[Mesa-dev] [PATCH V2 09/22] i965/gen9: Set vertical alignment for the miptree

[Anuj Phogat]

On Wed, May 27, 2015 at 12:04 PM, Ben Widawsky <ben at bwidawsk.net> wrote:
> On Fri, Apr 17, 2015 at 04:51:30PM -0700, Anuj Phogat wrote:
>> Signed-off-by: Anuj Phogat <anuj.phogat at gmail.com>
>> ---
>>  src/mesa/drivers/dri/i965/brw_tex_layout.c | 74 ++++++++++++++++++++++++++++++
>>  1 file changed, 74 insertions(+)
>>
>> diff --git a/src/mesa/drivers/dri/i965/brw_tex_layout.c b/src/mesa/drivers/dri/i965/brw_tex_layout.c
>> index 20dcc21..9342101 100644
>> --- a/src/mesa/drivers/dri/i965/brw_tex_layout.c
>> +++ b/src/mesa/drivers/dri/i965/brw_tex_layout.c
>> @@ -170,6 +170,69 @@ intel_horizontal_texture_alignment_unit(struct brw_context *brw,
>>  }
>>
>>  static unsigned int
>> +tr_mode_vertical_texture_alignment(struct brw_context *brw,
>> +                                   struct intel_mipmap_tree *mt)
>> +{
>> +   const unsigned *align_yf, *align_ys;
>> +   /* Bits per pixel/element. */
>> +   const unsigned bpp = _mesa_get_format_bytes(mt->format) * 8;
>> +
>> +   /* Vertical alignment tables for TRMODE_YF and TRMODE_YS. */
>> +   const unsigned align_2d_yf[] = {64, 32, 32, 16, 16};
>> +   const unsigned align_2d_ys[] = {256, 128, 128, 64, 64};
>> +   const unsigned align_3d_yf[] = {16, 16, 16, 8, 8};
>> +   const unsigned align_3d_ys[] = {32, 32, 32, 16, 16};
>> +   int i = 0;
>> +
>> +   assert(brw->gen >= 9 &&
>> +          mt->target != GL_TEXTURE_1D &&
>> +          mt->target != GL_TEXTURE_1D_ARRAY);
>> +
>> +   /* Alignment computations below assume bpp >= 8 and a power of 2. */
>> +   assert (bpp >= 8 && bpp <= 128 && (bpp & (bpp - 1)) == 0);
>> +
>> +   switch(mt->target) {
>> +   case GL_TEXTURE_2D:
>> +   case GL_TEXTURE_RECTANGLE:
>> +   case GL_TEXTURE_2D_ARRAY:
>> +   case GL_TEXTURE_CUBE_MAP:
>> +   case GL_TEXTURE_CUBE_MAP_ARRAY:
>> +   case GL_TEXTURE_2D_MULTISAMPLE:
>> +   case GL_TEXTURE_2D_MULTISAMPLE_ARRAY:
>> +      align_yf = align_2d_yf;
>> +      align_ys = align_2d_ys;
>> +      break;
>> +   case GL_TEXTURE_3D:
>> +      align_yf = align_3d_yf;
>> +      align_ys = align_3d_ys;
>> +      break;
>> +   default:
>> +      unreachable("not reached");
>> +   }
>> +
>> +   /* Compute array index. */
>> +   while (bpp >> (i + 4))
>> +      i++;
>> +
>> +   if (mt->num_samples > 1) {
>> +      const unsigned ms_align_div[] = {1, 2, 2, 4};
>> +      int j = 0;
>> +      /* num_samples must be power of 2. */
>> +      assert((mt->num_samples & (mt->num_samples -1)) == 0);
>> +
>> +      /* Compute array index. */
>> +      while (mt->num_samples >> (j + 2))
>> +         j++;
>> +      return mt->tr_mode == INTEL_MIPTREE_TRMODE_YF ?
>> +             align_yf[i] / ms_align_div[j] :
>> +             align_ys[i] / ms_align_div[j];
>> +   }
>> +
>> +   return (mt->tr_mode == INTEL_MIPTREE_TRMODE_YF ? align_yf[i]
>> +                                                  : align_ys[i]);
>> +}
>> +
>> +static unsigned int
>
> I think all the comments I had for horizontal alignments apply here too.
>
>>  intel_vertical_texture_alignment_unit(struct brw_context *brw,
>>                                        struct intel_mipmap_tree *mt)
>>  {
>> @@ -202,6 +265,17 @@ intel_vertical_texture_alignment_unit(struct brw_context *brw,
>>     if (mt->format == MESA_FORMAT_S_UINT8)
>>        return brw->gen >= 7 ? 8 : 4;
>>
>> +   /* Vertical texture alignment is not relevant for 1D texture targets. */
>> +   if (brw->gen >= 9 && (mt->target == GL_TEXTURE_1D ||
>> +       mt->target == GL_TEXTURE_1D_ARRAY))
>> +      return 1;
>
> I think this hunk can go away with Neil's patches since then.
>
right. I'll remove it.
>> +
>> +   if (mt->tr_mode != INTEL_MIPTREE_TRMODE_NONE) {
>> +      uint32_t align = tr_mode_vertical_texture_alignment(brw, mt);
>> +      /* XY_FAST_COPY_BLT doesn't support vertical alignment < 64 */
>> +      return align < 64 ? 64 : align;
>> +   }
>> +
>
> Like the HALIGN, I don't see this restriction, can you tell me where it is.
>
>>     /* Broadwell only supports VALIGN of 4, 8, and 16.  The BSpec says 4
>>      * should always be used, except for stencil buffers, which should be 8.
>>      */
>
> Reviewed-by: Ben Widawsky <ben at bwidawsk.net>