r600g: work around ddx over alignment
[mesa.git] / src / gallium / drivers / r600 / r600_texture.c
index 412efa5a24365d4c5f1db6a4d0fd6c6e4ecdec32..937345156c3a76ed86ff5bbeac136ab9c3d0a19e 100644 (file)
  *      Jerome Glisse
  *      Corbin Simpson
  */
+#include "r600_formats.h"
+#include "r600d.h"
+
 #include <errno.h>
-#include "pipe/p_screen.h"
-#include "util/u_format.h"
 #include "util/u_format_s3tc.h"
-#include "util/u_math.h"
-#include "util/u_inlines.h"
 #include "util/u_memory.h"
-#include "pipebuffer/pb_buffer.h"
-#include "r600_pipe.h"
-#include "r600_resource.h"
-#include "r600d.h"
-#include "r600_formats.h"
+
 
 /* Copy from a full GPU texture to a transfer's staging one. */
 static void r600_copy_to_staging_texture(struct pipe_context *ctx, struct r600_transfer *rtransfer)
 {
        struct pipe_transfer *transfer = (struct pipe_transfer*)rtransfer;
-       struct pipe_resource *texture = transfer->resource;
+       struct pipe_resource *dst = &rtransfer->staging->b.b;
+       struct pipe_resource *src = transfer->resource;
 
-       ctx->resource_copy_region(ctx, rtransfer->staging_texture,
-                               0, 0, 0, 0, texture, transfer->level,
-                               &transfer->box);
+       if (src->nr_samples <= 1) {
+               ctx->resource_copy_region(ctx, dst, 0, 0, 0, 0,
+                                         src, transfer->level, &transfer->box);
+       } else {
+               /* Resolve the resource. */
+               struct pipe_blit_info blit;
+
+               memset(&blit, 0, sizeof(blit));
+               blit.src.resource = src;
+               blit.src.format = src->format;
+               blit.src.level = transfer->level;
+               blit.src.box = transfer->box;
+               blit.dst.resource = dst;
+               blit.dst.format = dst->format;
+               blit.dst.box.width = transfer->box.width;
+               blit.dst.box.height = transfer->box.height;
+               blit.dst.box.depth = transfer->box.depth;
+               blit.mask = PIPE_MASK_RGBA;
+               blit.filter = PIPE_TEX_FILTER_NEAREST;
+
+               ctx->blit(ctx, &blit);
+       }
 }
 
-
 /* Copy from a transfer's staging texture to a full GPU one. */
 static void r600_copy_from_staging_texture(struct pipe_context *ctx, struct r600_transfer *rtransfer)
 {
@@ -56,487 +70,531 @@ static void r600_copy_from_staging_texture(struct pipe_context *ctx, struct r600
        struct pipe_resource *texture = transfer->resource;
        struct pipe_box sbox;
 
-       sbox.x = sbox.y = sbox.z = 0;
-       sbox.width = transfer->box.width;
-       sbox.height = transfer->box.height;
-       /* XXX that might be wrong */
-       sbox.depth = 1;
+       u_box_origin_2d(transfer->box.width, transfer->box.height, &sbox);
+
        ctx->resource_copy_region(ctx, texture, transfer->level,
                                  transfer->box.x, transfer->box.y, transfer->box.z,
-                                 rtransfer->staging_texture,
+                                 &rtransfer->staging->b.b,
                                  0, &sbox);
-
-       r600_flush(ctx, NULL, RADEON_FLUSH_ASYNC);
 }
 
-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.b.target) {
-       case PIPE_TEXTURE_3D:
-       case PIPE_TEXTURE_CUBE:
-       default:
-               return offset + layer * rtex->layer_size[level];
-       }
+       return rtex->surface.level[level].offset +
+              layer * rtex->surface.level[level].slice_size;
 }
 
-static unsigned r600_get_block_alignment(struct pipe_screen *screen,
-                                        enum pipe_format format,
-                                        unsigned array_mode)
+static int r600_init_surface(struct r600_screen *rscreen,
+                            struct radeon_surface *surface,
+                            const struct pipe_resource *ptex,
+                            unsigned array_mode,
+                            bool is_flushed_depth,
+                            bool from_ddx)
 {
-       struct r600_screen* rscreen = (struct r600_screen *)screen;
-       unsigned pixsize = util_format_get_blocksize(format);
-       int p_align;
+       const struct util_format_description *desc =
+               util_format_description(ptex->format);
+       bool is_depth, is_stencil;
+
+       is_depth = util_format_has_depth(desc);
+       is_stencil = util_format_has_stencil(desc);
+
+       surface->npix_x = ptex->width0;
+       surface->npix_y = ptex->height0;
+       surface->npix_z = ptex->depth0;
+       surface->blk_w = util_format_get_blockwidth(ptex->format);
+       surface->blk_h = util_format_get_blockheight(ptex->format);
+       surface->blk_d = 1;
+       surface->array_size = 1;
+       surface->last_level = ptex->last_level;
+
+       if (from_ddx) {
+               surface->npix_y = align(surface->npix_y, 8);
+       }
+
+       if (rscreen->chip_class >= EVERGREEN && !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;
+               }
+       }
 
-       switch(array_mode) {
+       surface->nsamples = ptex->nr_samples ? ptex->nr_samples : 1;
+       surface->flags = 0;
+
+       switch (array_mode) {
        case V_038000_ARRAY_1D_TILED_THIN1:
-               p_align = MAX2(8,
-                              ((rscreen->tiling_info.group_bytes / 8 / pixsize)));
+               surface->flags |= RADEON_SURF_SET(RADEON_SURF_MODE_1D, MODE);
                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;
+               surface->flags |= RADEON_SURF_SET(RADEON_SURF_MODE_2D, MODE);
                break;
        case V_038000_ARRAY_LINEAR_ALIGNED:
-               p_align = MAX2(64, rscreen->tiling_info.group_bytes / pixsize);
+               surface->flags |= RADEON_SURF_SET(RADEON_SURF_MODE_LINEAR_ALIGNED, MODE);
                break;
        case V_038000_ARRAY_LINEAR_GENERAL:
        default:
-               p_align = rscreen->tiling_info.group_bytes / pixsize;
+               surface->flags |= RADEON_SURF_SET(RADEON_SURF_MODE_LINEAR, MODE);
                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;
+       switch (ptex->target) {
+       case PIPE_TEXTURE_1D:
+               surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_1D, TYPE);
                break;
-       case V_038000_ARRAY_1D_TILED_THIN1:
-       case V_038000_ARRAY_LINEAR_ALIGNED:
-               h_align = 8;
+       case PIPE_TEXTURE_RECT:
+       case PIPE_TEXTURE_2D:
+               surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_2D, TYPE);
                break;
-       case V_038000_ARRAY_LINEAR_GENERAL:
-       default:
-               h_align = 1;
+       case PIPE_TEXTURE_3D:
+               surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_3D, TYPE);
+               break;
+       case PIPE_TEXTURE_1D_ARRAY:
+               surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_1D_ARRAY, TYPE);
+               surface->array_size = ptex->array_size;
+               break;
+       case PIPE_TEXTURE_2D_ARRAY:
+       case PIPE_TEXTURE_CUBE_ARRAY: /* cube array layout like 2d layout for now */
+               surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_2D_ARRAY, TYPE);
+               surface->array_size = ptex->array_size;
+               break;
+       case PIPE_TEXTURE_CUBE:
+               surface->flags |= RADEON_SURF_SET(RADEON_SURF_TYPE_CUBEMAP, TYPE);
                break;
+       case PIPE_BUFFER:
+       default:
+               return -EINVAL;
+       }
+       if (ptex->bind & PIPE_BIND_SCANOUT) {
+               surface->flags |= RADEON_SURF_SCANOUT;
        }
-       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;
+       if (!is_flushed_depth && is_depth) {
+               surface->flags |= RADEON_SURF_ZBUFFER;
 
-       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;
+               if (is_stencil) {
+                       surface->flags |= RADEON_SURF_SBUFFER |
+                                          RADEON_SURF_HAS_SBUFFER_MIPTREE;
+               }
        }
-       return b_align;
+       return 0;
 }
 
-static unsigned mip_minify(unsigned size, unsigned level)
+static int r600_setup_surface(struct pipe_screen *screen,
+                             struct r600_texture *rtex,
+                             unsigned pitch_in_bytes_override)
 {
-       unsigned val;
-       val = u_minify(size, level);
-       if (level > 0)
-               val = util_next_power_of_two(val);
-       return val;
+       struct pipe_resource *ptex = &rtex->resource.b.b;
+       struct r600_screen *rscreen = (struct r600_screen*)screen;
+       unsigned i;
+       int r;
+
+       r = rscreen->ws->surface_init(rscreen->ws, &rtex->surface);
+       if (r) {
+               return r;
+       }
+       rtex->size = rtex->surface.bo_size;
+       if (pitch_in_bytes_override && pitch_in_bytes_override != rtex->surface.level[0].pitch_bytes) {
+               /* old ddx on evergreen over estimate alignment for 1d, only 1 level
+                * for those
+                */
+               rtex->surface.level[0].nblk_x = pitch_in_bytes_override / rtex->surface.bpe;
+               rtex->surface.level[0].pitch_bytes = pitch_in_bytes_override;
+               rtex->surface.level[0].slice_size = pitch_in_bytes_override * rtex->surface.level[0].nblk_y;
+               if (rtex->surface.flags & RADEON_SURF_SBUFFER) {
+                       rtex->surface.stencil_offset =
+                       rtex->surface.stencil_level[0].offset = rtex->surface.level[0].slice_size;
+               }
+       }
+       for (i = 0; i <= ptex->last_level; i++) {
+               switch (rtex->surface.level[i].mode) {
+               case RADEON_SURF_MODE_LINEAR_ALIGNED:
+                       rtex->array_mode[i] = V_038000_ARRAY_LINEAR_ALIGNED;
+                       break;
+               case RADEON_SURF_MODE_1D:
+                       rtex->array_mode[i] = V_038000_ARRAY_1D_TILED_THIN1;
+                       break;
+               case RADEON_SURF_MODE_2D:
+                       rtex->array_mode[i] = V_038000_ARRAY_2D_TILED_THIN1;
+                       break;
+               default:
+               case RADEON_SURF_MODE_LINEAR:
+                       rtex->array_mode[i] = 0;
+                       break;
+               }
+       }
+       return 0;
 }
 
-static unsigned r600_texture_get_nblocksx(struct pipe_screen *screen,
-                                         struct r600_resource_texture *rtex,
-                                         unsigned level)
+static boolean r600_texture_get_handle(struct pipe_screen* screen,
+                                       struct pipe_resource *ptex,
+                                       struct winsys_handle *whandle)
 {
-       struct pipe_resource *ptex = &rtex->resource.b.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;
+       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;
 
-       width = mip_minify(ptex->width0, level);
-       nblocksx = util_format_get_nblocksx(rtex->real_format, width);
+       rscreen->ws->buffer_set_tiling(resource->buf,
+                                      NULL,
+                                      surface->level[0].mode >= RADEON_SURF_MODE_1D ?
+                                      RADEON_LAYOUT_TILED : RADEON_LAYOUT_LINEAR,
+                                      surface->level[0].mode >= RADEON_SURF_MODE_2D ?
+                                      RADEON_LAYOUT_TILED : RADEON_LAYOUT_LINEAR,
+                                      surface->bankw, surface->bankh,
+                                      surface->tile_split,
+                                      surface->stencil_tile_split,
+                                      surface->mtilea,
+                                      rtex->surface.level[0].pitch_bytes);
 
-       block_align = r600_get_block_alignment(screen, rtex->real_format,
-                                             rtex->array_mode[level]);
-       nblocksx = align(nblocksx, block_align);
-       return nblocksx;
+       return rscreen->ws->buffer_get_handle(resource->buf,
+                                             rtex->surface.level[0].pitch_bytes, whandle);
 }
 
-static unsigned r600_texture_get_nblocksy(struct pipe_screen *screen,
-                                         struct r600_resource_texture *rtex,
-                                         unsigned level)
+static void r600_texture_destroy(struct pipe_screen *screen,
+                                struct pipe_resource *ptex)
 {
-       struct pipe_resource *ptex = &rtex->resource.b.b.b;
-       unsigned height, tile_height;
+       struct r600_texture *rtex = (struct r600_texture*)ptex;
+       struct r600_resource *resource = &rtex->resource;
 
-       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]);
+       if (rtex->flushed_depth_texture)
+               pipe_resource_reference((struct pipe_resource **)&rtex->flushed_depth_texture, NULL);
 
-       /* 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;
+       pb_reference(&resource->buf, NULL);
+       FREE(rtex);
 }
 
-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.b;
+static const struct u_resource_vtbl r600_texture_vtbl;
 
-       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;
+/* 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 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;
+       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;
        }
-}
-
-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.b;
-       struct radeon *radeon = ((struct r600_screen*)screen)->radeon;
-       enum chip_class chipc = r600_get_family_class(radeon);
-       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;
+       /* R600-R700 errata? Anyway, this fixes colorbuffer corruption. */
+       if (rscreen->chip_class <= R700) {
+               fmask.bpe *= 2;
+       }
 
-               /* 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;
+       if (rscreen->chip_class >= EVERGREEN) {
+               fmask.bankh = nr_samples <= 4 ? 4 : 1;
+       }
 
-               offset += size;
+       if (rscreen->ws->surface_init(rscreen->ws, &fmask)) {
+               R600_ERR("Got error in surface_init while allocating FMASK.\n");
+               return;
        }
-       rtex->size = offset;
+       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;
 }
 
-/* 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)
+static void r600_texture_allocate_fmask(struct r600_screen *rscreen,
+                                       struct r600_texture *rtex)
 {
-       unsigned bind;
+       struct r600_fmask_info fmask;
 
-       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;
-
-       switch (res->usage) {
-       case PIPE_USAGE_STREAM:
-       case PIPE_USAGE_STAGING:
-               return FALSE;
+       r600_texture_get_fmask_info(rscreen, rtex,
+                                   rtex->resource.b.b.nr_samples, &fmask);
 
-       default:
-               return TRUE;
-       }
+       /* 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
 }
 
-static boolean r600_texture_get_handle(struct pipe_screen* screen,
-                                       struct pipe_resource *ptex,
-                                       struct winsys_handle *whandle)
+void r600_texture_get_cmask_info(struct r600_screen *rscreen,
+                                struct r600_texture *rtex,
+                                struct r600_cmask_info *out)
 {
-       struct r600_resource_texture *rtex = (struct r600_resource_texture*)ptex;
-       struct r600_resource *resource = &rtex->resource;
-       struct r600_screen *rscreen = (struct r600_screen*)screen;
-
-       return rscreen->ws->buffer_get_handle(resource->buf,
-                                             rtex->pitch_in_bytes[0], whandle);
+       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_destroy(struct pipe_screen *screen,
-                                struct pipe_resource *ptex)
+static void r600_texture_allocate_cmask(struct r600_screen *rscreen,
+                                       struct r600_texture *rtex)
 {
-       struct r600_resource_texture *rtex = (struct r600_resource_texture*)ptex;
-       struct r600_resource *resource = &rtex->resource;
+       struct r600_cmask_info cmask;
 
-       if (rtex->flushed_depth_texture)
-               pipe_resource_reference((struct pipe_resource **)&rtex->flushed_depth_texture, NULL);
+       r600_texture_get_cmask_info(rscreen, rtex, &cmask);
 
-       pb_reference(&resource->buf, NULL);
-       FREE(rtex);
+       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 const struct u_resource_vtbl r600_texture_vtbl =
-{
-       r600_texture_get_handle,        /* get_handle */
-       r600_texture_destroy,           /* resource_destroy */
-       r600_texture_get_transfer,      /* get_transfer */
-       r600_texture_transfer_destroy,  /* transfer_destroy */
-       r600_texture_transfer_map,      /* transfer_map */
-       u_default_transfer_flush_region,/* transfer_flush_region */
-       r600_texture_transfer_unmap,    /* transfer_unmap */
-       u_default_transfer_inline_write /* transfer_inline_write */
-};
+DEBUG_GET_ONCE_BOOL_OPTION(print_texdepth, "R600_PRINT_TEXDEPTH", FALSE);
 
-static struct r600_resource_texture *
+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)
+                          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;
 
        resource = &rtex->resource;
-       resource->b.b.b = *base;
-       resource->b.b.vtbl = &r600_texture_vtbl;
-       pipe_reference_init(&resource->b.b.b.reference, 1);
-       resource->b.b.b.screen = screen;
+       resource->b.b = *base;
+       resource->b.vtbl = &r600_texture_vtbl;
+       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->flags & R600_RESOURCE_FLAG_TRANSFER) &&
-           r600_get_family_class(((struct r600_screen*)screen)->radeon) >= EVERGREEN &&
-           util_format_is_depth_and_stencil(base->format)) {
-               struct pipe_resource stencil;
-               unsigned stencil_pitch_override = 0;
-
-               switch (base->format) {
-               case PIPE_FORMAT_Z24_UNORM_S8_USCALED:
-                       rtex->real_format = PIPE_FORMAT_Z24X8_UNORM;
-                       break;
-               case PIPE_FORMAT_S8_USCALED_Z24_UNORM:
-                       rtex->real_format = PIPE_FORMAT_X8Z24_UNORM;
-                       break;
-               case PIPE_FORMAT_Z32_FLOAT_S8X24_USCALED:
-                       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_USCALED ||
-                              base->format == PIPE_FORMAT_S8_USCALED_Z24_UNORM);
-                       stencil_pitch_override = pitch_in_bytes_override / 4;
-               }
+       /* 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));
 
-               /* Allocate the stencil buffer. */
-               stencil = *base;
-               stencil.format = PIPE_FORMAT_S8_USCALED;
-               rtex->stencil = r600_texture_create_object(screen, &stencil, array_mode,
-                                                          stencil_pitch_override,
-                                                          max_buffer_size, NULL, FALSE);
-               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->depth = 1;
-
-       r600_setup_miptree(screen, rtex, array_mode);
-
-       /* 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.b.last_level; i++)
-                       rtex->stencil->offset[i] += stencil_offset;
+       if (base->nr_samples > 1 && !rtex->is_depth && alloc_bo) {
+               r600_texture_allocate_cmask(rscreen, rtex);
+               r600_texture_allocate_fmask(rscreen, rtex);
+       }
 
-               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;
        }
 
+       /* Tiled depth textures utilize the non-displayable tile order. */
+       rtex->non_disp_tiling = rtex->is_depth && rtex->surface.level[0].mode >= RADEON_SURF_MODE_1D;
+
        /* Now create the backing buffer. */
        if (!buf && alloc_bo) {
-               struct pipe_resource *ptex = &rtex->resource.b.b.b;
-               unsigned base_align = r600_get_base_alignment(screen, ptex->format, array_mode);
+               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, base->usage)) {
-                       pipe_resource_reference((struct pipe_resource**)&rtex->stencil, NULL);
+               if (!r600_init_resource(rscreen, resource, rtex->size, base_align, FALSE, usage)) {
                        FREE(rtex);
                        return NULL;
                }
        } else if (buf) {
+               /* This is usually the window framebuffer. We want it in VRAM, always. */
                resource->buf = buf;
                resource->cs_buf = rscreen->ws->buffer_get_cs_handle(buf);
-               resource->domains = RADEON_DOMAIN_GTT | RADEON_DOMAIN_VRAM;
+               resource->domains = RADEON_DOMAIN_VRAM;
        }
 
-       if (rtex->stencil) {
-               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);
+       }
+
+       if (debug_get_option_print_texdepth() && rtex->is_depth && rtex->non_disp_tiling) {
+               printf("Texture: npix_x=%u, npix_y=%u, npix_z=%u, blk_w=%u, "
+                      "blk_h=%u, blk_d=%u, array_size=%u, last_level=%u, "
+                      "bpe=%u, nsamples=%u, flags=%u\n",
+                      rtex->surface.npix_x, rtex->surface.npix_y,
+                      rtex->surface.npix_z, rtex->surface.blk_w,
+                      rtex->surface.blk_h, rtex->surface.blk_d,
+                      rtex->surface.array_size, rtex->surface.last_level,
+                      rtex->surface.bpe, rtex->surface.nsamples,
+                      rtex->surface.flags);
+               if (rtex->surface.flags & RADEON_SURF_ZBUFFER) {
+                       for (int i = 0; i <= rtex->surface.last_level; i++) {
+                               printf("  Z %i: offset=%llu, slice_size=%llu, npix_x=%u, "
+                                      "npix_y=%u, npix_z=%u, nblk_x=%u, nblk_y=%u, "
+                                      "nblk_z=%u, pitch_bytes=%u, mode=%u\n",
+                                      i, (unsigned long long)rtex->surface.level[i].offset,
+                                      (unsigned long long)rtex->surface.level[i].slice_size,
+                                      rtex->surface.level[i].npix_x,
+                                      rtex->surface.level[i].npix_y,
+                                      rtex->surface.level[i].npix_z,
+                                      rtex->surface.level[i].nblk_x,
+                                      rtex->surface.level[i].nblk_y,
+                                      rtex->surface.level[i].nblk_z,
+                                      rtex->surface.level[i].pitch_bytes,
+                                      rtex->surface.level[i].mode);
+                       }
+               }
+               if (rtex->surface.flags & RADEON_SURF_SBUFFER) {
+                       for (int i = 0; i <= rtex->surface.last_level; i++) {
+                               printf("  S %i: offset=%llu, slice_size=%llu, npix_x=%u, "
+                                      "npix_y=%u, npix_z=%u, nblk_x=%u, nblk_y=%u, "
+                                      "nblk_z=%u, pitch_bytes=%u, mode=%u\n",
+                                      i, (unsigned long long)rtex->surface.stencil_level[i].offset,
+                                      (unsigned long long)rtex->surface.stencil_level[i].slice_size,
+                                      rtex->surface.stencil_level[i].npix_x,
+                                      rtex->surface.stencil_level[i].npix_y,
+                                      rtex->surface.stencil_level[i].npix_z,
+                                      rtex->surface.stencil_level[i].nblk_x,
+                                      rtex->surface.stencil_level[i].nblk_y,
+                                      rtex->surface.stencil_level[i].nblk_z,
+                                      rtex->surface.stencil_level[i].pitch_bytes,
+                                      rtex->surface.stencil_level[i].mode);
+                       }
+               }
        }
        return rtex;
 }
 
-DEBUG_GET_ONCE_BOOL_OPTION(tiling_enabled, "R600_TILING", FALSE);
-
 struct pipe_resource *r600_texture_create(struct pipe_screen *screen,
                                                const struct pipe_resource *templ)
 {
        struct r600_screen *rscreen = (struct r600_screen*)screen;
-       unsigned array_mode = 0;
+       struct radeon_surface surface;
+       const struct util_format_description *desc = util_format_description(templ->format);
+       unsigned array_mode;
+       int r;
+
+       /* Default tiling mode for staging textures. */
+       array_mode = V_038000_ARRAY_LINEAR_ALIGNED;
 
+       /* Tiling doesn't work with the 422 (SUBSAMPLED) formats. That's not an issue,
+        * because 422 formats are used for videos, which prefer linear buffers
+        * for fast uploads anyway. */
        if (!(templ->flags & R600_RESOURCE_FLAG_TRANSFER) &&
-           !(templ->bind & PIPE_BIND_SCANOUT)) {
-               if (util_format_is_compressed(templ->format)) {
-                       array_mode = V_038000_ARRAY_1D_TILED_THIN1;
-               }
-               else if (debug_get_option_tiling_enabled() &&
-                        rscreen->info.drm_minor >= 9 &&
-                        permit_hardware_blit(screen, templ)) {
+           desc->layout != UTIL_FORMAT_LAYOUT_SUBSAMPLED) {
+               if (!(templ->bind & PIPE_BIND_SCANOUT) &&
+                   templ->usage != PIPE_USAGE_STAGING &&
+                   templ->usage != PIPE_USAGE_STREAM &&
+                   templ->target != PIPE_TEXTURE_1D &&
+                   templ->target != PIPE_TEXTURE_1D_ARRAY &&
+                   templ->height0 > 3) {
                        array_mode = V_038000_ARRAY_2D_TILED_THIN1;
+               } else if (util_format_is_compressed(templ->format)) {
+                       array_mode = V_038000_ARRAY_1D_TILED_THIN1;
                }
        }
 
-       return (struct pipe_resource *)r600_texture_create_object(screen, templ, array_mode,
-                                                                 0, 0, NULL, TRUE);
+       r = r600_init_surface(rscreen, &surface, templ, array_mode,
+                             templ->flags & R600_RESOURCE_FLAG_FLUSHED_DEPTH,
+                             false);
+       if (r) {
+               return NULL;
+       }
+       r = rscreen->ws->surface_best(rscreen->ws, &surface);
+       if (r) {
+               return NULL;
+       }
+       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_surface *r600_create_surface_custom(struct pipe_context *pipe,
                                                struct pipe_resource *texture,
-                                               const struct pipe_surface *surf_tmpl)
+                                               const struct pipe_surface *templ,
+                                               unsigned width, unsigned height)
 {
-       struct r600_resource_texture *rtex = (struct r600_resource_texture*)texture;
        struct r600_surface *surface = CALLOC_STRUCT(r600_surface);
-       unsigned level = surf_tmpl->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;
-       /* XXX no offset */
-/*     offset = r600_texture_get_offset(rtex, level, surf_tmpl->u.tex.first_layer);*/
        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 = width;
+       surface->base.height = height;
+       surface->base.u = templ->u;
        return &surface->base;
 }
 
+static struct pipe_surface *r600_create_surface(struct pipe_context *pipe,
+                                               struct pipe_resource *texture,
+                                               const struct pipe_surface *templ)
+{
+       struct r600_texture *rtex = (struct r600_texture*)texture;
+       unsigned level = templ->u.tex.level;
+
+       return r600_create_surface_custom(pipe, texture, templ,
+                                         rtex->surface.level[level].npix_x,
+                                         rtex->surface.level[level].npix_y);
+}
+
 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);
 }
@@ -550,37 +608,50 @@ struct pipe_resource *r600_texture_from_handle(struct pipe_screen *screen,
        unsigned stride = 0;
        unsigned array_mode = 0;
        enum radeon_bo_layout micro, macro;
+       struct radeon_surface surface;
+       int r;
 
        /* Support only 2D textures without mipmaps */
        if ((templ->target != PIPE_TEXTURE_2D && templ->target != PIPE_TEXTURE_RECT) ||
              templ->depth0 != 1 || templ->last_level != 0)
                return NULL;
 
-       buf = rscreen->ws->buffer_from_handle(rscreen->ws, whandle, &stride, NULL);
+       buf = rscreen->ws->buffer_from_handle(rscreen->ws, whandle, &stride);
        if (!buf)
                return NULL;
 
-       rscreen->ws->buffer_get_tiling(buf, &micro, &macro);
+       rscreen->ws->buffer_get_tiling(buf, &micro, &macro,
+                                      &surface.bankw, &surface.bankh,
+                                      &surface.tile_split,
+                                      &surface.stencil_tile_split,
+                                      &surface.mtilea);
 
        if (macro == RADEON_LAYOUT_TILED)
                array_mode = V_0280A0_ARRAY_2D_TILED_THIN1;
        else if (micro == RADEON_LAYOUT_TILED)
                array_mode = V_0280A0_ARRAY_1D_TILED_THIN1;
        else
-               array_mode = 0;
+               array_mode = V_038000_ARRAY_LINEAR_ALIGNED;
 
-       return (struct pipe_resource *)r600_texture_create_object(screen, templ, array_mode,
-                                                                 stride, 0, buf, FALSE);
+       r = r600_init_surface(rscreen, &surface, templ, array_mode, false, true);
+       if (r) {
+               return NULL;
+       }
+       return (struct pipe_resource *)r600_texture_create_object(screen, templ,
+                                                                 stride, buf, FALSE, &surface);
 }
 
-int r600_texture_depth_flush(struct pipe_context *ctx,
-                            struct pipe_resource *texture, boolean just_create)
+bool r600_init_flushed_depth_texture(struct pipe_context *ctx,
+                                    struct pipe_resource *texture,
+                                    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_texture **flushed_depth_texture = staging ?
+                       staging : &rtex->flushed_depth_texture;
 
-       if (rtex->flushed_depth_texture)
-               goto out;
+       if (!staging && rtex->flushed_depth_texture)
+               return true; /* it's ready */
 
        resource.target = texture->target;
        resource.format = texture->format;
@@ -590,45 +661,44 @@ int r600_texture_depth_flush(struct pipe_context *ctx,
        resource.array_size = texture->array_size;
        resource.last_level = texture->last_level;
        resource.nr_samples = texture->nr_samples;
-       resource.usage = PIPE_USAGE_DYNAMIC;
-       resource.bind = texture->bind | PIPE_BIND_DEPTH_STENCIL;
-       resource.flags = R600_RESOURCE_FLAG_TRANSFER | texture->flags;
-
-       rtex->flushed_depth_texture = (struct r600_resource_texture *)ctx->screen->resource_create(ctx->screen, &resource);
-       if (rtex->flushed_depth_texture == NULL) {
-               R600_ERR("failed to create temporary texture to hold untiled copy\n");
-               return -ENOMEM;
-       }
+       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;
 
-       ((struct r600_resource_texture *)rtex->flushed_depth_texture)->is_flushing_texture = TRUE;
-out:
-       if (just_create)
-               return 0;
+       if (staging)
+               resource.flags |= R600_RESOURCE_FLAG_TRANSFER;
 
-       /* XXX: only do this if the depth texture has actually changed:
-        */
-       r600_blit_uncompress_depth(ctx, rtex);
-       return 0;
-}
+       *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 false;
+       }
 
-/* Needs adjustment for pixelformat:
- */
-static INLINE unsigned u_box_volume( const struct pipe_box *box )
-{
-       return box->width * box->depth * box->height;
-};
+       (*flushed_depth_texture)->is_flushing_texture = TRUE;
+       (*flushed_depth_texture)->non_disp_tiling = false;
+       return true;
+}
 
-struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
-                                               struct pipe_resource *texture,
-                                               unsigned level,
-                                               unsigned usage,
-                                               const struct pipe_box *box)
+static void *r600_texture_transfer_map(struct pipe_context *ctx,
+                                      struct pipe_resource *texture,
+                                      unsigned level,
+                                      unsigned usage,
+                                      const struct pipe_box *box,
+                                      struct pipe_transfer **ptransfer)
 {
-       struct r600_resource_texture *rtex = (struct r600_resource_texture*)texture;
+       struct r600_context *rctx = (struct r600_context*)ctx;
+       struct r600_texture *rtex = (struct r600_texture*)texture;
        struct pipe_resource resource;
        struct r600_transfer *trans;
-       int r;
        boolean use_staging_texture = FALSE;
+       enum pipe_format format = texture->format;
+       struct radeon_winsys_cs_handle *buf;
+       unsigned offset = 0;
+       char *map;
+
+       if ((texture->bind & PIPE_BIND_GLOBAL) && texture->target == PIPE_BUFFER) {
+               return r600_compute_global_transfer_map(ctx, texture, level, usage, box, ptransfer);
+       }
 
        /* We cannot map a tiled texture directly because the data is
         * in a different order, therefore we do detiling using a blit.
@@ -637,52 +707,60 @@ struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
         * the CPU is much happier reading out of cached system memory
         * than uncached VRAM.
         */
-       if (R600_TEX_IS_TILED(rtex, level))
-               use_staging_texture = TRUE;
-
-       if ((usage & PIPE_TRANSFER_READ) && u_box_volume(box) > 1024)
+       if (R600_TEX_IS_TILED(rtex, level)) {
                use_staging_texture = TRUE;
+       }
 
-       /* XXX: Use a staging texture for uploads if the underlying BO
-        * is busy.  No interface for checking that currently? so do
-        * it eagerly whenever the transfer doesn't require a readback
-        * and might block.
-        */
-       if ((usage & PIPE_TRANSFER_WRITE) &&
-                       !(usage & (PIPE_TRANSFER_READ |
-                                       PIPE_TRANSFER_DONTBLOCK |
-                                       PIPE_TRANSFER_UNSYNCHRONIZED)))
+       /* Use a staging texture for uploads if the underlying BO is busy. */
+       if (!(usage & PIPE_TRANSFER_READ) &&
+           (rctx->ws->cs_is_buffer_referenced(rctx->cs, rtex->resource.cs_buf, RADEON_USAGE_READWRITE) ||
+            rctx->ws->buffer_is_busy(rtex->resource.buf, RADEON_USAGE_READWRITE))) {
                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;
+       }
 
-       if (use_staging_texture && (usage & PIPE_TRANSFER_MAP_DIRECTLY))
+       if (use_staging_texture && (usage & PIPE_TRANSFER_MAP_DIRECTLY)) {
                return NULL;
+       }
 
        trans = CALLOC_STRUCT(r600_transfer);
        if (trans == NULL)
                return NULL;
-       pipe_resource_reference(&trans->transfer.resource, texture);
+       trans->transfer.resource = texture;
        trans->transfer.level = level;
        trans->transfer.usage = usage;
        trans->transfer.box = *box;
-       if (rtex->depth) {
+       if (rtex->is_depth) {
                /* XXX: only readback the rectangle which is being mapped?
                */
                /* XXX: when discard is true, no need to read back from depth texture
                */
-               r = r600_texture_depth_flush(ctx, texture, FALSE);
-               if (r < 0) {
+               struct r600_texture *staging_depth;
+
+               assert(rtex->resource.b.b.nr_samples <= 1);
+               if (rtex->resource.b.b.nr_samples > 1) {
+                       R600_ERR("mapping MSAA zbuffer unimplemented\n");
+                       FREE(trans);
+                       return NULL;
+               }
+
+               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 = rtex->flushed_depth_texture->pitch_in_bytes[level];
-               trans->offset = r600_texture_get_offset(rtex->flushed_depth_texture, level, box->z);
-               return &trans->transfer;
+
+               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;
        } else if (use_staging_texture) {
                resource.target = PIPE_TEXTURE_2D;
                resource.format = texture->format;
@@ -706,106 +784,85 @@ struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
                        resource.bind |= PIPE_BIND_SAMPLER_VIEW;
                }
                /* Create the temporary texture. */
-               trans->staging_texture = ctx->screen->resource_create(ctx->screen, &resource);
-               if (trans->staging_texture == NULL) {
+               trans->staging = (struct r600_resource*)ctx->screen->resource_create(ctx->screen, &resource);
+               if (trans->staging == NULL) {
                        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 =
-                       ((struct r600_resource_texture *)trans->staging_texture)->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. */
                        r600_flush(ctx, NULL, 0);
                }
-               return &trans->transfer;
-       }
-       trans->transfer.stride = rtex->pitch_in_bytes[level];
-       trans->transfer.layer_stride = rtex->layer_size[level];
-       trans->offset = r600_texture_get_offset(rtex, level, box->z);
-       return &trans->transfer;
-}
-
-void r600_texture_transfer_destroy(struct pipe_context *ctx,
-                                  struct pipe_transfer *transfer)
-{
-       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 (rtransfer->staging_texture) {
-               if (transfer->usage & PIPE_TRANSFER_WRITE) {
-                       r600_copy_from_staging_texture(ctx, rtransfer);
-               }
-               pipe_resource_reference(&rtransfer->staging_texture, NULL);
-       }
-
-       if (rtex->depth && !rtex->is_flushing_texture) {
-               if ((transfer->usage & PIPE_TRANSFER_WRITE) && rtex->flushed_depth_texture)
-                       r600_blit_push_depth(ctx, rtex);
+       } else {
+               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);
        }
 
-       pipe_resource_reference(&transfer->resource, NULL);
-       FREE(transfer);
-}
-
-void* r600_texture_transfer_map(struct pipe_context *ctx,
-                               struct pipe_transfer* transfer)
-{
-       struct r600_pipe_context *rctx = (struct r600_pipe_context *)ctx;
-       struct r600_transfer *rtransfer = (struct r600_transfer*)transfer;
-       struct pb_buffer *buf;
-       enum pipe_format format = transfer->resource->format;
-       unsigned offset = 0;
-       char *map;
-
-       if (rtransfer->staging_texture) {
-               buf = ((struct r600_resource *)rtransfer->staging_texture)->buf;
+       if (trans->staging) {
+               buf = ((struct r600_resource *)trans->staging)->cs_buf;
        } else {
-               struct r600_resource_texture *rtex = (struct r600_resource_texture*)transfer->resource;
-
-               if (rtex->flushed_depth_texture)
-                       buf = ((struct r600_resource *)rtex->flushed_depth_texture)->buf;
-               else
-                       buf = ((struct r600_resource *)transfer->resource)->buf;
-
-               offset = rtransfer->offset +
-                       transfer->box.y / util_format_get_blockheight(format) * transfer->stride +
-                       transfer->box.x / util_format_get_blockwidth(format) * util_format_get_blocksize(format);
+               buf = ((struct r600_resource *)texture)->cs_buf;
        }
 
-       if (!(map = rctx->ws->buffer_map(buf, rctx->ctx.cs, transfer->usage))) {
+       if (rtex->is_depth || !trans->staging)
+               offset = trans->offset +
+                       box->y / util_format_get_blockheight(format) * trans->transfer.stride +
+                       box->x / util_format_get_blockwidth(format) * util_format_get_blocksize(format);
+
+       if (!(map = rctx->ws->buffer_map(buf, rctx->cs, usage))) {
+               pipe_resource_reference((struct pipe_resource**)&trans->staging, NULL);
+               FREE(trans);
                return NULL;
        }
 
+       *ptransfer = &trans->transfer;
        return map + offset;
 }
 
-void r600_texture_transfer_unmap(struct pipe_context *ctx,
-                                struct pipe_transfer* transfer)
+static void r600_texture_transfer_unmap(struct pipe_context *ctx,
+                                       struct pipe_transfer* transfer)
 {
        struct r600_transfer *rtransfer = (struct r600_transfer*)transfer;
-       struct r600_pipe_context *rctx = (struct r600_pipe_context*)ctx;
-       struct pb_buffer *buf;
+       struct r600_context *rctx = (struct r600_context*)ctx;
+       struct radeon_winsys_cs_handle *buf;
+       struct pipe_resource *texture = transfer->resource;
+       struct r600_texture *rtex = (struct r600_texture*)texture;
 
-       if (rtransfer->staging_texture) {
-               buf = ((struct r600_resource *)rtransfer->staging_texture)->buf;
+       if ((transfer->resource->bind & PIPE_BIND_GLOBAL) && transfer->resource->target == PIPE_BUFFER) {
+               return r600_compute_global_transfer_unmap(ctx, transfer);
+       }
+
+       if (rtransfer->staging) {
+               buf = ((struct r600_resource *)rtransfer->staging)->cs_buf;
        } else {
-               struct r600_resource_texture *rtex = (struct r600_resource_texture*)transfer->resource;
+               buf = ((struct r600_resource *)transfer->resource)->cs_buf;
+       }
+       rctx->ws->buffer_unmap(buf);
 
-               if (rtex->flushed_depth_texture) {
-                       buf = ((struct r600_resource *)rtex->flushed_depth_texture)->buf;
+       if ((transfer->usage & PIPE_TRANSFER_WRITE) && rtransfer->staging) {
+               if (rtex->is_depth) {
+                       ctx->resource_copy_region(ctx, texture, transfer->level,
+                                                 transfer->box.x, transfer->box.y, transfer->box.z,
+                                                 &rtransfer->staging->b.b, transfer->level,
+                                                 &transfer->box);
                } else {
-                       buf = ((struct r600_resource *)transfer->resource)->buf;
+                       r600_copy_from_staging_texture(ctx, rtransfer);
                }
        }
-       rctx->ws->buffer_unmap(buf);
+
+       if (rtransfer->staging)
+               pipe_resource_reference((struct pipe_resource**)&rtransfer->staging, NULL);
+
+       FREE(transfer);
 }
 
-void r600_init_surface_functions(struct r600_pipe_context *r600)
+void r600_init_surface_functions(struct r600_context *r600)
 {
        r600->context.create_surface = r600_create_surface;
        r600->context.surface_destroy = r600_surface_destroy;
@@ -865,6 +922,7 @@ uint32_t r600_translate_texformat(struct pipe_screen *screen,
        const struct util_format_description *desc;
        boolean uniform = TRUE;
        static int r600_enable_s3tc = -1;
+       bool is_srgb_valid = FALSE;
 
        int i;
        const uint32_t sign_bit[4] = {
@@ -879,32 +937,34 @@ uint32_t r600_translate_texformat(struct pipe_screen *screen,
 
        /* Colorspace (return non-RGB formats directly). */
        switch (desc->colorspace) {
-               /* Depth stencil formats */
+       /* Depth stencil formats */
        case UTIL_FORMAT_COLORSPACE_ZS:
                switch (format) {
                case PIPE_FORMAT_Z16_UNORM:
                        result = FMT_16;
                        goto out_word4;
-               case PIPE_FORMAT_X24S8_USCALED:
+               case PIPE_FORMAT_X24S8_UINT:
                        word4 |= S_038010_NUM_FORMAT_ALL(V_038010_SQ_NUM_FORMAT_INT);
                case PIPE_FORMAT_Z24X8_UNORM:
-               case PIPE_FORMAT_Z24_UNORM_S8_USCALED:
+               case PIPE_FORMAT_Z24_UNORM_S8_UINT:
                        result = FMT_8_24;
                        goto out_word4;
-               case PIPE_FORMAT_S8X24_USCALED:
+               case PIPE_FORMAT_S8X24_UINT:
                        word4 |= S_038010_NUM_FORMAT_ALL(V_038010_SQ_NUM_FORMAT_INT);
                case PIPE_FORMAT_X8Z24_UNORM:
-               case PIPE_FORMAT_S8_USCALED_Z24_UNORM:
+               case PIPE_FORMAT_S8_UINT_Z24_UNORM:
                        result = FMT_24_8;
                        goto out_word4;
-               case PIPE_FORMAT_S8_USCALED:
+               case PIPE_FORMAT_S8_UINT:
                        result = FMT_8;
                        word4 |= S_038010_NUM_FORMAT_ALL(V_038010_SQ_NUM_FORMAT_INT);
                        goto out_word4;
                case PIPE_FORMAT_Z32_FLOAT:
                        result = FMT_32_FLOAT;
                        goto out_word4;
-               case PIPE_FORMAT_Z32_FLOAT_S8X24_USCALED:
+               case PIPE_FORMAT_X32_S8X24_UINT:
+                       word4 |= S_038010_NUM_FORMAT_ALL(V_038010_SQ_NUM_FORMAT_INT);
+               case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
                        result = FMT_X24_8_32_FLOAT;
                        goto out_word4;
                default:
@@ -919,7 +979,7 @@ uint32_t r600_translate_texformat(struct pipe_screen *screen,
                default:
                        break;
                }
-               goto out_unknown; /* TODO */
+               goto out_unknown; /* XXX */
 
        case UTIL_FORMAT_COLORSPACE_SRGB:
                word4 |= S_038010_FORCE_DEGAMMA(1);
@@ -976,14 +1036,32 @@ uint32_t r600_translate_texformat(struct pipe_screen *screen,
                case PIPE_FORMAT_DXT1_SRGB:
                case PIPE_FORMAT_DXT1_SRGBA:
                        result = FMT_BC1;
+                       is_srgb_valid = TRUE;
                        goto out_word4;
                case PIPE_FORMAT_DXT3_RGBA:
                case PIPE_FORMAT_DXT3_SRGBA:
                        result = FMT_BC2;
+                       is_srgb_valid = TRUE;
                        goto out_word4;
                case PIPE_FORMAT_DXT5_RGBA:
                case PIPE_FORMAT_DXT5_SRGBA:
                        result = FMT_BC3;
+                       is_srgb_valid = TRUE;
+                       goto out_word4;
+               default:
+                       goto out_unknown;
+               }
+       }
+
+       if (desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED) {
+               switch (format) {
+               case PIPE_FORMAT_R8G8_B8G8_UNORM:
+               case PIPE_FORMAT_G8R8_B8R8_UNORM:
+                       result = FMT_GB_GR;
+                       goto out_word4;
+               case PIPE_FORMAT_G8R8_G8B8_UNORM:
+               case PIPE_FORMAT_R8G8_R8B8_UNORM:
+                       result = FMT_BG_RG;
                        goto out_word4;
                default:
                        goto out_unknown;
@@ -1005,7 +1083,7 @@ uint32_t r600_translate_texformat(struct pipe_screen *screen,
                }
        }
 
-       /* R8G8Bx_SNORM - TODO CxV8U8 */
+       /* R8G8Bx_SNORM - XXX CxV8U8 */
 
        /* See whether the components are of the same size. */
        for (i = 1; i < desc->nr_channels; i++) {
@@ -1014,6 +1092,9 @@ uint32_t r600_translate_texformat(struct pipe_screen *screen,
 
        /* Non-uniform formats. */
        if (!uniform) {
+               if (desc->colorspace != UTIL_FORMAT_COLORSPACE_SRGB &&
+                   desc->channel[0].pure_integer)
+                       word4 |= S_038010_NUM_FORMAT_ALL(V_038010_SQ_NUM_FORMAT_INT);
                switch(desc->nr_channels) {
                case 3:
                        if (desc->channel[0].size == 5 &&
@@ -1064,7 +1145,7 @@ uint32_t r600_translate_texformat(struct pipe_screen *screen,
                }
 #endif
                if (desc->colorspace != UTIL_FORMAT_COLORSPACE_SRGB &&
-                   !desc->channel[i].normalized)
+                   desc->channel[i].pure_integer)
                        word4 |= S_038010_NUM_FORMAT_ALL(V_038010_SQ_NUM_FORMAT_INT);
 
                switch (desc->channel[i].size) {
@@ -1088,6 +1169,7 @@ uint32_t r600_translate_texformat(struct pipe_screen *screen,
                                goto out_word4;
                        case 4:
                                result = FMT_8_8_8_8;
+                               is_srgb_valid = TRUE;
                                goto out_word4;
                        }
                        goto out_unknown;
@@ -1151,6 +1233,9 @@ uint32_t r600_translate_texformat(struct pipe_screen *screen,
        }
 
 out_word4:
+
+       if (desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB && !is_srgb_valid)
+               return ~0;
        if (word4_p)
                *word4_p = word4;
        if (yuv_format_p)
@@ -1160,3 +1245,13 @@ out_unknown:
        /* R600_ERR("Unable to handle texformat %d %s\n", format, util_format_name(format)); */
        return ~0;
 }
+
+static const struct u_resource_vtbl r600_texture_vtbl =
+{
+       r600_texture_get_handle,        /* get_handle */
+       r600_texture_destroy,           /* resource_destroy */
+       r600_texture_transfer_map,      /* transfer_map */
+       NULL,                           /* transfer_flush_region */
+       r600_texture_transfer_unmap,    /* transfer_unmap */
+       NULL                            /* transfer_inline_write */
+};