freedreno/a6xx: Fix UBWC mipmap sizing.
[mesa.git] / src / freedreno / fdl / fd6_layout.c
index 0f8f14dfc2dd48cfc2b751878f4955bc773bdd4a..7b544fb7c6fd672e97bdd370bf36f74bf7e0a89c 100644 (file)
  * missing UBWC blockwidth/blockheight for npot+64 cpp
  * missing 96/128 CPP for 8x MSAA with 32_32_32/32_32_32_32
  */
-static const struct {
+static const struct tile_alignment {
        unsigned basealign;
        unsigned pitchalign;
        unsigned heightalign;
+       /* UBWC block width/height.  Used in size alignment, and calculating a
+        * descriptor's FLAG_BUFFER_LOG2W/H for mipmapping.
+        */
        uint8_t ubwc_blockwidth;
        uint8_t ubwc_blockheight;
 } tile_alignment[] = {
@@ -56,24 +59,31 @@ static const struct {
        [64] = { 256,  64, 16 },
 
        /* special cases for r8g8: */
-       [0]  = { 256, 64, 32, 16, 4 },
+       [0]  = { 256, 64, 32, 16, 8 },
 };
 
 #define RGB_TILE_WIDTH_ALIGNMENT 64
 #define RGB_TILE_HEIGHT_ALIGNMENT 16
 #define UBWC_PLANE_SIZE_ALIGNMENT 4096
 
-static int
-fdl6_pitchalign(struct fdl_layout *layout, int ta, int level)
+static const struct tile_alignment *
+fdl6_tile_alignment(struct fdl_layout *layout)
 {
-       const struct util_format_description *format_desc =
-               util_format_description(layout->format);
+       debug_assert(layout->cpp < ARRAY_SIZE(tile_alignment));
+
+       if ((layout->cpp == 2) && (util_format_get_nr_components(layout->format) == 2))
+               return &tile_alignment[0];
+       else
+               return &tile_alignment[layout->cpp];
+}
 
+static int
+fdl6_pitchalign(struct fdl_layout *layout, int level)
+{
        uint32_t pitchalign = 64;
        if (fdl_tile_mode(layout, level))
-               pitchalign = tile_alignment[ta].pitchalign;
-       if (format_desc->layout == UTIL_FORMAT_LAYOUT_ASTC)
-               pitchalign *= util_format_get_blockwidth(layout->format);
+               pitchalign = fdl6_tile_alignment(layout)->pitchalign;
+
        return pitchalign;
 }
 
@@ -104,46 +114,54 @@ fdl6_layout(struct fdl_layout *layout,
        if (tile_alignment[layout->cpp].ubwc_blockwidth == 0)
                layout->ubwc = false;
 
-       int ta = layout->cpp;
-
-       /* The z16/r16 formats seem to not play by the normal tiling rules: */
-       if ((layout->cpp == 2) && (util_format_get_nr_components(format) == 2))
-               ta = 0;
+       const struct tile_alignment *ta = fdl6_tile_alignment(layout);
 
        /* in layer_first layout, the level (slice) contains just one
         * layer (since in fact the layer contains the slices)
         */
        uint32_t layers_in_level = layout->layer_first ? 1 : array_size;
 
-       debug_assert(ta < ARRAY_SIZE(tile_alignment));
-       debug_assert(tile_alignment[ta].pitchalign);
+       debug_assert(ta->pitchalign);
 
        if (layout->tile_mode) {
-               layout->base_align = tile_alignment[ta].basealign;
+               layout->base_align = ta->basealign;
        } else {
                layout->base_align = 64;
        }
 
-       uint32_t pitch0 = util_align_npot(width0, fdl6_pitchalign(layout, ta, 0));
+       uint32_t pitch0 = util_align_npot(width0, fdl6_pitchalign(layout, 0));
+
+       uint32_t ubwc_width0 = width0;
+       uint32_t ubwc_height0 = height0;
+       if (mip_levels > 1) {
+               /* With mipmapping enabled, UBWC layout is power-of-two sized,
+                * specified in log2 width/height in the descriptors.
+                */
+               ubwc_width0 = util_next_power_of_two(width0);
+               ubwc_height0 = util_next_power_of_two(height0);
+       }
+       ubwc_width0 = align(DIV_ROUND_UP(ubwc_width0, ta->ubwc_blockwidth),
+                       RGB_TILE_WIDTH_ALIGNMENT);
+       ubwc_height0 = align(DIV_ROUND_UP(ubwc_height0, ta->ubwc_blockheight),
+                       RGB_TILE_HEIGHT_ALIGNMENT);
 
        for (uint32_t level = 0; level < mip_levels; level++) {
                uint32_t depth = u_minify(depth0, level);
                struct fdl_slice *slice = &layout->slices[level];
                struct fdl_slice *ubwc_slice = &layout->ubwc_slices[level];
                uint32_t tile_mode = fdl_tile_mode(layout, level);
-               uint32_t width, height;
+               uint32_t height;
 
                /* tiled levels of 3D textures are rounded up to PoT dimensions: */
                if (is_3d && tile_mode) {
-                       width = u_minify(util_next_power_of_two(width0), level);
                        height = u_minify(util_next_power_of_two(height0), level);
                } else {
-                       width = u_minify(width0, level);
                        height = u_minify(height0, level);
                }
 
+               uint32_t nblocksy = util_format_get_nblocksy(format, height);
                if (tile_mode)
-                       height = align(height, tile_alignment[ta].heightalign);
+                       nblocksy = align(nblocksy, ta->heightalign);
 
                /* The blits used for mem<->gmem work at a granularity of
                 * 32x32, which can cause faults due to over-fetch on the
@@ -153,14 +171,16 @@ fdl6_layout(struct fdl_layout *layout,
                 * may not be:
                 */
                if (level == mip_levels - 1)
-                       height = align(height, 32);
+                       nblocksy = align(nblocksy, 32);
 
-               slice->pitch = util_align_npot(u_minify(pitch0, level),
-                               fdl6_pitchalign(layout, ta, level));
+               uint32_t nblocksx =
+                       util_align_npot(util_format_get_nblocksx(format, u_minify(pitch0, level)),
+                                       fdl6_pitchalign(layout, level));
 
                slice->offset = layout->size;
-               uint32_t blocks = util_format_get_nblocks(format,
-                               slice->pitch, height);
+               uint32_t blocks = nblocksx * nblocksy;
+
+               slice->pitch = nblocksx * layout->cpp;
 
                /* 1d array and 2d array textures must all have the same layer size
                 * for each miplevel on a6xx. 3d textures can have different layer
@@ -184,22 +204,11 @@ fdl6_layout(struct fdl_layout *layout,
                        /* with UBWC every level is aligned to 4K */
                        layout->size = align(layout->size, 4096);
 
-                       uint32_t block_width = tile_alignment[ta].ubwc_blockwidth;
-                       uint32_t block_height = tile_alignment[ta].ubwc_blockheight;
-                       uint32_t meta_pitch = align(DIV_ROUND_UP(width, block_width),
+                       uint32_t meta_pitch = align(u_minify(ubwc_width0, level),
                                        RGB_TILE_WIDTH_ALIGNMENT);
-                       uint32_t meta_height = align(DIV_ROUND_UP(height, block_height),
+                       uint32_t meta_height = align(u_minify(ubwc_height0, level),
                                        RGB_TILE_HEIGHT_ALIGNMENT);
 
-                       /* it looks like mipmaps need alignment to power of two
-                        * TODO: needs testing with large npot textures
-                        * (needed for the first level?)
-                        */
-                       if (mip_levels > 1) {
-                               meta_pitch = util_next_power_of_two(meta_pitch);
-                               meta_height = util_next_power_of_two(meta_height);
-                       }
-
                        ubwc_slice->size0 = align(meta_pitch * meta_height, UBWC_PLANE_SIZE_ALIGNMENT);
                        ubwc_slice->pitch = meta_pitch;
                        ubwc_slice->offset = layout->ubwc_layer_size;
@@ -228,6 +237,7 @@ void
 fdl6_get_ubwc_blockwidth(struct fdl_layout *layout,
                uint32_t *blockwidth, uint32_t *blockheight)
 {
-       *blockwidth = tile_alignment[layout->cpp].ubwc_blockwidth;
-       *blockheight = tile_alignment[layout->cpp].ubwc_blockheight;
+       const struct tile_alignment *ta = fdl6_tile_alignment(layout);
+       *blockwidth = ta->ubwc_blockwidth;
+       *blockheight = ta->ubwc_blockheight;
 }