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 DEALINGS 25 * IN THE SOFTWARE. 26 */ 27 28 #include "util/mesa-sha1.h" 29 #include "util/u_atomic.h" 30 #include "radv_debug.h" 31 #include "radv_private.h" 32 #include "radv_shader.h" 33 #include "nir/nir.h" 34 #include "nir/nir_builder.h" 35 #include "spirv/nir_spirv.h" 36 #include "vk_util.h" 37 38 #include <llvm-c/Core.h> 39 #include <llvm-c/TargetMachine.h> 40 41 #include "sid.h" 42 #include "gfx9d.h" 43 #include "ac_binary.h" 44 #include "ac_llvm_util.h" 45 #include "ac_nir_to_llvm.h" 46 #include "vk_format.h" 47 #include "util/debug.h" 48 #include "ac_exp_param.h" 49 #include "ac_shader_util.h" 50 51 static void 52 radv_pipeline_destroy(struct radv_device *device, 53 struct radv_pipeline *pipeline, 54 const VkAllocationCallbacks* allocator) 55 { 56 for (unsigned i = 0; i < MESA_SHADER_STAGES; ++i) 57 if (pipeline->shaders[i]) 58 radv_shader_variant_destroy(device, pipeline->shaders[i]); 59 60 if (pipeline->gs_copy_shader) 61 radv_shader_variant_destroy(device, pipeline->gs_copy_shader); 62 63 vk_free2(&device->alloc, allocator, pipeline); 64 } 65 66 void radv_DestroyPipeline( 67 VkDevice _device, 68 VkPipeline _pipeline, 69 const VkAllocationCallbacks* pAllocator) 70 { 71 RADV_FROM_HANDLE(radv_device, device, _device); 72 RADV_FROM_HANDLE(radv_pipeline, pipeline, _pipeline); 73 74 if (!_pipeline) 75 return; 76 77 radv_pipeline_destroy(device, pipeline, pAllocator); 78 } 79 80 static void radv_dump_pipeline_stats(struct radv_device *device, struct radv_pipeline *pipeline) 81 { 82 int i; 83 84 for (i = 0; i < MESA_SHADER_STAGES; i++) { 85 if (!pipeline->shaders[i]) 86 continue; 87 88 radv_shader_dump_stats(device, pipeline->shaders[i], i, stderr); 89 } 90 } 91 92 static uint32_t get_hash_flags(struct radv_device *device) 93 { 94 uint32_t hash_flags = 0; 95 96 if (device->instance->debug_flags & RADV_DEBUG_UNSAFE_MATH) 97 hash_flags |= RADV_HASH_SHADER_UNSAFE_MATH; 98 if (device->instance->perftest_flags & RADV_PERFTEST_SISCHED) 99 hash_flags |= RADV_HASH_SHADER_SISCHED; 100 return hash_flags; 101 } 102 103 static VkResult 104 radv_pipeline_scratch_init(struct radv_device *device, 105 struct radv_pipeline *pipeline) 106 { 107 unsigned scratch_bytes_per_wave = 0; 108 unsigned max_waves = 0; 109 unsigned min_waves = 1; 110 111 for (int i = 0; i < MESA_SHADER_STAGES; ++i) { 112 if (pipeline->shaders[i]) { 113 unsigned max_stage_waves = device->scratch_waves; 114 115 scratch_bytes_per_wave = MAX2(scratch_bytes_per_wave, 116 pipeline->shaders[i]->config.scratch_bytes_per_wave); 117 118 max_stage_waves = MIN2(max_stage_waves, 119 4 * device->physical_device->rad_info.num_good_compute_units * 120 (256 / pipeline->shaders[i]->config.num_vgprs)); 121 max_waves = MAX2(max_waves, max_stage_waves); 122 } 123 } 124 125 if (pipeline->shaders[MESA_SHADER_COMPUTE]) { 126 unsigned group_size = pipeline->shaders[MESA_SHADER_COMPUTE]->info.cs.block_size[0] * 127 pipeline->shaders[MESA_SHADER_COMPUTE]->info.cs.block_size[1] * 128 pipeline->shaders[MESA_SHADER_COMPUTE]->info.cs.block_size[2]; 129 min_waves = MAX2(min_waves, round_up_u32(group_size, 64)); 130 } 131 132 if (scratch_bytes_per_wave) 133 max_waves = MIN2(max_waves, 0xffffffffu / scratch_bytes_per_wave); 134 135 if (scratch_bytes_per_wave && max_waves < min_waves) { 136 /* Not really true at this moment, but will be true on first 137 * execution. Avoid having hanging shaders. */ 138 return vk_error(VK_ERROR_OUT_OF_DEVICE_MEMORY); 139 } 140 pipeline->scratch_bytes_per_wave = scratch_bytes_per_wave; 141 pipeline->max_waves = max_waves; 142 return VK_SUCCESS; 143 } 144 145 static uint32_t si_translate_blend_function(VkBlendOp op) 146 { 147 switch (op) { 148 case VK_BLEND_OP_ADD: 149 return V_028780_COMB_DST_PLUS_SRC; 150 case VK_BLEND_OP_SUBTRACT: 151 return V_028780_COMB_SRC_MINUS_DST; 152 case VK_BLEND_OP_REVERSE_SUBTRACT: 153 return V_028780_COMB_DST_MINUS_SRC; 154 case VK_BLEND_OP_MIN: 155 return V_028780_COMB_MIN_DST_SRC; 156 case VK_BLEND_OP_MAX: 157 return V_028780_COMB_MAX_DST_SRC; 158 default: 159 return 0; 160 } 161 } 162 163 static uint32_t si_translate_blend_factor(VkBlendFactor factor) 164 { 165 switch (factor) { 166 case VK_BLEND_FACTOR_ZERO: 167 return V_028780_BLEND_ZERO; 168 case VK_BLEND_FACTOR_ONE: 169 return V_028780_BLEND_ONE; 170 case VK_BLEND_FACTOR_SRC_COLOR: 171 return V_028780_BLEND_SRC_COLOR; 172 case VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR: 173 return V_028780_BLEND_ONE_MINUS_SRC_COLOR; 174 case VK_BLEND_FACTOR_DST_COLOR: 175 return V_028780_BLEND_DST_COLOR; 176 case VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR: 177 return V_028780_BLEND_ONE_MINUS_DST_COLOR; 178 case VK_BLEND_FACTOR_SRC_ALPHA: 179 return V_028780_BLEND_SRC_ALPHA; 180 case VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA: 181 return V_028780_BLEND_ONE_MINUS_SRC_ALPHA; 182 case VK_BLEND_FACTOR_DST_ALPHA: 183 return V_028780_BLEND_DST_ALPHA; 184 case VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA: 185 return V_028780_BLEND_ONE_MINUS_DST_ALPHA; 186 case VK_BLEND_FACTOR_CONSTANT_COLOR: 187 return V_028780_BLEND_CONSTANT_COLOR; 188 case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_COLOR: 189 return V_028780_BLEND_ONE_MINUS_CONSTANT_COLOR; 190 case VK_BLEND_FACTOR_CONSTANT_ALPHA: 191 return V_028780_BLEND_CONSTANT_ALPHA; 192 case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_ALPHA: 193 return V_028780_BLEND_ONE_MINUS_CONSTANT_ALPHA; 194 case VK_BLEND_FACTOR_SRC_ALPHA_SATURATE: 195 return V_028780_BLEND_SRC_ALPHA_SATURATE; 196 case VK_BLEND_FACTOR_SRC1_COLOR: 197 return V_028780_BLEND_SRC1_COLOR; 198 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR: 199 return V_028780_BLEND_INV_SRC1_COLOR; 200 case VK_BLEND_FACTOR_SRC1_ALPHA: 201 return V_028780_BLEND_SRC1_ALPHA; 202 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA: 203 return V_028780_BLEND_INV_SRC1_ALPHA; 204 default: 205 return 0; 206 } 207 } 208 209 static uint32_t si_translate_blend_opt_function(VkBlendOp op) 210 { 211 switch (op) { 212 case VK_BLEND_OP_ADD: 213 return V_028760_OPT_COMB_ADD; 214 case VK_BLEND_OP_SUBTRACT: 215 return V_028760_OPT_COMB_SUBTRACT; 216 case VK_BLEND_OP_REVERSE_SUBTRACT: 217 return V_028760_OPT_COMB_REVSUBTRACT; 218 case VK_BLEND_OP_MIN: 219 return V_028760_OPT_COMB_MIN; 220 case VK_BLEND_OP_MAX: 221 return V_028760_OPT_COMB_MAX; 222 default: 223 return V_028760_OPT_COMB_BLEND_DISABLED; 224 } 225 } 226 227 static uint32_t si_translate_blend_opt_factor(VkBlendFactor factor, bool is_alpha) 228 { 229 switch (factor) { 230 case VK_BLEND_FACTOR_ZERO: 231 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_ALL; 232 case VK_BLEND_FACTOR_ONE: 233 return V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE; 234 case VK_BLEND_FACTOR_SRC_COLOR: 235 return is_alpha ? V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0 236 : V_028760_BLEND_OPT_PRESERVE_C1_IGNORE_C0; 237 case VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR: 238 return is_alpha ? V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1 239 : V_028760_BLEND_OPT_PRESERVE_C0_IGNORE_C1; 240 case VK_BLEND_FACTOR_SRC_ALPHA: 241 return V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0; 242 case VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA: 243 return V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1; 244 case VK_BLEND_FACTOR_SRC_ALPHA_SATURATE: 245 return is_alpha ? V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE 246 : V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0; 247 default: 248 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE; 249 } 250 } 251 252 /** 253 * Get rid of DST in the blend factors by commuting the operands: 254 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC) 255 */ 256 static void si_blend_remove_dst(unsigned *func, unsigned *src_factor, 257 unsigned *dst_factor, unsigned expected_dst, 258 unsigned replacement_src) 259 { 260 if (*src_factor == expected_dst && 261 *dst_factor == VK_BLEND_FACTOR_ZERO) { 262 *src_factor = VK_BLEND_FACTOR_ZERO; 263 *dst_factor = replacement_src; 264 265 /* Commuting the operands requires reversing subtractions. */ 266 if (*func == VK_BLEND_OP_SUBTRACT) 267 *func = VK_BLEND_OP_REVERSE_SUBTRACT; 268 else if (*func == VK_BLEND_OP_REVERSE_SUBTRACT) 269 *func = VK_BLEND_OP_SUBTRACT; 270 } 271 } 272 273 static bool si_blend_factor_uses_dst(unsigned factor) 274 { 275 return factor == VK_BLEND_FACTOR_DST_COLOR || 276 factor == VK_BLEND_FACTOR_DST_ALPHA || 277 factor == VK_BLEND_FACTOR_SRC_ALPHA_SATURATE || 278 factor == VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA || 279 factor == VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR; 280 } 281 282 static bool is_dual_src(VkBlendFactor factor) 283 { 284 switch (factor) { 285 case VK_BLEND_FACTOR_SRC1_COLOR: 286 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR: 287 case VK_BLEND_FACTOR_SRC1_ALPHA: 288 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA: 289 return true; 290 default: 291 return false; 292 } 293 } 294 295 static unsigned si_choose_spi_color_format(VkFormat vk_format, 296 bool blend_enable, 297 bool blend_need_alpha) 298 { 299 const struct vk_format_description *desc = vk_format_description(vk_format); 300 unsigned format, ntype, swap; 301 302 /* Alpha is needed for alpha-to-coverage. 303 * Blending may be with or without alpha. 304 */ 305 unsigned normal = 0; /* most optimal, may not support blending or export alpha */ 306 unsigned alpha = 0; /* exports alpha, but may not support blending */ 307 unsigned blend = 0; /* supports blending, but may not export alpha */ 308 unsigned blend_alpha = 0; /* least optimal, supports blending and exports alpha */ 309 310 format = radv_translate_colorformat(vk_format); 311 ntype = radv_translate_color_numformat(vk_format, desc, 312 vk_format_get_first_non_void_channel(vk_format)); 313 swap = radv_translate_colorswap(vk_format, false); 314 315 /* Choose the SPI color formats. These are required values for Stoney/RB+. 316 * Other chips have multiple choices, though they are not necessarily better. 317 */ 318 switch (format) { 319 case V_028C70_COLOR_5_6_5: 320 case V_028C70_COLOR_1_5_5_5: 321 case V_028C70_COLOR_5_5_5_1: 322 case V_028C70_COLOR_4_4_4_4: 323 case V_028C70_COLOR_10_11_11: 324 case V_028C70_COLOR_11_11_10: 325 case V_028C70_COLOR_8: 326 case V_028C70_COLOR_8_8: 327 case V_028C70_COLOR_8_8_8_8: 328 case V_028C70_COLOR_10_10_10_2: 329 case V_028C70_COLOR_2_10_10_10: 330 if (ntype == V_028C70_NUMBER_UINT) 331 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_UINT16_ABGR; 332 else if (ntype == V_028C70_NUMBER_SINT) 333 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_SINT16_ABGR; 334 else 335 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_FP16_ABGR; 336 break; 337 338 case V_028C70_COLOR_16: 339 case V_028C70_COLOR_16_16: 340 case V_028C70_COLOR_16_16_16_16: 341 if (ntype == V_028C70_NUMBER_UNORM || 342 ntype == V_028C70_NUMBER_SNORM) { 343 /* UNORM16 and SNORM16 don't support blending */ 344 if (ntype == V_028C70_NUMBER_UNORM) 345 normal = alpha = V_028714_SPI_SHADER_UNORM16_ABGR; 346 else 347 normal = alpha = V_028714_SPI_SHADER_SNORM16_ABGR; 348 349 /* Use 32 bits per channel for blending. */ 350 if (format == V_028C70_COLOR_16) { 351 if (swap == V_028C70_SWAP_STD) { /* R */ 352 blend = V_028714_SPI_SHADER_32_R; 353 blend_alpha = V_028714_SPI_SHADER_32_AR; 354 } else if (swap == V_028C70_SWAP_ALT_REV) /* A */ 355 blend = blend_alpha = V_028714_SPI_SHADER_32_AR; 356 else 357 assert(0); 358 } else if (format == V_028C70_COLOR_16_16) { 359 if (swap == V_028C70_SWAP_STD) { /* RG */ 360 blend = V_028714_SPI_SHADER_32_GR; 361 blend_alpha = V_028714_SPI_SHADER_32_ABGR; 362 } else if (swap == V_028C70_SWAP_ALT) /* RA */ 363 blend = blend_alpha = V_028714_SPI_SHADER_32_AR; 364 else 365 assert(0); 366 } else /* 16_16_16_16 */ 367 blend = blend_alpha = V_028714_SPI_SHADER_32_ABGR; 368 } else if (ntype == V_028C70_NUMBER_UINT) 369 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_UINT16_ABGR; 370 else if (ntype == V_028C70_NUMBER_SINT) 371 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_SINT16_ABGR; 372 else if (ntype == V_028C70_NUMBER_FLOAT) 373 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_FP16_ABGR; 374 else 375 assert(0); 376 break; 377 378 case V_028C70_COLOR_32: 379 if (swap == V_028C70_SWAP_STD) { /* R */ 380 blend = normal = V_028714_SPI_SHADER_32_R; 381 alpha = blend_alpha = V_028714_SPI_SHADER_32_AR; 382 } else if (swap == V_028C70_SWAP_ALT_REV) /* A */ 383 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_AR; 384 else 385 assert(0); 386 break; 387 388 case V_028C70_COLOR_32_32: 389 if (swap == V_028C70_SWAP_STD) { /* RG */ 390 blend = normal = V_028714_SPI_SHADER_32_GR; 391 alpha = blend_alpha = V_028714_SPI_SHADER_32_ABGR; 392 } else if (swap == V_028C70_SWAP_ALT) /* RA */ 393 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_AR; 394 else 395 assert(0); 396 break; 397 398 case V_028C70_COLOR_32_32_32_32: 399 case V_028C70_COLOR_8_24: 400 case V_028C70_COLOR_24_8: 401 case V_028C70_COLOR_X24_8_32_FLOAT: 402 alpha = blend = blend_alpha = normal = V_028714_SPI_SHADER_32_ABGR; 403 break; 404 405 default: 406 unreachable("unhandled blend format"); 407 } 408 409 if (blend_enable && blend_need_alpha) 410 return blend_alpha; 411 else if(blend_need_alpha) 412 return alpha; 413 else if(blend_enable) 414 return blend; 415 else 416 return normal; 417 } 418 419 static void 420 radv_pipeline_compute_spi_color_formats(struct radv_pipeline *pipeline, 421 const VkGraphicsPipelineCreateInfo *pCreateInfo, 422 uint32_t blend_enable, 423 uint32_t blend_need_alpha, 424 bool single_cb_enable, 425 bool blend_mrt0_is_dual_src) 426 { 427 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass); 428 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass; 429 struct radv_blend_state *blend = &pipeline->graphics.blend; 430 unsigned col_format = 0; 431 432 for (unsigned i = 0; i < (single_cb_enable ? 1 : subpass->color_count); ++i) { 433 unsigned cf; 434 435 if (subpass->color_attachments[i].attachment == VK_ATTACHMENT_UNUSED) { 436 cf = V_028714_SPI_SHADER_ZERO; 437 } else { 438 struct radv_render_pass_attachment *attachment = pass->attachments + subpass->color_attachments[i].attachment; 439 440 cf = si_choose_spi_color_format(attachment->format, 441 blend_enable & (1 << i), 442 blend_need_alpha & (1 << i)); 443 } 444 445 col_format |= cf << (4 * i); 446 } 447 448 blend->cb_shader_mask = ac_get_cb_shader_mask(col_format); 449 450 if (blend_mrt0_is_dual_src) 451 col_format |= (col_format & 0xf) << 4; 452 blend->spi_shader_col_format = col_format; 453 } 454 455 static bool 456 format_is_int8(VkFormat format) 457 { 458 const struct vk_format_description *desc = vk_format_description(format); 459 int channel = vk_format_get_first_non_void_channel(format); 460 461 return channel >= 0 && desc->channel[channel].pure_integer && 462 desc->channel[channel].size == 8; 463 } 464 465 static bool 466 format_is_int10(VkFormat format) 467 { 468 const struct vk_format_description *desc = vk_format_description(format); 469 470 if (desc->nr_channels != 4) 471 return false; 472 for (unsigned i = 0; i < 4; i++) { 473 if (desc->channel[i].pure_integer && desc->channel[i].size == 10) 474 return true; 475 } 476 return false; 477 } 478 479 unsigned radv_format_meta_fs_key(VkFormat format) 480 { 481 unsigned col_format = si_choose_spi_color_format(format, false, false) - 1; 482 bool is_int8 = format_is_int8(format); 483 bool is_int10 = format_is_int10(format); 484 485 return col_format + (is_int8 ? 3 : is_int10 ? 5 : 0); 486 } 487 488 static void 489 radv_pipeline_compute_get_int_clamp(const VkGraphicsPipelineCreateInfo *pCreateInfo, 490 unsigned *is_int8, unsigned *is_int10) 491 { 492 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass); 493 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass; 494 *is_int8 = 0; 495 *is_int10 = 0; 496 497 for (unsigned i = 0; i < subpass->color_count; ++i) { 498 struct radv_render_pass_attachment *attachment; 499 500 if (subpass->color_attachments[i].attachment == VK_ATTACHMENT_UNUSED) 501 continue; 502 503 attachment = pass->attachments + subpass->color_attachments[i].attachment; 504 505 if (format_is_int8(attachment->format)) 506 *is_int8 |= 1 << i; 507 if (format_is_int10(attachment->format)) 508 *is_int10 |= 1 << i; 509 } 510 } 511 512 static void 513 radv_pipeline_init_blend_state(struct radv_pipeline *pipeline, 514 const VkGraphicsPipelineCreateInfo *pCreateInfo, 515 const struct radv_graphics_pipeline_create_info *extra) 516 { 517 const VkPipelineColorBlendStateCreateInfo *vkblend = pCreateInfo->pColorBlendState; 518 const VkPipelineMultisampleStateCreateInfo *vkms = pCreateInfo->pMultisampleState; 519 struct radv_blend_state *blend = &pipeline->graphics.blend; 520 unsigned mode = V_028808_CB_NORMAL; 521 uint32_t blend_enable = 0, blend_need_alpha = 0; 522 bool blend_mrt0_is_dual_src = false; 523 int i; 524 bool single_cb_enable = false; 525 526 if (!vkblend) 527 return; 528 529 if (extra && extra->custom_blend_mode) { 530 single_cb_enable = true; 531 mode = extra->custom_blend_mode; 532 } 533 blend->cb_color_control = 0; 534 if (vkblend->logicOpEnable) 535 blend->cb_color_control |= S_028808_ROP3(vkblend->logicOp | (vkblend->logicOp << 4)); 536 else 537 blend->cb_color_control |= S_028808_ROP3(0xcc); 538 539 blend->db_alpha_to_mask = S_028B70_ALPHA_TO_MASK_OFFSET0(2) | 540 S_028B70_ALPHA_TO_MASK_OFFSET1(2) | 541 S_028B70_ALPHA_TO_MASK_OFFSET2(2) | 542 S_028B70_ALPHA_TO_MASK_OFFSET3(2); 543 544 if (vkms && vkms->alphaToCoverageEnable) { 545 blend->db_alpha_to_mask |= S_028B70_ALPHA_TO_MASK_ENABLE(1); 546 } 547 548 blend->cb_target_mask = 0; 549 for (i = 0; i < vkblend->attachmentCount; i++) { 550 const VkPipelineColorBlendAttachmentState *att = &vkblend->pAttachments[i]; 551 unsigned blend_cntl = 0; 552 unsigned srcRGB_opt, dstRGB_opt, srcA_opt, dstA_opt; 553 VkBlendOp eqRGB = att->colorBlendOp; 554 VkBlendFactor srcRGB = att->srcColorBlendFactor; 555 VkBlendFactor dstRGB = att->dstColorBlendFactor; 556 VkBlendOp eqA = att->alphaBlendOp; 557 VkBlendFactor srcA = att->srcAlphaBlendFactor; 558 VkBlendFactor dstA = att->dstAlphaBlendFactor; 559 560 blend->sx_mrt_blend_opt[i] = S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED) | S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED); 561 562 if (!att->colorWriteMask) 563 continue; 564 565 blend->cb_target_mask |= (unsigned)att->colorWriteMask << (4 * i); 566 if (!att->blendEnable) { 567 blend->cb_blend_control[i] = blend_cntl; 568 continue; 569 } 570 571 if (is_dual_src(srcRGB) || is_dual_src(dstRGB) || is_dual_src(srcA) || is_dual_src(dstA)) 572 if (i == 0) 573 blend_mrt0_is_dual_src = true; 574 575 if (eqRGB == VK_BLEND_OP_MIN || eqRGB == VK_BLEND_OP_MAX) { 576 srcRGB = VK_BLEND_FACTOR_ONE; 577 dstRGB = VK_BLEND_FACTOR_ONE; 578 } 579 if (eqA == VK_BLEND_OP_MIN || eqA == VK_BLEND_OP_MAX) { 580 srcA = VK_BLEND_FACTOR_ONE; 581 dstA = VK_BLEND_FACTOR_ONE; 582 } 583 584 /* Blending optimizations for RB+. 585 * These transformations don't change the behavior. 586 * 587 * First, get rid of DST in the blend factors: 588 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC) 589 */ 590 si_blend_remove_dst(&eqRGB, &srcRGB, &dstRGB, 591 VK_BLEND_FACTOR_DST_COLOR, 592 VK_BLEND_FACTOR_SRC_COLOR); 593 594 si_blend_remove_dst(&eqA, &srcA, &dstA, 595 VK_BLEND_FACTOR_DST_COLOR, 596 VK_BLEND_FACTOR_SRC_COLOR); 597 598 si_blend_remove_dst(&eqA, &srcA, &dstA, 599 VK_BLEND_FACTOR_DST_ALPHA, 600 VK_BLEND_FACTOR_SRC_ALPHA); 601 602 /* Look up the ideal settings from tables. */ 603 srcRGB_opt = si_translate_blend_opt_factor(srcRGB, false); 604 dstRGB_opt = si_translate_blend_opt_factor(dstRGB, false); 605 srcA_opt = si_translate_blend_opt_factor(srcA, true); 606 dstA_opt = si_translate_blend_opt_factor(dstA, true); 607 608 /* Handle interdependencies. */ 609 if (si_blend_factor_uses_dst(srcRGB)) 610 dstRGB_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE; 611 if (si_blend_factor_uses_dst(srcA)) 612 dstA_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE; 613 614 if (srcRGB == VK_BLEND_FACTOR_SRC_ALPHA_SATURATE && 615 (dstRGB == VK_BLEND_FACTOR_ZERO || 616 dstRGB == VK_BLEND_FACTOR_SRC_ALPHA || 617 dstRGB == VK_BLEND_FACTOR_SRC_ALPHA_SATURATE)) 618 dstRGB_opt = V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0; 619 620 /* Set the final value. */ 621 blend->sx_mrt_blend_opt[i] = 622 S_028760_COLOR_SRC_OPT(srcRGB_opt) | 623 S_028760_COLOR_DST_OPT(dstRGB_opt) | 624 S_028760_COLOR_COMB_FCN(si_translate_blend_opt_function(eqRGB)) | 625 S_028760_ALPHA_SRC_OPT(srcA_opt) | 626 S_028760_ALPHA_DST_OPT(dstA_opt) | 627 S_028760_ALPHA_COMB_FCN(si_translate_blend_opt_function(eqA)); 628 blend_cntl |= S_028780_ENABLE(1); 629 630 blend_cntl |= S_028780_COLOR_COMB_FCN(si_translate_blend_function(eqRGB)); 631 blend_cntl |= S_028780_COLOR_SRCBLEND(si_translate_blend_factor(srcRGB)); 632 blend_cntl |= S_028780_COLOR_DESTBLEND(si_translate_blend_factor(dstRGB)); 633 if (srcA != srcRGB || dstA != dstRGB || eqA != eqRGB) { 634 blend_cntl |= S_028780_SEPARATE_ALPHA_BLEND(1); 635 blend_cntl |= S_028780_ALPHA_COMB_FCN(si_translate_blend_function(eqA)); 636 blend_cntl |= S_028780_ALPHA_SRCBLEND(si_translate_blend_factor(srcA)); 637 blend_cntl |= S_028780_ALPHA_DESTBLEND(si_translate_blend_factor(dstA)); 638 } 639 blend->cb_blend_control[i] = blend_cntl; 640 641 blend_enable |= 1 << i; 642 643 if (srcRGB == VK_BLEND_FACTOR_SRC_ALPHA || 644 dstRGB == VK_BLEND_FACTOR_SRC_ALPHA || 645 srcRGB == VK_BLEND_FACTOR_SRC_ALPHA_SATURATE || 646 dstRGB == VK_BLEND_FACTOR_SRC_ALPHA_SATURATE || 647 srcRGB == VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA || 648 dstRGB == VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA) 649 blend_need_alpha |= 1 << i; 650 } 651 for (i = vkblend->attachmentCount; i < 8; i++) { 652 blend->cb_blend_control[i] = 0; 653 blend->sx_mrt_blend_opt[i] = S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED) | S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED); 654 } 655 656 /* disable RB+ for now */ 657 if (pipeline->device->physical_device->has_rbplus) 658 blend->cb_color_control |= S_028808_DISABLE_DUAL_QUAD(1); 659 660 if (blend->cb_target_mask) 661 blend->cb_color_control |= S_028808_MODE(mode); 662 else 663 blend->cb_color_control |= S_028808_MODE(V_028808_CB_DISABLE); 664 665 radv_pipeline_compute_spi_color_formats(pipeline, pCreateInfo, 666 blend_enable, blend_need_alpha, single_cb_enable, blend_mrt0_is_dual_src); 667 } 668 669 static uint32_t si_translate_stencil_op(enum VkStencilOp op) 670 { 671 switch (op) { 672 case VK_STENCIL_OP_KEEP: 673 return V_02842C_STENCIL_KEEP; 674 case VK_STENCIL_OP_ZERO: 675 return V_02842C_STENCIL_ZERO; 676 case VK_STENCIL_OP_REPLACE: 677 return V_02842C_STENCIL_REPLACE_TEST; 678 case VK_STENCIL_OP_INCREMENT_AND_CLAMP: 679 return V_02842C_STENCIL_ADD_CLAMP; 680 case VK_STENCIL_OP_DECREMENT_AND_CLAMP: 681 return V_02842C_STENCIL_SUB_CLAMP; 682 case VK_STENCIL_OP_INVERT: 683 return V_02842C_STENCIL_INVERT; 684 case VK_STENCIL_OP_INCREMENT_AND_WRAP: 685 return V_02842C_STENCIL_ADD_WRAP; 686 case VK_STENCIL_OP_DECREMENT_AND_WRAP: 687 return V_02842C_STENCIL_SUB_WRAP; 688 default: 689 return 0; 690 } 691 } 692 static void 693 radv_pipeline_init_depth_stencil_state(struct radv_pipeline *pipeline, 694 const VkGraphicsPipelineCreateInfo *pCreateInfo, 695 const struct radv_graphics_pipeline_create_info *extra) 696 { 697 const VkPipelineDepthStencilStateCreateInfo *vkds = pCreateInfo->pDepthStencilState; 698 struct radv_depth_stencil_state *ds = &pipeline->graphics.ds; 699 700 if (!vkds) 701 return; 702 703 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass); 704 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass; 705 if (subpass->depth_stencil_attachment.attachment == VK_ATTACHMENT_UNUSED) 706 return; 707 708 struct radv_render_pass_attachment *attachment = pass->attachments + subpass->depth_stencil_attachment.attachment; 709 bool has_depth_attachment = vk_format_is_depth(attachment->format); 710 bool has_stencil_attachment = vk_format_is_stencil(attachment->format); 711 712 if (has_depth_attachment) { 713 ds->db_depth_control = S_028800_Z_ENABLE(vkds->depthTestEnable ? 1 : 0) | 714 S_028800_Z_WRITE_ENABLE(vkds->depthWriteEnable ? 1 : 0) | 715 S_028800_ZFUNC(vkds->depthCompareOp) | 716 S_028800_DEPTH_BOUNDS_ENABLE(vkds->depthBoundsTestEnable ? 1 : 0); 717 } 718 719 if (has_stencil_attachment && vkds->stencilTestEnable) { 720 ds->db_depth_control |= S_028800_STENCIL_ENABLE(1) | S_028800_BACKFACE_ENABLE(1); 721 ds->db_depth_control |= S_028800_STENCILFUNC(vkds->front.compareOp); 722 ds->db_stencil_control |= S_02842C_STENCILFAIL(si_translate_stencil_op(vkds->front.failOp)); 723 ds->db_stencil_control |= S_02842C_STENCILZPASS(si_translate_stencil_op(vkds->front.passOp)); 724 ds->db_stencil_control |= S_02842C_STENCILZFAIL(si_translate_stencil_op(vkds->front.depthFailOp)); 725 726 ds->db_depth_control |= S_028800_STENCILFUNC_BF(vkds->back.compareOp); 727 ds->db_stencil_control |= S_02842C_STENCILFAIL_BF(si_translate_stencil_op(vkds->back.failOp)); 728 ds->db_stencil_control |= S_02842C_STENCILZPASS_BF(si_translate_stencil_op(vkds->back.passOp)); 729 ds->db_stencil_control |= S_02842C_STENCILZFAIL_BF(si_translate_stencil_op(vkds->back.depthFailOp)); 730 } 731 732 if (extra) { 733 734 ds->db_render_control |= S_028000_DEPTH_CLEAR_ENABLE(extra->db_depth_clear); 735 ds->db_render_control |= S_028000_STENCIL_CLEAR_ENABLE(extra->db_stencil_clear); 736 737 ds->db_render_control |= S_028000_RESUMMARIZE_ENABLE(extra->db_resummarize); 738 ds->db_render_control |= S_028000_DEPTH_COMPRESS_DISABLE(extra->db_flush_depth_inplace); 739 ds->db_render_control |= S_028000_STENCIL_COMPRESS_DISABLE(extra->db_flush_stencil_inplace); 740 ds->db_render_override2 |= S_028010_DISABLE_ZMASK_EXPCLEAR_OPTIMIZATION(extra->db_depth_disable_expclear); 741 ds->db_render_override2 |= S_028010_DISABLE_SMEM_EXPCLEAR_OPTIMIZATION(extra->db_stencil_disable_expclear); 742 } 743 } 744 745 static uint32_t si_translate_fill(VkPolygonMode func) 746 { 747 switch(func) { 748 case VK_POLYGON_MODE_FILL: 749 return V_028814_X_DRAW_TRIANGLES; 750 case VK_POLYGON_MODE_LINE: 751 return V_028814_X_DRAW_LINES; 752 case VK_POLYGON_MODE_POINT: 753 return V_028814_X_DRAW_POINTS; 754 default: 755 assert(0); 756 return V_028814_X_DRAW_POINTS; 757 } 758 } 759 static void 760 radv_pipeline_init_raster_state(struct radv_pipeline *pipeline, 761 const VkGraphicsPipelineCreateInfo *pCreateInfo) 762 { 763 const VkPipelineRasterizationStateCreateInfo *vkraster = pCreateInfo->pRasterizationState; 764 struct radv_raster_state *raster = &pipeline->graphics.raster; 765 766 raster->spi_interp_control = 767 S_0286D4_FLAT_SHADE_ENA(1) | 768 S_0286D4_PNT_SPRITE_ENA(1) | 769 S_0286D4_PNT_SPRITE_OVRD_X(V_0286D4_SPI_PNT_SPRITE_SEL_S) | 770 S_0286D4_PNT_SPRITE_OVRD_Y(V_0286D4_SPI_PNT_SPRITE_SEL_T) | 771 S_0286D4_PNT_SPRITE_OVRD_Z(V_0286D4_SPI_PNT_SPRITE_SEL_0) | 772 S_0286D4_PNT_SPRITE_OVRD_W(V_0286D4_SPI_PNT_SPRITE_SEL_1) | 773 S_0286D4_PNT_SPRITE_TOP_1(0); // vulkan is top to bottom - 1.0 at bottom 774 775 776 raster->pa_cl_clip_cntl = S_028810_PS_UCP_MODE(3) | 777 S_028810_DX_CLIP_SPACE_DEF(1) | // vulkan uses DX conventions. 778 S_028810_ZCLIP_NEAR_DISABLE(vkraster->depthClampEnable ? 1 : 0) | 779 S_028810_ZCLIP_FAR_DISABLE(vkraster->depthClampEnable ? 1 : 0) | 780 S_028810_DX_RASTERIZATION_KILL(vkraster->rasterizerDiscardEnable ? 1 : 0) | 781 S_028810_DX_LINEAR_ATTR_CLIP_ENA(1); 782 783 raster->pa_su_vtx_cntl = 784 S_028BE4_PIX_CENTER(1) | // TODO verify 785 S_028BE4_ROUND_MODE(V_028BE4_X_ROUND_TO_EVEN) | 786 S_028BE4_QUANT_MODE(V_028BE4_X_16_8_FIXED_POINT_1_256TH); 787 788 raster->pa_su_sc_mode_cntl = 789 S_028814_FACE(vkraster->frontFace) | 790 S_028814_CULL_FRONT(!!(vkraster->cullMode & VK_CULL_MODE_FRONT_BIT)) | 791 S_028814_CULL_BACK(!!(vkraster->cullMode & VK_CULL_MODE_BACK_BIT)) | 792 S_028814_POLY_MODE(vkraster->polygonMode != VK_POLYGON_MODE_FILL) | 793 S_028814_POLYMODE_FRONT_PTYPE(si_translate_fill(vkraster->polygonMode)) | 794 S_028814_POLYMODE_BACK_PTYPE(si_translate_fill(vkraster->polygonMode)) | 795 S_028814_POLY_OFFSET_FRONT_ENABLE(vkraster->depthBiasEnable ? 1 : 0) | 796 S_028814_POLY_OFFSET_BACK_ENABLE(vkraster->depthBiasEnable ? 1 : 0) | 797 S_028814_POLY_OFFSET_PARA_ENABLE(vkraster->depthBiasEnable ? 1 : 0); 798 799 } 800 801 static uint8_t radv_pipeline_get_ps_iter_samples(const VkPipelineMultisampleStateCreateInfo *vkms) 802 { 803 uint32_t num_samples = vkms->rasterizationSamples; 804 uint32_t ps_iter_samples = 1; 805 806 if (vkms->sampleShadingEnable) { 807 ps_iter_samples = ceil(vkms->minSampleShading * num_samples); 808 ps_iter_samples = util_next_power_of_two(ps_iter_samples); 809 } 810 return ps_iter_samples; 811 } 812 813 static void 814 radv_pipeline_init_multisample_state(struct radv_pipeline *pipeline, 815 const VkGraphicsPipelineCreateInfo *pCreateInfo) 816 { 817 const VkPipelineMultisampleStateCreateInfo *vkms = pCreateInfo->pMultisampleState; 818 struct radv_multisample_state *ms = &pipeline->graphics.ms; 819 unsigned num_tile_pipes = pipeline->device->physical_device->rad_info.num_tile_pipes; 820 int ps_iter_samples = 1; 821 uint32_t mask = 0xffff; 822 823 if (vkms) 824 ms->num_samples = vkms->rasterizationSamples; 825 else 826 ms->num_samples = 1; 827 828 if (vkms) 829 ps_iter_samples = radv_pipeline_get_ps_iter_samples(vkms); 830 if (vkms && !vkms->sampleShadingEnable && pipeline->shaders[MESA_SHADER_FRAGMENT]->info.info.ps.force_persample) { 831 ps_iter_samples = ms->num_samples; 832 } 833 834 ms->pa_sc_line_cntl = S_028BDC_DX10_DIAMOND_TEST_ENA(1); 835 ms->pa_sc_aa_config = 0; 836 ms->db_eqaa = S_028804_HIGH_QUALITY_INTERSECTIONS(1) | 837 S_028804_STATIC_ANCHOR_ASSOCIATIONS(1); 838 ms->pa_sc_mode_cntl_1 = 839 S_028A4C_WALK_FENCE_ENABLE(1) | //TODO linear dst fixes 840 S_028A4C_WALK_FENCE_SIZE(num_tile_pipes == 2 ? 2 : 3) | 841 /* always 1: */ 842 S_028A4C_WALK_ALIGN8_PRIM_FITS_ST(1) | 843 S_028A4C_SUPERTILE_WALK_ORDER_ENABLE(1) | 844 S_028A4C_TILE_WALK_ORDER_ENABLE(1) | 845 S_028A4C_MULTI_SHADER_ENGINE_PRIM_DISCARD_ENABLE(1) | 846 S_028A4C_FORCE_EOV_CNTDWN_ENABLE(1) | 847 S_028A4C_FORCE_EOV_REZ_ENABLE(1); 848 ms->pa_sc_mode_cntl_0 = S_028A48_ALTERNATE_RBS_PER_TILE(pipeline->device->physical_device->rad_info.chip_class >= GFX9) | 849 S_028A48_VPORT_SCISSOR_ENABLE(1); 850 851 if (ms->num_samples > 1) { 852 unsigned log_samples = util_logbase2(ms->num_samples); 853 unsigned log_ps_iter_samples = util_logbase2(ps_iter_samples); 854 ms->pa_sc_mode_cntl_0 |= S_028A48_MSAA_ENABLE(1); 855 ms->pa_sc_line_cntl |= S_028BDC_EXPAND_LINE_WIDTH(1); /* CM_R_028BDC_PA_SC_LINE_CNTL */ 856 ms->db_eqaa |= S_028804_MAX_ANCHOR_SAMPLES(log_samples) | 857 S_028804_PS_ITER_SAMPLES(log_ps_iter_samples) | 858 S_028804_MASK_EXPORT_NUM_SAMPLES(log_samples) | 859 S_028804_ALPHA_TO_MASK_NUM_SAMPLES(log_samples); 860 ms->pa_sc_aa_config |= S_028BE0_MSAA_NUM_SAMPLES(log_samples) | 861 S_028BE0_MAX_SAMPLE_DIST(radv_cayman_get_maxdist(log_samples)) | 862 S_028BE0_MSAA_EXPOSED_SAMPLES(log_samples); /* CM_R_028BE0_PA_SC_AA_CONFIG */ 863 ms->pa_sc_mode_cntl_1 |= S_028A4C_PS_ITER_SAMPLE(ps_iter_samples > 1); 864 if (ps_iter_samples > 1) 865 pipeline->graphics.spi_baryc_cntl |= S_0286E0_POS_FLOAT_LOCATION(2); 866 } 867 868 const struct VkPipelineRasterizationStateRasterizationOrderAMD *raster_order = 869 vk_find_struct_const(pCreateInfo->pRasterizationState->pNext, PIPELINE_RASTERIZATION_STATE_RASTERIZATION_ORDER_AMD); 870 if (raster_order && raster_order->rasterizationOrder == VK_RASTERIZATION_ORDER_RELAXED_AMD) { 871 ms->pa_sc_mode_cntl_1 |= S_028A4C_OUT_OF_ORDER_PRIMITIVE_ENABLE(1) | 872 S_028A4C_OUT_OF_ORDER_WATER_MARK(0x7); 873 } 874 875 if (vkms && vkms->pSampleMask) { 876 mask = vkms->pSampleMask[0] & 0xffff; 877 } 878 879 ms->pa_sc_aa_mask[0] = mask | (mask << 16); 880 ms->pa_sc_aa_mask[1] = mask | (mask << 16); 881 } 882 883 static bool 884 radv_prim_can_use_guardband(enum VkPrimitiveTopology topology) 885 { 886 switch (topology) { 887 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST: 888 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST: 889 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP: 890 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY: 891 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY: 892 return false; 893 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST: 894 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP: 895 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN: 896 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY: 897 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY: 898 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST: 899 return true; 900 default: 901 unreachable("unhandled primitive type"); 902 } 903 } 904 905 static uint32_t 906 si_translate_prim(enum VkPrimitiveTopology topology) 907 { 908 switch (topology) { 909 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST: 910 return V_008958_DI_PT_POINTLIST; 911 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST: 912 return V_008958_DI_PT_LINELIST; 913 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP: 914 return V_008958_DI_PT_LINESTRIP; 915 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST: 916 return V_008958_DI_PT_TRILIST; 917 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP: 918 return V_008958_DI_PT_TRISTRIP; 919 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN: 920 return V_008958_DI_PT_TRIFAN; 921 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY: 922 return V_008958_DI_PT_LINELIST_ADJ; 923 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY: 924 return V_008958_DI_PT_LINESTRIP_ADJ; 925 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY: 926 return V_008958_DI_PT_TRILIST_ADJ; 927 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY: 928 return V_008958_DI_PT_TRISTRIP_ADJ; 929 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST: 930 return V_008958_DI_PT_PATCH; 931 default: 932 assert(0); 933 return 0; 934 } 935 } 936 937 static uint32_t 938 si_conv_gl_prim_to_gs_out(unsigned gl_prim) 939 { 940 switch (gl_prim) { 941 case 0: /* GL_POINTS */ 942 return V_028A6C_OUTPRIM_TYPE_POINTLIST; 943 case 1: /* GL_LINES */ 944 case 3: /* GL_LINE_STRIP */ 945 case 0xA: /* GL_LINE_STRIP_ADJACENCY_ARB */ 946 case 0x8E7A: /* GL_ISOLINES */ 947 return V_028A6C_OUTPRIM_TYPE_LINESTRIP; 948 949 case 4: /* GL_TRIANGLES */ 950 case 0xc: /* GL_TRIANGLES_ADJACENCY_ARB */ 951 case 5: /* GL_TRIANGLE_STRIP */ 952 case 7: /* GL_QUADS */ 953 return V_028A6C_OUTPRIM_TYPE_TRISTRIP; 954 default: 955 assert(0); 956 return 0; 957 } 958 } 959 960 static uint32_t 961 si_conv_prim_to_gs_out(enum VkPrimitiveTopology topology) 962 { 963 switch (topology) { 964 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST: 965 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST: 966 return V_028A6C_OUTPRIM_TYPE_POINTLIST; 967 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST: 968 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP: 969 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY: 970 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY: 971 return V_028A6C_OUTPRIM_TYPE_LINESTRIP; 972 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST: 973 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP: 974 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN: 975 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY: 976 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY: 977 return V_028A6C_OUTPRIM_TYPE_TRISTRIP; 978 default: 979 assert(0); 980 return 0; 981 } 982 } 983 984 static unsigned si_map_swizzle(unsigned swizzle) 985 { 986 switch (swizzle) { 987 case VK_SWIZZLE_Y: 988 return V_008F0C_SQ_SEL_Y; 989 case VK_SWIZZLE_Z: 990 return V_008F0C_SQ_SEL_Z; 991 case VK_SWIZZLE_W: 992 return V_008F0C_SQ_SEL_W; 993 case VK_SWIZZLE_0: 994 return V_008F0C_SQ_SEL_0; 995 case VK_SWIZZLE_1: 996 return V_008F0C_SQ_SEL_1; 997 default: /* VK_SWIZZLE_X */ 998 return V_008F0C_SQ_SEL_X; 999 } 1000 } 1001 1002 1003 static unsigned radv_dynamic_state_mask(VkDynamicState state) 1004 { 1005 switch(state) { 1006 case VK_DYNAMIC_STATE_VIEWPORT: 1007 return RADV_DYNAMIC_VIEWPORT; 1008 case VK_DYNAMIC_STATE_SCISSOR: 1009 return RADV_DYNAMIC_SCISSOR; 1010 case VK_DYNAMIC_STATE_LINE_WIDTH: 1011 return RADV_DYNAMIC_LINE_WIDTH; 1012 case VK_DYNAMIC_STATE_DEPTH_BIAS: 1013 return RADV_DYNAMIC_DEPTH_BIAS; 1014 case VK_DYNAMIC_STATE_BLEND_CONSTANTS: 1015 return RADV_DYNAMIC_BLEND_CONSTANTS; 1016 case VK_DYNAMIC_STATE_DEPTH_BOUNDS: 1017 return RADV_DYNAMIC_DEPTH_BOUNDS; 1018 case VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK: 1019 return RADV_DYNAMIC_STENCIL_COMPARE_MASK; 1020 case VK_DYNAMIC_STATE_STENCIL_WRITE_MASK: 1021 return RADV_DYNAMIC_STENCIL_WRITE_MASK; 1022 case VK_DYNAMIC_STATE_STENCIL_REFERENCE: 1023 return RADV_DYNAMIC_STENCIL_REFERENCE; 1024 case VK_DYNAMIC_STATE_DISCARD_RECTANGLE_EXT: 1025 return RADV_DYNAMIC_DISCARD_RECTANGLE; 1026 default: 1027 unreachable("Unhandled dynamic state"); 1028 } 1029 } 1030 1031 static void 1032 radv_pipeline_init_dynamic_state(struct radv_pipeline *pipeline, 1033 const VkGraphicsPipelineCreateInfo *pCreateInfo) 1034 { 1035 uint32_t states = RADV_DYNAMIC_ALL; 1036 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass); 1037 struct radv_subpass *subpass = &pass->subpasses[pCreateInfo->subpass]; 1038 1039 pipeline->dynamic_state = default_dynamic_state; 1040 1041 if (pCreateInfo->pDynamicState) { 1042 /* Remove all of the states that are marked as dynamic */ 1043 uint32_t count = pCreateInfo->pDynamicState->dynamicStateCount; 1044 for (uint32_t s = 0; s < count; s++) 1045 states &= ~radv_dynamic_state_mask(pCreateInfo->pDynamicState->pDynamicStates[s]); 1046 } 1047 1048 struct radv_dynamic_state *dynamic = &pipeline->dynamic_state; 1049 1050 /* Section 9.2 of the Vulkan 1.0.15 spec says: 1051 * 1052 * pViewportState is [...] NULL if the pipeline 1053 * has rasterization disabled. 1054 */ 1055 if (!pCreateInfo->pRasterizationState->rasterizerDiscardEnable) { 1056 assert(pCreateInfo->pViewportState); 1057 1058 dynamic->viewport.count = pCreateInfo->pViewportState->viewportCount; 1059 if (states & RADV_DYNAMIC_VIEWPORT) { 1060 typed_memcpy(dynamic->viewport.viewports, 1061 pCreateInfo->pViewportState->pViewports, 1062 pCreateInfo->pViewportState->viewportCount); 1063 } 1064 1065 dynamic->scissor.count = pCreateInfo->pViewportState->scissorCount; 1066 if (states & RADV_DYNAMIC_SCISSOR) { 1067 typed_memcpy(dynamic->scissor.scissors, 1068 pCreateInfo->pViewportState->pScissors, 1069 pCreateInfo->pViewportState->scissorCount); 1070 } 1071 } 1072 1073 if (states & RADV_DYNAMIC_LINE_WIDTH) { 1074 assert(pCreateInfo->pRasterizationState); 1075 dynamic->line_width = pCreateInfo->pRasterizationState->lineWidth; 1076 } 1077 1078 if (states & RADV_DYNAMIC_DEPTH_BIAS) { 1079 assert(pCreateInfo->pRasterizationState); 1080 dynamic->depth_bias.bias = 1081 pCreateInfo->pRasterizationState->depthBiasConstantFactor; 1082 dynamic->depth_bias.clamp = 1083 pCreateInfo->pRasterizationState->depthBiasClamp; 1084 dynamic->depth_bias.slope = 1085 pCreateInfo->pRasterizationState->depthBiasSlopeFactor; 1086 } 1087 1088 /* Section 9.2 of the Vulkan 1.0.15 spec says: 1089 * 1090 * pColorBlendState is [...] NULL if the pipeline has rasterization 1091 * disabled or if the subpass of the render pass the pipeline is 1092 * created against does not use any color attachments. 1093 */ 1094 bool uses_color_att = false; 1095 for (unsigned i = 0; i < subpass->color_count; ++i) { 1096 if (subpass->color_attachments[i].attachment != VK_ATTACHMENT_UNUSED) { 1097 uses_color_att = true; 1098 break; 1099 } 1100 } 1101 1102 if (uses_color_att && states & RADV_DYNAMIC_BLEND_CONSTANTS) { 1103 assert(pCreateInfo->pColorBlendState); 1104 typed_memcpy(dynamic->blend_constants, 1105 pCreateInfo->pColorBlendState->blendConstants, 4); 1106 } 1107 1108 /* If there is no depthstencil attachment, then don't read 1109 * pDepthStencilState. The Vulkan spec states that pDepthStencilState may 1110 * be NULL in this case. Even if pDepthStencilState is non-NULL, there is 1111 * no need to override the depthstencil defaults in 1112 * radv_pipeline::dynamic_state when there is no depthstencil attachment. 1113 * 1114 * Section 9.2 of the Vulkan 1.0.15 spec says: 1115 * 1116 * pDepthStencilState is [...] NULL if the pipeline has rasterization 1117 * disabled or if the subpass of the render pass the pipeline is created 1118 * against does not use a depth/stencil attachment. 1119 */ 1120 if (!pCreateInfo->pRasterizationState->rasterizerDiscardEnable && 1121 subpass->depth_stencil_attachment.attachment != VK_ATTACHMENT_UNUSED) { 1122 assert(pCreateInfo->pDepthStencilState); 1123 1124 if (states & RADV_DYNAMIC_DEPTH_BOUNDS) { 1125 dynamic->depth_bounds.min = 1126 pCreateInfo->pDepthStencilState->minDepthBounds; 1127 dynamic->depth_bounds.max = 1128 pCreateInfo->pDepthStencilState->maxDepthBounds; 1129 } 1130 1131 if (states & RADV_DYNAMIC_STENCIL_COMPARE_MASK) { 1132 dynamic->stencil_compare_mask.front = 1133 pCreateInfo->pDepthStencilState->front.compareMask; 1134 dynamic->stencil_compare_mask.back = 1135 pCreateInfo->pDepthStencilState->back.compareMask; 1136 } 1137 1138 if (states & RADV_DYNAMIC_STENCIL_WRITE_MASK) { 1139 dynamic->stencil_write_mask.front = 1140 pCreateInfo->pDepthStencilState->front.writeMask; 1141 dynamic->stencil_write_mask.back = 1142 pCreateInfo->pDepthStencilState->back.writeMask; 1143 } 1144 1145 if (states & RADV_DYNAMIC_STENCIL_REFERENCE) { 1146 dynamic->stencil_reference.front = 1147 pCreateInfo->pDepthStencilState->front.reference; 1148 dynamic->stencil_reference.back = 1149 pCreateInfo->pDepthStencilState->back.reference; 1150 } 1151 } 1152 1153 const VkPipelineDiscardRectangleStateCreateInfoEXT *discard_rectangle_info = 1154 vk_find_struct_const(pCreateInfo->pNext, PIPELINE_DISCARD_RECTANGLE_STATE_CREATE_INFO_EXT); 1155 if (discard_rectangle_info) { 1156 dynamic->discard_rectangle.count = discard_rectangle_info->discardRectangleCount; 1157 typed_memcpy(dynamic->discard_rectangle.rectangles, 1158 discard_rectangle_info->pDiscardRectangles, 1159 discard_rectangle_info->discardRectangleCount); 1160 1161 unsigned mask = 0; 1162 1163 for (unsigned i = 0; i < (1u << MAX_DISCARD_RECTANGLES); ++i) { 1164 /* Interpret i as a bitmask, and then set the bit in the mask if 1165 * that combination of rectangles in which the pixel is contained 1166 * should pass the cliprect test. */ 1167 unsigned relevant_subset = i & ((1u << discard_rectangle_info->discardRectangleCount) - 1); 1168 1169 if (discard_rectangle_info->discardRectangleMode == VK_DISCARD_RECTANGLE_MODE_INCLUSIVE_EXT && 1170 !relevant_subset) 1171 continue; 1172 1173 if (discard_rectangle_info->discardRectangleMode == VK_DISCARD_RECTANGLE_MODE_EXCLUSIVE_EXT && 1174 relevant_subset) 1175 continue; 1176 1177 mask |= 1u << i; 1178 } 1179 pipeline->graphics.pa_sc_cliprect_rule = mask; 1180 } else { 1181 states &= ~RADV_DYNAMIC_DISCARD_RECTANGLE; 1182 1183 /* Allow from all rectangle combinations */ 1184 pipeline->graphics.pa_sc_cliprect_rule = 0xffff; 1185 } 1186 pipeline->dynamic_state.mask = states; 1187 } 1188 1189 static void calculate_gfx9_gs_info(const VkGraphicsPipelineCreateInfo *pCreateInfo, 1190 struct radv_pipeline *pipeline) 1191 { 1192 struct ac_shader_variant_info *gs_info = &pipeline->shaders[MESA_SHADER_GEOMETRY]->info; 1193 struct ac_es_output_info *es_info = radv_pipeline_has_tess(pipeline) ? 1194 &gs_info->tes.es_info : &gs_info->vs.es_info; 1195 unsigned gs_num_invocations = MAX2(gs_info->gs.invocations, 1); 1196 bool uses_adjacency; 1197 switch(pCreateInfo->pInputAssemblyState->topology) { 1198 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY: 1199 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY: 1200 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY: 1201 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY: 1202 uses_adjacency = true; 1203 break; 1204 default: 1205 uses_adjacency = false; 1206 break; 1207 } 1208 1209 /* All these are in dwords: */ 1210 /* We can't allow using the whole LDS, because GS waves compete with 1211 * other shader stages for LDS space. */ 1212 const unsigned max_lds_size = 8 * 1024; 1213 const unsigned esgs_itemsize = es_info->esgs_itemsize / 4; 1214 unsigned esgs_lds_size; 1215 1216 /* All these are per subgroup: */ 1217 const unsigned max_out_prims = 32 * 1024; 1218 const unsigned max_es_verts = 255; 1219 const unsigned ideal_gs_prims = 64; 1220 unsigned max_gs_prims, gs_prims; 1221 unsigned min_es_verts, es_verts, worst_case_es_verts; 1222 1223 if (uses_adjacency || gs_num_invocations > 1) 1224 max_gs_prims = 127 / gs_num_invocations; 1225 else 1226 max_gs_prims = 255; 1227 1228 /* MAX_PRIMS_PER_SUBGROUP = gs_prims * max_vert_out * gs_invocations. 1229 * Make sure we don't go over the maximum value. 1230 */ 1231 if (gs_info->gs.vertices_out > 0) { 1232 max_gs_prims = MIN2(max_gs_prims, 1233 max_out_prims / 1234 (gs_info->gs.vertices_out * gs_num_invocations)); 1235 } 1236 assert(max_gs_prims > 0); 1237 1238 /* If the primitive has adjacency, halve the number of vertices 1239 * that will be reused in multiple primitives. 1240 */ 1241 min_es_verts = gs_info->gs.vertices_in / (uses_adjacency ? 2 : 1); 1242 1243 gs_prims = MIN2(ideal_gs_prims, max_gs_prims); 1244 worst_case_es_verts = MIN2(min_es_verts * gs_prims, max_es_verts); 1245 1246 /* Compute ESGS LDS size based on the worst case number of ES vertices 1247 * needed to create the target number of GS prims per subgroup. 1248 */ 1249 esgs_lds_size = esgs_itemsize * worst_case_es_verts; 1250 1251 /* If total LDS usage is too big, refactor partitions based on ratio 1252 * of ESGS item sizes. 1253 */ 1254 if (esgs_lds_size > max_lds_size) { 1255 /* Our target GS Prims Per Subgroup was too large. Calculate 1256 * the maximum number of GS Prims Per Subgroup that will fit 1257 * into LDS, capped by the maximum that the hardware can support. 1258 */ 1259 gs_prims = MIN2((max_lds_size / (esgs_itemsize * min_es_verts)), 1260 max_gs_prims); 1261 assert(gs_prims > 0); 1262 worst_case_es_verts = MIN2(min_es_verts * gs_prims, 1263 max_es_verts); 1264 1265 esgs_lds_size = esgs_itemsize * worst_case_es_verts; 1266 assert(esgs_lds_size <= max_lds_size); 1267 } 1268 1269 /* Now calculate remaining ESGS information. */ 1270 if (esgs_lds_size) 1271 es_verts = MIN2(esgs_lds_size / esgs_itemsize, max_es_verts); 1272 else 1273 es_verts = max_es_verts; 1274 1275 /* Vertices for adjacency primitives are not always reused, so restore 1276 * it for ES_VERTS_PER_SUBGRP. 1277 */ 1278 min_es_verts = gs_info->gs.vertices_in; 1279 1280 /* For normal primitives, the VGT only checks if they are past the ES 1281 * verts per subgroup after allocating a full GS primitive and if they 1282 * are, kick off a new subgroup. But if those additional ES verts are 1283 * unique (e.g. not reused) we need to make sure there is enough LDS 1284 * space to account for those ES verts beyond ES_VERTS_PER_SUBGRP. 1285 */ 1286 es_verts -= min_es_verts - 1; 1287 1288 uint32_t es_verts_per_subgroup = es_verts; 1289 uint32_t gs_prims_per_subgroup = gs_prims; 1290 uint32_t gs_inst_prims_in_subgroup = gs_prims * gs_num_invocations; 1291 uint32_t max_prims_per_subgroup = gs_inst_prims_in_subgroup * gs_info->gs.vertices_out; 1292 pipeline->graphics.gs.lds_size = align(esgs_lds_size, 128) / 128; 1293 pipeline->graphics.gs.vgt_gs_onchip_cntl = 1294 S_028A44_ES_VERTS_PER_SUBGRP(es_verts_per_subgroup) | 1295 S_028A44_GS_PRIMS_PER_SUBGRP(gs_prims_per_subgroup) | 1296 S_028A44_GS_INST_PRIMS_IN_SUBGRP(gs_inst_prims_in_subgroup); 1297 pipeline->graphics.gs.vgt_gs_max_prims_per_subgroup = 1298 S_028A94_MAX_PRIMS_PER_SUBGROUP(max_prims_per_subgroup); 1299 pipeline->graphics.gs.vgt_esgs_ring_itemsize = esgs_itemsize; 1300 assert(max_prims_per_subgroup <= max_out_prims); 1301 } 1302 1303 static void 1304 calculate_gs_ring_sizes(struct radv_pipeline *pipeline) 1305 { 1306 struct radv_device *device = pipeline->device; 1307 unsigned num_se = device->physical_device->rad_info.max_se; 1308 unsigned wave_size = 64; 1309 unsigned max_gs_waves = 32 * num_se; /* max 32 per SE on GCN */ 1310 unsigned gs_vertex_reuse = 16 * num_se; /* GS_VERTEX_REUSE register (per SE) */ 1311 unsigned alignment = 256 * num_se; 1312 /* The maximum size is 63.999 MB per SE. */ 1313 unsigned max_size = ((unsigned)(63.999 * 1024 * 1024) & ~255) * num_se; 1314 struct ac_shader_variant_info *gs_info = &pipeline->shaders[MESA_SHADER_GEOMETRY]->info; 1315 struct ac_es_output_info *es_info; 1316 if (pipeline->device->physical_device->rad_info.chip_class >= GFX9) 1317 es_info = radv_pipeline_has_tess(pipeline) ? &gs_info->tes.es_info : &gs_info->vs.es_info; 1318 else 1319 es_info = radv_pipeline_has_tess(pipeline) ? 1320 &pipeline->shaders[MESA_SHADER_TESS_EVAL]->info.tes.es_info : 1321 &pipeline->shaders[MESA_SHADER_VERTEX]->info.vs.es_info; 1322 1323 /* Calculate the minimum size. */ 1324 unsigned min_esgs_ring_size = align(es_info->esgs_itemsize * gs_vertex_reuse * 1325 wave_size, alignment); 1326 /* These are recommended sizes, not minimum sizes. */ 1327 unsigned esgs_ring_size = max_gs_waves * 2 * wave_size * 1328 es_info->esgs_itemsize * gs_info->gs.vertices_in; 1329 unsigned gsvs_ring_size = max_gs_waves * 2 * wave_size * 1330 gs_info->gs.max_gsvs_emit_size * 1; // no streams in VK (gs->max_gs_stream + 1); 1331 1332 min_esgs_ring_size = align(min_esgs_ring_size, alignment); 1333 esgs_ring_size = align(esgs_ring_size, alignment); 1334 gsvs_ring_size = align(gsvs_ring_size, alignment); 1335 1336 if (pipeline->device->physical_device->rad_info.chip_class <= VI) 1337 pipeline->graphics.esgs_ring_size = CLAMP(esgs_ring_size, min_esgs_ring_size, max_size); 1338 1339 pipeline->graphics.gs.vgt_esgs_ring_itemsize = es_info->esgs_itemsize / 4; 1340 pipeline->graphics.gsvs_ring_size = MIN2(gsvs_ring_size, max_size); 1341 } 1342 1343 static void si_multiwave_lds_size_workaround(struct radv_device *device, 1344 unsigned *lds_size) 1345 { 1346 /* SPI barrier management bug: 1347 * Make sure we have at least 4k of LDS in use to avoid the bug. 1348 * It applies to workgroup sizes of more than one wavefront. 1349 */ 1350 if (device->physical_device->rad_info.family == CHIP_BONAIRE || 1351 device->physical_device->rad_info.family == CHIP_KABINI || 1352 device->physical_device->rad_info.family == CHIP_MULLINS) 1353 *lds_size = MAX2(*lds_size, 8); 1354 } 1355 1356 struct radv_shader_variant * 1357 radv_get_vertex_shader(struct radv_pipeline *pipeline) 1358 { 1359 if (pipeline->shaders[MESA_SHADER_VERTEX]) 1360 return pipeline->shaders[MESA_SHADER_VERTEX]; 1361 if (pipeline->shaders[MESA_SHADER_TESS_CTRL]) 1362 return pipeline->shaders[MESA_SHADER_TESS_CTRL]; 1363 return pipeline->shaders[MESA_SHADER_GEOMETRY]; 1364 } 1365 1366 static struct radv_shader_variant * 1367 radv_get_tess_eval_shader(struct radv_pipeline *pipeline) 1368 { 1369 if (pipeline->shaders[MESA_SHADER_TESS_EVAL]) 1370 return pipeline->shaders[MESA_SHADER_TESS_EVAL]; 1371 return pipeline->shaders[MESA_SHADER_GEOMETRY]; 1372 } 1373 1374 static void 1375 calculate_tess_state(struct radv_pipeline *pipeline, 1376 const VkGraphicsPipelineCreateInfo *pCreateInfo) 1377 { 1378 unsigned num_tcs_input_cp = pCreateInfo->pTessellationState->patchControlPoints; 1379 unsigned num_tcs_output_cp, num_tcs_inputs, num_tcs_outputs; 1380 unsigned num_tcs_patch_outputs; 1381 unsigned input_vertex_size, output_vertex_size, pervertex_output_patch_size; 1382 unsigned input_patch_size, output_patch_size, output_patch0_offset; 1383 unsigned lds_size, hardware_lds_size; 1384 unsigned perpatch_output_offset; 1385 unsigned num_patches; 1386 struct radv_tessellation_state *tess = &pipeline->graphics.tess; 1387 1388 /* This calculates how shader inputs and outputs among VS, TCS, and TES 1389 * are laid out in LDS. */ 1390 num_tcs_inputs = util_last_bit64(radv_get_vertex_shader(pipeline)->info.vs.outputs_written); 1391 1392 num_tcs_outputs = util_last_bit64(pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.outputs_written); //tcs->outputs_written 1393 num_tcs_output_cp = pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.tcs_vertices_out; //TCS VERTICES OUT 1394 num_tcs_patch_outputs = util_last_bit64(pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.tcs.patch_outputs_written); 1395 1396 /* Ensure that we only need one wave per SIMD so we don't need to check 1397 * resource usage. Also ensures that the number of tcs in and out 1398 * vertices per threadgroup are at most 256. 1399 */ 1400 input_vertex_size = num_tcs_inputs * 16; 1401 output_vertex_size = num_tcs_outputs * 16; 1402 1403 input_patch_size = num_tcs_input_cp * input_vertex_size; 1404 1405 pervertex_output_patch_size = num_tcs_output_cp * output_vertex_size; 1406 output_patch_size = pervertex_output_patch_size + num_tcs_patch_outputs * 16; 1407 /* Ensure that we only need one wave per SIMD so we don't need to check 1408 * resource usage. Also ensures that the number of tcs in and out 1409 * vertices per threadgroup are at most 256. 1410 */ 1411 num_patches = 64 / MAX2(num_tcs_input_cp, num_tcs_output_cp) * 4; 1412 1413 /* Make sure that the data fits in LDS. This assumes the shaders only 1414 * use LDS for the inputs and outputs. 1415 */ 1416 hardware_lds_size = pipeline->device->physical_device->rad_info.chip_class >= CIK ? 65536 : 32768; 1417 num_patches = MIN2(num_patches, hardware_lds_size / (input_patch_size + output_patch_size)); 1418 1419 /* Make sure the output data fits in the offchip buffer */ 1420 num_patches = MIN2(num_patches, 1421 (pipeline->device->tess_offchip_block_dw_size * 4) / 1422 output_patch_size); 1423 1424 /* Not necessary for correctness, but improves performance. The 1425 * specific value is taken from the proprietary driver. 1426 */ 1427 num_patches = MIN2(num_patches, 40); 1428 1429 /* SI bug workaround - limit LS-HS threadgroups to only one wave. */ 1430 if (pipeline->device->physical_device->rad_info.chip_class == SI) { 1431 unsigned one_wave = 64 / MAX2(num_tcs_input_cp, num_tcs_output_cp); 1432 num_patches = MIN2(num_patches, one_wave); 1433 } 1434 1435 output_patch0_offset = input_patch_size * num_patches; 1436 perpatch_output_offset = output_patch0_offset + pervertex_output_patch_size; 1437 1438 lds_size = output_patch0_offset + output_patch_size * num_patches; 1439 1440 if (pipeline->device->physical_device->rad_info.chip_class >= CIK) { 1441 assert(lds_size <= 65536); 1442 lds_size = align(lds_size, 512) / 512; 1443 } else { 1444 assert(lds_size <= 32768); 1445 lds_size = align(lds_size, 256) / 256; 1446 } 1447 si_multiwave_lds_size_workaround(pipeline->device, &lds_size); 1448 1449 tess->lds_size = lds_size; 1450 1451 tess->tcs_in_layout = (input_patch_size / 4) | 1452 ((input_vertex_size / 4) << 13); 1453 tess->tcs_out_layout = (output_patch_size / 4) | 1454 ((output_vertex_size / 4) << 13); 1455 tess->tcs_out_offsets = (output_patch0_offset / 16) | 1456 ((perpatch_output_offset / 16) << 16); 1457 tess->offchip_layout = (pervertex_output_patch_size * num_patches << 16) | 1458 (num_tcs_output_cp << 9) | num_patches; 1459 1460 tess->ls_hs_config = S_028B58_NUM_PATCHES(num_patches) | 1461 S_028B58_HS_NUM_INPUT_CP(num_tcs_input_cp) | 1462 S_028B58_HS_NUM_OUTPUT_CP(num_tcs_output_cp); 1463 tess->num_patches = num_patches; 1464 tess->num_tcs_input_cp = num_tcs_input_cp; 1465 1466 struct radv_shader_variant *tes = radv_get_tess_eval_shader(pipeline); 1467 unsigned type = 0, partitioning = 0, topology = 0, distribution_mode = 0; 1468 1469 switch (tes->info.tes.primitive_mode) { 1470 case GL_TRIANGLES: 1471 type = V_028B6C_TESS_TRIANGLE; 1472 break; 1473 case GL_QUADS: 1474 type = V_028B6C_TESS_QUAD; 1475 break; 1476 case GL_ISOLINES: 1477 type = V_028B6C_TESS_ISOLINE; 1478 break; 1479 } 1480 1481 switch (tes->info.tes.spacing) { 1482 case TESS_SPACING_EQUAL: 1483 partitioning = V_028B6C_PART_INTEGER; 1484 break; 1485 case TESS_SPACING_FRACTIONAL_ODD: 1486 partitioning = V_028B6C_PART_FRAC_ODD; 1487 break; 1488 case TESS_SPACING_FRACTIONAL_EVEN: 1489 partitioning = V_028B6C_PART_FRAC_EVEN; 1490 break; 1491 default: 1492 break; 1493 } 1494 1495 bool ccw = tes->info.tes.ccw; 1496 const VkPipelineTessellationDomainOriginStateCreateInfoKHR *domain_origin_state = 1497 vk_find_struct_const(pCreateInfo->pTessellationState, 1498 PIPELINE_TESSELLATION_DOMAIN_ORIGIN_STATE_CREATE_INFO_KHR); 1499 1500 if (domain_origin_state && domain_origin_state->domainOrigin != VK_TESSELLATION_DOMAIN_ORIGIN_UPPER_LEFT_KHR) 1501 ccw = !ccw; 1502 1503 if (tes->info.tes.point_mode) 1504 topology = V_028B6C_OUTPUT_POINT; 1505 else if (tes->info.tes.primitive_mode == GL_ISOLINES) 1506 topology = V_028B6C_OUTPUT_LINE; 1507 else if (ccw) 1508 topology = V_028B6C_OUTPUT_TRIANGLE_CCW; 1509 else 1510 topology = V_028B6C_OUTPUT_TRIANGLE_CW; 1511 1512 if (pipeline->device->has_distributed_tess) { 1513 if (pipeline->device->physical_device->rad_info.family == CHIP_FIJI || 1514 pipeline->device->physical_device->rad_info.family >= CHIP_POLARIS10) 1515 distribution_mode = V_028B6C_DISTRIBUTION_MODE_TRAPEZOIDS; 1516 else 1517 distribution_mode = V_028B6C_DISTRIBUTION_MODE_DONUTS; 1518 } else 1519 distribution_mode = V_028B6C_DISTRIBUTION_MODE_NO_DIST; 1520 1521 tess->tf_param = S_028B6C_TYPE(type) | 1522 S_028B6C_PARTITIONING(partitioning) | 1523 S_028B6C_TOPOLOGY(topology) | 1524 S_028B6C_DISTRIBUTION_MODE(distribution_mode); 1525 } 1526 1527 static const struct radv_prim_vertex_count prim_size_table[] = { 1528 [V_008958_DI_PT_NONE] = {0, 0}, 1529 [V_008958_DI_PT_POINTLIST] = {1, 1}, 1530 [V_008958_DI_PT_LINELIST] = {2, 2}, 1531 [V_008958_DI_PT_LINESTRIP] = {2, 1}, 1532 [V_008958_DI_PT_TRILIST] = {3, 3}, 1533 [V_008958_DI_PT_TRIFAN] = {3, 1}, 1534 [V_008958_DI_PT_TRISTRIP] = {3, 1}, 1535 [V_008958_DI_PT_LINELIST_ADJ] = {4, 4}, 1536 [V_008958_DI_PT_LINESTRIP_ADJ] = {4, 1}, 1537 [V_008958_DI_PT_TRILIST_ADJ] = {6, 6}, 1538 [V_008958_DI_PT_TRISTRIP_ADJ] = {6, 2}, 1539 [V_008958_DI_PT_RECTLIST] = {3, 3}, 1540 [V_008958_DI_PT_LINELOOP] = {2, 1}, 1541 [V_008958_DI_PT_POLYGON] = {3, 1}, 1542 [V_008958_DI_PT_2D_TRI_STRIP] = {0, 0}, 1543 }; 1544 1545 static struct ac_vs_output_info *get_vs_output_info(struct radv_pipeline *pipeline) 1546 { 1547 if (radv_pipeline_has_gs(pipeline)) 1548 return &pipeline->gs_copy_shader->info.vs.outinfo; 1549 else if (radv_pipeline_has_tess(pipeline)) 1550 return &pipeline->shaders[MESA_SHADER_TESS_EVAL]->info.tes.outinfo; 1551 else 1552 return &pipeline->shaders[MESA_SHADER_VERTEX]->info.vs.outinfo; 1553 } 1554 1555 static void calculate_vgt_gs_mode(struct radv_pipeline *pipeline) 1556 { 1557 struct ac_vs_output_info *outinfo = get_vs_output_info(pipeline); 1558 1559 pipeline->graphics.vgt_primitiveid_en = false; 1560 pipeline->graphics.vgt_gs_mode = 0; 1561 1562 if (radv_pipeline_has_gs(pipeline)) { 1563 struct radv_shader_variant *gs = 1564 pipeline->shaders[MESA_SHADER_GEOMETRY]; 1565 1566 pipeline->graphics.vgt_gs_mode = 1567 ac_vgt_gs_mode(gs->info.gs.vertices_out, 1568 pipeline->device->physical_device->rad_info.chip_class); 1569 } else if (outinfo->export_prim_id) { 1570 pipeline->graphics.vgt_gs_mode = S_028A40_MODE(V_028A40_GS_SCENARIO_A); 1571 pipeline->graphics.vgt_primitiveid_en = true; 1572 } 1573 } 1574 1575 static void calculate_vs_outinfo(struct radv_pipeline *pipeline) 1576 { 1577 struct ac_vs_output_info *outinfo = get_vs_output_info(pipeline); 1578 1579 unsigned clip_dist_mask, cull_dist_mask, total_mask; 1580 clip_dist_mask = outinfo->clip_dist_mask; 1581 cull_dist_mask = outinfo->cull_dist_mask; 1582 total_mask = clip_dist_mask | cull_dist_mask; 1583 1584 bool misc_vec_ena = outinfo->writes_pointsize || 1585 outinfo->writes_layer || 1586 outinfo->writes_viewport_index; 1587 pipeline->graphics.vs.pa_cl_vs_out_cntl = 1588 S_02881C_USE_VTX_POINT_SIZE(outinfo->writes_pointsize) | 1589 S_02881C_USE_VTX_RENDER_TARGET_INDX(outinfo->writes_layer) | 1590 S_02881C_USE_VTX_VIEWPORT_INDX(outinfo->writes_viewport_index) | 1591 S_02881C_VS_OUT_MISC_VEC_ENA(misc_vec_ena) | 1592 S_02881C_VS_OUT_MISC_SIDE_BUS_ENA(misc_vec_ena) | 1593 S_02881C_VS_OUT_CCDIST0_VEC_ENA((total_mask & 0x0f) != 0) | 1594 S_02881C_VS_OUT_CCDIST1_VEC_ENA((total_mask & 0xf0) != 0) | 1595 cull_dist_mask << 8 | 1596 clip_dist_mask; 1597 1598 pipeline->graphics.vs.spi_shader_pos_format = 1599 S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP) | 1600 S_02870C_POS1_EXPORT_FORMAT(outinfo->pos_exports > 1 ? 1601 V_02870C_SPI_SHADER_4COMP : 1602 V_02870C_SPI_SHADER_NONE) | 1603 S_02870C_POS2_EXPORT_FORMAT(outinfo->pos_exports > 2 ? 1604 V_02870C_SPI_SHADER_4COMP : 1605 V_02870C_SPI_SHADER_NONE) | 1606 S_02870C_POS3_EXPORT_FORMAT(outinfo->pos_exports > 3 ? 1607 V_02870C_SPI_SHADER_4COMP : 1608 V_02870C_SPI_SHADER_NONE); 1609 1610 pipeline->graphics.vs.spi_vs_out_config = S_0286C4_VS_EXPORT_COUNT(MAX2(1, outinfo->param_exports) - 1); 1611 /* only emitted on pre-VI */ 1612 pipeline->graphics.vs.vgt_reuse_off = S_028AB4_REUSE_OFF(outinfo->writes_viewport_index); 1613 } 1614 1615 static uint32_t offset_to_ps_input(uint32_t offset, bool flat_shade) 1616 { 1617 uint32_t ps_input_cntl; 1618 if (offset <= AC_EXP_PARAM_OFFSET_31) { 1619 ps_input_cntl = S_028644_OFFSET(offset); 1620 if (flat_shade) 1621 ps_input_cntl |= S_028644_FLAT_SHADE(1); 1622 } else { 1623 /* The input is a DEFAULT_VAL constant. */ 1624 assert(offset >= AC_EXP_PARAM_DEFAULT_VAL_0000 && 1625 offset <= AC_EXP_PARAM_DEFAULT_VAL_1111); 1626 offset -= AC_EXP_PARAM_DEFAULT_VAL_0000; 1627 ps_input_cntl = S_028644_OFFSET(0x20) | 1628 S_028644_DEFAULT_VAL(offset); 1629 } 1630 return ps_input_cntl; 1631 } 1632 1633 static void calculate_ps_inputs(struct radv_pipeline *pipeline) 1634 { 1635 struct radv_shader_variant *ps; 1636 struct ac_vs_output_info *outinfo = get_vs_output_info(pipeline); 1637 1638 ps = pipeline->shaders[MESA_SHADER_FRAGMENT]; 1639 1640 unsigned ps_offset = 0; 1641 1642 if (ps->info.fs.prim_id_input) { 1643 unsigned vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_PRIMITIVE_ID]; 1644 if (vs_offset != AC_EXP_PARAM_UNDEFINED) { 1645 pipeline->graphics.ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, true); 1646 ++ps_offset; 1647 } 1648 } 1649 1650 if (ps->info.fs.layer_input) { 1651 unsigned vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_LAYER]; 1652 if (vs_offset != AC_EXP_PARAM_UNDEFINED) 1653 pipeline->graphics.ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, true); 1654 else 1655 pipeline->graphics.ps_input_cntl[ps_offset] = offset_to_ps_input(AC_EXP_PARAM_DEFAULT_VAL_0000, true); 1656 ++ps_offset; 1657 } 1658 1659 if (ps->info.fs.has_pcoord) { 1660 unsigned val; 1661 val = S_028644_PT_SPRITE_TEX(1) | S_028644_OFFSET(0x20); 1662 pipeline->graphics.ps_input_cntl[ps_offset] = val; 1663 ps_offset++; 1664 } 1665 1666 for (unsigned i = 0; i < 32 && (1u << i) <= ps->info.fs.input_mask; ++i) { 1667 unsigned vs_offset; 1668 bool flat_shade; 1669 if (!(ps->info.fs.input_mask & (1u << i))) 1670 continue; 1671 1672 vs_offset = outinfo->vs_output_param_offset[VARYING_SLOT_VAR0 + i]; 1673 if (vs_offset == AC_EXP_PARAM_UNDEFINED) { 1674 pipeline->graphics.ps_input_cntl[ps_offset] = S_028644_OFFSET(0x20); 1675 ++ps_offset; 1676 continue; 1677 } 1678 1679 flat_shade = !!(ps->info.fs.flat_shaded_mask & (1u << ps_offset)); 1680 1681 pipeline->graphics.ps_input_cntl[ps_offset] = offset_to_ps_input(vs_offset, flat_shade); 1682 ++ps_offset; 1683 } 1684 1685 pipeline->graphics.ps_input_cntl_num = ps_offset; 1686 } 1687 1688 static void 1689 radv_link_shaders(struct radv_pipeline *pipeline, nir_shader **shaders) 1690 { 1691 nir_shader* ordered_shaders[MESA_SHADER_STAGES]; 1692 int shader_count = 0; 1693 1694 if(shaders[MESA_SHADER_FRAGMENT]) { 1695 ordered_shaders[shader_count++] = shaders[MESA_SHADER_FRAGMENT]; 1696 } 1697 if(shaders[MESA_SHADER_GEOMETRY]) { 1698 ordered_shaders[shader_count++] = shaders[MESA_SHADER_GEOMETRY]; 1699 } 1700 if(shaders[MESA_SHADER_TESS_EVAL]) { 1701 ordered_shaders[shader_count++] = shaders[MESA_SHADER_TESS_EVAL]; 1702 } 1703 if(shaders[MESA_SHADER_TESS_CTRL]) { 1704 ordered_shaders[shader_count++] = shaders[MESA_SHADER_TESS_CTRL]; 1705 } 1706 if(shaders[MESA_SHADER_VERTEX]) { 1707 ordered_shaders[shader_count++] = shaders[MESA_SHADER_VERTEX]; 1708 } 1709 1710 for (int i = 1; i < shader_count; ++i) { 1711 nir_lower_io_arrays_to_elements(ordered_shaders[i], 1712 ordered_shaders[i - 1]); 1713 1714 nir_remove_dead_variables(ordered_shaders[i], 1715 nir_var_shader_out); 1716 nir_remove_dead_variables(ordered_shaders[i - 1], 1717 nir_var_shader_in); 1718 1719 bool progress = nir_remove_unused_varyings(ordered_shaders[i], 1720 ordered_shaders[i - 1]); 1721 1722 if (progress) { 1723 if (nir_lower_global_vars_to_local(ordered_shaders[i])) { 1724 radv_lower_indirect_derefs(ordered_shaders[i], 1725 pipeline->device->physical_device); 1726 } 1727 radv_optimize_nir(ordered_shaders[i]); 1728 1729 if (nir_lower_global_vars_to_local(ordered_shaders[i - 1])) { 1730 radv_lower_indirect_derefs(ordered_shaders[i - 1], 1731 pipeline->device->physical_device); 1732 } 1733 radv_optimize_nir(ordered_shaders[i - 1]); 1734 } 1735 } 1736 } 1737 1738 1739 static struct radv_pipeline_key 1740 radv_generate_graphics_pipeline_key(struct radv_pipeline *pipeline, 1741 const VkGraphicsPipelineCreateInfo *pCreateInfo, 1742 bool has_view_index) 1743 { 1744 const VkPipelineVertexInputStateCreateInfo *input_state = 1745 pCreateInfo->pVertexInputState; 1746 struct radv_pipeline_key key; 1747 memset(&key, 0, sizeof(key)); 1748 1749 key.has_multiview_view_index = has_view_index; 1750 1751 uint32_t binding_input_rate = 0; 1752 for (unsigned i = 0; i < input_state->vertexBindingDescriptionCount; ++i) { 1753 if (input_state->pVertexBindingDescriptions[i].inputRate) 1754 binding_input_rate |= 1u << input_state->pVertexBindingDescriptions[i].binding; 1755 } 1756 1757 for (unsigned i = 0; i < input_state->vertexAttributeDescriptionCount; ++i) { 1758 unsigned binding; 1759 binding = input_state->pVertexAttributeDescriptions[i].binding; 1760 if (binding_input_rate & (1u << binding)) 1761 key.instance_rate_inputs |= 1u << input_state->pVertexAttributeDescriptions[i].location; 1762 } 1763 1764 if (pCreateInfo->pTessellationState) 1765 key.tess_input_vertices = pCreateInfo->pTessellationState->patchControlPoints; 1766 1767 1768 if (pCreateInfo->pMultisampleState && 1769 pCreateInfo->pMultisampleState->rasterizationSamples > 1) { 1770 uint32_t num_samples = pCreateInfo->pMultisampleState->rasterizationSamples; 1771 uint32_t ps_iter_samples = radv_pipeline_get_ps_iter_samples(pCreateInfo->pMultisampleState); 1772 key.multisample = true; 1773 key.log2_num_samples = util_logbase2(num_samples); 1774 key.log2_ps_iter_samples = util_logbase2(ps_iter_samples); 1775 } 1776 1777 key.col_format = pipeline->graphics.blend.spi_shader_col_format; 1778 if (pipeline->device->physical_device->rad_info.chip_class < VI) 1779 radv_pipeline_compute_get_int_clamp(pCreateInfo, &key.is_int8, &key.is_int10); 1780 1781 return key; 1782 } 1783 1784 static void 1785 radv_fill_shader_keys(struct ac_shader_variant_key *keys, 1786 const struct radv_pipeline_key *key, 1787 nir_shader **nir) 1788 { 1789 keys[MESA_SHADER_VERTEX].vs.instance_rate_inputs = key->instance_rate_inputs; 1790 1791 if (nir[MESA_SHADER_TESS_CTRL]) { 1792 keys[MESA_SHADER_VERTEX].vs.as_ls = true; 1793 keys[MESA_SHADER_TESS_CTRL].tcs.input_vertices = key->tess_input_vertices; 1794 keys[MESA_SHADER_TESS_CTRL].tcs.primitive_mode = nir[MESA_SHADER_TESS_EVAL]->info.tess.primitive_mode; 1795 1796 keys[MESA_SHADER_TESS_CTRL].tcs.tes_reads_tess_factors = !!(nir[MESA_SHADER_TESS_EVAL]->info.inputs_read & (VARYING_BIT_TESS_LEVEL_INNER | VARYING_BIT_TESS_LEVEL_OUTER)); 1797 } 1798 1799 if (nir[MESA_SHADER_GEOMETRY]) { 1800 if (nir[MESA_SHADER_TESS_CTRL]) 1801 keys[MESA_SHADER_TESS_EVAL].tes.as_es = true; 1802 else 1803 keys[MESA_SHADER_VERTEX].vs.as_es = true; 1804 } 1805 1806 for(int i = 0; i < MESA_SHADER_STAGES; ++i) 1807 keys[i].has_multiview_view_index = key->has_multiview_view_index; 1808 1809 keys[MESA_SHADER_FRAGMENT].fs.multisample = key->multisample; 1810 keys[MESA_SHADER_FRAGMENT].fs.col_format = key->col_format; 1811 keys[MESA_SHADER_FRAGMENT].fs.is_int8 = key->is_int8; 1812 keys[MESA_SHADER_FRAGMENT].fs.is_int10 = key->is_int10; 1813 keys[MESA_SHADER_FRAGMENT].fs.log2_ps_iter_samples = key->log2_ps_iter_samples; 1814 keys[MESA_SHADER_FRAGMENT].fs.log2_num_samples = key->log2_num_samples; 1815 } 1816 1817 static void 1818 merge_tess_info(struct shader_info *tes_info, 1819 const struct shader_info *tcs_info) 1820 { 1821 /* The Vulkan 1.0.38 spec, section 21.1 Tessellator says: 1822 * 1823 * "PointMode. Controls generation of points rather than triangles 1824 * or lines. This functionality defaults to disabled, and is 1825 * enabled if either shader stage includes the execution mode. 1826 * 1827 * and about Triangles, Quads, IsoLines, VertexOrderCw, VertexOrderCcw, 1828 * PointMode, SpacingEqual, SpacingFractionalEven, SpacingFractionalOdd, 1829 * and OutputVertices, it says: 1830 * 1831 * "One mode must be set in at least one of the tessellation 1832 * shader stages." 1833 * 1834 * So, the fields can be set in either the TCS or TES, but they must 1835 * agree if set in both. Our backend looks at TES, so bitwise-or in 1836 * the values from the TCS. 1837 */ 1838 assert(tcs_info->tess.tcs_vertices_out == 0 || 1839 tes_info->tess.tcs_vertices_out == 0 || 1840 tcs_info->tess.tcs_vertices_out == tes_info->tess.tcs_vertices_out); 1841 tes_info->tess.tcs_vertices_out |= tcs_info->tess.tcs_vertices_out; 1842 1843 assert(tcs_info->tess.spacing == TESS_SPACING_UNSPECIFIED || 1844 tes_info->tess.spacing == TESS_SPACING_UNSPECIFIED || 1845 tcs_info->tess.spacing == tes_info->tess.spacing); 1846 tes_info->tess.spacing |= tcs_info->tess.spacing; 1847 1848 assert(tcs_info->tess.primitive_mode == 0 || 1849 tes_info->tess.primitive_mode == 0 || 1850 tcs_info->tess.primitive_mode == tes_info->tess.primitive_mode); 1851 tes_info->tess.primitive_mode |= tcs_info->tess.primitive_mode; 1852 tes_info->tess.ccw |= tcs_info->tess.ccw; 1853 tes_info->tess.point_mode |= tcs_info->tess.point_mode; 1854 } 1855 1856 static 1857 void radv_create_shaders(struct radv_pipeline *pipeline, 1858 struct radv_device *device, 1859 struct radv_pipeline_cache *cache, 1860 struct radv_pipeline_key key, 1861 const VkPipelineShaderStageCreateInfo **pStages) 1862 { 1863 struct radv_shader_module fs_m = {0}; 1864 struct radv_shader_module *modules[MESA_SHADER_STAGES] = { 0, }; 1865 nir_shader *nir[MESA_SHADER_STAGES] = {0}; 1866 void *codes[MESA_SHADER_STAGES] = {0}; 1867 unsigned code_sizes[MESA_SHADER_STAGES] = {0}; 1868 struct ac_shader_variant_key keys[MESA_SHADER_STAGES] = {{{{0}}}}; 1869 unsigned char hash[20], gs_copy_hash[20]; 1870 1871 for (unsigned i = 0; i < MESA_SHADER_STAGES; ++i) { 1872 if (pStages[i]) { 1873 modules[i] = radv_shader_module_from_handle(pStages[i]->module); 1874 if (modules[i]->nir) 1875 _mesa_sha1_compute(modules[i]->nir->info.name, 1876 strlen(modules[i]->nir->info.name), 1877 modules[i]->sha1); 1878 } 1879 } 1880 1881 radv_hash_shaders(hash, pStages, pipeline->layout, &key, get_hash_flags(device)); 1882 memcpy(gs_copy_hash, hash, 20); 1883 gs_copy_hash[0] ^= 1; 1884 1885 if (modules[MESA_SHADER_GEOMETRY]) { 1886 struct radv_shader_variant *variants[MESA_SHADER_STAGES] = {0}; 1887 radv_create_shader_variants_from_pipeline_cache(device, cache, gs_copy_hash, variants); 1888 pipeline->gs_copy_shader = variants[MESA_SHADER_GEOMETRY]; 1889 } 1890 1891 if (radv_create_shader_variants_from_pipeline_cache(device, cache, hash, pipeline->shaders) && 1892 (!modules[MESA_SHADER_GEOMETRY] || pipeline->gs_copy_shader)) { 1893 for (unsigned i = 0; i < MESA_SHADER_STAGES; ++i) { 1894 if (pipeline->shaders[i]) 1895 pipeline->active_stages |= mesa_to_vk_shader_stage(i); 1896 } 1897 return; 1898 } 1899 1900 if (!modules[MESA_SHADER_FRAGMENT] && !modules[MESA_SHADER_COMPUTE]) { 1901 nir_builder fs_b; 1902 nir_builder_init_simple_shader(&fs_b, NULL, MESA_SHADER_FRAGMENT, NULL); 1903 fs_b.shader->info.name = ralloc_strdup(fs_b.shader, "noop_fs"); 1904 fs_m.nir = fs_b.shader; 1905 modules[MESA_SHADER_FRAGMENT] = &fs_m; 1906 } 1907 1908 /* Determine first and last stage. */ 1909 unsigned first = MESA_SHADER_STAGES; 1910 unsigned last = 0; 1911 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) { 1912 if (!pStages[i]) 1913 continue; 1914 if (first == MESA_SHADER_STAGES) 1915 first = i; 1916 last = i; 1917 } 1918 1919 int prev = -1; 1920 for (unsigned i = 0; i < MESA_SHADER_STAGES; ++i) { 1921 const VkPipelineShaderStageCreateInfo *stage = pStages[i]; 1922 1923 if (!modules[i]) 1924 continue; 1925 1926 nir[i] = radv_shader_compile_to_nir(device, modules[i], 1927 stage ? stage->pName : "main", i, 1928 stage ? stage->pSpecializationInfo : NULL); 1929 pipeline->active_stages |= mesa_to_vk_shader_stage(i); 1930 1931 /* We don't want to alter meta shaders IR directly so clone it 1932 * first. 1933 */ 1934 if (nir[i]->info.name) { 1935 nir[i] = nir_shader_clone(NULL, nir[i]); 1936 } 1937 1938 if (first != last) { 1939 nir_variable_mode mask = 0; 1940 1941 if (i != first) 1942 mask = mask | nir_var_shader_in; 1943 1944 if (i != last) 1945 mask = mask | nir_var_shader_out; 1946 1947 nir_lower_io_to_scalar_early(nir[i], mask); 1948 radv_optimize_nir(nir[i]); 1949 } 1950 1951 if (prev != -1) { 1952 nir_compact_varyings(nir[prev], nir[i], true); 1953 } 1954 prev = i; 1955 } 1956 1957 if (nir[MESA_SHADER_TESS_CTRL]) { 1958 nir_lower_tes_patch_vertices(nir[MESA_SHADER_TESS_EVAL], nir[MESA_SHADER_TESS_CTRL]->info.tess.tcs_vertices_out); 1959 merge_tess_info(&nir[MESA_SHADER_TESS_EVAL]->info, &nir[MESA_SHADER_TESS_CTRL]->info); 1960 } 1961 1962 radv_link_shaders(pipeline, nir); 1963 1964 for (int i = 0; i < MESA_SHADER_STAGES; ++i) { 1965 if (modules[i] && radv_can_dump_shader(device, modules[i])) 1966 nir_print_shader(nir[i], stderr); 1967 } 1968 1969 radv_fill_shader_keys(keys, &key, nir); 1970 1971 if (nir[MESA_SHADER_FRAGMENT]) { 1972 if (!pipeline->shaders[MESA_SHADER_FRAGMENT]) { 1973 pipeline->shaders[MESA_SHADER_FRAGMENT] = 1974 radv_shader_variant_create(device, modules[MESA_SHADER_FRAGMENT], &nir[MESA_SHADER_FRAGMENT], 1, 1975 pipeline->layout, keys + MESA_SHADER_FRAGMENT, 1976 &codes[MESA_SHADER_FRAGMENT], &code_sizes[MESA_SHADER_FRAGMENT]); 1977 } 1978 1979 /* TODO: These are no longer used as keys we should refactor this */ 1980 keys[MESA_SHADER_VERTEX].vs.export_prim_id = 1981 pipeline->shaders[MESA_SHADER_FRAGMENT]->info.fs.prim_id_input; 1982 keys[MESA_SHADER_TESS_EVAL].tes.export_prim_id = 1983 pipeline->shaders[MESA_SHADER_FRAGMENT]->info.fs.prim_id_input; 1984 } 1985 1986 if (device->physical_device->rad_info.chip_class >= GFX9 && modules[MESA_SHADER_TESS_CTRL]) { 1987 if (!pipeline->shaders[MESA_SHADER_TESS_CTRL]) { 1988 struct nir_shader *combined_nir[] = {nir[MESA_SHADER_VERTEX], nir[MESA_SHADER_TESS_CTRL]}; 1989 struct ac_shader_variant_key key = keys[MESA_SHADER_TESS_CTRL]; 1990 key.tcs.vs_key = keys[MESA_SHADER_VERTEX].vs; 1991 pipeline->shaders[MESA_SHADER_TESS_CTRL] = radv_shader_variant_create(device, modules[MESA_SHADER_TESS_CTRL], combined_nir, 2, 1992 pipeline->layout, 1993 &key, &codes[MESA_SHADER_TESS_CTRL], 1994 &code_sizes[MESA_SHADER_TESS_CTRL]); 1995 } 1996 modules[MESA_SHADER_VERTEX] = NULL; 1997 } 1998 1999 if (device->physical_device->rad_info.chip_class >= GFX9 && modules[MESA_SHADER_GEOMETRY]) { 2000 gl_shader_stage pre_stage = modules[MESA_SHADER_TESS_EVAL] ? MESA_SHADER_TESS_EVAL : MESA_SHADER_VERTEX; 2001 if (!pipeline->shaders[MESA_SHADER_GEOMETRY]) { 2002 struct nir_shader *combined_nir[] = {nir[pre_stage], nir[MESA_SHADER_GEOMETRY]}; 2003 pipeline->shaders[MESA_SHADER_GEOMETRY] = radv_shader_variant_create(device, modules[MESA_SHADER_GEOMETRY], combined_nir, 2, 2004 pipeline->layout, 2005 &keys[pre_stage] , &codes[MESA_SHADER_GEOMETRY], 2006 &code_sizes[MESA_SHADER_GEOMETRY]); 2007 } 2008 modules[pre_stage] = NULL; 2009 } 2010 2011 for (int i = 0; i < MESA_SHADER_STAGES; ++i) { 2012 if(modules[i] && !pipeline->shaders[i]) { 2013 pipeline->shaders[i] = radv_shader_variant_create(device, modules[i], &nir[i], 1, 2014 pipeline->layout, 2015 keys + i, &codes[i], 2016 &code_sizes[i]); 2017 } 2018 } 2019 2020 if(modules[MESA_SHADER_GEOMETRY]) { 2021 void *gs_copy_code = NULL; 2022 unsigned gs_copy_code_size = 0; 2023 if (!pipeline->gs_copy_shader) { 2024 pipeline->gs_copy_shader = radv_create_gs_copy_shader( 2025 device, nir[MESA_SHADER_GEOMETRY], &gs_copy_code, 2026 &gs_copy_code_size, 2027 keys[MESA_SHADER_GEOMETRY].has_multiview_view_index); 2028 } 2029 2030 if (pipeline->gs_copy_shader) { 2031 void *code[MESA_SHADER_STAGES] = {0}; 2032 unsigned code_size[MESA_SHADER_STAGES] = {0}; 2033 struct radv_shader_variant *variants[MESA_SHADER_STAGES] = {0}; 2034 2035 code[MESA_SHADER_GEOMETRY] = gs_copy_code; 2036 code_size[MESA_SHADER_GEOMETRY] = gs_copy_code_size; 2037 variants[MESA_SHADER_GEOMETRY] = pipeline->gs_copy_shader; 2038 2039 radv_pipeline_cache_insert_shaders(device, cache, 2040 gs_copy_hash, 2041 variants, 2042 (const void**)code, 2043 code_size); 2044 } 2045 free(gs_copy_code); 2046 } 2047 2048 radv_pipeline_cache_insert_shaders(device, cache, hash, pipeline->shaders, 2049 (const void**)codes, code_sizes); 2050 2051 for (int i = 0; i < MESA_SHADER_STAGES; ++i) { 2052 free(codes[i]); 2053 if (modules[i] && !pipeline->device->keep_shader_info) 2054 ralloc_free(nir[i]); 2055 } 2056 2057 if (fs_m.nir) 2058 ralloc_free(fs_m.nir); 2059 } 2060 2061 static uint32_t 2062 radv_pipeline_stage_to_user_data_0(struct radv_pipeline *pipeline, 2063 gl_shader_stage stage, enum chip_class chip_class) 2064 { 2065 bool has_gs = radv_pipeline_has_gs(pipeline); 2066 bool has_tess = radv_pipeline_has_tess(pipeline); 2067 switch (stage) { 2068 case MESA_SHADER_FRAGMENT: 2069 return R_00B030_SPI_SHADER_USER_DATA_PS_0; 2070 case MESA_SHADER_VERTEX: 2071 if (chip_class >= GFX9) { 2072 return has_tess ? R_00B430_SPI_SHADER_USER_DATA_LS_0 : 2073 has_gs ? R_00B330_SPI_SHADER_USER_DATA_ES_0 : 2074 R_00B130_SPI_SHADER_USER_DATA_VS_0; 2075 } 2076 if (has_tess) 2077 return R_00B530_SPI_SHADER_USER_DATA_LS_0; 2078 else 2079 return has_gs ? R_00B330_SPI_SHADER_USER_DATA_ES_0 : R_00B130_SPI_SHADER_USER_DATA_VS_0; 2080 case MESA_SHADER_GEOMETRY: 2081 return chip_class >= GFX9 ? R_00B330_SPI_SHADER_USER_DATA_ES_0 : 2082 R_00B230_SPI_SHADER_USER_DATA_GS_0; 2083 case MESA_SHADER_COMPUTE: 2084 return R_00B900_COMPUTE_USER_DATA_0; 2085 case MESA_SHADER_TESS_CTRL: 2086 return chip_class >= GFX9 ? R_00B430_SPI_SHADER_USER_DATA_LS_0 : 2087 R_00B430_SPI_SHADER_USER_DATA_HS_0; 2088 case MESA_SHADER_TESS_EVAL: 2089 if (chip_class >= GFX9) { 2090 return has_gs ? R_00B330_SPI_SHADER_USER_DATA_ES_0 : 2091 R_00B130_SPI_SHADER_USER_DATA_VS_0; 2092 } 2093 if (has_gs) 2094 return R_00B330_SPI_SHADER_USER_DATA_ES_0; 2095 else 2096 return R_00B130_SPI_SHADER_USER_DATA_VS_0; 2097 default: 2098 unreachable("unknown shader"); 2099 } 2100 } 2101 2102 struct radv_bin_size_entry { 2103 unsigned bpp; 2104 VkExtent2D extent; 2105 }; 2106 2107 static VkExtent2D 2108 radv_compute_bin_size(struct radv_pipeline *pipeline, const VkGraphicsPipelineCreateInfo *pCreateInfo) 2109 { 2110 static const struct radv_bin_size_entry color_size_table[][3][9] = { 2111 { 2112 /* One RB / SE */ 2113 { 2114 /* One shader engine */ 2115 { 0, {128, 128}}, 2116 { 1, { 64, 128}}, 2117 { 2, { 32, 128}}, 2118 { 3, { 16, 128}}, 2119 { 17, { 0, 0}}, 2120 { UINT_MAX, { 0, 0}}, 2121 }, 2122 { 2123 /* Two shader engines */ 2124 { 0, {128, 128}}, 2125 { 2, { 64, 128}}, 2126 { 3, { 32, 128}}, 2127 { 5, { 16, 128}}, 2128 { 17, { 0, 0}}, 2129 { UINT_MAX, { 0, 0}}, 2130 }, 2131 { 2132 /* Four shader engines */ 2133 { 0, {128, 128}}, 2134 { 3, { 64, 128}}, 2135 { 5, { 16, 128}}, 2136 { 17, { 0, 0}}, 2137 { UINT_MAX, { 0, 0}}, 2138 }, 2139 }, 2140 { 2141 /* Two RB / SE */ 2142 { 2143 /* One shader engine */ 2144 { 0, {128, 128}}, 2145 { 2, { 64, 128}}, 2146 { 3, { 32, 128}}, 2147 { 5, { 16, 128}}, 2148 { 33, { 0, 0}}, 2149 { UINT_MAX, { 0, 0}}, 2150 }, 2151 { 2152 /* Two shader engines */ 2153 { 0, {128, 128}}, 2154 { 3, { 64, 128}}, 2155 { 5, { 32, 128}}, 2156 { 9, { 16, 128}}, 2157 { 33, { 0, 0}}, 2158 { UINT_MAX, { 0, 0}}, 2159 }, 2160 { 2161 /* Four shader engines */ 2162 { 0, {256, 256}}, 2163 { 2, {128, 256}}, 2164 { 3, {128, 128}}, 2165 { 5, { 64, 128}}, 2166 { 9, { 16, 128}}, 2167 { 33, { 0, 0}}, 2168 { UINT_MAX, { 0, 0}}, 2169 }, 2170 }, 2171 { 2172 /* Four RB / SE */ 2173 { 2174 /* One shader engine */ 2175 { 0, {128, 256}}, 2176 { 2, {128, 128}}, 2177 { 3, { 64, 128}}, 2178 { 5, { 32, 128}}, 2179 { 9, { 16, 128}}, 2180 { 33, { 0, 0}}, 2181 { UINT_MAX, { 0, 0}}, 2182 }, 2183 { 2184 /* Two shader engines */ 2185 { 0, {256, 256}}, 2186 { 2, {128, 256}}, 2187 { 3, {128, 128}}, 2188 { 5, { 64, 128}}, 2189 { 9, { 32, 128}}, 2190 { 17, { 16, 128}}, 2191 { 33, { 0, 0}}, 2192 { UINT_MAX, { 0, 0}}, 2193 }, 2194 { 2195 /* Four shader engines */ 2196 { 0, {256, 512}}, 2197 { 2, {256, 256}}, 2198 { 3, {128, 256}}, 2199 { 5, {128, 128}}, 2200 { 9, { 64, 128}}, 2201 { 17, { 16, 128}}, 2202 { 33, { 0, 0}}, 2203 { UINT_MAX, { 0, 0}}, 2204 }, 2205 }, 2206 }; 2207 static const struct radv_bin_size_entry ds_size_table[][3][9] = { 2208 { 2209 // One RB / SE 2210 { 2211 // One shader engine 2212 { 0, {128, 256}}, 2213 { 2, {128, 128}}, 2214 { 4, { 64, 128}}, 2215 { 7, { 32, 128}}, 2216 { 13, { 16, 128}}, 2217 { 49, { 0, 0}}, 2218 { UINT_MAX, { 0, 0}}, 2219 }, 2220 { 2221 // Two shader engines 2222 { 0, {256, 256}}, 2223 { 2, {128, 256}}, 2224 { 4, {128, 128}}, 2225 { 7, { 64, 128}}, 2226 { 13, { 32, 128}}, 2227 { 25, { 16, 128}}, 2228 { 49, { 0, 0}}, 2229 { UINT_MAX, { 0, 0}}, 2230 }, 2231 { 2232 // Four shader engines 2233 { 0, {256, 512}}, 2234 { 2, {256, 256}}, 2235 { 4, {128, 256}}, 2236 { 7, {128, 128}}, 2237 { 13, { 64, 128}}, 2238 { 25, { 16, 128}}, 2239 { 49, { 0, 0}}, 2240 { UINT_MAX, { 0, 0}}, 2241 }, 2242 }, 2243 { 2244 // Two RB / SE 2245 { 2246 // One shader engine 2247 { 0, {256, 256}}, 2248 { 2, {128, 256}}, 2249 { 4, {128, 128}}, 2250 { 7, { 64, 128}}, 2251 { 13, { 32, 128}}, 2252 { 25, { 16, 128}}, 2253 { 97, { 0, 0}}, 2254 { UINT_MAX, { 0, 0}}, 2255 }, 2256 { 2257 // Two shader engines 2258 { 0, {256, 512}}, 2259 { 2, {256, 256}}, 2260 { 4, {128, 256}}, 2261 { 7, {128, 128}}, 2262 { 13, { 64, 128}}, 2263 { 25, { 32, 128}}, 2264 { 49, { 16, 128}}, 2265 { 97, { 0, 0}}, 2266 { UINT_MAX, { 0, 0}}, 2267 }, 2268 { 2269 // Four shader engines 2270 { 0, {512, 512}}, 2271 { 2, {256, 512}}, 2272 { 4, {256, 256}}, 2273 { 7, {128, 256}}, 2274 { 13, {128, 128}}, 2275 { 25, { 64, 128}}, 2276 { 49, { 16, 128}}, 2277 { 97, { 0, 0}}, 2278 { UINT_MAX, { 0, 0}}, 2279 }, 2280 }, 2281 { 2282 // Four RB / SE 2283 { 2284 // One shader engine 2285 { 0, {256, 512}}, 2286 { 2, {256, 256}}, 2287 { 4, {128, 256}}, 2288 { 7, {128, 128}}, 2289 { 13, { 64, 128}}, 2290 { 25, { 32, 128}}, 2291 { 49, { 16, 128}}, 2292 { UINT_MAX, { 0, 0}}, 2293 }, 2294 { 2295 // Two shader engines 2296 { 0, {512, 512}}, 2297 { 2, {256, 512}}, 2298 { 4, {256, 256}}, 2299 { 7, {128, 256}}, 2300 { 13, {128, 128}}, 2301 { 25, { 64, 128}}, 2302 { 49, { 32, 128}}, 2303 { 97, { 16, 128}}, 2304 { UINT_MAX, { 0, 0}}, 2305 }, 2306 { 2307 // Four shader engines 2308 { 0, {512, 512}}, 2309 { 4, {256, 512}}, 2310 { 7, {256, 256}}, 2311 { 13, {128, 256}}, 2312 { 25, {128, 128}}, 2313 { 49, { 64, 128}}, 2314 { 97, { 16, 128}}, 2315 { UINT_MAX, { 0, 0}}, 2316 }, 2317 }, 2318 }; 2319 2320 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass); 2321 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass; 2322 VkExtent2D extent = {512, 512}; 2323 2324 unsigned log_num_rb_per_se = 2325 util_logbase2_ceil(pipeline->device->physical_device->rad_info.num_render_backends / 2326 pipeline->device->physical_device->rad_info.max_se); 2327 unsigned log_num_se = util_logbase2_ceil(pipeline->device->physical_device->rad_info.max_se); 2328 2329 unsigned total_samples = 1u << G_028BE0_MSAA_NUM_SAMPLES(pipeline->graphics.ms.pa_sc_mode_cntl_1); 2330 unsigned ps_iter_samples = 1u << G_028804_PS_ITER_SAMPLES(pipeline->graphics.ms.db_eqaa); 2331 unsigned effective_samples = total_samples; 2332 unsigned cb_target_mask = pipeline->graphics.blend.cb_target_mask; 2333 unsigned color_bytes_per_pixel = 0; 2334 2335 for (unsigned i = 0; i < subpass->color_count; i++) { 2336 if (!(cb_target_mask & (0xf << (i * 4)))) 2337 continue; 2338 2339 if (subpass->color_attachments[i].attachment == VK_ATTACHMENT_UNUSED) 2340 continue; 2341 2342 VkFormat format = pass->attachments[subpass->color_attachments[i].attachment].format; 2343 color_bytes_per_pixel += vk_format_get_blocksize(format); 2344 } 2345 2346 /* MSAA images typically don't use all samples all the time. */ 2347 if (effective_samples >= 2 && ps_iter_samples <= 1) 2348 effective_samples = 2; 2349 color_bytes_per_pixel *= effective_samples; 2350 2351 const struct radv_bin_size_entry *color_entry = color_size_table[log_num_rb_per_se][log_num_se]; 2352 while(color_entry->bpp <= color_bytes_per_pixel) 2353 ++color_entry; 2354 2355 extent = color_entry->extent; 2356 2357 if (subpass->depth_stencil_attachment.attachment != VK_ATTACHMENT_UNUSED) { 2358 struct radv_render_pass_attachment *attachment = pass->attachments + subpass->depth_stencil_attachment.attachment; 2359 2360 /* Coefficients taken from AMDVLK */ 2361 unsigned depth_coeff = vk_format_is_depth(attachment->format) ? 5 : 0; 2362 unsigned stencil_coeff = vk_format_is_stencil(attachment->format) ? 1 : 0; 2363 unsigned ds_bytes_per_pixel = 4 * (depth_coeff + stencil_coeff) * total_samples; 2364 2365 const struct radv_bin_size_entry *ds_entry = ds_size_table[log_num_rb_per_se][log_num_se]; 2366 while(ds_entry->bpp <= ds_bytes_per_pixel) 2367 ++ds_entry; 2368 2369 extent.width = MIN2(extent.width, ds_entry->extent.width); 2370 extent.height = MIN2(extent.height, ds_entry->extent.height); 2371 } 2372 2373 return extent; 2374 } 2375 2376 static void 2377 radv_compute_binning_state(struct radv_pipeline *pipeline, const VkGraphicsPipelineCreateInfo *pCreateInfo) 2378 { 2379 pipeline->graphics.bin.pa_sc_binner_cntl_0 = 2380 S_028C44_BINNING_MODE(V_028C44_DISABLE_BINNING_USE_LEGACY_SC) | 2381 S_028C44_DISABLE_START_OF_PRIM(1); 2382 pipeline->graphics.bin.db_dfsm_control = S_028060_PUNCHOUT_MODE(V_028060_FORCE_OFF); 2383 2384 if (!pipeline->device->pbb_allowed) 2385 return; 2386 2387 VkExtent2D bin_size = radv_compute_bin_size(pipeline, pCreateInfo); 2388 if (!bin_size.width || !bin_size.height) 2389 return; 2390 2391 unsigned context_states_per_bin; /* allowed range: [1, 6] */ 2392 unsigned persistent_states_per_bin; /* allowed range: [1, 32] */ 2393 unsigned fpovs_per_batch; /* allowed range: [0, 255], 0 = unlimited */ 2394 2395 switch (pipeline->device->physical_device->rad_info.family) { 2396 case CHIP_VEGA10: 2397 context_states_per_bin = 1; 2398 persistent_states_per_bin = 1; 2399 fpovs_per_batch = 63; 2400 break; 2401 case CHIP_RAVEN: 2402 context_states_per_bin = 6; 2403 persistent_states_per_bin = 32; 2404 fpovs_per_batch = 63; 2405 break; 2406 default: 2407 unreachable("unhandled family while determining binning state."); 2408 } 2409 2410 pipeline->graphics.bin.pa_sc_binner_cntl_0 = 2411 S_028C44_BINNING_MODE(V_028C44_BINNING_ALLOWED) | 2412 S_028C44_BIN_SIZE_X(bin_size.width == 16) | 2413 S_028C44_BIN_SIZE_Y(bin_size.height == 16) | 2414 S_028C44_BIN_SIZE_X_EXTEND(util_logbase2(MAX2(bin_size.width, 32)) - 5) | 2415 S_028C44_BIN_SIZE_Y_EXTEND(util_logbase2(MAX2(bin_size.height, 32)) - 5) | 2416 S_028C44_CONTEXT_STATES_PER_BIN(context_states_per_bin - 1) | 2417 S_028C44_PERSISTENT_STATES_PER_BIN(persistent_states_per_bin - 1) | 2418 S_028C44_DISABLE_START_OF_PRIM(1) | 2419 S_028C44_FPOVS_PER_BATCH(fpovs_per_batch) | 2420 S_028C44_OPTIMAL_BIN_SELECTION(1); 2421 2422 /* DFSM is not implemented yet */ 2423 assert(!pipeline->device->dfsm_allowed); 2424 } 2425 2426 static VkResult 2427 radv_pipeline_init(struct radv_pipeline *pipeline, 2428 struct radv_device *device, 2429 struct radv_pipeline_cache *cache, 2430 const VkGraphicsPipelineCreateInfo *pCreateInfo, 2431 const struct radv_graphics_pipeline_create_info *extra, 2432 const VkAllocationCallbacks *alloc) 2433 { 2434 VkResult result; 2435 bool has_view_index = false; 2436 2437 RADV_FROM_HANDLE(radv_render_pass, pass, pCreateInfo->renderPass); 2438 struct radv_subpass *subpass = pass->subpasses + pCreateInfo->subpass; 2439 if (subpass->view_mask) 2440 has_view_index = true; 2441 if (alloc == NULL) 2442 alloc = &device->alloc; 2443 2444 pipeline->device = device; 2445 pipeline->layout = radv_pipeline_layout_from_handle(pCreateInfo->layout); 2446 assert(pipeline->layout); 2447 2448 radv_pipeline_init_dynamic_state(pipeline, pCreateInfo); 2449 radv_pipeline_init_blend_state(pipeline, pCreateInfo, extra); 2450 2451 const VkPipelineShaderStageCreateInfo *pStages[MESA_SHADER_STAGES] = { 0, }; 2452 for (uint32_t i = 0; i < pCreateInfo->stageCount; i++) { 2453 gl_shader_stage stage = ffs(pCreateInfo->pStages[i].stage) - 1; 2454 pStages[stage] = &pCreateInfo->pStages[i]; 2455 } 2456 2457 radv_create_shaders(pipeline, device, cache, 2458 radv_generate_graphics_pipeline_key(pipeline, pCreateInfo, has_view_index), 2459 pStages); 2460 2461 pipeline->graphics.spi_baryc_cntl = S_0286E0_FRONT_FACE_ALL_BITS(1); 2462 radv_pipeline_init_depth_stencil_state(pipeline, pCreateInfo, extra); 2463 radv_pipeline_init_raster_state(pipeline, pCreateInfo); 2464 radv_pipeline_init_multisample_state(pipeline, pCreateInfo); 2465 pipeline->graphics.prim = si_translate_prim(pCreateInfo->pInputAssemblyState->topology); 2466 pipeline->graphics.can_use_guardband = radv_prim_can_use_guardband(pCreateInfo->pInputAssemblyState->topology); 2467 2468 if (radv_pipeline_has_gs(pipeline)) { 2469 pipeline->graphics.gs_out = si_conv_gl_prim_to_gs_out(pipeline->shaders[MESA_SHADER_GEOMETRY]->info.gs.output_prim); 2470 pipeline->graphics.can_use_guardband = pipeline->graphics.gs_out == V_028A6C_OUTPRIM_TYPE_TRISTRIP; 2471 } else { 2472 pipeline->graphics.gs_out = si_conv_prim_to_gs_out(pCreateInfo->pInputAssemblyState->topology); 2473 } 2474 if (extra && extra->use_rectlist) { 2475 pipeline->graphics.prim = V_008958_DI_PT_RECTLIST; 2476 pipeline->graphics.gs_out = V_028A6C_OUTPRIM_TYPE_TRISTRIP; 2477 pipeline->graphics.can_use_guardband = true; 2478 } 2479 pipeline->graphics.prim_restart_enable = !!pCreateInfo->pInputAssemblyState->primitiveRestartEnable; 2480 /* prim vertex count will need TESS changes */ 2481 pipeline->graphics.prim_vertex_count = prim_size_table[pipeline->graphics.prim]; 2482 2483 /* Ensure that some export memory is always allocated, for two reasons: 2484 * 2485 * 1) Correctness: The hardware ignores the EXEC mask if no export 2486 * memory is allocated, so KILL and alpha test do not work correctly 2487 * without this. 2488 * 2) Performance: Every shader needs at least a NULL export, even when 2489 * it writes no color/depth output. The NULL export instruction 2490 * stalls without this setting. 2491 * 2492 * Don't add this to CB_SHADER_MASK. 2493 */ 2494 struct radv_shader_variant *ps = pipeline->shaders[MESA_SHADER_FRAGMENT]; 2495 if (!pipeline->graphics.blend.spi_shader_col_format) { 2496 if (!ps->info.fs.writes_z && 2497 !ps->info.fs.writes_stencil && 2498 !ps->info.fs.writes_sample_mask) 2499 pipeline->graphics.blend.spi_shader_col_format = V_028714_SPI_SHADER_32_R; 2500 } 2501 2502 unsigned z_order; 2503 pipeline->graphics.db_shader_control = 0; 2504 if (ps->info.fs.early_fragment_test || !ps->info.fs.writes_memory) 2505 z_order = V_02880C_EARLY_Z_THEN_LATE_Z; 2506 else 2507 z_order = V_02880C_LATE_Z; 2508 2509 pipeline->graphics.db_shader_control = 2510 S_02880C_Z_EXPORT_ENABLE(ps->info.fs.writes_z) | 2511 S_02880C_STENCIL_TEST_VAL_EXPORT_ENABLE(ps->info.fs.writes_stencil) | 2512 S_02880C_KILL_ENABLE(!!ps->info.fs.can_discard) | 2513 S_02880C_MASK_EXPORT_ENABLE(ps->info.fs.writes_sample_mask) | 2514 S_02880C_Z_ORDER(z_order) | 2515 S_02880C_DEPTH_BEFORE_SHADER(ps->info.fs.early_fragment_test) | 2516 S_02880C_EXEC_ON_HIER_FAIL(ps->info.fs.writes_memory) | 2517 S_02880C_EXEC_ON_NOOP(ps->info.fs.writes_memory); 2518 2519 if (pipeline->device->physical_device->has_rbplus) 2520 pipeline->graphics.db_shader_control |= S_02880C_DUAL_QUAD_DISABLE(1); 2521 2522 unsigned shader_z_format = 2523 ac_get_spi_shader_z_format(ps->info.fs.writes_z, 2524 ps->info.fs.writes_stencil, 2525 ps->info.fs.writes_sample_mask); 2526 pipeline->graphics.shader_z_format = shader_z_format; 2527 2528 calculate_vgt_gs_mode(pipeline); 2529 calculate_vs_outinfo(pipeline); 2530 calculate_ps_inputs(pipeline); 2531 2532 for (unsigned i = 0; i < MESA_SHADER_STAGES; i++) { 2533 if (pipeline->shaders[i]) { 2534 pipeline->need_indirect_descriptor_sets |= pipeline->shaders[i]->info.need_indirect_descriptor_sets; 2535 } 2536 } 2537 2538 uint32_t stages = 0; 2539 if (radv_pipeline_has_tess(pipeline)) { 2540 stages |= S_028B54_LS_EN(V_028B54_LS_STAGE_ON) | 2541 S_028B54_HS_EN(1) | S_028B54_DYNAMIC_HS(1); 2542 2543 if (radv_pipeline_has_gs(pipeline)) 2544 stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_DS) | 2545 S_028B54_GS_EN(1) | 2546 S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER); 2547 else 2548 stages |= S_028B54_VS_EN(V_028B54_VS_STAGE_DS); 2549 2550 } else if (radv_pipeline_has_gs(pipeline)) 2551 stages |= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL) | 2552 S_028B54_GS_EN(1) | 2553 S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER); 2554 2555 if (device->physical_device->rad_info.chip_class >= GFX9) 2556 stages |= S_028B54_MAX_PRIMGRP_IN_WAVE(2); 2557 2558 pipeline->graphics.vgt_shader_stages_en = stages; 2559 2560 if (radv_pipeline_has_gs(pipeline)) { 2561 calculate_gs_ring_sizes(pipeline); 2562 if (device->physical_device->rad_info.chip_class >= GFX9) 2563 calculate_gfx9_gs_info(pCreateInfo, pipeline); 2564 } 2565 2566 if (radv_pipeline_has_tess(pipeline)) { 2567 if (pipeline->graphics.prim == V_008958_DI_PT_PATCH) { 2568 pipeline->graphics.prim_vertex_count.min = pCreateInfo->pTessellationState->patchControlPoints; 2569 pipeline->graphics.prim_vertex_count.incr = 1; 2570 } 2571 calculate_tess_state(pipeline, pCreateInfo); 2572 } 2573 2574 if (radv_pipeline_has_tess(pipeline)) 2575 pipeline->graphics.primgroup_size = pipeline->graphics.tess.num_patches; 2576 else if (radv_pipeline_has_gs(pipeline)) 2577 pipeline->graphics.primgroup_size = 64; 2578 else 2579 pipeline->graphics.primgroup_size = 128; /* recommended without a GS */ 2580 2581 pipeline->graphics.partial_es_wave = false; 2582 if (pipeline->device->has_distributed_tess) { 2583 if (radv_pipeline_has_gs(pipeline)) { 2584 if (device->physical_device->rad_info.chip_class <= VI) 2585 pipeline->graphics.partial_es_wave = true; 2586 } 2587 } 2588 /* GS requirement. */ 2589 if (SI_GS_PER_ES / pipeline->graphics.primgroup_size >= pipeline->device->gs_table_depth - 3) 2590 pipeline->graphics.partial_es_wave = true; 2591 2592 pipeline->graphics.wd_switch_on_eop = false; 2593 if (device->physical_device->rad_info.chip_class >= CIK) { 2594 unsigned prim = pipeline->graphics.prim; 2595 /* WD_SWITCH_ON_EOP has no effect on GPUs with less than 2596 * 4 shader engines. Set 1 to pass the assertion below. 2597 * The other cases are hardware requirements. */ 2598 if (device->physical_device->rad_info.max_se < 4 || 2599 prim == V_008958_DI_PT_POLYGON || 2600 prim == V_008958_DI_PT_LINELOOP || 2601 prim == V_008958_DI_PT_TRIFAN || 2602 prim == V_008958_DI_PT_TRISTRIP_ADJ || 2603 (pipeline->graphics.prim_restart_enable && 2604 (device->physical_device->rad_info.family < CHIP_POLARIS10 || 2605 (prim != V_008958_DI_PT_POINTLIST && 2606 prim != V_008958_DI_PT_LINESTRIP && 2607 prim != V_008958_DI_PT_TRISTRIP)))) 2608 pipeline->graphics.wd_switch_on_eop = true; 2609 } 2610 2611 pipeline->graphics.ia_switch_on_eoi = false; 2612 if (pipeline->shaders[MESA_SHADER_FRAGMENT]->info.fs.prim_id_input) 2613 pipeline->graphics.ia_switch_on_eoi = true; 2614 if (radv_pipeline_has_gs(pipeline) && 2615 pipeline->shaders[MESA_SHADER_GEOMETRY]->info.info.uses_prim_id) 2616 pipeline->graphics.ia_switch_on_eoi = true; 2617 if (radv_pipeline_has_tess(pipeline)) { 2618 /* SWITCH_ON_EOI must be set if PrimID is used. */ 2619 if (pipeline->shaders[MESA_SHADER_TESS_CTRL]->info.info.uses_prim_id || 2620 radv_get_tess_eval_shader(pipeline)->info.info.uses_prim_id) 2621 pipeline->graphics.ia_switch_on_eoi = true; 2622 } 2623 2624 pipeline->graphics.partial_vs_wave = false; 2625 if (radv_pipeline_has_tess(pipeline)) { 2626 /* Bug with tessellation and GS on Bonaire and older 2 SE chips. */ 2627 if ((device->physical_device->rad_info.family == CHIP_TAHITI || 2628 device->physical_device->rad_info.family == CHIP_PITCAIRN || 2629 device->physical_device->rad_info.family == CHIP_BONAIRE) && 2630 radv_pipeline_has_gs(pipeline)) 2631 pipeline->graphics.partial_vs_wave = true; 2632 /* Needed for 028B6C_DISTRIBUTION_MODE != 0 */ 2633 if (device->has_distributed_tess) { 2634 if (radv_pipeline_has_gs(pipeline)) { 2635 if (device->physical_device->rad_info.family == CHIP_TONGA || 2636 device->physical_device->rad_info.family == CHIP_FIJI || 2637 device->physical_device->rad_info.family == CHIP_POLARIS10 || 2638 device->physical_device->rad_info.family == CHIP_POLARIS11 || 2639 device->physical_device->rad_info.family == CHIP_POLARIS12) 2640 pipeline->graphics.partial_vs_wave = true; 2641 } else { 2642 pipeline->graphics.partial_vs_wave = true; 2643 } 2644 } 2645 } 2646 2647 pipeline->graphics.base_ia_multi_vgt_param = 2648 S_028AA8_PRIMGROUP_SIZE(pipeline->graphics.primgroup_size - 1) | 2649 /* The following field was moved to VGT_SHADER_STAGES_EN in GFX9. */ 2650 S_028AA8_MAX_PRIMGRP_IN_WAVE(device->physical_device->rad_info.chip_class == VI ? 2 : 0) | 2651 S_030960_EN_INST_OPT_BASIC(device->physical_device->rad_info.chip_class >= GFX9) | 2652 S_030960_EN_INST_OPT_ADV(device->physical_device->rad_info.chip_class >= GFX9); 2653 2654 const VkPipelineVertexInputStateCreateInfo *vi_info = 2655 pCreateInfo->pVertexInputState; 2656 struct radv_vertex_elements_info *velems = &pipeline->vertex_elements; 2657 2658 for (uint32_t i = 0; i < vi_info->vertexAttributeDescriptionCount; i++) { 2659 const VkVertexInputAttributeDescription *desc = 2660 &vi_info->pVertexAttributeDescriptions[i]; 2661 unsigned loc = desc->location; 2662 const struct vk_format_description *format_desc; 2663 int first_non_void; 2664 uint32_t num_format, data_format; 2665 format_desc = vk_format_description(desc->format); 2666 first_non_void = vk_format_get_first_non_void_channel(desc->format); 2667 2668 num_format = radv_translate_buffer_numformat(format_desc, first_non_void); 2669 data_format = radv_translate_buffer_dataformat(format_desc, first_non_void); 2670 2671 velems->rsrc_word3[loc] = S_008F0C_DST_SEL_X(si_map_swizzle(format_desc->swizzle[0])) | 2672 S_008F0C_DST_SEL_Y(si_map_swizzle(format_desc->swizzle[1])) | 2673 S_008F0C_DST_SEL_Z(si_map_swizzle(format_desc->swizzle[2])) | 2674 S_008F0C_DST_SEL_W(si_map_swizzle(format_desc->swizzle[3])) | 2675 S_008F0C_NUM_FORMAT(num_format) | 2676 S_008F0C_DATA_FORMAT(data_format); 2677 velems->format_size[loc] = format_desc->block.bits / 8; 2678 velems->offset[loc] = desc->offset; 2679 velems->binding[loc] = desc->binding; 2680 velems->count = MAX2(velems->count, loc + 1); 2681 } 2682 2683 for (uint32_t i = 0; i < vi_info->vertexBindingDescriptionCount; i++) { 2684 const VkVertexInputBindingDescription *desc = 2685 &vi_info->pVertexBindingDescriptions[i]; 2686 2687 pipeline->binding_stride[desc->binding] = desc->stride; 2688 } 2689 2690 for (uint32_t i = 0; i < MESA_SHADER_STAGES; i++) 2691 pipeline->user_data_0[i] = radv_pipeline_stage_to_user_data_0(pipeline, i, device->physical_device->rad_info.chip_class); 2692 2693 struct ac_userdata_info *loc = radv_lookup_user_sgpr(pipeline, MESA_SHADER_VERTEX, 2694 AC_UD_VS_BASE_VERTEX_START_INSTANCE); 2695 if (loc->sgpr_idx != -1) { 2696 pipeline->graphics.vtx_base_sgpr = pipeline->user_data_0[MESA_SHADER_VERTEX]; 2697 pipeline->graphics.vtx_base_sgpr += loc->sgpr_idx * 4; 2698 if (radv_get_vertex_shader(pipeline)->info.info.vs.needs_draw_id) 2699 pipeline->graphics.vtx_emit_num = 3; 2700 else 2701 pipeline->graphics.vtx_emit_num = 2; 2702 } 2703 2704 pipeline->graphics.vtx_reuse_depth = 30; 2705 if (radv_pipeline_has_tess(pipeline) && 2706 radv_get_tess_eval_shader(pipeline)->info.tes.spacing == TESS_SPACING_FRACTIONAL_ODD) { 2707 pipeline->graphics.vtx_reuse_depth = 14; 2708 } 2709 2710 if (device->instance->debug_flags & RADV_DEBUG_DUMP_SHADER_STATS) { 2711 radv_dump_pipeline_stats(device, pipeline); 2712 } 2713 2714 radv_compute_binning_state(pipeline, pCreateInfo); 2715 2716 result = radv_pipeline_scratch_init(device, pipeline); 2717 return result; 2718 } 2719 2720 VkResult 2721 radv_graphics_pipeline_create( 2722 VkDevice _device, 2723 VkPipelineCache _cache, 2724 const VkGraphicsPipelineCreateInfo *pCreateInfo, 2725 const struct radv_graphics_pipeline_create_info *extra, 2726 const VkAllocationCallbacks *pAllocator, 2727 VkPipeline *pPipeline) 2728 { 2729 RADV_FROM_HANDLE(radv_device, device, _device); 2730 RADV_FROM_HANDLE(radv_pipeline_cache, cache, _cache); 2731 struct radv_pipeline *pipeline; 2732 VkResult result; 2733 2734 pipeline = vk_zalloc2(&device->alloc, pAllocator, sizeof(*pipeline), 8, 2735 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); 2736 if (pipeline == NULL) 2737 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); 2738 2739 result = radv_pipeline_init(pipeline, device, cache, 2740 pCreateInfo, extra, pAllocator); 2741 if (result != VK_SUCCESS) { 2742 radv_pipeline_destroy(device, pipeline, pAllocator); 2743 return result; 2744 } 2745 2746 *pPipeline = radv_pipeline_to_handle(pipeline); 2747 2748 return VK_SUCCESS; 2749 } 2750 2751 VkResult radv_CreateGraphicsPipelines( 2752 VkDevice _device, 2753 VkPipelineCache pipelineCache, 2754 uint32_t count, 2755 const VkGraphicsPipelineCreateInfo* pCreateInfos, 2756 const VkAllocationCallbacks* pAllocator, 2757 VkPipeline* pPipelines) 2758 { 2759 VkResult result = VK_SUCCESS; 2760 unsigned i = 0; 2761 2762 for (; i < count; i++) { 2763 VkResult r; 2764 r = radv_graphics_pipeline_create(_device, 2765 pipelineCache, 2766 &pCreateInfos[i], 2767 NULL, pAllocator, &pPipelines[i]); 2768 if (r != VK_SUCCESS) { 2769 result = r; 2770 pPipelines[i] = VK_NULL_HANDLE; 2771 } 2772 } 2773 2774 return result; 2775 } 2776 2777 static VkResult radv_compute_pipeline_create( 2778 VkDevice _device, 2779 VkPipelineCache _cache, 2780 const VkComputePipelineCreateInfo* pCreateInfo, 2781 const VkAllocationCallbacks* pAllocator, 2782 VkPipeline* pPipeline) 2783 { 2784 RADV_FROM_HANDLE(radv_device, device, _device); 2785 RADV_FROM_HANDLE(radv_pipeline_cache, cache, _cache); 2786 const VkPipelineShaderStageCreateInfo *pStages[MESA_SHADER_STAGES] = { 0, }; 2787 struct radv_pipeline *pipeline; 2788 VkResult result; 2789 2790 pipeline = vk_zalloc2(&device->alloc, pAllocator, sizeof(*pipeline), 8, 2791 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT); 2792 if (pipeline == NULL) 2793 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY); 2794 2795 pipeline->device = device; 2796 pipeline->layout = radv_pipeline_layout_from_handle(pCreateInfo->layout); 2797 assert(pipeline->layout); 2798 2799 pStages[MESA_SHADER_COMPUTE] = &pCreateInfo->stage; 2800 radv_create_shaders(pipeline, device, cache, (struct radv_pipeline_key) {0}, pStages); 2801 2802 pipeline->user_data_0[MESA_SHADER_COMPUTE] = radv_pipeline_stage_to_user_data_0(pipeline, MESA_SHADER_COMPUTE, device->physical_device->rad_info.chip_class); 2803 pipeline->need_indirect_descriptor_sets |= pipeline->shaders[MESA_SHADER_COMPUTE]->info.need_indirect_descriptor_sets; 2804 result = radv_pipeline_scratch_init(device, pipeline); 2805 if (result != VK_SUCCESS) { 2806 radv_pipeline_destroy(device, pipeline, pAllocator); 2807 return result; 2808 } 2809 2810 *pPipeline = radv_pipeline_to_handle(pipeline); 2811 2812 if (device->instance->debug_flags & RADV_DEBUG_DUMP_SHADER_STATS) { 2813 radv_dump_pipeline_stats(device, pipeline); 2814 } 2815 return VK_SUCCESS; 2816 } 2817 VkResult radv_CreateComputePipelines( 2818 VkDevice _device, 2819 VkPipelineCache pipelineCache, 2820 uint32_t count, 2821 const VkComputePipelineCreateInfo* pCreateInfos, 2822 const VkAllocationCallbacks* pAllocator, 2823 VkPipeline* pPipelines) 2824 { 2825 VkResult result = VK_SUCCESS; 2826 2827 unsigned i = 0; 2828 for (; i < count; i++) { 2829 VkResult r; 2830 r = radv_compute_pipeline_create(_device, pipelineCache, 2831 &pCreateInfos[i], 2832 pAllocator, &pPipelines[i]); 2833 if (r != VK_SUCCESS) { 2834 result = r; 2835 pPipelines[i] = VK_NULL_HANDLE; 2836 } 2837 } 2838 2839 return result; 2840 } 2841