anv: predicate emission of STATE_BASE_ADDRESS

Completely skip the stall & programming if the bindless address has
not changed. Only on Gfx12.5+ since previous generations also program
the binding table pool base address through STATE_BASE_ADDRESS.

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: José Roberto de Souza <jose.souza@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/29595>

src/intel/vulkan/anv_private.h

@@ -223,6 +223,12 @@ struct intel_perf_query_result;
  */
 #define ANV_PERF_QUERY_OFFSET_REG 0x2670 /* MI_ALU_REG14 */
 
+/* We reserve this MI ALU register to hold the last programmed bindless
+ * surface state base address so that we can predicate STATE_BASE_ADDRESS
+ * emissions if the address doesn't change.
+ */
+#define ANV_BINDLESS_SURFACE_BASE_ADDR_REG 0x2668 /* MI_ALU_REG13 */
+
 #define ANV_GRAPHICS_SHADER_STAGE_COUNT (MESA_SHADER_MESH + 1)
 
 /* RENDER_SURFACE_STATE is a bit smaller (48b) but since it is aligned to 64

src/intel/vulkan/genX_cmd_buffer.c

@@ -239,13 +239,21 @@ genX(cmd_buffer_emit_state_base_address)(struct anv_cmd_buffer *cmd_buffer)
        anv_cmd_buffer_is_video_queue(cmd_buffer))
       return;
 
+   struct anv_device *device = cmd_buffer->device;
+
    struct GENX(STATE_BASE_ADDRESS) sba = {};
    fill_state_base_addr(cmd_buffer, &sba);
 
-   /* If we are emitting a new state base address we probably need to re-emit
-    * binding tables.
-    */
-   cmd_buffer->state.descriptors_dirty |= ~0;
+#if GFX_VERx10 >= 125
+   struct mi_builder b;
+   mi_builder_init(&b, device->info, &cmd_buffer->batch);
+   mi_builder_set_mocs(&b, isl_mocs(&device->isl_dev, 0, false));
+   struct mi_goto_target t = MI_GOTO_TARGET_INIT;
+   mi_goto_if(&b,
+              mi_ieq(&b, mi_reg64(ANV_BINDLESS_SURFACE_BASE_ADDR_REG),
+                         mi_imm(sba.BindlessSurfaceStateBaseAddress.offset)),
+              &t);
+#endif
 
    /* Emit a render target cache flush.
     *
@@ -294,10 +302,6 @@ genX(cmd_buffer_emit_state_base_address)(struct anv_cmd_buffer *cmd_buffer)
       genX(flush_pipeline_select)(cmd_buffer, gfx12_wa_pipeline);
 #endif
 
-#if GFX_VERx10 >= 125
-   genX(cmd_buffer_emit_bt_pool_base_address)(cmd_buffer);
-#endif
-
    /* After re-setting the surface state base address, we have to do some
     * cache flushing so that the sampler engine will pick up the new
     * SURFACE_STATE objects and binding tables. From the Broadwell PRM,
@@ -400,6 +404,23 @@ genX(cmd_buffer_emit_state_base_address)(struct anv_cmd_buffer *cmd_buffer)
       cmd_buffer->state.compute.base.push_constants_data_dirty = true;
 #endif
    }
+
+#if GFX_VERx10 >= 125
+   assert(sba.BindlessSurfaceStateBaseAddress.offset != 0);
+   mi_store(&b, mi_reg64(ANV_BINDLESS_SURFACE_BASE_ADDR_REG),
+                mi_imm(sba.BindlessSurfaceStateBaseAddress.offset));
+
+   mi_goto_target(&b, &t);
+#endif
+
+#if GFX_VERx10 >= 125
+   genX(cmd_buffer_emit_bt_pool_base_address)(cmd_buffer);
+#endif
+
+   /* If we have emitted a new state base address we probably need to re-emit
+    * binding tables.
+    */
+   cmd_buffer->state.descriptors_dirty |= ~0;
 }
 
 void

src/intel/vulkan/genX_init_state.c

@@ -35,6 +35,8 @@
 #include "grl/genX_grl.h"
 #endif
 
+#include "genX_mi_builder.h"
+
 #include "vk_util.h"
 #include "vk_format.h"
 
@@ -309,7 +311,13 @@ init_common_queue_state(struct anv_queue *queue, struct anv_batch *batch)
       sba.L1CacheControl = L1CC_WB;
 #endif
    }
-#endif
+
+   struct mi_builder b;
+   mi_builder_init(&b, device->info, batch);
+
+   mi_store(&b, mi_reg64(ANV_BINDLESS_SURFACE_BASE_ADDR_REG),
+                mi_imm(device->physical->va.internal_surface_state_pool.addr));
+#endif /* GFX_VER >= 12 */
 
 #if GFX_VERx10 >= 125
    if (ANV_SUPPORT_RT && device->info->has_ray_tracing) {

src/intel/vulkan/genX_mi_builder.h

@@ -8,10 +8,11 @@
 #include "genxml/genX_pack.h"
 
 /* We reserve :
+ *    - GPR 13 for STATE_BASE_ADDRESS bindless surface base address
  *    - GPR 14 for perf queries
  *    - GPR 15 for conditional rendering
  */
-#define MI_BUILDER_NUM_ALLOC_GPRS 14
+#define MI_BUILDER_NUM_ALLOC_GPRS 13
 #define MI_BUILDER_CAN_WRITE_BATCH true
 /* Don't do any write check by default, we manually set it where it matters.
  */