Merge pull request #1006 from degasus/stream_buffer

GL renderer: Pick the streambuffer from citra and use them.

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -36,30 +36,21 @@ MICROPROFILE_DEFINE(OpenGL_Drawing, "OpenGL", "Drawing", MP_RGB(128, 128, 192));
 MICROPROFILE_DEFINE(OpenGL_Blits, "OpenGL", "Blits", MP_RGB(100, 100, 255));
 MICROPROFILE_DEFINE(OpenGL_CacheManagement, "OpenGL", "Cache Mgmt", MP_RGB(100, 255, 100));
 
-RasterizerOpenGL::RasterizerOpenGL(Core::Frontend::EmuWindow& window) : emu_window{window} {
+RasterizerOpenGL::RasterizerOpenGL(Core::Frontend::EmuWindow& window)
+    : emu_window{window}, stream_buffer(GL_ARRAY_BUFFER, STREAM_BUFFER_SIZE) {
     // Create sampler objects
     for (size_t i = 0; i < texture_samplers.size(); ++i) {
         texture_samplers[i].Create();
         state.texture_units[i].sampler = texture_samplers[i].sampler.handle;
     }
 
-    // Create SSBOs
-    for (size_t stage = 0; stage < ssbos.size(); ++stage) {
-        for (size_t buffer = 0; buffer < ssbos[stage].size(); ++buffer) {
-            ssbos[stage][buffer].Create();
-            state.draw.const_buffers[stage][buffer].ssbo = ssbos[stage][buffer].handle;
-        }
-    }
-
     GLint ext_num;
     glGetIntegerv(GL_NUM_EXTENSIONS, &ext_num);
     for (GLint i = 0; i < ext_num; i++) {
         const std::string_view extension{
             reinterpret_cast<const char*>(glGetStringi(GL_EXTENSIONS, i))};
 
-        if (extension == "GL_ARB_buffer_storage") {
-            has_ARB_buffer_storage = true;
-        } else if (extension == "GL_ARB_direct_state_access") {
+        if (extension == "GL_ARB_direct_state_access") {
             has_ARB_direct_state_access = true;
         } else if (extension == "GL_ARB_separate_shader_objects") {
             has_ARB_separate_shader_objects = true;
@@ -86,47 +77,31 @@ RasterizerOpenGL::RasterizerOpenGL(Core::Frontend::EmuWindow& window) : emu_wind
 
     hw_vao.Create();
 
-    stream_buffer = OGLStreamBuffer::MakeBuffer(has_ARB_buffer_storage, GL_ARRAY_BUFFER);
-    stream_buffer->Create(STREAM_BUFFER_SIZE, STREAM_BUFFER_SIZE / 2);
-    state.draw.vertex_buffer = stream_buffer->GetHandle();
+    state.draw.vertex_buffer = stream_buffer.GetHandle();
 
     shader_program_manager = std::make_unique<GLShader::ProgramManager>();
     state.draw.shader_program = 0;
     state.draw.vertex_array = hw_vao.handle;
     state.Apply();
 
-    glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, stream_buffer->GetHandle());
-
-    for (unsigned index = 0; index < uniform_buffers.size(); ++index) {
-        auto& buffer = uniform_buffers[index];
-        buffer.Create();
-        glBindBuffer(GL_UNIFORM_BUFFER, buffer.handle);
-        glBufferData(GL_UNIFORM_BUFFER, sizeof(GLShader::MaxwellUniformData), nullptr,
-                     GL_STREAM_COPY);
-        glBindBufferBase(GL_UNIFORM_BUFFER, index, buffer.handle);
-    }
+    glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, stream_buffer.GetHandle());
 
     glEnable(GL_BLEND);
 
+    glGetIntegerv(GL_UNIFORM_BUFFER_OFFSET_ALIGNMENT, &uniform_buffer_alignment);
+
     LOG_CRITICAL(Render_OpenGL, "Sync fixed function OpenGL state here!");
 }
 
-RasterizerOpenGL::~RasterizerOpenGL() {
-    if (stream_buffer != nullptr) {
-        state.draw.vertex_buffer = stream_buffer->GetHandle();
-        state.Apply();
-        stream_buffer->Release();
-    }
-}
+RasterizerOpenGL::~RasterizerOpenGL() {}
 
 std::pair<u8*, GLintptr> RasterizerOpenGL::SetupVertexArrays(u8* array_ptr,
                                                              GLintptr buffer_offset) {
     MICROPROFILE_SCOPE(OpenGL_VAO);
     const auto& regs = Core::System::GetInstance().GPU().Maxwell3D().regs;
-    const auto& memory_manager = Core::System::GetInstance().GPU().memory_manager;
 
     state.draw.vertex_array = hw_vao.handle;
-    state.draw.vertex_buffer = stream_buffer->GetHandle();
+    state.draw.vertex_buffer = stream_buffer.GetHandle();
     state.Apply();
 
     // Upload all guest vertex arrays sequentially to our buffer
@@ -141,16 +116,15 @@ std::pair<u8*, GLintptr> RasterizerOpenGL::SetupVertexArrays(u8* array_ptr,
         ASSERT(end > start);
         u64 size = end - start + 1;
 
-        // Copy vertex array data
-        Memory::ReadBlock(*memory_manager->GpuToCpuAddress(start), array_ptr, size);
+        GLintptr vertex_buffer_offset;
+        std::tie(array_ptr, buffer_offset, vertex_buffer_offset) =
+            UploadMemory(array_ptr, buffer_offset, start, size);
 
         // Bind the vertex array to the buffer at the current offset.
-        glBindVertexBuffer(index, stream_buffer->GetHandle(), buffer_offset, vertex_array.stride);
+        glBindVertexBuffer(index, stream_buffer.GetHandle(), vertex_buffer_offset,
+                           vertex_array.stride);
 
         ASSERT_MSG(vertex_array.divisor == 0, "Vertex buffer divisor unimplemented");
-
-        array_ptr += size;
-        buffer_offset += size;
     }
 
     // Use the vertex array as-is, assumes that the data is formatted correctly for OpenGL.
@@ -201,22 +175,12 @@ static GLShader::ProgramCode GetShaderProgramCode(Maxwell::ShaderProgram program
     return program_code;
 }
 
-void RasterizerOpenGL::SetupShaders(u8* buffer_ptr, GLintptr buffer_offset) {
-    // Helper function for uploading uniform data
-    const auto copy_buffer = [&](GLuint handle, GLintptr offset, GLsizeiptr size) {
-        if (has_ARB_direct_state_access) {
-            glCopyNamedBufferSubData(stream_buffer->GetHandle(), handle, offset, 0, size);
-        } else {
-            glBindBuffer(GL_COPY_WRITE_BUFFER, handle);
-            glCopyBufferSubData(GL_ARRAY_BUFFER, GL_COPY_WRITE_BUFFER, offset, 0, size);
-        }
-    };
-
+std::pair<u8*, GLintptr> RasterizerOpenGL::SetupShaders(u8* buffer_ptr, GLintptr buffer_offset) {
     auto& gpu = Core::System::GetInstance().GPU().Maxwell3D();
 
     // Next available bindpoints to use when uploading the const buffers and textures to the GLSL
     // shaders. The constbuffer bindpoint starts after the shader stage configuration bind points.
-    u32 current_constbuffer_bindpoint = static_cast<u32>(uniform_buffers.size());
+    u32 current_constbuffer_bindpoint = Tegra::Engines::Maxwell3D::Regs::MaxShaderStage;
     u32 current_texture_bindpoint = 0;
 
     for (size_t index = 0; index < Maxwell::MaxShaderProgram; ++index) {
@@ -228,22 +192,21 @@ void RasterizerOpenGL::SetupShaders(u8* buffer_ptr, GLintptr buffer_offset) {
             continue;
         }
 
+        std::tie(buffer_ptr, buffer_offset) =
+            AlignBuffer(buffer_ptr, buffer_offset, static_cast<size_t>(uniform_buffer_alignment));
+
         const size_t stage{index == 0 ? 0 : index - 1}; // Stage indices are 0 - 5
 
         GLShader::MaxwellUniformData ubo{};
         ubo.SetFromRegs(gpu.state.shader_stages[stage]);
         std::memcpy(buffer_ptr, &ubo, sizeof(ubo));
 
-        // Flush the buffer so that the GPU can see the data we just wrote.
-        glFlushMappedBufferRange(GL_ARRAY_BUFFER, buffer_offset, sizeof(ubo));
+        // Bind the buffer
+        glBindBufferRange(GL_UNIFORM_BUFFER, stage, stream_buffer.GetHandle(), buffer_offset,
+                          sizeof(ubo));
 
-        // Upload uniform data as one UBO per stage
-        const GLintptr ubo_offset = buffer_offset;
-        copy_buffer(uniform_buffers[stage].handle, ubo_offset,
-                    sizeof(GLShader::MaxwellUniformData));
-
-        buffer_ptr += sizeof(GLShader::MaxwellUniformData);
-        buffer_offset += sizeof(GLShader::MaxwellUniformData);
+        buffer_ptr += sizeof(ubo);
+        buffer_offset += sizeof(ubo);
 
         GLShader::ShaderSetup setup{GetShaderProgramCode(program)};
         GLShader::ShaderEntries shader_resources;
@@ -282,9 +245,9 @@ void RasterizerOpenGL::SetupShaders(u8* buffer_ptr, GLintptr buffer_offset) {
             static_cast<Maxwell::ShaderStage>(stage));
 
         // Configure the const buffers for this shader stage.
-        current_constbuffer_bindpoint =
-            SetupConstBuffers(static_cast<Maxwell::ShaderStage>(stage), gl_stage_program,
-                              current_constbuffer_bindpoint, shader_resources.const_buffer_entries);
+        std::tie(buffer_ptr, buffer_offset, current_constbuffer_bindpoint) = SetupConstBuffers(
+            buffer_ptr, buffer_offset, static_cast<Maxwell::ShaderStage>(stage), gl_stage_program,
+            current_constbuffer_bindpoint, shader_resources.const_buffer_entries);
 
         // Configure the textures for this shader stage.
         current_texture_bindpoint =
@@ -299,6 +262,8 @@ void RasterizerOpenGL::SetupShaders(u8* buffer_ptr, GLintptr buffer_offset) {
     }
 
     shader_program_manager->UseTrivialGeometryShader();
+
+    return {buffer_ptr, buffer_offset};
 }
 
 size_t RasterizerOpenGL::CalculateVertexArraysSize() const {
@@ -432,6 +397,31 @@ void RasterizerOpenGL::Clear() {
     }
 }
 
+std::pair<u8*, GLintptr> RasterizerOpenGL::AlignBuffer(u8* buffer_ptr, GLintptr buffer_offset,
+                                                       size_t alignment) {
+    // Align the offset, not the mapped pointer
+    GLintptr offset_aligned =
+        static_cast<GLintptr>(Common::AlignUp(static_cast<size_t>(buffer_offset), alignment));
+    return {buffer_ptr + (offset_aligned - buffer_offset), offset_aligned};
+}
+
+std::tuple<u8*, GLintptr, GLintptr> RasterizerOpenGL::UploadMemory(u8* buffer_ptr,
+                                                                   GLintptr buffer_offset,
+                                                                   Tegra::GPUVAddr gpu_addr,
+                                                                   size_t size, size_t alignment) {
+    std::tie(buffer_ptr, buffer_offset) = AlignBuffer(buffer_ptr, buffer_offset, alignment);
+    GLintptr uploaded_offset = buffer_offset;
+
+    const auto& memory_manager = Core::System::GetInstance().GPU().memory_manager;
+    const boost::optional<VAddr> cpu_addr{memory_manager->GpuToCpuAddress(gpu_addr)};
+    Memory::ReadBlock(*cpu_addr, buffer_ptr, size);
+
+    buffer_ptr += size;
+    buffer_offset += size;
+
+    return {buffer_ptr, buffer_offset, uploaded_offset};
+}
+
 void RasterizerOpenGL::DrawArrays() {
     if (accelerate_draw == AccelDraw::Disabled)
         return;
@@ -456,7 +446,7 @@ void RasterizerOpenGL::DrawArrays() {
     const u64 index_buffer_size{regs.index_array.count * regs.index_array.FormatSizeInBytes()};
     const unsigned vertex_num{is_indexed ? regs.index_array.count : regs.vertex_buffer.count};
 
-    state.draw.vertex_buffer = stream_buffer->GetHandle();
+    state.draw.vertex_buffer = stream_buffer.GetHandle();
     state.Apply();
 
     size_t buffer_size = CalculateVertexArraysSize();
@@ -466,41 +456,31 @@ void RasterizerOpenGL::DrawArrays() {
     }
 
     // Uniform space for the 5 shader stages
-    buffer_size = Common::AlignUp<size_t>(buffer_size, 4) +
-                  sizeof(GLShader::MaxwellUniformData) * Maxwell::MaxShaderStage;
+    buffer_size =
+        Common::AlignUp<size_t>(buffer_size, 4) +
+        (sizeof(GLShader::MaxwellUniformData) + uniform_buffer_alignment) * Maxwell::MaxShaderStage;
+
+    // Add space for at least 18 constant buffers
+    buffer_size += Maxwell::MaxConstBuffers * (MaxConstbufferSize + uniform_buffer_alignment);
 
     u8* buffer_ptr;
     GLintptr buffer_offset;
-    std::tie(buffer_ptr, buffer_offset) =
-        stream_buffer->Map(static_cast<GLsizeiptr>(buffer_size), 4);
+    std::tie(buffer_ptr, buffer_offset, std::ignore) =
+        stream_buffer.Map(static_cast<GLsizeiptr>(buffer_size), 4);
+    u8* buffer_ptr_base = buffer_ptr;
 
-    u8* offseted_buffer;
-    std::tie(offseted_buffer, buffer_offset) = SetupVertexArrays(buffer_ptr, buffer_offset);
-
-    offseted_buffer =
-        reinterpret_cast<u8*>(Common::AlignUp(reinterpret_cast<size_t>(offseted_buffer), 4));
-    buffer_offset = Common::AlignUp<size_t>(buffer_offset, 4);
+    std::tie(buffer_ptr, buffer_offset) = SetupVertexArrays(buffer_ptr, buffer_offset);
 
     // If indexed mode, copy the index buffer
     GLintptr index_buffer_offset = 0;
     if (is_indexed) {
-        const auto& memory_manager = Core::System::GetInstance().GPU().memory_manager;
-        const boost::optional<VAddr> index_data_addr{
-            memory_manager->GpuToCpuAddress(regs.index_array.StartAddress())};
-        Memory::ReadBlock(*index_data_addr, offseted_buffer, index_buffer_size);
-
-        index_buffer_offset = buffer_offset;
-        offseted_buffer += index_buffer_size;
-        buffer_offset += index_buffer_size;
+        std::tie(buffer_ptr, buffer_offset, index_buffer_offset) = UploadMemory(
+            buffer_ptr, buffer_offset, regs.index_array.StartAddress(), index_buffer_size);
     }
 
-    offseted_buffer =
-        reinterpret_cast<u8*>(Common::AlignUp(reinterpret_cast<size_t>(offseted_buffer), 4));
-    buffer_offset = Common::AlignUp<size_t>(buffer_offset, 4);
+    std::tie(buffer_ptr, buffer_offset) = SetupShaders(buffer_ptr, buffer_offset);
 
-    SetupShaders(offseted_buffer, buffer_offset);
-
-    stream_buffer->Unmap();
+    stream_buffer.Unmap(buffer_ptr - buffer_ptr_base);
 
     shader_program_manager->ApplyTo(state);
     state.Apply();
@@ -647,36 +627,23 @@ void RasterizerOpenGL::SamplerInfo::SyncWithConfig(const Tegra::Texture::TSCEntr
     }
 }
 
-u32 RasterizerOpenGL::SetupConstBuffers(Maxwell::ShaderStage stage, GLuint program,
-                                        u32 current_bindpoint,
-                                        const std::vector<GLShader::ConstBufferEntry>& entries) {
+std::tuple<u8*, GLintptr, u32> RasterizerOpenGL::SetupConstBuffers(
+    u8* buffer_ptr, GLintptr buffer_offset, Maxwell::ShaderStage stage, GLuint program,
+    u32 current_bindpoint, const std::vector<GLShader::ConstBufferEntry>& entries) {
     const auto& gpu = Core::System::GetInstance().GPU();
     const auto& maxwell3d = gpu.Maxwell3D();
 
-    // Reset all buffer draw state for this stage.
-    for (auto& buffer : state.draw.const_buffers[static_cast<size_t>(stage)]) {
-        buffer.bindpoint = 0;
-        buffer.enabled = false;
-    }
-
     // Upload only the enabled buffers from the 16 constbuffers of each shader stage
     const auto& shader_stage = maxwell3d.state.shader_stages[static_cast<size_t>(stage)];
 
     for (u32 bindpoint = 0; bindpoint < entries.size(); ++bindpoint) {
         const auto& used_buffer = entries[bindpoint];
         const auto& buffer = shader_stage.const_buffers[used_buffer.GetIndex()];
-        auto& buffer_draw_state =
-            state.draw.const_buffers[static_cast<size_t>(stage)][used_buffer.GetIndex()];
 
         if (!buffer.enabled) {
             continue;
         }
 
-        buffer_draw_state.enabled = true;
-        buffer_draw_state.bindpoint = current_bindpoint + bindpoint;
-
-        boost::optional<VAddr> addr = gpu.memory_manager->GpuToCpuAddress(buffer.address);
-
         size_t size = 0;
 
         if (used_buffer.IsIndirect()) {
@@ -698,25 +665,26 @@ u32 RasterizerOpenGL::SetupConstBuffers(Maxwell::ShaderStage stage, GLuint progr
         size = Common::AlignUp(size, sizeof(GLvec4));
         ASSERT_MSG(size <= MaxConstbufferSize, "Constbuffer too big");
 
-        std::vector<u8> data(size);
-        Memory::ReadBlock(*addr, data.data(), data.size());
+        GLintptr const_buffer_offset;
+        std::tie(buffer_ptr, buffer_offset, const_buffer_offset) =
+            UploadMemory(buffer_ptr, buffer_offset, buffer.address, size,
+                         static_cast<size_t>(uniform_buffer_alignment));
 
-        glBindBuffer(GL_UNIFORM_BUFFER, buffer_draw_state.ssbo);
-        glBufferData(GL_UNIFORM_BUFFER, data.size(), data.data(), GL_DYNAMIC_DRAW);
-        glBindBuffer(GL_UNIFORM_BUFFER, 0);
+        glBindBufferRange(GL_UNIFORM_BUFFER, current_bindpoint + bindpoint,
+                          stream_buffer.GetHandle(), const_buffer_offset, size);
 
         // Now configure the bindpoint of the buffer inside the shader
         const std::string buffer_name = used_buffer.GetName();
         const GLuint index =
             glGetProgramResourceIndex(program, GL_UNIFORM_BLOCK, buffer_name.c_str());
         if (index != GL_INVALID_INDEX) {
-            glUniformBlockBinding(program, index, buffer_draw_state.bindpoint);
+            glUniformBlockBinding(program, index, current_bindpoint + bindpoint);
         }
     }
 
     state.Apply();
 
-    return current_bindpoint + static_cast<u32>(entries.size());
+    return {buffer_ptr, buffer_offset, current_bindpoint + static_cast<u32>(entries.size())};
 }
 
 u32 RasterizerOpenGL::SetupTextures(Maxwell::ShaderStage stage, GLuint program, u32 current_unit,

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -7,6 +7,7 @@
 #include <array>
 #include <cstddef>
 #include <memory>
+#include <tuple>
 #include <utility>
 #include <vector>
 #include <glad/glad.h>
@@ -100,9 +101,10 @@ private:
      * @param entries Vector describing the buffers that are actually used in the guest shader.
      * @returns The next available bindpoint for use in the next shader stage.
      */
-    u32 SetupConstBuffers(Tegra::Engines::Maxwell3D::Regs::ShaderStage stage, GLuint program,
-                          u32 current_bindpoint,
-                          const std::vector<GLShader::ConstBufferEntry>& entries);
+    std::tuple<u8*, GLintptr, u32> SetupConstBuffers(
+        u8* buffer_ptr, GLintptr buffer_offset, Tegra::Engines::Maxwell3D::Regs::ShaderStage stage,
+        GLuint program, u32 current_bindpoint,
+        const std::vector<GLShader::ConstBufferEntry>& entries);
 
     /*
      * Configures the current textures to use for the draw command.
@@ -139,7 +141,6 @@ private:
     /// Syncs the blend state to match the guest state
     void SyncBlendState();
 
-    bool has_ARB_buffer_storage = false;
     bool has_ARB_direct_state_access = false;
     bool has_ARB_separate_shader_objects = false;
     bool has_ARB_vertex_attrib_binding = false;
@@ -155,22 +156,24 @@ private:
     OGLVertexArray hw_vao;
 
     std::array<SamplerInfo, GLShader::NumTextureSamplers> texture_samplers;
-    std::array<std::array<OGLBuffer, Tegra::Engines::Maxwell3D::Regs::MaxConstBuffers>,
-               Tegra::Engines::Maxwell3D::Regs::MaxShaderStage>
-        ssbos;
 
     static constexpr size_t STREAM_BUFFER_SIZE = 128 * 1024 * 1024;
-    std::unique_ptr<OGLStreamBuffer> stream_buffer;
+    OGLStreamBuffer stream_buffer;
     OGLBuffer uniform_buffer;
     OGLFramebuffer framebuffer;
+    GLint uniform_buffer_alignment;
 
     size_t CalculateVertexArraysSize() const;
 
     std::pair<u8*, GLintptr> SetupVertexArrays(u8* array_ptr, GLintptr buffer_offset);
 
-    std::array<OGLBuffer, Tegra::Engines::Maxwell3D::Regs::MaxShaderStage> uniform_buffers;
+    std::pair<u8*, GLintptr> SetupShaders(u8* buffer_ptr, GLintptr buffer_offset);
 
-    void SetupShaders(u8* buffer_ptr, GLintptr buffer_offset);
+    std::pair<u8*, GLintptr> AlignBuffer(u8* buffer_ptr, GLintptr buffer_offset, size_t alignment);
+
+    std::tuple<u8*, GLintptr, GLintptr> UploadMemory(u8* buffer_ptr, GLintptr buffer_offset,
+                                                     Tegra::GPUVAddr gpu_addr, size_t size,
+                                                     size_t alignment = 4);
 
     enum class AccelDraw { Disabled, Arrays, Indexed };
     AccelDraw accelerate_draw = AccelDraw::Disabled;

src/video_core/renderer_opengl/gl_state.cpp

@@ -203,21 +203,6 @@ void OpenGLState::Apply() const {
         }
     }
 
-    // Constbuffers
-    for (std::size_t stage = 0; stage < draw.const_buffers.size(); ++stage) {
-        for (std::size_t buffer_id = 0; buffer_id < draw.const_buffers[stage].size(); ++buffer_id) {
-            const auto& current = cur_state.draw.const_buffers[stage][buffer_id];
-            const auto& new_state = draw.const_buffers[stage][buffer_id];
-
-            if (current.enabled != new_state.enabled || current.bindpoint != new_state.bindpoint ||
-                current.ssbo != new_state.ssbo) {
-                if (new_state.enabled) {
-                    glBindBufferBase(GL_UNIFORM_BUFFER, new_state.bindpoint, new_state.ssbo);
-                }
-            }
-        }
-    }
-
     // Framebuffer
     if (draw.read_framebuffer != cur_state.draw.read_framebuffer) {
         glBindFramebuffer(GL_READ_FRAMEBUFFER, draw.read_framebuffer);

src/video_core/renderer_opengl/gl_state.h

@@ -119,12 +119,6 @@ public:
         GLuint uniform_buffer;   // GL_UNIFORM_BUFFER_BINDING
         GLuint shader_program;   // GL_CURRENT_PROGRAM
         GLuint program_pipeline; // GL_PROGRAM_PIPELINE_BINDING
-        struct ConstBufferConfig {
-            bool enabled = false;
-            GLuint bindpoint;
-            GLuint ssbo;
-        };
-        std::array<std::array<ConstBufferConfig, Regs::MaxConstBuffers>, 5> const_buffers;
     } draw;
 
     struct {

src/video_core/renderer_opengl/gl_stream_buffer.cpp

@@ -9,174 +9,91 @@
 #include "video_core/renderer_opengl/gl_state.h"
 #include "video_core/renderer_opengl/gl_stream_buffer.h"
 
-class OrphanBuffer : public OGLStreamBuffer {
-public:
-    explicit OrphanBuffer(GLenum target) : OGLStreamBuffer(target) {}
-    ~OrphanBuffer() override;
+OGLStreamBuffer::OGLStreamBuffer(GLenum target, GLsizeiptr size, bool prefer_coherent)
+    : gl_target(target), buffer_size(size) {
+    gl_buffer.Create();
+    glBindBuffer(gl_target, gl_buffer.handle);
 
-private:
-    void Create(size_t size, size_t sync_subdivide) override;
-    void Release() override;
+    GLsizeiptr allocate_size = size;
+    if (target == GL_ARRAY_BUFFER) {
+        // On AMD GPU there is a strange crash in indexed drawing. The crash happens when the buffer
+        // read position is near the end and is an out-of-bound access to the vertex buffer. This is
+        // probably a bug in the driver and is related to the usage of vec3<byte> attributes in the
+        // vertex array. Doubling the allocation size for the vertex buffer seems to avoid the
+        // crash.
+        allocate_size *= 2;
+    }
 
-    std::pair<u8*, GLintptr> Map(size_t size, size_t alignment) override;
-    void Unmap() override;
+    if (GLAD_GL_ARB_buffer_storage) {
+        persistent = true;
+        coherent = prefer_coherent;
+        GLbitfield flags =
+            GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT | (coherent ? GL_MAP_COHERENT_BIT : 0);
+        glBufferStorage(gl_target, allocate_size, nullptr, flags);
+        mapped_ptr = static_cast<u8*>(glMapBufferRange(
+            gl_target, 0, buffer_size, flags | (coherent ? 0 : GL_MAP_FLUSH_EXPLICIT_BIT)));
+    } else {
+        glBufferData(gl_target, allocate_size, nullptr, GL_STREAM_DRAW);
+    }
+}
 
-    std::vector<u8> data;
-};
-
-class StorageBuffer : public OGLStreamBuffer {
-public:
-    explicit StorageBuffer(GLenum target) : OGLStreamBuffer(target) {}
-    ~StorageBuffer() override;
-
-private:
-    void Create(size_t size, size_t sync_subdivide) override;
-    void Release() override;
-
-    std::pair<u8*, GLintptr> Map(size_t size, size_t alignment) override;
-    void Unmap() override;
-
-    struct Fence {
-        OGLSync sync;
-        size_t offset;
-    };
-    std::deque<Fence> head;
-    std::deque<Fence> tail;
-
-    u8* mapped_ptr;
-};
-
-OGLStreamBuffer::OGLStreamBuffer(GLenum target) {
-    gl_target = target;
+OGLStreamBuffer::~OGLStreamBuffer() {
+    if (persistent) {
+        glBindBuffer(gl_target, gl_buffer.handle);
+        glUnmapBuffer(gl_target);
+    }
+    gl_buffer.Release();
 }
 
 GLuint OGLStreamBuffer::GetHandle() const {
     return gl_buffer.handle;
 }
 
-std::unique_ptr<OGLStreamBuffer> OGLStreamBuffer::MakeBuffer(bool storage_buffer, GLenum target) {
-    if (storage_buffer) {
-        return std::make_unique<StorageBuffer>(target);
-    }
-    return std::make_unique<OrphanBuffer>(target);
+GLsizeiptr OGLStreamBuffer::GetSize() const {
+    return buffer_size;
 }
 
-OrphanBuffer::~OrphanBuffer() {
-    Release();
-}
-
-void OrphanBuffer::Create(size_t size, size_t /*sync_subdivide*/) {
-    buffer_pos = 0;
-    buffer_size = size;
-    data.resize(buffer_size);
-
-    if (gl_buffer.handle == 0) {
-        gl_buffer.Create();
-        glBindBuffer(gl_target, gl_buffer.handle);
-    }
-
-    glBufferData(gl_target, static_cast<GLsizeiptr>(buffer_size), nullptr, GL_STREAM_DRAW);
-}
-
-void OrphanBuffer::Release() {
-    gl_buffer.Release();
-}
-
-std::pair<u8*, GLintptr> OrphanBuffer::Map(size_t size, size_t alignment) {
-    buffer_pos = Common::AlignUp(buffer_pos, alignment);
-
-    if (buffer_pos + size > buffer_size) {
-        Create(std::max(buffer_size, size), 0);
-    }
-
-    mapped_size = size;
-    return std::make_pair(&data[buffer_pos], static_cast<GLintptr>(buffer_pos));
-}
-
-void OrphanBuffer::Unmap() {
-    glBufferSubData(gl_target, static_cast<GLintptr>(buffer_pos),
-                    static_cast<GLsizeiptr>(mapped_size), &data[buffer_pos]);
-    buffer_pos += mapped_size;
-}
-
-StorageBuffer::~StorageBuffer() {
-    Release();
-}
-
-void StorageBuffer::Create(size_t size, size_t sync_subdivide) {
-    if (gl_buffer.handle != 0)
-        return;
-
-    buffer_pos = 0;
-    buffer_size = size;
-    buffer_sync_subdivide = std::max<size_t>(sync_subdivide, 1);
-
-    gl_buffer.Create();
-    glBindBuffer(gl_target, gl_buffer.handle);
-
-    glBufferStorage(gl_target, static_cast<GLsizeiptr>(buffer_size), nullptr,
-                    GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
-    mapped_ptr = reinterpret_cast<u8*>(
-        glMapBufferRange(gl_target, 0, static_cast<GLsizeiptr>(buffer_size),
-                         GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT | GL_MAP_FLUSH_EXPLICIT_BIT));
-}
-
-void StorageBuffer::Release() {
-    if (gl_buffer.handle == 0)
-        return;
-
-    glUnmapBuffer(gl_target);
-
-    gl_buffer.Release();
-    head.clear();
-    tail.clear();
-}
-
-std::pair<u8*, GLintptr> StorageBuffer::Map(size_t size, size_t alignment) {
+std::tuple<u8*, GLintptr, bool> OGLStreamBuffer::Map(GLsizeiptr size, GLintptr alignment) {
     ASSERT(size <= buffer_size);
-
-    OGLSync sync;
-
-    buffer_pos = Common::AlignUp(buffer_pos, alignment);
-    size_t effective_offset = Common::AlignDown(buffer_pos, buffer_sync_subdivide);
-
-    if (!head.empty() &&
-        (effective_offset > head.back().offset || buffer_pos + size > buffer_size)) {
-        ASSERT(head.back().sync.handle == 0);
-        head.back().sync.Create();
-    }
-
-    if (buffer_pos + size > buffer_size) {
-        if (!tail.empty()) {
-            std::swap(sync, tail.back().sync);
-            tail.clear();
-        }
-        std::swap(tail, head);
-        buffer_pos = 0;
-        effective_offset = 0;
-    }
-
-    while (!tail.empty() && buffer_pos + size > tail.front().offset) {
-        std::swap(sync, tail.front().sync);
-        tail.pop_front();
-    }
-
-    if (sync.handle != 0) {
-        glClientWaitSync(sync.handle, GL_SYNC_FLUSH_COMMANDS_BIT, GL_TIMEOUT_IGNORED);
-        sync.Release();
-    }
-
-    if (head.empty() || effective_offset > head.back().offset) {
-        head.emplace_back();
-        head.back().offset = effective_offset;
-    }
-
+    ASSERT(alignment <= buffer_size);
     mapped_size = size;
-    return std::make_pair(&mapped_ptr[buffer_pos], static_cast<GLintptr>(buffer_pos));
+
+    if (alignment > 0) {
+        buffer_pos = Common::AlignUp<size_t>(buffer_pos, alignment);
+    }
+
+    bool invalidate = false;
+    if (buffer_pos + size > buffer_size) {
+        buffer_pos = 0;
+        invalidate = true;
+
+        if (persistent) {
+            glUnmapBuffer(gl_target);
+        }
+    }
+
+    if (invalidate | !persistent) {
+        GLbitfield flags = GL_MAP_WRITE_BIT | (persistent ? GL_MAP_PERSISTENT_BIT : 0) |
+                           (coherent ? GL_MAP_COHERENT_BIT : GL_MAP_FLUSH_EXPLICIT_BIT) |
+                           (invalidate ? GL_MAP_INVALIDATE_BUFFER_BIT : GL_MAP_UNSYNCHRONIZED_BIT);
+        mapped_ptr = static_cast<u8*>(
+            glMapBufferRange(gl_target, buffer_pos, buffer_size - buffer_pos, flags));
+        mapped_offset = buffer_pos;
+    }
+
+    return std::make_tuple(mapped_ptr + buffer_pos - mapped_offset, buffer_pos, invalidate);
 }
 
-void StorageBuffer::Unmap() {
-    glFlushMappedBufferRange(gl_target, static_cast<GLintptr>(buffer_pos),
-                             static_cast<GLsizeiptr>(mapped_size));
-    buffer_pos += mapped_size;
+void OGLStreamBuffer::Unmap(GLsizeiptr size) {
+    ASSERT(size <= mapped_size);
+
+    if (!coherent && size > 0) {
+        glFlushMappedBufferRange(gl_target, buffer_pos - mapped_offset, size);
+    }
+
+    if (!persistent) {
+        glUnmapBuffer(gl_target);
+    }
+
+    buffer_pos += size;
 }

src/video_core/renderer_opengl/gl_stream_buffer.h

@@ -2,35 +2,41 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
-#pragma once
-
-#include <memory>
+#include <tuple>
 #include <glad/glad.h>
 #include "common/common_types.h"
 #include "video_core/renderer_opengl/gl_resource_manager.h"
 
 class OGLStreamBuffer : private NonCopyable {
 public:
-    explicit OGLStreamBuffer(GLenum target);
-    virtual ~OGLStreamBuffer() = default;
-
-public:
-    static std::unique_ptr<OGLStreamBuffer> MakeBuffer(bool storage_buffer, GLenum target);
-
-    virtual void Create(size_t size, size_t sync_subdivide) = 0;
-    virtual void Release() {}
+    explicit OGLStreamBuffer(GLenum target, GLsizeiptr size, bool prefer_coherent = false);
+    ~OGLStreamBuffer();
 
     GLuint GetHandle() const;
+    GLsizeiptr GetSize() const;
 
-    virtual std::pair<u8*, GLintptr> Map(size_t size, size_t alignment) = 0;
-    virtual void Unmap() = 0;
+    /*
+     * Allocates a linear chunk of memory in the GPU buffer with at least "size" bytes
+     * and the optional alignment requirement.
+     * If the buffer is full, the whole buffer is reallocated which invalidates old chunks.
+     * The return values are the pointer to the new chunk, the offset within the buffer,
+     * and the invalidation flag for previous chunks.
+     * The actual used size must be specified on unmapping the chunk.
+     */
+    std::tuple<u8*, GLintptr, bool> Map(GLsizeiptr size, GLintptr alignment = 0);
 
-protected:
+    void Unmap(GLsizeiptr size);
+
+private:
     OGLBuffer gl_buffer;
     GLenum gl_target;
 
-    size_t buffer_pos = 0;
-    size_t buffer_size = 0;
-    size_t buffer_sync_subdivide = 0;
-    size_t mapped_size = 0;
+    bool coherent = false;
+    bool persistent = false;
+
+    GLintptr buffer_pos = 0;
+    GLsizeiptr buffer_size = 0;
+    GLintptr mapped_offset = 0;
+    GLsizeiptr mapped_size = 0;
+    u8* mapped_ptr = nullptr;
 };