tu: Split out viewport faking from per-view viewports

For VK_VALVE_fragment_density_map_layered, we need to split out whether
to enable per-view viewports and whether viewport 0 should be splatted
to all viewports. We also have to split this out for
VK_QCOM_multiview_per_view_viewport, where the splatting needs to be
disabled. To avoid conflicts between them, plumb through
"fake_single_viewport" which will be needed for the former extension and
needs to be modified for the latter extension.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35594>

src/freedreno/vulkan/tu_cmd_buffer.cc

@@ -4491,8 +4491,11 @@ tu_CmdBindPipeline(VkCommandBuffer commandBuffer,
          pipeline->shaders[MESA_SHADER_FRAGMENT]->fs.has_fdm;
    }
 
-   if (pipeline->program.per_view_viewport != cmd->state.per_view_viewport) {
+   if (pipeline->program.per_view_viewport != cmd->state.per_view_viewport ||
+       pipeline->program.fake_single_viewport != cmd->state.fake_single_viewport) {
       cmd->state.per_view_viewport = pipeline->program.per_view_viewport;
+      cmd->state.fake_single_viewport =
+         pipeline->program.fake_single_viewport;
       cmd->state.dirty |= TU_CMD_DIRTY_PER_VIEW_VIEWPORT;
    }
 

src/freedreno/vulkan/tu_cmd_buffer.h

@@ -289,6 +289,13 @@ struct tu_render_pass_state
    bool has_zpass_done_sample_count_write_in_rp;
    bool disable_gmem;
    bool sysmem_single_prim_mode;
+
+   /* This is set if, at any point in the render pass, we were not able to
+    * duplicate the viewport per-view due to the user using multiple viewports
+    * and instead we used the state from view 0 to transform each viewport. If
+    * this happens at any point then all views must contain the same FDM
+    * fragment size.
+    */
    bool shared_viewport;
 
    /* Track whether conditional predicate for COND_REG_EXEC is changed in draw_cs */
@@ -447,7 +454,10 @@ struct tu_cmd_state
    uint32_t max_vbs_bound;
 
    bool has_fdm;
+   /* See tu_pipeline::per_view_viewport */
    bool per_view_viewport;
+   /* See tu_pipeline::fake_single_viewport */
+   bool fake_single_viewport;
 
    /* saved states to re-emit in TU_CMD_DIRTY_DRAW_STATE case */
    struct tu_draw_state dynamic_state[TU_DYNAMIC_STATE_COUNT];

src/freedreno/vulkan/tu_pipeline.cc

@@ -2321,6 +2321,7 @@ tu_emit_program_state(struct tu_cs *sub_cs,
       !last_shader->writes_viewport &&
       shaders[MESA_SHADER_FRAGMENT]->fs.has_fdm &&
       dev->physical_device->info->a6xx.has_per_view_viewport;
+   prog->fake_single_viewport = prog->per_view_viewport;
    prog->writes_shading_rate = last_shader->writes_shading_rate;
    prog->reads_shading_rate = fs->reads_shading_rate;
    prog->accesses_smask = fs->reads_smask || fs->writes_smask;
@@ -2552,7 +2553,10 @@ tu6_emit_viewport(struct tu_cs *cs,
 struct apply_viewport_state {
    struct vk_viewport_state vp;
    struct vk_rasterization_state rs;
+   /* See tu_render_pass_state::shared_viewport */
    bool share_scale;
+   /* See tu_pipeline::fake_single_viewport */
+   bool fake_single_viewport;
 };
 
 /* It's a hardware restriction that the window offset (i.e. common_bin_offset)
@@ -2609,15 +2613,14 @@ fdm_apply_viewports(struct tu_cmd_buffer *cmd, struct tu_cs *cs, void *data,
        * same across all views, we can pick any view. However the number
        * of viewports and number of views is not guaranteed the same, so we
        * need to pick the 0'th view which always exists to be safe.
-       *
-       * Conversly, if we're not using shared scaling then the rasterizer in
-       * the original pipeline is using only the first viewport, so we need to
-       * replicate it across all viewports.
        */
       VkExtent2D frag_area = state->share_scale ? frag_areas[0] : frag_areas[i];
       VkRect2D bin = state->share_scale ? bins[0] : bins[i];
+      /* Implement fake_single_viewport by replicating viewport 0 across all
+       * views.
+       */
       VkViewport viewport =
-         state->share_scale ? state->vp.viewports[i] : state->vp.viewports[0];
+         state->fake_single_viewport ? state->vp.viewports[0] : state->vp.viewports[i];
       if (frag_area.width == 1 && frag_area.height == 1 &&
           common_bin_offset.x == bin.offset.x &&
           common_bin_offset.y == bin.offset.y) {
@@ -2653,6 +2656,7 @@ tu6_emit_viewport_fdm(struct tu_cs *cs, struct tu_cmd_buffer *cmd,
       .vp = *vp,
       .rs = *rs,
       .share_scale = !cmd->state.per_view_viewport,
+      .fake_single_viewport = cmd->state.fake_single_viewport,
    };
    if (!state.share_scale)
       state.vp.viewport_count = num_views;
@@ -2722,7 +2726,7 @@ fdm_apply_scissors(struct tu_cmd_buffer *cmd, struct tu_cs *cs, void *data,
       VkExtent2D frag_area = state->share_scale ? frag_areas[0] : frag_areas[i];
       VkRect2D bin = state->share_scale ? bins[0] : bins[i];
       VkRect2D scissor =
-         state->share_scale ? state->vp.scissors[i] : state->vp.scissors[0];
+         state->fake_single_viewport ? state->vp.scissors[0] : state->vp.scissors[i];
 
       /* Transform the scissor following the viewport. It's unclear how this
        * is supposed to handle cases where the scissor isn't aligned to the
@@ -2765,6 +2769,7 @@ tu6_emit_scissor_fdm(struct tu_cs *cs, struct tu_cmd_buffer *cmd,
    struct apply_viewport_state state = {
       .vp = *vp,
       .share_scale = !cmd->state.per_view_viewport,
+      .fake_single_viewport = cmd->state.fake_single_viewport,
    };
    if (!state.share_scale)
       state.vp.scissor_count = num_views;

src/freedreno/vulkan/tu_pipeline.h

@@ -109,7 +109,18 @@ struct tu_program_state
 
       unsigned dynamic_descriptor_offsets[MAX_SETS];
 
+      /* Whether the per-view-viewport feature should be enabled in HW. This
+       * implicitly adds gl_ViewIndex to gl_ViewportIndex so that from a HW
+       * point of view (but not necessarily the user's point of view!) there
+       * is a viewport per view.
+       */
       bool per_view_viewport;
+      /* If per_view_viewport is true and this is true, the app has provided
+       * a single viewport and we need to fake it by duplicating the viewport
+       * across views before transforming each viewport separately using FDM
+       * state.
+       */
+      bool fake_single_viewport;
       bool writes_shading_rate;
       bool reads_shading_rate;
       bool accesses_smask;