src/mesa/main/glthread.c

   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/glthread_marshal.h"
  38 #include "main/hash.h"
  39 #include "util/u_atomic.h"
  40 #include "util/u_thread.h"
  41
  42
  43 static void
  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    int pos = 0;
  49    int used = batch->used;
  50    uint8_t *buffer = batch->buffer;
  51
  52    _glapi_set_dispatch(ctx->CurrentServerDispatch);
  53
  54    while (pos < used) {
  55       const struct marshal_cmd_base *cmd =
  56          (const struct marshal_cmd_base *)&buffer[pos];
  57
  58       _mesa_unmarshal_dispatch[cmd->cmd_id](ctx, cmd);
  59       pos += cmd->cmd_size;
  60    }
  61
  62    assert(pos == used);
  63    batch->used = 0;
  64 }
  65
  66 static void
  67 glthread_thread_initialization(void *job, int thread_index)
  68 {
  69    struct gl_context *ctx = (struct gl_context*)job;
  70
  71    ctx->Driver.SetBackgroundContext(ctx, &ctx->GLThread->stats);
  72    _glapi_set_context(ctx);
  73 }
  74
  75 void
  76 _mesa_glthread_init(struct gl_context *ctx)
  77 {
  78    struct glthread_state *glthread = calloc(1, sizeof(*glthread));
  79
  80    if (!glthread)
  81       return;
  82
  83    if (!util_queue_init(&glthread->queue, "gl", MARSHAL_MAX_BATCHES - 2,
  84                         1, 0)) {
  85       free(glthread);
  86       return;
  87    }
  88
  89    glthread->VAOs = _mesa_NewHashTable();
  90    if (!glthread->VAOs) {
  91       util_queue_destroy(&glthread->queue);
  92       free(glthread);
  93       return;
  94    }
  95    glthread->CurrentVAO = &glthread->DefaultVAO;
  96
  97    ctx->MarshalExec = _mesa_create_marshal_table(ctx);
  98    if (!ctx->MarshalExec) {
  99       _mesa_DeleteHashTable(glthread->VAOs);
 100       util_queue_destroy(&glthread->queue);
 101       free(glthread);
 102       return;
 103    }
 104
 105    for (unsigned i = 0; i < MARSHAL_MAX_BATCHES; i++) {
 106       glthread->batches[i].ctx = ctx;
 107       util_queue_fence_init(&glthread->batches[i].fence);
 108    }
 109
 110    glthread->stats.queue = &glthread->queue;
 111    ctx->CurrentClientDispatch = ctx->MarshalExec;
 112    ctx->GLThread = glthread;
 113
 114    /* Execute the thread initialization function in the thread. */
 115    struct util_queue_fence fence;
 116    util_queue_fence_init(&fence);
 117    util_queue_add_job(&glthread->queue, ctx, &fence,
 118                       glthread_thread_initialization, NULL, 0);
 119    util_queue_fence_wait(&fence);
 120    util_queue_fence_destroy(&fence);
 121 }
 122
 123 static void
 124 free_vao(GLuint key, void *data, void *userData)
 125 {
 126    free(data);
 127 }
 128
 129 void
 130 _mesa_glthread_destroy(struct gl_context *ctx)
 131 {
 132    struct glthread_state *glthread = ctx->GLThread;
 133
 134    if (!glthread)
 135       return;
 136
 137    _mesa_glthread_finish(ctx);
 138    util_queue_destroy(&glthread->queue);
 139
 140    for (unsigned i = 0; i < MARSHAL_MAX_BATCHES; i++)
 141       util_queue_fence_destroy(&glthread->batches[i].fence);
 142
 143    _mesa_HashDeleteAll(glthread->VAOs, free_vao, NULL);
 144    _mesa_DeleteHashTable(glthread->VAOs);
 145
 146    free(glthread);
 147    ctx->GLThread = NULL;
 148
 149    _mesa_glthread_restore_dispatch(ctx, "destroy");
 150 }
 151
 152 void
 153 _mesa_glthread_restore_dispatch(struct gl_context *ctx, const char *func)
 154 {
 155    /* Remove ourselves from the dispatch table except if another ctx/thread
 156     * already installed a new dispatch table.
 157     *
 158     * Typically glxMakeCurrent will bind a new context (install new table) then
 159     * old context might be deleted.
 160     */
 161    if (_glapi_get_dispatch() == ctx->MarshalExec) {
 162        ctx->CurrentClientDispatch = ctx->CurrentServerDispatch;
 163        _glapi_set_dispatch(ctx->CurrentClientDispatch);
 164 #if 0
 165        printf("glthread disabled: %s\n", func);
 166 #endif
 167    }
 168 }
 169
 170 void
 171 _mesa_glthread_disable(struct gl_context *ctx, const char *func)
 172 {
 173    _mesa_glthread_finish_before(ctx, func);
 174    _mesa_glthread_restore_dispatch(ctx, func);
 175 }
 176
 177 void
 178 _mesa_glthread_flush_batch(struct gl_context *ctx)
 179 {
 180    struct glthread_state *glthread = ctx->GLThread;
 181    if (!glthread)
 182       return;
 183
 184    struct glthread_batch *next = &glthread->batches[glthread->next];
 185    if (!next->used)
 186       return;
 187
 188    /* Debug: execute the batch immediately from this thread.
 189     *
 190     * Note that glthread_unmarshal_batch() changes the dispatch table so we'll
 191     * need to restore it when it returns.
 192     */
 193    if (false) {
 194       glthread_unmarshal_batch(next, 0);
 195       _glapi_set_dispatch(ctx->CurrentClientDispatch);
 196       return;
 197    }
 198
 199    p_atomic_add(&glthread->stats.num_offloaded_items, next->used);
 200
 201    util_queue_add_job(&glthread->queue, next, &next->fence,
 202                       glthread_unmarshal_batch, NULL, 0);
 203    glthread->last = glthread->next;
 204    glthread->next = (glthread->next + 1) % MARSHAL_MAX_BATCHES;
 205 }
 206
 207 /**
 208  * Waits for all pending batches have been unmarshaled.
 209  *
 210  * This can be used by the main thread to synchronize access to the context,
 211  * since the worker thread will be idle after this.
 212  */
 213 void
 214 _mesa_glthread_finish(struct gl_context *ctx)
 215 {
 216    struct glthread_state *glthread = ctx->GLThread;
 217    if (!glthread)
 218       return;
 219
 220    /* If this is called from the worker thread, then we've hit a path that
 221     * might be called from either the main thread or the worker (such as some
 222     * dri interface entrypoints), in which case we don't need to actually
 223     * synchronize against ourself.
 224     */
 225    if (u_thread_is_self(glthread->queue.threads[0]))
 226       return;
 227
 228    struct glthread_batch *last = &glthread->batches[glthread->last];
 229    struct glthread_batch *next = &glthread->batches[glthread->next];
 230    bool synced = false;
 231
 232    if (!util_queue_fence_is_signalled(&last->fence)) {
 233       util_queue_fence_wait(&last->fence);
 234       synced = true;
 235    }
 236
 237    if (next->used) {
 238       p_atomic_add(&glthread->stats.num_direct_items, next->used);
 239
 240       /* Since glthread_unmarshal_batch changes the dispatch to direct,
 241        * restore it after it's done.
 242        */
 243       struct _glapi_table *dispatch = _glapi_get_dispatch();
 244       glthread_unmarshal_batch(next, 0);
 245       _glapi_set_dispatch(dispatch);
 246
 247       /* It's not a sync because we don't enqueue partial batches, but
 248        * it would be a sync if we did. So count it anyway.
 249        */
 250       synced = true;
 251    }
 252
 253    if (synced)
 254       p_atomic_inc(&glthread->stats.num_syncs);
 255 }
 256
 257 void
 258 _mesa_glthread_finish_before(struct gl_context *ctx, const char *func)
 259 {
 260    _mesa_glthread_finish(ctx);
 261
 262    /* Uncomment this if you want to know where glthread syncs. */
 263    /*printf("fallback to sync: %s\n", func);*/
 264 }