AlexIndustrial/mesa
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

tu/drm: Emulate combined gfx/sparse queues

Browse Source 
Create an extra sparse queue when creating a graphics queue, and send
sparse binds to that queue. Synchronize graphics commands against sparse
binds and vice versa using a pair of timeline semaphores.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37304>
This commit is contained in:
Connor Abbott
2025-09-11 11:41:53 -04:00
committed by Marge Bot
parent 6c296c1283
commit 0cc0e786e0
2 changed files with 127 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/freedreno/vulkan/tu_knl_drm_msm.cc
+122 -26
View File
@@ -335,30 +335,85 @@ msm_device_check_status(struct tu_device *device)
static int
msm_submitqueue_new(struct tu_device *dev, struct tu_queue *queue)
{
assert(queue->priority >= 0 &&
queue->priority < dev->physical_device->submitqueue_priority_count);
struct drm_msm_submitqueue req = {
.flags = queue->type == TU_QUEUE_SPARSE ? MSM_SUBMITQUEUE_VM_BIND :
(dev->physical_device->info->chip >= 7 &&
dev->physical_device->has_preemption ?
MSM_SUBMITQUEUE_ALLOW_PREEMPT : 0),
.prio = queue->priority,
};
int ret = 0;
int ret = drmCommandWriteRead(dev->fd,
DRM_MSM_SUBMITQUEUE_NEW, &req, sizeof(req));
if (ret)
return ret;
if (queue->type == TU_QUEUE_GFX) {
assert(queue->priority >= 0 &&
queue->priority < dev->physical_device->submitqueue_priority_count);
struct drm_msm_submitqueue req = {
.flags = (dev->physical_device->info->chip >= 7 &&
dev->physical_device->has_preemption ?
MSM_SUBMITQUEUE_ALLOW_PREEMPT : 0),
.prio = queue->priority,
};
ret = drmCommandWriteRead(dev->fd,
DRM_MSM_SUBMITQUEUE_NEW, &req, sizeof(req));
if (ret)
return ret;
queue->msm_queue_id = req.id;
}
if (queue->type == TU_QUEUE_SPARSE ||
dev->physical_device->has_vm_bind) {
struct drm_msm_submitqueue req = {
.flags = MSM_SUBMITQUEUE_VM_BIND,
.prio = 0,
};
ret = drmCommandWriteRead(dev->fd,
DRM_MSM_SUBMITQUEUE_NEW, &req, sizeof(req));
if (ret)
goto fail_sparse;
queue->sparse_queue_id = req.id;
}
if (queue->type == TU_QUEUE_GFX && dev->physical_device->has_vm_bind) {
ret = drmSyncobjCreate(dev->fd, 0, &queue->gfx_syncobj);
if (ret)
goto fail_gfx_syncobj;
ret = drmSyncobjCreate(dev->fd, 0, &queue->sparse_syncobj);
if (ret)
goto fail_sparse_syncobj;
}
queue->msm_queue_id = req.id;
return 0;
fail_sparse_syncobj:
drmSyncobjDestroy(dev->fd, queue->gfx_syncobj);
fail_gfx_syncobj:
if (queue->type == TU_QUEUE_SPARSE || dev->physical_device->has_vm_bind) {
drmCommandWrite(dev->fd, DRM_MSM_SUBMITQUEUE_CLOSE,
&queue->sparse_queue_id, sizeof(uint32_t));
}
fail_sparse:
if (queue->type == TU_QUEUE_GFX) {
drmCommandWrite(dev->fd, DRM_MSM_SUBMITQUEUE_CLOSE,
&queue->msm_queue_id, sizeof(uint32_t));
}
return ret;
}
static void
msm_submitqueue_close(struct tu_device *dev, struct tu_queue *queue)
{
drmCommandWrite(dev->fd, DRM_MSM_SUBMITQUEUE_CLOSE,
&queue->msm_queue_id, sizeof(uint32_t));
if (queue->msm_queue_id) {
drmCommandWrite(dev->fd, DRM_MSM_SUBMITQUEUE_CLOSE,
&queue->msm_queue_id, sizeof(uint32_t));
}
if (queue->sparse_queue_id) {
drmCommandWrite(dev->fd, DRM_MSM_SUBMITQUEUE_CLOSE,
&queue->sparse_queue_id, sizeof(uint32_t));
}
if (queue->type == TU_QUEUE_GFX && dev->physical_device->has_vm_bind) {
drmSyncobjDestroy(dev->fd, queue->sparse_syncobj);
drmSyncobjDestroy(dev->fd, queue->gfx_syncobj);
}
}
static void
@@ -1199,11 +1254,18 @@ msm_queue_submit(struct tu_queue *queue, void *_submit,
uint64_t start_ts = tu_perfetto_begin_submit();
#endif
uint32_t fence = 0;
uint64_t other_timepoint =
submit->binds.size != 0 ? queue->gfx_timepoint :
queue->sparse_timepoint;
uint32_t extra_wait_count =
(queue->type == TU_QUEUE_GFX && other_timepoint != 0) ? 1 : 0;
uint32_t extra_signal_count =
(queue->type == TU_QUEUE_GFX && has_vm_bind) ? 1 : 0;
/* Allocate without wait timeline semaphores */
in_syncobjs = (struct drm_msm_syncobj *) vk_zalloc(
&queue->device->vk.alloc,
wait_count * sizeof(*in_syncobjs), 8,
(wait_count + extra_wait_count) * sizeof(*in_syncobjs), 8,
VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
if (in_syncobjs == NULL) {
@@ -1214,7 +1276,7 @@ msm_queue_submit(struct tu_queue *queue, void *_submit,
/* Allocate with signal timeline semaphores considered */
out_syncobjs = (struct drm_msm_syncobj *) vk_zalloc(
&queue->device->vk.alloc,
signal_count * sizeof(*out_syncobjs), 8,
(signal_count + extra_signal_count) * sizeof(*out_syncobjs), 8,
VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
if (out_syncobjs == NULL) {
@@ -1242,7 +1304,22 @@ msm_queue_submit(struct tu_queue *queue, void *_submit,
};
}
if (queue->type == TU_QUEUE_SPARSE) {
if (queue->type == TU_QUEUE_SPARSE || submit->binds.size != 0) {
if (extra_wait_count) {
in_syncobjs[wait_count] = (struct drm_msm_syncobj) {
.handle = queue->gfx_syncobj,
.flags = 0,
.point = queue->gfx_timepoint,
};
}
if (extra_signal_count) {
out_syncobjs[signal_count] = (struct drm_msm_syncobj) {
.handle = queue->sparse_syncobj,
.flags = 0,
.point = ++queue->sparse_timepoint,
};
}
unsigned nr_ops = util_dynarray_num_elements(&submit->binds,
struct drm_msm_vm_bind_op);
@@ -1271,11 +1348,11 @@ msm_queue_submit(struct tu_queue *queue, void *_submit,
.flags = flags,
.nr_ops = nr_ops,
.fence_fd = queue->device->vm_bind_fence_fd,
.queue_id = queue->msm_queue_id,
.queue_id = queue->sparse_queue_id,
.in_syncobjs = (uint64_t)(uintptr_t)in_syncobjs,
.out_syncobjs = (uint64_t)(uintptr_t)out_syncobjs,
.nr_in_syncobjs = wait_count,
.nr_out_syncobjs = signal_count,
.nr_in_syncobjs = wait_count + extra_wait_count,
.nr_out_syncobjs = signal_count + extra_signal_count,
.syncobj_stride = sizeof(struct drm_msm_syncobj),
.op_stride = sizeof(struct drm_msm_vm_bind_op),
};
@@ -1337,9 +1414,13 @@ msm_queue_submit(struct tu_queue *queue, void *_submit,
.flags = flags,
.nr_ops = 1,
.fence_fd = queue->device->vm_bind_fence_fd,
.queue_id = queue->msm_queue_id,
.queue_id = queue->sparse_queue_id,
.in_syncobjs =
(uint64_t)(uintptr_t)&in_syncobjs[wait_count],
.out_syncobjs = (uint64_t)(uintptr_t)out_syncobjs,
.nr_out_syncobjs = last ? signal_count : 0,
.nr_in_syncobjs = extra_wait_count,
.nr_out_syncobjs = last ? signal_count +
extra_signal_count : 0,
.syncobj_stride = sizeof(struct drm_msm_syncobj),
.op_stride = sizeof(struct drm_msm_vm_bind_op),
.op = *op,
@@ -1369,6 +1450,21 @@ msm_queue_submit(struct tu_queue *queue, void *_submit,
flags |= MSM_SUBMIT_SYNCOBJ_OUT;
if (has_vm_bind) {
if (extra_wait_count) {
in_syncobjs[wait_count] = (struct drm_msm_syncobj) {
.handle = queue->sparse_syncobj,
.flags = 0,
.point = queue->sparse_timepoint,
};
}
if (extra_signal_count) {
out_syncobjs[signal_count] = (struct drm_msm_syncobj) {
.handle = queue->gfx_syncobj,
.flags = 0,
.point = ++queue->gfx_timepoint,
};
}
u_rwlock_rdlock(&queue->device->vm_bind_fence_lock);
if (queue->device->vm_bind_fence_fd != -1)
@@ -1422,8 +1518,8 @@ msm_queue_submit(struct tu_queue *queue, void *_submit,
.queueid = queue->msm_queue_id,
.in_syncobjs = (uint64_t)(uintptr_t)in_syncobjs,
.out_syncobjs = (uint64_t)(uintptr_t)out_syncobjs,
.nr_in_syncobjs = wait_count,
.nr_out_syncobjs = signal_count,
.nr_in_syncobjs = wait_count + extra_wait_count,
.nr_out_syncobjs = signal_count + extra_signal_count,
.syncobj_stride = sizeof(struct drm_msm_syncobj),
};
src/freedreno/vulkan/tu_queue.h
+5
View File
@@ -28,6 +28,11 @@ struct tu_queue
uint32_t msm_queue_id;
uint32_t priority;
uint32_t sparse_queue_id;
uint32_t sparse_syncobj, gfx_syncobj;
uint64_t sparse_timepoint, gfx_timepoint;
int fence; /* timestamp/fence of the last queue submission */
};
VK_DEFINE_HANDLE_CASTS(tu_queue, vk.base, VkQueue, VK_OBJECT_TYPE_QUEUE)