src/gallium/drivers/llvmpipe/lp_state_fs.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2009 VMware, Inc.
   4  * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
   5  * All Rights Reserved.
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * copy of this software and associated documentation files (the
   9  * "Software"), to deal in the Software without restriction, including
  10  * without limitation the rights to use, copy, modify, merge, publish,
  11  * distribute, sub license, and/or sell copies of the Software, and to
  12  * permit persons to whom the Software is furnished to do so, subject to
  13  * the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the
  16  * next paragraph) shall be included in all copies or substantial portions
  17  * of the Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  20  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  21  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  22  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
  23  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  24  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  25  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  26  *
  27  **************************************************************************/
  28
  29 /**
  30  * @file
  31  * Code generate the whole fragment pipeline.
  32  *
  33  * The fragment pipeline consists of the following stages:
  34  * - stipple (TBI)
  35  * - early depth test
  36  * - fragment shader
  37  * - alpha test
  38  * - depth/stencil test (stencil TBI)
  39  * - blending
  40  *
  41  * This file has only the glue to assembly the fragment pipeline.  The actual
  42  * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
  43  * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
  44  * muster the LLVM JIT execution engine to create a function that follows an
  45  * established binary interface and that can be called from C directly.
  46  *
  47  * A big source of complexity here is that we often want to run different
  48  * stages with different precisions and data types and precisions. For example,
  49  * the fragment shader needs typically to be done in floats, but the
  50  * depth/stencil test and blending is better done in the type that most closely
  51  * matches the depth/stencil and color buffer respectively.
  52  *
  53  * Since the width of a SIMD vector register stays the same regardless of the
  54  * element type, different types imply different number of elements, so we must
  55  * code generate more instances of the stages with larger types to be able to
  56  * feed/consume the stages with smaller types.
  57  *
  58  * @author Jose Fonseca <jfonseca@vmware.com>
  59  */
  60
  61 #include "pipe/p_defines.h"
  62 #include "util/u_memory.h"
  63 #include "util/u_format.h"
  64 #include "util/u_debug_dump.h"
  65 #include "pipe/internal/p_winsys_screen.h"
  66 #include "pipe/p_shader_tokens.h"
  67 #include "draw/draw_context.h"
  68 #include "tgsi/tgsi_dump.h"
  69 #include "tgsi/tgsi_scan.h"
  70 #include "tgsi/tgsi_parse.h"
  71 #include "lp_bld_type.h"
  72 #include "lp_bld_conv.h"
  73 #include "lp_bld_logic.h"
  74 #include "lp_bld_depth.h"
  75 #include "lp_bld_tgsi.h"
  76 #include "lp_bld_alpha.h"
  77 #include "lp_bld_blend.h"
  78 #include "lp_bld_swizzle.h"
  79 #include "lp_bld_flow.h"
  80 #include "lp_bld_debug.h"
  81 #include "lp_screen.h"
  82 #include "lp_context.h"
  83 #include "lp_state.h"
  84 #include "lp_quad.h"
  85
  86
  87 static const unsigned char quad_offset_x[4] = {0, 1, 0, 1};
  88 static const unsigned char quad_offset_y[4] = {0, 0, 1, 1};
  89
  90
  91 /**
  92  * Generate the position vectors.
  93  *
  94  * TODO: This should be called only once per fragment pipeline, for the first
  95  * quad, and the neighboring quad positions obtained by additions.
  96  *
  97  * Parameter x, y are the integer values with the quad upper left coordinates.
  98  */
  99 static void
 100 generate_pos(LLVMBuilderRef builder,
 101              LLVMValueRef x,
 102              LLVMValueRef y,
 103              LLVMValueRef a0_ptr,
 104              LLVMValueRef dadx_ptr,
 105              LLVMValueRef dady_ptr,
 106              LLVMValueRef *pos)
 107 {
 108    LLVMTypeRef int_elem_type = LLVMInt32Type();
 109    LLVMTypeRef int_vec_type = LLVMVectorType(int_elem_type, QUAD_SIZE);
 110    LLVMTypeRef elem_type = LLVMFloatType();
 111    LLVMTypeRef vec_type = LLVMVectorType(elem_type, QUAD_SIZE);
 112    LLVMValueRef x_offsets[QUAD_SIZE];
 113    LLVMValueRef y_offsets[QUAD_SIZE];
 114    unsigned chan;
 115    unsigned i;
 116
 117    /*
 118     * Derive from the quad's upper left scalar coordinates the coordinates for
 119     * all other quad pixels
 120     */
 121
 122    x = lp_build_broadcast(builder, int_vec_type, x);
 123    y = lp_build_broadcast(builder, int_vec_type, y);
 124
 125    for(i = 0; i < QUAD_SIZE; ++i) {
 126       x_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_x[i], 0);
 127       y_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_y[i], 0);
 128    }
 129
 130    x = LLVMBuildAdd(builder, x, LLVMConstVector(x_offsets, QUAD_SIZE), "");
 131    y = LLVMBuildAdd(builder, y, LLVMConstVector(y_offsets, QUAD_SIZE), "");
 132
 133    x = LLVMBuildSIToFP(builder, x, vec_type, "");
 134    y = LLVMBuildSIToFP(builder, y, vec_type, "");
 135
 136    pos[0] = x;
 137    pos[1] = y;
 138
 139    /*
 140     * Calculate z and w from the interpolation factors.
 141     */
 142
 143    for(chan = 2; chan < NUM_CHANNELS; ++chan) {
 144       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 145       LLVMValueRef a0   = LLVMBuildLoad(builder, LLVMBuildGEP(builder, a0_ptr,   &index, 1, ""), "");
 146       LLVMValueRef dadx = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dadx_ptr, &index, 1, ""), "");
 147       LLVMValueRef dady = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dady_ptr, &index, 1, ""), "");
 148       LLVMValueRef res;
 149       a0   = lp_build_broadcast(builder, vec_type, a0);
 150       dadx = lp_build_broadcast(builder, vec_type, dadx);
 151       dady = lp_build_broadcast(builder, vec_type, dady);
 152       res = a0;
 153       res = LLVMBuildAdd(builder, res, LLVMBuildMul(builder, dadx, x, ""), "");
 154       res = LLVMBuildAdd(builder, res, LLVMBuildMul(builder, dady, y, ""), "");
 155       pos[chan] = res;
 156    }
 157
 158    for(chan = 0; chan < NUM_CHANNELS; ++chan)
 159       lp_build_name(pos[chan], "pos.%c", "xyzw"[chan]);
 160 }
 161
 162
 163 /**
 164  * Generate the depth test.
 165  */
 166 static void
 167 generate_depth(struct llvmpipe_context *lp,
 168                LLVMBuilderRef builder,
 169                const struct pipe_depth_state *state,
 170                union lp_type src_type,
 171                struct lp_build_mask_context *mask,
 172                LLVMValueRef src,
 173                LLVMValueRef dst_ptr)
 174 {
 175    const struct util_format_description *format_desc;
 176    union lp_type dst_type;
 177
 178    if(!lp->framebuffer.zsbuf)
 179       return;
 180
 181    format_desc = util_format_description(lp->framebuffer.zsbuf->format);
 182    assert(format_desc);
 183
 184    /* Pick the depth type. */
 185    dst_type = lp_depth_type(format_desc, src_type.width*src_type.length);
 186
 187    /* FIXME: Cope with a depth test type with a different bit width. */
 188    assert(dst_type.width == src_type.width);
 189    assert(dst_type.length == src_type.length);
 190
 191 #if 1
 192    src = lp_build_clamped_float_to_unsigned_norm(builder,
 193                                                  src_type,
 194                                                  dst_type.width,
 195                                                  src);
 196 #else
 197    lp_build_conv(builder, src_type, dst_type, &src, 1, &src, 1);
 198 #endif
 199
 200    lp_build_depth_test(builder,
 201                        state,
 202                        dst_type,
 203                        format_desc,
 204                        mask,
 205                        src,
 206                        dst_ptr);
 207 }
 208
 209
 210 /**
 211  * Generate the fragment shader, depth/stencil test, and alpha tests.
 212  */
 213 static void
 214 generate_fs(struct llvmpipe_context *lp,
 215             struct lp_fragment_shader *shader,
 216             const struct lp_fragment_shader_variant_key *key,
 217             LLVMBuilderRef builder,
 218             union lp_type type,
 219             unsigned i,
 220             LLVMValueRef x,
 221             LLVMValueRef y,
 222             LLVMValueRef a0_ptr,
 223             LLVMValueRef dadx_ptr,
 224             LLVMValueRef dady_ptr,
 225             LLVMValueRef consts_ptr,
 226             LLVMValueRef *pmask,
 227             LLVMValueRef *color,
 228             LLVMValueRef depth_ptr,
 229             LLVMValueRef samplers_ptr)
 230 {
 231    const struct tgsi_token *tokens = shader->base.tokens;
 232    LLVMTypeRef elem_type;
 233    LLVMTypeRef vec_type;
 234    LLVMTypeRef int_vec_type;
 235    LLVMValueRef pos[NUM_CHANNELS];
 236    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
 237    struct lp_build_mask_context mask;
 238    boolean early_depth_test;
 239    unsigned attrib;
 240    unsigned chan;
 241
 242    elem_type = lp_build_elem_type(type);
 243    vec_type = lp_build_vec_type(type);
 244    int_vec_type = lp_build_int_vec_type(type);
 245
 246    generate_pos(builder, x, y, a0_ptr, dadx_ptr, dady_ptr, pos);
 247
 248    lp_build_mask_begin(&mask, builder, type, *pmask);
 249
 250    early_depth_test =
 251       lp->depth_stencil->depth.enabled &&
 252       lp->framebuffer.zsbuf &&
 253       !lp->depth_stencil->alpha.enabled &&
 254       !lp->fs->info.uses_kill &&
 255       !lp->fs->info.writes_z;
 256
 257    if(early_depth_test)
 258       generate_depth(lp, builder, &key->depth,
 259                           type, &mask,
 260                           pos[2], depth_ptr);
 261
 262    memset(outputs, 0, sizeof outputs);
 263
 264    lp_build_tgsi_soa(builder, tokens, type, &mask,
 265                      pos, a0_ptr, dadx_ptr, dady_ptr,
 266                      consts_ptr, outputs, samplers_ptr);
 267
 268    for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
 269       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 270          if(outputs[attrib][chan]) {
 271             lp_build_name(outputs[attrib][chan], "output%u.%u.%c", i, attrib, "xyzw"[chan]);
 272
 273             switch (shader->info.output_semantic_name[attrib]) {
 274             case TGSI_SEMANTIC_COLOR:
 275                {
 276                   unsigned cbuf = shader->info.output_semantic_index[attrib];
 277
 278                   lp_build_name(outputs[attrib][chan], "color%u.%u.%c", i, attrib, "rgba"[chan]);
 279
 280                   /* Alpha test */
 281                   /* XXX: should the alpha reference value be passed separately? */
 282                   if(cbuf == 0 && chan == 3)
 283                      lp_build_alpha_test(builder, &key->alpha, type,
 284                                          &mask,
 285                                          outputs[attrib][chan]);
 286
 287                   if(cbuf == 0)
 288                      color[chan] = outputs[attrib][chan];
 289
 290                   break;
 291                }
 292
 293             case TGSI_SEMANTIC_POSITION:
 294                if(chan == 2)
 295                   pos[2] = outputs[attrib][chan];
 296                break;
 297             }
 298          }
 299       }
 300    }
 301
 302    if(!early_depth_test)
 303       generate_depth(lp, builder, &key->depth,
 304                           type, &mask,
 305                           pos[2], depth_ptr);
 306
 307    lp_build_mask_end(&mask);
 308
 309    *pmask = mask.value;
 310
 311 }
 312
 313
 314 /**
 315  * Generate color blending and color output.
 316  */
 317 static void
 318 generate_blend(const struct pipe_blend_state *blend,
 319                LLVMBuilderRef builder,
 320                union lp_type type,
 321                LLVMValueRef mask,
 322                LLVMValueRef *src,
 323                LLVMValueRef const_ptr,
 324                LLVMValueRef dst_ptr)
 325 {
 326    struct lp_build_context bld;
 327    LLVMTypeRef vec_type;
 328    LLVMTypeRef int_vec_type;
 329    LLVMValueRef con[4];
 330    LLVMValueRef dst[4];
 331    LLVMValueRef res[4];
 332    unsigned chan;
 333
 334    vec_type = lp_build_vec_type(type);
 335    int_vec_type = lp_build_int_vec_type(type);
 336
 337    lp_build_context_init(&bld, builder, type);
 338
 339    for(chan = 0; chan < 4; ++chan) {
 340       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 341
 342       if(const_ptr)
 343          con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
 344       else
 345          con[chan] = LLVMGetUndef(vec_type); /* FIXME */
 346
 347       dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
 348
 349       lp_build_name(con[chan], "con.%c", "rgba"[chan]);
 350       lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
 351    }
 352
 353    lp_build_blend_soa(builder, blend, type, src, dst, con, res);
 354
 355    for(chan = 0; chan < 4; ++chan) {
 356       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 357       lp_build_name(res[chan], "res.%c", "rgba"[chan]);
 358       res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
 359       LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
 360    }
 361 }
 362
 363
 364 /**
 365  * Generate the runtime callable function for the whole fragment pipeline.
 366  */
 367 static struct lp_fragment_shader_variant *
 368 generate_fragment(struct llvmpipe_context *lp,
 369                   struct lp_fragment_shader *shader,
 370                   const struct lp_fragment_shader_variant_key *key)
 371 {
 372    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
 373    struct lp_fragment_shader_variant *variant;
 374    union lp_type fs_type;
 375    union lp_type blend_type;
 376    LLVMTypeRef fs_elem_type;
 377    LLVMTypeRef fs_vec_type;
 378    LLVMTypeRef fs_int_vec_type;
 379    LLVMTypeRef blend_vec_type;
 380    LLVMTypeRef blend_int_vec_type;
 381    LLVMTypeRef arg_types[9];
 382    LLVMTypeRef func_type;
 383    LLVMValueRef context_ptr;
 384    LLVMValueRef x;
 385    LLVMValueRef y;
 386    LLVMValueRef a0_ptr;
 387    LLVMValueRef dadx_ptr;
 388    LLVMValueRef dady_ptr;
 389    LLVMValueRef consts_ptr;
 390    LLVMValueRef mask_ptr;
 391    LLVMValueRef color_ptr;
 392    LLVMValueRef depth_ptr;
 393    LLVMValueRef samplers_ptr;
 394    LLVMBasicBlockRef block;
 395    LLVMBuilderRef builder;
 396    LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
 397    LLVMValueRef fs_out_color[NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
 398    LLVMValueRef blend_mask;
 399    LLVMValueRef blend_in_color[NUM_CHANNELS];
 400    LLVMValueRef fetch_texel;
 401    unsigned num_fs;
 402    unsigned i;
 403    unsigned chan;
 404
 405 #ifdef DEBUG
 406    tgsi_dump(shader->base.tokens, 0);
 407    if(key->depth.enabled) {
 408       debug_printf("depth.func = %s\n", debug_dump_func(key->depth.func, TRUE));
 409       debug_printf("depth.writemask = %u\n", key->depth.writemask);
 410       debug_printf("depth.occlusion_count = %u\n", key->depth.occlusion_count);
 411    }
 412    if(key->alpha.enabled) {
 413       debug_printf("alpha.func = %s\n", debug_dump_func(key->alpha.func, TRUE));
 414       debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
 415    }
 416    if(key->blend.logicop_enable) {
 417       debug_printf("blend.logicop_func = %u\n", key->blend.logicop_func);
 418    }
 419    else if(key->blend.blend_enable) {
 420       debug_printf("blend.rgb_func = %s\n",   debug_dump_blend_func  (key->blend.rgb_func, TRUE));
 421       debug_printf("rgb_src_factor = %s\n",   debug_dump_blend_factor(key->blend.rgb_src_factor, TRUE));
 422       debug_printf("rgb_dst_factor = %s\n",   debug_dump_blend_factor(key->blend.rgb_dst_factor, TRUE));
 423       debug_printf("alpha_func = %s\n",       debug_dump_blend_func  (key->blend.alpha_func, TRUE));
 424       debug_printf("alpha_src_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_src_factor, TRUE));
 425       debug_printf("alpha_dst_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_dst_factor, TRUE));
 426    }
 427    debug_printf("blend.colormask = 0x%x\n", key->blend.colormask);
 428 #endif
 429
 430    variant = CALLOC_STRUCT(lp_fragment_shader_variant);
 431    if(!variant)
 432       return NULL;
 433
 434    variant->shader = shader;
 435    memcpy(&variant->key, key, sizeof *key);
 436
 437    /* TODO: actually pick these based on the fs and color buffer
 438     * characteristics. */
 439
 440    fs_type.value = 0;
 441    fs_type.floating = TRUE; /* floating point values */
 442    fs_type.sign = TRUE;     /* values are signed */
 443    fs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
 444    fs_type.width = 32;      /* 32-bit float */
 445    fs_type.length = 4;      /* 4 element per vector */
 446    num_fs = 4;
 447
 448    blend_type.value = 0;
 449    blend_type.floating = FALSE; /* values are integers */
 450    blend_type.sign = FALSE;     /* values are unsigned */
 451    blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
 452    blend_type.width = 8;        /* 8-bit ubyte values */
 453    blend_type.length = 16;      /* 16 elements per vector */
 454
 455    /*
 456     * Generate the function prototype. Any change here must be reflected in
 457     * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
 458     */
 459
 460    fs_elem_type = lp_build_elem_type(fs_type);
 461    fs_vec_type = lp_build_vec_type(fs_type);
 462    fs_int_vec_type = lp_build_int_vec_type(fs_type);
 463
 464    blend_vec_type = lp_build_vec_type(blend_type);
 465    blend_int_vec_type = lp_build_int_vec_type(blend_type);
 466
 467    arg_types[0] = screen->context_ptr_type;            /* context */
 468    arg_types[1] = LLVMInt32Type();                     /* x */
 469    arg_types[2] = LLVMInt32Type();                     /* y */
 470    arg_types[3] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
 471    arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
 472    arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* dady */
 473    arg_types[6] = LLVMPointerType(fs_int_vec_type, 0); /* mask */
 474    arg_types[7] = LLVMPointerType(blend_vec_type, 0);  /* color */
 475    arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
 476
 477    func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
 478
 479    variant->function = LLVMAddFunction(screen->module, "shader", func_type);
 480    LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
 481    for(i = 0; i < Elements(arg_types); ++i)
 482       if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
 483          LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute);
 484
 485    context_ptr  = LLVMGetParam(variant->function, 0);
 486    x            = LLVMGetParam(variant->function, 1);
 487    y            = LLVMGetParam(variant->function, 2);
 488    a0_ptr       = LLVMGetParam(variant->function, 3);
 489    dadx_ptr     = LLVMGetParam(variant->function, 4);
 490    dady_ptr     = LLVMGetParam(variant->function, 5);
 491    mask_ptr     = LLVMGetParam(variant->function, 6);
 492    color_ptr    = LLVMGetParam(variant->function, 7);
 493    depth_ptr    = LLVMGetParam(variant->function, 8);
 494
 495    lp_build_name(context_ptr, "context");
 496    lp_build_name(x, "x");
 497    lp_build_name(y, "y");
 498    lp_build_name(a0_ptr, "a0");
 499    lp_build_name(dadx_ptr, "dadx");
 500    lp_build_name(dady_ptr, "dady");
 501    lp_build_name(mask_ptr, "mask");
 502    lp_build_name(color_ptr, "color");
 503    lp_build_name(depth_ptr, "depth");
 504
 505    /*
 506     * Function body
 507     */
 508
 509    block = LLVMAppendBasicBlock(variant->function, "entry");
 510    builder = LLVMCreateBuilder();
 511    LLVMPositionBuilderAtEnd(builder, block);
 512
 513    consts_ptr = lp_jit_context_constants(builder, context_ptr);
 514    samplers_ptr = lp_jit_context_samplers(builder, context_ptr);
 515
 516    for(i = 0; i < num_fs; ++i) {
 517       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
 518       LLVMValueRef out_color[NUM_CHANNELS];
 519       LLVMValueRef x_i;
 520       LLVMValueRef depth_ptr_i;
 521
 522       /* TODO: Reuse position interpolation */
 523       x_i = LLVMBuildAdd(builder, x, LLVMConstInt(LLVMInt32Type(), 2*i, 0), "");
 524
 525       fs_mask[i] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, mask_ptr, &index, 1, ""), "");
 526       depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
 527
 528       generate_fs(lp,
 529                   shader,
 530                   key,
 531                   builder,
 532                   fs_type,
 533                   i,
 534                   x_i,
 535                   y,
 536                   a0_ptr,
 537                   dadx_ptr,
 538                   dady_ptr,
 539                   consts_ptr,
 540                   &fs_mask[i],
 541                   out_color,
 542                   depth_ptr_i,
 543                   samplers_ptr);
 544
 545       for(chan = 0; chan < NUM_CHANNELS; ++chan)
 546          fs_out_color[chan][i] = out_color[chan];
 547    }
 548
 549    /*
 550     * Convert the fs's output color and mask to fit to the blending type.
 551     */
 552
 553    for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 554       lp_build_conv(builder, fs_type, blend_type,
 555                     fs_out_color[chan], num_fs,
 556                     &blend_in_color[chan], 1);
 557       lp_build_name(blend_in_color[chan], "color.%c", "rgba"[chan]);
 558    }
 559
 560    lp_build_conv_mask(builder, fs_type, blend_type,
 561                                fs_mask, num_fs,
 562                                &blend_mask, 1);
 563
 564    /*
 565     * Blending.
 566     */
 567
 568    generate_blend(&key->blend,
 569                   builder,
 570                   blend_type,
 571                   blend_mask,
 572                   blend_in_color,
 573                   NULL /* FIXME: blend_const_color */,
 574                   color_ptr);
 575
 576    LLVMBuildRetVoid(builder);
 577
 578    LLVMDisposeBuilder(builder);
 579
 580    /*
 581     * Translate the LLVM IR into machine code.
 582     */
 583
 584    LLVMRunFunctionPassManager(screen->pass, variant->function);
 585
 586 #ifdef DEBUG
 587    LLVMDumpValue(variant->function);
 588    debug_printf("\n");
 589 #endif
 590
 591    if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) {
 592       LLVMDumpValue(variant->function);
 593       abort();
 594    }
 595
 596    /* Tell where the fetch_texel function is, if the shader refers to it.
 597     * TODO: this should be done elsewhere.
 598     */
 599    fetch_texel = LLVMGetNamedFunction(screen->module, "fetch_texel");
 600    if(fetch_texel) {
 601       static boolean first_time = TRUE;
 602       if(first_time) {
 603          LLVMAddGlobalMapping(screen->engine, fetch_texel, lp_build_tgsi_fetch_texel_soa);
 604          first_time = FALSE;
 605       }
 606    }
 607
 608    variant->jit_function = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, variant->function);
 609
 610 #ifdef DEBUG
 611    lp_disassemble(variant->jit_function);
 612 #endif
 613
 614    variant->next = shader->variants;
 615    shader->variants = variant;
 616
 617    return variant;
 618 }
 619
 620
 621 void *
 622 llvmpipe_create_fs_state(struct pipe_context *pipe,
 623                          const struct pipe_shader_state *templ)
 624 {
 625    struct lp_fragment_shader *shader;
 626
 627    shader = CALLOC_STRUCT(lp_fragment_shader);
 628    if (!shader)
 629       return NULL;
 630
 631    /* get/save the summary info for this shader */
 632    tgsi_scan_shader(templ->tokens, &shader->info);
 633
 634    /* we need to keep a local copy of the tokens */
 635    shader->base.tokens = tgsi_dup_tokens(templ->tokens);
 636
 637    return shader;
 638 }
 639
 640
 641 void
 642 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
 643 {
 644    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 645
 646    llvmpipe->fs = (struct lp_fragment_shader *) fs;
 647
 648    llvmpipe->dirty |= LP_NEW_FS;
 649 }
 650
 651
 652 void
 653 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
 654 {
 655    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 656    struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
 657    struct lp_fragment_shader *shader = fs;
 658    struct lp_fragment_shader_variant *variant;
 659
 660    assert(fs != llvmpipe->fs);
 661
 662    variant = shader->variants;
 663    while(variant) {
 664       struct lp_fragment_shader_variant *next = variant->next;
 665
 666       if(variant->function) {
 667          if(variant->jit_function)
 668             LLVMFreeMachineCodeForFunction(screen->engine, variant->function);
 669          LLVMDeleteFunction(variant->function);
 670       }
 671
 672       FREE(variant);
 673
 674       variant = next;
 675    }
 676
 677    FREE((void *) shader->base.tokens);
 678    FREE(shader);
 679 }
 680
 681
 682
 683 void
 684 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
 685                              uint shader, uint index,
 686                              const struct pipe_constant_buffer *buf)
 687 {
 688    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 689
 690    assert(shader < PIPE_SHADER_TYPES);
 691    assert(index == 0);
 692
 693    /* note: reference counting */
 694    pipe_buffer_reference(&llvmpipe->constants[shader].buffer,
 695                          buf ? buf->buffer : NULL);
 696
 697    llvmpipe->dirty |= LP_NEW_CONSTANTS;
 698 }
 699
 700
 701 void
 702 llvmpipe_update_fs(struct llvmpipe_context *lp)
 703 {
 704    struct lp_fragment_shader *shader = lp->fs;
 705    struct lp_fragment_shader_variant_key key;
 706    struct lp_fragment_shader_variant *variant;
 707
 708    /* We need to generate several variants of the fragment pipeline to match
 709     * all the combinations of the contributing state atoms.
 710     *
 711     * TODO: there is actually no reason to tie this to context state -- the
 712     * generated code could be cached globally in the screen.
 713     */
 714
 715    memset(&key, 0, sizeof key);
 716    memcpy(&key.depth, &lp->depth_stencil->depth, sizeof &key.depth);
 717    memcpy(&key.alpha, &lp->depth_stencil->alpha, sizeof &key.alpha);
 718    memcpy(&key.blend, lp->blend, sizeof &key.blend);
 719
 720    variant = shader->variants;
 721    while(variant) {
 722       if(memcmp(&variant->key, &key, sizeof key) == 0)
 723          break;
 724
 725       variant = variant->next;
 726    }
 727
 728    if(!variant)
 729       variant = generate_fragment(lp, shader, &key);
 730
 731    shader->current = variant;
 732 }