turnip: Implement vkCmdBeginQuery for occlusion queries

src/freedreno/vulkan/tu_query.c

@@ -31,14 +31,29 @@
 #include <string.h>
 #include <unistd.h>
 
+#include "registers/adreno_pm4.xml.h"
+#include "registers/adreno_common.xml.h"
+#include "registers/a6xx.xml.h"
+
 #include "nir/nir_builder.h"
 
+#include "tu_cs.h"
+
 struct PACKED occlusion_query_slot {
    uint64_t available; /* 0 when unavailable, 1 when available */
    uint64_t begin;
    uint64_t end;
 };
 
+/* 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,
+      uint32_t index) {
+   return (pool->bo->iova) + (pool->stride * query) +
+         (index * sizeof(uint64_t));
+}
+
 VkResult
 tu_CreateQueryPool(VkDevice _device,
                    const VkQueryPoolCreateInfo *pCreateInfo,
@@ -142,12 +157,49 @@ tu_CmdResetQueryPool(VkCommandBuffer commandBuffer,
 {
 }
 
+static void
+emit_begin_occlusion_query(struct tu_cmd_buffer *cmdbuf,
+                           struct tu_cs *cs,
+                           struct tu_query_pool *pool,
+                           uint32_t query) {
+   tu_cs_reserve_space(cmdbuf->device, cs, 7);
+   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, query_address(pool, query, 1 /* begin */));
+
+   tu_cs_emit_pkt7(cs, CP_EVENT_WRITE, 1);
+   tu_cs_emit(cs, ZPASS_DONE);
+}
+
 void
 tu_CmdBeginQuery(VkCommandBuffer commandBuffer,
                  VkQueryPool queryPool,
                  uint32_t query,
                  VkQueryControlFlags flags)
 {
+   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:
+         /* In freedreno, there is no implementation difference between
+          * GL_SAMPLES_PASSED and GL_ANY_SAMPLES_PASSED, so we can similarly
+          * ignore the VK_QUERY_CONTROL_PRECISE_BIT flag here.
+          */
+         emit_begin_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");
+   }
+
+   tu_bo_list_add(&cmdbuf->bo_list, pool->bo, MSM_SUBMIT_BO_WRITE);
 }
 
 void