draw_gs.c revision b8e80941
1/************************************************************************** 2 * 3 * Copyright 2009 VMware, Inc. 4 * All Rights Reserved. 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a 7 * copy of this software and associated documentation files (the 8 * "Software"), to deal in the Software without restriction, including 9 * without limitation the rights to use, copy, modify, merge, publish, 10 * distribute, sub license, and/or sell copies of the Software, and to 11 * permit persons to whom the Software is furnished to do so, subject to 12 * the following conditions: 13 * 14 * The above copyright notice and this permission notice (including the 15 * next paragraph) shall be included in all copies or substantial portions 16 * of the Software. 17 * 18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS 19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. 21 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR 22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, 23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE 24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. 25 * 26 **************************************************************************/ 27 28#include "draw_gs.h" 29 30#include "draw_private.h" 31#include "draw_context.h" 32#ifdef HAVE_LLVM 33#include "draw_llvm.h" 34#endif 35 36#include "tgsi/tgsi_parse.h" 37#include "tgsi/tgsi_exec.h" 38 39#include "pipe/p_shader_tokens.h" 40 41#include "util/u_math.h" 42#include "util/u_memory.h" 43#include "util/u_prim.h" 44 45/* fixme: move it from here */ 46#define MAX_PRIMITIVES 64 47 48static inline int 49draw_gs_get_input_index(int semantic, int index, 50 const struct tgsi_shader_info *input_info) 51{ 52 int i; 53 const ubyte *input_semantic_names = input_info->output_semantic_name; 54 const ubyte *input_semantic_indices = input_info->output_semantic_index; 55 for (i = 0; i < PIPE_MAX_SHADER_OUTPUTS; i++) { 56 if (input_semantic_names[i] == semantic && 57 input_semantic_indices[i] == index) 58 return i; 59 } 60 return -1; 61} 62 63/** 64 * We execute geometry shaders in the SOA mode, so ideally we want to 65 * flush when the number of currently fetched primitives is equal to 66 * the number of elements in the SOA vector. This ensures that the 67 * throughput is optimized for the given vector instruction set. 68 */ 69static inline boolean 70draw_gs_should_flush(struct draw_geometry_shader *shader) 71{ 72 return (shader->fetched_prim_count == shader->vector_length); 73} 74 75/*#define DEBUG_OUTPUTS 1*/ 76static void 77tgsi_fetch_gs_outputs(struct draw_geometry_shader *shader, 78 unsigned stream, 79 unsigned num_primitives, 80 float (**p_output)[4]) 81{ 82 struct tgsi_exec_machine *machine = shader->machine; 83 unsigned prim_idx, j, slot; 84 float (*output)[4]; 85 86 output = *p_output; 87 88 /* Unswizzle all output results. 89 */ 90 91 for (prim_idx = 0; prim_idx < num_primitives; ++prim_idx) { 92 unsigned num_verts_per_prim = machine->Primitives[stream][prim_idx]; 93 unsigned prim_offset = machine->PrimitiveOffsets[stream][prim_idx]; 94 shader->stream[stream].primitive_lengths[prim_idx + shader->stream[stream].emitted_primitives] = 95 machine->Primitives[stream][prim_idx]; 96 shader->stream[stream].emitted_vertices += num_verts_per_prim; 97 98 for (j = 0; j < num_verts_per_prim; j++) { 99 int idx = prim_offset + j * shader->info.num_outputs; 100#ifdef DEBUG_OUTPUTS 101 debug_printf("%d/%d) Output vert:\n", stream, idx / shader->info.num_outputs); 102#endif 103 for (slot = 0; slot < shader->info.num_outputs; slot++) { 104 output[slot][0] = machine->Outputs[idx + slot].xyzw[0].f[0]; 105 output[slot][1] = machine->Outputs[idx + slot].xyzw[1].f[0]; 106 output[slot][2] = machine->Outputs[idx + slot].xyzw[2].f[0]; 107 output[slot][3] = machine->Outputs[idx + slot].xyzw[3].f[0]; 108#ifdef DEBUG_OUTPUTS 109 debug_printf("\t%d: %f %f %f %f\n", slot, 110 output[slot][0], 111 output[slot][1], 112 output[slot][2], 113 output[slot][3]); 114#endif 115 } 116 output = (float (*)[4])((char *)output + shader->vertex_size); 117 } 118 } 119 *p_output = output; 120 shader->stream[stream].emitted_primitives += num_primitives; 121} 122 123/*#define DEBUG_INPUTS 1*/ 124static void tgsi_fetch_gs_input(struct draw_geometry_shader *shader, 125 unsigned *indices, 126 unsigned num_vertices, 127 unsigned prim_idx) 128{ 129 struct tgsi_exec_machine *machine = shader->machine; 130 unsigned slot, i; 131 int vs_slot; 132 unsigned input_vertex_stride = shader->input_vertex_stride; 133 const float (*input_ptr)[4]; 134 135 input_ptr = shader->input; 136 137 for (i = 0; i < num_vertices; ++i) { 138 const float (*input)[4]; 139#if DEBUG_INPUTS 140 debug_printf("%d) vertex index = %d (prim idx = %d)\n", 141 i, indices[i], prim_idx); 142#endif 143 input = (const float (*)[4])( 144 (const char *)input_ptr + (indices[i] * input_vertex_stride)); 145 for (slot = 0, vs_slot = 0; slot < shader->info.num_inputs; ++slot) { 146 unsigned idx = i * TGSI_EXEC_MAX_INPUT_ATTRIBS + slot; 147 if (shader->info.input_semantic_name[slot] == TGSI_SEMANTIC_PRIMID) { 148 machine->Inputs[idx].xyzw[0].u[prim_idx] = shader->in_prim_idx; 149 machine->Inputs[idx].xyzw[1].u[prim_idx] = shader->in_prim_idx; 150 machine->Inputs[idx].xyzw[2].u[prim_idx] = shader->in_prim_idx; 151 machine->Inputs[idx].xyzw[3].u[prim_idx] = shader->in_prim_idx; 152 } else { 153 vs_slot = draw_gs_get_input_index( 154 shader->info.input_semantic_name[slot], 155 shader->info.input_semantic_index[slot], 156 shader->input_info); 157 if (vs_slot < 0) { 158 debug_printf("VS/GS signature mismatch!\n"); 159 machine->Inputs[idx].xyzw[0].f[prim_idx] = 0; 160 machine->Inputs[idx].xyzw[1].f[prim_idx] = 0; 161 machine->Inputs[idx].xyzw[2].f[prim_idx] = 0; 162 machine->Inputs[idx].xyzw[3].f[prim_idx] = 0; 163 } else { 164#if DEBUG_INPUTS 165 debug_printf("\tSlot = %d, vs_slot = %d, idx = %d:\n", 166 slot, vs_slot, idx); 167 assert(!util_is_inf_or_nan(input[vs_slot][0])); 168 assert(!util_is_inf_or_nan(input[vs_slot][1])); 169 assert(!util_is_inf_or_nan(input[vs_slot][2])); 170 assert(!util_is_inf_or_nan(input[vs_slot][3])); 171#endif 172 machine->Inputs[idx].xyzw[0].f[prim_idx] = input[vs_slot][0]; 173 machine->Inputs[idx].xyzw[1].f[prim_idx] = input[vs_slot][1]; 174 machine->Inputs[idx].xyzw[2].f[prim_idx] = input[vs_slot][2]; 175 machine->Inputs[idx].xyzw[3].f[prim_idx] = input[vs_slot][3]; 176#if DEBUG_INPUTS 177 debug_printf("\t\t%f %f %f %f\n", 178 machine->Inputs[idx].xyzw[0].f[prim_idx], 179 machine->Inputs[idx].xyzw[1].f[prim_idx], 180 machine->Inputs[idx].xyzw[2].f[prim_idx], 181 machine->Inputs[idx].xyzw[3].f[prim_idx]); 182#endif 183 ++vs_slot; 184 } 185 } 186 } 187 } 188} 189 190static void tgsi_gs_prepare(struct draw_geometry_shader *shader, 191 const void *constants[PIPE_MAX_CONSTANT_BUFFERS], 192 const unsigned constants_size[PIPE_MAX_CONSTANT_BUFFERS]) 193{ 194 struct tgsi_exec_machine *machine = shader->machine; 195 int j; 196 tgsi_exec_set_constant_buffers(machine, PIPE_MAX_CONSTANT_BUFFERS, 197 constants, constants_size); 198 199 if (shader->info.uses_invocationid) { 200 unsigned i = machine->SysSemanticToIndex[TGSI_SEMANTIC_INVOCATIONID]; 201 for (j = 0; j < TGSI_QUAD_SIZE; j++) 202 machine->SystemValue[i].xyzw[0].i[j] = shader->invocation_id; 203 } 204} 205 206static void tgsi_gs_run(struct draw_geometry_shader *shader, 207 unsigned input_primitives, 208 unsigned *out_prims) 209{ 210 struct tgsi_exec_machine *machine = shader->machine; 211 int i; 212 213 /* run interpreter */ 214 tgsi_exec_machine_run(machine, 0); 215 216 for (i = 0; i < 4; i++) { 217 int prim_i; 218 int prim_c; 219 switch (i) { 220 case 0: 221 prim_i = TGSI_EXEC_TEMP_PRIMITIVE_I; 222 prim_c = TGSI_EXEC_TEMP_PRIMITIVE_C; 223 break; 224 case 1: 225 prim_i = TGSI_EXEC_TEMP_PRIMITIVE_S1_I; 226 prim_c = TGSI_EXEC_TEMP_PRIMITIVE_S1_C; 227 break; 228 case 2: 229 prim_i = TGSI_EXEC_TEMP_PRIMITIVE_S2_I; 230 prim_c = TGSI_EXEC_TEMP_PRIMITIVE_S2_C; 231 break; 232 case 3: 233 prim_i = TGSI_EXEC_TEMP_PRIMITIVE_S3_I; 234 prim_c = TGSI_EXEC_TEMP_PRIMITIVE_S3_C; 235 break; 236 }; 237 238 out_prims[i] = machine->Temps[prim_i].xyzw[prim_c].u[0]; 239 } 240} 241 242#ifdef HAVE_LLVM 243 244static void 245llvm_fetch_gs_input(struct draw_geometry_shader *shader, 246 unsigned *indices, 247 unsigned num_vertices, 248 unsigned prim_idx) 249{ 250 unsigned slot, i; 251 int vs_slot; 252 unsigned input_vertex_stride = shader->input_vertex_stride; 253 const float (*input_ptr)[4]; 254 float (*input_data)[6][PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS][TGSI_NUM_CHANNELS] = &shader->gs_input->data; 255 256 shader->llvm_prim_ids[shader->fetched_prim_count] = shader->in_prim_idx; 257 258 input_ptr = shader->input; 259 260 for (i = 0; i < num_vertices; ++i) { 261 const float (*input)[4]; 262#if DEBUG_INPUTS 263 debug_printf("%d) vertex index = %d (prim idx = %d)\n", 264 i, indices[i], prim_idx); 265#endif 266 input = (const float (*)[4])( 267 (const char *)input_ptr + (indices[i] * input_vertex_stride)); 268 for (slot = 0, vs_slot = 0; slot < shader->info.num_inputs; ++slot) { 269 if (shader->info.input_semantic_name[slot] == TGSI_SEMANTIC_PRIMID) { 270 /* skip. we handle system values through gallivm */ 271 /* NOTE: If we hit this case here it's an ordinary input not a sv, 272 * even though it probably should be a sv. 273 * Not sure how to set it up as regular input however if that even, 274 * would make sense so hack around this later in gallivm. 275 */ 276 } else { 277 vs_slot = draw_gs_get_input_index( 278 shader->info.input_semantic_name[slot], 279 shader->info.input_semantic_index[slot], 280 shader->input_info); 281 if (vs_slot < 0) { 282 debug_printf("VS/GS signature mismatch!\n"); 283 (*input_data)[i][slot][0][prim_idx] = 0; 284 (*input_data)[i][slot][1][prim_idx] = 0; 285 (*input_data)[i][slot][2][prim_idx] = 0; 286 (*input_data)[i][slot][3][prim_idx] = 0; 287 } else { 288#if DEBUG_INPUTS 289 debug_printf("\tSlot = %d, vs_slot = %d, i = %d:\n", 290 slot, vs_slot, i); 291 assert(!util_is_inf_or_nan(input[vs_slot][0])); 292 assert(!util_is_inf_or_nan(input[vs_slot][1])); 293 assert(!util_is_inf_or_nan(input[vs_slot][2])); 294 assert(!util_is_inf_or_nan(input[vs_slot][3])); 295#endif 296 (*input_data)[i][slot][0][prim_idx] = input[vs_slot][0]; 297 (*input_data)[i][slot][1][prim_idx] = input[vs_slot][1]; 298 (*input_data)[i][slot][2][prim_idx] = input[vs_slot][2]; 299 (*input_data)[i][slot][3][prim_idx] = input[vs_slot][3]; 300#if DEBUG_INPUTS 301 debug_printf("\t\t%f %f %f %f\n", 302 (*input_data)[i][slot][0][prim_idx], 303 (*input_data)[i][slot][1][prim_idx], 304 (*input_data)[i][slot][2][prim_idx], 305 (*input_data)[i][slot][3][prim_idx]); 306#endif 307 ++vs_slot; 308 } 309 } 310 } 311 } 312} 313 314static void 315llvm_fetch_gs_outputs(struct draw_geometry_shader *shader, 316 unsigned stream, 317 unsigned num_primitives, 318 float (**p_output)[4]) 319{ 320 int total_verts = 0; 321 int vertex_count = 0; 322 int total_prims = 0; 323 int max_prims_per_invocation = 0; 324 char *output_ptr = (char*)shader->gs_output; 325 int i, j, prim_idx; 326 unsigned next_prim_boundary = shader->primitive_boundary; 327 328 for (i = 0; i < shader->vector_length; ++i) { 329 int prims = shader->llvm_emitted_primitives[i]; 330 total_prims += prims; 331 max_prims_per_invocation = MAX2(max_prims_per_invocation, prims); 332 } 333 for (i = 0; i < shader->vector_length; ++i) { 334 total_verts += shader->llvm_emitted_vertices[i]; 335 } 336 337 output_ptr += shader->stream[0].emitted_vertices * shader->vertex_size; 338 for (i = 0; i < shader->vector_length - 1; ++i) { 339 int current_verts = shader->llvm_emitted_vertices[i]; 340 int next_verts = shader->llvm_emitted_vertices[i + 1]; 341#if 0 342 int j; 343 for (j = 0; j < current_verts; ++j) { 344 struct vertex_header *vh = (struct vertex_header *) 345 (output_ptr + shader->vertex_size * (i * next_prim_boundary + j)); 346 debug_printf("--- %d) [%f, %f, %f, %f]\n", j + vertex_count, 347 vh->data[0][0], vh->data[0][1], vh->data[0][2], vh->data[0][3]); 348 349 } 350#endif 351 debug_assert(current_verts <= shader->max_output_vertices); 352 debug_assert(next_verts <= shader->max_output_vertices); 353 if (next_verts) { 354 memmove(output_ptr + (vertex_count + current_verts) * shader->vertex_size, 355 output_ptr + ((i + 1) * next_prim_boundary) * shader->vertex_size, 356 shader->vertex_size * next_verts); 357 } 358 vertex_count += current_verts; 359 } 360 361#if 0 362 { 363 int i; 364 for (i = 0; i < total_verts; ++i) { 365 struct vertex_header *vh = (struct vertex_header *)(output_ptr + shader->vertex_size * i); 366 debug_printf("%d) Vertex:\n", i); 367 for (j = 0; j < shader->info.num_outputs; ++j) { 368 unsigned *udata = (unsigned*)vh->data[j]; 369 debug_printf(" %d) [%f, %f, %f, %f] [%d, %d, %d, %d]\n", j, 370 vh->data[j][0], vh->data[j][1], vh->data[j][2], vh->data[j][3], 371 udata[0], udata[1], udata[2], udata[3]); 372 } 373 374 } 375 } 376#endif 377 378 prim_idx = 0; 379 for (i = 0; i < shader->vector_length; ++i) { 380 int num_prims = shader->llvm_emitted_primitives[i]; 381 for (j = 0; j < num_prims; ++j) { 382 int prim_length = 383 shader->llvm_prim_lengths[j][i]; 384 shader->stream[0].primitive_lengths[shader->stream[0].emitted_primitives + prim_idx] = 385 prim_length; 386 ++prim_idx; 387 } 388 } 389 390 shader->stream[0].emitted_primitives += total_prims; 391 shader->stream[0].emitted_vertices += total_verts; 392} 393 394static void 395llvm_gs_prepare(struct draw_geometry_shader *shader, 396 const void *constants[PIPE_MAX_CONSTANT_BUFFERS], 397 const unsigned constants_size[PIPE_MAX_CONSTANT_BUFFERS]) 398{ 399} 400 401static void 402llvm_gs_run(struct draw_geometry_shader *shader, 403 unsigned input_primitives, unsigned *out_prims) 404{ 405 unsigned ret; 406 char *input = (char*)shader->gs_output; 407 408 input += (shader->stream[0].emitted_vertices * shader->vertex_size); 409 410 ret = shader->current_variant->jit_func( 411 shader->jit_context, shader->gs_input->data, 412 (struct vertex_header*)input, 413 input_primitives, 414 shader->draw->instance_id, 415 shader->llvm_prim_ids, 416 shader->invocation_id); 417 418 *out_prims = ret; 419} 420 421#endif 422 423static void gs_flush(struct draw_geometry_shader *shader) 424{ 425 unsigned out_prim_count[TGSI_MAX_VERTEX_STREAMS]; 426 unsigned i; 427 unsigned input_primitives = shader->fetched_prim_count; 428 429 if (shader->draw->collect_statistics) { 430 shader->draw->statistics.gs_invocations += input_primitives; 431 } 432 433 debug_assert(input_primitives > 0 && 434 input_primitives <= 4); 435 436 shader->run(shader, input_primitives, out_prim_count); 437 for (i = 0; i < shader->num_vertex_streams; i++) { 438 shader->fetch_outputs(shader, i, out_prim_count[i], 439 &shader->stream[i].tmp_output); 440 } 441 442#if 0 443 for (i = 0; i < shader->num_vertex_streams; i++) { 444 debug_printf("stream %d: PRIM emitted prims = %d (verts=%d), cur prim count = %d\n", 445 i, 446 shader->stream[i].emitted_primitives, shader->stream[i].emitted_vertices, 447 out_prim_count[i]); 448 } 449#endif 450 451 shader->fetched_prim_count = 0; 452} 453 454static void gs_point(struct draw_geometry_shader *shader, 455 int idx) 456{ 457 unsigned indices[1]; 458 459 indices[0] = idx; 460 461 shader->fetch_inputs(shader, indices, 1, 462 shader->fetched_prim_count); 463 ++shader->in_prim_idx; 464 ++shader->fetched_prim_count; 465 466 if (draw_gs_should_flush(shader)) 467 gs_flush(shader); 468} 469 470static void gs_line(struct draw_geometry_shader *shader, 471 int i0, int i1) 472{ 473 unsigned indices[2]; 474 475 indices[0] = i0; 476 indices[1] = i1; 477 478 shader->fetch_inputs(shader, indices, 2, 479 shader->fetched_prim_count); 480 ++shader->in_prim_idx; 481 ++shader->fetched_prim_count; 482 483 if (draw_gs_should_flush(shader)) 484 gs_flush(shader); 485} 486 487static void gs_line_adj(struct draw_geometry_shader *shader, 488 int i0, int i1, int i2, int i3) 489{ 490 unsigned indices[4]; 491 492 indices[0] = i0; 493 indices[1] = i1; 494 indices[2] = i2; 495 indices[3] = i3; 496 497 shader->fetch_inputs(shader, indices, 4, 498 shader->fetched_prim_count); 499 ++shader->in_prim_idx; 500 ++shader->fetched_prim_count; 501 502 if (draw_gs_should_flush(shader)) 503 gs_flush(shader); 504} 505 506static void gs_tri(struct draw_geometry_shader *shader, 507 int i0, int i1, int i2) 508{ 509 unsigned indices[3]; 510 511 indices[0] = i0; 512 indices[1] = i1; 513 indices[2] = i2; 514 515 shader->fetch_inputs(shader, indices, 3, 516 shader->fetched_prim_count); 517 ++shader->in_prim_idx; 518 ++shader->fetched_prim_count; 519 520 if (draw_gs_should_flush(shader)) 521 gs_flush(shader); 522} 523 524static void gs_tri_adj(struct draw_geometry_shader *shader, 525 int i0, int i1, int i2, 526 int i3, int i4, int i5) 527{ 528 unsigned indices[6]; 529 530 indices[0] = i0; 531 indices[1] = i1; 532 indices[2] = i2; 533 indices[3] = i3; 534 indices[4] = i4; 535 indices[5] = i5; 536 537 shader->fetch_inputs(shader, indices, 6, 538 shader->fetched_prim_count); 539 ++shader->in_prim_idx; 540 ++shader->fetched_prim_count; 541 542 if (draw_gs_should_flush(shader)) 543 gs_flush(shader); 544} 545 546#define FUNC gs_run 547#define GET_ELT(idx) (idx) 548#include "draw_gs_tmp.h" 549 550 551#define FUNC gs_run_elts 552#define LOCAL_VARS const ushort *elts = input_prims->elts; 553#define GET_ELT(idx) (elts[idx]) 554#include "draw_gs_tmp.h" 555 556 557/** 558 * Execute geometry shader. 559 */ 560int draw_geometry_shader_run(struct draw_geometry_shader *shader, 561 const void *constants[PIPE_MAX_CONSTANT_BUFFERS], 562 const unsigned constants_size[PIPE_MAX_CONSTANT_BUFFERS], 563 const struct draw_vertex_info *input_verts, 564 const struct draw_prim_info *input_prim, 565 const struct tgsi_shader_info *input_info, 566 struct draw_vertex_info *output_verts, 567 struct draw_prim_info *output_prims ) 568{ 569 const float (*input)[4] = (const float (*)[4])input_verts->verts->data; 570 unsigned input_stride = input_verts->vertex_size; 571 unsigned num_outputs = draw_total_gs_outputs(shader->draw); 572 unsigned vertex_size = sizeof(struct vertex_header) + num_outputs * 4 * sizeof(float); 573 unsigned num_input_verts = input_prim->linear ? 574 input_verts->count : 575 input_prim->count; 576 unsigned num_in_primitives = 577 align( 578 MAX2(u_decomposed_prims_for_vertices(input_prim->prim, 579 num_input_verts), 580 u_decomposed_prims_for_vertices(shader->input_primitive, 581 num_input_verts)), 582 shader->vector_length); 583 unsigned max_out_prims = 584 u_decomposed_prims_for_vertices(shader->output_primitive, 585 shader->max_output_vertices) 586 * num_in_primitives; 587 /* we allocate exactly one extra vertex per primitive to allow the GS to emit 588 * overflown vertices into some area where they won't harm anyone */ 589 unsigned total_verts_per_buffer = shader->primitive_boundary * 590 num_in_primitives; 591 unsigned invocation; 592 int i; 593 //Assume at least one primitive 594 max_out_prims = MAX2(max_out_prims, 1); 595 596 for (i = 0; i < shader->num_vertex_streams; i++) { 597 /* write all the vertex data into all the streams */ 598 output_verts[i].vertex_size = vertex_size; 599 output_verts[i].stride = output_verts[i].vertex_size; 600 output_verts[i].verts = 601 (struct vertex_header *)MALLOC(output_verts[i].vertex_size * 602 total_verts_per_buffer * shader->num_invocations); 603 debug_assert(output_verts[i].verts); 604 } 605 606#if 0 607 debug_printf("%s count = %d (in prims # = %d)\n", 608 __FUNCTION__, num_input_verts, num_in_primitives); 609 debug_printf("\tlinear = %d, prim_info->count = %d\n", 610 input_prim->linear, input_prim->count); 611 debug_printf("\tprim pipe = %s, shader in = %s, shader out = %s\n" 612 u_prim_name(input_prim->prim), 613 u_prim_name(shader->input_primitive), 614 u_prim_name(shader->output_primitive)); 615 debug_printf("\tmaxv = %d, maxp = %d, primitive_boundary = %d, " 616 "vertex_size = %d, tverts = %d\n", 617 shader->max_output_vertices, max_out_prims, 618 shader->primitive_boundary, output_verts->vertex_size, 619 total_verts_per_buffer); 620#endif 621 622 for (i = 0; i < shader->num_vertex_streams; i++) { 623 shader->stream[i].emitted_vertices = 0; 624 shader->stream[i].emitted_primitives = 0; 625 FREE(shader->stream[i].primitive_lengths); 626 shader->stream[i].primitive_lengths = MALLOC(max_out_prims * sizeof(unsigned) * shader->num_invocations); 627 shader->stream[i].tmp_output = (float (*)[4])output_verts[i].verts->data; 628 } 629 shader->vertex_size = vertex_size; 630 shader->fetched_prim_count = 0; 631 shader->input_vertex_stride = input_stride; 632 shader->input = input; 633 shader->input_info = input_info; 634 635#ifdef HAVE_LLVM 636 if (shader->draw->llvm) { 637 shader->gs_output = output_verts[0].verts; 638 if (max_out_prims > shader->max_out_prims) { 639 unsigned i; 640 if (shader->llvm_prim_lengths) { 641 for (i = 0; i < shader->max_out_prims; ++i) { 642 align_free(shader->llvm_prim_lengths[i]); 643 } 644 FREE(shader->llvm_prim_lengths); 645 } 646 647 shader->llvm_prim_lengths = MALLOC(max_out_prims * sizeof(unsigned*)); 648 for (i = 0; i < max_out_prims; ++i) { 649 int vector_size = shader->vector_length * sizeof(unsigned); 650 shader->llvm_prim_lengths[i] = 651 align_malloc(vector_size, vector_size); 652 } 653 654 shader->max_out_prims = max_out_prims; 655 } 656 shader->jit_context->prim_lengths = shader->llvm_prim_lengths; 657 shader->jit_context->emitted_vertices = shader->llvm_emitted_vertices; 658 shader->jit_context->emitted_prims = shader->llvm_emitted_primitives; 659 } 660#endif 661 662 for (invocation = 0; invocation < shader->num_invocations; invocation++) { 663 shader->invocation_id = invocation; 664 665 shader->prepare(shader, constants, constants_size); 666 667 if (input_prim->linear) 668 gs_run(shader, input_prim, input_verts, 669 output_prims, output_verts); 670 else 671 gs_run_elts(shader, input_prim, input_verts, 672 output_prims, output_verts); 673 674 /* Flush the remaining primitives. Will happen if 675 * num_input_primitives % 4 != 0 676 */ 677 if (shader->fetched_prim_count > 0) { 678 gs_flush(shader); 679 } 680 debug_assert(shader->fetched_prim_count == 0); 681 } 682 683 /* Update prim_info: 684 */ 685 for (i = 0; i < shader->num_vertex_streams; i++) { 686 output_prims[i].linear = TRUE; 687 output_prims[i].elts = NULL; 688 output_prims[i].start = 0; 689 output_prims[i].count = shader->stream[i].emitted_vertices; 690 output_prims[i].prim = shader->output_primitive; 691 output_prims[i].flags = 0x0; 692 output_prims[i].primitive_lengths = shader->stream[i].primitive_lengths; 693 output_prims[i].primitive_count = shader->stream[i].emitted_primitives; 694 output_verts[i].count = shader->stream[i].emitted_vertices; 695 696 if (shader->draw->collect_statistics) { 697 unsigned j; 698 for (j = 0; j < shader->stream[i].emitted_primitives; ++j) { 699 shader->draw->statistics.gs_primitives += 700 u_decomposed_prims_for_vertices(shader->output_primitive, 701 shader->stream[i].primitive_lengths[j]); 702 } 703 } 704 } 705 706#if 0 707 debug_printf("GS finished\n"); 708 for (i = 0; i < 4; i++) 709 debug_printf("stream %d: prims = %d verts = %d\n", i, output_prims[i].primitive_count, output_verts[i].count); 710#endif 711 712 return 0; 713} 714 715void draw_geometry_shader_prepare(struct draw_geometry_shader *shader, 716 struct draw_context *draw) 717{ 718 boolean use_llvm = draw->llvm != NULL; 719 if (!use_llvm && shader && shader->machine->Tokens != shader->state.tokens) { 720 tgsi_exec_machine_bind_shader(shader->machine, 721 shader->state.tokens, 722 draw->gs.tgsi.sampler, 723 draw->gs.tgsi.image, 724 draw->gs.tgsi.buffer); 725 } 726} 727 728 729boolean 730draw_gs_init( struct draw_context *draw ) 731{ 732 if (!draw->llvm) { 733 draw->gs.tgsi.machine = tgsi_exec_machine_create(PIPE_SHADER_GEOMETRY); 734 735 for (unsigned i = 0; i < TGSI_MAX_VERTEX_STREAMS; i++) { 736 draw->gs.tgsi.machine->Primitives[i] = align_malloc( 737 MAX_PRIMITIVES * sizeof(struct tgsi_exec_vector), 16); 738 draw->gs.tgsi.machine->PrimitiveOffsets[i] = align_malloc( 739 MAX_PRIMITIVES * sizeof(struct tgsi_exec_vector), 16); 740 if (!draw->gs.tgsi.machine->Primitives[i] || !draw->gs.tgsi.machine->PrimitiveOffsets[i]) 741 return FALSE; 742 memset(draw->gs.tgsi.machine->Primitives[i], 0, 743 MAX_PRIMITIVES * sizeof(struct tgsi_exec_vector)); 744 memset(draw->gs.tgsi.machine->PrimitiveOffsets[i], 0, 745 MAX_PRIMITIVES * sizeof(struct tgsi_exec_vector)); 746 } 747 } 748 749 return TRUE; 750} 751 752void draw_gs_destroy( struct draw_context *draw ) 753{ 754 int i; 755 if (draw->gs.tgsi.machine) { 756 for (i = 0; i < TGSI_MAX_VERTEX_STREAMS; i++) { 757 align_free(draw->gs.tgsi.machine->Primitives[i]); 758 align_free(draw->gs.tgsi.machine->PrimitiveOffsets[i]); 759 } 760 tgsi_exec_machine_destroy(draw->gs.tgsi.machine); 761 } 762} 763 764struct draw_geometry_shader * 765draw_create_geometry_shader(struct draw_context *draw, 766 const struct pipe_shader_state *state) 767{ 768#ifdef HAVE_LLVM 769 boolean use_llvm = draw->llvm != NULL; 770 struct llvm_geometry_shader *llvm_gs = NULL; 771#endif 772 struct draw_geometry_shader *gs; 773 unsigned i; 774 775#ifdef HAVE_LLVM 776 if (use_llvm) { 777 llvm_gs = CALLOC_STRUCT(llvm_geometry_shader); 778 779 if (!llvm_gs) 780 return NULL; 781 782 gs = &llvm_gs->base; 783 784 make_empty_list(&llvm_gs->variants); 785 } else 786#endif 787 { 788 gs = CALLOC_STRUCT(draw_geometry_shader); 789 } 790 791 if (!gs) 792 return NULL; 793 794 gs->draw = draw; 795 gs->state = *state; 796 gs->state.tokens = tgsi_dup_tokens(state->tokens); 797 if (!gs->state.tokens) { 798 FREE(gs); 799 return NULL; 800 } 801 802 tgsi_scan_shader(state->tokens, &gs->info); 803 804 /* setup the defaults */ 805 gs->max_out_prims = 0; 806 807#ifdef HAVE_LLVM 808 if (use_llvm) { 809 /* TODO: change the input array to handle the following 810 vector length, instead of the currently hardcoded 811 TGSI_NUM_CHANNELS 812 gs->vector_length = lp_native_vector_width / 32;*/ 813 gs->vector_length = TGSI_NUM_CHANNELS; 814 } else 815#endif 816 { 817 gs->vector_length = 1; 818 } 819 820 gs->input_primitive = 821 gs->info.properties[TGSI_PROPERTY_GS_INPUT_PRIM]; 822 gs->output_primitive = 823 gs->info.properties[TGSI_PROPERTY_GS_OUTPUT_PRIM]; 824 gs->max_output_vertices = 825 gs->info.properties[TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES]; 826 gs->num_invocations = 827 gs->info.properties[TGSI_PROPERTY_GS_INVOCATIONS]; 828 if (!gs->max_output_vertices) 829 gs->max_output_vertices = 32; 830 831 /* Primitive boundary is bigger than max_output_vertices by one, because 832 * the specification says that the geometry shader should exit if the 833 * number of emitted vertices is bigger or equal to max_output_vertices and 834 * we can't do that because we're running in the SoA mode, which means that 835 * our storing routines will keep getting called on channels that have 836 * overflown. 837 * So we need some scratch area where we can keep writing the overflown 838 * vertices without overwriting anything important or crashing. 839 */ 840 gs->primitive_boundary = gs->max_output_vertices + 1; 841 842 gs->position_output = -1; 843 for (i = 0; i < gs->info.num_outputs; i++) { 844 if (gs->info.output_semantic_name[i] == TGSI_SEMANTIC_POSITION && 845 gs->info.output_semantic_index[i] == 0) 846 gs->position_output = i; 847 if (gs->info.output_semantic_name[i] == TGSI_SEMANTIC_VIEWPORT_INDEX) 848 gs->viewport_index_output = i; 849 if (gs->info.output_semantic_name[i] == TGSI_SEMANTIC_CLIPDIST) { 850 debug_assert(gs->info.output_semantic_index[i] < 851 PIPE_MAX_CLIP_OR_CULL_DISTANCE_ELEMENT_COUNT); 852 gs->ccdistance_output[gs->info.output_semantic_index[i]] = i; 853 } 854 } 855 856 gs->machine = draw->gs.tgsi.machine; 857 858 gs->num_vertex_streams = 1; 859 for (i = 0; i < gs->state.stream_output.num_outputs; i++) { 860 if (gs->state.stream_output.output[i].stream >= gs->num_vertex_streams) 861 gs->num_vertex_streams = gs->state.stream_output.output[i].stream + 1; 862 } 863 864#ifdef HAVE_LLVM 865 if (use_llvm) { 866 int vector_size = gs->vector_length * sizeof(float); 867 gs->gs_input = align_malloc(sizeof(struct draw_gs_inputs), 16); 868 memset(gs->gs_input, 0, sizeof(struct draw_gs_inputs)); 869 gs->llvm_prim_lengths = 0; 870 871 gs->llvm_emitted_primitives = align_malloc(vector_size, vector_size); 872 gs->llvm_emitted_vertices = align_malloc(vector_size, vector_size); 873 gs->llvm_prim_ids = align_malloc(vector_size, vector_size); 874 875 gs->fetch_outputs = llvm_fetch_gs_outputs; 876 gs->fetch_inputs = llvm_fetch_gs_input; 877 gs->prepare = llvm_gs_prepare; 878 gs->run = llvm_gs_run; 879 880 gs->jit_context = &draw->llvm->gs_jit_context; 881 882 883 llvm_gs->variant_key_size = 884 draw_gs_llvm_variant_key_size( 885 MAX2(gs->info.file_max[TGSI_FILE_SAMPLER]+1, 886 gs->info.file_max[TGSI_FILE_SAMPLER_VIEW]+1)); 887 } else 888#endif 889 { 890 gs->fetch_outputs = tgsi_fetch_gs_outputs; 891 gs->fetch_inputs = tgsi_fetch_gs_input; 892 gs->prepare = tgsi_gs_prepare; 893 gs->run = tgsi_gs_run; 894 } 895 896 return gs; 897} 898 899void draw_bind_geometry_shader(struct draw_context *draw, 900 struct draw_geometry_shader *dgs) 901{ 902 draw_do_flush(draw, DRAW_FLUSH_STATE_CHANGE); 903 904 if (dgs) { 905 draw->gs.geometry_shader = dgs; 906 draw->gs.num_gs_outputs = dgs->info.num_outputs; 907 draw->gs.position_output = dgs->position_output; 908 draw_geometry_shader_prepare(dgs, draw); 909 } 910 else { 911 draw->gs.geometry_shader = NULL; 912 draw->gs.num_gs_outputs = 0; 913 } 914} 915 916void draw_delete_geometry_shader(struct draw_context *draw, 917 struct draw_geometry_shader *dgs) 918{ 919 int i; 920 if (!dgs) { 921 return; 922 } 923#ifdef HAVE_LLVM 924 if (draw->llvm) { 925 struct llvm_geometry_shader *shader = llvm_geometry_shader(dgs); 926 struct draw_gs_llvm_variant_list_item *li; 927 928 li = first_elem(&shader->variants); 929 while(!at_end(&shader->variants, li)) { 930 struct draw_gs_llvm_variant_list_item *next = next_elem(li); 931 draw_gs_llvm_destroy_variant(li->base); 932 li = next; 933 } 934 935 assert(shader->variants_cached == 0); 936 937 if (dgs->llvm_prim_lengths) { 938 unsigned i; 939 for (i = 0; i < dgs->max_out_prims; ++i) { 940 align_free(dgs->llvm_prim_lengths[i]); 941 } 942 FREE(dgs->llvm_prim_lengths); 943 } 944 align_free(dgs->llvm_emitted_primitives); 945 align_free(dgs->llvm_emitted_vertices); 946 align_free(dgs->llvm_prim_ids); 947 948 align_free(dgs->gs_input); 949 } 950#endif 951 952 for (i = 0; i < TGSI_MAX_VERTEX_STREAMS; i++) 953 FREE(dgs->stream[i].primitive_lengths); 954 FREE((void*) dgs->state.tokens); 955 FREE(dgs); 956} 957 958 959#ifdef HAVE_LLVM 960void draw_gs_set_current_variant(struct draw_geometry_shader *shader, 961 struct draw_gs_llvm_variant *variant) 962{ 963 shader->current_variant = variant; 964} 965#endif 966 967/* 968 * Called at the very begin of the draw call with a new instance 969 * Used to reset state that should persist between primitive restart. 970 */ 971void 972draw_geometry_shader_new_instance(struct draw_geometry_shader *gs) 973{ 974 if (!gs) 975 return; 976 977 gs->in_prim_idx = 0; 978} 979