src/gallium/auxiliary/draw/draw_gs.c

   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 TUNGSTEN GRAPHICS 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
  48 static INLINE int
  49 draw_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    debug_assert(0);
  61    return -1;
  62 }
  63
  64 /**
  65  * We execute geometry shaders in the SOA mode, so ideally we want to
  66  * flush when the number of currently fetched primitives is equal to
  67  * the number of elements in the SOA vector. This ensures that the
  68  * throughput is optimized for the given vector instrunction set.
  69  */
  70 static INLINE boolean
  71 draw_gs_should_flush(struct draw_geometry_shader *shader)
  72 {
  73    return (shader->fetched_prim_count == shader->vector_length);
  74 }
  75
  76 /*#define DEBUG_OUTPUTS 1*/
  77 static void
  78 tgsi_fetch_gs_outputs(struct draw_geometry_shader *shader,
  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    unsigned current_idx = 0;
  85    float (*output)[4];
  86
  87    output = *p_output;
  88
  89    /* Unswizzle all output results.
  90     */
  91
  92    for (prim_idx = 0; prim_idx < num_primitives; ++prim_idx) {
  93       unsigned num_verts_per_prim = machine->Primitives[prim_idx];
  94       shader->primitive_lengths[prim_idx +   shader->emitted_primitives] =
  95          machine->Primitives[prim_idx];
  96       shader->emitted_vertices += num_verts_per_prim;
  97       for (j = 0; j < num_verts_per_prim; j++, current_idx++) {
  98          int idx = current_idx * shader->info.num_outputs;
  99 #ifdef DEBUG_OUTPUTS
 100          debug_printf("%d) Output vert:\n", idx / shader->info.num_outputs);
 101 #endif
 102          for (slot = 0; slot < shader->info.num_outputs; slot++) {
 103             output[slot][0] = machine->Outputs[idx + slot].xyzw[0].f[0];
 104             output[slot][1] = machine->Outputs[idx + slot].xyzw[1].f[0];
 105             output[slot][2] = machine->Outputs[idx + slot].xyzw[2].f[0];
 106             output[slot][3] = machine->Outputs[idx + slot].xyzw[3].f[0];
 107 #ifdef DEBUG_OUTPUTS
 108             debug_printf("\t%d: %f %f %f %f\n", slot,
 109                          output[slot][0],
 110                          output[slot][1],
 111                          output[slot][2],
 112                          output[slot][3]);
 113 #endif
 114             debug_assert(!util_is_inf_or_nan(output[slot][0]));
 115          }
 116          output = (float (*)[4])((char *)output + shader->vertex_size);
 117       }
 118    }
 119    *p_output = output;
 120    shader->emitted_primitives += num_primitives;
 121 }
 122
 123 /*#define DEBUG_INPUTS 1*/
 124 static 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, vs_slot, i;
 131    unsigned input_vertex_stride = shader->input_vertex_stride;
 132    const float (*input_ptr)[4];
 133
 134    input_ptr = shader->input;
 135
 136    for (i = 0; i < num_vertices; ++i) {
 137       const float (*input)[4];
 138 #if DEBUG_INPUTS
 139       debug_printf("%d) vertex index = %d (prim idx = %d)\n",
 140                    i, indices[i], prim_idx);
 141 #endif
 142       input = (const float (*)[4])(
 143          (const char *)input_ptr + (indices[i] * input_vertex_stride));
 144       for (slot = 0, vs_slot = 0; slot < shader->info.num_inputs; ++slot) {
 145          unsigned idx = i * TGSI_EXEC_MAX_INPUT_ATTRIBS + slot;
 146          if (shader->info.input_semantic_name[slot] == TGSI_SEMANTIC_PRIMID) {
 147             machine->Inputs[idx].xyzw[0].f[prim_idx] =
 148                (float)shader->in_prim_idx;
 149             machine->Inputs[idx].xyzw[1].f[prim_idx] =
 150                (float)shader->in_prim_idx;
 151             machine->Inputs[idx].xyzw[2].f[prim_idx] =
 152                (float)shader->in_prim_idx;
 153             machine->Inputs[idx].xyzw[3].f[prim_idx] =
 154                (float)shader->in_prim_idx;
 155          } else {
 156             vs_slot = draw_gs_get_input_index(
 157                         shader->info.input_semantic_name[slot],
 158                         shader->info.input_semantic_index[slot],
 159                         shader->input_info);
 160 #if DEBUG_INPUTS
 161             debug_printf("\tSlot = %d, vs_slot = %d, idx = %d:\n",
 162                          slot, vs_slot, idx);
 163             assert(!util_is_inf_or_nan(input[vs_slot][0]));
 164             assert(!util_is_inf_or_nan(input[vs_slot][1]));
 165             assert(!util_is_inf_or_nan(input[vs_slot][2]));
 166             assert(!util_is_inf_or_nan(input[vs_slot][3]));
 167 #endif
 168             machine->Inputs[idx].xyzw[0].f[prim_idx] = input[vs_slot][0];
 169             machine->Inputs[idx].xyzw[1].f[prim_idx] = input[vs_slot][1];
 170             machine->Inputs[idx].xyzw[2].f[prim_idx] = input[vs_slot][2];
 171             machine->Inputs[idx].xyzw[3].f[prim_idx] = input[vs_slot][3];
 172 #if DEBUG_INPUTS
 173             debug_printf("\t\t%f %f %f %f\n",
 174                          machine->Inputs[idx].xyzw[0].f[prim_idx],
 175                          machine->Inputs[idx].xyzw[1].f[prim_idx],
 176                          machine->Inputs[idx].xyzw[2].f[prim_idx],
 177                          machine->Inputs[idx].xyzw[3].f[prim_idx]);
 178 #endif
 179             ++vs_slot;
 180          }
 181       }
 182    }
 183 }
 184
 185 static void tgsi_gs_prepare(struct draw_geometry_shader *shader,
 186                             const void *constants[PIPE_MAX_CONSTANT_BUFFERS],
 187                             const unsigned constants_size[PIPE_MAX_CONSTANT_BUFFERS])
 188 {
 189    struct tgsi_exec_machine *machine = shader->machine;
 190
 191    tgsi_exec_set_constant_buffers(machine, PIPE_MAX_CONSTANT_BUFFERS,
 192                                   constants, constants_size);
 193 }
 194
 195 static unsigned tgsi_gs_run(struct draw_geometry_shader *shader,
 196                             unsigned input_primitives)
 197 {
 198    struct tgsi_exec_machine *machine = shader->machine;
 199
 200    tgsi_set_exec_mask(machine,
 201                       1,
 202                       input_primitives > 1,
 203                       input_primitives > 2,
 204                       input_primitives > 3);
 205
 206    /* run interpreter */
 207    tgsi_exec_machine_run(machine);
 208
 209    return
 210       machine->Temps[TGSI_EXEC_TEMP_PRIMITIVE_I].xyzw[TGSI_EXEC_TEMP_PRIMITIVE_C].u[0];
 211 }
 212
 213 #ifdef HAVE_LLVM
 214
 215 static void
 216 llvm_fetch_gs_input(struct draw_geometry_shader *shader,
 217                     unsigned *indices,
 218                     unsigned num_vertices,
 219                     unsigned prim_idx)
 220 {
 221    unsigned slot, vs_slot, i;
 222    unsigned input_vertex_stride = shader->input_vertex_stride;
 223    const float (*input_ptr)[4];
 224    float (*input_data)[6][PIPE_MAX_SHADER_INPUTS][TGSI_NUM_CHANNELS][TGSI_NUM_CHANNELS] = &shader->gs_input->data;
 225
 226    shader->llvm_prim_ids[shader->fetched_prim_count] =
 227       shader->in_prim_idx;
 228
 229    input_ptr = shader->input;
 230
 231    for (i = 0; i < num_vertices; ++i) {
 232       const float (*input)[4];
 233 #if DEBUG_INPUTS
 234       debug_printf("%d) vertex index = %d (prim idx = %d)\n",
 235                    i, indices[i], prim_idx);
 236 #endif
 237       input = (const float (*)[4])(
 238          (const char *)input_ptr + (indices[i] * input_vertex_stride));
 239       for (slot = 0, vs_slot = 0; slot < shader->info.num_inputs; ++slot) {
 240          if (shader->info.input_semantic_name[slot] == TGSI_SEMANTIC_PRIMID) {
 241             /* skip. we handle system values through gallivm */
 242          } else {
 243             vs_slot = draw_gs_get_input_index(
 244                         shader->info.input_semantic_name[slot],
 245                         shader->info.input_semantic_index[slot],
 246                         shader->input_info);
 247 #if DEBUG_INPUTS
 248             debug_printf("\tSlot = %d, vs_slot = %d, i = %d:\n",
 249                          slot, vs_slot, i);
 250             assert(!util_is_inf_or_nan(input[vs_slot][0]));
 251             assert(!util_is_inf_or_nan(input[vs_slot][1]));
 252             assert(!util_is_inf_or_nan(input[vs_slot][2]));
 253             assert(!util_is_inf_or_nan(input[vs_slot][3]));
 254 #endif
 255             (*input_data)[i][slot][0][prim_idx] = input[vs_slot][0];
 256             (*input_data)[i][slot][1][prim_idx] = input[vs_slot][1];
 257             (*input_data)[i][slot][2][prim_idx] = input[vs_slot][2];
 258             (*input_data)[i][slot][3][prim_idx] = input[vs_slot][3];
 259 #if DEBUG_INPUTS
 260             debug_printf("\t\t%f %f %f %f\n",
 261                          (*input_data)[i][slot][0][prim_idx],
 262                          (*input_data)[i][slot][1][prim_idx],
 263                          (*input_data)[i][slot][2][prim_idx],
 264                          (*input_data)[i][slot][3][prim_idx]);
 265 #endif
 266             ++vs_slot;
 267          }
 268       }
 269    }
 270 }
 271
 272 static void
 273 llvm_fetch_gs_outputs(struct draw_geometry_shader *shader,
 274                       unsigned num_primitives,
 275                       float (**p_output)[4])
 276 {
 277    int total_verts = 0;
 278    int vertex_count = 0;
 279    int total_prims = 0;
 280    int max_prims_per_invocation = 0;
 281    char *output_ptr = (char*)shader->gs_output;
 282    int i, j, prim_idx;
 283
 284    for (i = 0; i < shader->vector_length; ++i) {
 285       int prims = shader->llvm_emitted_primitives[i];
 286       total_prims += prims;
 287       max_prims_per_invocation = MAX2(max_prims_per_invocation, prims);
 288    }
 289    for (i = 0; i < shader->vector_length; ++i) {
 290       total_verts += shader->llvm_emitted_vertices[i];
 291    }
 292
 293
 294    output_ptr += shader->emitted_vertices * shader->vertex_size;
 295    for (i = 0; i < shader->vector_length - 1; ++i) {
 296       int current_verts = shader->llvm_emitted_vertices[i];
 297
 298       if (current_verts != shader->max_output_vertices) {
 299          memcpy(output_ptr + (vertex_count + current_verts) * shader->vertex_size,
 300                 output_ptr + (vertex_count + shader->max_output_vertices) * shader->vertex_size,
 301                 shader->vertex_size * (total_verts - vertex_count));
 302       }
 303       vertex_count += current_verts;
 304    }
 305
 306    prim_idx = 0;
 307    for (i = 0; i < shader->vector_length; ++i) {
 308       int num_prims = shader->llvm_emitted_primitives[i];
 309       for (j = 0; j < num_prims; ++j) {
 310          int prim_length =
 311             shader->llvm_prim_lengths[j][i];
 312          shader->primitive_lengths[shader->emitted_primitives + prim_idx] =
 313             prim_length;
 314          ++prim_idx;
 315       }
 316    }
 317
 318    shader->emitted_primitives += total_prims;
 319    shader->emitted_vertices += total_verts;
 320 }
 321
 322 static void
 323 llvm_gs_prepare(struct draw_geometry_shader *shader,
 324                 const void *constants[PIPE_MAX_CONSTANT_BUFFERS],
 325                 const unsigned constants_size[PIPE_MAX_CONSTANT_BUFFERS])
 326 {
 327 }
 328
 329 static unsigned
 330 llvm_gs_run(struct draw_geometry_shader *shader,
 331             unsigned input_primitives)
 332 {
 333    unsigned ret;
 334    char *input = (char*)shader->gs_output;
 335
 336    input += (shader->emitted_vertices * shader->vertex_size);
 337
 338    ret = shader->current_variant->jit_func(
 339       shader->jit_context, shader->gs_input->data,
 340       (struct vertex_header*)input,
 341       input_primitives,
 342       shader->draw->instance_id,
 343       shader->llvm_prim_ids);
 344
 345    return ret;
 346 }
 347
 348 #endif
 349
 350 static void gs_flush(struct draw_geometry_shader *shader)
 351 {
 352    unsigned out_prim_count;
 353
 354    unsigned input_primitives = shader->fetched_prim_count;
 355
 356    debug_assert(input_primitives > 0 &&
 357                 input_primitives <= 4);
 358
 359    out_prim_count = shader->run(shader, input_primitives);
 360    shader->fetch_outputs(shader, out_prim_count,
 361                          &shader->tmp_output);
 362
 363 #if 0
 364    debug_printf("PRIM emitted prims = %d (verts=%d), cur prim count = %d\n",
 365                 shader->emitted_primitives, shader->emitted_vertices,
 366                 out_prim_count);
 367 #endif
 368
 369    shader->fetched_prim_count = 0;
 370 }
 371
 372 static void gs_point(struct draw_geometry_shader *shader,
 373                      int idx)
 374 {
 375    unsigned indices[1];
 376
 377    indices[0] = idx;
 378
 379    shader->fetch_inputs(shader, indices, 1,
 380                         shader->fetched_prim_count);
 381    ++shader->in_prim_idx;
 382    ++shader->fetched_prim_count;
 383
 384    if (draw_gs_should_flush(shader))
 385       gs_flush(shader);
 386 }
 387
 388 static void gs_line(struct draw_geometry_shader *shader,
 389                     int i0, int i1)
 390 {
 391    unsigned indices[2];
 392
 393    indices[0] = i0;
 394    indices[1] = i1;
 395
 396    shader->fetch_inputs(shader, indices, 2,
 397                         shader->fetched_prim_count);
 398    ++shader->in_prim_idx;
 399    ++shader->fetched_prim_count;
 400
 401    if (draw_gs_should_flush(shader))
 402       gs_flush(shader);
 403 }
 404
 405 static void gs_line_adj(struct draw_geometry_shader *shader,
 406                         int i0, int i1, int i2, int i3)
 407 {
 408    unsigned indices[4];
 409
 410    indices[0] = i0;
 411    indices[1] = i1;
 412    indices[2] = i2;
 413    indices[3] = i3;
 414
 415    shader->fetch_inputs(shader, indices, 4,
 416                         shader->fetched_prim_count);
 417    ++shader->in_prim_idx;
 418    ++shader->fetched_prim_count;
 419
 420    if (draw_gs_should_flush(shader))
 421       gs_flush(shader);
 422 }
 423
 424 static void gs_tri(struct draw_geometry_shader *shader,
 425                    int i0, int i1, int i2)
 426 {
 427    unsigned indices[3];
 428
 429    indices[0] = i0;
 430    indices[1] = i1;
 431    indices[2] = i2;
 432
 433    shader->fetch_inputs(shader, indices, 3,
 434                         shader->fetched_prim_count);
 435    ++shader->in_prim_idx;
 436    ++shader->fetched_prim_count;
 437
 438    if (draw_gs_should_flush(shader))
 439       gs_flush(shader);
 440 }
 441
 442 static void gs_tri_adj(struct draw_geometry_shader *shader,
 443                        int i0, int i1, int i2,
 444                        int i3, int i4, int i5)
 445 {
 446    unsigned indices[6];
 447
 448    indices[0] = i0;
 449    indices[1] = i1;
 450    indices[2] = i2;
 451    indices[3] = i3;
 452    indices[4] = i4;
 453    indices[5] = i5;
 454
 455    shader->fetch_inputs(shader, indices, 6,
 456                         shader->fetched_prim_count);
 457    ++shader->in_prim_idx;
 458    ++shader->fetched_prim_count;
 459
 460    if (draw_gs_should_flush(shader))
 461       gs_flush(shader);
 462 }
 463
 464 #define FUNC         gs_run
 465 #define GET_ELT(idx) (idx)
 466 #include "draw_gs_tmp.h"
 467
 468
 469 #define FUNC         gs_run_elts
 470 #define LOCAL_VARS   const ushort *elts = input_prims->elts;
 471 #define GET_ELT(idx) (elts[idx])
 472 #include "draw_gs_tmp.h"
 473
 474
 475 /**
 476  * Execute geometry shader.
 477  */
 478 int draw_geometry_shader_run(struct draw_geometry_shader *shader,
 479                              const void *constants[PIPE_MAX_CONSTANT_BUFFERS],
 480                              const unsigned constants_size[PIPE_MAX_CONSTANT_BUFFERS],
 481                              const struct draw_vertex_info *input_verts,
 482                              const struct draw_prim_info *input_prim,
 483                              const struct tgsi_shader_info *input_info,
 484                              struct draw_vertex_info *output_verts,
 485                              struct draw_prim_info *output_prims )
 486 {
 487    const float (*input)[4] = (const float (*)[4])input_verts->verts->data;
 488    unsigned input_stride = input_verts->vertex_size;
 489    unsigned num_outputs = shader->info.num_outputs;
 490    unsigned vertex_size = sizeof(struct vertex_header) + num_outputs * 4 * sizeof(float);
 491    unsigned num_input_verts = input_prim->linear ?
 492       input_verts->count :
 493       input_prim->count;
 494    unsigned num_in_primitives =
 495       align(
 496          MAX2(u_gs_prims_for_vertices(input_prim->prim, num_input_verts),
 497               u_gs_prims_for_vertices(shader->input_primitive, num_input_verts)),
 498          shader->vector_length);
 499    unsigned max_out_prims = u_gs_prims_for_vertices(shader->output_primitive,
 500                                                     shader->max_output_vertices)
 501       * num_in_primitives;
 502
 503    //Assume at least one primitive
 504    max_out_prims = MAX2(max_out_prims, 1);
 505
 506
 507    output_verts->vertex_size = vertex_size;
 508    output_verts->stride = output_verts->vertex_size;
 509    output_verts->verts =
 510       (struct vertex_header *)MALLOC(output_verts->vertex_size *
 511                                      max_out_prims *
 512                                      shader->max_output_vertices);
 513
 514 #if 0
 515    debug_printf("%s count = %d (in prims # = %d)\n",
 516                 __FUNCTION__, num_input_verts, num_in_primitives);
 517    debug_printf("\tlinear = %d, prim_info->count = %d\n",
 518                 input_prim->linear, input_prim->count);
 519    debug_printf("\tprim pipe = %s, shader in = %s, shader out = %s, max out = %d\n",
 520                 u_prim_name(input_prim->prim),
 521                 u_prim_name(shader->input_primitive),
 522                 u_prim_name(shader->output_primitive),
 523                 shader->max_output_vertices);
 524 #endif
 525
 526    shader->emitted_vertices = 0;
 527    shader->emitted_primitives = 0;
 528    shader->vertex_size = vertex_size;
 529    shader->tmp_output = (float (*)[4])output_verts->verts->data;
 530    shader->in_prim_idx = 0;
 531    shader->fetched_prim_count = 0;
 532    shader->input_vertex_stride = input_stride;
 533    shader->input = input;
 534    shader->input_info = input_info;
 535    FREE(shader->primitive_lengths);
 536    shader->primitive_lengths = MALLOC(max_out_prims * sizeof(unsigned));
 537
 538
 539 #ifdef HAVE_LLVM
 540    if (draw_get_option_use_llvm()) {
 541       shader->gs_output = output_verts->verts;
 542       if (max_out_prims > shader->max_out_prims) {
 543          unsigned i;
 544          if (shader->llvm_prim_lengths) {
 545             for (i = 0; i < shader->max_out_prims; ++i) {
 546                align_free(shader->llvm_prim_lengths[i]);
 547             }
 548             FREE(shader->llvm_prim_lengths);
 549          }
 550
 551          shader->llvm_prim_lengths = MALLOC(max_out_prims * sizeof(unsigned*));
 552          for (i = 0; i < max_out_prims; ++i) {
 553             int vector_size = shader->vector_length * sizeof(unsigned);
 554             shader->llvm_prim_lengths[i] =
 555                align_malloc(vector_size, vector_size);
 556          }
 557
 558          shader->max_out_prims = max_out_prims;
 559       }
 560       shader->jit_context->prim_lengths = shader->llvm_prim_lengths;
 561       shader->jit_context->emitted_vertices = shader->llvm_emitted_vertices;
 562       shader->jit_context->emitted_prims = shader->llvm_emitted_primitives;
 563    }
 564 #endif
 565
 566    shader->prepare(shader, constants, constants_size);
 567
 568    if (input_prim->linear)
 569       gs_run(shader, input_prim, input_verts,
 570              output_prims, output_verts);
 571    else
 572       gs_run_elts(shader, input_prim, input_verts,
 573                   output_prims, output_verts);
 574
 575    /* Flush the remaining primitives. Will happen if
 576     * num_input_primitives % 4 != 0
 577     */
 578    if (shader->fetched_prim_count > 0) {
 579       gs_flush(shader);
 580    }
 581
 582    debug_assert(shader->fetched_prim_count == 0);
 583
 584    /* Update prim_info:
 585     */
 586    output_prims->linear = TRUE;
 587    output_prims->elts = NULL;
 588    output_prims->start = 0;
 589    output_prims->count = shader->emitted_vertices;
 590    output_prims->prim = shader->output_primitive;
 591    output_prims->flags = 0x0;
 592    output_prims->primitive_lengths = shader->primitive_lengths;
 593    output_prims->primitive_count = shader->emitted_primitives;
 594    output_verts->count = shader->emitted_vertices;
 595
 596 #if 0
 597    debug_printf("GS finished, prims = %d, verts = %d\n",
 598                 output_prims->primitive_count,
 599                 output_verts->count);
 600 #endif
 601
 602    return shader->emitted_vertices;
 603 }
 604
 605 void draw_geometry_shader_prepare(struct draw_geometry_shader *shader,
 606                                   struct draw_context *draw)
 607 {
 608    if (shader && shader->machine->Tokens != shader->state.tokens) {
 609       tgsi_exec_machine_bind_shader(shader->machine,
 610                                     shader->state.tokens,
 611                                     draw->gs.tgsi.sampler);
 612    }
 613 }
 614
 615
 616 boolean
 617 draw_gs_init( struct draw_context *draw )
 618 {
 619    draw->gs.tgsi.machine = tgsi_exec_machine_create();
 620    if (!draw->gs.tgsi.machine)
 621       return FALSE;
 622
 623    draw->gs.tgsi.machine->Primitives = align_malloc(
 624       MAX_PRIMITIVES * sizeof(struct tgsi_exec_vector), 16);
 625    if (!draw->gs.tgsi.machine->Primitives)
 626       return FALSE;
 627    memset(draw->gs.tgsi.machine->Primitives, 0,
 628           MAX_PRIMITIVES * sizeof(struct tgsi_exec_vector));
 629
 630    return TRUE;
 631 }
 632
 633 void draw_gs_destroy( struct draw_context *draw )
 634 {
 635    if (draw->gs.tgsi.machine) {
 636       align_free(draw->gs.tgsi.machine->Primitives);
 637       tgsi_exec_machine_destroy(draw->gs.tgsi.machine);
 638    }
 639 }
 640
 641 struct draw_geometry_shader *
 642 draw_create_geometry_shader(struct draw_context *draw,
 643                             const struct pipe_shader_state *state)
 644 {
 645 #ifdef HAVE_LLVM
 646    boolean use_llvm = draw_get_option_use_llvm();
 647    struct llvm_geometry_shader *llvm_gs;
 648 #endif
 649    struct draw_geometry_shader *gs;
 650    unsigned i;
 651
 652 #ifdef HAVE_LLVM
 653    if (use_llvm) {
 654       llvm_gs = CALLOC_STRUCT(llvm_geometry_shader);
 655
 656       if (llvm_gs == NULL)
 657          return NULL;
 658
 659       gs = &llvm_gs->base;
 660
 661       make_empty_list(&llvm_gs->variants);
 662    } else
 663 #endif
 664    {
 665       gs = CALLOC_STRUCT(draw_geometry_shader);
 666    }
 667
 668    if (!gs)
 669       return NULL;
 670
 671    gs->draw = draw;
 672    gs->state = *state;
 673    gs->state.tokens = tgsi_dup_tokens(state->tokens);
 674    if (!gs->state.tokens) {
 675       FREE(gs);
 676       return NULL;
 677    }
 678
 679    tgsi_scan_shader(state->tokens, &gs->info);
 680
 681    /* setup the defaults */
 682    gs->input_primitive = PIPE_PRIM_TRIANGLES;
 683    gs->output_primitive = PIPE_PRIM_TRIANGLE_STRIP;
 684    gs->max_output_vertices = 32;
 685    gs->max_out_prims = 0;
 686
 687 #ifdef HAVE_LLVM
 688    if (use_llvm) {
 689       /* TODO: change the input array to handle the following
 690          vector length, instead of the currently hardcoded
 691          TGSI_NUM_CHANNELS
 692       gs->vector_length = lp_native_vector_width / 32;*/
 693       gs->vector_length = TGSI_NUM_CHANNELS;
 694    } else
 695 #endif
 696    {
 697       gs->vector_length = 1;
 698    }
 699
 700    for (i = 0; i < gs->info.num_properties; ++i) {
 701       if (gs->info.properties[i].name ==
 702           TGSI_PROPERTY_GS_INPUT_PRIM)
 703          gs->input_primitive = gs->info.properties[i].data[0];
 704       else if (gs->info.properties[i].name ==
 705                TGSI_PROPERTY_GS_OUTPUT_PRIM)
 706          gs->output_primitive = gs->info.properties[i].data[0];
 707       else if (gs->info.properties[i].name ==
 708                TGSI_PROPERTY_GS_MAX_OUTPUT_VERTICES)
 709          gs->max_output_vertices = gs->info.properties[i].data[0];
 710    }
 711
 712    for (i = 0; i < gs->info.num_outputs; i++) {
 713       if (gs->info.output_semantic_name[i] == TGSI_SEMANTIC_POSITION &&
 714           gs->info.output_semantic_index[i] == 0)
 715          gs->position_output = i;
 716    }
 717
 718    gs->machine = draw->gs.tgsi.machine;
 719
 720 #ifdef HAVE_LLVM
 721    if (use_llvm) {
 722       int vector_size = gs->vector_length * sizeof(float);
 723       gs->gs_input = align_malloc(sizeof(struct draw_gs_inputs), 16);
 724       memset(gs->gs_input, 0, sizeof(struct draw_gs_inputs));
 725       gs->llvm_prim_lengths = 0;
 726
 727       gs->llvm_emitted_primitives = align_malloc(vector_size, vector_size);
 728       gs->llvm_emitted_vertices = align_malloc(vector_size, vector_size);
 729       gs->llvm_prim_ids = align_malloc(vector_size, vector_size);
 730
 731       gs->fetch_outputs = llvm_fetch_gs_outputs;
 732       gs->fetch_inputs = llvm_fetch_gs_input;
 733       gs->prepare = llvm_gs_prepare;
 734       gs->run = llvm_gs_run;
 735
 736       gs->jit_context = &draw->llvm->gs_jit_context;
 737
 738
 739       llvm_gs->variant_key_size =
 740          draw_gs_llvm_variant_key_size(
 741             MAX2(gs->info.file_max[TGSI_FILE_SAMPLER]+1,
 742                  gs->info.file_max[TGSI_FILE_SAMPLER_VIEW]+1));
 743    } else
 744 #endif
 745    {
 746       gs->fetch_outputs = tgsi_fetch_gs_outputs;
 747       gs->fetch_inputs = tgsi_fetch_gs_input;
 748       gs->prepare = tgsi_gs_prepare;
 749       gs->run = tgsi_gs_run;
 750    }
 751
 752    return gs;
 753 }
 754
 755 void draw_bind_geometry_shader(struct draw_context *draw,
 756                                struct draw_geometry_shader *dgs)
 757 {
 758    draw_do_flush(draw, DRAW_FLUSH_STATE_CHANGE);
 759
 760    if (dgs) {
 761       draw->gs.geometry_shader = dgs;
 762       draw->gs.num_gs_outputs = dgs->info.num_outputs;
 763       draw->gs.position_output = dgs->position_output;
 764       draw_geometry_shader_prepare(dgs, draw);
 765    }
 766    else {
 767       draw->gs.geometry_shader = NULL;
 768       draw->gs.num_gs_outputs = 0;
 769    }
 770 }
 771
 772 void draw_delete_geometry_shader(struct draw_context *draw,
 773                                  struct draw_geometry_shader *dgs)
 774 {
 775 #ifdef HAVE_LLVM
 776    if (draw_get_option_use_llvm()) {
 777       struct llvm_geometry_shader *shader = llvm_geometry_shader(dgs);
 778       struct draw_gs_llvm_variant_list_item *li;
 779
 780       li = first_elem(&shader->variants);
 781       while(!at_end(&shader->variants, li)) {
 782          struct draw_gs_llvm_variant_list_item *next = next_elem(li);
 783          draw_gs_llvm_destroy_variant(li->base);
 784          li = next;
 785       }
 786
 787       assert(shader->variants_cached == 0);
 788
 789       if (dgs->llvm_prim_lengths) {
 790          unsigned i;
 791          for (i = 0; i < dgs->max_out_prims; ++i) {
 792             align_free(dgs->llvm_prim_lengths[i]);
 793          }
 794          FREE(dgs->llvm_prim_lengths);
 795       }
 796       align_free(dgs->llvm_emitted_primitives);
 797       align_free(dgs->llvm_emitted_vertices);
 798       align_free(dgs->llvm_prim_ids);
 799
 800       align_free(dgs->gs_input);
 801    }
 802 #endif
 803
 804    FREE(dgs->primitive_lengths);
 805    FREE((void*) dgs->state.tokens);
 806    FREE(dgs);
 807 }
 808
 809
 810 #ifdef HAVE_LLVM
 811 void draw_gs_set_current_variant(struct draw_geometry_shader *shader,
 812                                  struct draw_gs_llvm_variant *variant)
 813 {
 814    shader->current_variant = variant;
 815 }
 816 #endif