Blender's GPU Accelerated UI, Decoupled from it's original codebase. Supports Rust, Python, JS, React, C & C++. Twice as good as GPUI
gpu_gl: create the GL context and device on a worker thread
nusaturne8548f804bd5
diff --git a/gpu/intern/gpu_gl.c b/gpu/intern/gpu_gl.cindex 474d1c0..d234c67 100644--- a/gpu/intern/gpu_gl.c+++ b/gpu/intern/gpu_gl.c@@ -14,15 +14,16 @@* platform, else the default display), else a GLX pbuffer.* libGL / libEGL / libX11 are loaded with dlopen on first use, never linked:* without them bui_gpu_create returns NULL and the software raster draws.- * The context is created on the calling thread (no worker thread as on D3D);- * bui_gpu_state is 1 right away, so the raster still switches over at its- * next poll, the same "pending" handoff the D3D device uses.+ * As on D3D, a worker thread creates the context and the device, so the first+ * frame does not wait for the driver. The worker then releases the context,+ * and the thread that draws makes it current at first use.** SPDX-License-Identifier: GPL-2.0-or-later */#if defined(BUI_GPU_GL) && !defined(_WIN32)# include <dlfcn.h>+# include <pthread.h># include <stddef.h># include <stdint.h># include <stdio.h>@@ -916,7 +917,11 @@ typedef struct TriV {enum { API_GLX = 1, API_EGL = 2 };struct BuiGpu {- int state; /* 1 ready (creation is synchronous; failures return NULL) */+ int state; /* 0 pending, 1 ready, -1 failed */+ pthread_t thread;+ int has_thread;+ XDisplay *host_dpy;+ XID host_win;/* context */int api;@@ -1418,15 +1423,23 @@ static int device_init(BuiGpu *g)return 1;}+static void context_release(BuiGpu *g)+{+ if (g->glx && p_glXGetCurrentContext() == g->glx) {+ p_glXMakeContextCurrent(g->xdpy, 0, 0, NULL);+ }+ if (g->egl && p_eglGetCurrentContext() == g->egl) {+ p_eglMakeCurrent(g->edpy, NULL, NULL, NULL);+ }+}+static void context_free(BuiGpu *g){/* destroying the context frees its objects; the window may already be gone,* so nothing is drawn or made current here */if (g->xdpy && (g->glx || g->pbuf)) {XErrHandler old = trap_begin(g->xdpy);- if (g->glx && p_glXGetCurrentContext() == g->glx) {- p_glXMakeContextCurrent(g->xdpy, 0, 0, NULL);- }+ context_release(g);if (g->glx) {p_glXDestroyContext(g->xdpy, g->glx);}@@ -1436,9 +1449,7 @@ static void context_free(BuiGpu *g)trap_end(g->xdpy, old);}if (g->egl) {- if (p_eglGetCurrentContext() == g->egl) {- p_eglMakeCurrent(g->edpy, NULL, NULL, NULL);- }+ context_release(g);if (g->esurf) {p_eglDestroySurface(g->edpy, g->esurf);}@@ -1454,41 +1465,60 @@ static void context_free(BuiGpu *g)g->api = 0;}+/* The GL libraries load on first use, so one worker creates a device at a time. */+static pthread_mutex_t g_create_lock = PTHREAD_MUTEX_INITIALIZER;++static void *device_thread(void *arg)+{+ BuiGpu *g = (BuiGpu *)arg;+ pthread_mutex_lock(&g_create_lock);+ int ok;+ if (g->host_win) { /* a window: GLX on the host's connection */+ ok = glx_create(g, g->host_dpy, g->host_win);+ }+ else { /* offscreen: EGL, else a GLX pbuffer */+ ok = egl_create(g);+ if (!ok) {+ context_free(g);+ ok = glx_create(g, NULL, 0);+ }+ }+ ok = ok && device_init(g);+ context_release(g);+ pthread_mutex_unlock(&g_create_lock);+ __atomic_store_n(&g->state, ok ? 1 : -1, __ATOMIC_RELEASE);+ return NULL;+}+BuiGpu *bui_gpu_create(void *handle){BuiGpu *g = (BuiGpu *)calloc(1, sizeof(BuiGpu));if (!g) {return NULL;}+ if (handle) {+ const BuiGpuX11Window *xw = (const BuiGpuX11Window *)handle;+ g->host_dpy = (XDisplay *)xw->display;+ g->host_win = (XID)xw->window;+ }g->mask_map = (Slot *)calloc(MASK_MAP, sizeof(Slot));g->img_map = (Slot *)calloc(IMG_MAP, sizeof(Slot));- int ok = g->mask_map && g->img_map;- if (ok) {- if (handle) { /* a window: GLX on the host's connection */- const BuiGpuX11Window *xw = (const BuiGpuX11Window *)handle;- ok = xw->window && glx_create(g, xw->display, (XID)xw->window);- }- else { /* offscreen: EGL, else a GLX pbuffer */- ok = egl_create(g);- if (!ok) {- context_free(g);- ok = glx_create(g, NULL, 0);- }- }- }- ok = ok && device_init(g);- if (!ok) {+ g->has_thread = (!handle || g->host_win) && g->mask_map && g->img_map &&+ pthread_create(&g->thread, NULL, device_thread, g) == 0;+ if (!g->has_thread) {bui_gpu_free(g);return NULL;}- g->state = 1;return g;}int bui_gpu_state(BuiGpu *g, int wait){- (void)wait; /* created synchronously: never pending */- return g->state;+ if (g->has_thread && (wait || __atomic_load_n(&g->state, __ATOMIC_ACQUIRE) != 0)) {+ pthread_join(g->thread, NULL);+ g->has_thread = 0;+ }+ return __atomic_load_n(&g->state, __ATOMIC_ACQUIRE);}void bui_gpu_free(BuiGpu *g)@@ -1496,6 +1526,9 @@ void bui_gpu_free(BuiGpu *g)if (!g) {return;}+ if (g->has_thread) {+ pthread_join(g->thread, NULL);+ }context_free(g); /* deletes the textures and buffers with it */free(g->dtex);free(g->inst);