+static void
+emit_end_xfb_query(struct tu_cmd_buffer *cmdbuf,
+ struct tu_query_pool *pool,
+ uint32_t query,
+ uint32_t stream_id)
+{
+ struct tu_cs *cs = cmdbuf->state.pass ? &cmdbuf->draw_cs : &cmdbuf->cs;
+
+ uint64_t end_iova = primitive_query_iova(pool, query, end[0], 0);
+ uint64_t result_written_iova = query_result_iova(pool, query, 0);
+ uint64_t result_generated_iova = query_result_iova(pool, query, 1);
+ uint64_t begin_written_iova = primitive_query_iova(pool, query, begin[stream_id], 0);
+ uint64_t begin_generated_iova = primitive_query_iova(pool, query, begin[stream_id], 1);
+ uint64_t end_written_iova = primitive_query_iova(pool, query, end[stream_id], 0);
+ uint64_t end_generated_iova = primitive_query_iova(pool, query, end[stream_id], 1);
+ uint64_t available_iova = query_available_iova(pool, query);
+
+ tu_cs_emit_regs(cs, A6XX_VPC_SO_STREAM_COUNTS_LO(end_iova));
+ tu6_emit_event_write(cmdbuf, cs, WRITE_PRIMITIVE_COUNTS, false);
+
+ tu_cs_emit_wfi(cs);
+ tu6_emit_event_write(cmdbuf, cs, CACHE_FLUSH_TS, true);
+
+ /* Set the count of written primitives */
+ tu_cs_emit_pkt7(cs, CP_MEM_TO_MEM, 9);
+ tu_cs_emit(cs, CP_MEM_TO_MEM_0_DOUBLE | CP_MEM_TO_MEM_0_NEG_C |
+ CP_MEM_TO_MEM_0_WAIT_FOR_MEM_WRITES | 0x80000000);
+ tu_cs_emit_qw(cs, result_written_iova);
+ tu_cs_emit_qw(cs, result_written_iova);
+ tu_cs_emit_qw(cs, end_written_iova);
+ tu_cs_emit_qw(cs, begin_written_iova);
+
+ tu6_emit_event_write(cmdbuf, cs, CACHE_FLUSH_TS, true);
+
+ /* Set the count of generated primitives */
+ tu_cs_emit_pkt7(cs, CP_MEM_TO_MEM, 9);
+ tu_cs_emit(cs, CP_MEM_TO_MEM_0_DOUBLE | CP_MEM_TO_MEM_0_NEG_C |
+ CP_MEM_TO_MEM_0_WAIT_FOR_MEM_WRITES | 0x80000000);
+ tu_cs_emit_qw(cs, result_generated_iova);
+ tu_cs_emit_qw(cs, result_generated_iova);
+ tu_cs_emit_qw(cs, end_generated_iova);
+ tu_cs_emit_qw(cs, begin_generated_iova);
+
+ /* Set the availability to 1 */
+ tu_cs_emit_pkt7(cs, CP_MEM_WRITE, 4);
+ tu_cs_emit_qw(cs, available_iova);
+ tu_cs_emit_qw(cs, 0x1);
+}
+