[Mesa-dev] [PATCH 4/6] gallium: add way for drivers to create fences without flushing
Rob Clark
robdclark at gmail.com
Mon Apr 4 13:58:33 UTC 2016
Hmm, so we could actually use EGL_SYNC_FLUSH_COMMANDS_BIT_KHR (which
is ignored when there is not a ctx bound).. from looking closer at
KHR_fence_sync this appears to be how things are expected to work.
ofc that means that you end up needing to keep both
ctx->fence_finish() and screen->fence_finish(), or maybe extend
screen->fence_finish() to take an optional ctx ptr.. neither of which
is the prettiest thing.
BR,
-R
On Mon, Apr 4, 2016 at 8:10 AM, Marek Olšák <maraeo at gmail.com> wrote:
> There is one problem with this: It doesn't allow calling
> pipe_context::fence_finish from another thread in a thread-safe
> manner, and it implies that fence_finish is generally NOT thread-safe.
> This thread safety is something I'd like to preserve.
>
> I would say that a flush flag telling the driver not to flush would be
> better, so that drivers which have per-screen fences and cheap flushes
> can ignore it.
>
> Marek
>
> On Fri, Apr 1, 2016 at 10:29 PM, Rob Clark <robdclark at gmail.com> wrote:
>> From: Rob Clark <robclark at freedesktop.org>
>>
>> Since current thing is kinda horrible for tilers. And that issue will
>> be even worse with EGL_ANDROID_native_fence_sync.
>>
>> Not wired up yet for gl syncobj, which can come later. For now we just
>> need this with EGL.
>>
>> Signed-off-by: Rob Clark <robclark at freedesktop.org>
>> ---
>> src/gallium/include/pipe/p_context.h | 24 ++++++++++++++++++++++++
>> src/gallium/state_trackers/dri/dri2.c | 29 ++++++++++++++++++++---------
>> 2 files changed, 44 insertions(+), 9 deletions(-)
>>
>> diff --git a/src/gallium/include/pipe/p_context.h b/src/gallium/include/pipe/p_context.h
>> index 1c97e82..02a946b 100644
>> --- a/src/gallium/include/pipe/p_context.h
>> +++ b/src/gallium/include/pipe/p_context.h
>> @@ -457,6 +457,30 @@ struct pipe_context {
>> unsigned flags);
>>
>> /**
>> + * Create a fence without necessarily flushing rendering. Note
>> + * that if the driver implements this, it must also implement
>> + * ctx->fence_finish() which will be used instead of
>> + * screen->fence_finish() to give the driver an opportunity to
>> + * flush.
>> + *
>> + * This allows drivers, in particular tilers, to defer flush
>> + * until someone actually wants to wait on a fence.
>> + *
>> + * \param fence if not NULL, an old fence to unref and transfer a
>> + * new fence reference to
>> + */
>> + void (*create_fence)(struct pipe_context *pipe,
>> + struct pipe_fence_handle **fence);
>> +
>> + /**
>> + * Wait for the fence to finish.
>> + * \param timeout in nanoseconds (may be PIPE_TIMEOUT_INFINITE).
>> + */
>> + boolean (*fence_finish)(struct pipe_context *pipe,
>> + struct pipe_fence_handle *fence,
>> + uint64_t timeout);
>> +
>> + /**
>> * Create a view on a texture to be used by a shader stage.
>> */
>> struct pipe_sampler_view * (*create_sampler_view)(struct pipe_context *ctx,
>> diff --git a/src/gallium/state_trackers/dri/dri2.c b/src/gallium/state_trackers/dri/dri2.c
>> index fb0a180..b66d885 100644
>> --- a/src/gallium/state_trackers/dri/dri2.c
>> +++ b/src/gallium/state_trackers/dri/dri2.c
>> @@ -1320,7 +1320,12 @@ dri2_create_fence(__DRIcontext *_ctx)
>> if (!fence)
>> return NULL;
>>
>> - ctx->flush(ctx, &fence->pipe_fence, 0);
>> + if (ctx->create_fence) {
>> + debug_assert(ctx->fence_finish);
>> + ctx->create_fence(ctx, &fence->pipe_fence);
>> + } else {
>> + ctx->flush(ctx, &fence->pipe_fence, 0);
>> + }
>>
>> if (!fence->pipe_fence) {
>> FREE(fence);
>> @@ -1376,27 +1381,33 @@ static GLboolean
>> dri2_client_wait_sync(__DRIcontext *_ctx, void *_fence, unsigned flags,
>> uint64_t timeout)
>> {
>> + struct pipe_context *ctx = dri_context(_ctx)->st->pipe;
>> struct dri2_fence *fence = (struct dri2_fence*)_fence;
>> struct dri_screen *driscreen = fence->driscreen;
>> struct pipe_screen *screen = driscreen->base.screen;
>> + struct pipe_fence_handle *pipe_fence = NULL;
>>
>> - /* No need to flush. The context was flushed when the fence was created. */
>> + /* No need to flush. The context was flushed when the fence was created,
>> + * or the ctx implements ctx->fence_finish() which will take care of
>> + * flushing if required
>> + */
>>
>> if (fence->pipe_fence)
>> - return screen->fence_finish(screen, fence->pipe_fence, timeout);
>> + pipe_fence = fence->pipe_fence;
>> else if (fence->cl_event) {
>> - struct pipe_fence_handle *pipe_fence =
>> - driscreen->opencl_dri_event_get_fence(fence->cl_event);
>> -
>> - if (pipe_fence)
>> - return screen->fence_finish(screen, pipe_fence, timeout);
>> - else
>> + pipe_fence = driscreen->opencl_dri_event_get_fence(fence->cl_event);
>> + if (!pipe_fence)
>> return driscreen->opencl_dri_event_wait(fence->cl_event, timeout);
>> }
>> else {
>> assert(0);
>> return false;
>> }
>> +
>> + if (ctx->fence_finish)
>> + return ctx->fence_finish(ctx, pipe_fence, timeout);
>> +
>> + return screen->fence_finish(screen, pipe_fence, timeout);
>> }
>>
>> static void
>> --
>> 2.5.5
>>
>> _______________________________________________
>> mesa-dev mailing list
>> mesa-dev at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
More information about the mesa-dev
mailing list