GPUBackend: Add new GPUBackend object to manage GL object allocations

This just set a global object responsible for allocating new objects
in a thread safe way without needing any GPUContext bound to this
thread.

This also introduce the GLContext which will contain all the GL related
functions for the current context.

# Conflicts:
#	source/blender/gpu/intern/gpu_context.cc
This commit is contained in:
Clément Foucault 2020-08-08 01:18:18 +02:00
parent 1dd7377596
commit cb7ea2ccfb
7 changed files with 347 additions and 170 deletions

View File

@ -52,59 +52,35 @@
#include <mutex>
#include <vector>
static std::vector<GLuint> orphaned_buffer_ids;
static std::vector<GLuint> orphaned_texture_ids;
static std::mutex orphans_mutex;
static std::mutex main_context_mutex;
using namespace blender::gpu;
static thread_local GPUContext *active_ctx = NULL;
static void orphans_add(GPUContext *ctx, std::vector<GLuint> *orphan_list, GLuint id)
{
std::mutex *mutex = (ctx) ? &ctx->orphans_mutex : &orphans_mutex;
/* -------------------------------------------------------------------- */
/** \name GPUContext methods
* \{ */
mutex->lock();
orphan_list->emplace_back(id);
mutex->unlock();
GPUContext::GPUContext()
{
thread_ = pthread_self();
is_active_ = false;
matrix_state = GPU_matrix_state_create();
}
static void orphans_clear(GPUContext *ctx)
GPUContext::~GPUContext()
{
/* need at least an active context */
BLI_assert(ctx);
/* context has been activated by another thread! */
BLI_assert(pthread_equal(pthread_self(), ctx->thread));
ctx->orphans_mutex.lock();
if (!ctx->orphaned_vertarray_ids.empty()) {
uint orphan_len = (uint)ctx->orphaned_vertarray_ids.size();
glDeleteVertexArrays(orphan_len, ctx->orphaned_vertarray_ids.data());
ctx->orphaned_vertarray_ids.clear();
}
if (!ctx->orphaned_framebuffer_ids.empty()) {
uint orphan_len = (uint)ctx->orphaned_framebuffer_ids.size();
glDeleteFramebuffers(orphan_len, ctx->orphaned_framebuffer_ids.data());
ctx->orphaned_framebuffer_ids.clear();
}
ctx->orphans_mutex.unlock();
orphans_mutex.lock();
if (!orphaned_buffer_ids.empty()) {
uint orphan_len = (uint)orphaned_buffer_ids.size();
glDeleteBuffers(orphan_len, orphaned_buffer_ids.data());
orphaned_buffer_ids.clear();
}
if (!orphaned_texture_ids.empty()) {
uint orphan_len = (uint)orphaned_texture_ids.size();
glDeleteTextures(orphan_len, orphaned_texture_ids.data());
orphaned_texture_ids.clear();
}
orphans_mutex.unlock();
GPU_matrix_state_discard(matrix_state);
}
bool GPUContext::is_active_on_thread(void)
{
return (this == active_ctx) && pthread_equal(pthread_self(), thread_);
}
/** \} */
/* -------------------------------------------------------------------- */
GPUContext *GPU_context_create(void *ghost_window)
{
if (gpu_backend_get() == NULL) {
@ -113,15 +89,7 @@ GPUContext *GPU_context_create(void *ghost_window)
}
GPUContext *ctx = gpu_backend_get()->context_alloc(ghost_window);
glGenVertexArrays(1, &ctx->default_vao);
if (ghost_window != NULL) {
ctx->default_framebuffer = GHOST_GetDefaultOpenGLFramebuffer((GHOST_WindowHandle)ghost_window);
}
else {
ctx->default_framebuffer = 0;
}
ctx->matrix_state = GPU_matrix_state_create();
GPU_context_active_set(ctx);
return ctx;
}
@ -129,21 +97,6 @@ GPUContext *GPU_context_create(void *ghost_window)
/* to be called after GPU_context_active_set(ctx_to_destroy) */
void GPU_context_discard(GPUContext *ctx)
{
/* Make sure no other thread has locked it. */
BLI_assert(ctx == active_ctx);
BLI_assert(pthread_equal(pthread_self(), ctx->thread));
BLI_assert(ctx->orphaned_vertarray_ids.empty());
#ifdef DEBUG
/* For now don't allow GPUFrameBuffers to be reuse in another ctx. */
BLI_assert(ctx->framebuffers.empty());
#endif
/* delete remaining vaos */
while (!ctx->batches.empty()) {
/* this removes the array entry */
GPU_batch_vao_cache_clear(*ctx->batches.begin());
}
GPU_matrix_state_discard(ctx->matrix_state);
glDeleteVertexArrays(1, &ctx->default_vao);
delete ctx;
active_ctx = NULL;
}
@ -151,22 +104,15 @@ void GPU_context_discard(GPUContext *ctx)
/* ctx can be NULL */
void GPU_context_active_set(GPUContext *ctx)
{
#if TRUST_NO_ONE
if (active_ctx) {
active_ctx->thread_is_used = false;
}
/* Make sure no other context is already bound to this thread. */
if (ctx) {
/* Make sure no other thread has locked it. */
assert(ctx->thread_is_used == false);
ctx->thread = pthread_self();
ctx->thread_is_used = true;
}
#endif
if (ctx) {
orphans_clear(ctx);
active_ctx->deactivate();
}
active_ctx = ctx;
if (ctx) {
ctx->activate();
}
}
GPUContext *GPU_context_active_get(void)
@ -177,23 +123,18 @@ GPUContext *GPU_context_active_get(void)
GLuint GPU_vao_default(void)
{
BLI_assert(active_ctx); /* need at least an active context */
BLI_assert(pthread_equal(
pthread_self(), active_ctx->thread)); /* context has been activated by another thread! */
return active_ctx->default_vao;
return static_cast<GLContext *>(active_ctx)->default_vao_;
}
GLuint GPU_framebuffer_default(void)
{
BLI_assert(active_ctx); /* need at least an active context */
BLI_assert(pthread_equal(
pthread_self(), active_ctx->thread)); /* context has been activated by another thread! */
return active_ctx->default_framebuffer;
return static_cast<GLContext *>(active_ctx)->default_framebuffer_;
}
GLuint GPU_vao_alloc(void)
{
GLuint new_vao_id = 0;
orphans_clear(active_ctx);
glGenVertexArrays(1, &new_vao_id);
return new_vao_id;
}
@ -201,7 +142,6 @@ GLuint GPU_vao_alloc(void)
GLuint GPU_fbo_alloc(void)
{
GLuint new_fbo_id = 0;
orphans_clear(active_ctx);
glGenFramebuffers(1, &new_fbo_id);
return new_fbo_id;
}
@ -209,7 +149,6 @@ GLuint GPU_fbo_alloc(void)
GLuint GPU_buf_alloc(void)
{
GLuint new_buffer_id = 0;
orphans_clear(active_ctx);
glGenBuffers(1, &new_buffer_id);
return new_buffer_id;
}
@ -217,51 +156,32 @@ GLuint GPU_buf_alloc(void)
GLuint GPU_tex_alloc(void)
{
GLuint new_texture_id = 0;
orphans_clear(active_ctx);
glGenTextures(1, &new_texture_id);
return new_texture_id;
}
void GPU_vao_free(GLuint vao_id, GPUContext *ctx)
{
BLI_assert(ctx);
if (ctx == active_ctx) {
glDeleteVertexArrays(1, &vao_id);
}
else {
orphans_add(ctx, &ctx->orphaned_vertarray_ids, vao_id);
}
static_cast<GLContext *>(ctx)->vao_free(vao_id);
}
void GPU_fbo_free(GLuint fbo_id, GPUContext *ctx)
{
BLI_assert(ctx);
if (ctx == active_ctx) {
glDeleteFramebuffers(1, &fbo_id);
}
else {
orphans_add(ctx, &ctx->orphaned_framebuffer_ids, fbo_id);
}
static_cast<GLContext *>(ctx)->fbo_free(fbo_id);
}
void GPU_buf_free(GLuint buf_id)
{
if (active_ctx) {
glDeleteBuffers(1, &buf_id);
}
else {
orphans_add(NULL, &orphaned_buffer_ids, buf_id);
}
/* TODO avoid using backend */
GPUBackend *backend = gpu_backend_get();
static_cast<GLBackend *>(backend)->buf_free(buf_id);
}
void GPU_tex_free(GLuint tex_id)
{
if (active_ctx) {
glDeleteTextures(1, &tex_id);
}
else {
orphans_add(NULL, &orphaned_texture_ids, tex_id);
}
/* TODO avoid using backend */
GPUBackend *backend = gpu_backend_get();
static_cast<GLBackend *>(backend)->tex_free(tex_id);
}
/* GPUBatch & GPUFrameBuffer contains respectively VAO & FBO indices
@ -271,26 +191,20 @@ void GPU_tex_free(GLuint tex_id)
void gpu_context_add_batch(GPUContext *ctx, GPUBatch *batch)
{
BLI_assert(ctx);
ctx->orphans_mutex.lock();
ctx->batches.emplace(batch);
ctx->orphans_mutex.unlock();
static_cast<GLContext *>(ctx)->batch_register(batch);
}
void gpu_context_remove_batch(GPUContext *ctx, GPUBatch *batch)
{
BLI_assert(ctx);
ctx->orphans_mutex.lock();
ctx->batches.erase(batch);
ctx->orphans_mutex.unlock();
static_cast<GLContext *>(ctx)->batch_unregister(batch);
}
void gpu_context_add_framebuffer(GPUContext *ctx, GPUFrameBuffer *fb)
{
#ifdef DEBUG
BLI_assert(ctx);
ctx->orphans_mutex.lock();
ctx->framebuffers.emplace(fb);
ctx->orphans_mutex.unlock();
static_cast<GLContext *>(ctx)->framebuffer_register(fb);
#else
UNUSED_VARS(ctx, fb);
#endif
@ -300,9 +214,7 @@ void gpu_context_remove_framebuffer(GPUContext *ctx, GPUFrameBuffer *fb)
{
#ifdef DEBUG
BLI_assert(ctx);
ctx->orphans_mutex.lock();
ctx->framebuffers.erase(fb);
ctx->orphans_mutex.unlock();
static_cast<GLContext *>(ctx)->framebuffer_unregister(fb);
#else
UNUSED_VARS(ctx, fb);
#endif
@ -324,6 +236,14 @@ struct GPUMatrixState *gpu_context_active_matrix_state_get()
return active_ctx->matrix_state;
}
/* -------------------------------------------------------------------- */
/** \name Main context global mutex
*
* Used to avoid crash on some old drivers.
* \{ */
static std::mutex main_context_mutex;
void GPU_context_main_lock(void)
{
main_context_mutex.lock();
@ -334,6 +254,8 @@ void GPU_context_main_unlock(void)
main_context_mutex.unlock();
}
/** \} */
/* -------------------------------------------------------------------- */
/** \name Backend selection
* \{ */
@ -358,7 +280,11 @@ void GPU_backend_init(eGPUBackendType backend_type)
void GPU_backend_exit(void)
{
delete g_backend;
if (g_backend) {
/* TODO assert no resource left. Currently UI textures are still not freed in their context
* correctly. */
delete g_backend;
}
}
GPUBackend *gpu_backend_get(void)

View File

@ -25,6 +25,8 @@
#pragma once
#include "MEM_guardedalloc.h"
#include "GPU_context.h"
/* TODO cleanup this ifdef */
@ -37,34 +39,29 @@
# include <vector>
struct GPUFrameBuffer;
struct GPUMatrixState;
struct GPUContext {
GLuint default_vao;
GLuint default_framebuffer;
GPUFrameBuffer *current_fbo;
std::unordered_set<GPUBatch *> batches; /* Batches that have VAOs from this context */
# ifdef DEBUG
std::unordered_set<GPUFrameBuffer *>
framebuffers; /* Framebuffers that have FBO from this context */
# endif
struct GPUMatrixState *matrix_state;
std::vector<GLuint> orphaned_vertarray_ids;
std::vector<GLuint> orphaned_framebuffer_ids;
std::mutex orphans_mutex; /* todo: try spinlock instead */
# if TRUST_NO_ONE
pthread_t thread; /* Thread on which this context is active. */
bool thread_is_used;
# endif
public:
/** State managment */
GPUFrameBuffer *current_fbo = NULL;
GPUMatrixState *matrix_state = NULL;
GPUContext()
{
# if TRUST_NO_ONE
thread_is_used = false;
# endif
current_fbo = 0;
};
protected:
/** Thread on which this context is active. */
pthread_t thread_;
bool is_active_;
virtual ~GPUContext(){};
public:
GPUContext();
virtual ~GPUContext();
virtual void activate(void) = 0;
virtual void deactivate(void) = 0;
bool is_active_on_thread(void);
MEM_CXX_CLASS_ALLOC_FUNCS("GPUContext")
};
#endif

View File

@ -93,8 +93,6 @@ void GPU_exit(void)
gpu_extensions_exit();
gpu_platform_exit(); /* must come last */
GPU_backend_exit();
initialized = false;
}

View File

@ -25,12 +25,33 @@
#include "gpu_backend.hh"
#include "BLI_vector.hh"
#include "gl_context.hh"
namespace blender {
namespace gpu {
class GLBackend : public GPUBackend {
private:
GLSharedOrphanLists shared_orphan_list_;
public:
GPUContext *context_alloc(void *ghost_window)
{
return new GLContext(ghost_window);
return new GLContext(ghost_window, shared_orphan_list_);
};
/* TODO remove */
void buf_free(GLuint buf_id);
void tex_free(GLuint tex_id);
void orphans_add(Vector<GLuint> &orphan_list, std::mutex &list_mutex, unsigned int id)
{
list_mutex.lock();
orphan_list.append(id);
list_mutex.unlock();
}
};
} // namespace gpu
} // namespace blender

View File

@ -30,24 +30,209 @@
#include "gpu_context_private.h"
#include "gl_backend.hh" /* TODO remove */
#include "gl_context.hh"
// TODO(fclem) this requires too much refactor for now.
// using namespace blender::gpu;
using namespace blender::gpu;
/* -------------------------------------------------------------------- */
/** \name Constructor / Destructor
* \{ */
GLContext::GLContext(void *ghost_window) : GPUContext()
GLContext::GLContext(void *ghost_window, GLSharedOrphanLists &shared_orphan_list)
: shared_orphan_list_(shared_orphan_list)
{
default_framebuffer_ = ghost_window ?
GHOST_GetDefaultOpenGLFramebuffer((GHOST_WindowHandle)ghost_window) :
0;
glGenVertexArrays(1, &default_vao_);
float data[4] = {0.0f, 0.0f, 0.0f, 1.0f};
glGenBuffers(1, &default_attr_vbo_);
glBindBuffer(GL_ARRAY_BUFFER, default_attr_vbo_);
glBufferData(GL_ARRAY_BUFFER, sizeof(data), data, GL_STATIC_DRAW);
glBindBuffer(GL_ARRAY_BUFFER, 0);
}
GLContext::~GLContext()
{
BLI_assert(orphaned_framebuffers_.is_empty());
BLI_assert(orphaned_vertarrays_.is_empty());
/* For now don't allow GPUFrameBuffers to be reuse in another context. */
BLI_assert(framebuffers_.is_empty());
/* Delete vaos so the batch can be reused in another context. */
for (GPUBatch *batch : batches_) {
GPU_batch_vao_cache_clear(batch);
}
glDeleteVertexArrays(1, &default_vao_);
glDeleteBuffers(1, &default_attr_vbo_);
}
/** \} */
/* -------------------------------------------------------------------- */
/** \name Activate / Deactivate context
* \{ */
void GLContext::activate(void)
{
/* Make sure no other context is already bound to this thread. */
BLI_assert(is_active_ == false);
is_active_ = true;
thread_ = pthread_self();
/* Clear accumulated orphans. */
orphans_clear();
}
void GLContext::deactivate(void)
{
is_active_ = false;
}
/** \} */
/* -------------------------------------------------------------------- */
/** \name Safe object deletion
*
* GPU objects can be freed when the context is not bound.
* In this case we delay the deletion until the context is bound again.
* \{ */
void GLSharedOrphanLists::orphans_clear(void)
{
/* Check if any context is active on this thread! */
BLI_assert(GPU_context_active_get());
lists_mutex.lock();
if (!buffers.is_empty()) {
glDeleteBuffers((uint)buffers.size(), buffers.data());
buffers.clear();
}
if (!textures.is_empty()) {
glDeleteTextures((uint)textures.size(), textures.data());
textures.clear();
}
lists_mutex.unlock();
};
void GLContext::orphans_clear(void)
{
/* Check if context has been activated by another thread! */
BLI_assert(this->is_active_on_thread());
lists_mutex_.lock();
if (!orphaned_vertarrays_.is_empty()) {
glDeleteVertexArrays((uint)orphaned_vertarrays_.size(), orphaned_vertarrays_.data());
orphaned_vertarrays_.clear();
}
if (!orphaned_framebuffers_.is_empty()) {
glDeleteFramebuffers((uint)orphaned_framebuffers_.size(), orphaned_framebuffers_.data());
orphaned_framebuffers_.clear();
}
lists_mutex_.unlock();
shared_orphan_list_.orphans_clear();
};
void GLContext::orphans_add(Vector<GLuint> &orphan_list, std::mutex &list_mutex, GLuint id)
{
list_mutex.lock();
orphan_list.append(id);
list_mutex.unlock();
}
void GLContext::vao_free(GLuint vao_id)
{
if (this == GPU_context_active_get()) {
glDeleteVertexArrays(1, &vao_id);
}
else {
orphans_add(orphaned_vertarrays_, lists_mutex_, vao_id);
}
}
void GLContext::fbo_free(GLuint fbo_id)
{
if (this == GPU_context_active_get()) {
glDeleteFramebuffers(1, &fbo_id);
}
else {
orphans_add(orphaned_framebuffers_, lists_mutex_, fbo_id);
}
}
void GLBackend::buf_free(GLuint buf_id)
{
/* Any context can free. */
if (GPU_context_active_get()) {
glDeleteBuffers(1, &buf_id);
}
else {
orphans_add(shared_orphan_list_.buffers, shared_orphan_list_.lists_mutex, buf_id);
}
}
void GLBackend::tex_free(GLuint tex_id)
{
/* Any context can free. */
if (GPU_context_active_get()) {
glDeleteTextures(1, &tex_id);
}
else {
orphans_add(shared_orphan_list_.textures, shared_orphan_list_.lists_mutex, tex_id);
}
}
/** \} */
/* -------------------------------------------------------------------- */
/** \name Linked object deletion
*
* These objects contain data that are stored per context. We
* need to do some cleanup if they are used accross context or if context
* is discarded.
* \{ */
void GLContext::batch_register(struct GPUBatch *batch)
{
lists_mutex_.lock();
batches_.add(batch);
lists_mutex_.unlock();
}
void GLContext::batch_unregister(struct GPUBatch *batch)
{
/* vao_cache_clear() can acquire lists_mutex_ so avoid deadlock. */
// reinterpret_cast<GLBatch *>(batch)->vao_cache_clear();
lists_mutex_.lock();
batches_.remove(batch);
lists_mutex_.unlock();
}
void GLContext::framebuffer_register(struct GPUFrameBuffer *fb)
{
#ifdef DEBUG
lists_mutex_.lock();
framebuffers_.add(fb);
lists_mutex_.unlock();
#else
UNUSED_VARS(fb);
#endif
}
void GLContext::framebuffer_unregister(struct GPUFrameBuffer *fb)
{
#ifdef DEBUG
lists_mutex_.lock();
framebuffers_.remove(fb);
lists_mutex_.unlock();
#else
UNUSED_VARS(fb);
#endif
}
/** \} */

View File

@ -25,6 +25,9 @@
#include "gpu_context_private.h"
#include "BLI_vector.hh"
#include "BLI_set.hh"
#include "glew-mx.h"
#include <iostream>
@ -32,19 +35,63 @@
#include <unordered_set>
#include <vector>
// TODO(fclem) this requires too much refactor for now.
// namespace blender {
// namespace gpu {
namespace blender {
namespace gpu {
class GLContext : public GPUContext {
class GLSharedOrphanLists {
public:
GLContext(void *ghost_window);
~GLContext();
/** Mutex for the bellow structures. */
std::mutex lists_mutex;
/** Buffers and textures are shared across context. Any context can free them. */
Vector<GLuint> textures;
Vector<GLuint> buffers;
private:
/** Default framebuffer object for some GL implementation. */
GLuint default_framebuffer_;
public:
void orphans_clear(void);
};
// } // namespace gpu
// } // namespace blender
class GLContext : public GPUContext {
/* TODO(fclem) these needs to become private. */
public:
/** Default VAO for procedural draw calls. */
GLuint default_vao_;
/** Default framebuffer object for some GL implementation. */
GLuint default_framebuffer_;
/** VBO for missing vertex attrib binding. Avoid undefined behavior on some implementation. */
GLuint default_attr_vbo_;
/**
* GPUBatch & GPUFramebuffer have references to the context they are from, in the case the
* context is destroyed, we need to remove any reference to it.
*/
Set<GPUBatch *> batches_;
Set<GPUFrameBuffer *> framebuffers_;
/** Mutex for the bellow structures. */
std::mutex lists_mutex_;
/** VertexArrays and framebuffers are not shared across context. */
Vector<GLuint> orphaned_vertarrays_;
Vector<GLuint> orphaned_framebuffers_;
/** GLBackend onws this data. */
GLSharedOrphanLists &shared_orphan_list_;
public:
GLContext(void *ghost_window, GLSharedOrphanLists &shared_orphan_list);
~GLContext();
void activate(void) override;
void deactivate(void) override;
/* TODO(fclem) these needs to become private. */
public:
void orphans_add(Vector<GLuint> &orphan_list, std::mutex &list_mutex, GLuint id);
void orphans_clear(void);
void vao_free(GLuint vao_id);
void fbo_free(GLuint fbo_id);
void batch_register(struct GPUBatch *batch);
void batch_unregister(struct GPUBatch *batch);
void framebuffer_register(struct GPUFrameBuffer *fb);
void framebuffer_unregister(struct GPUFrameBuffer *fb);
};
} // namespace gpu
} // namespace blender

View File

@ -121,6 +121,7 @@
#include "UI_interface.h"
#include "UI_resources.h"
#include "GPU_context.h"
#include "GPU_init_exit.h"
#include "GPU_material.h"
@ -634,6 +635,8 @@ void WM_exit_ex(bContext *C, const bool do_python)
RNA_exit(); /* should be after BPY_python_end so struct python slots are cleared */
GPU_backend_exit();
wm_ghost_exit();
CTX_free(C);