[Mesa-dev] [PATCH] radv: Unset ZRANGE_PRECISION when depth was zeroed

[Bas Nieuwenhuizen]

On Thu, Mar 8, 2018 at 12:59 PM, James Legg <jlegg at feralinteractive.com> wrote:
> This avoids bug 105396 somehow. I suspect it is a VI and GFX9 hardware
> bug which PAL calls WaTcCompatZRange, but I don't know for sure.
>
> In the VK_FORMAT_D32_SFLOAT case, TILE_STENCIL_DISABLE is not set for
> tc compatible image formats regardless of not having a stencil aspect.
> If TILE_STENCIL_DISABLE was set, ZRANGE_PRECISION would have no effect
> and the bug would occur again.
>
> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=105396
> CC: <mesa-stable at lists.freedesktop.org>
> CC: Dave Airlie <airlied at redhat.com>
> CC: Bas Nieuwenhuizen <bas at basnieuwenhuizen.nl>
> CC: Samuel Pitoiset <samuel.pitoiset at gmail.com>
> ---
>  src/amd/vulkan/radv_cmd_buffer.c | 52 +++++++++++++++++++++++++++++++++++++---
>  1 file changed, 49 insertions(+), 3 deletions(-)
>
> diff --git a/src/amd/vulkan/radv_cmd_buffer.c b/src/amd/vulkan/radv_cmd_buffer.c
> index 3e0ed0e9a9..89e31a0347 100644
> --- a/src/amd/vulkan/radv_cmd_buffer.c
> +++ b/src/amd/vulkan/radv_cmd_buffer.c
> @@ -915,6 +915,37 @@ radv_emit_fb_ds_state(struct radv_cmd_buffer *cmd_buffer,
>
>         }
>
> +       if (image->surface.htile_size)
> +       {
> +               /* If the last depth clear value was 0.0f, set ZRANGE_PRECISION
> +                * to 0 in dp_z_info for more accuracy with reverse depth; and
> +                * to avoid https://bugs.freedesktop.org/show_bug.cgi?id=105396.
> +                * Otherwise, we leave it set to 1.
> +                */
> +               radeon_emit(cmd_buffer->cs, PKT3(PKT3_COND_WRITE, 7, 0));
> +
> +               const uint32_t write_space = 0 << 8;    /* register */
> +               const uint32_t poll_space = 1 << 4;     /* memory */
> +               const uint32_t function = 3 << 0;       /* equal to the reference */
> +               const uint32_t options = write_space | poll_space | function;
> +               radeon_emit(cmd_buffer->cs, options);
> +
> +               /* poll address - location of the depth clear value */
> +               uint64_t va = radv_buffer_get_va(image->bo);
> +               va += image->offset + image->clear_value_offset;
> +               radeon_emit(cmd_buffer->cs, va);
> +               radeon_emit(cmd_buffer->cs, va >> 32);
> +
> +               radeon_emit(cmd_buffer->cs, fui(0.0f));         /* reference value */
> +               radeon_emit(cmd_buffer->cs, (uint32_t)-1);      /* comparison mask */
> +               radeon_emit(cmd_buffer->cs, R_028040_DB_Z_INFO >> 2); /* write address low */
> +               radeon_emit(cmd_buffer->cs, 0u);                /* write address high */
> +
> +               /* The value to write data when the condition passes */
> +               uint32_t db_z_info_clear_zero = db_z_info & C_028040_ZRANGE_PRECISION;
> +               radeon_emit(cmd_buffer->cs, db_z_info_clear_zero);
> +       }
> +
>         radeon_set_context_reg(cmd_buffer->cs, R_028B78_PA_SU_POLY_OFFSET_DB_FMT_CNTL,
>                                ds->pa_su_poly_offset_db_fmt_cntl);
>  }
> @@ -3479,7 +3510,8 @@ void radv_CmdEndRenderPass(
>
>  /*
>   * For HTILE we have the following interesting clear words:
> - *   0xfffff30f: Uncompressed, full depth range, for depth+stencil HTILE
> + *   0xfffff30f: Uncompressed, full depth range, for depth+stencil HTILE when ZRANGE_PRECISION is 1
> + *   0x0003f30f: Uncompressed, full depth range, for depth+stencil HTILE when ZRANGE_PRECISION is 0
>   *   0xfffc000f: Uncompressed, full depth range, for depth only HTILE.
>   *   0xfffffff0: Clear depth to 1.0
>   *   0x00000000: Clear depth to 0.0
> @@ -3528,8 +3560,22 @@ static void radv_handle_depth_image_transition(struct radv_cmd_buffer *cmd_buffe
>                 radv_initialize_htile(cmd_buffer, image, range, 0);
>         } else if (!radv_layout_is_htile_compressed(image, src_layout, src_queue_mask) &&
>                    radv_layout_is_htile_compressed(image, dst_layout, dst_queue_mask)) {
> -               uint32_t clear_value = vk_format_is_stencil(image->vk_format) ? 0xfffff30f : 0xfffc000f;
> -               radv_initialize_htile(cmd_buffer, image, range, clear_value);
> +               if (vk_format_is_stencil(image->vk_format)) {
> +                       /* The appropriate clear value depends on DB_Z_INFO's
> +                        * ZRANGE_PRECISION, which can vary depending on the
> +                        * last used clear value, which could be from another
> +                        * command buffer. Instead of picking the appropriate
> +                        * clear value on the GPU, resummarize accurately.
> +                        */
> +                       VkImageSubresourceRange local_range = *range;
> +                       local_range.aspectMask = VK_IMAGE_ASPECT_DEPTH_BIT;
> +                       local_range.baseMipLevel = 0;
> +                       local_range.levelCount = 1;
> +
> +                       radv_resummarize_depth_image_inplace(cmd_buffer, image, &local_range);

Can we instead of resummarizing, just do the reset + write the clear words?

Otherwise looks good to me, please tell if I should do it. I'd do it
in a follow up patch, except this is nominated to stable.

Apologies for the delay.

> +               } else {
> +                       radv_initialize_htile(cmd_buffer, image, range, 0xfffc000f);
> +               }
>         } else if (radv_layout_is_htile_compressed(image, src_layout, src_queue_mask) &&
>                    !radv_layout_is_htile_compressed(image, dst_layout, dst_queue_mask)) {
>                 VkImageSubresourceRange local_range = *range;
> --
> 2.14.3
>