Update the stream_buffer helper from Citra.

Please see https://github.com/citra-emu/citra/pull/3666 for more details.

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -36,7 +36,8 @@ 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();
@@ -57,9 +58,7 @@ RasterizerOpenGL::RasterizerOpenGL(Core::Frontend::EmuWindow& window) : emu_wind
         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;
@@ -86,16 +85,14 @@ 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());
+    glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, stream_buffer.GetHandle());
 
     for (unsigned index = 0; index < uniform_buffers.size(); ++index) {
         auto& buffer = uniform_buffers[index];
@@ -111,13 +108,7 @@ RasterizerOpenGL::RasterizerOpenGL(Core::Frontend::EmuWindow& window) : emu_wind
     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) {
@@ -126,7 +117,7 @@ std::pair<u8*, GLintptr> RasterizerOpenGL::SetupVertexArrays(u8* array_ptr,
     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
@@ -145,7 +136,7 @@ std::pair<u8*, GLintptr> RasterizerOpenGL::SetupVertexArrays(u8* array_ptr,
         Memory::ReadBlock(*memory_manager->GpuToCpuAddress(start), array_ptr, 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(), buffer_offset, vertex_array.stride);
 
         ASSERT_MSG(vertex_array.divisor == 0, "Vertex buffer divisor unimplemented");
 
@@ -205,7 +196,7 @@ 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);
+            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);
@@ -456,7 +447,7 @@ 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();
@@ -471,8 +462,8 @@ void RasterizerOpenGL::DrawArrays() {
 
     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* offseted_buffer;
     std::tie(offseted_buffer, buffer_offset) = SetupVertexArrays(buffer_ptr, buffer_offset);
@@ -500,7 +491,8 @@ void RasterizerOpenGL::DrawArrays() {
 
     SetupShaders(offseted_buffer, buffer_offset);
 
-    stream_buffer->Unmap();
+    // TODO: Don't use buffer_size here, use the updated buffer_offset.
+    stream_buffer.Unmap(buffer_size);
 
     shader_program_manager->ApplyTo(state);
     state.Apply();

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -139,7 +139,6 @@ 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;
@@ -160,7 +159,7 @@ private:
         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;
 

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;
-
-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;
-
-    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;
-}
-
-GLuint OGLStreamBuffer::GetHandle() const {
-    return gl_buffer.handle;
-}
+OGLStreamBuffer::OGLStreamBuffer(GLenum target, GLsizeiptr size, bool prefer_coherent)
+    : gl_target(target), buffer_size(size) {
+    gl_buffer.Create();
+    glBindBuffer(gl_target, gl_buffer.handle);
 
-std::unique_ptr<OGLStreamBuffer> OGLStreamBuffer::MakeBuffer(bool storage_buffer, GLenum target) {
-    if (storage_buffer) {
-        return std::make_unique<StorageBuffer>(target);
+    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;
     }
-    return std::make_unique<OrphanBuffer>(target);
-}
 
-OrphanBuffer::~OrphanBuffer() {
-    Release();
+    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);
+    }
 }
 
-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();
+OGLStreamBuffer::~OGLStreamBuffer() {
+    if (persistent) {
         glBindBuffer(gl_target, gl_buffer.handle);
+        glUnmapBuffer(gl_target);
     }
-
-    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();
+GLuint OGLStreamBuffer::GetHandle() const {
+    return gl_buffer.handle;
 }
 
-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));
+GLsizeiptr OGLStreamBuffer::GetSize() const {
+    return buffer_size;
 }
 
-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);
+    ASSERT(alignment <= buffer_size);
+    mapped_size = 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 (alignment > 0) {
+        buffer_pos = Common::AlignUp<size_t>(buffer_pos, alignment);
     }
 
+    bool invalidate = false;
     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;
-    }
+        invalidate = true;
 
-    while (!tail.empty() && buffer_pos + size > tail.front().offset) {
-        std::swap(sync, tail.front().sync);
-        tail.pop_front();
+        if (persistent) {
+            glUnmapBuffer(gl_target);
+        }
     }
 
-    if (sync.handle != 0) {
-        glClientWaitSync(sync.handle, GL_SYNC_FLUSH_COMMANDS_BIT, GL_TIMEOUT_IGNORED);
-        sync.Release();
+    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;
     }
 
-    if (head.empty() || effective_offset > head.back().offset) {
-        head.emplace_back();
-        head.back().offset = effective_offset;
+    return std::make_tuple(mapped_ptr + buffer_pos - mapped_offset, buffer_pos, invalidate);
+}
+
+void OGLStreamBuffer::Unmap(GLsizeiptr size) {
+    ASSERT(size <= mapped_size);
+
+    if (!coherent && size > 0) {
+        glFlushMappedBufferRange(gl_target, buffer_pos - mapped_offset, size);
     }
 
-    mapped_size = size;
-    return std::make_pair(&mapped_ptr[buffer_pos], static_cast<GLintptr>(buffer_pos));
-}
+    if (!persistent) {
+        glUnmapBuffer(gl_target);
+    }
 
-void StorageBuffer::Unmap() {
-    glFlushMappedBufferRange(gl_target, static_cast<GLintptr>(buffer_pos),
-                             static_cast<GLsizeiptr>(mapped_size));
-    buffer_pos += mapped_size;
+    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;
+
+    /*
+     * 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);
 
-    virtual std::pair<u8*, GLintptr> Map(size_t size, size_t alignment) = 0;
-    virtual void Unmap() = 0;
+    void Unmap(GLsizeiptr size);
 
-protected:
+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;
 };