Files
test/source/blender/gpu/intern/gpu_context.cc
Jacques Lucke b7a1325c3c BLI: use blender::Mutex by default which wraps tbb::mutex
This patch adds a new `BLI_mutex.hh` header which adds `blender::Mutex` as alias
for either `tbb::mutex` or `std::mutex` depending on whether TBB is enabled.

Description copied from the patch:
```
/**
 * blender::Mutex should be used as the default mutex in Blender. It implements a subset of the API
 * of std::mutex but has overall better guaranteed properties. It can be used with RAII helpers
 * like std::lock_guard. However, it is not compatible with e.g. std::condition_variable. So one
 * still has to use std::mutex for that case.
 *
 * The mutex provided by TBB has these properties:
 * - It's as fast as a spin-lock in the non-contended case, i.e. when no other thread is trying to
 *   lock the mutex at the same time.
 * - In the contended case, it spins a couple of times but then blocks to avoid draining system
 *   resources by spinning for a long time.
 * - It's only 1 byte large, compared to e.g. 40 bytes when using the std::mutex of GCC. This makes
 *   it more feasible to have many smaller mutexes which can improve scalability of algorithms
 *   compared to using fewer larger mutexes. Also it just reduces "memory slop" across Blender.
 * - It is *not* a fair mutex, i.e. it's not guaranteed that a thread will ever be able to lock the
 *   mutex when there are always more than one threads that try to lock it. In the majority of
 *   cases, using a fair mutex just causes extra overhead without any benefit. std::mutex is not
 *   guaranteed to be fair either.
 */
 ```

The performance benchmark suggests that the impact is negilible in almost
all cases. The only benchmarks that show interesting behavior are the once
testing foreach zones in Geometry Nodes. These tests are explicitly testing
overhead, which I still have to reduce over time. So it's not unexpected that
changing the mutex has an impact there. What's interesting is that on macos the
performance improves a lot while on linux it gets worse. Since that overhead
should eventually be removed almost entirely, I don't really consider that
blocking.

Links:
* Documentation of different mutex flavors in TBB:
  https://www.intel.com/content/www/us/en/docs/onetbb/developer-guide-api-reference/2021-12/mutex-flavors.html
* Older implementation of a similar mutex by me:
  https://archive.blender.org/developer/differential/0016/0016711/index.html
* Interesting read regarding how a mutex can be this small:
  https://webkit.org/blog/6161/locking-in-webkit/

Pull Request: https://projects.blender.org/blender/blender/pulls/138370
2025-05-07 04:53:16 +02:00

607 lines
14 KiB
C++

/* SPDX-FileCopyrightText: 2016 by Mike Erwin. All rights reserved.
*
* SPDX-License-Identifier: GPL-2.0-or-later */
/** \file
* \ingroup gpu
*
* Manage GL vertex array IDs in a thread-safe way
* Use these instead of glGenBuffers & its friends
* - alloc must be called from a thread that is bound
* to the context that will be used for drawing with
* this VAO.
* - free can be called from any thread
*/
#include "BKE_global.hh"
#include "BLI_assert.h"
#include "BLI_threads.h"
#include "BLI_vector_set.hh"
#include "DNA_userdef_types.h"
#include "GHOST_C-api.h"
#include "GHOST_Types.h"
#include "GPU_context.hh"
#include "GPU_batch.hh"
#include "gpu_backend.hh"
#include "gpu_context_private.hh"
#include "gpu_matrix_private.hh"
#include "gpu_private.hh"
#include "gpu_shader_private.hh"
#ifdef WITH_VULKAN_BACKEND
# include "vk_backend.hh"
#endif
#ifdef WITH_OPENGL_BACKEND
# include "gl_backend.hh"
# include "gl_context.hh"
#endif
#ifdef WITH_METAL_BACKEND
# include "mtl_backend.hh"
#endif
#include "dummy_backend.hh"
#include "draw_debug.hh"
#include <mutex>
using namespace blender::gpu;
static thread_local Context *active_ctx = nullptr;
static blender::Mutex backend_users_mutex;
static int num_backend_users = 0;
static void gpu_backend_create();
static void gpu_backend_discard();
/* -------------------------------------------------------------------- */
/** \name gpu::Context methods
* \{ */
namespace blender::gpu {
int Context::context_counter = 0;
Context::Context()
{
thread_ = pthread_self();
is_active_ = false;
matrix_state = GPU_matrix_state_create();
texture_pool = new TexturePool();
context_id = Context::context_counter;
Context::context_counter++;
}
Context::~Context()
{
/* Derived class should have called free_resources already. */
BLI_assert(front_left == nullptr);
BLI_assert(back_left == nullptr);
BLI_assert(front_right == nullptr);
BLI_assert(back_right == nullptr);
BLI_assert(texture_pool == nullptr);
GPU_matrix_state_discard(matrix_state);
GPU_BATCH_DISCARD_SAFE(procedural_points_batch);
GPU_BATCH_DISCARD_SAFE(procedural_lines_batch);
GPU_BATCH_DISCARD_SAFE(procedural_triangles_batch);
GPU_BATCH_DISCARD_SAFE(procedural_triangle_strips_batch);
GPU_VERTBUF_DISCARD_SAFE(dummy_vbo);
delete state_manager;
delete imm;
}
void Context::free_resources()
{
delete front_left;
delete back_left;
delete front_right;
delete back_right;
front_left = nullptr;
back_left = nullptr;
front_right = nullptr;
back_right = nullptr;
delete texture_pool;
texture_pool = nullptr;
}
bool Context::is_active_on_thread()
{
return (this == active_ctx) && pthread_equal(pthread_self(), thread_);
}
Context *Context::get()
{
return active_ctx;
}
VertBuf *Context::dummy_vbo_get()
{
if (this->dummy_vbo) {
return this->dummy_vbo;
}
/* TODO(fclem): get rid of this dummy VBO. */
GPUVertFormat format = {0};
GPU_vertformat_attr_add(&format, "dummy", GPU_COMP_F32, 1, GPU_FETCH_FLOAT);
this->dummy_vbo = GPU_vertbuf_create_with_format(format);
GPU_vertbuf_data_alloc(*this->dummy_vbo, 1);
return this->dummy_vbo;
}
Batch *Context::procedural_points_batch_get()
{
if (procedural_points_batch) {
return procedural_points_batch;
}
procedural_points_batch = GPU_batch_create(GPU_PRIM_POINTS, dummy_vbo_get(), nullptr);
return procedural_points_batch;
}
Batch *Context::procedural_lines_batch_get()
{
if (procedural_lines_batch) {
return procedural_lines_batch;
}
procedural_lines_batch = GPU_batch_create(GPU_PRIM_LINES, dummy_vbo_get(), nullptr);
return procedural_lines_batch;
}
Batch *Context::procedural_triangles_batch_get()
{
if (procedural_triangles_batch) {
return procedural_triangles_batch;
}
procedural_triangles_batch = GPU_batch_create(GPU_PRIM_TRIS, dummy_vbo_get(), nullptr);
return procedural_triangles_batch;
}
Batch *Context::procedural_triangle_strips_batch_get()
{
if (procedural_triangle_strips_batch) {
return procedural_triangle_strips_batch;
}
procedural_triangle_strips_batch = GPU_batch_create(
GPU_PRIM_TRI_STRIP, dummy_vbo_get(), nullptr);
return procedural_triangle_strips_batch;
}
} // namespace blender::gpu
/** \} */
/* -------------------------------------------------------------------- */
GPUContext *GPU_context_create(void *ghost_window, void *ghost_context)
{
{
std::scoped_lock lock(backend_users_mutex);
if (num_backend_users == 0) {
/* Automatically create backend when first context is created. */
gpu_backend_create();
}
num_backend_users++;
}
Context *ctx = GPUBackend::get()->context_alloc(ghost_window, ghost_context);
GPU_context_active_set(wrap(ctx));
blender::draw::DebugDraw::get().acquire();
return wrap(ctx);
}
void GPU_context_discard(GPUContext *ctx_)
{
Context *ctx = unwrap(ctx_);
BLI_assert(active_ctx == ctx);
blender::draw::DebugDraw::get().release();
GPUBackend *backend = GPUBackend::get();
/* Flush any remaining printf while making sure we are inside render boundaries. */
backend->render_begin();
printf_end(ctx);
backend->render_end();
delete ctx;
active_ctx = nullptr;
{
std::scoped_lock lock(backend_users_mutex);
num_backend_users--;
BLI_assert(num_backend_users >= 0);
if (num_backend_users == 0) {
/* Discard backend when last context is discarded. */
gpu_backend_discard();
}
}
}
void GPU_context_active_set(GPUContext *ctx_)
{
Context *ctx = unwrap(ctx_);
if (active_ctx) {
active_ctx->deactivate();
}
active_ctx = ctx;
if (ctx) {
ctx->activate();
}
}
GPUContext *GPU_context_active_get()
{
return wrap(Context::get());
}
void GPU_context_begin_frame(GPUContext *ctx)
{
blender::gpu::Context *_ctx = unwrap(ctx);
if (_ctx) {
_ctx->begin_frame();
}
}
void GPU_context_end_frame(GPUContext *ctx)
{
blender::gpu::Context *_ctx = unwrap(ctx);
if (_ctx) {
_ctx->end_frame();
}
}
/* -------------------------------------------------------------------- */
/** \name Main context global mutex
*
* Used to avoid crash on some old drivers.
* \{ */
static blender::Mutex main_context_mutex;
void GPU_context_main_lock()
{
main_context_mutex.lock();
}
void GPU_context_main_unlock()
{
main_context_mutex.unlock();
}
/** \} */
/* -------------------------------------------------------------------- */
/** \name GPU Begin/end work blocks
*
* Used to explicitly define a per-frame block within which GPU work will happen.
* Used for global autoreleasepool flushing in Metal
* \{ */
void GPU_render_begin()
{
GPUBackend *backend = GPUBackend::get();
BLI_assert(backend);
/* WORKAROUND: Currently a band-aid for the heist production. Has no side effect for GL backend
* but should be fixed for Metal. */
if (backend) {
backend->render_begin();
printf_begin(active_ctx);
}
}
void GPU_render_end()
{
GPUBackend *backend = GPUBackend::get();
BLI_assert(backend);
if (backend) {
printf_end(active_ctx);
backend->render_end();
}
}
void GPU_render_step(bool force_resource_release)
{
GPUBackend *backend = GPUBackend::get();
BLI_assert(backend);
if (backend) {
printf_end(active_ctx);
backend->render_step(force_resource_release);
printf_begin(active_ctx);
}
}
/** \} */
/* -------------------------------------------------------------------- */
/** \name Backend selection
* \{ */
static eGPUBackendType g_backend_type = GPU_BACKEND_OPENGL;
static std::optional<eGPUBackendType> g_backend_type_override = std::nullopt;
static std::optional<bool> g_backend_type_supported = std::nullopt;
static GPUBackend *g_backend = nullptr;
static GHOST_SystemHandle g_ghost_system = nullptr;
void GPU_backend_ghost_system_set(void *ghost_system_handle)
{
g_ghost_system = reinterpret_cast<GHOST_SystemHandle>(ghost_system_handle);
}
void *GPU_backend_ghost_system_get()
{
return g_ghost_system;
}
void GPU_backend_type_selection_set(const eGPUBackendType backend)
{
g_backend_type = backend;
g_backend_type_supported = std::nullopt;
}
eGPUBackendType GPU_backend_type_selection_get()
{
return g_backend_type;
}
void GPU_backend_type_selection_set_override(const eGPUBackendType backend_type)
{
g_backend_type_override = backend_type;
}
bool GPU_backend_type_selection_is_overridden()
{
return g_backend_type_override.has_value();
}
bool GPU_backend_type_selection_detect()
{
blender::VectorSet<eGPUBackendType> backends_to_check;
if (g_backend_type_override.has_value()) {
backends_to_check.add(*g_backend_type_override);
}
#if defined(WITH_OPENGL_BACKEND)
backends_to_check.add(GPU_BACKEND_OPENGL);
#elif defined(WITH_METAL_BACKEND)
backends_to_check.add(GPU_BACKEND_METAL);
#endif
#if defined(WITH_VULKAN_BACKEND)
backends_to_check.add(GPU_BACKEND_VULKAN);
#endif
for (const eGPUBackendType backend_type : backends_to_check) {
GPU_backend_type_selection_set(backend_type);
if (GPU_backend_supported()) {
return true;
}
G.f |= G_FLAG_GPU_BACKEND_FALLBACK;
}
GPU_backend_type_selection_set(GPU_BACKEND_NONE);
return false;
}
static bool gpu_backend_supported()
{
switch (g_backend_type) {
case GPU_BACKEND_OPENGL:
#ifdef WITH_OPENGL_BACKEND
return true;
#else
return false;
#endif
case GPU_BACKEND_VULKAN:
#ifdef WITH_VULKAN_BACKEND
return VKBackend::is_supported();
#else
return false;
#endif
case GPU_BACKEND_METAL:
#ifdef WITH_METAL_BACKEND
return MTLBackend::metal_is_supported();
#else
return false;
#endif
case GPU_BACKEND_NONE:
return true;
default:
BLI_assert(false && "No backend specified");
return false;
}
}
bool GPU_backend_supported()
{
if (!g_backend_type_supported.has_value()) {
g_backend_type_supported = gpu_backend_supported();
}
return *g_backend_type_supported;
}
static void gpu_backend_create()
{
BLI_assert(g_backend == nullptr);
BLI_assert(GPU_backend_supported());
switch (g_backend_type) {
#ifdef WITH_OPENGL_BACKEND
case GPU_BACKEND_OPENGL:
g_backend = MEM_new<GLBackend>(__func__);
break;
#endif
#ifdef WITH_VULKAN_BACKEND
case GPU_BACKEND_VULKAN:
g_backend = MEM_new<VKBackend>(__func__);
break;
#endif
#ifdef WITH_METAL_BACKEND
case GPU_BACKEND_METAL:
g_backend = MEM_new<MTLBackend>(__func__);
break;
#endif
case GPU_BACKEND_NONE:
g_backend = MEM_new<DummyBackend>(__func__);
break;
default:
BLI_assert(0);
break;
}
}
void gpu_backend_init_resources()
{
BLI_assert(g_backend);
g_backend->init_resources();
}
void gpu_backend_delete_resources()
{
BLI_assert(g_backend);
g_backend->delete_resources();
}
void gpu_backend_discard()
{
/* TODO: assert no resource left. */
MEM_delete(g_backend);
g_backend = nullptr;
}
eGPUBackendType GPU_backend_get_type()
{
#ifdef WITH_OPENGL_BACKEND
if (g_backend && dynamic_cast<GLBackend *>(g_backend) != nullptr) {
return GPU_BACKEND_OPENGL;
}
#endif
#ifdef WITH_METAL_BACKEND
if (g_backend && dynamic_cast<MTLBackend *>(g_backend) != nullptr) {
return GPU_BACKEND_METAL;
}
#endif
#ifdef WITH_VULKAN_BACKEND
if (g_backend && dynamic_cast<VKBackend *>(g_backend) != nullptr) {
return GPU_BACKEND_VULKAN;
}
#endif
return GPU_BACKEND_NONE;
}
GPUBackend *GPUBackend::get()
{
return g_backend;
}
/** \} */
/* -------------------------------------------------------------------- */
/** \name GPUSecondaryContext
* \{ */
static GHOST_TDrawingContextType ghost_context_type()
{
switch (GPU_backend_type_selection_get()) {
#ifdef WITH_OPENGL_BACKEND
case GPU_BACKEND_OPENGL:
return GHOST_kDrawingContextTypeOpenGL;
#endif
#ifdef WITH_VULKAN_BACKEND
case GPU_BACKEND_VULKAN:
return GHOST_kDrawingContextTypeVulkan;
#endif
#ifdef WITH_METAL_BACKEND
case GPU_BACKEND_METAL:
return GHOST_kDrawingContextTypeMetal;
#endif
default:
BLI_assert_unreachable();
return GHOST_kDrawingContextTypeNone;
}
}
GPUSecondaryContext::GPUSecondaryContext()
{
/* Contexts can only be created on the main thread. */
BLI_assert(BLI_thread_is_main());
GHOST_ContextHandle main_thread_ghost_context = GHOST_GetActiveGPUContext();
GPUContext *main_thread_gpu_context = GPU_context_active_get();
/* GPU settings for context creation. */
GHOST_GPUSettings gpu_settings = {0};
gpu_settings.context_type = ghost_context_type();
if (G.debug & G_DEBUG_GPU) {
gpu_settings.flags |= GHOST_gpuDebugContext;
}
gpu_settings.preferred_device.index = U.gpu_preferred_index;
gpu_settings.preferred_device.vendor_id = U.gpu_preferred_vendor_id;
gpu_settings.preferred_device.device_id = U.gpu_preferred_device_id;
/* Grab the system handle. */
GHOST_SystemHandle ghost_system = reinterpret_cast<GHOST_SystemHandle>(
GPU_backend_ghost_system_get());
BLI_assert(ghost_system);
/* Create a Ghost GPU Context using the system handle. */
ghost_context_ = GHOST_CreateGPUContext(ghost_system, gpu_settings);
BLI_assert(ghost_context_);
/* Create a GPU context for the secondary thread to use. */
gpu_context_ = GPU_context_create(nullptr, ghost_context_);
BLI_assert(gpu_context_);
/* Release the Ghost GPU Context from this thread. */
GHOST_TSuccess success = GHOST_ReleaseGPUContext(
reinterpret_cast<GHOST_ContextHandle>(ghost_context_));
BLI_assert(success);
UNUSED_VARS_NDEBUG(success);
/* Restore the main thread contexts.
* (required as the above context creation also makes it active). */
GHOST_ActivateGPUContext(main_thread_ghost_context);
GPU_context_active_set(main_thread_gpu_context);
}
GPUSecondaryContext::~GPUSecondaryContext()
{
/* Contexts should be destructed on the thread they were activated. */
BLI_assert(!BLI_thread_is_main());
GPU_context_discard(gpu_context_);
GHOST_ReleaseGPUContext(reinterpret_cast<GHOST_ContextHandle>(ghost_context_));
GHOST_SystemHandle ghost_system = reinterpret_cast<GHOST_SystemHandle>(
GPU_backend_ghost_system_get());
BLI_assert(ghost_system);
GHOST_DisposeGPUContext(ghost_system, reinterpret_cast<GHOST_ContextHandle>(ghost_context_));
}
void GPUSecondaryContext::activate()
{
/* Contexts need to be activated in the thread they're going to be used. */
BLI_assert(!BLI_thread_is_main());
GHOST_ActivateGPUContext(reinterpret_cast<GHOST_ContextHandle>(ghost_context_));
GPU_context_active_set(gpu_context_);
}
/** \} */