4

v3dv: handle wait semaphores in the first job by queue

 1 year ago
source link: https://gitlab.freedesktop.org/mesa/mesa/-/commit/0ab98612ef6860751f6a5115229e561eab7bb890
Go to the source link to view the article. You can view the picture content, updated content and better typesetting reading experience. If the link is broken, please click the button below to view the snapshot at that time.

Commit 0ab98612

authored 7 months ago by

Melissa Wen

Browse files

v3dv: handle wait semaphores in the first job by queue 


With multiple semaphore support, we can improve the way we handle
wait semaphores considering different job types and cmd buffer
batch scenarios, that means:

- A GPU job depends on wait semaphores whenever it is the first job
submitted to a queue in this command buffer batch (the `first` flag
for the job's queue type is set).
- For the first CPU job, if there are wait semaphores, we should
wait for the CPU and GPU being idle to process the job.
Signed-off-by: Melissa Wen <[email protected]>
Reviewed-by: Iago Toral Quiroga <[email protected]>
Part-of: <!13178>

1 merge request!13178v3dv: add support to multiple wait and signal semaphores

Showing
......@@ -687,8 +687,8 @@ v3dv_cmd_buffer_finish_job(struct v3dv_cmd_buffer *cmd_buffer)
}
}
static bool
job_type_is_gpu(struct v3dv_job *job)
bool
v3dv_job_type_is_gpu(struct v3dv_job *job)
{
switch (job->type) {
case V3DV_JOB_TYPE_GPU_CL:
......@@ -713,7 +713,7 @@ cmd_buffer_serialize_job_if_needed(struct v3dv_cmd_buffer *cmd_buffer,
/* Serialization only affects GPU jobs, CPU jobs are always automatically
* serialized.
*/
if (!job_type_is_gpu(job))
if (!v3dv_job_type_is_gpu(job))
return;
job->serialize = true;
......
......@@ -1148,6 +1148,8 @@ void v3dv_job_start_frame(struct v3dv_job *job,
uint8_t max_internal_bpp,
bool msaa);
bool v3dv_job_type_is_gpu(struct v3dv_job *job);
struct v3dv_job *
v3dv_job_clone_in_cmd_buffer(struct v3dv_job *job,
struct v3dv_cmd_buffer *cmd_buffer);
......
......@@ -668,15 +668,23 @@ multisync_free(struct v3dv_device *device,
static struct drm_v3d_sem *
set_in_syncs(struct v3dv_device *device,
struct v3dv_job *job,
enum v3dv_queue_type queue,
uint32_t *count,
struct v3dv_submit_info_semaphores *sems_info)
{
/* If we are serializing a job in a cmd buffer, we are already making it
* wait until the last job submitted to each queue completes before
* running, so in that case we can skip waiting for any additional
* semaphores.
uint32_t n_sems = 0;
/* If this is the first job submitted to a given GPU queue in this cmd buf
* batch, it has to wait on wait semaphores (if any) before running.
*/
if (device->last_job_syncs.first[queue])
n_sems = sems_info->wait_sem_count;
/* If we don't need to wait on wait semaphores but the serialize flag is
* set, this job waits for completion of all GPU jobs submitted in any
* queue V3DV_QUEUE_(CL/TFU/CSD) before running.
*/
*count = job->serialize ? 3 : sems_info->wait_sem_count;
*count = n_sems == 0 && job->serialize ? 3 : n_sems;
if (!*count)
return NULL;
......@@ -688,7 +696,7 @@ set_in_syncs(struct v3dv_device *device,
if (!syncs)
return NULL;
if (!job->serialize) {
if (n_sems) {
for (int i = 0; i < *count; i++) {
struct v3dv_semaphore *sem =
v3dv_semaphore_from_handle(sems_info->wait_sems[i]);
......@@ -768,7 +776,8 @@ set_multisync(struct drm_v3d_multi_sync *ms,
{
uint32_t out_sync_count = 0, in_sync_count = 0;
in_syncs = set_in_syncs(device, job, &in_sync_count, sems_info);
in_syncs = set_in_syncs(device, job, queue_sync,
&in_sync_count, sems_info);
if (!in_syncs && in_sync_count)
goto fail;
......@@ -787,6 +796,8 @@ set_multisync(struct drm_v3d_multi_sync *ms,
ms->in_sync_count = in_sync_count;
ms->in_syncs = (uintptr_t)(void *)in_syncs;
device->last_job_syncs.first[queue_sync] = false;
return;
fail:
......@@ -1045,6 +1056,30 @@ queue_submit_job(struct v3dv_queue *queue,
{
assert(job);
/* CPU jobs typically execute explicit waits before they are processed. For
* example, a query reset CPU job will explicitly wait for the queries
* being unused before proceeding, etc. However, if we have any wait
* semaphores, we need to honour that too for the first CPU job we process
* in the command buffer batch. We do that by waiting for idle to ensure
* that any previous work has been completed, at which point any wait
* semaphores must be signalled, and we never need to do this again for the
* same batch.
*/
if (!v3dv_job_type_is_gpu(job) && sems_info->wait_sem_count) {
v3dv_QueueWaitIdle(v3dv_queue_to_handle(&job->device->queue));
#ifdef DEBUG
/* Loop through wait sems and check they are all signalled */
for (int i = 0; i < sems_info->wait_sem_count; i++) {
int render_fd = queue->device->pdevice->render_fd;
struct v3dv_semaphore *sem =
v3dv_semaphore_from_handle(sems_info->wait_sems[i]);
int ret = drmSyncobjWait(render_fd, &sem->sync, 1, 0, 0, NULL);
assert(ret == 0);
}
#endif
sems_info->wait_sem_count = 0;
}
switch (job->type) {
case V3DV_JOB_TYPE_GPU_CL:
return handle_cl_job(queue, job, sems_info);
......@@ -1212,6 +1247,12 @@ queue_submit_cmd_buffer_batch(struct v3dv_queue *queue,
.signal_sems = (VkSemaphore *) pSubmit->pSignalSemaphores,
};
/* In the beginning of a cmd buffer batch, we set all last_job_syncs as
* first. It helps to determine wait semaphores conditions.
*/
for (unsigned i = 0; i < V3DV_QUEUE_COUNT; i++)
queue->device->last_job_syncs.first[i] = true;
/* Even if we don't have any actual work to submit we still need to wait
* on the wait semaphores and signal the signal semaphores and fence, so
* in this scenario we just submit a trivial no-op job so we don't have
......
Please register or sign in to comment

report

Recommend

About Joyk


Aggregate valuable and interesting links.
Joyk means Joy of geeK