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/marshal.h"
  38 #include "util/u_atomic.h"
  39 #include "util/u_thread.h"
  40
  41
  42 static void
  43 glthread_unmarshal_batch(void *job, int thread_index)
  44 {
  45    struct glthread_batch *batch = (struct glthread_batch*)job;
  46    struct gl_context *ctx = batch->ctx;
  47    size_t pos = 0;
  48
  49    _glapi_set_dispatch(ctx->CurrentServerDispatch);
  50
  51    while (pos < batch->used) {
  52       const struct marshal_cmd_base *cmd =
  53          (const struct marshal_cmd_base *)&batch->buffer[pos];
  54
  55       _mesa_unmarshal_dispatch[cmd->cmd_id](ctx, cmd);
  56       pos += cmd->cmd_size;
  57    }
  58
  59    assert(pos == batch->used);
  60    batch->used = 0;
  61 }
  62
  63 static void
  64 glthread_thread_initialization(void *job, int thread_index)
  65 {
  66    struct gl_context *ctx = (struct gl_context*)job;
  67
  68    ctx->Driver.SetBackgroundContext(ctx, &ctx->GLThread->stats);
  69    _glapi_set_context(ctx);
  70 }
  71
  72 void
  73 _mesa_glthread_init(struct gl_context *ctx)
  74 {
  75    struct glthread_state *glthread = calloc(1, sizeof(*glthread));
  76
  77    if (!glthread)
  78       return;
  79
  80    if (!util_queue_init(&glthread->queue, "gl", MARSHAL_MAX_BATCHES - 2,
  81                         1, 0)) {
  82       free(glthread);
  83       return;
  84    }
  85
  86    ctx->MarshalExec = _mesa_create_marshal_table(ctx);
  87    if (!ctx->MarshalExec) {
  88       util_queue_destroy(&glthread->queue);
  89       free(glthread);
  90       return;
  91    }
  92
  93    for (unsigned i = 0; i < MARSHAL_MAX_BATCHES; i++) {
  94       glthread->batches[i].ctx = ctx;
  95       util_queue_fence_init(&glthread->batches[i].fence);
  96    }
  97
  98    glthread->stats.queue = &glthread->queue;
  99    ctx->CurrentClientDispatch = ctx->MarshalExec;
 100    ctx->GLThread = glthread;
 101
 102    /* Execute the thread initialization function in the thread. */
 103    struct util_queue_fence fence;
 104    util_queue_fence_init(&fence);
 105    util_queue_add_job(&glthread->queue, ctx, &fence,
 106                       glthread_thread_initialization, NULL, 0);
 107    util_queue_fence_wait(&fence);
 108    util_queue_fence_destroy(&fence);
 109 }
 110
 111 void
 112 _mesa_glthread_destroy(struct gl_context *ctx)
 113 {
 114    struct glthread_state *glthread = ctx->GLThread;
 115
 116    if (!glthread)
 117       return;
 118
 119    _mesa_glthread_finish(ctx);
 120    util_queue_destroy(&glthread->queue);
 121
 122    for (unsigned i = 0; i < MARSHAL_MAX_BATCHES; i++)
 123       util_queue_fence_destroy(&glthread->batches[i].fence);
 124
 125    free(glthread);
 126    ctx->GLThread = NULL;
 127
 128    _mesa_glthread_restore_dispatch(ctx, "destroy");
 129 }
 130
 131 void
 132 _mesa_glthread_restore_dispatch(struct gl_context *ctx, const char *func)
 133 {
 134    /* Remove ourselves from the dispatch table except if another ctx/thread
 135     * already installed a new dispatch table.
 136     *
 137     * Typically glxMakeCurrent will bind a new context (install new table) then
 138     * old context might be deleted.
 139     */
 140    if (_glapi_get_dispatch() == ctx->MarshalExec) {
 141        ctx->CurrentClientDispatch = ctx->CurrentServerDispatch;
 142        _glapi_set_dispatch(ctx->CurrentClientDispatch);
 143 #if 0
 144        printf("glthread disabled: %s\n", func);
 145 #endif
 146    }
 147 }
 148
 149 void
 150 _mesa_glthread_flush_batch(struct gl_context *ctx)
 151 {
 152    struct glthread_state *glthread = ctx->GLThread;
 153    if (!glthread)
 154       return;
 155
 156    struct glthread_batch *next = &glthread->batches[glthread->next];
 157    if (!next->used)
 158       return;
 159
 160    /* Debug: execute the batch immediately from this thread.
 161     *
 162     * Note that glthread_unmarshal_batch() changes the dispatch table so we'll
 163     * need to restore it when it returns.
 164     */
 165    if (false) {
 166       glthread_unmarshal_batch(next, 0);
 167       _glapi_set_dispatch(ctx->CurrentClientDispatch);
 168       return;
 169    }
 170
 171    p_atomic_add(&glthread->stats.num_offloaded_items, next->used);
 172
 173    util_queue_add_job(&glthread->queue, next, &next->fence,
 174                       glthread_unmarshal_batch, NULL, 0);
 175    glthread->last = glthread->next;
 176    glthread->next = (glthread->next + 1) % MARSHAL_MAX_BATCHES;
 177 }
 178
 179 /**
 180  * Waits for all pending batches have been unmarshaled.
 181  *
 182  * This can be used by the main thread to synchronize access to the context,
 183  * since the worker thread will be idle after this.
 184  */
 185 void
 186 _mesa_glthread_finish(struct gl_context *ctx)
 187 {
 188    struct glthread_state *glthread = ctx->GLThread;
 189    if (!glthread)
 190       return;
 191
 192    /* If this is called from the worker thread, then we've hit a path that
 193     * might be called from either the main thread or the worker (such as some
 194     * dri interface entrypoints), in which case we don't need to actually
 195     * synchronize against ourself.
 196     */
 197    if (u_thread_is_self(glthread->queue.threads[0]))
 198       return;
 199
 200    struct glthread_batch *last = &glthread->batches[glthread->last];
 201    struct glthread_batch *next = &glthread->batches[glthread->next];
 202    bool synced = false;
 203
 204    if (!util_queue_fence_is_signalled(&last->fence)) {
 205       util_queue_fence_wait(&last->fence);
 206       synced = true;
 207    }
 208
 209    if (next->used) {
 210       p_atomic_add(&glthread->stats.num_direct_items, next->used);
 211
 212       /* Since glthread_unmarshal_batch changes the dispatch to direct,
 213        * restore it after it's done.
 214        */
 215       struct _glapi_table *dispatch = _glapi_get_dispatch();
 216       glthread_unmarshal_batch(next, 0);
 217       _glapi_set_dispatch(dispatch);
 218
 219       /* It's not a sync because we don't enqueue partial batches, but
 220        * it would be a sync if we did. So count it anyway.
 221        */
 222       synced = true;
 223    }
 224
 225    if (synced)
 226       p_atomic_inc(&glthread->stats.num_syncs);
 227 }