[Mesa-dev] [PATCH 14/22] i965/vec4: fix double_to_single() for IVB/VLV

[Matt Turner]

On Thu, Jan 5, 2017 at 5:07 AM, Samuel Iglesias Gonsálvez
<siglesias at igalia.com> wrote:
> From: "Juan A. Suarez Romero" <jasuarez at igalia.com>
>
> In the generator we must generate slightly different code for
> Ivybridge/Valleview, because of the way the stride works in
> this hardware.
> ---
>  src/mesa/drivers/dri/i965/brw_vec4_generator.cpp | 26 +++++++++++++++++++++---
>  1 file changed, 23 insertions(+), 3 deletions(-)
>
> diff --git a/src/mesa/drivers/dri/i965/brw_vec4_generator.cpp b/src/mesa/drivers/dri/i965/brw_vec4_generator.cpp
> index 0eaa91b..a68e14c 100644
> --- a/src/mesa/drivers/dri/i965/brw_vec4_generator.cpp
> +++ b/src/mesa/drivers/dri/i965/brw_vec4_generator.cpp
> @@ -1936,13 +1936,28 @@ generate_code(struct brw_codegen *p,
>
>           brw_set_default_access_mode(p, BRW_ALIGN_1);
>
> -         dst.hstride = BRW_HORIZONTAL_STRIDE_2;
> +         /* When converting from DF->F, we set destination's stride as 2 as an
> +          * aligment requirement. But in IVB/VLV, each DF implicitly writes

Typo: alignment

> +          * two floats, being the first one the converted value. So we don't
> +          * need to explicitly set stride 2, but 1.
> +          */
> +         if (devinfo->gen == 7 && !devinfo->is_haswell)
> +            dst.hstride = BRW_HORIZONTAL_STRIDE_1;
> +         else
> +            dst.hstride = BRW_HORIZONTAL_STRIDE_2;
> +
>           dst.width = BRW_WIDTH_4;
>           src[0].vstride = BRW_VERTICAL_STRIDE_4;
>           src[0].width = BRW_WIDTH_4;
>           brw_MOV(p, dst, src[0]);
>
>           struct brw_reg dst_as_src = dst;
> +         /* As we have set horizontal stride 1 instead of 2 in IVB/VLV, we
> +          * need to fix it here to have the expected value.
> +          */
> +         if (devinfo->gen == 7 && !devinfo->is_haswell)
> +            dst_as_src.hstride = BRW_HORIZONTAL_STRIDE_2;
> +
>           dst.hstride = BRW_HORIZONTAL_STRIDE_1;
>           dst.width = BRW_WIDTH_8;
>           brw_MOV(p, dst, dst_as_src);
> @@ -1965,8 +1980,13 @@ generate_code(struct brw_codegen *p,
>           src[0].width = BRW_WIDTH_4;
>           brw_MOV(p, tmp, src[0]);
>
> -         tmp.vstride = BRW_VERTICAL_STRIDE_8;
> -         tmp.hstride = BRW_HORIZONTAL_STRIDE_2;
> +         if (devinfo->gen == 7 && !devinfo->is_haswell) {
> +            tmp.vstride = BRW_VERTICAL_STRIDE_4;
> +            tmp.hstride = BRW_HORIZONTAL_STRIDE_1;
> +         } else {
> +            tmp.vstride = BRW_VERTICAL_STRIDE_8;
> +            tmp.hstride = BRW_HORIZONTAL_STRIDE_2;
> +         }

With the patch I sent to replace 09/22, there should be no changes
needed to VEC4_OPCODE_TO_DOUBLE. :)

Please change double_to_single() to VEC4_OPCODE_FROM_DOUBLE in the title.