vc4: Add a blitter path using just the render thread.

This accelerates the path for generating the shadow tiled texture when
asked to sample from a raster texture (typical in glamor).

diff --git a/src/gallium/drivers/vc4/vc4_blit.c b/src/gallium/drivers/vc4/vc4_blit.c
index 5c98fb612e0d6688cf78d78e03b8d14bc6ef9894..4f87189fd2e29d5817278fd5806d988fde7c9da5 100644 (file)
--- a/src/gallium/drivers/vc4/vc4_blit.c
+++ b/src/gallium/drivers/vc4/vc4_blit.c
@@ -26,6 +26,130 @@
 #include "util/u_blitter.h"
 #include "vc4_context.h"
 
+static void
+vc4_tile_blit_color_rcl(struct vc4_context *vc4,
+                        struct vc4_surface *dst_surf,
+                        struct vc4_surface *src_surf)
+{
+        struct vc4_resource *src = vc4_resource(src_surf->base.texture);
+        struct vc4_resource *dst = vc4_resource(dst_surf->base.texture);
+
+        uint32_t min_x_tile = 0;
+        uint32_t min_y_tile = 0;
+        uint32_t max_x_tile = (dst_surf->base.width - 1) / 64;
+        uint32_t max_y_tile = (dst_surf->base.height - 1) / 64;
+        uint32_t xtiles = max_x_tile - min_x_tile + 1;
+        uint32_t ytiles = max_y_tile - min_y_tile + 1;
+        uint32_t reloc_size = 9;
+        uint32_t config_size = 11 + reloc_size;
+        uint32_t loadstore_size = 7 + reloc_size;
+        uint32_t tilecoords_size = 3;
+        cl_ensure_space(&vc4->rcl,
+                        config_size +
+                        xtiles * ytiles * (loadstore_size * 2 +
+                                           tilecoords_size * 1));
+        cl_ensure_space(&vc4->bo_handles, 2 * sizeof(uint32_t));
+        cl_ensure_space(&vc4->bo_pointers, 2 * sizeof(struct vc4_bo *));
+
+        cl_start_reloc(&vc4->rcl, 1);
+        cl_u8(&vc4->rcl, VC4_PACKET_TILE_RENDERING_MODE_CONFIG);
+        cl_reloc(vc4, &vc4->rcl, dst->bo, dst_surf->offset);
+        cl_u16(&vc4->rcl, dst_surf->base.width);
+        cl_u16(&vc4->rcl, dst_surf->base.height);
+        cl_u16(&vc4->rcl, ((dst_surf->tiling <<
+                            VC4_RENDER_CONFIG_MEMORY_FORMAT_SHIFT) |
+                           (vc4_rt_format_is_565(dst_surf->base.format) ?
+                            VC4_RENDER_CONFIG_FORMAT_BGR565 :
+                            VC4_RENDER_CONFIG_FORMAT_RGBA8888)));
+
+        uint32_t src_hindex = vc4_gem_hindex(vc4, src->bo);
+
+        for (int y = min_y_tile; y <= max_y_tile; y++) {
+                for (int x = min_x_tile; x <= max_x_tile; x++) {
+                        bool end_of_frame = (x == max_x_tile &&
+                                             y == max_y_tile);
+
+                        cl_start_reloc(&vc4->rcl, 1);
+                        cl_u8(&vc4->rcl, VC4_PACKET_LOAD_TILE_BUFFER_GENERAL);
+                        cl_u8(&vc4->rcl,
+                              VC4_LOADSTORE_TILE_BUFFER_COLOR |
+                              (src_surf->tiling <<
+                               VC4_LOADSTORE_TILE_BUFFER_FORMAT_SHIFT));
+                        cl_u8(&vc4->rcl,
+                              vc4_rt_format_is_565(src_surf->base.format) ?
+                              VC4_LOADSTORE_TILE_BUFFER_BGR565 :
+                              VC4_LOADSTORE_TILE_BUFFER_RGBA8888);
+                        cl_reloc_hindex(&vc4->rcl, src_hindex,
+                                        src_surf->offset);
+
+                        cl_u8(&vc4->rcl, VC4_PACKET_TILE_COORDINATES);
+                        cl_u8(&vc4->rcl, x);
+                        cl_u8(&vc4->rcl, y);
+
+                        if (end_of_frame) {
+                                cl_u8(&vc4->rcl,
+                                      VC4_PACKET_STORE_MS_TILE_BUFFER_AND_EOF);
+                        } else {
+                                cl_u8(&vc4->rcl,
+                                      VC4_PACKET_STORE_MS_TILE_BUFFER);
+                        }
+                }
+        }
+
+        vc4->draw_min_x = 0;
+        vc4->draw_min_y = 0;
+        vc4->draw_max_x = dst_surf->base.width;
+        vc4->draw_max_y = dst_surf->base.height;
+
+        dst->writes++;
+        vc4->needs_flush = true;
+}
+
+static struct vc4_surface *
+vc4_get_blit_surface(struct pipe_context *pctx,
+                     struct pipe_resource *prsc, unsigned level)
+{
+        struct pipe_surface tmpl;
+
+        memset(&tmpl, 0, sizeof(tmpl));
+        tmpl.format = prsc->format;
+        tmpl.u.tex.level = level;
+        tmpl.u.tex.first_layer = 0;
+        tmpl.u.tex.last_layer = 0;
+
+        return vc4_surface(pctx->create_surface(pctx, prsc, &tmpl));
+}
+
+static bool
+vc4_tile_blit(struct pipe_context *pctx, const struct pipe_blit_info *info)
+{
+        struct vc4_context *vc4 = vc4_context(pctx);
+
+        if ((info->mask & PIPE_MASK_RGBA) == 0)
+                return false;
+
+        if (info->dst.box.x != 0 || info->dst.box.y != 0 ||
+            info->src.box.x != 0 || info->src.box.y != 0 ||
+            info->dst.box.width != info->src.box.width ||
+            info->dst.box.height != info->src.box.height) {
+                return false;
+        }
+
+        struct vc4_surface *dst_surf =
+                vc4_get_blit_surface(pctx, info->dst.resource, info->dst.level);
+        struct vc4_surface *src_surf =
+                vc4_get_blit_surface(pctx, info->src.resource, info->src.level);
+
+        vc4_flush(pctx);
+        vc4_tile_blit_color_rcl(vc4, dst_surf, src_surf);
+        vc4_job_submit(vc4);
+
+        pctx->surface_destroy(pctx, &dst_surf->base);
+        pctx->surface_destroy(pctx, &src_surf->base);
+
+        return true;
+}
+
 static bool
 vc4_render_blit(struct pipe_context *ctx, struct pipe_blit_info *info)
 {
@@ -77,6 +201,9 @@ vc4_blit(struct pipe_context *pctx, const struct pipe_blit_info *blit_info)
                 return;
         }
 
+        if (vc4_tile_blit(pctx, blit_info))
+                return;
+
         if (util_try_blit_via_copy_region(pctx, &info)) {
                 return; /* done */
         }