pvr: Implement vkCmdBeginQuery API.

Signed-off-by: Rajnesh Kanwal <rajnesh.kanwal@imgtec.com>
Reviewed-by: Frank Binns <frank.binns@imgtec.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/19371>

src/imagination/vulkan/pvr_cmd_buffer.c

@@ -139,6 +139,8 @@ static void pvr_cmd_buffer_free_resources(struct pvr_cmd_buffer *cmd_buffer)
    vk_free(&cmd_buffer->vk.pool->alloc,
            cmd_buffer->state.render_pass_info.clear_values);
 
+   util_dynarray_fini(&cmd_buffer->state.query_indices);
+
    pvr_cmd_buffer_free_sub_cmds(cmd_buffer);
 
    list_for_each_entry_safe (struct pvr_bo, bo, &cmd_buffer->bo_list, link) {
@@ -1421,7 +1423,7 @@ pvr_compute_generate_fence(struct pvr_cmd_buffer *cmd_buffer,
    pvr_compute_generate_control_stream(csb, sub_cmd, &info);
 }
 
-static VkResult pvr_cmd_buffer_end_sub_cmd(struct pvr_cmd_buffer *cmd_buffer)
+VkResult pvr_cmd_buffer_end_sub_cmd(struct pvr_cmd_buffer *cmd_buffer)
 {
    struct pvr_cmd_buffer_state *state = &cmd_buffer->state;
    struct pvr_sub_cmd *sub_cmd = state->current_sub_cmd;
@@ -1595,8 +1597,8 @@ pvr_cmd_uses_deferred_cs_cmds(const struct pvr_cmd_buffer *const cmd_buffer)
              deferred_control_stream_flags;
 }
 
-static VkResult pvr_cmd_buffer_start_sub_cmd(struct pvr_cmd_buffer *cmd_buffer,
+VkResult pvr_cmd_buffer_start_sub_cmd(struct pvr_cmd_buffer *cmd_buffer,
-                                             enum pvr_sub_cmd_type type)
+                                      enum pvr_sub_cmd_type type)
 {
    struct pvr_cmd_buffer_state *state = &cmd_buffer->state;
    struct pvr_device *device = cmd_buffer->device;
@@ -2472,8 +2474,12 @@ VkResult pvr_BeginCommandBuffer(VkCommandBuffer commandBuffer,
          pvr_cmd_buffer_start_sub_cmd(cmd_buffer, PVR_SUB_CMD_TYPE_GRAPHICS);
       if (result != VK_SUCCESS)
          return result;
+
+      state->vis_test_enabled = inheritance_info->occlusionQueryEnable;
    }
 
+   util_dynarray_init(&state->query_indices, NULL);
+
    memset(state->barriers_needed,
           0xFF,
           sizeof(*state->barriers_needed) * ARRAY_SIZE(state->barriers_needed));
@@ -3604,6 +3610,11 @@ pvr_setup_isp_faces_and_control(struct pvr_cmd_buffer *const cmd_buffer,
       ispctl.dbenable = !rasterizer_discard &&
                         dynamic_state->rs.depth_bias.enable &&
                         obj_type == PVRX(TA_OBJTYPE_TRIANGLE);
+      if (!rasterizer_discard && cmd_buffer->state.vis_test_enabled) {
+         ispctl.vistest = true;
+         ispctl.visreg = cmd_buffer->state.vis_reg;
+      }
+
       ispctl.scenable = !rasterizer_discard;
 
       ppp_state->isp.control_struct = ispctl;
@@ -4526,7 +4537,7 @@ pvr_emit_dirty_ppp_state(struct pvr_cmd_buffer *const cmd_buffer,
                           MESA_VK_DYNAMIC_DS_STENCIL_WRITE_MASK) ||
               BITSET_TEST(dynamic_state->dirty,
                           MESA_VK_DYNAMIC_RS_LINE_WIDTH) ||
-              state->dirty.isp_userpass) {
+              state->dirty.isp_userpass || state->dirty.vis_test) {
       pvr_setup_isp_faces_and_control(cmd_buffer, NULL);
    }
 
@@ -4829,6 +4840,8 @@ static VkResult pvr_validate_draw_state(struct pvr_cmd_buffer *cmd_buffer)
    sub_cmd->frag_uses_texture_rw |= fragment_state->uses_texture_rw;
    sub_cmd->vertex_uses_texture_rw |= vertex_state->uses_texture_rw;
 
+   sub_cmd->job.get_vis_results = state->vis_test_enabled;
+
    fstencil_keep =
       (dynamic_state->ds.stencil.front.op.fail == VK_STENCIL_OP_KEEP) &&
       (dynamic_state->ds.stencil.front.op.pass == VK_STENCIL_OP_KEEP);
@@ -4937,6 +4950,7 @@ static VkResult pvr_validate_draw_state(struct pvr_cmd_buffer *cmd_buffer)
    state->dirty.gfx_pipeline_binding = false;
    state->dirty.isp_userpass = false;
    state->dirty.vertex_bindings = false;
+   state->dirty.vis_test = false;
 
    return VK_SUCCESS;
 }
@@ -5701,6 +5715,9 @@ pvr_execute_graphics_cmd_buffer(struct pvr_cmd_buffer *cmd_buffer,
    struct pvr_sub_cmd *primary_sub_cmd = state->current_sub_cmd;
    VkResult result;
 
+   /* Inherited queries are not supported. */
+   assert(!state->vis_test_enabled);
+
    if (list_is_empty(&sec_cmd_buffer->sub_cmds))
       return;
 
@@ -6431,6 +6448,11 @@ VkResult pvr_EndCommandBuffer(VkCommandBuffer commandBuffer)
    if (state->status != VK_SUCCESS)
       return state->status;
 
+   /* TODO: We should be freeing all the resources, allocated for recording,
+    * here.
+    */
+   util_dynarray_fini(&state->query_indices);
+
    result = pvr_cmd_buffer_end_sub_cmd(cmd_buffer);
    if (result != VK_SUCCESS)
       return result;

src/imagination/vulkan/pvr_job_render.c

@@ -1623,6 +1623,9 @@ pvr_render_job_ws_fragment_state_init(struct pvr_render_ctx *ctx,
 
    if (job->frag_uses_atomic_ops)
       state->flags |= PVR_WINSYS_FRAG_FLAG_SINGLE_CORE;
+
+   if (job->get_vis_results)
+      state->flags |= PVR_WINSYS_FRAG_FLAG_GET_VIS_RESULTS;
 }
 
 static void pvr_render_job_ws_submit_info_init(

src/imagination/vulkan/pvr_job_render.h

@@ -51,6 +51,7 @@ struct pvr_render_job {
    bool disable_compute_overlap;
    bool enable_bg_tag;
    bool process_empty_tiles;
+   bool get_vis_results;
 
    uint32_t pds_pixel_event_data_offset;
 

src/imagination/vulkan/pvr_private.h

@@ -684,6 +684,11 @@ struct pvr_sub_cmd_gfx {
    bool modifies_depth;
    bool modifies_stencil;
 
+   bool barrier_store;
+   bool barrier_load;
+
+   const struct pvr_query_pool *query_pool;
+
    /* Control stream builder object */
    struct pvr_csb control_stream;
 
@@ -961,6 +966,8 @@ struct pvr_cmd_buffer_state {
        */
       bool draw_base_instance : 1;
       bool draw_variant : 1;
+
+      bool vis_test;
    } dirty;
 
    struct pvr_cmd_buffer_draw_state draw_state;
@@ -970,6 +977,12 @@ struct pvr_cmd_buffer_state {
       const struct pvr_pds_info *info;
    } pds_shader;
 
+   const struct pvr_query_pool *query_pool;
+   bool vis_test_enabled;
+   uint32_t vis_reg;
+
+   struct util_dynarray query_indices;
+
    uint32_t max_shared_regs;
 
    /* Address of data segment for vertex attrib upload program. */
@@ -1554,6 +1567,10 @@ VkResult pvr_device_tile_buffer_ensure_cap(struct pvr_device *device,
                                            uint32_t capacity,
                                            uint32_t size_in_bytes);
 
+VkResult pvr_cmd_buffer_start_sub_cmd(struct pvr_cmd_buffer *cmd_buffer,
+                                      enum pvr_sub_cmd_type type);
+VkResult pvr_cmd_buffer_end_sub_cmd(struct pvr_cmd_buffer *cmd_buffer);
+
 void pvr_compute_update_shared_private(
    struct pvr_cmd_buffer *cmd_buffer,
    struct pvr_sub_cmd_compute *const sub_cmd,

src/imagination/vulkan/pvr_query.c

@@ -230,7 +230,51 @@ void pvr_CmdBeginQuery(VkCommandBuffer commandBuffer,
                        uint32_t query,
                        VkQueryControlFlags flags)
 {
-   assert(!"Unimplemented");
+   PVR_FROM_HANDLE(pvr_cmd_buffer, cmd_buffer, commandBuffer);
+   struct pvr_cmd_buffer_state *state = &cmd_buffer->state;
+   PVR_FROM_HANDLE(pvr_query_pool, pool, queryPool);
+
+   PVR_CHECK_COMMAND_BUFFER_BUILDING_STATE(cmd_buffer);
+
+   /* Occlusion queries can't be nested. */
+   assert(!state->vis_test_enabled);
+
+   if (state->current_sub_cmd) {
+      assert(state->current_sub_cmd->type == PVR_SUB_CMD_TYPE_GRAPHICS);
+
+      if (!state->current_sub_cmd->gfx.query_pool) {
+         state->current_sub_cmd->gfx.query_pool = pool;
+      } else if (state->current_sub_cmd->gfx.query_pool != pool) {
+         VkResult result;
+
+         /* Kick render. */
+         state->current_sub_cmd->gfx.barrier_store = true;
+
+         result = pvr_cmd_buffer_end_sub_cmd(cmd_buffer);
+         if (result != VK_SUCCESS)
+            return;
+
+         result =
+            pvr_cmd_buffer_start_sub_cmd(cmd_buffer, PVR_SUB_CMD_TYPE_GRAPHICS);
+         if (result != VK_SUCCESS)
+            return;
+
+         /* Use existing render setup, but load color attachments from HW
+          * BGOBJ.
+          */
+         state->current_sub_cmd->gfx.barrier_load = true;
+         state->current_sub_cmd->gfx.barrier_store = false;
+         state->current_sub_cmd->gfx.query_pool = pool;
+      }
+   }
+
+   state->query_pool = pool;
+   state->vis_test_enabled = true;
+   state->vis_reg = query;
+   state->dirty.vis_test = true;
+
+   /* Add the index to the list for this render. */
+   util_dynarray_append(&state->query_indices, __typeof__(query), query);
 }
 
 void pvr_CmdEndQuery(VkCommandBuffer commandBuffer,

src/imagination/vulkan/winsys/pvr_winsys.h

@@ -338,6 +338,7 @@ struct pvr_winsys_compute_submit_info {
 #define PVR_WINSYS_FRAG_FLAG_STENCIL_BUFFER_PRESENT BITFIELD_BIT(1U)
 #define PVR_WINSYS_FRAG_FLAG_PREVENT_CDM_OVERLAP BITFIELD_BIT(2U)
 #define PVR_WINSYS_FRAG_FLAG_SINGLE_CORE BITFIELD_BIT(3U)
+#define PVR_WINSYS_FRAG_FLAG_GET_VIS_RESULTS BITFIELD_BIT(4U)
 
 struct pvr_winsys_render_submit_info {
    struct pvr_winsys_rt_dataset *rt_dataset;

src/imagination/vulkan/winsys/pvrsrvkm/pvr_srv_job_render.c

@@ -659,6 +659,9 @@ static void pvr_srv_fragment_cmd_init(
 
    if (state->flags & PVR_WINSYS_FRAG_FLAG_SINGLE_CORE)
       cmd->flags |= ROGUE_FWIF_RENDERFLAGS_SINGLE_CORE;
+
+   if (state->flags & PVR_WINSYS_FRAG_FLAG_GET_VIS_RESULTS)
+      cmd->flags |= ROGUE_FWIF_RENDERFLAGS_GETVISRESULTS;
 }
 
 VkResult pvr_srv_winsys_render_submit(