Merge pull request #3666 from degasus/vertex_streaming

renderer_opengl: Rewrite stream buffer.

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -30,7 +30,8 @@ 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() : shader_dirty(true) {
+RasterizerOpenGL::RasterizerOpenGL()
+    : shader_dirty(true), vertex_buffer(GL_ARRAY_BUFFER, VERTEX_BUFFER_SIZE) {
     // Clipping plane 0 is always enabled for PICA fixed clip plane z <= 0
     state.clip_distance[0] = true;
 
@@ -46,13 +47,11 @@ RasterizerOpenGL::RasterizerOpenGL() : shader_dirty(true) {
     texture_cube.Create();
 
     // Generate VBO, VAO and UBO
-    vertex_buffer = OGLStreamBuffer::MakeBuffer(GLAD_GL_ARB_buffer_storage, GL_ARRAY_BUFFER);
-    vertex_buffer->Create(VERTEX_BUFFER_SIZE, VERTEX_BUFFER_SIZE / 2);
     vertex_array.Create();
     uniform_buffer.Create();
 
     state.draw.vertex_array = vertex_array.handle;
-    state.draw.vertex_buffer = vertex_buffer->GetHandle();
+    state.draw.vertex_buffer = vertex_buffer.GetHandle();
     state.draw.uniform_buffer = uniform_buffer.handle;
     state.Apply();
 
@@ -499,14 +498,16 @@ void RasterizerOpenGL::DrawTriangles() {
     state.Apply();
 
     // Draw the vertex batch
-    size_t max_vertices = 3 * (VERTEX_BUFFER_SIZE / (3 * sizeof(HardwareVertex)));
+    size_t max_vertices = 3 * (vertex_buffer.GetSize() / (3 * sizeof(HardwareVertex)));
     for (size_t base_vertex = 0; base_vertex < vertex_batch.size(); base_vertex += max_vertices) {
         size_t vertices = std::min(max_vertices, vertex_batch.size() - base_vertex);
         size_t vertex_size = vertices * sizeof(HardwareVertex);
-        auto map = vertex_buffer->Map(vertex_size, 1);
+        u8* vbo;
-        memcpy(map.first, vertex_batch.data() + base_vertex, vertex_size);
+        GLintptr offset;
-        vertex_buffer->Unmap();
+        std::tie(vbo, offset, std::ignore) = vertex_buffer.Map(vertex_size, sizeof(HardwareVertex));
-        glDrawArrays(GL_TRIANGLES, map.second / sizeof(HardwareVertex), (GLsizei)vertices);
+        memcpy(vbo, vertex_batch.data() + base_vertex, vertex_size);
+        vertex_buffer.Unmap(vertex_size);
+        glDrawArrays(GL_TRIANGLES, offset / sizeof(HardwareVertex), (GLsizei)vertices);
     }
 
     // Disable scissor test

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -239,8 +239,8 @@ private:
 
     std::array<SamplerInfo, 3> texture_samplers;
     OGLVertexArray vertex_array;
-    static constexpr size_t VERTEX_BUFFER_SIZE = 128 * 1024 * 1024;
+    static constexpr size_t VERTEX_BUFFER_SIZE = 32 * 1024 * 1024;
-    std::unique_ptr<OGLStreamBuffer> vertex_buffer;
+    OGLStreamBuffer vertex_buffer;
     OGLBuffer uniform_buffer;
     OGLFramebuffer framebuffer;
 

src/video_core/renderer_opengl/gl_resource_manager.h

@@ -225,39 +225,6 @@ public:
     GLuint handle = 0;
 };
 
-class OGLSync : private NonCopyable {
-public:
-    OGLSync() = default;
-
-    OGLSync(OGLSync&& o) : handle(std::exchange(o.handle, nullptr)) {}
-
-    ~OGLSync() {
-        Release();
-    }
-    OGLSync& operator=(OGLSync&& o) {
-        Release();
-        handle = std::exchange(o.handle, nullptr);
-        return *this;
-    }
-
-    /// Creates a new internal OpenGL resource and stores the handle
-    void Create() {
-        if (handle != 0)
-            return;
-        handle = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
-    }
-
-    /// Deletes the internal OpenGL resource
-    void Release() {
-        if (handle == 0)
-            return;
-        glDeleteSync(handle);
-        handle = 0;
-    }
-
-    GLsync handle = 0;
-};
-
 class OGLVertexArray : private NonCopyable {
 public:
     OGLVertexArray() = default;

src/video_core/renderer_opengl/gl_stream_buffer.cpp

@@ -9,174 +9,81 @@
 #include "video_core/renderer_opengl/gl_state.h"
 #include "video_core/renderer_opengl/gl_stream_buffer.h"
 
-class OrphanBuffer : public OGLStreamBuffer {
+OGLStreamBuffer::OGLStreamBuffer(GLenum target, GLsizeiptr size, bool prefer_coherent)
-public:
+    : gl_target(target), buffer_size(size) {
-    explicit OrphanBuffer(GLenum target) : OGLStreamBuffer(target) {}
+    gl_buffer.Create();
-    ~OrphanBuffer() override;
+    glBindBuffer(gl_target, gl_buffer.handle);
 
-private:
+    if (GLAD_GL_ARB_buffer_storage) {
-    void Create(size_t size, size_t sync_subdivide) override;
+        persistent = true;
-    void Release() override;
+        coherent = prefer_coherent;
+        GLbitfield flags =
+            GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT | (coherent ? GL_MAP_COHERENT_BIT : 0);
+        glBufferStorage(gl_target, buffer_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, buffer_size, nullptr, GL_STREAM_DRAW);
+    }
+}
 
-    std::pair<u8*, GLintptr> Map(size_t size, size_t alignment) override;
+OGLStreamBuffer::~OGLStreamBuffer() {
-    void Unmap() override;
+    if (persistent) {
-
+        glBindBuffer(gl_target, gl_buffer.handle);
-    std::vector<u8> data;
+        glUnmapBuffer(gl_target);
-};
+    }
-
+    gl_buffer.Release();
-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;
 }
 
-std::unique_ptr<OGLStreamBuffer> OGLStreamBuffer::MakeBuffer(bool storage_buffer, GLenum target) {
+GLsizeiptr OGLStreamBuffer::GetSize() const {
-    if (storage_buffer) {
+    return buffer_size;
-        return std::make_unique<StorageBuffer>(target);
-    }
-    return std::make_unique<OrphanBuffer>(target);
 }
 
-OrphanBuffer::~OrphanBuffer() {
+std::tuple<u8*, GLintptr, bool> OGLStreamBuffer::Map(GLsizeiptr size, GLintptr alignment) {
-    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) {
     ASSERT(size <= buffer_size);
-
+    ASSERT(alignment <= 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;
-    }
-
     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);
     }
 
-void StorageBuffer::Unmap() {
+    bool invalidate = false;
-    glFlushMappedBufferRange(gl_target, static_cast<GLintptr>(buffer_pos),
+    if (buffer_pos + size > buffer_size) {
-                             static_cast<GLsizeiptr>(mapped_size));
+        buffer_pos = 0;
-    buffer_pos += mapped_size;
+        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 OGLStreamBuffer::Unmap(GLsizeiptr size) {
+    ASSERT(size <= mapped_size);
+
+    if (!coherent) {
+        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

@@ -2,33 +2,41 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
-#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);
+    explicit OGLStreamBuffer(GLenum target, GLsizeiptr size, bool prefer_coherent = false);
-    virtual ~OGLStreamBuffer() = default;
+    ~OGLStreamBuffer();
-
-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() {}
 
     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;
+    bool coherent = false;
-    size_t buffer_size = 0;
+    bool persistent = false;
-    size_t buffer_sync_subdivide = 0;
+
-    size_t mapped_size = 0;
+    GLintptr buffer_pos = 0;
+    GLsizeiptr buffer_size = 0;
+    GLintptr mapped_offset = 0;
+    GLsizeiptr mapped_size = 0;
+    u8* mapped_ptr = nullptr;
 };