1 /* 2 * Copyright 2011 Intel Corporation 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice (including the next 12 * paragraph) shall be included in all copies or substantial portions of the 13 * Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER 21 * DEALINGS IN THE SOFTWARE. 22 */ 23 24 /** 25 * \file lower_varyings_to_packed.cpp 26 * 27 * This lowering pass generates GLSL code that manually packs varyings into 28 * vec4 slots, for the benefit of back-ends that don't support packed varyings 29 * natively. 30 * 31 * For example, the following shader: 32 * 33 * out mat3x2 foo; // location=4, location_frac=0 34 * out vec3 bar[2]; // location=5, location_frac=2 35 * 36 * main() 37 * { 38 * ... 39 * } 40 * 41 * Is rewritten to: 42 * 43 * mat3x2 foo; 44 * vec3 bar[2]; 45 * out vec4 packed4; // location=4, location_frac=0 46 * out vec4 packed5; // location=5, location_frac=0 47 * out vec4 packed6; // location=6, location_frac=0 48 * 49 * main() 50 * { 51 * ... 52 * packed4.xy = foo[0]; 53 * packed4.zw = foo[1]; 54 * packed5.xy = foo[2]; 55 * packed5.zw = bar[0].xy; 56 * packed6.x = bar[0].z; 57 * packed6.yzw = bar[1]; 58 * } 59 * 60 * This lowering pass properly handles "double parking" of a varying vector 61 * across two varying slots. For example, in the code above, two of the 62 * components of bar[0] are stored in packed5, and the remaining component is 63 * stored in packed6. 64 * 65 * Note that in theory, the extra instructions may cause some loss of 66 * performance. However, hopefully in most cases the performance loss will 67 * either be absorbed by a later optimization pass, or it will be offset by 68 * memory bandwidth savings (because fewer varyings are used). 69 * 70 * This lowering pass also packs flat floats, ints, and uints together, by 71 * using ivec4 as the base type of flat "varyings", and using appropriate 72 * casts to convert floats and uints into ints. 73 * 74 * This lowering pass also handles varyings whose type is a struct or an array 75 * of struct. Structs are packed in order and with no gaps, so there may be a 76 * performance penalty due to structure elements being double-parked. 77 * 78 * Lowering of geometry shader inputs is slightly more complex, since geometry 79 * inputs are always arrays, so we need to lower arrays to arrays. For 80 * example, the following input: 81 * 82 * in struct Foo { 83 * float f; 84 * vec3 v; 85 * vec2 a[2]; 86 * } arr[3]; // location=4, location_frac=0 87 * 88 * Would get lowered like this if it occurred in a fragment shader: 89 * 90 * struct Foo { 91 * float f; 92 * vec3 v; 93 * vec2 a[2]; 94 * } arr[3]; 95 * in vec4 packed4; // location=4, location_frac=0 96 * in vec4 packed5; // location=5, location_frac=0 97 * in vec4 packed6; // location=6, location_frac=0 98 * in vec4 packed7; // location=7, location_frac=0 99 * in vec4 packed8; // location=8, location_frac=0 100 * in vec4 packed9; // location=9, location_frac=0 101 * 102 * main() 103 * { 104 * arr[0].f = packed4.x; 105 * arr[0].v = packed4.yzw; 106 * arr[0].a[0] = packed5.xy; 107 * arr[0].a[1] = packed5.zw; 108 * arr[1].f = packed6.x; 109 * arr[1].v = packed6.yzw; 110 * arr[1].a[0] = packed7.xy; 111 * arr[1].a[1] = packed7.zw; 112 * arr[2].f = packed8.x; 113 * arr[2].v = packed8.yzw; 114 * arr[2].a[0] = packed9.xy; 115 * arr[2].a[1] = packed9.zw; 116 * ... 117 * } 118 * 119 * But it would get lowered like this if it occurred in a geometry shader: 120 * 121 * struct Foo { 122 * float f; 123 * vec3 v; 124 * vec2 a[2]; 125 * } arr[3]; 126 * in vec4 packed4[3]; // location=4, location_frac=0 127 * in vec4 packed5[3]; // location=5, location_frac=0 128 * 129 * main() 130 * { 131 * arr[0].f = packed4[0].x; 132 * arr[0].v = packed4[0].yzw; 133 * arr[0].a[0] = packed5[0].xy; 134 * arr[0].a[1] = packed5[0].zw; 135 * arr[1].f = packed4[1].x; 136 * arr[1].v = packed4[1].yzw; 137 * arr[1].a[0] = packed5[1].xy; 138 * arr[1].a[1] = packed5[1].zw; 139 * arr[2].f = packed4[2].x; 140 * arr[2].v = packed4[2].yzw; 141 * arr[2].a[0] = packed5[2].xy; 142 * arr[2].a[1] = packed5[2].zw; 143 * ... 144 * } 145 */ 146 147 #include "glsl_symbol_table.h" 148 #include "ir.h" 149 #include "ir_builder.h" 150 #include "ir_optimization.h" 151 #include "program/prog_instruction.h" 152 153 using namespace ir_builder; 154 155 namespace { 156 157 /** 158 * Visitor that performs varying packing. For each varying declared in the 159 * shader, this visitor determines whether it needs to be packed. If so, it 160 * demotes it to an ordinary global, creates new packed varyings, and 161 * generates assignments to convert between the original varying and the 162 * packed varying. 163 */ 164 class lower_packed_varyings_visitor 165 { 166 public: 167 lower_packed_varyings_visitor(void *mem_ctx, 168 unsigned locations_used, 169 const uint8_t *components, 170 ir_variable_mode mode, 171 unsigned gs_input_vertices, 172 exec_list *out_instructions, 173 exec_list *out_variables, 174 bool disable_varying_packing, 175 bool xfb_enabled); 176 177 void run(struct gl_linked_shader *shader); 178 179 private: 180 void bitwise_assign_pack(ir_rvalue *lhs, ir_rvalue *rhs); 181 void bitwise_assign_unpack(ir_rvalue *lhs, ir_rvalue *rhs); 182 unsigned lower_rvalue(ir_rvalue *rvalue, unsigned fine_location, 183 ir_variable *unpacked_var, const char *name, 184 bool gs_input_toplevel, unsigned vertex_index); 185 unsigned lower_arraylike(ir_rvalue *rvalue, unsigned array_size, 186 unsigned fine_location, 187 ir_variable *unpacked_var, const char *name, 188 bool gs_input_toplevel, unsigned vertex_index); 189 ir_dereference *get_packed_varying_deref(unsigned location, 190 ir_variable *unpacked_var, 191 const char *name, 192 unsigned vertex_index); 193 bool needs_lowering(ir_variable *var); 194 195 /** 196 * Memory context used to allocate new instructions for the shader. 197 */ 198 void * const mem_ctx; 199 200 /** 201 * Number of generic varying slots which are used by this shader. This is 202 * used to allocate temporary intermediate data structures. If any varying 203 * used by this shader has a location greater than or equal to 204 * VARYING_SLOT_VAR0 + locations_used, an assertion will fire. 205 */ 206 const unsigned locations_used; 207 208 const uint8_t* components; 209 210 /** 211 * Array of pointers to the packed varyings that have been created for each 212 * generic varying slot. NULL entries in this array indicate varying slots 213 * for which a packed varying has not been created yet. 214 */ 215 ir_variable **packed_varyings; 216 217 /** 218 * Type of varying which is being lowered in this pass (either 219 * ir_var_shader_in or ir_var_shader_out). 220 */ 221 const ir_variable_mode mode; 222 223 /** 224 * If we are currently lowering geometry shader inputs, the number of input 225 * vertices the geometry shader accepts. Otherwise zero. 226 */ 227 const unsigned gs_input_vertices; 228 229 /** 230 * Exec list into which the visitor should insert the packing instructions. 231 * Caller provides this list; it should insert the instructions into the 232 * appropriate place in the shader once the visitor has finished running. 233 */ 234 exec_list *out_instructions; 235 236 /** 237 * Exec list into which the visitor should insert any new variables. 238 */ 239 exec_list *out_variables; 240 241 bool disable_varying_packing; 242 bool xfb_enabled; 243 }; 244 245 } /* anonymous namespace */ 246 247 lower_packed_varyings_visitor::lower_packed_varyings_visitor( 248 void *mem_ctx, unsigned locations_used, const uint8_t *components, 249 ir_variable_mode mode, 250 unsigned gs_input_vertices, exec_list *out_instructions, 251 exec_list *out_variables, bool disable_varying_packing, 252 bool xfb_enabled) 253 : mem_ctx(mem_ctx), 254 locations_used(locations_used), 255 components(components), 256 packed_varyings((ir_variable **) 257 rzalloc_array_size(mem_ctx, sizeof(*packed_varyings), 258 locations_used)), 259 mode(mode), 260 gs_input_vertices(gs_input_vertices), 261 out_instructions(out_instructions), 262 out_variables(out_variables), 263 disable_varying_packing(disable_varying_packing), 264 xfb_enabled(xfb_enabled) 265 { 266 } 267 268 void 269 lower_packed_varyings_visitor::run(struct gl_linked_shader *shader) 270 { 271 foreach_in_list(ir_instruction, node, shader->ir) { 272 ir_variable *var = node->as_variable(); 273 if (var == NULL) 274 continue; 275 276 if (var->data.mode != this->mode || 277 var->data.location < VARYING_SLOT_VAR0 || 278 !this->needs_lowering(var)) 279 continue; 280 281 /* This lowering pass is only capable of packing floats and ints 282 * together when their interpolation mode is "flat". Treat integers as 283 * being flat when the interpolation mode is none. 284 */ 285 assert(var->data.interpolation == INTERP_MODE_FLAT || 286 var->data.interpolation == INTERP_MODE_NONE || 287 !var->type->contains_integer()); 288 289 /* Clone the variable for program resource list before 290 * it gets modified and lost. 291 */ 292 if (!shader->packed_varyings) 293 shader->packed_varyings = new (shader) exec_list; 294 295 shader->packed_varyings->push_tail(var->clone(shader, NULL)); 296 297 /* Change the old varying into an ordinary global. */ 298 assert(var->data.mode != ir_var_temporary); 299 var->data.mode = ir_var_auto; 300 301 /* Create a reference to the old varying. */ 302 ir_dereference_variable *deref 303 = new(this->mem_ctx) ir_dereference_variable(var); 304 305 /* Recursively pack or unpack it. */ 306 this->lower_rvalue(deref, var->data.location * 4 + var->data.location_frac, var, 307 var->name, this->gs_input_vertices != 0, 0); 308 } 309 } 310 311 #define SWIZZLE_ZWZW MAKE_SWIZZLE4(SWIZZLE_Z, SWIZZLE_W, SWIZZLE_Z, SWIZZLE_W) 312 313 /** 314 * Make an ir_assignment from \c rhs to \c lhs, performing appropriate 315 * bitcasts if necessary to match up types. 316 * 317 * This function is called when packing varyings. 318 */ 319 void 320 lower_packed_varyings_visitor::bitwise_assign_pack(ir_rvalue *lhs, 321 ir_rvalue *rhs) 322 { 323 if (lhs->type->base_type != rhs->type->base_type) { 324 /* Since we only mix types in flat varyings, and we always store flat 325 * varyings as type ivec4, we need only produce conversions from (uint 326 * or float) to int. 327 */ 328 assert(lhs->type->base_type == GLSL_TYPE_INT); 329 switch (rhs->type->base_type) { 330 case GLSL_TYPE_UINT: 331 rhs = new(this->mem_ctx) 332 ir_expression(ir_unop_u2i, lhs->type, rhs); 333 break; 334 case GLSL_TYPE_FLOAT: 335 rhs = new(this->mem_ctx) 336 ir_expression(ir_unop_bitcast_f2i, lhs->type, rhs); 337 break; 338 case GLSL_TYPE_DOUBLE: 339 assert(rhs->type->vector_elements <= 2); 340 if (rhs->type->vector_elements == 2) { 341 ir_variable *t = new(mem_ctx) ir_variable(lhs->type, "pack", ir_var_temporary); 342 343 assert(lhs->type->vector_elements == 4); 344 this->out_variables->push_tail(t); 345 this->out_instructions->push_tail( 346 assign(t, u2i(expr(ir_unop_unpack_double_2x32, swizzle_x(rhs->clone(mem_ctx, NULL)))), 0x3)); 347 this->out_instructions->push_tail( 348 assign(t, u2i(expr(ir_unop_unpack_double_2x32, swizzle_y(rhs))), 0xc)); 349 rhs = deref(t).val; 350 } else { 351 rhs = u2i(expr(ir_unop_unpack_double_2x32, rhs)); 352 } 353 break; 354 case GLSL_TYPE_INT64: 355 assert(rhs->type->vector_elements <= 2); 356 if (rhs->type->vector_elements == 2) { 357 ir_variable *t = new(mem_ctx) ir_variable(lhs->type, "pack", ir_var_temporary); 358 359 assert(lhs->type->vector_elements == 4); 360 this->out_variables->push_tail(t); 361 this->out_instructions->push_tail( 362 assign(t, expr(ir_unop_unpack_int_2x32, swizzle_x(rhs->clone(mem_ctx, NULL))), 0x3)); 363 this->out_instructions->push_tail( 364 assign(t, expr(ir_unop_unpack_int_2x32, swizzle_y(rhs)), 0xc)); 365 rhs = deref(t).val; 366 } else { 367 rhs = expr(ir_unop_unpack_int_2x32, rhs); 368 } 369 break; 370 case GLSL_TYPE_UINT64: 371 assert(rhs->type->vector_elements <= 2); 372 if (rhs->type->vector_elements == 2) { 373 ir_variable *t = new(mem_ctx) ir_variable(lhs->type, "pack", ir_var_temporary); 374 375 assert(lhs->type->vector_elements == 4); 376 this->out_variables->push_tail(t); 377 this->out_instructions->push_tail( 378 assign(t, u2i(expr(ir_unop_unpack_uint_2x32, swizzle_x(rhs->clone(mem_ctx, NULL)))), 0x3)); 379 this->out_instructions->push_tail( 380 assign(t, u2i(expr(ir_unop_unpack_uint_2x32, swizzle_y(rhs))), 0xc)); 381 rhs = deref(t).val; 382 } else { 383 rhs = u2i(expr(ir_unop_unpack_uint_2x32, rhs)); 384 } 385 break; 386 case GLSL_TYPE_SAMPLER: 387 rhs = u2i(expr(ir_unop_unpack_sampler_2x32, rhs)); 388 break; 389 case GLSL_TYPE_IMAGE: 390 rhs = u2i(expr(ir_unop_unpack_image_2x32, rhs)); 391 break; 392 default: 393 assert(!"Unexpected type conversion while lowering varyings"); 394 break; 395 } 396 } 397 this->out_instructions->push_tail(new (this->mem_ctx) ir_assignment(lhs, rhs)); 398 } 399 400 401 /** 402 * Make an ir_assignment from \c rhs to \c lhs, performing appropriate 403 * bitcasts if necessary to match up types. 404 * 405 * This function is called when unpacking varyings. 406 */ 407 void 408 lower_packed_varyings_visitor::bitwise_assign_unpack(ir_rvalue *lhs, 409 ir_rvalue *rhs) 410 { 411 if (lhs->type->base_type != rhs->type->base_type) { 412 /* Since we only mix types in flat varyings, and we always store flat 413 * varyings as type ivec4, we need only produce conversions from int to 414 * (uint or float). 415 */ 416 assert(rhs->type->base_type == GLSL_TYPE_INT); 417 switch (lhs->type->base_type) { 418 case GLSL_TYPE_UINT: 419 rhs = new(this->mem_ctx) 420 ir_expression(ir_unop_i2u, lhs->type, rhs); 421 break; 422 case GLSL_TYPE_FLOAT: 423 rhs = new(this->mem_ctx) 424 ir_expression(ir_unop_bitcast_i2f, lhs->type, rhs); 425 break; 426 case GLSL_TYPE_DOUBLE: 427 assert(lhs->type->vector_elements <= 2); 428 if (lhs->type->vector_elements == 2) { 429 ir_variable *t = new(mem_ctx) ir_variable(lhs->type, "unpack", ir_var_temporary); 430 assert(rhs->type->vector_elements == 4); 431 this->out_variables->push_tail(t); 432 this->out_instructions->push_tail( 433 assign(t, expr(ir_unop_pack_double_2x32, i2u(swizzle_xy(rhs->clone(mem_ctx, NULL)))), 0x1)); 434 this->out_instructions->push_tail( 435 assign(t, expr(ir_unop_pack_double_2x32, i2u(swizzle(rhs->clone(mem_ctx, NULL), SWIZZLE_ZWZW, 2))), 0x2)); 436 rhs = deref(t).val; 437 } else { 438 rhs = expr(ir_unop_pack_double_2x32, i2u(rhs)); 439 } 440 break; 441 case GLSL_TYPE_INT64: 442 assert(lhs->type->vector_elements <= 2); 443 if (lhs->type->vector_elements == 2) { 444 ir_variable *t = new(mem_ctx) ir_variable(lhs->type, "unpack", ir_var_temporary); 445 assert(rhs->type->vector_elements == 4); 446 this->out_variables->push_tail(t); 447 this->out_instructions->push_tail( 448 assign(t, expr(ir_unop_pack_int_2x32, swizzle_xy(rhs->clone(mem_ctx, NULL))), 0x1)); 449 this->out_instructions->push_tail( 450 assign(t, expr(ir_unop_pack_int_2x32, swizzle(rhs->clone(mem_ctx, NULL), SWIZZLE_ZWZW, 2)), 0x2)); 451 rhs = deref(t).val; 452 } else { 453 rhs = expr(ir_unop_pack_int_2x32, rhs); 454 } 455 break; 456 case GLSL_TYPE_UINT64: 457 assert(lhs->type->vector_elements <= 2); 458 if (lhs->type->vector_elements == 2) { 459 ir_variable *t = new(mem_ctx) ir_variable(lhs->type, "unpack", ir_var_temporary); 460 assert(rhs->type->vector_elements == 4); 461 this->out_variables->push_tail(t); 462 this->out_instructions->push_tail( 463 assign(t, expr(ir_unop_pack_uint_2x32, i2u(swizzle_xy(rhs->clone(mem_ctx, NULL)))), 0x1)); 464 this->out_instructions->push_tail( 465 assign(t, expr(ir_unop_pack_uint_2x32, i2u(swizzle(rhs->clone(mem_ctx, NULL), SWIZZLE_ZWZW, 2))), 0x2)); 466 rhs = deref(t).val; 467 } else { 468 rhs = expr(ir_unop_pack_uint_2x32, i2u(rhs)); 469 } 470 break; 471 case GLSL_TYPE_SAMPLER: 472 rhs = new(mem_ctx) 473 ir_expression(ir_unop_pack_sampler_2x32, lhs->type, i2u(rhs)); 474 break; 475 case GLSL_TYPE_IMAGE: 476 rhs = new(mem_ctx) 477 ir_expression(ir_unop_pack_image_2x32, lhs->type, i2u(rhs)); 478 break; 479 default: 480 assert(!"Unexpected type conversion while lowering varyings"); 481 break; 482 } 483 } 484 this->out_instructions->push_tail(new(this->mem_ctx) ir_assignment(lhs, rhs)); 485 } 486 487 488 /** 489 * Recursively pack or unpack the given varying (or portion of a varying) by 490 * traversing all of its constituent vectors. 491 * 492 * \param fine_location is the location where the first constituent vector 493 * should be packed--the word "fine" indicates that this location is expressed 494 * in multiples of a float, rather than multiples of a vec4 as is used 495 * elsewhere in Mesa. 496 * 497 * \param gs_input_toplevel should be set to true if we are lowering geometry 498 * shader inputs, and we are currently lowering the whole input variable 499 * (i.e. we are lowering the array whose index selects the vertex). 500 * 501 * \param vertex_index: if we are lowering geometry shader inputs, and the 502 * level of the array that we are currently lowering is *not* the top level, 503 * then this indicates which vertex we are currently lowering. Otherwise it 504 * is ignored. 505 * 506 * \return the location where the next constituent vector (after this one) 507 * should be packed. 508 */ 509 unsigned 510 lower_packed_varyings_visitor::lower_rvalue(ir_rvalue *rvalue, 511 unsigned fine_location, 512 ir_variable *unpacked_var, 513 const char *name, 514 bool gs_input_toplevel, 515 unsigned vertex_index) 516 { 517 unsigned dmul = rvalue->type->is_64bit() ? 2 : 1; 518 /* When gs_input_toplevel is set, we should be looking at a geometry shader 519 * input array. 520 */ 521 assert(!gs_input_toplevel || rvalue->type->is_array()); 522 523 if (rvalue->type->is_record()) { 524 for (unsigned i = 0; i < rvalue->type->length; i++) { 525 if (i != 0) 526 rvalue = rvalue->clone(this->mem_ctx, NULL); 527 const char *field_name = rvalue->type->fields.structure[i].name; 528 ir_dereference_record *dereference_record = new(this->mem_ctx) 529 ir_dereference_record(rvalue, field_name); 530 char *deref_name 531 = ralloc_asprintf(this->mem_ctx, "%s.%s", name, field_name); 532 fine_location = this->lower_rvalue(dereference_record, fine_location, 533 unpacked_var, deref_name, false, 534 vertex_index); 535 } 536 return fine_location; 537 } else if (rvalue->type->is_array()) { 538 /* Arrays are packed/unpacked by considering each array element in 539 * sequence. 540 */ 541 return this->lower_arraylike(rvalue, rvalue->type->array_size(), 542 fine_location, unpacked_var, name, 543 gs_input_toplevel, vertex_index); 544 } else if (rvalue->type->is_matrix()) { 545 /* Matrices are packed/unpacked by considering each column vector in 546 * sequence. 547 */ 548 return this->lower_arraylike(rvalue, rvalue->type->matrix_columns, 549 fine_location, unpacked_var, name, 550 false, vertex_index); 551 } else if (rvalue->type->vector_elements * dmul + 552 fine_location % 4 > 4) { 553 /* This vector is going to be "double parked" across two varying slots, 554 * so handle it as two separate assignments. For doubles, a dvec3/dvec4 555 * can end up being spread over 3 slots. However the second splitting 556 * will happen later, here we just always want to split into 2. 557 */ 558 unsigned left_components, right_components; 559 unsigned left_swizzle_values[4] = { 0, 0, 0, 0 }; 560 unsigned right_swizzle_values[4] = { 0, 0, 0, 0 }; 561 char left_swizzle_name[4] = { 0, 0, 0, 0 }; 562 char right_swizzle_name[4] = { 0, 0, 0, 0 }; 563 564 left_components = 4 - fine_location % 4; 565 if (rvalue->type->is_64bit()) { 566 /* We might actually end up with 0 left components! */ 567 left_components /= 2; 568 } 569 right_components = rvalue->type->vector_elements - left_components; 570 571 for (unsigned i = 0; i < left_components; i++) { 572 left_swizzle_values[i] = i; 573 left_swizzle_name[i] = "xyzw"[i]; 574 } 575 for (unsigned i = 0; i < right_components; i++) { 576 right_swizzle_values[i] = i + left_components; 577 right_swizzle_name[i] = "xyzw"[i + left_components]; 578 } 579 ir_swizzle *left_swizzle = new(this->mem_ctx) 580 ir_swizzle(rvalue, left_swizzle_values, left_components); 581 ir_swizzle *right_swizzle = new(this->mem_ctx) 582 ir_swizzle(rvalue->clone(this->mem_ctx, NULL), right_swizzle_values, 583 right_components); 584 char *left_name 585 = ralloc_asprintf(this->mem_ctx, "%s.%s", name, left_swizzle_name); 586 char *right_name 587 = ralloc_asprintf(this->mem_ctx, "%s.%s", name, right_swizzle_name); 588 if (left_components) 589 fine_location = this->lower_rvalue(left_swizzle, fine_location, 590 unpacked_var, left_name, false, 591 vertex_index); 592 else 593 /* Top up the fine location to the next slot */ 594 fine_location++; 595 return this->lower_rvalue(right_swizzle, fine_location, unpacked_var, 596 right_name, false, vertex_index); 597 } else { 598 /* No special handling is necessary; pack the rvalue into the 599 * varying. 600 */ 601 unsigned swizzle_values[4] = { 0, 0, 0, 0 }; 602 unsigned components = rvalue->type->vector_elements * dmul; 603 unsigned location = fine_location / 4; 604 unsigned location_frac = fine_location % 4; 605 for (unsigned i = 0; i < components; ++i) 606 swizzle_values[i] = i + location_frac; 607 ir_dereference *packed_deref = 608 this->get_packed_varying_deref(location, unpacked_var, name, 609 vertex_index); 610 if (unpacked_var->data.stream != 0) { 611 assert(unpacked_var->data.stream < 4); 612 ir_variable *packed_var = packed_deref->variable_referenced(); 613 for (unsigned i = 0; i < components; ++i) { 614 packed_var->data.stream |= 615 unpacked_var->data.stream << (2 * (location_frac + i)); 616 } 617 } 618 ir_swizzle *swizzle = new(this->mem_ctx) 619 ir_swizzle(packed_deref, swizzle_values, components); 620 if (this->mode == ir_var_shader_out) { 621 this->bitwise_assign_pack(swizzle, rvalue); 622 } else { 623 this->bitwise_assign_unpack(rvalue, swizzle); 624 } 625 return fine_location + components; 626 } 627 } 628 629 /** 630 * Recursively pack or unpack a varying for which we need to iterate over its 631 * constituent elements, accessing each one using an ir_dereference_array. 632 * This takes care of both arrays and matrices, since ir_dereference_array 633 * treats a matrix like an array of its column vectors. 634 * 635 * \param gs_input_toplevel should be set to true if we are lowering geometry 636 * shader inputs, and we are currently lowering the whole input variable 637 * (i.e. we are lowering the array whose index selects the vertex). 638 * 639 * \param vertex_index: if we are lowering geometry shader inputs, and the 640 * level of the array that we are currently lowering is *not* the top level, 641 * then this indicates which vertex we are currently lowering. Otherwise it 642 * is ignored. 643 */ 644 unsigned 645 lower_packed_varyings_visitor::lower_arraylike(ir_rvalue *rvalue, 646 unsigned array_size, 647 unsigned fine_location, 648 ir_variable *unpacked_var, 649 const char *name, 650 bool gs_input_toplevel, 651 unsigned vertex_index) 652 { 653 for (unsigned i = 0; i < array_size; i++) { 654 if (i != 0) 655 rvalue = rvalue->clone(this->mem_ctx, NULL); 656 ir_constant *constant = new(this->mem_ctx) ir_constant(i); 657 ir_dereference_array *dereference_array = new(this->mem_ctx) 658 ir_dereference_array(rvalue, constant); 659 if (gs_input_toplevel) { 660 /* Geometry shader inputs are a special case. Instead of storing 661 * each element of the array at a different location, all elements 662 * are at the same location, but with a different vertex index. 663 */ 664 (void) this->lower_rvalue(dereference_array, fine_location, 665 unpacked_var, name, false, i); 666 } else { 667 char *subscripted_name 668 = ralloc_asprintf(this->mem_ctx, "%s[%d]", name, i); 669 fine_location = 670 this->lower_rvalue(dereference_array, fine_location, 671 unpacked_var, subscripted_name, 672 false, vertex_index); 673 } 674 } 675 return fine_location; 676 } 677 678 /** 679 * Retrieve the packed varying corresponding to the given varying location. 680 * If no packed varying has been created for the given varying location yet, 681 * create it and add it to the shader before returning it. 682 * 683 * The newly created varying inherits its interpolation parameters from \c 684 * unpacked_var. Its base type is ivec4 if we are lowering a flat varying, 685 * vec4 otherwise. 686 * 687 * \param vertex_index: if we are lowering geometry shader inputs, then this 688 * indicates which vertex we are currently lowering. Otherwise it is ignored. 689 */ 690 ir_dereference * 691 lower_packed_varyings_visitor::get_packed_varying_deref( 692 unsigned location, ir_variable *unpacked_var, const char *name, 693 unsigned vertex_index) 694 { 695 unsigned slot = location - VARYING_SLOT_VAR0; 696 assert(slot < locations_used); 697 if (this->packed_varyings[slot] == NULL) { 698 char *packed_name = ralloc_asprintf(this->mem_ctx, "packed:%s", name); 699 const glsl_type *packed_type; 700 assert(components[slot] != 0); 701 if (unpacked_var->is_interpolation_flat()) 702 packed_type = glsl_type::get_instance(GLSL_TYPE_INT, components[slot], 1); 703 else 704 packed_type = glsl_type::get_instance(GLSL_TYPE_FLOAT, components[slot], 1); 705 if (this->gs_input_vertices != 0) { 706 packed_type = 707 glsl_type::get_array_instance(packed_type, 708 this->gs_input_vertices); 709 } 710 ir_variable *packed_var = new(this->mem_ctx) 711 ir_variable(packed_type, packed_name, this->mode); 712 if (this->gs_input_vertices != 0) { 713 /* Prevent update_array_sizes() from messing with the size of the 714 * array. 715 */ 716 packed_var->data.max_array_access = this->gs_input_vertices - 1; 717 } 718 packed_var->data.centroid = unpacked_var->data.centroid; 719 packed_var->data.sample = unpacked_var->data.sample; 720 packed_var->data.patch = unpacked_var->data.patch; 721 packed_var->data.interpolation = 722 packed_type->without_array() == glsl_type::ivec4_type 723 ? unsigned(INTERP_MODE_FLAT) : unpacked_var->data.interpolation; 724 packed_var->data.location = location; 725 packed_var->data.precision = unpacked_var->data.precision; 726 packed_var->data.always_active_io = unpacked_var->data.always_active_io; 727 packed_var->data.stream = 1u << 31; 728 unpacked_var->insert_before(packed_var); 729 this->packed_varyings[slot] = packed_var; 730 } else { 731 /* For geometry shader inputs, only update the packed variable name the 732 * first time we visit each component. 733 */ 734 if (this->gs_input_vertices == 0 || vertex_index == 0) { 735 ir_variable *var = this->packed_varyings[slot]; 736 737 if (var->is_name_ralloced()) 738 ralloc_asprintf_append((char **) &var->name, ",%s", name); 739 else 740 var->name = ralloc_asprintf(var, "%s,%s", var->name, name); 741 } 742 } 743 744 ir_dereference *deref = new(this->mem_ctx) 745 ir_dereference_variable(this->packed_varyings[slot]); 746 if (this->gs_input_vertices != 0) { 747 /* When lowering GS inputs, the packed variable is an array, so we need 748 * to dereference it using vertex_index. 749 */ 750 ir_constant *constant = new(this->mem_ctx) ir_constant(vertex_index); 751 deref = new(this->mem_ctx) ir_dereference_array(deref, constant); 752 } 753 return deref; 754 } 755 756 bool 757 lower_packed_varyings_visitor::needs_lowering(ir_variable *var) 758 { 759 /* Things composed of vec4's, varyings with explicitly assigned 760 * locations or varyings marked as must_be_shader_input (which might be used 761 * by interpolateAt* functions) shouldn't be lowered. Everything else can be. 762 */ 763 if (var->data.explicit_location || var->data.must_be_shader_input) 764 return false; 765 766 /* Override disable_varying_packing if the var is only used by transform 767 * feedback. Also override it if transform feedback is enabled and the 768 * variable is an array, struct or matrix as the elements of these types 769 * will always have the same interpolation and therefore are safe to pack. 770 */ 771 const glsl_type *type = var->type; 772 if (disable_varying_packing && !var->data.is_xfb_only && 773 !((type->is_array() || type->is_record() || type->is_matrix()) && 774 xfb_enabled)) 775 return false; 776 777 type = type->without_array(); 778 if (type->vector_elements == 4 && !type->is_64bit()) 779 return false; 780 return true; 781 } 782 783 784 /** 785 * Visitor that splices varying packing code before every use of EmitVertex() 786 * in a geometry shader. 787 */ 788 class lower_packed_varyings_gs_splicer : public ir_hierarchical_visitor 789 { 790 public: 791 explicit lower_packed_varyings_gs_splicer(void *mem_ctx, 792 const exec_list *instructions); 793 794 virtual ir_visitor_status visit_leave(ir_emit_vertex *ev); 795 796 private: 797 /** 798 * Memory context used to allocate new instructions for the shader. 799 */ 800 void * const mem_ctx; 801 802 /** 803 * Instructions that should be spliced into place before each EmitVertex() 804 * call. 805 */ 806 const exec_list *instructions; 807 }; 808 809 810 lower_packed_varyings_gs_splicer::lower_packed_varyings_gs_splicer( 811 void *mem_ctx, const exec_list *instructions) 812 : mem_ctx(mem_ctx), instructions(instructions) 813 { 814 } 815 816 817 ir_visitor_status 818 lower_packed_varyings_gs_splicer::visit_leave(ir_emit_vertex *ev) 819 { 820 foreach_in_list(ir_instruction, ir, this->instructions) { 821 ev->insert_before(ir->clone(this->mem_ctx, NULL)); 822 } 823 return visit_continue; 824 } 825 826 /** 827 * Visitor that splices varying packing code before every return. 828 */ 829 class lower_packed_varyings_return_splicer : public ir_hierarchical_visitor 830 { 831 public: 832 explicit lower_packed_varyings_return_splicer(void *mem_ctx, 833 const exec_list *instructions); 834 835 virtual ir_visitor_status visit_leave(ir_return *ret); 836 837 private: 838 /** 839 * Memory context used to allocate new instructions for the shader. 840 */ 841 void * const mem_ctx; 842 843 /** 844 * Instructions that should be spliced into place before each return. 845 */ 846 const exec_list *instructions; 847 }; 848 849 850 lower_packed_varyings_return_splicer::lower_packed_varyings_return_splicer( 851 void *mem_ctx, const exec_list *instructions) 852 : mem_ctx(mem_ctx), instructions(instructions) 853 { 854 } 855 856 857 ir_visitor_status 858 lower_packed_varyings_return_splicer::visit_leave(ir_return *ret) 859 { 860 foreach_in_list(ir_instruction, ir, this->instructions) { 861 ret->insert_before(ir->clone(this->mem_ctx, NULL)); 862 } 863 return visit_continue; 864 } 865 866 void 867 lower_packed_varyings(void *mem_ctx, unsigned locations_used, 868 const uint8_t *components, 869 ir_variable_mode mode, unsigned gs_input_vertices, 870 gl_linked_shader *shader, bool disable_varying_packing, 871 bool xfb_enabled) 872 { 873 exec_list *instructions = shader->ir; 874 ir_function *main_func = shader->symbols->get_function("main"); 875 exec_list void_parameters; 876 ir_function_signature *main_func_sig 877 = main_func->matching_signature(NULL, &void_parameters, false); 878 exec_list new_instructions, new_variables; 879 lower_packed_varyings_visitor visitor(mem_ctx, 880 locations_used, 881 components, 882 mode, 883 gs_input_vertices, 884 &new_instructions, 885 &new_variables, 886 disable_varying_packing, 887 xfb_enabled); 888 visitor.run(shader); 889 if (mode == ir_var_shader_out) { 890 if (shader->Stage == MESA_SHADER_GEOMETRY) { 891 /* For geometry shaders, outputs need to be lowered before each call 892 * to EmitVertex() 893 */ 894 lower_packed_varyings_gs_splicer splicer(mem_ctx, &new_instructions); 895 896 /* Add all the variables in first. */ 897 main_func_sig->body.get_head_raw()->insert_before(&new_variables); 898 899 /* Now update all the EmitVertex instances */ 900 splicer.run(instructions); 901 } else { 902 /* For other shader types, outputs need to be lowered before each 903 * return statement and at the end of main() 904 */ 905 906 lower_packed_varyings_return_splicer splicer(mem_ctx, &new_instructions); 907 908 main_func_sig->body.get_head_raw()->insert_before(&new_variables); 909 910 splicer.run(instructions); 911 912 /* Lower outputs at the end of main() if the last instruction is not 913 * a return statement 914 */ 915 if (((ir_instruction*)instructions->get_tail())->ir_type != ir_type_return) { 916 main_func_sig->body.append_list(&new_instructions); 917 } 918 } 919 } else { 920 /* Shader inputs need to be lowered at the beginning of main() */ 921 main_func_sig->body.get_head_raw()->insert_before(&new_instructions); 922 main_func_sig->body.get_head_raw()->insert_before(&new_variables); 923 } 924 } 925