anv/cmd_buffer: Rework surface relocation helpers

[mesa.git] / src / intel / vulkan / genX_cmd_buffer.c

/*
 * 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 "anv_private.h"
#include "vk_format_info.h"
#include "vk_util.h"

#include "common/gen_l3_config.h"
#include "genxml/gen_macros.h"
#include "genxml/genX_pack.h"

static void
emit_lrm(struct anv_batch *batch,
         uint32_t reg, struct anv_bo *bo, uint32_t offset)
{
   anv_batch_emit(batch, GENX(MI_LOAD_REGISTER_MEM), lrm) {
      lrm.RegisterAddress  = reg;
      lrm.MemoryAddress    = (struct anv_address) { bo, offset };
   }
}

static void
emit_lri(struct anv_batch *batch, uint32_t reg, uint32_t imm)
{
   anv_batch_emit(batch, GENX(MI_LOAD_REGISTER_IMM), lri) {
      lri.RegisterOffset   = reg;
      lri.DataDWord        = imm;
   }
}

#if GEN_IS_HASWELL || GEN_GEN >= 8
static void
emit_lrr(struct anv_batch *batch, uint32_t dst, uint32_t src)
{
   anv_batch_emit(batch, GENX(MI_LOAD_REGISTER_REG), lrr) {
      lrr.SourceRegisterAddress        = src;
      lrr.DestinationRegisterAddress   = dst;
   }
}
#endif

void
genX(cmd_buffer_emit_state_base_address)(struct anv_cmd_buffer *cmd_buffer)
{
   struct anv_device *device = cmd_buffer->device;

   /* Emit a render target cache flush.
    *
    * This isn't documented anywhere in the PRM.  However, it seems to be
    * necessary prior to changing the surface state base adress.  Without
    * this, we get GPU hangs when using multi-level command buffers which
    * clear depth, reset state base address, and then go render stuff.
    */
   anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
      pc.DCFlushEnable = true;
      pc.RenderTargetCacheFlushEnable = true;
      pc.CommandStreamerStallEnable = true;
   }

   anv_batch_emit(&cmd_buffer->batch, GENX(STATE_BASE_ADDRESS), sba) {
      sba.GeneralStateBaseAddress = (struct anv_address) { NULL, 0 };
      sba.GeneralStateMemoryObjectControlState = GENX(MOCS);
      sba.GeneralStateBaseAddressModifyEnable = true;

      sba.SurfaceStateBaseAddress =
         anv_cmd_buffer_surface_base_address(cmd_buffer);
      sba.SurfaceStateMemoryObjectControlState = GENX(MOCS);
      sba.SurfaceStateBaseAddressModifyEnable = true;

      sba.DynamicStateBaseAddress =
         (struct anv_address) { &device->dynamic_state_pool.block_pool.bo, 0 };
      sba.DynamicStateMemoryObjectControlState = GENX(MOCS);
      sba.DynamicStateBaseAddressModifyEnable = true;

      sba.IndirectObjectBaseAddress = (struct anv_address) { NULL, 0 };
      sba.IndirectObjectMemoryObjectControlState = GENX(MOCS);
      sba.IndirectObjectBaseAddressModifyEnable = true;

      sba.InstructionBaseAddress =
         (struct anv_address) { &device->instruction_state_pool.block_pool.bo, 0 };
      sba.InstructionMemoryObjectControlState = GENX(MOCS);
      sba.InstructionBaseAddressModifyEnable = true;

#  if (GEN_GEN >= 8)
      /* Broadwell requires that we specify a buffer size for a bunch of
       * these fields.  However, since we will be growing the BO's live, we
       * just set them all to the maximum.
       */
      sba.GeneralStateBufferSize                = 0xfffff;
      sba.GeneralStateBufferSizeModifyEnable    = true;
      sba.DynamicStateBufferSize                = 0xfffff;
      sba.DynamicStateBufferSizeModifyEnable    = true;
      sba.IndirectObjectBufferSize              = 0xfffff;
      sba.IndirectObjectBufferSizeModifyEnable  = true;
      sba.InstructionBufferSize                 = 0xfffff;
      sba.InstructionBuffersizeModifyEnable     = true;
#  endif
   }

   /* 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, GENX(PIPE_CONTROL), pc) {
      pc.TextureCacheInvalidationEnable = true;
      pc.ConstantCacheInvalidationEnable = true;
      pc.StateCacheInvalidationEnable = true;
   }
}

static void
add_surface_reloc(struct anv_cmd_buffer *cmd_buffer,
                  struct anv_state state, struct anv_address addr)
{
   const struct isl_device *isl_dev = &cmd_buffer->device->isl_dev;

   VkResult result =
      anv_reloc_list_add(&cmd_buffer->surface_relocs, &cmd_buffer->pool->alloc,
                         state.offset + isl_dev->ss.addr_offset,
                         addr.bo, addr.offset);
   if (result != VK_SUCCESS)
      anv_batch_set_error(&cmd_buffer->batch, result);
}

static void
add_surface_state_relocs(struct anv_cmd_buffer *cmd_buffer,
                         struct anv_surface_state state)
{
   const struct isl_device *isl_dev = &cmd_buffer->device->isl_dev;

   assert(!anv_address_is_null(state.address));
   add_surface_reloc(cmd_buffer, state.state, state.address);

   if (!anv_address_is_null(state.aux_address)) {
      VkResult result =
         anv_reloc_list_add(&cmd_buffer->surface_relocs,
                            &cmd_buffer->pool->alloc,
                            state.state.offset + isl_dev->ss.aux_addr_offset,
                            state.aux_address.bo, state.aux_address.offset);
      if (result != VK_SUCCESS)
         anv_batch_set_error(&cmd_buffer->batch, result);
   }

   if (!anv_address_is_null(state.clear_address)) {
      VkResult result =
         anv_reloc_list_add(&cmd_buffer->surface_relocs,
                            &cmd_buffer->pool->alloc,
                            state.state.offset +
                            isl_dev->ss.clear_color_state_offset,
                            state.clear_address.bo, state.clear_address.offset);
      if (result != VK_SUCCESS)
         anv_batch_set_error(&cmd_buffer->batch, result);
   }
}

static void
color_attachment_compute_aux_usage(struct anv_device * device,
                                   struct anv_cmd_state * cmd_state,
                                   uint32_t att, VkRect2D render_area,
                                   union isl_color_value *fast_clear_color)
{
   struct anv_attachment_state *att_state = &cmd_state->attachments[att];
   struct anv_image_view *iview = cmd_state->framebuffer->attachments[att];

   assert(iview->n_planes == 1);

   if (iview->planes[0].isl.base_array_layer >=
       anv_image_aux_layers(iview->image, VK_IMAGE_ASPECT_COLOR_BIT,
                            iview->planes[0].isl.base_level)) {
      /* There is no aux buffer which corresponds to the level and layer(s)
       * being accessed.
       */
      att_state->aux_usage = ISL_AUX_USAGE_NONE;
      att_state->input_aux_usage = ISL_AUX_USAGE_NONE;
      att_state->fast_clear = false;
      return;
   }

   att_state->aux_usage =
      anv_layout_to_aux_usage(&device->info, iview->image,
                              VK_IMAGE_ASPECT_COLOR_BIT,
                              VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL);

   /* If we don't have aux, then we should have returned early in the layer
    * check above.  If we got here, we must have something.
    */
   assert(att_state->aux_usage != ISL_AUX_USAGE_NONE);

   if (att_state->aux_usage == ISL_AUX_USAGE_CCS_E ||
       att_state->aux_usage == ISL_AUX_USAGE_MCS) {
      att_state->input_aux_usage = att_state->aux_usage;
   } else {
      /* From the Sky Lake PRM, RENDER_SURFACE_STATE::AuxiliarySurfaceMode:
       *
       *    "If Number of Multisamples is MULTISAMPLECOUNT_1, AUX_CCS_D
       *    setting is only allowed if Surface Format supported for Fast
       *    Clear. In addition, if the surface is bound to the sampling
       *    engine, Surface Format must be supported for Render Target
       *    Compression for surfaces bound to the sampling engine."
       *
       * In other words, we can only sample from a fast-cleared image if it
       * also supports color compression.
       */
      if (isl_format_supports_ccs_e(&device->info, iview->planes[0].isl.format)) {
         att_state->input_aux_usage = ISL_AUX_USAGE_CCS_D;

         /* While fast-clear resolves and partial resolves are fairly cheap in the
          * case where you render to most of the pixels, full resolves are not
          * because they potentially involve reading and writing the entire
          * framebuffer.  If we can't texture with CCS_E, we should leave it off and
          * limit ourselves to fast clears.
          */
         if (cmd_state->pass->attachments[att].first_subpass_layout ==
             VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL) {
            anv_perf_warn(device->instance, iview->image,
                          "Not temporarily enabling CCS_E.");
         }
      } else {
         att_state->input_aux_usage = ISL_AUX_USAGE_NONE;
      }
   }

   assert(iview->image->planes[0].aux_surface.isl.usage &
          (ISL_SURF_USAGE_CCS_BIT | ISL_SURF_USAGE_MCS_BIT));

   union isl_color_value clear_color = {};
   anv_clear_color_from_att_state(&clear_color, att_state, iview);

   att_state->clear_color_is_zero_one =
      isl_color_value_is_zero_one(clear_color, iview->planes[0].isl.format);
   att_state->clear_color_is_zero =
      isl_color_value_is_zero(clear_color, iview->planes[0].isl.format);

   if (att_state->pending_clear_aspects == VK_IMAGE_ASPECT_COLOR_BIT) {
      /* Start by getting the fast clear type.  We use the first subpass
       * layout here because we don't want to fast-clear if the first subpass
       * to use the attachment can't handle fast-clears.
       */
      enum anv_fast_clear_type fast_clear_type =
         anv_layout_to_fast_clear_type(&device->info, iview->image,
                                       VK_IMAGE_ASPECT_COLOR_BIT,
                                       cmd_state->pass->attachments[att].first_subpass_layout);
      switch (fast_clear_type) {
      case ANV_FAST_CLEAR_NONE:
         att_state->fast_clear = false;
         break;
      case ANV_FAST_CLEAR_DEFAULT_VALUE:
         att_state->fast_clear = att_state->clear_color_is_zero;
         break;
      case ANV_FAST_CLEAR_ANY:
         att_state->fast_clear = true;
         break;
      }

      /* Potentially, we could do partial fast-clears but doing so has crazy
       * alignment restrictions.  It's easier to just restrict to full size
       * fast clears for now.
       */
      if (render_area.offset.x != 0 ||
          render_area.offset.y != 0 ||
          render_area.extent.width != iview->extent.width ||
          render_area.extent.height != iview->extent.height)
         att_state->fast_clear = false;

      /* On Broadwell and earlier, we can only handle 0/1 clear colors */
      if (GEN_GEN <= 8 && !att_state->clear_color_is_zero_one)
         att_state->fast_clear = false;

      /* We only allow fast clears to the first slice of an image (level 0,
       * layer 0) and only for the entire slice.  This guarantees us that, at
       * any given time, there is only one clear color on any given image at
       * any given time.  At the time of our testing (Jan 17, 2018), there
       * were no known applications which would benefit from fast-clearing
       * more than just the first slice.
       */
      if (att_state->fast_clear &&
          (iview->planes[0].isl.base_level > 0 ||
           iview->planes[0].isl.base_array_layer > 0)) {
         anv_perf_warn(device->instance, iview->image,
                       "Rendering with multi-lod or multi-layer framebuffer "
                       "with LOAD_OP_LOAD and baseMipLevel > 0 or "
                       "baseArrayLayer > 0.  Not fast clearing.");
         att_state->fast_clear = false;
      } else if (att_state->fast_clear && cmd_state->framebuffer->layers > 1) {
         anv_perf_warn(device->instance, iview->image,
                       "Rendering to a multi-layer framebuffer with "
                       "LOAD_OP_CLEAR.  Only fast-clearing the first slice");
      }

      if (att_state->fast_clear)
         *fast_clear_color = clear_color;
   } else {
      att_state->fast_clear = false;
   }
}

static void
depth_stencil_attachment_compute_aux_usage(struct anv_device *device,
                                           struct anv_cmd_state *cmd_state,
                                           uint32_t att, VkRect2D render_area)
{
   struct anv_render_pass_attachment *pass_att =
      &cmd_state->pass->attachments[att];
   struct anv_attachment_state *att_state = &cmd_state->attachments[att];
   struct anv_image_view *iview = cmd_state->framebuffer->attachments[att];

   /* These will be initialized after the first subpass transition. */
   att_state->aux_usage = ISL_AUX_USAGE_NONE;
   att_state->input_aux_usage = ISL_AUX_USAGE_NONE;

   if (GEN_GEN == 7) {
      /* We don't do any HiZ or depth fast-clears on gen7 yet */
      att_state->fast_clear = false;
      return;
   }

   if (!(att_state->pending_clear_aspects & VK_IMAGE_ASPECT_DEPTH_BIT)) {
      /* If we're just clearing stencil, we can always HiZ clear */
      att_state->fast_clear = true;
      return;
   }

   /* Default to false for now */
   att_state->fast_clear = false;

   /* We must have depth in order to have HiZ */
   if (!(iview->image->aspects & VK_IMAGE_ASPECT_DEPTH_BIT))
      return;

   const enum isl_aux_usage first_subpass_aux_usage =
      anv_layout_to_aux_usage(&device->info, iview->image,
                              VK_IMAGE_ASPECT_DEPTH_BIT,
                              pass_att->first_subpass_layout);
   if (first_subpass_aux_usage != ISL_AUX_USAGE_HIZ)
      return;

   if (!blorp_can_hiz_clear_depth(GEN_GEN,
                                  iview->planes[0].isl.format,
                                  iview->image->samples,
                                  render_area.offset.x,
                                  render_area.offset.y,
                                  render_area.offset.x +
                                  render_area.extent.width,
                                  render_area.offset.y +
                                  render_area.extent.height))
      return;

   if (att_state->clear_value.depthStencil.depth != ANV_HZ_FC_VAL)
      return;

   if (GEN_GEN == 8 && anv_can_sample_with_hiz(&device->info, iview->image)) {
      /* Only gen9+ supports returning ANV_HZ_FC_VAL when sampling a
       * fast-cleared portion of a HiZ buffer. Testing has revealed that Gen8
       * only supports returning 0.0f. Gens prior to gen8 do not support this
       * feature at all.
       */
      return;
   }

   /* If we got here, then we can fast clear */
   att_state->fast_clear = true;
}

static bool
need_input_attachment_state(const struct anv_render_pass_attachment *att)
{
   if (!(att->usage & VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT))
      return false;

   /* We only allocate input attachment states for color surfaces. Compression
    * is not yet enabled for depth textures and stencil doesn't allow
    * compression so we can just use the texture surface state from the view.
    */
   return vk_format_is_color(att->format);
}

/* Transitions a HiZ-enabled depth buffer from one layout to another. Unless
 * the initial layout is undefined, the HiZ buffer and depth buffer will
 * represent the same data at the end of this operation.
 */
static void
transition_depth_buffer(struct anv_cmd_buffer *cmd_buffer,
                        const struct anv_image *image,
                        VkImageLayout initial_layout,
                        VkImageLayout final_layout)
{
   const bool hiz_enabled = ISL_AUX_USAGE_HIZ ==
      anv_layout_to_aux_usage(&cmd_buffer->device->info, image,
                              VK_IMAGE_ASPECT_DEPTH_BIT, initial_layout);
   const bool enable_hiz = ISL_AUX_USAGE_HIZ ==
      anv_layout_to_aux_usage(&cmd_buffer->device->info, image,
                              VK_IMAGE_ASPECT_DEPTH_BIT, final_layout);

   enum isl_aux_op hiz_op;
   if (hiz_enabled && !enable_hiz) {
      hiz_op = ISL_AUX_OP_FULL_RESOLVE;
   } else if (!hiz_enabled && enable_hiz) {
      hiz_op = ISL_AUX_OP_AMBIGUATE;
   } else {
      assert(hiz_enabled == enable_hiz);
      /* If the same buffer will be used, no resolves are necessary. */
      hiz_op = ISL_AUX_OP_NONE;
   }

   if (hiz_op != ISL_AUX_OP_NONE)
      anv_image_hiz_op(cmd_buffer, image, VK_IMAGE_ASPECT_DEPTH_BIT,
                       0, 0, 1, hiz_op);
}

#define MI_PREDICATE_SRC0  0x2400
#define MI_PREDICATE_SRC1  0x2408

static void
set_image_compressed_bit(struct anv_cmd_buffer *cmd_buffer,
                         const struct anv_image *image,
                         VkImageAspectFlagBits aspect,
                         uint32_t level,
                         uint32_t base_layer, uint32_t layer_count,
                         bool compressed)
{
   uint32_t plane = anv_image_aspect_to_plane(image->aspects, aspect);

   /* We only have compression tracking for CCS_E */
   if (image->planes[plane].aux_usage != ISL_AUX_USAGE_CCS_E)
      return;

   for (uint32_t a = 0; a < layer_count; a++) {
      uint32_t layer = base_layer + a;
      anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_DATA_IMM), sdi) {
         sdi.Address = anv_image_get_compression_state_addr(cmd_buffer->device,
                                                            image, aspect,
                                                            level, layer);
         sdi.ImmediateData = compressed ? UINT32_MAX : 0;
      }
   }
}

static void
set_image_fast_clear_state(struct anv_cmd_buffer *cmd_buffer,
                           const struct anv_image *image,
                           VkImageAspectFlagBits aspect,
                           enum anv_fast_clear_type fast_clear)
{
   anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_DATA_IMM), sdi) {
      sdi.Address = anv_image_get_fast_clear_type_addr(cmd_buffer->device,
                                                       image, aspect);
      sdi.ImmediateData = fast_clear;
   }

   /* Whenever we have fast-clear, we consider that slice to be compressed.
    * This makes building predicates much easier.
    */
   if (fast_clear != ANV_FAST_CLEAR_NONE)
      set_image_compressed_bit(cmd_buffer, image, aspect, 0, 0, 1, true);
}

#if GEN_IS_HASWELL || GEN_GEN >= 8
static inline uint32_t
mi_alu(uint32_t opcode, uint32_t operand1, uint32_t operand2)
{
   struct GENX(MI_MATH_ALU_INSTRUCTION) instr = {
      .ALUOpcode = opcode,
      .Operand1 = operand1,
      .Operand2 = operand2,
   };

   uint32_t dw;
   GENX(MI_MATH_ALU_INSTRUCTION_pack)(NULL, &dw, &instr);

   return dw;
}
#endif

#define CS_GPR(n) (0x2600 + (n) * 8)

/* This is only really practical on haswell and above because it requires
 * MI math in order to get it correct.
 */
#if GEN_GEN >= 8 || GEN_IS_HASWELL
static void
anv_cmd_compute_resolve_predicate(struct anv_cmd_buffer *cmd_buffer,
                                  const struct anv_image *image,
                                  VkImageAspectFlagBits aspect,
                                  uint32_t level, uint32_t array_layer,
                                  enum isl_aux_op resolve_op,
                                  enum anv_fast_clear_type fast_clear_supported)
{
   struct anv_address fast_clear_type_addr =
      anv_image_get_fast_clear_type_addr(cmd_buffer->device, image, aspect);

   /* Name some registers */
   const int image_fc_reg = MI_ALU_REG0;
   const int fc_imm_reg = MI_ALU_REG1;
   const int pred_reg = MI_ALU_REG2;

   uint32_t *dw;

   if (resolve_op == ISL_AUX_OP_FULL_RESOLVE) {
      /* In this case, we're doing a full resolve which means we want the
       * resolve to happen if any compression (including fast-clears) is
       * present.
       *
       * In order to simplify the logic a bit, we make the assumption that,
       * if the first slice has been fast-cleared, it is also marked as
       * compressed.  See also set_image_fast_clear_state.
       */
      struct anv_address compression_state_addr =
         anv_image_get_compression_state_addr(cmd_buffer->device, image,
                                              aspect, level, array_layer);
      anv_batch_emit(&cmd_buffer->batch, GENX(MI_LOAD_REGISTER_MEM), lrm) {
         lrm.RegisterAddress  = MI_PREDICATE_SRC0;
         lrm.MemoryAddress    = compression_state_addr;
      }
      anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_DATA_IMM), sdi) {
         sdi.Address       = compression_state_addr;
         sdi.ImmediateData = 0;
      }

      if (level == 0 && array_layer == 0) {
         /* If the predicate is true, we want to write 0 to the fast clear type
          * and, if it's false, leave it alone.  We can do this by writing
          *
          * clear_type = clear_type & ~predicate;
          */
         anv_batch_emit(&cmd_buffer->batch, GENX(MI_LOAD_REGISTER_MEM), lrm) {
            lrm.RegisterAddress  = CS_GPR(image_fc_reg);
            lrm.MemoryAddress    = fast_clear_type_addr;
         }
         anv_batch_emit(&cmd_buffer->batch, GENX(MI_LOAD_REGISTER_REG), lrr) {
            lrr.DestinationRegisterAddress   = CS_GPR(pred_reg);
            lrr.SourceRegisterAddress        = MI_PREDICATE_SRC0;
         }

         dw = anv_batch_emitn(&cmd_buffer->batch, 5, GENX(MI_MATH));
         dw[1] = mi_alu(MI_ALU_LOAD, MI_ALU_SRCA, image_fc_reg);
         dw[2] = mi_alu(MI_ALU_LOADINV, MI_ALU_SRCB, pred_reg);
         dw[3] = mi_alu(MI_ALU_AND, 0, 0);
         dw[4] = mi_alu(MI_ALU_STORE, image_fc_reg, MI_ALU_ACCU);

         anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_REGISTER_MEM), srm) {
            srm.MemoryAddress    = fast_clear_type_addr;
            srm.RegisterAddress  = CS_GPR(image_fc_reg);
         }
      }
   } else if (level == 0 && array_layer == 0) {
      /* In this case, we are doing a partial resolve to get rid of fast-clear
       * colors.  We don't care about the compression state but we do care
       * about how much fast clear is allowed by the final layout.
       */
      assert(resolve_op == ISL_AUX_OP_PARTIAL_RESOLVE);
      assert(fast_clear_supported < ANV_FAST_CLEAR_ANY);

      anv_batch_emit(&cmd_buffer->batch, GENX(MI_LOAD_REGISTER_MEM), lrm) {
         lrm.RegisterAddress  = CS_GPR(image_fc_reg);
         lrm.MemoryAddress    = fast_clear_type_addr;
      }
      emit_lri(&cmd_buffer->batch, CS_GPR(image_fc_reg) + 4, 0);

      emit_lri(&cmd_buffer->batch, CS_GPR(fc_imm_reg), fast_clear_supported);
      emit_lri(&cmd_buffer->batch, CS_GPR(fc_imm_reg) + 4, 0);

      /* We need to compute (fast_clear_supported < image->fast_clear).
       * We do this by subtracting and storing the carry bit.
       */
      dw = anv_batch_emitn(&cmd_buffer->batch, 5, GENX(MI_MATH));
      dw[1] = mi_alu(MI_ALU_LOAD, MI_ALU_SRCA, fc_imm_reg);
      dw[2] = mi_alu(MI_ALU_LOAD, MI_ALU_SRCB, image_fc_reg);
      dw[3] = mi_alu(MI_ALU_SUB, 0, 0);
      dw[4] = mi_alu(MI_ALU_STORE, pred_reg, MI_ALU_CF);

      /* Store the predicate */
      emit_lrr(&cmd_buffer->batch, MI_PREDICATE_SRC0, CS_GPR(pred_reg));

      /* If the predicate is true, we want to write 0 to the fast clear type
       * and, if it's false, leave it alone.  We can do this by writing
       *
       * clear_type = clear_type & ~predicate;
       */
      dw = anv_batch_emitn(&cmd_buffer->batch, 5, GENX(MI_MATH));
      dw[1] = mi_alu(MI_ALU_LOAD, MI_ALU_SRCA, image_fc_reg);
      dw[2] = mi_alu(MI_ALU_LOADINV, MI_ALU_SRCB, pred_reg);
      dw[3] = mi_alu(MI_ALU_AND, 0, 0);
      dw[4] = mi_alu(MI_ALU_STORE, image_fc_reg, MI_ALU_ACCU);

      anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_REGISTER_MEM), srm) {
         srm.RegisterAddress  = CS_GPR(image_fc_reg);
         srm.MemoryAddress    = fast_clear_type_addr;
      }
   } else {
      /* In this case, we're trying to do a partial resolve on a slice that
       * doesn't have clear color.  There's nothing to do.
       */
      assert(resolve_op == ISL_AUX_OP_PARTIAL_RESOLVE);
      return;
   }

   /* We use the first half of src0 for the actual predicate.  Set the second
    * half of src0 and all of src1 to 0 as the predicate operation will be
    * doing an implicit src0 != src1.
    */
   emit_lri(&cmd_buffer->batch, MI_PREDICATE_SRC0 + 4, 0);
   emit_lri(&cmd_buffer->batch, MI_PREDICATE_SRC1    , 0);
   emit_lri(&cmd_buffer->batch, MI_PREDICATE_SRC1 + 4, 0);

   anv_batch_emit(&cmd_buffer->batch, GENX(MI_PREDICATE), mip) {
      mip.LoadOperation    = LOAD_LOADINV;
      mip.CombineOperation = COMBINE_SET;
      mip.CompareOperation = COMPARE_SRCS_EQUAL;
   }
}
#endif /* GEN_GEN >= 8 || GEN_IS_HASWELL */

#if GEN_GEN <= 8
static void
anv_cmd_simple_resolve_predicate(struct anv_cmd_buffer *cmd_buffer,
                                 const struct anv_image *image,
                                 VkImageAspectFlagBits aspect,
                                 uint32_t level, uint32_t array_layer,
                                 enum isl_aux_op resolve_op,
                                 enum anv_fast_clear_type fast_clear_supported)
{
   struct anv_address fast_clear_type_addr =
      anv_image_get_fast_clear_type_addr(cmd_buffer->device, image, aspect);

   /* This only works for partial resolves and only when the clear color is
    * all or nothing.  On the upside, this emits less command streamer code
    * and works on Ivybridge and Bay Trail.
    */
   assert(resolve_op == ISL_AUX_OP_PARTIAL_RESOLVE);
   assert(fast_clear_supported != ANV_FAST_CLEAR_ANY);

   /* We don't support fast clears on anything other than the first slice. */
   if (level > 0 || array_layer > 0)
      return;

   /* On gen8, we don't have a concept of default clear colors because we
    * can't sample from CCS surfaces.  It's enough to just load the fast clear
    * state into the predicate register.
    */
   anv_batch_emit(&cmd_buffer->batch, GENX(MI_LOAD_REGISTER_MEM), lrm) {
      lrm.RegisterAddress  = MI_PREDICATE_SRC0;
      lrm.MemoryAddress    = fast_clear_type_addr;
   }
   anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_DATA_IMM), sdi) {
      sdi.Address          = fast_clear_type_addr;
      sdi.ImmediateData    = 0;
   }

   /* We use the first half of src0 for the actual predicate.  Set the second
    * half of src0 and all of src1 to 0 as the predicate operation will be
    * doing an implicit src0 != src1.
    */
   emit_lri(&cmd_buffer->batch, MI_PREDICATE_SRC0 + 4, 0);
   emit_lri(&cmd_buffer->batch, MI_PREDICATE_SRC1    , 0);
   emit_lri(&cmd_buffer->batch, MI_PREDICATE_SRC1 + 4, 0);

   anv_batch_emit(&cmd_buffer->batch, GENX(MI_PREDICATE), mip) {
      mip.LoadOperation    = LOAD_LOADINV;
      mip.CombineOperation = COMBINE_SET;
      mip.CompareOperation = COMPARE_SRCS_EQUAL;
   }
}
#endif /* GEN_GEN <= 8 */

static void
anv_cmd_predicated_ccs_resolve(struct anv_cmd_buffer *cmd_buffer,
                               const struct anv_image *image,
                               VkImageAspectFlagBits aspect,
                               uint32_t level, uint32_t array_layer,
                               enum isl_aux_op resolve_op,
                               enum anv_fast_clear_type fast_clear_supported)
{
   const uint32_t plane = anv_image_aspect_to_plane(image->aspects, aspect);

#if GEN_GEN >= 9
   anv_cmd_compute_resolve_predicate(cmd_buffer, image,
                                     aspect, level, array_layer,
                                     resolve_op, fast_clear_supported);
#else /* GEN_GEN <= 8 */
   anv_cmd_simple_resolve_predicate(cmd_buffer, image,
                                    aspect, level, array_layer,
                                    resolve_op, fast_clear_supported);
#endif

   /* CCS_D only supports full resolves and BLORP will assert on us if we try
    * to do a partial resolve on a CCS_D surface.
    */
   if (resolve_op == ISL_AUX_OP_PARTIAL_RESOLVE &&
       image->planes[plane].aux_usage == ISL_AUX_USAGE_NONE)
      resolve_op = ISL_AUX_OP_FULL_RESOLVE;

   anv_image_ccs_op(cmd_buffer, image, aspect, level,
                    array_layer, 1, resolve_op, NULL, true);
}

static void
anv_cmd_predicated_mcs_resolve(struct anv_cmd_buffer *cmd_buffer,
                               const struct anv_image *image,
                               VkImageAspectFlagBits aspect,
                               uint32_t array_layer,
                               enum isl_aux_op resolve_op,
                               enum anv_fast_clear_type fast_clear_supported)
{
   assert(aspect == VK_IMAGE_ASPECT_COLOR_BIT);
   assert(resolve_op == ISL_AUX_OP_PARTIAL_RESOLVE);

#if GEN_GEN >= 8 || GEN_IS_HASWELL
   anv_cmd_compute_resolve_predicate(cmd_buffer, image,
                                     aspect, 0, array_layer,
                                     resolve_op, fast_clear_supported);

   anv_image_mcs_op(cmd_buffer, image, aspect,
                    array_layer, 1, resolve_op, NULL, true);
#else
   unreachable("MCS resolves are unsupported on Ivybridge and Bay Trail");
#endif
}

void
genX(cmd_buffer_mark_image_written)(struct anv_cmd_buffer *cmd_buffer,
                                    const struct anv_image *image,
                                    VkImageAspectFlagBits aspect,
                                    enum isl_aux_usage aux_usage,
                                    uint32_t level,
                                    uint32_t base_layer,
                                    uint32_t layer_count)
{
   /* The aspect must be exactly one of the image aspects. */
   assert(_mesa_bitcount(aspect) == 1 && (aspect & image->aspects));

   /* The only compression types with more than just fast-clears are MCS,
    * CCS_E, and HiZ.  With HiZ we just trust the layout and don't actually
    * track the current fast-clear and compression state.  This leaves us
    * with just MCS and CCS_E.
    */
   if (aux_usage != ISL_AUX_USAGE_CCS_E &&
       aux_usage != ISL_AUX_USAGE_MCS)
      return;

   set_image_compressed_bit(cmd_buffer, image, aspect,
                            level, base_layer, layer_count, true);
}

static void
init_fast_clear_color(struct anv_cmd_buffer *cmd_buffer,
                      const struct anv_image *image,
                      VkImageAspectFlagBits aspect)
{
   assert(cmd_buffer && image);
   assert(image->aspects & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV);

   set_image_fast_clear_state(cmd_buffer, image, aspect,
                              ANV_FAST_CLEAR_NONE);

   /* The fast clear value dword(s) will be copied into a surface state object.
    * Ensure that the restrictions of the fields in the dword(s) are followed.
    *
    * CCS buffers on SKL+ can have any value set for the clear colors.
    */
   if (image->samples == 1 && GEN_GEN >= 9)
      return;

   /* Other combinations of auxiliary buffers and platforms require specific
    * values in the clear value dword(s).
    */
   struct anv_address addr =
      anv_image_get_clear_color_addr(cmd_buffer->device, image, aspect);

   if (GEN_GEN >= 9) {
      for (unsigned i = 0; i < 4; i++) {
         anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_DATA_IMM), sdi) {
            sdi.Address = addr;
            sdi.Address.offset += i * 4;
            /* MCS buffers on SKL+ can only have 1/0 clear colors. */
            assert(image->samples > 1);
            sdi.ImmediateData = 0;
         }
      }
   } else {
      anv_batch_emit(&cmd_buffer->batch, GENX(MI_STORE_DATA_IMM), sdi) {
         sdi.Address = addr;
         if (GEN_GEN >= 8 || GEN_IS_HASWELL) {
            /* Pre-SKL, the dword containing the clear values also contains
             * other fields, so we need to initialize those fields to match the
             * values that would be in a color attachment.
             */
            sdi.ImmediateData = ISL_CHANNEL_SELECT_RED   << 25 |
                                ISL_CHANNEL_SELECT_GREEN << 22 |
                                ISL_CHANNEL_SELECT_BLUE  << 19 |
                                ISL_CHANNEL_SELECT_ALPHA << 16;
         } else if (GEN_GEN == 7) {
            /* On IVB, the dword containing the clear values also contains
             * other fields that must be zero or can be zero.
             */
            sdi.ImmediateData = 0;
         }
      }
   }
}

/* Copy the fast-clear value dword(s) between a surface state object and an
 * image's fast clear state buffer.
 */
static void
genX(copy_fast_clear_dwords)(struct anv_cmd_buffer *cmd_buffer,
                             struct anv_state surface_state,
                             const struct anv_image *image,
                             VkImageAspectFlagBits aspect,
                             bool copy_from_surface_state)
{
   assert(cmd_buffer && image);
   assert(image->aspects & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV);

   struct anv_bo *ss_bo =
      &cmd_buffer->device->surface_state_pool.block_pool.bo;
   uint32_t ss_clear_offset = surface_state.offset +
      cmd_buffer->device->isl_dev.ss.clear_value_offset;
   const struct anv_address entry_addr =
      anv_image_get_clear_color_addr(cmd_buffer->device, image, aspect);
   unsigned copy_size = cmd_buffer->device->isl_dev.ss.clear_value_size;

   if (copy_from_surface_state) {
      genX(cmd_buffer_mi_memcpy)(cmd_buffer, entry_addr.bo, entry_addr.offset,
                                 ss_bo, ss_clear_offset, copy_size);
   } else {
      genX(cmd_buffer_mi_memcpy)(cmd_buffer, ss_bo, ss_clear_offset,
                                 entry_addr.bo, entry_addr.offset, copy_size);

      /* Updating a surface state object may require that the state cache be
       * invalidated. From the SKL PRM, Shared Functions -> State -> State
       * Caching:
       *
       *    Whenever the RENDER_SURFACE_STATE object in memory pointed to by
       *    the Binding Table Pointer (BTP) and Binding Table Index (BTI) is
       *    modified [...], the L1 state cache must be invalidated to ensure
       *    the new surface or sampler state is fetched from system memory.
       *
       * In testing, SKL doesn't actually seem to need this, but HSW does.
       */
      cmd_buffer->state.pending_pipe_bits |=
         ANV_PIPE_STATE_CACHE_INVALIDATE_BIT;
   }
}

/**
 * @brief Transitions a color buffer from one layout to another.
 *
 * See section 6.1.1. Image Layout Transitions of the Vulkan 1.0.50 spec for
 * more information.
 *
 * @param level_count VK_REMAINING_MIP_LEVELS isn't supported.
 * @param layer_count VK_REMAINING_ARRAY_LAYERS isn't supported. For 3D images,
 *                    this represents the maximum layers to transition at each
 *                    specified miplevel.
 */
static void
transition_color_buffer(struct anv_cmd_buffer *cmd_buffer,
                        const struct anv_image *image,
                        VkImageAspectFlagBits aspect,
                        const uint32_t base_level, uint32_t level_count,
                        uint32_t base_layer, uint32_t layer_count,
                        VkImageLayout initial_layout,
                        VkImageLayout final_layout)
{
   const struct gen_device_info *devinfo = &cmd_buffer->device->info;
   /* Validate the inputs. */
   assert(cmd_buffer);
   assert(image && image->aspects & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV);
   /* These values aren't supported for simplicity's sake. */
   assert(level_count != VK_REMAINING_MIP_LEVELS &&
          layer_count != VK_REMAINING_ARRAY_LAYERS);
   /* Ensure the subresource range is valid. */
   uint64_t last_level_num = base_level + level_count;
   const uint32_t max_depth = anv_minify(image->extent.depth, base_level);
   UNUSED const uint32_t image_layers = MAX2(image->array_size, max_depth);
   assert((uint64_t)base_layer + layer_count  <= image_layers);
   assert(last_level_num <= image->levels);
   /* The spec disallows these final layouts. */
   assert(final_layout != VK_IMAGE_LAYOUT_UNDEFINED &&
          final_layout != VK_IMAGE_LAYOUT_PREINITIALIZED);

   /* No work is necessary if the layout stays the same or if this subresource
    * range lacks auxiliary data.
    */
   if (initial_layout == final_layout)
      return;

   uint32_t plane = anv_image_aspect_to_plane(image->aspects, aspect);

   if (image->planes[plane].shadow_surface.isl.size > 0 &&
       final_layout == VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL) {
      /* This surface is a linear compressed image with a tiled shadow surface
       * for texturing.  The client is about to use it in READ_ONLY_OPTIMAL so
       * we need to ensure the shadow copy is up-to-date.
       */
      assert(image->aspects == VK_IMAGE_ASPECT_COLOR_BIT);
      assert(image->planes[plane].surface.isl.tiling == ISL_TILING_LINEAR);
      assert(image->planes[plane].shadow_surface.isl.tiling != ISL_TILING_LINEAR);
      assert(isl_format_is_compressed(image->planes[plane].surface.isl.format));
      assert(plane == 0);
      anv_image_copy_to_shadow(cmd_buffer, image,
                               base_level, level_count,
                               base_layer, layer_count);
   }

   if (base_layer >= anv_image_aux_layers(image, aspect, base_level))
      return;

   assert(image->tiling == VK_IMAGE_TILING_OPTIMAL);

   if (initial_layout == VK_IMAGE_LAYOUT_UNDEFINED ||
       initial_layout == VK_IMAGE_LAYOUT_PREINITIALIZED) {
      /* A subresource in the undefined layout may have been aliased and
       * populated with any arrangement of bits. Therefore, we must initialize
       * the related aux buffer and clear buffer entry with desirable values.
       * An initial layout of PREINITIALIZED is the same as UNDEFINED for
       * images with VK_IMAGE_TILING_OPTIMAL.
       *
       * Initialize the relevant clear buffer entries.
       */
      if (base_level == 0 && base_layer == 0)
         init_fast_clear_color(cmd_buffer, image, aspect);

      /* Initialize the aux buffers to enable correct rendering.  In order to
       * ensure that things such as storage images work correctly, aux buffers
       * need to be initialized to valid data.
       *
       * Having an aux buffer with invalid data is a problem for two reasons:
       *
       *  1) Having an invalid value in the buffer can confuse the hardware.
       *     For instance, with CCS_E on SKL, a two-bit CCS value of 2 is
       *     invalid and leads to the hardware doing strange things.  It
       *     doesn't hang as far as we can tell but rendering corruption can
       *     occur.
       *
       *  2) If this transition is into the GENERAL layout and we then use the
       *     image as a storage image, then we must have the aux buffer in the
       *     pass-through state so that, if we then go to texture from the
       *     image, we get the results of our storage image writes and not the
       *     fast clear color or other random data.
       *
       * For CCS both of the problems above are real demonstrable issues.  In
       * that case, the only thing we can do is to perform an ambiguate to
       * transition the aux surface into the pass-through state.
       *
       * For MCS, (2) is never an issue because we don't support multisampled
       * storage images.  In theory, issue (1) is a problem with MCS but we've
       * never seen it in the wild.  For 4x and 16x, all bit patters could, in
       * theory, be interpreted as something but we don't know that all bit
       * patterns are actually valid.  For 2x and 8x, you could easily end up
       * with the MCS referring to an invalid plane because not all bits of
       * the MCS value are actually used.  Even though we've never seen issues
       * in the wild, it's best to play it safe and initialize the MCS.  We
       * can use a fast-clear for MCS because we only ever touch from render
       * and texture (no image load store).
       */
      if (image->samples == 1) {
         for (uint32_t l = 0; l < level_count; l++) {
            const uint32_t level = base_level + l;

            uint32_t aux_layers = anv_image_aux_layers(image, aspect, level);
            if (base_layer >= aux_layers)
               break; /* We will only get fewer layers as level increases */
            uint32_t level_layer_count =
               MIN2(layer_count, aux_layers - base_layer);

            anv_image_ccs_op(cmd_buffer, image, aspect, level,
                             base_layer, level_layer_count,
                             ISL_AUX_OP_AMBIGUATE, NULL, false);

            if (image->planes[plane].aux_usage == ISL_AUX_USAGE_CCS_E) {
               set_image_compressed_bit(cmd_buffer, image, aspect,
                                        level, base_layer, level_layer_count,
                                        false);
            }
         }
      } else {
         if (image->samples == 4 || image->samples == 16) {
            anv_perf_warn(cmd_buffer->device->instance, image,
                          "Doing a potentially unnecessary fast-clear to "
                          "define an MCS buffer.");
         }

         assert(base_level == 0 && level_count == 1);
         anv_image_mcs_op(cmd_buffer, image, aspect,
                          base_layer, layer_count,
                          ISL_AUX_OP_FAST_CLEAR, NULL, false);
      }
      return;
   }

   const enum isl_aux_usage initial_aux_usage =
      anv_layout_to_aux_usage(devinfo, image, aspect, initial_layout);
   const enum isl_aux_usage final_aux_usage =
      anv_layout_to_aux_usage(devinfo, image, aspect, final_layout);

   /* The current code assumes that there is no mixing of CCS_E and CCS_D.
    * We can handle transitions between CCS_D/E to and from NONE.  What we
    * don't yet handle is switching between CCS_E and CCS_D within a given
    * image.  Doing so in a performant way requires more detailed aux state
    * tracking such as what is done in i965.  For now, just assume that we
    * only have one type of compression.
    */
   assert(initial_aux_usage == ISL_AUX_USAGE_NONE ||
          final_aux_usage == ISL_AUX_USAGE_NONE ||
          initial_aux_usage == final_aux_usage);

   /* If initial aux usage is NONE, there is nothing to resolve */
   if (initial_aux_usage == ISL_AUX_USAGE_NONE)
      return;

   enum isl_aux_op resolve_op = ISL_AUX_OP_NONE;

   /* If the initial layout supports more fast clear than the final layout
    * then we need at least a partial resolve.
    */
   const enum anv_fast_clear_type initial_fast_clear =
      anv_layout_to_fast_clear_type(devinfo, image, aspect, initial_layout);
   const enum anv_fast_clear_type final_fast_clear =
      anv_layout_to_fast_clear_type(devinfo, image, aspect, final_layout);
   if (final_fast_clear < initial_fast_clear)
      resolve_op = ISL_AUX_OP_PARTIAL_RESOLVE;

   if (initial_aux_usage == ISL_AUX_USAGE_CCS_E &&
       final_aux_usage != ISL_AUX_USAGE_CCS_E)
      resolve_op = ISL_AUX_OP_FULL_RESOLVE;

   if (resolve_op == ISL_AUX_OP_NONE)
      return;

   /* Perform a resolve to synchronize data between the main and aux buffer.
    * Before we begin, we must satisfy the cache flushing requirement specified
    * in the Sky Lake PRM Vol. 7, "MCS Buffer for Render Target(s)":
    *
    *    Any transition from any value in {Clear, Render, Resolve} to a
    *    different value in {Clear, Render, Resolve} requires end of pipe
    *    synchronization.
    *
    * We perform a flush of the write cache before and after the clear and
    * resolve operations to meet this requirement.
    *
    * Unlike other drawing, fast clear operations are not properly
    * synchronized. The first PIPE_CONTROL here likely ensures that the
    * contents of the previous render or clear hit the render target before we
    * resolve and the second likely ensures that the resolve is complete before
    * we do any more rendering or clearing.
    */
   cmd_buffer->state.pending_pipe_bits |=
      ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT | ANV_PIPE_CS_STALL_BIT;

   for (uint32_t l = 0; l < level_count; l++) {
      uint32_t level = base_level + l;

      uint32_t aux_layers = anv_image_aux_layers(image, aspect, level);
      if (base_layer >= aux_layers)
         break; /* We will only get fewer layers as level increases */
      uint32_t level_layer_count =
         MIN2(layer_count, aux_layers - base_layer);

      for (uint32_t a = 0; a < level_layer_count; a++) {
         uint32_t array_layer = base_layer + a;
         if (image->samples == 1) {
            anv_cmd_predicated_ccs_resolve(cmd_buffer, image, aspect,
                                           level, array_layer, resolve_op,
                                           final_fast_clear);
         } else {
            anv_cmd_predicated_mcs_resolve(cmd_buffer, image, aspect,
                                           array_layer, resolve_op,
                                           final_fast_clear);
         }
      }
   }

   cmd_buffer->state.pending_pipe_bits |=
      ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT | ANV_PIPE_CS_STALL_BIT;
}

/**
 * Setup anv_cmd_state::attachments for vkCmdBeginRenderPass.
 */
static VkResult
genX(cmd_buffer_setup_attachments)(struct anv_cmd_buffer *cmd_buffer,
                                   struct anv_render_pass *pass,
                                   const VkRenderPassBeginInfo *begin)
{
   const struct isl_device *isl_dev = &cmd_buffer->device->isl_dev;
   struct anv_cmd_state *state = &cmd_buffer->state;

   vk_free(&cmd_buffer->pool->alloc, state->attachments);

   if (pass->attachment_count > 0) {
      state->attachments = vk_alloc(&cmd_buffer->pool->alloc,
                                    pass->attachment_count *
                                         sizeof(state->attachments[0]),
                                    8, VK_SYSTEM_ALLOCATION_SCOPE_OBJECT);
      if (state->attachments == NULL) {
         /* Propagate VK_ERROR_OUT_OF_HOST_MEMORY to vkEndCommandBuffer */
         return anv_batch_set_error(&cmd_buffer->batch,
                                    VK_ERROR_OUT_OF_HOST_MEMORY);
      }
   } else {
      state->attachments = NULL;
   }

   /* Reserve one for the NULL state. */
   unsigned num_states = 1;
   for (uint32_t i = 0; i < pass->attachment_count; ++i) {
      if (vk_format_is_color(pass->attachments[i].format))
         num_states++;

      if (need_input_attachment_state(&pass->attachments[i]))
         num_states++;
   }

   const uint32_t ss_stride = align_u32(isl_dev->ss.size, isl_dev->ss.align);
   state->render_pass_states =
      anv_state_stream_alloc(&cmd_buffer->surface_state_stream,
                             num_states * ss_stride, isl_dev->ss.align);

   struct anv_state next_state = state->render_pass_states;
   next_state.alloc_size = isl_dev->ss.size;

   state->null_surface_state = next_state;
   next_state.offset += ss_stride;
   next_state.map += ss_stride;

   for (uint32_t i = 0; i < pass->attachment_count; ++i) {
      if (vk_format_is_color(pass->attachments[i].format)) {
         state->attachments[i].color.state = next_state;
         next_state.offset += ss_stride;
         next_state.map += ss_stride;
      }

      if (need_input_attachment_state(&pass->attachments[i])) {
         state->attachments[i].input.state = next_state;
         next_state.offset += ss_stride;
         next_state.map += ss_stride;
      }
   }
   assert(next_state.offset == state->render_pass_states.offset +
                               state->render_pass_states.alloc_size);

   if (begin) {
      ANV_FROM_HANDLE(anv_framebuffer, framebuffer, begin->framebuffer);
      assert(pass->attachment_count == framebuffer->attachment_count);

      isl_null_fill_state(isl_dev, state->null_surface_state.map,
                          isl_extent3d(framebuffer->width,
                                       framebuffer->height,
                                       framebuffer->layers));

      for (uint32_t i = 0; i < pass->attachment_count; ++i) {
         struct anv_render_pass_attachment *att = &pass->attachments[i];
         VkImageAspectFlags att_aspects = vk_format_aspects(att->format);
         VkImageAspectFlags clear_aspects = 0;
         VkImageAspectFlags load_aspects = 0;

         if (att_aspects & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV) {
            /* color attachment */
            if (att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
               clear_aspects |= VK_IMAGE_ASPECT_COLOR_BIT;
            } else if (att->load_op == VK_ATTACHMENT_LOAD_OP_LOAD) {
               load_aspects |= VK_IMAGE_ASPECT_COLOR_BIT;
            }
         } else {
            /* depthstencil attachment */
            if (att_aspects & VK_IMAGE_ASPECT_DEPTH_BIT) {
               if (att->load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
                  clear_aspects |= VK_IMAGE_ASPECT_DEPTH_BIT;
               } else if (att->load_op == VK_ATTACHMENT_LOAD_OP_LOAD) {
                  load_aspects |= VK_IMAGE_ASPECT_DEPTH_BIT;
               }
            }
            if (att_aspects & VK_IMAGE_ASPECT_STENCIL_BIT) {
               if (att->stencil_load_op == VK_ATTACHMENT_LOAD_OP_CLEAR) {
                  clear_aspects |= VK_IMAGE_ASPECT_STENCIL_BIT;
               } else if (att->stencil_load_op == VK_ATTACHMENT_LOAD_OP_LOAD) {
                  load_aspects |= VK_IMAGE_ASPECT_STENCIL_BIT;
               }
            }
         }

         state->attachments[i].current_layout = att->initial_layout;
         state->attachments[i].pending_clear_aspects = clear_aspects;
         state->attachments[i].pending_load_aspects = load_aspects;
         if (clear_aspects)
            state->attachments[i].clear_value = begin->pClearValues[i];

         struct anv_image_view *iview = framebuffer->attachments[i];
         anv_assert(iview->vk_format == att->format);

         const uint32_t num_layers = iview->planes[0].isl.array_len;
         state->attachments[i].pending_clear_views = (1 << num_layers) - 1;

         union isl_color_value clear_color = { .u32 = { 0, } };
         if (att_aspects & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV) {
            anv_assert(iview->n_planes == 1);
            assert(att_aspects == VK_IMAGE_ASPECT_COLOR_BIT);
            color_attachment_compute_aux_usage(cmd_buffer->device,
                                               state, i, begin->renderArea,
                                               &clear_color);

            anv_image_fill_surface_state(cmd_buffer->device,
                                         iview->image,
                                         VK_IMAGE_ASPECT_COLOR_BIT,
                                         &iview->planes[0].isl,
                                         ISL_SURF_USAGE_RENDER_TARGET_BIT,
                                         state->attachments[i].aux_usage,
                                         &clear_color,
                                         0,
                                         &state->attachments[i].color,
                                         NULL);

            add_surface_state_relocs(cmd_buffer, state->attachments[i].color);
         } else {
            depth_stencil_attachment_compute_aux_usage(cmd_buffer->device,
                                                       state, i,
                                                       begin->renderArea);
         }

         if (need_input_attachment_state(&pass->attachments[i])) {
            anv_image_fill_surface_state(cmd_buffer->device,
                                         iview->image,
                                         VK_IMAGE_ASPECT_COLOR_BIT,
                                         &iview->planes[0].isl,
                                         ISL_SURF_USAGE_TEXTURE_BIT,
                                         state->attachments[i].input_aux_usage,
                                         &clear_color,
                                         0,
                                         &state->attachments[i].input,
                                         NULL);

            add_surface_state_relocs(cmd_buffer, state->attachments[i].input);
         }
      }
   }

   return VK_SUCCESS;
}

VkResult
genX(BeginCommandBuffer)(
    VkCommandBuffer                             commandBuffer,
    const VkCommandBufferBeginInfo*             pBeginInfo)
{
   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);

   /* If this is the first vkBeginCommandBuffer, we must *initialize* the
    * command buffer's state. Otherwise, we must *reset* its state. In both
    * cases we reset it.
    *
    * From the Vulkan 1.0 spec:
    *
    *    If a command buffer is in the executable state and the command buffer
    *    was allocated from a command pool with the
    *    VK_COMMAND_POOL_CREATE_RESET_COMMAND_BUFFER_BIT flag set, then
    *    vkBeginCommandBuffer implicitly resets the command buffer, behaving
    *    as if vkResetCommandBuffer had been called with
    *    VK_COMMAND_BUFFER_RESET_RELEASE_RESOURCES_BIT not set. It then puts
    *    the command buffer in the recording state.
    */
   anv_cmd_buffer_reset(cmd_buffer);

   cmd_buffer->usage_flags = pBeginInfo->flags;

   assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY ||
          !(cmd_buffer->usage_flags & VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT));

   genX(cmd_buffer_emit_state_base_address)(cmd_buffer);

   /* We sometimes store vertex data in the dynamic state buffer for blorp
    * operations and our dynamic state stream may re-use data from previous
    * command buffers.  In order to prevent stale cache data, we flush the VF
    * cache.  We could do this on every blorp call but that's not really
    * needed as all of the data will get written by the CPU prior to the GPU
    * executing anything.  The chances are fairly high that they will use
    * blorp at least once per primary command buffer so it shouldn't be
    * wasted.
    */
   if (cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY)
      cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_VF_CACHE_INVALIDATE_BIT;

   /* We send an "Indirect State Pointers Disable" packet at
    * EndCommandBuffer, so all push contant packets are ignored during a
    * context restore. Documentation says after that command, we need to
    * emit push constants again before any rendering operation. So we
    * flag them dirty here to make sure they get emitted.
    */
   cmd_buffer->state.push_constants_dirty |= VK_SHADER_STAGE_ALL_GRAPHICS;

   VkResult result = VK_SUCCESS;
   if (cmd_buffer->usage_flags &
       VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT) {
      assert(pBeginInfo->pInheritanceInfo);
      cmd_buffer->state.pass =
         anv_render_pass_from_handle(pBeginInfo->pInheritanceInfo->renderPass);
      cmd_buffer->state.subpass =
         &cmd_buffer->state.pass->subpasses[pBeginInfo->pInheritanceInfo->subpass];

      /* This is optional in the inheritance info. */
      cmd_buffer->state.framebuffer =
         anv_framebuffer_from_handle(pBeginInfo->pInheritanceInfo->framebuffer);

      result = genX(cmd_buffer_setup_attachments)(cmd_buffer,
                                                  cmd_buffer->state.pass, NULL);

      /* Record that HiZ is enabled if we can. */
      if (cmd_buffer->state.framebuffer) {
         const struct anv_image_view * const iview =
            anv_cmd_buffer_get_depth_stencil_view(cmd_buffer);

         if (iview) {
            VkImageLayout layout =
                cmd_buffer->state.subpass->depth_stencil_attachment.layout;

            enum isl_aux_usage aux_usage =
               anv_layout_to_aux_usage(&cmd_buffer->device->info, iview->image,
                                       VK_IMAGE_ASPECT_DEPTH_BIT, layout);

            cmd_buffer->state.hiz_enabled = aux_usage == ISL_AUX_USAGE_HIZ;
         }
      }

      cmd_buffer->state.gfx.dirty |= ANV_CMD_DIRTY_RENDER_TARGETS;
   }

   return result;
}

/* From the PRM, Volume 2a:
 *
 *    "Indirect State Pointers Disable
 *
 *    At the completion of the post-sync operation associated with this pipe
 *    control packet, the indirect state pointers in the hardware are
 *    considered invalid; the indirect pointers are not saved in the context.
 *    If any new indirect state commands are executed in the command stream
 *    while the pipe control is pending, the new indirect state commands are
 *    preserved.
 *
 *    [DevIVB+]: Using Invalidate State Pointer (ISP) only inhibits context
 *    restoring of Push Constant (3DSTATE_CONSTANT_*) commands. Push Constant
 *    commands are only considered as Indirect State Pointers. Once ISP is
 *    issued in a context, SW must initialize by programming push constant
 *    commands for all the shaders (at least to zero length) before attempting
 *    any rendering operation for the same context."
 *
 * 3DSTATE_CONSTANT_* packets are restored during a context restore,
 * even though they point to a BO that has been already unreferenced at
 * the end of the previous batch buffer. This has been fine so far since
 * we are protected by these scratch page (every address not covered by
 * a BO should be pointing to the scratch page). But on CNL, it is
 * causing a GPU hang during context restore at the 3DSTATE_CONSTANT_*
 * instruction.
 *
 * The flag "Indirect State Pointers Disable" in PIPE_CONTROL tells the
 * hardware to ignore previous 3DSTATE_CONSTANT_* packets during a
 * context restore, so the mentioned hang doesn't happen. However,
 * software must program push constant commands for all stages prior to
 * rendering anything. So we flag them dirty in BeginCommandBuffer.
 *
 * Finally, we also make sure to stall at pixel scoreboard to make sure the
 * constants have been loaded into the EUs prior to disable the push constants
 * so that it doesn't hang a previous 3DPRIMITIVE.
 */
static void
emit_isp_disable(struct anv_cmd_buffer *cmd_buffer)
{
   anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
         pc.StallAtPixelScoreboard = true;
         pc.CommandStreamerStallEnable = true;
   }
   anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
         pc.IndirectStatePointersDisable = true;
         pc.CommandStreamerStallEnable = true;
   }
}

VkResult
genX(EndCommandBuffer)(
    VkCommandBuffer                             commandBuffer)
{
   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);

   if (anv_batch_has_error(&cmd_buffer->batch))
      return cmd_buffer->batch.status;

   /* We want every command buffer to start with the PMA fix in a known state,
    * so we disable it at the end of the command buffer.
    */
   genX(cmd_buffer_enable_pma_fix)(cmd_buffer, false);

   genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);

   emit_isp_disable(cmd_buffer);

   anv_cmd_buffer_end_batch_buffer(cmd_buffer);

   return VK_SUCCESS;
}

void
genX(CmdExecuteCommands)(
    VkCommandBuffer                             commandBuffer,
    uint32_t                                    commandBufferCount,
    const VkCommandBuffer*                      pCmdBuffers)
{
   ANV_FROM_HANDLE(anv_cmd_buffer, primary, commandBuffer);

   assert(primary->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);

   if (anv_batch_has_error(&primary->batch))
      return;

   /* The secondary command buffers will assume that the PMA fix is disabled
    * when they begin executing.  Make sure this is true.
    */
   genX(cmd_buffer_enable_pma_fix)(primary, false);

   /* The secondary command buffer doesn't know which textures etc. have been
    * flushed prior to their execution.  Apply those flushes now.
    */
   genX(cmd_buffer_apply_pipe_flushes)(primary);

   for (uint32_t i = 0; i < commandBufferCount; i++) {
      ANV_FROM_HANDLE(anv_cmd_buffer, secondary, pCmdBuffers[i]);

      assert(secondary->level == VK_COMMAND_BUFFER_LEVEL_SECONDARY);
      assert(!anv_batch_has_error(&secondary->batch));

      if (secondary->usage_flags &
          VK_COMMAND_BUFFER_USAGE_RENDER_PASS_CONTINUE_BIT) {
         /* If we're continuing a render pass from the primary, we need to
          * copy the surface states for the current subpass into the storage
          * we allocated for them in BeginCommandBuffer.
          */
         struct anv_bo *ss_bo =
            &primary->device->surface_state_pool.block_pool.bo;
         struct anv_state src_state = primary->state.render_pass_states;
         struct anv_state dst_state = secondary->state.render_pass_states;
         assert(src_state.alloc_size == dst_state.alloc_size);

         genX(cmd_buffer_so_memcpy)(primary, ss_bo, dst_state.offset,
                                    ss_bo, src_state.offset,
                                    src_state.alloc_size);
      }

      anv_cmd_buffer_add_secondary(primary, secondary);
   }

   /* The secondary may have selected a different pipeline (3D or compute) and
    * may have changed the current L3$ configuration.  Reset our tracking
    * variables to invalid values to ensure that we re-emit these in the case
    * where we do any draws or compute dispatches from the primary after the
    * secondary has returned.
    */
   primary->state.current_pipeline = UINT32_MAX;
   primary->state.current_l3_config = NULL;

   /* Each of the secondary command buffers will use its own state base
    * address.  We need to re-emit state base address for the primary after
    * all of the secondaries are done.
    *
    * TODO: Maybe we want to make this a dirty bit to avoid extra state base
    * address calls?
    */
   genX(cmd_buffer_emit_state_base_address)(primary);
}

#define IVB_L3SQCREG1_SQGHPCI_DEFAULT     0x00730000
#define VLV_L3SQCREG1_SQGHPCI_DEFAULT     0x00d30000
#define HSW_L3SQCREG1_SQGHPCI_DEFAULT     0x00610000

/**
 * Program the hardware to use the specified L3 configuration.
 */
void
genX(cmd_buffer_config_l3)(struct anv_cmd_buffer *cmd_buffer,
                           const struct gen_l3_config *cfg)
{
   assert(cfg);
   if (cfg == cmd_buffer->state.current_l3_config)
      return;

   if (unlikely(INTEL_DEBUG & DEBUG_L3)) {
      intel_logd("L3 config transition: ");
      gen_dump_l3_config(cfg, stderr);
   }

   const bool has_slm = cfg->n[GEN_L3P_SLM];

   /* According to the hardware docs, the L3 partitioning can only be changed
    * while the pipeline is completely drained and the caches are flushed,
    * which involves a first PIPE_CONTROL flush which stalls the pipeline...
    */
   anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
      pc.DCFlushEnable = true;
      pc.PostSyncOperation = NoWrite;
      pc.CommandStreamerStallEnable = true;
   }

   /* ...followed by a second pipelined PIPE_CONTROL that initiates
    * invalidation of the relevant caches.  Note that because RO invalidation
    * happens at the top of the pipeline (i.e. right away as the PIPE_CONTROL
    * command is processed by the CS) we cannot combine it with the previous
    * stalling flush as the hardware documentation suggests, because that
    * would cause the CS to stall on previous rendering *after* RO
    * invalidation and wouldn't prevent the RO caches from being polluted by
    * concurrent rendering before the stall completes.  This intentionally
    * doesn't implement the SKL+ hardware workaround suggesting to enable CS
    * stall on PIPE_CONTROLs with the texture cache invalidation bit set for
    * GPGPU workloads because the previous and subsequent PIPE_CONTROLs
    * already guarantee that there is no concurrent GPGPU kernel execution
    * (see SKL HSD 2132585).
    */
   anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
      pc.TextureCacheInvalidationEnable = true;
      pc.ConstantCacheInvalidationEnable = true;
      pc.InstructionCacheInvalidateEnable = true;
      pc.StateCacheInvalidationEnable = true;
      pc.PostSyncOperation = NoWrite;
   }

   /* Now send a third stalling flush to make sure that invalidation is
    * complete when the L3 configuration registers are modified.
    */
   anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
      pc.DCFlushEnable = true;
      pc.PostSyncOperation = NoWrite;
      pc.CommandStreamerStallEnable = true;
   }

#if GEN_GEN >= 8

   assert(!cfg->n[GEN_L3P_IS] && !cfg->n[GEN_L3P_C] && !cfg->n[GEN_L3P_T]);

   uint32_t l3cr;
   anv_pack_struct(&l3cr, GENX(L3CNTLREG),
                   .SLMEnable = has_slm,
                   .URBAllocation = cfg->n[GEN_L3P_URB],
                   .ROAllocation = cfg->n[GEN_L3P_RO],
                   .DCAllocation = cfg->n[GEN_L3P_DC],
                   .AllAllocation = cfg->n[GEN_L3P_ALL]);

   /* Set up the L3 partitioning. */
   emit_lri(&cmd_buffer->batch, GENX(L3CNTLREG_num), l3cr);

#else

   const bool has_dc = cfg->n[GEN_L3P_DC] || cfg->n[GEN_L3P_ALL];
   const bool has_is = cfg->n[GEN_L3P_IS] || cfg->n[GEN_L3P_RO] ||
                       cfg->n[GEN_L3P_ALL];
   const bool has_c = cfg->n[GEN_L3P_C] || cfg->n[GEN_L3P_RO] ||
                      cfg->n[GEN_L3P_ALL];
   const bool has_t = cfg->n[GEN_L3P_T] || cfg->n[GEN_L3P_RO] ||
                      cfg->n[GEN_L3P_ALL];

   assert(!cfg->n[GEN_L3P_ALL]);

   /* When enabled SLM only uses a portion of the L3 on half of the banks,
    * the matching space on the remaining banks has to be allocated to a
    * client (URB for all validated configurations) set to the
    * lower-bandwidth 2-bank address hashing mode.
    */
   const struct gen_device_info *devinfo = &cmd_buffer->device->info;
   const bool urb_low_bw = has_slm && !devinfo->is_baytrail;
   assert(!urb_low_bw || cfg->n[GEN_L3P_URB] == cfg->n[GEN_L3P_SLM]);

   /* Minimum number of ways that can be allocated to the URB. */
   MAYBE_UNUSED const unsigned n0_urb = devinfo->is_baytrail ? 32 : 0;
   assert(cfg->n[GEN_L3P_URB] >= n0_urb);

   uint32_t l3sqcr1, l3cr2, l3cr3;
   anv_pack_struct(&l3sqcr1, GENX(L3SQCREG1),
                   .ConvertDC_UC = !has_dc,
                   .ConvertIS_UC = !has_is,
                   .ConvertC_UC = !has_c,
                   .ConvertT_UC = !has_t);
   l3sqcr1 |=
      GEN_IS_HASWELL ? HSW_L3SQCREG1_SQGHPCI_DEFAULT :
      devinfo->is_baytrail ? VLV_L3SQCREG1_SQGHPCI_DEFAULT :
      IVB_L3SQCREG1_SQGHPCI_DEFAULT;

   anv_pack_struct(&l3cr2, GENX(L3CNTLREG2),
                   .SLMEnable = has_slm,
                   .URBLowBandwidth = urb_low_bw,
                   .URBAllocation = cfg->n[GEN_L3P_URB] - n0_urb,
#if !GEN_IS_HASWELL
                   .ALLAllocation = cfg->n[GEN_L3P_ALL],
#endif
                   .ROAllocation = cfg->n[GEN_L3P_RO],
                   .DCAllocation = cfg->n[GEN_L3P_DC]);

   anv_pack_struct(&l3cr3, GENX(L3CNTLREG3),
                   .ISAllocation = cfg->n[GEN_L3P_IS],
                   .ISLowBandwidth = 0,
                   .CAllocation = cfg->n[GEN_L3P_C],
                   .CLowBandwidth = 0,
                   .TAllocation = cfg->n[GEN_L3P_T],
                   .TLowBandwidth = 0);

   /* Set up the L3 partitioning. */
   emit_lri(&cmd_buffer->batch, GENX(L3SQCREG1_num), l3sqcr1);
   emit_lri(&cmd_buffer->batch, GENX(L3CNTLREG2_num), l3cr2);
   emit_lri(&cmd_buffer->batch, GENX(L3CNTLREG3_num), l3cr3);

#if GEN_IS_HASWELL
   if (cmd_buffer->device->instance->physicalDevice.cmd_parser_version >= 4) {
      /* Enable L3 atomics on HSW if we have a DC partition, otherwise keep
       * them disabled to avoid crashing the system hard.
       */
      uint32_t scratch1, chicken3;
      anv_pack_struct(&scratch1, GENX(SCRATCH1),
                      .L3AtomicDisable = !has_dc);
      anv_pack_struct(&chicken3, GENX(CHICKEN3),
                      .L3AtomicDisableMask = true,
                      .L3AtomicDisable = !has_dc);
      emit_lri(&cmd_buffer->batch, GENX(SCRATCH1_num), scratch1);
      emit_lri(&cmd_buffer->batch, GENX(CHICKEN3_num), chicken3);
   }
#endif

#endif

   cmd_buffer->state.current_l3_config = cfg;
}

void
genX(cmd_buffer_apply_pipe_flushes)(struct anv_cmd_buffer *cmd_buffer)
{
   enum anv_pipe_bits bits = cmd_buffer->state.pending_pipe_bits;

   /* Flushes are pipelined while invalidations are handled immediately.
    * Therefore, if we're flushing anything then we need to schedule a stall
    * before any invalidations can happen.
    */
   if (bits & ANV_PIPE_FLUSH_BITS)
      bits |= ANV_PIPE_NEEDS_CS_STALL_BIT;

   /* If we're going to do an invalidate and we have a pending CS stall that
    * has yet to be resolved, we do the CS stall now.
    */
   if ((bits & ANV_PIPE_INVALIDATE_BITS) &&
       (bits & ANV_PIPE_NEEDS_CS_STALL_BIT)) {
      bits |= ANV_PIPE_CS_STALL_BIT;
      bits &= ~ANV_PIPE_NEEDS_CS_STALL_BIT;
   }

   if (bits & (ANV_PIPE_FLUSH_BITS | ANV_PIPE_CS_STALL_BIT)) {
      anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pipe) {
         pipe.DepthCacheFlushEnable = bits & ANV_PIPE_DEPTH_CACHE_FLUSH_BIT;
         pipe.DCFlushEnable = bits & ANV_PIPE_DATA_CACHE_FLUSH_BIT;
         pipe.RenderTargetCacheFlushEnable =
            bits & ANV_PIPE_RENDER_TARGET_CACHE_FLUSH_BIT;

         pipe.DepthStallEnable = bits & ANV_PIPE_DEPTH_STALL_BIT;
         pipe.CommandStreamerStallEnable = bits & ANV_PIPE_CS_STALL_BIT;
         pipe.StallAtPixelScoreboard = bits & ANV_PIPE_STALL_AT_SCOREBOARD_BIT;

         /*
          * According to the Broadwell documentation, any PIPE_CONTROL with the
          * "Command Streamer Stall" bit set must also have another bit set,
          * with five different options:
          *
          *  - Render Target Cache Flush
          *  - Depth Cache Flush
          *  - Stall at Pixel Scoreboard
          *  - Post-Sync Operation
          *  - Depth Stall
          *  - DC Flush Enable
          *
          * I chose "Stall at Pixel Scoreboard" since that's what we use in
          * mesa and it seems to work fine. The choice is fairly arbitrary.
          */
         if ((bits & ANV_PIPE_CS_STALL_BIT) &&
             !(bits & (ANV_PIPE_FLUSH_BITS | ANV_PIPE_DEPTH_STALL_BIT |
                       ANV_PIPE_STALL_AT_SCOREBOARD_BIT)))
            pipe.StallAtPixelScoreboard = true;
      }

      bits &= ~(ANV_PIPE_FLUSH_BITS | ANV_PIPE_CS_STALL_BIT);
   }

   if (bits & ANV_PIPE_INVALIDATE_BITS) {
      anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pipe) {
         pipe.StateCacheInvalidationEnable =
            bits & ANV_PIPE_STATE_CACHE_INVALIDATE_BIT;
         pipe.ConstantCacheInvalidationEnable =
            bits & ANV_PIPE_CONSTANT_CACHE_INVALIDATE_BIT;
         pipe.VFCacheInvalidationEnable =
            bits & ANV_PIPE_VF_CACHE_INVALIDATE_BIT;
         pipe.TextureCacheInvalidationEnable =
            bits & ANV_PIPE_TEXTURE_CACHE_INVALIDATE_BIT;
         pipe.InstructionCacheInvalidateEnable =
            bits & ANV_PIPE_INSTRUCTION_CACHE_INVALIDATE_BIT;
      }

      bits &= ~ANV_PIPE_INVALIDATE_BITS;
   }

   cmd_buffer->state.pending_pipe_bits = bits;
}

void genX(CmdPipelineBarrier)(
    VkCommandBuffer                             commandBuffer,
    VkPipelineStageFlags                        srcStageMask,
    VkPipelineStageFlags                        destStageMask,
    VkBool32                                    byRegion,
    uint32_t                                    memoryBarrierCount,
    const VkMemoryBarrier*                      pMemoryBarriers,
    uint32_t                                    bufferMemoryBarrierCount,
    const VkBufferMemoryBarrier*                pBufferMemoryBarriers,
    uint32_t                                    imageMemoryBarrierCount,
    const VkImageMemoryBarrier*                 pImageMemoryBarriers)
{
   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);

   /* 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.
    */
   VkAccessFlags src_flags = 0;
   VkAccessFlags dst_flags = 0;

   for (uint32_t i = 0; i < memoryBarrierCount; i++) {
      src_flags |= pMemoryBarriers[i].srcAccessMask;
      dst_flags |= pMemoryBarriers[i].dstAccessMask;
   }

   for (uint32_t i = 0; i < bufferMemoryBarrierCount; i++) {
      src_flags |= pBufferMemoryBarriers[i].srcAccessMask;
      dst_flags |= pBufferMemoryBarriers[i].dstAccessMask;
   }

   for (uint32_t i = 0; i < imageMemoryBarrierCount; i++) {
      src_flags |= pImageMemoryBarriers[i].srcAccessMask;
      dst_flags |= pImageMemoryBarriers[i].dstAccessMask;
      ANV_FROM_HANDLE(anv_image, image, pImageMemoryBarriers[i].image);
      const VkImageSubresourceRange *range =
         &pImageMemoryBarriers[i].subresourceRange;

      if (range->aspectMask & VK_IMAGE_ASPECT_DEPTH_BIT) {
         transition_depth_buffer(cmd_buffer, image,
                                 pImageMemoryBarriers[i].oldLayout,
                                 pImageMemoryBarriers[i].newLayout);
      } else if (range->aspectMask & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV) {
         VkImageAspectFlags color_aspects =
            anv_image_expand_aspects(image, range->aspectMask);
         uint32_t aspect_bit;

         uint32_t base_layer, layer_count;
         if (image->type == VK_IMAGE_TYPE_3D) {
            base_layer = 0;
            layer_count = anv_minify(image->extent.depth, range->baseMipLevel);
         } else {
            base_layer = range->baseArrayLayer;
            layer_count = anv_get_layerCount(image, range);
         }

         anv_foreach_image_aspect_bit(aspect_bit, image, color_aspects) {
            transition_color_buffer(cmd_buffer, image, 1UL << aspect_bit,
                                    range->baseMipLevel,
                                    anv_get_levelCount(image, range),
                                    base_layer, layer_count,
                                    pImageMemoryBarriers[i].oldLayout,
                                    pImageMemoryBarriers[i].newLayout);
         }
      }
   }

   cmd_buffer->state.pending_pipe_bits |=
      anv_pipe_flush_bits_for_access_flags(src_flags) |
      anv_pipe_invalidate_bits_for_access_flags(dst_flags);
}

static void
cmd_buffer_alloc_push_constants(struct anv_cmd_buffer *cmd_buffer)
{
   VkShaderStageFlags stages =
      cmd_buffer->state.gfx.base.pipeline->active_stages;

   /* In order to avoid thrash, we assume that vertex and fragment stages
    * always exist.  In the rare case where one is missing *and* the other
    * uses push concstants, this may be suboptimal.  However, avoiding stalls
    * seems more important.
    */
   stages |= VK_SHADER_STAGE_FRAGMENT_BIT | VK_SHADER_STAGE_VERTEX_BIT;

   if (stages == cmd_buffer->state.push_constant_stages)
      return;

#if GEN_GEN >= 8
   const unsigned push_constant_kb = 32;
#elif GEN_IS_HASWELL
   const unsigned push_constant_kb = cmd_buffer->device->info.gt == 3 ? 32 : 16;
#else
   const unsigned push_constant_kb = 16;
#endif

   const unsigned num_stages =
      _mesa_bitcount(stages & VK_SHADER_STAGE_ALL_GRAPHICS);
   unsigned size_per_stage = push_constant_kb / num_stages;

   /* Broadwell+ and Haswell gt3 require that the push constant sizes be in
    * units of 2KB.  Incidentally, these are the same platforms that have
    * 32KB worth of push constant space.
    */
   if (push_constant_kb == 32)
      size_per_stage &= ~1u;

   uint32_t kb_used = 0;
   for (int i = MESA_SHADER_VERTEX; i < MESA_SHADER_FRAGMENT; i++) {
      unsigned push_size = (stages & (1 << i)) ? size_per_stage : 0;
      anv_batch_emit(&cmd_buffer->batch,
                     GENX(3DSTATE_PUSH_CONSTANT_ALLOC_VS), alloc) {
         alloc._3DCommandSubOpcode  = 18 + i;
         alloc.ConstantBufferOffset = (push_size > 0) ? kb_used : 0;
         alloc.ConstantBufferSize   = push_size;
      }
      kb_used += push_size;
   }

   anv_batch_emit(&cmd_buffer->batch,
                  GENX(3DSTATE_PUSH_CONSTANT_ALLOC_PS), alloc) {
      alloc.ConstantBufferOffset = kb_used;
      alloc.ConstantBufferSize = push_constant_kb - kb_used;
   }

   cmd_buffer->state.push_constant_stages = stages;

   /* From the BDW PRM for 3DSTATE_PUSH_CONSTANT_ALLOC_VS:
    *
    *    "The 3DSTATE_CONSTANT_VS must be reprogrammed prior to
    *    the next 3DPRIMITIVE command after programming the
    *    3DSTATE_PUSH_CONSTANT_ALLOC_VS"
    *
    * Since 3DSTATE_PUSH_CONSTANT_ALLOC_VS is programmed as part of
    * pipeline setup, we need to dirty push constants.
    */
   cmd_buffer->state.push_constants_dirty |= VK_SHADER_STAGE_ALL_GRAPHICS;
}

static const struct anv_descriptor *
anv_descriptor_for_binding(const struct anv_cmd_pipeline_state *pipe_state,
                           const struct anv_pipeline_binding *binding)
{
   assert(binding->set < MAX_SETS);
   const struct anv_descriptor_set *set =
      pipe_state->descriptors[binding->set];
   const uint32_t offset =
      set->layout->binding[binding->binding].descriptor_index;
   return &set->descriptors[offset + binding->index];
}

static uint32_t
dynamic_offset_for_binding(const struct anv_cmd_pipeline_state *pipe_state,
                           const struct anv_pipeline_binding *binding)
{
   assert(binding->set < MAX_SETS);
   const struct anv_descriptor_set *set =
      pipe_state->descriptors[binding->set];

   uint32_t dynamic_offset_idx =
      pipe_state->layout->set[binding->set].dynamic_offset_start +
      set->layout->binding[binding->binding].dynamic_offset_index +
      binding->index;

   return pipe_state->dynamic_offsets[dynamic_offset_idx];
}

static VkResult
emit_binding_table(struct anv_cmd_buffer *cmd_buffer,
                   gl_shader_stage stage,
                   struct anv_state *bt_state)
{
   struct anv_subpass *subpass = cmd_buffer->state.subpass;
   struct anv_cmd_pipeline_state *pipe_state;
   struct anv_pipeline *pipeline;
   uint32_t bias, state_offset;

   switch (stage) {
   case  MESA_SHADER_COMPUTE:
      pipe_state = &cmd_buffer->state.compute.base;
      bias = 1;
      break;
   default:
      pipe_state = &cmd_buffer->state.gfx.base;
      bias = 0;
      break;
   }
   pipeline = pipe_state->pipeline;

   if (!anv_pipeline_has_stage(pipeline, stage)) {
      *bt_state = (struct anv_state) { 0, };
      return VK_SUCCESS;
   }

   struct anv_pipeline_bind_map *map = &pipeline->shaders[stage]->bind_map;
   if (bias + map->surface_count == 0) {
      *bt_state = (struct anv_state) { 0, };
      return VK_SUCCESS;
   }

   *bt_state = anv_cmd_buffer_alloc_binding_table(cmd_buffer,
                                                  bias + map->surface_count,
                                                  &state_offset);
   uint32_t *bt_map = bt_state->map;

   if (bt_state->map == NULL)
      return VK_ERROR_OUT_OF_DEVICE_MEMORY;

   if (stage == MESA_SHADER_COMPUTE &&
       get_cs_prog_data(pipeline)->uses_num_work_groups) {
      struct anv_state surface_state;
      surface_state =
         anv_cmd_buffer_alloc_surface_state(cmd_buffer);

      const enum isl_format format =
         anv_isl_format_for_descriptor_type(VK_DESCRIPTOR_TYPE_STORAGE_BUFFER);
      anv_fill_buffer_surface_state(cmd_buffer->device, surface_state,
                                    format,
                                    cmd_buffer->state.compute.num_workgroups,
                                    12, 1);

      bt_map[0] = surface_state.offset + state_offset;
      add_surface_reloc(cmd_buffer, surface_state,
                        cmd_buffer->state.compute.num_workgroups);
   }

   if (map->surface_count == 0)
      goto out;

   if (map->image_count > 0) {
      VkResult result =
         anv_cmd_buffer_ensure_push_constant_field(cmd_buffer, stage, images);
      if (result != VK_SUCCESS)
         return result;

      cmd_buffer->state.push_constants_dirty |= 1 << stage;
   }

   uint32_t image = 0;
   for (uint32_t s = 0; s < map->surface_count; s++) {
      struct anv_pipeline_binding *binding = &map->surface_to_descriptor[s];

      struct anv_state surface_state;

      if (binding->set == ANV_DESCRIPTOR_SET_COLOR_ATTACHMENTS) {
         /* Color attachment binding */
         assert(stage == MESA_SHADER_FRAGMENT);
         assert(binding->binding == 0);
         if (binding->index < subpass->color_count) {
            const unsigned att =
               subpass->color_attachments[binding->index].attachment;

            /* From the Vulkan 1.0.46 spec:
             *
             *    "If any color or depth/stencil attachments are
             *    VK_ATTACHMENT_UNUSED, then no writes occur for those
             *    attachments."
             */
            if (att == VK_ATTACHMENT_UNUSED) {
               surface_state = cmd_buffer->state.null_surface_state;
            } else {
               surface_state = cmd_buffer->state.attachments[att].color.state;
            }
         } else {
            surface_state = cmd_buffer->state.null_surface_state;
         }

         bt_map[bias + s] = surface_state.offset + state_offset;
         continue;
      }

      const struct anv_descriptor *desc =
         anv_descriptor_for_binding(pipe_state, binding);

      switch (desc->type) {
      case VK_DESCRIPTOR_TYPE_SAMPLER:
         /* Nothing for us to do here */
         continue;

      case VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER:
      case VK_DESCRIPTOR_TYPE_SAMPLED_IMAGE: {
         struct anv_surface_state sstate =
            (desc->layout == VK_IMAGE_LAYOUT_GENERAL) ?
            desc->image_view->planes[binding->plane].general_sampler_surface_state :
            desc->image_view->planes[binding->plane].optimal_sampler_surface_state;
         surface_state = sstate.state;
         assert(surface_state.alloc_size);
         add_surface_state_relocs(cmd_buffer, sstate);
         break;
      }
      case VK_DESCRIPTOR_TYPE_INPUT_ATTACHMENT:
         assert(stage == MESA_SHADER_FRAGMENT);
         if ((desc->image_view->aspect_mask & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV) == 0) {
            /* For depth and stencil input attachments, we treat it like any
             * old texture that a user may have bound.
             */
            struct anv_surface_state sstate =
               (desc->layout == VK_IMAGE_LAYOUT_GENERAL) ?
               desc->image_view->planes[binding->plane].general_sampler_surface_state :
               desc->image_view->planes[binding->plane].optimal_sampler_surface_state;
            surface_state = sstate.state;
            assert(surface_state.alloc_size);
            add_surface_state_relocs(cmd_buffer, sstate);
         } else {
            /* For color input attachments, we create the surface state at
             * vkBeginRenderPass time so that we can include aux and clear
             * color information.
             */
            assert(binding->input_attachment_index < subpass->input_count);
            const unsigned subpass_att = binding->input_attachment_index;
            const unsigned att = subpass->input_attachments[subpass_att].attachment;
            surface_state = cmd_buffer->state.attachments[att].input.state;
         }
         break;

      case VK_DESCRIPTOR_TYPE_STORAGE_IMAGE: {
         struct anv_surface_state sstate = (binding->write_only)
            ? desc->image_view->planes[binding->plane].writeonly_storage_surface_state
            : desc->image_view->planes[binding->plane].storage_surface_state;
         surface_state = sstate.state;
         assert(surface_state.alloc_size);
         add_surface_state_relocs(cmd_buffer, sstate);

         struct brw_image_param *image_param =
            &cmd_buffer->state.push_constants[stage]->images[image++];

         *image_param = desc->image_view->planes[binding->plane].storage_image_param;
         image_param->surface_idx = bias + s;
         break;
      }

      case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER:
      case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER:
      case VK_DESCRIPTOR_TYPE_UNIFORM_TEXEL_BUFFER:
         surface_state = desc->buffer_view->surface_state;
         assert(surface_state.alloc_size);
         add_surface_reloc(cmd_buffer, surface_state,
                           desc->buffer_view->address);
         break;

      case VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC:
      case VK_DESCRIPTOR_TYPE_STORAGE_BUFFER_DYNAMIC: {
         /* Compute the offset within the buffer */
         uint32_t dynamic_offset =
            dynamic_offset_for_binding(pipe_state, binding);
         uint64_t offset = desc->offset + dynamic_offset;
         /* Clamp to the buffer size */
         offset = MIN2(offset, desc->buffer->size);
         /* Clamp the range to the buffer size */
         uint32_t range = MIN2(desc->range, desc->buffer->size - offset);

         struct anv_address address =
            anv_address_add(desc->buffer->address, offset);

         surface_state =
            anv_state_stream_alloc(&cmd_buffer->surface_state_stream, 64, 64);
         enum isl_format format =
            anv_isl_format_for_descriptor_type(desc->type);

         anv_fill_buffer_surface_state(cmd_buffer->device, surface_state,
                                       format, address, range, 1);
         add_surface_reloc(cmd_buffer, surface_state, address);
         break;
      }

      case VK_DESCRIPTOR_TYPE_STORAGE_TEXEL_BUFFER:
         surface_state = (binding->write_only)
            ? desc->buffer_view->writeonly_storage_surface_state
            : desc->buffer_view->storage_surface_state;
         assert(surface_state.alloc_size);
         add_surface_reloc(cmd_buffer, surface_state,
                           desc->buffer_view->address);

         struct brw_image_param *image_param =
            &cmd_buffer->state.push_constants[stage]->images[image++];

         *image_param = desc->buffer_view->storage_image_param;
         image_param->surface_idx = bias + s;
         break;

      default:
         assert(!"Invalid descriptor type");
         continue;
      }

      bt_map[bias + s] = surface_state.offset + state_offset;
   }
   assert(image == map->image_count);

 out:
   anv_state_flush(cmd_buffer->device, *bt_state);

#if GEN_GEN >= 11
   /* The PIPE_CONTROL command description says:
    *
    *    "Whenever a Binding Table Index (BTI) used by a Render Taget Message
    *     points to a different RENDER_SURFACE_STATE, SW must issue a Render
    *     Target Cache Flush by enabling this bit. When render target flush
    *     is set due to new association of BTI, PS Scoreboard Stall bit must
    *     be set in this packet."
    *
    * FINISHME: Currently we shuffle around the surface states in the binding
    * table based on if they are getting used or not. So, we've to do below
    * pipe control flush for every binding table upload. Make changes so
    * that we do it only when we modify render target surface states.
    */
   anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
      pc.RenderTargetCacheFlushEnable  = true;
      pc.StallAtPixelScoreboard        = true;
   }
#endif

   return VK_SUCCESS;
}

static VkResult
emit_samplers(struct anv_cmd_buffer *cmd_buffer,
              gl_shader_stage stage,
              struct anv_state *state)
{
   struct anv_cmd_pipeline_state *pipe_state =
      stage == MESA_SHADER_COMPUTE ? &cmd_buffer->state.compute.base :
                                     &cmd_buffer->state.gfx.base;
   struct anv_pipeline *pipeline = pipe_state->pipeline;

   if (!anv_pipeline_has_stage(pipeline, stage)) {
      *state = (struct anv_state) { 0, };
      return VK_SUCCESS;
   }

   struct anv_pipeline_bind_map *map = &pipeline->shaders[stage]->bind_map;
   if (map->sampler_count == 0) {
      *state = (struct anv_state) { 0, };
      return VK_SUCCESS;
   }

   uint32_t size = map->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 s = 0; s < map->sampler_count; s++) {
      struct anv_pipeline_binding *binding = &map->sampler_to_descriptor[s];
      const struct anv_descriptor *desc =
         anv_descriptor_for_binding(pipe_state, binding);

      if (desc->type != VK_DESCRIPTOR_TYPE_SAMPLER &&
          desc->type != VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER)
         continue;

      struct anv_sampler *sampler = desc->sampler;

      /* This can happen if we have an unfilled slot since TYPE_SAMPLER
       * happens to be zero.
       */
      if (sampler == NULL)
         continue;

      memcpy(state->map + (s * 16),
             sampler->state[binding->plane], sizeof(sampler->state[0]));
   }

   anv_state_flush(cmd_buffer->device, *state);

   return VK_SUCCESS;
}

static uint32_t
flush_descriptor_sets(struct anv_cmd_buffer *cmd_buffer)
{
   struct anv_pipeline *pipeline = cmd_buffer->state.gfx.base.pipeline;

   VkShaderStageFlags dirty = cmd_buffer->state.descriptors_dirty &
                              pipeline->active_stages;

   VkResult result = VK_SUCCESS;
   anv_foreach_stage(s, dirty) {
      result = emit_samplers(cmd_buffer, s, &cmd_buffer->state.samplers[s]);
      if (result != VK_SUCCESS)
         break;
      result = emit_binding_table(cmd_buffer, s,
                                  &cmd_buffer->state.binding_tables[s]);
      if (result != VK_SUCCESS)
         break;
   }

   if (result != VK_SUCCESS) {
      assert(result == VK_ERROR_OUT_OF_DEVICE_MEMORY);

      result = anv_cmd_buffer_new_binding_table_block(cmd_buffer);
      if (result != VK_SUCCESS)
         return 0;

      /* Re-emit state base addresses so we get the new surface state base
       * address before we start emitting binding tables etc.
       */
      genX(cmd_buffer_emit_state_base_address)(cmd_buffer);

      /* Re-emit all active binding tables */
      dirty |= pipeline->active_stages;
      anv_foreach_stage(s, dirty) {
         result = emit_samplers(cmd_buffer, s, &cmd_buffer->state.samplers[s]);
         if (result != VK_SUCCESS) {
            anv_batch_set_error(&cmd_buffer->batch, result);
            return 0;
         }
         result = emit_binding_table(cmd_buffer, s,
                                     &cmd_buffer->state.binding_tables[s]);
         if (result != VK_SUCCESS) {
            anv_batch_set_error(&cmd_buffer->batch, result);
            return 0;
         }
      }
   }

   cmd_buffer->state.descriptors_dirty &= ~dirty;

   return dirty;
}

static void
cmd_buffer_emit_descriptor_pointers(struct anv_cmd_buffer *cmd_buffer,
                                    uint32_t stages)
{
   static const uint32_t sampler_state_opcodes[] = {
      [MESA_SHADER_VERTEX]                      = 43,
      [MESA_SHADER_TESS_CTRL]                   = 44, /* HS */
      [MESA_SHADER_TESS_EVAL]                   = 45, /* DS */
      [MESA_SHADER_GEOMETRY]                    = 46,
      [MESA_SHADER_FRAGMENT]                    = 47,
      [MESA_SHADER_COMPUTE]                     = 0,
   };

   static const uint32_t binding_table_opcodes[] = {
      [MESA_SHADER_VERTEX]                      = 38,
      [MESA_SHADER_TESS_CTRL]                   = 39,
      [MESA_SHADER_TESS_EVAL]                   = 40,
      [MESA_SHADER_GEOMETRY]                    = 41,
      [MESA_SHADER_FRAGMENT]                    = 42,
      [MESA_SHADER_COMPUTE]                     = 0,
   };

   anv_foreach_stage(s, stages) {
      assert(s < ARRAY_SIZE(binding_table_opcodes));
      assert(binding_table_opcodes[s] > 0);

      if (cmd_buffer->state.samplers[s].alloc_size > 0) {
         anv_batch_emit(&cmd_buffer->batch,
                        GENX(3DSTATE_SAMPLER_STATE_POINTERS_VS), ssp) {
            ssp._3DCommandSubOpcode = sampler_state_opcodes[s];
            ssp.PointertoVSSamplerState = cmd_buffer->state.samplers[s].offset;
         }
      }

      /* Always emit binding table pointers if we're asked to, since on SKL
       * this is what flushes push constants. */
      anv_batch_emit(&cmd_buffer->batch,
                     GENX(3DSTATE_BINDING_TABLE_POINTERS_VS), btp) {
         btp._3DCommandSubOpcode = binding_table_opcodes[s];
         btp.PointertoVSBindingTable = cmd_buffer->state.binding_tables[s].offset;
      }
   }
}

static void
cmd_buffer_flush_push_constants(struct anv_cmd_buffer *cmd_buffer,
                                VkShaderStageFlags dirty_stages)
{
   const struct anv_cmd_graphics_state *gfx_state = &cmd_buffer->state.gfx;
   const struct anv_pipeline *pipeline = gfx_state->base.pipeline;

   static const uint32_t push_constant_opcodes[] = {
      [MESA_SHADER_VERTEX]                      = 21,
      [MESA_SHADER_TESS_CTRL]                   = 25, /* HS */
      [MESA_SHADER_TESS_EVAL]                   = 26, /* DS */
      [MESA_SHADER_GEOMETRY]                    = 22,
      [MESA_SHADER_FRAGMENT]                    = 23,
      [MESA_SHADER_COMPUTE]                     = 0,
   };

   VkShaderStageFlags flushed = 0;

   anv_foreach_stage(stage, dirty_stages) {
      assert(stage < ARRAY_SIZE(push_constant_opcodes));
      assert(push_constant_opcodes[stage] > 0);

      anv_batch_emit(&cmd_buffer->batch, GENX(3DSTATE_CONSTANT_VS), c) {
         c._3DCommandSubOpcode = push_constant_opcodes[stage];

         if (anv_pipeline_has_stage(pipeline, stage)) {
#if GEN_GEN >= 8 || GEN_IS_HASWELL
            const struct brw_stage_prog_data *prog_data =
               pipeline->shaders[stage]->prog_data;
            const struct anv_pipeline_bind_map *bind_map =
               &pipeline->shaders[stage]->bind_map;

            /* The Skylake PRM contains the following restriction:
             *
             *    "The driver must ensure The following case does not occur
             *     without a flush to the 3D engine: 3DSTATE_CONSTANT_* with
             *     buffer 3 read length equal to zero committed followed by a
             *     3DSTATE_CONSTANT_* with buffer 0 read length not equal to
             *     zero committed."
             *
             * To avoid this, we program the buffers in the highest slots.
             * This way, slot 0 is only used if slot 3 is also used.
             */
            int n = 3;

            for (int i = 3; i >= 0; i--) {
               const struct brw_ubo_range *range = &prog_data->ubo_ranges[i];
               if (range->length == 0)
                  continue;

               const unsigned surface =
                  prog_data->binding_table.ubo_start + range->block;

               assert(surface <= bind_map->surface_count);
               const struct anv_pipeline_binding *binding =
                  &bind_map->surface_to_descriptor[surface];

               const struct anv_descriptor *desc =
                  anv_descriptor_for_binding(&gfx_state->base, binding);

               struct anv_address read_addr;
               uint32_t read_len;
               if (desc->type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER) {
                  read_len = MIN2(range->length,
                     DIV_ROUND_UP(desc->buffer_view->range, 32) - range->start);
                  read_addr = anv_address_add(desc->buffer_view->address,
                                              range->start * 32);
               } else {
                  assert(desc->type == VK_DESCRIPTOR_TYPE_UNIFORM_BUFFER_DYNAMIC);

                  uint32_t dynamic_offset =
                     dynamic_offset_for_binding(&gfx_state->base, binding);
                  uint32_t buf_offset =
                     MIN2(desc->offset + dynamic_offset, desc->buffer->size);
                  uint32_t buf_range =
                     MIN2(desc->range, desc->buffer->size - buf_offset);

                  read_len = MIN2(range->length,
                     DIV_ROUND_UP(buf_range, 32) - range->start);
                  read_addr = anv_address_add(desc->buffer->address,
                                              buf_offset + range->start * 32);
               }

               if (read_len > 0) {
                  c.ConstantBody.Buffer[n] = read_addr;
                  c.ConstantBody.ReadLength[n] = read_len;
                  n--;
               }
            }

            struct anv_state state =
               anv_cmd_buffer_push_constants(cmd_buffer, stage);

            if (state.alloc_size > 0) {
               c.ConstantBody.Buffer[n] = (struct anv_address) {
                  .bo = &cmd_buffer->device->dynamic_state_pool.block_pool.bo,
                  .offset = state.offset,
               };
               c.ConstantBody.ReadLength[n] =
                  DIV_ROUND_UP(state.alloc_size, 32);
            }
#else
            /* For Ivy Bridge, the push constants packets have a different
             * rule that would require us to iterate in the other direction
             * and possibly mess around with dynamic state base address.
             * Don't bother; just emit regular push constants at n = 0.
             */
            struct anv_state state =
               anv_cmd_buffer_push_constants(cmd_buffer, stage);

            if (state.alloc_size > 0) {
               c.ConstantBody.Buffer[0].offset = state.offset,
               c.ConstantBody.ReadLength[0] =
                  DIV_ROUND_UP(state.alloc_size, 32);
            }
#endif
         }
      }

      flushed |= mesa_to_vk_shader_stage(stage);
   }

   cmd_buffer->state.push_constants_dirty &= ~flushed;
}

void
genX(cmd_buffer_flush_state)(struct anv_cmd_buffer *cmd_buffer)
{
   struct anv_pipeline *pipeline = cmd_buffer->state.gfx.base.pipeline;
   uint32_t *p;

   uint32_t vb_emit = cmd_buffer->state.gfx.vb_dirty & pipeline->vb_used;

   assert((pipeline->active_stages & VK_SHADER_STAGE_COMPUTE_BIT) == 0);

   genX(cmd_buffer_config_l3)(cmd_buffer, pipeline->urb.l3_config);

   genX(flush_pipeline_select_3d)(cmd_buffer);

   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,
                          GENX(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 GENX(VERTEX_BUFFER_STATE) state = {
            .VertexBufferIndex = vb,

#if GEN_GEN >= 8
            .MemoryObjectControlState = GENX(MOCS),
#else
            .BufferAccessType = pipeline->instancing_enable[vb] ? INSTANCEDATA : VERTEXDATA,
            /* Our implementation of VK_KHR_multiview uses instancing to draw
             * the different views.  If the client asks for instancing, we
             * need to use the Instance Data Step Rate to ensure that we
             * repeat the client's per-instance data once for each view.
             */
            .InstanceDataStepRate = anv_subpass_view_count(pipeline->subpass),
            .VertexBufferMemoryObjectControlState = GENX(MOCS),
#endif

            .AddressModifyEnable = true,
            .BufferPitch = pipeline->binding_stride[vb],
            .BufferStartingAddress = anv_address_add(buffer->address, offset),

#if GEN_GEN >= 8
            .BufferSize = buffer->size - offset
#else
            .EndAddress = anv_address_add(buffer->address, buffer->size - 1),
#endif
         };

         GENX(VERTEX_BUFFER_STATE_pack)(&cmd_buffer->batch, &p[1 + i * 4], &state);
         i++;
      }
   }

   cmd_buffer->state.gfx.vb_dirty &= ~vb_emit;

   if (cmd_buffer->state.gfx.dirty & ANV_CMD_DIRTY_PIPELINE) {
      anv_batch_emit_batch(&cmd_buffer->batch, &pipeline->batch);

      /* The exact descriptor layout is pulled from the pipeline, so we need
       * to re-emit binding tables on every pipeline change.
       */
      cmd_buffer->state.descriptors_dirty |= pipeline->active_stages;

      /* If the pipeline changed, we may need to re-allocate push constant
       * space in the URB.
       */
      cmd_buffer_alloc_push_constants(cmd_buffer);
   }

#if GEN_GEN <= 7
   if (cmd_buffer->state.descriptors_dirty & VK_SHADER_STAGE_VERTEX_BIT ||
       cmd_buffer->state.push_constants_dirty & VK_SHADER_STAGE_VERTEX_BIT) {
      /* From the IVB PRM Vol. 2, Part 1, Section 3.2.1:
       *
       *    "A PIPE_CONTROL with Post-Sync Operation set to 1h and a depth
       *    stall needs to be sent just prior to any 3DSTATE_VS,
       *    3DSTATE_URB_VS, 3DSTATE_CONSTANT_VS,
       *    3DSTATE_BINDING_TABLE_POINTER_VS,
       *    3DSTATE_SAMPLER_STATE_POINTER_VS command.  Only one
       *    PIPE_CONTROL needs to be sent before any combination of VS
       *    associated 3DSTATE."
       */
      anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
         pc.DepthStallEnable  = true;
         pc.PostSyncOperation = WriteImmediateData;
         pc.Address           =
            (struct anv_address) { &cmd_buffer->device->workaround_bo, 0 };
      }
   }
#endif

   /* Render targets live in the same binding table as fragment descriptors */
   if (cmd_buffer->state.gfx.dirty & ANV_CMD_DIRTY_RENDER_TARGETS)
      cmd_buffer->state.descriptors_dirty |= VK_SHADER_STAGE_FRAGMENT_BIT;

   /* We emit the binding tables and sampler tables first, then emit push
    * constants and then finally emit binding table and sampler table
    * pointers.  It has to happen in this order, since emitting the binding
    * tables may change the push constants (in case of storage images). After
    * emitting push constants, on SKL+ we have to emit the corresponding
    * 3DSTATE_BINDING_TABLE_POINTER_* for the push constants to take effect.
    */
   uint32_t dirty = 0;
   if (cmd_buffer->state.descriptors_dirty)
      dirty = flush_descriptor_sets(cmd_buffer);

   if (dirty || cmd_buffer->state.push_constants_dirty) {
      /* Because we're pushing UBOs, we have to push whenever either
       * descriptors or push constants is dirty.
       */
      dirty |= cmd_buffer->state.push_constants_dirty;
      dirty &= ANV_STAGE_MASK & VK_SHADER_STAGE_ALL_GRAPHICS;
      cmd_buffer_flush_push_constants(cmd_buffer, dirty);
   }

   if (dirty)
      cmd_buffer_emit_descriptor_pointers(cmd_buffer, dirty);

   if (cmd_buffer->state.gfx.dirty & ANV_CMD_DIRTY_DYNAMIC_VIEWPORT)
      gen8_cmd_buffer_emit_viewport(cmd_buffer);

   if (cmd_buffer->state.gfx.dirty & (ANV_CMD_DIRTY_DYNAMIC_VIEWPORT |
                                  ANV_CMD_DIRTY_PIPELINE)) {
      gen8_cmd_buffer_emit_depth_viewport(cmd_buffer,
                                          pipeline->depth_clamp_enable);
   }

   if (cmd_buffer->state.gfx.dirty & ANV_CMD_DIRTY_DYNAMIC_SCISSOR)
      gen7_cmd_buffer_emit_scissor(cmd_buffer);

   genX(cmd_buffer_flush_dynamic_state)(cmd_buffer);

   genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
}

static void
emit_vertex_bo(struct anv_cmd_buffer *cmd_buffer,
               struct anv_address addr,
               uint32_t size, uint32_t index)
{
   uint32_t *p = anv_batch_emitn(&cmd_buffer->batch, 5,
                                 GENX(3DSTATE_VERTEX_BUFFERS));

   GENX(VERTEX_BUFFER_STATE_pack)(&cmd_buffer->batch, p + 1,
      &(struct GENX(VERTEX_BUFFER_STATE)) {
         .VertexBufferIndex = index,
         .AddressModifyEnable = true,
         .BufferPitch = 0,
#if (GEN_GEN >= 8)
         .MemoryObjectControlState = GENX(MOCS),
         .BufferStartingAddress = addr,
         .BufferSize = size
#else
         .VertexBufferMemoryObjectControlState = GENX(MOCS),
         .BufferStartingAddress = addr,
         .EndAddress = anv_address_add(addr, size),
#endif
      });
}

static void
emit_base_vertex_instance_bo(struct anv_cmd_buffer *cmd_buffer,
                             struct anv_address addr)
{
   emit_vertex_bo(cmd_buffer, addr, 8, ANV_SVGS_VB_INDEX);
}

static void
emit_base_vertex_instance(struct anv_cmd_buffer *cmd_buffer,
                          uint32_t base_vertex, uint32_t base_instance)
{
   struct anv_state id_state =
      anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, 8, 4);

   ((uint32_t *)id_state.map)[0] = base_vertex;
   ((uint32_t *)id_state.map)[1] = base_instance;

   anv_state_flush(cmd_buffer->device, id_state);

   struct anv_address addr = {
      .bo = &cmd_buffer->device->dynamic_state_pool.block_pool.bo,
      .offset = id_state.offset,
   };

   emit_base_vertex_instance_bo(cmd_buffer, addr);
}

static void
emit_draw_index(struct anv_cmd_buffer *cmd_buffer, uint32_t draw_index)
{
   struct anv_state state =
      anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, 4, 4);

   ((uint32_t *)state.map)[0] = draw_index;

   anv_state_flush(cmd_buffer->device, state);

   struct anv_address addr = {
      .bo = &cmd_buffer->device->dynamic_state_pool.block_pool.bo,
      .offset = state.offset,
   };

   emit_vertex_bo(cmd_buffer, addr, 4, ANV_DRAWID_VB_INDEX);
}

void genX(CmdDraw)(
    VkCommandBuffer                             commandBuffer,
    uint32_t                                    vertexCount,
    uint32_t                                    instanceCount,
    uint32_t                                    firstVertex,
    uint32_t                                    firstInstance)
{
   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
   struct anv_pipeline *pipeline = cmd_buffer->state.gfx.base.pipeline;
   const struct brw_vs_prog_data *vs_prog_data = get_vs_prog_data(pipeline);

   if (anv_batch_has_error(&cmd_buffer->batch))
      return;

   genX(cmd_buffer_flush_state)(cmd_buffer);

   if (vs_prog_data->uses_firstvertex ||
       vs_prog_data->uses_baseinstance)
      emit_base_vertex_instance(cmd_buffer, firstVertex, firstInstance);
   if (vs_prog_data->uses_drawid)
      emit_draw_index(cmd_buffer, 0);

   /* Our implementation of VK_KHR_multiview uses instancing to draw the
    * different views.  We need to multiply instanceCount by the view count.
    */
   instanceCount *= anv_subpass_view_count(cmd_buffer->state.subpass);

   anv_batch_emit(&cmd_buffer->batch, GENX(3DPRIMITIVE), prim) {
      prim.VertexAccessType         = SEQUENTIAL;
      prim.PrimitiveTopologyType    = pipeline->topology;
      prim.VertexCountPerInstance   = vertexCount;
      prim.StartVertexLocation      = firstVertex;
      prim.InstanceCount            = instanceCount;
      prim.StartInstanceLocation    = firstInstance;
      prim.BaseVertexLocation       = 0;
   }
}

void genX(CmdDrawIndexed)(
    VkCommandBuffer                             commandBuffer,
    uint32_t                                    indexCount,
    uint32_t                                    instanceCount,
    uint32_t                                    firstIndex,
    int32_t                                     vertexOffset,
    uint32_t                                    firstInstance)
{
   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
   struct anv_pipeline *pipeline = cmd_buffer->state.gfx.base.pipeline;
   const struct brw_vs_prog_data *vs_prog_data = get_vs_prog_data(pipeline);

   if (anv_batch_has_error(&cmd_buffer->batch))
      return;

   genX(cmd_buffer_flush_state)(cmd_buffer);

   if (vs_prog_data->uses_firstvertex ||
       vs_prog_data->uses_baseinstance)
      emit_base_vertex_instance(cmd_buffer, vertexOffset, firstInstance);
   if (vs_prog_data->uses_drawid)
      emit_draw_index(cmd_buffer, 0);

   /* Our implementation of VK_KHR_multiview uses instancing to draw the
    * different views.  We need to multiply instanceCount by the view count.
    */
   instanceCount *= anv_subpass_view_count(cmd_buffer->state.subpass);

   anv_batch_emit(&cmd_buffer->batch, GENX(3DPRIMITIVE), prim) {
      prim.VertexAccessType         = RANDOM;
      prim.PrimitiveTopologyType    = pipeline->topology;
      prim.VertexCountPerInstance   = indexCount;
      prim.StartVertexLocation      = firstIndex;
      prim.InstanceCount            = instanceCount;
      prim.StartInstanceLocation    = firstInstance;
      prim.BaseVertexLocation       = vertexOffset;
   }
}

/* 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

/* MI_MATH only exists on Haswell+ */
#if GEN_IS_HASWELL || GEN_GEN >= 8

/* Emit dwords to multiply GPR0 by N */
static void
build_alu_multiply_gpr0(uint32_t *dw, unsigned *dw_count, uint32_t N)
{
   VK_OUTARRAY_MAKE(out, dw, dw_count);

#define append_alu(opcode, operand1, operand2) \
   vk_outarray_append(&out, alu_dw) *alu_dw = mi_alu(opcode, operand1, operand2)

   assert(N > 0);
   unsigned top_bit = 31 - __builtin_clz(N);
   for (int i = top_bit - 1; i >= 0; i--) {
      /* We get our initial data in GPR0 and we write the final data out to
       * GPR0 but we use GPR1 as our scratch register.
       */
      unsigned src_reg = i == top_bit - 1 ? MI_ALU_REG0 : MI_ALU_REG1;
      unsigned dst_reg = i == 0 ? MI_ALU_REG0 : MI_ALU_REG1;

      /* Shift the current value left by 1 */
      append_alu(MI_ALU_LOAD, MI_ALU_SRCA, src_reg);
      append_alu(MI_ALU_LOAD, MI_ALU_SRCB, src_reg);
      append_alu(MI_ALU_ADD, 0, 0);

      if (N & (1 << i)) {
         /* Store ACCU to R1 and add R0 to R1 */
         append_alu(MI_ALU_STORE, MI_ALU_REG1, MI_ALU_ACCU);
         append_alu(MI_ALU_LOAD, MI_ALU_SRCA, MI_ALU_REG0);
         append_alu(MI_ALU_LOAD, MI_ALU_SRCB, MI_ALU_REG1);
         append_alu(MI_ALU_ADD, 0, 0);
      }

      append_alu(MI_ALU_STORE, dst_reg, MI_ALU_ACCU);
   }

#undef append_alu
}

static void
emit_mul_gpr0(struct anv_batch *batch, uint32_t N)
{
   uint32_t num_dwords;
   build_alu_multiply_gpr0(NULL, &num_dwords, N);

   uint32_t *dw = anv_batch_emitn(batch, 1 + num_dwords, GENX(MI_MATH));
   build_alu_multiply_gpr0(dw + 1, &num_dwords, N);
}

#endif /* GEN_IS_HASWELL || GEN_GEN >= 8 */

static void
load_indirect_parameters(struct anv_cmd_buffer *cmd_buffer,
                         struct anv_address addr,
                         bool indexed)
{
   struct anv_batch *batch = &cmd_buffer->batch;

   emit_lrm(batch, GEN7_3DPRIM_VERTEX_COUNT, addr.bo, addr.offset);

   unsigned view_count = anv_subpass_view_count(cmd_buffer->state.subpass);
   if (view_count > 1) {
#if GEN_IS_HASWELL || GEN_GEN >= 8
      emit_lrm(batch, CS_GPR(0), addr.bo, addr.offset + 4);
      emit_mul_gpr0(batch, view_count);
      emit_lrr(batch, GEN7_3DPRIM_INSTANCE_COUNT, CS_GPR(0));
#else
      anv_finishme("Multiview + indirect draw requires MI_MATH; "
                   "MI_MATH is not supported on Ivy Bridge");
      emit_lrm(batch, GEN7_3DPRIM_INSTANCE_COUNT, addr.bo, addr.offset + 4);
#endif
   } else {
      emit_lrm(batch, GEN7_3DPRIM_INSTANCE_COUNT, addr.bo, addr.offset + 4);
   }

   emit_lrm(batch, GEN7_3DPRIM_START_VERTEX, addr.bo, addr.offset + 8);

   if (indexed) {
      emit_lrm(batch, GEN7_3DPRIM_BASE_VERTEX, addr.bo, addr.offset + 12);
      emit_lrm(batch, GEN7_3DPRIM_START_INSTANCE, addr.bo, addr.offset + 16);
   } else {
      emit_lrm(batch, GEN7_3DPRIM_START_INSTANCE, addr.bo, addr.offset + 12);
      emit_lri(batch, GEN7_3DPRIM_BASE_VERTEX, 0);
   }
}

void genX(CmdDrawIndirect)(
    VkCommandBuffer                             commandBuffer,
    VkBuffer                                    _buffer,
    VkDeviceSize                                offset,
    uint32_t                                    drawCount,
    uint32_t                                    stride)
{
   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
   ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
   struct anv_pipeline *pipeline = cmd_buffer->state.gfx.base.pipeline;
   const struct brw_vs_prog_data *vs_prog_data = get_vs_prog_data(pipeline);

   if (anv_batch_has_error(&cmd_buffer->batch))
      return;

   genX(cmd_buffer_flush_state)(cmd_buffer);

   for (uint32_t i = 0; i < drawCount; i++) {
      struct anv_address draw = anv_address_add(buffer->address, offset);

      if (vs_prog_data->uses_firstvertex ||
          vs_prog_data->uses_baseinstance)
         emit_base_vertex_instance_bo(cmd_buffer, anv_address_add(draw, 8));
      if (vs_prog_data->uses_drawid)
         emit_draw_index(cmd_buffer, i);

      load_indirect_parameters(cmd_buffer, draw, false);

      anv_batch_emit(&cmd_buffer->batch, GENX(3DPRIMITIVE), prim) {
         prim.IndirectParameterEnable  = true;
         prim.VertexAccessType         = SEQUENTIAL;
         prim.PrimitiveTopologyType    = pipeline->topology;
      }

      offset += stride;
   }
}

void genX(CmdDrawIndexedIndirect)(
    VkCommandBuffer                             commandBuffer,
    VkBuffer                                    _buffer,
    VkDeviceSize                                offset,
    uint32_t                                    drawCount,
    uint32_t                                    stride)
{
   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
   ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
   struct anv_pipeline *pipeline = cmd_buffer->state.gfx.base.pipeline;
   const struct brw_vs_prog_data *vs_prog_data = get_vs_prog_data(pipeline);

   if (anv_batch_has_error(&cmd_buffer->batch))
      return;

   genX(cmd_buffer_flush_state)(cmd_buffer);

   for (uint32_t i = 0; i < drawCount; i++) {
      struct anv_address draw = anv_address_add(buffer->address, offset);

      /* TODO: We need to stomp base vertex to 0 somehow */
      if (vs_prog_data->uses_firstvertex ||
          vs_prog_data->uses_baseinstance)
         emit_base_vertex_instance_bo(cmd_buffer, anv_address_add(draw, 12));
      if (vs_prog_data->uses_drawid)
         emit_draw_index(cmd_buffer, i);

      load_indirect_parameters(cmd_buffer, draw, true);

      anv_batch_emit(&cmd_buffer->batch, GENX(3DPRIMITIVE), prim) {
         prim.IndirectParameterEnable  = true;
         prim.VertexAccessType         = RANDOM;
         prim.PrimitiveTopologyType    = pipeline->topology;
      }

      offset += stride;
   }
}

static VkResult
flush_compute_descriptor_set(struct anv_cmd_buffer *cmd_buffer)
{
   struct anv_pipeline *pipeline = cmd_buffer->state.compute.base.pipeline;
   struct anv_state surfaces = { 0, }, samplers = { 0, };
   VkResult result;

   result = emit_binding_table(cmd_buffer, MESA_SHADER_COMPUTE, &surfaces);
   if (result != VK_SUCCESS) {
      assert(result == VK_ERROR_OUT_OF_DEVICE_MEMORY);

      result = anv_cmd_buffer_new_binding_table_block(cmd_buffer);
      if (result != VK_SUCCESS)
         return result;

      /* Re-emit state base addresses so we get the new surface state base
       * address before we start emitting binding tables etc.
       */
      genX(cmd_buffer_emit_state_base_address)(cmd_buffer);

      result = emit_binding_table(cmd_buffer, MESA_SHADER_COMPUTE, &surfaces);
      if (result != VK_SUCCESS) {
         anv_batch_set_error(&cmd_buffer->batch, result);
         return result;
      }
   }

   result = emit_samplers(cmd_buffer, MESA_SHADER_COMPUTE, &samplers);
   if (result != VK_SUCCESS) {
      anv_batch_set_error(&cmd_buffer->batch, result);
      return result;
   }

   uint32_t iface_desc_data_dw[GENX(INTERFACE_DESCRIPTOR_DATA_length)];
   struct GENX(INTERFACE_DESCRIPTOR_DATA) desc = {
      .BindingTablePointer = surfaces.offset,
      .SamplerStatePointer = samplers.offset,
   };
   GENX(INTERFACE_DESCRIPTOR_DATA_pack)(NULL, iface_desc_data_dw, &desc);

   struct anv_state state =
      anv_cmd_buffer_merge_dynamic(cmd_buffer, iface_desc_data_dw,
                                   pipeline->interface_descriptor_data,
                                   GENX(INTERFACE_DESCRIPTOR_DATA_length),
                                   64);

   uint32_t size = GENX(INTERFACE_DESCRIPTOR_DATA_length) * sizeof(uint32_t);
   anv_batch_emit(&cmd_buffer->batch,
                  GENX(MEDIA_INTERFACE_DESCRIPTOR_LOAD), mid) {
      mid.InterfaceDescriptorTotalLength        = size;
      mid.InterfaceDescriptorDataStartAddress   = state.offset;
   }

   return VK_SUCCESS;
}

void
genX(cmd_buffer_flush_compute_state)(struct anv_cmd_buffer *cmd_buffer)
{
   struct anv_pipeline *pipeline = cmd_buffer->state.compute.base.pipeline;
   MAYBE_UNUSED VkResult result;

   assert(pipeline->active_stages == VK_SHADER_STAGE_COMPUTE_BIT);

   genX(cmd_buffer_config_l3)(cmd_buffer, pipeline->urb.l3_config);

   genX(flush_pipeline_select_gpgpu)(cmd_buffer);

   if (cmd_buffer->state.compute.pipeline_dirty) {
      /* From the Sky Lake PRM Vol 2a, MEDIA_VFE_STATE:
       *
       *    "A stalling PIPE_CONTROL is required before MEDIA_VFE_STATE unless
       *    the only bits that are changed are scoreboard related: Scoreboard
       *    Enable, Scoreboard Type, Scoreboard Mask, Scoreboard * Delta. For
       *    these scoreboard related states, a MEDIA_STATE_FLUSH is
       *    sufficient."
       */
      cmd_buffer->state.pending_pipe_bits |= ANV_PIPE_CS_STALL_BIT;
      genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);

      anv_batch_emit_batch(&cmd_buffer->batch, &pipeline->batch);
   }

   if ((cmd_buffer->state.descriptors_dirty & VK_SHADER_STAGE_COMPUTE_BIT) ||
       cmd_buffer->state.compute.pipeline_dirty) {
      /* FIXME: figure out descriptors for gen7 */
      result = flush_compute_descriptor_set(cmd_buffer);
      if (result != VK_SUCCESS)
         return;

      cmd_buffer->state.descriptors_dirty &= ~VK_SHADER_STAGE_COMPUTE_BIT;
   }

   if (cmd_buffer->state.push_constants_dirty & VK_SHADER_STAGE_COMPUTE_BIT) {
      struct anv_state push_state =
         anv_cmd_buffer_cs_push_constants(cmd_buffer);

      if (push_state.alloc_size) {
         anv_batch_emit(&cmd_buffer->batch, GENX(MEDIA_CURBE_LOAD), curbe) {
            curbe.CURBETotalDataLength    = push_state.alloc_size;
            curbe.CURBEDataStartAddress   = push_state.offset;
         }
      }
   }

   cmd_buffer->state.compute.pipeline_dirty = false;

   genX(cmd_buffer_apply_pipe_flushes)(cmd_buffer);
}

#if GEN_GEN == 7

static VkResult
verify_cmd_parser(const struct anv_device *device,
                  int required_version,
                  const char *function)
{
   if (device->instance->physicalDevice.cmd_parser_version < required_version) {
      return vk_errorf(device->instance, device->instance,
                       VK_ERROR_FEATURE_NOT_PRESENT,
                       "cmd parser version %d is required for %s",
                       required_version, function);
   } else {
      return VK_SUCCESS;
   }
}

#endif

static void
anv_cmd_buffer_push_base_group_id(struct anv_cmd_buffer *cmd_buffer,
                                  uint32_t baseGroupX,
                                  uint32_t baseGroupY,
                                  uint32_t baseGroupZ)
{
   if (anv_batch_has_error(&cmd_buffer->batch))
      return;

   VkResult result =
      anv_cmd_buffer_ensure_push_constant_field(cmd_buffer, MESA_SHADER_COMPUTE,
                                                base_work_group_id);
   if (result != VK_SUCCESS) {
      cmd_buffer->batch.status = result;
      return;
   }

   struct anv_push_constants *push =
      cmd_buffer->state.push_constants[MESA_SHADER_COMPUTE];
   if (push->base_work_group_id[0] != baseGroupX ||
       push->base_work_group_id[1] != baseGroupY ||
       push->base_work_group_id[2] != baseGroupZ) {
      push->base_work_group_id[0] = baseGroupX;
      push->base_work_group_id[1] = baseGroupY;
      push->base_work_group_id[2] = baseGroupZ;

      cmd_buffer->state.push_constants_dirty |= VK_SHADER_STAGE_COMPUTE_BIT;
   }
}

void genX(CmdDispatch)(
    VkCommandBuffer                             commandBuffer,
    uint32_t                                    x,
    uint32_t                                    y,
    uint32_t                                    z)
{
   genX(CmdDispatchBase)(commandBuffer, 0, 0, 0, x, y, z);
}

void genX(CmdDispatchBase)(
    VkCommandBuffer                             commandBuffer,
    uint32_t                                    baseGroupX,
    uint32_t                                    baseGroupY,
    uint32_t                                    baseGroupZ,
    uint32_t                                    groupCountX,
    uint32_t                                    groupCountY,
    uint32_t                                    groupCountZ)
{
   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
   struct anv_pipeline *pipeline = cmd_buffer->state.compute.base.pipeline;
   const struct brw_cs_prog_data *prog_data = get_cs_prog_data(pipeline);

   anv_cmd_buffer_push_base_group_id(cmd_buffer, baseGroupX,
                                     baseGroupY, baseGroupZ);

   if (anv_batch_has_error(&cmd_buffer->batch))
      return;

   if (prog_data->uses_num_work_groups) {
      struct anv_state state =
         anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, 12, 4);
      uint32_t *sizes = state.map;
      sizes[0] = groupCountX;
      sizes[1] = groupCountY;
      sizes[2] = groupCountZ;
      anv_state_flush(cmd_buffer->device, state);
      cmd_buffer->state.compute.num_workgroups = (struct anv_address) {
         .bo = &cmd_buffer->device->dynamic_state_pool.block_pool.bo,
         .offset = state.offset,
      };
   }

   genX(cmd_buffer_flush_compute_state)(cmd_buffer);

   anv_batch_emit(&cmd_buffer->batch, GENX(GPGPU_WALKER), ggw) {
      ggw.SIMDSize                     = prog_data->simd_size / 16;
      ggw.ThreadDepthCounterMaximum    = 0;
      ggw.ThreadHeightCounterMaximum   = 0;
      ggw.ThreadWidthCounterMaximum    = prog_data->threads - 1;
      ggw.ThreadGroupIDXDimension      = groupCountX;
      ggw.ThreadGroupIDYDimension      = groupCountY;
      ggw.ThreadGroupIDZDimension      = groupCountZ;
      ggw.RightExecutionMask           = pipeline->cs_right_mask;
      ggw.BottomExecutionMask          = 0xffffffff;
   }

   anv_batch_emit(&cmd_buffer->batch, GENX(MEDIA_STATE_FLUSH), msf);
}

#define GPGPU_DISPATCHDIMX 0x2500
#define GPGPU_DISPATCHDIMY 0x2504
#define GPGPU_DISPATCHDIMZ 0x2508

void genX(CmdDispatchIndirect)(
    VkCommandBuffer                             commandBuffer,
    VkBuffer                                    _buffer,
    VkDeviceSize                                offset)
{
   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
   ANV_FROM_HANDLE(anv_buffer, buffer, _buffer);
   struct anv_pipeline *pipeline = cmd_buffer->state.compute.base.pipeline;
   const struct brw_cs_prog_data *prog_data = get_cs_prog_data(pipeline);
   struct anv_address addr = anv_address_add(buffer->address, offset);
   struct anv_batch *batch = &cmd_buffer->batch;

   anv_cmd_buffer_push_base_group_id(cmd_buffer, 0, 0, 0);

#if GEN_GEN == 7
   /* Linux 4.4 added command parser version 5 which allows the GPGPU
    * indirect dispatch registers to be written.
    */
   if (verify_cmd_parser(cmd_buffer->device, 5,
                         "vkCmdDispatchIndirect") != VK_SUCCESS)
      return;
#endif

   if (prog_data->uses_num_work_groups)
      cmd_buffer->state.compute.num_workgroups = addr;

   genX(cmd_buffer_flush_compute_state)(cmd_buffer);

   emit_lrm(batch, GPGPU_DISPATCHDIMX, addr.bo, addr.offset);
   emit_lrm(batch, GPGPU_DISPATCHDIMY, addr.bo, addr.offset + 4);
   emit_lrm(batch, GPGPU_DISPATCHDIMZ, addr.bo, addr.offset + 8);

#if GEN_GEN <= 7
   /* Clear upper 32-bits of SRC0 and all 64-bits of SRC1 */
   emit_lri(batch, MI_PREDICATE_SRC0 + 4, 0);
   emit_lri(batch, MI_PREDICATE_SRC1 + 0, 0);
   emit_lri(batch, MI_PREDICATE_SRC1 + 4, 0);

   /* Load compute_dispatch_indirect_x_size into SRC0 */
   emit_lrm(batch, MI_PREDICATE_SRC0, addr.bo, addr.offset + 0);

   /* predicate = (compute_dispatch_indirect_x_size == 0); */
   anv_batch_emit(batch, GENX(MI_PREDICATE), mip) {
      mip.LoadOperation    = LOAD_LOAD;
      mip.CombineOperation = COMBINE_SET;
      mip.CompareOperation = COMPARE_SRCS_EQUAL;
   }

   /* Load compute_dispatch_indirect_y_size into SRC0 */
   emit_lrm(batch, MI_PREDICATE_SRC0, addr.bo, addr.offset + 4);

   /* predicate |= (compute_dispatch_indirect_y_size == 0); */
   anv_batch_emit(batch, GENX(MI_PREDICATE), mip) {
      mip.LoadOperation    = LOAD_LOAD;
      mip.CombineOperation = COMBINE_OR;
      mip.CompareOperation = COMPARE_SRCS_EQUAL;
   }

   /* Load compute_dispatch_indirect_z_size into SRC0 */
   emit_lrm(batch, MI_PREDICATE_SRC0, addr.bo, addr.offset + 8);

   /* predicate |= (compute_dispatch_indirect_z_size == 0); */
   anv_batch_emit(batch, GENX(MI_PREDICATE), mip) {
      mip.LoadOperation    = LOAD_LOAD;
      mip.CombineOperation = COMBINE_OR;
      mip.CompareOperation = COMPARE_SRCS_EQUAL;
   }

   /* predicate = !predicate; */
#define COMPARE_FALSE                           1
   anv_batch_emit(batch, GENX(MI_PREDICATE), mip) {
      mip.LoadOperation    = LOAD_LOADINV;
      mip.CombineOperation = COMBINE_OR;
      mip.CompareOperation = COMPARE_FALSE;
   }
#endif

   anv_batch_emit(batch, GENX(GPGPU_WALKER), ggw) {
      ggw.IndirectParameterEnable      = true;
      ggw.PredicateEnable              = GEN_GEN <= 7;
      ggw.SIMDSize                     = prog_data->simd_size / 16;
      ggw.ThreadDepthCounterMaximum    = 0;
      ggw.ThreadHeightCounterMaximum   = 0;
      ggw.ThreadWidthCounterMaximum    = prog_data->threads - 1;
      ggw.RightExecutionMask           = pipeline->cs_right_mask;
      ggw.BottomExecutionMask          = 0xffffffff;
   }

   anv_batch_emit(batch, GENX(MEDIA_STATE_FLUSH), msf);
}

static void
genX(flush_pipeline_select)(struct anv_cmd_buffer *cmd_buffer,
                            uint32_t pipeline)
{
   UNUSED const struct gen_device_info *devinfo = &cmd_buffer->device->info;

   if (cmd_buffer->state.current_pipeline == pipeline)
      return;

#if GEN_GEN >= 8 && GEN_GEN < 10
   /* From the Broadwell PRM, Volume 2a: Instructions, PIPELINE_SELECT:
    *
    *   Software must clear the COLOR_CALC_STATE Valid field in
    *   3DSTATE_CC_STATE_POINTERS command prior to send a PIPELINE_SELECT
    *   with Pipeline Select set to GPGPU.
    *
    * The internal hardware docs recommend the same workaround for Gen9
    * hardware too.
    */
   if (pipeline == GPGPU)
      anv_batch_emit(&cmd_buffer->batch, GENX(3DSTATE_CC_STATE_POINTERS), t);
#endif

   /* From "BXML » GT » MI » vol1a GPU Overview » [Instruction]
    * PIPELINE_SELECT [DevBWR+]":
    *
    *   Project: DEVSNB+
    *
    *   Software must ensure all the write caches are flushed through a
    *   stalling PIPE_CONTROL command followed by another PIPE_CONTROL
    *   command to invalidate read only caches prior to programming
    *   MI_PIPELINE_SELECT command to change the Pipeline Select Mode.
    */
   anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
      pc.RenderTargetCacheFlushEnable  = true;
      pc.DepthCacheFlushEnable         = true;
      pc.DCFlushEnable                 = true;
      pc.PostSyncOperation             = NoWrite;
      pc.CommandStreamerStallEnable    = true;
   }

   anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pc) {
      pc.TextureCacheInvalidationEnable   = true;
      pc.ConstantCacheInvalidationEnable  = true;
      pc.StateCacheInvalidationEnable     = true;
      pc.InstructionCacheInvalidateEnable = true;
      pc.PostSyncOperation                = NoWrite;
   }

   anv_batch_emit(&cmd_buffer->batch, GENX(PIPELINE_SELECT), ps) {
#if GEN_GEN >= 9
      ps.MaskBits = 3;
#endif
      ps.PipelineSelection = pipeline;
   }

#if GEN_GEN == 9
   if (devinfo->is_geminilake) {
      /* Project: DevGLK
       *
       * "This chicken bit works around a hardware issue with barrier logic
       *  encountered when switching between GPGPU and 3D pipelines.  To
       *  workaround the issue, this mode bit should be set after a pipeline
       *  is selected."
       */
      uint32_t scec;
      anv_pack_struct(&scec, GENX(SLICE_COMMON_ECO_CHICKEN1),
                      .GLKBarrierMode =
                          pipeline == GPGPU ? GLK_BARRIER_MODE_GPGPU
                                            : GLK_BARRIER_MODE_3D_HULL,
                      .GLKBarrierModeMask = 1);
      emit_lri(&cmd_buffer->batch, GENX(SLICE_COMMON_ECO_CHICKEN1_num), scec);
   }
#endif

   cmd_buffer->state.current_pipeline = pipeline;
}

void
genX(flush_pipeline_select_3d)(struct anv_cmd_buffer *cmd_buffer)
{
   genX(flush_pipeline_select)(cmd_buffer, _3D);
}

void
genX(flush_pipeline_select_gpgpu)(struct anv_cmd_buffer *cmd_buffer)
{
   genX(flush_pipeline_select)(cmd_buffer, GPGPU);
}

void
genX(cmd_buffer_emit_gen7_depth_flush)(struct anv_cmd_buffer *cmd_buffer)
{
   if (GEN_GEN >= 8)
      return;

   /* From the Haswell PRM, documentation for 3DSTATE_DEPTH_BUFFER:
    *
    *    "Restriction: Prior to changing Depth/Stencil Buffer state (i.e., any
    *    combination of 3DSTATE_DEPTH_BUFFER, 3DSTATE_CLEAR_PARAMS,
    *    3DSTATE_STENCIL_BUFFER, 3DSTATE_HIER_DEPTH_BUFFER) SW must first
    *    issue a pipelined depth stall (PIPE_CONTROL with Depth Stall bit
    *    set), followed by a pipelined depth cache flush (PIPE_CONTROL with
    *    Depth Flush Bit set, followed by another pipelined depth stall
    *    (PIPE_CONTROL with Depth Stall Bit set), unless SW can otherwise
    *    guarantee that the pipeline from WM onwards is already flushed (e.g.,
    *    via a preceding MI_FLUSH)."
    */
   anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pipe) {
      pipe.DepthStallEnable = true;
   }
   anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pipe) {
      pipe.DepthCacheFlushEnable = true;
   }
   anv_batch_emit(&cmd_buffer->batch, GENX(PIPE_CONTROL), pipe) {
      pipe.DepthStallEnable = true;
   }
}

static void
cmd_buffer_emit_depth_stencil(struct anv_cmd_buffer *cmd_buffer)
{
   struct anv_device *device = cmd_buffer->device;
   const struct anv_image_view *iview =
      anv_cmd_buffer_get_depth_stencil_view(cmd_buffer);
   const struct anv_image *image = iview ? iview->image : NULL;

   /* FIXME: Width and Height are wrong */

   genX(cmd_buffer_emit_gen7_depth_flush)(cmd_buffer);

   uint32_t *dw = anv_batch_emit_dwords(&cmd_buffer->batch,
                                        device->isl_dev.ds.size / 4);
   if (dw == NULL)
      return;

   struct isl_depth_stencil_hiz_emit_info info = {
      .mocs = device->default_mocs,
   };

   if (iview)
      info.view = &iview->planes[0].isl;

   if (image && (image->aspects & VK_IMAGE_ASPECT_DEPTH_BIT)) {
      uint32_t depth_plane =
         anv_image_aspect_to_plane(image->aspects, VK_IMAGE_ASPECT_DEPTH_BIT);
      const struct anv_surface *surface = &image->planes[depth_plane].surface;

      info.depth_surf = &surface->isl;

      info.depth_address =
         anv_batch_emit_reloc(&cmd_buffer->batch,
                              dw + device->isl_dev.ds.depth_offset / 4,
                              image->planes[depth_plane].bo,
                              image->planes[depth_plane].bo_offset +
                              surface->offset);

      const uint32_t ds =
         cmd_buffer->state.subpass->depth_stencil_attachment.attachment;
      info.hiz_usage = cmd_buffer->state.attachments[ds].aux_usage;
      if (info.hiz_usage == ISL_AUX_USAGE_HIZ) {
         info.hiz_surf = &image->planes[depth_plane].aux_surface.isl;

         info.hiz_address =
            anv_batch_emit_reloc(&cmd_buffer->batch,
                                 dw + device->isl_dev.ds.hiz_offset / 4,
                                 image->planes[depth_plane].bo,
                                 image->planes[depth_plane].bo_offset +
                                 image->planes[depth_plane].aux_surface.offset);

         info.depth_clear_value = ANV_HZ_FC_VAL;
      }
   }

   if (image && (image->aspects & VK_IMAGE_ASPECT_STENCIL_BIT)) {
      uint32_t stencil_plane =
         anv_image_aspect_to_plane(image->aspects, VK_IMAGE_ASPECT_STENCIL_BIT);
      const struct anv_surface *surface = &image->planes[stencil_plane].surface;

      info.stencil_surf = &surface->isl;

      info.stencil_address =
         anv_batch_emit_reloc(&cmd_buffer->batch,
                              dw + device->isl_dev.ds.stencil_offset / 4,
                              image->planes[stencil_plane].bo,
                              image->planes[stencil_plane].bo_offset + surface->offset);
   }

   isl_emit_depth_stencil_hiz_s(&device->isl_dev, dw, &info);

   cmd_buffer->state.hiz_enabled = info.hiz_usage == ISL_AUX_USAGE_HIZ;
}

/**
 * This ANDs the view mask of the current subpass with the pending clear
 * views in the attachment to get the mask of views active in the subpass
 * that still need to be cleared.
 */
static inline uint32_t
get_multiview_subpass_clear_mask(const struct anv_cmd_state *cmd_state,
                                 const struct anv_attachment_state *att_state)
{
   return cmd_state->subpass->view_mask & att_state->pending_clear_views;
}

static inline bool
do_first_layer_clear(const struct anv_cmd_state *cmd_state,
                     const struct anv_attachment_state *att_state)
{
   if (!cmd_state->subpass->view_mask)
      return true;

   uint32_t pending_clear_mask =
      get_multiview_subpass_clear_mask(cmd_state, att_state);

   return pending_clear_mask & 1;
}

static inline bool
current_subpass_is_last_for_attachment(const struct anv_cmd_state *cmd_state,
                                       uint32_t att_idx)
{
   const uint32_t last_subpass_idx =
      cmd_state->pass->attachments[att_idx].last_subpass_idx;
   const struct anv_subpass *last_subpass =
      &cmd_state->pass->subpasses[last_subpass_idx];
   return last_subpass == cmd_state->subpass;
}

static void
cmd_buffer_begin_subpass(struct anv_cmd_buffer *cmd_buffer,
                         uint32_t subpass_id)
{
   struct anv_cmd_state *cmd_state = &cmd_buffer->state;
   struct anv_subpass *subpass = &cmd_state->pass->subpasses[subpass_id];
   cmd_state->subpass = subpass;

   cmd_buffer->state.gfx.dirty |= ANV_CMD_DIRTY_RENDER_TARGETS;

   /* Our implementation of VK_KHR_multiview uses instancing to draw the
    * different views.  If the client asks for instancing, we need to use the
    * Instance Data Step Rate to ensure that we repeat the client's
    * per-instance data once for each view.  Since this bit is in
    * VERTEX_BUFFER_STATE on gen7, we need to dirty vertex buffers at the top
    * of each subpass.
    */
   if (GEN_GEN == 7)
      cmd_buffer->state.gfx.vb_dirty |= ~0;

   /* It is possible to start a render pass with an old pipeline.  Because the
    * render pass and subpass index are both baked into the pipeline, this is
    * highly unlikely.  In order to do so, it requires that you have a render
    * pass with a single subpass and that you use that render pass twice
    * back-to-back and use the same pipeline at the start of the second render
    * pass as at the end of the first.  In order to avoid unpredictable issues
    * with this edge case, we just dirty the pipeline at the start of every
    * subpass.
    */
   cmd_buffer->state.gfx.dirty |= ANV_CMD_DIRTY_PIPELINE;

   /* Accumulate any subpass flushes that need to happen before the subpass */
   cmd_buffer->state.pending_pipe_bits |=
      cmd_buffer->state.pass->subpass_flushes[subpass_id];

   VkRect2D render_area = cmd_buffer->state.render_area;
   struct anv_framebuffer *fb = cmd_buffer->state.framebuffer;

   bool is_multiview = subpass->view_mask != 0;

   for (uint32_t i = 0; i < subpass->attachment_count; ++i) {
      const uint32_t a = subpass->attachments[i].attachment;
      if (a == VK_ATTACHMENT_UNUSED)
         continue;

      assert(a < cmd_state->pass->attachment_count);
      struct anv_attachment_state *att_state = &cmd_state->attachments[a];

      struct anv_image_view *iview = fb->attachments[a];
      const struct anv_image *image = iview->image;

      /* A resolve is necessary before use as an input attachment if the clear
       * color or auxiliary buffer usage isn't supported by the sampler.
       */
      const bool input_needs_resolve =
            (att_state->fast_clear && !att_state->clear_color_is_zero_one) ||
            att_state->input_aux_usage != att_state->aux_usage;

      VkImageLayout target_layout;
      if (iview->aspect_mask & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV &&
          !input_needs_resolve) {
         /* Layout transitions before the final only help to enable sampling
          * as an input attachment. If the input attachment supports sampling
          * using the auxiliary surface, we can skip such transitions by
          * making the target layout one that is CCS-aware.
          */
         target_layout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL;
      } else {
         target_layout = subpass->attachments[i].layout;
      }

      if (image->aspects & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV) {
         assert(image->aspects == VK_IMAGE_ASPECT_COLOR_BIT);

         uint32_t base_layer, layer_count;
         if (image->type == VK_IMAGE_TYPE_3D) {
            base_layer = 0;
            layer_count = anv_minify(iview->image->extent.depth,
                                     iview->planes[0].isl.base_level);
         } else {
            base_layer = iview->planes[0].isl.base_array_layer;
            layer_count = fb->layers;
         }

         transition_color_buffer(cmd_buffer, image, VK_IMAGE_ASPECT_COLOR_BIT,
                                 iview->planes[0].isl.base_level, 1,
                                 base_layer, layer_count,
                                 att_state->current_layout, target_layout);
      } else if (image->aspects & VK_IMAGE_ASPECT_DEPTH_BIT) {
         transition_depth_buffer(cmd_buffer, image,
                                 att_state->current_layout, target_layout);
         att_state->aux_usage =
            anv_layout_to_aux_usage(&cmd_buffer->device->info, image,
                                    VK_IMAGE_ASPECT_DEPTH_BIT, target_layout);
      }
      att_state->current_layout = target_layout;

      if (att_state->pending_clear_aspects & VK_IMAGE_ASPECT_COLOR_BIT) {
         assert(att_state->pending_clear_aspects == VK_IMAGE_ASPECT_COLOR_BIT);

         /* Multi-planar images are not supported as attachments */
         assert(image->aspects == VK_IMAGE_ASPECT_COLOR_BIT);
         assert(image->n_planes == 1);

         uint32_t base_clear_layer = iview->planes[0].isl.base_array_layer;
         uint32_t clear_layer_count = fb->layers;

         if (att_state->fast_clear &&
             do_first_layer_clear(cmd_state, att_state)) {
            /* We only support fast-clears on the first layer */
            assert(iview->planes[0].isl.base_level == 0);
            assert(iview->planes[0].isl.base_array_layer == 0);

            union isl_color_value clear_color = {};
            anv_clear_color_from_att_state(&clear_color, att_state, iview);
            if (iview->image->samples == 1) {
               anv_image_ccs_op(cmd_buffer, image, VK_IMAGE_ASPECT_COLOR_BIT,
                                0, 0, 1, ISL_AUX_OP_FAST_CLEAR,
                                &clear_color,
                                false);
            } else {
               anv_image_mcs_op(cmd_buffer, image, VK_IMAGE_ASPECT_COLOR_BIT,
                                0, 1, ISL_AUX_OP_FAST_CLEAR,
                                &clear_color,
                                false);
            }
            base_clear_layer++;
            clear_layer_count--;
            if (is_multiview)
               att_state->pending_clear_views &= ~1;

            if (att_state->clear_color_is_zero) {
               /* This image has the auxiliary buffer enabled. We can mark the
                * subresource as not needing a resolve because the clear color
                * will match what's in every RENDER_SURFACE_STATE object when
                * it's being used for sampling.
                */
               set_image_fast_clear_state(cmd_buffer, iview->image,
                                          VK_IMAGE_ASPECT_COLOR_BIT,
                                          ANV_FAST_CLEAR_DEFAULT_VALUE);
            } else {
               set_image_fast_clear_state(cmd_buffer, iview->image,
                                          VK_IMAGE_ASPECT_COLOR_BIT,
                                          ANV_FAST_CLEAR_ANY);
            }
         }

         /* From the VkFramebufferCreateInfo spec:
          *
          * "If the render pass uses multiview, then layers must be one and each
          *  attachment requires a number of layers that is greater than the
          *  maximum bit index set in the view mask in the subpasses in which it
          *  is used."
          *
          * So if multiview is active we ignore the number of layers in the
          * framebuffer and instead we honor the view mask from the subpass.
          */
         if (is_multiview) {
            assert(image->n_planes == 1);
            uint32_t pending_clear_mask =
               get_multiview_subpass_clear_mask(cmd_state, att_state);

            uint32_t layer_idx;
            for_each_bit(layer_idx, pending_clear_mask) {
               uint32_t layer =
                  iview->planes[0].isl.base_array_layer + layer_idx;

               anv_image_clear_color(cmd_buffer, image,
                                     VK_IMAGE_ASPECT_COLOR_BIT,
                                     att_state->aux_usage,
                                     iview->planes[0].isl.format,
                                     iview->planes[0].isl.swizzle,
                                     iview->planes[0].isl.base_level,
                                     layer, 1,
                                     render_area,
                                     vk_to_isl_color(att_state->clear_value.color));
            }

            att_state->pending_clear_views &= ~pending_clear_mask;
         } else if (clear_layer_count > 0) {
            assert(image->n_planes == 1);
            anv_image_clear_color(cmd_buffer, image, VK_IMAGE_ASPECT_COLOR_BIT,
                                  att_state->aux_usage,
                                  iview->planes[0].isl.format,
                                  iview->planes[0].isl.swizzle,
                                  iview->planes[0].isl.base_level,
                                  base_clear_layer, clear_layer_count,
                                  render_area,
                                  vk_to_isl_color(att_state->clear_value.color));
         }
      } else if (att_state->pending_clear_aspects & (VK_IMAGE_ASPECT_DEPTH_BIT |
                                                     VK_IMAGE_ASPECT_STENCIL_BIT)) {
         if (att_state->fast_clear && !is_multiview) {
            /* We currently only support HiZ for single-layer images */
            if (att_state->pending_clear_aspects & VK_IMAGE_ASPECT_DEPTH_BIT) {
               assert(iview->image->planes[0].aux_usage == ISL_AUX_USAGE_HIZ);
               assert(iview->planes[0].isl.base_level == 0);
               assert(iview->planes[0].isl.base_array_layer == 0);
               assert(fb->layers == 1);
            }

            anv_image_hiz_clear(cmd_buffer, image,
                                att_state->pending_clear_aspects,
                                iview->planes[0].isl.base_level,
                                iview->planes[0].isl.base_array_layer,
                                fb->layers, render_area,
                                att_state->clear_value.depthStencil.stencil);
         } else if (is_multiview) {
            uint32_t pending_clear_mask =
              get_multiview_subpass_clear_mask(cmd_state, att_state);

            uint32_t layer_idx;
            for_each_bit(layer_idx, pending_clear_mask) {
               uint32_t layer =
                  iview->planes[0].isl.base_array_layer + layer_idx;

               anv_image_clear_depth_stencil(cmd_buffer, image,
                                             att_state->pending_clear_aspects,
                                             att_state->aux_usage,
                                             iview->planes[0].isl.base_level,
                                             layer, 1,
                                             render_area,
                                             att_state->clear_value.depthStencil.depth,
                                             att_state->clear_value.depthStencil.stencil);
            }

            att_state->pending_clear_views &= ~pending_clear_mask;
         } else {
            anv_image_clear_depth_stencil(cmd_buffer, image,
                                          att_state->pending_clear_aspects,
                                          att_state->aux_usage,
                                          iview->planes[0].isl.base_level,
                                          iview->planes[0].isl.base_array_layer,
                                          fb->layers, render_area,
                                          att_state->clear_value.depthStencil.depth,
                                          att_state->clear_value.depthStencil.stencil);
         }
      } else  {
         assert(att_state->pending_clear_aspects == 0);
      }

      if (GEN_GEN < 10 &&
          (att_state->pending_load_aspects & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV) &&
          image->planes[0].aux_surface.isl.size > 0 &&
          iview->planes[0].isl.base_level == 0 &&
          iview->planes[0].isl.base_array_layer == 0) {
         if (att_state->aux_usage != ISL_AUX_USAGE_NONE) {
            genX(copy_fast_clear_dwords)(cmd_buffer, att_state->color.state,
                                         image, VK_IMAGE_ASPECT_COLOR_BIT,
                                         false /* copy to ss */);
         }

         if (need_input_attachment_state(&cmd_state->pass->attachments[a]) &&
             att_state->input_aux_usage != ISL_AUX_USAGE_NONE) {
            genX(copy_fast_clear_dwords)(cmd_buffer, att_state->input.state,
                                         image, VK_IMAGE_ASPECT_COLOR_BIT,
                                         false /* copy to ss */);
         }
      }

      if (subpass->attachments[i].usage ==
          VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT) {
         /* We assume that if we're starting a subpass, we're going to do some
          * rendering so we may end up with compressed data.
          */
         genX(cmd_buffer_mark_image_written)(cmd_buffer, iview->image,
                                             VK_IMAGE_ASPECT_COLOR_BIT,
                                             att_state->aux_usage,
                                             iview->planes[0].isl.base_level,
                                             iview->planes[0].isl.base_array_layer,
                                             fb->layers);
      } else if (subpass->attachments[i].usage ==
                 VK_IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT) {
         /* We may be writing depth or stencil so we need to mark the surface.
          * Unfortunately, there's no way to know at this point whether the
          * depth or stencil tests used will actually write to the surface.
          *
          * Even though stencil may be plane 1, it always shares a base_level
          * with depth.
          */
         const struct isl_view *ds_view = &iview->planes[0].isl;
         if (iview->aspect_mask & VK_IMAGE_ASPECT_DEPTH_BIT) {
            genX(cmd_buffer_mark_image_written)(cmd_buffer, image,
                                                VK_IMAGE_ASPECT_DEPTH_BIT,
                                                att_state->aux_usage,
                                                ds_view->base_level,
                                                ds_view->base_array_layer,
                                                fb->layers);
         }
         if (iview->aspect_mask & VK_IMAGE_ASPECT_STENCIL_BIT) {
            /* Even though stencil may be plane 1, it always shares a
             * base_level with depth.
             */
            genX(cmd_buffer_mark_image_written)(cmd_buffer, image,
                                                VK_IMAGE_ASPECT_STENCIL_BIT,
                                                ISL_AUX_USAGE_NONE,
                                                ds_view->base_level,
                                                ds_view->base_array_layer,
                                                fb->layers);
         }
      }

      /* If multiview is enabled, then we are only done clearing when we no
       * longer have pending layers to clear, or when we have processed the
       * last subpass that uses this attachment.
       */
      if (!is_multiview ||
          att_state->pending_clear_views == 0 ||
          current_subpass_is_last_for_attachment(cmd_state, a)) {
         att_state->pending_clear_aspects = 0;
      }

      att_state->pending_load_aspects = 0;
   }

   cmd_buffer_emit_depth_stencil(cmd_buffer);
}

static void
cmd_buffer_end_subpass(struct anv_cmd_buffer *cmd_buffer)
{
   struct anv_cmd_state *cmd_state = &cmd_buffer->state;
   struct anv_subpass *subpass = cmd_state->subpass;
   uint32_t subpass_id = anv_get_subpass_id(&cmd_buffer->state);

   anv_cmd_buffer_resolve_subpass(cmd_buffer);

   struct anv_framebuffer *fb = cmd_buffer->state.framebuffer;
   for (uint32_t i = 0; i < subpass->attachment_count; ++i) {
      const uint32_t a = subpass->attachments[i].attachment;
      if (a == VK_ATTACHMENT_UNUSED)
         continue;

      if (cmd_state->pass->attachments[a].last_subpass_idx != subpass_id)
         continue;

      assert(a < cmd_state->pass->attachment_count);
      struct anv_attachment_state *att_state = &cmd_state->attachments[a];
      struct anv_image_view *iview = fb->attachments[a];
      const struct anv_image *image = iview->image;

      /* Transition the image into the final layout for this render pass */
      VkImageLayout target_layout =
         cmd_state->pass->attachments[a].final_layout;

      if (image->aspects & VK_IMAGE_ASPECT_ANY_COLOR_BIT_ANV) {
         assert(image->aspects == VK_IMAGE_ASPECT_COLOR_BIT);

         uint32_t base_layer, layer_count;
         if (image->type == VK_IMAGE_TYPE_3D) {
            base_layer = 0;
            layer_count = anv_minify(iview->image->extent.depth,
                                     iview->planes[0].isl.base_level);
         } else {
            base_layer = iview->planes[0].isl.base_array_layer;
            layer_count = fb->layers;
         }

         transition_color_buffer(cmd_buffer, image, VK_IMAGE_ASPECT_COLOR_BIT,
                                 iview->planes[0].isl.base_level, 1,
                                 base_layer, layer_count,
                                 att_state->current_layout, target_layout);
      } else if (image->aspects & VK_IMAGE_ASPECT_DEPTH_BIT) {
         transition_depth_buffer(cmd_buffer, image,
                                 att_state->current_layout, target_layout);
      }
   }

   /* Accumulate any subpass flushes that need to happen after the subpass.
    * Yes, they do get accumulated twice in the NextSubpass case but since
    * genX_CmdNextSubpass just calls end/begin back-to-back, we just end up
    * ORing the bits in twice so it's harmless.
    */
   cmd_buffer->state.pending_pipe_bits |=
      cmd_buffer->state.pass->subpass_flushes[subpass_id + 1];
}

void genX(CmdBeginRenderPass)(
    VkCommandBuffer                             commandBuffer,
    const VkRenderPassBeginInfo*                pRenderPassBegin,
    VkSubpassContents                           contents)
{
   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
   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;
   cmd_buffer->state.render_area = pRenderPassBegin->renderArea;
   VkResult result =
      genX(cmd_buffer_setup_attachments)(cmd_buffer, pass, pRenderPassBegin);

   /* If we failed to setup the attachments we should not try to go further */
   if (result != VK_SUCCESS) {
      assert(anv_batch_has_error(&cmd_buffer->batch));
      return;
   }

   genX(flush_pipeline_select_3d)(cmd_buffer);

   cmd_buffer_begin_subpass(cmd_buffer, 0);
}

void genX(CmdNextSubpass)(
    VkCommandBuffer                             commandBuffer,
    VkSubpassContents                           contents)
{
   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);

   if (anv_batch_has_error(&cmd_buffer->batch))
      return;

   assert(cmd_buffer->level == VK_COMMAND_BUFFER_LEVEL_PRIMARY);

   uint32_t prev_subpass = anv_get_subpass_id(&cmd_buffer->state);
   cmd_buffer_end_subpass(cmd_buffer);
   cmd_buffer_begin_subpass(cmd_buffer, prev_subpass + 1);
}

void genX(CmdEndRenderPass)(
    VkCommandBuffer                             commandBuffer)
{
   ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);

   if (anv_batch_has_error(&cmd_buffer->batch))
      return;

   cmd_buffer_end_subpass(cmd_buffer);

   cmd_buffer->state.hiz_enabled = false;

#ifndef NDEBUG
   anv_dump_add_framebuffer(cmd_buffer, cmd_buffer->state.framebuffer);
#endif

   /* Remove references to render pass specific state. This enables us to
    * detect whether or not we're in a renderpass.
    */
   cmd_buffer->state.framebuffer = NULL;
   cmd_buffer->state.pass = NULL;
   cmd_buffer->state.subpass = NULL;
}