diff --git a/.pick_status.json b/.pick_status.json
index b0f0c1ebab3..eb2f3525388 100644
--- a/.pick_status.json
+++ b/.pick_status.json
@@ -1114,7 +1114,7 @@
         "description": "anv: Add missing L3 flushes",
         "nominated": true,
         "nomination_type": 1,
-        "resolution": 0,
+        "resolution": 1,
         "main_sha": null,
         "because_sha": null,
         "notes": null
diff --git a/src/intel/vulkan/genX_acceleration_structure.c b/src/intel/vulkan/genX_acceleration_structure.c
index 4c3a68fcef5..ffee6108547 100644
--- a/src/intel/vulkan/genX_acceleration_structure.c
+++ b/src/intel/vulkan/genX_acceleration_structure.c
@@ -429,6 +429,15 @@ anv_encode_as(VkCommandBuffer commandBuffer,
 
    anv_CmdPushConstants2KHR(commandBuffer, &push_info);
 
+   /* L1/L2 caches flushes should have been dealt with by pipeline barriers.
+    * Unfortunately some platforms require L3 flush because CS (reading the
+    * ir_internal_node_count paramters from vk_ir_header) is not L3 coherent.
+    */
+   if (!ANV_DEVINFO_HAS_COHERENT_L3_CS(cmd_buffer->device->info)) {
+      anv_add_pending_pipe_bits(cmd_buffer, ANV_PIPE_DATA_CACHE_FLUSH_BIT,
+                                "ir internal node count for dispatch");
+   }
+
    struct anv_address indirect_addr =
       anv_address_from_u64(intermediate_header_addr +
                             offsetof(struct vk_ir_header, ir_internal_node_count));
@@ -485,9 +494,6 @@ anv_init_header(VkCommandBuffer commandBuffer,
 
    VkDeviceAddress header_addr = vk_acceleration_structure_get_va(dst);
 
-   UNUSED size_t base = offsetof(struct anv_accel_struct_header,
-                                 copy_dispatch_size);
-
    uint32_t instance_count = geometry_type == VK_GEOMETRY_TYPE_INSTANCES_KHR ?
                              leaf_count : 0;
 
@@ -497,13 +503,6 @@ anv_init_header(VkCommandBuffer commandBuffer,
        */
       vk_barrier_compute_w_to_compute_r(commandBuffer);
 
-      /* VK_BUILD_ACCELERATION_STRUCTURE_ALLOW_COMPACTION_BIT_KHR is set, so we
-       * want to populate header.compacted_size with the compacted size, which
-       * needs to be calculated by using ir_header.dst_node_offset, which we'll
-       * access in the header.comp.
-       */
-      base = offsetof(struct anv_accel_struct_header, instance_count);
-
       VkPipeline pipeline;
       VkPipelineLayout layout;
       get_pipeline_spv(device, "header", header_spv, sizeof(header_spv),
@@ -530,6 +529,19 @@ anv_init_header(VkCommandBuffer commandBuffer,
    } else {
       vk_barrier_compute_w_to_host_r(commandBuffer);
 
+      /* L1/L2 caches flushes should have been dealt with by pipeline barriers.
+       * Unfortunately some platforms require L3 flush because CS (reading the
+       * dispatch size paramters) is not L3 coherent.
+       */
+      if (!ANV_DEVINFO_HAS_COHERENT_L3_CS(cmd_buffer->device->info)) {
+         anv_add_pending_pipe_bits(cmd_buffer, ANV_PIPE_DATA_CACHE_FLUSH_BIT,
+                                   "copy dispatch size for dispatch");
+         genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
+      }
+
+      size_t base = offsetof(struct anv_accel_struct_header,
+                             copy_dispatch_size);
+
       struct anv_accel_struct_header header = {};
 
       header.instance_count = instance_count;