[PATCH 0/1] MR9811: wined3d/gl: Split UBOs to separate chunks.

[Matteo Bruni (＠Mystral)]

It matters a lot on Nvidia proprietary drivers, where binding to
GL_UNIFORM_BUFFER makes a very significant difference.

Wine-Bug: https://bugs.winehq.org/show_bug.cgi?id=56167

-- 
https://gitlab.winehq.org/wine/wine/-/merge_requests/9811

[PATCH 1/1] wined3d/gl: Split UBOs to separate chunks.

[Matteo Bruni]

From: Matteo Bruni <[e-mail]>

It matters a lot on Nvidia proprietary drivers, where binding to
GL_UNIFORM_BUFFER makes a very significant difference.

Wine-Bug: https://bugs.winehq.org/show_bug.cgi?id=56167
---
 dlls/wined3d/context_gl.c | 16 ++++++++++------
 dlls/wined3d/device.c     | 18 +++++++++++++++---
 dlls/wined3d/wined3d_gl.h |  1 +
 3 files changed, 26 insertions(+), 9 deletions(-)

diff --git a/dlls/wined3d/context_gl.c b/dlls/wined3d/context_gl.c
index ee0335ca3c4..50ef9d25b50 100644
--- a/dlls/wined3d/context_gl.c
+++ b/dlls/wined3d/context_gl.c
@@ -2664,6 +2664,7 @@ GLuint wined3d_context_gl_allocate_vram_chunk_buffer(struct wined3d_context_gl *
 {
     const struct wined3d_gl_info *gl_info = context_gl->gl_info;
     GLbitfield flags;
+    GLenum binding;
     GLuint id = 0;
 
     TRACE("context_gl %p, pool %u, size %Iu.\n", context_gl, pool, size);
@@ -2674,12 +2675,13 @@ GLuint wined3d_context_gl_allocate_vram_chunk_buffer(struct wined3d_context_gl *
         checkGLcall("buffer object creation");
         return 0;
     }
-    wined3d_context_gl_bind_bo(context_gl, GL_PIXEL_UNPACK_BUFFER, id);
+    binding = wined3d_device_gl_get_memory_type_binding(pool);
+    wined3d_context_gl_bind_bo(context_gl, binding, id);
 
     flags = wined3d_device_gl_get_memory_type_flags(pool) | GL_DYNAMIC_STORAGE_BIT;
     if (flags & (GL_MAP_READ_BIT | GL_MAP_WRITE_BIT))
         flags |= GL_MAP_PERSISTENT_BIT | GL_MAP_COHERENT_BIT;
-    GL_EXTCALL(glBufferStorage(GL_PIXEL_UNPACK_BUFFER, size, NULL, flags));
+    GL_EXTCALL(glBufferStorage(binding, size, NULL, flags));
 
     checkGLcall("buffer object creation");
 
@@ -2701,14 +2703,15 @@ static void *wined3d_allocator_chunk_gl_map(struct wined3d_allocator_chunk_gl *c
     if (!chunk_gl->c.map_ptr)
     {
         unsigned int flags = wined3d_device_gl_get_memory_type_flags(chunk_gl->memory_type) & ~GL_CLIENT_STORAGE_BIT;
+        GLenum binding = wined3d_device_gl_get_memory_type_binding(chunk_gl->memory_type);
 
         flags |= GL_MAP_PERSISTENT_BIT | GL_MAP_COHERENT_BIT;
         if (!(flags & GL_MAP_READ_BIT))
             flags |= GL_MAP_UNSYNCHRONIZED_BIT;
         if (flags & GL_MAP_WRITE_BIT)
             flags |= GL_MAP_FLUSH_EXPLICIT_BIT;
-        wined3d_context_gl_bind_bo(context_gl, GL_PIXEL_UNPACK_BUFFER, chunk_gl->gl_buffer);
-        chunk_gl->c.map_ptr = GL_EXTCALL(glMapBufferRange(GL_PIXEL_UNPACK_BUFFER,
+        wined3d_context_gl_bind_bo(context_gl, binding, chunk_gl->gl_buffer);
+        chunk_gl->c.map_ptr = GL_EXTCALL(glMapBufferRange(binding,
                 0, WINED3D_ALLOCATOR_CHUNK_SIZE, flags));
         if (!chunk_gl->c.map_ptr)
         {
@@ -2731,6 +2734,7 @@ static void *wined3d_allocator_chunk_gl_map(struct wined3d_allocator_chunk_gl *c
 static void wined3d_allocator_chunk_gl_unmap(struct wined3d_allocator_chunk_gl *chunk_gl,
         struct wined3d_context_gl *context_gl)
 {
+    GLenum binding = wined3d_device_gl_get_memory_type_binding(chunk_gl->memory_type);
     const struct wined3d_gl_info *gl_info = context_gl->gl_info;
 
     TRACE("chunk_gl %p, context_gl %p.\n", chunk_gl, context_gl);
@@ -2739,8 +2743,8 @@ static void wined3d_allocator_chunk_gl_unmap(struct wined3d_allocator_chunk_gl *
 
     if (!--chunk_gl->c.map_count)
     {
-        wined3d_context_gl_bind_bo(context_gl, GL_PIXEL_UNPACK_BUFFER, chunk_gl->gl_buffer);
-        GL_EXTCALL(glUnmapBuffer(GL_PIXEL_UNPACK_BUFFER));
+        wined3d_context_gl_bind_bo(context_gl, binding, chunk_gl->gl_buffer);
+        GL_EXTCALL(glUnmapBuffer(binding));
         chunk_gl->c.map_ptr = NULL;
 
         adapter_adjust_mapped_memory(context_gl->c.device->adapter, -WINED3D_ALLOCATOR_CHUNK_SIZE);
diff --git a/dlls/wined3d/device.c b/dlls/wined3d/device.c
index 974c195f76b..3feb7e0e4e2 100644
--- a/dlls/wined3d/device.c
+++ b/dlls/wined3d/device.c
@@ -1243,6 +1243,7 @@ static const struct wined3d_allocator_ops wined3d_allocator_gl_ops =
 static const struct
 {
     GLbitfield flags;
+    GLenum binding;
 }
 gl_memory_types[] =
 {
@@ -1255,15 +1256,21 @@ gl_memory_types[] =
     {GL_CLIENT_STORAGE_BIT | GL_MAP_READ_BIT},
     {GL_CLIENT_STORAGE_BIT | GL_MAP_WRITE_BIT},
     {GL_CLIENT_STORAGE_BIT | GL_MAP_READ_BIT | GL_MAP_WRITE_BIT},
+
+    {0, GL_UNIFORM_BUFFER},
+    {GL_MAP_READ_BIT, GL_UNIFORM_BUFFER},
+    {GL_MAP_WRITE_BIT, GL_UNIFORM_BUFFER},
+    {GL_MAP_READ_BIT | GL_MAP_WRITE_BIT, GL_UNIFORM_BUFFER},
 };
 
-static unsigned int wined3d_device_gl_find_memory_type(GLbitfield flags)
+static unsigned int wined3d_device_gl_find_memory_type(GLbitfield flags, GLenum binding)
 {
     unsigned int i;
 
     for (i = 0; i < ARRAY_SIZE(gl_memory_types); ++i)
     {
-        if (gl_memory_types[i].flags == flags)
+        if (gl_memory_types[i].flags == flags
+                && (binding != GL_UNIFORM_BUFFER || gl_memory_types[i].binding == binding))
             return i;
     }
 
@@ -1276,6 +1283,11 @@ GLbitfield wined3d_device_gl_get_memory_type_flags(unsigned int memory_type_idx)
     return gl_memory_types[memory_type_idx].flags;
 }
 
+GLenum wined3d_device_gl_get_memory_type_binding(unsigned int memory_type_idx)
+{
+    return gl_memory_types[memory_type_idx].binding ? gl_memory_types[memory_type_idx].binding : GL_ARRAY_BUFFER;
+}
+
 static struct wined3d_allocator_block *wined3d_device_gl_allocate_memory(struct wined3d_device_gl *device_gl,
         struct wined3d_context_gl *context_gl, unsigned int memory_type, GLsizeiptr size, GLuint *id)
 {
@@ -1336,7 +1348,7 @@ bool wined3d_device_gl_create_bo(struct wined3d_device_gl *device_gl, struct win
         GLsizeiptr size, GLenum binding, GLenum usage, bool coherent, GLbitfield flags, struct wined3d_bo_gl *bo)
 {
     const struct wined3d_gl_info *gl_info = &wined3d_adapter_gl(device_gl->d.adapter)->gl_info;
-    unsigned int memory_type_idx = wined3d_device_gl_find_memory_type(flags);
+    unsigned int memory_type_idx = wined3d_device_gl_find_memory_type(flags, binding);
     struct wined3d_allocator_block *memory = NULL;
     GLsizeiptr buffer_offset = 0;
     GLuint id = 0;
diff --git a/dlls/wined3d/wined3d_gl.h b/dlls/wined3d/wined3d_gl.h
index 2c22c0a9c3c..00c82c7ac0a 100644
--- a/dlls/wined3d/wined3d_gl.h
+++ b/dlls/wined3d/wined3d_gl.h
@@ -920,6 +920,7 @@ void wined3d_device_gl_create_primary_opengl_context_cs(void *object);
 void wined3d_device_gl_delete_opengl_contexts_cs(void *object);
 HDC wined3d_device_gl_get_backup_dc(struct wined3d_device_gl *device_gl);
 GLbitfield wined3d_device_gl_get_memory_type_flags(unsigned int memory_type_idx);
+GLenum wined3d_device_gl_get_memory_type_binding(unsigned int memory_type_idx);
 
 GLbitfield wined3d_resource_gl_map_flags(const struct wined3d_bo_gl *bo, DWORD d3d_flags);
 GLenum wined3d_resource_gl_legacy_map_flags(DWORD d3d_flags);
-- 
GitLab

https://gitlab.winehq.org/wine/wine/-/merge_requests/9811

[Matteo Bruni (＠Mystral)]

I did write a patch on top of this to make it more generic, but I don't think it's really an improvement. Attaching it for reference.

[0001-wined3d-gl-Generalize-the-split-binding-handling.txt](/uploads/7dc8f91763d7577bb529adba88db7ae3/0001-wined3d-gl-Generalize-the-split-binding-handling.txt)

-- 
https://gitlab.winehq.org/wine/wine/-/merge_requests/9811#note_125977

[Elizabeth Figura (＠zfigura)]

I think we need the CLIENT_STORAGE_BIT pools here, don't we? I'm getting an assert in d2d1 without them. Or was the intent to use the non-uniform pools for those?

-- 
https://gitlab.winehq.org/wine/wine/-/merge_requests/9811#note_125983