r600g: atomize stencil ref state
[mesa.git] / src / gallium / drivers / r600 / r600_texture.c
index cbb9e5a9f48bef70ac6ff8cd7d82c177d2b6d351..6de3d6a8645aae6a2678dc295864a73e25bcb427 100644 (file)
@@ -58,181 +58,15 @@ static void r600_copy_from_staging_texture(struct pipe_context *ctx, struct r600
                                  0, &sbox);
 }
 
-unsigned r600_texture_get_offset(struct r600_resource_texture *rtex,
+unsigned r600_texture_get_offset(struct r600_texture *rtex,
                                        unsigned level, unsigned layer)
 {
-       unsigned offset = rtex->offset[level];
-
-       switch (rtex->resource.b.b.target) {
-       case PIPE_TEXTURE_3D:
-       case PIPE_TEXTURE_CUBE:
-       default:
-               return offset + layer * rtex->layer_size[level];
-       }
-}
-
-static unsigned r600_get_block_alignment(struct pipe_screen *screen,
-                                        enum pipe_format format,
-                                        unsigned array_mode)
-{
-       struct r600_screen* rscreen = (struct r600_screen *)screen;
-       unsigned pixsize = util_format_get_blocksize(format);
-       int p_align;
-
-       switch(array_mode) {
-       case V_038000_ARRAY_1D_TILED_THIN1:
-               p_align = MAX2(8,
-                              ((rscreen->tiling_info.group_bytes / 8 / pixsize)));
-               break;
-       case V_038000_ARRAY_2D_TILED_THIN1:
-               p_align = MAX2(rscreen->tiling_info.num_banks,
-                              (((rscreen->tiling_info.group_bytes / 8 / pixsize)) *
-                               rscreen->tiling_info.num_banks)) * 8;
-               break;
-       case V_038000_ARRAY_LINEAR_ALIGNED:
-               p_align = MAX2(64, rscreen->tiling_info.group_bytes / pixsize);
-               break;
-       case V_038000_ARRAY_LINEAR_GENERAL:
-       default:
-               p_align = rscreen->tiling_info.group_bytes / pixsize;
-               break;
-       }
-       return p_align;
-}
-
-static unsigned r600_get_height_alignment(struct pipe_screen *screen,
-                                         unsigned array_mode)
-{
-       struct r600_screen* rscreen = (struct r600_screen *)screen;
-       int h_align;
-
-       switch (array_mode) {
-       case V_038000_ARRAY_2D_TILED_THIN1:
-               h_align = rscreen->tiling_info.num_channels * 8;
-               break;
-       case V_038000_ARRAY_1D_TILED_THIN1:
-       case V_038000_ARRAY_LINEAR_ALIGNED:
-               h_align = 8;
-               break;
-       case V_038000_ARRAY_LINEAR_GENERAL:
-       default:
-               h_align = 1;
-               break;
-       }
-       return h_align;
-}
-
-static unsigned r600_get_base_alignment(struct pipe_screen *screen,
-                                       enum pipe_format format,
-                                       unsigned array_mode)
-{
-       struct r600_screen* rscreen = (struct r600_screen *)screen;
-       unsigned pixsize = util_format_get_blocksize(format);
-       int p_align = r600_get_block_alignment(screen, format, array_mode);
-       int h_align = r600_get_height_alignment(screen, array_mode);
-       int b_align;
-
-       switch (array_mode) {
-       case V_038000_ARRAY_2D_TILED_THIN1:
-               b_align = MAX2(rscreen->tiling_info.num_banks * rscreen->tiling_info.num_channels * 8 * 8 * pixsize,
-                              p_align * pixsize * h_align);
-               break;
-       case V_038000_ARRAY_1D_TILED_THIN1:
-       case V_038000_ARRAY_LINEAR_ALIGNED:
-       case V_038000_ARRAY_LINEAR_GENERAL:
-       default:
-               b_align = rscreen->tiling_info.group_bytes;
-               break;
-       }
-       return b_align;
+       return rtex->surface.level[level].offset +
+              layer * rtex->surface.level[level].slice_size;
 }
 
-static unsigned mip_minify(unsigned size, unsigned level)
-{
-       unsigned val;
-       val = u_minify(size, level);
-       if (level > 0)
-               val = util_next_power_of_two(val);
-       return val;
-}
-
-static unsigned r600_texture_get_nblocksx(struct pipe_screen *screen,
-                                         struct r600_resource_texture *rtex,
-                                         unsigned level)
-{
-       struct pipe_resource *ptex = &rtex->resource.b.b;
-       unsigned nblocksx, block_align, width;
-       unsigned blocksize = util_format_get_blocksize(rtex->real_format);
-
-       if (rtex->pitch_override)
-               return rtex->pitch_override / blocksize;
-
-       width = mip_minify(ptex->width0, level);
-       nblocksx = util_format_get_nblocksx(rtex->real_format, width);
-
-       block_align = r600_get_block_alignment(screen, rtex->real_format,
-                                             rtex->array_mode[level]);
-       nblocksx = align(nblocksx, block_align);
-       return nblocksx;
-}
-
-static unsigned r600_texture_get_nblocksy(struct pipe_screen *screen,
-                                         struct r600_resource_texture *rtex,
-                                         unsigned level)
-{
-       struct pipe_resource *ptex = &rtex->resource.b.b;
-       unsigned height, tile_height;
-
-       height = mip_minify(ptex->height0, level);
-       height = util_format_get_nblocksy(rtex->real_format, height);
-       tile_height = r600_get_height_alignment(screen,
-                                               rtex->array_mode[level]);
-
-       /* XXX Hack around an alignment issue. Less tests fail with this.
-        *
-        * The thing is depth-stencil buffers should be tiled, i.e.
-        * the alignment should be >=8. If I make them tiled, stencil starts
-        * working because it no longer overlaps with the depth buffer
-        * in memory, but texturing like drawpix-stencil breaks. */
-       if (util_format_is_depth_or_stencil(rtex->real_format) && tile_height < 8)
-               tile_height = 8;
-
-       height = align(height, tile_height);
-       return height;
-}
-
-static void r600_texture_set_array_mode(struct pipe_screen *screen,
-                                       struct r600_resource_texture *rtex,
-                                       unsigned level, unsigned array_mode)
-{
-       struct pipe_resource *ptex = &rtex->resource.b.b;
-
-       switch (array_mode) {
-       case V_0280A0_ARRAY_LINEAR_GENERAL:
-       case V_0280A0_ARRAY_LINEAR_ALIGNED:
-       case V_0280A0_ARRAY_1D_TILED_THIN1:
-       default:
-               rtex->array_mode[level] = array_mode;
-               break;
-       case V_0280A0_ARRAY_2D_TILED_THIN1:
-       {
-               unsigned w, h, tile_height, tile_width;
-
-               tile_height = r600_get_height_alignment(screen, array_mode);
-               tile_width = r600_get_block_alignment(screen, rtex->real_format, array_mode);
-
-               w = mip_minify(ptex->width0, level);
-               h = mip_minify(ptex->height0, level);
-               if (w <= tile_width || h <= tile_height)
-                       rtex->array_mode[level] = V_0280A0_ARRAY_1D_TILED_THIN1;
-               else
-                       rtex->array_mode[level] = array_mode;
-       }
-       break;
-       }
-}
-
-static int r600_init_surface(struct radeon_surface *surface,
+static int r600_init_surface(struct r600_screen *rscreen,
+                            struct radeon_surface *surface,
                             const struct pipe_resource *ptex,
                             unsigned array_mode,
                             bool is_transfer, bool is_flushed_depth)
@@ -252,13 +86,22 @@ static int r600_init_surface(struct radeon_surface *surface,
        surface->blk_d = 1;
        surface->array_size = 1;
        surface->last_level = ptex->last_level;
-       surface->bpe = util_format_get_blocksize(ptex->format);
-       /* align byte per element on dword */
-       if (surface->bpe == 3) {
-               surface->bpe = 4;
+
+       if (rscreen->chip_class >= EVERGREEN &&
+           !is_transfer && !is_flushed_depth &&
+           ptex->format == PIPE_FORMAT_Z32_FLOAT_S8X24_UINT) {
+               surface->bpe = 4; /* stencil is allocated separately on evergreen */
+       } else {
+               surface->bpe = util_format_get_blocksize(ptex->format);
+               /* align byte per element on dword */
+               if (surface->bpe == 3) {
+                       surface->bpe = 4;
+               }
        }
-       surface->nsamples = 1;
+
+       surface->nsamples = ptex->nr_samples ? ptex->nr_samples : 1;
        surface->flags = 0;
+
        switch (array_mode) {
        case V_038000_ARRAY_1D_TILED_THIN1:
                surface->flags |= RADEON_SURF_SET(RADEON_SURF_MODE_1D, MODE);
@@ -315,8 +158,7 @@ static int r600_init_surface(struct radeon_surface *surface,
 }
 
 static int r600_setup_surface(struct pipe_screen *screen,
-                             struct r600_resource_texture *rtex,
-                             unsigned array_mode,
+                             struct r600_texture *rtex,
                              unsigned pitch_in_bytes_override)
 {
        struct pipe_resource *ptex = &rtex->resource.b.b;
@@ -341,9 +183,6 @@ static int r600_setup_surface(struct pipe_screen *screen,
                }
        }
        for (i = 0; i <= ptex->last_level; i++) {
-               rtex->offset[i] = rtex->surface.level[i].offset;
-               rtex->layer_size[i] = rtex->surface.level[i].slice_size;
-               rtex->pitch_in_bytes[i] = rtex->surface.level[i].pitch_bytes;
                switch (rtex->surface.level[i].mode) {
                case RADEON_SURF_MODE_LINEAR_ALIGNED:
                        rtex->array_mode[i] = V_038000_ARRAY_LINEAR_ALIGNED;
@@ -363,92 +202,11 @@ static int r600_setup_surface(struct pipe_screen *screen,
        return 0;
 }
 
-static void r600_setup_miptree(struct pipe_screen *screen,
-                              struct r600_resource_texture *rtex,
-                              unsigned array_mode)
-{
-       struct pipe_resource *ptex = &rtex->resource.b.b;
-       enum chip_class chipc = ((struct r600_screen*)screen)->chip_class;
-       unsigned size, layer_size, i, offset;
-       unsigned nblocksx, nblocksy;
-
-       for (i = 0, offset = 0; i <= ptex->last_level; i++) {
-               unsigned blocksize = util_format_get_blocksize(rtex->real_format);
-               unsigned base_align = r600_get_base_alignment(screen, rtex->real_format, array_mode);
-
-               r600_texture_set_array_mode(screen, rtex, i, array_mode);
-
-               nblocksx = r600_texture_get_nblocksx(screen, rtex, i);
-               nblocksy = r600_texture_get_nblocksy(screen, rtex, i);
-
-               if (chipc >= EVERGREEN && array_mode == V_038000_ARRAY_LINEAR_GENERAL)
-                       layer_size = align(nblocksx, 64) * nblocksy * blocksize;
-               else
-                       layer_size = nblocksx * nblocksy * blocksize;
-
-               if (ptex->target == PIPE_TEXTURE_CUBE) {
-                       if (chipc >= R700)
-                               size = layer_size * 8;
-                       else
-                               size = layer_size * 6;
-               }
-               else if (ptex->target == PIPE_TEXTURE_3D)
-                       size = layer_size * u_minify(ptex->depth0, i);
-               else
-                       size = layer_size * ptex->array_size;
-
-               /* align base image and start of miptree */
-               if ((i == 0) || (i == 1))
-                       offset = align(offset, base_align);
-               rtex->offset[i] = offset;
-               rtex->layer_size[i] = layer_size;
-               rtex->pitch_in_blocks[i] = nblocksx; /* CB talks in elements */
-               rtex->pitch_in_bytes[i] = nblocksx * blocksize;
-
-               offset += size;
-       }
-       rtex->size = offset;
-}
-
-/* Figure out whether u_blitter will fallback to a transfer operation.
- * If so, don't use a staging resource.
- */
-static boolean permit_hardware_blit(struct pipe_screen *screen,
-                                       const struct pipe_resource *res)
-{
-       unsigned bind;
-
-       if (util_format_is_depth_or_stencil(res->format))
-               bind = PIPE_BIND_DEPTH_STENCIL;
-       else
-               bind = PIPE_BIND_RENDER_TARGET;
-
-       /* hackaround for S3TC */
-       if (util_format_is_compressed(res->format))
-               return TRUE;
-
-       if (!screen->is_format_supported(screen,
-                               res->format,
-                               res->target,
-                               res->nr_samples,
-                                bind))
-               return FALSE;
-
-       if (!screen->is_format_supported(screen,
-                               res->format,
-                               res->target,
-                               res->nr_samples,
-                                PIPE_BIND_SAMPLER_VIEW))
-               return FALSE;
-
-       return TRUE;
-}
-
 static boolean r600_texture_get_handle(struct pipe_screen* screen,
                                        struct pipe_resource *ptex,
                                        struct winsys_handle *whandle)
 {
-       struct r600_resource_texture *rtex = (struct r600_resource_texture*)ptex;
+       struct r600_texture *rtex = (struct r600_texture*)ptex;
        struct r600_resource *resource = &rtex->resource;
        struct radeon_surface *surface = &rtex->surface;
        struct r600_screen *rscreen = (struct r600_screen*)screen;
@@ -463,24 +221,21 @@ static boolean r600_texture_get_handle(struct pipe_screen* screen,
                                       surface->tile_split,
                                       surface->stencil_tile_split,
                                       surface->mtilea,
-                                      rtex->pitch_in_bytes[0]);
+                                      rtex->surface.level[0].pitch_bytes);
 
        return rscreen->ws->buffer_get_handle(resource->buf,
-                                             rtex->pitch_in_bytes[0], whandle);
+                                             rtex->surface.level[0].pitch_bytes, whandle);
 }
 
 static void r600_texture_destroy(struct pipe_screen *screen,
                                 struct pipe_resource *ptex)
 {
-       struct r600_resource_texture *rtex = (struct r600_resource_texture*)ptex;
+       struct r600_texture *rtex = (struct r600_texture*)ptex;
        struct r600_resource *resource = &rtex->resource;
 
        if (rtex->flushed_depth_texture)
                pipe_resource_reference((struct pipe_resource **)&rtex->flushed_depth_texture, NULL);
 
-       if (rtex->stencil)
-               pipe_resource_reference((struct pipe_resource **)&rtex->stencil, NULL);
-
        pb_reference(&resource->buf, NULL);
        FREE(rtex);
 }
@@ -497,22 +252,143 @@ static const struct u_resource_vtbl r600_texture_vtbl =
        NULL                            /* transfer_inline_write */
 };
 
-static struct r600_resource_texture *
+/* The number of samples can be specified independently of the texture. */
+void r600_texture_get_fmask_info(struct r600_screen *rscreen,
+                                struct r600_texture *rtex,
+                                unsigned nr_samples,
+                                struct r600_fmask_info *out)
+{
+       /* FMASK is allocated pretty much like an ordinary texture.
+        * Here we use bpe in the units of bits, not bytes. */
+       struct radeon_surface fmask = rtex->surface;
+
+       switch (nr_samples) {
+       case 2:
+               /* This should be 8,1, but we should set nsamples > 1
+                * for the allocator to treat it as a multisample surface.
+                * Let's set 4,2 then. */
+       case 4:
+               fmask.bpe = 4;
+               fmask.nsamples = 2;
+               break;
+       case 8:
+               fmask.bpe = 8;
+               fmask.nsamples = 4;
+               break;
+       case 16:
+               fmask.bpe = 16;
+               fmask.nsamples = 4;
+               break;
+       default:
+               R600_ERR("Invalid sample count for FMASK allocation.\n");
+               return;
+       }
+
+       /* R600-R700 errata? Anyway, this fixes colorbuffer corruption. */
+       if (rscreen->chip_class <= R700) {
+               fmask.bpe *= 2;
+       }
+
+       if (rscreen->chip_class >= EVERGREEN) {
+               fmask.bankh = nr_samples <= 4 ? 4 : 1;
+       }
+
+       if (rscreen->ws->surface_init(rscreen->ws, &fmask)) {
+               R600_ERR("Got error in surface_init while allocating FMASK.\n");
+               return;
+       }
+       assert(fmask.level[0].mode == RADEON_SURF_MODE_2D);
+
+       out->bank_height = fmask.bankh;
+       out->alignment = MAX2(256, fmask.bo_alignment);
+       out->size = (fmask.bo_size + 7) / 8;
+}
+
+static void r600_texture_allocate_fmask(struct r600_screen *rscreen,
+                                       struct r600_texture *rtex)
+{
+       struct r600_fmask_info fmask;
+
+       r600_texture_get_fmask_info(rscreen, rtex,
+                                   rtex->resource.b.b.nr_samples, &fmask);
+
+       /* Reserve space for FMASK while converting bits back to bytes. */
+       rtex->fmask_bank_height = fmask.bank_height;
+       rtex->fmask_offset = align(rtex->size, fmask.alignment);
+       rtex->fmask_size = fmask.size;
+       rtex->size = rtex->fmask_offset + rtex->fmask_size;
+#if 0
+       printf("FMASK width=%u, height=%i, bits=%u, size=%u\n",
+              fmask.npix_x, fmask.npix_y, fmask.bpe * fmask.nsamples, rtex->fmask_size);
+#endif
+}
+
+void r600_texture_get_cmask_info(struct r600_screen *rscreen,
+                                struct r600_texture *rtex,
+                                struct r600_cmask_info *out)
+{
+       unsigned cmask_tile_width = 8;
+       unsigned cmask_tile_height = 8;
+       unsigned cmask_tile_elements = cmask_tile_width * cmask_tile_height;
+       unsigned element_bits = 4;
+       unsigned cmask_cache_bits = 1024;
+       unsigned num_pipes = rscreen->tiling_info.num_channels;
+       unsigned pipe_interleave_bytes = rscreen->tiling_info.group_bytes;
+
+       unsigned elements_per_macro_tile = (cmask_cache_bits / element_bits) * num_pipes;
+       unsigned pixels_per_macro_tile = elements_per_macro_tile * cmask_tile_elements;
+       unsigned sqrt_pixels_per_macro_tile = sqrt(pixels_per_macro_tile);
+       unsigned macro_tile_width = util_next_power_of_two(sqrt_pixels_per_macro_tile);
+       unsigned macro_tile_height = pixels_per_macro_tile / macro_tile_width;
+
+       unsigned pitch_elements = align(rtex->surface.npix_x, macro_tile_width);
+       unsigned height = align(rtex->surface.npix_y, macro_tile_height);
+
+       unsigned base_align = num_pipes * pipe_interleave_bytes;
+       unsigned slice_bytes =
+               ((pitch_elements * height * element_bits + 7) / 8) / cmask_tile_elements;
+
+       assert(macro_tile_width % 128 == 0);
+       assert(macro_tile_height % 128 == 0);
+
+       out->slice_tile_max = ((pitch_elements * height) / (128*128)) - 1;
+       out->alignment = MAX2(256, base_align);
+       out->size = rtex->surface.array_size * align(slice_bytes, base_align);
+}
+
+static void r600_texture_allocate_cmask(struct r600_screen *rscreen,
+                                       struct r600_texture *rtex)
+{
+       struct r600_cmask_info cmask;
+
+       r600_texture_get_cmask_info(rscreen, rtex, &cmask);
+
+       rtex->cmask_slice_tile_max = cmask.slice_tile_max;
+       rtex->cmask_offset = align(rtex->size, cmask.alignment);
+       rtex->cmask_size = cmask.size;
+       rtex->size = rtex->cmask_offset + rtex->cmask_size;
+#if 0
+       printf("CMASK: macro tile width = %u, macro tile height = %u, "
+              "pitch elements = %u, height = %u, slice tile max = %u\n",
+              macro_tile_width, macro_tile_height, pitch_elements, height,
+              rtex->cmask_slice_tile_max);
+#endif
+}
+
+static struct r600_texture *
 r600_texture_create_object(struct pipe_screen *screen,
                           const struct pipe_resource *base,
-                          unsigned array_mode,
                           unsigned pitch_in_bytes_override,
-                          unsigned max_buffer_size,
                           struct pb_buffer *buf,
                           boolean alloc_bo,
                           struct radeon_surface *surface)
 {
-       struct r600_resource_texture *rtex;
+       struct r600_texture *rtex;
        struct r600_resource *resource;
        struct r600_screen *rscreen = (struct r600_screen*)screen;
        int r;
 
-       rtex = CALLOC_STRUCT(r600_resource_texture);
+       rtex = CALLOC_STRUCT(r600_texture);
        if (rtex == NULL)
                return NULL;
 
@@ -522,94 +398,35 @@ r600_texture_create_object(struct pipe_screen *screen,
        pipe_reference_init(&resource->b.b.reference, 1);
        resource->b.b.screen = screen;
        rtex->pitch_override = pitch_in_bytes_override;
-       rtex->real_format = base->format;
 
-       /* We must split depth and stencil into two separate buffers on Evergreen. */
-       if ((base->bind & PIPE_BIND_DEPTH_STENCIL) &&
-           ((struct r600_screen*)screen)->chip_class >= EVERGREEN &&
-           util_format_is_depth_and_stencil(base->format) &&
-           !rscreen->use_surface_alloc) {
-               struct pipe_resource stencil;
-               unsigned stencil_pitch_override = 0;
+       /* don't include stencil-only formats which we don't support for rendering */
+       rtex->is_depth = util_format_has_depth(util_format_description(rtex->resource.b.b.format));
 
-               switch (base->format) {
-               case PIPE_FORMAT_Z24_UNORM_S8_UINT:
-                       rtex->real_format = PIPE_FORMAT_Z24X8_UNORM;
-                       break;
-               case PIPE_FORMAT_S8_UINT_Z24_UNORM:
-                       rtex->real_format = PIPE_FORMAT_X8Z24_UNORM;
-                       break;
-               case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
-                       rtex->real_format = PIPE_FORMAT_Z32_FLOAT;
-                       break;
-               default:
-                       assert(0);
-                       FREE(rtex);
-                       return NULL;
-               }
-
-               /* Divide the pitch in bytes by 4 for stencil, because it has a smaller pixel size. */
-               if (pitch_in_bytes_override) {
-                       assert(base->format == PIPE_FORMAT_Z24_UNORM_S8_UINT ||
-                              base->format == PIPE_FORMAT_S8_UINT_Z24_UNORM);
-                       stencil_pitch_override = pitch_in_bytes_override / 4;
-               }
-
-               /* Allocate the stencil buffer. */
-               stencil = *base;
-               stencil.format = PIPE_FORMAT_S8_UINT;
-               rtex->stencil = r600_texture_create_object(screen, &stencil, array_mode,
-                                                          stencil_pitch_override,
-                                                          max_buffer_size, NULL, FALSE, surface);
-               if (!rtex->stencil) {
-                       FREE(rtex);
-                       return NULL;
-               }
-               /* Proceed in creating the depth buffer. */
+       rtex->surface = *surface;
+       r = r600_setup_surface(screen, rtex,
+                              pitch_in_bytes_override);
+       if (r) {
+               FREE(rtex);
+               return NULL;
        }
 
-       /* only mark depth textures the HW can hit as depth textures */
-       if (util_format_is_depth_or_stencil(rtex->real_format) &&
-                       permit_hardware_blit(screen, base))
-               rtex->is_depth = true;
-
-       r600_setup_miptree(screen, rtex, array_mode);
-       if (rscreen->use_surface_alloc) {
-               rtex->surface = *surface;
-               r = r600_setup_surface(screen, rtex, array_mode,
-                                      pitch_in_bytes_override);
-               if (r) {
-                       FREE(rtex);
-                       return NULL;
-               }
+       if (base->nr_samples > 1 && !rtex->is_depth && alloc_bo) {
+               r600_texture_allocate_cmask(rscreen, rtex);
+               r600_texture_allocate_fmask(rscreen, rtex);
        }
 
-       /* If we initialized separate stencil for Evergreen. place it after depth. */
-       if (rtex->stencil) {
-               unsigned stencil_align, stencil_offset;
-
-               stencil_align = r600_get_base_alignment(screen, rtex->stencil->real_format, array_mode);
-               stencil_offset = align(rtex->size, stencil_align);
-
-               for (unsigned i = 0; i <= rtex->stencil->resource.b.b.last_level; i++)
-                       rtex->stencil->offset[i] += stencil_offset;
-
-               rtex->size = stencil_offset + rtex->stencil->size;
+       if (!rtex->is_depth && base->nr_samples > 1 &&
+           (!rtex->fmask_size || !rtex->cmask_size)) {
+               FREE(rtex);
+               return NULL;
        }
 
        /* Now create the backing buffer. */
        if (!buf && alloc_bo) {
-               struct pipe_resource *ptex = &rtex->resource.b.b;
-               unsigned base_align = r600_get_base_alignment(screen, ptex->format, array_mode);
-
-               if (rscreen->use_surface_alloc) {
-                       base_align = rtex->surface.bo_alignment;
-               } else if (util_format_is_depth_or_stencil(rtex->real_format)) {
-                       /* ugly work around depth buffer need stencil room at end of bo */
-                       rtex->size += ptex->width0 * ptex->height0;
-               }
-               if (!r600_init_resource(rscreen, resource, rtex->size, base_align, base->bind, base->usage)) {
-                       pipe_resource_reference((struct pipe_resource**)&rtex->stencil, NULL);
+               unsigned base_align = rtex->surface.bo_alignment;
+               unsigned usage = R600_TEX_IS_TILED(rtex, 0) ? PIPE_USAGE_STATIC : base->usage;
+
+               if (!r600_init_resource(rscreen, resource, rtex->size, base_align, base->bind, usage)) {
                        FREE(rtex);
                        return NULL;
                }
@@ -619,10 +436,11 @@ r600_texture_create_object(struct pipe_screen *screen,
                resource->domains = RADEON_DOMAIN_GTT | RADEON_DOMAIN_VRAM;
        }
 
-       if (rtex->stencil) {
-               pb_reference(&rtex->stencil->resource.buf, rtex->resource.buf);
-               rtex->stencil->resource.cs_buf = rtex->resource.cs_buf;
-               rtex->stencil->resource.domains = rtex->resource.domains;
+       if (rtex->cmask_size) {
+               /* Initialize the cmask to 0xCC (= compressed state). */
+               char *ptr = rscreen->ws->buffer_map(resource->cs_buf, NULL, PIPE_TRANSFER_WRITE);
+               memset(ptr + rtex->cmask_offset, 0xCC, rtex->cmask_size);
+               rscreen->ws->buffer_unmap(resource->cs_buf);
        }
        return rtex;
 }
@@ -636,18 +454,20 @@ struct pipe_resource *r600_texture_create(struct pipe_screen *screen,
        int r;
 
        if (!(templ->flags & R600_RESOURCE_FLAG_TRANSFER)) {
-               if (rscreen->use_surface_alloc &&
-                   !(templ->bind & PIPE_BIND_SCANOUT) &&
+               if (!(templ->bind & PIPE_BIND_SCANOUT) &&
                    templ->usage != PIPE_USAGE_STAGING &&
-                   templ->usage != PIPE_USAGE_STREAM &&
-                   permit_hardware_blit(screen, templ)) {
+                   templ->usage != PIPE_USAGE_STREAM) {
                        array_mode = V_038000_ARRAY_2D_TILED_THIN1;
                } else if (util_format_is_compressed(templ->format)) {
                        array_mode = V_038000_ARRAY_1D_TILED_THIN1;
                }
        }
 
-       r = r600_init_surface(&surface, templ, array_mode,
+       /* XXX tiling is broken for the 422 formats */
+       if (util_format_description(templ->format)->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED)
+               array_mode = V_038000_ARRAY_LINEAR_ALIGNED;
+
+       r = r600_init_surface(rscreen, &surface, templ, array_mode,
                              templ->flags & R600_RESOURCE_FLAG_TRANSFER,
                              templ->flags & R600_RESOURCE_FLAG_FLUSHED_DEPTH);
        if (r) {
@@ -657,41 +477,38 @@ struct pipe_resource *r600_texture_create(struct pipe_screen *screen,
        if (r) {
                return NULL;
        }
-       return (struct pipe_resource *)r600_texture_create_object(screen, templ, array_mode,
-                                                                 0, 0, NULL, TRUE, &surface);
+       return (struct pipe_resource *)r600_texture_create_object(screen, templ,
+                                                                 0, NULL, TRUE, &surface);
 }
 
 static struct pipe_surface *r600_create_surface(struct pipe_context *pipe,
                                                struct pipe_resource *texture,
-                                               const struct pipe_surface *surf_tmpl)
+                                               const struct pipe_surface *templ)
 {
-       struct r600_resource_texture *rtex = (struct r600_resource_texture*)texture;
+       struct r600_texture *rtex = (struct r600_texture*)texture;
        struct r600_surface *surface = CALLOC_STRUCT(r600_surface);
-       unsigned level = surf_tmpl->u.tex.level;
+       unsigned level = templ->u.tex.level;
 
-       assert(surf_tmpl->u.tex.first_layer == surf_tmpl->u.tex.last_layer);
+       assert(templ->u.tex.first_layer == templ->u.tex.last_layer);
        if (surface == NULL)
                return NULL;
        pipe_reference_init(&surface->base.reference, 1);
        pipe_resource_reference(&surface->base.texture, texture);
        surface->base.context = pipe;
-       surface->base.format = surf_tmpl->format;
-       surface->base.width = mip_minify(texture->width0, level);
-       surface->base.height = mip_minify(texture->height0, level);
-       surface->base.usage = surf_tmpl->usage;
-       surface->base.texture = texture;
-       surface->base.u.tex.first_layer = surf_tmpl->u.tex.first_layer;
-       surface->base.u.tex.last_layer = surf_tmpl->u.tex.last_layer;
-       surface->base.u.tex.level = level;
-
-       surface->aligned_height = r600_texture_get_nblocksy(pipe->screen,
-                                                           rtex, level);
+       surface->base.format = templ->format;
+       surface->base.width = rtex->surface.level[level].npix_x;
+       surface->base.height = rtex->surface.level[level].npix_y;
+       surface->base.usage = templ->usage;
+       surface->base.u = templ->u;
        return &surface->base;
 }
 
 static void r600_surface_destroy(struct pipe_context *pipe,
                                 struct pipe_surface *surface)
 {
+       struct r600_surface *surf = (struct r600_surface*)surface;
+       pipe_resource_reference((struct pipe_resource**)&surf->cb_buffer_fmask, NULL);
+       pipe_resource_reference((struct pipe_resource**)&surf->cb_buffer_cmask, NULL);
        pipe_resource_reference(&surface->texture, NULL);
        FREE(surface);
 }
@@ -730,25 +547,25 @@ struct pipe_resource *r600_texture_from_handle(struct pipe_screen *screen,
        else
                array_mode = 0;
 
-       r = r600_init_surface(&surface, templ, array_mode, false, false);
+       r = r600_init_surface(rscreen, &surface, templ, array_mode, false, false);
        if (r) {
                return NULL;
        }
-       return (struct pipe_resource *)r600_texture_create_object(screen, templ, array_mode,
-                                                                 stride, 0, buf, FALSE, &surface);
+       return (struct pipe_resource *)r600_texture_create_object(screen, templ,
+                                                                 stride, buf, FALSE, &surface);
 }
 
-void r600_init_flushed_depth_texture(struct pipe_context *ctx,
+bool r600_init_flushed_depth_texture(struct pipe_context *ctx,
                                     struct pipe_resource *texture,
-                                    struct r600_resource_texture **staging)
+                                    struct r600_texture **staging)
 {
-       struct r600_resource_texture *rtex = (struct r600_resource_texture*)texture;
+       struct r600_texture *rtex = (struct r600_texture*)texture;
        struct pipe_resource resource;
-       struct r600_resource_texture **flushed_depth_texture = staging ?
+       struct r600_texture **flushed_depth_texture = staging ?
                        staging : &rtex->flushed_depth_texture;
 
        if (!staging && rtex->flushed_depth_texture)
-               return; /* it's ready */
+               return true; /* it's ready */
 
        resource.target = texture->target;
        resource.format = texture->format;
@@ -758,48 +575,21 @@ void r600_init_flushed_depth_texture(struct pipe_context *ctx,
        resource.array_size = texture->array_size;
        resource.last_level = texture->last_level;
        resource.nr_samples = texture->nr_samples;
-       resource.usage = staging ? PIPE_USAGE_DYNAMIC : PIPE_USAGE_DEFAULT;
+       resource.usage = staging ? PIPE_USAGE_STAGING : PIPE_USAGE_STATIC;
        resource.bind = texture->bind & ~PIPE_BIND_DEPTH_STENCIL;
        resource.flags = texture->flags | R600_RESOURCE_FLAG_FLUSHED_DEPTH;
 
        if (staging)
                resource.flags |= R600_RESOURCE_FLAG_TRANSFER;
 
-       *flushed_depth_texture = (struct r600_resource_texture *)ctx->screen->resource_create(ctx->screen, &resource);
+       *flushed_depth_texture = (struct r600_texture *)ctx->screen->resource_create(ctx->screen, &resource);
        if (*flushed_depth_texture == NULL) {
                R600_ERR("failed to create temporary texture to hold flushed depth\n");
-               return;
+               return false;
        }
 
        (*flushed_depth_texture)->is_flushing_texture = TRUE;
-
-}
-
-void r600_texture_depth_flush(struct pipe_context *ctx,
-                             struct pipe_resource *texture,
-                             struct r600_resource_texture **staging,
-                             unsigned first_level, unsigned last_level,
-                             unsigned first_layer, unsigned last_layer)
-{
-       struct r600_resource_texture *rtex = (struct r600_resource_texture*)texture;
-
-       r600_init_flushed_depth_texture(ctx, texture, staging);
-
-       if (staging) {
-               if (!*staging)
-                       return; /* error */
-
-               r600_blit_uncompress_depth(ctx, rtex, *staging,
-                                          first_level, last_level,
-                                          first_layer, last_layer);
-       } else {
-               if (!rtex->flushed_depth_texture)
-                       return; /* error */
-
-               r600_blit_uncompress_depth(ctx, rtex, NULL,
-                                          first_level, last_level,
-                                          first_layer, last_layer);
-       }
+       return true;
 }
 
 /* Needs adjustment for pixelformat:
@@ -816,7 +606,7 @@ struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
                                                const struct pipe_box *box)
 {
        struct r600_context *rctx = (struct r600_context*)ctx;
-       struct r600_resource_texture *rtex = (struct r600_resource_texture*)texture;
+       struct r600_texture *rtex = (struct r600_texture*)texture;
        struct pipe_resource resource;
        struct r600_transfer *trans;
        boolean use_staging_texture = FALSE;
@@ -842,8 +632,7 @@ struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
                use_staging_texture = TRUE;
        }
 
-       if (!permit_hardware_blit(ctx->screen, texture) ||
-               (texture->flags & R600_RESOURCE_FLAG_TRANSFER)) {
+       if (texture->flags & R600_RESOURCE_FLAG_TRANSFER) {
                use_staging_texture = FALSE;
        }
 
@@ -863,18 +652,21 @@ struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
                */
                /* XXX: when discard is true, no need to read back from depth texture
                */
-               struct r600_resource_texture *staging_depth;
+               struct r600_texture *staging_depth;
 
-               r600_texture_depth_flush(ctx, texture, &staging_depth,
-                                        level, level,
-                                        box->z, box->z + box->depth - 1);
-               if (!staging_depth) {
+               if (!r600_init_flushed_depth_texture(ctx, texture, &staging_depth)) {
                        R600_ERR("failed to create temporary texture to hold untiled copy\n");
                        pipe_resource_reference(&trans->transfer.resource, NULL);
                        FREE(trans);
                        return NULL;
                }
-               trans->transfer.stride = staging_depth->pitch_in_bytes[level];
+
+               r600_blit_decompress_depth(ctx, rtex, staging_depth,
+                                          level, level,
+                                          box->z, box->z + box->depth - 1,
+                                          0, 0);
+
+               trans->transfer.stride = staging_depth->surface.level[level].pitch_bytes;
                trans->offset = r600_texture_get_offset(staging_depth, level, box->z);
                trans->staging = (struct r600_resource*)staging_depth;
                return &trans->transfer;
@@ -910,7 +702,7 @@ struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
                }
 
                trans->transfer.stride =
-                       ((struct r600_resource_texture *)trans->staging)->pitch_in_bytes[0];
+                       ((struct r600_texture *)trans->staging)->surface.level[0].pitch_bytes;
                if (usage & PIPE_TRANSFER_READ) {
                        r600_copy_to_staging_texture(ctx, trans);
                        /* Always referenced in the blit. */
@@ -918,8 +710,8 @@ struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
                }
                return &trans->transfer;
        }
-       trans->transfer.stride = rtex->pitch_in_bytes[level];
-       trans->transfer.layer_stride = rtex->layer_size[level];
+       trans->transfer.stride = rtex->surface.level[level].pitch_bytes;
+       trans->transfer.layer_stride = rtex->surface.level[level].slice_size;
        trans->offset = r600_texture_get_offset(rtex, level, box->z);
        return &trans->transfer;
 }
@@ -929,21 +721,15 @@ void r600_texture_transfer_destroy(struct pipe_context *ctx,
 {
        struct r600_transfer *rtransfer = (struct r600_transfer*)transfer;
        struct pipe_resource *texture = transfer->resource;
-       struct r600_resource_texture *rtex = (struct r600_resource_texture*)texture;
-
-       if (rtex->is_depth) {
-               if ((transfer->usage & PIPE_TRANSFER_WRITE) && rtransfer->staging) {
-                       struct pipe_box sbox;
-
-                       u_box_origin_2d(texture->width0, texture->height0, &sbox);
+       struct r600_texture *rtex = (struct r600_texture*)texture;
 
+       if ((transfer->usage & PIPE_TRANSFER_WRITE) && rtransfer->staging) {
+               if (rtex->is_depth) {
                        ctx->resource_copy_region(ctx, texture, transfer->level,
-                                                 0, 0, transfer->box.z,
+                                                 transfer->box.x, transfer->box.y, transfer->box.z,
                                                  &rtransfer->staging->b.b, transfer->level,
-                                                 &sbox);
-               }
-       } else if (rtransfer->staging) {
-               if (transfer->usage & PIPE_TRANSFER_WRITE) {
+                                                 &transfer->box);
+               } else {
                        r600_copy_from_staging_texture(ctx, rtransfer);
                }
        }
@@ -961,8 +747,8 @@ void* r600_texture_transfer_map(struct pipe_context *ctx,
        struct r600_context *rctx = (struct r600_context *)ctx;
        struct r600_transfer *rtransfer = (struct r600_transfer*)transfer;
        struct radeon_winsys_cs_handle *buf;
-       struct r600_resource_texture *rtex =
-                       (struct r600_resource_texture*)transfer->resource;
+       struct r600_texture *rtex =
+                       (struct r600_texture*)transfer->resource;
        enum pipe_format format = transfer->resource->format;
        unsigned offset = 0;
        char *map;