radeon/winsys: add offset support for BO import/export
[mesa.git] / src / gallium / drivers / radeon / r600_texture.c
index 81f32c0f4f3410cc53f084f46d5c5542991374fe..15818aaae6f519867cfdf9f64b63d24636b30334 100644 (file)
@@ -28,6 +28,7 @@
 #include "r600_cs.h"
 #include "util/u_format.h"
 #include "util/u_memory.h"
+#include "util/u_pack_color.h"
 #include <errno.h>
 #include <inttypes.h>
 
@@ -79,12 +80,8 @@ static void r600_copy_to_staging_texture(struct pipe_context *ctx, struct r600_t
                return;
        }
 
-       if (!rctx->dma_copy(ctx, dst, 0, 0, 0, 0,
-                             src, transfer->level,
-                             &transfer->box)) {
-               ctx->resource_copy_region(ctx, dst, 0, 0, 0, 0,
-                                         src, transfer->level, &transfer->box);
-       }
+       rctx->dma_copy(ctx, dst, 0, 0, 0, 0, src, transfer->level,
+                      &transfer->box);
 }
 
 /* Copy from a transfer's staging texture to a full GPU one. */
@@ -105,13 +102,9 @@ static void r600_copy_from_staging_texture(struct pipe_context *ctx, struct r600
                return;
        }
 
-       if (!rctx->dma_copy(ctx, dst, transfer->level,
-                             transfer->box.x, transfer->box.y, transfer->box.z,
-                             src, 0, &sbox)) {
-               ctx->resource_copy_region(ctx, dst, transfer->level,
-                                         transfer->box.x, transfer->box.y, transfer->box.z,
-                                         src, 0, &sbox);
-       }
+       rctx->dma_copy(ctx, dst, transfer->level,
+                      transfer->box.x, transfer->box.y, transfer->box.z,
+                      src, 0, &sbox);
 }
 
 static unsigned r600_texture_get_offset(struct r600_texture *rtex, unsigned level,
@@ -126,7 +119,7 @@ static unsigned r600_texture_get_offset(struct r600_texture *rtex, unsigned leve
 }
 
 static int r600_init_surface(struct r600_common_screen *rscreen,
-                            struct radeon_surface *surface,
+                            struct radeon_surf *surface,
                             const struct pipe_resource *ptex,
                             unsigned array_mode,
                             bool is_flushed_depth)
@@ -208,9 +201,11 @@ static int r600_init_surface(struct r600_common_screen *rscreen,
 
 static int r600_setup_surface(struct pipe_screen *screen,
                              struct r600_texture *rtex,
-                             unsigned pitch_in_bytes_override)
+                             unsigned pitch_in_bytes_override,
+                             unsigned offset)
 {
        struct r600_common_screen *rscreen = (struct r600_common_screen*)screen;
+       unsigned i;
        int r;
 
        r = rscreen->ws->surface_init(rscreen->ws, &rtex->surface);
@@ -232,33 +227,154 @@ static int r600_setup_surface(struct pipe_screen *screen,
                        rtex->surface.stencil_level[0].offset = rtex->surface.level[0].slice_size;
                }
        }
+
+       if (offset) {
+               for (i = 0; i < Elements(rtex->surface.level); ++i)
+                       rtex->surface.level[i].offset += offset;
+       }
        return 0;
 }
 
+static void r600_texture_init_metadata(struct r600_texture *rtex,
+                                      struct radeon_bo_metadata *metadata)
+{
+       struct radeon_surf *surface = &rtex->surface;
+
+       memset(metadata, 0, sizeof(*metadata));
+       metadata->microtile = surface->level[0].mode >= RADEON_SURF_MODE_1D ?
+                                  RADEON_LAYOUT_TILED : RADEON_LAYOUT_LINEAR;
+       metadata->macrotile = surface->level[0].mode >= RADEON_SURF_MODE_2D ?
+                                  RADEON_LAYOUT_TILED : RADEON_LAYOUT_LINEAR;
+       metadata->pipe_config = surface->pipe_config;
+       metadata->bankw = surface->bankw;
+       metadata->bankh = surface->bankh;
+       metadata->tile_split = surface->tile_split;
+       metadata->stencil_tile_split = surface->stencil_tile_split;
+       metadata->mtilea = surface->mtilea;
+       metadata->num_banks = surface->num_banks;
+       metadata->stride = surface->level[0].pitch_bytes;
+       metadata->scanout = (surface->flags & RADEON_SURF_SCANOUT) != 0;
+}
+
+static void r600_dirty_all_framebuffer_states(struct r600_common_screen *rscreen)
+{
+       p_atomic_inc(&rscreen->dirty_fb_counter);
+}
+
+static void r600_eliminate_fast_color_clear(struct r600_common_screen *rscreen,
+                                     struct r600_texture *rtex)
+{
+       struct pipe_context *ctx = rscreen->aux_context;
+
+       pipe_mutex_lock(rscreen->aux_context_lock);
+       ctx->flush_resource(ctx, &rtex->resource.b.b);
+       ctx->flush(ctx, NULL, 0);
+       pipe_mutex_unlock(rscreen->aux_context_lock);
+}
+
+static void r600_texture_disable_cmask(struct r600_common_screen *rscreen,
+                                      struct r600_texture *rtex)
+{
+       if (!rtex->cmask.size)
+               return;
+
+       assert(rtex->resource.b.b.nr_samples <= 1);
+
+       /* Disable CMASK. */
+       memset(&rtex->cmask, 0, sizeof(rtex->cmask));
+       rtex->cmask.base_address_reg = rtex->resource.gpu_address >> 8;
+
+       if (rscreen->chip_class >= SI)
+               rtex->cb_color_info &= ~SI_S_028C70_FAST_CLEAR(1);
+       else
+               rtex->cb_color_info &= ~EG_S_028C70_FAST_CLEAR(1);
+
+       if (rtex->cmask_buffer != &rtex->resource)
+           pipe_resource_reference((struct pipe_resource**)&rtex->cmask_buffer, NULL);
+
+       /* Notify all contexts about the change. */
+       r600_dirty_all_framebuffer_states(rscreen);
+       p_atomic_inc(&rscreen->compressed_colortex_counter);
+}
+
+static void r600_texture_disable_dcc(struct r600_common_screen *rscreen,
+                                    struct r600_texture *rtex)
+{
+       struct r600_common_context *rctx =
+               (struct r600_common_context *)rscreen->aux_context;
+
+       if (!rtex->dcc_offset)
+               return;
+
+       /* Decompress DCC. */
+       pipe_mutex_lock(rscreen->aux_context_lock);
+       rctx->decompress_dcc(&rctx->b, rtex);
+       rctx->b.flush(&rctx->b, NULL, 0);
+       pipe_mutex_unlock(rscreen->aux_context_lock);
+
+       /* Disable DCC. */
+       rtex->dcc_offset = 0;
+       rtex->cb_color_info &= ~VI_S_028C70_DCC_ENABLE(1);
+
+       /* Notify all contexts about the change. */
+       r600_dirty_all_framebuffer_states(rscreen);
+}
+
 static boolean r600_texture_get_handle(struct pipe_screen* screen,
-                                      struct pipe_resource *ptex,
-                                      struct winsys_handle *whandle)
+                                      struct pipe_resource *resource,
+                                      struct winsys_handle *whandle,
+                                       unsigned usage)
 {
-       struct r600_texture *rtex = (struct r600_texture*)ptex;
-       struct r600_resource *resource = &rtex->resource;
-       struct radeon_surface *surface = &rtex->surface;
        struct r600_common_screen *rscreen = (struct r600_common_screen*)screen;
+       struct r600_resource *res = (struct r600_resource*)resource;
+       struct r600_texture *rtex = (struct r600_texture*)resource;
+       struct radeon_bo_metadata metadata;
+
+       /* This is not supported now, but it might be required for OpenCL
+        * interop in the future.
+        */
+       if (resource->target != PIPE_BUFFER &&
+           (resource->nr_samples > 1 || rtex->is_depth))
+               return NULL;
 
-       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,
-                                      surface->level[0].pitch_bytes,
-                                      (surface->flags & RADEON_SURF_SCANOUT) != 0);
-
-       return rscreen->ws->buffer_get_handle(resource->buf,
-                                               surface->level[0].pitch_bytes, whandle);
+       if (!res->is_shared) {
+               res->is_shared = true;
+               res->external_usage = usage;
+
+               if (resource->target != PIPE_BUFFER) {
+                       /* Since shader image stores don't support DCC on VI,
+                        * disable it for external clients that want write
+                        * access.
+                        */
+                       if (usage & PIPE_HANDLE_USAGE_WRITE)
+                               r600_texture_disable_dcc(rscreen, rtex);
+
+                       if (!(usage & PIPE_HANDLE_USAGE_EXPLICIT_FLUSH)) {
+                               /* Eliminate fast clear (both CMASK and DCC) */
+                               r600_eliminate_fast_color_clear(rscreen, rtex);
+
+                               /* Disable CMASK if flush_resource isn't going
+                                * to be called.
+                                */
+                               r600_texture_disable_cmask(rscreen, rtex);
+                       }
+
+                       /* Set metadata. */
+                       r600_texture_init_metadata(rtex, &metadata);
+                       if (rscreen->query_opaque_metadata)
+                               rscreen->query_opaque_metadata(rscreen, rtex,
+                                                              &metadata);
+
+                       rscreen->ws->buffer_set_metadata(res->buf, &metadata);
+               }
+       } else {
+               assert(res->external_usage == usage);
+       }
+
+       return rscreen->ws->buffer_get_handle(res->buf,
+                                             rtex->surface.level[0].pitch_bytes,
+                                             rtex->surface.level[0].offset,
+                                             whandle);
 }
 
 static void r600_texture_destroy(struct pipe_screen *screen,
@@ -287,7 +403,7 @@ void r600_texture_get_fmask_info(struct r600_common_screen *rscreen,
                                 struct r600_fmask_info *out)
 {
        /* FMASK is allocated like an ordinary texture. */
-       struct radeon_surface fmask = rtex->surface;
+       struct radeon_surf fmask = rtex->surface;
 
        memset(out, 0, sizeof(*out));
 
@@ -296,6 +412,12 @@ void r600_texture_get_fmask_info(struct r600_common_screen *rscreen,
        fmask.nsamples = 1;
        fmask.flags |= RADEON_SURF_FMASK;
 
+       /* Force 2D tiling if it wasn't set. This may occur when creating
+        * FMASK for MSAA resolve on R6xx. On R6xx, the single-sample
+        * destination buffer must have an FMASK too. */
+       fmask.flags = RADEON_SURF_CLR(fmask.flags, MODE);
+       fmask.flags |= RADEON_SURF_SET(RADEON_SURF_MODE_2D, MODE);
+
        if (rscreen->chip_class >= SI) {
                fmask.flags |= RADEON_SURF_HAS_TILE_MODE_INDEX;
        }
@@ -335,7 +457,7 @@ void r600_texture_get_fmask_info(struct r600_common_screen *rscreen,
                out->slice_tile_max -= 1;
 
        out->tile_mode_index = fmask.tiling_index[0];
-       out->pitch = fmask.level[0].nblk_x;
+       out->pitch_in_pixels = fmask.level[0].nblk_x;
        out->bank_height = fmask.bankh;
        out->alignment = MAX2(256, fmask.bo_alignment);
        out->size = fmask.bo_size;
@@ -360,8 +482,8 @@ void r600_texture_get_cmask_info(struct r600_common_screen *rscreen,
        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 num_pipes = rscreen->info.num_tile_pipes;
+       unsigned pipe_interleave_bytes = rscreen->info.pipe_interleave_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;
@@ -379,17 +501,22 @@ void r600_texture_get_cmask_info(struct r600_common_screen *rscreen,
        assert(macro_tile_width % 128 == 0);
        assert(macro_tile_height % 128 == 0);
 
+       out->pitch = pitch_elements;
+       out->height = height;
+       out->xalign = macro_tile_width;
+       out->yalign = macro_tile_height;
        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);
+       out->size = (util_max_layer(&rtex->resource.b.b, 0) + 1) *
+                   align(slice_bytes, base_align);
 }
 
 static void si_texture_get_cmask_info(struct r600_common_screen *rscreen,
                                      struct r600_texture *rtex,
                                      struct r600_cmask_info *out)
 {
-       unsigned pipe_interleave_bytes = rscreen->tiling_info.group_bytes;
-       unsigned num_pipes = rscreen->tiling_info.num_channels;
+       unsigned pipe_interleave_bytes = rscreen->info.pipe_interleave_bytes;
+       unsigned num_pipes = rscreen->info.num_tile_pipes;
        unsigned cl_width, cl_height;
 
        switch (num_pipes) {
@@ -423,12 +550,17 @@ static void si_texture_get_cmask_info(struct r600_common_screen *rscreen,
        /* Each element of CMASK is a nibble. */
        unsigned slice_bytes = slice_elements / 2;
 
+       out->pitch = width;
+       out->height = height;
+       out->xalign = cl_width * 8;
+       out->yalign = cl_height * 8;
        out->slice_tile_max = (width * height) / (128*128);
        if (out->slice_tile_max)
                out->slice_tile_max -= 1;
 
        out->alignment = MAX2(256, base_align);
-       out->size = rtex->surface.array_size * align(slice_bytes, base_align);
+       out->size = (util_max_layer(&rtex->resource.b.b, 0) + 1) *
+                   align(slice_bytes, base_align);
 }
 
 static void r600_texture_allocate_cmask(struct r600_common_screen *rscreen,
@@ -442,36 +574,78 @@ static void r600_texture_allocate_cmask(struct r600_common_screen *rscreen,
 
        rtex->cmask.offset = align(rtex->size, rtex->cmask.alignment);
        rtex->size = rtex->cmask.offset + rtex->cmask.size;
+
+       if (rscreen->chip_class >= SI)
+               rtex->cb_color_info |= SI_S_028C70_FAST_CLEAR(1);
+       else
+               rtex->cb_color_info |= EG_S_028C70_FAST_CLEAR(1);
 }
 
-void r600_texture_init_cmask(struct r600_common_screen *rscreen,
-                            struct r600_texture *rtex)
+static void r600_texture_alloc_cmask_separate(struct r600_common_screen *rscreen,
+                                             struct r600_texture *rtex)
 {
+       if (rtex->cmask_buffer)
+                return;
+
        assert(rtex->cmask.size == 0);
 
-       r600_texture_get_cmask_info(rscreen, rtex, &rtex->cmask);
+       if (rscreen->chip_class >= SI) {
+               si_texture_get_cmask_info(rscreen, rtex, &rtex->cmask);
+       } else {
+               r600_texture_get_cmask_info(rscreen, rtex, &rtex->cmask);
+       }
 
        rtex->cmask_buffer = (struct r600_resource *)
                pipe_buffer_create(&rscreen->b, PIPE_BIND_CUSTOM,
                                   PIPE_USAGE_DEFAULT, rtex->cmask.size);
        if (rtex->cmask_buffer == NULL) {
                rtex->cmask.size = 0;
+               return;
        }
+
+       /* update colorbuffer state bits */
+       rtex->cmask.base_address_reg = rtex->cmask_buffer->gpu_address >> 8;
+
+       if (rscreen->chip_class >= SI)
+               rtex->cb_color_info |= SI_S_028C70_FAST_CLEAR(1);
+       else
+               rtex->cb_color_info |= EG_S_028C70_FAST_CLEAR(1);
+
+       p_atomic_inc(&rscreen->compressed_colortex_counter);
 }
 
-static unsigned si_texture_htile_alloc_size(struct r600_common_screen *rscreen,
+static unsigned r600_texture_get_htile_size(struct r600_common_screen *rscreen,
                                            struct r600_texture *rtex)
 {
        unsigned cl_width, cl_height, width, height;
        unsigned slice_elements, slice_bytes, pipe_interleave_bytes, base_align;
-       unsigned num_pipes = rscreen->tiling_info.num_channels;
+       unsigned num_pipes = rscreen->info.num_tile_pipes;
 
-       /* HTILE doesn't work with 1D tiling (there's massive corruption
-        * in glxgears). */
-       if (rtex->surface.level[0].mode != RADEON_SURF_MODE_2D)
+       if (rscreen->chip_class <= EVERGREEN &&
+           rscreen->info.drm_major == 2 && rscreen->info.drm_minor < 26)
                return 0;
 
+       /* HW bug on R6xx. */
+       if (rscreen->chip_class == R600 &&
+           (rtex->surface.level[0].npix_x > 7680 ||
+            rtex->surface.level[0].npix_y > 7680))
+               return 0;
+
+       /* HTILE is broken with 1D tiling on old kernels and CIK. */
+       if (rscreen->chip_class >= CIK &&
+           rtex->surface.level[0].mode == RADEON_SURF_MODE_1D &&
+           rscreen->info.drm_major == 2 && rscreen->info.drm_minor < 38)
+               return 0;
+
+       /* Overalign HTILE on Stoney to fix piglit/depthstencil-render-miplevels 585. */
+       if (rscreen->family == CHIP_STONEY)
+               num_pipes = 4;
+
        switch (num_pipes) {
+       case 1:
+               cl_width = 32;
+               cl_height = 16;
+               break;
        case 2:
                cl_width = 32;
                cl_height = 32;
@@ -499,51 +673,26 @@ static unsigned si_texture_htile_alloc_size(struct r600_common_screen *rscreen,
        slice_elements = (width * height) / (8 * 8);
        slice_bytes = slice_elements * 4;
 
-       pipe_interleave_bytes = rscreen->tiling_info.group_bytes;
+       pipe_interleave_bytes = rscreen->info.pipe_interleave_bytes;
        base_align = num_pipes * pipe_interleave_bytes;
 
-       return rtex->surface.array_size * align(slice_bytes, base_align);
-}
+       rtex->htile.pitch = width;
+       rtex->htile.height = height;
+       rtex->htile.xalign = cl_width * 8;
+       rtex->htile.yalign = cl_height * 8;
 
-static unsigned r600_texture_htile_alloc_size(struct r600_common_screen *rscreen,
-                                             struct r600_texture *rtex)
-{
-       unsigned sw = rtex->surface.level[0].nblk_x * rtex->surface.blk_w;
-       unsigned sh = rtex->surface.level[0].nblk_y * rtex->surface.blk_h;
-       unsigned npipes = rscreen->info.r600_num_tile_pipes;
-       unsigned htile_size;
-
-       /* XXX also use it for other texture targets */
-       if (rscreen->info.drm_minor < 26 ||
-           rtex->resource.b.b.target != PIPE_TEXTURE_2D ||
-           rtex->surface.level[0].nblk_x < 32 ||
-           rtex->surface.level[0].nblk_y < 32) {
-               return 0;
-       }
-
-       /* this alignment and htile size only apply to linear htile buffer */
-       sw = align(sw, 16 << 3);
-       sh = align(sh, npipes << 3);
-       htile_size = (sw >> 3) * (sh >> 3) * 4;
-       /* must be aligned with 2K * npipes */
-       htile_size = align(htile_size, (2 << 10) * npipes);
-       return htile_size;
+       return (util_max_layer(&rtex->resource.b.b, 0) + 1) *
+               align(slice_bytes, base_align);
 }
 
 static void r600_texture_allocate_htile(struct r600_common_screen *rscreen,
                                        struct r600_texture *rtex)
 {
-       unsigned htile_size;
-       if (rscreen->chip_class >= SI) {
-               htile_size = si_texture_htile_alloc_size(rscreen, rtex);
-       } else {
-               htile_size = r600_texture_htile_alloc_size(rscreen, rtex);
-       }
+       unsigned htile_size = r600_texture_get_htile_size(rscreen, rtex);
 
        if (!htile_size)
                return;
 
-       /* XXX don't allocate it separately */
        rtex->htile_buffer = (struct r600_resource*)
                             pipe_buffer_create(&rscreen->b, PIPE_BIND_CUSTOM,
                                                PIPE_USAGE_DEFAULT, htile_size);
@@ -552,7 +701,96 @@ static void r600_texture_allocate_htile(struct r600_common_screen *rscreen,
                 * without htile buffer */
                R600_ERR("Failed to create buffer object for htile buffer.\n");
        } else {
-               r600_screen_clear_buffer(rscreen, &rtex->htile_buffer->b.b, 0, htile_size, 0);
+               r600_screen_clear_buffer(rscreen, &rtex->htile_buffer->b.b, 0,
+                                        htile_size, 0, true);
+       }
+}
+
+void r600_print_texture_info(struct r600_texture *rtex, FILE *f)
+{
+       int i;
+
+       fprintf(f, "  Info: 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=0x%x, %s\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, util_format_short_name(rtex->resource.b.b.format));
+
+       fprintf(f, "  Layout: size=%"PRIu64", alignment=%"PRIu64", bankw=%u, "
+               "bankh=%u, nbanks=%u, mtilea=%u, tilesplit=%u, pipeconfig=%u, scanout=%u\n",
+               rtex->surface.bo_size, rtex->surface.bo_alignment, rtex->surface.bankw,
+               rtex->surface.bankh, rtex->surface.num_banks, rtex->surface.mtilea,
+               rtex->surface.tile_split, rtex->surface.pipe_config,
+               (rtex->surface.flags & RADEON_SURF_SCANOUT) != 0);
+
+       if (rtex->fmask.size)
+               fprintf(f, "  FMask: offset=%u, size=%u, alignment=%u, pitch_in_pixels=%u, "
+                       "bankh=%u, slice_tile_max=%u, tile_mode_index=%u\n",
+                       rtex->fmask.offset, rtex->fmask.size, rtex->fmask.alignment,
+                       rtex->fmask.pitch_in_pixels, rtex->fmask.bank_height,
+                       rtex->fmask.slice_tile_max, rtex->fmask.tile_mode_index);
+
+       if (rtex->cmask.size)
+               fprintf(f, "  CMask: offset=%u, size=%u, alignment=%u, pitch=%u, "
+                       "height=%u, xalign=%u, yalign=%u, slice_tile_max=%u\n",
+                       rtex->cmask.offset, rtex->cmask.size, rtex->cmask.alignment,
+                       rtex->cmask.pitch, rtex->cmask.height, rtex->cmask.xalign,
+                       rtex->cmask.yalign, rtex->cmask.slice_tile_max);
+
+       if (rtex->htile_buffer)
+               fprintf(f, "  HTile: size=%u, alignment=%u, pitch=%u, height=%u, "
+                       "xalign=%u, yalign=%u\n",
+                       rtex->htile_buffer->b.b.width0,
+                       rtex->htile_buffer->buf->alignment, rtex->htile.pitch,
+                       rtex->htile.height, rtex->htile.xalign, rtex->htile.yalign);
+
+       if (rtex->dcc_offset) {
+               fprintf(f, "  DCC: offset=%u, size=%"PRIu64", alignment=%"PRIu64"\n",
+                       rtex->dcc_offset, rtex->surface.dcc_size,
+                       rtex->surface.dcc_alignment);
+               for (i = 0; i <= rtex->surface.last_level; i++)
+                       fprintf(f, "  DCCLevel[%i]: offset=%"PRIu64"\n",
+                               i, rtex->surface.level[i].dcc_offset);
+       }
+
+       for (i = 0; i <= rtex->surface.last_level; i++)
+               fprintf(f, "  Level[%i]: offset=%"PRIu64", slice_size=%"PRIu64", "
+                       "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,
+                       u_minify(rtex->resource.b.b.width0, i),
+                       u_minify(rtex->resource.b.b.height0, i),
+                       u_minify(rtex->resource.b.b.depth0, i),
+                       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 (i = 0; i <= rtex->surface.last_level; i++) {
+                       fprintf(f, "  StencilLayout: tilesplit=%u\n",
+                               rtex->surface.stencil_tile_split);
+                       fprintf(f, "  StencilLevel[%i]: offset=%"PRIu64", "
+                               "slice_size=%"PRIu64", 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,
+                               u_minify(rtex->resource.b.b.width0, i),
+                               u_minify(rtex->resource.b.b.height0, i),
+                               u_minify(rtex->resource.b.b.depth0, i),
+                               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);
+               }
        }
 }
 
@@ -561,15 +799,16 @@ static struct r600_texture *
 r600_texture_create_object(struct pipe_screen *screen,
                           const struct pipe_resource *base,
                           unsigned pitch_in_bytes_override,
+                          unsigned offset,
                           struct pb_buffer *buf,
-                          struct radeon_surface *surface)
+                          struct radeon_surf *surface)
 {
        struct r600_texture *rtex;
        struct r600_resource *resource;
        struct r600_common_screen *rscreen = (struct r600_common_screen*)screen;
 
        rtex = CALLOC_STRUCT(r600_texture);
-       if (rtex == NULL)
+       if (!rtex)
                return NULL;
 
        resource = &rtex->resource;
@@ -577,13 +816,12 @@ r600_texture_create_object(struct pipe_screen *screen,
        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;
 
        /* 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));
 
        rtex->surface = *surface;
-       if (r600_setup_surface(screen, rtex, pitch_in_bytes_override)) {
+       if (r600_setup_surface(screen, rtex, pitch_in_bytes_override, offset)) {
                FREE(rtex);
                return NULL;
        }
@@ -596,7 +834,7 @@ r600_texture_create_object(struct pipe_screen *screen,
        if (rtex->is_depth) {
                if (!(base->flags & (R600_RESOURCE_FLAG_TRANSFER |
                                     R600_RESOURCE_FLAG_FLUSHED_DEPTH)) &&
-                   (rscreen->debug_flags & DBG_HYPERZ)) {
+                   !(rscreen->debug_flags & DBG_NO_HYPERZ)) {
 
                        r600_texture_allocate_htile(rscreen, rtex);
                }
@@ -612,79 +850,59 @@ r600_texture_create_object(struct pipe_screen *screen,
                                return NULL;
                        }
                }
+
+               if (!buf && rtex->surface.dcc_size &&
+                   !(rscreen->debug_flags & DBG_NO_DCC)) {
+                       /* Reserve space for the DCC buffer. */
+                       rtex->dcc_offset = align(rtex->size, rtex->surface.dcc_alignment);
+                       rtex->size = rtex->dcc_offset + rtex->surface.dcc_size;
+                       rtex->cb_color_info |= VI_S_028C70_DCC_ENABLE(1);
+               }
        }
 
        /* Now create the backing buffer. */
        if (!buf) {
                if (!r600_init_resource(rscreen, resource, rtex->size,
-                                       rtex->surface.bo_alignment, FALSE)) {
+                                       rtex->surface.bo_alignment, TRUE)) {
                        FREE(rtex);
                        return NULL;
                }
        } else {
                resource->buf = buf;
-               resource->cs_buf = rscreen->ws->buffer_get_cs_handle(buf);
-               resource->domains = RADEON_DOMAIN_GTT | RADEON_DOMAIN_VRAM;
+               resource->gpu_address = rscreen->ws->buffer_get_virtual_address(resource->buf);
+               resource->domains = rscreen->ws->buffer_get_initial_domain(resource->buf);
        }
 
        if (rtex->cmask.size) {
                /* Initialize the cmask to 0xCC (= compressed state). */
                r600_screen_clear_buffer(rscreen, &rtex->cmask_buffer->b.b,
-                                        rtex->cmask.offset, rtex->cmask.size, 0xCCCCCCCC);
+                                        rtex->cmask.offset, rtex->cmask.size,
+                                        0xCCCCCCCC, true);
        }
+       if (rtex->dcc_offset) {
+               r600_screen_clear_buffer(rscreen, &rtex->resource.b.b,
+                                        rtex->dcc_offset,
+                                        rtex->surface.dcc_size,
+                                        0xFFFFFFFF, true);
+       }
+
+       /* Initialize the CMASK base register value. */
+       rtex->cmask.base_address_reg =
+               (rtex->resource.gpu_address + rtex->cmask.offset) >> 8;
 
        if (rscreen->debug_flags & DBG_VM) {
-               fprintf(stderr, "VM start=0x%"PRIu64"  end=0x%"PRIu64" | Texture %ix%ix%i, %i levels, %i samples, %s\n",
-                       r600_resource_va(screen, &rtex->resource.b.b),
-                       r600_resource_va(screen, &rtex->resource.b.b) + rtex->resource.buf->size,
+               fprintf(stderr, "VM start=0x%"PRIX64"  end=0x%"PRIX64" | Texture %ix%ix%i, %i levels, %i samples, %s\n",
+                       rtex->resource.gpu_address,
+                       rtex->resource.gpu_address + rtex->resource.buf->size,
                        base->width0, base->height0, util_max_layer(base, 0)+1, base->last_level+1,
                        base->nr_samples ? base->nr_samples : 1, util_format_short_name(base->format));
        }
 
-       if (rscreen->debug_flags & DBG_TEX ||
-           (rtex->resource.b.b.last_level > 0 && rscreen->debug_flags & DBG_TEXMIP)) {
-               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=0x%x, %s\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, util_format_short_name(base->format));
-               for (int i = 0; i <= rtex->surface.last_level; i++) {
-                       printf("  L %i: offset=%"PRIu64", slice_size=%"PRIu64", 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,
-                              u_minify(rtex->resource.b.b.width0, i),
-                              u_minify(rtex->resource.b.b.height0, i),
-                              u_minify(rtex->resource.b.b.depth0, i),
-                              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=%"PRIu64", slice_size=%"PRIu64", 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,
-                                      u_minify(rtex->resource.b.b.width0, i),
-                                      u_minify(rtex->resource.b.b.height0, i),
-                                      u_minify(rtex->resource.b.b.depth0, i),
-                                      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);
-                       }
-               }
+       if (rscreen->debug_flags & DBG_TEX) {
+               puts("Texture:");
+               r600_print_texture_info(rtex, stdout);
        }
+
        return rtex;
 }
 
@@ -692,6 +910,7 @@ static unsigned r600_choose_tiling(struct r600_common_screen *rscreen,
                                   const struct pipe_resource *templ)
 {
        const struct util_format_description *desc = util_format_description(templ->format);
+       bool force_tiling = templ->flags & R600_RESOURCE_FLAG_FORCE_TILING;
 
        /* MSAA resources must be 2D tiled. */
        if (templ->nr_samples > 1)
@@ -701,13 +920,25 @@ static unsigned r600_choose_tiling(struct r600_common_screen *rscreen,
        if (templ->flags & R600_RESOURCE_FLAG_TRANSFER)
                return RADEON_SURF_MODE_LINEAR_ALIGNED;
 
+       /* r600g: force tiling on TEXTURE_2D and TEXTURE_3D compute resources. */
+       if (rscreen->chip_class >= R600 && rscreen->chip_class <= CAYMAN &&
+           (templ->bind & PIPE_BIND_COMPUTE_RESOURCE) &&
+           (templ->target == PIPE_TEXTURE_2D ||
+            templ->target == PIPE_TEXTURE_3D))
+               force_tiling = true;
+
        /* Handle common candidates for the linear mode.
         * Compressed textures must always be tiled. */
-       if (!(templ->flags & R600_RESOURCE_FLAG_FORCE_TILING) &&
-           !util_format_is_compressed(templ->format)) {
-               /* Tiling doesn't work with the 422 (SUBSAMPLED) formats on R600-Cayman. */
-               if (rscreen->chip_class <= CAYMAN &&
-                   desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED)
+       if (!force_tiling && !util_format_is_compressed(templ->format)) {
+               /* Not everything can be linear, so we cannot enforce it
+                * for all textures. */
+               if ((rscreen->debug_flags & DBG_NO_TILING) &&
+                   (!util_format_is_depth_or_stencil(templ->format) ||
+                    !(templ->flags & R600_RESOURCE_FLAG_FLUSHED_DEPTH)))
+                       return RADEON_SURF_MODE_LINEAR_ALIGNED;
+
+               /* Tiling doesn't work with the 422 (SUBSAMPLED) formats on R600+. */
+               if (desc->layout == UTIL_FORMAT_LAYOUT_SUBSAMPLED)
                        return RADEON_SURF_MODE_LINEAR_ALIGNED;
 
                /* Cursors are linear on SI.
@@ -732,7 +963,8 @@ static unsigned r600_choose_tiling(struct r600_common_screen *rscreen,
        }
 
        /* Make small textures 1D tiled. */
-       if (templ->width0 <= 16 || templ->height0 <= 16)
+       if (templ->width0 <= 16 || templ->height0 <= 16 ||
+           (rscreen->debug_flags & DBG_NO_2D_TILING))
                return RADEON_SURF_MODE_1D;
 
        /* The allocator will switch to 1D if needed. */
@@ -743,7 +975,7 @@ struct pipe_resource *r600_texture_create(struct pipe_screen *screen,
                                          const struct pipe_resource *templ)
 {
        struct r600_common_screen *rscreen = (struct r600_common_screen*)screen;
-       struct radeon_surface surface = {0};
+       struct radeon_surf surface = {0};
        int r;
 
        r = r600_init_surface(rscreen, &surface, templ,
@@ -756,41 +988,44 @@ struct pipe_resource *r600_texture_create(struct pipe_screen *screen,
        if (r) {
                return NULL;
        }
-       return (struct pipe_resource *)r600_texture_create_object(screen, templ,
+       return (struct pipe_resource *)r600_texture_create_object(screen, templ, 0,
                                                                  0, NULL, &surface);
 }
 
 static struct pipe_resource *r600_texture_from_handle(struct pipe_screen *screen,
                                                      const struct pipe_resource *templ,
-                                                     struct winsys_handle *whandle)
+                                                     struct winsys_handle *whandle,
+                                                      unsigned usage)
 {
        struct r600_common_screen *rscreen = (struct r600_common_screen*)screen;
        struct pb_buffer *buf = NULL;
-       unsigned stride = 0;
+       unsigned stride = 0, offset = 0;
        unsigned array_mode;
-       enum radeon_bo_layout micro, macro;
-       struct radeon_surface surface;
-       bool scanout;
+       struct radeon_surf surface;
        int r;
+       struct radeon_bo_metadata metadata = {};
+       struct r600_texture *rtex;
 
        /* 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);
+       buf = rscreen->ws->buffer_from_handle(rscreen->ws, whandle, &stride, &offset);
        if (!buf)
                return NULL;
 
-       rscreen->ws->buffer_get_tiling(buf, &micro, &macro,
-                                      &surface.bankw, &surface.bankh,
-                                      &surface.tile_split,
-                                      &surface.stencil_tile_split,
-                                      &surface.mtilea, &scanout);
+       rscreen->ws->buffer_get_metadata(buf, &metadata);
 
-       if (macro == RADEON_LAYOUT_TILED)
+       surface.bankw = metadata.bankw;
+       surface.bankh = metadata.bankh;
+       surface.tile_split = metadata.tile_split;
+       surface.stencil_tile_split = metadata.stencil_tile_split;
+       surface.mtilea = metadata.mtilea;
+
+       if (metadata.macrotile == RADEON_LAYOUT_TILED)
                array_mode = RADEON_SURF_MODE_2D;
-       else if (micro == RADEON_LAYOUT_TILED)
+       else if (metadata.microtile == RADEON_LAYOUT_TILED)
                array_mode = RADEON_SURF_MODE_1D;
        else
                array_mode = RADEON_SURF_MODE_LINEAR_ALIGNED;
@@ -800,11 +1035,17 @@ static struct pipe_resource *r600_texture_from_handle(struct pipe_screen *screen
                return NULL;
        }
 
-       if (scanout)
+       if (metadata.scanout)
                surface.flags |= RADEON_SURF_SCANOUT;
 
-       return (struct pipe_resource *)r600_texture_create_object(screen, templ,
-                                                                 stride, buf, &surface);
+       rtex = r600_texture_create_object(screen, templ, stride,
+                                         offset, buf, &surface);
+       if (!rtex)
+               return NULL;
+
+       rtex->resource.is_shared = true;
+       rtex->resource.external_usage = usage;
+       return &rtex->resource.b.b;
 }
 
 bool r600_init_flushed_depth_texture(struct pipe_context *ctx,
@@ -905,19 +1146,16 @@ static void *r600_texture_transfer_map(struct pipe_context *ctx,
         * the CPU is much happier reading out of cached system memory
         * than uncached VRAM.
         */
-       if (rtex->surface.level[level].mode >= RADEON_SURF_MODE_1D)
+       if (rtex->surface.level[0].mode >= RADEON_SURF_MODE_1D) {
                use_staging_texture = TRUE;
-
-       /* Untiled buffers in VRAM, which is slow for CPU reads and writes */
-       if (!(usage & PIPE_TRANSFER_MAP_DIRECTLY) &&
+       } else if ((usage & PIPE_TRANSFER_READ) && !(usage & PIPE_TRANSFER_MAP_DIRECTLY) &&
            (rtex->resource.domains == RADEON_DOMAIN_VRAM)) {
+               /* Untiled buffers in VRAM, which is slow for CPU reads */
                use_staging_texture = TRUE;
-       }
-
-       /* Use a staging texture for uploads if the underlying BO is busy. */
-       if (!(usage & PIPE_TRANSFER_READ) &&
-           (r600_rings_is_buffer_referenced(rctx, rtex->resource.cs_buf, RADEON_USAGE_READWRITE) ||
-            rctx->ws->buffer_is_busy(rtex->resource.buf, RADEON_USAGE_READWRITE))) {
+       } else if (!(usage & PIPE_TRANSFER_READ) &&
+           (r600_rings_is_buffer_referenced(rctx, rtex->resource.buf, RADEON_USAGE_READWRITE) ||
+            !rctx->ws->buffer_wait(rtex->resource.buf, 0, RADEON_USAGE_READWRITE))) {
+               /* Use a staging texture for uploads if the underlying BO is busy. */
                use_staging_texture = TRUE;
        }
 
@@ -930,7 +1168,7 @@ static void *r600_texture_transfer_map(struct pipe_context *ctx,
        }
 
        trans = CALLOC_STRUCT(r600_transfer);
-       if (trans == NULL)
+       if (!trans)
                return NULL;
        trans->transfer.resource = texture;
        trans->transfer.level = level;
@@ -963,11 +1201,16 @@ static void *r600_texture_transfer_map(struct pipe_context *ctx,
 
                        if (usage & PIPE_TRANSFER_READ) {
                                struct pipe_resource *temp = ctx->screen->resource_create(ctx->screen, &resource);
+                               if (!temp) {
+                                       R600_ERR("failed to create a temporary depth texture\n");
+                                       FREE(trans);
+                                       return NULL;
+                               }
 
                                r600_copy_region_with_blit(ctx, temp, 0, 0, 0, 0, texture, level, box);
                                rctx->blit_decompress_depth(ctx, (struct r600_texture*)temp, staging_depth,
                                                            0, 0, 0, box->depth, 0, 0);
-                               pipe_resource_reference((struct pipe_resource**)&temp, NULL);
+                               pipe_resource_reference(&temp, NULL);
                        }
                }
                else {
@@ -996,10 +1239,12 @@ static void *r600_texture_transfer_map(struct pipe_context *ctx,
 
                r600_init_temp_resource_from_box(&resource, texture, box, level,
                                                 R600_RESOURCE_FLAG_TRANSFER);
+               resource.usage = (usage & PIPE_TRANSFER_READ) ?
+                       PIPE_USAGE_STAGING : PIPE_USAGE_STREAM;
 
                /* Create the temporary texture. */
                staging = (struct r600_texture*)ctx->screen->resource_create(ctx->screen, &resource);
-               if (staging == NULL) {
+               if (!staging) {
                        R600_ERR("failed to create temporary texture to hold untiled copy\n");
                        FREE(trans);
                        return NULL;
@@ -1019,6 +1264,8 @@ static void *r600_texture_transfer_map(struct pipe_context *ctx,
 
        if (trans->staging) {
                buf = trans->staging;
+               if (!rtex->is_depth && !(usage & PIPE_TRANSFER_READ))
+                       usage |= PIPE_TRANSFER_UNSYNCHRONIZED;
        } else {
                buf = &rtex->resource;
        }
@@ -1037,18 +1284,9 @@ static void r600_texture_transfer_unmap(struct pipe_context *ctx,
                                        struct pipe_transfer* transfer)
 {
        struct r600_transfer *rtransfer = (struct r600_transfer*)transfer;
-       struct r600_common_context *rctx = (struct r600_common_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) {
-               buf = rtransfer->staging->cs_buf;
-       } else {
-               buf = r600_resource(transfer->resource)->cs_buf;
-       }
-       rctx->ws->buffer_unmap(buf);
-
        if ((transfer->usage & PIPE_TRANSFER_WRITE) && rtransfer->staging) {
                if (rtex->is_depth && rtex->resource.b.b.nr_samples <= 1) {
                        ctx->resource_copy_region(ctx, texture, transfer->level,
@@ -1071,7 +1309,7 @@ static const struct u_resource_vtbl r600_texture_vtbl =
        NULL,                           /* get_handle */
        r600_texture_destroy,           /* resource_destroy */
        r600_texture_transfer_map,      /* transfer_map */
-       NULL,                           /* transfer_flush_region */
+       u_default_transfer_flush_region, /* transfer_flush_region */
        r600_texture_transfer_unmap,    /* transfer_unmap */
        NULL                            /* transfer_inline_write */
 };
@@ -1083,7 +1321,7 @@ struct pipe_surface *r600_create_surface_custom(struct pipe_context *pipe,
 {
        struct r600_surface *surface = CALLOC_STRUCT(r600_surface);
 
-       if (surface == NULL)
+       if (!surface)
                return NULL;
 
        assert(templ->u.tex.first_layer <= util_max_layer(texture, templ->u.tex.level));
@@ -1104,10 +1342,30 @@ static struct pipe_surface *r600_create_surface(struct pipe_context *pipe,
                                                const struct pipe_surface *templ)
 {
        unsigned level = templ->u.tex.level;
+       unsigned width = u_minify(tex->width0, level);
+       unsigned height = u_minify(tex->height0, level);
+
+       if (tex->target != PIPE_BUFFER && templ->format != tex->format) {
+               const struct util_format_description *tex_desc
+                       = util_format_description(tex->format);
+               const struct util_format_description *templ_desc
+                       = util_format_description(templ->format);
+
+               assert(tex_desc->block.bits == templ_desc->block.bits);
+
+               /* Adjust size of surface if and only if the block width or
+                * height is changed. */
+               if (tex_desc->block.width != templ_desc->block.width ||
+                   tex_desc->block.height != templ_desc->block.height) {
+                       unsigned nblks_x = util_format_get_nblocksx(tex->format, width);
+                       unsigned nblks_y = util_format_get_nblocksy(tex->format, height);
+
+                       width = nblks_x * templ_desc->block.width;
+                       height = nblks_y * templ_desc->block.height;
+               }
+       }
 
-       return r600_create_surface_custom(pipe, tex, templ,
-                                         u_minify(tex->width0, level),
-                                         u_minify(tex->height0, level));
+       return r600_create_surface_custom(pipe, tex, templ, width, height);
 }
 
 static void r600_surface_destroy(struct pipe_context *pipe,
@@ -1167,13 +1425,215 @@ unsigned r600_translate_colorswap(enum pipe_format format)
                        return V_0280A0_SWAP_STD_REV; /* WZYX */
                else if (HAS_SWIZZLE(1,Y) && HAS_SWIZZLE(2,X))
                        return V_0280A0_SWAP_ALT; /* ZYXW */
-               else if (HAS_SWIZZLE(1,X) && HAS_SWIZZLE(2,Y))
-                       return V_0280A0_SWAP_ALT_REV; /* WXYZ */
+               else if (HAS_SWIZZLE(1,Z) && HAS_SWIZZLE(2,W))
+                       return V_0280A0_SWAP_ALT_REV; /* YZWX */
                break;
        }
        return ~0U;
 }
 
+static void evergreen_set_clear_color(struct r600_texture *rtex,
+                                     enum pipe_format surface_format,
+                                     const union pipe_color_union *color)
+{
+       union util_color uc;
+
+       memset(&uc, 0, sizeof(uc));
+
+       if (util_format_is_pure_uint(surface_format)) {
+               util_format_write_4ui(surface_format, color->ui, 0, &uc, 0, 0, 0, 1, 1);
+       } else if (util_format_is_pure_sint(surface_format)) {
+               util_format_write_4i(surface_format, color->i, 0, &uc, 0, 0, 0, 1, 1);
+       } else {
+               util_pack_color(color->f, surface_format, &uc);
+       }
+
+       memcpy(rtex->color_clear_value, &uc, 2 * sizeof(uint32_t));
+}
+
+static void vi_get_fast_clear_parameters(enum pipe_format surface_format,
+                                        const union pipe_color_union *color,
+                                        uint32_t* reset_value,
+                                        bool* clear_words_needed)
+{
+       bool values[4] = {};
+       int i;
+       bool main_value = false;
+       bool extra_value = false;
+       int extra_channel;
+       const struct util_format_description *desc = util_format_description(surface_format);
+
+       *clear_words_needed = true;
+       *reset_value = 0x20202020U;
+
+       /* If we want to clear without needing a fast clear eliminate step, we
+        * can set each channel to 0 or 1 (or 0/max for integer formats). We
+        * have two sets of flags, one for the last or first channel(extra) and
+        * one for the other channels(main).
+        */
+
+       if (surface_format == PIPE_FORMAT_R11G11B10_FLOAT ||
+           surface_format == PIPE_FORMAT_B5G6R5_UNORM ||
+           surface_format == PIPE_FORMAT_B5G6R5_SRGB) {
+               extra_channel = -1;
+       } else if (desc->layout == UTIL_FORMAT_LAYOUT_PLAIN) {
+               if(r600_translate_colorswap(surface_format) <= 1)
+                       extra_channel = desc->nr_channels - 1;
+               else
+                       extra_channel = 0;
+       } else
+               return;
+
+       for (i = 0; i < 4; ++i) {
+               int index = desc->swizzle[i] - UTIL_FORMAT_SWIZZLE_X;
+
+               if (desc->swizzle[i] < UTIL_FORMAT_SWIZZLE_X ||
+                   desc->swizzle[i] > UTIL_FORMAT_SWIZZLE_W)
+                       continue;
+
+               if (util_format_is_pure_sint(surface_format)) {
+                       values[i] = color->i[i] != 0;
+                       if (color->i[i] != 0 && color->i[i] != INT32_MAX)
+                               return;
+               } else if (util_format_is_pure_uint(surface_format)) {
+                       values[i] = color->ui[i] != 0U;
+                       if (color->ui[i] != 0U && color->ui[i] != UINT32_MAX)
+                               return;
+               } else {
+                       values[i] = color->f[i] != 0.0F;
+                       if (color->f[i] != 0.0F && color->f[i] != 1.0F)
+                               return;
+               }
+
+               if (index == extra_channel)
+                       extra_value = values[i];
+               else
+                       main_value = values[i];
+       }
+
+       for (int i = 0; i < 4; ++i)
+               if (values[i] != main_value &&
+                   desc->swizzle[i] - UTIL_FORMAT_SWIZZLE_X != extra_channel &&
+                   desc->swizzle[i] >= UTIL_FORMAT_SWIZZLE_X &&
+                   desc->swizzle[i] <= UTIL_FORMAT_SWIZZLE_W)
+                       return;
+
+       *clear_words_needed = false;
+       if (main_value)
+               *reset_value |= 0x80808080U;
+
+       if (extra_value)
+               *reset_value |= 0x40404040U;
+}
+
+void evergreen_do_fast_color_clear(struct r600_common_context *rctx,
+                                  struct pipe_framebuffer_state *fb,
+                                  struct r600_atom *fb_state,
+                                  unsigned *buffers, unsigned *dirty_cbufs,
+                                  const union pipe_color_union *color)
+{
+       int i;
+
+       /* This function is broken in BE, so just disable this path for now */
+#ifdef PIPE_ARCH_BIG_ENDIAN
+       return;
+#endif
+
+       if (rctx->render_cond)
+               return;
+
+       for (i = 0; i < fb->nr_cbufs; i++) {
+               struct r600_texture *tex;
+               unsigned clear_bit = PIPE_CLEAR_COLOR0 << i;
+
+               if (!fb->cbufs[i])
+                       continue;
+
+               /* if this colorbuffer is not being cleared */
+               if (!(*buffers & clear_bit))
+                       continue;
+
+               tex = (struct r600_texture *)fb->cbufs[i]->texture;
+
+               /* 128-bit formats are unusupported */
+               if (util_format_get_blocksizebits(fb->cbufs[i]->format) > 64) {
+                       continue;
+               }
+
+               /* the clear is allowed if all layers are bound */
+               if (fb->cbufs[i]->u.tex.first_layer != 0 ||
+                   fb->cbufs[i]->u.tex.last_layer != util_max_layer(&tex->resource.b.b, 0)) {
+                       continue;
+               }
+
+               /* cannot clear mipmapped textures */
+               if (fb->cbufs[i]->texture->last_level != 0) {
+                       continue;
+               }
+
+               /* only supported on tiled surfaces */
+               if (tex->surface.level[0].mode < RADEON_SURF_MODE_1D) {
+                       continue;
+               }
+
+               /* shared textures can't use fast clear without an explicit flush,
+                * because there is no way to communicate the clear color among
+                * all clients
+                */
+               if (tex->resource.is_shared &&
+                   !(tex->resource.external_usage & PIPE_HANDLE_USAGE_EXPLICIT_FLUSH))
+                       continue;
+
+               /* fast color clear with 1D tiling doesn't work on old kernels and CIK */
+               if (tex->surface.level[0].mode == RADEON_SURF_MODE_1D &&
+                   rctx->chip_class >= CIK &&
+                   rctx->screen->info.drm_major == 2 &&
+                   rctx->screen->info.drm_minor < 38) {
+                       continue;
+               }
+
+               if (tex->dcc_offset) {
+                       uint32_t reset_value;
+                       bool clear_words_needed;
+
+                       if (rctx->screen->debug_flags & DBG_NO_DCC_CLEAR)
+                               continue;
+
+                       vi_get_fast_clear_parameters(fb->cbufs[i]->format, color, &reset_value, &clear_words_needed);
+
+                       rctx->clear_buffer(&rctx->b, &tex->resource.b.b,
+                                          tex->dcc_offset, tex->surface.dcc_size,
+                                          reset_value, true);
+
+                       if (clear_words_needed)
+                               tex->dirty_level_mask |= 1 << fb->cbufs[i]->u.tex.level;
+               } else {
+                       /* Stoney/RB+ doesn't work with CMASK fast clear. */
+                       if (rctx->family == CHIP_STONEY)
+                               continue;
+
+                       /* ensure CMASK is enabled */
+                       r600_texture_alloc_cmask_separate(rctx->screen, tex);
+                       if (tex->cmask.size == 0) {
+                               continue;
+                       }
+
+                       /* Do the fast clear. */
+                       rctx->clear_buffer(&rctx->b, &tex->cmask_buffer->b.b,
+                                       tex->cmask.offset, tex->cmask.size, 0, true);
+
+                       tex->dirty_level_mask |= 1 << fb->cbufs[i]->u.tex.level;
+               }
+
+               evergreen_set_clear_color(tex, fb->cbufs[i]->format, color);
+
+               if (dirty_cbufs)
+                       *dirty_cbufs |= 1 << i;
+               rctx->set_atom_dirty(rctx, fb_state, true);
+               *buffers &= ~clear_bit;
+       }
+}
+
 void r600_init_screen_texture_functions(struct r600_common_screen *rscreen)
 {
        rscreen->b.resource_from_handle = r600_texture_from_handle;