1/* 2 * Copyright © 2016 Red Hat. 3 * Copyright © 2016 Bas Nieuwenhuizen 4 * 5 * based in part on anv driver which is: 6 * Copyright © 2015 Intel Corporation 7 * 8 * Permission is hereby granted, free of charge, to any person obtaining a 9 * copy of this software and associated documentation files (the "Software"), 10 * to deal in the Software without restriction, including without limitation 11 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 12 * and/or sell copies of the Software, and to permit persons to whom the 13 * Software is furnished to do so, subject to the following conditions: 14 * 15 * The above copyright notice and this permission notice (including the next 16 * paragraph) shall be included in all copies or substantial portions of the 17 * Software. 18 * 19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER 25 * DEALINGS IN THE SOFTWARE. 26 */ 27 28#include "tu_private.h" 29#include "fdl/fd6_format_table.h" 30 31#include "util/debug.h" 32#include "util/u_atomic.h" 33#include "util/format/u_format.h" 34#include "vk_format.h" 35#include "vk_util.h" 36#include "drm-uapi/drm_fourcc.h" 37 38#include "tu_cs.h" 39 40static uint32_t 41tu6_plane_count(VkFormat format) 42{ 43 switch (format) { 44 default: 45 return 1; 46 case VK_FORMAT_G8_B8R8_2PLANE_420_UNORM: 47 case VK_FORMAT_D32_SFLOAT_S8_UINT: 48 return 2; 49 case VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM: 50 return 3; 51 } 52} 53 54static VkFormat 55tu6_plane_format(VkFormat format, uint32_t plane) 56{ 57 switch (format) { 58 case VK_FORMAT_G8_B8R8_2PLANE_420_UNORM: 59 /* note: with UBWC, and Y plane UBWC is different from R8_UNORM */ 60 return plane ? VK_FORMAT_R8G8_UNORM : VK_FORMAT_R8_UNORM; 61 case VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM: 62 return VK_FORMAT_R8_UNORM; 63 case VK_FORMAT_D32_SFLOAT_S8_UINT: 64 return plane ? VK_FORMAT_S8_UINT : VK_FORMAT_D32_SFLOAT; 65 default: 66 return format; 67 } 68} 69 70static uint32_t 71tu6_plane_index(VkFormat format, VkImageAspectFlags aspect_mask) 72{ 73 switch (aspect_mask) { 74 default: 75 return 0; 76 case VK_IMAGE_ASPECT_PLANE_1_BIT: 77 return 1; 78 case VK_IMAGE_ASPECT_PLANE_2_BIT: 79 return 2; 80 case VK_IMAGE_ASPECT_STENCIL_BIT: 81 return format == VK_FORMAT_D32_SFLOAT_S8_UINT; 82 } 83} 84 85static void 86compose_swizzle(unsigned char *swiz, const VkComponentMapping *mapping) 87{ 88 unsigned char src_swiz[4] = { swiz[0], swiz[1], swiz[2], swiz[3] }; 89 VkComponentSwizzle vk_swiz[4] = { 90 mapping->r, mapping->g, mapping->b, mapping->a 91 }; 92 for (int i = 0; i < 4; i++) { 93 switch (vk_swiz[i]) { 94 case VK_COMPONENT_SWIZZLE_IDENTITY: 95 swiz[i] = src_swiz[i]; 96 break; 97 case VK_COMPONENT_SWIZZLE_R...VK_COMPONENT_SWIZZLE_A: 98 swiz[i] = src_swiz[vk_swiz[i] - VK_COMPONENT_SWIZZLE_R]; 99 break; 100 case VK_COMPONENT_SWIZZLE_ZERO: 101 swiz[i] = A6XX_TEX_ZERO; 102 break; 103 case VK_COMPONENT_SWIZZLE_ONE: 104 swiz[i] = A6XX_TEX_ONE; 105 break; 106 default: 107 unreachable("unexpected swizzle"); 108 } 109 } 110} 111 112static uint32_t 113tu6_texswiz(const VkComponentMapping *comps, 114 const struct tu_sampler_ycbcr_conversion *conversion, 115 VkFormat format, 116 VkImageAspectFlagBits aspect_mask, 117 bool has_z24uint_s8uint) 118{ 119 unsigned char swiz[4] = { 120 A6XX_TEX_X, A6XX_TEX_Y, A6XX_TEX_Z, A6XX_TEX_W, 121 }; 122 123 switch (format) { 124 case VK_FORMAT_G8B8G8R8_422_UNORM: 125 case VK_FORMAT_B8G8R8G8_422_UNORM: 126 case VK_FORMAT_G8_B8R8_2PLANE_420_UNORM: 127 case VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM: 128 swiz[0] = A6XX_TEX_Z; 129 swiz[1] = A6XX_TEX_X; 130 swiz[2] = A6XX_TEX_Y; 131 break; 132 case VK_FORMAT_BC1_RGB_UNORM_BLOCK: 133 case VK_FORMAT_BC1_RGB_SRGB_BLOCK: 134 /* same hardware format is used for BC1_RGB / BC1_RGBA */ 135 swiz[3] = A6XX_TEX_ONE; 136 break; 137 case VK_FORMAT_D24_UNORM_S8_UINT: 138 if (aspect_mask == VK_IMAGE_ASPECT_STENCIL_BIT) { 139 if (!has_z24uint_s8uint) { 140 /* using FMT6_8_8_8_8_UINT, so need to pick out the W channel and 141 * swizzle (0,0,1) in the rest (see "Conversion to RGBA"). 142 */ 143 swiz[0] = A6XX_TEX_W; 144 swiz[1] = A6XX_TEX_ZERO; 145 swiz[2] = A6XX_TEX_ZERO; 146 swiz[3] = A6XX_TEX_ONE; 147 } else { 148 /* using FMT6_Z24_UINT_S8_UINT, which is (d, s, 0, 1), so need to 149 * swizzle away the d. 150 */ 151 swiz[0] = A6XX_TEX_Y; 152 swiz[1] = A6XX_TEX_ZERO; 153 } 154 } 155 break; 156 default: 157 break; 158 } 159 160 compose_swizzle(swiz, comps); 161 if (conversion) 162 compose_swizzle(swiz, &conversion->components); 163 164 return A6XX_TEX_CONST_0_SWIZ_X(swiz[0]) | 165 A6XX_TEX_CONST_0_SWIZ_Y(swiz[1]) | 166 A6XX_TEX_CONST_0_SWIZ_Z(swiz[2]) | 167 A6XX_TEX_CONST_0_SWIZ_W(swiz[3]); 168} 169 170void 171tu_cs_image_ref(struct tu_cs *cs, const struct tu_image_view *iview, uint32_t layer) 172{ 173 tu_cs_emit(cs, iview->PITCH); 174 tu_cs_emit(cs, iview->layer_size >> 6); 175 tu_cs_emit_qw(cs, iview->base_addr + iview->layer_size * layer); 176} 177 178void 179tu_cs_image_stencil_ref(struct tu_cs *cs, const struct tu_image_view *iview, uint32_t layer) 180{ 181 tu_cs_emit(cs, iview->stencil_PITCH); 182 tu_cs_emit(cs, iview->stencil_layer_size >> 6); 183 tu_cs_emit_qw(cs, iview->stencil_base_addr + iview->stencil_layer_size * layer); 184} 185 186void 187tu_cs_image_ref_2d(struct tu_cs *cs, const struct tu_image_view *iview, uint32_t layer, bool src) 188{ 189 tu_cs_emit_qw(cs, iview->base_addr + iview->layer_size * layer); 190 /* SP_PS_2D_SRC_PITCH has shifted pitch field */ 191 tu_cs_emit(cs, iview->PITCH << (src ? 9 : 0)); 192} 193 194void 195tu_cs_image_flag_ref(struct tu_cs *cs, const struct tu_image_view *iview, uint32_t layer) 196{ 197 tu_cs_emit_qw(cs, iview->ubwc_addr + iview->ubwc_layer_size * layer); 198 tu_cs_emit(cs, iview->FLAG_BUFFER_PITCH); 199} 200 201void 202tu_image_view_init(struct tu_image_view *iview, 203 const VkImageViewCreateInfo *pCreateInfo, 204 bool has_z24uint_s8uint) 205{ 206 TU_FROM_HANDLE(tu_image, image, pCreateInfo->image); 207 const VkImageSubresourceRange *range = &pCreateInfo->subresourceRange; 208 VkFormat format = pCreateInfo->format; 209 VkImageAspectFlagBits aspect_mask = pCreateInfo->subresourceRange.aspectMask; 210 211 const struct VkSamplerYcbcrConversionInfo *ycbcr_conversion = 212 vk_find_struct_const(pCreateInfo->pNext, SAMPLER_YCBCR_CONVERSION_INFO); 213 const struct tu_sampler_ycbcr_conversion *conversion = ycbcr_conversion ? 214 tu_sampler_ycbcr_conversion_from_handle(ycbcr_conversion->conversion) : NULL; 215 216 iview->image = image; 217 218 memset(iview->descriptor, 0, sizeof(iview->descriptor)); 219 220 struct fdl_layout *layout = 221 &image->layout[tu6_plane_index(image->vk_format, aspect_mask)]; 222 223 uint32_t width = u_minify(layout->width0, range->baseMipLevel); 224 uint32_t height = u_minify(layout->height0, range->baseMipLevel); 225 uint32_t storage_depth = tu_get_layerCount(image, range); 226 if (pCreateInfo->viewType == VK_IMAGE_VIEW_TYPE_3D) { 227 storage_depth = u_minify(image->layout[0].depth0, range->baseMipLevel); 228 } 229 230 uint32_t depth = storage_depth; 231 if (pCreateInfo->viewType == VK_IMAGE_VIEW_TYPE_CUBE || 232 pCreateInfo->viewType == VK_IMAGE_VIEW_TYPE_CUBE_ARRAY) { 233 /* Cubes are treated as 2D arrays for storage images, so only divide the 234 * depth by 6 for the texture descriptor. 235 */ 236 depth /= 6; 237 } 238 239 uint64_t base_addr = image->bo->iova + image->bo_offset + 240 fdl_surface_offset(layout, range->baseMipLevel, range->baseArrayLayer); 241 uint64_t ubwc_addr = image->bo->iova + image->bo_offset + 242 fdl_ubwc_offset(layout, range->baseMipLevel, range->baseArrayLayer); 243 244 uint32_t pitch = fdl_pitch(layout, range->baseMipLevel); 245 uint32_t ubwc_pitch = fdl_ubwc_pitch(layout, range->baseMipLevel); 246 uint32_t layer_size = fdl_layer_stride(layout, range->baseMipLevel); 247 248 if (aspect_mask != VK_IMAGE_ASPECT_COLOR_BIT) 249 format = tu6_plane_format(format, tu6_plane_index(format, aspect_mask)); 250 251 struct tu_native_format fmt = tu6_format_texture(format, layout->tile_mode); 252 /* note: freedreno layout assumes no TILE_ALL bit for non-UBWC color formats 253 * this means smaller mipmap levels have a linear tile mode. 254 * Depth/stencil formats have non-linear tile mode. 255 */ 256 fmt.tile_mode = fdl_tile_mode(layout, range->baseMipLevel); 257 258 bool ubwc_enabled = fdl_ubwc_enabled(layout, range->baseMipLevel); 259 260 bool is_d24s8 = (format == VK_FORMAT_D24_UNORM_S8_UINT || 261 format == VK_FORMAT_X8_D24_UNORM_PACK32); 262 263 if (is_d24s8 && ubwc_enabled) 264 fmt.fmt = FMT6_Z24_UNORM_S8_UINT_AS_R8G8B8A8; 265 266 unsigned fmt_tex = fmt.fmt; 267 if (is_d24s8) { 268 if (aspect_mask & VK_IMAGE_ASPECT_DEPTH_BIT) 269 fmt_tex = FMT6_Z24_UNORM_S8_UINT; 270 if (aspect_mask == VK_IMAGE_ASPECT_STENCIL_BIT) 271 fmt_tex = has_z24uint_s8uint ? FMT6_Z24_UINT_S8_UINT : FMT6_8_8_8_8_UINT; 272 /* TODO: also use this format with storage descriptor ? */ 273 } 274 275 iview->descriptor[0] = 276 A6XX_TEX_CONST_0_TILE_MODE(fmt.tile_mode) | 277 COND(vk_format_is_srgb(format), A6XX_TEX_CONST_0_SRGB) | 278 A6XX_TEX_CONST_0_FMT(fmt_tex) | 279 A6XX_TEX_CONST_0_SAMPLES(tu_msaa_samples(layout->nr_samples)) | 280 A6XX_TEX_CONST_0_SWAP(fmt.swap) | 281 tu6_texswiz(&pCreateInfo->components, conversion, format, aspect_mask, has_z24uint_s8uint) | 282 A6XX_TEX_CONST_0_MIPLVLS(tu_get_levelCount(image, range) - 1); 283 iview->descriptor[1] = A6XX_TEX_CONST_1_WIDTH(width) | A6XX_TEX_CONST_1_HEIGHT(height); 284 iview->descriptor[2] = 285 A6XX_TEX_CONST_2_PITCHALIGN(layout->pitchalign - 6) | 286 A6XX_TEX_CONST_2_PITCH(pitch) | 287 A6XX_TEX_CONST_2_TYPE(tu6_tex_type(pCreateInfo->viewType, false)); 288 iview->descriptor[3] = A6XX_TEX_CONST_3_ARRAY_PITCH(layer_size); 289 iview->descriptor[4] = base_addr; 290 iview->descriptor[5] = (base_addr >> 32) | A6XX_TEX_CONST_5_DEPTH(depth); 291 292 if (layout->tile_all) 293 iview->descriptor[3] |= A6XX_TEX_CONST_3_TILE_ALL; 294 295 if (format == VK_FORMAT_G8_B8R8_2PLANE_420_UNORM || 296 format == VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM) { 297 /* chroma offset re-uses MIPLVLS bits */ 298 assert(tu_get_levelCount(image, range) == 1); 299 if (conversion) { 300 if (conversion->chroma_offsets[0] == VK_CHROMA_LOCATION_MIDPOINT) 301 iview->descriptor[0] |= A6XX_TEX_CONST_0_CHROMA_MIDPOINT_X; 302 if (conversion->chroma_offsets[1] == VK_CHROMA_LOCATION_MIDPOINT) 303 iview->descriptor[0] |= A6XX_TEX_CONST_0_CHROMA_MIDPOINT_Y; 304 } 305 306 uint64_t base_addr[3]; 307 308 iview->descriptor[3] |= A6XX_TEX_CONST_3_TILE_ALL; 309 if (ubwc_enabled) { 310 iview->descriptor[3] |= A6XX_TEX_CONST_3_FLAG; 311 /* no separate ubwc base, image must have the expected layout */ 312 for (uint32_t i = 0; i < 3; i++) { 313 base_addr[i] = image->bo->iova + image->bo_offset + 314 fdl_ubwc_offset(&image->layout[i], range->baseMipLevel, range->baseArrayLayer); 315 } 316 } else { 317 for (uint32_t i = 0; i < 3; i++) { 318 base_addr[i] = image->bo->iova + image->bo_offset + 319 fdl_surface_offset(&image->layout[i], range->baseMipLevel, range->baseArrayLayer); 320 } 321 } 322 323 iview->descriptor[4] = base_addr[0]; 324 iview->descriptor[5] |= base_addr[0] >> 32; 325 iview->descriptor[6] = 326 A6XX_TEX_CONST_6_PLANE_PITCH(fdl_pitch(&image->layout[1], range->baseMipLevel)); 327 iview->descriptor[7] = base_addr[1]; 328 iview->descriptor[8] = base_addr[1] >> 32; 329 iview->descriptor[9] = base_addr[2]; 330 iview->descriptor[10] = base_addr[2] >> 32; 331 332 assert(pCreateInfo->viewType != VK_IMAGE_VIEW_TYPE_3D); 333 return; 334 } 335 336 if (ubwc_enabled) { 337 uint32_t block_width, block_height; 338 fdl6_get_ubwc_blockwidth(layout, &block_width, &block_height); 339 340 iview->descriptor[3] |= A6XX_TEX_CONST_3_FLAG; 341 iview->descriptor[7] = ubwc_addr; 342 iview->descriptor[8] = ubwc_addr >> 32; 343 iview->descriptor[9] |= A6XX_TEX_CONST_9_FLAG_BUFFER_ARRAY_PITCH(layout->ubwc_layer_size >> 2); 344 iview->descriptor[10] |= 345 A6XX_TEX_CONST_10_FLAG_BUFFER_PITCH(ubwc_pitch) | 346 A6XX_TEX_CONST_10_FLAG_BUFFER_LOGW(util_logbase2_ceil(DIV_ROUND_UP(width, block_width))) | 347 A6XX_TEX_CONST_10_FLAG_BUFFER_LOGH(util_logbase2_ceil(DIV_ROUND_UP(height, block_height))); 348 } 349 350 if (pCreateInfo->viewType == VK_IMAGE_VIEW_TYPE_3D) { 351 iview->descriptor[3] |= 352 A6XX_TEX_CONST_3_MIN_LAYERSZ(layout->slices[image->level_count - 1].size0); 353 } 354 355 iview->SP_PS_2D_SRC_INFO = A6XX_SP_PS_2D_SRC_INFO( 356 .color_format = fmt.fmt, 357 .tile_mode = fmt.tile_mode, 358 .color_swap = fmt.swap, 359 .flags = ubwc_enabled, 360 .srgb = vk_format_is_srgb(format), 361 .samples = tu_msaa_samples(layout->nr_samples), 362 .samples_average = layout->nr_samples > 1 && 363 !vk_format_is_int(format) && 364 !vk_format_is_depth_or_stencil(format), 365 .unk20 = 1, 366 .unk22 = 1).value; 367 iview->SP_PS_2D_SRC_SIZE = 368 A6XX_SP_PS_2D_SRC_SIZE(.width = width, .height = height).value; 369 370 /* note: these have same encoding for MRT and 2D (except 2D PITCH src) */ 371 iview->PITCH = A6XX_RB_DEPTH_BUFFER_PITCH(pitch).value; 372 iview->FLAG_BUFFER_PITCH = A6XX_RB_DEPTH_FLAG_BUFFER_PITCH( 373 .pitch = ubwc_pitch, .array_pitch = layout->ubwc_layer_size >> 2).value; 374 375 iview->base_addr = base_addr; 376 iview->ubwc_addr = ubwc_addr; 377 iview->layer_size = layer_size; 378 iview->ubwc_layer_size = layout->ubwc_layer_size; 379 380 /* Don't set fields that are only used for attachments/blit dest if COLOR 381 * is unsupported. 382 */ 383 if (!tu6_format_color_supported(format)) 384 return; 385 386 struct tu_native_format cfmt = tu6_format_color(format, layout->tile_mode); 387 cfmt.tile_mode = fmt.tile_mode; 388 389 if (is_d24s8 && ubwc_enabled) 390 cfmt.fmt = FMT6_Z24_UNORM_S8_UINT_AS_R8G8B8A8; 391 392 memset(iview->storage_descriptor, 0, sizeof(iview->storage_descriptor)); 393 394 iview->storage_descriptor[0] = 395 A6XX_IBO_0_FMT(fmt.fmt) | 396 A6XX_IBO_0_TILE_MODE(fmt.tile_mode); 397 iview->storage_descriptor[1] = 398 A6XX_IBO_1_WIDTH(width) | 399 A6XX_IBO_1_HEIGHT(height); 400 iview->storage_descriptor[2] = 401 A6XX_IBO_2_PITCH(pitch) | 402 A6XX_IBO_2_TYPE(tu6_tex_type(pCreateInfo->viewType, true)); 403 iview->storage_descriptor[3] = A6XX_IBO_3_ARRAY_PITCH(layer_size); 404 405 iview->storage_descriptor[4] = base_addr; 406 iview->storage_descriptor[5] = (base_addr >> 32) | A6XX_IBO_5_DEPTH(storage_depth); 407 408 if (ubwc_enabled) { 409 iview->storage_descriptor[3] |= A6XX_IBO_3_FLAG | A6XX_IBO_3_UNK27; 410 iview->storage_descriptor[7] |= ubwc_addr; 411 iview->storage_descriptor[8] |= ubwc_addr >> 32; 412 iview->storage_descriptor[9] = A6XX_IBO_9_FLAG_BUFFER_ARRAY_PITCH(layout->ubwc_layer_size >> 2); 413 iview->storage_descriptor[10] = 414 A6XX_IBO_10_FLAG_BUFFER_PITCH(ubwc_pitch); 415 } 416 417 iview->extent.width = width; 418 iview->extent.height = height; 419 iview->need_y2_align = 420 (fmt.tile_mode == TILE6_LINEAR && range->baseMipLevel != image->level_count - 1); 421 422 iview->ubwc_enabled = ubwc_enabled; 423 424 iview->RB_MRT_BUF_INFO = A6XX_RB_MRT_BUF_INFO(0, 425 .color_tile_mode = cfmt.tile_mode, 426 .color_format = cfmt.fmt, 427 .color_swap = cfmt.swap).value; 428 429 iview->SP_FS_MRT_REG = A6XX_SP_FS_MRT_REG(0, 430 .color_format = cfmt.fmt, 431 .color_sint = vk_format_is_sint(format), 432 .color_uint = vk_format_is_uint(format)).value; 433 434 iview->RB_2D_DST_INFO = A6XX_RB_2D_DST_INFO( 435 .color_format = cfmt.fmt, 436 .tile_mode = cfmt.tile_mode, 437 .color_swap = cfmt.swap, 438 .flags = ubwc_enabled, 439 .srgb = vk_format_is_srgb(format)).value; 440 441 iview->RB_BLIT_DST_INFO = A6XX_RB_BLIT_DST_INFO( 442 .tile_mode = cfmt.tile_mode, 443 .samples = tu_msaa_samples(layout->nr_samples), 444 .color_format = cfmt.fmt, 445 .color_swap = cfmt.swap, 446 .flags = ubwc_enabled).value; 447 448 if (image->vk_format == VK_FORMAT_D32_SFLOAT_S8_UINT) { 449 layout = &image->layout[1]; 450 iview->stencil_base_addr = image->bo->iova + image->bo_offset + 451 fdl_surface_offset(layout, range->baseMipLevel, range->baseArrayLayer); 452 iview->stencil_layer_size = fdl_layer_stride(layout, range->baseMipLevel); 453 iview->stencil_PITCH = A6XX_RB_STENCIL_BUFFER_PITCH(fdl_pitch(layout, range->baseMipLevel)).value; 454 } 455} 456 457bool 458ubwc_possible(VkFormat format, VkImageType type, VkImageUsageFlags usage, 459 VkImageUsageFlags stencil_usage, const struct fd_dev_info *info, 460 VkSampleCountFlagBits samples) 461{ 462 /* no UBWC with compressed formats, E5B9G9R9, S8_UINT 463 * (S8_UINT because separate stencil doesn't have UBWC-enable bit) 464 */ 465 if (vk_format_is_compressed(format) || 466 format == VK_FORMAT_E5B9G9R9_UFLOAT_PACK32 || 467 format == VK_FORMAT_S8_UINT) 468 return false; 469 470 if (!info->a6xx.has_8bpp_ubwc && 471 (format == VK_FORMAT_R8_UNORM || 472 format == VK_FORMAT_R8_SNORM || 473 format == VK_FORMAT_R8_UINT || 474 format == VK_FORMAT_R8_SINT || 475 format == VK_FORMAT_R8_SRGB)) 476 return false; 477 478 if (type == VK_IMAGE_TYPE_3D) { 479 tu_finishme("UBWC with 3D textures"); 480 return false; 481 } 482 483 /* Disable UBWC for storage images. 484 * 485 * The closed GL driver skips UBWC for storage images (and additionally 486 * uses linear for writeonly images). We seem to have image tiling working 487 * in freedreno in general, so turnip matches that. freedreno also enables 488 * UBWC on images, but it's not really tested due to the lack of 489 * UBWC-enabled mipmaps in freedreno currently. Just match the closed GL 490 * behavior of no UBWC. 491 */ 492 if ((usage | stencil_usage) & VK_IMAGE_USAGE_STORAGE_BIT) 493 return false; 494 495 /* Disable UBWC for D24S8 on A630 in some cases 496 * 497 * VK_IMAGE_ASPECT_STENCIL_BIT image view requires to be able to sample 498 * from the stencil component as UINT, however no format allows this 499 * on a630 (the special FMT6_Z24_UINT_S8_UINT format is missing) 500 * 501 * It must be sampled as FMT6_8_8_8_8_UINT, which is not UBWC-compatible 502 * 503 * Additionally, the special AS_R8G8B8A8 format is broken without UBWC, 504 * so we have to fallback to 8_8_8_8_UNORM when UBWC is disabled 505 */ 506 if (!info->a6xx.has_z24uint_s8uint && 507 format == VK_FORMAT_D24_UNORM_S8_UINT && 508 (stencil_usage & (VK_IMAGE_USAGE_SAMPLED_BIT | VK_IMAGE_USAGE_INPUT_ATTACHMENT_BIT))) 509 return false; 510 511 if (!info->a6xx.has_z24uint_s8uint && samples > VK_SAMPLE_COUNT_1_BIT) 512 return false; 513 514 return true; 515} 516 517VKAPI_ATTR VkResult VKAPI_CALL 518tu_CreateImage(VkDevice _device, 519 const VkImageCreateInfo *pCreateInfo, 520 const VkAllocationCallbacks *alloc, 521 VkImage *pImage) 522{ 523 TU_FROM_HANDLE(tu_device, device, _device); 524 uint64_t modifier = DRM_FORMAT_MOD_INVALID; 525 const VkSubresourceLayout *plane_layouts = NULL; 526 struct tu_image *image; 527 528 if (pCreateInfo->tiling == VK_IMAGE_TILING_DRM_FORMAT_MODIFIER_EXT) { 529 const VkImageDrmFormatModifierListCreateInfoEXT *mod_info = 530 vk_find_struct_const(pCreateInfo->pNext, 531 IMAGE_DRM_FORMAT_MODIFIER_LIST_CREATE_INFO_EXT); 532 const VkImageDrmFormatModifierExplicitCreateInfoEXT *drm_explicit_info = 533 vk_find_struct_const(pCreateInfo->pNext, 534 IMAGE_DRM_FORMAT_MODIFIER_EXPLICIT_CREATE_INFO_EXT); 535 536 assert(mod_info || drm_explicit_info); 537 538 if (mod_info) { 539 modifier = DRM_FORMAT_MOD_LINEAR; 540 for (unsigned i = 0; i < mod_info->drmFormatModifierCount; i++) { 541 if (mod_info->pDrmFormatModifiers[i] == DRM_FORMAT_MOD_QCOM_COMPRESSED) 542 modifier = DRM_FORMAT_MOD_QCOM_COMPRESSED; 543 } 544 } else { 545 modifier = drm_explicit_info->drmFormatModifier; 546 assert(modifier == DRM_FORMAT_MOD_LINEAR || 547 modifier == DRM_FORMAT_MOD_QCOM_COMPRESSED); 548 plane_layouts = drm_explicit_info->pPlaneLayouts; 549 } 550 } else { 551 const struct wsi_image_create_info *wsi_info = 552 vk_find_struct_const(pCreateInfo->pNext, WSI_IMAGE_CREATE_INFO_MESA); 553 if (wsi_info && wsi_info->scanout) 554 modifier = DRM_FORMAT_MOD_LINEAR; 555 } 556 557#ifdef ANDROID 558 const VkNativeBufferANDROID *gralloc_info = 559 vk_find_struct_const(pCreateInfo->pNext, NATIVE_BUFFER_ANDROID); 560 int dma_buf; 561 if (gralloc_info) { 562 VkResult result = tu_gralloc_info(device, gralloc_info, &dma_buf, &modifier); 563 if (result != VK_SUCCESS) 564 return result; 565 } 566#endif 567 568 image = vk_object_zalloc(&device->vk, alloc, sizeof(*image), 569 VK_OBJECT_TYPE_IMAGE); 570 if (!image) 571 return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY); 572 573 const VkExternalMemoryImageCreateInfo *external_info = 574 vk_find_struct_const(pCreateInfo->pNext, EXTERNAL_MEMORY_IMAGE_CREATE_INFO); 575 image->shareable = external_info != NULL; 576 577 image->vk_format = pCreateInfo->format; 578 image->level_count = pCreateInfo->mipLevels; 579 image->layer_count = pCreateInfo->arrayLayers; 580 581 enum a6xx_tile_mode tile_mode = TILE6_3; 582 bool ubwc_enabled = 583 !(device->physical_device->instance->debug_flags & TU_DEBUG_NOUBWC); 584 585 /* use linear tiling if requested */ 586 if (pCreateInfo->tiling == VK_IMAGE_TILING_LINEAR || modifier == DRM_FORMAT_MOD_LINEAR) { 587 tile_mode = TILE6_LINEAR; 588 ubwc_enabled = false; 589 } 590 591 /* Mutable images can be reinterpreted as any other compatible format. 592 * This is a problem with UBWC (compression for different formats is different), 593 * but also tiling ("swap" affects how tiled formats are stored in memory) 594 * Depth and stencil formats cannot be reintepreted as another format, and 595 * cannot be linear with sysmem rendering, so don't fall back for those. 596 * 597 * TODO: 598 * - if the fmt_list contains only formats which are swapped, but compatible 599 * with each other (B8G8R8A8_UNORM and B8G8R8A8_UINT for example), then 600 * tiling is still possible 601 * - figure out which UBWC compressions are compatible to keep it enabled 602 */ 603 if ((pCreateInfo->flags & VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT) && 604 !vk_format_is_depth_or_stencil(image->vk_format)) { 605 const VkImageFormatListCreateInfo *fmt_list = 606 vk_find_struct_const(pCreateInfo->pNext, IMAGE_FORMAT_LIST_CREATE_INFO); 607 bool may_be_swapped = true; 608 if (fmt_list) { 609 may_be_swapped = false; 610 for (uint32_t i = 0; i < fmt_list->viewFormatCount; i++) { 611 if (tu6_format_texture(fmt_list->pViewFormats[i], TILE6_LINEAR).swap) { 612 may_be_swapped = true; 613 break; 614 } 615 } 616 } 617 if (may_be_swapped) 618 tile_mode = TILE6_LINEAR; 619 ubwc_enabled = false; 620 } 621 622 const VkImageStencilUsageCreateInfo *stencil_usage_info = 623 vk_find_struct_const(pCreateInfo->pNext, IMAGE_STENCIL_USAGE_CREATE_INFO); 624 625 if (!ubwc_possible(image->vk_format, pCreateInfo->imageType, pCreateInfo->usage, 626 stencil_usage_info ? stencil_usage_info->stencilUsage : pCreateInfo->usage, 627 device->physical_device->info, pCreateInfo->samples)) 628 ubwc_enabled = false; 629 630 /* expect UBWC enabled if we asked for it */ 631 assert(modifier != DRM_FORMAT_MOD_QCOM_COMPRESSED || ubwc_enabled); 632 633 for (uint32_t i = 0; i < tu6_plane_count(image->vk_format); i++) { 634 struct fdl_layout *layout = &image->layout[i]; 635 VkFormat format = tu6_plane_format(image->vk_format, i); 636 uint32_t width0 = pCreateInfo->extent.width; 637 uint32_t height0 = pCreateInfo->extent.height; 638 639 if (i > 0) { 640 switch (image->vk_format) { 641 case VK_FORMAT_G8_B8R8_2PLANE_420_UNORM: 642 case VK_FORMAT_G8_B8_R8_3PLANE_420_UNORM: 643 /* half width/height on chroma planes */ 644 width0 = (width0 + 1) >> 1; 645 height0 = (height0 + 1) >> 1; 646 break; 647 case VK_FORMAT_D32_SFLOAT_S8_UINT: 648 /* no UBWC for separate stencil */ 649 ubwc_enabled = false; 650 break; 651 default: 652 break; 653 } 654 } 655 656 struct fdl_explicit_layout plane_layout; 657 658 if (plane_layouts) { 659 /* only expect simple 2D images for now */ 660 if (pCreateInfo->mipLevels != 1 || 661 pCreateInfo->arrayLayers != 1 || 662 pCreateInfo->extent.depth != 1) 663 goto invalid_layout; 664 665 plane_layout.offset = plane_layouts[i].offset; 666 plane_layout.pitch = plane_layouts[i].rowPitch; 667 /* note: use plane_layouts[0].arrayPitch to support array formats */ 668 } 669 670 layout->tile_mode = tile_mode; 671 layout->ubwc = ubwc_enabled; 672 673 if (!fdl6_layout(layout, vk_format_to_pipe_format(format), 674 pCreateInfo->samples, 675 width0, height0, 676 pCreateInfo->extent.depth, 677 pCreateInfo->mipLevels, 678 pCreateInfo->arrayLayers, 679 pCreateInfo->imageType == VK_IMAGE_TYPE_3D, 680 plane_layouts ? &plane_layout : NULL)) { 681 assert(plane_layouts); /* can only fail with explicit layout */ 682 goto invalid_layout; 683 } 684 685 /* fdl6_layout can't take explicit offset without explicit pitch 686 * add offset manually for extra layouts for planes 687 */ 688 if (!plane_layouts && i > 0) { 689 uint32_t offset = ALIGN_POT(image->total_size, 4096); 690 for (int i = 0; i < pCreateInfo->mipLevels; i++) { 691 layout->slices[i].offset += offset; 692 layout->ubwc_slices[i].offset += offset; 693 } 694 layout->size += offset; 695 } 696 697 image->total_size = MAX2(image->total_size, layout->size); 698 } 699 700 const struct util_format_description *desc = util_format_description(image->layout[0].format); 701 if (util_format_has_depth(desc) && !(device->instance->debug_flags & TU_DEBUG_NOLRZ)) 702 { 703 /* Depth plane is the first one */ 704 struct fdl_layout *layout = &image->layout[0]; 705 unsigned width = layout->width0; 706 unsigned height = layout->height0; 707 708 /* LRZ buffer is super-sampled */ 709 switch (layout->nr_samples) { 710 case 4: 711 width *= 2; 712 FALLTHROUGH; 713 case 2: 714 height *= 2; 715 break; 716 default: 717 break; 718 } 719 720 unsigned lrz_pitch = align(DIV_ROUND_UP(width, 8), 32); 721 unsigned lrz_height = align(DIV_ROUND_UP(height, 8), 16); 722 723 image->lrz_height = lrz_height; 724 image->lrz_pitch = lrz_pitch; 725 image->lrz_offset = image->total_size; 726 unsigned lrz_size = lrz_pitch * lrz_height * 2; 727 image->total_size += lrz_size; 728 } 729 730 *pImage = tu_image_to_handle(image); 731 732#ifdef ANDROID 733 if (gralloc_info) 734 return tu_import_memory_from_gralloc_handle(_device, dma_buf, alloc, *pImage); 735#endif 736 return VK_SUCCESS; 737 738invalid_layout: 739 vk_object_free(&device->vk, alloc, image); 740 return vk_error(device, VK_ERROR_INVALID_DRM_FORMAT_MODIFIER_PLANE_LAYOUT_EXT); 741} 742 743VKAPI_ATTR void VKAPI_CALL 744tu_DestroyImage(VkDevice _device, 745 VkImage _image, 746 const VkAllocationCallbacks *pAllocator) 747{ 748 TU_FROM_HANDLE(tu_device, device, _device); 749 TU_FROM_HANDLE(tu_image, image, _image); 750 751 if (!image) 752 return; 753 754#ifdef ANDROID 755 if (image->owned_memory != VK_NULL_HANDLE) 756 tu_FreeMemory(_device, image->owned_memory, pAllocator); 757#endif 758 759 vk_object_free(&device->vk, pAllocator, image); 760} 761 762VKAPI_ATTR void VKAPI_CALL 763tu_GetImageSubresourceLayout(VkDevice _device, 764 VkImage _image, 765 const VkImageSubresource *pSubresource, 766 VkSubresourceLayout *pLayout) 767{ 768 TU_FROM_HANDLE(tu_image, image, _image); 769 770 struct fdl_layout *layout = 771 &image->layout[tu6_plane_index(image->vk_format, pSubresource->aspectMask)]; 772 const struct fdl_slice *slice = layout->slices + pSubresource->mipLevel; 773 774 pLayout->offset = 775 fdl_surface_offset(layout, pSubresource->mipLevel, pSubresource->arrayLayer); 776 pLayout->rowPitch = fdl_pitch(layout, pSubresource->mipLevel); 777 pLayout->arrayPitch = fdl_layer_stride(layout, pSubresource->mipLevel); 778 pLayout->depthPitch = slice->size0; 779 pLayout->size = pLayout->depthPitch * layout->depth0; 780 781 if (fdl_ubwc_enabled(layout, pSubresource->mipLevel)) { 782 /* UBWC starts at offset 0 */ 783 pLayout->offset = 0; 784 /* UBWC scanout won't match what the kernel wants if we have levels/layers */ 785 assert(image->level_count == 1 && image->layer_count == 1); 786 } 787} 788 789VKAPI_ATTR VkResult VKAPI_CALL 790tu_GetImageDrmFormatModifierPropertiesEXT( 791 VkDevice device, 792 VkImage _image, 793 VkImageDrmFormatModifierPropertiesEXT* pProperties) 794{ 795 TU_FROM_HANDLE(tu_image, image, _image); 796 797 /* TODO invent a modifier for tiled but not UBWC buffers */ 798 799 if (!image->layout[0].tile_mode) 800 pProperties->drmFormatModifier = DRM_FORMAT_MOD_LINEAR; 801 else if (image->layout[0].ubwc_layer_size) 802 pProperties->drmFormatModifier = DRM_FORMAT_MOD_QCOM_COMPRESSED; 803 else 804 pProperties->drmFormatModifier = DRM_FORMAT_MOD_INVALID; 805 806 return VK_SUCCESS; 807} 808 809 810VKAPI_ATTR VkResult VKAPI_CALL 811tu_CreateImageView(VkDevice _device, 812 const VkImageViewCreateInfo *pCreateInfo, 813 const VkAllocationCallbacks *pAllocator, 814 VkImageView *pView) 815{ 816 TU_FROM_HANDLE(tu_device, device, _device); 817 struct tu_image_view *view; 818 819 view = vk_object_alloc(&device->vk, pAllocator, sizeof(*view), 820 VK_OBJECT_TYPE_IMAGE_VIEW); 821 if (view == NULL) 822 return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY); 823 824 tu_image_view_init(view, pCreateInfo, device->physical_device->info->a6xx.has_z24uint_s8uint); 825 826 *pView = tu_image_view_to_handle(view); 827 828 return VK_SUCCESS; 829} 830 831VKAPI_ATTR void VKAPI_CALL 832tu_DestroyImageView(VkDevice _device, 833 VkImageView _iview, 834 const VkAllocationCallbacks *pAllocator) 835{ 836 TU_FROM_HANDLE(tu_device, device, _device); 837 TU_FROM_HANDLE(tu_image_view, iview, _iview); 838 839 if (!iview) 840 return; 841 842 vk_object_free(&device->vk, pAllocator, iview); 843} 844 845void 846tu_buffer_view_init(struct tu_buffer_view *view, 847 struct tu_device *device, 848 const VkBufferViewCreateInfo *pCreateInfo) 849{ 850 TU_FROM_HANDLE(tu_buffer, buffer, pCreateInfo->buffer); 851 852 view->buffer = buffer; 853 854 enum VkFormat vfmt = pCreateInfo->format; 855 enum pipe_format pfmt = vk_format_to_pipe_format(vfmt); 856 const struct tu_native_format fmt = tu6_format_texture(vfmt, TILE6_LINEAR); 857 858 uint32_t range; 859 if (pCreateInfo->range == VK_WHOLE_SIZE) 860 range = buffer->size - pCreateInfo->offset; 861 else 862 range = pCreateInfo->range; 863 uint32_t elements = range / util_format_get_blocksize(pfmt); 864 865 static const VkComponentMapping components = { 866 .r = VK_COMPONENT_SWIZZLE_R, 867 .g = VK_COMPONENT_SWIZZLE_G, 868 .b = VK_COMPONENT_SWIZZLE_B, 869 .a = VK_COMPONENT_SWIZZLE_A, 870 }; 871 872 uint64_t iova = tu_buffer_iova(buffer) + pCreateInfo->offset; 873 874 memset(&view->descriptor, 0, sizeof(view->descriptor)); 875 876 view->descriptor[0] = 877 A6XX_TEX_CONST_0_TILE_MODE(TILE6_LINEAR) | 878 A6XX_TEX_CONST_0_SWAP(fmt.swap) | 879 A6XX_TEX_CONST_0_FMT(fmt.fmt) | 880 A6XX_TEX_CONST_0_MIPLVLS(0) | 881 tu6_texswiz(&components, NULL, vfmt, VK_IMAGE_ASPECT_COLOR_BIT, false); 882 COND(vk_format_is_srgb(vfmt), A6XX_TEX_CONST_0_SRGB); 883 view->descriptor[1] = 884 A6XX_TEX_CONST_1_WIDTH(elements & MASK(15)) | 885 A6XX_TEX_CONST_1_HEIGHT(elements >> 15); 886 view->descriptor[2] = 887 A6XX_TEX_CONST_2_UNK4 | 888 A6XX_TEX_CONST_2_UNK31; 889 view->descriptor[4] = iova; 890 view->descriptor[5] = iova >> 32; 891} 892 893VKAPI_ATTR VkResult VKAPI_CALL 894tu_CreateBufferView(VkDevice _device, 895 const VkBufferViewCreateInfo *pCreateInfo, 896 const VkAllocationCallbacks *pAllocator, 897 VkBufferView *pView) 898{ 899 TU_FROM_HANDLE(tu_device, device, _device); 900 struct tu_buffer_view *view; 901 902 view = vk_object_alloc(&device->vk, pAllocator, sizeof(*view), 903 VK_OBJECT_TYPE_BUFFER_VIEW); 904 if (!view) 905 return vk_error(device, VK_ERROR_OUT_OF_HOST_MEMORY); 906 907 tu_buffer_view_init(view, device, pCreateInfo); 908 909 *pView = tu_buffer_view_to_handle(view); 910 911 return VK_SUCCESS; 912} 913 914VKAPI_ATTR void VKAPI_CALL 915tu_DestroyBufferView(VkDevice _device, 916 VkBufferView bufferView, 917 const VkAllocationCallbacks *pAllocator) 918{ 919 TU_FROM_HANDLE(tu_device, device, _device); 920 TU_FROM_HANDLE(tu_buffer_view, view, bufferView); 921 922 if (!view) 923 return; 924 925 vk_object_free(&device->vk, pAllocator, view); 926} 927