[Mesa-dev] [PATCH V2 08/11] anv/cmd_buffer: Add code for performing HZ operations

Nanley Chery nanleychery at gmail.com
Tue Oct 4 18:32:19 UTC 2016


On Mon, Oct 03, 2016 at 05:23:27PM -0700, Jason Ekstrand wrote:
> On Mon, Sep 26, 2016 at 5:10 PM, Nanley Chery <nanleychery at gmail.com> wrote:
> 
> > Create a function that performs one of three HiZ operations -
> > depth/stencil clears, HiZ resolve, and depth resolves.
> >
> > Signed-off-by: Nanley Chery <nanley.g.chery at intel.com>
> >
> > ---
> >
> > v2. Add documentation
> >     Fix the alignment check
> >     Don't minify clear rectangle (Jason)
> >     Use blorp enums (Jason)
> >     Enable depth stalls and flushes
> >     Use full RT rectangle for resolve ops
> >     Add stencil clear todo
> >
> >  src/intel/vulkan/anv_genX.h        |   3 +
> >  src/intel/vulkan/gen7_cmd_buffer.c |   6 ++
> >  src/intel/vulkan/gen8_cmd_buffer.c | 167 ++++++++++++++++++++++++++++++
> > +++++++
> >  3 files changed, 176 insertions(+)
> >
> > diff --git a/src/intel/vulkan/anv_genX.h b/src/intel/vulkan/anv_genX.h
> > index 02e79c2..ad3bec9 100644
> > --- a/src/intel/vulkan/anv_genX.h
> > +++ b/src/intel/vulkan/anv_genX.h
> > @@ -58,6 +58,9 @@ genX(emit_urb_setup)(struct anv_device *device, struct
> > anv_batch *batch,
> >                       unsigned vs_entry_size, unsigned gs_entry_size,
> >                       const struct gen_l3_config *l3_config);
> >
> > +void genX(cmd_buffer_do_hz_op)(struct anv_cmd_buffer *cmd_buffer,
> > +                               enum blorp_hiz_op op);
> > +
> >  VkResult
> >  genX(graphics_pipeline_create)(VkDevice _device,
> >                                 struct anv_pipeline_cache *cache,
> > diff --git a/src/intel/vulkan/gen7_cmd_buffer.c
> > b/src/intel/vulkan/gen7_cmd_buffer.c
> > index b627ef0..78b5ac7 100644
> > --- a/src/intel/vulkan/gen7_cmd_buffer.c
> > +++ b/src/intel/vulkan/gen7_cmd_buffer.c
> > @@ -323,6 +323,12 @@ genX(cmd_buffer_flush_dynamic_state)(struct
> > anv_cmd_buffer *cmd_buffer)
> >     cmd_buffer->state.dirty = 0;
> >  }
> >
> > +void
> > +genX(cmd_buffer_do_hz_op)(struct anv_cmd_buffer *cmd_buffer,
> > +                          enum blorp_hiz_op op)
> > +{
> >
> 
> This should have an anv_finishme in it.
> 
> 

I'll add this in V3.

> > +}
> > +
> >  void genX(CmdSetEvent)(
> >      VkCommandBuffer                             commandBuffer,
> >      VkEvent                                     event,
> > diff --git a/src/intel/vulkan/gen8_cmd_buffer.c
> > b/src/intel/vulkan/gen8_cmd_buffer.c
> > index 7058608..a13413c 100644
> > --- a/src/intel/vulkan/gen8_cmd_buffer.c
> > +++ b/src/intel/vulkan/gen8_cmd_buffer.c
> > @@ -399,6 +399,173 @@ genX(cmd_buffer_flush_compute_state)(struct
> > anv_cmd_buffer *cmd_buffer)
> >     genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
> >  }
> >
> > +
> > +/**
> > + * Emit the HZ_OP packet in the sequence specified by the BDW PRM section
> > + * entitled: "Optimized Depth Buffer Clear and/or Stencil Buffer Clear."
> > + *
> > + * \todo Enable Stencil Buffer-only clears
> > + */
> > +void
> > +genX(cmd_buffer_do_hz_op)(struct anv_cmd_buffer *cmd_buffer,
> > +                          enum blorp_hiz_op op)
> > +{
> > +   struct anv_cmd_state *cmd_state = &cmd_buffer->state;
> > +   const struct anv_image_view *iview =
> > +      anv_cmd_buffer_get_depth_stencil_view(cmd_buffer);
> > +
> > +   if (iview == NULL || !anv_image_has_hiz(iview->image))
> > +      return;
> > +
> > +   const uint32_t ds = cmd_state->subpass->depth_stencil_attachment;
> > +   const bool full_surface_op =
> > +             cmd_state->render_area.extent.width == iview->extent.width
> > &&
> > +             cmd_state->render_area.extent.height ==
> > iview->extent.height;
> >
> 
> It's probably a bit redundant, but we might as well check
> render_area.offset == 0.  I realize that, from API requirements, if the
> extents match then offset must be 0, but it's not incredibly obvious and
> the check won't hurt that much.
> 
> 

Sure. I'll add an assertion and a comment.

> > +
> > +   /* Validate that we can perform the HZ operation and that it's
> > necessary. */
> > +   switch (op) {
> > +   case BLORP_HIZ_OP_DEPTH_CLEAR:
> > +      if (cmd_buffer->state.pass->attachments[ds].load_op !=
> > +          VK_ATTACHMENT_LOAD_OP_CLEAR)
> > +         return;
> > +
> > +      /* Apply alignment restrictions. Despite the BDW PRM mentioning
> > this is
> > +       * only needed for a depth buffer surface type of D16_UNORM, testing
> > +       * showed it to be necessary for other depth formats as well
> > +       * (e.g., D32_FLOAT).
> > +       */
> > +      if (!full_surface_op) {
> > +
> > +         struct isl_extent2d px_dim;
> >
> 
> Would it be better to call this hiz_block_size_px?  That follows the ISL
> naming convention a bit better.
> 
> 

The variable naming corresponds to the table headers below. I'll add
a comment about this.

> > +#if GEN_GEN == 8
> >
> 
> Mind making this <= 8?  I know it's in a gen8+ file, but <= 8 makes it
> clear that the else case is >= 9 and not != 8.
> 
> 

I see your point. I'll change the #else to #elif GEN_GEN >=9 as
I don't want to make the if case unspecific to make the else
case specific.

> > +         /* Pre-SKL, HiZ has an 8x4 sample block. As the number of samples
> > +          * increases, the number of pixels representable by this block
> > +          * decreases by a factor of the sample dimensions. Sample
> > dimensions
> > +          * scale following the MSAA interleaved pattern.
> > +          *
> > +          * Sample|Sample|Pixel
> > +          * Count |Dim   |Dim
> > +          * ===================
> > +          *    1  | 1x1  | 8x4
> > +          *    2  | 2x1  | 4x4
> > +          *    4  | 2x2  | 4x2
> > +          *    8  | 4x2  | 2x2
> > +          *   16  | 4x4  | 2x1
> > +          *
> > +          * Table: Pixel Dimensions in a HiZ Sample Block Pre-SKL
> > +          */
> > +         const struct isl_extent2d sa_dim =
> >
> 
> Maybe call this px_size_sa?
> 
> 

I'll add a comment about this as well.

> > +            isl_get_interleaved_msaa_px_size_sa(iview->image->samples);
> > +         px_dim.w = 8 / sa_dim.w;
> > +         px_dim.h = 4 / sa_dim.h;
> > +#else
> > +         /* SKL+, the sample block becomes a "pixel block" so the expected
> > +          * pixel dimension is a constant 8x4 px for all sample counts.
> > +          */
> > +         px_dim = (struct isl_extent2d) { .w = 8, .h = 4};
> > +#endif
> > +
> > +         /* Fast depth clears clear an entire sample block at a time. As a
> > +          * result, the rectangle must be aligned to the pixel dimensions
> > of
> > +          * a sample block for a successful operation.
> > +          */
> > +         if (cmd_state->render_area.offset.x % px_dim.w ||
> > +             cmd_state->render_area.offset.y % px_dim.h ||
> > +             cmd_state->render_area.extent.width % px_dim.w ||
> > +             cmd_state->render_area.extent.height % px_dim.h)
> > +            return;
> >
> 
> What about the case where the offset is aligned and we go to the far right
> or bottom of the image but the image size is not aligned?  I think this
> case should be possible too.
> 
> 

Great point! This will be updated in V3.

> > +      }
> > +      break;
> > +   case BLORP_HIZ_OP_DEPTH_RESOLVE:
> > +      if (cmd_buffer->state.pass->attachments[ds].store_op !=
> > +          VK_ATTACHMENT_STORE_OP_STORE)
> > +         return;
> > +      break;
> > +   case BLORP_HIZ_OP_HIZ_RESOLVE:
> > +      if (cmd_buffer->state.pass->attachments[ds].load_op !=
> > +          VK_ATTACHMENT_LOAD_OP_LOAD)
> > +         return;
> > +      break;
> > +   case BLORP_HIZ_OP_NONE:
> > +      unreachable("Invalid HiZ OP");
> > +      break;
> > +   }
> > +
> > +   anv_batch_emit(&cmd_buffer->batch, GENX(3DSTATE_WM_HZ_OP), hzp) {
> > +      switch (op) {
> > +      case BLORP_HIZ_OP_DEPTH_CLEAR:
> > +         hzp.StencilBufferClearEnable = VK_IMAGE_ASPECT_STENCIL_BIT &
> > +                            cmd_state->attachments[ds].
> > pending_clear_aspects;
> > +         hzp.DepthBufferClearEnable = VK_IMAGE_ASPECT_DEPTH_BIT &
> > +                            cmd_state->attachments[ds].
> > pending_clear_aspects;
> > +         hzp.FullSurfaceDepthandStencilClear = full_surface_op;
> > +         hzp.StencilClearValue = 0xff &
> > +                   cmd_state->attachments[ds].clear_value.depthStencil.
> > stencil;
> >
> 
> nit: I would find having "clear_value & 0xff" and all on the second line
> more readable.
> 
> 

Updated in V3.

> > +
> > +         /* Mark aspects as cleared */
> > +         cmd_state->attachments[ds].pending_clear_aspects = 0;
> > +         break;
> > +      case BLORP_HIZ_OP_DEPTH_RESOLVE:
> > +         hzp.DepthBufferResolveEnable = true;
> > +         break;
> > +      case BLORP_HIZ_OP_HIZ_RESOLVE:
> > +         hzp.HierarchicalDepthBufferResolveEnable = true;
> > +         break;
> > +      case BLORP_HIZ_OP_NONE:
> > +         unreachable("Invalid HiZ OP");
> > +         break;
> > +      }
> > +
> > +      if (op != BLORP_HIZ_OP_DEPTH_CLEAR) {
> > +         /* The Optimized HiZ resolve rectangle must be the size of the
> > full RT
> > +          * and aligned to 8x4. The non-optimized Depth resolve rectangle
> > must
> > +          * be the size of the full RT. The same alignment is assumed to
> > be
> > +          * required.
> > +          *
> > +          * TODO:
> > +          * Consider changing halign of non-D16 depth formats to 8 as mip
> > 2 may
> > +          * get clobbered.
> >
> 
> This shouldn't be needed on BDW+
> 

I'll remove the TODO.

> Most of my comments were just nits.
> 
> Reviewed-by: Jason Ekstrand <jason at jlekstrand.net>
> 
> 

Thanks! I'll add a (v2) next your Rb. I'll also wait for your
confirmation that it still applies on v3 after it's been
released.

> > +          */
> > +         hzp.ClearRectangleXMax = align_u32(iview->extent.width, 8);
> > +         hzp.ClearRectangleYMax = align_u32(iview->extent.height, 4);
> >
> +      } else {
> > +         /* This clear rectangle is aligned */
> > +         hzp.ClearRectangleXMin = cmd_state->render_area.offset.x;
> > +         hzp.ClearRectangleYMin = cmd_state->render_area.offset.y;
> > +         hzp.ClearRectangleXMax = cmd_state->render_area.offset.x +
> > +                                  cmd_state->render_area.extent.width;
> > +         hzp.ClearRectangleYMax = cmd_state->render_area.offset.y +
> > +                                  cmd_state->render_area.extent.height;
> > +      }
> > +
> > +
> > +      /* Due to a hardware issue, this bit MBZ */
> > +      hzp.ScissorRectangleEnable = false;
> > +      hzp.NumberofMultisamples = ffs(iview->image->samples) - 1;
> > +      hzp.SampleMask = 0xFFFF;
> > +   }
> > +
> > +   anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
> > +      pc.PostSyncOperation = WriteImmediateData;
> > +      pc.Address =
> > +         (struct anv_address){ &cmd_buffer->device->workaround_bo, 0 };
> > +   }
> > +
> > +   anv_batch_emit(&cmd_buffer->batch, GENX(3DSTATE_WM_HZ_OP), hzp);
> > +
> > +   /* TODO:
> > +    * Determine if a DepthCacheFlushEnable and DepthStallEnable is really
> > +    * necessary for non-full_surface_op clears. Performing a HZ op without
> > +    * this pipecontrol showed no impact on rendering results.
> > +    */
> > +   if (!full_surface_op && op == BLORP_HIZ_OP_DEPTH_CLEAR) {
> > +      anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
> > +         pc.DepthStallEnable = true;
> > +         pc.DepthCacheFlushEnable = true;
> > +      }
> > +   }
> > +}
> > +
> >  void genX(CmdSetEvent)(
> >      VkCommandBuffer                             commandBuffer,
> >      VkEvent                                     _event,
> > --
> > 2.10.0
> >
> > _______________________________________________
> > mesa-dev mailing list
> > mesa-dev at lists.freedesktop.org
> > https://lists.freedesktop.org/mailman/listinfo/mesa-dev
> >


More information about the mesa-dev mailing list