gpuav: Update debug label regions management #8886
Conversation
|
CI Vulkan-ValidationLayers build queued with queue ID 305756. |
| std::string debug_region_name; | ||
| if (!cb_state.GetLabelCommands().empty()) { | ||
| const uint32_t last_label_command_i = static_cast<uint32_t>(cb_state.GetLabelCommands().size() - 1); | ||
| debug_region_name = cb_state.GetDebugRegionName(cb_state.GetLabelCommands(), last_label_command_i); |
There was a problem hiding this comment.
This is how I use GetDebugRegionName, anything looking wrong @artem-lunarg ?
There was a problem hiding this comment.
Looks good. Not sure what is the scope in gpu-av. If gpu-av needs also to consider global label state then GetDebugRegionName() provides initial_label_stack parameter and it can be used to add that support too.
Label commands are not limited to a single command buffer, in general. EndLabel can be in different command buffer. So after specific queue submission the queue can have current non-empty label name (per-queue state). It can be used as initial_label_stack to prepend when replaying label commands from current command buffer.
Not asking about change here (could be non-trivial change), just for awareness, in case you need this.
There was a problem hiding this comment.
Because this label stuff is really not the obvious, one example. It's fine to start command buffer with EndLabel command (I think I found this in real app too). This assumes that BeginLabel was in another command buffer. So queue-level state tracking helps with this.
NegativeSyncValReporting tests with DebugRegion in the name have various examples
There was a problem hiding this comment.
Ultimately we want the same level of support as sync val, so this is valuable info! I need to look at how you do those things and will probably ping you again on this topic
There was a problem hiding this comment.
Just refreshed memory how queue label state is tracked. It's vvl::Queue and cmdbuf_label_stack field.
Example how it is used in syncval:
One difference of tracking queue label state comparing to command buffer label tracking is that on queue level it's enough to keep std::vector<string> of nested labels, in contrast to LabelCommand in the command buffer. The reason is because after command buffer is submitted we know exact label hierarchy (just replay label commands from entire submitted command buffer) so we just save result of this replay.
|
CI Vulkan-ValidationLayers build # 18062 running. |
|
CI Vulkan-ValidationLayers build # 18062 failed. |
arno-lunarg
force-pushed
the
arno-debug-label-gpu-av
branch
from
ecce40d
to
7f10927
Compare
|
CI Vulkan-ValidationLayers build queued with queue ID 306110. |
|
CI Vulkan-ValidationLayers build # 18069 running. |
|
CI Vulkan-ValidationLayers build # 18069 failed. |
Took some of what Artem did for syncval Add debug label regions for debug printf Still need to add support for secondary command buffers, not so easy
arno-lunarg
force-pushed
the
arno-debug-label-gpu-av
branch
from
7f10927
to
fe1eb19
Compare
|
CI Vulkan-ValidationLayers build queued with queue ID 306142. |
|
CI Vulkan-ValidationLayers build # 18070 running. |
|
CI Vulkan-ValidationLayers build # 18070 failed. |
|
CI Vulkan-ValidationLayers build queued with queue ID 306976. |
|
CI Vulkan-ValidationLayers build # 18088 running. |
|
CI Vulkan-ValidationLayers build # 18088 passed. |
Took some of what Artem did for syncval
Add debug label regions for debug printf
Still need to add support for secondary command buffers, not so easy