vk: Re-name command buffer implementation files

Previously, the command buffer implementation was split between
anv_cmd_buffer.c and anv_cmd_emit.c.  However, this naming convention was
confusing because none of the Vulkan entrypoints for anv_cmd_buffer were
actually in anv_cmd_buffer.c.  This changes it so that anv_cmd_buffer.c is
what you think it is and the internals are in anv_batch_chain.c.

4 files changed, 2087 insertions, 2087 deletions

diff --git a/src/vulkan/Makefile.am b/src/vulkan/Makefile.am
index 6d1212c532c..c816f97034f 100644
--- a/src/vulkan/Makefile.am
+++ b/src/vulkan/Makefile.am
@@ -57,7 +57,7 @@ libvulkan_la_SOURCES =                                  \
 	anv_allocator.c                                 \
 	anv_aub.c                                       \
 	anv_cmd_buffer.c                                \
-	anv_cmd_emit.c                                  \
+	anv_batch_chain.c                               \
 	anv_compiler.cpp                                \
 	anv_device.c                                    \
 	anv_entrypoints.c                               \
diff --git a/src/vulkan/anv_batch_chain.c b/src/vulkan/anv_batch_chain.c
new file mode 100644
index 00000000000..2f09248acee
--- /dev/null
+++ b/src/vulkan/anv_batch_chain.c
@@ -0,0 +1,926 @@
+/*
+ * Copyright © 2015 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include <assert.h>
+#include <stdbool.h>
+#include <string.h>
+#include <unistd.h>
+#include <fcntl.h>
+
+#include "anv_private.h"
+
+/** \file anv_batch_chain.c
+ *
+ * This file contains functions related to anv_cmd_buffer as a data
+ * structure.  This involves everything required to create and destroy
+ * the actual batch buffers as well as link them together and handle
+ * relocations and surface state.  It specifically does *not* contain any
+ * handling of actual vkCmd calls beyond vkCmdExecuteCommands.
+ */
+
+/*-----------------------------------------------------------------------*
+ * Functions related to anv_reloc_list
+ *-----------------------------------------------------------------------*/
+
+static VkResult
+anv_reloc_list_init_clone(struct anv_reloc_list *list,
+                          struct anv_device *device,
+                          const struct anv_reloc_list *other_list)
+{
+   if (other_list) {
+      list->num_relocs = other_list->num_relocs;
+      list->array_length = other_list->array_length;
+   } else {
+      list->num_relocs = 0;
+      list->array_length = 256;
+   }
+
+   list->relocs =
+      anv_device_alloc(device, list->array_length * sizeof(*list->relocs), 8,
+                       VK_SYSTEM_ALLOC_TYPE_INTERNAL);
+
+   if (list->relocs == NULL)
+      return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+   list->reloc_bos =
+      anv_device_alloc(device, list->array_length * sizeof(*list->reloc_bos), 8,
+                       VK_SYSTEM_ALLOC_TYPE_INTERNAL);
+
+   if (list->reloc_bos == NULL) {
+      anv_device_free(device, list->relocs);
+      return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+   }
+
+   if (other_list) {
+      memcpy(list->relocs, other_list->relocs,
+             list->array_length * sizeof(*list->relocs));
+      memcpy(list->reloc_bos, other_list->reloc_bos,
+             list->array_length * sizeof(*list->reloc_bos));
+   }
+
+   return VK_SUCCESS;
+}
+
+VkResult
+anv_reloc_list_init(struct anv_reloc_list *list, struct anv_device *device)
+{
+   return anv_reloc_list_init_clone(list, device, NULL);
+}
+
+void
+anv_reloc_list_finish(struct anv_reloc_list *list, struct anv_device *device)
+{
+   anv_device_free(device, list->relocs);
+   anv_device_free(device, list->reloc_bos);
+}
+
+static VkResult
+anv_reloc_list_grow(struct anv_reloc_list *list, struct anv_device *device,
+                    size_t num_additional_relocs)
+{
+   if (list->num_relocs + num_additional_relocs <= list->array_length)
+      return VK_SUCCESS;
+
+   size_t new_length = list->array_length * 2;
+   while (new_length < list->num_relocs + num_additional_relocs)
+      new_length *= 2;
+
+   struct drm_i915_gem_relocation_entry *new_relocs =
+      anv_device_alloc(device, new_length * sizeof(*list->relocs), 8,
+                       VK_SYSTEM_ALLOC_TYPE_INTERNAL);
+   if (new_relocs == NULL)
+      return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+   struct anv_bo **new_reloc_bos =
+      anv_device_alloc(device, new_length * sizeof(*list->reloc_bos), 8,
+                       VK_SYSTEM_ALLOC_TYPE_INTERNAL);
+   if (new_relocs == NULL) {
+      anv_device_free(device, new_relocs);
+      return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+   }
+
+   memcpy(new_relocs, list->relocs, list->num_relocs * sizeof(*list->relocs));
+   memcpy(new_reloc_bos, list->reloc_bos,
+          list->num_relocs * sizeof(*list->reloc_bos));
+
+   anv_device_free(device, list->relocs);
+   anv_device_free(device, list->reloc_bos);
+
+   list->array_length = new_length;
+   list->relocs = new_relocs;
+   list->reloc_bos = new_reloc_bos;
+
+   return VK_SUCCESS;
+}
+
+uint64_t
+anv_reloc_list_add(struct anv_reloc_list *list, struct anv_device *device,
+                   uint32_t offset, struct anv_bo *target_bo, uint32_t delta)
+{
+   struct drm_i915_gem_relocation_entry *entry;
+   int index;
+
+   anv_reloc_list_grow(list, device, 1);
+   /* TODO: Handle failure */
+
+   /* XXX: Can we use I915_EXEC_HANDLE_LUT? */
+   index = list->num_relocs++;
+   list->reloc_bos[index] = target_bo;
+   entry = &list->relocs[index];
+   entry->target_handle = target_bo->gem_handle;
+   entry->delta = delta;
+   entry->offset = offset;
+   entry->presumed_offset = target_bo->offset;
+   entry->read_domains = 0;
+   entry->write_domain = 0;
+
+   return target_bo->offset + delta;
+}
+
+static void
+anv_reloc_list_append(struct anv_reloc_list *list, struct anv_device *device,
+                      struct anv_reloc_list *other, uint32_t offset)
+{
+   anv_reloc_list_grow(list, device, other->num_relocs);
+   /* TODO: Handle failure */
+
+   memcpy(&list->relocs[list->num_relocs], &other->relocs[0],
+          other->num_relocs * sizeof(other->relocs[0]));
+   memcpy(&list->reloc_bos[list->num_relocs], &other->reloc_bos[0],
+          other->num_relocs * sizeof(other->reloc_bos[0]));
+
+   for (uint32_t i = 0; i < other->num_relocs; i++)
+      list->relocs[i + list->num_relocs].offset += offset;
+
+   list->num_relocs += other->num_relocs;
+}
+
+/*-----------------------------------------------------------------------*
+ * Functions related to anv_batch
+ *-----------------------------------------------------------------------*/
+
+void *
+anv_batch_emit_dwords(struct anv_batch *batch, int num_dwords)
+{
+   if (batch->next + num_dwords * 4 > batch->end)
+      batch->extend_cb(batch, batch->user_data);
+
+   void *p = batch->next;
+
+   batch->next += num_dwords * 4;
+   assert(batch->next <= batch->end);
+
+   return p;
+}
+
+uint64_t
+anv_batch_emit_reloc(struct anv_batch *batch,
+                     void *location, struct anv_bo *bo, uint32_t delta)
+{
+   return anv_reloc_list_add(batch->relocs, batch->device,
+                             location - batch->start, bo, delta);
+}
+
+void
+anv_batch_emit_batch(struct anv_batch *batch, struct anv_batch *other)
+{
+   uint32_t size, offset;
+
+   size = other->next - other->start;
+   assert(size % 4 == 0);
+
+   if (batch->next + size > batch->end)
+      batch->extend_cb(batch, batch->user_data);
+
+   assert(batch->next + size <= batch->end);
+
+   VG(VALGRIND_CHECK_MEM_IS_DEFINED(other->start, size));
+   memcpy(batch->next, other->start, size);
+
+   offset = batch->next - batch->start;
+   anv_reloc_list_append(batch->relocs, batch->device,
+                         other->relocs, offset);
+
+   batch->next += size;
+}
+
+/*-----------------------------------------------------------------------*
+ * Functions related to anv_batch_bo
+ *-----------------------------------------------------------------------*/
+
+static VkResult
+anv_batch_bo_create(struct anv_device *device, struct anv_batch_bo **bbo_out)
+{
+   VkResult result;
+
+   struct anv_batch_bo *bbo =
+      anv_device_alloc(device, sizeof(*bbo), 8, VK_SYSTEM_ALLOC_TYPE_INTERNAL);
+   if (bbo == NULL)
+      return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+   result = anv_bo_pool_alloc(&device->batch_bo_pool, &bbo->bo);
+   if (result != VK_SUCCESS)
+      goto fail_alloc;
+
+   result = anv_reloc_list_init(&bbo->relocs, device);
+   if (result != VK_SUCCESS)
+      goto fail_bo_alloc;
+
+   *bbo_out = bbo;
+
+   return VK_SUCCESS;
+
+ fail_bo_alloc:
+   anv_bo_pool_free(&device->batch_bo_pool, &bbo->bo);
+ fail_alloc:
+   anv_device_free(device, bbo);
+
+   return result;
+}
+
+static VkResult
+anv_batch_bo_clone(struct anv_device *device,
+                   const struct anv_batch_bo *other_bbo,
+                   struct anv_batch_bo **bbo_out)
+{
+   VkResult result;
+
+   struct anv_batch_bo *bbo =
+      anv_device_alloc(device, sizeof(*bbo), 8, VK_SYSTEM_ALLOC_TYPE_INTERNAL);
+   if (bbo == NULL)
+      return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+   result = anv_bo_pool_alloc(&device->batch_bo_pool, &bbo->bo);
+   if (result != VK_SUCCESS)
+      goto fail_alloc;
+
+   result = anv_reloc_list_init_clone(&bbo->relocs, device, &other_bbo->relocs);
+   if (result != VK_SUCCESS)
+      goto fail_bo_alloc;
+
+   bbo->length = other_bbo->length;
+   memcpy(bbo->bo.map, other_bbo->bo.map, other_bbo->length);
+
+   *bbo_out = bbo;
+
+   return VK_SUCCESS;
+
+ fail_bo_alloc:
+   anv_bo_pool_free(&device->batch_bo_pool, &bbo->bo);
+ fail_alloc:
+   anv_device_free(device, bbo);
+
+   return result;
+}
+
+static void
+anv_batch_bo_start(struct anv_batch_bo *bbo, struct anv_batch *batch,
+                   size_t batch_padding)
+{
+   batch->next = batch->start = bbo->bo.map;
+   batch->end = bbo->bo.map + bbo->bo.size - batch_padding;
+   batch->relocs = &bbo->relocs;
+   bbo->relocs.num_relocs = 0;
+}
+
+static void
+anv_batch_bo_continue(struct anv_batch_bo *bbo, struct anv_batch *batch,
+                      size_t batch_padding)
+{
+   batch->start = bbo->bo.map;
+   batch->next = bbo->bo.map + bbo->length;
+   batch->end = bbo->bo.map + bbo->bo.size - batch_padding;
+   batch->relocs = &bbo->relocs;
+}
+
+static void
+anv_batch_bo_finish(struct anv_batch_bo *bbo, struct anv_batch *batch)
+{
+   assert(batch->start == bbo->bo.map);
+   bbo->length = batch->next - batch->start;
+   VG(VALGRIND_CHECK_MEM_IS_DEFINED(batch->start, bbo->length));
+}
+
+static void
+anv_batch_bo_destroy(struct anv_batch_bo *bbo, struct anv_device *device)
+{
+   anv_reloc_list_finish(&bbo->relocs, device);
+   anv_bo_pool_free(&device->batch_bo_pool, &bbo->bo);
+   anv_device_free(device, bbo);
+}
+
+static VkResult
+anv_batch_bo_list_clone(const struct list_head *list, struct anv_device *device,
+                        struct list_head *new_list)
+{
+   VkResult result = VK_SUCCESS;
+
+   list_inithead(new_list);
+
+   struct anv_batch_bo *prev_bbo = NULL;
+   list_for_each_entry(struct anv_batch_bo, bbo, list, link) {
+      struct anv_batch_bo *new_bbo;
+      result = anv_batch_bo_clone(device, bbo, &new_bbo);
+      if (result != VK_SUCCESS)
+         break;
+      list_addtail(&new_bbo->link, new_list);
+
+      if (prev_bbo) {
+         /* As we clone this list of batch_bo's, they chain one to the
+          * other using MI_BATCH_BUFFER_START commands.  We need to fix up
+          * those relocations as we go.  Fortunately, this is pretty easy
+          * as it will always be the last relocation in the list.
+          */
+         uint32_t last_idx = prev_bbo->relocs.num_relocs - 1;
+         assert(prev_bbo->relocs.reloc_bos[last_idx] == &bbo->bo);
+         prev_bbo->relocs.reloc_bos[last_idx] = &new_bbo->bo;
+      }
+
+      prev_bbo = new_bbo;
+   }
+
+   if (result != VK_SUCCESS) {
+      list_for_each_entry_safe(struct anv_batch_bo, bbo, new_list, link)
+         anv_batch_bo_destroy(bbo, device);
+   }
+
+   return result;
+}
+
+/*-----------------------------------------------------------------------*
+ * Functions related to anv_batch_bo
+ *-----------------------------------------------------------------------*/
+
+static inline struct anv_batch_bo *
+anv_cmd_buffer_current_batch_bo(struct anv_cmd_buffer *cmd_buffer)
+{
+   return LIST_ENTRY(struct anv_batch_bo, cmd_buffer->batch_bos.prev, link);
+}
+
+static inline struct anv_batch_bo *
+anv_cmd_buffer_current_surface_bbo(struct anv_cmd_buffer *cmd_buffer)
+{
+   return LIST_ENTRY(struct anv_batch_bo, cmd_buffer->surface_bos.prev, link);
+}
+
+struct anv_bo *
+anv_cmd_buffer_current_surface_bo(struct anv_cmd_buffer *cmd_buffer)
+{
+   return &anv_cmd_buffer_current_surface_bbo(cmd_buffer)->bo;
+}
+
+struct anv_reloc_list *
+anv_cmd_buffer_current_surface_relocs(struct anv_cmd_buffer *cmd_buffer)
+{
+   return &anv_cmd_buffer_current_surface_bbo(cmd_buffer)->relocs;
+}
+
+static void
+cmd_buffer_chain_to_batch_bo(struct anv_cmd_buffer *cmd_buffer,
+                             struct anv_batch_bo *bbo)
+{
+   struct anv_batch *batch = &cmd_buffer->batch;
+   struct anv_batch_bo *current_bbo =
+      anv_cmd_buffer_current_batch_bo(cmd_buffer);
+
+   /* We set the end of the batch a little short so we would be sure we
+    * have room for the chaining command.  Since we're about to emit the
+    * chaining command, let's set it back where it should go.
+    */
+   batch->end += GEN8_MI_BATCH_BUFFER_START_length * 4;
+   assert(batch->end == current_bbo->bo.map + current_bbo->bo.size);
+
+   anv_batch_emit(batch, GEN8_MI_BATCH_BUFFER_START,
+      GEN8_MI_BATCH_BUFFER_START_header,
+      ._2ndLevelBatchBuffer = _1stlevelbatch,
+      .AddressSpaceIndicator = ASI_PPGTT,
+      .BatchBufferStartAddress = { &bbo->bo, 0 },
+   );
+
+   anv_batch_bo_finish(current_bbo, batch);
+}
+
+static VkResult
+anv_cmd_buffer_chain_batch(struct anv_batch *batch, void *_data)
+{
+   struct anv_cmd_buffer *cmd_buffer = _data;
+   struct anv_batch_bo *new_bbo;
+
+   VkResult result = anv_batch_bo_create(cmd_buffer->device, &new_bbo);
+   if (result != VK_SUCCESS)
+      return result;
+
+   struct anv_batch_bo **seen_bbo = anv_vector_add(&cmd_buffer->seen_bbos);
+   if (seen_bbo == NULL) {
+      anv_batch_bo_destroy(new_bbo, cmd_buffer->device);
+      return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+   }
+   *seen_bbo = new_bbo;
+
+   cmd_buffer_chain_to_batch_bo(cmd_buffer, new_bbo);
+
+   list_addtail(&new_bbo->link, &cmd_buffer->batch_bos);
+
+   anv_batch_bo_start(new_bbo, batch, GEN8_MI_BATCH_BUFFER_START_length * 4);
+
+   return VK_SUCCESS;
+}
+
+struct anv_state
+anv_cmd_buffer_alloc_surface_state(struct anv_cmd_buffer *cmd_buffer,
+                                   uint32_t size, uint32_t alignment)
+{
+   struct anv_bo *surface_bo =
+      anv_cmd_buffer_current_surface_bo(cmd_buffer);
+   struct anv_state state;
+
+   state.offset = align_u32(cmd_buffer->surface_next, alignment);
+   if (state.offset + size > surface_bo->size)
+      return (struct anv_state) { 0 };
+
+   state.map = surface_bo->map + state.offset;
+   state.alloc_size = size;
+   cmd_buffer->surface_next = state.offset + size;
+
+   assert(state.offset + size <= surface_bo->size);
+
+   return state;
+}
+
+struct anv_state
+anv_cmd_buffer_alloc_dynamic_state(struct anv_cmd_buffer *cmd_buffer,
+                                   uint32_t size, uint32_t alignment)
+{
+   return anv_state_stream_alloc(&cmd_buffer->dynamic_state_stream,
+                                 size, alignment);
+}
+
+VkResult
+anv_cmd_buffer_new_surface_state_bo(struct anv_cmd_buffer *cmd_buffer)
+{
+   struct anv_batch_bo *new_bbo, *old_bbo =
+      anv_cmd_buffer_current_surface_bbo(cmd_buffer);
+
+   /* Finish off the old buffer */
+   old_bbo->length = cmd_buffer->surface_next;
+
+   VkResult result = anv_batch_bo_create(cmd_buffer->device, &new_bbo);
+   if (result != VK_SUCCESS)
+      return result;
+
+   struct anv_batch_bo **seen_bbo = anv_vector_add(&cmd_buffer->seen_bbos);
+   if (seen_bbo == NULL) {
+      anv_batch_bo_destroy(new_bbo, cmd_buffer->device);
+      return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+   }
+   *seen_bbo = new_bbo;
+
+   cmd_buffer->surface_next = 1;
+
+   list_addtail(&new_bbo->link, &cmd_buffer->surface_bos);
+
+   return VK_SUCCESS;
+}
+
+VkResult
+anv_cmd_buffer_init_batch_bo_chain(struct anv_cmd_buffer *cmd_buffer)
+{
+   struct anv_batch_bo *batch_bo, *surface_bbo;
+   struct anv_device *device = cmd_buffer->device;
+   VkResult result;
+
+   list_inithead(&cmd_buffer->batch_bos);
+   list_inithead(&cmd_buffer->surface_bos);
+
+   result = anv_batch_bo_create(device, &batch_bo);
+   if (result != VK_SUCCESS)
+      return result;
+
+   list_addtail(&batch_bo->link, &cmd_buffer->batch_bos);
+
+   cmd_buffer->batch.device = device;
+   cmd_buffer->batch.extend_cb = anv_cmd_buffer_chain_batch;
+   cmd_buffer->batch.user_data = cmd_buffer;
+
+   anv_batch_bo_start(batch_bo, &cmd_buffer->batch,
+                      GEN8_MI_BATCH_BUFFER_START_length * 4);
+
+   result = anv_batch_bo_create(device, &surface_bbo);
+   if (result != VK_SUCCESS)
+      goto fail_batch_bo;
+
+   list_addtail(&surface_bbo->link, &cmd_buffer->surface_bos);
+
+   int success = anv_vector_init(&cmd_buffer->seen_bbos,
+                                 sizeof(struct anv_bo *),
+                                 8 * sizeof(struct anv_bo *));
+   if (!success)
+      goto fail_surface_bo;
+
+   *(struct anv_batch_bo **)anv_vector_add(&cmd_buffer->seen_bbos) = batch_bo;
+   *(struct anv_batch_bo **)anv_vector_add(&cmd_buffer->seen_bbos) = surface_bbo;
+
+   /* Start surface_next at 1 so surface offset 0 is invalid. */
+   cmd_buffer->surface_next = 1;
+
+   cmd_buffer->execbuf2.objects = NULL;
+   cmd_buffer->execbuf2.bos = NULL;
+   cmd_buffer->execbuf2.array_length = 0;
+
+   return VK_SUCCESS;
+
+ fail_surface_bo:
+   anv_batch_bo_destroy(surface_bbo, device);
+ fail_batch_bo:
+   anv_batch_bo_destroy(batch_bo, device);
+
+   return result;
+}
+
+void
+anv_cmd_buffer_fini_batch_bo_chain(struct anv_cmd_buffer *cmd_buffer)
+{
+   struct anv_device *device = cmd_buffer->device;
+
+   anv_vector_finish(&cmd_buffer->seen_bbos);
+
+   /* Destroy all of the batch buffers */
+   list_for_each_entry_safe(struct anv_batch_bo, bbo,
+                            &cmd_buffer->batch_bos, link) {
+      anv_batch_bo_destroy(bbo, device);
+   }
+
+   /* Destroy all of the surface state buffers */
+   list_for_each_entry_safe(struct anv_batch_bo, bbo,
+                            &cmd_buffer->surface_bos, link) {
+      anv_batch_bo_destroy(bbo, device);
+   }
+
+   anv_device_free(device, cmd_buffer->execbuf2.objects);
+   anv_device_free(device, cmd_buffer->execbuf2.bos);
+}
+
+void
+anv_cmd_buffer_reset_batch_bo_chain(struct anv_cmd_buffer *cmd_buffer)
+{
+   struct anv_device *device = cmd_buffer->device;
+
+   /* Delete all but the first batch bo */
+   assert(!list_empty(&cmd_buffer->batch_bos));
+   while (cmd_buffer->batch_bos.next != cmd_buffer->batch_bos.prev) {
+      struct anv_batch_bo *bbo = anv_cmd_buffer_current_batch_bo(cmd_buffer);
+      list_del(&bbo->link);
+      anv_batch_bo_destroy(bbo, device);
+   }
+   assert(!list_empty(&cmd_buffer->batch_bos));
+
+   anv_batch_bo_start(anv_cmd_buffer_current_batch_bo(cmd_buffer),
+                      &cmd_buffer->batch,
+                      GEN8_MI_BATCH_BUFFER_START_length * 4);
+
+   /* Delete all but the first batch bo */
+   assert(!list_empty(&cmd_buffer->batch_bos));
+   while (cmd_buffer->surface_bos.next != cmd_buffer->surface_bos.prev) {
+      struct anv_batch_bo *bbo = anv_cmd_buffer_current_surface_bbo(cmd_buffer);
+      list_del(&bbo->link);
+      anv_batch_bo_destroy(bbo, device);
+   }
+   assert(!list_empty(&cmd_buffer->batch_bos));
+
+   anv_cmd_buffer_current_surface_bbo(cmd_buffer)->relocs.num_relocs = 0;
+
+   cmd_buffer->surface_next = 1;
+
+   /* Reset the list of seen buffers */
+   cmd_buffer->seen_bbos.head = 0;
+   cmd_buffer->seen_bbos.tail = 0;
+
+   *(struct anv_batch_bo **)anv_vector_add(&cmd_buffer->seen_bbos) =
+      anv_cmd_buffer_current_batch_bo(cmd_buffer);
+   *(struct anv_batch_bo **)anv_vector_add(&cmd_buffer->seen_bbos) =
+      anv_cmd_buffer_current_surface_bbo(cmd_buffer);
+}
+
+void
+anv_cmd_buffer_end_batch_buffer(struct anv_cmd_buffer *cmd_buffer)
+{
+   struct anv_batch_bo *batch_bo = anv_cmd_buffer_current_batch_bo(cmd_buffer);
+   struct anv_batch_bo *surface_bbo =
+      anv_cmd_buffer_current_surface_bbo(cmd_buffer);
+
+   if (cmd_buffer->level == VK_CMD_BUFFER_LEVEL_PRIMARY) {
+      anv_batch_emit(&cmd_buffer->batch, GEN8_MI_BATCH_BUFFER_END);
+
+      /* Round batch up to an even number of dwords. */
+      if ((cmd_buffer->batch.next - cmd_buffer->batch.start) & 4)
+         anv_batch_emit(&cmd_buffer->batch, GEN8_MI_NOOP);
+
+      cmd_buffer->exec_mode = ANV_CMD_BUFFER_EXEC_MODE_PRIMARY;
+   } else {
+      /* If this is a secondary command buffer, we need to determine the
+       * mode in which it will be executed with vkExecuteCommands.  We
+       * determine this statically here so that this stays in sync with the
+       * actual ExecuteCommands implementation.
+       */
+      if ((cmd_buffer->batch_bos.next == cmd_buffer->batch_bos.prev) &&
+          (anv_cmd_buffer_current_batch_bo(cmd_buffer)->length <
+           ANV_CMD_BUFFER_BATCH_SIZE / 2)) {
+         /* If the secondary has exactly one batch buffer in its list *and*
+          * that batch buffer is less than half of the maximum size, we're
+          * probably better of simply copying it into our batch.
+          */
+         cmd_buffer->exec_mode = ANV_CMD_BUFFER_EXEC_MODE_EMIT;
+      } else if (cmd_buffer->opt_flags &
+                 VK_CMD_BUFFER_OPTIMIZE_NO_SIMULTANEOUS_USE_BIT) {
+         cmd_buffer->exec_mode = ANV_CMD_BUFFER_EXEC_MODE_CHAIN;
+
+         /* For chaining mode, we need to increment the number of
+          * relocations.  This is because, when we chain, we need to add
+          * an MI_BATCH_BUFFER_START command.  Adding this command will
+          * also add a relocation.  In order to handle theis we'll
+          * increment it here and decrement it right before adding the
+          * MI_BATCH_BUFFER_START command.
+          */
+         anv_cmd_buffer_current_batch_bo(cmd_buffer)->relocs.num_relocs++;
+      } else {
+         cmd_buffer->exec_mode = ANV_CMD_BUFFER_EXEC_MODE_COPY_AND_CHAIN;
+      }
+   }
+
+   anv_batch_bo_finish(batch_bo, &cmd_buffer->batch);
+
+   surface_bbo->length = cmd_buffer->surface_next;
+}
+
+static inline VkResult
+anv_cmd_buffer_add_seen_bbos(struct anv_cmd_buffer *cmd_buffer,
+                             struct list_head *list)
+{
+   list_for_each_entry(struct anv_batch_bo, bbo, list, link) {
+      struct anv_batch_bo **bbo_ptr = anv_vector_add(&cmd_buffer->seen_bbos);
+      if (bbo_ptr == NULL)
+         return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+      *bbo_ptr = bbo;
+   }
+
+   return VK_SUCCESS;
+}
+
+void
+anv_cmd_buffer_add_secondary(struct anv_cmd_buffer *primary,
+                             struct anv_cmd_buffer *secondary)
+{
+   switch (secondary->exec_mode) {
+   case ANV_CMD_BUFFER_EXEC_MODE_EMIT:
+      anv_batch_emit_batch(&primary->batch, &secondary->batch);
+      break;
+   case ANV_CMD_BUFFER_EXEC_MODE_CHAIN: {
+      struct anv_batch_bo *first_bbo =
+         list_first_entry(&secondary->batch_bos, struct anv_batch_bo, link);
+      struct anv_batch_bo *last_bbo =
+         list_last_entry(&secondary->batch_bos, struct anv_batch_bo, link);
+
+      anv_batch_emit(&primary->batch, GEN8_MI_BATCH_BUFFER_START,
+         GEN8_MI_BATCH_BUFFER_START_header,
+         ._2ndLevelBatchBuffer = _1stlevelbatch,
+         .AddressSpaceIndicator = ASI_PPGTT,
+         .BatchBufferStartAddress = { &first_bbo->bo, 0 },
+      );
+
+      struct anv_batch_bo *this_bbo = anv_cmd_buffer_current_batch_bo(primary);
+      assert(primary->batch.start == this_bbo->bo.map);
+      uint32_t offset = primary->batch.next - primary->batch.start;
+
+      struct GEN8_MI_BATCH_BUFFER_START ret = {
+         GEN8_MI_BATCH_BUFFER_START_header,
+         ._2ndLevelBatchBuffer = _1stlevelbatch,
+         .AddressSpaceIndicator = ASI_PPGTT,
+         .BatchBufferStartAddress = { &this_bbo->bo, offset },
+      };
+      last_bbo->relocs.num_relocs++;
+      GEN8_MI_BATCH_BUFFER_START_pack(&secondary->batch,
+                                      last_bbo->bo.map + last_bbo->length,
+                                      &ret);
+
+      anv_cmd_buffer_add_seen_bbos(primary, &secondary->batch_bos);
+      break;
+   }
+   case ANV_CMD_BUFFER_EXEC_MODE_COPY_AND_CHAIN: {
+      struct list_head copy_list;
+      VkResult result = anv_batch_bo_list_clone(&secondary->batch_bos,
+                                                secondary->device,
+                                                &copy_list);
+      if (result != VK_SUCCESS)
+         return; /* FIXME */
+
+      anv_cmd_buffer_add_seen_bbos(primary, &copy_list);
+
+      struct anv_batch_bo *first_bbo =
+         list_first_entry(&copy_list, struct anv_batch_bo, link);
+      struct anv_batch_bo *last_bbo =
+         list_last_entry(&copy_list, struct anv_batch_bo, link);
+
+      cmd_buffer_chain_to_batch_bo(primary, first_bbo);
+
+      list_splicetail(&copy_list, &primary->batch_bos);
+
+      anv_batch_bo_continue(last_bbo, &primary->batch,
+                            GEN8_MI_BATCH_BUFFER_START_length * 4);
+
+      anv_cmd_buffer_emit_state_base_address(primary);
+      break;
+   }
+   default:
+      assert(!"Invalid execution mode");
+   }
+
+   /* Mark the surface buffer from the secondary as seen */
+   anv_cmd_buffer_add_seen_bbos(primary, &secondary->surface_bos);
+}
+
+static VkResult
+anv_cmd_buffer_add_bo(struct anv_cmd_buffer *cmd_buffer,
+                      struct anv_bo *bo,
+                      struct anv_reloc_list *relocs)
+{
+   struct drm_i915_gem_exec_object2 *obj = NULL;
+
+   if (bo->index < cmd_buffer->execbuf2.bo_count &&
+       cmd_buffer->execbuf2.bos[bo->index] == bo)
+      obj = &cmd_buffer->execbuf2.objects[bo->index];
+
+   if (obj == NULL) {
+      /* We've never seen this one before.  Add it to the list and assign
+       * an id that we can use later.
+       */
+      if (cmd_buffer->execbuf2.bo_count >= cmd_buffer->execbuf2.array_length) {
+         uint32_t new_len = cmd_buffer->execbuf2.objects ?
+                            cmd_buffer->execbuf2.array_length * 2 : 64;
+
+         struct drm_i915_gem_exec_object2 *new_objects =
+            anv_device_alloc(cmd_buffer->device, new_len * sizeof(*new_objects),
+                             8, VK_SYSTEM_ALLOC_TYPE_INTERNAL);
+         if (new_objects == NULL)
+            return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+
+         struct anv_bo **new_bos =
+            anv_device_alloc(cmd_buffer->device, new_len * sizeof(*new_bos),
+                             8, VK_SYSTEM_ALLOC_TYPE_INTERNAL);
+         if (new_objects == NULL) {
+            anv_device_free(cmd_buffer->device, new_objects);
+            return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+         }
+
+         if (cmd_buffer->execbuf2.objects) {
+            memcpy(new_objects, cmd_buffer->execbuf2.objects,
+                   cmd_buffer->execbuf2.bo_count * sizeof(*new_objects));
+            memcpy(new_bos, cmd_buffer->execbuf2.bos,
+                   cmd_buffer->execbuf2.bo_count * sizeof(*new_bos));
+         }
+
+         cmd_buffer->execbuf2.objects = new_objects;
+         cmd_buffer->execbuf2.bos = new_bos;
+         cmd_buffer->execbuf2.array_length = new_len;
+      }
+
+      assert(cmd_buffer->execbuf2.bo_count < cmd_buffer->execbuf2.array_length);
+
+      bo->index = cmd_buffer->execbuf2.bo_count++;
+      obj = &cmd_buffer->execbuf2.objects[bo->index];
+      cmd_buffer->execbuf2.bos[bo->index] = bo;
+
+      obj->handle = bo->gem_handle;
+      obj->relocation_count = 0;
+      obj->relocs_ptr = 0;
+      obj->alignment = 0;
+      obj->offset = bo->offset;
+      obj->flags = 0;
+      obj->rsvd1 = 0;
+      obj->rsvd2 = 0;
+   }
+
+   if (relocs != NULL && obj->relocation_count == 0) {
+      /* This is the first time we've ever seen a list of relocations for
+       * this BO.  Go ahead and set the relocations and then walk the list
+       * of relocations and add them all.
+       */
+      obj->relocation_count = relocs->num_relocs;
+      obj->relocs_ptr = (uintptr_t) relocs->relocs;
+
+      for (size_t i = 0; i < relocs->num_relocs; i++)
+         anv_cmd_buffer_add_bo(cmd_buffer, relocs->reloc_bos[i], NULL);
+   }
+
+   return VK_SUCCESS;
+}
+
+static void
+anv_cmd_buffer_process_relocs(struct anv_cmd_buffer *cmd_buffer,
+                              struct anv_reloc_list *list)
+{
+   struct anv_bo *bo;
+
+   /* If the kernel supports I915_EXEC_NO_RELOC, it will compare offset in
+    * struct drm_i915_gem_exec_object2 against the bos current offset and if
+    * all bos haven't moved it will skip relocation processing alltogether.
+    * If I915_EXEC_NO_RELOC is not supported, the kernel ignores the incoming
+    * value of offset so we can set it either way.  For that to work we need
+    * to make sure all relocs use the same presumed offset.
+    */
+
+   for (size_t i = 0; i < list->num_relocs; i++) {
+      bo = list->reloc_bos[i];
+      if (bo->offset != list->relocs[i].presumed_offset)
+         cmd_buffer->execbuf2.need_reloc = true;
+
+      list->relocs[i].target_handle = bo->index;
+   }
+}
+
+void
+anv_cmd_buffer_prepare_execbuf(struct anv_cmd_buffer *cmd_buffer)
+{
+   struct anv_batch *batch = &cmd_buffer->batch;
+
+   cmd_buffer->execbuf2.bo_count = 0;
+   cmd_buffer->execbuf2.need_reloc = false;
+
+   /* First, we walk over all of the bos we've seen and add them and their
+    * relocations to the validate list.
+    */
+   struct anv_batch_bo **bbo;
+   anv_vector_foreach(bbo, &cmd_buffer->seen_bbos)
+      anv_cmd_buffer_add_bo(cmd_buffer, &(*bbo)->bo, &(*bbo)->relocs);
+
+   struct anv_batch_bo *first_batch_bo =
+      list_first_entry(&cmd_buffer->batch_bos, struct anv_batch_bo, link);
+
+   /* The kernel requires that the last entry in the validation list be the
+    * batch buffer to execute.  We can simply swap the element
+    * corresponding to the first batch_bo in the chain with the last
+    * element in the list.
+    */
+   if (first_batch_bo->bo.index != cmd_buffer->execbuf2.bo_count - 1) {
+      uint32_t idx = first_batch_bo->bo.index;
+
+      struct drm_i915_gem_exec_object2 tmp_obj =
+         cmd_buffer->execbuf2.objects[idx];
+      assert(cmd_buffer->execbuf2.bos[idx] == &first_batch_bo->bo);
+
+      cmd_buffer->execbuf2.objects[idx] =
+         cmd_buffer->execbuf2.objects[cmd_buffer->execbuf2.bo_count - 1];
+      cmd_buffer->execbuf2.bos[idx] =
+         cmd_buffer->execbuf2.bos[cmd_buffer->execbuf2.bo_count - 1];
+      cmd_buffer->execbuf2.bos[idx]->index = idx;
+
+      cmd_buffer->execbuf2.objects[cmd_buffer->execbuf2.bo_count - 1] = tmp_obj;
+      cmd_buffer->execbuf2.bos[cmd_buffer->execbuf2.bo_count - 1] =
+         &first_batch_bo->bo;
+      first_batch_bo->bo.index = cmd_buffer->execbuf2.bo_count - 1;
+   }
+
+   /* Now we go through and fixup all of the relocation lists to point to
+    * the correct indices in the object array.  We have to do this after we
+    * reorder the list above as some of the indices may have changed.
+    */
+   anv_vector_foreach(bbo, &cmd_buffer->seen_bbos)
+      anv_cmd_buffer_process_relocs(cmd_buffer, &(*bbo)->relocs);
+
+   cmd_buffer->execbuf2.execbuf = (struct drm_i915_gem_execbuffer2) {
+      .buffers_ptr = (uintptr_t) cmd_buffer->execbuf2.objects,
+      .buffer_count = cmd_buffer->execbuf2.bo_count,
+      .batch_start_offset = 0,
+      .batch_len = batch->next - batch->start,
+      .cliprects_ptr = 0,
+      .num_cliprects = 0,
+      .DR1 = 0,
+      .DR4 = 0,
+      .flags = I915_EXEC_HANDLE_LUT | I915_EXEC_RENDER,
+      .rsvd1 = cmd_buffer->device->context_id,
+      .rsvd2 = 0,
+   };
+
+   if (!cmd_buffer->execbuf2.need_reloc)
+      cmd_buffer->execbuf2.execbuf.flags |= I915_EXEC_NO_RELOC;
+}
diff --git a/src/vulkan/anv_cmd_buffer.c b/src/vulkan/anv_cmd_buffer.c
index 28a3af7a9b8..3b9e67fdd0f 100644
--- a/src/vulkan/anv_cmd_buffer.c
+++ b/src/vulkan/anv_cmd_buffer.c
@@ -31,896 +31,1395 @@
 
 /** \file anv_cmd_buffer.c
  *
- * This file contains functions related to anv_cmd_buffer as a data
- * structure.  This involves everything required to create and destroy
- * the actual batch buffers as well as link them together and handle
- * relocations and surface state.  It specifically does *not* contain any
- * handling of actual vkCmd calls beyond vkCmdExecuteCommands.
+ * This file contains all of the stuff for emitting commands into a command
+ * buffer.  This includes implementations of most of the vkCmd*
+ * entrypoints.  This file is concerned entirely with state emission and
+ * not with the command buffer data structure itself.  As far as this file
+ * is concerned, most of anv_cmd_buffer is magic.
  */
 
-/*-----------------------------------------------------------------------*
- * Functions related to anv_reloc_list
- *-----------------------------------------------------------------------*/
-
-static VkResult
-anv_reloc_list_init_clone(struct anv_reloc_list *list,
-                          struct anv_device *device,
-                          const struct anv_reloc_list *other_list)
+static void
+anv_cmd_state_init(struct anv_cmd_state *state)
 {
-   if (other_list) {
-      list->num_relocs = other_list->num_relocs;
-      list->array_length = other_list->array_length;
-   } else {
-      list->num_relocs = 0;
-      list->array_length = 256;
-   }
+   state->rs_state = NULL;
+   state->vp_state = NULL;
+   state->cb_state = NULL;
+   state->ds_state = NULL;
+   memset(&state->state_vf, 0, sizeof(state->state_vf));
+   memset(&state->descriptors, 0, sizeof(state->descriptors));
+
+   state->dirty = 0;
+   state->vb_dirty = 0;
+   state->descriptors_dirty = 0;
+   state->pipeline = NULL;
+   state->vp_state = NULL;
+   state->rs_state = NULL;
+   state->ds_state = NULL;
+}
 
-   list->relocs =
-      anv_device_alloc(device, list->array_length * sizeof(*list->relocs), 8,
-                       VK_SYSTEM_ALLOC_TYPE_INTERNAL);
+VkResult anv_CreateCommandBuffer(
+    VkDevice                                    _device,
+    const VkCmdBufferCreateInfo*                pCreateInfo,
+    VkCmdBuffer*                                pCmdBuffer)
+{
+   ANV_FROM_HANDLE(anv_device, device, _device);
+   ANV_FROM_HANDLE(anv_cmd_pool, pool, pCreateInfo->cmdPool);
+   struct anv_cmd_buffer *cmd_buffer;
+   VkResult result;
 
-   if (list->relocs == NULL)
+   cmd_buffer = anv_device_alloc(device, sizeof(*cmd_buffer), 8,
+                                 VK_SYSTEM_ALLOC_TYPE_API_OBJECT);
+   if (cmd_buffer == NULL)
       return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 
-   list->reloc_bos =
-      anv_device_alloc(device, list->array_length * sizeof(*list->reloc_bos), 8,
-                       VK_SYSTEM_ALLOC_TYPE_INTERNAL);
+   cmd_buffer->device = device;
 
-   if (list->reloc_bos == NULL) {
-      anv_device_free(device, list->relocs);
-      return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
-   }
+   result = anv_cmd_buffer_init_batch_bo_chain(cmd_buffer);
+   if (result != VK_SUCCESS)
+      goto fail;
 
-   if (other_list) {
-      memcpy(list->relocs, other_list->relocs,
-             list->array_length * sizeof(*list->relocs));
-      memcpy(list->reloc_bos, other_list->reloc_bos,
-             list->array_length * sizeof(*list->reloc_bos));
-   }
+   anv_state_stream_init(&cmd_buffer->surface_state_stream,
+                         &device->surface_state_block_pool);
+   anv_state_stream_init(&cmd_buffer->dynamic_state_stream,
+                         &device->dynamic_state_block_pool);
+
+   cmd_buffer->level = pCreateInfo->level;
+   cmd_buffer->opt_flags = 0;
+
+   anv_cmd_state_init(&cmd_buffer->state);
+
+   list_addtail(&cmd_buffer->pool_link, &pool->cmd_buffers);
+
+   *pCmdBuffer = anv_cmd_buffer_to_handle(cmd_buffer);
 
    return VK_SUCCESS;
+
+ fail: anv_device_free(device, cmd_buffer);
+
+   return result;
 }
 
-VkResult
-anv_reloc_list_init(struct anv_reloc_list *list, struct anv_device *device)
+VkResult anv_DestroyCommandBuffer(
+    VkDevice                                    _device,
+    VkCmdBuffer                                 _cmd_buffer)
 {
-   return anv_reloc_list_init_clone(list, device, NULL);
+   ANV_FROM_HANDLE(anv_device, device, _device);
+   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, _cmd_buffer);
+
+   list_del(&cmd_buffer->pool_link);
+
+   anv_cmd_buffer_fini_batch_bo_chain(cmd_buffer);
+
+   anv_state_stream_finish(&cmd_buffer->surface_state_stream);
+   anv_state_stream_finish(&cmd_buffer->dynamic_state_stream);
+   anv_device_free(device, cmd_buffer);
+
+   return VK_SUCCESS;
+}
+
+VkResult anv_ResetCommandBuffer(
+    VkCmdBuffer                                 cmdBuffer,
+    VkCmdBufferResetFlags                       flags)
+{
+   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
+
+   anv_cmd_buffer_reset_batch_bo_chain(cmd_buffer);
+
+   anv_cmd_state_init(&cmd_buffer->state);
+
+   return VK_SUCCESS;
 }
 
 void
-anv_reloc_list_finish(struct anv_reloc_list *list, struct anv_device *device)
+anv_cmd_buffer_emit_state_base_address(struct anv_cmd_buffer *cmd_buffer)
 {
-   anv_device_free(device, list->relocs);
-   anv_device_free(device, list->reloc_bos);
+   struct anv_device *device = cmd_buffer->device;
+   struct anv_bo *scratch_bo = NULL;
+
+   cmd_buffer->state.scratch_size = device->scratch_block_pool.size;
+   if (cmd_buffer->state.scratch_size > 0)
+      scratch_bo = &device->scratch_block_pool.bo;
+
+   anv_batch_emit(&cmd_buffer->batch, GEN8_STATE_BASE_ADDRESS,
+                  .GeneralStateBaseAddress = { scratch_bo, 0 },
+                  .GeneralStateMemoryObjectControlState = GEN8_MOCS,
+                  .GeneralStateBaseAddressModifyEnable = true,
+                  .GeneralStateBufferSize = 0xfffff,
+                  .GeneralStateBufferSizeModifyEnable = true,
+
+                  .SurfaceStateBaseAddress = { anv_cmd_buffer_current_surface_bo(cmd_buffer), 0 },
+                  .SurfaceStateMemoryObjectControlState = GEN8_MOCS,
+                  .SurfaceStateBaseAddressModifyEnable = true,
+
+                  .DynamicStateBaseAddress = { &device->dynamic_state_block_pool.bo, 0 },
+                  .DynamicStateMemoryObjectControlState = GEN8_MOCS,
+                  .DynamicStateBaseAddressModifyEnable = true,
+                  .DynamicStateBufferSize = 0xfffff,
+                  .DynamicStateBufferSizeModifyEnable = true,
+
+                  .IndirectObjectBaseAddress = { NULL, 0 },
+                  .IndirectObjectMemoryObjectControlState = GEN8_MOCS,
+                  .IndirectObjectBaseAddressModifyEnable = true,
+                  .IndirectObjectBufferSize = 0xfffff,
+                  .IndirectObjectBufferSizeModifyEnable = true,
+
+                  .InstructionBaseAddress = { &device->instruction_block_pool.bo, 0 },
+                  .InstructionMemoryObjectControlState = GEN8_MOCS,
+                  .InstructionBaseAddressModifyEnable = true,
+                  .InstructionBufferSize = 0xfffff,
+                  .InstructionBuffersizeModifyEnable = true);
+
+   /* After re-setting the surface state base address, we have to do some
+    * cache flusing so that the sampler engine will pick up the new
+    * SURFACE_STATE objects and binding tables. From the Broadwell PRM,
+    * Shared Function > 3D Sampler > State > State Caching (page 96):
+    *
+    *    Coherency with system memory in the state cache, like the texture
+    *    cache is handled partially by software. It is expected that the
+    *    command stream or shader will issue Cache Flush operation or
+    *    Cache_Flush sampler message to ensure that the L1 cache remains
+    *    coherent with system memory.
+    *
+    *    [...]
+    *
+    *    Whenever the value of the Dynamic_State_Base_Addr,
+    *    Surface_State_Base_Addr are altered, the L1 state cache must be
+    *    invalidated to ensure the new surface or sampler state is fetched
+    *    from system memory.
+    *
+    * The PIPE_CONTROL command has a "State Cache Invalidation Enable" bit
+    * which, according the PIPE_CONTROL instruction documentation in the
+    * Broadwell PRM:
+    *
+    *    Setting this bit is independent of any other bit in this packet.
+    *    This bit controls the invalidation of the L1 and L2 state caches
+    *    at the top of the pipe i.e. at the parsing time.
+    *
+    * Unfortunately, experimentation seems to indicate that state cache
+    * invalidation through a PIPE_CONTROL does nothing whatsoever in
+    * regards to surface state and binding tables.  In stead, it seems that
+    * invalidating the texture cache is what is actually needed.
+    *
+    * XXX:  As far as we have been able to determine through
+    * experimentation, shows that flush the texture cache appears to be
+    * sufficient.  The theory here is that all of the sampling/rendering
+    * units cache the binding table in the texture cache.  However, we have
+    * yet to be able to actually confirm this.
+    */
+   anv_batch_emit(&cmd_buffer->batch, GEN8_PIPE_CONTROL,
+                  .TextureCacheInvalidationEnable = true);
 }
 
-static VkResult
-anv_reloc_list_grow(struct anv_reloc_list *list, struct anv_device *device,
-                    size_t num_additional_relocs)
+VkResult anv_BeginCommandBuffer(
+    VkCmdBuffer                                 cmdBuffer,
+    const VkCmdBufferBeginInfo*                 pBeginInfo)
 {
-   if (list->num_relocs + num_additional_relocs <= list->array_length)
-      return VK_SUCCESS;
+   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
 
-   size_t new_length = list->array_length * 2;
-   while (new_length < list->num_relocs + num_additional_relocs)
-      new_length *= 2;
+   cmd_buffer->opt_flags = pBeginInfo->flags;
 
-   struct drm_i915_gem_relocation_entry *new_relocs =
-      anv_device_alloc(device, new_length * sizeof(*list->relocs), 8,
-                       VK_SYSTEM_ALLOC_TYPE_INTERNAL);
-   if (new_relocs == NULL)
-      return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+   if (cmd_buffer->level == VK_CMD_BUFFER_LEVEL_SECONDARY) {
+      cmd_buffer->state.framebuffer =
+         anv_framebuffer_from_handle(pBeginInfo->framebuffer);
+      cmd_buffer->state.pass =
+         anv_render_pass_from_handle(pBeginInfo->renderPass);
 
-   struct anv_bo **new_reloc_bos =
-      anv_device_alloc(device, new_length * sizeof(*list->reloc_bos), 8,
-                       VK_SYSTEM_ALLOC_TYPE_INTERNAL);
-   if (new_relocs == NULL) {
-      anv_device_free(device, new_relocs);
-      return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+      /* FIXME: We shouldn't be starting on the first subpass */
+      anv_cmd_buffer_begin_subpass(cmd_buffer,
+                                   &cmd_buffer->state.pass->subpasses[0]);
    }
 
-   memcpy(new_relocs, list->relocs, list->num_relocs * sizeof(*list->relocs));
-   memcpy(new_reloc_bos, list->reloc_bos,
-          list->num_relocs * sizeof(*list->reloc_bos));
-
-   anv_device_free(device, list->relocs);
-   anv_device_free(device, list->reloc_bos);
-
-   list->array_length = new_length;
-   list->relocs = new_relocs;
-   list->reloc_bos = new_reloc_bos;
+   anv_cmd_buffer_emit_state_base_address(cmd_buffer);
+   cmd_buffer->state.current_pipeline = UINT32_MAX;
 
    return VK_SUCCESS;
 }
 
-uint64_t
-anv_reloc_list_add(struct anv_reloc_list *list, struct anv_device *device,
-                   uint32_t offset, struct anv_bo *target_bo, uint32_t delta)
+VkResult anv_EndCommandBuffer(
+    VkCmdBuffer                                 cmdBuffer)
 {
-   struct drm_i915_gem_relocation_entry *entry;
-   int index;
+   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
+   struct anv_device *device = cmd_buffer->device;
 
-   anv_reloc_list_grow(list, device, 1);
-   /* TODO: Handle failure */
+   anv_cmd_buffer_end_batch_buffer(cmd_buffer);
 
-   /* XXX: Can we use I915_EXEC_HANDLE_LUT? */
-   index = list->num_relocs++;
-   list->reloc_bos[index] = target_bo;
-   entry = &list->relocs[index];
-   entry->target_handle = target_bo->gem_handle;
-   entry->delta = delta;
-   entry->offset = offset;
-   entry->presumed_offset = target_bo->offset;
-   entry->read_domains = 0;
-   entry->write_domain = 0;
+   if (cmd_buffer->level == VK_CMD_BUFFER_LEVEL_PRIMARY) {
+      /* The algorithm used to compute the validate list is not threadsafe as
+       * it uses the bo->index field.  We have to lock the device around it.
+       * Fortunately, the chances for contention here are probably very low.
+       */
+      pthread_mutex_lock(&device->mutex);
+      anv_cmd_buffer_prepare_execbuf(cmd_buffer);
+      pthread_mutex_unlock(&device->mutex);
+   }
 
-   return target_bo->offset + delta;
+   return VK_SUCCESS;
 }
 
-static void
-anv_reloc_list_append(struct anv_reloc_list *list, struct anv_device *device,
-                      struct anv_reloc_list *other, uint32_t offset)
+void anv_CmdBindPipeline(
+    VkCmdBuffer                                 cmdBuffer,
+    VkPipelineBindPoint                         pipelineBindPoint,
+    VkPipeline                                  _pipeline)
 {
-   anv_reloc_list_grow(list, device, other->num_relocs);
-   /* TODO: Handle failure */
+   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
+   ANV_FROM_HANDLE(anv_pipeline, pipeline, _pipeline);
 
-   memcpy(&list->relocs[list->num_relocs], &other->relocs[0],
-          other->num_relocs * sizeof(other->relocs[0]));
-   memcpy(&list->reloc_bos[list->num_relocs], &other->reloc_bos[0],
-          other->num_relocs * sizeof(other->reloc_bos[0]));
+   switch (pipelineBindPoint) {
+   case VK_PIPELINE_BIND_POINT_COMPUTE:
+      cmd_buffer->state.compute_pipeline = pipeline;
+      cmd_buffer->state.compute_dirty |= ANV_CMD_BUFFER_PIPELINE_DIRTY;
+      break;
 
-   for (uint32_t i = 0; i < other->num_relocs; i++)
-      list->relocs[i + list->num_relocs].offset += offset;
+   case VK_PIPELINE_BIND_POINT_GRAPHICS:
+      cmd_buffer->state.pipeline = pipeline;
+      cmd_buffer->state.vb_dirty |= pipeline->vb_used;
+      cmd_buffer->state.dirty |= ANV_CMD_BUFFER_PIPELINE_DIRTY;
+      break;
 
-   list->num_relocs += other->num_relocs;
+   default:
+      assert(!"invalid bind point");
+      break;
+   }
 }
 
-/*-----------------------------------------------------------------------*
- * Functions related to anv_batch
- *-----------------------------------------------------------------------*/
+void anv_CmdBindDynamicViewportState(
+    VkCmdBuffer                                 cmdBuffer,
+    VkDynamicViewportState                      dynamicViewportState)
+{
+   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
+   ANV_FROM_HANDLE(anv_dynamic_vp_state, vp_state, dynamicViewportState);
 
-void *
-anv_batch_emit_dwords(struct anv_batch *batch, int num_dwords)
+   cmd_buffer->state.vp_state = vp_state;
+   cmd_buffer->state.dirty |= ANV_CMD_BUFFER_VP_DIRTY;
+}
+
+void anv_CmdBindDynamicRasterState(
+    VkCmdBuffer                                 cmdBuffer,
+    VkDynamicRasterState                        dynamicRasterState)
 {
-   if (batch->next + num_dwords * 4 > batch->end)
-      batch->extend_cb(batch, batch->user_data);
+   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
+   ANV_FROM_HANDLE(anv_dynamic_rs_state, rs_state, dynamicRasterState);
 
-   void *p = batch->next;
+   cmd_buffer->state.rs_state = rs_state;
+   cmd_buffer->state.dirty |= ANV_CMD_BUFFER_RS_DIRTY;
+}
 
-   batch->next += num_dwords * 4;
-   assert(batch->next <= batch->end);
+void anv_CmdBindDynamicColorBlendState(
+    VkCmdBuffer                                 cmdBuffer,
+    VkDynamicColorBlendState                    dynamicColorBlendState)
+{
+   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
+   ANV_FROM_HANDLE(anv_dynamic_cb_state, cb_state, dynamicColorBlendState);
 
-   return p;
+   cmd_buffer->state.cb_state = cb_state;
+   cmd_buffer->state.dirty |= ANV_CMD_BUFFER_CB_DIRTY;
 }
 
-uint64_t
-anv_batch_emit_reloc(struct anv_batch *batch,
-                     void *location, struct anv_bo *bo, uint32_t delta)
+void anv_CmdBindDynamicDepthStencilState(
+    VkCmdBuffer                                 cmdBuffer,
+    VkDynamicDepthStencilState                  dynamicDepthStencilState)
 {
-   return anv_reloc_list_add(batch->relocs, batch->device,
-                             location - batch->start, bo, delta);
+   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
+   ANV_FROM_HANDLE(anv_dynamic_ds_state, ds_state, dynamicDepthStencilState);
+
+   cmd_buffer->state.ds_state = ds_state;
+   cmd_buffer->state.dirty |= ANV_CMD_BUFFER_DS_DIRTY;
 }
 
-void
-anv_batch_emit_batch(struct anv_batch *batch, struct anv_batch *other)
+void anv_CmdBindDescriptorSets(
+    VkCmdBuffer                                 cmdBuffer,
+    VkPipelineBindPoint                         pipelineBindPoint,
+    VkPipelineLayout                            _layout,
+    uint32_t                                    firstSet,
+    uint32_t                                    setCount,
+    const VkDescriptorSet*                      pDescriptorSets,
+    uint32_t                                    dynamicOffsetCount,
+    const uint32_t*                             pDynamicOffsets)
 {
-   uint32_t size, offset;
+   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
+   ANV_FROM_HANDLE(anv_pipeline_layout, layout, _layout);
+   struct anv_descriptor_set_layout *set_layout;
 
-   size = other->next - other->start;
-   assert(size % 4 == 0);
+   assert(firstSet + setCount < MAX_SETS);
 
-   if (batch->next + size > batch->end)
-      batch->extend_cb(batch, batch->user_data);
+   uint32_t dynamic_slot = 0;
+   for (uint32_t i = 0; i < setCount; i++) {
+      ANV_FROM_HANDLE(anv_descriptor_set, set, pDescriptorSets[i]);
+      set_layout = layout->set[firstSet + i].layout;
 
-   assert(batch->next + size <= batch->end);
+      cmd_buffer->state.descriptors[firstSet + i].set = set;
 
-   VG(VALGRIND_CHECK_MEM_IS_DEFINED(other->start, size));
-   memcpy(batch->next, other->start, size);
+      assert(set_layout->num_dynamic_buffers <
+             ARRAY_SIZE(cmd_buffer->state.descriptors[0].dynamic_offsets));
+      memcpy(cmd_buffer->state.descriptors[firstSet + i].dynamic_offsets,
+             pDynamicOffsets + dynamic_slot,
+             set_layout->num_dynamic_buffers * sizeof(*pDynamicOffsets));
 
-   offset = batch->next - batch->start;
-   anv_reloc_list_append(batch->relocs, batch->device,
-                         other->relocs, offset);
+      cmd_buffer->state.descriptors_dirty |= set_layout->shader_stages;
 
-   batch->next += size;
+      dynamic_slot += set_layout->num_dynamic_buffers;
+   }
 }
 
-/*-----------------------------------------------------------------------*
- * Functions related to anv_batch_bo
- *-----------------------------------------------------------------------*/
-
-static VkResult
-anv_batch_bo_create(struct anv_device *device, struct anv_batch_bo **bbo_out)
+void anv_CmdBindIndexBuffer(
+    VkCmdBuffer                                 cmdBuffer,
+    VkBuffer                                    _buffer,
+    VkDeviceSize                                offset,
+    VkIndexType                                 indexType)
 {
-   VkResult result;
+   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
+   ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
 
-   struct anv_batch_bo *bbo =
-      anv_device_alloc(device, sizeof(*bbo), 8, VK_SYSTEM_ALLOC_TYPE_INTERNAL);
-   if (bbo == NULL)
-      return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+   static const uint32_t vk_to_gen_index_type[] = {
+      [VK_INDEX_TYPE_UINT16]                    = INDEX_WORD,
+      [VK_INDEX_TYPE_UINT32]                    = INDEX_DWORD,
+   };
 
-   result = anv_bo_pool_alloc(&device->batch_bo_pool, &bbo->bo);
-   if (result != VK_SUCCESS)
-      goto fail_alloc;
+   struct GEN8_3DSTATE_VF vf = {
+      GEN8_3DSTATE_VF_header,
+      .CutIndex = (indexType == VK_INDEX_TYPE_UINT16) ? UINT16_MAX : UINT32_MAX,
+   };
+   GEN8_3DSTATE_VF_pack(NULL, cmd_buffer->state.state_vf, &vf);
 
-   result = anv_reloc_list_init(&bbo->relocs, device);
-   if (result != VK_SUCCESS)
-      goto fail_bo_alloc;
+   cmd_buffer->state.dirty |= ANV_CMD_BUFFER_INDEX_BUFFER_DIRTY;
 
-   *bbo_out = bbo;
+   anv_batch_emit(&cmd_buffer->batch, GEN8_3DSTATE_INDEX_BUFFER,
+                  .IndexFormat = vk_to_gen_index_type[indexType],
+                  .MemoryObjectControlState = GEN8_MOCS,
+                  .BufferStartingAddress = { buffer->bo, buffer->offset + offset },
+                  .BufferSize = buffer->size - offset);
+}
 
-   return VK_SUCCESS;
+void anv_CmdBindVertexBuffers(
+    VkCmdBuffer                                 cmdBuffer,
+    uint32_t                                    startBinding,
+    uint32_t                                    bindingCount,
+    const VkBuffer*                             pBuffers,
+    const VkDeviceSize*                         pOffsets)
+{
+   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
+   struct anv_vertex_binding *vb = cmd_buffer->state.vertex_bindings;
 
- fail_bo_alloc:
-   anv_bo_pool_free(&device->batch_bo_pool, &bbo->bo);
- fail_alloc:
-   anv_device_free(device, bbo);
+   /* We have to defer setting up vertex buffer since we need the buffer
+    * stride from the pipeline. */
 
-   return result;
+   assert(startBinding + bindingCount < MAX_VBS);
+   for (uint32_t i = 0; i < bindingCount; i++) {
+      vb[startBinding + i].buffer = anv_buffer_from_handle(pBuffers[i]);
+      vb[startBinding + i].offset = pOffsets[i];
+      cmd_buffer->state.vb_dirty |= 1 << (startBinding + i);
+   }
 }
 
 static VkResult
-anv_batch_bo_clone(struct anv_device *device,
-                   const struct anv_batch_bo *other_bbo,
-                   struct anv_batch_bo **bbo_out)
+cmd_buffer_emit_binding_table(struct anv_cmd_buffer *cmd_buffer,
+                              unsigned stage, struct anv_state *bt_state)
 {
-   VkResult result;
+   struct anv_framebuffer *fb = cmd_buffer->state.framebuffer;
+   struct anv_subpass *subpass = cmd_buffer->state.subpass;
+   struct anv_pipeline_layout *layout;
+   uint32_t attachments, bias, size;
+
+   if (stage == VK_SHADER_STAGE_COMPUTE)
+      layout = cmd_buffer->state.compute_pipeline->layout;
+   else
+      layout = cmd_buffer->state.pipeline->layout;
+
+   if (stage == VK_SHADER_STAGE_FRAGMENT) {
+      bias = MAX_RTS;
+      attachments = subpass->color_count;
+   } else {
+      bias = 0;
+      attachments = 0;
+   }
 
-   struct anv_batch_bo *bbo =
-      anv_device_alloc(device, sizeof(*bbo), 8, VK_SYSTEM_ALLOC_TYPE_INTERNAL);
-   if (bbo == NULL)
-      return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+   /* This is a little awkward: layout can be NULL but we still have to
+    * allocate and set a binding table for the PS stage for render
+    * targets. */
+   uint32_t surface_count = layout ? layout->stage[stage].surface_count : 0;
 
-   result = anv_bo_pool_alloc(&device->batch_bo_pool, &bbo->bo);
-   if (result != VK_SUCCESS)
-      goto fail_alloc;
+   if (attachments + surface_count == 0)
+      return VK_SUCCESS;
 
-   result = anv_reloc_list_init_clone(&bbo->relocs, device, &other_bbo->relocs);
-   if (result != VK_SUCCESS)
-      goto fail_bo_alloc;
+   size = (bias + surface_count) * sizeof(uint32_t);
+   *bt_state = anv_cmd_buffer_alloc_surface_state(cmd_buffer, size, 32);
+   uint32_t *bt_map = bt_state->map;
 
-   bbo->length = other_bbo->length;
-   memcpy(bbo->bo.map, other_bbo->bo.map, other_bbo->length);
+   if (bt_state->map == NULL)
+      return VK_ERROR_OUT_OF_DEVICE_MEMORY;
 
-   *bbo_out = bbo;
+   /* This is highly annoying.  The Vulkan spec puts the depth-stencil
+    * attachments in with the color attachments.  Unfortunately, thanks to
+    * other aspects of the API, we cana't really saparate them before this
+    * point.  Therefore, we have to walk all of the attachments but only
+    * put the color attachments into the binding table.
+    */
+   for (uint32_t a = 0; a < attachments; a++) {
+      const struct anv_attachment_view *attachment =
+         fb->attachments[subpass->color_attachments[a]];
 
-   return VK_SUCCESS;
+      assert(attachment->attachment_type == ANV_ATTACHMENT_VIEW_TYPE_COLOR);
+      const struct anv_color_attachment_view *view =
+         (const struct anv_color_attachment_view *)attachment;
 
- fail_bo_alloc:
-   anv_bo_pool_free(&device->batch_bo_pool, &bbo->bo);
- fail_alloc:
-   anv_device_free(device, bbo);
+      struct anv_state state =
+         anv_cmd_buffer_alloc_surface_state(cmd_buffer, 64, 64);
 
-   return result;
-}
+      if (state.map == NULL)
+         return VK_ERROR_OUT_OF_DEVICE_MEMORY;
 
-static void
-anv_batch_bo_start(struct anv_batch_bo *bbo, struct anv_batch *batch,
-                   size_t batch_padding)
-{
-   batch->next = batch->start = bbo->bo.map;
-   batch->end = bbo->bo.map + bbo->bo.size - batch_padding;
-   batch->relocs = &bbo->relocs;
-   bbo->relocs.num_relocs = 0;
+      memcpy(state.map, view->view.surface_state.map, 64);
+
+      /* The address goes in dwords 8 and 9 of the SURFACE_STATE */
+      *(uint64_t *)(state.map + 8 * 4) =
+         anv_reloc_list_add(anv_cmd_buffer_current_surface_relocs(cmd_buffer),
+                            cmd_buffer->device,
+                            state.offset + 8 * 4,
+                            view->view.bo, view->view.offset);
+
+      bt_map[a] = state.offset;
+   }
+
+   if (layout == NULL)
+      return VK_SUCCESS;
+
+   for (uint32_t set = 0; set < layout->num_sets; set++) {
+      struct anv_descriptor_set_binding *d = &cmd_buffer->state.descriptors[set];
+      struct anv_descriptor_set_layout *set_layout = layout->set[set].layout;
+      struct anv_descriptor_slot *surface_slots =
+         set_layout->stage[stage].surface_start;
+
+      uint32_t start = bias + layout->set[set].surface_start[stage];
+
+      for (uint32_t b = 0; b < set_layout->stage[stage].surface_count; b++) {
+         struct anv_surface_view *view =
+            d->set->descriptors[surface_slots[b].index].view;
+
+         if (!view)
+            continue;
+
+         struct anv_state state =
+            anv_cmd_buffer_alloc_surface_state(cmd_buffer, 64, 64);
+
+         if (state.map == NULL)
+            return VK_ERROR_OUT_OF_DEVICE_MEMORY;
+
+         uint32_t offset;
+         if (surface_slots[b].dynamic_slot >= 0) {
+            uint32_t dynamic_offset =
+               d->dynamic_offsets[surface_slots[b].dynamic_slot];
+
+            offset = view->offset + dynamic_offset;
+            anv_fill_buffer_surface_state(state.map, view->format, offset,
+                                          view->range - dynamic_offset);
+         } else {
+            offset = view->offset;
+            memcpy(state.map, view->surface_state.map, 64);
+         }
+
+         /* The address goes in dwords 8 and 9 of the SURFACE_STATE */
+         *(uint64_t *)(state.map + 8 * 4) =
+            anv_reloc_list_add(anv_cmd_buffer_current_surface_relocs(cmd_buffer),
+                               cmd_buffer->device,
+                               state.offset + 8 * 4,
+                               view->bo, offset);
+
+         bt_map[start + b] = state.offset;
+      }
+   }
+
+   return VK_SUCCESS;
 }
 
-static void
-anv_batch_bo_continue(struct anv_batch_bo *bbo, struct anv_batch *batch,
-                      size_t batch_padding)
+static VkResult
+cmd_buffer_emit_samplers(struct anv_cmd_buffer *cmd_buffer,
+                         unsigned stage, struct anv_state *state)
 {
-   batch->start = bbo->bo.map;
-   batch->next = bbo->bo.map + bbo->length;
-   batch->end = bbo->bo.map + bbo->bo.size - batch_padding;
-   batch->relocs = &bbo->relocs;
+   struct anv_pipeline_layout *layout;
+   uint32_t sampler_count;
+
+   if (stage == VK_SHADER_STAGE_COMPUTE)
+      layout = cmd_buffer->state.compute_pipeline->layout;
+   else
+      layout = cmd_buffer->state.pipeline->layout;
+
+   sampler_count = layout ? layout->stage[stage].sampler_count : 0;
+   if (sampler_count == 0)
+      return VK_SUCCESS;
+
+   uint32_t size = sampler_count * 16;
+   *state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, size, 32);
+
+   if (state->map == NULL)
+      return VK_ERROR_OUT_OF_DEVICE_MEMORY;
+
+   for (uint32_t set = 0; set < layout->num_sets; set++) {
+      struct anv_descriptor_set_binding *d = &cmd_buffer->state.descriptors[set];
+      struct anv_descriptor_set_layout *set_layout = layout->set[set].layout;
+      struct anv_descriptor_slot *sampler_slots =
+         set_layout->stage[stage].sampler_start;
+
+      uint32_t start = layout->set[set].sampler_start[stage];
+
+      for (uint32_t b = 0; b < set_layout->stage[stage].sampler_count; b++) {
+         struct anv_sampler *sampler =
+            d->set->descriptors[sampler_slots[b].index].sampler;
+
+         if (!sampler)
+            continue;
+
+         memcpy(state->map + (start + b) * 16,
+                sampler->state, sizeof(sampler->state));
+      }
+   }
+
+   return VK_SUCCESS;
 }
 
-static void
-anv_batch_bo_finish(struct anv_batch_bo *bbo, struct anv_batch *batch)
+static VkResult
+flush_descriptor_set(struct anv_cmd_buffer *cmd_buffer, uint32_t stage)
 {
-   assert(batch->start == bbo->bo.map);
-   bbo->length = batch->next - batch->start;
-   VG(VALGRIND_CHECK_MEM_IS_DEFINED(batch->start, bbo->length));
+   struct anv_state surfaces = { 0, }, samplers = { 0, };
+   VkResult result;
+
+   result = cmd_buffer_emit_samplers(cmd_buffer, stage, &samplers);
+   if (result != VK_SUCCESS)
+      return result;
+   result = cmd_buffer_emit_binding_table(cmd_buffer, stage, &surfaces);
+   if (result != VK_SUCCESS)
+      return result;
+
+   static const uint32_t sampler_state_opcodes[] = {
+      [VK_SHADER_STAGE_VERTEX]                  = 43,
+      [VK_SHADER_STAGE_TESS_CONTROL]            = 44, /* HS */
+      [VK_SHADER_STAGE_TESS_EVALUATION]         = 45, /* DS */
+      [VK_SHADER_STAGE_GEOMETRY]                = 46,
+      [VK_SHADER_STAGE_FRAGMENT]                = 47,
+      [VK_SHADER_STAGE_COMPUTE]                 = 0,
+   };
+
+   static const uint32_t binding_table_opcodes[] = {
+      [VK_SHADER_STAGE_VERTEX]                  = 38,
+      [VK_SHADER_STAGE_TESS_CONTROL]            = 39,
+      [VK_SHADER_STAGE_TESS_EVALUATION]         = 40,
+      [VK_SHADER_STAGE_GEOMETRY]                = 41,
+      [VK_SHADER_STAGE_FRAGMENT]                = 42,
+      [VK_SHADER_STAGE_COMPUTE]                 = 0,
+   };
+
+   if (samplers.alloc_size > 0) {
+      anv_batch_emit(&cmd_buffer->batch,
+                     GEN8_3DSTATE_SAMPLER_STATE_POINTERS_VS,
+                     ._3DCommandSubOpcode  = sampler_state_opcodes[stage],
+                     .PointertoVSSamplerState = samplers.offset);
+   }
+
+   if (surfaces.alloc_size > 0) {
+      anv_batch_emit(&cmd_buffer->batch,
+                     GEN8_3DSTATE_BINDING_TABLE_POINTERS_VS,
+                     ._3DCommandSubOpcode  = binding_table_opcodes[stage],
+                     .PointertoVSBindingTable = surfaces.offset);
+   }
+
+   return VK_SUCCESS;
 }
 
 static void
-anv_batch_bo_destroy(struct anv_batch_bo *bbo, struct anv_device *device)
+flush_descriptor_sets(struct anv_cmd_buffer *cmd_buffer)
 {
-   anv_reloc_list_finish(&bbo->relocs, device);
-   anv_bo_pool_free(&device->batch_bo_pool, &bbo->bo);
-   anv_device_free(device, bbo);
-}
+   uint32_t s, dirty = cmd_buffer->state.descriptors_dirty &
+                       cmd_buffer->state.pipeline->active_stages;
 
-static VkResult
-anv_batch_bo_list_clone(const struct list_head *list, struct anv_device *device,
-                        struct list_head *new_list)
-{
    VkResult result = VK_SUCCESS;
-
-   list_inithead(new_list);
-
-   struct anv_batch_bo *prev_bbo = NULL;
-   list_for_each_entry(struct anv_batch_bo, bbo, list, link) {
-      struct anv_batch_bo *new_bbo;
-      result = anv_batch_bo_clone(device, bbo, &new_bbo);
+   for_each_bit(s, dirty) {
+      result = flush_descriptor_set(cmd_buffer, s);
       if (result != VK_SUCCESS)
          break;
-      list_addtail(&new_bbo->link, new_list);
-
-      if (prev_bbo) {
-         /* As we clone this list of batch_bo's, they chain one to the
-          * other using MI_BATCH_BUFFER_START commands.  We need to fix up
-          * those relocations as we go.  Fortunately, this is pretty easy
-          * as it will always be the last relocation in the list.
-          */
-         uint32_t last_idx = prev_bbo->relocs.num_relocs - 1;
-         assert(prev_bbo->relocs.reloc_bos[last_idx] == &bbo->bo);
-         prev_bbo->relocs.reloc_bos[last_idx] = &new_bbo->bo;
-      }
-
-      prev_bbo = new_bbo;
    }
 
    if (result != VK_SUCCESS) {
-      list_for_each_entry_safe(struct anv_batch_bo, bbo, new_list, link)
-         anv_batch_bo_destroy(bbo, device);
-   }
+      assert(result == VK_ERROR_OUT_OF_DEVICE_MEMORY);
 
-   return result;
-}
+      result = anv_cmd_buffer_new_surface_state_bo(cmd_buffer);
+      assert(result == VK_SUCCESS);
 
-/*-----------------------------------------------------------------------*
- * Functions related to anv_batch_bo
- *-----------------------------------------------------------------------*/
+      /* Re-emit state base addresses so we get the new surface state base
+       * address before we start emitting binding tables etc.
+       */
+      anv_cmd_buffer_emit_state_base_address(cmd_buffer);
 
-static inline struct anv_batch_bo *
-anv_cmd_buffer_current_batch_bo(struct anv_cmd_buffer *cmd_buffer)
-{
-   return LIST_ENTRY(struct anv_batch_bo, cmd_buffer->batch_bos.prev, link);
-}
+      /* Re-emit all active binding tables */
+      for_each_bit(s, cmd_buffer->state.pipeline->active_stages) {
+         result = flush_descriptor_set(cmd_buffer, s);
 
-static inline struct anv_batch_bo *
-anv_cmd_buffer_current_surface_bbo(struct anv_cmd_buffer *cmd_buffer)
-{
-   return LIST_ENTRY(struct anv_batch_bo, cmd_buffer->surface_bos.prev, link);
-}
+         /* It had better succeed this time */
+         assert(result == VK_SUCCESS);
+      }
+   }
 
-struct anv_bo *
-anv_cmd_buffer_current_surface_bo(struct anv_cmd_buffer *cmd_buffer)
-{
-   return &anv_cmd_buffer_current_surface_bbo(cmd_buffer)->bo;
+   cmd_buffer->state.descriptors_dirty &= ~cmd_buffer->state.pipeline->active_stages;
 }
 
-struct anv_reloc_list *
-anv_cmd_buffer_current_surface_relocs(struct anv_cmd_buffer *cmd_buffer)
+static struct anv_state
+anv_cmd_buffer_emit_dynamic(struct anv_cmd_buffer *cmd_buffer,
+                             uint32_t *a, uint32_t dwords, uint32_t alignment)
 {
-   return &anv_cmd_buffer_current_surface_bbo(cmd_buffer)->relocs;
+   struct anv_state state;
+
+   state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
+                                              dwords * 4, alignment);
+   memcpy(state.map, a, dwords * 4);
+
+   VG(VALGRIND_CHECK_MEM_IS_DEFINED(state.map, dwords * 4));
+
+   return state;
 }
 
-static void
-cmd_buffer_chain_to_batch_bo(struct anv_cmd_buffer *cmd_buffer,
-                             struct anv_batch_bo *bbo)
+static struct anv_state
+anv_cmd_buffer_merge_dynamic(struct anv_cmd_buffer *cmd_buffer,
+                             uint32_t *a, uint32_t *b,
+                             uint32_t dwords, uint32_t alignment)
 {
-   struct anv_batch *batch = &cmd_buffer->batch;
-   struct anv_batch_bo *current_bbo =
-      anv_cmd_buffer_current_batch_bo(cmd_buffer);
+   struct anv_state state;
+   uint32_t *p;
 
-   /* We set the end of the batch a little short so we would be sure we
-    * have room for the chaining command.  Since we're about to emit the
-    * chaining command, let's set it back where it should go.
-    */
-   batch->end += GEN8_MI_BATCH_BUFFER_START_length * 4;
-   assert(batch->end == current_bbo->bo.map + current_bbo->bo.size);
+   state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
+                                              dwords * 4, alignment);
+   p = state.map;
+   for (uint32_t i = 0; i < dwords; i++)
+      p[i] = a[i] | b[i];
 
-   anv_batch_emit(batch, GEN8_MI_BATCH_BUFFER_START,
-      GEN8_MI_BATCH_BUFFER_START_header,
-      ._2ndLevelBatchBuffer = _1stlevelbatch,
-      .AddressSpaceIndicator = ASI_PPGTT,
-      .BatchBufferStartAddress = { &bbo->bo, 0 },
-   );
+   VG(VALGRIND_CHECK_MEM_IS_DEFINED(p, dwords * 4));
 
-   anv_batch_bo_finish(current_bbo, batch);
+   return state;
 }
 
 static VkResult
-anv_cmd_buffer_chain_batch(struct anv_batch *batch, void *_data)
+flush_compute_descriptor_set(struct anv_cmd_buffer *cmd_buffer)
 {
-   struct anv_cmd_buffer *cmd_buffer = _data;
-   struct anv_batch_bo *new_bbo;
+   struct anv_device *device = cmd_buffer->device;
+   struct anv_pipeline *pipeline = cmd_buffer->state.compute_pipeline;
+   struct anv_state surfaces = { 0, }, samplers = { 0, };
+   VkResult result;
 
-   VkResult result = anv_batch_bo_create(cmd_buffer->device, &new_bbo);
+   result = cmd_buffer_emit_samplers(cmd_buffer,
+                                     VK_SHADER_STAGE_COMPUTE, &samplers);
+   if (result != VK_SUCCESS)
+      return result;
+   result = cmd_buffer_emit_binding_table(cmd_buffer,
+                                          VK_SHADER_STAGE_COMPUTE, &surfaces);
    if (result != VK_SUCCESS)
       return result;
 
-   struct anv_batch_bo **seen_bbo = anv_vector_add(&cmd_buffer->seen_bbos);
-   if (seen_bbo == NULL) {
-      anv_batch_bo_destroy(new_bbo, cmd_buffer->device);
-      return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
-   }
-   *seen_bbo = new_bbo;
+   struct GEN8_INTERFACE_DESCRIPTOR_DATA desc = {
+      .KernelStartPointer = pipeline->cs_simd,
+      .KernelStartPointerHigh = 0,
+      .BindingTablePointer = surfaces.offset,
+      .BindingTableEntryCount = 0,
+      .SamplerStatePointer = samplers.offset,
+      .SamplerCount = 0,
+      .NumberofThreadsinGPGPUThreadGroup = 0 /* FIXME: Really? */
+   };
 
-   cmd_buffer_chain_to_batch_bo(cmd_buffer, new_bbo);
+   uint32_t size = GEN8_INTERFACE_DESCRIPTOR_DATA_length * sizeof(uint32_t);
+   struct anv_state state =
+      anv_state_pool_alloc(&device->dynamic_state_pool, size, 64);
 
-   list_addtail(&new_bbo->link, &cmd_buffer->batch_bos);
+   GEN8_INTERFACE_DESCRIPTOR_DATA_pack(NULL, state.map, &desc);
 
-   anv_batch_bo_start(new_bbo, batch, GEN8_MI_BATCH_BUFFER_START_length * 4);
+   anv_batch_emit(&cmd_buffer->batch, GEN8_MEDIA_INTERFACE_DESCRIPTOR_LOAD,
+                  .InterfaceDescriptorTotalLength = size,
+                  .InterfaceDescriptorDataStartAddress = state.offset);
 
    return VK_SUCCESS;
 }
 
-struct anv_state
-anv_cmd_buffer_alloc_surface_state(struct anv_cmd_buffer *cmd_buffer,
-                                   uint32_t size, uint32_t alignment)
+static void
+anv_cmd_buffer_flush_compute_state(struct anv_cmd_buffer *cmd_buffer)
 {
-   struct anv_bo *surface_bo =
-      anv_cmd_buffer_current_surface_bo(cmd_buffer);
-   struct anv_state state;
+   struct anv_pipeline *pipeline = cmd_buffer->state.compute_pipeline;
+   VkResult result;
 
-   state.offset = align_u32(cmd_buffer->surface_next, alignment);
-   if (state.offset + size > surface_bo->size)
-      return (struct anv_state) { 0 };
+   assert(pipeline->active_stages == VK_SHADER_STAGE_COMPUTE_BIT);
 
-   state.map = surface_bo->map + state.offset;
-   state.alloc_size = size;
-   cmd_buffer->surface_next = state.offset + size;
+   if (cmd_buffer->state.current_pipeline != GPGPU) {
+      anv_batch_emit(&cmd_buffer->batch, GEN8_PIPELINE_SELECT,
+                     .PipelineSelection = GPGPU);
+      cmd_buffer->state.current_pipeline = GPGPU;
+   }
 
-   assert(state.offset + size <= surface_bo->size);
+   if (cmd_buffer->state.compute_dirty & ANV_CMD_BUFFER_PIPELINE_DIRTY)
+      anv_batch_emit_batch(&cmd_buffer->batch, &pipeline->batch);
 
-   return state;
-}
+   if ((cmd_buffer->state.descriptors_dirty & VK_SHADER_STAGE_COMPUTE_BIT) ||
+       (cmd_buffer->state.compute_dirty & ANV_CMD_BUFFER_PIPELINE_DIRTY)) {
+      result = flush_compute_descriptor_set(cmd_buffer);
+      assert(result == VK_SUCCESS);
+      cmd_buffer->state.descriptors_dirty &= ~VK_SHADER_STAGE_COMPUTE;
+   }
 
-struct anv_state
-anv_cmd_buffer_alloc_dynamic_state(struct anv_cmd_buffer *cmd_buffer,
-                                   uint32_t size, uint32_t alignment)
-{
-   return anv_state_stream_alloc(&cmd_buffer->dynamic_state_stream,
-                                 size, alignment);
+   cmd_buffer->state.compute_dirty = 0;
 }
 
-VkResult
-anv_cmd_buffer_new_surface_state_bo(struct anv_cmd_buffer *cmd_buffer)
+static void
+anv_cmd_buffer_flush_state(struct anv_cmd_buffer *cmd_buffer)
 {
-   struct anv_batch_bo *new_bbo, *old_bbo =
-      anv_cmd_buffer_current_surface_bbo(cmd_buffer);
+   struct anv_pipeline *pipeline = cmd_buffer->state.pipeline;
+   uint32_t *p;
 
-   /* Finish off the old buffer */
-   old_bbo->length = cmd_buffer->surface_next;
+   uint32_t vb_emit = cmd_buffer->state.vb_dirty & pipeline->vb_used;
 
-   VkResult result = anv_batch_bo_create(cmd_buffer->device, &new_bbo);
-   if (result != VK_SUCCESS)
-      return result;
+   assert((pipeline->active_stages & VK_SHADER_STAGE_COMPUTE_BIT) == 0);
 
-   struct anv_batch_bo **seen_bbo = anv_vector_add(&cmd_buffer->seen_bbos);
-   if (seen_bbo == NULL) {
-      anv_batch_bo_destroy(new_bbo, cmd_buffer->device);
-      return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+   if (cmd_buffer->state.current_pipeline != _3D) {
+      anv_batch_emit(&cmd_buffer->batch, GEN8_PIPELINE_SELECT,
+                     .PipelineSelection = _3D);
+      cmd_buffer->state.current_pipeline = _3D;
    }
-   *seen_bbo = new_bbo;
 
-   cmd_buffer->surface_next = 1;
+   if (vb_emit) {
+      const uint32_t num_buffers = __builtin_popcount(vb_emit);
+      const uint32_t num_dwords = 1 + num_buffers * 4;
+
+      p = anv_batch_emitn(&cmd_buffer->batch, num_dwords,
+                          GEN8_3DSTATE_VERTEX_BUFFERS);
+      uint32_t vb, i = 0;
+      for_each_bit(vb, vb_emit) {
+         struct anv_buffer *buffer = cmd_buffer->state.vertex_bindings[vb].buffer;
+         uint32_t offset = cmd_buffer->state.vertex_bindings[vb].offset;
+
+         struct GEN8_VERTEX_BUFFER_STATE state = {
+            .VertexBufferIndex = vb,
+            .MemoryObjectControlState = GEN8_MOCS,
+            .AddressModifyEnable = true,
+            .BufferPitch = pipeline->binding_stride[vb],
+            .BufferStartingAddress = { buffer->bo, buffer->offset + offset },
+            .BufferSize = buffer->size - offset
+         };
+
+         GEN8_VERTEX_BUFFER_STATE_pack(&cmd_buffer->batch, &p[1 + i * 4], &state);
+         i++;
+      }
+   }
 
-   list_addtail(&new_bbo->link, &cmd_buffer->surface_bos);
+   if (cmd_buffer->state.dirty & ANV_CMD_BUFFER_PIPELINE_DIRTY) {
+      /* If somebody compiled a pipeline after starting a command buffer the
+       * scratch bo may have grown since we started this cmd buffer (and
+       * emitted STATE_BASE_ADDRESS).  If we're binding that pipeline now,
+       * reemit STATE_BASE_ADDRESS so that we use the bigger scratch bo. */
+      if (cmd_buffer->state.scratch_size < pipeline->total_scratch)
+         anv_cmd_buffer_emit_state_base_address(cmd_buffer);
 
-   return VK_SUCCESS;
-}
+      anv_batch_emit_batch(&cmd_buffer->batch, &pipeline->batch);
+   }
 
-VkResult
-anv_cmd_buffer_init_batch_bo_chain(struct anv_cmd_buffer *cmd_buffer)
-{
-   struct anv_batch_bo *batch_bo, *surface_bbo;
-   struct anv_device *device = cmd_buffer->device;
-   VkResult result;
+   if (cmd_buffer->state.descriptors_dirty)
+      flush_descriptor_sets(cmd_buffer);
+
+   if (cmd_buffer->state.dirty & ANV_CMD_BUFFER_VP_DIRTY) {
+      struct anv_dynamic_vp_state *vp_state = cmd_buffer->state.vp_state;
+      anv_batch_emit(&cmd_buffer->batch, GEN8_3DSTATE_SCISSOR_STATE_POINTERS,
+                     .ScissorRectPointer = vp_state->scissor.offset);
+      anv_batch_emit(&cmd_buffer->batch, GEN8_3DSTATE_VIEWPORT_STATE_POINTERS_CC,
+                     .CCViewportPointer = vp_state->cc_vp.offset);
+      anv_batch_emit(&cmd_buffer->batch, GEN8_3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP,
+                     .SFClipViewportPointer = vp_state->sf_clip_vp.offset);
+   }
+
+   if (cmd_buffer->state.dirty & (ANV_CMD_BUFFER_PIPELINE_DIRTY |
+                                  ANV_CMD_BUFFER_RS_DIRTY)) {
+      anv_batch_emit_merge(&cmd_buffer->batch,
+                           cmd_buffer->state.rs_state->state_sf,
+                           pipeline->state_sf);
+      anv_batch_emit_merge(&cmd_buffer->batch,
+                           cmd_buffer->state.rs_state->state_raster,
+                           pipeline->state_raster);
+   }
 
-   list_inithead(&cmd_buffer->batch_bos);
-   list_inithead(&cmd_buffer->surface_bos);
+   if (cmd_buffer->state.ds_state &&
+       (cmd_buffer->state.dirty & (ANV_CMD_BUFFER_PIPELINE_DIRTY |
+                                   ANV_CMD_BUFFER_DS_DIRTY))) {
+      anv_batch_emit_merge(&cmd_buffer->batch,
+                           cmd_buffer->state.ds_state->state_wm_depth_stencil,
+                           pipeline->state_wm_depth_stencil);
+   }
 
-   result = anv_batch_bo_create(device, &batch_bo);
-   if (result != VK_SUCCESS)
-      return result;
+   if (cmd_buffer->state.dirty & (ANV_CMD_BUFFER_CB_DIRTY |
+                                  ANV_CMD_BUFFER_DS_DIRTY)) {
+      struct anv_state state;
+      if (cmd_buffer->state.ds_state == NULL)
+         state = anv_cmd_buffer_emit_dynamic(cmd_buffer,
+                                             cmd_buffer->state.cb_state->state_color_calc,
+                                             GEN8_COLOR_CALC_STATE_length, 64);
+      else if (cmd_buffer->state.cb_state == NULL)
+         state = anv_cmd_buffer_emit_dynamic(cmd_buffer,
+                                             cmd_buffer->state.ds_state->state_color_calc,
+                                             GEN8_COLOR_CALC_STATE_length, 64);
+      else
+         state = anv_cmd_buffer_merge_dynamic(cmd_buffer,
+                                              cmd_buffer->state.ds_state->state_color_calc,
+                                              cmd_buffer->state.cb_state->state_color_calc,
+                                              GEN8_COLOR_CALC_STATE_length, 64);
+
+      anv_batch_emit(&cmd_buffer->batch,
+                     GEN8_3DSTATE_CC_STATE_POINTERS,
+                     .ColorCalcStatePointer = state.offset,
+                     .ColorCalcStatePointerValid = true);
+   }
 
-   list_addtail(&batch_bo->link, &cmd_buffer->batch_bos);
+   if (cmd_buffer->state.dirty & (ANV_CMD_BUFFER_PIPELINE_DIRTY |
+                                  ANV_CMD_BUFFER_INDEX_BUFFER_DIRTY)) {
+      anv_batch_emit_merge(&cmd_buffer->batch,
+                           cmd_buffer->state.state_vf, pipeline->state_vf);
+   }
 
-   cmd_buffer->batch.device = device;
-   cmd_buffer->batch.extend_cb = anv_cmd_buffer_chain_batch;
-   cmd_buffer->batch.user_data = cmd_buffer;
+   cmd_buffer->state.vb_dirty &= ~vb_emit;
+   cmd_buffer->state.dirty = 0;
+}
 
-   anv_batch_bo_start(batch_bo, &cmd_buffer->batch,
-                      GEN8_MI_BATCH_BUFFER_START_length * 4);
+void anv_CmdDraw(
+    VkCmdBuffer                                 cmdBuffer,
+    uint32_t                                    firstVertex,
+    uint32_t                                    vertexCount,
+    uint32_t                                    firstInstance,
+    uint32_t                                    instanceCount)
+{
+   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
 
-   result = anv_batch_bo_create(device, &surface_bbo);
-   if (result != VK_SUCCESS)
-      goto fail_batch_bo;
+   anv_cmd_buffer_flush_state(cmd_buffer);
 
-   list_addtail(&surface_bbo->link, &cmd_buffer->surface_bos);
+   anv_batch_emit(&cmd_buffer->batch, GEN8_3DPRIMITIVE,
+                  .VertexAccessType = SEQUENTIAL,
+                  .VertexCountPerInstance = vertexCount,
+                  .StartVertexLocation = firstVertex,
+                  .InstanceCount = instanceCount,
+                  .StartInstanceLocation = firstInstance,
+                  .BaseVertexLocation = 0);
+}
 
-   int success = anv_vector_init(&cmd_buffer->seen_bbos,
-                                 sizeof(struct anv_bo *),
-                                 8 * sizeof(struct anv_bo *));
-   if (!success)
-      goto fail_surface_bo;
+void anv_CmdDrawIndexed(
+    VkCmdBuffer                                 cmdBuffer,
+    uint32_t                                    firstIndex,
+    uint32_t                                    indexCount,
+    int32_t                                     vertexOffset,
+    uint32_t                                    firstInstance,
+    uint32_t                                    instanceCount)
+{
+   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
 
-   *(struct anv_batch_bo **)anv_vector_add(&cmd_buffer->seen_bbos) = batch_bo;
-   *(struct anv_batch_bo **)anv_vector_add(&cmd_buffer->seen_bbos) = surface_bbo;
+   anv_cmd_buffer_flush_state(cmd_buffer);
 
-   /* Start surface_next at 1 so surface offset 0 is invalid. */
-   cmd_buffer->surface_next = 1;
+   anv_batch_emit(&cmd_buffer->batch, GEN8_3DPRIMITIVE,
+                  .VertexAccessType = RANDOM,
+                  .VertexCountPerInstance = indexCount,
+                  .StartVertexLocation = firstIndex,
+                  .InstanceCount = instanceCount,
+                  .StartInstanceLocation = firstInstance,
+                  .BaseVertexLocation = vertexOffset);
+}
 
-   cmd_buffer->execbuf2.objects = NULL;
-   cmd_buffer->execbuf2.bos = NULL;
-   cmd_buffer->execbuf2.array_length = 0;
+static void
+anv_batch_lrm(struct anv_batch *batch,
+              uint32_t reg, struct anv_bo *bo, uint32_t offset)
+{
+   anv_batch_emit(batch, GEN8_MI_LOAD_REGISTER_MEM,
+                  .RegisterAddress = reg,
+                  .MemoryAddress = { bo, offset });
+}
 
-   return VK_SUCCESS;
+static void
+anv_batch_lri(struct anv_batch *batch, uint32_t reg, uint32_t imm)
+{
+   anv_batch_emit(batch, GEN8_MI_LOAD_REGISTER_IMM,
+                  .RegisterOffset = reg,
+                  .DataDWord = imm);
+}
 
- fail_surface_bo:
-   anv_batch_bo_destroy(surface_bbo, device);
- fail_batch_bo:
-   anv_batch_bo_destroy(batch_bo, device);
+/* Auto-Draw / Indirect Registers */
+#define GEN7_3DPRIM_END_OFFSET          0x2420
+#define GEN7_3DPRIM_START_VERTEX        0x2430
+#define GEN7_3DPRIM_VERTEX_COUNT        0x2434
+#define GEN7_3DPRIM_INSTANCE_COUNT      0x2438
+#define GEN7_3DPRIM_START_INSTANCE      0x243C
+#define GEN7_3DPRIM_BASE_VERTEX         0x2440
+
+void anv_CmdDrawIndirect(
+    VkCmdBuffer                                 cmdBuffer,
+    VkBuffer                                    _buffer,
+    VkDeviceSize                                offset,
+    uint32_t                                    count,
+    uint32_t                                    stride)
+{
+   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
+   ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
+   struct anv_bo *bo = buffer->bo;
+   uint32_t bo_offset = buffer->offset + offset;
+
+   anv_cmd_buffer_flush_state(cmd_buffer);
+
+   anv_batch_lrm(&cmd_buffer->batch, GEN7_3DPRIM_VERTEX_COUNT, bo, bo_offset);
+   anv_batch_lrm(&cmd_buffer->batch, GEN7_3DPRIM_INSTANCE_COUNT, bo, bo_offset + 4);
+   anv_batch_lrm(&cmd_buffer->batch, GEN7_3DPRIM_START_VERTEX, bo, bo_offset + 8);
+   anv_batch_lrm(&cmd_buffer->batch, GEN7_3DPRIM_START_INSTANCE, bo, bo_offset + 12);
+   anv_batch_lri(&cmd_buffer->batch, GEN7_3DPRIM_BASE_VERTEX, 0);
+
+   anv_batch_emit(&cmd_buffer->batch, GEN8_3DPRIMITIVE,
+                  .IndirectParameterEnable = true,
+                  .VertexAccessType = SEQUENTIAL);
+}
 
-   return result;
+void anv_CmdDrawIndexedIndirect(
+    VkCmdBuffer                                 cmdBuffer,
+    VkBuffer                                    _buffer,
+    VkDeviceSize                                offset,
+    uint32_t                                    count,
+    uint32_t                                    stride)
+{
+   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
+   ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
+   struct anv_bo *bo = buffer->bo;
+   uint32_t bo_offset = buffer->offset + offset;
+
+   anv_cmd_buffer_flush_state(cmd_buffer);
+
+   anv_batch_lrm(&cmd_buffer->batch, GEN7_3DPRIM_VERTEX_COUNT, bo, bo_offset);
+   anv_batch_lrm(&cmd_buffer->batch, GEN7_3DPRIM_INSTANCE_COUNT, bo, bo_offset + 4);
+   anv_batch_lrm(&cmd_buffer->batch, GEN7_3DPRIM_START_VERTEX, bo, bo_offset + 8);
+   anv_batch_lrm(&cmd_buffer->batch, GEN7_3DPRIM_BASE_VERTEX, bo, bo_offset + 12);
+   anv_batch_lrm(&cmd_buffer->batch, GEN7_3DPRIM_START_INSTANCE, bo, bo_offset + 16);
+
+   anv_batch_emit(&cmd_buffer->batch, GEN8_3DPRIMITIVE,
+                  .IndirectParameterEnable = true,
+                  .VertexAccessType = RANDOM);
 }
 
-void
-anv_cmd_buffer_fini_batch_bo_chain(struct anv_cmd_buffer *cmd_buffer)
+void anv_CmdDispatch(
+    VkCmdBuffer                                 cmdBuffer,
+    uint32_t                                    x,
+    uint32_t                                    y,
+    uint32_t                                    z)
 {
-   struct anv_device *device = cmd_buffer->device;
+   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
+   struct anv_pipeline *pipeline = cmd_buffer->state.compute_pipeline;
+   struct brw_cs_prog_data *prog_data = &pipeline->cs_prog_data;
+
+   anv_cmd_buffer_flush_compute_state(cmd_buffer);
+
+   anv_batch_emit(&cmd_buffer->batch, GEN8_GPGPU_WALKER,
+                  .SIMDSize = prog_data->simd_size / 16,
+                  .ThreadDepthCounterMaximum = 0,
+                  .ThreadHeightCounterMaximum = 0,
+                  .ThreadWidthCounterMaximum = pipeline->cs_thread_width_max,
+                  .ThreadGroupIDXDimension = x,
+                  .ThreadGroupIDYDimension = y,
+                  .ThreadGroupIDZDimension = z,
+                  .RightExecutionMask = pipeline->cs_right_mask,
+                  .BottomExecutionMask = 0xffffffff);
+
+   anv_batch_emit(&cmd_buffer->batch, GEN8_MEDIA_STATE_FLUSH);
+}
 
-   anv_vector_finish(&cmd_buffer->seen_bbos);
+#define GPGPU_DISPATCHDIMX 0x2500
+#define GPGPU_DISPATCHDIMY 0x2504
+#define GPGPU_DISPATCHDIMZ 0x2508
 
-   /* Destroy all of the batch buffers */
-   list_for_each_entry_safe(struct anv_batch_bo, bbo,
-                            &cmd_buffer->batch_bos, link) {
-      anv_batch_bo_destroy(bbo, device);
-   }
+void anv_CmdDispatchIndirect(
+    VkCmdBuffer                                 cmdBuffer,
+    VkBuffer                                    _buffer,
+    VkDeviceSize                                offset)
+{
+   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
+   ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
+   struct anv_pipeline *pipeline = cmd_buffer->state.compute_pipeline;
+   struct brw_cs_prog_data *prog_data = &pipeline->cs_prog_data;
+   struct anv_bo *bo = buffer->bo;
+   uint32_t bo_offset = buffer->offset + offset;
+
+   anv_cmd_buffer_flush_compute_state(cmd_buffer);
+
+   anv_batch_lrm(&cmd_buffer->batch, GPGPU_DISPATCHDIMX, bo, bo_offset);
+   anv_batch_lrm(&cmd_buffer->batch, GPGPU_DISPATCHDIMY, bo, bo_offset + 4);
+   anv_batch_lrm(&cmd_buffer->batch, GPGPU_DISPATCHDIMZ, bo, bo_offset + 8);
+
+   anv_batch_emit(&cmd_buffer->batch, GEN8_GPGPU_WALKER,
+                  .IndirectParameterEnable = true,
+                  .SIMDSize = prog_data->simd_size / 16,
+                  .ThreadDepthCounterMaximum = 0,
+                  .ThreadHeightCounterMaximum = 0,
+                  .ThreadWidthCounterMaximum = pipeline->cs_thread_width_max,
+                  .RightExecutionMask = pipeline->cs_right_mask,
+                  .BottomExecutionMask = 0xffffffff);
+
+   anv_batch_emit(&cmd_buffer->batch, GEN8_MEDIA_STATE_FLUSH);
+}
 
-   /* Destroy all of the surface state buffers */
-   list_for_each_entry_safe(struct anv_batch_bo, bbo,
-                            &cmd_buffer->surface_bos, link) {
-      anv_batch_bo_destroy(bbo, device);
-   }
+void anv_CmdSetEvent(
+    VkCmdBuffer                                 cmdBuffer,
+    VkEvent                                     event,
+    VkPipelineStageFlags                        stageMask)
+{
+   stub();
+}
 
-   anv_device_free(device, cmd_buffer->execbuf2.objects);
-   anv_device_free(device, cmd_buffer->execbuf2.bos);
+void anv_CmdResetEvent(
+    VkCmdBuffer                                 cmdBuffer,
+    VkEvent                                     event,
+    VkPipelineStageFlags                        stageMask)
+{
+   stub();
 }
 
-void
-anv_cmd_buffer_reset_batch_bo_chain(struct anv_cmd_buffer *cmd_buffer)
+void anv_CmdWaitEvents(
+    VkCmdBuffer                                 cmdBuffer,
+    uint32_t                                    eventCount,
+    const VkEvent*                              pEvents,
+    VkPipelineStageFlags                        srcStageMask,
+    VkPipelineStageFlags                        destStageMask,
+    uint32_t                                    memBarrierCount,
+    const void* const*                          ppMemBarriers)
 {
-   struct anv_device *device = cmd_buffer->device;
+   stub();
+}
 
-   /* Delete all but the first batch bo */
-   assert(!list_empty(&cmd_buffer->batch_bos));
-   while (cmd_buffer->batch_bos.next != cmd_buffer->batch_bos.prev) {
-      struct anv_batch_bo *bbo = anv_cmd_buffer_current_batch_bo(cmd_buffer);
-      list_del(&bbo->link);
-      anv_batch_bo_destroy(bbo, device);
-   }
-   assert(!list_empty(&cmd_buffer->batch_bos));
+void anv_CmdPipelineBarrier(
+    VkCmdBuffer                                 cmdBuffer,
+    VkPipelineStageFlags                        srcStageMask,
+    VkPipelineStageFlags                        destStageMask,
+    VkBool32                                    byRegion,
+    uint32_t                                    memBarrierCount,
+    const void* const*                          ppMemBarriers)
+{
+   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
+   uint32_t b, *dw;
 
-   anv_batch_bo_start(anv_cmd_buffer_current_batch_bo(cmd_buffer),
-                      &cmd_buffer->batch,
-                      GEN8_MI_BATCH_BUFFER_START_length * 4);
+   struct GEN8_PIPE_CONTROL cmd = {
+      GEN8_PIPE_CONTROL_header,
+      .PostSyncOperation = NoWrite,
+   };
 
-   /* Delete all but the first batch bo */
-   assert(!list_empty(&cmd_buffer->batch_bos));
-   while (cmd_buffer->surface_bos.next != cmd_buffer->surface_bos.prev) {
-      struct anv_batch_bo *bbo = anv_cmd_buffer_current_surface_bbo(cmd_buffer);
-      list_del(&bbo->link);
-      anv_batch_bo_destroy(bbo, device);
+   /* XXX: I think waitEvent is a no-op on our HW.  We should verify that. */
+
+   if (anv_clear_mask(&srcStageMask, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT)) {
+      /* This is just what PIPE_CONTROL does */
    }
-   assert(!list_empty(&cmd_buffer->batch_bos));
 
-   anv_cmd_buffer_current_surface_bbo(cmd_buffer)->relocs.num_relocs = 0;
+   if (anv_clear_mask(&srcStageMask,
+                      VK_PIPELINE_STAGE_DRAW_INDIRECT_BIT |
+                      VK_PIPELINE_STAGE_VERTEX_INPUT_BIT |
+                      VK_PIPELINE_STAGE_VERTEX_SHADER_BIT |
+                      VK_PIPELINE_STAGE_TESS_CONTROL_SHADER_BIT |
+                      VK_PIPELINE_STAGE_TESS_EVALUATION_SHADER_BIT |
+                      VK_PIPELINE_STAGE_GEOMETRY_SHADER_BIT |
+                      VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT |
+                      VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT |
+                      VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT |
+                      VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT)) {
+      cmd.StallAtPixelScoreboard = true;
+   }
 
-   cmd_buffer->surface_next = 1;
 
-   /* Reset the list of seen buffers */
-   cmd_buffer->seen_bbos.head = 0;
-   cmd_buffer->seen_bbos.tail = 0;
+   if (anv_clear_mask(&srcStageMask,
+                      VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT |
+                      VK_PIPELINE_STAGE_TRANSFER_BIT |
+                      VK_PIPELINE_STAGE_TRANSITION_BIT)) {
+      cmd.CommandStreamerStallEnable = true;
+   }
 
-   *(struct anv_batch_bo **)anv_vector_add(&cmd_buffer->seen_bbos) =
-      anv_cmd_buffer_current_batch_bo(cmd_buffer);
-   *(struct anv_batch_bo **)anv_vector_add(&cmd_buffer->seen_bbos) =
-      anv_cmd_buffer_current_surface_bbo(cmd_buffer);
-}
+   if (anv_clear_mask(&srcStageMask, VK_PIPELINE_STAGE_HOST_BIT)) {
+      anv_finishme("VK_PIPE_EVENT_CPU_SIGNAL_BIT");
+   }
 
-void
-anv_cmd_buffer_end_batch_buffer(struct anv_cmd_buffer *cmd_buffer)
-{
-   struct anv_batch_bo *batch_bo = anv_cmd_buffer_current_batch_bo(cmd_buffer);
-   struct anv_batch_bo *surface_bbo =
-      anv_cmd_buffer_current_surface_bbo(cmd_buffer);
+   /* On our hardware, all stages will wait for execution as needed. */
+   (void)destStageMask;
 
-   if (cmd_buffer->level == VK_CMD_BUFFER_LEVEL_PRIMARY) {
-      anv_batch_emit(&cmd_buffer->batch, GEN8_MI_BATCH_BUFFER_END);
+   /* We checked all known VkPipeEventFlags. */
+   anv_assert(srcStageMask == 0);
 
-      /* Round batch up to an even number of dwords. */
-      if ((cmd_buffer->batch.next - cmd_buffer->batch.start) & 4)
-         anv_batch_emit(&cmd_buffer->batch, GEN8_MI_NOOP);
+   /* XXX: Right now, we're really dumb and just flush whatever categories
+    * the app asks for.  One of these days we may make this a bit better
+    * but right now that's all the hardware allows for in most areas.
+    */
+   VkMemoryOutputFlags out_flags = 0;
+   VkMemoryInputFlags in_flags = 0;
+
+   for (uint32_t i = 0; i < memBarrierCount; i++) {
+      const struct anv_common *common = ppMemBarriers[i];
+      switch (common->sType) {
+      case VK_STRUCTURE_TYPE_MEMORY_BARRIER: {
+         ANV_COMMON_TO_STRUCT(VkMemoryBarrier, barrier, common);
+         out_flags |= barrier->outputMask;
+         in_flags |= barrier->inputMask;
+         break;
+      }
+      case VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER: {
+         ANV_COMMON_TO_STRUCT(VkBufferMemoryBarrier, barrier, common);
+         out_flags |= barrier->outputMask;
+         in_flags |= barrier->inputMask;
+         break;
+      }
+      case VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER: {
+         ANV_COMMON_TO_STRUCT(VkImageMemoryBarrier, barrier, common);
+         out_flags |= barrier->outputMask;
+         in_flags |= barrier->inputMask;
+         break;
+      }
+      default:
+         unreachable("Invalid memory barrier type");
+      }
+   }
 
-      cmd_buffer->exec_mode = ANV_CMD_BUFFER_EXEC_MODE_PRIMARY;
-   } else {
-      /* If this is a secondary command buffer, we need to determine the
-       * mode in which it will be executed with vkExecuteCommands.  We
-       * determine this statically here so that this stays in sync with the
-       * actual ExecuteCommands implementation.
-       */
-      if ((cmd_buffer->batch_bos.next == cmd_buffer->batch_bos.prev) &&
-          (anv_cmd_buffer_current_batch_bo(cmd_buffer)->length <
-           ANV_CMD_BUFFER_BATCH_SIZE / 2)) {
-         /* If the secondary has exactly one batch buffer in its list *and*
-          * that batch buffer is less than half of the maximum size, we're
-          * probably better of simply copying it into our batch.
-          */
-         cmd_buffer->exec_mode = ANV_CMD_BUFFER_EXEC_MODE_EMIT;
-      } else if (cmd_buffer->opt_flags &
-                 VK_CMD_BUFFER_OPTIMIZE_NO_SIMULTANEOUS_USE_BIT) {
-         cmd_buffer->exec_mode = ANV_CMD_BUFFER_EXEC_MODE_CHAIN;
-
-         /* For chaining mode, we need to increment the number of
-          * relocations.  This is because, when we chain, we need to add
-          * an MI_BATCH_BUFFER_START command.  Adding this command will
-          * also add a relocation.  In order to handle theis we'll
-          * increment it here and decrement it right before adding the
-          * MI_BATCH_BUFFER_START command.
-          */
-         anv_cmd_buffer_current_batch_bo(cmd_buffer)->relocs.num_relocs++;
-      } else {
-         cmd_buffer->exec_mode = ANV_CMD_BUFFER_EXEC_MODE_COPY_AND_CHAIN;
+   for_each_bit(b, out_flags) {
+      switch ((VkMemoryOutputFlags)(1 << b)) {
+      case VK_MEMORY_OUTPUT_HOST_WRITE_BIT:
+         break; /* FIXME: Little-core systems */
+      case VK_MEMORY_OUTPUT_SHADER_WRITE_BIT:
+         cmd.DCFlushEnable = true;
+         break;
+      case VK_MEMORY_OUTPUT_COLOR_ATTACHMENT_BIT:
+         cmd.RenderTargetCacheFlushEnable = true;
+         break;
+      case VK_MEMORY_OUTPUT_DEPTH_STENCIL_ATTACHMENT_BIT:
+         cmd.DepthCacheFlushEnable = true;
+         break;
+      case VK_MEMORY_OUTPUT_TRANSFER_BIT:
+         cmd.RenderTargetCacheFlushEnable = true;
+         cmd.DepthCacheFlushEnable = true;
+         break;
+      default:
+         unreachable("Invalid memory output flag");
       }
    }
 
-   anv_batch_bo_finish(batch_bo, &cmd_buffer->batch);
+   for_each_bit(b, out_flags) {
+      switch ((VkMemoryInputFlags)(1 << b)) {
+      case VK_MEMORY_INPUT_HOST_READ_BIT:
+         break; /* FIXME: Little-core systems */
+      case VK_MEMORY_INPUT_INDIRECT_COMMAND_BIT:
+      case VK_MEMORY_INPUT_INDEX_FETCH_BIT:
+      case VK_MEMORY_INPUT_VERTEX_ATTRIBUTE_FETCH_BIT:
+         cmd.VFCacheInvalidationEnable = true;
+         break;
+      case VK_MEMORY_INPUT_UNIFORM_READ_BIT:
+         cmd.ConstantCacheInvalidationEnable = true;
+         /* fallthrough */
+      case VK_MEMORY_INPUT_SHADER_READ_BIT:
+         cmd.DCFlushEnable = true;
+         cmd.TextureCacheInvalidationEnable = true;
+         break;
+      case VK_MEMORY_INPUT_COLOR_ATTACHMENT_BIT:
+      case VK_MEMORY_INPUT_DEPTH_STENCIL_ATTACHMENT_BIT:
+         break; /* XXX: Hunh? */
+      case VK_MEMORY_INPUT_TRANSFER_BIT:
+         cmd.TextureCacheInvalidationEnable = true;
+         break;
+      }
+   }
 
-   surface_bbo->length = cmd_buffer->surface_next;
+   dw = anv_batch_emit_dwords(&cmd_buffer->batch, GEN8_PIPE_CONTROL_length);
+   GEN8_PIPE_CONTROL_pack(&cmd_buffer->batch, dw, &cmd);
 }
 
-static inline VkResult
-anv_cmd_buffer_add_seen_bbos(struct anv_cmd_buffer *cmd_buffer,
-                             struct list_head *list)
+void anv_CmdPushConstants(
+    VkCmdBuffer                                 cmdBuffer,
+    VkPipelineLayout                            layout,
+    VkShaderStageFlags                          stageFlags,
+    uint32_t                                    start,
+    uint32_t                                    length,
+    const void*                                 values)
 {
-   list_for_each_entry(struct anv_batch_bo, bbo, list, link) {
-      struct anv_batch_bo **bbo_ptr = anv_vector_add(&cmd_buffer->seen_bbos);
-      if (bbo_ptr == NULL)
-         return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
+   stub();
+}
 
-      *bbo_ptr = bbo;
+static void
+anv_cmd_buffer_emit_depth_stencil(struct anv_cmd_buffer *cmd_buffer)
+{
+   struct anv_subpass *subpass = cmd_buffer->state.subpass;
+   struct anv_framebuffer *fb = cmd_buffer->state.framebuffer;
+   const struct anv_depth_stencil_view *view;
+
+   static const struct anv_depth_stencil_view null_view =
+      { .depth_format = D16_UNORM, .depth_stride = 0, .stencil_stride = 0 };
+
+   if (subpass->depth_stencil_attachment != VK_ATTACHMENT_UNUSED) {
+      const struct anv_attachment_view *aview =
+         fb->attachments[subpass->depth_stencil_attachment];
+      assert(aview->attachment_type == ANV_ATTACHMENT_VIEW_TYPE_DEPTH_STENCIL);
+      view = (const struct anv_depth_stencil_view *)aview;
+   } else {
+      view = &null_view;
    }
 
-   return VK_SUCCESS;
+   /* FIXME: Implement the PMA stall W/A */
+   /* FIXME: Width and Height are wrong */
+
+   anv_batch_emit(&cmd_buffer->batch, GEN8_3DSTATE_DEPTH_BUFFER,
+                  .SurfaceType = SURFTYPE_2D,
+                  .DepthWriteEnable = view->depth_stride > 0,
+                  .StencilWriteEnable = view->stencil_stride > 0,
+                  .HierarchicalDepthBufferEnable = false,
+                  .SurfaceFormat = view->depth_format,
+                  .SurfacePitch = view->depth_stride > 0 ? view->depth_stride - 1 : 0,
+                  .SurfaceBaseAddress = { view->bo,  view->depth_offset },
+                  .Height = cmd_buffer->state.framebuffer->height - 1,
+                  .Width = cmd_buffer->state.framebuffer->width - 1,
+                  .LOD = 0,
+                  .Depth = 1 - 1,
+                  .MinimumArrayElement = 0,
+                  .DepthBufferObjectControlState = GEN8_MOCS,
+                  .RenderTargetViewExtent = 1 - 1,
+                  .SurfaceQPitch = view->depth_qpitch >> 2);
+
+   /* Disable hierarchial depth buffers. */
+   anv_batch_emit(&cmd_buffer->batch, GEN8_3DSTATE_HIER_DEPTH_BUFFER);
+
+   anv_batch_emit(&cmd_buffer->batch, GEN8_3DSTATE_STENCIL_BUFFER,
+                  .StencilBufferEnable = view->stencil_stride > 0,
+                  .StencilBufferObjectControlState = GEN8_MOCS,
+                  .SurfacePitch = view->stencil_stride > 0 ? view->stencil_stride - 1 : 0,
+                  .SurfaceBaseAddress = { view->bo, view->stencil_offset },
+                  .SurfaceQPitch = view->stencil_qpitch >> 2);
+
+   /* Clear the clear params. */
+   anv_batch_emit(&cmd_buffer->batch, GEN8_3DSTATE_CLEAR_PARAMS);
 }
 
 void
-anv_cmd_buffer_add_secondary(struct anv_cmd_buffer *primary,
-                             struct anv_cmd_buffer *secondary)
+anv_cmd_buffer_begin_subpass(struct anv_cmd_buffer *cmd_buffer,
+                             struct anv_subpass *subpass)
 {
-   switch (secondary->exec_mode) {
-   case ANV_CMD_BUFFER_EXEC_MODE_EMIT:
-      anv_batch_emit_batch(&primary->batch, &secondary->batch);
-      break;
-   case ANV_CMD_BUFFER_EXEC_MODE_CHAIN: {
-      struct anv_batch_bo *first_bbo =
-         list_first_entry(&secondary->batch_bos, struct anv_batch_bo, link);
-      struct anv_batch_bo *last_bbo =
-         list_last_entry(&secondary->batch_bos, struct anv_batch_bo, link);
-
-      anv_batch_emit(&primary->batch, GEN8_MI_BATCH_BUFFER_START,
-         GEN8_MI_BATCH_BUFFER_START_header,
-         ._2ndLevelBatchBuffer = _1stlevelbatch,
-         .AddressSpaceIndicator = ASI_PPGTT,
-         .BatchBufferStartAddress = { &first_bbo->bo, 0 },
-      );
-
-      struct anv_batch_bo *this_bbo = anv_cmd_buffer_current_batch_bo(primary);
-      assert(primary->batch.start == this_bbo->bo.map);
-      uint32_t offset = primary->batch.next - primary->batch.start;
-
-      struct GEN8_MI_BATCH_BUFFER_START ret = {
-         GEN8_MI_BATCH_BUFFER_START_header,
-         ._2ndLevelBatchBuffer = _1stlevelbatch,
-         .AddressSpaceIndicator = ASI_PPGTT,
-         .BatchBufferStartAddress = { &this_bbo->bo, offset },
-      };
-      last_bbo->relocs.num_relocs++;
-      GEN8_MI_BATCH_BUFFER_START_pack(&secondary->batch,
-                                      last_bbo->bo.map + last_bbo->length,
-                                      &ret);
-
-      anv_cmd_buffer_add_seen_bbos(primary, &secondary->batch_bos);
-      break;
-   }
-   case ANV_CMD_BUFFER_EXEC_MODE_COPY_AND_CHAIN: {
-      struct list_head copy_list;
-      VkResult result = anv_batch_bo_list_clone(&secondary->batch_bos,
-                                                secondary->device,
-                                                &copy_list);
-      if (result != VK_SUCCESS)
-         return; /* FIXME */
+   cmd_buffer->state.subpass = subpass;
 
-      anv_cmd_buffer_add_seen_bbos(primary, &copy_list);
+   cmd_buffer->state.descriptors_dirty |= VK_SHADER_STAGE_FRAGMENT_BIT;
 
-      struct anv_batch_bo *first_bbo =
-         list_first_entry(&copy_list, struct anv_batch_bo, link);
-      struct anv_batch_bo *last_bbo =
-         list_last_entry(&copy_list, struct anv_batch_bo, link);
+   anv_cmd_buffer_emit_depth_stencil(cmd_buffer);
+}
 
-      cmd_buffer_chain_to_batch_bo(primary, first_bbo);
+void anv_CmdBeginRenderPass(
+    VkCmdBuffer                                 cmdBuffer,
+    const VkRenderPassBeginInfo*                pRenderPassBegin,
+    VkRenderPassContents                        contents)
+{
+   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
+   ANV_FROM_HANDLE(anv_render_pass, pass, pRenderPassBegin->renderPass);
+   ANV_FROM_HANDLE(anv_framebuffer, framebuffer, pRenderPassBegin->framebuffer);
 
-      list_splicetail(&copy_list, &primary->batch_bos);
+   cmd_buffer->state.framebuffer = framebuffer;
+   cmd_buffer->state.pass = pass;
 
-      anv_batch_bo_continue(last_bbo, &primary->batch,
-                            GEN8_MI_BATCH_BUFFER_START_length * 4);
+   const VkRect2D *render_area = &pRenderPassBegin->renderArea;
 
-      anv_cmd_buffer_emit_state_base_address(primary);
-      break;
-   }
-   default:
-      assert(!"Invalid execution mode");
-   }
+   anv_batch_emit(&cmd_buffer->batch, GEN8_3DSTATE_DRAWING_RECTANGLE,
+                  .ClippedDrawingRectangleYMin = render_area->offset.y,
+                  .ClippedDrawingRectangleXMin = render_area->offset.x,
+                  .ClippedDrawingRectangleYMax =
+                     render_area->offset.y + render_area->extent.height - 1,
+                  .ClippedDrawingRectangleXMax =
+                     render_area->offset.x + render_area->extent.width - 1,
+                  .DrawingRectangleOriginY = 0,
+                  .DrawingRectangleOriginX = 0);
+
+   anv_cmd_buffer_clear_attachments(cmd_buffer, pass,
+                                    pRenderPassBegin->pAttachmentClearValues);
 
-   /* Mark the surface buffer from the secondary as seen */
-   anv_cmd_buffer_add_seen_bbos(primary, &secondary->surface_bos);
+   anv_cmd_buffer_begin_subpass(cmd_buffer, pass->subpasses);
 }
 
-static VkResult
-anv_cmd_buffer_add_bo(struct anv_cmd_buffer *cmd_buffer,
-                      struct anv_bo *bo,
-                      struct anv_reloc_list *relocs)
+void anv_CmdNextSubpass(
+    VkCmdBuffer                                 cmdBuffer,
+    VkRenderPassContents                        contents)
 {
-   struct drm_i915_gem_exec_object2 *obj = NULL;
+   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
 
-   if (bo->index < cmd_buffer->execbuf2.bo_count &&
-       cmd_buffer->execbuf2.bos[bo->index] == bo)
-      obj = &cmd_buffer->execbuf2.objects[bo->index];
+   assert(cmd_buffer->level == VK_CMD_BUFFER_LEVEL_PRIMARY);
 
-   if (obj == NULL) {
-      /* We've never seen this one before.  Add it to the list and assign
-       * an id that we can use later.
-       */
-      if (cmd_buffer->execbuf2.bo_count >= cmd_buffer->execbuf2.array_length) {
-         uint32_t new_len = cmd_buffer->execbuf2.objects ?
-                            cmd_buffer->execbuf2.array_length * 2 : 64;
-
-         struct drm_i915_gem_exec_object2 *new_objects =
-            anv_device_alloc(cmd_buffer->device, new_len * sizeof(*new_objects),
-                             8, VK_SYSTEM_ALLOC_TYPE_INTERNAL);
-         if (new_objects == NULL)
-            return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
-
-         struct anv_bo **new_bos =
-            anv_device_alloc(cmd_buffer->device, new_len * sizeof(*new_bos),
-                             8, VK_SYSTEM_ALLOC_TYPE_INTERNAL);
-         if (new_objects == NULL) {
-            anv_device_free(cmd_buffer->device, new_objects);
-            return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
-         }
+   anv_cmd_buffer_begin_subpass(cmd_buffer, cmd_buffer->state.subpass + 1);
+}
 
-         if (cmd_buffer->execbuf2.objects) {
-            memcpy(new_objects, cmd_buffer->execbuf2.objects,
-                   cmd_buffer->execbuf2.bo_count * sizeof(*new_objects));
-            memcpy(new_bos, cmd_buffer->execbuf2.bos,
-                   cmd_buffer->execbuf2.bo_count * sizeof(*new_bos));
-         }
+void anv_CmdEndRenderPass(
+    VkCmdBuffer                                 cmdBuffer)
+{
+   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
 
-         cmd_buffer->execbuf2.objects = new_objects;
-         cmd_buffer->execbuf2.bos = new_bos;
-         cmd_buffer->execbuf2.array_length = new_len;
-      }
+   /* Emit a flushing pipe control at the end of a pass.  This is kind of a
+    * hack but it ensures that render targets always actually get written.
+    * Eventually, we should do flushing based on image format transitions
+    * or something of that nature.
+    */
+   anv_batch_emit(&cmd_buffer->batch, GEN8_PIPE_CONTROL,
+                  .PostSyncOperation = NoWrite,
+                  .RenderTargetCacheFlushEnable = true,
+                  .InstructionCacheInvalidateEnable = true,
+                  .DepthCacheFlushEnable = true,
+                  .VFCacheInvalidationEnable = true,
+                  .TextureCacheInvalidationEnable = true,
+                  .CommandStreamerStallEnable = true);
+}
 
-      assert(cmd_buffer->execbuf2.bo_count < cmd_buffer->execbuf2.array_length);
+void anv_CmdExecuteCommands(
+    VkCmdBuffer                                 cmdBuffer,
+    uint32_t                                    cmdBuffersCount,
+    const VkCmdBuffer*                          pCmdBuffers)
+{
+   ANV_FROM_HANDLE(anv_cmd_buffer, primary, cmdBuffer);
 
-      bo->index = cmd_buffer->execbuf2.bo_count++;
-      obj = &cmd_buffer->execbuf2.objects[bo->index];
-      cmd_buffer->execbuf2.bos[bo->index] = bo;
+   assert(primary->level == VK_CMD_BUFFER_LEVEL_PRIMARY);
 
-      obj->handle = bo->gem_handle;
-      obj->relocation_count = 0;
-      obj->relocs_ptr = 0;
-      obj->alignment = 0;
-      obj->offset = bo->offset;
-      obj->flags = 0;
-      obj->rsvd1 = 0;
-      obj->rsvd2 = 0;
-   }
+   anv_assert(primary->state.subpass == &primary->state.pass->subpasses[0]);
 
-   if (relocs != NULL && obj->relocation_count == 0) {
-      /* This is the first time we've ever seen a list of relocations for
-       * this BO.  Go ahead and set the relocations and then walk the list
-       * of relocations and add them all.
-       */
-      obj->relocation_count = relocs->num_relocs;
-      obj->relocs_ptr = (uintptr_t) relocs->relocs;
+   for (uint32_t i = 0; i < cmdBuffersCount; i++) {
+      ANV_FROM_HANDLE(anv_cmd_buffer, secondary, pCmdBuffers[i]);
 
-      for (size_t i = 0; i < relocs->num_relocs; i++)
-         anv_cmd_buffer_add_bo(cmd_buffer, relocs->reloc_bos[i], NULL);
-   }
+      assert(secondary->level == VK_CMD_BUFFER_LEVEL_SECONDARY);
 
-   return VK_SUCCESS;
+      anv_cmd_buffer_add_secondary(primary, secondary);
+   }
 }
 
-static void
-anv_cmd_buffer_process_relocs(struct anv_cmd_buffer *cmd_buffer,
-                              struct anv_reloc_list *list)
+VkResult anv_CreateCommandPool(
+    VkDevice                                    _device,
+    const VkCmdPoolCreateInfo*                  pCreateInfo,
+    VkCmdPool*                                  pCmdPool)
 {
-   struct anv_bo *bo;
+   ANV_FROM_HANDLE(anv_device, device, _device);
+   struct anv_cmd_pool *pool;
 
-   /* If the kernel supports I915_EXEC_NO_RELOC, it will compare offset in
-    * struct drm_i915_gem_exec_object2 against the bos current offset and if
-    * all bos haven't moved it will skip relocation processing alltogether.
-    * If I915_EXEC_NO_RELOC is not supported, the kernel ignores the incoming
-    * value of offset so we can set it either way.  For that to work we need
-    * to make sure all relocs use the same presumed offset.
-    */
+   pool = anv_device_alloc(device, sizeof(*pool), 8,
+                           VK_SYSTEM_ALLOC_TYPE_API_OBJECT);
+   if (pool == NULL)
+      return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
 
-   for (size_t i = 0; i < list->num_relocs; i++) {
-      bo = list->reloc_bos[i];
-      if (bo->offset != list->relocs[i].presumed_offset)
-         cmd_buffer->execbuf2.need_reloc = true;
+   list_inithead(&pool->cmd_buffers);
 
-      list->relocs[i].target_handle = bo->index;
-   }
+   *pCmdPool = anv_cmd_pool_to_handle(pool);
+
+   return VK_SUCCESS;
 }
 
-void
-anv_cmd_buffer_prepare_execbuf(struct anv_cmd_buffer *cmd_buffer)
+VkResult anv_DestroyCommandPool(
+    VkDevice                                    _device,
+    VkCmdPool                                   cmdPool)
 {
-   struct anv_batch *batch = &cmd_buffer->batch;
+   ANV_FROM_HANDLE(anv_device, device, _device);
+   ANV_FROM_HANDLE(anv_cmd_pool, pool, cmdPool);
 
-   cmd_buffer->execbuf2.bo_count = 0;
-   cmd_buffer->execbuf2.need_reloc = false;
-
-   /* First, we walk over all of the bos we've seen and add them and their
-    * relocations to the validate list.
-    */
-   struct anv_batch_bo **bbo;
-   anv_vector_foreach(bbo, &cmd_buffer->seen_bbos)
-      anv_cmd_buffer_add_bo(cmd_buffer, &(*bbo)->bo, &(*bbo)->relocs);
+   anv_ResetCommandPool(_device, cmdPool, 0);
 
-   struct anv_batch_bo *first_batch_bo =
-      list_first_entry(&cmd_buffer->batch_bos, struct anv_batch_bo, link);
+   anv_device_free(device, pool);
 
-   /* The kernel requires that the last entry in the validation list be the
-    * batch buffer to execute.  We can simply swap the element
-    * corresponding to the first batch_bo in the chain with the last
-    * element in the list.
-    */
-   if (first_batch_bo->bo.index != cmd_buffer->execbuf2.bo_count - 1) {
-      uint32_t idx = first_batch_bo->bo.index;
-
-      struct drm_i915_gem_exec_object2 tmp_obj =
-         cmd_buffer->execbuf2.objects[idx];
-      assert(cmd_buffer->execbuf2.bos[idx] == &first_batch_bo->bo);
+   return VK_SUCCESS;
+}
 
-      cmd_buffer->execbuf2.objects[idx] =
-         cmd_buffer->execbuf2.objects[cmd_buffer->execbuf2.bo_count - 1];
-      cmd_buffer->execbuf2.bos[idx] =
-         cmd_buffer->execbuf2.bos[cmd_buffer->execbuf2.bo_count - 1];
-      cmd_buffer->execbuf2.bos[idx]->index = idx;
+VkResult anv_ResetCommandPool(
+    VkDevice                                    device,
+    VkCmdPool                                   cmdPool,
+    VkCmdPoolResetFlags                         flags)
+{
+   ANV_FROM_HANDLE(anv_cmd_pool, pool, cmdPool);
 
-      cmd_buffer->execbuf2.objects[cmd_buffer->execbuf2.bo_count - 1] = tmp_obj;
-      cmd_buffer->execbuf2.bos[cmd_buffer->execbuf2.bo_count - 1] =
-         &first_batch_bo->bo;
-      first_batch_bo->bo.index = cmd_buffer->execbuf2.bo_count - 1;
+   list_for_each_entry_safe(struct anv_cmd_buffer, cmd_buffer,
+                            &pool->cmd_buffers, pool_link) {
+      anv_DestroyCommandBuffer(device, anv_cmd_buffer_to_handle(cmd_buffer));
    }
 
-   /* Now we go through and fixup all of the relocation lists to point to
-    * the correct indices in the object array.  We have to do this after we
-    * reorder the list above as some of the indices may have changed.
-    */
-   anv_vector_foreach(bbo, &cmd_buffer->seen_bbos)
-      anv_cmd_buffer_process_relocs(cmd_buffer, &(*bbo)->relocs);
-
-   cmd_buffer->execbuf2.execbuf = (struct drm_i915_gem_execbuffer2) {
-      .buffers_ptr = (uintptr_t) cmd_buffer->execbuf2.objects,
-      .buffer_count = cmd_buffer->execbuf2.bo_count,
-      .batch_start_offset = 0,
-      .batch_len = batch->next - batch->start,
-      .cliprects_ptr = 0,
-      .num_cliprects = 0,
-      .DR1 = 0,
-      .DR4 = 0,
-      .flags = I915_EXEC_HANDLE_LUT | I915_EXEC_RENDER,
-      .rsvd1 = cmd_buffer->device->context_id,
-      .rsvd2 = 0,
-   };
-
-   if (!cmd_buffer->execbuf2.need_reloc)
-      cmd_buffer->execbuf2.execbuf.flags |= I915_EXEC_NO_RELOC;
+   return VK_SUCCESS;
 }
diff --git a/src/vulkan/anv_cmd_emit.c b/src/vulkan/anv_cmd_emit.c
deleted file mode 100644
index 3b9e67fdd0f..00000000000
--- a/src/vulkan/anv_cmd_emit.c
+++ /dev/null
@@ -1,1425 +0,0 @@
-/*
- * Copyright © 2015 Intel Corporation
- *
- * Permission is hereby granted, free of charge, to any person obtaining a
- * copy of this software and associated documentation files (the "Software"),
- * to deal in the Software without restriction, including without limitation
- * the rights to use, copy, modify, merge, publish, distribute, sublicense,
- * and/or sell copies of the Software, and to permit persons to whom the
- * Software is furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice (including the next
- * paragraph) shall be included in all copies or substantial portions of the
- * Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
- * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
- * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
- * IN THE SOFTWARE.
- */
-
-#include <assert.h>
-#include <stdbool.h>
-#include <string.h>
-#include <unistd.h>
-#include <fcntl.h>
-
-#include "anv_private.h"
-
-/** \file anv_cmd_buffer.c
- *
- * This file contains all of the stuff for emitting commands into a command
- * buffer.  This includes implementations of most of the vkCmd*
- * entrypoints.  This file is concerned entirely with state emission and
- * not with the command buffer data structure itself.  As far as this file
- * is concerned, most of anv_cmd_buffer is magic.
- */
-
-static void
-anv_cmd_state_init(struct anv_cmd_state *state)
-{
-   state->rs_state = NULL;
-   state->vp_state = NULL;
-   state->cb_state = NULL;
-   state->ds_state = NULL;
-   memset(&state->state_vf, 0, sizeof(state->state_vf));
-   memset(&state->descriptors, 0, sizeof(state->descriptors));
-
-   state->dirty = 0;
-   state->vb_dirty = 0;
-   state->descriptors_dirty = 0;
-   state->pipeline = NULL;
-   state->vp_state = NULL;
-   state->rs_state = NULL;
-   state->ds_state = NULL;
-}
-
-VkResult anv_CreateCommandBuffer(
-    VkDevice                                    _device,
-    const VkCmdBufferCreateInfo*                pCreateInfo,
-    VkCmdBuffer*                                pCmdBuffer)
-{
-   ANV_FROM_HANDLE(anv_device, device, _device);
-   ANV_FROM_HANDLE(anv_cmd_pool, pool, pCreateInfo->cmdPool);
-   struct anv_cmd_buffer *cmd_buffer;
-   VkResult result;
-
-   cmd_buffer = anv_device_alloc(device, sizeof(*cmd_buffer), 8,
-                                 VK_SYSTEM_ALLOC_TYPE_API_OBJECT);
-   if (cmd_buffer == NULL)
-      return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
-
-   cmd_buffer->device = device;
-
-   result = anv_cmd_buffer_init_batch_bo_chain(cmd_buffer);
-   if (result != VK_SUCCESS)
-      goto fail;
-
-   anv_state_stream_init(&cmd_buffer->surface_state_stream,
-                         &device->surface_state_block_pool);
-   anv_state_stream_init(&cmd_buffer->dynamic_state_stream,
-                         &device->dynamic_state_block_pool);
-
-   cmd_buffer->level = pCreateInfo->level;
-   cmd_buffer->opt_flags = 0;
-
-   anv_cmd_state_init(&cmd_buffer->state);
-
-   list_addtail(&cmd_buffer->pool_link, &pool->cmd_buffers);
-
-   *pCmdBuffer = anv_cmd_buffer_to_handle(cmd_buffer);
-
-   return VK_SUCCESS;
-
- fail: anv_device_free(device, cmd_buffer);
-
-   return result;
-}
-
-VkResult anv_DestroyCommandBuffer(
-    VkDevice                                    _device,
-    VkCmdBuffer                                 _cmd_buffer)
-{
-   ANV_FROM_HANDLE(anv_device, device, _device);
-   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, _cmd_buffer);
-
-   list_del(&cmd_buffer->pool_link);
-
-   anv_cmd_buffer_fini_batch_bo_chain(cmd_buffer);
-
-   anv_state_stream_finish(&cmd_buffer->surface_state_stream);
-   anv_state_stream_finish(&cmd_buffer->dynamic_state_stream);
-   anv_device_free(device, cmd_buffer);
-
-   return VK_SUCCESS;
-}
-
-VkResult anv_ResetCommandBuffer(
-    VkCmdBuffer                                 cmdBuffer,
-    VkCmdBufferResetFlags                       flags)
-{
-   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
-
-   anv_cmd_buffer_reset_batch_bo_chain(cmd_buffer);
-
-   anv_cmd_state_init(&cmd_buffer->state);
-
-   return VK_SUCCESS;
-}
-
-void
-anv_cmd_buffer_emit_state_base_address(struct anv_cmd_buffer *cmd_buffer)
-{
-   struct anv_device *device = cmd_buffer->device;
-   struct anv_bo *scratch_bo = NULL;
-
-   cmd_buffer->state.scratch_size = device->scratch_block_pool.size;
-   if (cmd_buffer->state.scratch_size > 0)
-      scratch_bo = &device->scratch_block_pool.bo;
-
-   anv_batch_emit(&cmd_buffer->batch, GEN8_STATE_BASE_ADDRESS,
-                  .GeneralStateBaseAddress = { scratch_bo, 0 },
-                  .GeneralStateMemoryObjectControlState = GEN8_MOCS,
-                  .GeneralStateBaseAddressModifyEnable = true,
-                  .GeneralStateBufferSize = 0xfffff,
-                  .GeneralStateBufferSizeModifyEnable = true,
-
-                  .SurfaceStateBaseAddress = { anv_cmd_buffer_current_surface_bo(cmd_buffer), 0 },
-                  .SurfaceStateMemoryObjectControlState = GEN8_MOCS,
-                  .SurfaceStateBaseAddressModifyEnable = true,
-
-                  .DynamicStateBaseAddress = { &device->dynamic_state_block_pool.bo, 0 },
-                  .DynamicStateMemoryObjectControlState = GEN8_MOCS,
-                  .DynamicStateBaseAddressModifyEnable = true,
-                  .DynamicStateBufferSize = 0xfffff,
-                  .DynamicStateBufferSizeModifyEnable = true,
-
-                  .IndirectObjectBaseAddress = { NULL, 0 },
-                  .IndirectObjectMemoryObjectControlState = GEN8_MOCS,
-                  .IndirectObjectBaseAddressModifyEnable = true,
-                  .IndirectObjectBufferSize = 0xfffff,
-                  .IndirectObjectBufferSizeModifyEnable = true,
-
-                  .InstructionBaseAddress = { &device->instruction_block_pool.bo, 0 },
-                  .InstructionMemoryObjectControlState = GEN8_MOCS,
-                  .InstructionBaseAddressModifyEnable = true,
-                  .InstructionBufferSize = 0xfffff,
-                  .InstructionBuffersizeModifyEnable = true);
-
-   /* After re-setting the surface state base address, we have to do some
-    * cache flusing so that the sampler engine will pick up the new
-    * SURFACE_STATE objects and binding tables. From the Broadwell PRM,
-    * Shared Function > 3D Sampler > State > State Caching (page 96):
-    *
-    *    Coherency with system memory in the state cache, like the texture
-    *    cache is handled partially by software. It is expected that the
-    *    command stream or shader will issue Cache Flush operation or
-    *    Cache_Flush sampler message to ensure that the L1 cache remains
-    *    coherent with system memory.
-    *
-    *    [...]
-    *
-    *    Whenever the value of the Dynamic_State_Base_Addr,
-    *    Surface_State_Base_Addr are altered, the L1 state cache must be
-    *    invalidated to ensure the new surface or sampler state is fetched
-    *    from system memory.
-    *
-    * The PIPE_CONTROL command has a "State Cache Invalidation Enable" bit
-    * which, according the PIPE_CONTROL instruction documentation in the
-    * Broadwell PRM:
-    *
-    *    Setting this bit is independent of any other bit in this packet.
-    *    This bit controls the invalidation of the L1 and L2 state caches
-    *    at the top of the pipe i.e. at the parsing time.
-    *
-    * Unfortunately, experimentation seems to indicate that state cache
-    * invalidation through a PIPE_CONTROL does nothing whatsoever in
-    * regards to surface state and binding tables.  In stead, it seems that
-    * invalidating the texture cache is what is actually needed.
-    *
-    * XXX:  As far as we have been able to determine through
-    * experimentation, shows that flush the texture cache appears to be
-    * sufficient.  The theory here is that all of the sampling/rendering
-    * units cache the binding table in the texture cache.  However, we have
-    * yet to be able to actually confirm this.
-    */
-   anv_batch_emit(&cmd_buffer->batch, GEN8_PIPE_CONTROL,
-                  .TextureCacheInvalidationEnable = true);
-}
-
-VkResult anv_BeginCommandBuffer(
-    VkCmdBuffer                                 cmdBuffer,
-    const VkCmdBufferBeginInfo*                 pBeginInfo)
-{
-   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
-
-   cmd_buffer->opt_flags = pBeginInfo->flags;
-
-   if (cmd_buffer->level == VK_CMD_BUFFER_LEVEL_SECONDARY) {
-      cmd_buffer->state.framebuffer =
-         anv_framebuffer_from_handle(pBeginInfo->framebuffer);
-      cmd_buffer->state.pass =
-         anv_render_pass_from_handle(pBeginInfo->renderPass);
-
-      /* FIXME: We shouldn't be starting on the first subpass */
-      anv_cmd_buffer_begin_subpass(cmd_buffer,
-                                   &cmd_buffer->state.pass->subpasses[0]);
-   }
-
-   anv_cmd_buffer_emit_state_base_address(cmd_buffer);
-   cmd_buffer->state.current_pipeline = UINT32_MAX;
-
-   return VK_SUCCESS;
-}
-
-VkResult anv_EndCommandBuffer(
-    VkCmdBuffer                                 cmdBuffer)
-{
-   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
-   struct anv_device *device = cmd_buffer->device;
-
-   anv_cmd_buffer_end_batch_buffer(cmd_buffer);
-
-   if (cmd_buffer->level == VK_CMD_BUFFER_LEVEL_PRIMARY) {
-      /* The algorithm used to compute the validate list is not threadsafe as
-       * it uses the bo->index field.  We have to lock the device around it.
-       * Fortunately, the chances for contention here are probably very low.
-       */
-      pthread_mutex_lock(&device->mutex);
-      anv_cmd_buffer_prepare_execbuf(cmd_buffer);
-      pthread_mutex_unlock(&device->mutex);
-   }
-
-   return VK_SUCCESS;
-}
-
-void anv_CmdBindPipeline(
-    VkCmdBuffer                                 cmdBuffer,
-    VkPipelineBindPoint                         pipelineBindPoint,
-    VkPipeline                                  _pipeline)
-{
-   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
-   ANV_FROM_HANDLE(anv_pipeline, pipeline, _pipeline);
-
-   switch (pipelineBindPoint) {
-   case VK_PIPELINE_BIND_POINT_COMPUTE:
-      cmd_buffer->state.compute_pipeline = pipeline;
-      cmd_buffer->state.compute_dirty |= ANV_CMD_BUFFER_PIPELINE_DIRTY;
-      break;
-
-   case VK_PIPELINE_BIND_POINT_GRAPHICS:
-      cmd_buffer->state.pipeline = pipeline;
-      cmd_buffer->state.vb_dirty |= pipeline->vb_used;
-      cmd_buffer->state.dirty |= ANV_CMD_BUFFER_PIPELINE_DIRTY;
-      break;
-
-   default:
-      assert(!"invalid bind point");
-      break;
-   }
-}
-
-void anv_CmdBindDynamicViewportState(
-    VkCmdBuffer                                 cmdBuffer,
-    VkDynamicViewportState                      dynamicViewportState)
-{
-   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
-   ANV_FROM_HANDLE(anv_dynamic_vp_state, vp_state, dynamicViewportState);
-
-   cmd_buffer->state.vp_state = vp_state;
-   cmd_buffer->state.dirty |= ANV_CMD_BUFFER_VP_DIRTY;
-}
-
-void anv_CmdBindDynamicRasterState(
-    VkCmdBuffer                                 cmdBuffer,
-    VkDynamicRasterState                        dynamicRasterState)
-{
-   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
-   ANV_FROM_HANDLE(anv_dynamic_rs_state, rs_state, dynamicRasterState);
-
-   cmd_buffer->state.rs_state = rs_state;
-   cmd_buffer->state.dirty |= ANV_CMD_BUFFER_RS_DIRTY;
-}
-
-void anv_CmdBindDynamicColorBlendState(
-    VkCmdBuffer                                 cmdBuffer,
-    VkDynamicColorBlendState                    dynamicColorBlendState)
-{
-   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
-   ANV_FROM_HANDLE(anv_dynamic_cb_state, cb_state, dynamicColorBlendState);
-
-   cmd_buffer->state.cb_state = cb_state;
-   cmd_buffer->state.dirty |= ANV_CMD_BUFFER_CB_DIRTY;
-}
-
-void anv_CmdBindDynamicDepthStencilState(
-    VkCmdBuffer                                 cmdBuffer,
-    VkDynamicDepthStencilState                  dynamicDepthStencilState)
-{
-   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
-   ANV_FROM_HANDLE(anv_dynamic_ds_state, ds_state, dynamicDepthStencilState);
-
-   cmd_buffer->state.ds_state = ds_state;
-   cmd_buffer->state.dirty |= ANV_CMD_BUFFER_DS_DIRTY;
-}
-
-void anv_CmdBindDescriptorSets(
-    VkCmdBuffer                                 cmdBuffer,
-    VkPipelineBindPoint                         pipelineBindPoint,
-    VkPipelineLayout                            _layout,
-    uint32_t                                    firstSet,
-    uint32_t                                    setCount,
-    const VkDescriptorSet*                      pDescriptorSets,
-    uint32_t                                    dynamicOffsetCount,
-    const uint32_t*                             pDynamicOffsets)
-{
-   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
-   ANV_FROM_HANDLE(anv_pipeline_layout, layout, _layout);
-   struct anv_descriptor_set_layout *set_layout;
-
-   assert(firstSet + setCount < MAX_SETS);
-
-   uint32_t dynamic_slot = 0;
-   for (uint32_t i = 0; i < setCount; i++) {
-      ANV_FROM_HANDLE(anv_descriptor_set, set, pDescriptorSets[i]);
-      set_layout = layout->set[firstSet + i].layout;
-
-      cmd_buffer->state.descriptors[firstSet + i].set = set;
-
-      assert(set_layout->num_dynamic_buffers <
-             ARRAY_SIZE(cmd_buffer->state.descriptors[0].dynamic_offsets));
-      memcpy(cmd_buffer->state.descriptors[firstSet + i].dynamic_offsets,
-             pDynamicOffsets + dynamic_slot,
-             set_layout->num_dynamic_buffers * sizeof(*pDynamicOffsets));
-
-      cmd_buffer->state.descriptors_dirty |= set_layout->shader_stages;
-
-      dynamic_slot += set_layout->num_dynamic_buffers;
-   }
-}
-
-void anv_CmdBindIndexBuffer(
-    VkCmdBuffer                                 cmdBuffer,
-    VkBuffer                                    _buffer,
-    VkDeviceSize                                offset,
-    VkIndexType                                 indexType)
-{
-   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
-   ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
-
-   static const uint32_t vk_to_gen_index_type[] = {
-      [VK_INDEX_TYPE_UINT16]                    = INDEX_WORD,
-      [VK_INDEX_TYPE_UINT32]                    = INDEX_DWORD,
-   };
-
-   struct GEN8_3DSTATE_VF vf = {
-      GEN8_3DSTATE_VF_header,
-      .CutIndex = (indexType == VK_INDEX_TYPE_UINT16) ? UINT16_MAX : UINT32_MAX,
-   };
-   GEN8_3DSTATE_VF_pack(NULL, cmd_buffer->state.state_vf, &vf);
-
-   cmd_buffer->state.dirty |= ANV_CMD_BUFFER_INDEX_BUFFER_DIRTY;
-
-   anv_batch_emit(&cmd_buffer->batch, GEN8_3DSTATE_INDEX_BUFFER,
-                  .IndexFormat = vk_to_gen_index_type[indexType],
-                  .MemoryObjectControlState = GEN8_MOCS,
-                  .BufferStartingAddress = { buffer->bo, buffer->offset + offset },
-                  .BufferSize = buffer->size - offset);
-}
-
-void anv_CmdBindVertexBuffers(
-    VkCmdBuffer                                 cmdBuffer,
-    uint32_t                                    startBinding,
-    uint32_t                                    bindingCount,
-    const VkBuffer*                             pBuffers,
-    const VkDeviceSize*                         pOffsets)
-{
-   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
-   struct anv_vertex_binding *vb = cmd_buffer->state.vertex_bindings;
-
-   /* We have to defer setting up vertex buffer since we need the buffer
-    * stride from the pipeline. */
-
-   assert(startBinding + bindingCount < MAX_VBS);
-   for (uint32_t i = 0; i < bindingCount; i++) {
-      vb[startBinding + i].buffer = anv_buffer_from_handle(pBuffers[i]);
-      vb[startBinding + i].offset = pOffsets[i];
-      cmd_buffer->state.vb_dirty |= 1 << (startBinding + i);
-   }
-}
-
-static VkResult
-cmd_buffer_emit_binding_table(struct anv_cmd_buffer *cmd_buffer,
-                              unsigned stage, struct anv_state *bt_state)
-{
-   struct anv_framebuffer *fb = cmd_buffer->state.framebuffer;
-   struct anv_subpass *subpass = cmd_buffer->state.subpass;
-   struct anv_pipeline_layout *layout;
-   uint32_t attachments, bias, size;
-
-   if (stage == VK_SHADER_STAGE_COMPUTE)
-      layout = cmd_buffer->state.compute_pipeline->layout;
-   else
-      layout = cmd_buffer->state.pipeline->layout;
-
-   if (stage == VK_SHADER_STAGE_FRAGMENT) {
-      bias = MAX_RTS;
-      attachments = subpass->color_count;
-   } else {
-      bias = 0;
-      attachments = 0;
-   }
-
-   /* This is a little awkward: layout can be NULL but we still have to
-    * allocate and set a binding table for the PS stage for render
-    * targets. */
-   uint32_t surface_count = layout ? layout->stage[stage].surface_count : 0;
-
-   if (attachments + surface_count == 0)
-      return VK_SUCCESS;
-
-   size = (bias + surface_count) * sizeof(uint32_t);
-   *bt_state = anv_cmd_buffer_alloc_surface_state(cmd_buffer, size, 32);
-   uint32_t *bt_map = bt_state->map;
-
-   if (bt_state->map == NULL)
-      return VK_ERROR_OUT_OF_DEVICE_MEMORY;
-
-   /* This is highly annoying.  The Vulkan spec puts the depth-stencil
-    * attachments in with the color attachments.  Unfortunately, thanks to
-    * other aspects of the API, we cana't really saparate them before this
-    * point.  Therefore, we have to walk all of the attachments but only
-    * put the color attachments into the binding table.
-    */
-   for (uint32_t a = 0; a < attachments; a++) {
-      const struct anv_attachment_view *attachment =
-         fb->attachments[subpass->color_attachments[a]];
-
-      assert(attachment->attachment_type == ANV_ATTACHMENT_VIEW_TYPE_COLOR);
-      const struct anv_color_attachment_view *view =
-         (const struct anv_color_attachment_view *)attachment;
-
-      struct anv_state state =
-         anv_cmd_buffer_alloc_surface_state(cmd_buffer, 64, 64);
-
-      if (state.map == NULL)
-         return VK_ERROR_OUT_OF_DEVICE_MEMORY;
-
-      memcpy(state.map, view->view.surface_state.map, 64);
-
-      /* The address goes in dwords 8 and 9 of the SURFACE_STATE */
-      *(uint64_t *)(state.map + 8 * 4) =
-         anv_reloc_list_add(anv_cmd_buffer_current_surface_relocs(cmd_buffer),
-                            cmd_buffer->device,
-                            state.offset + 8 * 4,
-                            view->view.bo, view->view.offset);
-
-      bt_map[a] = state.offset;
-   }
-
-   if (layout == NULL)
-      return VK_SUCCESS;
-
-   for (uint32_t set = 0; set < layout->num_sets; set++) {
-      struct anv_descriptor_set_binding *d = &cmd_buffer->state.descriptors[set];
-      struct anv_descriptor_set_layout *set_layout = layout->set[set].layout;
-      struct anv_descriptor_slot *surface_slots =
-         set_layout->stage[stage].surface_start;
-
-      uint32_t start = bias + layout->set[set].surface_start[stage];
-
-      for (uint32_t b = 0; b < set_layout->stage[stage].surface_count; b++) {
-         struct anv_surface_view *view =
-            d->set->descriptors[surface_slots[b].index].view;
-
-         if (!view)
-            continue;
-
-         struct anv_state state =
-            anv_cmd_buffer_alloc_surface_state(cmd_buffer, 64, 64);
-
-         if (state.map == NULL)
-            return VK_ERROR_OUT_OF_DEVICE_MEMORY;
-
-         uint32_t offset;
-         if (surface_slots[b].dynamic_slot >= 0) {
-            uint32_t dynamic_offset =
-               d->dynamic_offsets[surface_slots[b].dynamic_slot];
-
-            offset = view->offset + dynamic_offset;
-            anv_fill_buffer_surface_state(state.map, view->format, offset,
-                                          view->range - dynamic_offset);
-         } else {
-            offset = view->offset;
-            memcpy(state.map, view->surface_state.map, 64);
-         }
-
-         /* The address goes in dwords 8 and 9 of the SURFACE_STATE */
-         *(uint64_t *)(state.map + 8 * 4) =
-            anv_reloc_list_add(anv_cmd_buffer_current_surface_relocs(cmd_buffer),
-                               cmd_buffer->device,
-                               state.offset + 8 * 4,
-                               view->bo, offset);
-
-         bt_map[start + b] = state.offset;
-      }
-   }
-
-   return VK_SUCCESS;
-}
-
-static VkResult
-cmd_buffer_emit_samplers(struct anv_cmd_buffer *cmd_buffer,
-                         unsigned stage, struct anv_state *state)
-{
-   struct anv_pipeline_layout *layout;
-   uint32_t sampler_count;
-
-   if (stage == VK_SHADER_STAGE_COMPUTE)
-      layout = cmd_buffer->state.compute_pipeline->layout;
-   else
-      layout = cmd_buffer->state.pipeline->layout;
-
-   sampler_count = layout ? layout->stage[stage].sampler_count : 0;
-   if (sampler_count == 0)
-      return VK_SUCCESS;
-
-   uint32_t size = sampler_count * 16;
-   *state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, size, 32);
-
-   if (state->map == NULL)
-      return VK_ERROR_OUT_OF_DEVICE_MEMORY;
-
-   for (uint32_t set = 0; set < layout->num_sets; set++) {
-      struct anv_descriptor_set_binding *d = &cmd_buffer->state.descriptors[set];
-      struct anv_descriptor_set_layout *set_layout = layout->set[set].layout;
-      struct anv_descriptor_slot *sampler_slots =
-         set_layout->stage[stage].sampler_start;
-
-      uint32_t start = layout->set[set].sampler_start[stage];
-
-      for (uint32_t b = 0; b < set_layout->stage[stage].sampler_count; b++) {
-         struct anv_sampler *sampler =
-            d->set->descriptors[sampler_slots[b].index].sampler;
-
-         if (!sampler)
-            continue;
-
-         memcpy(state->map + (start + b) * 16,
-                sampler->state, sizeof(sampler->state));
-      }
-   }
-
-   return VK_SUCCESS;
-}
-
-static VkResult
-flush_descriptor_set(struct anv_cmd_buffer *cmd_buffer, uint32_t stage)
-{
-   struct anv_state surfaces = { 0, }, samplers = { 0, };
-   VkResult result;
-
-   result = cmd_buffer_emit_samplers(cmd_buffer, stage, &samplers);
-   if (result != VK_SUCCESS)
-      return result;
-   result = cmd_buffer_emit_binding_table(cmd_buffer, stage, &surfaces);
-   if (result != VK_SUCCESS)
-      return result;
-
-   static const uint32_t sampler_state_opcodes[] = {
-      [VK_SHADER_STAGE_VERTEX]                  = 43,
-      [VK_SHADER_STAGE_TESS_CONTROL]            = 44, /* HS */
-      [VK_SHADER_STAGE_TESS_EVALUATION]         = 45, /* DS */
-      [VK_SHADER_STAGE_GEOMETRY]                = 46,
-      [VK_SHADER_STAGE_FRAGMENT]                = 47,
-      [VK_SHADER_STAGE_COMPUTE]                 = 0,
-   };
-
-   static const uint32_t binding_table_opcodes[] = {
-      [VK_SHADER_STAGE_VERTEX]                  = 38,
-      [VK_SHADER_STAGE_TESS_CONTROL]            = 39,
-      [VK_SHADER_STAGE_TESS_EVALUATION]         = 40,
-      [VK_SHADER_STAGE_GEOMETRY]                = 41,
-      [VK_SHADER_STAGE_FRAGMENT]                = 42,
-      [VK_SHADER_STAGE_COMPUTE]                 = 0,
-   };
-
-   if (samplers.alloc_size > 0) {
-      anv_batch_emit(&cmd_buffer->batch,
-                     GEN8_3DSTATE_SAMPLER_STATE_POINTERS_VS,
-                     ._3DCommandSubOpcode  = sampler_state_opcodes[stage],
-                     .PointertoVSSamplerState = samplers.offset);
-   }
-
-   if (surfaces.alloc_size > 0) {
-      anv_batch_emit(&cmd_buffer->batch,
-                     GEN8_3DSTATE_BINDING_TABLE_POINTERS_VS,
-                     ._3DCommandSubOpcode  = binding_table_opcodes[stage],
-                     .PointertoVSBindingTable = surfaces.offset);
-   }
-
-   return VK_SUCCESS;
-}
-
-static void
-flush_descriptor_sets(struct anv_cmd_buffer *cmd_buffer)
-{
-   uint32_t s, dirty = cmd_buffer->state.descriptors_dirty &
-                       cmd_buffer->state.pipeline->active_stages;
-
-   VkResult result = VK_SUCCESS;
-   for_each_bit(s, dirty) {
-      result = flush_descriptor_set(cmd_buffer, s);
-      if (result != VK_SUCCESS)
-         break;
-   }
-
-   if (result != VK_SUCCESS) {
-      assert(result == VK_ERROR_OUT_OF_DEVICE_MEMORY);
-
-      result = anv_cmd_buffer_new_surface_state_bo(cmd_buffer);
-      assert(result == VK_SUCCESS);
-
-      /* Re-emit state base addresses so we get the new surface state base
-       * address before we start emitting binding tables etc.
-       */
-      anv_cmd_buffer_emit_state_base_address(cmd_buffer);
-
-      /* Re-emit all active binding tables */
-      for_each_bit(s, cmd_buffer->state.pipeline->active_stages) {
-         result = flush_descriptor_set(cmd_buffer, s);
-
-         /* It had better succeed this time */
-         assert(result == VK_SUCCESS);
-      }
-   }
-
-   cmd_buffer->state.descriptors_dirty &= ~cmd_buffer->state.pipeline->active_stages;
-}
-
-static struct anv_state
-anv_cmd_buffer_emit_dynamic(struct anv_cmd_buffer *cmd_buffer,
-                             uint32_t *a, uint32_t dwords, uint32_t alignment)
-{
-   struct anv_state state;
-
-   state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
-                                              dwords * 4, alignment);
-   memcpy(state.map, a, dwords * 4);
-
-   VG(VALGRIND_CHECK_MEM_IS_DEFINED(state.map, dwords * 4));
-
-   return state;
-}
-
-static struct anv_state
-anv_cmd_buffer_merge_dynamic(struct anv_cmd_buffer *cmd_buffer,
-                             uint32_t *a, uint32_t *b,
-                             uint32_t dwords, uint32_t alignment)
-{
-   struct anv_state state;
-   uint32_t *p;
-
-   state = anv_cmd_buffer_alloc_dynamic_state(cmd_buffer,
-                                              dwords * 4, alignment);
-   p = state.map;
-   for (uint32_t i = 0; i < dwords; i++)
-      p[i] = a[i] | b[i];
-
-   VG(VALGRIND_CHECK_MEM_IS_DEFINED(p, dwords * 4));
-
-   return state;
-}
-
-static VkResult
-flush_compute_descriptor_set(struct anv_cmd_buffer *cmd_buffer)
-{
-   struct anv_device *device = cmd_buffer->device;
-   struct anv_pipeline *pipeline = cmd_buffer->state.compute_pipeline;
-   struct anv_state surfaces = { 0, }, samplers = { 0, };
-   VkResult result;
-
-   result = cmd_buffer_emit_samplers(cmd_buffer,
-                                     VK_SHADER_STAGE_COMPUTE, &samplers);
-   if (result != VK_SUCCESS)
-      return result;
-   result = cmd_buffer_emit_binding_table(cmd_buffer,
-                                          VK_SHADER_STAGE_COMPUTE, &surfaces);
-   if (result != VK_SUCCESS)
-      return result;
-
-   struct GEN8_INTERFACE_DESCRIPTOR_DATA desc = {
-      .KernelStartPointer = pipeline->cs_simd,
-      .KernelStartPointerHigh = 0,
-      .BindingTablePointer = surfaces.offset,
-      .BindingTableEntryCount = 0,
-      .SamplerStatePointer = samplers.offset,
-      .SamplerCount = 0,
-      .NumberofThreadsinGPGPUThreadGroup = 0 /* FIXME: Really? */
-   };
-
-   uint32_t size = GEN8_INTERFACE_DESCRIPTOR_DATA_length * sizeof(uint32_t);
-   struct anv_state state =
-      anv_state_pool_alloc(&device->dynamic_state_pool, size, 64);
-
-   GEN8_INTERFACE_DESCRIPTOR_DATA_pack(NULL, state.map, &desc);
-
-   anv_batch_emit(&cmd_buffer->batch, GEN8_MEDIA_INTERFACE_DESCRIPTOR_LOAD,
-                  .InterfaceDescriptorTotalLength = size,
-                  .InterfaceDescriptorDataStartAddress = state.offset);
-
-   return VK_SUCCESS;
-}
-
-static void
-anv_cmd_buffer_flush_compute_state(struct anv_cmd_buffer *cmd_buffer)
-{
-   struct anv_pipeline *pipeline = cmd_buffer->state.compute_pipeline;
-   VkResult result;
-
-   assert(pipeline->active_stages == VK_SHADER_STAGE_COMPUTE_BIT);
-
-   if (cmd_buffer->state.current_pipeline != GPGPU) {
-      anv_batch_emit(&cmd_buffer->batch, GEN8_PIPELINE_SELECT,
-                     .PipelineSelection = GPGPU);
-      cmd_buffer->state.current_pipeline = GPGPU;
-   }
-
-   if (cmd_buffer->state.compute_dirty & ANV_CMD_BUFFER_PIPELINE_DIRTY)
-      anv_batch_emit_batch(&cmd_buffer->batch, &pipeline->batch);
-
-   if ((cmd_buffer->state.descriptors_dirty & VK_SHADER_STAGE_COMPUTE_BIT) ||
-       (cmd_buffer->state.compute_dirty & ANV_CMD_BUFFER_PIPELINE_DIRTY)) {
-      result = flush_compute_descriptor_set(cmd_buffer);
-      assert(result == VK_SUCCESS);
-      cmd_buffer->state.descriptors_dirty &= ~VK_SHADER_STAGE_COMPUTE;
-   }
-
-   cmd_buffer->state.compute_dirty = 0;
-}
-
-static void
-anv_cmd_buffer_flush_state(struct anv_cmd_buffer *cmd_buffer)
-{
-   struct anv_pipeline *pipeline = cmd_buffer->state.pipeline;
-   uint32_t *p;
-
-   uint32_t vb_emit = cmd_buffer->state.vb_dirty & pipeline->vb_used;
-
-   assert((pipeline->active_stages & VK_SHADER_STAGE_COMPUTE_BIT) == 0);
-
-   if (cmd_buffer->state.current_pipeline != _3D) {
-      anv_batch_emit(&cmd_buffer->batch, GEN8_PIPELINE_SELECT,
-                     .PipelineSelection = _3D);
-      cmd_buffer->state.current_pipeline = _3D;
-   }
-
-   if (vb_emit) {
-      const uint32_t num_buffers = __builtin_popcount(vb_emit);
-      const uint32_t num_dwords = 1 + num_buffers * 4;
-
-      p = anv_batch_emitn(&cmd_buffer->batch, num_dwords,
-                          GEN8_3DSTATE_VERTEX_BUFFERS);
-      uint32_t vb, i = 0;
-      for_each_bit(vb, vb_emit) {
-         struct anv_buffer *buffer = cmd_buffer->state.vertex_bindings[vb].buffer;
-         uint32_t offset = cmd_buffer->state.vertex_bindings[vb].offset;
-
-         struct GEN8_VERTEX_BUFFER_STATE state = {
-            .VertexBufferIndex = vb,
-            .MemoryObjectControlState = GEN8_MOCS,
-            .AddressModifyEnable = true,
-            .BufferPitch = pipeline->binding_stride[vb],
-            .BufferStartingAddress = { buffer->bo, buffer->offset + offset },
-            .BufferSize = buffer->size - offset
-         };
-
-         GEN8_VERTEX_BUFFER_STATE_pack(&cmd_buffer->batch, &p[1 + i * 4], &state);
-         i++;
-      }
-   }
-
-   if (cmd_buffer->state.dirty & ANV_CMD_BUFFER_PIPELINE_DIRTY) {
-      /* If somebody compiled a pipeline after starting a command buffer the
-       * scratch bo may have grown since we started this cmd buffer (and
-       * emitted STATE_BASE_ADDRESS).  If we're binding that pipeline now,
-       * reemit STATE_BASE_ADDRESS so that we use the bigger scratch bo. */
-      if (cmd_buffer->state.scratch_size < pipeline->total_scratch)
-         anv_cmd_buffer_emit_state_base_address(cmd_buffer);
-
-      anv_batch_emit_batch(&cmd_buffer->batch, &pipeline->batch);
-   }
-
-   if (cmd_buffer->state.descriptors_dirty)
-      flush_descriptor_sets(cmd_buffer);
-
-   if (cmd_buffer->state.dirty & ANV_CMD_BUFFER_VP_DIRTY) {
-      struct anv_dynamic_vp_state *vp_state = cmd_buffer->state.vp_state;
-      anv_batch_emit(&cmd_buffer->batch, GEN8_3DSTATE_SCISSOR_STATE_POINTERS,
-                     .ScissorRectPointer = vp_state->scissor.offset);
-      anv_batch_emit(&cmd_buffer->batch, GEN8_3DSTATE_VIEWPORT_STATE_POINTERS_CC,
-                     .CCViewportPointer = vp_state->cc_vp.offset);
-      anv_batch_emit(&cmd_buffer->batch, GEN8_3DSTATE_VIEWPORT_STATE_POINTERS_SF_CLIP,
-                     .SFClipViewportPointer = vp_state->sf_clip_vp.offset);
-   }
-
-   if (cmd_buffer->state.dirty & (ANV_CMD_BUFFER_PIPELINE_DIRTY |
-                                  ANV_CMD_BUFFER_RS_DIRTY)) {
-      anv_batch_emit_merge(&cmd_buffer->batch,
-                           cmd_buffer->state.rs_state->state_sf,
-                           pipeline->state_sf);
-      anv_batch_emit_merge(&cmd_buffer->batch,
-                           cmd_buffer->state.rs_state->state_raster,
-                           pipeline->state_raster);
-   }
-
-   if (cmd_buffer->state.ds_state &&
-       (cmd_buffer->state.dirty & (ANV_CMD_BUFFER_PIPELINE_DIRTY |
-                                   ANV_CMD_BUFFER_DS_DIRTY))) {
-      anv_batch_emit_merge(&cmd_buffer->batch,
-                           cmd_buffer->state.ds_state->state_wm_depth_stencil,
-                           pipeline->state_wm_depth_stencil);
-   }
-
-   if (cmd_buffer->state.dirty & (ANV_CMD_BUFFER_CB_DIRTY |
-                                  ANV_CMD_BUFFER_DS_DIRTY)) {
-      struct anv_state state;
-      if (cmd_buffer->state.ds_state == NULL)
-         state = anv_cmd_buffer_emit_dynamic(cmd_buffer,
-                                             cmd_buffer->state.cb_state->state_color_calc,
-                                             GEN8_COLOR_CALC_STATE_length, 64);
-      else if (cmd_buffer->state.cb_state == NULL)
-         state = anv_cmd_buffer_emit_dynamic(cmd_buffer,
-                                             cmd_buffer->state.ds_state->state_color_calc,
-                                             GEN8_COLOR_CALC_STATE_length, 64);
-      else
-         state = anv_cmd_buffer_merge_dynamic(cmd_buffer,
-                                              cmd_buffer->state.ds_state->state_color_calc,
-                                              cmd_buffer->state.cb_state->state_color_calc,
-                                              GEN8_COLOR_CALC_STATE_length, 64);
-
-      anv_batch_emit(&cmd_buffer->batch,
-                     GEN8_3DSTATE_CC_STATE_POINTERS,
-                     .ColorCalcStatePointer = state.offset,
-                     .ColorCalcStatePointerValid = true);
-   }
-
-   if (cmd_buffer->state.dirty & (ANV_CMD_BUFFER_PIPELINE_DIRTY |
-                                  ANV_CMD_BUFFER_INDEX_BUFFER_DIRTY)) {
-      anv_batch_emit_merge(&cmd_buffer->batch,
-                           cmd_buffer->state.state_vf, pipeline->state_vf);
-   }
-
-   cmd_buffer->state.vb_dirty &= ~vb_emit;
-   cmd_buffer->state.dirty = 0;
-}
-
-void anv_CmdDraw(
-    VkCmdBuffer                                 cmdBuffer,
-    uint32_t                                    firstVertex,
-    uint32_t                                    vertexCount,
-    uint32_t                                    firstInstance,
-    uint32_t                                    instanceCount)
-{
-   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
-
-   anv_cmd_buffer_flush_state(cmd_buffer);
-
-   anv_batch_emit(&cmd_buffer->batch, GEN8_3DPRIMITIVE,
-                  .VertexAccessType = SEQUENTIAL,
-                  .VertexCountPerInstance = vertexCount,
-                  .StartVertexLocation = firstVertex,
-                  .InstanceCount = instanceCount,
-                  .StartInstanceLocation = firstInstance,
-                  .BaseVertexLocation = 0);
-}
-
-void anv_CmdDrawIndexed(
-    VkCmdBuffer                                 cmdBuffer,
-    uint32_t                                    firstIndex,
-    uint32_t                                    indexCount,
-    int32_t                                     vertexOffset,
-    uint32_t                                    firstInstance,
-    uint32_t                                    instanceCount)
-{
-   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
-
-   anv_cmd_buffer_flush_state(cmd_buffer);
-
-   anv_batch_emit(&cmd_buffer->batch, GEN8_3DPRIMITIVE,
-                  .VertexAccessType = RANDOM,
-                  .VertexCountPerInstance = indexCount,
-                  .StartVertexLocation = firstIndex,
-                  .InstanceCount = instanceCount,
-                  .StartInstanceLocation = firstInstance,
-                  .BaseVertexLocation = vertexOffset);
-}
-
-static void
-anv_batch_lrm(struct anv_batch *batch,
-              uint32_t reg, struct anv_bo *bo, uint32_t offset)
-{
-   anv_batch_emit(batch, GEN8_MI_LOAD_REGISTER_MEM,
-                  .RegisterAddress = reg,
-                  .MemoryAddress = { bo, offset });
-}
-
-static void
-anv_batch_lri(struct anv_batch *batch, uint32_t reg, uint32_t imm)
-{
-   anv_batch_emit(batch, GEN8_MI_LOAD_REGISTER_IMM,
-                  .RegisterOffset = reg,
-                  .DataDWord = imm);
-}
-
-/* Auto-Draw / Indirect Registers */
-#define GEN7_3DPRIM_END_OFFSET          0x2420
-#define GEN7_3DPRIM_START_VERTEX        0x2430
-#define GEN7_3DPRIM_VERTEX_COUNT        0x2434
-#define GEN7_3DPRIM_INSTANCE_COUNT      0x2438
-#define GEN7_3DPRIM_START_INSTANCE      0x243C
-#define GEN7_3DPRIM_BASE_VERTEX         0x2440
-
-void anv_CmdDrawIndirect(
-    VkCmdBuffer                                 cmdBuffer,
-    VkBuffer                                    _buffer,
-    VkDeviceSize                                offset,
-    uint32_t                                    count,
-    uint32_t                                    stride)
-{
-   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
-   ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
-   struct anv_bo *bo = buffer->bo;
-   uint32_t bo_offset = buffer->offset + offset;
-
-   anv_cmd_buffer_flush_state(cmd_buffer);
-
-   anv_batch_lrm(&cmd_buffer->batch, GEN7_3DPRIM_VERTEX_COUNT, bo, bo_offset);
-   anv_batch_lrm(&cmd_buffer->batch, GEN7_3DPRIM_INSTANCE_COUNT, bo, bo_offset + 4);
-   anv_batch_lrm(&cmd_buffer->batch, GEN7_3DPRIM_START_VERTEX, bo, bo_offset + 8);
-   anv_batch_lrm(&cmd_buffer->batch, GEN7_3DPRIM_START_INSTANCE, bo, bo_offset + 12);
-   anv_batch_lri(&cmd_buffer->batch, GEN7_3DPRIM_BASE_VERTEX, 0);
-
-   anv_batch_emit(&cmd_buffer->batch, GEN8_3DPRIMITIVE,
-                  .IndirectParameterEnable = true,
-                  .VertexAccessType = SEQUENTIAL);
-}
-
-void anv_CmdDrawIndexedIndirect(
-    VkCmdBuffer                                 cmdBuffer,
-    VkBuffer                                    _buffer,
-    VkDeviceSize                                offset,
-    uint32_t                                    count,
-    uint32_t                                    stride)
-{
-   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
-   ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
-   struct anv_bo *bo = buffer->bo;
-   uint32_t bo_offset = buffer->offset + offset;
-
-   anv_cmd_buffer_flush_state(cmd_buffer);
-
-   anv_batch_lrm(&cmd_buffer->batch, GEN7_3DPRIM_VERTEX_COUNT, bo, bo_offset);
-   anv_batch_lrm(&cmd_buffer->batch, GEN7_3DPRIM_INSTANCE_COUNT, bo, bo_offset + 4);
-   anv_batch_lrm(&cmd_buffer->batch, GEN7_3DPRIM_START_VERTEX, bo, bo_offset + 8);
-   anv_batch_lrm(&cmd_buffer->batch, GEN7_3DPRIM_BASE_VERTEX, bo, bo_offset + 12);
-   anv_batch_lrm(&cmd_buffer->batch, GEN7_3DPRIM_START_INSTANCE, bo, bo_offset + 16);
-
-   anv_batch_emit(&cmd_buffer->batch, GEN8_3DPRIMITIVE,
-                  .IndirectParameterEnable = true,
-                  .VertexAccessType = RANDOM);
-}
-
-void anv_CmdDispatch(
-    VkCmdBuffer                                 cmdBuffer,
-    uint32_t                                    x,
-    uint32_t                                    y,
-    uint32_t                                    z)
-{
-   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
-   struct anv_pipeline *pipeline = cmd_buffer->state.compute_pipeline;
-   struct brw_cs_prog_data *prog_data = &pipeline->cs_prog_data;
-
-   anv_cmd_buffer_flush_compute_state(cmd_buffer);
-
-   anv_batch_emit(&cmd_buffer->batch, GEN8_GPGPU_WALKER,
-                  .SIMDSize = prog_data->simd_size / 16,
-                  .ThreadDepthCounterMaximum = 0,
-                  .ThreadHeightCounterMaximum = 0,
-                  .ThreadWidthCounterMaximum = pipeline->cs_thread_width_max,
-                  .ThreadGroupIDXDimension = x,
-                  .ThreadGroupIDYDimension = y,
-                  .ThreadGroupIDZDimension = z,
-                  .RightExecutionMask = pipeline->cs_right_mask,
-                  .BottomExecutionMask = 0xffffffff);
-
-   anv_batch_emit(&cmd_buffer->batch, GEN8_MEDIA_STATE_FLUSH);
-}
-
-#define GPGPU_DISPATCHDIMX 0x2500
-#define GPGPU_DISPATCHDIMY 0x2504
-#define GPGPU_DISPATCHDIMZ 0x2508
-
-void anv_CmdDispatchIndirect(
-    VkCmdBuffer                                 cmdBuffer,
-    VkBuffer                                    _buffer,
-    VkDeviceSize                                offset)
-{
-   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
-   ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
-   struct anv_pipeline *pipeline = cmd_buffer->state.compute_pipeline;
-   struct brw_cs_prog_data *prog_data = &pipeline->cs_prog_data;
-   struct anv_bo *bo = buffer->bo;
-   uint32_t bo_offset = buffer->offset + offset;
-
-   anv_cmd_buffer_flush_compute_state(cmd_buffer);
-
-   anv_batch_lrm(&cmd_buffer->batch, GPGPU_DISPATCHDIMX, bo, bo_offset);
-   anv_batch_lrm(&cmd_buffer->batch, GPGPU_DISPATCHDIMY, bo, bo_offset + 4);
-   anv_batch_lrm(&cmd_buffer->batch, GPGPU_DISPATCHDIMZ, bo, bo_offset + 8);
-
-   anv_batch_emit(&cmd_buffer->batch, GEN8_GPGPU_WALKER,
-                  .IndirectParameterEnable = true,
-                  .SIMDSize = prog_data->simd_size / 16,
-                  .ThreadDepthCounterMaximum = 0,
-                  .ThreadHeightCounterMaximum = 0,
-                  .ThreadWidthCounterMaximum = pipeline->cs_thread_width_max,
-                  .RightExecutionMask = pipeline->cs_right_mask,
-                  .BottomExecutionMask = 0xffffffff);
-
-   anv_batch_emit(&cmd_buffer->batch, GEN8_MEDIA_STATE_FLUSH);
-}
-
-void anv_CmdSetEvent(
-    VkCmdBuffer                                 cmdBuffer,
-    VkEvent                                     event,
-    VkPipelineStageFlags                        stageMask)
-{
-   stub();
-}
-
-void anv_CmdResetEvent(
-    VkCmdBuffer                                 cmdBuffer,
-    VkEvent                                     event,
-    VkPipelineStageFlags                        stageMask)
-{
-   stub();
-}
-
-void anv_CmdWaitEvents(
-    VkCmdBuffer                                 cmdBuffer,
-    uint32_t                                    eventCount,
-    const VkEvent*                              pEvents,
-    VkPipelineStageFlags                        srcStageMask,
-    VkPipelineStageFlags                        destStageMask,
-    uint32_t                                    memBarrierCount,
-    const void* const*                          ppMemBarriers)
-{
-   stub();
-}
-
-void anv_CmdPipelineBarrier(
-    VkCmdBuffer                                 cmdBuffer,
-    VkPipelineStageFlags                        srcStageMask,
-    VkPipelineStageFlags                        destStageMask,
-    VkBool32                                    byRegion,
-    uint32_t                                    memBarrierCount,
-    const void* const*                          ppMemBarriers)
-{
-   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
-   uint32_t b, *dw;
-
-   struct GEN8_PIPE_CONTROL cmd = {
-      GEN8_PIPE_CONTROL_header,
-      .PostSyncOperation = NoWrite,
-   };
-
-   /* XXX: I think waitEvent is a no-op on our HW.  We should verify that. */
-
-   if (anv_clear_mask(&srcStageMask, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT)) {
-      /* This is just what PIPE_CONTROL does */
-   }
-
-   if (anv_clear_mask(&srcStageMask,
-                      VK_PIPELINE_STAGE_DRAW_INDIRECT_BIT |
-                      VK_PIPELINE_STAGE_VERTEX_INPUT_BIT |
-                      VK_PIPELINE_STAGE_VERTEX_SHADER_BIT |
-                      VK_PIPELINE_STAGE_TESS_CONTROL_SHADER_BIT |
-                      VK_PIPELINE_STAGE_TESS_EVALUATION_SHADER_BIT |
-                      VK_PIPELINE_STAGE_GEOMETRY_SHADER_BIT |
-                      VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT |
-                      VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT |
-                      VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT |
-                      VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT)) {
-      cmd.StallAtPixelScoreboard = true;
-   }
-
-
-   if (anv_clear_mask(&srcStageMask,
-                      VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT |
-                      VK_PIPELINE_STAGE_TRANSFER_BIT |
-                      VK_PIPELINE_STAGE_TRANSITION_BIT)) {
-      cmd.CommandStreamerStallEnable = true;
-   }
-
-   if (anv_clear_mask(&srcStageMask, VK_PIPELINE_STAGE_HOST_BIT)) {
-      anv_finishme("VK_PIPE_EVENT_CPU_SIGNAL_BIT");
-   }
-
-   /* On our hardware, all stages will wait for execution as needed. */
-   (void)destStageMask;
-
-   /* We checked all known VkPipeEventFlags. */
-   anv_assert(srcStageMask == 0);
-
-   /* XXX: Right now, we're really dumb and just flush whatever categories
-    * the app asks for.  One of these days we may make this a bit better
-    * but right now that's all the hardware allows for in most areas.
-    */
-   VkMemoryOutputFlags out_flags = 0;
-   VkMemoryInputFlags in_flags = 0;
-
-   for (uint32_t i = 0; i < memBarrierCount; i++) {
-      const struct anv_common *common = ppMemBarriers[i];
-      switch (common->sType) {
-      case VK_STRUCTURE_TYPE_MEMORY_BARRIER: {
-         ANV_COMMON_TO_STRUCT(VkMemoryBarrier, barrier, common);
-         out_flags |= barrier->outputMask;
-         in_flags |= barrier->inputMask;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_BUFFER_MEMORY_BARRIER: {
-         ANV_COMMON_TO_STRUCT(VkBufferMemoryBarrier, barrier, common);
-         out_flags |= barrier->outputMask;
-         in_flags |= barrier->inputMask;
-         break;
-      }
-      case VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER: {
-         ANV_COMMON_TO_STRUCT(VkImageMemoryBarrier, barrier, common);
-         out_flags |= barrier->outputMask;
-         in_flags |= barrier->inputMask;
-         break;
-      }
-      default:
-         unreachable("Invalid memory barrier type");
-      }
-   }
-
-   for_each_bit(b, out_flags) {
-      switch ((VkMemoryOutputFlags)(1 << b)) {
-      case VK_MEMORY_OUTPUT_HOST_WRITE_BIT:
-         break; /* FIXME: Little-core systems */
-      case VK_MEMORY_OUTPUT_SHADER_WRITE_BIT:
-         cmd.DCFlushEnable = true;
-         break;
-      case VK_MEMORY_OUTPUT_COLOR_ATTACHMENT_BIT:
-         cmd.RenderTargetCacheFlushEnable = true;
-         break;
-      case VK_MEMORY_OUTPUT_DEPTH_STENCIL_ATTACHMENT_BIT:
-         cmd.DepthCacheFlushEnable = true;
-         break;
-      case VK_MEMORY_OUTPUT_TRANSFER_BIT:
-         cmd.RenderTargetCacheFlushEnable = true;
-         cmd.DepthCacheFlushEnable = true;
-         break;
-      default:
-         unreachable("Invalid memory output flag");
-      }
-   }
-
-   for_each_bit(b, out_flags) {
-      switch ((VkMemoryInputFlags)(1 << b)) {
-      case VK_MEMORY_INPUT_HOST_READ_BIT:
-         break; /* FIXME: Little-core systems */
-      case VK_MEMORY_INPUT_INDIRECT_COMMAND_BIT:
-      case VK_MEMORY_INPUT_INDEX_FETCH_BIT:
-      case VK_MEMORY_INPUT_VERTEX_ATTRIBUTE_FETCH_BIT:
-         cmd.VFCacheInvalidationEnable = true;
-         break;
-      case VK_MEMORY_INPUT_UNIFORM_READ_BIT:
-         cmd.ConstantCacheInvalidationEnable = true;
-         /* fallthrough */
-      case VK_MEMORY_INPUT_SHADER_READ_BIT:
-         cmd.DCFlushEnable = true;
-         cmd.TextureCacheInvalidationEnable = true;
-         break;
-      case VK_MEMORY_INPUT_COLOR_ATTACHMENT_BIT:
-      case VK_MEMORY_INPUT_DEPTH_STENCIL_ATTACHMENT_BIT:
-         break; /* XXX: Hunh? */
-      case VK_MEMORY_INPUT_TRANSFER_BIT:
-         cmd.TextureCacheInvalidationEnable = true;
-         break;
-      }
-   }
-
-   dw = anv_batch_emit_dwords(&cmd_buffer->batch, GEN8_PIPE_CONTROL_length);
-   GEN8_PIPE_CONTROL_pack(&cmd_buffer->batch, dw, &cmd);
-}
-
-void anv_CmdPushConstants(
-    VkCmdBuffer                                 cmdBuffer,
-    VkPipelineLayout                            layout,
-    VkShaderStageFlags                          stageFlags,
-    uint32_t                                    start,
-    uint32_t                                    length,
-    const void*                                 values)
-{
-   stub();
-}
-
-static void
-anv_cmd_buffer_emit_depth_stencil(struct anv_cmd_buffer *cmd_buffer)
-{
-   struct anv_subpass *subpass = cmd_buffer->state.subpass;
-   struct anv_framebuffer *fb = cmd_buffer->state.framebuffer;
-   const struct anv_depth_stencil_view *view;
-
-   static const struct anv_depth_stencil_view null_view =
-      { .depth_format = D16_UNORM, .depth_stride = 0, .stencil_stride = 0 };
-
-   if (subpass->depth_stencil_attachment != VK_ATTACHMENT_UNUSED) {
-      const struct anv_attachment_view *aview =
-         fb->attachments[subpass->depth_stencil_attachment];
-      assert(aview->attachment_type == ANV_ATTACHMENT_VIEW_TYPE_DEPTH_STENCIL);
-      view = (const struct anv_depth_stencil_view *)aview;
-   } else {
-      view = &null_view;
-   }
-
-   /* FIXME: Implement the PMA stall W/A */
-   /* FIXME: Width and Height are wrong */
-
-   anv_batch_emit(&cmd_buffer->batch, GEN8_3DSTATE_DEPTH_BUFFER,
-                  .SurfaceType = SURFTYPE_2D,
-                  .DepthWriteEnable = view->depth_stride > 0,
-                  .StencilWriteEnable = view->stencil_stride > 0,
-                  .HierarchicalDepthBufferEnable = false,
-                  .SurfaceFormat = view->depth_format,
-                  .SurfacePitch = view->depth_stride > 0 ? view->depth_stride - 1 : 0,
-                  .SurfaceBaseAddress = { view->bo,  view->depth_offset },
-                  .Height = cmd_buffer->state.framebuffer->height - 1,
-                  .Width = cmd_buffer->state.framebuffer->width - 1,
-                  .LOD = 0,
-                  .Depth = 1 - 1,
-                  .MinimumArrayElement = 0,
-                  .DepthBufferObjectControlState = GEN8_MOCS,
-                  .RenderTargetViewExtent = 1 - 1,
-                  .SurfaceQPitch = view->depth_qpitch >> 2);
-
-   /* Disable hierarchial depth buffers. */
-   anv_batch_emit(&cmd_buffer->batch, GEN8_3DSTATE_HIER_DEPTH_BUFFER);
-
-   anv_batch_emit(&cmd_buffer->batch, GEN8_3DSTATE_STENCIL_BUFFER,
-                  .StencilBufferEnable = view->stencil_stride > 0,
-                  .StencilBufferObjectControlState = GEN8_MOCS,
-                  .SurfacePitch = view->stencil_stride > 0 ? view->stencil_stride - 1 : 0,
-                  .SurfaceBaseAddress = { view->bo, view->stencil_offset },
-                  .SurfaceQPitch = view->stencil_qpitch >> 2);
-
-   /* Clear the clear params. */
-   anv_batch_emit(&cmd_buffer->batch, GEN8_3DSTATE_CLEAR_PARAMS);
-}
-
-void
-anv_cmd_buffer_begin_subpass(struct anv_cmd_buffer *cmd_buffer,
-                             struct anv_subpass *subpass)
-{
-   cmd_buffer->state.subpass = subpass;
-
-   cmd_buffer->state.descriptors_dirty |= VK_SHADER_STAGE_FRAGMENT_BIT;
-
-   anv_cmd_buffer_emit_depth_stencil(cmd_buffer);
-}
-
-void anv_CmdBeginRenderPass(
-    VkCmdBuffer                                 cmdBuffer,
-    const VkRenderPassBeginInfo*                pRenderPassBegin,
-    VkRenderPassContents                        contents)
-{
-   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
-   ANV_FROM_HANDLE(anv_render_pass, pass, pRenderPassBegin->renderPass);
-   ANV_FROM_HANDLE(anv_framebuffer, framebuffer, pRenderPassBegin->framebuffer);
-
-   cmd_buffer->state.framebuffer = framebuffer;
-   cmd_buffer->state.pass = pass;
-
-   const VkRect2D *render_area = &pRenderPassBegin->renderArea;
-
-   anv_batch_emit(&cmd_buffer->batch, GEN8_3DSTATE_DRAWING_RECTANGLE,
-                  .ClippedDrawingRectangleYMin = render_area->offset.y,
-                  .ClippedDrawingRectangleXMin = render_area->offset.x,
-                  .ClippedDrawingRectangleYMax =
-                     render_area->offset.y + render_area->extent.height - 1,
-                  .ClippedDrawingRectangleXMax =
-                     render_area->offset.x + render_area->extent.width - 1,
-                  .DrawingRectangleOriginY = 0,
-                  .DrawingRectangleOriginX = 0);
-
-   anv_cmd_buffer_clear_attachments(cmd_buffer, pass,
-                                    pRenderPassBegin->pAttachmentClearValues);
-
-   anv_cmd_buffer_begin_subpass(cmd_buffer, pass->subpasses);
-}
-
-void anv_CmdNextSubpass(
-    VkCmdBuffer                                 cmdBuffer,
-    VkRenderPassContents                        contents)
-{
-   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
-
-   assert(cmd_buffer->level == VK_CMD_BUFFER_LEVEL_PRIMARY);
-
-   anv_cmd_buffer_begin_subpass(cmd_buffer, cmd_buffer->state.subpass + 1);
-}
-
-void anv_CmdEndRenderPass(
-    VkCmdBuffer                                 cmdBuffer)
-{
-   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, cmdBuffer);
-
-   /* Emit a flushing pipe control at the end of a pass.  This is kind of a
-    * hack but it ensures that render targets always actually get written.
-    * Eventually, we should do flushing based on image format transitions
-    * or something of that nature.
-    */
-   anv_batch_emit(&cmd_buffer->batch, GEN8_PIPE_CONTROL,
-                  .PostSyncOperation = NoWrite,
-                  .RenderTargetCacheFlushEnable = true,
-                  .InstructionCacheInvalidateEnable = true,
-                  .DepthCacheFlushEnable = true,
-                  .VFCacheInvalidationEnable = true,
-                  .TextureCacheInvalidationEnable = true,
-                  .CommandStreamerStallEnable = true);
-}
-
-void anv_CmdExecuteCommands(
-    VkCmdBuffer                                 cmdBuffer,
-    uint32_t                                    cmdBuffersCount,
-    const VkCmdBuffer*                          pCmdBuffers)
-{
-   ANV_FROM_HANDLE(anv_cmd_buffer, primary, cmdBuffer);
-
-   assert(primary->level == VK_CMD_BUFFER_LEVEL_PRIMARY);
-
-   anv_assert(primary->state.subpass == &primary->state.pass->subpasses[0]);
-
-   for (uint32_t i = 0; i < cmdBuffersCount; i++) {
-      ANV_FROM_HANDLE(anv_cmd_buffer, secondary, pCmdBuffers[i]);
-
-      assert(secondary->level == VK_CMD_BUFFER_LEVEL_SECONDARY);
-
-      anv_cmd_buffer_add_secondary(primary, secondary);
-   }
-}
-
-VkResult anv_CreateCommandPool(
-    VkDevice                                    _device,
-    const VkCmdPoolCreateInfo*                  pCreateInfo,
-    VkCmdPool*                                  pCmdPool)
-{
-   ANV_FROM_HANDLE(anv_device, device, _device);
-   struct anv_cmd_pool *pool;
-
-   pool = anv_device_alloc(device, sizeof(*pool), 8,
-                           VK_SYSTEM_ALLOC_TYPE_API_OBJECT);
-   if (pool == NULL)
-      return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY);
-
-   list_inithead(&pool->cmd_buffers);
-
-   *pCmdPool = anv_cmd_pool_to_handle(pool);
-
-   return VK_SUCCESS;
-}
-
-VkResult anv_DestroyCommandPool(
-    VkDevice                                    _device,
-    VkCmdPool                                   cmdPool)
-{
-   ANV_FROM_HANDLE(anv_device, device, _device);
-   ANV_FROM_HANDLE(anv_cmd_pool, pool, cmdPool);
-
-   anv_ResetCommandPool(_device, cmdPool, 0);
-
-   anv_device_free(device, pool);
-
-   return VK_SUCCESS;
-}
-
-VkResult anv_ResetCommandPool(
-    VkDevice                                    device,
-    VkCmdPool                                   cmdPool,
-    VkCmdPoolResetFlags                         flags)
-{
-   ANV_FROM_HANDLE(anv_cmd_pool, pool, cmdPool);
-
-   list_for_each_entry_safe(struct anv_cmd_buffer, cmd_buffer,
-                            &pool->cmd_buffers, pool_link) {
-      anv_DestroyCommandBuffer(device, anv_cmd_buffer_to_handle(cmd_buffer));
-   }
-
-   return VK_SUCCESS;
-}