On Wed, Jan 24, 2018 at 04:15:00PM +0200, Pohjolainen, Topi wrote:
> On Fri, Jan 19, 2018 at 03:47:31PM -0800, Jason Ekstrand wrote:
> > This pass performs an "ambiguate" operation on a CCS-compressed surface
> > by manually writing zeros into the CCS. On gen8+, ISL gives us a fairly
> > detailed notion of how the CCS is laid out so this is fairly simple to
> > do. On gen7, the CCS tiling is quite crazy but that isn't an issue
> > because we can only do CCS on single-slice images so we can just blast
> > over the entire CCS buffer if we want to.
> > ---
> > src/intel/blorp/blorp.h | 5 ++
> > src/intel/blorp/blorp_clear.c | 149
> > ++++++++++++++++++++++++++++++++++++++++++
> > 2 files changed, 154 insertions(+)
> >
> > diff --git a/src/intel/blorp/blorp.h b/src/intel/blorp/blorp.h
> > index a1dd571..478a9af 100644
> > --- a/src/intel/blorp/blorp.h
> > +++ b/src/intel/blorp/blorp.h
> > @@ -204,6 +204,11 @@ blorp_ccs_resolve(struct blorp_batch *batch,
> > enum blorp_fast_clear_op resolve_op);
> >
> > void
> > +blorp_ccs_ambiguate(struct blorp_batch *batch,
> > + struct blorp_surf *surf,
> > + uint32_t level, uint32_t layer);
> > +
> > +void
> > blorp_mcs_partial_resolve(struct blorp_batch *batch,
> > struct blorp_surf *surf,
> > enum isl_format format,
> > diff --git a/src/intel/blorp/blorp_clear.c b/src/intel/blorp/blorp_clear.c
> > index 8e7bc9f..fa2abd9 100644
> > --- a/src/intel/blorp/blorp_clear.c
> > +++ b/src/intel/blorp/blorp_clear.c
> > @@ -881,3 +881,152 @@ blorp_mcs_partial_resolve(struct blorp_batch *batch,
> >
> > batch->blorp->exec(batch, ¶ms);
> > }
> > +
> > +/** Clear a CCS to the "uncompressed" state
> > + *
> > + * This pass is the CCS equivalent of a "HiZ resolve". It sets the CCS
> > values
> > + * for a given layer/level of a surface to 0x0 which is the "uncompressed"
> > + * state which tells the sampler to go look at the main surface.
> > + */
> > +void
> > +blorp_ccs_ambiguate(struct blorp_batch *batch,
> > + struct blorp_surf *surf,
> > + uint32_t level, uint32_t layer)
> > +{
> > + struct blorp_params params;
> > + blorp_params_init(¶ms);
> > +
> > + assert(ISL_DEV_GEN(batch->blorp->isl_dev) >= 7);
> > +
> > + const struct isl_format_layout *aux_fmtl =
> > + isl_format_get_layout(surf->aux_surf->format);
> > + assert(aux_fmtl->txc == ISL_TXC_CCS);
> > +
> > + params.dst = (struct brw_blorp_surface_info) {
> > + .enabled = true,
> > + .addr = surf->aux_addr,
> > + .view = {
> > + .usage = ISL_SURF_USAGE_RENDER_TARGET_BIT,
> > + .format = ISL_FORMAT_R32G32B32A32_UINT,
> > + .base_level = 0,
> > + .base_array_layer = 0,
> > + .levels = 1,
> > + .array_len = 1,
> > + .swizzle = ISL_SWIZZLE_IDENTITY,
> > + },
> > + };
> > +
> > + uint32_t z = 0;
> > + if (surf->surf->dim == ISL_SURF_DIM_3D) {
> > + z = layer;
> > + layer = 0;
> > + }
> > +
> > + uint32_t offset_B, x_offset_el, y_offset_el;
> > + isl_surf_get_image_offset_el(surf->aux_surf, level, layer, z,
> > + &x_offset_el, &y_offset_el);
> > + isl_tiling_get_intratile_offset_el(surf->aux_surf->tiling,
> > aux_fmtl->bpb,
> > + surf->aux_surf->row_pitch,
> > + x_offset_el, y_offset_el,
> > + &offset_B, &x_offset_el,
> > &y_offset_el);
> > + params.dst.addr.offset += offset_B;
> > +
> > + const uint32_t width_px = minify(surf->surf->logical_level0_px.width,
> > level);
> > + const uint32_t height_px = minify(surf->surf->logical_level0_px.height,
> > level);
> > + const uint32_t width_el = DIV_ROUND_UP(width_px, aux_fmtl->bw);
> > + const uint32_t height_el = DIV_ROUND_UP(height_px, aux_fmtl->bh);
> > +
> > + struct isl_tile_info ccs_tile_info;
> > + isl_surf_get_tile_info(surf->aux_surf, &ccs_tile_info);
> > +
> > + /* We're going to map it as a regular RGBA32_UINT surface. We need to
> > + * downscale a good deal. We start by computing the area on the CCS to
> > + * clear in units of Y-tiled cache lines.
> > + */
> > + uint32_t x_offset_y_cl, y_offset_y_cl, width_y_cl, height_y_cl;
> > + if (ISL_DEV_GEN(batch->blorp->isl_dev) >= 8) {
> > + /* From the Sky Lake PRM Vol. 12 in the section on planes:
> > + *
> > + * "The Color Control Surface (CCS) contains the compression
> > status
> > + * of the cache-line pairs. The compression state of the
> > cache-line
> > + * pair is specified by 2 bits in the CCS. Each CCS cache-line
> > + * represents an area on the main surface of 16x16 sets of 128
> > byte
> > + * Y-tiled cache-line-pairs. CCS is always Y tiled."
> > + *
> > + * Each 2-bit surface element in the CCS corresponds to a single
> > + * cache-line pair in the main surface. This means that 16x16 el
> > block
> > + * in the CCS maps to a Y-tiled cache line. Fortunately, CCS layouts
> > + * are calculated with a very large alignment so we can round up to a
> > + * whole cache line without worrying about overdraw.
> > + */
> > +
> > + /* On Broadwell and above, a CCS tile is the same as a Y tile when
> > + * viewed at the cache-line granularity. Fortunately, the horizontal
> > + * and vertical alignment requirements of the CCS are such that we
> > can
> > + * align to an entire cache line without worrying about crossing over
> > + * from one LOD to another.
> > + */
> > + const uint32_t scale_x = ccs_tile_info.logical_extent_el.w / 8;
> > + const uint32_t scale_y = ccs_tile_info.logical_extent_el.h / 8;
>
> These are now CCS tile dimensions in number of bytes, right?
>
> > + assert(surf->aux_surf->image_alignment_el.w % scale_x == 0);
> > + assert(surf->aux_surf->image_alignment_el.h % scale_y == 0);
> > +
> > + assert(x_offset_el % scale_x == 0 && y_offset_el % scale_y == 0);
> > + x_offset_y_cl = x_offset_el / scale_x;
> > + y_offset_y_cl = y_offset_el / scale_y;
> > + width_y_cl = DIV_ROUND_UP(width_el, scale_x);
> > + height_y_cl = DIV_ROUND_UP(height_el, scale_y);
>
> This talks about cache lines but I'm reading that these *_cl are actually
> numbers in CCS tiles. What am I missing/misunderstanding?
>
> > + } else {
> > + /* On gen7, the CCS tiling is not so nice. However, there we are
> > + * guaranteed that we only have a single level and slice so we don't
> > + * have to worry about it and can just align to a whole tile.
> > + */
> > + assert(x_offset_el == 0 && y_offset_el == 0);
> > + const uint32_t width_tl =
> > + DIV_ROUND_UP(width_el, ccs_tile_info.logical_extent_el.w);
> > + const uint32_t height_tl =
> > + DIV_ROUND_UP(height_el, ccs_tile_info.logical_extent_el.h);
> > + x_offset_y_cl = 0;
> > + y_offset_y_cl = 0;
> > + width_y_cl = width_tl * 8;
> > + height_y_cl = height_tl * 8;
>
> I am probably badly off in the weeds, and just need to ask. Here I'm thinking
> that width_tl, for example, is something in number of bits and width_tl
> becomes width in CCS tiles.
>
> > + }
> > +
> > + /* We're going to use a RGBA32 format so as to write data as quickly as
> > + * possible. A y-tiled cache line will then be 1x4 px.
> > + */
> > + const uint32_t x_offset_rgba_px = x_offset_y_cl;
> > + const uint32_t y_offset_rgba_px = y_offset_y_cl * 4;
> > + const uint32_t width_rgba_px = width_y_cl;
> > + const uint32_t height_rgba_px = height_y_cl * 4;
>
> And my confusion continues here. Numbers denoted as *_cl represent something
> four times as small as pixels. I'd thought one would need to divide the number
> instead of multiplying it. One is about to represent the same amount with
> larger units (pixels).
Right, this bit here is my bad. For some reason I treated RGBA32 as taking 256
bytes (I have no idea why). So this part makes perfect sense.
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
