[RFC PATCH v4 22/42] drm/vkms: Use s32 for internal color pipeline precision

[Pekka Paalanen]

On Mon, 26 Feb 2024 16:10:36 -0500
Harry Wentland <harry.wentland at amd.com> wrote:

> Certain operations require us to preserve values below 0.0 and
> above 1.0 (0x0 and 0xffff respectively in 16 bpc unorm). One
> such operation is a BT709 encoding operation followed by its
> decoding operation, or the reverse.
> 
> We'll use s32 values as intermediate in and outputs of our
> color operations, for the operations where it matters.
> 
> For now this won't apply to LUT operations. We might want to
> update those to work on s32 as well, but it's unclear how
> that should work for unorm LUT definitions. We'll revisit
> that once we add LUT + CTM tests.
> 
> In order to allow for this we'll also invert the nesting of our
> colorop processing loops. We now use the pixel iteration loop
> on the outside and the colorop iteration on the inside.

I always go through the same thought process:

- putting the pixel iteration loop outermost is going to be horrible
  for performance

- maybe should turn the temporary line buffer to argb_s32 for all of
  VKMS

- that's going to be a lot of memory traffic... maybe your way is not
  horrible for performance

- maybe processing pixel by pixel is good, if you can prepare the
  operation in advance, so you don't need to analyze colorops again and
  again on each pixel

- eh, maybe it's not a gain after all, needs benchmarking

My comment on patch 17 was like the step 0 of this train of thought. I
just always get the same comments in my mind when seeing the same code
again.


Thanks,
pq

> 
> v4:
>  - Clarify that we're packing 16-bit UNORM into s32, not
>    converting values to a different representation (Pekka)
> 
> v3:
>  - Use new colorop->next pointer
> 
> Signed-off-by: Harry Wentland <harry.wentland at amd.com>
> ---
>  drivers/gpu/drm/vkms/vkms_composer.c | 57 +++++++++++++++++++++-------
>  drivers/gpu/drm/vkms/vkms_drv.h      |  4 ++
>  2 files changed, 48 insertions(+), 13 deletions(-)
> 
> diff --git a/drivers/gpu/drm/vkms/vkms_composer.c b/drivers/gpu/drm/vkms/vkms_composer.c
> index 25b786b49db0..d2101fa55aa3 100644
> --- a/drivers/gpu/drm/vkms/vkms_composer.c
> +++ b/drivers/gpu/drm/vkms/vkms_composer.c
> @@ -164,7 +164,7 @@ static void apply_lut(const struct vkms_crtc_state *crtc_state, struct line_buff
>  	}
>  }
>  
> -static void apply_colorop(struct pixel_argb_u16 *pixel, struct drm_colorop *colorop)
> +static void apply_colorop(struct pixel_argb_s32 *pixel, struct drm_colorop *colorop)
>  {
>  	/* TODO is this right? */
>  	struct drm_colorop_state *colorop_state = colorop->state;
> @@ -191,25 +191,56 @@ static void apply_colorop(struct pixel_argb_u16 *pixel, struct drm_colorop *colo
>  
>  static void pre_blend_color_transform(const struct vkms_plane_state *plane_state, struct line_buffer *output_buffer)
>  {
> -	struct drm_colorop *colorop = plane_state->base.base.color_pipeline;
> +	struct drm_colorop *colorop;
> +	struct pixel_argb_s32 pixel;
>  
> -	while (colorop) {
> -		struct drm_colorop_state *colorop_state;
> +	for (size_t x = 0; x < output_buffer->n_pixels; x++) {
>  
> -		if (!colorop)
> -			return;
> +		/*
> +		 * Some operations, such as applying a BT709 encoding matrix,
> +		 * followed by a decoding matrix, require that we preserve
> +		 * values above 1.0 and below 0.0 until the end of the pipeline.
> +		 *
> +		 * Pack the 16-bit UNORM values into s32 to give us head-room to
> +		 * avoid clipping until we're at the end of the pipeline. Clip
> +		 * intentionally at the end of the pipeline before packing
> +		 * UNORM values back into u16.
> +		 */
> +		pixel.a = output_buffer->pixels[x].a;
> +		pixel.r = output_buffer->pixels[x].r;
> +		pixel.g = output_buffer->pixels[x].g;
> +		pixel.b = output_buffer->pixels[x].b;
>  
> -		/* TODO this is probably wrong */
> -		colorop_state = colorop->state;
> +		colorop = plane_state->base.base.color_pipeline;
> +		while (colorop) {
> +			struct drm_colorop_state *colorop_state;
>  
> -		if (!colorop_state)
> -			return;
> +			if (!colorop)
> +				return;
> +
> +			/* TODO this is probably wrong */
> +			colorop_state = colorop->state;
> +
> +			if (!colorop_state)
> +				return;
>  
> -		for (size_t x = 0; x < output_buffer->n_pixels; x++)
>  			if (!colorop_state->bypass)
> -				apply_colorop(&output_buffer->pixels[x], colorop);
> +				apply_colorop(&pixel, colorop);
>  
> -		colorop = colorop->next;
> +			colorop = colorop->next;
> +		}
> +
> +		/* clamp pixel */
> +		pixel.a = max(min(pixel.a, 0xffff), 0x0);
> +		pixel.r = max(min(pixel.r, 0xffff), 0x0);
> +		pixel.g = max(min(pixel.g, 0xffff), 0x0);
> +		pixel.b = max(min(pixel.b, 0xffff), 0x0);
> +
> +		/* put back to output_buffer */
> +		output_buffer->pixels[x].a = pixel.a;
> +		output_buffer->pixels[x].r = pixel.r;
> +		output_buffer->pixels[x].g = pixel.g;
> +		output_buffer->pixels[x].b = pixel.b;
>  	}
>  }
>  
> diff --git a/drivers/gpu/drm/vkms/vkms_drv.h b/drivers/gpu/drm/vkms/vkms_drv.h
> index 2bcc24c196a2..fadb7685a360 100644
> --- a/drivers/gpu/drm/vkms/vkms_drv.h
> +++ b/drivers/gpu/drm/vkms/vkms_drv.h
> @@ -36,6 +36,10 @@ struct vkms_frame_info {
>  	unsigned int cpp;
>  };
>  
> +struct pixel_argb_s32 {
> +	s32 a, r, g, b;
> +};
> +
>  struct pixel_argb_u16 {
>  	u16 a, r, g, b;
>  };

-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 833 bytes
Desc: OpenPGP digital signature
URL: <https://lists.freedesktop.org/archives/amd-gfx/attachments/20240312/eeb490d5/attachment.sig>