radeonsi: add llvm processor names for CIK
[mesa.git] / src / gallium / drivers / radeonsi / r600_blit.c
index 4406204b5129836b8d23457b02e4b109601e79d7..34f14bae03479349679fb4e8bdd49041aea235e4 100644 (file)
@@ -39,6 +39,9 @@ enum r600_blitter_op /* bitmask */
        R600_COPY          = R600_SAVE_FRAMEBUFFER | R600_SAVE_TEXTURES |
                             R600_DISABLE_RENDER_COND,
 
+       R600_BLIT          = R600_SAVE_FRAMEBUFFER | R600_SAVE_TEXTURES |
+                            R600_DISABLE_RENDER_COND,
+
        R600_DECOMPRESS    = R600_SAVE_FRAMEBUFFER | R600_DISABLE_RENDER_COND,
 };
 
@@ -58,9 +61,7 @@ static void r600_blitter_begin(struct pipe_context *ctx, enum r600_blitter_op op
        if (rctx->queued.named.viewport) {
                util_blitter_save_viewport(rctx->blitter, &rctx->queued.named.viewport->viewport);
        }
-       util_blitter_save_vertex_buffers(rctx->blitter,
-                                        rctx->nr_vertex_buffers,
-                                        rctx->vertex_buffer);
+       util_blitter_save_vertex_buffer_slot(rctx->blitter, rctx->vertex_buffer);
        util_blitter_save_so_targets(rctx->blitter, rctx->num_so_targets,
                                     (struct pipe_stream_output_target**)rctx->so_targets);
 
@@ -79,8 +80,9 @@ static void r600_blitter_begin(struct pipe_context *ctx, enum r600_blitter_op op
 
        if ((op & R600_DISABLE_RENDER_COND) && rctx->current_render_cond) {
                rctx->saved_render_cond = rctx->current_render_cond;
+               rctx->saved_render_cond_cond = rctx->current_render_cond_cond;
                rctx->saved_render_cond_mode = rctx->current_render_cond_mode;
-               rctx->context.render_condition(&rctx->context, NULL, 0);
+               rctx->context.render_condition(&rctx->context, NULL, FALSE, 0);
        }
 
 }
@@ -91,103 +93,147 @@ static void r600_blitter_end(struct pipe_context *ctx)
        if (rctx->saved_render_cond) {
                rctx->context.render_condition(&rctx->context,
                                               rctx->saved_render_cond,
+                                              rctx->saved_render_cond_cond,
                                               rctx->saved_render_cond_mode);
                rctx->saved_render_cond = NULL;
        }
        r600_context_queries_resume(rctx);
 }
 
-static unsigned u_num_layers(struct pipe_resource *r, unsigned level)
-{
-       switch (r->target) {
-       case PIPE_TEXTURE_CUBE:
-               return 6;
-       case PIPE_TEXTURE_3D:
-               return u_minify(r->depth0, level);
-       case PIPE_TEXTURE_1D_ARRAY:
-               return r->array_size;
-       case PIPE_TEXTURE_2D_ARRAY:
-               return r->array_size;
-       default:
-               return 1;
-       }
-}
-
-void si_blit_uncompress_depth(struct pipe_context *ctx, struct r600_resource_texture *texture)
+void si_blit_uncompress_depth(struct pipe_context *ctx,
+               struct r600_resource_texture *texture,
+               struct r600_resource_texture *staging,
+               unsigned first_level, unsigned last_level,
+               unsigned first_layer, unsigned last_layer)
 {
        struct r600_context *rctx = (struct r600_context *)ctx;
-       unsigned layer, level;
+       unsigned layer, level, checked_last_layer, max_layer;
        float depth = 1.0f;
+       const struct util_format_description *desc;
+       void *custom_dsa;
+       struct r600_resource_texture *flushed_depth_texture = staging ?
+                       staging : texture->flushed_depth_texture;
 
-       if (!texture->dirty_db)
+       if (!staging && !texture->dirty_db_mask)
                return;
 
-       for (level = 0; level <= texture->resource.b.b.last_level; level++) {
-               unsigned num_layers = u_num_layers(&texture->resource.b.b, level);
+       desc = util_format_description(flushed_depth_texture->resource.b.b.format);
+       switch (util_format_has_depth(desc) | util_format_has_stencil(desc) << 1) {
+       default:
+               assert(!"No depth or stencil to uncompress");
+       case 3:
+               custom_dsa = rctx->custom_dsa_flush_depth_stencil;
+               break;
+       case 2:
+               custom_dsa = rctx->custom_dsa_flush_stencil;
+               break;
+       case 1:
+               custom_dsa = rctx->custom_dsa_flush_depth;
+               break;
+       }
+
+       for (level = first_level; level <= last_level; level++) {
+               if (!staging && !(texture->dirty_db_mask & (1 << level)))
+                       continue;
+
+               /* The smaller the mipmap level, the less layers there are
+                * as far as 3D textures are concerned. */
+               max_layer = util_max_layer(&texture->resource.b.b, level);
+               checked_last_layer = last_layer < max_layer ? last_layer : max_layer;
 
-               for (layer = 0; layer < num_layers; layer++) {
+               for (layer = first_layer; layer <= checked_last_layer; layer++) {
                        struct pipe_surface *zsurf, *cbsurf, surf_tmpl;
 
                        surf_tmpl.format = texture->real_format;
                        surf_tmpl.u.tex.level = level;
                        surf_tmpl.u.tex.first_layer = layer;
                        surf_tmpl.u.tex.last_layer = layer;
-                       surf_tmpl.usage = PIPE_BIND_DEPTH_STENCIL;
 
                        zsurf = ctx->create_surface(ctx, &texture->resource.b.b, &surf_tmpl);
 
-                       surf_tmpl.format = texture->flushed_depth_texture->real_format;
-                       surf_tmpl.usage = PIPE_BIND_RENDER_TARGET;
+                       surf_tmpl.format = flushed_depth_texture->real_format;
                        cbsurf = ctx->create_surface(ctx,
-                                       (struct pipe_resource*)texture->flushed_depth_texture, &surf_tmpl);
+                                       (struct pipe_resource*)flushed_depth_texture, &surf_tmpl);
 
                        r600_blitter_begin(ctx, R600_DECOMPRESS);
-                       util_blitter_custom_depth_stencil(rctx->blitter, zsurf, cbsurf, ~0, rctx->custom_dsa_flush, depth);
+                       util_blitter_custom_depth_stencil(rctx->blitter, zsurf, cbsurf, ~0, custom_dsa, depth);
                        r600_blitter_end(ctx);
 
                        pipe_surface_reference(&zsurf, NULL);
                        pipe_surface_reference(&cbsurf, NULL);
                }
-       }
 
-       texture->dirty_db = FALSE;
+               /* The texture will always be dirty if some layers aren't flushed.
+                * I don't think this case can occur though. */
+               if (!staging && first_layer == 0 && last_layer == max_layer) {
+                       texture->dirty_db_mask &= ~(1 << level);
+               }
+       }
 }
 
-void si_flush_depth_textures(struct r600_context *rctx)
+static void si_blit_decompress_depth_in_place(struct r600_context *rctx,
+                                              struct r600_resource_texture *texture,
+                                              unsigned first_level, unsigned last_level,
+                                              unsigned first_layer, unsigned last_layer)
 {
-       unsigned int i;
+       struct pipe_surface *zsurf, surf_tmpl = {{0}};
+       unsigned layer, max_layer, checked_last_layer, level;
 
-       /* FIXME: This handles fragment shader textures only. */
+       surf_tmpl.format = texture->resource.b.b.format;
 
-       for (i = 0; i < rctx->ps_samplers.n_views; ++i) {
-               struct si_pipe_sampler_view *view;
-               struct r600_resource_texture *tex;
+       for (level = first_level; level <= last_level; level++) {
+               if (!(texture->dirty_db_mask & (1 << level)))
+                       continue;
 
-               view = rctx->ps_samplers.views[i];
-               if (!view) continue;
+               surf_tmpl.u.tex.level = level;
 
-               tex = (struct r600_resource_texture *)view->base.texture;
-               if (!tex->depth)
-                       continue;
+               /* The smaller the mipmap level, the less layers there are
+                * as far as 3D textures are concerned. */
+               max_layer = util_max_layer(&texture->resource.b.b, level);
+               checked_last_layer = last_layer < max_layer ? last_layer : max_layer;
 
-               if (tex->is_flushing_texture)
-                       continue;
+               for (layer = first_layer; layer <= checked_last_layer; layer++) {
+                       surf_tmpl.u.tex.first_layer = layer;
+                       surf_tmpl.u.tex.last_layer = layer;
+
+                       zsurf = rctx->context.create_surface(&rctx->context, &texture->resource.b.b, &surf_tmpl);
+
+                       r600_blitter_begin(&rctx->context, R600_DECOMPRESS);
+                       util_blitter_custom_depth_stencil(rctx->blitter, zsurf, NULL, ~0,
+                                                         rctx->custom_dsa_flush_inplace,
+                                                         1.0f);
+                       r600_blitter_end(&rctx->context);
+
+                       pipe_surface_reference(&zsurf, NULL);
+               }
 
-               si_blit_uncompress_depth(&rctx->context, tex);
+               /* 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) {
+                       texture->dirty_db_mask &= ~(1 << level);
+               }
        }
+}
+
+void si_flush_depth_textures(struct r600_context *rctx,
+                            struct r600_textures_info *textures)
+{
+       unsigned i;
 
-       /* also check CB here */
-       for (i = 0; i < rctx->framebuffer.nr_cbufs; i++) {
+       for (i = 0; i < textures->n_views; ++i) {
+               struct pipe_sampler_view *view;
                struct r600_resource_texture *tex;
-               tex = (struct r600_resource_texture *)rctx->framebuffer.cbufs[i]->texture;
 
-               if (!tex->depth)
-                       continue;
+               view = &textures->views[i]->base;
+               if (!view) continue;
 
-               if (tex->is_flushing_texture)
+               tex = (struct r600_resource_texture *)view->texture;
+               if (!tex->is_depth || tex->is_flushing_texture)
                        continue;
 
-               si_blit_uncompress_depth(&rctx->context, tex);
+               si_blit_decompress_depth_in_place(rctx, tex,
+                                                 view->u.tex.first_level, view->u.tex.last_level,
+                                                 0, util_max_layer(&tex->resource.b.b, view->u.tex.first_level));
        }
 }
 
@@ -200,8 +246,7 @@ static void r600_clear(struct pipe_context *ctx, unsigned buffers,
 
        r600_blitter_begin(ctx, R600_CLEAR);
        util_blitter_clear(rctx->blitter, fb->width, fb->height,
-                          fb->nr_cbufs, buffers, fb->nr_cbufs ? fb->cbufs[0]->format : PIPE_FORMAT_NONE,
-                          color, depth, stencil);
+                          buffers, color, depth, stencil);
        r600_blitter_end(ctx);
 }
 
@@ -235,25 +280,6 @@ static void r600_clear_depth_stencil(struct pipe_context *ctx,
        r600_blitter_end(ctx);
 }
 
-
-
-/* Copy a block of pixels from one surface to another using HW. */
-static void r600_hw_copy_region(struct pipe_context *ctx,
-                               struct pipe_resource *dst,
-                               unsigned dst_level,
-                               unsigned dstx, unsigned dsty, unsigned dstz,
-                               struct pipe_resource *src,
-                               unsigned src_level,
-                               const struct pipe_box *src_box)
-{
-       struct r600_context *rctx = (struct r600_context *)ctx;
-
-       r600_blitter_begin(ctx, R600_COPY);
-       util_blitter_copy_texture(rctx->blitter, dst, dst_level, ~0, dstx, dsty, dstz,
-                                 src, src_level, 0, src_box);
-       r600_blitter_end(ctx);
-}
-
 struct texture_orig_info {
        unsigned format;
        unsigned width0;
@@ -269,7 +295,6 @@ static void r600_compressed_to_blittable(struct pipe_resource *tex,
                                   struct texture_orig_info *orig)
 {
        struct r600_resource_texture *rtex = (struct r600_resource_texture*)tex;
-       struct r600_screen *rscreen = (struct r600_screen *)tex->screen;
        unsigned pixsize = util_format_get_blocksize(rtex->real_format);
        int new_format;
        int new_height, new_width;
@@ -299,12 +324,29 @@ static void r600_compressed_to_blittable(struct pipe_resource *tex,
        rtex->surface.level[level].npix_y = util_format_get_nblocksy(orig->format, orig->npix_y);
 }
 
-static void r600_reset_blittable_to_compressed(struct pipe_resource *tex,
-                                              unsigned level,
-                                              struct texture_orig_info *orig)
+static void r600_change_format(struct pipe_resource *tex,
+                              unsigned level,
+                              struct texture_orig_info *orig,
+                              enum pipe_format format)
+{
+       struct r600_resource_texture *rtex = (struct r600_resource_texture*)tex;
+
+       orig->format = tex->format;
+       orig->width0 = tex->width0;
+       orig->height0 = tex->height0;
+       orig->npix0_x = rtex->surface.level[0].npix_x;
+       orig->npix0_y = rtex->surface.level[0].npix_y;
+       orig->npix_x = rtex->surface.level[level].npix_x;
+       orig->npix_y = rtex->surface.level[level].npix_y;
+
+       tex->format = format;
+}
+
+static void r600_reset_blittable_to_orig(struct pipe_resource *tex,
+                                        unsigned level,
+                                        struct texture_orig_info *orig)
 {
        struct r600_resource_texture *rtex = (struct r600_resource_texture*)tex;
-       struct r600_screen *rscreen = (struct r600_screen *)tex->screen;
 
        tex->format = orig->format;
        tex->width0 = orig->width0;
@@ -323,10 +365,11 @@ static void r600_resource_copy_region(struct pipe_context *ctx,
                                      unsigned src_level,
                                      const struct pipe_box *src_box)
 {
+       struct r600_context *rctx = (struct r600_context *)ctx;
        struct r600_resource_texture *rsrc = (struct r600_resource_texture*)src;
        struct texture_orig_info orig_info[2];
        struct pipe_box sbox;
-       const struct pipe_box *psbox;
+       const struct pipe_box *psbox = src_box;
        boolean restore_orig[2];
 
        memset(orig_info, 0, sizeof(orig_info));
@@ -338,12 +381,17 @@ static void r600_resource_copy_region(struct pipe_context *ctx,
                return;
        }
 
-       if (rsrc->depth && !rsrc->is_flushing_texture)
-               r600_texture_depth_flush(ctx, src, FALSE);
+       /* This must be done before entering u_blitter to avoid recursion. */
+       if (rsrc->is_depth && !rsrc->is_flushing_texture) {
+               si_blit_decompress_depth_in_place(rctx, rsrc,
+                                                 src_level, src_level,
+                                                 src_box->z, src_box->z + src_box->depth - 1);
+       }
 
        restore_orig[0] = restore_orig[1] = FALSE;
 
-       if (util_format_is_compressed(src->format)) {
+       if (util_format_is_compressed(src->format) &&
+           util_format_is_compressed(dst->format)) {
                r600_compressed_to_blittable(src, src_level, &orig_info[0]);
                restore_orig[0] = TRUE;
                sbox.x = util_format_get_nblocksx(orig_info[0].format, src_box->x);
@@ -353,25 +401,94 @@ static void r600_resource_copy_region(struct pipe_context *ctx,
                sbox.height = util_format_get_nblocksy(orig_info[0].format, src_box->height);
                sbox.depth = src_box->depth;
                psbox=&sbox;
-       } else
-               psbox=src_box;
 
-       if (util_format_is_compressed(dst->format)) {
                r600_compressed_to_blittable(dst, dst_level, &orig_info[1]);
                restore_orig[1] = TRUE;
                /* translate the dst box as well */
                dstx = util_format_get_nblocksx(orig_info[1].format, dstx);
                dsty = util_format_get_nblocksy(orig_info[1].format, dsty);
+       } else if (!util_blitter_is_copy_supported(rctx->blitter, dst, src,
+                                                  PIPE_MASK_RGBAZS)) {
+               unsigned blocksize = util_format_get_blocksize(src->format);
+
+               switch (blocksize) {
+               case 1:
+                       r600_change_format(src, src_level, &orig_info[0],
+                                          PIPE_FORMAT_R8_UNORM);
+                       r600_change_format(dst, dst_level, &orig_info[1],
+                                          PIPE_FORMAT_R8_UNORM);
+                       break;
+               case 2:
+                       r600_change_format(src, src_level, &orig_info[0],
+                                          PIPE_FORMAT_R8G8_UNORM);
+                       r600_change_format(dst, dst_level, &orig_info[1],
+                                          PIPE_FORMAT_R8G8_UNORM);
+                       break;
+               case 4:
+                       r600_change_format(src, src_level, &orig_info[0],
+                                          PIPE_FORMAT_R8G8B8A8_UNORM);
+                       r600_change_format(dst, dst_level, &orig_info[1],
+                                          PIPE_FORMAT_R8G8B8A8_UNORM);
+                       break;
+               case 8:
+                       r600_change_format(src, src_level, &orig_info[0],
+                                          PIPE_FORMAT_R16G16B16A16_UINT);
+                       r600_change_format(dst, dst_level, &orig_info[1],
+                                          PIPE_FORMAT_R16G16B16A16_UINT);
+                       break;
+               case 16:
+                       r600_change_format(src, src_level, &orig_info[0],
+                                          PIPE_FORMAT_R32G32B32A32_UINT);
+                       r600_change_format(dst, dst_level, &orig_info[1],
+                                          PIPE_FORMAT_R32G32B32A32_UINT);
+                       break;
+               default:
+                       fprintf(stderr, "Unhandled format %s with blocksize %u\n",
+                               util_format_short_name(src->format), blocksize);
+                       assert(0);
+               }
+               restore_orig[0] = TRUE;
+               restore_orig[1] = TRUE;
        }
 
-       r600_hw_copy_region(ctx, dst, dst_level, dstx, dsty, dstz,
-                           src, src_level, psbox);
+       r600_blitter_begin(ctx, R600_COPY);
+       util_blitter_copy_texture(rctx->blitter, dst, dst_level, dstx, dsty, dstz,
+                                 src, src_level, psbox, PIPE_MASK_RGBAZS, TRUE);
+       r600_blitter_end(ctx);
 
        if (restore_orig[0])
-               r600_reset_blittable_to_compressed(src, src_level, &orig_info[0]);
+               r600_reset_blittable_to_orig(src, src_level, &orig_info[0]);
 
        if (restore_orig[1])
-               r600_reset_blittable_to_compressed(dst, dst_level, &orig_info[1]);
+               r600_reset_blittable_to_orig(dst, dst_level, &orig_info[1]);
+}
+
+static void si_blit(struct pipe_context *ctx,
+                      const struct pipe_blit_info *info)
+{
+       struct r600_context *rctx = (struct r600_context*)ctx;
+       struct r600_resource_texture *rsrc = (struct r600_resource_texture*)info->src.resource;
+
+       assert(util_blitter_is_blit_supported(rctx->blitter, info));
+
+       if (info->src.resource->nr_samples > 1 &&
+           info->dst.resource->nr_samples <= 1 &&
+           !util_format_is_depth_or_stencil(info->src.resource->format) &&
+           !util_format_is_pure_integer(info->src.resource->format)) {
+               debug_printf("radeonsi: color resolve is unimplemented\n");
+               return;
+       }
+
+       if (rsrc->is_depth && !rsrc->is_flushing_texture) {
+               si_blit_decompress_depth_in_place(rctx, rsrc,
+                                                 info->src.level, info->src.level,
+                                                 info->src.box.z,
+                                                 info->src.box.z + info->src.box.depth - 1);
+       }
+
+       r600_blitter_begin(ctx, R600_BLIT);
+       util_blitter_blit(rctx->blitter, info);
+       r600_blitter_end(ctx);
 }
 
 void si_init_blit_functions(struct r600_context *rctx)
@@ -380,20 +497,5 @@ void si_init_blit_functions(struct r600_context *rctx)
        rctx->context.clear_render_target = r600_clear_render_target;
        rctx->context.clear_depth_stencil = r600_clear_depth_stencil;
        rctx->context.resource_copy_region = r600_resource_copy_region;
-}
-
-void r600_blit_push_depth(struct pipe_context *ctx, struct r600_resource_texture *texture)
-{
-       struct pipe_box sbox;
-
-       sbox.x = sbox.y = sbox.z = 0;
-       sbox.width = texture->resource.b.b.width0;
-       sbox.height = texture->resource.b.b.height0;
-       /* XXX that might be wrong */
-       sbox.depth = 1;
-
-       r600_hw_copy_region(ctx, (struct pipe_resource *)texture, 0,
-                           0, 0, 0,
-                           (struct pipe_resource *)texture->flushed_depth_texture, 0,
-                           &sbox);
+       rctx->context.blit = si_blit;
 }