_VERTEX("glBufferSubData"); - LL_PROFILE_ZONE_NUM(start); - LL_PROFILE_ZONE_NUM(end); - LL_PROFILE_ZONE_NUM(end-start); - - constexpr U32 block_size = 65536; + // on OS X, flush_vbo doesn't actually write to the GL buffer, so be sure to call + // _mapBuffer to tag the buffer for flushing to GL + _mapBuffer(); + LL_PROFILE_ZONE_NAMED_CATEGORY_VERTEX("vb memcpy"); + STOP_GLERROR; + // copy into mapped buffer + memcpy(dst+start, data, end-start+1); + } + else + { + llassert(target == GL_ARRAY_BUFFER ? sGLRenderBuffer == mGLBuffer : sGLRenderIndices == mGLIndices); - for (U32 i = start; i <= end; i += block_size) + // skip mapped data and stream to GPU via glBufferSubData + if (end != 0) { - //LL_PROFILE_ZONE_NAMED_CATEGORY_VERTEX("glBufferSubData block"); - //LL_PROFILE_GPU_ZONE("glBufferSubData"); - U32 tend = llmin(i + block_size, end); - U32 size = tend - i + 1; - glBufferSubData(target, i, size, (U8*) data + (i-start)); + LL_PROFILE_ZONE_NAMED_CATEGORY_VERTEX("glBufferSubData"); + LL_PROFILE_ZONE_NUM(start); + LL_PROFILE_ZONE_NUM(end); + LL_PROFILE_ZONE_NUM(end-start); + + constexpr U32 block_size = 65536; + + for (U32 i = start; i <= end; i += block_size) + { + //LL_PROFILE_ZONE_NAMED_CATEGORY_VERTEX("glBufferSubData block"); + //LL_PROFILE_GPU_ZONE("glBufferSubData"); + U32 tend = llmin(i + block_size, end); + U32 size = tend - i + 1; + glBufferSubData(target, i, size, (U8*) data + (i-start)); + } } } -#endif } void LLVertexBuffer::unmapBuffer() @@ -1364,114 +1410,116 @@ void LLVertexBuffer::_unmapBuffer() } }; -#if LL_DARWIN - STOP_GLERROR; - if (mMappedData) - { - if (mGLBuffer) - { - glDeleteBuffers(1, &mGLBuffer); - } - mGLBuffer = gen_buffer(); - glBindBuffer(GL_ARRAY_BUFFER, mGLBuffer); - sGLRenderBuffer = mGLBuffer; - glBufferData(GL_ARRAY_BUFFER, mSize, mMappedData, GL_STATIC_DRAW); - } - else if (mGLBuffer != sGLRenderBuffer) - { - glBindBuffer(GL_ARRAY_BUFFER, mGLBuffer); - sGLRenderBuffer = mGLBuffer; - } - STOP_GLERROR; - - if (mMappedIndexData) + if (gGLManager.mIsApple) { - if (mGLIndices) + STOP_GLERROR; + if (mMappedData) { - glDeleteBuffers(1, &mGLIndices); + if (mGLBuffer) + { + delete_buffers(1, &mGLBuffer); + } + mGLBuffer = gen_buffer(); + glBindBuffer(GL_ARRAY_BUFFER, mGLBuffer); + sGLRenderBuffer = mGLBuffer; + glBufferData(GL_ARRAY_BUFFER, mSize, mMappedData, GL_STATIC_DRAW); } - - mGLIndices = gen_buffer(); - glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, mGLIndices); - sGLRenderIndices = mGLIndices; - - glBufferData(GL_ELEMENT_ARRAY_BUFFER, mIndicesSize, mMappedIndexData, GL_STATIC_DRAW); - } - else if (mGLIndices != sGLRenderIndices) - { - glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, mGLIndices); - sGLRenderIndices = mGLIndices; - } - STOP_GLERROR; -#else - - if (!mMappedVertexRegions.empty()) - { - LL_PROFILE_ZONE_NAMED_CATEGORY_VERTEX("unmapBuffer - vertex"); - - if (sGLRenderBuffer != mGLBuffer) + else if (mGLBuffer != sGLRenderBuffer) { glBindBuffer(GL_ARRAY_BUFFER, mGLBuffer); sGLRenderBuffer = mGLBuffer; } + STOP_GLERROR; - U32 start = 0; - U32 end = 0; - - std::sort(mMappedVertexRegions.begin(), mMappedVertexRegions.end(), SortMappedRegion()); - - for (U32 i = 0; i < mMappedVertexRegions.size(); ++i) + if (mMappedIndexData) { - const MappedRegion& region = mMappedVertexRegions[i]; - if (region.mStart == end + 1) - { - end = region.mEnd; - } - else + if (mGLIndices) { - flush_vbo(GL_ARRAY_BUFFER, start, end, (U8*)mMappedData + start, mMappedData); - start = region.mStart; - end = region.mEnd; + delete_buffers(1, &mGLIndices); } - } - - flush_vbo(GL_ARRAY_BUFFER, start, end, (U8*)mMappedData + start, mMappedData); - mMappedVertexRegions.clear(); - } - if (!mMappedIndexRegions.empty()) - { - LL_PROFILE_ZONE_NAMED_CATEGORY_VERTEX("unmapBuffer - index"); + mGLIndices = gen_buffer(); + glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, mGLIndices); + sGLRenderIndices = mGLIndices; - if (mGLIndices != sGLRenderIndices) + glBufferData(GL_ELEMENT_ARRAY_BUFFER, mIndicesSize, mMappedIndexData, GL_STATIC_DRAW); + } + else if (mGLIndices != sGLRenderIndices) { glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, mGLIndices); sGLRenderIndices = mGLIndices; } - U32 start = 0; - U32 end = 0; + STOP_GLERROR; + } + else + { + if (!mMappedVertexRegions.empty()) + { + LL_PROFILE_ZONE_NAMED_CATEGORY_VERTEX("unmapBuffer - vertex"); + + if (sGLRenderBuffer != mGLBuffer) + { + glBindBuffer(GL_ARRAY_BUFFER, mGLBuffer); + sGLRenderBuffer = mGLBuffer; + } - std::sort(mMappedIndexRegions.begin(), mMappedIndexRegions.end(), SortMappedRegion()); + U32 start = 0; + U32 end = 0; - for (U32 i = 0; i < mMappedIndexRegions.size(); ++i) + std::sort(mMappedVertexRegions.begin(), mMappedVertexRegions.end(), SortMappedRegion()); + + for (U32 i = 0; i < mMappedVertexRegions.size(); ++i) + { + const MappedRegion& region = mMappedVertexRegions[i]; + if (region.mStart == end + 1) + { + end = region.mEnd; + } + else + { + flush_vbo(GL_ARRAY_BUFFER, start, end, (U8*)mMappedData + start, mMappedData); + start = region.mStart; + end = region.mEnd; + } + } + + flush_vbo(GL_ARRAY_BUFFER, start, end, (U8*)mMappedData + start, mMappedData); + mMappedVertexRegions.clear(); + } + + if (!mMappedIndexRegions.empty()) { - const MappedRegion& region = mMappedIndexRegions[i]; - if (region.mStart == end + 1) + LL_PROFILE_ZONE_NAMED_CATEGORY_VERTEX("unmapBuffer - index"); + + if (mGLIndices != sGLRenderIndices) { - end = region.mEnd; + glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, mGLIndices); + sGLRenderIndices = mGLIndices; } - else + U32 start = 0; + U32 end = 0; + + std::sort(mMappedIndexRegions.begin(), mMappedIndexRegions.end(), SortMappedRegion()); + + for (U32 i = 0; i < mMappedIndexRegions.size(); ++i) { - flush_vbo(GL_ELEMENT_ARRAY_BUFFER, start, end, (U8*)mMappedIndexData + start, mMappedIndexData); - start = region.mStart; - end = region.mEnd; + const MappedRegion& region = mMappedIndexRegions[i]; + if (region.mStart == end + 1) + { + end = region.mEnd; + } + else + { + flush_vbo(GL_ELEMENT_ARRAY_BUFFER, start, end, (U8*)mMappedIndexData + start, mMappedIndexData); + start = region.mStart; + end = region.mEnd; + } } - } - flush_vbo(GL_ELEMENT_ARRAY_BUFFER, start, end, (U8*)mMappedIndexData + start, mMappedIndexData); - mMappedIndexRegions.clear(); + flush_vbo(GL_ELEMENT_ARRAY_BUFFER, start, end, (U8*)mMappedIndexData + start, mMappedIndexData); + mMappedIndexRegions.clear(); + } } -#endif } //---------------------------------------------------------------------------- -- cgit v1.2.3