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    sse_movss(cp->func, data, x86_make_disp(src_ptr, 8));
  58    /* data = z ? ? ? */
  59    sse_shufps(cp->func, data, aos_get_internal_xmm( cp, IMM_IDENTITY ), SHUF(X,Y,Z,W) );
  60    /* data = z ? 0 1 */
  61    sse_shufps(cp->func, data, data, SHUF(Y,Z,X,W) );
  62    /* data = ? 0 z 1 */
  63    sse_movlps(cp->func, data, src_ptr);
  64    /* data = x y z 1 */
  65 }
  66
  67 static void emit_load_R32G32( struct aos_compilation *cp,
  68                            struct x86_reg data,
  69                            struct x86_reg src_ptr )
  70 {
  71    sse_movups(cp->func, data, aos_get_internal_xmm( cp, IMM_IDENTITY ) );
  72    sse_movlps(cp->func, data, src_ptr);
  73 }
  74
  75
  76 static void emit_load_R32( struct aos_compilation *cp,
  77                            struct x86_reg data,
  78                            struct x86_reg src_ptr )
  79 {
  80    sse_movss(cp->func, data, src_ptr);
  81    sse_orps(cp->func, data, aos_get_internal_xmm( cp, IMM_IDENTITY ) );
  82 }
  83
  84
  85 static void emit_load_R8G8B8A8_UNORM( struct aos_compilation *cp,
  86                                        struct x86_reg data,
  87                                        struct x86_reg src_ptr )
  88 {
  89    sse_movss(cp->func, data, src_ptr);
  90    sse2_punpcklbw(cp->func, data, aos_get_internal_xmm( cp, IMM_IDENTITY ));
  91    sse2_punpcklbw(cp->func, data, aos_get_internal_xmm( cp, IMM_IDENTITY ));
  92    sse2_cvtdq2ps(cp->func, data, data);
  93    sse_mulps(cp->func, data, aos_get_internal(cp, IMM_INV_255));
  94 }
  95
  96
  97
  98 /* Extended swizzles?  Maybe later.
  99  */
 100 static void emit_swizzle( struct aos_compilation *cp,
 101                           struct x86_reg dest,
 102                           struct x86_reg src,
 103                           ubyte shuffle )
 104 {
 105    sse_shufps(cp->func, dest, src, shuffle);
 106 }
 107
 108
 109
 110 static boolean get_buffer_ptr( struct aos_compilation *cp,
 111                                boolean linear,
 112                                unsigned buf_idx,
 113                                struct x86_reg elt,
 114                                struct x86_reg ptr)
 115 {
 116    struct x86_reg buf = x86_make_disp(aos_get_x86( cp, 0, X86_BUFFERS ),
 117                                       buf_idx * sizeof(struct aos_buffer));
 118
 119    struct x86_reg buf_stride = x86_make_disp(buf,
 120                                              Offset(struct aos_buffer, stride));
 121    if (linear) {
 122       struct x86_reg buf_ptr = x86_make_disp(buf,
 123                                              Offset(struct aos_buffer, ptr));
 124
 125
 126       /* Calculate pointer to current attrib:
 127        */
 128       x86_mov(cp->func, ptr, buf_ptr);
 129       x86_mov(cp->func, elt, buf_stride);
 130       x86_add(cp->func, elt, ptr);
 131       sse_prefetchnta(cp->func, x86_deref(elt));
 132       x86_mov(cp->func, buf_ptr, elt);
 133    }
 134    else {
 135       struct x86_reg buf_base_ptr = x86_make_disp(buf,
 136                                                   Offset(struct aos_buffer, base_ptr));
 137
 138
 139       /* Calculate pointer to current attrib:
 140        */
 141       x86_mov(cp->func, ptr, buf_stride);
 142       x86_imul(cp->func, ptr, elt);
 143       x86_add(cp->func, ptr, buf_base_ptr);
 144    }
 145
 146    cp->insn_counter++;
 147
 148    return TRUE;
 149 }
 150
 151
 152 static boolean load_input( struct aos_compilation *cp,
 153                            unsigned idx,
 154                            struct x86_reg bufptr )
 155 {
 156    unsigned format = cp->vaos->base.key.element[idx].in.format;
 157    unsigned offset = cp->vaos->base.key.element[idx].in.offset;
 158    struct x86_reg dataXMM = aos_get_xmm_reg(cp);
 159
 160    /* Figure out source pointer address:
 161     */
 162    struct x86_reg src = x86_make_disp(bufptr, offset);
 163
 164    aos_adopt_xmm_reg( cp,
 165                       dataXMM,
 166                       TGSI_FILE_INPUT,
 167                       idx,
 168                       TRUE );
 169
 170    switch (format) {
 171    case PIPE_FORMAT_R32_FLOAT:
 172       emit_load_R32(cp, dataXMM, src);
 173       break;
 174    case PIPE_FORMAT_R32G32_FLOAT:
 175       emit_load_R32G32(cp, dataXMM, src);
 176       break;
 177    case PIPE_FORMAT_R32G32B32_FLOAT:
 178       emit_load_R32G32B32(cp, dataXMM, src);
 179       break;
 180    case PIPE_FORMAT_R32G32B32A32_FLOAT:
 181       emit_load_R32G32B32A32(cp, dataXMM, src);
 182       break;
 183    case PIPE_FORMAT_B8G8R8A8_UNORM:
 184       emit_load_R8G8B8A8_UNORM(cp, dataXMM, src);
 185       emit_swizzle(cp, dataXMM, dataXMM, SHUF(Z,Y,X,W));
 186       break;
 187    case PIPE_FORMAT_R8G8B8A8_UNORM:
 188       emit_load_R8G8B8A8_UNORM(cp, dataXMM, src);
 189       break;
 190    default:
 191       ERROR(cp, "unhandled input format");
 192       return FALSE;
 193    }
 194
 195    return TRUE;
 196 }
 197
 198 static boolean load_inputs( struct aos_compilation *cp,
 199                             unsigned buffer,
 200                             struct x86_reg ptr )
 201 {
 202    unsigned i;
 203
 204    for (i = 0; i < cp->vaos->base.key.nr_inputs; i++) {
 205       if (cp->vaos->base.key.element[i].in.buffer == buffer) {
 206
 207          if (!load_input( cp, i, ptr ))
 208             return FALSE;
 209
 210          cp->insn_counter++;
 211       }
 212    }
 213
 214    return TRUE;
 215 }
 216
 217 boolean aos_init_inputs( struct aos_compilation *cp, boolean linear )
 218 {
 219    unsigned i;
 220    for (i = 0; i < cp->vaos->nr_vb; i++) {
 221       struct x86_reg buf = x86_make_disp(aos_get_x86( cp, 0, X86_BUFFERS ),
 222                                          i * sizeof(struct aos_buffer));
 223
 224       struct x86_reg buf_base_ptr = x86_make_disp(buf,
 225                                                   Offset(struct aos_buffer, base_ptr));
 226
 227       if (cp->vaos->base.key.const_vbuffers & (1<<i)) {
 228          struct x86_reg ptr = cp->tmp_EAX;
 229
 230          x86_mov(cp->func, ptr, buf_base_ptr);
 231
 232          /* Load all inputs for this constant vertex buffer
 233           */
 234          load_inputs( cp, i, x86_deref(ptr) );
 235
 236          /* Then just force them out to aos_machine.input[]
 237           */
 238          aos_spill_all( cp );
 239
 240       }
 241       else if (linear) {
 242
 243          struct x86_reg elt = cp->idx_EBX;
 244          struct x86_reg ptr = cp->tmp_EAX;
 245
 246          struct x86_reg buf_stride = x86_make_disp(buf,
 247                                                    Offset(struct aos_buffer, stride));
 248
 249          struct x86_reg buf_ptr = x86_make_disp(buf,
 250                                                 Offset(struct aos_buffer, ptr));
 251
 252
 253          /* Calculate pointer to current attrib:
 254           */
 255          x86_mov(cp->func, ptr, buf_stride);
 256          x86_imul(cp->func, ptr, elt);
 257          x86_add(cp->func, ptr, buf_base_ptr);
 258
 259
 260          /* In the linear case, keep the buffer pointer instead of the
 261           * index number.
 262           */
 263          if (cp->vaos->nr_vb == 1)
 264             x86_mov( cp->func, elt, ptr );
 265          else
 266             x86_mov( cp->func, buf_ptr, ptr );
 267
 268          cp->insn_counter++;
 269       }
 270    }
 271
 272    return TRUE;
 273 }
 274
 275 boolean aos_fetch_inputs( struct aos_compilation *cp, boolean linear )
 276 {
 277    unsigned j;
 278
 279    for (j = 0; j < cp->vaos->nr_vb; j++) {
 280       if (cp->vaos->base.key.const_vbuffers & (1<<j)) {
 281          /* just retreive pre-transformed input */
 282       }
 283       else if (linear && cp->vaos->nr_vb == 1) {
 284          load_inputs( cp, 0, cp->idx_EBX );
 285       }
 286       else {
 287          struct x86_reg elt = linear ? cp->idx_EBX : x86_deref(cp->idx_EBX);
 288          struct x86_reg ptr = cp->tmp_EAX;
 289
 290          if (!get_buffer_ptr( cp, linear, j, elt, ptr ))
 291             return FALSE;
 292
 293          if (!load_inputs( cp, j, ptr ))
 294             return FALSE;
 295       }
 296    }
 297
 298    return TRUE;
 299 }
 300
 301 boolean aos_incr_inputs( struct aos_compilation *cp, boolean linear )
 302 {
 303    if (linear && cp->vaos->nr_vb == 1) {
 304       struct x86_reg stride = x86_make_disp(aos_get_x86( cp, 0, X86_BUFFERS ),
 305                                             (0 * sizeof(struct aos_buffer) +
 306                                              Offset(struct aos_buffer, stride)));
 307
 308       x86_add(cp->func, cp->idx_EBX, stride);
 309       sse_prefetchnta(cp->func, x86_deref(cp->idx_EBX));
 310    }
 311    else if (linear) {
 312       /* Nothing to do */
 313    }
 314    else {
 315       x86_lea(cp->func, cp->idx_EBX, x86_make_disp(cp->idx_EBX, 4));
 316    }
 317
 318    return TRUE;
 319 }
 320
 321
 322
 323
 324
 325
 326 static void emit_store_R32G32B32A32( struct aos_compilation *cp,
 327                                      struct x86_reg dst_ptr,
 328                                      struct x86_reg dataXMM )
 329 {
 330    sse_movups(cp->func, dst_ptr, dataXMM);
 331 }
 332
 333 static void emit_store_R32G32B32( struct aos_compilation *cp,
 334                                   struct x86_reg dst_ptr,
 335                                   struct x86_reg dataXMM )
 336 {
 337    sse_movlps(cp->func, dst_ptr, dataXMM);
 338    sse_shufps(cp->func, dataXMM, dataXMM, SHUF(Z,Z,Z,Z) ); /* NOTE! destructive */
 339    sse_movss(cp->func, x86_make_disp(dst_ptr,8), dataXMM);
 340 }
 341
 342 static void emit_store_R32G32( struct aos_compilation *cp,
 343                                struct x86_reg dst_ptr,
 344                                struct x86_reg dataXMM )
 345 {
 346    sse_movlps(cp->func, dst_ptr, dataXMM);
 347 }
 348
 349 static void emit_store_R32( struct aos_compilation *cp,
 350                             struct x86_reg dst_ptr,
 351                             struct x86_reg dataXMM )
 352 {
 353    sse_movss(cp->func, dst_ptr, dataXMM);
 354 }
 355
 356
 357
 358 static void emit_store_R8G8B8A8_UNORM( struct aos_compilation *cp,
 359                                        struct x86_reg dst_ptr,
 360                                        struct x86_reg dataXMM )
 361 {
 362    sse_mulps(cp->func, dataXMM, aos_get_internal(cp, IMM_255));
 363    sse2_cvtps2dq(cp->func, dataXMM, dataXMM);
 364    sse2_packssdw(cp->func, dataXMM, dataXMM);
 365    sse2_packuswb(cp->func, dataXMM, dataXMM);
 366    sse_movss(cp->func, dst_ptr, dataXMM);
 367 }
 368
 369
 370
 371
 372
 373 static boolean emit_output( struct aos_compilation *cp,
 374                             struct x86_reg ptr,
 375                             struct x86_reg dataXMM,
 376                             unsigned format )
 377 {
 378    switch (format) {
 379    case EMIT_1F:
 380    case EMIT_1F_PSIZE:
 381       emit_store_R32(cp, ptr, dataXMM);
 382       break;
 383    case EMIT_2F:
 384       emit_store_R32G32(cp, ptr, dataXMM);
 385       break;
 386    case EMIT_3F:
 387       emit_store_R32G32B32(cp, ptr, dataXMM);
 388       break;
 389    case EMIT_4F:
 390       emit_store_R32G32B32A32(cp, ptr, dataXMM);
 391       break;
 392    case EMIT_4UB:
 393       if (1) {
 394          emit_swizzle(cp, dataXMM, dataXMM, SHUF(Z,Y,X,W));
 395          emit_store_R8G8B8A8_UNORM(cp, ptr, dataXMM);
 396       }
 397       else {
 398          emit_store_R8G8B8A8_UNORM(cp, ptr, dataXMM);
 399       }
 400       break;
 401    default:
 402       ERROR(cp, "unhandled output format");
 403       return FALSE;
 404    }
 405
 406    return TRUE;
 407 }
 408
 409
 410
 411 boolean aos_emit_outputs( struct aos_compilation *cp )
 412 {
 413    unsigned i;
 414
 415    for (i = 0; i < cp->vaos->base.key.nr_outputs; i++) {
 416       unsigned format = cp->vaos->base.key.element[i].out.format;
 417       unsigned offset = cp->vaos->base.key.element[i].out.offset;
 418       unsigned vs_output = cp->vaos->base.key.element[i].out.vs_output;
 419
 420       struct x86_reg data;
 421
 422       if (format == EMIT_1F_PSIZE) {
 423          data = aos_get_internal_xmm( cp, IMM_PSIZE );
 424       }
 425       else {
 426          data = aos_get_shader_reg( cp,
 427                                     TGSI_FILE_OUTPUT,
 428                                     vs_output );
 429       }
 430
 431       if (data.file != file_XMM) {
 432          struct x86_reg tmp = aos_get_xmm_reg( cp );
 433          sse_movups(cp->func, tmp, data);
 434          data = tmp;
 435       }
 436
 437       if (!emit_output( cp,
 438                         x86_make_disp( cp->outbuf_ECX, offset ),
 439                         data,
 440                         format ))
 441          return FALSE;
 442
 443       aos_release_xmm_reg( cp, data.idx );
 444
 445       cp->insn_counter++;
 446    }
 447
 448    return TRUE;
 449 }
 450
 451 #endif