2 * Copyright © 2012 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
26 * Support functions for the glthread feature of Mesa.
28 * In multicore systems, many applications end up CPU-bound with about half
29 * their time spent inside their rendering thread and half inside Mesa. To
30 * alleviate this, we put a shim layer in Mesa at the GL dispatch level that
31 * quickly logs the GL commands to a buffer to be processed by a worker
35 #include "main/mtypes.h"
36 #include "main/glthread.h"
37 #include "main/glthread_marshal.h"
38 #include "main/hash.h"
39 #include "util/u_atomic.h"
40 #include "util/u_thread.h"
44 glthread_unmarshal_batch(void *job
, int thread_index
)
46 struct glthread_batch
*batch
= (struct glthread_batch
*)job
;
47 struct gl_context
*ctx
= batch
->ctx
;
49 int used
= batch
->used
;
50 uint8_t *buffer
= batch
->buffer
;
52 _glapi_set_dispatch(ctx
->CurrentServerDispatch
);
55 const struct marshal_cmd_base
*cmd
=
56 (const struct marshal_cmd_base
*)&buffer
[pos
];
58 _mesa_unmarshal_dispatch
[cmd
->cmd_id
](ctx
, cmd
);
67 glthread_thread_initialization(void *job
, int thread_index
)
69 struct gl_context
*ctx
= (struct gl_context
*)job
;
71 ctx
->Driver
.SetBackgroundContext(ctx
, &ctx
->GLThread
.stats
);
72 _glapi_set_context(ctx
);
76 _mesa_glthread_init(struct gl_context
*ctx
)
78 struct glthread_state
*glthread
= &ctx
->GLThread
;
80 assert(!glthread
->enabled
);
82 if (!util_queue_init(&glthread
->queue
, "gl", MARSHAL_MAX_BATCHES
- 2,
87 glthread
->VAOs
= _mesa_NewHashTable();
88 if (!glthread
->VAOs
) {
89 util_queue_destroy(&glthread
->queue
);
92 glthread
->CurrentVAO
= &glthread
->DefaultVAO
;
94 ctx
->MarshalExec
= _mesa_create_marshal_table(ctx
);
95 if (!ctx
->MarshalExec
) {
96 _mesa_DeleteHashTable(glthread
->VAOs
);
97 util_queue_destroy(&glthread
->queue
);
101 for (unsigned i
= 0; i
< MARSHAL_MAX_BATCHES
; i
++) {
102 glthread
->batches
[i
].ctx
= ctx
;
103 util_queue_fence_init(&glthread
->batches
[i
].fence
);
105 glthread
->next_batch
= &glthread
->batches
[glthread
->next
];
107 glthread
->enabled
= true;
108 glthread
->stats
.queue
= &glthread
->queue
;
110 glthread
->SupportsBufferUploads
=
111 ctx
->Const
.BufferCreateMapUnsynchronizedThreadSafe
&&
112 ctx
->Const
.AllowMappedBuffersDuringExecution
;
114 ctx
->CurrentClientDispatch
= ctx
->MarshalExec
;
116 /* Execute the thread initialization function in the thread. */
117 struct util_queue_fence fence
;
118 util_queue_fence_init(&fence
);
119 util_queue_add_job(&glthread
->queue
, ctx
, &fence
,
120 glthread_thread_initialization
, NULL
, 0);
121 util_queue_fence_wait(&fence
);
122 util_queue_fence_destroy(&fence
);
126 free_vao(GLuint key
, void *data
, void *userData
)
132 _mesa_glthread_destroy(struct gl_context
*ctx
)
134 struct glthread_state
*glthread
= &ctx
->GLThread
;
136 if (!glthread
->enabled
)
139 _mesa_glthread_finish(ctx
);
140 util_queue_destroy(&glthread
->queue
);
142 for (unsigned i
= 0; i
< MARSHAL_MAX_BATCHES
; i
++)
143 util_queue_fence_destroy(&glthread
->batches
[i
].fence
);
145 _mesa_HashDeleteAll(glthread
->VAOs
, free_vao
, NULL
);
146 _mesa_DeleteHashTable(glthread
->VAOs
);
148 ctx
->GLThread
.enabled
= false;
150 _mesa_glthread_restore_dispatch(ctx
, "destroy");
154 _mesa_glthread_restore_dispatch(struct gl_context
*ctx
, const char *func
)
156 /* Remove ourselves from the dispatch table except if another ctx/thread
157 * already installed a new dispatch table.
159 * Typically glxMakeCurrent will bind a new context (install new table) then
160 * old context might be deleted.
162 if (_glapi_get_dispatch() == ctx
->MarshalExec
) {
163 ctx
->CurrentClientDispatch
= ctx
->CurrentServerDispatch
;
164 _glapi_set_dispatch(ctx
->CurrentClientDispatch
);
166 printf("glthread disabled: %s\n", func
);
172 _mesa_glthread_disable(struct gl_context
*ctx
, const char *func
)
174 _mesa_glthread_finish_before(ctx
, func
);
175 _mesa_glthread_restore_dispatch(ctx
, func
);
179 _mesa_glthread_flush_batch(struct gl_context
*ctx
)
181 struct glthread_state
*glthread
= &ctx
->GLThread
;
182 if (!glthread
->enabled
)
185 struct glthread_batch
*next
= glthread
->next_batch
;
189 /* Debug: execute the batch immediately from this thread.
191 * Note that glthread_unmarshal_batch() changes the dispatch table so we'll
192 * need to restore it when it returns.
195 glthread_unmarshal_batch(next
, 0);
196 _glapi_set_dispatch(ctx
->CurrentClientDispatch
);
200 p_atomic_add(&glthread
->stats
.num_offloaded_items
, next
->used
);
202 util_queue_add_job(&glthread
->queue
, next
, &next
->fence
,
203 glthread_unmarshal_batch
, NULL
, 0);
204 glthread
->last
= glthread
->next
;
205 glthread
->next
= (glthread
->next
+ 1) % MARSHAL_MAX_BATCHES
;
206 glthread
->next_batch
= &glthread
->batches
[glthread
->next
];
210 * Waits for all pending batches have been unmarshaled.
212 * This can be used by the main thread to synchronize access to the context,
213 * since the worker thread will be idle after this.
216 _mesa_glthread_finish(struct gl_context
*ctx
)
218 struct glthread_state
*glthread
= &ctx
->GLThread
;
219 if (!glthread
->enabled
)
222 /* If this is called from the worker thread, then we've hit a path that
223 * might be called from either the main thread or the worker (such as some
224 * dri interface entrypoints), in which case we don't need to actually
225 * synchronize against ourself.
227 if (u_thread_is_self(glthread
->queue
.threads
[0]))
230 struct glthread_batch
*last
= &glthread
->batches
[glthread
->last
];
231 struct glthread_batch
*next
= glthread
->next_batch
;
234 if (!util_queue_fence_is_signalled(&last
->fence
)) {
235 util_queue_fence_wait(&last
->fence
);
240 p_atomic_add(&glthread
->stats
.num_direct_items
, next
->used
);
242 /* Since glthread_unmarshal_batch changes the dispatch to direct,
243 * restore it after it's done.
245 struct _glapi_table
*dispatch
= _glapi_get_dispatch();
246 glthread_unmarshal_batch(next
, 0);
247 _glapi_set_dispatch(dispatch
);
249 /* It's not a sync because we don't enqueue partial batches, but
250 * it would be a sync if we did. So count it anyway.
256 p_atomic_inc(&glthread
->stats
.num_syncs
);
260 _mesa_glthread_finish_before(struct gl_context
*ctx
, const char *func
)
262 _mesa_glthread_finish(ctx
);
264 /* Uncomment this if you want to know where glthread syncs. */
265 /*printf("fallback to sync: %s\n", func);*/