src/gallium/auxiliary/draw/draw_vs_aos_io.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
   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
  29 #include "util/u_memory.h"
  30 #include "pipe/p_shader_tokens.h"
  31 #include "tgsi/tgsi_parse.h"
  32 #include "tgsi/tgsi_util.h"
  33 #include "tgsi/tgsi_exec.h"
  34 #include "draw_vs.h"
  35 #include "draw_vs_aos.h"
  36 #include "draw_vertex.h"
  37
  38 #include "rtasm/rtasm_x86sse.h"
  39
  40 #ifdef PIPE_ARCH_X86
  41
  42 /* Note - don't yet have to worry about interacting with the code in
  43  * draw_vs_aos.c as there is no intermingling of generated code...
  44  * That may have to change, we'll see.
  45  */
  46 static void emit_load_R32G32B32A32( struct aos_compilation *cp,
  47                                     struct x86_reg data,
  48                                     struct x86_reg src_ptr )
  49 {
  50    sse_movups(cp->func, data, src_ptr);
  51 }
  52
  53 static void emit_load_R32G32B32( struct aos_compilation *cp,
  54                                  struct x86_reg data,
  55                                  struct x86_reg src_ptr )
  56 {
  57 #if 1
  58    sse_movss(cp->func, data, x86_make_disp(src_ptr, 8));
  59    /* data = z ? ? ? */
  60    sse_shufps(cp->func, data, aos_get_internal_xmm( cp, IMM_IDENTITY ), SHUF(X,Y,Z,W) );
  61    /* data = z ? 0 1 */
  62    sse_shufps(cp->func, data, data, SHUF(Y,Z,X,W) );
  63    /* data = ? 0 z 1 */
  64    sse_movlps(cp->func, data, src_ptr);
  65    /* data = x y z 1 */
  66 #else
  67    sse_movups(cp->func, data, src_ptr);
  68    /* data = x y z ? */
  69    sse2_pshufd(cp->func, data, data, SHUF(W,X,Y,Z) );
  70    /* data = ? x y z */
  71    sse_movss(cp->func, data, aos_get_internal_xmm( cp, IMM_ONES ) );
  72    /* data = 1 x y z */
  73    sse2_pshufd(cp->func, data, data, SHUF(Y,Z,W,X) );
  74    /* data = x y z 1 */
  75 #endif
  76 }
  77
  78 static void emit_load_R32G32( struct aos_compilation *cp,
  79                            struct x86_reg data,
  80                            struct x86_reg src_ptr )
  81 {
  82    sse_movups(cp->func, data, aos_get_internal_xmm( cp, IMM_IDENTITY ) );
  83    sse_movlps(cp->func, data, src_ptr);
  84 }
  85
  86
  87 static void emit_load_R32( struct aos_compilation *cp,
  88                            struct x86_reg data,
  89                            struct x86_reg src_ptr )
  90 {
  91    sse_movss(cp->func, data, src_ptr);
  92    sse_orps(cp->func, data, aos_get_internal_xmm( cp, IMM_IDENTITY ) );
  93 }
  94
  95
  96 static void emit_load_R8G8B8A8_UNORM( struct aos_compilation *cp,
  97                                        struct x86_reg data,
  98                                        struct x86_reg src_ptr )
  99 {
 100    sse_movss(cp->func, data, src_ptr);
 101    sse2_punpcklbw(cp->func, data, aos_get_internal_xmm( cp, IMM_IDENTITY ));
 102    sse2_punpcklbw(cp->func, data, aos_get_internal_xmm( cp, IMM_IDENTITY ));
 103    sse2_cvtdq2ps(cp->func, data, data);
 104    sse_mulps(cp->func, data, aos_get_internal(cp, IMM_INV_255));
 105 }
 106
 107
 108
 109 /* Extended swizzles?  Maybe later.
 110  */
 111 static void emit_swizzle( struct aos_compilation *cp,
 112                           struct x86_reg dest,
 113                           struct x86_reg src,
 114                           ubyte shuffle )
 115 {
 116    sse_shufps(cp->func, dest, src, shuffle);
 117 }
 118
 119
 120
 121 static boolean get_buffer_ptr( struct aos_compilation *cp,
 122                                boolean linear,
 123                                unsigned buf_idx,
 124                                struct x86_reg elt,
 125                                struct x86_reg ptr)
 126 {
 127    struct x86_reg buf = x86_make_disp(aos_get_x86( cp, 0, X86_BUFFERS ),
 128                                       buf_idx * sizeof(struct aos_buffer));
 129
 130    struct x86_reg buf_stride = x86_make_disp(buf,
 131                                              Offset(struct aos_buffer, stride));
 132    if (linear) {
 133       struct x86_reg buf_ptr = x86_make_disp(buf,
 134                                              Offset(struct aos_buffer, ptr));
 135
 136
 137       /* Calculate pointer to current attrib:
 138        */
 139       x86_mov(cp->func, ptr, buf_ptr);
 140       x86_mov(cp->func, elt, buf_stride);
 141       x86_add(cp->func, elt, ptr);
 142       if (buf_idx == 0) sse_prefetchnta(cp->func, x86_make_disp(elt, 192));
 143       x86_mov(cp->func, buf_ptr, elt);
 144    }
 145    else {
 146       struct x86_reg buf_base_ptr = x86_make_disp(buf,
 147                                                   Offset(struct aos_buffer, base_ptr));
 148
 149
 150       /* Calculate pointer to current attrib:
 151        */
 152       x86_mov(cp->func, ptr, buf_stride);
 153       x86_imul(cp->func, ptr, elt);
 154       x86_add(cp->func, ptr, buf_base_ptr);
 155    }
 156
 157    cp->insn_counter++;
 158
 159    return TRUE;
 160 }
 161
 162
 163 static boolean load_input( struct aos_compilation *cp,
 164                            unsigned idx,
 165                            struct x86_reg bufptr )
 166 {
 167    unsigned format = cp->vaos->base.key.element[idx].in.format;
 168    unsigned offset = cp->vaos->base.key.element[idx].in.offset;
 169    struct x86_reg dataXMM = aos_get_xmm_reg(cp);
 170
 171    /* Figure out source pointer address:
 172     */
 173    struct x86_reg src = x86_make_disp(bufptr, offset);
 174
 175    aos_adopt_xmm_reg( cp,
 176                       dataXMM,
 177                       TGSI_FILE_INPUT,
 178                       idx,
 179                       TRUE );
 180
 181    switch (format) {
 182    case PIPE_FORMAT_R32_FLOAT:
 183       emit_load_R32(cp, dataXMM, src);
 184       break;
 185    case PIPE_FORMAT_R32G32_FLOAT:
 186       emit_load_R32G32(cp, dataXMM, src);
 187       break;
 188    case PIPE_FORMAT_R32G32B32_FLOAT:
 189       emit_load_R32G32B32(cp, dataXMM, src);
 190       break;
 191    case PIPE_FORMAT_R32G32B32A32_FLOAT:
 192       emit_load_R32G32B32A32(cp, dataXMM, src);
 193       break;
 194    case PIPE_FORMAT_A8R8G8B8_UNORM:
 195       emit_load_R8G8B8A8_UNORM(cp, dataXMM, src);
 196       emit_swizzle(cp, dataXMM, dataXMM, SHUF(Z,Y,X,W));
 197       break;
 198    case PIPE_FORMAT_R8G8B8A8_UNORM:
 199       emit_load_R8G8B8A8_UNORM(cp, dataXMM, src);
 200       break;
 201    default:
 202       AOS_ERROR(cp, "unhandled input format");
 203       return FALSE;
 204    }
 205
 206    return TRUE;
 207 }
 208
 209 static boolean load_inputs( struct aos_compilation *cp,
 210                             unsigned buffer,
 211                             struct x86_reg ptr )
 212 {
 213    unsigned i;
 214
 215    for (i = 0; i < cp->vaos->base.key.nr_inputs; i++) {
 216       if (cp->vaos->base.key.element[i].in.buffer == buffer) {
 217
 218          if (!load_input( cp, i, ptr ))
 219             return FALSE;
 220
 221          cp->insn_counter++;
 222       }
 223    }
 224
 225    return TRUE;
 226 }
 227
 228 boolean aos_init_inputs( struct aos_compilation *cp, boolean linear )
 229 {
 230    unsigned i;
 231    for (i = 0; i < cp->vaos->nr_vb; i++) {
 232       struct x86_reg buf = x86_make_disp(aos_get_x86( cp, 0, X86_BUFFERS ),
 233                                          i * sizeof(struct aos_buffer));
 234
 235       struct x86_reg buf_base_ptr = x86_make_disp(buf,
 236                                                   Offset(struct aos_buffer, base_ptr));
 237
 238       if (cp->vaos->base.key.const_vbuffers & (1<<i)) {
 239          struct x86_reg ptr = cp->tmp_EAX;
 240
 241          x86_mov(cp->func, ptr, buf_base_ptr);
 242
 243          /* Load all inputs for this constant vertex buffer
 244           */
 245          load_inputs( cp, i, x86_deref(ptr) );
 246
 247          /* Then just force them out to aos_machine.input[]
 248           */
 249          aos_spill_all( cp );
 250
 251       }
 252       else if (linear) {
 253
 254          struct x86_reg elt = cp->idx_EBX;
 255          struct x86_reg ptr = cp->tmp_EAX;
 256
 257          struct x86_reg buf_stride = x86_make_disp(buf,
 258                                                    Offset(struct aos_buffer, stride));
 259
 260          struct x86_reg buf_ptr = x86_make_disp(buf,
 261                                                 Offset(struct aos_buffer, ptr));
 262
 263
 264          /* Calculate pointer to current attrib:
 265           */
 266          x86_mov(cp->func, ptr, buf_stride);
 267          x86_imul(cp->func, ptr, elt);
 268          x86_add(cp->func, ptr, buf_base_ptr);
 269
 270
 271          /* In the linear case, keep the buffer pointer instead of the
 272           * index number.
 273           */
 274          if (cp->vaos->nr_vb == 1)
 275             x86_mov( cp->func, elt, ptr );
 276          else
 277             x86_mov( cp->func, buf_ptr, ptr );
 278
 279          cp->insn_counter++;
 280       }
 281    }
 282
 283    return TRUE;
 284 }
 285
 286 boolean aos_fetch_inputs( struct aos_compilation *cp, boolean linear )
 287 {
 288    unsigned j;
 289
 290    for (j = 0; j < cp->vaos->nr_vb; j++) {
 291       if (cp->vaos->base.key.const_vbuffers & (1<<j)) {
 292          /* just retreive pre-transformed input */
 293       }
 294       else if (linear && cp->vaos->nr_vb == 1) {
 295          load_inputs( cp, 0, cp->idx_EBX );
 296       }
 297       else {
 298          struct x86_reg elt = linear ? cp->idx_EBX : x86_deref(cp->idx_EBX);
 299          struct x86_reg ptr = cp->tmp_EAX;
 300
 301          if (!get_buffer_ptr( cp, linear, j, elt, ptr ))
 302             return FALSE;
 303
 304          if (!load_inputs( cp, j, ptr ))
 305             return FALSE;
 306       }
 307    }
 308
 309    return TRUE;
 310 }
 311
 312 boolean aos_incr_inputs( struct aos_compilation *cp, boolean linear )
 313 {
 314    if (linear && cp->vaos->nr_vb == 1) {
 315       struct x86_reg stride = x86_make_disp(aos_get_x86( cp, 0, X86_BUFFERS ),
 316                                             (0 * sizeof(struct aos_buffer) +
 317                                              Offset(struct aos_buffer, stride)));
 318
 319       x86_add(cp->func, cp->idx_EBX, stride);
 320       sse_prefetchnta(cp->func, x86_make_disp(cp->idx_EBX, 192));
 321    }
 322    else if (linear) {
 323       /* Nothing to do */
 324    }
 325    else {
 326       x86_lea(cp->func, cp->idx_EBX, x86_make_disp(cp->idx_EBX, 4));
 327    }
 328
 329    return TRUE;
 330 }
 331
 332
 333
 334
 335
 336
 337 static void emit_store_R32G32B32A32( struct aos_compilation *cp,
 338                                      struct x86_reg dst_ptr,
 339                                      struct x86_reg dataXMM )
 340 {
 341    sse_movups(cp->func, dst_ptr, dataXMM);
 342 }
 343
 344 static void emit_store_R32G32B32( struct aos_compilation *cp,
 345                                   struct x86_reg dst_ptr,
 346                                   struct x86_reg dataXMM )
 347 {
 348    sse_movlps(cp->func, dst_ptr, dataXMM);
 349    sse_shufps(cp->func, dataXMM, dataXMM, SHUF(Z,Z,Z,Z) ); /* NOTE! destructive */
 350    sse_movss(cp->func, x86_make_disp(dst_ptr,8), dataXMM);
 351 }
 352
 353 static void emit_store_R32G32( struct aos_compilation *cp,
 354                                struct x86_reg dst_ptr,
 355                                struct x86_reg dataXMM )
 356 {
 357    sse_movlps(cp->func, dst_ptr, dataXMM);
 358 }
 359
 360 static void emit_store_R32( struct aos_compilation *cp,
 361                             struct x86_reg dst_ptr,
 362                             struct x86_reg dataXMM )
 363 {
 364    sse_movss(cp->func, dst_ptr, dataXMM);
 365 }
 366
 367
 368
 369 static void emit_store_R8G8B8A8_UNORM( struct aos_compilation *cp,
 370                                        struct x86_reg dst_ptr,
 371                                        struct x86_reg dataXMM )
 372 {
 373    sse_mulps(cp->func, dataXMM, aos_get_internal(cp, IMM_255));
 374    sse2_cvtps2dq(cp->func, dataXMM, dataXMM);
 375    sse2_packssdw(cp->func, dataXMM, dataXMM);
 376    sse2_packuswb(cp->func, dataXMM, dataXMM);
 377    sse_movss(cp->func, dst_ptr, dataXMM);
 378 }
 379
 380
 381
 382
 383
 384 static boolean emit_output( struct aos_compilation *cp,
 385                             struct x86_reg ptr,
 386                             struct x86_reg dataXMM,
 387                             unsigned format )
 388 {
 389    switch (format) {
 390    case EMIT_1F:
 391    case EMIT_1F_PSIZE:
 392       emit_store_R32(cp, ptr, dataXMM);
 393       break;
 394    case EMIT_2F:
 395       emit_store_R32G32(cp, ptr, dataXMM);
 396       break;
 397    case EMIT_3F:
 398       emit_store_R32G32B32(cp, ptr, dataXMM);
 399       break;
 400    case EMIT_4F:
 401       emit_store_R32G32B32A32(cp, ptr, dataXMM);
 402       break;
 403    case EMIT_4UB:
 404       emit_store_R8G8B8A8_UNORM(cp, ptr, dataXMM);
 405       break;
 406    case EMIT_4UB_BGRA:
 407       emit_swizzle(cp, dataXMM, dataXMM, SHUF(Z,Y,X,W));
 408       emit_store_R8G8B8A8_UNORM(cp, ptr, dataXMM);
 409       break;
 410    default:
 411       AOS_ERROR(cp, "unhandled output format");
 412       return FALSE;
 413    }
 414
 415    return TRUE;
 416 }
 417
 418
 419
 420 boolean aos_emit_outputs( struct aos_compilation *cp )
 421 {
 422    unsigned i;
 423
 424    for (i = 0; i < cp->vaos->base.key.nr_outputs; i++) {
 425       unsigned format = cp->vaos->base.key.element[i].out.format;
 426       unsigned offset = cp->vaos->base.key.element[i].out.offset;
 427       unsigned vs_output = cp->vaos->base.key.element[i].out.vs_output;
 428
 429       struct x86_reg data;
 430
 431       if (format == EMIT_1F_PSIZE) {
 432          data = aos_get_internal_xmm( cp, IMM_PSIZE );
 433       }
 434       else {
 435          data = aos_get_shader_reg( cp,
 436                                     TGSI_FILE_OUTPUT,
 437                                     vs_output );
 438       }
 439
 440       if (data.file != file_XMM) {
 441          struct x86_reg tmp = aos_get_xmm_reg( cp );
 442          sse_movaps(cp->func, tmp, data);
 443          data = tmp;
 444       }
 445
 446       if (!emit_output( cp,
 447                         x86_make_disp( cp->outbuf_ECX, offset ),
 448                         data,
 449                         format ))
 450          return FALSE;
 451
 452       aos_release_xmm_reg( cp, data.idx );
 453
 454       cp->insn_counter++;
 455    }
 456
 457    return TRUE;
 458 }
 459
 460 #endif