src/intel/vulkan/anv_meta_blit.c

   1 /*
   2  * Copyright © 2015 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include "anv_meta.h"
  25 #include "nir/nir_builder.h"
  26
  27 struct blit_region {
  28    VkOffset3D src_offset;
  29    VkExtent3D src_extent;
  30    VkOffset3D dest_offset;
  31    VkExtent3D dest_extent;
  32 };
  33
  34 static nir_shader *
  35 build_nir_vertex_shader(void)
  36 {
  37    const struct glsl_type *vec4 = glsl_vec4_type();
  38    nir_builder b;
  39
  40    nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_VERTEX, NULL);
  41    b.shader->info.name = ralloc_strdup(b.shader, "meta_blit_vs");
  42
  43    nir_variable *pos_in = nir_variable_create(b.shader, nir_var_shader_in,
  44                                               vec4, "a_pos");
  45    pos_in->data.location = VERT_ATTRIB_GENERIC0;
  46    nir_variable *pos_out = nir_variable_create(b.shader, nir_var_shader_out,
  47                                                vec4, "gl_Position");
  48    pos_out->data.location = VARYING_SLOT_POS;
  49    nir_copy_var(&b, pos_out, pos_in);
  50
  51    nir_variable *tex_pos_in = nir_variable_create(b.shader, nir_var_shader_in,
  52                                                   vec4, "a_tex_pos");
  53    tex_pos_in->data.location = VERT_ATTRIB_GENERIC1;
  54    nir_variable *tex_pos_out = nir_variable_create(b.shader, nir_var_shader_out,
  55                                                    vec4, "v_tex_pos");
  56    tex_pos_out->data.location = VARYING_SLOT_VAR0;
  57    tex_pos_out->data.interpolation = INTERP_QUALIFIER_SMOOTH;
  58    nir_copy_var(&b, tex_pos_out, tex_pos_in);
  59
  60    return b.shader;
  61 }
  62
  63 static nir_shader *
  64 build_nir_copy_fragment_shader(enum glsl_sampler_dim tex_dim)
  65 {
  66    const struct glsl_type *vec4 = glsl_vec4_type();
  67    nir_builder b;
  68
  69    nir_builder_init_simple_shader(&b, NULL, MESA_SHADER_FRAGMENT, NULL);
  70    b.shader->info.name = ralloc_strdup(b.shader, "meta_blit_fs");
  71
  72    nir_variable *tex_pos_in = nir_variable_create(b.shader, nir_var_shader_in,
  73                                                   vec4, "v_tex_pos");
  74    tex_pos_in->data.location = VARYING_SLOT_VAR0;
  75
  76    /* Swizzle the array index which comes in as Z coordinate into the right
  77     * position.
  78     */
  79    unsigned swz[] = { 0, (tex_dim == GLSL_SAMPLER_DIM_1D ? 2 : 1), 2 };
  80    nir_ssa_def *const tex_pos =
  81       nir_swizzle(&b, nir_load_var(&b, tex_pos_in), swz,
  82                   (tex_dim == GLSL_SAMPLER_DIM_1D ? 2 : 3), false);
  83
  84    const struct glsl_type *sampler_type =
  85       glsl_sampler_type(tex_dim, false, tex_dim != GLSL_SAMPLER_DIM_3D,
  86                         glsl_get_base_type(vec4));
  87    nir_variable *sampler = nir_variable_create(b.shader, nir_var_uniform,
  88                                                sampler_type, "s_tex");
  89    sampler->data.descriptor_set = 0;
  90    sampler->data.binding = 0;
  91
  92    nir_tex_instr *tex = nir_tex_instr_create(b.shader, 1);
  93    tex->sampler_dim = tex_dim;
  94    tex->op = nir_texop_tex;
  95    tex->src[0].src_type = nir_tex_src_coord;
  96    tex->src[0].src = nir_src_for_ssa(tex_pos);
  97    tex->dest_type = nir_type_float; /* TODO */
  98    tex->is_array = glsl_sampler_type_is_array(sampler_type);
  99    tex->coord_components = tex_pos->num_components;
 100    tex->texture = nir_deref_var_create(tex, sampler);
 101    tex->sampler = nir_deref_var_create(tex, sampler);
 102
 103    nir_ssa_dest_init(&tex->instr, &tex->dest, 4, "tex");
 104    nir_builder_instr_insert(&b, &tex->instr);
 105
 106    nir_variable *color_out = nir_variable_create(b.shader, nir_var_shader_out,
 107                                                  vec4, "f_color");
 108    color_out->data.location = FRAG_RESULT_DATA0;
 109    nir_store_var(&b, color_out, &tex->dest.ssa, 4);
 110
 111    return b.shader;
 112 }
 113
 114 static void
 115 meta_prepare_blit(struct anv_cmd_buffer *cmd_buffer,
 116                   struct anv_meta_saved_state *saved_state)
 117 {
 118    anv_meta_save(saved_state, cmd_buffer,
 119                  (1 << VK_DYNAMIC_STATE_VIEWPORT));
 120 }
 121
 122 void
 123 anv_meta_begin_blit2d(struct anv_cmd_buffer *cmd_buffer,
 124                       struct anv_meta_saved_state *save)
 125 {
 126    meta_prepare_blit(cmd_buffer, save);
 127 }
 128
 129
 130 /* Returns the user-provided VkBufferImageCopy::imageOffset in units of
 131  * elements rather than texels. One element equals one texel or one block
 132  * if Image is uncompressed or compressed, respectively.
 133  */
 134 static struct VkOffset3D
 135 meta_region_offset_el(const struct anv_image * image,
 136                       const struct VkOffset3D * offset)
 137 {
 138    const struct isl_format_layout * isl_layout = image->format->isl_layout;
 139    return (VkOffset3D) {
 140       .x = offset->x / isl_layout->bw,
 141       .y = offset->y / isl_layout->bh,
 142       .z = offset->z / isl_layout->bd,
 143    };
 144 }
 145
 146 /* Returns the user-provided VkBufferImageCopy::imageExtent in units of
 147  * elements rather than texels. One element equals one texel or one block
 148  * if Image is uncompressed or compressed, respectively.
 149  */
 150 static struct VkExtent3D
 151 meta_region_extent_el(const VkFormat format,
 152                       const struct VkExtent3D * extent)
 153 {
 154    const struct isl_format_layout * isl_layout =
 155       anv_format_for_vk_format(format)->isl_layout;
 156    return (VkExtent3D) {
 157       .width  = DIV_ROUND_UP(extent->width , isl_layout->bw),
 158       .height = DIV_ROUND_UP(extent->height, isl_layout->bh),
 159       .depth  = DIV_ROUND_UP(extent->depth , isl_layout->bd),
 160    };
 161 }
 162
 163 static void
 164 meta_emit_blit(struct anv_cmd_buffer *cmd_buffer,
 165                struct anv_image *src_image,
 166                struct anv_image_view *src_iview,
 167                VkOffset3D src_offset,
 168                VkExtent3D src_extent,
 169                struct anv_image *dest_image,
 170                struct anv_image_view *dest_iview,
 171                VkOffset3D dest_offset,
 172                VkExtent3D dest_extent,
 173                VkFilter blit_filter)
 174 {
 175    struct anv_device *device = cmd_buffer->device;
 176
 177    struct blit_vb_data {
 178       float pos[2];
 179       float tex_coord[3];
 180    } *vb_data;
 181
 182    assert(src_image->samples == dest_image->samples);
 183
 184    unsigned vb_size = sizeof(struct anv_vue_header) + 3 * sizeof(*vb_data);
 185
 186    struct anv_state vb_state =
 187       anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, vb_size, 16);
 188    memset(vb_state.map, 0, sizeof(struct anv_vue_header));
 189    vb_data = vb_state.map + sizeof(struct anv_vue_header);
 190
 191    vb_data[0] = (struct blit_vb_data) {
 192       .pos = {
 193          dest_offset.x + dest_extent.width,
 194          dest_offset.y + dest_extent.height,
 195       },
 196       .tex_coord = {
 197          (float)(src_offset.x + src_extent.width) / (float)src_iview->extent.width,
 198          (float)(src_offset.y + src_extent.height) / (float)src_iview->extent.height,
 199          (float)src_offset.z / (float)src_iview->extent.depth,
 200       },
 201    };
 202
 203    vb_data[1] = (struct blit_vb_data) {
 204       .pos = {
 205          dest_offset.x,
 206          dest_offset.y + dest_extent.height,
 207       },
 208       .tex_coord = {
 209          (float)src_offset.x / (float)src_iview->extent.width,
 210          (float)(src_offset.y + src_extent.height) / (float)src_iview->extent.height,
 211          (float)src_offset.z / (float)src_iview->extent.depth,
 212       },
 213    };
 214
 215    vb_data[2] = (struct blit_vb_data) {
 216       .pos = {
 217          dest_offset.x,
 218          dest_offset.y,
 219       },
 220       .tex_coord = {
 221          (float)src_offset.x / (float)src_iview->extent.width,
 222          (float)src_offset.y / (float)src_iview->extent.height,
 223          (float)src_offset.z / (float)src_iview->extent.depth,
 224       },
 225    };
 226
 227    anv_state_clflush(vb_state);
 228
 229    struct anv_buffer vertex_buffer = {
 230       .device = device,
 231       .size = vb_size,
 232       .bo = &device->dynamic_state_block_pool.bo,
 233       .offset = vb_state.offset,
 234    };
 235
 236    anv_CmdBindVertexBuffers(anv_cmd_buffer_to_handle(cmd_buffer), 0, 2,
 237       (VkBuffer[]) {
 238          anv_buffer_to_handle(&vertex_buffer),
 239          anv_buffer_to_handle(&vertex_buffer)
 240       },
 241       (VkDeviceSize[]) {
 242          0,
 243          sizeof(struct anv_vue_header),
 244       });
 245
 246    VkSampler sampler;
 247    ANV_CALL(CreateSampler)(anv_device_to_handle(device),
 248       &(VkSamplerCreateInfo) {
 249          .sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO,
 250          .magFilter = blit_filter,
 251          .minFilter = blit_filter,
 252       }, &cmd_buffer->pool->alloc, &sampler);
 253
 254    VkDescriptorPool desc_pool;
 255    anv_CreateDescriptorPool(anv_device_to_handle(device),
 256       &(const VkDescriptorPoolCreateInfo) {
 257          .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_POOL_CREATE_INFO,
 258          .pNext = NULL,
 259          .flags = 0,
 260          .maxSets = 1,
 261          .poolSizeCount = 1,
 262          .pPoolSizes = (VkDescriptorPoolSize[]) {
 263             {
 264                .type = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
 265                .descriptorCount = 1
 266             },
 267          }
 268       }, &cmd_buffer->pool->alloc, &desc_pool);
 269
 270    VkDescriptorSet set;
 271    anv_AllocateDescriptorSets(anv_device_to_handle(device),
 272       &(VkDescriptorSetAllocateInfo) {
 273          .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_ALLOCATE_INFO,
 274          .descriptorPool = desc_pool,
 275          .descriptorSetCount = 1,
 276          .pSetLayouts = &device->meta_state.blit.ds_layout
 277       }, &set);
 278
 279    anv_UpdateDescriptorSets(anv_device_to_handle(device),
 280       1, /* writeCount */
 281       (VkWriteDescriptorSet[]) {
 282          {
 283             .sType = VK_STRUCTURE_TYPE_WRITE_DESCRIPTOR_SET,
 284             .dstSet = set,
 285             .dstBinding = 0,
 286             .dstArrayElement = 0,
 287             .descriptorCount = 1,
 288             .descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
 289             .pImageInfo = (VkDescriptorImageInfo[]) {
 290                {
 291                   .sampler = sampler,
 292                   .imageView = anv_image_view_to_handle(src_iview),
 293                   .imageLayout = VK_IMAGE_LAYOUT_GENERAL,
 294                },
 295             }
 296          }
 297       }, 0, NULL);
 298
 299    VkFramebuffer fb;
 300    anv_CreateFramebuffer(anv_device_to_handle(device),
 301       &(VkFramebufferCreateInfo) {
 302          .sType = VK_STRUCTURE_TYPE_FRAMEBUFFER_CREATE_INFO,
 303          .attachmentCount = 1,
 304          .pAttachments = (VkImageView[]) {
 305             anv_image_view_to_handle(dest_iview),
 306          },
 307          .width = dest_iview->extent.width,
 308          .height = dest_iview->extent.height,
 309          .layers = 1
 310       }, &cmd_buffer->pool->alloc, &fb);
 311
 312    ANV_CALL(CmdBeginRenderPass)(anv_cmd_buffer_to_handle(cmd_buffer),
 313       &(VkRenderPassBeginInfo) {
 314          .sType = VK_STRUCTURE_TYPE_RENDER_PASS_BEGIN_INFO,
 315          .renderPass = device->meta_state.blit.render_pass,
 316          .framebuffer = fb,
 317          .renderArea = {
 318             .offset = { dest_offset.x, dest_offset.y },
 319             .extent = { dest_extent.width, dest_extent.height },
 320          },
 321          .clearValueCount = 0,
 322          .pClearValues = NULL,
 323       }, VK_SUBPASS_CONTENTS_INLINE);
 324
 325    VkPipeline pipeline;
 326
 327    switch (src_image->type) {
 328    case VK_IMAGE_TYPE_1D:
 329       pipeline = device->meta_state.blit.pipeline_1d_src;
 330       break;
 331    case VK_IMAGE_TYPE_2D:
 332       pipeline = device->meta_state.blit.pipeline_2d_src;
 333       break;
 334    case VK_IMAGE_TYPE_3D:
 335       pipeline = device->meta_state.blit.pipeline_3d_src;
 336       break;
 337    default:
 338       unreachable(!"bad VkImageType");
 339    }
 340
 341    if (cmd_buffer->state.pipeline != anv_pipeline_from_handle(pipeline)) {
 342       anv_CmdBindPipeline(anv_cmd_buffer_to_handle(cmd_buffer),
 343                           VK_PIPELINE_BIND_POINT_GRAPHICS, pipeline);
 344    }
 345
 346    anv_CmdSetViewport(anv_cmd_buffer_to_handle(cmd_buffer), 0, 1,
 347                       &(VkViewport) {
 348                         .x = 0.0f,
 349                         .y = 0.0f,
 350                         .width = dest_iview->extent.width,
 351                         .height = dest_iview->extent.height,
 352                         .minDepth = 0.0f,
 353                         .maxDepth = 1.0f,
 354                       });
 355
 356    anv_CmdBindDescriptorSets(anv_cmd_buffer_to_handle(cmd_buffer),
 357                              VK_PIPELINE_BIND_POINT_GRAPHICS,
 358                              device->meta_state.blit.pipeline_layout, 0, 1,
 359                              &set, 0, NULL);
 360
 361    ANV_CALL(CmdDraw)(anv_cmd_buffer_to_handle(cmd_buffer), 3, 1, 0, 0);
 362
 363    ANV_CALL(CmdEndRenderPass)(anv_cmd_buffer_to_handle(cmd_buffer));
 364
 365    /* At the point where we emit the draw call, all data from the
 366     * descriptor sets, etc. has been used.  We are free to delete it.
 367     */
 368    anv_DestroyDescriptorPool(anv_device_to_handle(device),
 369                              desc_pool, &cmd_buffer->pool->alloc);
 370    anv_DestroySampler(anv_device_to_handle(device), sampler,
 371                       &cmd_buffer->pool->alloc);
 372    anv_DestroyFramebuffer(anv_device_to_handle(device), fb,
 373                           &cmd_buffer->pool->alloc);
 374 }
 375
 376 static void
 377 meta_finish_blit(struct anv_cmd_buffer *cmd_buffer,
 378                  const struct anv_meta_saved_state *saved_state)
 379 {
 380    anv_meta_restore(saved_state, cmd_buffer);
 381 }
 382
 383 void
 384 anv_meta_end_blit2d(struct anv_cmd_buffer *cmd_buffer,
 385                     struct anv_meta_saved_state *save)
 386 {
 387    meta_finish_blit(cmd_buffer, save);
 388 }
 389
 390 static VkFormat
 391 vk_format_for_size(int bs)
 392 {
 393    /* The choice of UNORM and UINT formats is very intentional here.  Most of
 394     * the time, we want to use a UINT format to avoid any rounding error in
 395     * the blit.  For stencil blits, R8_UINT is required by the hardware.
 396     * (It's the only format allowed in conjunction with W-tiling.)  Also we
 397     * intentionally use the 4-channel formats whenever we can.  This is so
 398     * that, when we do a RGB <-> RGBX copy, the two formats will line up even
 399     * though one of them is 3/4 the size of the other.  The choice of UNORM
 400     * vs. UINT is also very intentional because Haswell doesn't handle 8 or
 401     * 16-bit RGB UINT formats at all so we have to use UNORM there.
 402     * Fortunately, the only time we should ever use two different formats in
 403     * the table below is for RGB -> RGBA blits and so we will never have any
 404     * UNORM/UINT mismatch.
 405     */
 406    switch (bs) {
 407    case 1: return VK_FORMAT_R8_UINT;
 408    case 2: return VK_FORMAT_R8G8_UINT;
 409    case 3: return VK_FORMAT_R8G8B8_UNORM;
 410    case 4: return VK_FORMAT_R8G8B8A8_UNORM;
 411    case 6: return VK_FORMAT_R16G16B16_UNORM;
 412    case 8: return VK_FORMAT_R16G16B16A16_UNORM;
 413    case 12: return VK_FORMAT_R32G32B32_UINT;
 414    case 16: return VK_FORMAT_R32G32B32A32_UINT;
 415    default:
 416       unreachable("Invalid format block size");
 417    }
 418 }
 419
 420 static struct anv_meta_blit2d_surf
 421 blit_surf_for_image(const struct anv_image* image,
 422                     const struct isl_surf *img_isl_surf)
 423 {
 424    return (struct anv_meta_blit2d_surf) {
 425             .bo = image->bo,
 426             .tiling = img_isl_surf->tiling,
 427             .base_offset = image->offset,
 428             .bs = isl_format_get_layout(img_isl_surf->format)->bs,
 429             .pitch = isl_surf_get_row_pitch(img_isl_surf),
 430    };
 431 }
 432
 433 void
 434 anv_meta_blit2d(struct anv_cmd_buffer *cmd_buffer,
 435                 struct anv_meta_blit2d_surf *src,
 436                 struct anv_meta_blit2d_surf *dst,
 437                 unsigned num_rects,
 438                 struct anv_meta_blit2d_rect *rects)
 439 {
 440    VkDevice vk_device = anv_device_to_handle(cmd_buffer->device);
 441    VkFormat src_format = vk_format_for_size(src->bs);
 442    VkFormat dst_format = vk_format_for_size(dst->bs);
 443
 444    for (unsigned r = 0; r < num_rects; ++r) {
 445
 446       /* Create VkImages */
 447       VkImageCreateInfo image_info = {
 448          .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
 449          .imageType = VK_IMAGE_TYPE_2D,
 450          .format = 0, /* TEMPLATE */
 451          .extent = {
 452             .width = 0, /* TEMPLATE */
 453             /* Pad to highest tile height to compensate for a vertical intratile offset */
 454             .height = MIN(rects[r].height + 64, 1 << 14),
 455             .depth = 1,
 456          },
 457          .mipLevels = 1,
 458          .arrayLayers = 1,
 459          .samples = 1,
 460          .tiling = 0, /* TEMPLATE */
 461          .usage = 0, /* TEMPLATE */
 462       };
 463       struct anv_image_create_info anv_image_info = {
 464          .vk_info = &image_info,
 465          .isl_tiling_flags = 0, /* TEMPLATE */
 466       };
 467
 468       anv_image_info.isl_tiling_flags = 1 << src->tiling;
 469       image_info.tiling = anv_image_info.isl_tiling_flags == ISL_TILING_LINEAR_BIT ?
 470                         VK_IMAGE_TILING_LINEAR : VK_IMAGE_TILING_OPTIMAL;
 471       image_info.usage = VK_IMAGE_USAGE_SAMPLED_BIT;
 472       image_info.format = src_format,
 473       image_info.extent.width = src->pitch / src->bs;
 474       VkImage src_image;
 475       anv_image_create(vk_device, &anv_image_info,
 476                      &cmd_buffer->pool->alloc, &src_image);
 477
 478       anv_image_info.isl_tiling_flags = 1 << dst->tiling;
 479       image_info.tiling = anv_image_info.isl_tiling_flags == ISL_TILING_LINEAR_BIT ?
 480                         VK_IMAGE_TILING_LINEAR : VK_IMAGE_TILING_OPTIMAL;
 481       image_info.usage = VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT;
 482       image_info.format = dst_format,
 483       image_info.extent.width = dst->pitch / dst->bs;
 484       VkImage dst_image;
 485       anv_image_create(vk_device, &anv_image_info,
 486                      &cmd_buffer->pool->alloc, &dst_image);
 487
 488       /* We could use a vk call to bind memory, but that would require
 489       * creating a dummy memory object etc. so there's really no point.
 490       */
 491       anv_image_from_handle(src_image)->bo = src->bo;
 492       anv_image_from_handle(src_image)->offset = src->base_offset;
 493       anv_image_from_handle(dst_image)->bo = dst->bo;
 494       anv_image_from_handle(dst_image)->offset = dst->base_offset;
 495
 496       /* Create VkImageViews */
 497       VkImageViewCreateInfo iview_info = {
 498          .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
 499          .image = 0, /* TEMPLATE */
 500          .viewType = VK_IMAGE_VIEW_TYPE_2D,
 501          .format = 0, /* TEMPLATE */
 502          .subresourceRange = {
 503             .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
 504             .baseMipLevel = 0,
 505             .levelCount = 1,
 506             .baseArrayLayer = 0,
 507             .layerCount = 1
 508          },
 509       };
 510       uint32_t img_o = 0;
 511
 512       iview_info.image = src_image;
 513       iview_info.format = src_format;
 514       VkOffset3D src_offset_el = {0};
 515       isl_surf_get_image_intratile_offset_el_xy(&cmd_buffer->device->isl_dev,
 516                                                 &anv_image_from_handle(src_image)->
 517                                                    color_surface.isl,
 518                                                 rects[r].src_x,
 519                                                 rects[r].src_y,
 520                                                 &img_o,
 521                                                 (uint32_t*)&src_offset_el.x,
 522                                                 (uint32_t*)&src_offset_el.y);
 523
 524       struct anv_image_view src_iview;
 525       anv_image_view_init(&src_iview, cmd_buffer->device,
 526          &iview_info, cmd_buffer, img_o, VK_IMAGE_USAGE_SAMPLED_BIT);
 527
 528       iview_info.image = dst_image;
 529       iview_info.format = dst_format;
 530       VkOffset3D dst_offset_el = {0};
 531       isl_surf_get_image_intratile_offset_el_xy(&cmd_buffer->device->isl_dev,
 532                                                 &anv_image_from_handle(dst_image)->
 533                                                    color_surface.isl,
 534                                                 rects[r].dst_x,
 535                                                 rects[r].dst_y,
 536                                                 &img_o,
 537                                                 (uint32_t*)&dst_offset_el.x,
 538                                                 (uint32_t*)&dst_offset_el.y);
 539       struct anv_image_view dst_iview;
 540       anv_image_view_init(&dst_iview, cmd_buffer->device,
 541          &iview_info, cmd_buffer, img_o, VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT);
 542
 543       /* Perform blit */
 544       meta_emit_blit(cmd_buffer,
 545                      anv_image_from_handle(src_image),
 546                      &src_iview,
 547                      src_offset_el,
 548                      (VkExtent3D){rects[r].width, rects[r].height, 1},
 549                      anv_image_from_handle(dst_image),
 550                      &dst_iview,
 551                      dst_offset_el,
 552                      (VkExtent3D){rects[r].width, rects[r].height, 1},
 553                      VK_FILTER_NEAREST);
 554
 555       anv_DestroyImage(vk_device, src_image, &cmd_buffer->pool->alloc);
 556       anv_DestroyImage(vk_device, dst_image, &cmd_buffer->pool->alloc);
 557    }
 558 }
 559
 560 static void
 561 do_buffer_copy(struct anv_cmd_buffer *cmd_buffer,
 562                struct anv_bo *src, uint64_t src_offset,
 563                struct anv_bo *dest, uint64_t dest_offset,
 564                int width, int height, int bs)
 565 {
 566    VkDevice vk_device = anv_device_to_handle(cmd_buffer->device);
 567    VkFormat copy_format = vk_format_for_size(bs);
 568
 569    VkImageCreateInfo image_info = {
 570       .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
 571       .imageType = VK_IMAGE_TYPE_2D,
 572       .format = copy_format,
 573       .extent = {
 574          .width = width,
 575          .height = height,
 576          .depth = 1,
 577       },
 578       .mipLevels = 1,
 579       .arrayLayers = 1,
 580       .samples = 1,
 581       .tiling = VK_IMAGE_TILING_LINEAR,
 582       .usage = 0,
 583       .flags = 0,
 584    };
 585
 586    VkImage src_image;
 587    image_info.usage = VK_IMAGE_USAGE_SAMPLED_BIT;
 588    anv_CreateImage(vk_device, &image_info,
 589                    &cmd_buffer->pool->alloc, &src_image);
 590
 591    VkImage dest_image;
 592    image_info.usage = VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT;
 593    anv_CreateImage(vk_device, &image_info,
 594                    &cmd_buffer->pool->alloc, &dest_image);
 595
 596    /* We could use a vk call to bind memory, but that would require
 597     * creating a dummy memory object etc. so there's really no point.
 598     */
 599    anv_image_from_handle(src_image)->bo = src;
 600    anv_image_from_handle(src_image)->offset = src_offset;
 601    anv_image_from_handle(dest_image)->bo = dest;
 602    anv_image_from_handle(dest_image)->offset = dest_offset;
 603
 604    VkImageViewCreateInfo iview_info = {
 605       .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
 606       .image = 0, /* TEMPLATE */
 607       .viewType = VK_IMAGE_VIEW_TYPE_2D,
 608       .format = copy_format,
 609       .subresourceRange = {
 610          .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
 611          .baseMipLevel = 0,
 612          .levelCount = 1,
 613          .baseArrayLayer = 0,
 614          .layerCount = 1
 615       },
 616    };
 617
 618    struct anv_image_view src_iview;
 619    iview_info.image = src_image;
 620    anv_image_view_init(&src_iview, cmd_buffer->device,
 621       &iview_info, cmd_buffer, 0, VK_IMAGE_USAGE_SAMPLED_BIT);
 622
 623    struct anv_image_view dest_iview;
 624    iview_info.image = dest_image;
 625    anv_image_view_init(&dest_iview, cmd_buffer->device,
 626       &iview_info, cmd_buffer, 0, VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT);
 627
 628    meta_emit_blit(cmd_buffer,
 629                   anv_image_from_handle(src_image),
 630                   &src_iview,
 631                   (VkOffset3D) { 0, 0, 0 },
 632                   (VkExtent3D) { width, height, 1 },
 633                   anv_image_from_handle(dest_image),
 634                   &dest_iview,
 635                   (VkOffset3D) { 0, 0, 0 },
 636                   (VkExtent3D) { width, height, 1 },
 637                   VK_FILTER_NEAREST);
 638
 639    anv_DestroyImage(vk_device, src_image, &cmd_buffer->pool->alloc);
 640    anv_DestroyImage(vk_device, dest_image, &cmd_buffer->pool->alloc);
 641 }
 642
 643 void anv_CmdCopyBuffer(
 644     VkCommandBuffer                             commandBuffer,
 645     VkBuffer                                    srcBuffer,
 646     VkBuffer                                    destBuffer,
 647     uint32_t                                    regionCount,
 648     const VkBufferCopy*                         pRegions)
 649 {
 650    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 651    ANV_FROM_HANDLE(anv_buffer, src_buffer, srcBuffer);
 652    ANV_FROM_HANDLE(anv_buffer, dest_buffer, destBuffer);
 653
 654    struct anv_meta_saved_state saved_state;
 655
 656    meta_prepare_blit(cmd_buffer, &saved_state);
 657
 658    for (unsigned r = 0; r < regionCount; r++) {
 659       uint64_t src_offset = src_buffer->offset + pRegions[r].srcOffset;
 660       uint64_t dest_offset = dest_buffer->offset + pRegions[r].dstOffset;
 661       uint64_t copy_size = pRegions[r].size;
 662
 663       /* First, we compute the biggest format that can be used with the
 664        * given offsets and size.
 665        */
 666       int bs = 16;
 667
 668       int fs = ffs(src_offset) - 1;
 669       if (fs != -1)
 670          bs = MIN2(bs, 1 << fs);
 671       assert(src_offset % bs == 0);
 672
 673       fs = ffs(dest_offset) - 1;
 674       if (fs != -1)
 675          bs = MIN2(bs, 1 << fs);
 676       assert(dest_offset % bs == 0);
 677
 678       fs = ffs(pRegions[r].size) - 1;
 679       if (fs != -1)
 680          bs = MIN2(bs, 1 << fs);
 681       assert(pRegions[r].size % bs == 0);
 682
 683       /* This is maximum possible width/height our HW can handle */
 684       uint64_t max_surface_dim = 1 << 14;
 685
 686       /* First, we make a bunch of max-sized copies */
 687       uint64_t max_copy_size = max_surface_dim * max_surface_dim * bs;
 688       while (copy_size >= max_copy_size) {
 689          do_buffer_copy(cmd_buffer, src_buffer->bo, src_offset,
 690                         dest_buffer->bo, dest_offset,
 691                         max_surface_dim, max_surface_dim, bs);
 692          copy_size -= max_copy_size;
 693          src_offset += max_copy_size;
 694          dest_offset += max_copy_size;
 695       }
 696
 697       uint64_t height = copy_size / (max_surface_dim * bs);
 698       assert(height < max_surface_dim);
 699       if (height != 0) {
 700          uint64_t rect_copy_size = height * max_surface_dim * bs;
 701          do_buffer_copy(cmd_buffer, src_buffer->bo, src_offset,
 702                         dest_buffer->bo, dest_offset,
 703                         max_surface_dim, height, bs);
 704          copy_size -= rect_copy_size;
 705          src_offset += rect_copy_size;
 706          dest_offset += rect_copy_size;
 707       }
 708
 709       if (copy_size != 0) {
 710          do_buffer_copy(cmd_buffer, src_buffer->bo, src_offset,
 711                         dest_buffer->bo, dest_offset,
 712                         copy_size / bs, 1, bs);
 713       }
 714    }
 715
 716    meta_finish_blit(cmd_buffer, &saved_state);
 717 }
 718
 719 void anv_CmdUpdateBuffer(
 720     VkCommandBuffer                             commandBuffer,
 721     VkBuffer                                    dstBuffer,
 722     VkDeviceSize                                dstOffset,
 723     VkDeviceSize                                dataSize,
 724     const uint32_t*                             pData)
 725 {
 726    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 727    ANV_FROM_HANDLE(anv_buffer, dst_buffer, dstBuffer);
 728    struct anv_meta_saved_state saved_state;
 729
 730    meta_prepare_blit(cmd_buffer, &saved_state);
 731
 732    /* We can't quite grab a full block because the state stream needs a
 733     * little data at the top to build its linked list.
 734     */
 735    const uint32_t max_update_size =
 736       cmd_buffer->device->dynamic_state_block_pool.block_size - 64;
 737
 738    assert(max_update_size < (1 << 14) * 4);
 739
 740    while (dataSize) {
 741       const uint32_t copy_size = MIN2(dataSize, max_update_size);
 742
 743       struct anv_state tmp_data =
 744          anv_cmd_buffer_alloc_dynamic_state(cmd_buffer, copy_size, 64);
 745
 746       memcpy(tmp_data.map, pData, copy_size);
 747
 748       int bs;
 749       if ((copy_size & 15) == 0 && (dstOffset & 15) == 0) {
 750          bs = 16;
 751       } else if ((copy_size & 7) == 0 && (dstOffset & 7) == 0) {
 752          bs = 8;
 753       } else {
 754          assert((copy_size & 3) == 0 && (dstOffset & 3) == 0);
 755          bs = 4;
 756       }
 757
 758       do_buffer_copy(cmd_buffer,
 759                      &cmd_buffer->device->dynamic_state_block_pool.bo,
 760                      tmp_data.offset,
 761                      dst_buffer->bo, dst_buffer->offset + dstOffset,
 762                      copy_size / bs, 1, bs);
 763
 764       dataSize -= copy_size;
 765       dstOffset += copy_size;
 766       pData = (void *)pData + copy_size;
 767    }
 768
 769    meta_finish_blit(cmd_buffer, &saved_state);
 770 }
 771
 772 static VkFormat
 773 choose_iview_format(struct anv_image *image, VkImageAspectFlagBits aspect)
 774 {
 775    assert(__builtin_popcount(aspect) == 1);
 776
 777    struct isl_surf *surf =
 778       &anv_image_get_surface_for_aspect_mask(image, aspect)->isl;
 779
 780    /* vkCmdCopyImage behaves like memcpy. Therefore we choose identical UINT
 781     * formats for the source and destination image views.
 782     *
 783     * From the Vulkan spec (2015-12-30):
 784     *
 785     *    vkCmdCopyImage performs image copies in a similar manner to a host
 786     *    memcpy. It does not perform general-purpose conversions such as
 787     *    scaling, resizing, blending, color-space conversion, or format
 788     *    conversions.  Rather, it simply copies raw image data. vkCmdCopyImage
 789     *    can copy between images with different formats, provided the formats
 790     *    are compatible as defined below.
 791     *
 792     *    [The spec later defines compatibility as having the same number of
 793     *    bytes per block].
 794     */
 795    return vk_format_for_size(isl_format_layouts[surf->format].bs);
 796 }
 797
 798 static VkFormat
 799 choose_buffer_format(VkFormat format, VkImageAspectFlagBits aspect)
 800 {
 801    assert(__builtin_popcount(aspect) == 1);
 802
 803    /* vkCmdCopy* commands behave like memcpy. Therefore we choose
 804     * compatable UINT formats for the source and destination image views.
 805     *
 806     * For the buffer, we go back to the original image format and get a
 807     * the format as if it were linear.  This way, for RGB formats, we get
 808     * an RGB format here even if the tiled image is RGBA. XXX: This doesn't
 809     * work if the buffer is the destination.
 810     */
 811    enum isl_format linear_format = anv_get_isl_format(format, aspect,
 812                                                       VK_IMAGE_TILING_LINEAR,
 813                                                       NULL);
 814
 815    return vk_format_for_size(isl_format_layouts[linear_format].bs);
 816 }
 817
 818 void anv_CmdCopyImage(
 819     VkCommandBuffer                             commandBuffer,
 820     VkImage                                     srcImage,
 821     VkImageLayout                               srcImageLayout,
 822     VkImage                                     destImage,
 823     VkImageLayout                               destImageLayout,
 824     uint32_t                                    regionCount,
 825     const VkImageCopy*                          pRegions)
 826 {
 827    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 828    ANV_FROM_HANDLE(anv_image, src_image, srcImage);
 829    ANV_FROM_HANDLE(anv_image, dest_image, destImage);
 830    struct anv_meta_saved_state saved_state;
 831
 832    /* From the Vulkan 1.0 spec:
 833     *
 834     *    vkCmdCopyImage can be used to copy image data between multisample
 835     *    images, but both images must have the same number of samples.
 836     */
 837    assert(src_image->samples == dest_image->samples);
 838
 839    meta_prepare_blit(cmd_buffer, &saved_state);
 840
 841    for (unsigned r = 0; r < regionCount; r++) {
 842       assert(pRegions[r].srcSubresource.aspectMask ==
 843              pRegions[r].dstSubresource.aspectMask);
 844
 845       VkImageAspectFlags aspect = pRegions[r].srcSubresource.aspectMask;
 846
 847       VkFormat src_format = choose_iview_format(src_image, aspect);
 848       VkFormat dst_format = choose_iview_format(dest_image, aspect);
 849
 850       struct anv_image_view src_iview;
 851       anv_image_view_init(&src_iview, cmd_buffer->device,
 852          &(VkImageViewCreateInfo) {
 853             .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
 854             .image = srcImage,
 855             .viewType = anv_meta_get_view_type(src_image),
 856             .format = src_format,
 857             .subresourceRange = {
 858                .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
 859                .baseMipLevel = pRegions[r].srcSubresource.mipLevel,
 860                .levelCount = 1,
 861                .baseArrayLayer = pRegions[r].srcSubresource.baseArrayLayer,
 862                .layerCount = pRegions[r].dstSubresource.layerCount,
 863             },
 864          },
 865          cmd_buffer, 0, VK_IMAGE_USAGE_SAMPLED_BIT);
 866
 867       const uint32_t dest_base_array_slice =
 868          anv_meta_get_iview_layer(dest_image, &pRegions[r].dstSubresource,
 869                                   &pRegions[r].dstOffset);
 870
 871
 872       unsigned num_slices_3d = pRegions[r].extent.depth;
 873       unsigned num_slices_array = pRegions[r].dstSubresource.layerCount;
 874       unsigned slice_3d = 0;
 875       unsigned slice_array = 0;
 876       while (slice_3d < num_slices_3d && slice_array < num_slices_array) {
 877          VkOffset3D src_offset = pRegions[r].srcOffset;
 878          src_offset.z += slice_3d + slice_array;
 879
 880          uint32_t img_x = 0;
 881          uint32_t img_y = 0;
 882          uint32_t img_o = 0;
 883          if (isl_format_is_compressed(dest_image->format->isl_format))
 884             isl_surf_get_image_intratile_offset_el(&cmd_buffer->device->isl_dev,
 885                                                    &dest_image->color_surface.isl,
 886                                                    pRegions[r].dstSubresource.mipLevel,
 887                                                    pRegions[r].dstSubresource.baseArrayLayer + slice_array,
 888                                                    pRegions[r].dstOffset.z + slice_3d,
 889                                                    &img_o, &img_x, &img_y);
 890
 891          VkOffset3D dest_offset_el = meta_region_offset_el(dest_image, &pRegions[r].dstOffset);
 892          dest_offset_el.x += img_x;
 893          dest_offset_el.y += img_y;
 894          dest_offset_el.z = 0;
 895
 896          struct anv_image_view dest_iview;
 897          anv_image_view_init(&dest_iview, cmd_buffer->device,
 898             &(VkImageViewCreateInfo) {
 899                .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
 900                .image = destImage,
 901                .viewType = anv_meta_get_view_type(dest_image),
 902                .format = dst_format,
 903                .subresourceRange = {
 904                   .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
 905                   .baseMipLevel = pRegions[r].dstSubresource.mipLevel,
 906                   .levelCount = 1,
 907                   .baseArrayLayer = dest_base_array_slice +
 908                                     slice_array + slice_3d,
 909                   .layerCount = 1
 910                },
 911             },
 912             cmd_buffer, img_o, VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT);
 913
 914          const VkExtent3D img_extent_el = meta_region_extent_el(dest_image->vk_format,
 915                                                                 &pRegions[r].extent);
 916
 917          meta_emit_blit(cmd_buffer,
 918                         src_image, &src_iview,
 919                         src_offset,
 920                         img_extent_el,
 921                         dest_image, &dest_iview,
 922                         dest_offset_el,
 923                         img_extent_el,
 924                         VK_FILTER_NEAREST);
 925
 926          if (dest_image->type == VK_IMAGE_TYPE_3D)
 927             slice_3d++;
 928          else
 929             slice_array++;
 930       }
 931    }
 932
 933    meta_finish_blit(cmd_buffer, &saved_state);
 934 }
 935
 936 void anv_CmdBlitImage(
 937     VkCommandBuffer                             commandBuffer,
 938     VkImage                                     srcImage,
 939     VkImageLayout                               srcImageLayout,
 940     VkImage                                     destImage,
 941     VkImageLayout                               destImageLayout,
 942     uint32_t                                    regionCount,
 943     const VkImageBlit*                          pRegions,
 944     VkFilter                                    filter)
 945
 946 {
 947    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
 948    ANV_FROM_HANDLE(anv_image, src_image, srcImage);
 949    ANV_FROM_HANDLE(anv_image, dest_image, destImage);
 950    struct anv_meta_saved_state saved_state;
 951
 952    /* From the Vulkan 1.0 spec:
 953     *
 954     *    vkCmdBlitImage must not be used for multisampled source or
 955     *    destination images. Use vkCmdResolveImage for this purpose.
 956     */
 957    assert(src_image->samples == 1);
 958    assert(dest_image->samples == 1);
 959
 960    anv_finishme("respect VkFilter");
 961
 962    meta_prepare_blit(cmd_buffer, &saved_state);
 963
 964    for (unsigned r = 0; r < regionCount; r++) {
 965       struct anv_image_view src_iview;
 966       anv_image_view_init(&src_iview, cmd_buffer->device,
 967          &(VkImageViewCreateInfo) {
 968             .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
 969             .image = srcImage,
 970             .viewType = anv_meta_get_view_type(src_image),
 971             .format = src_image->vk_format,
 972             .subresourceRange = {
 973                .aspectMask = pRegions[r].srcSubresource.aspectMask,
 974                .baseMipLevel = pRegions[r].srcSubresource.mipLevel,
 975                .levelCount = 1,
 976                .baseArrayLayer = pRegions[r].srcSubresource.baseArrayLayer,
 977                .layerCount = 1
 978             },
 979          },
 980          cmd_buffer, 0, VK_IMAGE_USAGE_SAMPLED_BIT);
 981
 982       const VkOffset3D dest_offset = {
 983          .x = pRegions[r].dstOffsets[0].x,
 984          .y = pRegions[r].dstOffsets[0].y,
 985          .z = 0,
 986       };
 987
 988       if (pRegions[r].dstOffsets[1].x < pRegions[r].dstOffsets[0].x ||
 989           pRegions[r].dstOffsets[1].y < pRegions[r].dstOffsets[0].y ||
 990           pRegions[r].srcOffsets[1].x < pRegions[r].srcOffsets[0].x ||
 991           pRegions[r].srcOffsets[1].y < pRegions[r].srcOffsets[0].y)
 992          anv_finishme("FINISHME: Allow flipping in blits");
 993
 994       const VkExtent3D dest_extent = {
 995          .width = pRegions[r].dstOffsets[1].x - pRegions[r].dstOffsets[0].x,
 996          .height = pRegions[r].dstOffsets[1].y - pRegions[r].dstOffsets[0].y,
 997       };
 998
 999       const VkExtent3D src_extent = {
1000          .width = pRegions[r].srcOffsets[1].x - pRegions[r].srcOffsets[0].x,
1001          .height = pRegions[r].srcOffsets[1].y - pRegions[r].srcOffsets[0].y,
1002       };
1003
1004       const uint32_t dest_array_slice =
1005          anv_meta_get_iview_layer(dest_image, &pRegions[r].dstSubresource,
1006                                   &pRegions[r].dstOffsets[0]);
1007
1008       if (pRegions[r].srcSubresource.layerCount > 1)
1009          anv_finishme("FINISHME: copy multiple array layers");
1010
1011       if (pRegions[r].srcOffsets[0].z + 1 != pRegions[r].srcOffsets[1].z ||
1012           pRegions[r].dstOffsets[0].z + 1 != pRegions[r].dstOffsets[1].z)
1013          anv_finishme("FINISHME: copy multiple depth layers");
1014
1015       struct anv_image_view dest_iview;
1016       anv_image_view_init(&dest_iview, cmd_buffer->device,
1017          &(VkImageViewCreateInfo) {
1018             .sType = VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO,
1019             .image = destImage,
1020             .viewType = anv_meta_get_view_type(dest_image),
1021             .format = dest_image->vk_format,
1022             .subresourceRange = {
1023                .aspectMask = VK_IMAGE_ASPECT_COLOR_BIT,
1024                .baseMipLevel = pRegions[r].dstSubresource.mipLevel,
1025                .levelCount = 1,
1026                .baseArrayLayer = dest_array_slice,
1027                .layerCount = 1
1028             },
1029          },
1030          cmd_buffer, 0, VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT);
1031
1032       meta_emit_blit(cmd_buffer,
1033                      src_image, &src_iview,
1034                      pRegions[r].srcOffsets[0], src_extent,
1035                      dest_image, &dest_iview,
1036                      dest_offset, dest_extent,
1037                      filter);
1038    }
1039
1040    meta_finish_blit(cmd_buffer, &saved_state);
1041 }
1042
1043 static void
1044 meta_copy_buffer_to_image(struct anv_cmd_buffer *cmd_buffer,
1045                           struct anv_buffer* buffer,
1046                           struct anv_image* image,
1047                           uint32_t regionCount,
1048                           const VkBufferImageCopy* pRegions,
1049                           bool forward)
1050 {
1051    struct anv_meta_saved_state saved_state;
1052
1053    /* The Vulkan 1.0 spec says "dstImage must have a sample count equal to
1054     * VK_SAMPLE_COUNT_1_BIT."
1055     */
1056    assert(image->samples == 1);
1057
1058    anv_meta_begin_blit2d(cmd_buffer, &saved_state);
1059
1060    for (unsigned r = 0; r < regionCount; r++) {
1061
1062       /* Start creating blit rect */
1063       const VkOffset3D img_offset_el = meta_region_offset_el(image, &pRegions[r].imageOffset);
1064       const VkExtent3D bufferExtent = {
1065          .width = pRegions[r].bufferRowLength,
1066          .height = pRegions[r].bufferImageHeight,
1067       };
1068       const VkExtent3D buf_extent_el = meta_region_extent_el(image->vk_format, &bufferExtent);
1069       const VkExtent3D img_extent_el = meta_region_extent_el(image->vk_format,
1070                                                                &pRegions[r].imageExtent);
1071       struct anv_meta_blit2d_rect rect = {
1072          .width = MAX2(buf_extent_el.width, img_extent_el.width),
1073          .height = MAX2(buf_extent_el.height, img_extent_el.height),
1074       };
1075
1076       /* Create blit surfaces */
1077       VkImageAspectFlags aspect = pRegions[r].imageSubresource.aspectMask;
1078       const struct isl_surf *img_isl_surf =
1079          &anv_image_get_surface_for_aspect_mask(image, aspect)->isl;
1080       struct anv_meta_blit2d_surf img_bsurf = blit_surf_for_image(image, img_isl_surf);
1081       struct anv_meta_blit2d_surf buf_bsurf = {
1082          .bo = buffer->bo,
1083          .tiling = ISL_TILING_LINEAR,
1084          .base_offset = buffer->offset + pRegions[r].bufferOffset,
1085          .bs = forward ? image->format->isl_layout->bs : img_bsurf.bs,
1086          .pitch = rect.width * buf_bsurf.bs,
1087       };
1088
1089       /* Set direction-dependent variables */
1090       struct anv_meta_blit2d_surf *dst_bsurf = forward ? &img_bsurf : &buf_bsurf;
1091       struct anv_meta_blit2d_surf *src_bsurf = forward ? &buf_bsurf : &img_bsurf;
1092       uint32_t *x_offset = forward ? &rect.dst_x : &rect.src_x;
1093       uint32_t *y_offset = forward ? &rect.dst_y : &rect.src_y;
1094
1095       /* Loop through each 3D or array slice */
1096       unsigned num_slices_3d = pRegions[r].imageExtent.depth;
1097       unsigned num_slices_array = pRegions[r].imageSubresource.layerCount;
1098       unsigned slice_3d = 0;
1099       unsigned slice_array = 0;
1100       while (slice_3d < num_slices_3d && slice_array < num_slices_array) {
1101
1102          /* Finish creating blit rect */
1103          isl_surf_get_image_offset_el(img_isl_surf,
1104                                     pRegions[r].imageSubresource.mipLevel,
1105                                     pRegions[r].imageSubresource.baseArrayLayer + slice_array,
1106                                     pRegions[r].imageOffset.z + slice_3d,
1107                                     x_offset,
1108                                     y_offset);
1109          *x_offset += img_offset_el.x;
1110          *y_offset += img_offset_el.y;
1111
1112          /* Perform Blit */
1113          anv_meta_blit2d(cmd_buffer,
1114                         src_bsurf,
1115                         dst_bsurf,
1116                         1,
1117                         &rect);
1118
1119          /* Once we've done the blit, all of the actual information about
1120           * the image is embedded in the command buffer so we can just
1121           * increment the offset directly in the image effectively
1122           * re-binding it to different backing memory.
1123           */
1124          buf_bsurf.base_offset += rect.width * rect.height * buf_bsurf.bs;
1125
1126          if (image->type == VK_IMAGE_TYPE_3D)
1127             slice_3d++;
1128          else
1129             slice_array++;
1130       }
1131    }
1132    anv_meta_end_blit2d(cmd_buffer, &saved_state);
1133 }
1134
1135 static struct anv_image *
1136 make_image_for_buffer(VkDevice vk_device, VkBuffer vk_buffer, VkFormat format,
1137                       VkImageUsageFlags usage,
1138                       VkImageType image_type,
1139                       const VkAllocationCallbacks *alloc,
1140                       const VkBufferImageCopy *copy)
1141 {
1142    ANV_FROM_HANDLE(anv_buffer, buffer, vk_buffer);
1143
1144    VkExtent3D extent = copy->imageExtent;
1145    if (copy->bufferRowLength)
1146       extent.width = copy->bufferRowLength;
1147    if (copy->bufferImageHeight)
1148       extent.height = copy->bufferImageHeight;
1149    extent.depth = 1;
1150    extent = meta_region_extent_el(format, &extent);
1151
1152    VkImageAspectFlags aspect = copy->imageSubresource.aspectMask;
1153    VkFormat buffer_format = choose_buffer_format(format, aspect);
1154
1155    VkImage vk_image;
1156    VkResult result = anv_CreateImage(vk_device,
1157       &(VkImageCreateInfo) {
1158          .sType = VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO,
1159          .imageType = VK_IMAGE_TYPE_2D,
1160          .format = buffer_format,
1161          .extent = extent,
1162          .mipLevels = 1,
1163          .arrayLayers = 1,
1164          .samples = 1,
1165          .tiling = VK_IMAGE_TILING_LINEAR,
1166          .usage = usage,
1167          .flags = 0,
1168       }, alloc, &vk_image);
1169    assert(result == VK_SUCCESS);
1170
1171    ANV_FROM_HANDLE(anv_image, image, vk_image);
1172
1173    /* We could use a vk call to bind memory, but that would require
1174     * creating a dummy memory object etc. so there's really no point.
1175     */
1176    image->bo = buffer->bo;
1177    image->offset = buffer->offset + copy->bufferOffset;
1178
1179    return image;
1180 }
1181
1182 void anv_CmdCopyBufferToImage(
1183     VkCommandBuffer                             commandBuffer,
1184     VkBuffer                                    srcBuffer,
1185     VkImage                                     destImage,
1186     VkImageLayout                               destImageLayout,
1187     uint32_t                                    regionCount,
1188     const VkBufferImageCopy*                    pRegions)
1189 {
1190    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
1191    ANV_FROM_HANDLE(anv_image, dest_image, destImage);
1192    ANV_FROM_HANDLE(anv_buffer, src_buffer, srcBuffer);
1193
1194    meta_copy_buffer_to_image(cmd_buffer, src_buffer, dest_image,
1195                              regionCount, pRegions, true);
1196 }
1197
1198 void anv_CmdCopyImageToBuffer(
1199     VkCommandBuffer                             commandBuffer,
1200     VkImage                                     srcImage,
1201     VkImageLayout                               srcImageLayout,
1202     VkBuffer                                    destBuffer,
1203     uint32_t                                    regionCount,
1204     const VkBufferImageCopy*                    pRegions)
1205 {
1206    ANV_FROM_HANDLE(anv_cmd_buffer, cmd_buffer, commandBuffer);
1207    ANV_FROM_HANDLE(anv_image, src_image, srcImage);
1208    ANV_FROM_HANDLE(anv_buffer, dst_buffer, destBuffer);
1209
1210    meta_copy_buffer_to_image(cmd_buffer, dst_buffer, src_image,
1211                              regionCount, pRegions, false);
1212 }
1213
1214 void
1215 anv_device_finish_meta_blit_state(struct anv_device *device)
1216 {
1217    anv_DestroyRenderPass(anv_device_to_handle(device),
1218                          device->meta_state.blit.render_pass,
1219                          &device->meta_state.alloc);
1220    anv_DestroyPipeline(anv_device_to_handle(device),
1221                        device->meta_state.blit.pipeline_1d_src,
1222                        &device->meta_state.alloc);
1223    anv_DestroyPipeline(anv_device_to_handle(device),
1224                        device->meta_state.blit.pipeline_2d_src,
1225                        &device->meta_state.alloc);
1226    anv_DestroyPipeline(anv_device_to_handle(device),
1227                        device->meta_state.blit.pipeline_3d_src,
1228                        &device->meta_state.alloc);
1229    anv_DestroyPipelineLayout(anv_device_to_handle(device),
1230                              device->meta_state.blit.pipeline_layout,
1231                              &device->meta_state.alloc);
1232    anv_DestroyDescriptorSetLayout(anv_device_to_handle(device),
1233                                   device->meta_state.blit.ds_layout,
1234                                   &device->meta_state.alloc);
1235 }
1236
1237 VkResult
1238 anv_device_init_meta_blit_state(struct anv_device *device)
1239 {
1240    VkResult result;
1241
1242    result = anv_CreateRenderPass(anv_device_to_handle(device),
1243       &(VkRenderPassCreateInfo) {
1244          .sType = VK_STRUCTURE_TYPE_RENDER_PASS_CREATE_INFO,
1245          .attachmentCount = 1,
1246          .pAttachments = &(VkAttachmentDescription) {
1247             .format = VK_FORMAT_UNDEFINED, /* Our shaders don't care */
1248             .loadOp = VK_ATTACHMENT_LOAD_OP_LOAD,
1249             .storeOp = VK_ATTACHMENT_STORE_OP_STORE,
1250             .initialLayout = VK_IMAGE_LAYOUT_GENERAL,
1251             .finalLayout = VK_IMAGE_LAYOUT_GENERAL,
1252          },
1253          .subpassCount = 1,
1254          .pSubpasses = &(VkSubpassDescription) {
1255             .pipelineBindPoint = VK_PIPELINE_BIND_POINT_GRAPHICS,
1256             .inputAttachmentCount = 0,
1257             .colorAttachmentCount = 1,
1258             .pColorAttachments = &(VkAttachmentReference) {
1259                .attachment = 0,
1260                .layout = VK_IMAGE_LAYOUT_GENERAL,
1261             },
1262             .pResolveAttachments = NULL,
1263             .pDepthStencilAttachment = &(VkAttachmentReference) {
1264                .attachment = VK_ATTACHMENT_UNUSED,
1265                .layout = VK_IMAGE_LAYOUT_GENERAL,
1266             },
1267             .preserveAttachmentCount = 1,
1268             .pPreserveAttachments = (uint32_t[]) { 0 },
1269          },
1270          .dependencyCount = 0,
1271       }, &device->meta_state.alloc, &device->meta_state.blit.render_pass);
1272    if (result != VK_SUCCESS)
1273       goto fail;
1274
1275    /* We don't use a vertex shader for blitting, but instead build and pass
1276     * the VUEs directly to the rasterization backend.  However, we do need
1277     * to provide GLSL source for the vertex shader so that the compiler
1278     * does not dead-code our inputs.
1279     */
1280    struct anv_shader_module vs = {
1281       .nir = build_nir_vertex_shader(),
1282    };
1283
1284    struct anv_shader_module fs_1d = {
1285       .nir = build_nir_copy_fragment_shader(GLSL_SAMPLER_DIM_1D),
1286    };
1287
1288    struct anv_shader_module fs_2d = {
1289       .nir = build_nir_copy_fragment_shader(GLSL_SAMPLER_DIM_2D),
1290    };
1291
1292    struct anv_shader_module fs_3d = {
1293       .nir = build_nir_copy_fragment_shader(GLSL_SAMPLER_DIM_3D),
1294    };
1295
1296    VkPipelineVertexInputStateCreateInfo vi_create_info = {
1297       .sType = VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO,
1298       .vertexBindingDescriptionCount = 2,
1299       .pVertexBindingDescriptions = (VkVertexInputBindingDescription[]) {
1300          {
1301             .binding = 0,
1302             .stride = 0,
1303             .inputRate = VK_VERTEX_INPUT_RATE_VERTEX
1304          },
1305          {
1306             .binding = 1,
1307             .stride = 5 * sizeof(float),
1308             .inputRate = VK_VERTEX_INPUT_RATE_VERTEX
1309          },
1310       },
1311       .vertexAttributeDescriptionCount = 3,
1312       .pVertexAttributeDescriptions = (VkVertexInputAttributeDescription[]) {
1313          {
1314             /* VUE Header */
1315             .location = 0,
1316             .binding = 0,
1317             .format = VK_FORMAT_R32G32B32A32_UINT,
1318             .offset = 0
1319          },
1320          {
1321             /* Position */
1322             .location = 1,
1323             .binding = 1,
1324             .format = VK_FORMAT_R32G32_SFLOAT,
1325             .offset = 0
1326          },
1327          {
1328             /* Texture Coordinate */
1329             .location = 2,
1330             .binding = 1,
1331             .format = VK_FORMAT_R32G32B32_SFLOAT,
1332             .offset = 8
1333          }
1334       }
1335    };
1336
1337    VkDescriptorSetLayoutCreateInfo ds_layout_info = {
1338       .sType = VK_STRUCTURE_TYPE_DESCRIPTOR_SET_LAYOUT_CREATE_INFO,
1339       .bindingCount = 1,
1340       .pBindings = (VkDescriptorSetLayoutBinding[]) {
1341          {
1342             .binding = 0,
1343             .descriptorType = VK_DESCRIPTOR_TYPE_COMBINED_IMAGE_SAMPLER,
1344             .descriptorCount = 1,
1345             .stageFlags = VK_SHADER_STAGE_FRAGMENT_BIT,
1346             .pImmutableSamplers = NULL
1347          },
1348       }
1349    };
1350    result = anv_CreateDescriptorSetLayout(anv_device_to_handle(device),
1351                                           &ds_layout_info,
1352                                           &device->meta_state.alloc,
1353                                           &device->meta_state.blit.ds_layout);
1354    if (result != VK_SUCCESS)
1355       goto fail_render_pass;
1356
1357    result = anv_CreatePipelineLayout(anv_device_to_handle(device),
1358       &(VkPipelineLayoutCreateInfo) {
1359          .sType = VK_STRUCTURE_TYPE_PIPELINE_LAYOUT_CREATE_INFO,
1360          .setLayoutCount = 1,
1361          .pSetLayouts = &device->meta_state.blit.ds_layout,
1362       },
1363       &device->meta_state.alloc, &device->meta_state.blit.pipeline_layout);
1364    if (result != VK_SUCCESS)
1365       goto fail_descriptor_set_layout;
1366
1367    VkPipelineShaderStageCreateInfo pipeline_shader_stages[] = {
1368       {
1369          .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
1370          .stage = VK_SHADER_STAGE_VERTEX_BIT,
1371          .module = anv_shader_module_to_handle(&vs),
1372          .pName = "main",
1373          .pSpecializationInfo = NULL
1374       }, {
1375          .sType = VK_STRUCTURE_TYPE_PIPELINE_SHADER_STAGE_CREATE_INFO,
1376          .stage = VK_SHADER_STAGE_FRAGMENT_BIT,
1377          .module = VK_NULL_HANDLE, /* TEMPLATE VALUE! FILL ME IN! */
1378          .pName = "main",
1379          .pSpecializationInfo = NULL
1380       },
1381    };
1382
1383    const VkGraphicsPipelineCreateInfo vk_pipeline_info = {
1384       .sType = VK_STRUCTURE_TYPE_GRAPHICS_PIPELINE_CREATE_INFO,
1385       .stageCount = ARRAY_SIZE(pipeline_shader_stages),
1386       .pStages = pipeline_shader_stages,
1387       .pVertexInputState = &vi_create_info,
1388       .pInputAssemblyState = &(VkPipelineInputAssemblyStateCreateInfo) {
1389          .sType = VK_STRUCTURE_TYPE_PIPELINE_INPUT_ASSEMBLY_STATE_CREATE_INFO,
1390          .topology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP,
1391          .primitiveRestartEnable = false,
1392       },
1393       .pViewportState = &(VkPipelineViewportStateCreateInfo) {
1394          .sType = VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO,
1395          .viewportCount = 1,
1396          .scissorCount = 1,
1397       },
1398       .pRasterizationState = &(VkPipelineRasterizationStateCreateInfo) {
1399          .sType = VK_STRUCTURE_TYPE_PIPELINE_RASTERIZATION_STATE_CREATE_INFO,
1400          .rasterizerDiscardEnable = false,
1401          .polygonMode = VK_POLYGON_MODE_FILL,
1402          .cullMode = VK_CULL_MODE_NONE,
1403          .frontFace = VK_FRONT_FACE_COUNTER_CLOCKWISE
1404       },
1405       .pMultisampleState = &(VkPipelineMultisampleStateCreateInfo) {
1406          .sType = VK_STRUCTURE_TYPE_PIPELINE_MULTISAMPLE_STATE_CREATE_INFO,
1407          .rasterizationSamples = 1,
1408          .sampleShadingEnable = false,
1409          .pSampleMask = (VkSampleMask[]) { UINT32_MAX },
1410       },
1411       .pColorBlendState = &(VkPipelineColorBlendStateCreateInfo) {
1412          .sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO,
1413          .attachmentCount = 1,
1414          .pAttachments = (VkPipelineColorBlendAttachmentState []) {
1415             { .colorWriteMask =
1416                  VK_COLOR_COMPONENT_A_BIT |
1417                  VK_COLOR_COMPONENT_R_BIT |
1418                  VK_COLOR_COMPONENT_G_BIT |
1419                  VK_COLOR_COMPONENT_B_BIT },
1420          }
1421       },
1422       .pDynamicState = &(VkPipelineDynamicStateCreateInfo) {
1423          .sType = VK_STRUCTURE_TYPE_PIPELINE_DYNAMIC_STATE_CREATE_INFO,
1424          .dynamicStateCount = 9,
1425          .pDynamicStates = (VkDynamicState[]) {
1426             VK_DYNAMIC_STATE_VIEWPORT,
1427             VK_DYNAMIC_STATE_SCISSOR,
1428             VK_DYNAMIC_STATE_LINE_WIDTH,
1429             VK_DYNAMIC_STATE_DEPTH_BIAS,
1430             VK_DYNAMIC_STATE_BLEND_CONSTANTS,
1431             VK_DYNAMIC_STATE_DEPTH_BOUNDS,
1432             VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK,
1433             VK_DYNAMIC_STATE_STENCIL_WRITE_MASK,
1434             VK_DYNAMIC_STATE_STENCIL_REFERENCE,
1435          },
1436       },
1437       .flags = 0,
1438       .layout = device->meta_state.blit.pipeline_layout,
1439       .renderPass = device->meta_state.blit.render_pass,
1440       .subpass = 0,
1441    };
1442
1443    const struct anv_graphics_pipeline_create_info anv_pipeline_info = {
1444       .color_attachment_count = -1,
1445       .use_repclear = false,
1446       .disable_viewport = true,
1447       .disable_scissor = true,
1448       .disable_vs = true,
1449       .use_rectlist = true
1450    };
1451
1452    pipeline_shader_stages[1].module = anv_shader_module_to_handle(&fs_1d);
1453    result = anv_graphics_pipeline_create(anv_device_to_handle(device),
1454       VK_NULL_HANDLE,
1455       &vk_pipeline_info, &anv_pipeline_info,
1456       &device->meta_state.alloc, &device->meta_state.blit.pipeline_1d_src);
1457    if (result != VK_SUCCESS)
1458       goto fail_pipeline_layout;
1459
1460    pipeline_shader_stages[1].module = anv_shader_module_to_handle(&fs_2d);
1461    result = anv_graphics_pipeline_create(anv_device_to_handle(device),
1462       VK_NULL_HANDLE,
1463       &vk_pipeline_info, &anv_pipeline_info,
1464       &device->meta_state.alloc, &device->meta_state.blit.pipeline_2d_src);
1465    if (result != VK_SUCCESS)
1466       goto fail_pipeline_1d;
1467
1468    pipeline_shader_stages[1].module = anv_shader_module_to_handle(&fs_3d);
1469    result = anv_graphics_pipeline_create(anv_device_to_handle(device),
1470       VK_NULL_HANDLE,
1471       &vk_pipeline_info, &anv_pipeline_info,
1472       &device->meta_state.alloc, &device->meta_state.blit.pipeline_3d_src);
1473    if (result != VK_SUCCESS)
1474       goto fail_pipeline_2d;
1475
1476    ralloc_free(vs.nir);
1477    ralloc_free(fs_1d.nir);
1478    ralloc_free(fs_2d.nir);
1479    ralloc_free(fs_3d.nir);
1480
1481    return VK_SUCCESS;
1482
1483  fail_pipeline_2d:
1484    anv_DestroyPipeline(anv_device_to_handle(device),
1485                        device->meta_state.blit.pipeline_2d_src,
1486                        &device->meta_state.alloc);
1487
1488  fail_pipeline_1d:
1489    anv_DestroyPipeline(anv_device_to_handle(device),
1490                        device->meta_state.blit.pipeline_1d_src,
1491                        &device->meta_state.alloc);
1492
1493  fail_pipeline_layout:
1494    anv_DestroyPipelineLayout(anv_device_to_handle(device),
1495                              device->meta_state.blit.pipeline_layout,
1496                              &device->meta_state.alloc);
1497  fail_descriptor_set_layout:
1498    anv_DestroyDescriptorSetLayout(anv_device_to_handle(device),
1499                                   device->meta_state.blit.ds_layout,
1500                                   &device->meta_state.alloc);
1501  fail_render_pass:
1502    anv_DestroyRenderPass(anv_device_to_handle(device),
1503                          device->meta_state.blit.render_pass,
1504                          &device->meta_state.alloc);
1505
1506    ralloc_free(vs.nir);
1507    ralloc_free(fs_1d.nir);
1508    ralloc_free(fs_2d.nir);
1509    ralloc_free(fs_3d.nir);
1510  fail:
1511    return result;
1512 }