diff --git a/src/freedreno/vulkan/tu_cmd_buffer.c b/src/freedreno/vulkan/tu_cmd_buffer.c
index 1a4c3d7fd7d..99ad052a1bc 100644
--- a/src/freedreno/vulkan/tu_cmd_buffer.c
+++ b/src/freedreno/vulkan/tu_cmd_buffer.c
@@ -1670,27 +1670,14 @@ tu_create_cmd_buffer(struct tu_device *device,
    if (result != VK_SUCCESS)
       goto fail_scratch_bo;
 
-#define VSC_DATA_SIZE(pitch)  ((pitch) * 32 + 0x100)  /* extra size to store VSC_SIZE */
-#define VSC_DATA2_SIZE(pitch) ((pitch) * 32)
-
-   /* TODO: resize on overflow or compute a max size from # of vertices in renderpass?? */
-   cmd_buffer->vsc_data_pitch = 0x440 * 4;
-   cmd_buffer->vsc_data2_pitch = 0x1040 * 4;
-
-   result = tu_bo_init_new(device, &cmd_buffer->vsc_data, VSC_DATA_SIZE(cmd_buffer->vsc_data_pitch));
-   if (result != VK_SUCCESS)
-      goto fail_vsc_data;
-
-   result = tu_bo_init_new(device, &cmd_buffer->vsc_data2, VSC_DATA2_SIZE(cmd_buffer->vsc_data2_pitch));
-   if (result != VK_SUCCESS)
-      goto fail_vsc_data2;
+   /* TODO: resize on overflow */
+   cmd_buffer->vsc_data_pitch = device->vsc_data_pitch;
+   cmd_buffer->vsc_data2_pitch = device->vsc_data2_pitch;
+   cmd_buffer->vsc_data = device->vsc_data;
+   cmd_buffer->vsc_data2 = device->vsc_data2;
 
    return VK_SUCCESS;
 
-fail_vsc_data2:
-   tu_bo_finish(cmd_buffer->device, &cmd_buffer->vsc_data);
-fail_vsc_data:
-   tu_bo_finish(cmd_buffer->device, &cmd_buffer->scratch_bo);
 fail_scratch_bo:
    list_del(&cmd_buffer->pool_link);
    return result;
@@ -1700,8 +1687,6 @@ static void
 tu_cmd_buffer_destroy(struct tu_cmd_buffer *cmd_buffer)
 {
    tu_bo_finish(cmd_buffer->device, &cmd_buffer->scratch_bo);
-   tu_bo_finish(cmd_buffer->device, &cmd_buffer->vsc_data);
-   tu_bo_finish(cmd_buffer->device, &cmd_buffer->vsc_data2);
 
    list_del(&cmd_buffer->pool_link);
 
diff --git a/src/freedreno/vulkan/tu_device.c b/src/freedreno/vulkan/tu_device.c
index df846dac63b..b68f0f4540f 100644
--- a/src/freedreno/vulkan/tu_device.c
+++ b/src/freedreno/vulkan/tu_device.c
@@ -1074,7 +1074,7 @@ tu_CreateDevice(VkPhysicalDevice physicalDevice,
          8, VK_SYSTEM_ALLOCATION_SCOPE_DEVICE);
       if (!device->queues[qfi]) {
          result = VK_ERROR_OUT_OF_HOST_MEMORY;
-         goto fail;
+         goto fail_queues;
       }
 
       memset(device->queues[qfi], 0,
@@ -1086,13 +1086,27 @@ tu_CreateDevice(VkPhysicalDevice physicalDevice,
          result = tu_queue_init(device, &device->queues[qfi][q], qfi, q,
                                 queue_create->flags);
          if (result != VK_SUCCESS)
-            goto fail;
+            goto fail_queues;
       }
    }
 
    device->compiler = ir3_compiler_create(NULL, physical_device->gpu_id);
    if (!device->compiler)
-      goto fail;
+      goto fail_queues;
+
+#define VSC_DATA_SIZE(pitch)  ((pitch) * 32 + 0x100)  /* extra size to store VSC_SIZE */
+#define VSC_DATA2_SIZE(pitch) ((pitch) * 32)
+
+   device->vsc_data_pitch = 0x440 * 4;
+   device->vsc_data2_pitch = 0x1040 * 4;
+
+   result = tu_bo_init_new(device, &device->vsc_data, VSC_DATA_SIZE(device->vsc_data_pitch));
+   if (result != VK_SUCCESS)
+      goto fail_vsc_data;
+
+   result = tu_bo_init_new(device, &device->vsc_data2, VSC_DATA2_SIZE(device->vsc_data2_pitch));
+   if (result != VK_SUCCESS)
+      goto fail_vsc_data2;
 
    VkPipelineCacheCreateInfo ci;
    ci.sType = VK_STRUCTURE_TYPE_PIPELINE_CACHE_CREATE_INFO;
@@ -1104,14 +1118,23 @@ tu_CreateDevice(VkPhysicalDevice physicalDevice,
    result =
       tu_CreatePipelineCache(tu_device_to_handle(device), &ci, NULL, &pc);
    if (result != VK_SUCCESS)
-      goto fail;
+      goto fail_pipeline_cache;
 
    device->mem_cache = tu_pipeline_cache_from_handle(pc);
 
    *pDevice = tu_device_to_handle(device);
    return VK_SUCCESS;
 
-fail:
+fail_pipeline_cache:
+   tu_bo_finish(device, &device->vsc_data2);
+
+fail_vsc_data2:
+   tu_bo_finish(device, &device->vsc_data);
+
+fail_vsc_data:
+   ralloc_free(device->compiler);
+
+fail_queues:
    for (unsigned i = 0; i < TU_MAX_QUEUE_FAMILIES; i++) {
       for (unsigned q = 0; q < device->queue_count[i]; q++)
          tu_queue_finish(&device->queues[i][q]);
@@ -1119,9 +1142,6 @@ fail:
          vk_free(&device->alloc, device->queues[i]);
    }
 
-   if (device->compiler)
-      ralloc_free(device->compiler);
-
    vk_free(&device->alloc, device);
    return result;
 }
@@ -1134,6 +1154,9 @@ tu_DestroyDevice(VkDevice _device, const VkAllocationCallbacks *pAllocator)
    if (!device)
       return;
 
+   tu_bo_finish(device, &device->vsc_data);
+   tu_bo_finish(device, &device->vsc_data2);
+
    for (unsigned i = 0; i < TU_MAX_QUEUE_FAMILIES; i++) {
       for (unsigned q = 0; q < device->queue_count[i]; q++)
          tu_queue_finish(&device->queues[i][q]);
diff --git a/src/freedreno/vulkan/tu_private.h b/src/freedreno/vulkan/tu_private.h
index 3a1bb331b9d..e0e76bd179e 100644
--- a/src/freedreno/vulkan/tu_private.h
+++ b/src/freedreno/vulkan/tu_private.h
@@ -453,6 +453,14 @@ struct tu_queue
    struct tu_fence submit_fence;
 };
 
+struct tu_bo
+{
+   uint32_t gem_handle;
+   uint64_t size;
+   uint64_t iova;
+   void *map;
+};
+
 struct tu_device
 {
    VK_LOADER_DATA _loader_data;
@@ -473,20 +481,17 @@ struct tu_device
    /* Backup in-memory cache to be used if the app doesn't provide one */
    struct tu_pipeline_cache *mem_cache;
 
+   struct tu_bo vsc_data;
+   struct tu_bo vsc_data2;
+   uint32_t vsc_data_pitch;
+   uint32_t vsc_data2_pitch;
+
    struct list_head shader_slabs;
    mtx_t shader_slab_mutex;
 
    struct tu_device_extension_table enabled_extensions;
 };
 
-struct tu_bo
-{
-   uint32_t gem_handle;
-   uint64_t size;
-   uint64_t iova;
-   void *map;
-};
-
 VkResult
 tu_bo_init_new(struct tu_device *dev, struct tu_bo *bo, uint64_t size);
 VkResult