src/gallium/auxiliary/translate/translate_sse.c

   1 /*
   2  * Copyright 2003 Tungsten Graphics, inc.
   3  * All Rights Reserved.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * on the rights to use, copy, modify, merge, publish, distribute, sub
   9  * license, and/or sell copies of the Software, and to permit persons to whom
  10  * the Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.  IN NO EVENT SHALL
  19  * TUNGSTEN GRAPHICS AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  20  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  21  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  22  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  23  *
  24  * Authors:
  25  *    Keith Whitwell <keithw@tungstengraphics.com>
  26  */
  27
  28
  29 #include "pipe/p_config.h"
  30 #include "pipe/p_compiler.h"
  31 #include "util/u_memory.h"
  32 #include "util/u_math.h"
  33
  34 #include "translate.h"
  35
  36
  37 #if defined(PIPE_ARCH_X86)
  38
  39 #include "rtasm/rtasm_cpu.h"
  40 #include "rtasm/rtasm_x86sse.h"
  41
  42
  43 #define X    0
  44 #define Y    1
  45 #define Z    2
  46 #define W    3
  47
  48
  49 typedef void (PIPE_CDECL *run_func)( struct translate *translate,
  50                                      unsigned start,
  51                                      unsigned count,
  52                                      unsigned instance_id,
  53                                      void *output_buffer );
  54
  55 typedef void (PIPE_CDECL *run_elts_func)( struct translate *translate,
  56                                           const unsigned *elts,
  57                                           unsigned count,
  58                                           unsigned instance_id,
  59                                           void *output_buffer );
  60
  61 struct translate_buffer {
  62    const void *base_ptr;
  63    unsigned stride;
  64 };
  65
  66 struct translate_buffer_varient {
  67    unsigned buffer_index;
  68    unsigned instance_divisor;
  69    void *ptr;                    /* updated either per vertex or per instance */
  70 };
  71
  72
  73 struct translate_sse {
  74    struct translate translate;
  75
  76    struct x86_function linear_func;
  77    struct x86_function elt_func;
  78    struct x86_function *func;
  79
  80    boolean loaded_identity;
  81    boolean loaded_255;
  82    boolean loaded_inv_255;
  83
  84    float identity[4];
  85    float float_255[4];
  86    float inv_255[4];
  87
  88    struct translate_buffer buffer[PIPE_MAX_ATTRIBS];
  89    unsigned nr_buffers;
  90
  91    /* Multiple buffer varients can map to a single buffer. */
  92    struct translate_buffer_varient buffer_varient[PIPE_MAX_ATTRIBS];
  93    unsigned nr_buffer_varients;
  94
  95    /* Multiple elements can map to a single buffer varient. */
  96    unsigned element_to_buffer_varient[PIPE_MAX_ATTRIBS];
  97
  98    boolean use_instancing;
  99    unsigned instance_id;
 100
 101    run_func      gen_run;
 102    run_elts_func gen_run_elts;
 103
 104    /* these are actually known values, but putting them in a struct
 105     * like this is helpful to keep them in sync across the file.
 106     */
 107    struct x86_reg tmp_EAX;
 108    struct x86_reg idx_EBX;     /* either start+i or &elt[i] */
 109    struct x86_reg outbuf_ECX;
 110    struct x86_reg machine_EDX;
 111    struct x86_reg count_ESI;    /* decrements to zero */
 112 };
 113
 114 static int get_offset( const void *a, const void *b )
 115 {
 116    return (const char *)b - (const char *)a;
 117 }
 118
 119
 120
 121 static struct x86_reg get_identity( struct translate_sse *p )
 122 {
 123    struct x86_reg reg = x86_make_reg(file_XMM, 6);
 124
 125    if (!p->loaded_identity) {
 126       p->loaded_identity = TRUE;
 127       p->identity[0] = 0;
 128       p->identity[1] = 0;
 129       p->identity[2] = 0;
 130       p->identity[3] = 1;
 131
 132       sse_movups(p->func, reg,
 133                  x86_make_disp(p->machine_EDX,
 134                                get_offset(p, &p->identity[0])));
 135    }
 136
 137    return reg;
 138 }
 139
 140 static struct x86_reg get_255( struct translate_sse *p )
 141 {
 142    struct x86_reg reg = x86_make_reg(file_XMM, 7);
 143
 144    if (!p->loaded_255) {
 145       p->loaded_255 = TRUE;
 146       p->float_255[0] =
 147          p->float_255[1] =
 148          p->float_255[2] =
 149          p->float_255[3] = 255.0f;
 150
 151       sse_movups(p->func, reg,
 152                  x86_make_disp(p->machine_EDX,
 153                                get_offset(p, &p->float_255[0])));
 154    }
 155
 156    return reg;
 157 }
 158
 159 static struct x86_reg get_inv_255( struct translate_sse *p )
 160 {
 161    struct x86_reg reg = x86_make_reg(file_XMM, 5);
 162
 163    if (!p->loaded_inv_255) {
 164       p->loaded_inv_255 = TRUE;
 165       p->inv_255[0] =
 166          p->inv_255[1] =
 167          p->inv_255[2] =
 168          p->inv_255[3] = 1.0f / 255.0f;
 169
 170       sse_movups(p->func, reg,
 171                  x86_make_disp(p->machine_EDX,
 172                                get_offset(p, &p->inv_255[0])));
 173    }
 174
 175    return reg;
 176 }
 177
 178
 179 static void emit_load_R32G32B32A32( struct translate_sse *p,
 180                                     struct x86_reg data,
 181                                     struct x86_reg arg0 )
 182 {
 183    sse_movups(p->func, data, arg0);
 184 }
 185
 186 static void emit_load_R32G32B32( struct translate_sse *p,
 187                                  struct x86_reg data,
 188                                  struct x86_reg arg0 )
 189 {
 190    /* Have to jump through some hoops:
 191     *
 192     * c 0 0 0
 193     * c 0 0 1
 194     * 0 0 c 1
 195     * a b c 1
 196     */
 197    sse_movss(p->func, data, x86_make_disp(arg0, 8));
 198    sse_shufps(p->func, data, get_identity(p), SHUF(X,Y,Z,W) );
 199    sse_shufps(p->func, data, data, SHUF(Y,Z,X,W) );
 200    sse_movlps(p->func, data, arg0);
 201 }
 202
 203 static void emit_load_R32G32( struct translate_sse *p,
 204                            struct x86_reg data,
 205                            struct x86_reg arg0 )
 206 {
 207    /* 0 0 0 1
 208     * a b 0 1
 209     */
 210    sse_movups(p->func, data, get_identity(p) );
 211    sse_movlps(p->func, data, arg0);
 212 }
 213
 214
 215 static void emit_load_R32( struct translate_sse *p,
 216                            struct x86_reg data,
 217                            struct x86_reg arg0 )
 218 {
 219    /* a 0 0 0
 220     * a 0 0 1
 221     */
 222    sse_movss(p->func, data, arg0);
 223    sse_orps(p->func, data, get_identity(p) );
 224 }
 225
 226
 227 static void emit_load_R8G8B8A8_UNORM( struct translate_sse *p,
 228                                        struct x86_reg data,
 229                                        struct x86_reg src )
 230 {
 231
 232    /* Load and unpack twice:
 233     */
 234    sse_movss(p->func, data, src);
 235    sse2_punpcklbw(p->func, data, get_identity(p));
 236    sse2_punpcklbw(p->func, data, get_identity(p));
 237
 238    /* Convert to float:
 239     */
 240    sse2_cvtdq2ps(p->func, data, data);
 241
 242
 243    /* Scale by 1/255.0
 244     */
 245    sse_mulps(p->func, data, get_inv_255(p));
 246 }
 247
 248
 249
 250
 251 static void emit_store_R32G32B32A32( struct translate_sse *p,
 252                                      struct x86_reg dest,
 253                                      struct x86_reg dataXMM )
 254 {
 255    sse_movups(p->func, dest, dataXMM);
 256 }
 257
 258 static void emit_store_R32G32B32( struct translate_sse *p,
 259                                   struct x86_reg dest,
 260                                   struct x86_reg dataXMM )
 261 {
 262    /* Emit two, shuffle, emit one.
 263     */
 264    sse_movlps(p->func, dest, dataXMM);
 265    sse_shufps(p->func, dataXMM, dataXMM, SHUF(Z,Z,Z,Z) ); /* NOTE! destructive */
 266    sse_movss(p->func, x86_make_disp(dest,8), dataXMM);
 267 }
 268
 269 static void emit_store_R32G32( struct translate_sse *p,
 270                                struct x86_reg dest,
 271                                struct x86_reg dataXMM )
 272 {
 273    sse_movlps(p->func, dest, dataXMM);
 274 }
 275
 276 static void emit_store_R32( struct translate_sse *p,
 277                             struct x86_reg dest,
 278                             struct x86_reg dataXMM )
 279 {
 280    sse_movss(p->func, dest, dataXMM);
 281 }
 282
 283
 284
 285 static void emit_store_R8G8B8A8_UNORM( struct translate_sse *p,
 286                                        struct x86_reg dest,
 287                                        struct x86_reg dataXMM )
 288 {
 289    /* Scale by 255.0
 290     */
 291    sse_mulps(p->func, dataXMM, get_255(p));
 292
 293    /* Pack and emit:
 294     */
 295    sse2_cvtps2dq(p->func, dataXMM, dataXMM);
 296    sse2_packssdw(p->func, dataXMM, dataXMM);
 297    sse2_packuswb(p->func, dataXMM, dataXMM);
 298    sse_movss(p->func, dest, dataXMM);
 299 }
 300
 301
 302
 303
 304
 305 /* Extended swizzles?  Maybe later.
 306  */
 307 static void emit_swizzle( struct translate_sse *p,
 308                           struct x86_reg dest,
 309                           struct x86_reg src,
 310                           unsigned char shuffle )
 311 {
 312    sse_shufps(p->func, dest, src, shuffle);
 313 }
 314
 315
 316 static boolean translate_attr( struct translate_sse *p,
 317                                const struct translate_element *a,
 318                                struct x86_reg srcECX,
 319                                struct x86_reg dstEAX)
 320 {
 321    struct x86_reg dataXMM = x86_make_reg(file_XMM, 0);
 322
 323    switch (a->input_format) {
 324    case PIPE_FORMAT_R32_FLOAT:
 325       emit_load_R32(p, dataXMM, srcECX);
 326       break;
 327    case PIPE_FORMAT_R32G32_FLOAT:
 328       emit_load_R32G32(p, dataXMM, srcECX);
 329       break;
 330    case PIPE_FORMAT_R32G32B32_FLOAT:
 331       emit_load_R32G32B32(p, dataXMM, srcECX);
 332       break;
 333    case PIPE_FORMAT_R32G32B32A32_FLOAT:
 334       emit_load_R32G32B32A32(p, dataXMM, srcECX);
 335       break;
 336    case PIPE_FORMAT_B8G8R8A8_UNORM:
 337       emit_load_R8G8B8A8_UNORM(p, dataXMM, srcECX);
 338       emit_swizzle(p, dataXMM, dataXMM, SHUF(Z,Y,X,W));
 339       break;
 340    case PIPE_FORMAT_R8G8B8A8_UNORM:
 341       emit_load_R8G8B8A8_UNORM(p, dataXMM, srcECX);
 342       break;
 343    default:
 344       return FALSE;
 345    }
 346
 347    switch (a->output_format) {
 348    case PIPE_FORMAT_R32_FLOAT:
 349       emit_store_R32(p, dstEAX, dataXMM);
 350       break;
 351    case PIPE_FORMAT_R32G32_FLOAT:
 352       emit_store_R32G32(p, dstEAX, dataXMM);
 353       break;
 354    case PIPE_FORMAT_R32G32B32_FLOAT:
 355       emit_store_R32G32B32(p, dstEAX, dataXMM);
 356       break;
 357    case PIPE_FORMAT_R32G32B32A32_FLOAT:
 358       emit_store_R32G32B32A32(p, dstEAX, dataXMM);
 359       break;
 360    case PIPE_FORMAT_B8G8R8A8_UNORM:
 361       emit_swizzle(p, dataXMM, dataXMM, SHUF(Z,Y,X,W));
 362       emit_store_R8G8B8A8_UNORM(p, dstEAX, dataXMM);
 363       break;
 364    case PIPE_FORMAT_R8G8B8A8_UNORM:
 365       emit_store_R8G8B8A8_UNORM(p, dstEAX, dataXMM);
 366       break;
 367    default:
 368       return FALSE;
 369    }
 370
 371    return TRUE;
 372 }
 373
 374
 375 static boolean init_inputs( struct translate_sse *p,
 376                             boolean linear )
 377 {
 378    unsigned i;
 379    if (linear) {
 380       struct x86_reg instance_id = x86_make_disp(p->machine_EDX,
 381                                                  get_offset(p, &p->instance_id));
 382
 383       for (i = 0; i < p->nr_buffer_varients; i++) {
 384          struct translate_buffer_varient *varient = &p->buffer_varient[i];
 385          struct translate_buffer *buffer = &p->buffer[varient->buffer_index];
 386          struct x86_reg buf_stride   = x86_make_disp(p->machine_EDX,
 387                                                      get_offset(p, &buffer->stride));
 388          struct x86_reg buf_ptr      = x86_make_disp(p->machine_EDX,
 389                                                      get_offset(p, &varient->ptr));
 390          struct x86_reg buf_base_ptr = x86_make_disp(p->machine_EDX,
 391                                                      get_offset(p, &buffer->base_ptr));
 392          struct x86_reg elt = p->idx_EBX;
 393          struct x86_reg tmp_EAX = p->tmp_EAX;
 394
 395          /* Calculate pointer to first attrib:
 396           *   base_ptr + stride * index, where index depends on instance divisor
 397           */
 398          if (varient->instance_divisor) {
 399             /* Our index is instance ID divided by instance divisor.
 400              */
 401             x86_mov(p->func, tmp_EAX, instance_id);
 402
 403             if (varient->instance_divisor != 1) {
 404                struct x86_reg tmp_EDX = p->machine_EDX;
 405                struct x86_reg tmp_ECX = p->outbuf_ECX;
 406
 407                /* TODO: Add x86_shr() to rtasm and use it whenever
 408                 *       instance divisor is power of two.
 409                 */
 410
 411                x86_push(p->func, tmp_EDX);
 412                x86_push(p->func, tmp_ECX);
 413                x86_xor(p->func, tmp_EDX, tmp_EDX);
 414                x86_mov_reg_imm(p->func, tmp_ECX, varient->instance_divisor);
 415                x86_div(p->func, tmp_ECX);    /* EAX = EDX:EAX / ECX */
 416                x86_pop(p->func, tmp_ECX);
 417                x86_pop(p->func, tmp_EDX);
 418             }
 419          } else {
 420             x86_mov(p->func, tmp_EAX, elt);
 421          }
 422          x86_imul(p->func, tmp_EAX, buf_stride);
 423          x86_add(p->func, tmp_EAX, buf_base_ptr);
 424
 425
 426          /* In the linear case, keep the buffer pointer instead of the
 427           * index number.
 428           */
 429          if (p->nr_buffer_varients == 1)
 430             x86_mov(p->func, elt, tmp_EAX);
 431          else
 432             x86_mov(p->func, buf_ptr, tmp_EAX);
 433       }
 434    }
 435
 436    return TRUE;
 437 }
 438
 439
 440 static struct x86_reg get_buffer_ptr( struct translate_sse *p,
 441                                       boolean linear,
 442                                       unsigned var_idx,
 443                                       struct x86_reg elt )
 444 {
 445    if (linear && p->nr_buffer_varients == 1) {
 446       return p->idx_EBX;
 447    }
 448    else if (linear) {
 449       struct x86_reg ptr = p->tmp_EAX;
 450       struct x86_reg buf_ptr =
 451          x86_make_disp(p->machine_EDX,
 452                        get_offset(p, &p->buffer_varient[var_idx].ptr));
 453
 454       x86_mov(p->func, ptr, buf_ptr);
 455       return ptr;
 456    }
 457    else {
 458       struct x86_reg ptr = p->tmp_EAX;
 459       const struct translate_buffer_varient *varient = &p->buffer_varient[var_idx];
 460
 461       struct x86_reg buf_stride =
 462          x86_make_disp(p->machine_EDX,
 463                        get_offset(p, &p->buffer[varient->buffer_index].stride));
 464
 465       struct x86_reg buf_base_ptr =
 466          x86_make_disp(p->machine_EDX,
 467                        get_offset(p, &p->buffer[varient->buffer_index].base_ptr));
 468
 469
 470
 471       /* Calculate pointer to current attrib:
 472        */
 473       x86_mov(p->func, ptr, buf_stride);
 474       x86_imul(p->func, ptr, elt);
 475       x86_add(p->func, ptr, buf_base_ptr);
 476       return ptr;
 477    }
 478 }
 479
 480
 481
 482 static boolean incr_inputs( struct translate_sse *p,
 483                             boolean linear )
 484 {
 485    if (linear && p->nr_buffer_varients == 1) {
 486       struct x86_reg stride = x86_make_disp(p->machine_EDX,
 487                                             get_offset(p, &p->buffer[0].stride));
 488
 489       if (p->buffer_varient[0].instance_divisor == 0) {
 490          x86_add(p->func, p->idx_EBX, stride);
 491          sse_prefetchnta(p->func, x86_make_disp(p->idx_EBX, 192));
 492       }
 493    }
 494    else if (linear) {
 495       unsigned i;
 496
 497       /* Is this worthwhile??
 498        */
 499       for (i = 0; i < p->nr_buffer_varients; i++) {
 500          struct translate_buffer_varient *varient = &p->buffer_varient[i];
 501          struct x86_reg buf_ptr = x86_make_disp(p->machine_EDX,
 502                                                 get_offset(p, &varient->ptr));
 503          struct x86_reg buf_stride = x86_make_disp(p->machine_EDX,
 504                                                    get_offset(p, &p->buffer[varient->buffer_index].stride));
 505
 506          if (varient->instance_divisor == 0) {
 507             x86_mov(p->func, p->tmp_EAX, buf_ptr);
 508             x86_add(p->func, p->tmp_EAX, buf_stride);
 509             if (i == 0) sse_prefetchnta(p->func, x86_make_disp(p->tmp_EAX, 192));
 510             x86_mov(p->func, buf_ptr, p->tmp_EAX);
 511          }
 512       }
 513    }
 514    else {
 515       x86_lea(p->func, p->idx_EBX, x86_make_disp(p->idx_EBX, 4));
 516    }
 517
 518    return TRUE;
 519 }
 520
 521
 522 /* Build run( struct translate *machine,
 523  *            unsigned start,
 524  *            unsigned count,
 525  *            void *output_buffer )
 526  * or
 527  *  run_elts( struct translate *machine,
 528  *            unsigned *elts,
 529  *            unsigned count,
 530  *            void *output_buffer )
 531  *
 532  *  Lots of hardcoding
 533  *
 534  * EAX -- pointer to current output vertex
 535  * ECX -- pointer to current attribute
 536  *
 537  */
 538 static boolean build_vertex_emit( struct translate_sse *p,
 539                                   struct x86_function *func,
 540                                   boolean linear )
 541 {
 542    int fixup, label;
 543    unsigned j;
 544
 545    p->tmp_EAX       = x86_make_reg(file_REG32, reg_AX);
 546    p->idx_EBX       = x86_make_reg(file_REG32, reg_BX);
 547    p->outbuf_ECX    = x86_make_reg(file_REG32, reg_CX);
 548    p->machine_EDX   = x86_make_reg(file_REG32, reg_DX);
 549    p->count_ESI     = x86_make_reg(file_REG32, reg_SI);
 550
 551    p->func = func;
 552    p->loaded_inv_255 = FALSE;
 553    p->loaded_255 = FALSE;
 554    p->loaded_identity = FALSE;
 555
 556    x86_init_func(p->func);
 557
 558    /* Push a few regs?
 559     */
 560    x86_push(p->func, p->idx_EBX);
 561    x86_push(p->func, p->count_ESI);
 562
 563    /* Load arguments into regs:
 564     */
 565    x86_mov(p->func, p->machine_EDX, x86_fn_arg(p->func, 1));
 566    x86_mov(p->func, p->idx_EBX, x86_fn_arg(p->func, 2));
 567    x86_mov(p->func, p->count_ESI, x86_fn_arg(p->func, 3));
 568    x86_mov(p->func, p->outbuf_ECX, x86_fn_arg(p->func, 5));
 569
 570    /* Load instance ID.
 571     */
 572    if (p->use_instancing) {
 573       x86_mov(p->func,
 574               p->tmp_EAX,
 575               x86_fn_arg(p->func, 4));
 576       x86_mov(p->func,
 577               x86_make_disp(p->machine_EDX, get_offset(p, &p->instance_id)),
 578               p->tmp_EAX);
 579    }
 580
 581    /* Get vertex count, compare to zero
 582     */
 583    x86_xor(p->func, p->tmp_EAX, p->tmp_EAX);
 584    x86_cmp(p->func, p->count_ESI, p->tmp_EAX);
 585    fixup = x86_jcc_forward(p->func, cc_E);
 586
 587    /* always load, needed or not:
 588     */
 589    init_inputs(p, linear);
 590
 591    /* Note address for loop jump
 592     */
 593    label = x86_get_label(p->func);
 594    {
 595       struct x86_reg elt = linear ? p->idx_EBX : x86_deref(p->idx_EBX);
 596       int last_varient = -1;
 597       struct x86_reg vb;
 598
 599       for (j = 0; j < p->translate.key.nr_elements; j++) {
 600          const struct translate_element *a = &p->translate.key.element[j];
 601          unsigned varient = p->element_to_buffer_varient[j];
 602
 603          /* Figure out source pointer address:
 604           */
 605          if (varient != last_varient) {
 606             last_varient = varient;
 607             vb = get_buffer_ptr(p, linear, varient, elt);
 608          }
 609
 610          if (!translate_attr( p, a,
 611                               x86_make_disp(vb, a->input_offset),
 612                               x86_make_disp(p->outbuf_ECX, a->output_offset)))
 613             return FALSE;
 614       }
 615
 616       /* Next output vertex:
 617        */
 618       x86_lea(p->func,
 619               p->outbuf_ECX,
 620               x86_make_disp(p->outbuf_ECX,
 621                             p->translate.key.output_stride));
 622
 623       /* Incr index
 624        */
 625       incr_inputs( p, linear );
 626    }
 627
 628    /* decr count, loop if not zero
 629     */
 630    x86_dec(p->func, p->count_ESI);
 631    x86_jcc(p->func, cc_NZ, label);
 632
 633    /* Exit mmx state?
 634     */
 635    if (p->func->need_emms)
 636       mmx_emms(p->func);
 637
 638    /* Land forward jump here:
 639     */
 640    x86_fixup_fwd_jump(p->func, fixup);
 641
 642    /* Pop regs and return
 643     */
 644
 645    x86_pop(p->func, p->count_ESI);
 646    x86_pop(p->func, p->idx_EBX);
 647    x86_ret(p->func);
 648
 649    return TRUE;
 650 }
 651
 652
 653
 654
 655
 656
 657
 658 static void translate_sse_set_buffer( struct translate *translate,
 659                                 unsigned buf,
 660                                 const void *ptr,
 661                                 unsigned stride )
 662 {
 663    struct translate_sse *p = (struct translate_sse *)translate;
 664
 665    if (buf < p->nr_buffers) {
 666       p->buffer[buf].base_ptr = (char *)ptr;
 667       p->buffer[buf].stride = stride;
 668    }
 669
 670    if (0) debug_printf("%s %d/%d: %p %d\n",
 671                        __FUNCTION__, buf,
 672                        p->nr_buffers,
 673                        ptr, stride);
 674 }
 675
 676
 677 static void translate_sse_release( struct translate *translate )
 678 {
 679    struct translate_sse *p = (struct translate_sse *)translate;
 680
 681    x86_release_func( &p->linear_func );
 682    x86_release_func( &p->elt_func );
 683
 684    FREE(p);
 685 }
 686
 687 static void PIPE_CDECL translate_sse_run_elts( struct translate *translate,
 688                               const unsigned *elts,
 689                               unsigned count,
 690                               void *output_buffer )
 691 {
 692    struct translate_sse *p = (struct translate_sse *)translate;
 693
 694    p->gen_run_elts( translate,
 695                     elts,
 696                     count,
 697                     0,
 698                     output_buffer );
 699 }
 700
 701 static void PIPE_CDECL translate_sse_run( struct translate *translate,
 702                          unsigned start,
 703                          unsigned count,
 704                          unsigned instance_id,
 705                          void *output_buffer )
 706 {
 707    struct translate_sse *p = (struct translate_sse *)translate;
 708
 709    p->gen_run( translate,
 710                start,
 711                count,
 712                instance_id,
 713                output_buffer );
 714 }
 715
 716
 717 struct translate *translate_sse2_create( const struct translate_key *key )
 718 {
 719    struct translate_sse *p = NULL;
 720    unsigned i;
 721
 722    if (!rtasm_cpu_has_sse() || !rtasm_cpu_has_sse2())
 723       goto fail;
 724
 725    p = CALLOC_STRUCT( translate_sse );
 726    if (p == NULL)
 727       goto fail;
 728
 729    p->translate.key = *key;
 730    p->translate.release = translate_sse_release;
 731    p->translate.set_buffer = translate_sse_set_buffer;
 732    p->translate.run_elts = translate_sse_run_elts;
 733    p->translate.run = translate_sse_run;
 734
 735    for (i = 0; i < key->nr_elements; i++) {
 736       unsigned j;
 737
 738       p->nr_buffers = MAX2( p->nr_buffers, key->element[i].input_buffer + 1 );
 739
 740       if (key->element[i].instance_divisor) {
 741          p->use_instancing = TRUE;
 742       }
 743
 744       /*
 745        * Map vertex element to vertex buffer varient.
 746        */
 747       for (j = 0; j < p->nr_buffer_varients; j++) {
 748          if (p->buffer_varient[j].buffer_index == key->element[i].input_buffer &&
 749              p->buffer_varient[j].instance_divisor == key->element[i].instance_divisor) {
 750             break;
 751          }
 752       }
 753       if (j == p->nr_buffer_varients) {
 754          p->buffer_varient[j].buffer_index = key->element[i].input_buffer;
 755          p->buffer_varient[j].instance_divisor = key->element[i].instance_divisor;
 756          p->nr_buffer_varients++;
 757       }
 758       p->element_to_buffer_varient[i] = j;
 759    }
 760
 761    if (0) debug_printf("nr_buffers: %d\n", p->nr_buffers);
 762
 763    if (!build_vertex_emit(p, &p->linear_func, TRUE))
 764       goto fail;
 765
 766    if (!build_vertex_emit(p, &p->elt_func, FALSE))
 767       goto fail;
 768
 769    p->gen_run = (run_func)x86_get_func(&p->linear_func);
 770    if (p->gen_run == NULL)
 771       goto fail;
 772
 773    p->gen_run_elts = (run_elts_func)x86_get_func(&p->elt_func);
 774    if (p->gen_run_elts == NULL)
 775       goto fail;
 776
 777    return &p->translate;
 778
 779  fail:
 780    if (p)
 781       translate_sse_release( &p->translate );
 782
 783    return NULL;
 784 }
 785
 786
 787
 788 #else
 789
 790 struct translate *translate_sse2_create( const struct translate_key *key )
 791 {
 792    return NULL;
 793 }
 794
 795 #endif