Async Workgroup Update

1
Async Workgroup Update

• Barthold Lichtenbelt

2
Goals

• Provide synchronization framework for OpenGL
• Provide base functionality as defined in NV_fence
  and GL2_async_core
• Build a framework for future, more complex,
  functionality, some of which discussed in
  GL2_async_core
• Initially support CPU lt-gt GPU synchronization
• Support synchronization across multiple OpenGL
  contexts
• Resulted in GL_ARB_sync spec
• Finished April 2006
• Posted draft to opengl.org for feedback
• Not quite official ARB extension yet

3
Functionality overview

• ARB_sync provides synchronization primitives
• Can be tested, set and waited upon
• Specifically, a Fence Synchronization Object
  and corresponding Fence command
• Fence completion allows for partial glFinish
• All commands prior to the fence are forced to
  complete before control is returned to caller
• Fence Sync Objects can be shared across contexts
• Allows for synchronization of OpenGL command
  streams across contexts
• New data type GLtime represents intervals in
  nanoseconds
• 64 bit integer, same encoding as UST counter in
  OpenML
• Accuracy implementation dependent, precision in
  nanoseconds
• If you have used the Windows Event model, this
  will feel familiar

4
Synchronization model in ARB_sync 1/2

• A sync object is a primitive used for
  synchronization between CPU and GPU, CPU, or
  something else.
• Sync object has state type, condition, status
• A sync objects status can be signaled or
  non-signaled
• when created status is signaled unless a flag is
  set in which case it is non-signaled
• A fence sync object is a specific type of sync
  object
• Provides partial finish semantics
• Only type of sync object currently defined
• A fence is a token inserted in the GL command
  stream
• A sync object is not inserted into the command
  stream
• Fence has no state
• A fence is associated with a fence sync object.
• Multiple fences can be associated with the same
  sync object
• When a fence is inserted in the command stream,
  the status of its sync object is set to
  non-signaled
• A fence, once completed, will set the status of
  its sync object to signaled

5
Synchronization model in ARB_sync 2/2

• A wait function waits on a sync object, not on a
  fence
• A poll function polls a sync object, not a fence
• A wait function called on a sync object in the
  non-signaled state will block. It unblocks when
  the sync object transitions to the signaled
  state.

6
Example RTT with two contexts

• Context A
• Sync_objectA glCreateSync(attrib)
• ltrender to texture that context B needsgt
• glFence(sync_objectA)
• glFlush() // prevent deadlock
• Context B
• glClientWaitSync(sync_objectA,0,GL_FOREVER)
• glBindTexture(.) // Just rendered
• ltrender using texturegt

7
OS specific functionality

• Convert sync object to the window system native
  event primitive
• Allows applications to synchronize all events in
  a system using one API
• All operations on ltsyncgt are reflected in OS
  event and vice-versa
• Both ltsyncgt and the OS event are valid to use in
  your code
• On windows, convert to an Event
• HANDLE wglConvertSyncToEvent(object sync)
• Need to specify, when sync object is created,
  that it can be converted to OS event
• Separate extension WGL_ARB_sync_event
• On Unix, convert to a file-descriptor, x-event or
  semaphore?
• Still TBD

8
Possible future functionality

• Add a WaitForMultipleSync(uint sync_objects, .)
  command
• Synchronize with multiple sync objects at once
• Add a payload to a fence
• For example, the time it completed
• Allow one GPU stream to wait for another GPU
  stream
• WaitSync(sync_object)
• A sync object whose status will pulse with every
  vblank
• A sync object that can signal when data binding
  has completed
• As opposed to when rendering has completed using
  the data

9
Example Streaming video processing

• Loop
• Draw frame 1 // To a FBO, for example
• glFence(sync_object1) // inserts a fence in the
  command stream
• Draw frame 2
• glFence(sync_object2)
• while (glClientWaitSync(sync_object1,0,0)!GL_ALRE
  ADY_SIGNALED)
• ltDo some useful workgt // App uses CPU cycles
  instead of blocking
• Read back data in frame 1
• while (glClientWaitSync(sync_object1,0,0)!GL_ALRE
  ADY_SIGNALED)
• ltDo some useful workgt // App uses CPU cycles
  instead of blocking
• Read back data in frame 2

10
Variation with asynchronous read back

• Loop
• Draw frame 1 // To a FBO, for example
• Read back frame 1 into PBO 1 // Asynchronous
  readback
• glFence(sync_object1) // Inserts a fence in the
  command stream
• Draw frame 2
• Read back frame 2 into PBO 2
• glFence(sync_object2)
• while (glClientWaitSync(sync_object1,0,0)!GL_ALRE
  ADY_SIGNALED)
• ltDo some useful workgt // App uses CPU cycles
  instead of blocking
• glMapBuffer() // Access the data of frame 1 in
  PBO 1
• while (glClientWaitSync(sync_object1,0,0)!GL_ALRE
  ADY_SIGNALED)
• ltDo some useful workgt // App uses CPU cycles
  instead of blocking
• glMapBuffer() // Access the data of frame 2 in
  PBO 2

11
Differences with GL_NV_Fence

• No separation of sync objects and fences in
  NV_Fence
• NV version only has fence objects
• Fence object has state
• Creation of sync object and inserting a fence in
  one command
• SetFenceNV creates and inserts a fence (old
  object model)
• NV Fence objects not shared across contexts

12
API Overview 1/2

• Create a sync attribute object
• object CreateSyncAttrib()
• SYNC_TYPE has to be FENCE
• SYNC_CONDITION has to be SYNC_PRIOR_COMMANDS_COMP
  LETE
• SYNC_STATUS SIGNALED or UNSIGNALED
• Create the sync object
• object CreateSync(object attrib)
• Insert a fence, associated with a sync object,
  into command stream
• void Fence(object sync)

13
API Overview 2/2

• Wait or test the status of a fence sync object
• enum ClientWaitSync(object sync, uint flags,
• time timeout)
• Blocks until sync is signalled or timeout expired
• If timeout 0, does not block, returns the
  status of sync
• If timeout FOREVER, call does not timeout
• Optionally will flush before blocking
• Returns 3 values ALREADY_SIGNALED,
  TIMEOUT_EXPIRED, CONDITION_SATISFIED
• Signal or unsignal a sync object
• void SignalSync(object sync, enum mode)
• If status transitions from unsignaled to
  signaled, ClientWaitSync will unblock

14
Example Streaming video processing

• Loop
• Draw frame 1 // To a FBO, for example
• glFence(sync_object1) // inserts a fence in the
  command stream
• Draw frame 2
• glFence(sync_object2)
• while (glClientWaitSync(sync_object1,0,0)!GL_ALRE
  ADY_SIGNALED)
• ltDo some useful workgt // App uses CPU cycles
  instead of blocking
• Read back data in frame 1
• while (glClientWaitSync(sync_object1,0,0)!GL_ALRE
  ADY_SIGNALED)
• ltDo some useful workgt // App uses CPU cycles
  instead of blocking
• Read back data in frame 2

15
Variation with asynchronous read back

• Loop
• Draw frame 1 // To a FBO, for example
• Read back frame 1 into PBO 1 // Asynchronous
  readback
• glFence(sync_object1) // Inserts a fence in the
  command stream
• Draw frame 2
• Read back frame 2 into PBO 2
• glFence(sync_object2)
• while (glClientWaitSync(sync_object1,0,0)!GL_ALRE
  ADY_SIGNALED)
• ltDo some useful workgt // App uses CPU cycles
  instead of blocking
• glMapBuffer() // Access the data of frame 1 in
  PBO 1
• while (glClientWaitSync(sync_object1,0,0)!GL_ALRE
  ADY_SIGNALED)
• ltDo some useful workgt // App uses CPU cycles
  instead of blocking
• glMapBuffer() // Access the data of frame 2 in
  PBO 2