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