Observed Behavior
The current default behavior for chrome traces is to put host API calls from each thread on their own "track", and device commands for each OpenCL command-queue on their own "track". This is useful for most multi-threaded applications that create a small pool of threads and command-queues, but it's bad for some other applications that create (and destroy) a lot of threads or command-queues.
For completeness, note that there are also controls for chrome performance timing "in stages" or "per kernel", but for this request we can assume that these controls are not set and hence not are not applicable.
Desired Behavior
For these classes of applications, add a control to "collapse" all device execution to a single track (per-device?). This provides a single device code synopsis regardless of how many command-queues are created. Consider "collapsing" all host API calls to a single track also, though device code execution is the number one priority.
Steps to Reproduce
Run an OpenCL application that creates and enqueues work to multiple command-queues targeting a single device with ChromePerformanceTiming enabled. Observe that each command-queue has its own "track".
Observed Behavior
The current default behavior for chrome traces is to put host API calls from each thread on their own "track", and device commands for each OpenCL command-queue on their own "track". This is useful for most multi-threaded applications that create a small pool of threads and command-queues, but it's bad for some other applications that create (and destroy) a lot of threads or command-queues.
For completeness, note that there are also controls for chrome performance timing "in stages" or "per kernel", but for this request we can assume that these controls are not set and hence not are not applicable.
Desired Behavior
For these classes of applications, add a control to "collapse" all device execution to a single track (per-device?). This provides a single device code synopsis regardless of how many command-queues are created. Consider "collapsing" all host API calls to a single track also, though device code execution is the number one priority.
Steps to Reproduce
Run an OpenCL application that creates and enqueues work to multiple command-queues targeting a single device with ChromePerformanceTiming enabled. Observe that each command-queue has its own "track".