radeonsi: emit TA_BC_BASE_ADDR_HI for border color on CIK
[mesa.git] / src / gallium / drivers / radeonsi / r600_texture.c
index 38ff36df1e96b9839a3f54b494eb13f673caf0bb..8992f9a1fa29b2683b091fc09c29e04d194e8673 100644 (file)
@@ -42,12 +42,11 @@ static void r600_copy_to_staging_texture(struct pipe_context *ctx, struct r600_t
        struct pipe_transfer *transfer = (struct pipe_transfer*)rtransfer;
        struct pipe_resource *texture = transfer->resource;
 
-       ctx->resource_copy_region(ctx, rtransfer->staging_texture,
+       ctx->resource_copy_region(ctx, rtransfer->staging,
                                0, 0, 0, 0, texture, transfer->level,
                                &transfer->box);
 }
 
-
 /* 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)
 {
@@ -55,195 +54,34 @@ 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_3d(0, 0, 0, transfer->box.width, transfer->box.height, transfer->box.depth, &sbox);
+
        ctx->resource_copy_region(ctx, texture, transfer->level,
                                  transfer->box.x, transfer->box.y, transfer->box.z,
-                                 rtransfer->staging_texture,
+                                 rtransfer->staging,
                                  0, &sbox);
 }
 
-unsigned r600_texture_get_offset(struct r600_resource_texture *rtex,
+static unsigned r600_texture_get_offset(struct r600_resource_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_009910_ARRAY_1D_TILED_THIN1:
-               p_align = MAX2(8,
-                              ((rscreen->tiling_info.group_bytes / 8 / pixsize)));
-               break;
-       case V_009910_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_009910_ARRAY_LINEAR_ALIGNED:
-               p_align = MAX2(64, rscreen->tiling_info.group_bytes / pixsize);
-               break;
-       case V_009910_ARRAY_LINEAR_GENERAL:
-       default:
-               p_align = rscreen->tiling_info.group_bytes / pixsize;
-               break;
-       }
-       return p_align;
+       return rtex->surface.level[level].offset +
+              layer * rtex->surface.level[level].slice_size;
 }
 
-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_009910_ARRAY_2D_TILED_THIN1:
-               h_align = rscreen->tiling_info.num_channels * 8;
-               break;
-       case V_009910_ARRAY_1D_TILED_THIN1:
-       case V_009910_ARRAY_LINEAR_ALIGNED:
-               h_align = 8;
-               break;
-       case V_009910_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_009910_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_009910_ARRAY_1D_TILED_THIN1:
-       case V_009910_ARRAY_LINEAR_ALIGNED:
-       case V_009910_ARRAY_LINEAR_GENERAL:
-       default:
-               b_align = rscreen->tiling_info.group_bytes;
-               break;
-       }
-       return b_align;
-}
-
-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)
+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)
 {
-       struct pipe_resource *ptex = &rtex->resource.b.b;
-
-       switch (array_mode) {
-       case V_009910_ARRAY_LINEAR_GENERAL:
-       case V_009910_ARRAY_LINEAR_ALIGNED:
-       case V_009910_ARRAY_1D_TILED_THIN1:
-       default:
-               rtex->array_mode[level] = array_mode;
-               break;
-       case V_009910_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);
+       const struct util_format_description *desc =
+               util_format_description(ptex->format);
+       bool is_depth, is_stencil;
 
-               w = mip_minify(ptex->width0, level);
-               h = mip_minify(ptex->height0, level);
-               if (w <= tile_width || h <= tile_height)
-                       rtex->array_mode[level] = V_009910_ARRAY_1D_TILED_THIN1;
-               else
-                       rtex->array_mode[level] = array_mode;
-       }
-       break;
-       }
-}
+       is_depth = util_format_has_depth(desc);
+       is_stencil = util_format_has_stencil(desc);
 
-static int r600_init_surface(struct radeon_surface *surface,
-                            const struct pipe_resource *ptex,
-                            unsigned array_mode)
-{
        surface->npix_x = ptex->width0;
        surface->npix_y = ptex->height0;
        surface->npix_z = ptex->depth0;
@@ -252,11 +90,18 @@ 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 (!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->flags = 0;
        switch (array_mode) {
@@ -303,11 +148,15 @@ static int r600_init_surface(struct radeon_surface *surface,
        if (ptex->bind & PIPE_BIND_SCANOUT) {
                surface->flags |= RADEON_SURF_SCANOUT;
        }
-       if (util_format_is_depth_and_stencil(ptex->format)) {
+
+       if (!is_flushed_depth && is_depth) {
                surface->flags |= RADEON_SURF_ZBUFFER;
-               surface->flags |= RADEON_SURF_SBUFFER;
+               if (is_stencil) {
+                       surface->flags |= RADEON_SURF_SBUFFER |
+                               RADEON_SURF_HAS_SBUFFER_MIPTREE;
+               }
        }
-
+       surface->flags |= RADEON_SURF_HAS_TILE_MODE_INDEX;
        return 0;
 }
 
@@ -316,21 +165,13 @@ static int r600_setup_surface(struct pipe_screen *screen,
                              unsigned array_mode,
                              unsigned pitch_in_bytes_override)
 {
-       struct pipe_resource *ptex = &rtex->resource.b.b;
        struct r600_screen *rscreen = (struct r600_screen*)screen;
-       unsigned i;
        int r;
 
-       if (util_format_is_depth_or_stencil(rtex->real_format)) {
-               rtex->surface.flags |= RADEON_SURF_ZBUFFER;
-               rtex->surface.flags |= RADEON_SURF_SBUFFER;
-       }
-
        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
@@ -339,118 +180,13 @@ static int r600_setup_surface(struct pipe_screen *screen,
                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.level[0].slice_size;
-               }
-       }
-       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_009910_ARRAY_LINEAR_ALIGNED;
-                       break;
-               case RADEON_SURF_MODE_1D:
-                       rtex->array_mode[i] = V_009910_ARRAY_1D_TILED_THIN1;
-                       break;
-               case RADEON_SURF_MODE_2D:
-                       rtex->array_mode[i] = V_009910_ARRAY_2D_TILED_THIN1;
-                       break;
-               default:
-               case RADEON_SURF_MODE_LINEAR:
-                       rtex->array_mode[i] = 0;
-                       break;
+                       rtex->surface.stencil_offset =
+                       rtex->surface.stencil_level[0].offset = rtex->surface.level[0].slice_size;
                }
        }
        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 (array_mode == V_009910_ARRAY_LINEAR_GENERAL)
-                       layer_size = align(nblocksx, 64) * nblocksy * blocksize;
-               else
-                       layer_size = nblocksx * nblocksy * blocksize;
-
-               if (ptex->target == PIPE_TEXTURE_CUBE) {
-                       if (chipc >= CAYMAN)
-                               size = layer_size * 8;
-               }
-               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;
-
-       switch (res->usage) {
-       case PIPE_USAGE_STREAM:
-       case PIPE_USAGE_STAGING:
-               return FALSE;
-
-       default:
-               return TRUE;
-       }
-}
-
 static boolean r600_texture_get_handle(struct pipe_screen* screen,
                                        struct pipe_resource *ptex,
                                        struct winsys_handle *whandle)
@@ -470,10 +206,10 @@ static boolean r600_texture_get_handle(struct pipe_screen* screen,
                                       surface->tile_split,
                                       surface->stencil_tile_split,
                                       surface->mtilea,
-                                      rtex->pitch_in_bytes[0]);
+                                      surface->level[0].pitch_bytes);
 
        return rscreen->ws->buffer_get_handle(resource->buf,
-                                             rtex->pitch_in_bytes[0], whandle);
+                                             surface->level[0].pitch_bytes, whandle);
 }
 
 static void r600_texture_destroy(struct pipe_screen *screen,
@@ -483,27 +219,205 @@ static void r600_texture_destroy(struct pipe_screen *screen,
        struct si_resource *resource = &rtex->resource;
 
        if (rtex->flushed_depth_texture)
-               si_resource_reference(&rtex->flushed_depth_texture, NULL);
-
-       if (rtex->stencil)
-               si_resource_reference(&rtex->stencil, NULL);
+               si_resource_reference((struct si_resource **)&rtex->flushed_depth_texture, NULL);
 
        pb_reference(&resource->buf, NULL);
        FREE(rtex);
 }
 
+static void *si_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_context *rctx = (struct r600_context *)ctx;
+       struct r600_resource_texture *rtex = (struct r600_resource_texture*)texture;
+       struct r600_transfer *trans;
+       boolean use_staging_texture = FALSE;
+       struct radeon_winsys_cs_handle *buf;
+       enum pipe_format format = texture->format;
+       unsigned offset = 0;
+       char *map;
+
+       /* We cannot map a tiled texture directly because the data is
+        * in a different order, therefore we do detiling using a blit.
+        *
+        * Also, use a temporary in GTT memory for read transfers, as
+        * the CPU is much happier reading out of cached system memory
+        * than uncached VRAM.
+        */
+       if (rtex->surface.level[level].mode != RADEON_SURF_MODE_LINEAR_ALIGNED &&
+           rtex->surface.level[level].mode != RADEON_SURF_MODE_LINEAR)
+               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_staging_texture = TRUE;
+
+       if (texture->flags & R600_RESOURCE_FLAG_TRANSFER)
+               use_staging_texture = FALSE;
+
+       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.level = level;
+       trans->transfer.usage = usage;
+       trans->transfer.box = *box;
+       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
+               */
+               struct r600_resource_texture *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;
+               }
+               si_blit_uncompress_depth(ctx, rtex, staging_depth,
+                                        level, level,
+                                        box->z, box->z + box->depth - 1);
+               trans->transfer.stride = staging_depth->surface.level[level].pitch_bytes;
+               trans->transfer.layer_stride = staging_depth->surface.level[level].slice_size;
+               trans->offset = r600_texture_get_offset(staging_depth, level, box->z);
+
+               trans->staging = &staging_depth->resource.b.b;
+       } else if (use_staging_texture) {
+               struct pipe_resource resource;
+               struct r600_resource_texture *staging;
+
+               memset(&resource, 0, sizeof(resource));
+               resource.format = texture->format;
+               resource.width0 = box->width;
+               resource.height0 = box->height;
+               resource.depth0 = 1;
+               resource.array_size = 1;
+               resource.usage = PIPE_USAGE_STAGING;
+               resource.flags = R600_RESOURCE_FLAG_TRANSFER;
+
+               /* We must set the correct texture target and dimensions if needed for a 3D transfer. */
+               if (box->depth > 1 && util_max_layer(texture, level) > 0)
+                       resource.target = texture->target;
+               else
+                       resource.target = PIPE_TEXTURE_2D;
+
+               switch (resource.target) {
+               case PIPE_TEXTURE_1D_ARRAY:
+               case PIPE_TEXTURE_2D_ARRAY:
+               case PIPE_TEXTURE_CUBE_ARRAY:
+                       resource.array_size = box->depth;
+                       break;
+               case PIPE_TEXTURE_3D:
+                       resource.depth0 = box->depth;
+                       break;
+               default:;
+               }
+               /* Create the temporary texture. */
+               staging = (struct r600_resource_texture*)ctx->screen->resource_create(ctx->screen, &resource);
+               if (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->staging = &staging->resource.b.b;
+               trans->transfer.stride = staging->surface.level[0].pitch_bytes;
+               trans->transfer.layer_stride = staging->surface.level[0].slice_size;
+               if (usage & PIPE_TRANSFER_READ) {
+                       r600_copy_to_staging_texture(ctx, trans);
+                       /* Always referenced in the blit. */
+                       radeonsi_flush(ctx, NULL, 0);
+               }
+       } 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);
+       }
+
+       if (trans->staging) {
+               buf = si_resource(trans->staging)->cs_buf;
+       } else {
+               buf = rtex->resource.cs_buf;
+       }
+
+       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(&trans->staging, NULL);
+               pipe_resource_reference(&trans->transfer.resource, NULL);
+               FREE(trans);
+               return NULL;
+       }
+
+       *ptransfer = &trans->transfer;
+       return map + offset;
+}
+
+static void si_texture_transfer_unmap(struct pipe_context *ctx,
+                                     struct pipe_transfer* transfer)
+{
+       struct r600_transfer *rtransfer = (struct r600_transfer*)transfer;
+       struct r600_context *rctx = (struct r600_context*)ctx;
+       struct radeon_winsys_cs_handle *buf;
+       struct pipe_resource *texture = transfer->resource;
+       struct r600_resource_texture *rtex = (struct r600_resource_texture*)texture;
+
+       if (rtransfer->staging) {
+               buf = si_resource(rtransfer->staging)->cs_buf;
+       } else {
+               buf = si_resource(transfer->resource)->cs_buf;
+       }
+       rctx->ws->buffer_unmap(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,
+                                                 &si_resource(rtransfer->staging)->b.b, transfer->level,
+                                                 &transfer->box);
+               } else {
+                       r600_copy_from_staging_texture(ctx, rtransfer);
+               }
+       }
+
+       if (rtransfer->staging)
+               pipe_resource_reference((struct pipe_resource**)&rtransfer->staging, NULL);
+
+       pipe_resource_reference(&transfer->resource, NULL);
+       FREE(transfer);
+}
+
 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 */
+       si_texture_transfer_map,        /* transfer_map */
        u_default_transfer_flush_region,/* transfer_flush_region */
-       r600_texture_transfer_unmap,    /* transfer_unmap */
+       si_texture_transfer_unmap,      /* transfer_unmap */
        NULL    /* transfer_inline_write */
 };
 
+DEBUG_GET_ONCE_BOOL_OPTION(print_texdepth, "RADEON_PRINT_TEXDEPTH", FALSE);
+
 static struct r600_resource_texture *
 r600_texture_create_object(struct pipe_screen *screen,
                           const struct pipe_resource *base,
@@ -531,11 +445,9 @@ r600_texture_create_object(struct pipe_screen *screen,
        rtex->pitch_override = pitch_in_bytes_override;
        rtex->real_format = base->format;
 
-       /* 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;
+       /* 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));
 
-       r600_setup_miptree(screen, rtex, array_mode);
        rtex->surface = *surface;
        r = r600_setup_surface(screen, rtex, array_mode, pitch_in_bytes_override);
        if (r) {
@@ -543,27 +455,13 @@ r600_texture_create_object(struct pipe_screen *screen,
                return NULL;
        }
 
-       /* 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;
-       }
-
        /* 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);
+               unsigned base_align = rtex->surface.bo_alignment;
+               unsigned size = rtex->surface.bo_size;
 
                base_align = rtex->surface.bo_alignment;
-               if (!r600_init_resource(rscreen, resource, rtex->size, base_align, base->bind, base->usage)) {
-                       si_resource_reference(&rtex->stencil, NULL);
+               if (!si_init_resource(rscreen, resource, size, base_align, FALSE, base->usage)) {
                        FREE(rtex);
                        return NULL;
                }
@@ -573,30 +471,73 @@ 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 (debug_get_option_print_texdepth() && rtex->is_depth) {
+               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, rtex->surface.level[i].offset,
+                                      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, rtex->surface.stencil_level[i].offset,
+                                      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;
 }
 
-struct pipe_resource *r600_texture_create(struct pipe_screen *screen,
-                                               const struct pipe_resource *templ)
+struct pipe_resource *si_texture_create(struct pipe_screen *screen,
+                                       const struct pipe_resource *templ)
 {
        struct r600_screen *rscreen = (struct r600_screen*)screen;
        struct radeon_surface surface;
-       unsigned array_mode = 0;
+       unsigned array_mode = V_009910_ARRAY_LINEAR_ALIGNED;
        int r;
 
        if (!(templ->flags & R600_RESOURCE_FLAG_TRANSFER) &&
            !(templ->bind & PIPE_BIND_SCANOUT)) {
-               if (permit_hardware_blit(screen, templ)) {
+               if (util_format_is_compressed(templ->format)) {
+                       array_mode = V_009910_ARRAY_1D_TILED_THIN1;
+               } else {
                        array_mode = V_009910_ARRAY_2D_TILED_THIN1;
                }
        }
 
-       r = r600_init_surface(&surface, templ, array_mode);
+       r = r600_init_surface(rscreen, &surface, templ, array_mode,
+                             templ->flags & R600_RESOURCE_FLAG_FLUSHED_DEPTH);
        if (r) {
                return NULL;
        }
@@ -616,6 +557,8 @@ static struct pipe_surface *r600_create_surface(struct pipe_context *pipe,
        struct r600_surface *surface = CALLOC_STRUCT(r600_surface);
        unsigned level = surf_tmpl->u.tex.level;
 
+       assert(surf_tmpl->u.tex.first_layer <= util_max_layer(texture, surf_tmpl->u.tex.level));
+       assert(surf_tmpl->u.tex.last_layer <= util_max_layer(texture, surf_tmpl->u.tex.level));
        assert(surf_tmpl->u.tex.first_layer == surf_tmpl->u.tex.last_layer);
        if (surface == NULL)
                return NULL;
@@ -625,16 +568,13 @@ static struct pipe_surface *r600_create_surface(struct pipe_context *pipe,
        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.width = rtex->surface.level[level].npix_x;
+       surface->base.height = rtex->surface.level[level].npix_y;
        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);
        return &surface->base;
 }
 
@@ -645,14 +585,14 @@ static void r600_surface_destroy(struct pipe_context *pipe,
        FREE(surface);
 }
 
-struct pipe_resource *r600_texture_from_handle(struct pipe_screen *screen,
-                                              const struct pipe_resource *templ,
-                                              struct winsys_handle *whandle)
+struct pipe_resource *si_texture_from_handle(struct pipe_screen *screen,
+                                            const struct pipe_resource *templ,
+                                            struct winsys_handle *whandle)
 {
        struct r600_screen *rscreen = (struct r600_screen*)screen;
        struct pb_buffer *buf = NULL;
        unsigned stride = 0;
-       unsigned array_mode = 0;
+       unsigned array_mode = V_009910_ARRAY_LINEAR_ALIGNED;
        enum radeon_bo_layout micro, macro;
        struct radeon_surface surface;
        int r;
@@ -677,24 +617,29 @@ struct pipe_resource *r600_texture_from_handle(struct pipe_screen *screen,
        else if (micro == RADEON_LAYOUT_TILED)
                array_mode = V_009910_ARRAY_1D_TILED_THIN1;
        else
-               array_mode = 0;
+               array_mode = V_009910_ARRAY_LINEAR_ALIGNED;
 
-       r = r600_init_surface(&surface, templ, array_mode);
+       r = r600_init_surface(rscreen, &surface, templ, array_mode, false);
        if (r) {
                return NULL;
        }
+       /* always set the scanout flags */
+       surface.flags |= RADEON_SURF_SCANOUT;
        return (struct pipe_resource *)r600_texture_create_object(screen, templ, array_mode,
                                                                  stride, 0, 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_resource_texture **staging)
 {
        struct r600_resource_texture *rtex = (struct r600_resource_texture*)texture;
        struct pipe_resource resource;
+       struct r600_resource_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;
@@ -704,222 +649,26 @@ 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;
-       }
-
-       ((struct r600_resource_texture *)rtex->flushed_depth_texture)->is_flushing_texture = TRUE;
-out:
-       if (just_create)
-               return 0;
-
-       /* XXX: only do this if the depth texture has actually changed:
-        */
-       r600_blit_uncompress_depth(ctx, rtex);
-       return 0;
-}
-
-/* Needs adjustment for pixelformat:
- */
-static INLINE unsigned u_box_volume( const struct pipe_box *box )
-{
-       return box->width * box->depth * box->height;
-};
-
-struct pipe_transfer* r600_texture_get_transfer(struct pipe_context *ctx,
-                                               struct pipe_resource *texture,
-                                               unsigned level,
-                                               unsigned usage,
-                                               const struct pipe_box *box)
-{
-       struct r600_resource_texture *rtex = (struct r600_resource_texture*)texture;
-       struct pipe_resource resource;
-       struct r600_transfer *trans;
-       int r;
-       boolean use_staging_texture = FALSE;
-
-       /* We cannot map a tiled texture directly because the data is
-        * in a different order, therefore we do detiling using a blit.
-        *
-        * Also, use a temporary in GTT memory for read transfers, as
-        * 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;
+       resource.usage = staging ? PIPE_USAGE_DYNAMIC : PIPE_USAGE_DEFAULT;
+       resource.bind = texture->bind & ~PIPE_BIND_DEPTH_STENCIL;
+       resource.flags = texture->flags | R600_RESOURCE_FLAG_FLUSHED_DEPTH;
 
-       if ((usage & PIPE_TRANSFER_READ) && u_box_volume(box) > 1024)
-               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_staging_texture = TRUE;
-
-       if (!permit_hardware_blit(ctx->screen, texture) ||
-               (texture->flags & R600_RESOURCE_FLAG_TRANSFER))
-               use_staging_texture = FALSE;
-
-       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.level = level;
-       trans->transfer.usage = usage;
-       trans->transfer.box = *box;
-       if (rtex->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) {
-                       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;
-       } else if (use_staging_texture) {
-               resource.target = PIPE_TEXTURE_2D;
-               resource.format = texture->format;
-               resource.width0 = box->width;
-               resource.height0 = box->height;
-               resource.depth0 = 1;
-               resource.array_size = 1;
-               resource.last_level = 0;
-               resource.nr_samples = 0;
-               resource.usage = PIPE_USAGE_STAGING;
-               resource.bind = 0;
-               resource.flags = R600_RESOURCE_FLAG_TRANSFER;
-               /* For texture reading, the temporary (detiled) texture is used as
-                * a render target when blitting from a tiled texture. */
-               if (usage & PIPE_TRANSFER_READ) {
-                       resource.bind |= PIPE_BIND_RENDER_TARGET;
-               }
-               /* For texture writing, the temporary texture is used as a sampler
-                * when blitting into a tiled texture. */
-               if (usage & PIPE_TRANSFER_WRITE) {
-                       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) {
-                       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];
-               if (usage & PIPE_TRANSFER_READ) {
-                       r600_copy_to_staging_texture(ctx, trans);
-                       /* Always referenced in the blit. */
-                       radeonsi_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);
-       }
-
-       pipe_resource_reference(&transfer->resource, NULL);
-       FREE(transfer);
-}
-
-void* r600_texture_transfer_map(struct pipe_context *ctx,
-                               struct pipe_transfer* transfer)
-{
-       struct r600_context *rctx = (struct r600_context *)ctx;
-       struct r600_transfer *rtransfer = (struct r600_transfer*)transfer;
-       struct radeon_winsys_cs_handle *buf;
-       enum pipe_format format = transfer->resource->format;
-       unsigned offset = 0;
-       char *map;
-
-       if (rtransfer->staging_texture) {
-               buf = si_resource(rtransfer->staging_texture)->cs_buf;
-       } else {
-               struct r600_resource_texture *rtex = (struct r600_resource_texture*)transfer->resource;
-
-               if (rtex->flushed_depth_texture)
-                       buf = si_resource(rtex->flushed_depth_texture)->cs_buf;
-               else
-                       buf = si_resource(transfer->resource)->cs_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);
-       }
+       if (staging)
+               resource.flags |= R600_RESOURCE_FLAG_TRANSFER;
+       else
+               rtex->dirty_db_mask = (1 << (resource.last_level+1)) - 1;
 
-       if (!(map = rctx->ws->buffer_map(buf, rctx->cs, transfer->usage))) {
-               return NULL;
+       *flushed_depth_texture = (struct r600_resource_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;
        }
 
-       return map + offset;
-}
-
-void r600_texture_transfer_unmap(struct pipe_context *ctx,
-                                struct pipe_transfer* transfer)
-{
-       struct r600_transfer *rtransfer = (struct r600_transfer*)transfer;
-       struct r600_context *rctx = (struct r600_context*)ctx;
-       struct radeon_winsys_cs_handle *buf;
-
-       if (rtransfer->staging_texture) {
-               buf = si_resource(rtransfer->staging_texture)->cs_buf;
-       } else {
-               struct r600_resource_texture *rtex = (struct r600_resource_texture*)transfer->resource;
-
-               if (rtex->flushed_depth_texture) {
-                       buf = si_resource(rtex->flushed_depth_texture)->cs_buf;
-               } else {
-                       buf = si_resource(transfer->resource)->cs_buf;
-               }
-       }
-       rctx->ws->buffer_unmap(buf);
+       (*flushed_depth_texture)->is_flushing_texture = TRUE;
+       return true;
 }
 
-void r600_init_surface_functions(struct r600_context *r600)
+void si_init_surface_functions(struct r600_context *r600)
 {
        r600->context.create_surface = r600_create_surface;
        r600->context.surface_destroy = r600_surface_destroy;