turnip: Implement vkCmdEndQuery for occlusion queries

src/freedreno/vulkan/tu_query.c

 /* Returns the IOVA of a given uint64_t field in a given slot of a query
  * pool.
  */
-static uint64_t query_address(struct tu_query_pool *pool, uint32_t query,
+static uint64_t query_iova(struct tu_query_pool *pool, uint32_t query,
       uint32_t index) {
-   return (pool->bo->iova) + (pool->stride * query) +
-         (index * sizeof(uint64_t));
+   return pool->bo->iova + pool->stride * query + index * sizeof(uint64_t);
 }
 
 VkResult
    tu_cs_emit(cs, A6XX_RB_SAMPLE_COUNT_CONTROL_COPY);
 
    tu_cs_emit_pkt4(cs, REG_A6XX_RB_SAMPLE_COUNT_ADDR_LO, 2);
-   tu_cs_emit_qw(cs, query_address(pool, query, 1 /* begin */));
+   tu_cs_emit_qw(cs, query_iova(pool, query, 1 /* begin */));
 
    tu_cs_emit_pkt7(cs, CP_EVENT_WRITE, 1);
    tu_cs_emit(cs, ZPASS_DONE);
    tu_bo_list_add(&cmdbuf->bo_list, pool->bo, MSM_SUBMIT_BO_WRITE);
 }
 
+static void
+emit_end_occlusion_query(struct tu_cmd_buffer *cmdbuf,
+                         struct tu_cs *cs,
+                         struct tu_query_pool *pool,
+                         uint32_t query) {
+   /* Ending an occlusion query happens in a few steps:
+    *    1) Set the query->end to UINT64_MAX.
+    *    2) Set up the SAMPLE_COUNT registers and trigger a CP_EVENT_WRITE to
+    *       write the current sample count value into query->end.
+    *    3) Since (2) is asynchronous, wait until query->end is not equal to
+    *       UINT64_MAX to set the query's availability bit via CP_WAIT_REG_MEM.
+    */
+   uint64_t end_iova = query_iova(pool, query, 2 /* end */);
+   tu_cs_reserve_space(cmdbuf->device, cs, 25);
+   tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 4);
+   tu_cs_emit_qw(cs, end_iova);
+   tu_cs_emit_qw(cs, 0xffffffffffffffffull);
+   tu_cs_emit_pkt7(cs, CP_WAIT_MEM_WRITES, 0);
+
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_SAMPLE_COUNT_CONTROL, 1);
+   tu_cs_emit(cs, A6XX_RB_SAMPLE_COUNT_CONTROL_COPY);
+   tu_cs_emit_pkt4(cs, REG_A6XX_RB_SAMPLE_COUNT_ADDR_LO, 2);
+   tu_cs_emit_qw(cs, end_iova);
+   tu_cs_emit_pkt7(cs, CP_EVENT_WRITE, 1);
+   tu_cs_emit(cs, ZPASS_DONE);
+
+   tu_cs_emit_pkt7(cs, CP_WAIT_REG_MEM, 6);
+   tu_cs_emit(cs, 0x00000014); // XXX
+   tu_cs_emit_qw(cs, end_iova);
+   tu_cs_emit_qw(cs, 0xffffffffffffffffull);
+   tu_cs_emit(cs, 0x00000010); // XXX
+   tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 4);
+   tu_cs_emit_qw(cs, query_iova(pool, query, 0 /* available */));
+   tu_cs_emit_qw(cs, 0x1);
+}
+
 void
 tu_CmdEndQuery(VkCommandBuffer commandBuffer,
                VkQueryPool queryPool,
                uint32_t query)
 {
+   TU_FROM_HANDLE(tu_cmd_buffer, cmdbuf, commandBuffer);
+   TU_FROM_HANDLE(tu_query_pool, pool, queryPool);
+   struct tu_cs *cs = &cmdbuf->cs;
+   assert(query < pool->size);
+
+   switch (pool->type) {
+      case VK_QUERY_TYPE_OCCLUSION:
+         emit_end_occlusion_query(cmdbuf, cs, pool, query);
+         break;
+      case VK_QUERY_TYPE_PIPELINE_STATISTICS:
+      case VK_QUERY_TYPE_TIMESTAMP:
+         unreachable("Unimplemented query type");
+      default:
+         assert(!"Invalid query type");
+   }
 }
 
 void