1 /*
2 * Copyright © 2012 Intel Corporation
3 *
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:
10 *
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
13 * Software.
14 *
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
21 * IN THE SOFTWARE.
22 */
23
24 /** @file glthread.c
25 *
26 * Support functions for the glthread feature of Mesa.
27 *
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
32 * thread.
33 */
34
35 #include "main/mtypes.h"
36 #include "main/glthread.h"
37 #include "main/marshal.h"
38 #include "main/marshal_generated.h"
39 #include "util/u_atomic.h"
40 #include "util/u_thread.h"
41
42
43 static void
glthread_unmarshal_batch(void * job,int thread_index)44 glthread_unmarshal_batch(void *job, int thread_index)
45 {
46 struct glthread_batch *batch = (struct glthread_batch*)job;
47 struct gl_context *ctx = batch->ctx;
48 size_t pos = 0;
49
50 _glapi_set_dispatch(ctx->CurrentServerDispatch);
51
52 while (pos < batch->used)
53 pos += _mesa_unmarshal_dispatch_cmd(ctx, &batch->buffer[pos]);
54
55 assert(pos == batch->used);
56 batch->used = 0;
57 }
58
59 static void
glthread_thread_initialization(void * job,int thread_index)60 glthread_thread_initialization(void *job, int thread_index)
61 {
62 struct gl_context *ctx = (struct gl_context*)job;
63
64 ctx->Driver.SetBackgroundContext(ctx, &ctx->GLThread->stats);
65 _glapi_set_context(ctx);
66 }
67
68 void
_mesa_glthread_init(struct gl_context * ctx)69 _mesa_glthread_init(struct gl_context *ctx)
70 {
71 struct glthread_state *glthread = calloc(1, sizeof(*glthread));
72
73 if (!glthread)
74 return;
75
76 if (!util_queue_init(&glthread->queue, "glthread", MARSHAL_MAX_BATCHES - 2,
77 1, 0)) {
78 free(glthread);
79 return;
80 }
81
82 ctx->MarshalExec = _mesa_create_marshal_table(ctx);
83 if (!ctx->MarshalExec) {
84 util_queue_destroy(&glthread->queue);
85 free(glthread);
86 return;
87 }
88
89 for (unsigned i = 0; i < MARSHAL_MAX_BATCHES; i++) {
90 glthread->batches[i].ctx = ctx;
91 util_queue_fence_init(&glthread->batches[i].fence);
92 }
93
94 glthread->stats.queue = &glthread->queue;
95 ctx->CurrentClientDispatch = ctx->MarshalExec;
96 ctx->GLThread = glthread;
97
98 /* Execute the thread initialization function in the thread. */
99 struct util_queue_fence fence;
100 util_queue_fence_init(&fence);
101 util_queue_add_job(&glthread->queue, ctx, &fence,
102 glthread_thread_initialization, NULL);
103 util_queue_fence_wait(&fence);
104 util_queue_fence_destroy(&fence);
105 }
106
107 void
_mesa_glthread_destroy(struct gl_context * ctx)108 _mesa_glthread_destroy(struct gl_context *ctx)
109 {
110 struct glthread_state *glthread = ctx->GLThread;
111
112 if (!glthread)
113 return;
114
115 _mesa_glthread_finish(ctx);
116 util_queue_destroy(&glthread->queue);
117
118 for (unsigned i = 0; i < MARSHAL_MAX_BATCHES; i++)
119 util_queue_fence_destroy(&glthread->batches[i].fence);
120
121 free(glthread);
122 ctx->GLThread = NULL;
123
124 _mesa_glthread_restore_dispatch(ctx);
125 }
126
127 void
_mesa_glthread_restore_dispatch(struct gl_context * ctx)128 _mesa_glthread_restore_dispatch(struct gl_context *ctx)
129 {
130 /* Remove ourselves from the dispatch table except if another ctx/thread
131 * already installed a new dispatch table.
132 *
133 * Typically glxMakeCurrent will bind a new context (install new table) then
134 * old context might be deleted.
135 */
136 if (_glapi_get_dispatch() == ctx->MarshalExec) {
137 ctx->CurrentClientDispatch = ctx->CurrentServerDispatch;
138 _glapi_set_dispatch(ctx->CurrentClientDispatch);
139 }
140 }
141
142 void
_mesa_glthread_flush_batch(struct gl_context * ctx)143 _mesa_glthread_flush_batch(struct gl_context *ctx)
144 {
145 struct glthread_state *glthread = ctx->GLThread;
146 if (!glthread)
147 return;
148
149 struct glthread_batch *next = &glthread->batches[glthread->next];
150 if (!next->used)
151 return;
152
153 /* Debug: execute the batch immediately from this thread.
154 *
155 * Note that glthread_unmarshal_batch() changes the dispatch table so we'll
156 * need to restore it when it returns.
157 */
158 if (false) {
159 glthread_unmarshal_batch(next, 0);
160 _glapi_set_dispatch(ctx->CurrentClientDispatch);
161 return;
162 }
163
164 p_atomic_add(&glthread->stats.num_offloaded_items, next->used);
165
166 util_queue_add_job(&glthread->queue, next, &next->fence,
167 glthread_unmarshal_batch, NULL);
168 glthread->last = glthread->next;
169 glthread->next = (glthread->next + 1) % MARSHAL_MAX_BATCHES;
170 }
171
172 /**
173 * Waits for all pending batches have been unmarshaled.
174 *
175 * This can be used by the main thread to synchronize access to the context,
176 * since the worker thread will be idle after this.
177 */
178 void
_mesa_glthread_finish(struct gl_context * ctx)179 _mesa_glthread_finish(struct gl_context *ctx)
180 {
181 struct glthread_state *glthread = ctx->GLThread;
182 if (!glthread)
183 return;
184
185 /* If this is called from the worker thread, then we've hit a path that
186 * might be called from either the main thread or the worker (such as some
187 * dri interface entrypoints), in which case we don't need to actually
188 * synchronize against ourself.
189 */
190 if (u_thread_is_self(glthread->queue.threads[0]))
191 return;
192
193 struct glthread_batch *last = &glthread->batches[glthread->last];
194 struct glthread_batch *next = &glthread->batches[glthread->next];
195 bool synced = false;
196
197 if (!util_queue_fence_is_signalled(&last->fence)) {
198 util_queue_fence_wait(&last->fence);
199 synced = true;
200 }
201
202 if (next->used) {
203 p_atomic_add(&glthread->stats.num_direct_items, next->used);
204
205 /* Since glthread_unmarshal_batch changes the dispatch to direct,
206 * restore it after it's done.
207 */
208 struct _glapi_table *dispatch = _glapi_get_dispatch();
209 glthread_unmarshal_batch(next, 0);
210 _glapi_set_dispatch(dispatch);
211
212 /* It's not a sync because we don't enqueue partial batches, but
213 * it would be a sync if we did. So count it anyway.
214 */
215 synced = true;
216 }
217
218 if (synced)
219 p_atomic_inc(&glthread->stats.num_syncs);
220 }
221