radeonsi: tell LLVM not to remove s_barrier instructions
[mesa.git] / src / gallium / drivers / r600 / r600_blit.c
index edf972650f9a48932e557dc0c997cb393abc1f31..c52492e8c2d3b6a2e72a1d8508012d0acf471599 100644 (file)
@@ -275,10 +275,9 @@ void r600_decompress_depth_textures(struct r600_context *rctx,
                rview = (struct r600_pipe_sampler_view*)view;
 
                tex = (struct r600_texture *)view->texture;
-               assert(tex->is_depth && !tex->is_flushing_texture);
+               assert(tex->db_compatible);
 
-               if (rctx->b.chip_class >= EVERGREEN ||
-                   r600_can_read_depth(tex)) {
+               if (r600_can_sample_zs(tex, rview->is_stencil_sampler)) {
                        r600_blit_decompress_depth_in_place(rctx, tex,
                                                   rview->is_stencil_sampler,
                                                   view->u.tex.first_level, view->u.tex.last_level,
@@ -373,9 +372,8 @@ static bool r600_decompress_subresource(struct pipe_context *ctx,
        struct r600_context *rctx = (struct r600_context *)ctx;
        struct r600_texture *rtex = (struct r600_texture*)tex;
 
-       if (rtex->is_depth && !rtex->is_flushing_texture) {
-               if (rctx->b.chip_class >= EVERGREEN ||
-                   r600_can_read_depth(rtex)) {
+       if (rtex->db_compatible) {
+               if (r600_can_sample_zs(rtex, false)) {
                        r600_blit_decompress_depth_in_place(rctx, rtex, false,
                                                   level, level,
                                                   first_layer, last_layer);
@@ -475,11 +473,13 @@ static void r600_clear_render_target(struct pipe_context *ctx,
                                     struct pipe_surface *dst,
                                     const union pipe_color_union *color,
                                     unsigned dstx, unsigned dsty,
-                                    unsigned width, unsigned height)
+                                    unsigned width, unsigned height,
+                                    bool render_condition_enabled)
 {
        struct r600_context *rctx = (struct r600_context *)ctx;
 
-       r600_blitter_begin(ctx, R600_CLEAR_SURFACE);
+       r600_blitter_begin(ctx, R600_CLEAR_SURFACE |
+                          (render_condition_enabled ? 0 : R600_DISABLE_RENDER_COND));
        util_blitter_clear_render_target(rctx->blitter, dst, color,
                                         dstx, dsty, width, height);
        r600_blitter_end(ctx);
@@ -491,11 +491,13 @@ static void r600_clear_depth_stencil(struct pipe_context *ctx,
                                     double depth,
                                     unsigned stencil,
                                     unsigned dstx, unsigned dsty,
-                                    unsigned width, unsigned height)
+                                    unsigned width, unsigned height,
+                                    bool render_condition_enabled)
 {
        struct r600_context *rctx = (struct r600_context *)ctx;
 
-       r600_blitter_begin(ctx, R600_CLEAR_SURFACE);
+       r600_blitter_begin(ctx, R600_CLEAR_SURFACE |
+                          (render_condition_enabled ? 0 : R600_DISABLE_RENDER_COND));
        util_blitter_clear_depth_stencil(rctx->blitter, dst, clear_flags, depth, stencil,
                                         dstx, dsty, width, height);
        r600_blitter_end(ctx);
@@ -724,7 +726,10 @@ void r600_resource_copy_region(struct pipe_context *ctx,
                }
        }
 
-       dst_view = r600_create_surface_custom(ctx, dst, &dst_templ, dst_width, dst_height);
+       dst_view = r600_create_surface_custom(ctx, dst, &dst_templ,
+                                             /* we don't care about these two for r600g */
+                                             dst->width0, dst->height0,
+                                             dst_width, dst_height);
 
        if (rctx->b.chip_class >= EVERGREEN) {
                src_view = evergreen_create_sampler_view_custom(ctx, src, &src_templ,
@@ -790,7 +795,7 @@ static bool do_hardware_msaa_resolve(struct pipe_context *ctx,
            info->src.box.width == dst_width &&
            info->src.box.height == dst_height &&
            info->src.box.depth == 1 &&
-           dst->surface.level[info->dst.level].mode >= RADEON_SURF_MODE_1D &&
+           dst->surface.u.legacy.level[info->dst.level].mode >= RADEON_SURF_MODE_1D &&
            (!dst->cmask.size || !dst->dirty_level_mask) /* dst cannot be fast-cleared */) {
                r600_blitter_begin(ctx, R600_COLOR_RESOLVE |
                                   (info->render_condition_enable ? 0 : R600_DISABLE_RENDER_COND));
@@ -848,11 +853,30 @@ static void r600_blit(struct pipe_context *ctx,
                       const struct pipe_blit_info *info)
 {
        struct r600_context *rctx = (struct r600_context*)ctx;
+       struct r600_texture *rdst = (struct r600_texture *)info->dst.resource;
 
        if (do_hardware_msaa_resolve(ctx, info)) {
                return;
        }
 
+       /* Using SDMA for copying to a linear texture in GTT is much faster.
+        * This improves DRI PRIME performance.
+        *
+        * resource_copy_region can't do this yet, because dma_copy calls it
+        * on failure (recursion).
+        */
+       if (rdst->surface.u.legacy.level[info->dst.level].mode ==
+           RADEON_SURF_MODE_LINEAR_ALIGNED &&
+           rctx->b.dma_copy &&
+           util_can_blit_via_copy_region(info, false)) {
+               rctx->b.dma_copy(ctx, info->dst.resource, info->dst.level,
+                                info->dst.box.x, info->dst.box.y,
+                                info->dst.box.z,
+                                info->src.resource, info->src.level,
+                                &info->src.box);
+               return;
+       }
+
        assert(util_blitter_is_blit_supported(rctx->blitter, info));
 
        /* The driver doesn't decompress resources automatically while