radeonsi: move VS_STATE.LS_OUT_PATCH_SIZE a few bits higher to make space there
[mesa.git] / src / gallium / drivers / radeonsi / si_blit.c
index 1cbd26f46e024da6a0a0efe0e0ae61ab1cc0df3a..5ac68406f7fb467f717a2d322f5de3e69b523866 100644 (file)
@@ -25,7 +25,7 @@
 
 #include "si_pipe.h"
 #include "si_compute.h"
-#include "util/u_format.h"
+#include "util/format/u_format.h"
 #include "util/u_log.h"
 #include "util/u_surface.h"
 
@@ -58,7 +58,11 @@ void si_blitter_begin(struct si_context *sctx, enum si_blitter_op op)
                util_blitter_save_stencil_ref(sctx->blitter, &sctx->stencil_ref.state);
                util_blitter_save_fragment_shader(sctx->blitter, sctx->ps_shader.cso);
                util_blitter_save_sample_mask(sctx->blitter, sctx->sample_mask);
-               util_blitter_save_scissor(sctx->blitter, &sctx->scissors.states[0]);
+               util_blitter_save_scissor(sctx->blitter, &sctx->scissors[0]);
+               util_blitter_save_window_rectangles(sctx->blitter,
+                                                   sctx->window_rectangles_include,
+                                                   sctx->num_window_rectangles,
+                                                   sctx->window_rectangles);
        }
 
        if (op & SI_SAVE_FRAMEBUFFER)
@@ -75,16 +79,27 @@ void si_blitter_begin(struct si_context *sctx, enum si_blitter_op op)
 
        if (op & SI_DISABLE_RENDER_COND)
                sctx->render_cond_force_off = true;
+
+       if (sctx->screen->dpbb_allowed) {
+               sctx->dpbb_force_off = true;
+               si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
+       }
 }
 
 void si_blitter_end(struct si_context *sctx)
 {
+       if (sctx->screen->dpbb_allowed) {
+               sctx->dpbb_force_off = false;
+               si_mark_atom_dirty(sctx, &sctx->atoms.s.dpbb_state);
+       }
+
        sctx->render_cond_force_off = false;
 
        /* Restore shader pointers because the VS blit shader changed all
         * non-global VS user SGPRs. */
        sctx->shader_pointers_dirty |= SI_DESCS_SHADER_MASK(VERTEX);
        sctx->vertex_buffer_pointer_dirty = sctx->vb_descriptors_buffer != NULL;
+       sctx->vertex_buffer_user_sgprs_dirty = sctx->num_vertex_elements > 0;
        si_mark_atom_dirty(sctx, &sctx->atoms.s.shader_pointers);
 }
 
@@ -95,8 +110,8 @@ static unsigned u_max_sample(struct pipe_resource *r)
 
 static unsigned
 si_blit_dbcb_copy(struct si_context *sctx,
-                 struct r600_texture *src,
-                 struct r600_texture *dst,
+                 struct si_texture *src,
+                 struct si_texture *dst,
                  unsigned planes, unsigned level_mask,
                  unsigned first_layer, unsigned last_layer,
                  unsigned first_sample, unsigned last_sample)
@@ -166,36 +181,11 @@ si_blit_dbcb_copy(struct si_context *sctx,
        return fully_copied_levels;
 }
 
-void si_blit_decompress_depth(struct pipe_context *ctx,
-                             struct r600_texture *texture,
-                             struct r600_texture *staging,
-                             unsigned first_level, unsigned last_level,
-                             unsigned first_layer, unsigned last_layer,
-                             unsigned first_sample, unsigned last_sample)
-{
-       const struct util_format_description *desc;
-       unsigned planes = 0;
-
-       assert(staging != NULL && "use si_blit_decompress_zs_in_place instead");
-
-       desc = util_format_description(staging->buffer.b.b.format);
-
-       if (util_format_has_depth(desc))
-               planes |= PIPE_MASK_Z;
-       if (util_format_has_stencil(desc))
-               planes |= PIPE_MASK_S;
-
-       si_blit_dbcb_copy(
-               (struct si_context *)ctx, texture, staging, planes,
-               u_bit_consecutive(first_level, last_level - first_level + 1),
-               first_layer, last_layer, first_sample, last_sample);
-}
-
 /* Helper function for si_blit_decompress_zs_in_place.
  */
 static void
 si_blit_decompress_zs_planes_in_place(struct si_context *sctx,
-                                     struct r600_texture *texture,
+                                     struct si_texture *texture,
                                      unsigned planes, unsigned level_mask,
                                      unsigned first_layer, unsigned last_layer)
 {
@@ -264,7 +254,7 @@ si_blit_decompress_zs_planes_in_place(struct si_context *sctx,
  */
 static void
 si_blit_decompress_zs_in_place(struct si_context *sctx,
-                              struct r600_texture *texture,
+                              struct si_texture *texture,
                               unsigned levels_z, unsigned levels_s,
                               unsigned first_layer, unsigned last_layer)
 {
@@ -298,7 +288,7 @@ si_blit_decompress_zs_in_place(struct si_context *sctx,
 
 static void
 si_decompress_depth(struct si_context *sctx,
-                   struct r600_texture *tex,
+                   struct si_texture *tex,
                    unsigned required_planes,
                    unsigned first_level, unsigned last_level,
                    unsigned first_layer, unsigned last_layer)
@@ -341,8 +331,8 @@ si_decompress_depth(struct si_context *sctx,
         */
        if (copy_planes &&
            (tex->flushed_depth_texture ||
-            si_init_flushed_depth_texture(&sctx->b, &tex->buffer.b.b, NULL))) {
-               struct r600_texture *dst = tex->flushed_depth_texture;
+            si_init_flushed_depth_texture(&sctx->b, &tex->buffer.b.b))) {
+               struct si_texture *dst = tex->flushed_depth_texture;
                unsigned fully_copied_levels;
                unsigned levels = 0;
 
@@ -372,8 +362,9 @@ si_decompress_depth(struct si_context *sctx,
        }
 
        if (inplace_planes) {
-               bool has_htile = si_htile_enabled(tex, first_level);
-               bool tc_compat_htile = vi_tc_compat_htile_enabled(tex, first_level);
+               bool has_htile = si_htile_enabled(tex, first_level, inplace_planes);
+               bool tc_compat_htile = vi_tc_compat_htile_enabled(tex, first_level,
+                                                                 inplace_planes);
 
                /* Don't decompress if there is no HTILE or when HTILE is
                 * TC-compatible. */
@@ -407,7 +398,7 @@ si_decompress_depth(struct si_context *sctx,
         */
        if (copy_planes && tex->buffer.b.b.nr_samples > 1)
                si_make_CB_shader_coherent(sctx, tex->buffer.b.b.nr_samples,
-                                          false);
+                                          false, true /* no DCC */);
 }
 
 static void
@@ -420,7 +411,7 @@ si_decompress_sampler_depth_textures(struct si_context *sctx,
        while (mask) {
                struct pipe_sampler_view *view;
                struct si_sampler_view *sview;
-               struct r600_texture *tex;
+               struct si_texture *tex;
 
                i = u_bit_scan(&mask);
 
@@ -428,7 +419,7 @@ si_decompress_sampler_depth_textures(struct si_context *sctx,
                assert(view);
                sview = (struct si_sampler_view*)view;
 
-               tex = (struct r600_texture *)view->texture;
+               tex = (struct si_texture *)view->texture;
                assert(tex->db_compatible);
 
                si_decompress_depth(sctx, tex,
@@ -439,10 +430,11 @@ si_decompress_sampler_depth_textures(struct si_context *sctx,
 }
 
 static void si_blit_decompress_color(struct si_context *sctx,
-               struct r600_texture *rtex,
-               unsigned first_level, unsigned last_level,
-               unsigned first_layer, unsigned last_layer,
-               bool need_dcc_decompress)
+                                    struct si_texture *tex,
+                                    unsigned first_level, unsigned last_level,
+                                    unsigned first_layer, unsigned last_layer,
+                                    bool need_dcc_decompress,
+                                    bool need_fmask_expand)
 {
        void* custom_blend;
        unsigned layer, checked_last_layer, max_layer;
@@ -450,9 +442,9 @@ static void si_blit_decompress_color(struct si_context *sctx,
                u_bit_consecutive(first_level, last_level - first_level + 1);
 
        if (!need_dcc_decompress)
-               level_mask &= rtex->dirty_level_mask;
+               level_mask &= tex->dirty_level_mask;
        if (!level_mask)
-               return;
+               goto expand_fmask;
 
        if (unlikely(sctx->log))
                u_log_printf(sctx->log,
@@ -463,14 +455,14 @@ static void si_blit_decompress_color(struct si_context *sctx,
        if (need_dcc_decompress) {
                custom_blend = sctx->custom_blend_dcc_decompress;
 
-               assert(rtex->dcc_offset);
+               assert(tex->surface.dcc_offset);
 
                /* disable levels without DCC */
                for (int i = first_level; i <= last_level; i++) {
-                       if (!vi_dcc_enabled(rtex, i))
+                       if (!vi_dcc_enabled(tex, i))
                                level_mask &= ~(1 << i);
                }
-       } else if (rtex->fmask.size) {
+       } else if (tex->surface.fmask_size) {
                custom_blend = sctx->custom_blend_fmask_decompress;
        } else {
                custom_blend = sctx->custom_blend_eliminate_fastclear;
@@ -483,17 +475,17 @@ static void si_blit_decompress_color(struct si_context *sctx,
 
                /* The smaller the mipmap level, the less layers there are
                 * as far as 3D textures are concerned. */
-               max_layer = util_max_layer(&rtex->buffer.b.b, level);
+               max_layer = util_max_layer(&tex->buffer.b.b, level);
                checked_last_layer = MIN2(last_layer, max_layer);
 
                for (layer = first_layer; layer <= checked_last_layer; layer++) {
                        struct pipe_surface *cbsurf, surf_tmpl;
 
-                       surf_tmpl.format = rtex->buffer.b.b.format;
+                       surf_tmpl.format = tex->buffer.b.b.format;
                        surf_tmpl.u.tex.level = level;
                        surf_tmpl.u.tex.first_layer = layer;
                        surf_tmpl.u.tex.last_layer = layer;
-                       cbsurf = sctx->b.create_surface(&sctx->b, &rtex->buffer.b.b, &surf_tmpl);
+                       cbsurf = sctx->b.create_surface(&sctx->b, &tex->buffer.b.b, &surf_tmpl);
 
                        /* Required before and after FMASK and DCC_DECOMPRESS. */
                        if (custom_blend == sctx->custom_blend_fmask_decompress ||
@@ -514,26 +506,34 @@ static void si_blit_decompress_color(struct si_context *sctx,
                /* The texture will always be dirty if some layers aren't flushed.
                 * I don't think this case occurs often though. */
                if (first_layer == 0 && last_layer >= max_layer) {
-                       rtex->dirty_level_mask &= ~(1 << level);
+                       tex->dirty_level_mask &= ~(1 << level);
                }
        }
 
        sctx->decompression_enabled = false;
-       si_make_CB_shader_coherent(sctx, rtex->buffer.b.b.nr_samples,
-                                  vi_dcc_enabled(rtex, first_level));
+       si_make_CB_shader_coherent(sctx, tex->buffer.b.b.nr_samples,
+                                  vi_dcc_enabled(tex, first_level),
+                                  tex->surface.u.gfx9.dcc.pipe_aligned);
+
+expand_fmask:
+       if (need_fmask_expand && tex->surface.fmask_offset && !tex->fmask_is_identity) {
+               si_compute_expand_fmask(&sctx->b, &tex->buffer.b.b);
+               tex->fmask_is_identity = true;
+       }
 }
 
 static void
-si_decompress_color_texture(struct si_context *sctx, struct r600_texture *tex,
-                           unsigned first_level, unsigned last_level)
+si_decompress_color_texture(struct si_context *sctx, struct si_texture *tex,
+                           unsigned first_level, unsigned last_level,
+                           bool need_fmask_expand)
 {
        /* CMASK or DCC can be discarded and we can still end up here. */
-       if (!tex->cmask.size && !tex->fmask.size && !tex->dcc_offset)
+       if (!tex->cmask_buffer && !tex->surface.fmask_size && !tex->surface.dcc_offset)
                return;
 
        si_blit_decompress_color(sctx, tex, first_level, last_level, 0,
                                 util_max_layer(&tex->buffer.b.b, first_level),
-                                false);
+                                false, need_fmask_expand);
 }
 
 static void
@@ -545,17 +545,17 @@ si_decompress_sampler_color_textures(struct si_context *sctx,
 
        while (mask) {
                struct pipe_sampler_view *view;
-               struct r600_texture *tex;
+               struct si_texture *tex;
 
                i = u_bit_scan(&mask);
 
                view = textures->views[i];
                assert(view);
 
-               tex = (struct r600_texture *)view->texture;
+               tex = (struct si_texture *)view->texture;
 
                si_decompress_color_texture(sctx, tex, view->u.tex.first_level,
-                                           view->u.tex.last_level);
+                                           view->u.tex.last_level, false);
        }
 }
 
@@ -568,22 +568,23 @@ si_decompress_image_color_textures(struct si_context *sctx,
 
        while (mask) {
                const struct pipe_image_view *view;
-               struct r600_texture *tex;
+               struct si_texture *tex;
 
                i = u_bit_scan(&mask);
 
                view = &images->views[i];
                assert(view->resource->target != PIPE_BUFFER);
 
-               tex = (struct r600_texture *)view->resource;
+               tex = (struct si_texture *)view->resource;
 
                si_decompress_color_texture(sctx, tex, view->u.tex.level,
-                                           view->u.tex.level);
+                                           view->u.tex.level,
+                                           view->access & PIPE_IMAGE_ACCESS_WRITE);
        }
 }
 
 static void si_check_render_feedback_texture(struct si_context *sctx,
-                                            struct r600_texture *tex,
+                                            struct si_texture *tex,
                                             unsigned first_level,
                                             unsigned last_level,
                                             unsigned first_layer,
@@ -591,18 +592,18 @@ static void si_check_render_feedback_texture(struct si_context *sctx,
 {
        bool render_feedback = false;
 
-       if (!tex->dcc_offset)
+       if (!tex->surface.dcc_offset)
                return;
 
        for (unsigned j = 0; j < sctx->framebuffer.state.nr_cbufs; ++j) {
-               struct r600_surface * surf;
+               struct si_surface * surf;
 
                if (!sctx->framebuffer.state.cbufs[j])
                        continue;
 
-               surf = (struct r600_surface*)sctx->framebuffer.state.cbufs[j];
+               surf = (struct si_surface*)sctx->framebuffer.state.cbufs[j];
 
-               if (tex == (struct r600_texture *)surf->base.texture &&
+               if (tex == (struct si_texture *)surf->base.texture &&
                    surf->base.u.tex.level >= first_level &&
                    surf->base.u.tex.level <= last_level &&
                    surf->base.u.tex.first_layer <= last_layer &&
@@ -623,7 +624,7 @@ static void si_check_render_feedback_textures(struct si_context *sctx,
 
        while (mask) {
                const struct pipe_sampler_view *view;
-               struct r600_texture *tex;
+               struct si_texture *tex;
 
                unsigned i = u_bit_scan(&mask);
 
@@ -631,7 +632,7 @@ static void si_check_render_feedback_textures(struct si_context *sctx,
                if(view->texture->target == PIPE_BUFFER)
                        continue;
 
-               tex = (struct r600_texture *)view->texture;
+               tex = (struct si_texture *)view->texture;
 
                si_check_render_feedback_texture(sctx, tex,
                                                 view->u.tex.first_level,
@@ -648,7 +649,7 @@ static void si_check_render_feedback_images(struct si_context *sctx,
 
        while (mask) {
                const struct pipe_image_view *view;
-               struct r600_texture *tex;
+               struct si_texture *tex;
 
                unsigned i = u_bit_scan(&mask);
 
@@ -656,7 +657,7 @@ static void si_check_render_feedback_images(struct si_context *sctx,
                if (view->resource->target == PIPE_BUFFER)
                        continue;
 
-               tex = (struct r600_texture *)view->resource;
+               tex = (struct si_texture *)view->resource;
 
                si_check_render_feedback_texture(sctx, tex,
                                                 view->u.tex.level,
@@ -671,13 +672,13 @@ static void si_check_render_feedback_resident_textures(struct si_context *sctx)
        util_dynarray_foreach(&sctx->resident_tex_handles,
                              struct si_texture_handle *, tex_handle) {
                struct pipe_sampler_view *view;
-               struct r600_texture *tex;
+               struct si_texture *tex;
 
                view = (*tex_handle)->view;
                if (view->texture->target == PIPE_BUFFER)
                        continue;
 
-               tex = (struct r600_texture *)view->texture;
+               tex = (struct si_texture *)view->texture;
 
                si_check_render_feedback_texture(sctx, tex,
                                                 view->u.tex.first_level,
@@ -692,13 +693,13 @@ static void si_check_render_feedback_resident_images(struct si_context *sctx)
        util_dynarray_foreach(&sctx->resident_img_handles,
                              struct si_image_handle *, img_handle) {
                struct pipe_image_view *view;
-               struct r600_texture *tex;
+               struct si_texture *tex;
 
                view = &(*img_handle)->view;
                if (view->resource->target == PIPE_BUFFER)
                        continue;
 
-               tex = (struct r600_texture *)view->resource;
+               tex = (struct si_texture *)view->resource;
 
                si_check_render_feedback_texture(sctx, tex,
                                                 view->u.tex.level,
@@ -735,17 +736,17 @@ static void si_decompress_resident_textures(struct si_context *sctx)
        util_dynarray_foreach(&sctx->resident_tex_needs_color_decompress,
                              struct si_texture_handle *, tex_handle) {
                struct pipe_sampler_view *view = (*tex_handle)->view;
-               struct r600_texture *tex = (struct r600_texture *)view->texture;
+               struct si_texture *tex = (struct si_texture *)view->texture;
 
                si_decompress_color_texture(sctx, tex, view->u.tex.first_level,
-                                           view->u.tex.last_level);
+                                           view->u.tex.last_level, false);
        }
 
        util_dynarray_foreach(&sctx->resident_tex_needs_depth_decompress,
                              struct si_texture_handle *, tex_handle) {
                struct pipe_sampler_view *view = (*tex_handle)->view;
                struct si_sampler_view *sview = (struct si_sampler_view *)view;
-               struct r600_texture *tex = (struct r600_texture *)view->texture;
+               struct si_texture *tex = (struct si_texture *)view->texture;
 
                si_decompress_depth(sctx, tex,
                        sview->is_stencil_sampler ? PIPE_MASK_S : PIPE_MASK_Z,
@@ -759,10 +760,11 @@ static void si_decompress_resident_images(struct si_context *sctx)
        util_dynarray_foreach(&sctx->resident_img_needs_color_decompress,
                              struct si_image_handle *, img_handle) {
                struct pipe_image_view *view = &(*img_handle)->view;
-               struct r600_texture *tex = (struct r600_texture *)view->resource;
+               struct si_texture *tex = (struct si_texture *)view->resource;
 
                si_decompress_color_texture(sctx, tex, view->u.tex.level,
-                                           view->u.tex.level);
+                                           view->u.tex.level,
+                                           view->access & PIPE_IMAGE_ACCESS_WRITE);
        }
 }
 
@@ -805,16 +807,16 @@ void si_decompress_textures(struct si_context *sctx, unsigned shader_mask)
                if (sctx->ps_uses_fbfetch) {
                        struct pipe_surface *cb0 = sctx->framebuffer.state.cbufs[0];
                        si_decompress_color_texture(sctx,
-                                                   (struct r600_texture*)cb0->texture,
+                                                   (struct si_texture*)cb0->texture,
                                                    cb0->u.tex.first_layer,
-                                                   cb0->u.tex.last_layer);
+                                                   cb0->u.tex.last_layer, false);
                }
 
                si_check_render_feedback(sctx);
        } else if (shader_mask & (1 << PIPE_SHADER_COMPUTE)) {
-               if (sctx->cs_shader_state.program->uses_bindless_samplers)
+               if (sctx->cs_shader_state.program->sel.info.uses_bindless_samplers)
                        si_decompress_resident_textures(sctx);
-               if (sctx->cs_shader_state.program->uses_bindless_images)
+               if (sctx->cs_shader_state.program->sel.info.uses_bindless_images)
                        si_decompress_resident_images(sctx);
        }
 }
@@ -829,12 +831,12 @@ static void si_decompress_subresource(struct pipe_context *ctx,
                                      unsigned first_layer, unsigned last_layer)
 {
        struct si_context *sctx = (struct si_context *)ctx;
-       struct r600_texture *rtex = (struct r600_texture*)tex;
+       struct si_texture *stex = (struct si_texture*)tex;
 
-       if (rtex->db_compatible) {
+       if (stex->db_compatible) {
                planes &= PIPE_MASK_Z | PIPE_MASK_S;
 
-               if (!rtex->surface.has_stencil)
+               if (!stex->surface.has_stencil)
                        planes &= ~PIPE_MASK_S;
 
                /* If we've rendered into the framebuffer and it's a blitting
@@ -846,10 +848,10 @@ static void si_decompress_subresource(struct pipe_context *ctx,
                    sctx->framebuffer.state.zsbuf->texture == tex)
                        si_update_fb_dirtiness_after_rendering(sctx);
 
-               si_decompress_depth(sctx, rtex, planes,
+               si_decompress_depth(sctx, stex, planes,
                                    level, level,
                                    first_layer, last_layer);
-       } else if (rtex->fmask.size || rtex->cmask.size || rtex->dcc_offset) {
+       } else if (stex->surface.fmask_size || stex->cmask_buffer || stex->surface.dcc_offset) {
                /* If we've rendered into the framebuffer and it's a blitting
                 * source, make sure the decompression pass is invoked
                 * by dirtying the framebuffer.
@@ -863,8 +865,8 @@ static void si_decompress_subresource(struct pipe_context *ctx,
                        }
                }
 
-               si_blit_decompress_color(sctx, rtex, level, level,
-                                        first_layer, last_layer, false);
+               si_blit_decompress_color(sctx, stex, level, level,
+                                        first_layer, last_layer, false, false);
        }
 }
 
@@ -887,7 +889,8 @@ void si_resource_copy_region(struct pipe_context *ctx,
                             const struct pipe_box *src_box)
 {
        struct si_context *sctx = (struct si_context *)ctx;
-       struct r600_texture *rsrc = (struct r600_texture*)src;
+       struct si_texture *ssrc = (struct si_texture*)src;
+       struct si_texture *sdst = (struct si_texture*)dst;
        struct pipe_surface *dst_view, dst_templ;
        struct pipe_sampler_view src_templ, *src_view;
        unsigned dst_width, dst_height, src_width0, src_height0;
@@ -896,7 +899,18 @@ void si_resource_copy_region(struct pipe_context *ctx,
 
        /* Handle buffers first. */
        if (dst->target == PIPE_BUFFER && src->target == PIPE_BUFFER) {
-               si_copy_buffer(sctx, dst, src, dstx, src_box->x, src_box->width, 0);
+               si_copy_buffer(sctx, dst, src, dstx, src_box->x, src_box->width);
+               return;
+       }
+
+       if (!util_format_is_compressed(src->format) &&
+           !util_format_is_compressed(dst->format) &&
+           !util_format_is_depth_or_stencil(src->format) &&
+           src->nr_samples <= 1 &&
+           !sdst->surface.dcc_offset &&
+           !(dst->target != src->target &&
+             (src->target == PIPE_TEXTURE_1D_ARRAY || dst->target == PIPE_TEXTURE_1D_ARRAY))) {
+               si_compute_copy_image(sctx, dst, dst_level, src, src_level, dstx, dsty, dstz, src_box);
                return;
        }
 
@@ -919,7 +933,7 @@ void si_resource_copy_region(struct pipe_context *ctx,
 
        if (util_format_is_compressed(src->format) ||
            util_format_is_compressed(dst->format)) {
-               unsigned blocksize = rsrc->surface.bpe;
+               unsigned blocksize = ssrc->surface.bpe;
 
                if (blocksize == 8)
                        src_templ.format = PIPE_FORMAT_R16G16B16A16_UINT; /* 64-bit block */
@@ -962,7 +976,7 @@ void si_resource_copy_region(struct pipe_context *ctx,
                        sbox.width = util_format_get_nblocksx(src->format, src_box->width);
                        src_box = &sbox;
                } else {
-                       unsigned blocksize = rsrc->surface.bpe;
+                       unsigned blocksize = ssrc->surface.bpe;
 
                        switch (blocksize) {
                        case 1:
@@ -998,36 +1012,8 @@ void si_resource_copy_region(struct pipe_context *ctx,
         * Note that some chips avoid this issue by using SDMA.
         */
        if (util_format_is_snorm8(dst_templ.format)) {
-               switch (dst_templ.format) {
-               case PIPE_FORMAT_R8_SNORM:
-                       dst_templ.format = src_templ.format = PIPE_FORMAT_R8_SINT;
-                       break;
-               case PIPE_FORMAT_R8G8_SNORM:
-                       dst_templ.format = src_templ.format = PIPE_FORMAT_R8G8_SINT;
-                       break;
-               case PIPE_FORMAT_R8G8B8X8_SNORM:
-                       dst_templ.format = src_templ.format = PIPE_FORMAT_R8G8B8X8_SINT;
-                       break;
-               case PIPE_FORMAT_R8G8B8A8_SNORM:
-               /* There are no SINT variants for ABGR and XBGR, so we have to use RGBA. */
-               case PIPE_FORMAT_A8B8G8R8_SNORM:
-               case PIPE_FORMAT_X8B8G8R8_SNORM:
-                       dst_templ.format = src_templ.format = PIPE_FORMAT_R8G8B8A8_SINT;
-                       break;
-               case PIPE_FORMAT_A8_SNORM:
-                       dst_templ.format = src_templ.format = PIPE_FORMAT_A8_SINT;
-                       break;
-               case PIPE_FORMAT_L8_SNORM:
-                       dst_templ.format = src_templ.format = PIPE_FORMAT_L8_SINT;
-                       break;
-               case PIPE_FORMAT_L8A8_SNORM:
-                       dst_templ.format = src_templ.format = PIPE_FORMAT_L8A8_SINT;
-                       break;
-               case PIPE_FORMAT_I8_SNORM:
-                       dst_templ.format = src_templ.format = PIPE_FORMAT_I8_SINT;
-                       break;
-               default:; /* fall through */
-               }
+               dst_templ.format = src_templ.format =
+                       util_format_snorm8_to_sint8(dst_templ.format);
        }
 
        vi_disable_dcc_if_incompatible_format(sctx, dst, dst_level,
@@ -1078,16 +1064,16 @@ static void si_do_CB_resolve(struct si_context *sctx,
        si_blitter_end(sctx);
 
        /* Flush caches for possible texturing. */
-       si_make_CB_shader_coherent(sctx, 1, false);
+       si_make_CB_shader_coherent(sctx, 1, false, true /* no DCC */);
 }
 
 static bool do_hardware_msaa_resolve(struct pipe_context *ctx,
                                     const struct pipe_blit_info *info)
 {
        struct si_context *sctx = (struct si_context*)ctx;
-       struct r600_texture *src = (struct r600_texture*)info->src.resource;
-       struct r600_texture *dst = (struct r600_texture*)info->dst.resource;
-       MAYBE_UNUSED struct r600_texture *rtmp;
+       struct si_texture *src = (struct si_texture*)info->src.resource;
+       struct si_texture *dst = (struct si_texture*)info->dst.resource;
+       ASSERTED struct si_texture *stmp;
        unsigned dst_width = u_minify(info->dst.resource->width0, info->dst.level);
        unsigned dst_height = u_minify(info->dst.resource->height0, info->dst.level);
        enum pipe_format format = info->src.format;
@@ -1129,12 +1115,18 @@ static bool do_hardware_msaa_resolve(struct pipe_context *ctx,
            info->src.box.height == dst_height &&
            info->src.box.depth == 1 &&
            !dst->surface.is_linear &&
-           (!dst->cmask.size || !dst->dirty_level_mask)) { /* dst cannot be fast-cleared */
+           (!dst->cmask_buffer || !dst->dirty_level_mask)) { /* dst cannot be fast-cleared */
                /* Check the last constraint. */
                if (src->surface.micro_tile_mode != dst->surface.micro_tile_mode) {
                        /* The next fast clear will switch to this mode to
                         * get direct hw resolve next time if the mode is
                         * different now.
+                        *
+                        * TODO-GFX10: This does not work in GFX10 because MSAA
+                        * is restricted to 64KB_R_X and 64KB_Z_X swizzle modes.
+                        * In some cases we could change the swizzle of the
+                        * destination texture instead, but the more general
+                        * solution is to implement compute shader resolve.
                         */
                        src->last_msaa_resolve_target_micro_mode =
                                dst->surface.micro_tile_mode;
@@ -1146,13 +1138,10 @@ static bool do_hardware_msaa_resolve(struct pipe_context *ctx,
                 * This is still the fastest codepath even with this clear.
                 */
                if (vi_dcc_enabled(dst, info->dst.level)) {
-                       /* TODO: Implement per-level DCC clears for GFX9. */
-                       if (sctx->chip_class >= GFX9 &&
-                           info->dst.resource->last_level != 0)
+                       if (!vi_dcc_clear_level(sctx, dst, info->dst.level,
+                                               DCC_UNCOMPRESSED))
                                goto resolve_to_temp;
 
-                       vi_dcc_clear_level(sctx, dst, info->dst.level,
-                                          0xFFFFFFFF);
                        dst->dirty_level_mask &= ~(1 << info->dst.level);
                }
 
@@ -1174,11 +1163,14 @@ resolve_to_temp:
        templ.depth0 = 1;
        templ.array_size = 1;
        templ.usage = PIPE_USAGE_DEFAULT;
-       templ.flags = SI_RESOURCE_FLAG_FORCE_TILING |
+       templ.flags = SI_RESOURCE_FLAG_FORCE_MSAA_TILING |
+                     SI_RESOURCE_FLAG_FORCE_MICRO_TILE_MODE |
+                     SI_RESOURCE_FLAG_MICRO_TILE_MODE_SET(src->surface.micro_tile_mode) |
                      SI_RESOURCE_FLAG_DISABLE_DCC;
 
        /* The src and dst microtile modes must be the same. */
-       if (src->surface.micro_tile_mode == RADEON_MICRO_MODE_DISPLAY)
+       if (sctx->chip_class <= GFX8 &&
+           src->surface.micro_tile_mode == RADEON_MICRO_MODE_DISPLAY)
                templ.bind = PIPE_BIND_SCANOUT;
        else
                templ.bind = 0;
@@ -1186,10 +1178,10 @@ resolve_to_temp:
        tmp = ctx->screen->resource_create(ctx->screen, &templ);
        if (!tmp)
                return false;
-       rtmp = (struct r600_texture*)tmp;
+       stmp = (struct si_texture*)tmp;
 
-       assert(!rtmp->surface.is_linear);
-       assert(src->surface.micro_tile_mode == rtmp->surface.micro_tile_mode);
+       assert(!stmp->surface.is_linear);
+       assert(src->surface.micro_tile_mode == stmp->surface.micro_tile_mode);
 
        /* resolve */
        si_do_CB_resolve(sctx, info, tmp, 0, 0, format);
@@ -1212,7 +1204,7 @@ static void si_blit(struct pipe_context *ctx,
                    const struct pipe_blit_info *info)
 {
        struct si_context *sctx = (struct si_context*)ctx;
-       struct r600_texture *rdst = (struct r600_texture *)info->dst.resource;
+       struct si_texture *dst = (struct si_texture *)info->dst.resource;
 
        if (do_hardware_msaa_resolve(ctx, info)) {
                return;
@@ -1224,8 +1216,7 @@ static void si_blit(struct pipe_context *ctx,
         * resource_copy_region can't do this yet, because dma_copy calls it
         * on failure (recursion).
         */
-       if (rdst->surface.is_linear &&
-           sctx->dma_copy &&
+       if (dst->surface.is_linear &&
            util_can_blit_via_copy_region(info, false)) {
                sctx->dma_copy(ctx, info->dst.resource, info->dst.level,
                                 info->dst.box.x, info->dst.box.y,
@@ -1245,12 +1236,12 @@ static void si_blit(struct pipe_context *ctx,
        vi_disable_dcc_if_incompatible_format(sctx, info->dst.resource,
                                              info->dst.level,
                                              info->dst.format);
-       si_decompress_subresource(ctx, info->src.resource, info->mask,
+       si_decompress_subresource(ctx, info->src.resource, PIPE_MASK_RGBAZS,
                                  info->src.level,
                                  info->src.box.z,
                                  info->src.box.z + info->src.box.depth - 1);
 
-       if (sctx->screen->debug_flags & DBG(FORCE_DMA) &&
+       if (sctx->screen->debug_flags & DBG(FORCE_SDMA) &&
            util_try_blit_via_copy_region(ctx, info))
                return;
 
@@ -1260,14 +1251,14 @@ static void si_blit(struct pipe_context *ctx,
        si_blitter_end(sctx);
 }
 
-static boolean si_generate_mipmap(struct pipe_context *ctx,
-                                 struct pipe_resource *tex,
-                                 enum pipe_format format,
-                                 unsigned base_level, unsigned last_level,
-                                 unsigned first_layer, unsigned last_layer)
+static bool si_generate_mipmap(struct pipe_context *ctx,
+                              struct pipe_resource *tex,
+                              enum pipe_format format,
+                              unsigned base_level, unsigned last_level,
+                              unsigned first_layer, unsigned last_layer)
 {
        struct si_context *sctx = (struct si_context*)ctx;
-       struct r600_texture *rtex = (struct r600_texture *)tex;
+       struct si_texture *stex = (struct si_texture *)tex;
 
        if (!util_blitter_is_copy_supported(sctx->blitter, tex, tex))
                return false;
@@ -1281,10 +1272,10 @@ static boolean si_generate_mipmap(struct pipe_context *ctx,
 
        /* Clear dirty_level_mask for the levels that will be overwritten. */
        assert(base_level < last_level);
-       rtex->dirty_level_mask &= ~u_bit_consecutive(base_level + 1,
+       stex->dirty_level_mask &= ~u_bit_consecutive(base_level + 1,
                                                     last_level - base_level);
 
-       sctx->generate_mipmap_for_depth = rtex->is_depth;
+       sctx->generate_mipmap_for_depth = stex->is_depth;
 
        si_blitter_begin(sctx, SI_BLIT | SI_DISABLE_RENDER_COND);
        util_blitter_generate_mipmap(sctx->blitter, tex, format,
@@ -1300,43 +1291,78 @@ static void si_flush_resource(struct pipe_context *ctx,
                              struct pipe_resource *res)
 {
        struct si_context *sctx = (struct si_context*)ctx;
-       struct r600_texture *rtex = (struct r600_texture*)res;
+       struct si_texture *tex = (struct si_texture*)res;
 
        assert(res->target != PIPE_BUFFER);
-       assert(!rtex->dcc_separate_buffer || rtex->dcc_gather_statistics);
+       assert(!tex->dcc_separate_buffer || tex->dcc_gather_statistics);
 
        /* st/dri calls flush twice per frame (not a bug), this prevents double
         * decompression. */
-       if (rtex->dcc_separate_buffer && !rtex->separate_dcc_dirty)
+       if (tex->dcc_separate_buffer && !tex->separate_dcc_dirty)
                return;
 
-       if (!rtex->is_depth && (rtex->cmask.size || rtex->dcc_offset)) {
-               si_blit_decompress_color(sctx, rtex, 0, res->last_level,
+       if (!tex->is_depth && (tex->cmask_buffer || tex->surface.dcc_offset)) {
+               si_blit_decompress_color(sctx, tex, 0, res->last_level,
                                         0, util_max_layer(res, 0),
-                                        rtex->dcc_separate_buffer != NULL);
+                                        tex->dcc_separate_buffer != NULL, false);
+
+               if (tex->surface.display_dcc_offset && tex->displayable_dcc_dirty) {
+                       si_retile_dcc(sctx, tex);
+                       tex->displayable_dcc_dirty = false;
+               }
        }
 
        /* Always do the analysis even if DCC is disabled at the moment. */
-       if (rtex->dcc_gather_statistics && rtex->separate_dcc_dirty) {
-               rtex->separate_dcc_dirty = false;
-               vi_separate_dcc_process_and_reset_stats(ctx, rtex);
+       if (tex->dcc_gather_statistics) {
+               bool separate_dcc_dirty = tex->separate_dcc_dirty;
+
+               /* If the color buffer hasn't been unbound and fast clear hasn't
+                * been used, separate_dcc_dirty is false, but there may have been
+                * new rendering. Check if the color buffer is bound and assume
+                * it's dirty.
+                *
+                * Note that DRI2 never unbinds window colorbuffers, which means
+                * the DCC pipeline statistics query would never be re-set and would
+                * keep adding new results until all free memory is exhausted if we
+                * didn't do this.
+                */
+               if (!separate_dcc_dirty) {
+                       for (unsigned i = 0; i < sctx->framebuffer.state.nr_cbufs; i++) {
+                               if (sctx->framebuffer.state.cbufs[i] &&
+                                   sctx->framebuffer.state.cbufs[i]->texture == res) {
+                                       separate_dcc_dirty = true;
+                                       break;
+                               }
+                       }
+               }
+
+               if (separate_dcc_dirty) {
+                       tex->separate_dcc_dirty = false;
+                       vi_separate_dcc_process_and_reset_stats(ctx, tex);
+               }
        }
 }
 
-void si_decompress_dcc(struct si_context *sctx, struct r600_texture *rtex)
+void si_decompress_dcc(struct si_context *sctx, struct si_texture *tex)
 {
-       if (!rtex->dcc_offset)
+       /* If graphics is disabled, we can't decompress DCC, but it shouldn't
+        * be compressed either. The caller should simply discard it.
+        */
+       if (!tex->surface.dcc_offset || !sctx->has_graphics)
                return;
 
-       si_blit_decompress_color(sctx, rtex, 0, rtex->buffer.b.b.last_level,
-                                0, util_max_layer(&rtex->buffer.b.b, 0),
-                                true);
+       si_blit_decompress_color(sctx, tex, 0, tex->buffer.b.b.last_level,
+                                0, util_max_layer(&tex->buffer.b.b, 0),
+                                true, false);
 }
 
 void si_init_blit_functions(struct si_context *sctx)
 {
        sctx->b.resource_copy_region = si_resource_copy_region;
-       sctx->b.blit = si_blit;
-       sctx->b.flush_resource = si_flush_resource;
-       sctx->b.generate_mipmap = si_generate_mipmap;
+
+       if (sctx->has_graphics) {
+               sctx->b.blit = si_blit;
+               sctx->b.flush_resource = si_flush_resource;
+               sctx->b.generate_mipmap = si_generate_mipmap;
+       }
 }