src/gallium/auxiliary/draw/draw_pipe_aapoint.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2008 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  * AA point stage:  AA points are converted to quads and rendered with a
  30  * special fragment shader.  Another approach would be to use a texture
  31  * map image of a point, but experiments indicate the quality isn't nearly
  32  * as good as this approach.
  33  *
  34  * Note: this looks a lot like draw_aaline.c but there's actually little
  35  * if any code that can be shared.
  36  *
  37  * Authors:  Brian Paul
  38  */
  39
  40
  41 #include "pipe/p_context.h"
  42 #include "pipe/p_defines.h"
  43 #include "pipe/p_shader_tokens.h"
  44
  45 #include "tgsi/tgsi_transform.h"
  46 #include "tgsi/tgsi_dump.h"
  47
  48 #include "util/u_math.h"
  49 #include "util/u_memory.h"
  50
  51 #include "draw_context.h"
  52 #include "draw_vs.h"
  53 #include "draw_pipe.h"
  54
  55
  56 /** Approx number of new tokens for instructions in aa_transform_inst() */
  57 #define NUM_NEW_TOKENS 200
  58
  59
  60 /*
  61  * Enabling NORMALIZE might give _slightly_ better results.
  62  * Basically, it controls whether we compute distance as d=sqrt(x*x+y*y) or
  63  * d=x*x+y*y.  Since we're working with a unit circle, the later seems
  64  * close enough and saves some costly instructions.
  65  */
  66 #define NORMALIZE 0
  67
  68
  69 /**
  70  * Subclass of pipe_shader_state to carry extra fragment shader info.
  71  */
  72 struct aapoint_fragment_shader
  73 {
  74    struct pipe_shader_state state;
  75    void *driver_fs;   /**< the regular shader */
  76    void *aapoint_fs;  /**< the aa point-augmented shader */
  77    int generic_attrib; /**< The generic input attrib/texcoord we'll use */
  78 };
  79
  80
  81 /**
  82  * Subclass of draw_stage
  83  */
  84 struct aapoint_stage
  85 {
  86    struct draw_stage stage;
  87
  88    /** half of pipe_rasterizer_state::point_size */
  89    float radius;
  90
  91    /** vertex attrib slot containing point size */
  92    int psize_slot;
  93
  94    /** this is the vertex attrib slot for the new texcoords */
  95    uint tex_slot;
  96
  97    /** vertex attrib slot containing position */
  98    uint pos_slot;
  99
 100    /** Currently bound fragment shader */
 101    struct aapoint_fragment_shader *fs;
 102
 103    /*
 104     * Driver interface/override functions
 105     */
 106    void * (*driver_create_fs_state)(struct pipe_context *,
 107                                     const struct pipe_shader_state *);
 108    void (*driver_bind_fs_state)(struct pipe_context *, void *);
 109    void (*driver_delete_fs_state)(struct pipe_context *, void *);
 110 };
 111
 112
 113
 114 /**
 115  * Subclass of tgsi_transform_context, used for transforming the
 116  * user's fragment shader to add the special AA instructions.
 117  */
 118 struct aa_transform_context {
 119    struct tgsi_transform_context base;
 120    uint tempsUsed;  /**< bitmask */
 121    int colorOutput; /**< which output is the primary color */
 122    int maxInput, maxGeneric;  /**< max input index found */
 123    int tmp0, colorTemp;  /**< temp registers */
 124    boolean firstInstruction;
 125 };
 126
 127
 128 /**
 129  * TGSI declaration transform callback.
 130  * Look for two free temp regs and available input reg for new texcoords.
 131  */
 132 static void
 133 aa_transform_decl(struct tgsi_transform_context *ctx,
 134                   struct tgsi_full_declaration *decl)
 135 {
 136    struct aa_transform_context *aactx = (struct aa_transform_context *) ctx;
 137
 138    if (decl->Declaration.File == TGSI_FILE_OUTPUT &&
 139        decl->Semantic.Name == TGSI_SEMANTIC_COLOR &&
 140        decl->Semantic.Index == 0) {
 141       aactx->colorOutput = decl->Range.First;
 142    }
 143    else if (decl->Declaration.File == TGSI_FILE_INPUT) {
 144       if ((int) decl->Range.Last > aactx->maxInput)
 145          aactx->maxInput = decl->Range.Last;
 146       if (decl->Semantic.Name == TGSI_SEMANTIC_GENERIC &&
 147            (int) decl->Semantic.Index > aactx->maxGeneric) {
 148          aactx->maxGeneric = decl->Semantic.Index;
 149       }
 150    }
 151    else if (decl->Declaration.File == TGSI_FILE_TEMPORARY) {
 152       uint i;
 153       for (i = decl->Range.First;
 154            i <= decl->Range.Last; i++) {
 155          aactx->tempsUsed |= (1 << i);
 156       }
 157    }
 158
 159    ctx->emit_declaration(ctx, decl);
 160 }
 161
 162
 163 /**
 164  * TGSI instruction transform callback.
 165  * Replace writes to result.color w/ a temp reg.
 166  * Upon END instruction, insert texture sampling code for antialiasing.
 167  */
 168 static void
 169 aa_transform_inst(struct tgsi_transform_context *ctx,
 170                   struct tgsi_full_instruction *inst)
 171 {
 172    struct aa_transform_context *aactx = (struct aa_transform_context *) ctx;
 173    struct tgsi_full_instruction newInst;
 174
 175    if (aactx->firstInstruction) {
 176       /* emit our new declarations before the first instruction */
 177
 178       struct tgsi_full_declaration decl;
 179       const int texInput = aactx->maxInput + 1;
 180       int tmp0;
 181       uint i;
 182
 183       /* find two free temp regs */
 184       for (i = 0; i < 32; i++) {
 185          if ((aactx->tempsUsed & (1 << i)) == 0) {
 186             /* found a free temp */
 187             if (aactx->tmp0 < 0)
 188                aactx->tmp0 = i;
 189             else if (aactx->colorTemp < 0)
 190                aactx->colorTemp = i;
 191             else
 192                break;
 193          }
 194       }
 195
 196       assert(aactx->colorTemp != aactx->tmp0);
 197
 198       tmp0 = aactx->tmp0;
 199
 200       /* declare new generic input/texcoord */
 201       decl = tgsi_default_full_declaration();
 202       decl.Declaration.File = TGSI_FILE_INPUT;
 203       /* XXX this could be linear... */
 204       decl.Declaration.Interpolate = TGSI_INTERPOLATE_PERSPECTIVE;
 205       decl.Declaration.Semantic = 1;
 206       decl.Semantic.Name = TGSI_SEMANTIC_GENERIC;
 207       decl.Semantic.Index = aactx->maxGeneric + 1;
 208       decl.Range.First =
 209       decl.Range.Last = texInput;
 210       ctx->emit_declaration(ctx, &decl);
 211
 212       /* declare new temp regs */
 213       decl = tgsi_default_full_declaration();
 214       decl.Declaration.File = TGSI_FILE_TEMPORARY;
 215       decl.Range.First =
 216       decl.Range.Last = tmp0;
 217       ctx->emit_declaration(ctx, &decl);
 218
 219       decl = tgsi_default_full_declaration();
 220       decl.Declaration.File = TGSI_FILE_TEMPORARY;
 221       decl.Range.First =
 222       decl.Range.Last = aactx->colorTemp;
 223       ctx->emit_declaration(ctx, &decl);
 224
 225       aactx->firstInstruction = FALSE;
 226
 227
 228       /*
 229        * Emit code to compute fragment coverage, kill if outside point radius
 230        *
 231        * Temp reg0 usage:
 232        *  t0.x = distance of fragment from center point
 233        *  t0.y = boolean, is t0.x > 1.0, also misc temp usage
 234        *  t0.z = temporary for computing 1/(1-k) value
 235        *  t0.w = final coverage value
 236        */
 237
 238       /* MUL t0.xy, tex, tex;  # compute x^2, y^2 */
 239       newInst = tgsi_default_full_instruction();
 240       newInst.Instruction.Opcode = TGSI_OPCODE_MUL;
 241       newInst.Instruction.NumDstRegs = 1;
 242       newInst.Dst[0].Register.File = TGSI_FILE_TEMPORARY;
 243       newInst.Dst[0].Register.Index = tmp0;
 244       newInst.Dst[0].Register.WriteMask = TGSI_WRITEMASK_XY;
 245       newInst.Instruction.NumSrcRegs = 2;
 246       newInst.Src[0].Register.File = TGSI_FILE_INPUT;
 247       newInst.Src[0].Register.Index = texInput;
 248       newInst.Src[1].Register.File = TGSI_FILE_INPUT;
 249       newInst.Src[1].Register.Index = texInput;
 250       ctx->emit_instruction(ctx, &newInst);
 251
 252       /* ADD t0.x, t0.x, t0.y;  # x^2 + y^2 */
 253       newInst = tgsi_default_full_instruction();
 254       newInst.Instruction.Opcode = TGSI_OPCODE_ADD;
 255       newInst.Instruction.NumDstRegs = 1;
 256       newInst.Dst[0].Register.File = TGSI_FILE_TEMPORARY;
 257       newInst.Dst[0].Register.Index = tmp0;
 258       newInst.Dst[0].Register.WriteMask = TGSI_WRITEMASK_X;
 259       newInst.Instruction.NumSrcRegs = 2;
 260       newInst.Src[0].Register.File = TGSI_FILE_TEMPORARY;
 261       newInst.Src[0].Register.Index = tmp0;
 262       newInst.Src[0].Register.SwizzleX = TGSI_SWIZZLE_X;
 263       newInst.Src[1].Register.File = TGSI_FILE_TEMPORARY;
 264       newInst.Src[1].Register.Index = tmp0;
 265       newInst.Src[1].Register.SwizzleX = TGSI_SWIZZLE_Y;
 266       ctx->emit_instruction(ctx, &newInst);
 267
 268 #if NORMALIZE  /* OPTIONAL normalization of length */
 269       /* RSQ t0.x, t0.x; */
 270       newInst = tgsi_default_full_instruction();
 271       newInst.Instruction.Opcode = TGSI_OPCODE_RSQ;
 272       newInst.Instruction.NumDstRegs = 1;
 273       newInst.Dst[0].Register.File = TGSI_FILE_TEMPORARY;
 274       newInst.Dst[0].Register.Index = tmp0;
 275       newInst.Dst[0].Register.WriteMask = TGSI_WRITEMASK_X;
 276       newInst.Instruction.NumSrcRegs = 1;
 277       newInst.Src[0].Register.File = TGSI_FILE_TEMPORARY;
 278       newInst.Src[0].Register.Index = tmp0;
 279       ctx->emit_instruction(ctx, &newInst);
 280
 281       /* RCP t0.x, t0.x; */
 282       newInst = tgsi_default_full_instruction();
 283       newInst.Instruction.Opcode = TGSI_OPCODE_RCP;
 284       newInst.Instruction.NumDstRegs = 1;
 285       newInst.Dst[0].Register.File = TGSI_FILE_TEMPORARY;
 286       newInst.Dst[0].Register.Index = tmp0;
 287       newInst.Dst[0].Register.WriteMask = TGSI_WRITEMASK_X;
 288       newInst.Instruction.NumSrcRegs = 1;
 289       newInst.Src[0].Register.File = TGSI_FILE_TEMPORARY;
 290       newInst.Src[0].Register.Index = tmp0;
 291       ctx->emit_instruction(ctx, &newInst);
 292 #endif
 293
 294       /* SGT t0.y, t0.xxxx, tex.wwww;  # bool b = d > 1 (NOTE tex.w == 1) */
 295       newInst = tgsi_default_full_instruction();
 296       newInst.Instruction.Opcode = TGSI_OPCODE_SGT;
 297       newInst.Instruction.NumDstRegs = 1;
 298       newInst.Dst[0].Register.File = TGSI_FILE_TEMPORARY;
 299       newInst.Dst[0].Register.Index = tmp0;
 300       newInst.Dst[0].Register.WriteMask = TGSI_WRITEMASK_Y;
 301       newInst.Instruction.NumSrcRegs = 2;
 302       newInst.Src[0].Register.File = TGSI_FILE_TEMPORARY;
 303       newInst.Src[0].Register.Index = tmp0;
 304       newInst.Src[0].Register.SwizzleY = TGSI_SWIZZLE_X;
 305       newInst.Src[1].Register.File = TGSI_FILE_INPUT;
 306       newInst.Src[1].Register.Index = texInput;
 307       newInst.Src[1].Register.SwizzleY = TGSI_SWIZZLE_W;
 308       ctx->emit_instruction(ctx, &newInst);
 309
 310       /* KIL -tmp0.yyyy;   # if -tmp0.y < 0, KILL */
 311       newInst = tgsi_default_full_instruction();
 312       newInst.Instruction.Opcode = TGSI_OPCODE_KIL;
 313       newInst.Instruction.NumDstRegs = 0;
 314       newInst.Instruction.NumSrcRegs = 1;
 315       newInst.Src[0].Register.File = TGSI_FILE_TEMPORARY;
 316       newInst.Src[0].Register.Index = tmp0;
 317       newInst.Src[0].Register.SwizzleX = TGSI_SWIZZLE_Y;
 318       newInst.Src[0].Register.SwizzleY = TGSI_SWIZZLE_Y;
 319       newInst.Src[0].Register.SwizzleZ = TGSI_SWIZZLE_Y;
 320       newInst.Src[0].Register.SwizzleW = TGSI_SWIZZLE_Y;
 321       newInst.Src[0].Register.Negate = 1;
 322       ctx->emit_instruction(ctx, &newInst);
 323
 324
 325       /* compute coverage factor = (1-d)/(1-k) */
 326
 327       /* SUB t0.z, tex.w, tex.z;  # m = 1 - k */
 328       newInst = tgsi_default_full_instruction();
 329       newInst.Instruction.Opcode = TGSI_OPCODE_SUB;
 330       newInst.Instruction.NumDstRegs = 1;
 331       newInst.Dst[0].Register.File = TGSI_FILE_TEMPORARY;
 332       newInst.Dst[0].Register.Index = tmp0;
 333       newInst.Dst[0].Register.WriteMask = TGSI_WRITEMASK_Z;
 334       newInst.Instruction.NumSrcRegs = 2;
 335       newInst.Src[0].Register.File = TGSI_FILE_INPUT;
 336       newInst.Src[0].Register.Index = texInput;
 337       newInst.Src[0].Register.SwizzleZ = TGSI_SWIZZLE_W;
 338       newInst.Src[1].Register.File = TGSI_FILE_INPUT;
 339       newInst.Src[1].Register.Index = texInput;
 340       newInst.Src[1].Register.SwizzleZ = TGSI_SWIZZLE_Z;
 341       ctx->emit_instruction(ctx, &newInst);
 342
 343       /* RCP t0.z, t0.z;  # t0.z = 1 / m */
 344       newInst = tgsi_default_full_instruction();
 345       newInst.Instruction.Opcode = TGSI_OPCODE_RCP;
 346       newInst.Instruction.NumDstRegs = 1;
 347       newInst.Dst[0].Register.File = TGSI_FILE_TEMPORARY;
 348       newInst.Dst[0].Register.Index = tmp0;
 349       newInst.Dst[0].Register.WriteMask = TGSI_WRITEMASK_Z;
 350       newInst.Instruction.NumSrcRegs = 1;
 351       newInst.Src[0].Register.File = TGSI_FILE_TEMPORARY;
 352       newInst.Src[0].Register.Index = tmp0;
 353       newInst.Src[0].Register.SwizzleX = TGSI_SWIZZLE_Z;
 354       ctx->emit_instruction(ctx, &newInst);
 355
 356       /* SUB t0.y, 1, t0.x;  # d = 1 - d */
 357       newInst = tgsi_default_full_instruction();
 358       newInst.Instruction.Opcode = TGSI_OPCODE_SUB;
 359       newInst.Instruction.NumDstRegs = 1;
 360       newInst.Dst[0].Register.File = TGSI_FILE_TEMPORARY;
 361       newInst.Dst[0].Register.Index = tmp0;
 362       newInst.Dst[0].Register.WriteMask = TGSI_WRITEMASK_Y;
 363       newInst.Instruction.NumSrcRegs = 2;
 364       newInst.Src[0].Register.File = TGSI_FILE_INPUT;
 365       newInst.Src[0].Register.Index = texInput;
 366       newInst.Src[0].Register.SwizzleY = TGSI_SWIZZLE_W;
 367       newInst.Src[1].Register.File = TGSI_FILE_TEMPORARY;
 368       newInst.Src[1].Register.Index = tmp0;
 369       newInst.Src[1].Register.SwizzleY = TGSI_SWIZZLE_X;
 370       ctx->emit_instruction(ctx, &newInst);
 371
 372       /* MUL t0.w, t0.y, t0.z;   # coverage = d * m */
 373       newInst = tgsi_default_full_instruction();
 374       newInst.Instruction.Opcode = TGSI_OPCODE_MUL;
 375       newInst.Instruction.NumDstRegs = 1;
 376       newInst.Dst[0].Register.File = TGSI_FILE_TEMPORARY;
 377       newInst.Dst[0].Register.Index = tmp0;
 378       newInst.Dst[0].Register.WriteMask = TGSI_WRITEMASK_W;
 379       newInst.Instruction.NumSrcRegs = 2;
 380       newInst.Src[0].Register.File = TGSI_FILE_TEMPORARY;
 381       newInst.Src[0].Register.Index = tmp0;
 382       newInst.Src[0].Register.SwizzleW = TGSI_SWIZZLE_Y;
 383       newInst.Src[1].Register.File = TGSI_FILE_TEMPORARY;
 384       newInst.Src[1].Register.Index = tmp0;
 385       newInst.Src[1].Register.SwizzleW = TGSI_SWIZZLE_Z;
 386       ctx->emit_instruction(ctx, &newInst);
 387
 388       /* SLE t0.y, t0.x, tex.z;  # bool b = distance <= k */
 389       newInst = tgsi_default_full_instruction();
 390       newInst.Instruction.Opcode = TGSI_OPCODE_SLE;
 391       newInst.Instruction.NumDstRegs = 1;
 392       newInst.Dst[0].Register.File = TGSI_FILE_TEMPORARY;
 393       newInst.Dst[0].Register.Index = tmp0;
 394       newInst.Dst[0].Register.WriteMask = TGSI_WRITEMASK_Y;
 395       newInst.Instruction.NumSrcRegs = 2;
 396       newInst.Src[0].Register.File = TGSI_FILE_TEMPORARY;
 397       newInst.Src[0].Register.Index = tmp0;
 398       newInst.Src[0].Register.SwizzleY = TGSI_SWIZZLE_X;
 399       newInst.Src[1].Register.File = TGSI_FILE_INPUT;
 400       newInst.Src[1].Register.Index = texInput;
 401       newInst.Src[1].Register.SwizzleY = TGSI_SWIZZLE_Z;
 402       ctx->emit_instruction(ctx, &newInst);
 403
 404       /* CMP t0.w, -t0.y, tex.w, t0.w;
 405        *  # if -t0.y < 0 then
 406        *       t0.w = 1
 407        *    else
 408        *       t0.w = t0.w
 409        */
 410       newInst = tgsi_default_full_instruction();
 411       newInst.Instruction.Opcode = TGSI_OPCODE_CMP;
 412       newInst.Instruction.NumDstRegs = 1;
 413       newInst.Dst[0].Register.File = TGSI_FILE_TEMPORARY;
 414       newInst.Dst[0].Register.Index = tmp0;
 415       newInst.Dst[0].Register.WriteMask = TGSI_WRITEMASK_W;
 416       newInst.Instruction.NumSrcRegs = 3;
 417       newInst.Src[0].Register.File = TGSI_FILE_TEMPORARY;
 418       newInst.Src[0].Register.Index = tmp0;
 419       newInst.Src[0].Register.SwizzleX = TGSI_SWIZZLE_Y;
 420       newInst.Src[0].Register.SwizzleY = TGSI_SWIZZLE_Y;
 421       newInst.Src[0].Register.SwizzleZ = TGSI_SWIZZLE_Y;
 422       newInst.Src[0].Register.SwizzleW = TGSI_SWIZZLE_Y;
 423       newInst.Src[0].Register.Negate = 1;
 424       newInst.Src[1].Register.File = TGSI_FILE_INPUT;
 425       newInst.Src[1].Register.Index = texInput;
 426       newInst.Src[1].Register.SwizzleX = TGSI_SWIZZLE_W;
 427       newInst.Src[1].Register.SwizzleY = TGSI_SWIZZLE_W;
 428       newInst.Src[1].Register.SwizzleZ = TGSI_SWIZZLE_W;
 429       newInst.Src[1].Register.SwizzleW = TGSI_SWIZZLE_W;
 430       newInst.Src[2].Register.File = TGSI_FILE_TEMPORARY;
 431       newInst.Src[2].Register.Index = tmp0;
 432       newInst.Src[2].Register.SwizzleX = TGSI_SWIZZLE_W;
 433       newInst.Src[2].Register.SwizzleY = TGSI_SWIZZLE_W;
 434       newInst.Src[2].Register.SwizzleZ = TGSI_SWIZZLE_W;
 435       newInst.Src[2].Register.SwizzleW = TGSI_SWIZZLE_W;
 436       ctx->emit_instruction(ctx, &newInst);
 437
 438    }
 439
 440    if (inst->Instruction.Opcode == TGSI_OPCODE_END) {
 441       /* add alpha modulation code at tail of program */
 442
 443       /* MOV result.color.xyz, colorTemp; */
 444       newInst = tgsi_default_full_instruction();
 445       newInst.Instruction.Opcode = TGSI_OPCODE_MOV;
 446       newInst.Instruction.NumDstRegs = 1;
 447       newInst.Dst[0].Register.File = TGSI_FILE_OUTPUT;
 448       newInst.Dst[0].Register.Index = aactx->colorOutput;
 449       newInst.Dst[0].Register.WriteMask = TGSI_WRITEMASK_XYZ;
 450       newInst.Instruction.NumSrcRegs = 1;
 451       newInst.Src[0].Register.File = TGSI_FILE_TEMPORARY;
 452       newInst.Src[0].Register.Index = aactx->colorTemp;
 453       ctx->emit_instruction(ctx, &newInst);
 454
 455       /* MUL result.color.w, colorTemp, tmp0.w; */
 456       newInst = tgsi_default_full_instruction();
 457       newInst.Instruction.Opcode = TGSI_OPCODE_MUL;
 458       newInst.Instruction.NumDstRegs = 1;
 459       newInst.Dst[0].Register.File = TGSI_FILE_OUTPUT;
 460       newInst.Dst[0].Register.Index = aactx->colorOutput;
 461       newInst.Dst[0].Register.WriteMask = TGSI_WRITEMASK_W;
 462       newInst.Instruction.NumSrcRegs = 2;
 463       newInst.Src[0].Register.File = TGSI_FILE_TEMPORARY;
 464       newInst.Src[0].Register.Index = aactx->colorTemp;
 465       newInst.Src[1].Register.File = TGSI_FILE_TEMPORARY;
 466       newInst.Src[1].Register.Index = aactx->tmp0;
 467       ctx->emit_instruction(ctx, &newInst);
 468    }
 469    else {
 470       /* Not an END instruction.
 471        * Look for writes to result.color and replace with colorTemp reg.
 472        */
 473       uint i;
 474
 475       for (i = 0; i < inst->Instruction.NumDstRegs; i++) {
 476          struct tgsi_full_dst_register *dst = &inst->Dst[i];
 477          if (dst->Register.File == TGSI_FILE_OUTPUT &&
 478              dst->Register.Index == aactx->colorOutput) {
 479             dst->Register.File = TGSI_FILE_TEMPORARY;
 480             dst->Register.Index = aactx->colorTemp;
 481          }
 482       }
 483    }
 484
 485    ctx->emit_instruction(ctx, inst);
 486 }
 487
 488
 489 /**
 490  * Generate the frag shader we'll use for drawing AA points.
 491  * This will be the user's shader plus some texture/modulate instructions.
 492  */
 493 static boolean
 494 generate_aapoint_fs(struct aapoint_stage *aapoint)
 495 {
 496    const struct pipe_shader_state *orig_fs = &aapoint->fs->state;
 497    struct pipe_shader_state aapoint_fs;
 498    struct aa_transform_context transform;
 499    const uint newLen = tgsi_num_tokens(orig_fs->tokens) + NUM_NEW_TOKENS;
 500    struct pipe_context *pipe = aapoint->stage.draw->pipe;
 501
 502    aapoint_fs = *orig_fs; /* copy to init */
 503    aapoint_fs.tokens = tgsi_alloc_tokens(newLen);
 504    if (aapoint_fs.tokens == NULL)
 505       return FALSE;
 506
 507    memset(&transform, 0, sizeof(transform));
 508    transform.colorOutput = -1;
 509    transform.maxInput = -1;
 510    transform.maxGeneric = -1;
 511    transform.colorTemp = -1;
 512    transform.tmp0 = -1;
 513    transform.firstInstruction = TRUE;
 514    transform.base.transform_instruction = aa_transform_inst;
 515    transform.base.transform_declaration = aa_transform_decl;
 516
 517    tgsi_transform_shader(orig_fs->tokens,
 518                          (struct tgsi_token *) aapoint_fs.tokens,
 519                          newLen, &transform.base);
 520
 521 #if 0 /* DEBUG */
 522    printf("draw_aapoint, orig shader:\n");
 523    tgsi_dump(orig_fs->tokens, 0);
 524    printf("draw_aapoint, new shader:\n");
 525    tgsi_dump(aapoint_fs.tokens, 0);
 526 #endif
 527
 528    aapoint->fs->aapoint_fs
 529       = aapoint->driver_create_fs_state(pipe, &aapoint_fs);
 530    if (aapoint->fs->aapoint_fs == NULL)
 531       goto fail;
 532
 533    aapoint->fs->generic_attrib = transform.maxGeneric + 1;
 534    FREE((void *)aapoint_fs.tokens);
 535    return TRUE;
 536
 537 fail:
 538    FREE((void *)aapoint_fs.tokens);
 539    return FALSE;
 540 }
 541
 542
 543 /**
 544  * When we're about to draw our first AA point in a batch, this function is
 545  * called to tell the driver to bind our modified fragment shader.
 546  */
 547 static boolean
 548 bind_aapoint_fragment_shader(struct aapoint_stage *aapoint)
 549 {
 550    struct draw_context *draw = aapoint->stage.draw;
 551    struct pipe_context *pipe = draw->pipe;
 552
 553    if (!aapoint->fs->aapoint_fs &&
 554        !generate_aapoint_fs(aapoint))
 555       return FALSE;
 556
 557    draw->suspend_flushing = TRUE;
 558    aapoint->driver_bind_fs_state(pipe, aapoint->fs->aapoint_fs);
 559    draw->suspend_flushing = FALSE;
 560
 561    return TRUE;
 562 }
 563
 564
 565
 566 static INLINE struct aapoint_stage *
 567 aapoint_stage( struct draw_stage *stage )
 568 {
 569    return (struct aapoint_stage *) stage;
 570 }
 571
 572
 573
 574
 575 /**
 576  * Draw an AA point by drawing a quad.
 577  */
 578 static void
 579 aapoint_point(struct draw_stage *stage, struct prim_header *header)
 580 {
 581    const struct aapoint_stage *aapoint = aapoint_stage(stage);
 582    struct prim_header tri;
 583    struct vertex_header *v[4];
 584    const uint tex_slot = aapoint->tex_slot;
 585    const uint pos_slot = aapoint->pos_slot;
 586    float radius, *pos, *tex;
 587    uint i;
 588    float k;
 589
 590    if (aapoint->psize_slot >= 0) {
 591       radius = 0.5f * header->v[0]->data[aapoint->psize_slot][0];
 592    }
 593    else {
 594       radius = aapoint->radius;
 595    }
 596
 597    /*
 598     * Note: the texcoords (generic attrib, really) we use are special:
 599     * The S and T components simply vary from -1 to +1.
 600     * The R component is k, below.
 601     * The Q component is 1.0 and will used as a handy constant in the
 602     * fragment shader.
 603     */
 604
 605    /*
 606     * k is the threshold distance from the point's center at which
 607     * we begin alpha attenuation (the coverage value).
 608     * Operating within a unit circle, we'll compute the fragment's
 609     * distance 'd' from the center point using the texcoords.
 610     * IF d > 1.0 THEN
 611     *    KILL fragment
 612     * ELSE IF d > k THEN
 613     *    compute coverage in [0,1] proportional to d in [k, 1].
 614     * ELSE
 615     *    coverage = 1.0;  // full coverage
 616     * ENDIF
 617     *
 618     * Note: the ELSEIF and ELSE clauses are actually implemented with CMP to
 619     * avoid using IF/ELSE/ENDIF TGSI opcodes.
 620     */
 621
 622 #if !NORMALIZE
 623    k = 1.0f / radius;
 624    k = 1.0f - 2.0f * k + k * k;
 625 #else
 626    k = 1.0f - 1.0f / radius;
 627 #endif
 628
 629    /* allocate/dup new verts */
 630    for (i = 0; i < 4; i++) {
 631       v[i] = dup_vert(stage, header->v[0], i);
 632    }
 633
 634    /* new verts */
 635    pos = v[0]->data[pos_slot];
 636    pos[0] -= radius;
 637    pos[1] -= radius;
 638
 639    pos = v[1]->data[pos_slot];
 640    pos[0] += radius;
 641    pos[1] -= radius;
 642
 643    pos = v[2]->data[pos_slot];
 644    pos[0] += radius;
 645    pos[1] += radius;
 646
 647    pos = v[3]->data[pos_slot];
 648    pos[0] -= radius;
 649    pos[1] += radius;
 650
 651    /* new texcoords */
 652    tex = v[0]->data[tex_slot];
 653    ASSIGN_4V(tex, -1, -1, k, 1);
 654
 655    tex = v[1]->data[tex_slot];
 656    ASSIGN_4V(tex,  1, -1, k, 1);
 657
 658    tex = v[2]->data[tex_slot];
 659    ASSIGN_4V(tex,  1,  1, k, 1);
 660
 661    tex = v[3]->data[tex_slot];
 662    ASSIGN_4V(tex, -1,  1, k, 1);
 663
 664    /* emit 2 tris for the quad strip */
 665    tri.v[0] = v[0];
 666    tri.v[1] = v[1];
 667    tri.v[2] = v[2];
 668    stage->next->tri( stage->next, &tri );
 669
 670    tri.v[0] = v[0];
 671    tri.v[1] = v[2];
 672    tri.v[2] = v[3];
 673    stage->next->tri( stage->next, &tri );
 674 }
 675
 676
 677 static void
 678 aapoint_first_point(struct draw_stage *stage, struct prim_header *header)
 679 {
 680    auto struct aapoint_stage *aapoint = aapoint_stage(stage);
 681    struct draw_context *draw = stage->draw;
 682    struct pipe_context *pipe = draw->pipe;
 683    const struct pipe_rasterizer_state *rast = draw->rasterizer;
 684    void *r;
 685
 686    assert(draw->rasterizer->point_smooth);
 687
 688    if (draw->rasterizer->point_size <= 2.0)
 689       aapoint->radius = 1.0;
 690    else
 691       aapoint->radius = 0.5f * draw->rasterizer->point_size;
 692
 693    /*
 694     * Bind (generate) our fragprog.
 695     */
 696    bind_aapoint_fragment_shader(aapoint);
 697
 698    /* update vertex attrib info */
 699    aapoint->tex_slot = draw_current_shader_outputs(draw);
 700    assert(aapoint->tex_slot > 0); /* output[0] is vertex pos */
 701
 702    aapoint->pos_slot = draw_current_shader_position_output(draw);
 703
 704    /* allocate the extra post-transformed vertex attribute */
 705    (void) draw_alloc_extra_vertex_attrib(draw, TGSI_SEMANTIC_GENERIC,
 706                                          aapoint->fs->generic_attrib);
 707
 708    /* find psize slot in post-transform vertex */
 709    aapoint->psize_slot = -1;
 710    if (draw->rasterizer->point_size_per_vertex) {
 711       /* find PSIZ vertex output */
 712       const struct draw_vertex_shader *vs = draw->vs.vertex_shader;
 713       uint i;
 714       for (i = 0; i < vs->info.num_outputs; i++) {
 715          if (vs->info.output_semantic_name[i] == TGSI_SEMANTIC_PSIZE) {
 716             aapoint->psize_slot = i;
 717             break;
 718          }
 719       }
 720    }
 721
 722    draw->suspend_flushing = TRUE;
 723
 724    /* Disable triangle culling, stippling, unfilled mode etc. */
 725    r = draw_get_rasterizer_no_cull(draw, rast->scissor, rast->flatshade);
 726    pipe->bind_rasterizer_state(pipe, r);
 727
 728    draw->suspend_flushing = FALSE;
 729
 730    /* now really draw first point */
 731    stage->point = aapoint_point;
 732    stage->point(stage, header);
 733 }
 734
 735
 736 static void
 737 aapoint_flush(struct draw_stage *stage, unsigned flags)
 738 {
 739    struct draw_context *draw = stage->draw;
 740    struct aapoint_stage *aapoint = aapoint_stage(stage);
 741    struct pipe_context *pipe = draw->pipe;
 742
 743    stage->point = aapoint_first_point;
 744    stage->next->flush( stage->next, flags );
 745
 746    /* restore original frag shader */
 747    draw->suspend_flushing = TRUE;
 748    aapoint->driver_bind_fs_state(pipe, aapoint->fs->driver_fs);
 749
 750    /* restore original rasterizer state */
 751    if (draw->rast_handle) {
 752       pipe->bind_rasterizer_state(pipe, draw->rast_handle);
 753    }
 754
 755    draw->suspend_flushing = FALSE;
 756
 757    draw_remove_extra_vertex_attribs(draw);
 758 }
 759
 760
 761 static void
 762 aapoint_reset_stipple_counter(struct draw_stage *stage)
 763 {
 764    stage->next->reset_stipple_counter( stage->next );
 765 }
 766
 767
 768 static void
 769 aapoint_destroy(struct draw_stage *stage)
 770 {
 771    draw_free_temp_verts( stage );
 772    FREE( stage );
 773 }
 774
 775
 776 static struct aapoint_stage *
 777 draw_aapoint_stage(struct draw_context *draw)
 778 {
 779    struct aapoint_stage *aapoint = CALLOC_STRUCT(aapoint_stage);
 780    if (aapoint == NULL)
 781       goto fail;
 782
 783    aapoint->stage.draw = draw;
 784    aapoint->stage.name = "aapoint";
 785    aapoint->stage.next = NULL;
 786    aapoint->stage.point = aapoint_first_point;
 787    aapoint->stage.line = draw_pipe_passthrough_line;
 788    aapoint->stage.tri = draw_pipe_passthrough_tri;
 789    aapoint->stage.flush = aapoint_flush;
 790    aapoint->stage.reset_stipple_counter = aapoint_reset_stipple_counter;
 791    aapoint->stage.destroy = aapoint_destroy;
 792
 793    if (!draw_alloc_temp_verts( &aapoint->stage, 4 ))
 794       goto fail;
 795
 796    return aapoint;
 797
 798  fail:
 799    if (aapoint)
 800       aapoint->stage.destroy(&aapoint->stage);
 801
 802    return NULL;
 803
 804 }
 805
 806
 807 static struct aapoint_stage *
 808 aapoint_stage_from_pipe(struct pipe_context *pipe)
 809 {
 810    struct draw_context *draw = (struct draw_context *) pipe->draw;
 811    return aapoint_stage(draw->pipeline.aapoint);
 812 }
 813
 814
 815 /**
 816  * This function overrides the driver's create_fs_state() function and
 817  * will typically be called by the state tracker.
 818  */
 819 static void *
 820 aapoint_create_fs_state(struct pipe_context *pipe,
 821                        const struct pipe_shader_state *fs)
 822 {
 823    struct aapoint_stage *aapoint = aapoint_stage_from_pipe(pipe);
 824    struct aapoint_fragment_shader *aafs = CALLOC_STRUCT(aapoint_fragment_shader);
 825    if (aafs == NULL)
 826       return NULL;
 827
 828    aafs->state = *fs;
 829
 830    /* pass-through */
 831    aafs->driver_fs = aapoint->driver_create_fs_state(pipe, fs);
 832
 833    return aafs;
 834 }
 835
 836
 837 static void
 838 aapoint_bind_fs_state(struct pipe_context *pipe, void *fs)
 839 {
 840    struct aapoint_stage *aapoint = aapoint_stage_from_pipe(pipe);
 841    struct aapoint_fragment_shader *aafs = (struct aapoint_fragment_shader *) fs;
 842    /* save current */
 843    aapoint->fs = aafs;
 844    /* pass-through */
 845    aapoint->driver_bind_fs_state(pipe,
 846                                  (aafs ? aafs->driver_fs : NULL));
 847 }
 848
 849
 850 static void
 851 aapoint_delete_fs_state(struct pipe_context *pipe, void *fs)
 852 {
 853    struct aapoint_stage *aapoint = aapoint_stage_from_pipe(pipe);
 854    struct aapoint_fragment_shader *aafs = (struct aapoint_fragment_shader *) fs;
 855
 856    /* pass-through */
 857    aapoint->driver_delete_fs_state(pipe, aafs->driver_fs);
 858
 859    if (aafs->aapoint_fs)
 860       aapoint->driver_delete_fs_state(pipe, aafs->aapoint_fs);
 861
 862    FREE(aafs);
 863 }
 864
 865
 866 /**
 867  * Called by drivers that want to install this AA point prim stage
 868  * into the draw module's pipeline.  This will not be used if the
 869  * hardware has native support for AA points.
 870  */
 871 boolean
 872 draw_install_aapoint_stage(struct draw_context *draw,
 873                            struct pipe_context *pipe)
 874 {
 875    struct aapoint_stage *aapoint;
 876
 877    pipe->draw = (void *) draw;
 878
 879    /*
 880     * Create / install AA point drawing / prim stage
 881     */
 882    aapoint = draw_aapoint_stage( draw );
 883    if (aapoint == NULL)
 884       return FALSE;
 885
 886    /* save original driver functions */
 887    aapoint->driver_create_fs_state = pipe->create_fs_state;
 888    aapoint->driver_bind_fs_state = pipe->bind_fs_state;
 889    aapoint->driver_delete_fs_state = pipe->delete_fs_state;
 890
 891    /* override the driver's functions */
 892    pipe->create_fs_state = aapoint_create_fs_state;
 893    pipe->bind_fs_state = aapoint_bind_fs_state;
 894    pipe->delete_fs_state = aapoint_delete_fs_state;
 895
 896    draw->pipeline.aapoint = &aapoint->stage;
 897
 898    return TRUE;
 899 }