anv/allocator: Set the BO flags in bo_cache_alloc/import
[mesa.git] / src / intel / vulkan / anv_private.h
index c6fab614a90961da45856d0a9784dbd5f7319d24..b47420f89ecf88b4562b5cb18f52d38c7d5aaabd 100644 (file)
 #endif
 
 #include "common/gen_clflush.h"
-#include "common/gen_device_info.h"
+#include "common/gen_gem.h"
+#include "dev/gen_device_info.h"
 #include "blorp/blorp.h"
 #include "compiler/brw_compiler.h"
 #include "util/macros.h"
+#include "util/hash_table.h"
 #include "util/list.h"
+#include "util/set.h"
 #include "util/u_atomic.h"
 #include "util/u_vector.h"
+#include "util/vma.h"
 #include "vk_alloc.h"
 #include "vk_debug_report.h"
 
@@ -79,6 +83,55 @@ struct gen_l3_config;
 #include "common/intel_log.h"
 #include "wsi_common.h"
 
+/* anv Virtual Memory Layout
+ * =========================
+ *
+ * When the anv driver is determining the virtual graphics addresses of memory
+ * objects itself using the softpin mechanism, the following memory ranges
+ * will be used.
+ *
+ * Three special considerations to notice:
+ *
+ * (1) the dynamic state pool is located within the same 4 GiB as the low
+ * heap. This is to work around a VF cache issue described in a comment in
+ * anv_physical_device_init_heaps.
+ *
+ * (2) the binding table pool is located at lower addresses than the surface
+ * state pool, within a 4 GiB range. This allows surface state base addresses
+ * to cover both binding tables (16 bit offsets) and surface states (32 bit
+ * offsets).
+ *
+ * (3) the last 4 GiB of the address space is withheld from the high
+ * heap. Various hardware units will read past the end of an object for
+ * various reasons. This healthy margin prevents reads from wrapping around
+ * 48-bit addresses.
+ */
+#define LOW_HEAP_MIN_ADDRESS               0x000000001000ULL /* 4 KiB */
+#define LOW_HEAP_MAX_ADDRESS               0x0000bfffffffULL
+#define DYNAMIC_STATE_POOL_MIN_ADDRESS     0x0000c0000000ULL /* 3 GiB */
+#define DYNAMIC_STATE_POOL_MAX_ADDRESS     0x0000ffffffffULL
+#define BINDING_TABLE_POOL_MIN_ADDRESS     0x000100000000ULL /* 4 GiB */
+#define BINDING_TABLE_POOL_MAX_ADDRESS     0x00013fffffffULL
+#define SURFACE_STATE_POOL_MIN_ADDRESS     0x000140000000ULL /* 5 GiB */
+#define SURFACE_STATE_POOL_MAX_ADDRESS     0x00017fffffffULL
+#define INSTRUCTION_STATE_POOL_MIN_ADDRESS 0x000180000000ULL /* 6 GiB */
+#define INSTRUCTION_STATE_POOL_MAX_ADDRESS 0x0001bfffffffULL
+#define HIGH_HEAP_MIN_ADDRESS              0x0001c0000000ULL /* 7 GiB */
+#define HIGH_HEAP_MAX_ADDRESS              0xfffeffffffffULL
+
+#define LOW_HEAP_SIZE               \
+   (LOW_HEAP_MAX_ADDRESS - LOW_HEAP_MIN_ADDRESS + 1)
+#define HIGH_HEAP_SIZE              \
+   (HIGH_HEAP_MAX_ADDRESS - HIGH_HEAP_MIN_ADDRESS + 1)
+#define DYNAMIC_STATE_POOL_SIZE     \
+   (DYNAMIC_STATE_POOL_MAX_ADDRESS - DYNAMIC_STATE_POOL_MIN_ADDRESS + 1)
+#define BINDING_TABLE_POOL_SIZE     \
+   (BINDING_TABLE_POOL_MAX_ADDRESS - BINDING_TABLE_POOL_MIN_ADDRESS + 1)
+#define SURFACE_STATE_POOL_SIZE     \
+   (SURFACE_STATE_POOL_MAX_ADDRESS - SURFACE_STATE_POOL_MIN_ADDRESS + 1)
+#define INSTRUCTION_STATE_POOL_SIZE \
+   (INSTRUCTION_STATE_POOL_MAX_ADDRESS - INSTRUCTION_STATE_POOL_MIN_ADDRESS + 1)
+
 /* Allowing different clear colors requires us to perform a depth resolve at
  * the end of certain render passes. This is because while slow clears store
  * the clear color in the HiZ buffer, fast clears (without a resolve) don't.
@@ -104,6 +157,25 @@ struct gen_l3_config;
 #define MAX_IMAGES 8
 #define MAX_PUSH_DESCRIPTORS 32 /* Minimum requirement */
 
+/* The kernel relocation API has a limitation of a 32-bit delta value
+ * applied to the address before it is written which, in spite of it being
+ * unsigned, is treated as signed .  Because of the way that this maps to
+ * the Vulkan API, we cannot handle an offset into a buffer that does not
+ * fit into a signed 32 bits.  The only mechanism we have for dealing with
+ * this at the moment is to limit all VkDeviceMemory objects to a maximum
+ * of 2GB each.  The Vulkan spec allows us to do this:
+ *
+ *    "Some platforms may have a limit on the maximum size of a single
+ *    allocation. For example, certain systems may fail to create
+ *    allocations with a size greater than or equal to 4GB. Such a limit is
+ *    implementation-dependent, and if such a failure occurs then the error
+ *    VK_ERROR_OUT_OF_DEVICE_MEMORY should be returned."
+ *
+ * We don't use vk_error here because it's not an error so much as an
+ * indication to the application that the allocation is too large.
+ */
+#define MAX_MEMORY_ALLOCATION_SIZE (1ull << 31)
+
 #define ANV_SVGS_VB_INDEX    MAX_VBS
 #define ANV_DRAWID_VB_INDEX (MAX_VBS + 1)
 
@@ -547,6 +619,12 @@ struct anv_block_pool {
 
    struct anv_bo bo;
 
+   /* The address where the start of the pool is pinned. The various bos that
+    * are created as the pool grows will have addresses in the range
+    * [start_address, start_address + BLOCK_POOL_MEMFD_SIZE).
+    */
+   uint64_t start_address;
+
    /* The offset from the start of the bo to the "center" of the block
     * pool.  Pointers to allocated blocks are given by
     * bo.map + center_bo_offset + offsets.
@@ -643,6 +721,7 @@ struct anv_state_stream {
  */
 VkResult anv_block_pool_init(struct anv_block_pool *pool,
                              struct anv_device *device,
+                             uint64_t start_address,
                              uint32_t initial_size,
                              uint64_t bo_flags);
 void anv_block_pool_finish(struct anv_block_pool *pool);
@@ -653,6 +732,7 @@ int32_t anv_block_pool_alloc_back(struct anv_block_pool *pool,
 
 VkResult anv_state_pool_init(struct anv_state_pool *pool,
                              struct anv_device *device,
+                             uint64_t start_address,
                              uint32_t block_size,
                              uint64_t bo_flags);
 void anv_state_pool_finish(struct anv_state_pool *pool);
@@ -715,10 +795,12 @@ VkResult anv_bo_cache_init(struct anv_bo_cache *cache);
 void anv_bo_cache_finish(struct anv_bo_cache *cache);
 VkResult anv_bo_cache_alloc(struct anv_device *device,
                             struct anv_bo_cache *cache,
-                            uint64_t size, struct anv_bo **bo);
+                            uint64_t size, uint64_t bo_flags,
+                            struct anv_bo **bo);
 VkResult anv_bo_cache_import(struct anv_device *device,
                              struct anv_bo_cache *cache,
-                             int fd, struct anv_bo **bo);
+                             int fd, uint64_t bo_flags,
+                             struct anv_bo **bo);
 VkResult anv_bo_cache_export(struct anv_device *device,
                              struct anv_bo_cache *cache,
                              struct anv_bo *bo_in, int *fd_out);
@@ -749,6 +831,7 @@ struct anv_physical_device {
 
     struct anv_instance *                       instance;
     uint32_t                                    chipset_id;
+    bool                                        no_hw;
     char                                        path[20];
     const char *                                name;
     struct gen_device_info                      info;
@@ -769,6 +852,8 @@ struct anv_physical_device {
     bool                                        has_exec_fence;
     bool                                        has_syncobj;
     bool                                        has_syncobj_wait;
+    bool                                        has_context_priority;
+    bool                                        use_softpin;
 
     struct anv_device_extension_table           supported_extensions;
 
@@ -817,7 +902,7 @@ struct anv_queue {
 
     struct anv_device *                         device;
 
-    struct anv_state_pool *                     pool;
+    VkDeviceQueueCreateFlags                    flags;
 };
 
 struct anv_pipeline_cache {
@@ -852,6 +937,7 @@ struct anv_device {
 
     struct anv_instance *                       instance;
     uint32_t                                    chipset_id;
+    bool                                        no_hw;
     struct gen_device_info                      info;
     struct isl_device                           isl_dev;
     int                                         context_id;
@@ -861,16 +947,24 @@ struct anv_device {
     struct anv_device_extension_table           enabled_extensions;
     struct anv_dispatch_table                   dispatch;
 
+    pthread_mutex_t                             vma_mutex;
+    struct util_vma_heap                        vma_lo;
+    struct util_vma_heap                        vma_hi;
+    uint64_t                                    vma_lo_available;
+    uint64_t                                    vma_hi_available;
+
     struct anv_bo_pool                          batch_bo_pool;
 
     struct anv_bo_cache                         bo_cache;
 
     struct anv_state_pool                       dynamic_state_pool;
     struct anv_state_pool                       instruction_state_pool;
+    struct anv_state_pool                       binding_table_pool;
     struct anv_state_pool                       surface_state_pool;
 
     struct anv_bo                               workaround_bo;
     struct anv_bo                               trivial_batch_bo;
+    struct anv_bo                               hiz_clear_bo;
 
     struct anv_pipeline_cache                   blorp_shader_cache;
     struct blorp_context                        blorp;
@@ -888,6 +982,29 @@ struct anv_device {
     bool                                        lost;
 };
 
+static inline struct anv_state_pool *
+anv_binding_table_pool(struct anv_device *device)
+{
+   if (device->instance->physicalDevice.use_softpin)
+      return &device->binding_table_pool;
+   else
+      return &device->surface_state_pool;
+}
+
+static inline struct anv_state
+anv_binding_table_pool_alloc(struct anv_device *device) {
+   if (device->instance->physicalDevice.use_softpin)
+      return anv_state_pool_alloc(&device->binding_table_pool,
+                                  device->binding_table_pool.block_size, 0);
+   else
+      return anv_state_pool_alloc_back(&device->surface_state_pool);
+}
+
+static inline void
+anv_binding_table_pool_free(struct anv_device *device, struct anv_state state) {
+   anv_state_pool_free(anv_binding_table_pool(device), state);
+}
+
 static void inline
 anv_state_flush(struct anv_device *device, struct anv_state state)
 {
@@ -921,14 +1038,16 @@ int anv_gem_execbuffer(struct anv_device *device,
 int anv_gem_set_tiling(struct anv_device *device, uint32_t gem_handle,
                        uint32_t stride, uint32_t tiling);
 int anv_gem_create_context(struct anv_device *device);
+bool anv_gem_has_context_priority(int fd);
 int anv_gem_destroy_context(struct anv_device *device, int context);
+int anv_gem_set_context_param(int fd, int context, uint32_t param,
+                              uint64_t value);
 int anv_gem_get_context_param(int fd, int context, uint32_t param,
                               uint64_t *value);
 int anv_gem_get_param(int fd, uint32_t param);
 int anv_gem_get_tiling(struct anv_device *device, uint32_t gem_handle);
 bool anv_gem_get_bit6_swizzle(int fd, uint32_t tiling);
 int anv_gem_get_aperture(int fd, uint64_t *size);
-bool anv_gem_supports_48b_addresses(int fd);
 int anv_gem_gpu_get_reset_stats(struct anv_device *device,
                                 uint32_t *active, uint32_t *pending);
 int anv_gem_handle_to_fd(struct anv_device *device, uint32_t gem_handle);
@@ -951,6 +1070,9 @@ int anv_gem_syncobj_wait(struct anv_device *device,
                          uint32_t *handles, uint32_t num_handles,
                          int64_t abs_timeout_ns, bool wait_all);
 
+bool anv_vma_alloc(struct anv_device *device, struct anv_bo *bo);
+void anv_vma_free(struct anv_device *device, struct anv_bo *bo);
+
 VkResult anv_bo_init_new(struct anv_bo *bo, struct anv_device *device, uint64_t size);
 
 struct anv_reloc_list {
@@ -958,6 +1080,7 @@ struct anv_reloc_list {
    uint32_t                                     array_length;
    struct drm_i915_gem_relocation_entry *       relocs;
    struct anv_bo **                             reloc_bos;
+   struct set *                                 deps;
 };
 
 VkResult anv_reloc_list_init(struct anv_reloc_list *list,
@@ -1034,6 +1157,46 @@ struct anv_address {
    uint32_t offset;
 };
 
+#define ANV_NULL_ADDRESS ((struct anv_address) { NULL, 0 })
+
+static inline bool
+anv_address_is_null(struct anv_address addr)
+{
+   return addr.bo == NULL && addr.offset == 0;
+}
+
+static inline uint64_t
+anv_address_physical(struct anv_address addr)
+{
+   if (addr.bo && (addr.bo->flags & EXEC_OBJECT_PINNED))
+      return gen_canonical_address(addr.bo->offset + addr.offset);
+   else
+      return gen_canonical_address(addr.offset);
+}
+
+static inline struct anv_address
+anv_address_add(struct anv_address addr, uint64_t offset)
+{
+   addr.offset += offset;
+   return addr;
+}
+
+static inline void
+write_reloc(const struct anv_device *device, void *p, uint64_t v, bool flush)
+{
+   unsigned reloc_size = 0;
+   if (device->info.gen >= 8) {
+      reloc_size = sizeof(uint64_t);
+      *(uint64_t *)p = gen_canonical_address(v);
+   } else {
+      reloc_size = sizeof(uint32_t);
+      *(uint32_t *)p = v;
+   }
+
+   if (flush && !device->info.has_llc)
+      gen_flush_range(p, reloc_size);
+}
+
 static inline uint64_t
 _anv_combine_address(struct anv_batch *batch, void *location,
                      const struct anv_address address, uint32_t delta)
@@ -1148,6 +1311,17 @@ _anv_combine_address(struct anv_batch *batch, void *location,
       .IndextoMOCSTables                           = 1  \
    }
 
+/* Ice Lake MOCS defines are duplicates of Skylake MOCS defines. */
+#define GEN11_MOCS (struct GEN11_MEMORY_OBJECT_CONTROL_STATE) {  \
+      /* TC=LLC/eLLC, LeCC=WB, LRUM=3, L3CC=WB */              \
+      .IndextoMOCSTables                           = 2         \
+   }
+
+#define GEN11_MOCS_PTE {                                 \
+      /* TC=LLC/eLLC, LeCC=WB, LRUM=3, L3CC=WB */       \
+      .IndextoMOCSTables                           = 1  \
+   }
+
 struct anv_device_memory {
    struct anv_bo *                              bo;
    struct anv_memory_type *                     type;
@@ -1199,6 +1373,9 @@ struct anv_descriptor_set_binding_layout {
 };
 
 struct anv_descriptor_set_layout {
+   /* Descriptor set layouts can be destroyed at almost any time */
+   uint32_t ref_cnt;
+
    /* Number of bindings in this descriptor set */
    uint16_t binding_count;
 
@@ -1218,6 +1395,22 @@ struct anv_descriptor_set_layout {
    struct anv_descriptor_set_binding_layout binding[0];
 };
 
+static inline void
+anv_descriptor_set_layout_ref(struct anv_descriptor_set_layout *layout)
+{
+   assert(layout && layout->ref_cnt >= 1);
+   p_atomic_inc(&layout->ref_cnt);
+}
+
+static inline void
+anv_descriptor_set_layout_unref(struct anv_device *device,
+                                struct anv_descriptor_set_layout *layout)
+{
+   assert(layout && layout->ref_cnt >= 1);
+   if (p_atomic_dec_zero(&layout->ref_cnt))
+      vk_free(&device->alloc, layout);
+}
+
 struct anv_descriptor {
    VkDescriptorType type;
 
@@ -1239,7 +1432,7 @@ struct anv_descriptor {
 };
 
 struct anv_descriptor_set {
-   const struct anv_descriptor_set_layout *layout;
+   struct anv_descriptor_set_layout *layout;
    uint32_t size;
    uint32_t buffer_count;
    struct anv_buffer_view *buffer_views;
@@ -1248,10 +1441,10 @@ struct anv_descriptor_set {
 
 struct anv_buffer_view {
    enum isl_format format; /**< VkBufferViewCreateInfo::format */
-   struct anv_bo *bo;
-   uint32_t offset; /**< Offset into bo. */
    uint64_t range; /**< VkBufferViewCreateInfo::range */
 
+   struct anv_address address;
+
    struct anv_state surface_state;
    struct anv_state storage_surface_state;
    struct anv_state writeonly_storage_surface_state;
@@ -1363,7 +1556,7 @@ anv_descriptor_set_write_template(struct anv_descriptor_set *set,
 VkResult
 anv_descriptor_set_create(struct anv_device *device,
                           struct anv_descriptor_pool *pool,
-                          const struct anv_descriptor_set_layout *layout,
+                          struct anv_descriptor_set_layout *layout,
                           struct anv_descriptor_set **out_set);
 
 void
@@ -1418,8 +1611,7 @@ struct anv_buffer {
    VkBufferUsageFlags                           usage;
 
    /* Set when bound */
-   struct anv_bo *                              bo;
-   VkDeviceSize                                 offset;
+   struct anv_address                           address;
 };
 
 static inline uint64_t
@@ -1551,13 +1743,13 @@ anv_pipe_invalidate_bits_for_access_flags(VkAccessFlags flags)
 
 #define VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV (         \
    VK_IMAGE_ASPECT_COLOR_BIT | \
-   VK_IMAGE_ASPECT_PLANE_0_BIT_KHR | \
-   VK_IMAGE_ASPECT_PLANE_1_BIT_KHR | \
-   VK_IMAGE_ASPECT_PLANE_2_BIT_KHR)
+   VK_IMAGE_ASPECT_PLANE_0_BIT | \
+   VK_IMAGE_ASPECT_PLANE_1_BIT | \
+   VK_IMAGE_ASPECT_PLANE_2_BIT)
 #define VK_IMAGE_ASPECT_PLANES_BITS_ANV ( \
-   VK_IMAGE_ASPECT_PLANE_0_BIT_KHR | \
-   VK_IMAGE_ASPECT_PLANE_1_BIT_KHR | \
-   VK_IMAGE_ASPECT_PLANE_2_BIT_KHR)
+   VK_IMAGE_ASPECT_PLANE_0_BIT | \
+   VK_IMAGE_ASPECT_PLANE_1_BIT | \
+   VK_IMAGE_ASPECT_PLANE_2_BIT)
 
 struct anv_vertex_binding {
    struct anv_buffer *                          buffer;
@@ -1577,6 +1769,9 @@ struct anv_push_constants {
    /* Push constant data provided by the client through vkPushConstants */
    uint8_t client_data[MAX_PUSH_CONSTANTS_SIZE];
 
+   /* Used for vkCmdDispatchBase */
+   uint32_t base_work_group_id[3];
+
    /* Image data for image_load_store on pre-SKL */
    struct brw_image_param images[MAX_IMAGES];
 };
@@ -1635,15 +1830,20 @@ struct anv_surface_state {
     *
     * This address is relative to the start of the BO.
     */
-   uint64_t address;
+   struct anv_address address;
    /* Address of the aux surface, if any
     *
-    * This field is 0 if and only if no aux surface exists.
+    * This field is ANV_NULL_ADDRESS if and only if no aux surface exists.
     *
-    * This address is relative to the start of the BO.  On gen7, the bottom 12
-    * bits of this address include extra aux information.
+    * With the exception of gen8, the bottom 12 bits of this address' offset
+    * include extra aux information.
     */
-   uint64_t aux_address;
+   struct anv_address aux_address;
+   /* Address of the clear color, if any
+    *
+    * This address is relative to the start of the BO.
+    */
+   struct anv_address clear_address;
 };
 
 /**
@@ -1659,10 +1859,19 @@ struct anv_attachment_state {
 
    VkImageLayout                                current_layout;
    VkImageAspectFlags                           pending_clear_aspects;
+   VkImageAspectFlags                           pending_load_aspects;
    bool                                         fast_clear;
    VkClearValue                                 clear_value;
    bool                                         clear_color_is_zero_one;
    bool                                         clear_color_is_zero;
+
+   /* When multiview is active, attachments with a renderpass clear
+    * operation have their respective layers cleared on the first
+    * subpass that uses them, and only in that subpass. We keep track
+    * of this using a bitfield to indicate which layers of an attachment
+    * have not been cleared yet when multiview is active.
+    */
+   uint32_t                                     pending_clear_views;
 };
 
 /** State tracking for particular pipeline bind point
@@ -1675,6 +1884,7 @@ struct anv_attachment_state {
  */
 struct anv_cmd_pipeline_state {
    struct anv_pipeline *pipeline;
+   struct anv_pipeline_layout *layout;
 
    struct anv_descriptor_set *descriptors[MAX_SETS];
    uint32_t dynamic_offsets[MAX_DYNAMIC_BUFFERS];
@@ -1696,6 +1906,12 @@ struct anv_cmd_graphics_state {
    uint32_t vb_dirty;
 
    struct anv_dynamic_state dynamic;
+
+   struct {
+      struct anv_buffer *index_buffer;
+      uint32_t index_type; /**< 3DSTATE_INDEX_BUFFER.IndexFormat */
+      uint32_t index_offset;
+   } gen7;
 };
 
 /** State tracking for compute pipeline
@@ -1769,12 +1985,6 @@ struct anv_cmd_state {
     * is one of the states in render_pass_states.
     */
    struct anv_state                             null_surface_state;
-
-   struct {
-      struct anv_buffer *                       index_buffer;
-      uint32_t                                  index_type; /**< 3DSTATE_INDEX_BUFFER.IndexFormat */
-      uint32_t                                  index_offset;
-   } gen7;
 };
 
 struct anv_cmd_pool {
@@ -1903,7 +2113,6 @@ anv_cmd_buffer_push_constants(struct anv_cmd_buffer *cmd_buffer,
 struct anv_state
 anv_cmd_buffer_cs_push_constants(struct anv_cmd_buffer *cmd_buffer);
 
-void anv_cmd_buffer_clear_subpass(struct anv_cmd_buffer *cmd_buffer);
 void anv_cmd_buffer_resolve_subpass(struct anv_cmd_buffer *cmd_buffer);
 
 const struct anv_image_view *
@@ -2124,7 +2333,6 @@ struct anv_pipeline {
    struct anv_dynamic_state                     dynamic_state;
 
    struct anv_subpass *                         subpass;
-   struct anv_pipeline_layout *                 layout;
 
    bool                                         needs_data_cache;
 
@@ -2251,15 +2459,15 @@ anv_image_aspect_to_plane(VkImageAspectFlags image_aspects,
    switch (aspect_mask) {
    case VK_IMAGE_ASPECT_COLOR_BIT:
    case VK_IMAGE_ASPECT_DEPTH_BIT:
-   case VK_IMAGE_ASPECT_PLANE_0_BIT_KHR:
+   case VK_IMAGE_ASPECT_PLANE_0_BIT:
       return 0;
    case VK_IMAGE_ASPECT_STENCIL_BIT:
       if ((image_aspects & VK_IMAGE_ASPECT_DEPTH_BIT) == 0)
          return 0;
       /* Fall-through */
-   case VK_IMAGE_ASPECT_PLANE_1_BIT_KHR:
+   case VK_IMAGE_ASPECT_PLANE_1_BIT:
       return 1;
-   case VK_IMAGE_ASPECT_PLANE_2_BIT_KHR:
+   case VK_IMAGE_ASPECT_PLANE_2_BIT:
       return 2;
    default:
       /* Purposefully assert with depth/stencil aspects. */
@@ -2275,11 +2483,15 @@ anv_image_aspect_get_planes(VkImageAspectFlags aspect_mask)
    if (aspect_mask & (VK_IMAGE_ASPECT_COLOR_BIT |
                       VK_IMAGE_ASPECT_DEPTH_BIT |
                       VK_IMAGE_ASPECT_STENCIL_BIT |
-                      VK_IMAGE_ASPECT_PLANE_0_BIT_KHR))
+                      VK_IMAGE_ASPECT_PLANE_0_BIT))
       planes++;
-   if (aspect_mask & VK_IMAGE_ASPECT_PLANE_1_BIT_KHR)
+   if (aspect_mask & VK_IMAGE_ASPECT_PLANE_1_BIT)
       planes++;
-   if (aspect_mask & VK_IMAGE_ASPECT_PLANE_2_BIT_KHR)
+   if (aspect_mask & VK_IMAGE_ASPECT_PLANE_2_BIT)
+      planes++;
+
+   if ((aspect_mask & VK_IMAGE_ASPECT_DEPTH_BIT) != 0 &&
+       (aspect_mask & VK_IMAGE_ASPECT_STENCIL_BIT) != 0)
       planes++;
 
    return planes;
@@ -2291,7 +2503,7 @@ anv_plane_to_aspect(VkImageAspectFlags image_aspects,
 {
    if (image_aspects & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV) {
       if (_mesa_bitcount(image_aspects) > 1)
-         return VK_IMAGE_ASPECT_PLANE_0_BIT_KHR << plane;
+         return VK_IMAGE_ASPECT_PLANE_0_BIT << plane;
       return VK_IMAGE_ASPECT_COLOR_BIT;
    }
    if (image_aspects & VK_IMAGE_ASPECT_DEPTH_BIT)
@@ -2373,8 +2585,18 @@ struct anv_image {
    VkImageUsageFlags usage; /**< Superset of VkImageCreateInfo::usage. */
    VkImageTiling tiling; /** VkImageCreateInfo::tiling */
 
+   /** True if this is needs to be bound to an appropriately tiled BO.
+    *
+    * When not using modifiers, consumers such as X11, Wayland, and KMS need
+    * the tiling passed via I915_GEM_SET_TILING.  When exporting these buffers
+    * we require a dedicated allocation so that we can know to allocate a
+    * tiled buffer.
+    */
+   bool needs_set_tiling;
+
    /**
-    * DRM format modifier for this image or DRM_FORMAT_MOD_INVALID.
+    * Must be DRM_FORMAT_MOD_INVALID unless tiling is
+    * VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT.
     */
    uint64_t drm_format_mod;
 
@@ -2461,8 +2683,7 @@ struct anv_image {
       /**
        * BO associated with this plane, set when bound.
        */
-      struct anv_bo *bo;
-      VkDeviceSize bo_offset;
+      struct anv_address address;
 
       /**
        * When destroying the image, also free the bo.
@@ -2471,6 +2692,16 @@ struct anv_image {
    } planes[3];
 };
 
+/* The ordering of this enum is important */
+enum anv_fast_clear_type {
+   /** Image does not have/support any fast-clear blocks */
+   ANV_FAST_CLEAR_NONE = 0,
+   /** Image has/supports fast-clear but only to the default value */
+   ANV_FAST_CLEAR_DEFAULT_VALUE = 1,
+   /** Image has/supports fast-clear with an arbitrary fast-clear value */
+   ANV_FAST_CLEAR_ANY = 2,
+};
+
 /* Returns the number of auxiliary buffer levels attached to an image. */
 static inline uint8_t
 anv_image_aux_levels(const struct anv_image * const image,
@@ -2504,47 +2735,56 @@ anv_image_aux_layers(const struct anv_image * const image,
    }
 }
 
-static inline unsigned
-anv_fast_clear_state_entry_size(const struct anv_device *device)
+static inline struct anv_address
+anv_image_get_clear_color_addr(const struct anv_device *device,
+                               const struct anv_image *image,
+                               VkImageAspectFlagBits aspect)
 {
-   assert(device);
-   /* Entry contents:
-    *   +--------------------------------------------+
-    *   | clear value dword(s) | needs resolve dword |
-    *   +--------------------------------------------+
-    */
+   assert(image->aspects & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV);
 
-   /* Ensure that the needs resolve dword is in fact dword-aligned to enable
-    * GPU memcpy operations.
-    */
-   assert(device->isl_dev.ss.clear_value_size % 4 == 0);
-   return device->isl_dev.ss.clear_value_size + 4;
+   uint32_t plane = anv_image_aspect_to_plane(image->aspects, aspect);
+   return anv_address_add(image->planes[plane].address,
+                          image->planes[plane].fast_clear_state_offset);
 }
 
 static inline struct anv_address
-anv_image_get_clear_color_addr(const struct anv_device *device,
-                               const struct anv_image *image,
-                               VkImageAspectFlagBits aspect,
-                               unsigned level)
+anv_image_get_fast_clear_type_addr(const struct anv_device *device,
+                                   const struct anv_image *image,
+                                   VkImageAspectFlagBits aspect)
 {
-   uint32_t plane = anv_image_aspect_to_plane(image->aspects, aspect);
-   return (struct anv_address) {
-      .bo = image->planes[plane].bo,
-      .offset = image->planes[plane].bo_offset +
-                image->planes[plane].fast_clear_state_offset +
-                anv_fast_clear_state_entry_size(device) * level,
-   };
+   struct anv_address addr =
+      anv_image_get_clear_color_addr(device, image, aspect);
+
+   const unsigned clear_color_state_size = device->info.gen >= 10 ?
+      device->isl_dev.ss.clear_color_state_size :
+      device->isl_dev.ss.clear_value_size;
+   addr.offset += clear_color_state_size;
+   return addr;
 }
 
 static inline struct anv_address
-anv_image_get_needs_resolve_addr(const struct anv_device *device,
-                                 const struct anv_image *image,
-                                 VkImageAspectFlagBits aspect,
-                                 unsigned level)
+anv_image_get_compression_state_addr(const struct anv_device *device,
+                                     const struct anv_image *image,
+                                     VkImageAspectFlagBits aspect,
+                                     uint32_t level, uint32_t array_layer)
 {
+   assert(level < anv_image_aux_levels(image, aspect));
+   assert(array_layer < anv_image_aux_layers(image, aspect, level));
+   UNUSED uint32_t plane = anv_image_aspect_to_plane(image->aspects, aspect);
+   assert(image->planes[plane].aux_usage == ISL_AUX_USAGE_CCS_E);
+
    struct anv_address addr =
-      anv_image_get_clear_color_addr(device, image, aspect, level);
-   addr.offset += device->isl_dev.ss.clear_value_size;
+      anv_image_get_fast_clear_type_addr(device, image, aspect);
+   addr.offset += 4; /* Go past the fast clear type */
+
+   if (image->type == VK_IMAGE_TYPE_3D) {
+      for (uint32_t l = 0; l < level; l++)
+         addr.offset += anv_minify(image->extent.depth, l) * 4;
+   } else {
+      addr.offset += level * image->array_size * 4;
+   }
+   addr.offset += array_layer * 4;
+
    return addr;
 }
 
@@ -2563,23 +2803,58 @@ anv_can_sample_with_hiz(const struct gen_device_info * const devinfo,
 }
 
 void
-anv_gen8_hiz_op_resolve(struct anv_cmd_buffer *cmd_buffer,
-                        const struct anv_image *image,
-                        enum blorp_hiz_op op);
-void
-anv_ccs_resolve(struct anv_cmd_buffer * const cmd_buffer,
-                const struct anv_image * const image,
-                VkImageAspectFlagBits aspect,
-                const uint8_t level,
-                const uint32_t start_layer, const uint32_t layer_count,
-                const enum blorp_fast_clear_op op);
+anv_cmd_buffer_mark_image_written(struct anv_cmd_buffer *cmd_buffer,
+                                  const struct anv_image *image,
+                                  VkImageAspectFlagBits aspect,
+                                  enum isl_aux_usage aux_usage,
+                                  uint32_t level,
+                                  uint32_t base_layer,
+                                  uint32_t layer_count);
 
 void
-anv_image_fast_clear(struct anv_cmd_buffer *cmd_buffer,
-                     const struct anv_image *image,
-                     VkImageAspectFlagBits aspect,
-                     const uint32_t base_level, const uint32_t level_count,
-                     const uint32_t base_layer, uint32_t layer_count);
+anv_image_clear_color(struct anv_cmd_buffer *cmd_buffer,
+                      const struct anv_image *image,
+                      VkImageAspectFlagBits aspect,
+                      enum isl_aux_usage aux_usage,
+                      enum isl_format format, struct isl_swizzle swizzle,
+                      uint32_t level, uint32_t base_layer, uint32_t layer_count,
+                      VkRect2D area, union isl_color_value clear_color);
+void
+anv_image_clear_depth_stencil(struct anv_cmd_buffer *cmd_buffer,
+                              const struct anv_image *image,
+                              VkImageAspectFlags aspects,
+                              enum isl_aux_usage depth_aux_usage,
+                              uint32_t level,
+                              uint32_t base_layer, uint32_t layer_count,
+                              VkRect2D area,
+                              float depth_value, uint8_t stencil_value);
+void
+anv_image_hiz_op(struct anv_cmd_buffer *cmd_buffer,
+                 const struct anv_image *image,
+                 VkImageAspectFlagBits aspect, uint32_t level,
+                 uint32_t base_layer, uint32_t layer_count,
+                 enum isl_aux_op hiz_op);
+void
+anv_image_hiz_clear(struct anv_cmd_buffer *cmd_buffer,
+                    const struct anv_image *image,
+                    VkImageAspectFlags aspects,
+                    uint32_t level,
+                    uint32_t base_layer, uint32_t layer_count,
+                    VkRect2D area, uint8_t stencil_value);
+void
+anv_image_mcs_op(struct anv_cmd_buffer *cmd_buffer,
+                 const struct anv_image *image,
+                 VkImageAspectFlagBits aspect,
+                 uint32_t base_layer, uint32_t layer_count,
+                 enum isl_aux_op mcs_op, union isl_color_value *clear_value,
+                 bool predicate);
+void
+anv_image_ccs_op(struct anv_cmd_buffer *cmd_buffer,
+                 const struct anv_image *image,
+                 VkImageAspectFlagBits aspect, uint32_t level,
+                 uint32_t base_layer, uint32_t layer_count,
+                 enum isl_aux_op ccs_op, union isl_color_value *clear_value,
+                 bool predicate);
 
 void
 anv_image_copy_to_shadow(struct anv_cmd_buffer *cmd_buffer,
@@ -2593,6 +2868,12 @@ anv_layout_to_aux_usage(const struct gen_device_info * const devinfo,
                         const VkImageAspectFlagBits aspect,
                         const VkImageLayout layout);
 
+enum anv_fast_clear_type
+anv_layout_to_fast_clear_type(const struct gen_device_info * const devinfo,
+                              const struct anv_image * const image,
+                              const VkImageAspectFlagBits aspect,
+                              const VkImageLayout layout);
+
 /* This is defined as a macro so that it works for both
  * VkImageSubresourceRange and VkImageSubresourceLayers
  */
@@ -2760,16 +3041,36 @@ anv_sanitize_image_offset(const VkImageType imageType,
 void anv_fill_buffer_surface_state(struct anv_device *device,
                                    struct anv_state state,
                                    enum isl_format format,
-                                   uint32_t offset, uint32_t range,
-                                   uint32_t stride);
+                                   struct anv_address address,
+                                   uint32_t range, uint32_t stride);
+
+static inline void
+anv_clear_color_from_att_state(union isl_color_value *clear_color,
+                               const struct anv_attachment_state *att_state,
+                               const struct anv_image_view *iview)
+{
+   const struct isl_format_layout *view_fmtl =
+      isl_format_get_layout(iview->planes[0].isl.format);
+
+#define COPY_CLEAR_COLOR_CHANNEL(c, i) \
+   if (view_fmtl->channels.c.bits) \
+      clear_color->u32[i] = att_state->clear_value.color.uint32[i]
+
+   COPY_CLEAR_COLOR_CHANNEL(r, 0);
+   COPY_CLEAR_COLOR_CHANNEL(g, 1);
+   COPY_CLEAR_COLOR_CHANNEL(b, 2);
+   COPY_CLEAR_COLOR_CHANNEL(a, 3);
+
+#undef COPY_CLEAR_COLOR_CHANNEL
+}
 
 
 struct anv_ycbcr_conversion {
    const struct anv_format *        format;
-   VkSamplerYcbcrModelConversionKHR ycbcr_model;
-   VkSamplerYcbcrRangeKHR           ycbcr_range;
+   VkSamplerYcbcrModelConversion    ycbcr_model;
+   VkSamplerYcbcrRange              ycbcr_range;
    VkComponentSwizzle               mapping[4];
-   VkChromaLocationKHR              chroma_offsets[2];
+   VkChromaLocation                 chroma_offsets[2];
    VkFilter                         chroma_filter;
    bool                             chroma_reconstruction;
 };
@@ -2789,6 +3090,12 @@ struct anv_framebuffer {
    struct anv_image_view *                      attachments[0];
 };
 
+struct anv_subpass_attachment {
+   VkImageUsageFlagBits usage;
+   uint32_t attachment;
+   VkImageLayout layout;
+};
+
 struct anv_subpass {
    uint32_t                                     attachment_count;
 
@@ -2796,14 +3103,14 @@ struct anv_subpass {
     * A pointer to all attachment references used in this subpass.
     * Only valid if ::attachment_count > 0.
     */
-   VkAttachmentReference *                      attachments;
+   struct anv_subpass_attachment *              attachments;
    uint32_t                                     input_count;
-   VkAttachmentReference *                      input_attachments;
+   struct anv_subpass_attachment *              input_attachments;
    uint32_t                                     color_count;
-   VkAttachmentReference *                      color_attachments;
-   VkAttachmentReference *                      resolve_attachments;
+   struct anv_subpass_attachment *              color_attachments;
+   struct anv_subpass_attachment *              resolve_attachments;
 
-   VkAttachmentReference                        depth_stencil_attachment;
+   struct anv_subpass_attachment                depth_stencil_attachment;
 
    uint32_t                                     view_mask;
 
@@ -2958,7 +3265,7 @@ ANV_DEFINE_NONDISP_HANDLE_CASTS(anv_sampler, VkSampler)
 ANV_DEFINE_NONDISP_HANDLE_CASTS(anv_semaphore, VkSemaphore)
 ANV_DEFINE_NONDISP_HANDLE_CASTS(anv_shader_module, VkShaderModule)
 ANV_DEFINE_NONDISP_HANDLE_CASTS(vk_debug_report_callback, VkDebugReportCallbackEXT)
-ANV_DEFINE_NONDISP_HANDLE_CASTS(anv_ycbcr_conversion, VkSamplerYcbcrConversionKHR)
+ANV_DEFINE_NONDISP_HANDLE_CASTS(anv_ycbcr_conversion, VkSamplerYcbcrConversion)
 
 /* Gen-specific function declarations */
 #ifdef genX
@@ -2979,6 +3286,9 @@ ANV_DEFINE_NONDISP_HANDLE_CASTS(anv_ycbcr_conversion, VkSamplerYcbcrConversionKH
 #  define genX(x) gen10_##x
 #  include "anv_genX.h"
 #  undef genX
+#  define genX(x) gen11_##x
+#  include "anv_genX.h"
+#  undef genX
 #endif
 
 #endif /* ANV_PRIVATE_H */