gl_rasterizer: Skip VB upload if the state is clean.

src/video_core/command_processor.cpp

@@ -34,6 +34,9 @@ MICROPROFILE_DEFINE(ProcessCommandLists, "GPU", "Execute command buffer", MP_RGB
 void GPU::ProcessCommandLists(const std::vector<CommandListHeader>& commands) {
     MICROPROFILE_SCOPE(ProcessCommandLists);
 
+    // On entering GPU code, assume all memory may be touched by the ARM core.
+    maxwell_3d->dirty_flags.OnMemoryWrite();
+
     auto WriteReg = [this](u32 method, u32 subchannel, u32 value, u32 remaining_params) {
         LOG_TRACE(HW_GPU,
                   "Processing method {:08X} on subchannel {} value "

src/video_core/engines/fermi_2d.cpp

@@ -2,8 +2,10 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#include "core/core.h"
 #include "core/memory.h"
 #include "video_core/engines/fermi_2d.h"
+#include "video_core/engines/maxwell_3d.h"
 #include "video_core/rasterizer_interface.h"
 #include "video_core/textures/decoders.h"
 
@@ -47,6 +49,9 @@ void Fermi2D::HandleSurfaceCopy() {
     u32 dst_bytes_per_pixel = RenderTargetBytesPerPixel(regs.dst.format);
 
     if (!rasterizer.AccelerateSurfaceCopy(regs.src, regs.dst)) {
+        // All copies here update the main memory, so mark all rasterizer states as invalid.
+        Core::System::GetInstance().GPU().Maxwell3D().dirty_flags.OnMemoryWrite();
+
         rasterizer.FlushRegion(source_cpu, src_bytes_per_pixel * regs.src.width * regs.src.height);
         // We have to invalidate the destination region to evict any outdated surfaces from the
         // cache. We do this before actually writing the new data because the destination address

src/video_core/engines/kepler_memory.cpp

@@ -3,8 +3,10 @@
 // Refer to the license.txt file included.
 
 #include "common/logging/log.h"
+#include "core/core.h"
 #include "core/memory.h"
 #include "video_core/engines/kepler_memory.h"
+#include "video_core/engines/maxwell_3d.h"
 #include "video_core/rasterizer_interface.h"
 
 namespace Tegra::Engines {
@@ -47,6 +49,7 @@ void KeplerMemory::ProcessData(u32 data) {
     rasterizer.InvalidateRegion(dest_address, sizeof(u32));
 
     Memory::Write32(dest_address, data);
+    Core::System::GetInstance().GPU().Maxwell3D().dirty_flags.OnMemoryWrite();
 
     state.write_offset++;
 }

src/video_core/engines/maxwell_3d.cpp

@@ -123,10 +123,24 @@ void Maxwell3D::WriteReg(u32 method, u32 value, u32 remaining_params) {
 
     if (regs.reg_array[method] != value) {
         regs.reg_array[method] = value;
+        // Vertex format
         if (method >= MAXWELL3D_REG_INDEX(vertex_attrib_format) &&
             method < MAXWELL3D_REG_INDEX(vertex_attrib_format) + regs.vertex_attrib_format.size()) {
             dirty_flags.vertex_attrib_format = true;
         }
+
+        // Vertex buffer
+        if (method >= MAXWELL3D_REG_INDEX(vertex_array) &&
+            method < MAXWELL3D_REG_INDEX(vertex_array) + 4 * 32) {
+            dirty_flags.vertex_array |= 1u << ((method - MAXWELL3D_REG_INDEX(vertex_array)) >> 2);
+        } else if (method >= MAXWELL3D_REG_INDEX(vertex_array_limit) &&
+                   method < MAXWELL3D_REG_INDEX(vertex_array_limit) + 2 * 32) {
+            dirty_flags.vertex_array |=
+                1u << ((method - MAXWELL3D_REG_INDEX(vertex_array_limit)) >> 1);
+        } else if (method >= MAXWELL3D_REG_INDEX(instanced_arrays) &&
+                   method < MAXWELL3D_REG_INDEX(instanced_arrays) + 32) {
+            dirty_flags.vertex_array |= 1u << (method - MAXWELL3D_REG_INDEX(instanced_arrays));
+        }
     }
 
     switch (method) {
@@ -258,6 +272,7 @@ void Maxwell3D::ProcessQueryGet() {
             query_result.timestamp = CoreTiming::GetTicks();
             Memory::WriteBlock(*address, &query_result, sizeof(query_result));
         }
+        dirty_flags.OnMemoryWrite();
         break;
     }
     default:
@@ -334,6 +349,7 @@ void Maxwell3D::ProcessCBData(u32 value) {
         memory_manager.GpuToCpuAddress(buffer_address + regs.const_buffer.cb_pos);
 
     Memory::Write32(*address, value);
+    dirty_flags.OnMemoryWrite();
 
     // Increment the current buffer position.
     regs.const_buffer.cb_pos = regs.const_buffer.cb_pos + 4;

src/video_core/engines/maxwell_3d.h

@@ -1014,6 +1014,11 @@ public:
 
     struct DirtyFlags {
         bool vertex_attrib_format = true;
+        u32 vertex_array = 0xFFFFFFFF;
+
+        void OnMemoryWrite() {
+            vertex_array = 0xFFFFFFFF;
+        }
     };
 
     DirtyFlags dirty_flags;

src/video_core/engines/maxwell_dma.cpp

@@ -2,7 +2,9 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#include "core/core.h"
 #include "core/memory.h"
+#include "video_core/engines/maxwell_3d.h"
 #include "video_core/engines/maxwell_dma.h"
 #include "video_core/rasterizer_interface.h"
 #include "video_core/textures/decoders.h"
@@ -54,6 +56,9 @@ void MaxwellDMA::HandleCopy() {
         return;
     }
 
+    // All copies here update the main memory, so mark all rasterizer states as invalid.
+    Core::System::GetInstance().GPU().Maxwell3D().dirty_flags.OnMemoryWrite();
+
     if (regs.exec.is_dst_linear && regs.exec.is_src_linear) {
         // When the enable_2d bit is disabled, the copy is performed as if we were copying a 1D
         // buffer of length `x_count`, otherwise we copy a 2D image of dimensions (x_count,

src/video_core/renderer_opengl/gl_buffer_cache.cpp

@@ -76,7 +76,7 @@ std::tuple<u8*, GLintptr> OGLBufferCache::ReserveMemory(std::size_t size, std::s
     return std::make_tuple(uploaded_ptr, uploaded_offset);
 }
 
-void OGLBufferCache::Map(std::size_t max_size) {
+bool OGLBufferCache::Map(std::size_t max_size) {
     bool invalidate;
     std::tie(buffer_ptr, buffer_offset_base, invalidate) =
         stream_buffer.Map(static_cast<GLsizeiptr>(max_size), 4);
@@ -85,6 +85,7 @@ void OGLBufferCache::Map(std::size_t max_size) {
     if (invalidate) {
         InvalidateAll();
     }
+    return invalidate;
 }
 
 void OGLBufferCache::Unmap() {

src/video_core/renderer_opengl/gl_buffer_cache.h

@@ -50,7 +50,7 @@ public:
     /// Reserves memory to be used by host's CPU. Returns mapped address and offset.
     std::tuple<u8*, GLintptr> ReserveMemory(std::size_t size, std::size_t alignment = 4);
 
-    void Map(std::size_t max_size);
+    bool Map(std::size_t max_size);
     void Unmap();
 
     GLuint GetHandle() const;

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -183,15 +183,25 @@ void RasterizerOpenGL::SetupVertexFormat() {
     }
     state.draw.vertex_array = VAO.handle;
     state.ApplyVertexBufferState();
+
+    // Rebinding the VAO invalidates the vertex buffer bindings.
+    gpu.dirty_flags.vertex_array = 0xFFFFFFFF;
 }
 
 void RasterizerOpenGL::SetupVertexBuffer() {
-    MICROPROFILE_SCOPE(OpenGL_VB);
+    auto& gpu = Core::System::GetInstance().GPU().Maxwell3D();
-    const auto& gpu = Core::System::GetInstance().GPU().Maxwell3D();
     const auto& regs = gpu.regs;
 
+    if (!gpu.dirty_flags.vertex_array)
+        return;
+
+    MICROPROFILE_SCOPE(OpenGL_VB);
+
     // Upload all guest vertex arrays sequentially to our buffer
     for (u32 index = 0; index < Maxwell::NumVertexArrays; ++index) {
+        if (~gpu.dirty_flags.vertex_array & (1u << index))
+            continue;
+
         const auto& vertex_array = regs.vertex_array[index];
         if (!vertex_array.IsEnabled())
             continue;
@@ -218,6 +228,8 @@ void RasterizerOpenGL::SetupVertexBuffer() {
 
     // Implicit set by glBindVertexBuffer. Stupid glstate handling...
     state.draw.vertex_buffer = buffer_cache.GetHandle();
+
+    gpu.dirty_flags.vertex_array = 0;
 }
 
 DrawParameters RasterizerOpenGL::SetupDraw() {
@@ -575,7 +587,7 @@ void RasterizerOpenGL::DrawArrays() {
         return;
 
     MICROPROFILE_SCOPE(OpenGL_Drawing);
-    const auto& gpu = Core::System::GetInstance().GPU().Maxwell3D();
+    auto& gpu = Core::System::GetInstance().GPU().Maxwell3D();
     const auto& regs = gpu.regs;
 
     ScopeAcquireGLContext acquire_context{emu_window};
@@ -626,7 +638,11 @@ void RasterizerOpenGL::DrawArrays() {
     // Add space for at least 18 constant buffers
     buffer_size += Maxwell::MaxConstBuffers * (MaxConstbufferSize + uniform_buffer_alignment);
 
-    buffer_cache.Map(buffer_size);
+    bool invalidate = buffer_cache.Map(buffer_size);
+    if (invalidate) {
+        // As all cached buffers are invalidated, we need to recheck their state.
+        gpu.dirty_flags.vertex_attrib_format = 0xFFFFFFFF;
+    }
 
     SetupVertexFormat();
     SetupVertexBuffer();