[Intel-gfx] [PATCH v1 2/4] drm/i915: Add provision to extend Golden context batch

[Chris Wilson]

On Fri, Jul 17, 2015 at 05:08:52PM +0100, Arun Siluvery wrote:
> The Golden batch carries 3D state at the beginning so that HW starts with
> a known state. It is carried as a binary blob which is auto-generated from
> source. The idea was it would be easier to maintain and keep the complexity
> out of the kernel which makes sense as we don't really touch it. However if
> you really need to update it then you need to update generator source and
> keep the binary blob in sync with it.
> 
> There is a need to patch this in bxt to send one additional command to enable
> a feature. A solution was to patch the binary data with some additional
> data structures (included as part of auto-generator source) but it was
> unnecessarily complicated.
> 
> Chris suggested the idea of having a secondary batch and execute two batch
> buffers. It has clear advantages as we needn't touch the base golden batch,
> can customize secondary/auxiliary batch depending on Gen and can be carried
> in the driver with no dependencies.
> 
> This patch adds support for this auxiliary batch which is inserted at the
> end of golden batch and is completely independent from it. Thanks to Mika
> for the preliminary review.
> 
> Cc: Mika Kuoppala <mika.kuoppala at intel.com>
> Cc: Chris Wilson <chris at chris-wilson.co.uk>
> Cc: Armin Reese <armin.c.reese at intel.com>
> Signed-off-by: Arun Siluvery <arun.siluvery at linux.intel.com>
> ---
>  drivers/gpu/drm/i915/i915_gem_render_state.c | 27 +++++++++++++++++++++++++++
>  drivers/gpu/drm/i915/i915_gem_render_state.h |  2 ++
>  drivers/gpu/drm/i915/intel_lrc.c             |  6 ++++++
>  3 files changed, 35 insertions(+)
> 
> diff --git a/drivers/gpu/drm/i915/i915_gem_render_state.c b/drivers/gpu/drm/i915/i915_gem_render_state.c
> index b6492fe..b86e382 100644
> --- a/drivers/gpu/drm/i915/i915_gem_render_state.c
> +++ b/drivers/gpu/drm/i915/i915_gem_render_state.c
> @@ -73,6 +73,15 @@ free_gem:
>  	return ret;
>  }
>  
> +#define OUT_BATCH(batch, i, val)				\
> +	do {							\
> +		if (WARN_ON((i) >= PAGE_SIZE / sizeof(u32))) {	\

We have to be slightly more careful here, as we don't have the full page
available since we put render state into the high arena of the golden
bb. Something like WARN_ON(i > PAGE/sizeof(u32) || (batch)[i]) should
suffice.

> @@ -110,6 +119,15 @@ static int render_state_setup(struct render_state *so)
>  
>  		d[i++] = s;
>  	}
> +
> +	while (i % CACHELINE_DWORDS)
> +		OUT_BATCH(d, i, MI_NOOP);
> +
> +	so->aux_batch_offset = i * sizeof(u32);
> +
> +	OUT_BATCH(d, i, MI_BATCH_BUFFER_END);
> +	so->aux_batch_size = (i * sizeof(u32)) - so->aux_batch_offset;

Strictly, and if we are passing the batch length we are being strictly
conformant, then the aux_batch_size must be a multiple of 8.

> +
>  	kunmap(page);
>  
>  	ret = i915_gem_object_set_to_gtt_domain(so->obj, false);
> @@ -128,6 +146,8 @@ err_out:
>  	return ret;
>  }
>  
> +#undef OUT_BATCH
> +
>  void i915_gem_render_state_fini(struct render_state *so)
>  {
>  	i915_gem_object_ggtt_unpin(so->obj);
> @@ -176,6 +196,13 @@ int i915_gem_render_state_init(struct drm_i915_gem_request *req)
>  	if (ret)
>  		goto out;
>  
Then we need only execute this BB if so.aux_batch_size > 8

> +	ret = req->ring->dispatch_execbuffer(req,
> +					     (so.ggtt_offset + so.aux_batch_offset),
> +					     so.aux_batch_size,
> +					     I915_DISPATCH_SECURE);
> +	if (ret)
> +		goto out;
> +
>  	i915_vma_move_to_active(i915_gem_obj_to_ggtt(so.obj), req);
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre