src/gallium/auxiliary/draw/draw_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_util.h"
  42 #include "pipe/p_inlines.h"
  43 #include "pipe/p_context.h"
  44 #include "pipe/p_defines.h"
  45 #include "pipe/p_shader_tokens.h"
  46
  47 #include "tgsi/util/tgsi_transform.h"
  48 #include "tgsi/util/tgsi_dump.h"
  49
  50 #include "draw_context.h"
  51 #include "draw_private.h"
  52
  53
  54 /*
  55  * Enabling NORMALIZE might give _slightly_ better results.
  56  * Basically, it controls whether we compute distance as d=sqrt(x*x+y*y) or
  57  * d=x*x+y*y.  Since we're working with a unit circle, the later seems
  58  * close enough and saves some costly instructions.
  59  */
  60 #define NORMALIZE 0
  61
  62
  63 /**
  64  * Subclass of pipe_shader_state to carry extra fragment shader info.
  65  */
  66 struct aapoint_fragment_shader
  67 {
  68    struct pipe_shader_state state;
  69    void *driver_fs;   /**< the regular shader */
  70    void *aapoint_fs;  /**< the aa point-augmented shader */
  71    int generic_attrib; /**< The generic input attrib/texcoord we'll use */
  72 };
  73
  74
  75 /**
  76  * Subclass of draw_stage
  77  */
  78 struct aapoint_stage
  79 {
  80    struct draw_stage stage;
  81
  82    int psize_slot;
  83    float radius;
  84
  85    /** this is the vertex attrib slot for the new texcoords */
  86    uint tex_slot;
  87
  88    /*
  89     * Currently bound state
  90     */
  91    struct aapoint_fragment_shader *fs;
  92
  93    /*
  94     * Driver interface/override functions
  95     */
  96    void * (*driver_create_fs_state)(struct pipe_context *,
  97                                     const struct pipe_shader_state *);
  98    void (*driver_bind_fs_state)(struct pipe_context *, void *);
  99    void (*driver_delete_fs_state)(struct pipe_context *, void *);
 100
 101    struct pipe_context *pipe;
 102 };
 103
 104
 105
 106 /**
 107  * Subclass of tgsi_transform_context, used for transforming the
 108  * user's fragment shader to add the special AA instructions.
 109  */
 110 struct aa_transform_context {
 111    struct tgsi_transform_context base;
 112    uint tempsUsed;  /**< bitmask */
 113    int colorOutput; /**< which output is the primary color */
 114    int maxInput, maxGeneric;  /**< max input index found */
 115    int tmp0, colorTemp;  /**< temp registers */
 116    boolean firstInstruction;
 117 };
 118
 119
 120 /**
 121  * TGSI declaration transform callback.
 122  * Look for two free temp regs and available input reg for new texcoords.
 123  */
 124 static void
 125 aa_transform_decl(struct tgsi_transform_context *ctx,
 126                   struct tgsi_full_declaration *decl)
 127 {
 128    struct aa_transform_context *aactx = (struct aa_transform_context *) ctx;
 129
 130    if (decl->Declaration.File == TGSI_FILE_OUTPUT &&
 131        decl->Semantic.SemanticName == TGSI_SEMANTIC_COLOR &&
 132        decl->Semantic.SemanticIndex == 0) {
 133       aactx->colorOutput = decl->u.DeclarationRange.First;
 134    }
 135    else if (decl->Declaration.File == TGSI_FILE_INPUT) {
 136       if ((int) decl->u.DeclarationRange.Last > aactx->maxInput)
 137          aactx->maxInput = decl->u.DeclarationRange.Last;
 138       if (decl->Semantic.SemanticName == TGSI_SEMANTIC_GENERIC &&
 139            (int) decl->Semantic.SemanticIndex > aactx->maxGeneric) {
 140          aactx->maxGeneric = decl->Semantic.SemanticIndex;
 141       }
 142    }
 143    else if (decl->Declaration.File == TGSI_FILE_TEMPORARY) {
 144       uint i;
 145       for (i = decl->u.DeclarationRange.First;
 146            i <= decl->u.DeclarationRange.Last; i++) {
 147          aactx->tempsUsed |= (1 << i);
 148       }
 149    }
 150
 151    ctx->emit_declaration(ctx, decl);
 152 }
 153
 154
 155 /**
 156  * TGSI instruction transform callback.
 157  * Replace writes to result.color w/ a temp reg.
 158  * Upon END instruction, insert texture sampling code for antialiasing.
 159  */
 160 static void
 161 aa_transform_inst(struct tgsi_transform_context *ctx,
 162                   struct tgsi_full_instruction *inst)
 163 {
 164    struct aa_transform_context *aactx = (struct aa_transform_context *) ctx;
 165    struct tgsi_full_instruction newInst;
 166
 167    if (aactx->firstInstruction) {
 168       /* emit our new declarations before the first instruction */
 169
 170       struct tgsi_full_declaration decl;
 171       const int texInput = aactx->maxInput + 1;
 172       int tmp0;
 173       uint i;
 174
 175       /* find two free temp regs */
 176       for (i = 0; i < 32; i++) {
 177          if ((aactx->tempsUsed & (1 << i)) == 0) {
 178             /* found a free temp */
 179             if (aactx->tmp0 < 0)
 180                aactx->tmp0 = i;
 181             else if (aactx->colorTemp < 0)
 182                aactx->colorTemp = i;
 183             else
 184                break;
 185          }
 186       }
 187
 188       assert(aactx->colorTemp != aactx->tmp0);
 189
 190       tmp0 = aactx->tmp0;
 191
 192       /* declare new generic input/texcoord */
 193       decl = tgsi_default_full_declaration();
 194       decl.Declaration.File = TGSI_FILE_INPUT;
 195       decl.Declaration.Semantic = 1;
 196       decl.Semantic.SemanticName = TGSI_SEMANTIC_GENERIC;
 197       decl.Semantic.SemanticIndex = aactx->maxGeneric + 1;
 198       decl.Declaration.Interpolate = 1;
 199       /* XXX this could be linear... */
 200       decl.Interpolation.Interpolate = TGSI_INTERPOLATE_PERSPECTIVE;
 201       decl.u.DeclarationRange.First =
 202       decl.u.DeclarationRange.Last = texInput;
 203       ctx->emit_declaration(ctx, &decl);
 204
 205       /* declare new temp regs */
 206       decl = tgsi_default_full_declaration();
 207       decl.Declaration.File = TGSI_FILE_TEMPORARY;
 208       decl.u.DeclarationRange.First =
 209       decl.u.DeclarationRange.Last = tmp0;
 210       ctx->emit_declaration(ctx, &decl);
 211
 212       decl = tgsi_default_full_declaration();
 213       decl.Declaration.File = TGSI_FILE_TEMPORARY;
 214       decl.u.DeclarationRange.First =
 215       decl.u.DeclarationRange.Last = aactx->colorTemp;
 216       ctx->emit_declaration(ctx, &decl);
 217
 218       aactx->firstInstruction = FALSE;
 219
 220
 221       /*
 222        * Emit code to compute fragment coverage, kill if outside point radius
 223        *
 224        * Temp reg0 usage:
 225        *  t0.x = distance of fragment from center point
 226        *  t0.y = boolean, is t0.x > 1.0, also misc temp usage
 227        *  t0.z = temporary for computing 1/(1-k) value
 228        *  t0.w = final coverage value
 229        */
 230
 231       /* MUL t0.xy, tex, tex;  # compute x^2, y^2 */
 232       newInst = tgsi_default_full_instruction();
 233       newInst.Instruction.Opcode = TGSI_OPCODE_MUL;
 234       newInst.Instruction.NumDstRegs = 1;
 235       newInst.FullDstRegisters[0].DstRegister.File = TGSI_FILE_TEMPORARY;
 236       newInst.FullDstRegisters[0].DstRegister.Index = tmp0;
 237       newInst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_XY;
 238       newInst.Instruction.NumSrcRegs = 2;
 239       newInst.FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_INPUT;
 240       newInst.FullSrcRegisters[0].SrcRegister.Index = texInput;
 241       newInst.FullSrcRegisters[1].SrcRegister.File = TGSI_FILE_INPUT;
 242       newInst.FullSrcRegisters[1].SrcRegister.Index = texInput;
 243       ctx->emit_instruction(ctx, &newInst);
 244
 245       /* ADD t0.x, t0.x, t0.y;  # x^2 + y^2 */
 246       newInst = tgsi_default_full_instruction();
 247       newInst.Instruction.Opcode = TGSI_OPCODE_ADD;
 248       newInst.Instruction.NumDstRegs = 1;
 249       newInst.FullDstRegisters[0].DstRegister.File = TGSI_FILE_TEMPORARY;
 250       newInst.FullDstRegisters[0].DstRegister.Index = tmp0;
 251       newInst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_X;
 252       newInst.Instruction.NumSrcRegs = 2;
 253       newInst.FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_TEMPORARY;
 254       newInst.FullSrcRegisters[0].SrcRegister.Index = tmp0;
 255       newInst.FullSrcRegisters[0].SrcRegister.SwizzleX = TGSI_SWIZZLE_X;
 256       newInst.FullSrcRegisters[1].SrcRegister.File = TGSI_FILE_TEMPORARY;
 257       newInst.FullSrcRegisters[1].SrcRegister.Index = tmp0;
 258       newInst.FullSrcRegisters[1].SrcRegister.SwizzleX = TGSI_SWIZZLE_Y;
 259       ctx->emit_instruction(ctx, &newInst);
 260
 261 #if NORMALIZE  /* OPTIONAL normalization of length */
 262       /* RSQ t0.x, t0.x; */
 263       newInst = tgsi_default_full_instruction();
 264       newInst.Instruction.Opcode = TGSI_OPCODE_RSQ;
 265       newInst.Instruction.NumDstRegs = 1;
 266       newInst.FullDstRegisters[0].DstRegister.File = TGSI_FILE_TEMPORARY;
 267       newInst.FullDstRegisters[0].DstRegister.Index = tmp0;
 268       newInst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_X;
 269       newInst.Instruction.NumSrcRegs = 1;
 270       newInst.FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_TEMPORARY;
 271       newInst.FullSrcRegisters[0].SrcRegister.Index = tmp0;
 272       ctx->emit_instruction(ctx, &newInst);
 273
 274       /* RCP t0.x, t0.x; */
 275       newInst = tgsi_default_full_instruction();
 276       newInst.Instruction.Opcode = TGSI_OPCODE_RCP;
 277       newInst.Instruction.NumDstRegs = 1;
 278       newInst.FullDstRegisters[0].DstRegister.File = TGSI_FILE_TEMPORARY;
 279       newInst.FullDstRegisters[0].DstRegister.Index = tmp0;
 280       newInst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_X;
 281       newInst.Instruction.NumSrcRegs = 1;
 282       newInst.FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_TEMPORARY;
 283       newInst.FullSrcRegisters[0].SrcRegister.Index = tmp0;
 284       ctx->emit_instruction(ctx, &newInst);
 285 #endif
 286
 287       /* SGT t0.y, t0.xxxx, t0.wwww;  # bool b = d > 1 (NOTE t0.w == 1) */
 288       newInst = tgsi_default_full_instruction();
 289       newInst.Instruction.Opcode = TGSI_OPCODE_SGT;
 290       newInst.Instruction.NumDstRegs = 1;
 291       newInst.FullDstRegisters[0].DstRegister.File = TGSI_FILE_TEMPORARY;
 292       newInst.FullDstRegisters[0].DstRegister.Index = tmp0;
 293       newInst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_Y;
 294       newInst.Instruction.NumSrcRegs = 2;
 295       newInst.FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_TEMPORARY;
 296       newInst.FullSrcRegisters[0].SrcRegister.Index = tmp0;
 297       newInst.FullSrcRegisters[0].SrcRegister.SwizzleY = TGSI_SWIZZLE_X;
 298       newInst.FullSrcRegisters[1].SrcRegister.File = TGSI_FILE_INPUT;
 299       newInst.FullSrcRegisters[1].SrcRegister.Index = texInput;
 300       newInst.FullSrcRegisters[1].SrcRegister.SwizzleY = TGSI_SWIZZLE_W;
 301       ctx->emit_instruction(ctx, &newInst);
 302
 303       /* KILP -t0.yyyy;   # if b, KILL */
 304       newInst = tgsi_default_full_instruction();
 305       newInst.Instruction.Opcode = TGSI_OPCODE_KILP;
 306       newInst.Instruction.NumDstRegs = 0;
 307       newInst.Instruction.NumSrcRegs = 1;
 308       newInst.FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_TEMPORARY;
 309       newInst.FullSrcRegisters[0].SrcRegister.Index = tmp0;
 310       newInst.FullSrcRegisters[0].SrcRegister.SwizzleX = TGSI_SWIZZLE_Y;
 311       newInst.FullSrcRegisters[0].SrcRegister.SwizzleY = TGSI_SWIZZLE_Y;
 312       newInst.FullSrcRegisters[0].SrcRegister.SwizzleZ = TGSI_SWIZZLE_Y;
 313       newInst.FullSrcRegisters[0].SrcRegister.SwizzleW = TGSI_SWIZZLE_Y;
 314       newInst.FullSrcRegisters[0].SrcRegister.Negate = 1;
 315       ctx->emit_instruction(ctx, &newInst);
 316
 317
 318       /* compute coverage factor = (1-d)/(1-k) */
 319
 320       /* SUB t0.z, tex.w, tex.z;  # m = 1 - k */
 321       newInst = tgsi_default_full_instruction();
 322       newInst.Instruction.Opcode = TGSI_OPCODE_SUB;
 323       newInst.Instruction.NumDstRegs = 1;
 324       newInst.FullDstRegisters[0].DstRegister.File = TGSI_FILE_TEMPORARY;
 325       newInst.FullDstRegisters[0].DstRegister.Index = tmp0;
 326       newInst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_Z;
 327       newInst.Instruction.NumSrcRegs = 2;
 328       newInst.FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_INPUT;
 329       newInst.FullSrcRegisters[0].SrcRegister.Index = texInput;
 330       newInst.FullSrcRegisters[0].SrcRegister.SwizzleZ = TGSI_SWIZZLE_W;
 331       newInst.FullSrcRegisters[1].SrcRegister.File = TGSI_FILE_INPUT;
 332       newInst.FullSrcRegisters[1].SrcRegister.Index = texInput;
 333       newInst.FullSrcRegisters[1].SrcRegister.SwizzleZ = TGSI_SWIZZLE_Z;
 334       ctx->emit_instruction(ctx, &newInst);
 335
 336       /* RCP t0.z, t0.z;  # t0.z = 1 / m */
 337       newInst = tgsi_default_full_instruction();
 338       newInst.Instruction.Opcode = TGSI_OPCODE_RCP;
 339       newInst.Instruction.NumDstRegs = 1;
 340       newInst.FullDstRegisters[0].DstRegister.File = TGSI_FILE_TEMPORARY;
 341       newInst.FullDstRegisters[0].DstRegister.Index = tmp0;
 342       newInst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_Z;
 343       newInst.Instruction.NumSrcRegs = 1;
 344       newInst.FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_TEMPORARY;
 345       newInst.FullSrcRegisters[0].SrcRegister.Index = tmp0;
 346       newInst.FullSrcRegisters[0].SrcRegister.SwizzleX = TGSI_SWIZZLE_Z;
 347       ctx->emit_instruction(ctx, &newInst);
 348
 349       /* SUB t0.y, 1, t0.x;  # d = 1 - d */
 350       newInst = tgsi_default_full_instruction();
 351       newInst.Instruction.Opcode = TGSI_OPCODE_SUB;
 352       newInst.Instruction.NumDstRegs = 1;
 353       newInst.FullDstRegisters[0].DstRegister.File = TGSI_FILE_TEMPORARY;
 354       newInst.FullDstRegisters[0].DstRegister.Index = tmp0;
 355       newInst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_Y;
 356       newInst.Instruction.NumSrcRegs = 2;
 357       newInst.FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_INPUT;
 358       newInst.FullSrcRegisters[0].SrcRegister.Index = texInput;
 359       newInst.FullSrcRegisters[0].SrcRegister.SwizzleY = TGSI_SWIZZLE_W;
 360       newInst.FullSrcRegisters[1].SrcRegister.File = TGSI_FILE_TEMPORARY;
 361       newInst.FullSrcRegisters[1].SrcRegister.Index = tmp0;
 362       newInst.FullSrcRegisters[1].SrcRegister.SwizzleY = TGSI_SWIZZLE_X;
 363       ctx->emit_instruction(ctx, &newInst);
 364
 365       /* MUL t0.w, t0.y, t0.z;   # coverage = d * m */
 366       newInst = tgsi_default_full_instruction();
 367       newInst.Instruction.Opcode = TGSI_OPCODE_MUL;
 368       newInst.Instruction.NumDstRegs = 1;
 369       newInst.FullDstRegisters[0].DstRegister.File = TGSI_FILE_TEMPORARY;
 370       newInst.FullDstRegisters[0].DstRegister.Index = tmp0;
 371       newInst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_W;
 372       newInst.Instruction.NumSrcRegs = 2;
 373       newInst.FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_TEMPORARY;
 374       newInst.FullSrcRegisters[0].SrcRegister.Index = tmp0;
 375       newInst.FullSrcRegisters[0].SrcRegister.SwizzleW = TGSI_SWIZZLE_Y;
 376       newInst.FullSrcRegisters[1].SrcRegister.File = TGSI_FILE_TEMPORARY;
 377       newInst.FullSrcRegisters[1].SrcRegister.Index = tmp0;
 378       newInst.FullSrcRegisters[1].SrcRegister.SwizzleW = TGSI_SWIZZLE_Z;
 379       ctx->emit_instruction(ctx, &newInst);
 380
 381       /* SLE t0.y, t0.x, tex.z;  # bool b = distance <= k */
 382       newInst = tgsi_default_full_instruction();
 383       newInst.Instruction.Opcode = TGSI_OPCODE_SLE;
 384       newInst.Instruction.NumDstRegs = 1;
 385       newInst.FullDstRegisters[0].DstRegister.File = TGSI_FILE_TEMPORARY;
 386       newInst.FullDstRegisters[0].DstRegister.Index = tmp0;
 387       newInst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_Y;
 388       newInst.Instruction.NumSrcRegs = 2;
 389       newInst.FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_TEMPORARY;
 390       newInst.FullSrcRegisters[0].SrcRegister.Index = tmp0;
 391       newInst.FullSrcRegisters[0].SrcRegister.SwizzleY = TGSI_SWIZZLE_X;
 392       newInst.FullSrcRegisters[1].SrcRegister.File = TGSI_FILE_INPUT;
 393       newInst.FullSrcRegisters[1].SrcRegister.Index = texInput;
 394       newInst.FullSrcRegisters[1].SrcRegister.SwizzleY = TGSI_SWIZZLE_Z;
 395       ctx->emit_instruction(ctx, &newInst);
 396
 397       /* CMP t0.w, -t0.y, tex.w, t0.w;
 398        *  # if -t0.y < 0 then
 399        *       t0.w = 1
 400        *    else
 401        *       t0.w = t0.w
 402        */
 403       newInst = tgsi_default_full_instruction();
 404       newInst.Instruction.Opcode = TGSI_OPCODE_CMP;
 405       newInst.Instruction.NumDstRegs = 1;
 406       newInst.FullDstRegisters[0].DstRegister.File = TGSI_FILE_TEMPORARY;
 407       newInst.FullDstRegisters[0].DstRegister.Index = tmp0;
 408       newInst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_W;
 409       newInst.Instruction.NumSrcRegs = 3;
 410       newInst.FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_TEMPORARY;
 411       newInst.FullSrcRegisters[0].SrcRegister.Index = tmp0;
 412       newInst.FullSrcRegisters[0].SrcRegister.SwizzleX = TGSI_SWIZZLE_Y;
 413       newInst.FullSrcRegisters[0].SrcRegister.SwizzleY = TGSI_SWIZZLE_Y;
 414       newInst.FullSrcRegisters[0].SrcRegister.SwizzleZ = TGSI_SWIZZLE_Y;
 415       newInst.FullSrcRegisters[0].SrcRegister.SwizzleW = TGSI_SWIZZLE_Y;
 416       newInst.FullSrcRegisters[0].SrcRegister.Negate = 1;
 417       newInst.FullSrcRegisters[1].SrcRegister.File = TGSI_FILE_INPUT;
 418       newInst.FullSrcRegisters[1].SrcRegister.Index = texInput;
 419       newInst.FullSrcRegisters[1].SrcRegister.SwizzleX = TGSI_SWIZZLE_W;
 420       newInst.FullSrcRegisters[1].SrcRegister.SwizzleY = TGSI_SWIZZLE_W;
 421       newInst.FullSrcRegisters[1].SrcRegister.SwizzleZ = TGSI_SWIZZLE_W;
 422       newInst.FullSrcRegisters[1].SrcRegister.SwizzleW = TGSI_SWIZZLE_W;
 423       newInst.FullSrcRegisters[2].SrcRegister.File = TGSI_FILE_TEMPORARY;
 424       newInst.FullSrcRegisters[2].SrcRegister.Index = tmp0;
 425       newInst.FullSrcRegisters[2].SrcRegister.SwizzleX = TGSI_SWIZZLE_W;
 426       newInst.FullSrcRegisters[2].SrcRegister.SwizzleY = TGSI_SWIZZLE_W;
 427       newInst.FullSrcRegisters[2].SrcRegister.SwizzleZ = TGSI_SWIZZLE_W;
 428       newInst.FullSrcRegisters[2].SrcRegister.SwizzleW = TGSI_SWIZZLE_W;
 429       ctx->emit_instruction(ctx, &newInst);
 430
 431    }
 432
 433    if (inst->Instruction.Opcode == TGSI_OPCODE_END) {
 434       /* add alpha modulation code at tail of program */
 435
 436       /* MOV result.color.xyz, colorTemp; */
 437       newInst = tgsi_default_full_instruction();
 438       newInst.Instruction.Opcode = TGSI_OPCODE_MOV;
 439       newInst.Instruction.NumDstRegs = 1;
 440       newInst.FullDstRegisters[0].DstRegister.File = TGSI_FILE_OUTPUT;
 441       newInst.FullDstRegisters[0].DstRegister.Index = aactx->colorOutput;
 442       newInst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_XYZ;
 443       newInst.Instruction.NumSrcRegs = 1;
 444       newInst.FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_TEMPORARY;
 445       newInst.FullSrcRegisters[0].SrcRegister.Index = aactx->colorTemp;
 446       ctx->emit_instruction(ctx, &newInst);
 447
 448       /* MUL result.color.w, colorTemp, tmp0.w; */
 449       newInst = tgsi_default_full_instruction();
 450       newInst.Instruction.Opcode = TGSI_OPCODE_MUL;
 451       newInst.Instruction.NumDstRegs = 1;
 452       newInst.FullDstRegisters[0].DstRegister.File = TGSI_FILE_OUTPUT;
 453       newInst.FullDstRegisters[0].DstRegister.Index = aactx->colorOutput;
 454       newInst.FullDstRegisters[0].DstRegister.WriteMask = TGSI_WRITEMASK_W;
 455       newInst.Instruction.NumSrcRegs = 2;
 456       newInst.FullSrcRegisters[0].SrcRegister.File = TGSI_FILE_TEMPORARY;
 457       newInst.FullSrcRegisters[0].SrcRegister.Index = aactx->colorTemp;
 458       newInst.FullSrcRegisters[1].SrcRegister.File = TGSI_FILE_TEMPORARY;
 459       newInst.FullSrcRegisters[1].SrcRegister.Index = aactx->tmp0;
 460       ctx->emit_instruction(ctx, &newInst);
 461    }
 462    else {
 463       /* Not an END instruction.
 464        * Look for writes to result.color and replace with colorTemp reg.
 465        */
 466       uint i;
 467
 468       for (i = 0; i < inst->Instruction.NumDstRegs; i++) {
 469          struct tgsi_full_dst_register *dst = &inst->FullDstRegisters[i];
 470          if (dst->DstRegister.File == TGSI_FILE_OUTPUT &&
 471              dst->DstRegister.Index == aactx->colorOutput) {
 472             dst->DstRegister.File = TGSI_FILE_TEMPORARY;
 473             dst->DstRegister.Index = aactx->colorTemp;
 474          }
 475       }
 476    }
 477
 478    ctx->emit_instruction(ctx, inst);
 479 }
 480
 481
 482 /**
 483  * Generate the frag shader we'll use for drawing AA lines.
 484  * This will be the user's shader plus some texture/modulate instructions.
 485  */
 486 static void
 487 generate_aapoint_fs(struct aapoint_stage *aapoint)
 488 {
 489    const struct pipe_shader_state *orig_fs = &aapoint->fs->state;
 490    struct pipe_shader_state aapoint_fs;
 491    struct aa_transform_context transform;
 492
 493 #define MAX 1000
 494
 495    aapoint_fs = *orig_fs; /* copy to init */
 496    aapoint_fs.tokens = MALLOC(sizeof(struct tgsi_token) * MAX);
 497
 498    memset(&transform, 0, sizeof(transform));
 499    transform.colorOutput = -1;
 500    transform.maxInput = -1;
 501    transform.maxGeneric = -1;
 502    transform.colorTemp = -1;
 503    transform.tmp0 = -1;
 504    transform.firstInstruction = TRUE;
 505    transform.base.transform_instruction = aa_transform_inst;
 506    transform.base.transform_declaration = aa_transform_decl;
 507
 508    tgsi_transform_shader(orig_fs->tokens,
 509                          (struct tgsi_token *) aapoint_fs.tokens,
 510                          MAX, &transform.base);
 511
 512 #if 0 /* DEBUG */
 513    printf("draw_aapoint, orig shader:\n");
 514    tgsi_dump(orig_fs->tokens, 0);
 515    printf("draw_aapoint, new shader:\n");
 516    tgsi_dump(aapoint_fs.tokens, 0);
 517 #endif
 518
 519    aapoint->fs->aapoint_fs
 520       = aapoint->driver_create_fs_state(aapoint->pipe, &aapoint_fs);
 521
 522    aapoint->fs->generic_attrib = transform.maxGeneric + 1;
 523 }
 524
 525
 526 /**
 527  * When we're about to draw our first AA line in a batch, this function is
 528  * called to tell the driver to bind our modified fragment shader.
 529  */
 530 static void
 531 bind_aapoint_fragment_shader(struct aapoint_stage *aapoint)
 532 {
 533    if (!aapoint->fs->aapoint_fs) {
 534       generate_aapoint_fs(aapoint);
 535    }
 536    aapoint->driver_bind_fs_state(aapoint->pipe, aapoint->fs->aapoint_fs);
 537 }
 538
 539
 540
 541 static INLINE struct aapoint_stage *
 542 aapoint_stage( struct draw_stage *stage )
 543 {
 544    return (struct aapoint_stage *) stage;
 545 }
 546
 547
 548 static void
 549 passthrough_line(struct draw_stage *stage, struct prim_header *header)
 550 {
 551    stage->next->line(stage->next, header);
 552 }
 553
 554
 555 static void
 556 passthrough_tri(struct draw_stage *stage, struct prim_header *header)
 557 {
 558    stage->next->tri(stage->next, header);
 559 }
 560
 561
 562 /**
 563  * Draw an AA point by drawing a quad.
 564  */
 565 static void
 566 aapoint_point(struct draw_stage *stage, struct prim_header *header)
 567 {
 568    const struct aapoint_stage *aapoint = aapoint_stage(stage);
 569    struct prim_header tri;
 570    struct vertex_header *v[4];
 571    uint texPos = aapoint->tex_slot;
 572    float radius, *pos, *tex;
 573    uint i;
 574    float k;
 575
 576    if (aapoint->psize_slot >= 0) {
 577       radius = 0.5f * header->v[0]->data[aapoint->psize_slot][0];
 578    }
 579    else {
 580       radius = aapoint->radius;
 581    }
 582
 583    /*
 584     * Note: the texcoords (generic attrib, really) we use are special:
 585     * The S and T components simply vary from -1 to +1.
 586     * The R component is k, below.
 587     * The Q component is 1.0 and will used as a handy constant in the
 588     * fragment shader.
 589     */
 590
 591    /*
 592     * k is the threshold distance from the point's center at which
 593     * we begin alpha attenuation (the coverage value).
 594     * Operating within a unit circle, we'll compute the fragment's
 595     * distance 'd' from the center point using the texcoords.
 596     * IF d > 1.0 THEN
 597     *    KILL fragment
 598     * ELSE IF d > k THEN
 599     *    compute coverage in [0,1] proportional to d in [k, 1].
 600     * ELSE
 601     *    coverage = 1.0;  // full coverage
 602     * ENDIF
 603     *
 604     * Note: the ELSEIF and ELSE clauses are actually implemented with CMP to
 605     * avoid using IF/ELSE/ENDIF TGSI opcodes.
 606     */
 607
 608 #if !NORMALIZE
 609    k = 1.0f / radius;
 610    k = 1.0f - 2.0f * k + k * k;
 611 #else
 612    k = 1.0f - 1.0f / radius;
 613 #endif
 614
 615    /* allocate/dup new verts */
 616    for (i = 0; i < 4; i++) {
 617       v[i] = dup_vert(stage, header->v[0], i);
 618    }
 619
 620    /* new verts */
 621    pos = v[0]->data[0];
 622    pos[0] -= radius;
 623    pos[1] -= radius;
 624
 625    pos = v[1]->data[0];
 626    pos[0] += radius;
 627    pos[1] -= radius;
 628
 629    pos = v[2]->data[0];
 630    pos[0] += radius;
 631    pos[1] += radius;
 632
 633    pos = v[3]->data[0];
 634    pos[0] -= radius;
 635    pos[1] += radius;
 636
 637    /* new texcoords */
 638    tex = v[0]->data[texPos];
 639    ASSIGN_4V(tex, -1, -1, k, 1);
 640
 641    tex = v[1]->data[texPos];
 642    ASSIGN_4V(tex,  1, -1, k, 1);
 643
 644    tex = v[2]->data[texPos];
 645    ASSIGN_4V(tex,  1,  1, k, 1);
 646
 647    tex = v[3]->data[texPos];
 648    ASSIGN_4V(tex, -1,  1, k, 1);
 649
 650    /* emit 2 tris for the quad strip */
 651    tri.v[0] = v[0];
 652    tri.v[1] = v[1];
 653    tri.v[2] = v[2];
 654    stage->next->tri( stage->next, &tri );
 655
 656    tri.v[0] = v[0];
 657    tri.v[1] = v[2];
 658    tri.v[2] = v[3];
 659    stage->next->tri( stage->next, &tri );
 660 }
 661
 662
 663 static void
 664 aapoint_first_point(struct draw_stage *stage, struct prim_header *header)
 665 {
 666    auto struct aapoint_stage *aapoint = aapoint_stage(stage);
 667    struct draw_context *draw = stage->draw;
 668
 669    assert(draw->rasterizer->point_smooth);
 670
 671    if (draw->rasterizer->point_size <= 2.0)
 672       aapoint->radius = 1.0;
 673    else
 674       aapoint->radius = 0.5f * draw->rasterizer->point_size;
 675
 676    /*
 677     * Bind (generate) our fragprog.
 678     */
 679    bind_aapoint_fragment_shader(aapoint);
 680
 681    /* update vertex attrib info */
 682    aapoint->tex_slot = draw->num_vs_outputs;
 683    assert(aapoint->tex_slot > 0); /* output[0] is vertex pos */
 684
 685    draw->extra_vp_outputs.semantic_name = TGSI_SEMANTIC_GENERIC;
 686    draw->extra_vp_outputs.semantic_index = aapoint->fs->generic_attrib;
 687    draw->extra_vp_outputs.slot = aapoint->tex_slot;
 688
 689    /* find psize slot in post-transform vertex */
 690    aapoint->psize_slot = -1;
 691    if (draw->rasterizer->point_size_per_vertex) {
 692       /* find PSIZ vertex output */
 693       const struct draw_vertex_shader *vs = draw->vertex_shader;
 694       uint i;
 695       for (i = 0; i < vs->info.num_outputs; i++) {
 696          if (vs->info.output_semantic_name[i] == TGSI_SEMANTIC_PSIZE) {
 697             aapoint->psize_slot = i;
 698             break;
 699          }
 700       }
 701    }
 702
 703    /* now really draw first line */
 704    stage->point = aapoint_point;
 705    stage->point(stage, header);
 706 }
 707
 708
 709 static void
 710 aapoint_flush(struct draw_stage *stage, unsigned flags)
 711 {
 712    struct draw_context *draw = stage->draw;
 713    struct aapoint_stage *aapoint = aapoint_stage(stage);
 714    struct pipe_context *pipe = aapoint->pipe;
 715
 716    stage->point = aapoint_first_point;
 717    stage->next->flush( stage->next, flags );
 718
 719    /* restore original frag shader */
 720    aapoint->driver_bind_fs_state(pipe, aapoint->fs->driver_fs);
 721
 722    draw->extra_vp_outputs.slot = 0;
 723 }
 724
 725
 726 static void
 727 aapoint_reset_stipple_counter(struct draw_stage *stage)
 728 {
 729    stage->next->reset_stipple_counter( stage->next );
 730 }
 731
 732
 733 static void
 734 aapoint_destroy(struct draw_stage *stage)
 735 {
 736    draw_free_temp_verts( stage );
 737    FREE( stage );
 738 }
 739
 740
 741 static struct aapoint_stage *
 742 draw_aapoint_stage(struct draw_context *draw)
 743 {
 744    struct aapoint_stage *aapoint = CALLOC_STRUCT(aapoint_stage);
 745
 746    draw_alloc_temp_verts( &aapoint->stage, 4 );
 747
 748    aapoint->stage.draw = draw;
 749    aapoint->stage.next = NULL;
 750    aapoint->stage.point = aapoint_first_point;
 751    aapoint->stage.line = passthrough_line;
 752    aapoint->stage.tri = passthrough_tri;
 753    aapoint->stage.flush = aapoint_flush;
 754    aapoint->stage.reset_stipple_counter = aapoint_reset_stipple_counter;
 755    aapoint->stage.destroy = aapoint_destroy;
 756
 757    return aapoint;
 758 }
 759
 760
 761 static struct aapoint_stage *
 762 aapoint_stage_from_pipe(struct pipe_context *pipe)
 763 {
 764    struct draw_context *draw = (struct draw_context *) pipe->draw;
 765    return aapoint_stage(draw->pipeline.aapoint);
 766 }
 767
 768
 769 /**
 770  * This function overrides the driver's create_fs_state() function and
 771  * will typically be called by the state tracker.
 772  */
 773 static void *
 774 aapoint_create_fs_state(struct pipe_context *pipe,
 775                        const struct pipe_shader_state *fs)
 776 {
 777    struct aapoint_stage *aapoint = aapoint_stage_from_pipe(pipe);
 778    struct aapoint_fragment_shader *aafs = CALLOC_STRUCT(aapoint_fragment_shader);
 779
 780    if (aafs) {
 781       aafs->state = *fs;
 782
 783       /* pass-through */
 784       aafs->driver_fs = aapoint->driver_create_fs_state(aapoint->pipe, fs);
 785    }
 786
 787    return aafs;
 788 }
 789
 790
 791 static void
 792 aapoint_bind_fs_state(struct pipe_context *pipe, void *fs)
 793 {
 794    struct aapoint_stage *aapoint = aapoint_stage_from_pipe(pipe);
 795    struct aapoint_fragment_shader *aafs = (struct aapoint_fragment_shader *) fs;
 796    /* save current */
 797    aapoint->fs = aafs;
 798    /* pass-through */
 799    aapoint->driver_bind_fs_state(aapoint->pipe,
 800                                  (aafs ? aafs->driver_fs : NULL));
 801 }
 802
 803
 804 static void
 805 aapoint_delete_fs_state(struct pipe_context *pipe, void *fs)
 806 {
 807    struct aapoint_stage *aapoint = aapoint_stage_from_pipe(pipe);
 808    struct aapoint_fragment_shader *aafs = (struct aapoint_fragment_shader *) fs;
 809    /* pass-through */
 810    aapoint->driver_delete_fs_state(aapoint->pipe, aafs->driver_fs);
 811    FREE(aafs);
 812 }
 813
 814
 815 /**
 816  * Called by drivers that want to install this AA point prim stage
 817  * into the draw module's pipeline.  This will not be used if the
 818  * hardware has native support for AA points.
 819  */
 820 void
 821 draw_install_aapoint_stage(struct draw_context *draw,
 822                            struct pipe_context *pipe)
 823 {
 824    struct aapoint_stage *aapoint;
 825
 826    pipe->draw = (void *) draw;
 827
 828    /*
 829     * Create / install AA point drawing / prim stage
 830     */
 831    aapoint = draw_aapoint_stage( draw );
 832    assert(aapoint);
 833    draw->pipeline.aapoint = &aapoint->stage;
 834
 835    aapoint->pipe = pipe;
 836
 837    /* save original driver functions */
 838    aapoint->driver_create_fs_state = pipe->create_fs_state;
 839    aapoint->driver_bind_fs_state = pipe->bind_fs_state;
 840    aapoint->driver_delete_fs_state = pipe->delete_fs_state;
 841
 842    /* override the driver's functions */
 843    pipe->create_fs_state = aapoint_create_fs_state;
 844    pipe->bind_fs_state = aapoint_bind_fs_state;
 845    pipe->delete_fs_state = aapoint_delete_fs_state;
 846 }