gallium/util: import the multithreaded job queue from amdgpu winsys (v2)
authorMarek Olšák <marek.olsak@amd.com>
Sat, 11 Jun 2016 11:10:49 +0000 (13:10 +0200)
committerMarek Olšák <marek.olsak@amd.com>
Wed, 15 Jun 2016 19:07:34 +0000 (21:07 +0200)
v2: rename the event to util_queue_fence

Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
src/gallium/auxiliary/Makefile.sources
src/gallium/auxiliary/util/u_queue.c [new file with mode: 0644]
src/gallium/auxiliary/util/u_queue.h [new file with mode: 0644]
src/gallium/winsys/amdgpu/drm/amdgpu_cs.c
src/gallium/winsys/amdgpu/drm/amdgpu_cs.h
src/gallium/winsys/amdgpu/drm/amdgpu_winsys.c
src/gallium/winsys/amdgpu/drm/amdgpu_winsys.h

index 7b3853ef2bd3a233b6abcc295ed8d8b9c12aa32c..ab58358a1a3c5af34623a3f568145809e6cceaa8 100644 (file)
@@ -274,6 +274,8 @@ C_SOURCES := \
        util/u_pstipple.c \
        util/u_pstipple.h \
        util/u_pwr8.h \
+       util/u_queue.c \
+       util/u_queue.h \
        util/u_range.h \
        util/u_rect.h \
        util/u_resource.c \
diff --git a/src/gallium/auxiliary/util/u_queue.c b/src/gallium/auxiliary/util/u_queue.c
new file mode 100644 (file)
index 0000000..8e58414
--- /dev/null
@@ -0,0 +1,129 @@
+/*
+ * Copyright © 2016 Advanced Micro Devices, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining
+ * a copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
+ * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS
+ * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ */
+
+#include "u_queue.h"
+
+static PIPE_THREAD_ROUTINE(util_queue_thread_func, param)
+{
+   struct util_queue *queue = (struct util_queue*)param;
+   unsigned i;
+
+   while (1) {
+      struct util_queue_job job;
+
+      pipe_semaphore_wait(&queue->queued);
+      if (queue->kill_thread)
+         break;
+
+      pipe_mutex_lock(queue->lock);
+      job = queue->jobs[0];
+      for (i = 1; i < queue->num_jobs; i++)
+         queue->jobs[i - 1] = queue->jobs[i];
+      queue->jobs[--queue->num_jobs].job = NULL;
+      pipe_mutex_unlock(queue->lock);
+
+      pipe_semaphore_signal(&queue->has_space);
+
+      if (job.job) {
+         queue->execute_job(job.job);
+         pipe_semaphore_signal(&job.fence->done);
+      }
+   }
+
+   /* signal remaining jobs before terminating */
+   pipe_mutex_lock(queue->lock);
+   for (i = 0; i < queue->num_jobs; i++) {
+      pipe_semaphore_signal(&queue->jobs[i].fence->done);
+      queue->jobs[i].job = NULL;
+   }
+   queue->num_jobs = 0;
+   pipe_mutex_unlock(queue->lock);
+   return 0;
+}
+
+void
+util_queue_init(struct util_queue *queue,
+                void (*execute_job)(void *))
+{
+   memset(queue, 0, sizeof(*queue));
+   queue->execute_job = execute_job;
+   pipe_mutex_init(queue->lock);
+   pipe_semaphore_init(&queue->has_space, ARRAY_SIZE(queue->jobs));
+   pipe_semaphore_init(&queue->queued, 0);
+   queue->thread = pipe_thread_create(util_queue_thread_func, queue);
+}
+
+void
+util_queue_destroy(struct util_queue *queue)
+{
+   queue->kill_thread = 1;
+   pipe_semaphore_signal(&queue->queued);
+   pipe_thread_wait(queue->thread);
+   pipe_semaphore_destroy(&queue->has_space);
+   pipe_semaphore_destroy(&queue->queued);
+   pipe_mutex_destroy(queue->lock);
+}
+
+void
+util_queue_fence_init(struct util_queue_fence *fence)
+{
+   pipe_semaphore_init(&fence->done, 1);
+}
+
+void
+util_queue_fence_destroy(struct util_queue_fence *fence)
+{
+   pipe_semaphore_destroy(&fence->done);
+}
+
+void
+util_queue_add_job(struct util_queue *queue,
+                   void *job,
+                   struct util_queue_fence *fence)
+{
+   /* Set the semaphore to "busy". */
+   pipe_semaphore_wait(&fence->done);
+
+   /* if the queue is full, wait until there is space */
+   pipe_semaphore_wait(&queue->has_space);
+
+   pipe_mutex_lock(queue->lock);
+   assert(queue->num_jobs < ARRAY_SIZE(queue->jobs));
+   queue->jobs[queue->num_jobs].job = job;
+   queue->jobs[queue->num_jobs].fence = fence;
+   queue->num_jobs++;
+   pipe_mutex_unlock(queue->lock);
+   pipe_semaphore_signal(&queue->queued);
+}
+
+void
+util_queue_job_wait(struct util_queue_fence *fence)
+{
+   /* wait and set the semaphore to "busy" */
+   pipe_semaphore_wait(&fence->done);
+   /* set the semaphore to "idle" */
+   pipe_semaphore_signal(&fence->done);
+}
diff --git a/src/gallium/auxiliary/util/u_queue.h b/src/gallium/auxiliary/util/u_queue.h
new file mode 100644 (file)
index 0000000..db5a266
--- /dev/null
@@ -0,0 +1,80 @@
+/*
+ * Copyright © 2016 Advanced Micro Devices, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining
+ * a copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
+ * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+ * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS
+ * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ */
+
+/* Job queue with execution in a separate thread.
+ *
+ * Jobs can be added from any thread. After that, the wait call can be used
+ * to wait for completion of the job.
+ */
+
+#ifndef U_QUEUE_H
+#define U_QUEUE_H
+
+#include "os/os_thread.h"
+
+/* Job completion fence.
+ * Put this into your job structure.
+ */
+struct util_queue_fence {
+   pipe_semaphore done;
+};
+
+struct util_queue_job {
+   void *job;
+   struct util_queue_fence *fence;
+};
+
+/* Put this into your context. */
+struct util_queue {
+   pipe_mutex lock;
+   pipe_semaphore has_space;
+   pipe_semaphore queued;
+   pipe_thread thread;
+   int kill_thread;
+   int num_jobs;
+   struct util_queue_job jobs[8];
+   void (*execute_job)(void *job);
+};
+
+void util_queue_init(struct util_queue *queue,
+                     void (*execute_job)(void *));
+void util_queue_destroy(struct util_queue *queue);
+void util_queue_fence_init(struct util_queue_fence *fence);
+void util_queue_fence_destroy(struct util_queue_fence *fence);
+
+void util_queue_add_job(struct util_queue *queue,
+                        void *job,
+                        struct util_queue_fence *fence);
+void util_queue_job_wait(struct util_queue_fence *fence);
+
+/* util_queue needs to be cleared to zeroes for this to work */
+static inline bool
+util_queue_is_initialized(struct util_queue *queue)
+{
+   return queue->thread != 0;
+}
+
+#endif
index fefa5d6db535cb548f66406576d70d4a66be5bbc..4a7302ab2cb095d943fe0a165b26e7ba68151557 100644 (file)
@@ -605,7 +605,7 @@ amdgpu_cs_create(struct radeon_winsys_ctx *rwctx,
       return NULL;
    }
 
-   pipe_semaphore_init(&cs->flush_completed, 1);
+   util_queue_fence_init(&cs->flush_completed);
 
    cs->ctx = ctx;
    cs->flush_cs = flush;
@@ -872,8 +872,9 @@ static void amdgpu_add_fence_dependencies(struct amdgpu_cs *acs)
    }
 }
 
-void amdgpu_cs_submit_ib(struct amdgpu_cs *acs)
+void amdgpu_cs_submit_ib(void *job)
 {
+   struct amdgpu_cs *acs = (struct amdgpu_cs*)job;
    struct amdgpu_winsys *ws = acs->ctx->ws;
    struct amdgpu_cs_context *cs = acs->cst;
    int i, r;
@@ -957,14 +958,11 @@ cleanup:
 void amdgpu_cs_sync_flush(struct radeon_winsys_cs *rcs)
 {
    struct amdgpu_cs *cs = amdgpu_cs(rcs);
+   struct amdgpu_winsys *ws = cs->ctx->ws;
 
    /* Wait for any pending ioctl of this CS to complete. */
-   if (cs->ctx->ws->thread) {
-      /* wait and set the semaphore to "busy" */
-      pipe_semaphore_wait(&cs->flush_completed);
-      /* set the semaphore to "idle" */
-      pipe_semaphore_signal(&cs->flush_completed);
-   }
+   if (util_queue_is_initialized(&ws->cs_queue))
+      util_queue_job_wait(&cs->flush_completed);
 }
 
 DEBUG_GET_ONCE_BOOL_OPTION(noop, "RADEON_NOOP", FALSE)
@@ -1052,10 +1050,9 @@ static void amdgpu_cs_flush(struct radeon_winsys_cs *rcs,
       cs->cst = cur;
 
       /* Submit. */
-      if (ws->thread && (flags & RADEON_FLUSH_ASYNC)) {
-         /* Set the semaphore to "busy". */
-         pipe_semaphore_wait(&cs->flush_completed);
-         amdgpu_ws_queue_cs(ws, cs);
+      if ((flags & RADEON_FLUSH_ASYNC) &&
+          util_queue_is_initialized(&ws->cs_queue)) {
+         util_queue_add_job(&ws->cs_queue, cs, &cs->flush_completed);
       } else {
          amdgpu_cs_submit_ib(cs);
       }
@@ -1077,7 +1074,7 @@ static void amdgpu_cs_destroy(struct radeon_winsys_cs *rcs)
    struct amdgpu_cs *cs = amdgpu_cs(rcs);
 
    amdgpu_cs_sync_flush(rcs);
-   pipe_semaphore_destroy(&cs->flush_completed);
+   util_queue_fence_destroy(&cs->flush_completed);
    p_atomic_dec(&cs->ctx->ws->num_cs);
    pb_reference(&cs->main.big_ib_buffer, NULL);
    FREE(cs->main.base.prev);
index cc1516cfbf203d2365eca7261787d27e517246d8..354e403fa36d9006a498f7d3480276609ab48a99 100644 (file)
@@ -111,7 +111,7 @@ struct amdgpu_cs {
    void (*flush_cs)(void *ctx, unsigned flags, struct pipe_fence_handle **fence);
    void *flush_data;
 
-   pipe_semaphore flush_completed;
+   struct util_queue_fence flush_completed;
 };
 
 struct amdgpu_fence {
@@ -218,6 +218,6 @@ bool amdgpu_fence_wait(struct pipe_fence_handle *fence, uint64_t timeout,
                        bool absolute);
 void amdgpu_cs_sync_flush(struct radeon_winsys_cs *rcs);
 void amdgpu_cs_init_functions(struct amdgpu_winsys *ws);
-void amdgpu_cs_submit_ib(struct amdgpu_cs *cs);
+void amdgpu_cs_submit_ib(void *job);
 
 #endif
index 701622184c8af52df4ad09f2c0ac268d41b35f46..7ef35298aa8a33250e5398b44f3e660bea453efd 100644 (file)
@@ -308,14 +308,9 @@ static void amdgpu_winsys_destroy(struct radeon_winsys *rws)
 {
    struct amdgpu_winsys *ws = (struct amdgpu_winsys*)rws;
 
-   if (ws->thread) {
-      ws->kill_thread = 1;
-      pipe_semaphore_signal(&ws->cs_queued);
-      pipe_thread_wait(ws->thread);
-   }
-   pipe_semaphore_destroy(&ws->cs_queue_has_space);
-   pipe_semaphore_destroy(&ws->cs_queued);
-   pipe_mutex_destroy(ws->cs_queue_lock);
+   if (util_queue_is_initialized(&ws->cs_queue))
+      util_queue_destroy(&ws->cs_queue);
+
    pipe_mutex_destroy(ws->bo_fence_lock);
    pb_cache_deinit(&ws->bo_cache);
    pipe_mutex_destroy(ws->global_bo_list_lock);
@@ -400,53 +395,7 @@ static int compare_dev(void *key1, void *key2)
    return key1 != key2;
 }
 
-void amdgpu_ws_queue_cs(struct amdgpu_winsys *ws, struct amdgpu_cs *cs)
-{
-   pipe_semaphore_wait(&ws->cs_queue_has_space);
-
-   pipe_mutex_lock(ws->cs_queue_lock);
-   assert(ws->num_enqueued_cs < ARRAY_SIZE(ws->cs_queue));
-   ws->cs_queue[ws->num_enqueued_cs++] = cs;
-   pipe_mutex_unlock(ws->cs_queue_lock);
-   pipe_semaphore_signal(&ws->cs_queued);
-}
-
-static PIPE_THREAD_ROUTINE(amdgpu_cs_thread_func, param)
-{
-   struct amdgpu_winsys *ws = (struct amdgpu_winsys *)param;
-   struct amdgpu_cs *cs;
-   unsigned i;
-
-   while (1) {
-      pipe_semaphore_wait(&ws->cs_queued);
-      if (ws->kill_thread)
-         break;
-
-      pipe_mutex_lock(ws->cs_queue_lock);
-      cs = ws->cs_queue[0];
-      for (i = 1; i < ws->num_enqueued_cs; i++)
-         ws->cs_queue[i - 1] = ws->cs_queue[i];
-      ws->cs_queue[--ws->num_enqueued_cs] = NULL;
-      pipe_mutex_unlock(ws->cs_queue_lock);
-
-      pipe_semaphore_signal(&ws->cs_queue_has_space);
-
-      if (cs) {
-         amdgpu_cs_submit_ib(cs);
-         pipe_semaphore_signal(&cs->flush_completed);
-      }
-   }
-   pipe_mutex_lock(ws->cs_queue_lock);
-   for (i = 0; i < ws->num_enqueued_cs; i++) {
-      pipe_semaphore_signal(&ws->cs_queue[i]->flush_completed);
-      ws->cs_queue[i] = NULL;
-   }
-   pipe_mutex_unlock(ws->cs_queue_lock);
-   return 0;
-}
-
 DEBUG_GET_ONCE_BOOL_OPTION(thread, "RADEON_THREAD", TRUE)
-static PIPE_THREAD_ROUTINE(amdgpu_cs_thread_func, param);
 
 static bool amdgpu_winsys_unref(struct radeon_winsys *rws)
 {
@@ -541,14 +490,10 @@ amdgpu_winsys_create(int fd, radeon_screen_create_t screen_create)
 
    LIST_INITHEAD(&ws->global_bo_list);
    pipe_mutex_init(ws->global_bo_list_lock);
-   pipe_mutex_init(ws->cs_queue_lock);
    pipe_mutex_init(ws->bo_fence_lock);
 
-   pipe_semaphore_init(&ws->cs_queue_has_space, ARRAY_SIZE(ws->cs_queue));
-   pipe_semaphore_init(&ws->cs_queued, 0);
-
    if (sysconf(_SC_NPROCESSORS_ONLN) > 1 && debug_get_option_thread())
-      ws->thread = pipe_thread_create(amdgpu_cs_thread_func, ws);
+      util_queue_init(&ws->cs_queue, amdgpu_cs_submit_ib);
 
    /* Create the screen at the end. The winsys must be initialized
     * completely.
index d6734f7bfbce55d61e3779f839490372b69a6d84..b13a17e119c5e749cc55c898a8c2e2fe4d0fa93c 100644 (file)
@@ -35,7 +35,7 @@
 #include "pipebuffer/pb_cache.h"
 #include "gallium/drivers/radeon/radeon_winsys.h"
 #include "addrlib/addrinterface.h"
-#include "os/os_thread.h"
+#include "util/u_queue.h"
 #include <amdgpu.h>
 
 struct amdgpu_cs;
@@ -59,13 +59,7 @@ struct amdgpu_winsys {
    struct radeon_info info;
 
    /* multithreaded IB submission */
-   pipe_mutex cs_queue_lock;
-   pipe_semaphore cs_queue_has_space;
-   pipe_semaphore cs_queued;
-   pipe_thread thread;
-   int kill_thread;
-   int num_enqueued_cs;
-   struct amdgpu_cs *cs_queue[8];
+   struct util_queue cs_queue;
 
    struct amdgpu_gpu_info amdinfo;
    ADDR_HANDLE addrlib;
@@ -84,7 +78,6 @@ amdgpu_winsys(struct radeon_winsys *base)
    return (struct amdgpu_winsys*)base;
 }
 
-void amdgpu_ws_queue_cs(struct amdgpu_winsys *ws, struct amdgpu_cs *cs);
 void amdgpu_surface_init_functions(struct amdgpu_winsys *ws);
 ADDR_HANDLE amdgpu_addr_create(struct amdgpu_winsys *ws);