[Mesa-dev] [PATCH v2 1/1] clover: Wait for requested operation if blocking flag is set
Francisco Jerez
currojerez at riseup.net
Tue Aug 15 19:00:48 UTC 2017
Jan Vesely <jan.vesely at rutgers.edu> writes:
> On Sat, 2017-08-12 at 20:14 -0700, Francisco Jerez wrote:
>> Jan Vesely <jan.vesely at rutgers.edu> writes:
>>
>> > On Sat, 2017-08-05 at 12:34 -0700, Francisco Jerez wrote:
>> > > Francisco Jerez <currojerez at riseup.net> writes:
>> > >
>> > > > Jan Vesely <jan.vesely at rutgers.edu> writes:
>> > > >
>> > > > > Hi,
>> > > > >
>> > > > > thanks for detailed explanation. I indeed missed the writeBuffer part
>> > > > > in specs.
>> > > > >
>> > > > > On Wed, 2017-08-02 at 15:05 -0700, Francisco Jerez wrote:
>> > > > > > These changes are somewhat redundant and potentially
>> > > > > > performance-impacting, the reason is that in the OpenCL API,
>> > > > > > clEnqueueWrite* commands are specified to block until the memory
>> > > > > > provided by the application as origin can be reused safely (i.e. until
>> > > > > > soft_copy_op()() runs), not necessarily until the transfer to graphics
>> > > > > > memory has completed (which is what hard_event::wait() will wait for).
>> > > > > > OTOH reads and maps as implemented by soft_copy_op and friends are
>> > > > > > essentially blocking so the wait() call is redundant in most cases.
>> > > > >
>> > > > > That explains a noticeable slowdown running piglit with these changes.
>> > > > > I'm not sure about the read part though. I expected it to be basically
>> > > > > a noop, but it changes behaviour.
>> > > >
>> > > > I think this change would have slowed you down the most whenever the
>> > > > mapping operation performed by soft_copy_op() is able to proceed
>> > > > immediately, either because the buffer is idle (so the driver doesn't
>> > > > stall on transfer_map()) *or* because the driver is trying to be smart
>> > > > and creates a bounce buffer where data can be copied into immediately
>> > > > without stalling, then inserts a pipelined GPU copy from the bounce
>> > > > buffer into the real buffer. With this patch you will stall on the GPU
>> > > > copy regardless (and whatever other work was already on the command
>> > > > stream which might be substantial), even though it wouldn't have been
>> > > > necessary in any of these cases.
>> > > >
>> > > > > Adding clGetEventInfo(CL_EVENT_COMMAND_EXECUTION_STATUS) after a
>> > > > > blocking read in one of the piglit tests surprisingly returns
>> > > > > CL_QUEUED.
>> > > > >
>> > > >
>> > > > Hmm, yeah, that seems kind of debatable behaviour, although it's
>> > > > definitely legit for writes, not quite sure for reads... I believe the
>> > > > reason why that happens is because the CPU copy proceeds very quickly
>> > > > (due to the reasons mentioned in the last paragraph), but the hard_event
>> > > > is still associated with a pipe_fence synchronous with the GPU's command
>> > > > stream, so it won't get signalled until the GPU catches up.
>> >
>> > Hi,
>> > sorry for the delay, last week was submission week...
>> >
>>
>> No worries.
>>
>> > The part that I'm still missing is what kind of GPU work needs to be
>> > done after clEnqueueRead*(). I assume all necessary work is completed
>> > before I can access the data.
>> > Also CL_QUEUED status was surprising. since we performed at least some
>> > of the work (we got the data), I'd expect CL_RUNNING or CL_SUBMITTED at
>> > least.
>> >
>>
>> The lag is not due to GPU work that needs to be performed after the
>> clEnqueueRead call, but due to GPU work that may precede it in the
>> command stream: Because clover doesn't know when the last time was that
>> the buffer was referenced by GPU work, the event is associated with a
>> fence synchronous with the current batch (which obviously won't signal
>> before any of the GPU work that actually referenced the buffer
>> completes). However the pipe driver has a more accurate idea of when
>> the buffer was used last, so the transfer_map() operation is likely to
>> complete more quickly than the CL event status changes to CL_COMPLETE.
>> The reason you're seeing CL_QUEUED rather than CL_SUBMITTED is because
>> the read operation didn't even need to flush the current batch, so
>> there's no fence associated with the CL event object yet.
>
> thanks. so the event is waiting for the current batch, even if the
> buffer access is done out of order. The question is, why do we use the
> fence at all? If I understood correctly mapping the buffer will be
> delayed by the pipe driver if needed, so we don't really need the
> fence. Am I missing something?
>
The current CL event object implementation needs a gallium fence object
in order to find out what status it's in and in order to implement
waiting. The problem is that it doesn't have any fence object available
that is close "enough" to the actual GPU operation that modified the
buffer for the last time. There are several more or less hairy ways to
improve this, e.g. using a soft_event that's manually signaled when
triggered instead of a hard_event (though I'm not sure whether this
would break other CL APIs that may rely on hard_event functionality),
using a different event subclass that doesn't use fences and is instead
considered immediately signaled as soon as it's triggered, using the
current hard_event implementation with a dummy known-signaled pipe
fence, or adding per-buffer fence tracking so each memory object is able
to remember which fence was the last to reference the buffer in the
command stream. Most of these would involve substantial work relative
to the slight benefit achieved (which is dependent on your definition of
the fence being "close enough" to the relevant GPU work).
> Jan
>
>>
>> > > >
>> > > > > The specs don't mention use of events with blocking read, but it does
>> > > > > say that "When the read command has completed, the contents of the
>> > > > > buffer that ptr points to can be used by the application." in the non-
>> > > > > blocking section. I'd say that the expectation is for the event to be
>> > > > > CL_COMPLETE after blocking read (at least beignet/pocl/intel-cpu-sdk
>> > > > > follow this).
>> > > > >
>> > > > > >
>> > > > > > The only reason why it might be useful to behave differently on blocking
>> > > > > > transfers is that the application may have specified a user event in the
>> > > > > > event dependency list, which may cause the soft_copy_op() call to be
>> > > > > > delayed until the application signals the user event. In order to fix
>> > > > > > that it should really only be necessary to wait for the event action to
>> > > > > > be executed, not necessarily its associated GPU work.
>> > > > >
>> > > > > I think that another use is that non-blocking writes do not create an
>> > > > > extra copy of the buffer. Thus
>> > > > > clEnqueueWriteBuffer(...,cl_false, ev, ...)
>> > > > > clWaitForEvents(ev)
>> > > > > is more memory efficient.
>> > > > >
>> > > > > >
>> > > > > > Last time this issue came up (yeah it's not the first time) I proposed
>> > > > > > the patches below to add a mechanism to wait for the event action only,
>> > > > > > feel free to include it as PATCH 0.1 and 0.2 of this series (it's been a
>> > > > > > while so they may no longer apply cleanly).
>> > > > >
>> > > > > I think we can just add comments explaining why the blocking argument
>> > > > > is ignored, until someone chooses to fix this problem
>> > > >
>> > > > I think the problem is definitely worth fixing, and it shouldn't really
>> > > > take more effort than adding comments explaining the current behaviour
>> > > > ;), basically just add a bunch of 'if (blocking)
>> > > > hev().wait_signalled();' where the spec requires it, roughly as you had
>> > > > been doing in this patch, but wait_signalled() should only stall on the
>> > > > CPU work associated with the event, which should give you the same
>> > > > performance as the current approach.
>> >
>> > I can send a patch that replaces wait() -> wait_signalled()
>> >
>>
>> Thanks :)
>>
>> > > >
>> > > > > and/or to
>> > > > > implement proper non-blocking variants (would std::async work for
>> > > > > trivial cases like ReadBuffer?)
>> > > > >
>> > >
>> > > Hm, and to answer this question -- Yeah, std::async would probably work,
>> > > but I'm not certain whether it would actually perform better than the
>> > > current approach, because on the one hand the actual DMA-ing of the
>> > > buffer is likely to happen quasi-asynchronously already assuming the
>> > > driver is competent, and OTOH because spawning a new thread for the copy
>> > > would introduce additional overhead that might defeat your purpose
>> > > unless the copy is very large -- Only experimentation will tell whether
>> > > it pays off.
>> >
>> > it was just a speculation. it looks like Vedran is interested in
>> > implementing non-blocking reads/writes[0] so I'll leave it to him. r600
>> > has bigger problems elsewhere atm.
>> >
>>
>> Yeah, I'm aware of his work, I suspect the above are the reasons why he
>> got rather mixed performance results from his changes.
>>
>> > thanks,
>> > Jan
>> >
>> > [0]https://bugs.freedesktop.org/show_bug.cgi?id=100199
>> >
>> > >
>> > > > > thanks,
>> > > > > Jan
>> > > > >
>> > > > > >
>> > > > > > Thank you.
>> > > > > >
>> > > > > > Jan Vesely <jan.vesely at rutgers.edu> writes:
>> > > > > >
>> > > > > > > v2: wait in map_buffer and map_image as well
>> > > > > > >
>> > > > > > > Signed-off-by: Jan Vesely <jan.vesely at rutgers.edu>
>> > > > > > > ---
>> > > > > > > Hi Aaron,
>> > > > > > >
>> > > > > > > yes, I think you're right, we should wait in Map* as well.
>> > > > > > > If nothing else it's consistent, even if passing the flag to
>> > > > > > > add_map might make it unnecessary (haven't actually checked).
>> > > > > > >
>> > > > > > > thanks,
>> > > > > > > Jan
>> > > > > > >
>> > > > > > > src/gallium/state_trackers/clover/api/transfer.cpp | 30 ++++++++++++++++++++--
>> > > > > > > 1 file changed, 28 insertions(+), 2 deletions(-)
>> > > > > > >
>> > > > > > > diff --git a/src/gallium/state_trackers/clover/api/transfer.cpp b/src/gallium/state_trackers/clover/api/transfer.cpp
>> > > > > > > index f7046253be..729a34590e 100644
>> > > > > > > --- a/src/gallium/state_trackers/clover/api/transfer.cpp
>> > > > > > > +++ b/src/gallium/state_trackers/clover/api/transfer.cpp
>> > > > > > > @@ -295,6 +295,9 @@ clEnqueueReadBuffer(cl_command_queue d_q, cl_mem d_mem, cl_bool blocking,
>> > > > > > > &mem, obj_origin, obj_pitch,
>> > > > > > > region));
>> > > > > > >
>> > > > > > > + if (blocking)
>> > > > > > > + hev().wait();
>> > > > > > > +
>> > > > > > > ret_object(rd_ev, hev);
>> > > > > > > return CL_SUCCESS;
>> > > > > > >
>> > > > > > > @@ -325,6 +328,9 @@ clEnqueueWriteBuffer(cl_command_queue d_q, cl_mem d_mem, cl_bool blocking,
>> > > > > > > ptr, {}, obj_pitch,
>> > > > > > > region));
>> > > > > > >
>> > > > > > > + if (blocking)
>> > > > > > > + hev().wait();
>> > > > > > > +
>> > > > > > > ret_object(rd_ev, hev);
>> > > > > > > return CL_SUCCESS;
>> > > > > > >
>> > > > > > > @@ -362,6 +368,9 @@ clEnqueueReadBufferRect(cl_command_queue d_q, cl_mem d_mem, cl_bool blocking,
>> > > > > > > &mem, obj_origin, obj_pitch,
>> > > > > > > region));
>> > > > > > >
>> > > > > > > + if (blocking)
>> > > > > > > + hev().wait();
>> > > > > > > +
>> > > > > > > ret_object(rd_ev, hev);
>> > > > > > > return CL_SUCCESS;
>> > > > > > >
>> > > > > > > @@ -399,6 +408,9 @@ clEnqueueWriteBufferRect(cl_command_queue d_q, cl_mem d_mem, cl_bool blocking,
>> > > > > > > ptr, host_origin, host_pitch,
>> > > > > > > region));
>> > > > > > >
>> > > > > > > + if (blocking)
>> > > > > > > + hev().wait();
>> > > > > > > +
>> > > > > > > ret_object(rd_ev, hev);
>> > > > > > > return CL_SUCCESS;
>> > > > > > >
>> > > > > > > @@ -504,6 +516,9 @@ clEnqueueReadImage(cl_command_queue d_q, cl_mem d_mem, cl_bool blocking,
>> > > > > > > &img, src_origin, src_pitch,
>> > > > > > > region));
>> > > > > > >
>> > > > > > > + if (blocking)
>> > > > > > > + hev().wait();
>> > > > > > > +
>> > > > > > > ret_object(rd_ev, hev);
>> > > > > > > return CL_SUCCESS;
>> > > > > > >
>> > > > > > > @@ -538,6 +553,9 @@ clEnqueueWriteImage(cl_command_queue d_q, cl_mem d_mem, cl_bool blocking,
>> > > > > > > ptr, {}, src_pitch,
>> > > > > > > region));
>> > > > > > >
>> > > > > > > + if (blocking)
>> > > > > > > + hev().wait();
>> > > > > > > +
>> > > > > > > ret_object(rd_ev, hev);
>> > > > > > > return CL_SUCCESS;
>> > > > > > >
>> > > > > > > @@ -667,7 +685,11 @@ clEnqueueMapBuffer(cl_command_queue d_q, cl_mem d_mem, cl_bool blocking,
>> > > > > > >
>> > > > > > > void *map = mem.resource(q).add_map(q, flags, blocking, obj_origin, region);
>> > > > > > >
>> > > > > > > - ret_object(rd_ev, create<hard_event>(q, CL_COMMAND_MAP_BUFFER, deps));
>> > > > > > > + auto hev = create<hard_event>(q, CL_COMMAND_MAP_BUFFER, deps);
>> > > > > > > + if (blocking)
>> > > > > > > + hev().wait();
>> > > > > > > +
>> > > > > > > + ret_object(rd_ev, hev);
>> > > > > > > ret_error(r_errcode, CL_SUCCESS);
>> > > > > > > return map;
>> > > > > > >
>> > > > > > > @@ -695,7 +717,11 @@ clEnqueueMapImage(cl_command_queue d_q, cl_mem d_mem, cl_bool blocking,
>> > > > > > >
>> > > > > > > void *map = img.resource(q).add_map(q, flags, blocking, origin, region);
>> > > > > > >
>> > > > > > > - ret_object(rd_ev, create<hard_event>(q, CL_COMMAND_MAP_IMAGE, deps));
>> > > > > > > + auto hev = create<hard_event>(q, CL_COMMAND_MAP_IMAGE, deps);
>> > > > > > > + if (blocking)
>> > > > > > > + hev().wait();
>> > > > > > > +
>> > > > > > > + ret_object(rd_ev, hev);
>> > > > > > > ret_error(r_errcode, CL_SUCCESS);
>> > > > > > > return map;
>> > > > > > >
>> > > > > > > --
>> > > > > > > 2.13.3
>> > > > > >
>> > > > > > _______________________________________________
>> > > > > > mesa-dev mailing list
>> > > > > > mesa-dev at lists.freedesktop.org
>> > > > > > https://lists.freedesktop.org/mailman/listinfo/mesa-dev
>> > > > >
>> > > > > --
>> > > > > Jan Vesely <jan.vesely at rutgers.edu>
>> >
>> > --
>> > Jan Vesely <jan.vesely at rutgers.edu>
>>
>> _______________________________________________
>> mesa-dev mailing list
>> mesa-dev at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
>
> --
> Jan Vesely <jan.vesely at rutgers.edu>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 212 bytes
Desc: not available
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20170815/0ff286ce/attachment-0001.sig>
More information about the mesa-dev
mailing list