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_const.h"
  73 #include "lp_bld_conv.h"
  74 #include "lp_bld_intr.h"
  75 #include "lp_bld_logic.h"
  76 #include "lp_bld_depth.h"
  77 #include "lp_bld_interp.h"
  78 #include "lp_bld_tgsi.h"
  79 #include "lp_bld_alpha.h"
  80 #include "lp_bld_blend.h"
  81 #include "lp_bld_swizzle.h"
  82 #include "lp_bld_flow.h"
  83 #include "lp_bld_debug.h"
  84 #include "lp_screen.h"
  85 #include "lp_context.h"
  86 #include "lp_state.h"
  87 #include "lp_quad.h"
  88
  89
  90 static const unsigned char quad_offset_x[4] = {0, 1, 0, 1};
  91 static const unsigned char quad_offset_y[4] = {0, 0, 1, 1};
  92
  93
  94 /*
  95  * Derive from the quad's upper left scalar coordinates the coordinates for
  96  * all other quad pixels
  97  */
  98 static void
  99 generate_pos0(LLVMBuilderRef builder,
 100               LLVMValueRef x,
 101               LLVMValueRef y,
 102               LLVMValueRef *x0,
 103               LLVMValueRef *y0)
 104 {
 105    LLVMTypeRef int_elem_type = LLVMInt32Type();
 106    LLVMTypeRef int_vec_type = LLVMVectorType(int_elem_type, QUAD_SIZE);
 107    LLVMTypeRef elem_type = LLVMFloatType();
 108    LLVMTypeRef vec_type = LLVMVectorType(elem_type, QUAD_SIZE);
 109    LLVMValueRef x_offsets[QUAD_SIZE];
 110    LLVMValueRef y_offsets[QUAD_SIZE];
 111    unsigned i;
 112
 113    x = lp_build_broadcast(builder, int_vec_type, x);
 114    y = lp_build_broadcast(builder, int_vec_type, y);
 115
 116    for(i = 0; i < QUAD_SIZE; ++i) {
 117       x_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_x[i], 0);
 118       y_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_y[i], 0);
 119    }
 120
 121    x = LLVMBuildAdd(builder, x, LLVMConstVector(x_offsets, QUAD_SIZE), "");
 122    y = LLVMBuildAdd(builder, y, LLVMConstVector(y_offsets, QUAD_SIZE), "");
 123
 124    *x0 = LLVMBuildSIToFP(builder, x, vec_type, "");
 125    *y0 = LLVMBuildSIToFP(builder, y, vec_type, "");
 126 }
 127
 128
 129 /**
 130  * Generate the depth test.
 131  */
 132 static void
 133 generate_depth(struct llvmpipe_context *lp,
 134                LLVMBuilderRef builder,
 135                const struct pipe_depth_state *state,
 136                union lp_type src_type,
 137                struct lp_build_mask_context *mask,
 138                LLVMValueRef src,
 139                LLVMValueRef dst_ptr)
 140 {
 141    const struct util_format_description *format_desc;
 142    union lp_type dst_type;
 143
 144    if(!lp->framebuffer.zsbuf)
 145       return;
 146
 147    format_desc = util_format_description(lp->framebuffer.zsbuf->format);
 148    assert(format_desc);
 149
 150    /* Pick the depth type. */
 151    dst_type = lp_depth_type(format_desc, src_type.width*src_type.length);
 152
 153    /* FIXME: Cope with a depth test type with a different bit width. */
 154    assert(dst_type.width == src_type.width);
 155    assert(dst_type.length == src_type.length);
 156
 157 #if 1
 158    src = lp_build_clamped_float_to_unsigned_norm(builder,
 159                                                  src_type,
 160                                                  dst_type.width,
 161                                                  src);
 162 #else
 163    lp_build_conv(builder, src_type, dst_type, &src, 1, &src, 1);
 164 #endif
 165
 166    lp_build_depth_test(builder,
 167                        state,
 168                        dst_type,
 169                        format_desc,
 170                        mask,
 171                        src,
 172                        dst_ptr);
 173 }
 174
 175
 176 struct build_fetch_texel_context
 177 {
 178    LLVMValueRef context_ptr;
 179
 180    LLVMValueRef samplers_ptr;
 181
 182    /** Coords/texels store */
 183    LLVMValueRef store_ptr;
 184 };
 185
 186
 187 void PIPE_CDECL
 188 lp_fetch_texel_soa( struct tgsi_sampler **samplers,
 189                     uint32_t unit,
 190                     float *store )
 191 {
 192    struct tgsi_sampler *sampler = samplers[unit];
 193
 194 #if 0
 195    uint j;
 196
 197    debug_printf("%s sampler: %p (%p) store: %p\n",
 198                 __FUNCTION__,
 199                 sampler, *sampler,
 200                 store );
 201
 202    debug_printf("lodbias %f\n", store[12]);
 203
 204    for (j = 0; j < 4; j++)
 205       debug_printf("sample %d texcoord %f %f\n",
 206                    j,
 207                    store[0+j],
 208                    store[4+j]);
 209 #endif
 210
 211    {
 212       float rgba[NUM_CHANNELS][QUAD_SIZE];
 213       sampler->get_samples(sampler,
 214                            &store[0],
 215                            &store[4],
 216                            &store[8],
 217                            0.0f, /*store[12],  lodbias */
 218                            rgba);
 219       memcpy(store, rgba, sizeof rgba);
 220    }
 221
 222 #if 0
 223    for (j = 0; j < 4; j++)
 224       debug_printf("sample %d result %f %f %f %f\n",
 225                    j,
 226                    store[0+j],
 227                    store[4+j],
 228                    store[8+j],
 229                    store[12+j]);
 230 #endif
 231 }
 232
 233
 234 static void
 235 emit_fetch_texel( LLVMBuilderRef builder,
 236                   void *context,
 237                   unsigned unit,
 238                   unsigned num_coords,
 239                   const LLVMValueRef *coords,
 240                   LLVMValueRef lodbias,
 241                   LLVMValueRef *texel)
 242 {
 243    struct build_fetch_texel_context *bld = context;
 244    LLVMTypeRef vec_type = LLVMTypeOf(coords[0]);
 245    LLVMValueRef args[3];
 246    unsigned i;
 247
 248    if(!bld->samplers_ptr)
 249       bld->samplers_ptr = lp_jit_context_samplers(builder, bld->context_ptr);
 250
 251    if(!bld->store_ptr)
 252       bld->store_ptr = LLVMBuildArrayAlloca(builder,
 253                                             vec_type,
 254                                             LLVMConstInt(LLVMInt32Type(), 4, 0),
 255                                             "texel_store");
 256
 257    for (i = 0; i < num_coords; i++) {
 258       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
 259       LLVMValueRef coord_ptr = LLVMBuildGEP(builder, bld->store_ptr, &index, 1, "");
 260       LLVMBuildStore(builder, coords[i], coord_ptr);
 261    }
 262
 263    args[0] = bld->samplers_ptr;
 264    args[1] = LLVMConstInt(LLVMInt32Type(), unit, 0);
 265    args[2] = bld->store_ptr;
 266
 267    lp_build_intrinsic(builder, "fetch_texel", LLVMVoidType(), args, 3);
 268
 269    for (i = 0; i < NUM_CHANNELS; ++i) {
 270       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
 271       LLVMValueRef texel_ptr = LLVMBuildGEP(builder, bld->store_ptr, &index, 1, "");
 272       texel[i] = LLVMBuildLoad(builder, texel_ptr, "");
 273    }
 274 }
 275
 276
 277 /**
 278  * Generate the fragment shader, depth/stencil test, and alpha tests.
 279  */
 280 static void
 281 generate_fs(struct llvmpipe_context *lp,
 282             struct lp_fragment_shader *shader,
 283             const struct lp_fragment_shader_variant_key *key,
 284             LLVMBuilderRef builder,
 285             union lp_type type,
 286             LLVMValueRef context_ptr,
 287             unsigned i,
 288             const struct lp_build_interp_soa_context *interp,
 289             struct build_fetch_texel_context *sampler,
 290             LLVMValueRef *pmask,
 291             LLVMValueRef *color,
 292             LLVMValueRef depth_ptr)
 293 {
 294    const struct tgsi_token *tokens = shader->base.tokens;
 295    LLVMTypeRef elem_type;
 296    LLVMTypeRef vec_type;
 297    LLVMTypeRef int_vec_type;
 298    LLVMValueRef consts_ptr;
 299    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
 300    LLVMValueRef z = interp->pos[2];
 301    struct lp_build_mask_context mask;
 302    boolean early_depth_test;
 303    unsigned attrib;
 304    unsigned chan;
 305
 306    elem_type = lp_build_elem_type(type);
 307    vec_type = lp_build_vec_type(type);
 308    int_vec_type = lp_build_int_vec_type(type);
 309
 310    consts_ptr = lp_jit_context_constants(builder, context_ptr);
 311
 312    lp_build_mask_begin(&mask, builder, type, *pmask);
 313
 314    early_depth_test =
 315       lp->depth_stencil->depth.enabled &&
 316       lp->framebuffer.zsbuf &&
 317       !lp->depth_stencil->alpha.enabled &&
 318       !lp->fs->info.uses_kill &&
 319       !lp->fs->info.writes_z;
 320
 321    if(early_depth_test)
 322       generate_depth(lp, builder, &key->depth,
 323                      type, &mask,
 324                      z, depth_ptr);
 325
 326    memset(outputs, 0, sizeof outputs);
 327
 328    lp_build_tgsi_soa(builder, tokens, type, &mask,
 329                      consts_ptr, interp->pos, interp->inputs,
 330                      outputs, emit_fetch_texel, sampler);
 331
 332    for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
 333       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 334          if(outputs[attrib][chan]) {
 335             lp_build_name(outputs[attrib][chan], "output%u.%u.%c", i, attrib, "xyzw"[chan]);
 336
 337             switch (shader->info.output_semantic_name[attrib]) {
 338             case TGSI_SEMANTIC_COLOR:
 339                {
 340                   unsigned cbuf = shader->info.output_semantic_index[attrib];
 341
 342                   lp_build_name(outputs[attrib][chan], "color%u.%u.%c", i, attrib, "rgba"[chan]);
 343
 344                   /* Alpha test */
 345                   /* XXX: should the alpha reference value be passed separately? */
 346                   if(cbuf == 0 && chan == 3) {
 347                      LLVMValueRef alpha = outputs[attrib][chan];
 348                      LLVMValueRef alpha_ref_value;
 349                      alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
 350                      alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
 351                      lp_build_alpha_test(builder, &key->alpha, type,
 352                                          &mask, alpha, alpha_ref_value);
 353                   }
 354
 355                   if(cbuf == 0)
 356                      color[chan] = outputs[attrib][chan];
 357
 358                   break;
 359                }
 360
 361             case TGSI_SEMANTIC_POSITION:
 362                if(chan == 2)
 363                   z = outputs[attrib][chan];
 364                break;
 365             }
 366          }
 367       }
 368    }
 369
 370    if(!early_depth_test)
 371       generate_depth(lp, builder, &key->depth,
 372                      type, &mask,
 373                      z, depth_ptr);
 374
 375    lp_build_mask_end(&mask);
 376
 377    *pmask = mask.value;
 378
 379 }
 380
 381
 382 /**
 383  * Generate color blending and color output.
 384  */
 385 static void
 386 generate_blend(const struct pipe_blend_state *blend,
 387                LLVMBuilderRef builder,
 388                union lp_type type,
 389                LLVMValueRef context_ptr,
 390                LLVMValueRef mask,
 391                LLVMValueRef *src,
 392                LLVMValueRef dst_ptr)
 393 {
 394    struct lp_build_context bld;
 395    LLVMTypeRef vec_type;
 396    LLVMTypeRef int_vec_type;
 397    LLVMValueRef const_ptr;
 398    LLVMValueRef con[4];
 399    LLVMValueRef dst[4];
 400    LLVMValueRef res[4];
 401    unsigned chan;
 402
 403    vec_type = lp_build_vec_type(type);
 404    int_vec_type = lp_build_int_vec_type(type);
 405
 406    lp_build_context_init(&bld, builder, type);
 407
 408    const_ptr = lp_jit_context_blend_color(builder, context_ptr);
 409    const_ptr = LLVMBuildBitCast(builder, const_ptr,
 410                                 LLVMPointerType(vec_type, 0), "");
 411
 412    for(chan = 0; chan < 4; ++chan) {
 413       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 414       con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
 415
 416       dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
 417
 418       lp_build_name(con[chan], "con.%c", "rgba"[chan]);
 419       lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
 420    }
 421
 422    lp_build_blend_soa(builder, blend, type, src, dst, con, res);
 423
 424    for(chan = 0; chan < 4; ++chan) {
 425       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 426       lp_build_name(res[chan], "res.%c", "rgba"[chan]);
 427       res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
 428       LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
 429    }
 430 }
 431
 432
 433 /**
 434  * Generate the runtime callable function for the whole fragment pipeline.
 435  */
 436 static struct lp_fragment_shader_variant *
 437 generate_fragment(struct llvmpipe_context *lp,
 438                   struct lp_fragment_shader *shader,
 439                   const struct lp_fragment_shader_variant_key *key)
 440 {
 441    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
 442    struct lp_fragment_shader_variant *variant;
 443    union lp_type fs_type;
 444    union lp_type blend_type;
 445    LLVMTypeRef fs_elem_type;
 446    LLVMTypeRef fs_vec_type;
 447    LLVMTypeRef fs_int_vec_type;
 448    LLVMTypeRef blend_vec_type;
 449    LLVMTypeRef blend_int_vec_type;
 450    LLVMTypeRef arg_types[9];
 451    LLVMTypeRef func_type;
 452    LLVMValueRef context_ptr;
 453    LLVMValueRef x;
 454    LLVMValueRef y;
 455    LLVMValueRef a0_ptr;
 456    LLVMValueRef dadx_ptr;
 457    LLVMValueRef dady_ptr;
 458    LLVMValueRef mask_ptr;
 459    LLVMValueRef color_ptr;
 460    LLVMValueRef depth_ptr;
 461    LLVMBasicBlockRef block;
 462    LLVMBuilderRef builder;
 463    LLVMValueRef x0;
 464    LLVMValueRef y0;
 465    struct build_fetch_texel_context sampler;
 466    struct lp_build_interp_soa_context interp;
 467    LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
 468    LLVMValueRef fs_out_color[NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
 469    LLVMValueRef blend_mask;
 470    LLVMValueRef blend_in_color[NUM_CHANNELS];
 471    unsigned num_fs;
 472    unsigned i;
 473    unsigned chan;
 474
 475 #ifdef DEBUG
 476    tgsi_dump(shader->base.tokens, 0);
 477    if(key->depth.enabled) {
 478       debug_printf("depth.func = %s\n", debug_dump_func(key->depth.func, TRUE));
 479       debug_printf("depth.writemask = %u\n", key->depth.writemask);
 480       debug_printf("depth.occlusion_count = %u\n", key->depth.occlusion_count);
 481    }
 482    if(key->alpha.enabled) {
 483       debug_printf("alpha.func = %s\n", debug_dump_func(key->alpha.func, TRUE));
 484       debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
 485    }
 486    if(key->blend.logicop_enable) {
 487       debug_printf("blend.logicop_func = %u\n", key->blend.logicop_func);
 488    }
 489    else if(key->blend.blend_enable) {
 490       debug_printf("blend.rgb_func = %s\n",   debug_dump_blend_func  (key->blend.rgb_func, TRUE));
 491       debug_printf("rgb_src_factor = %s\n",   debug_dump_blend_factor(key->blend.rgb_src_factor, TRUE));
 492       debug_printf("rgb_dst_factor = %s\n",   debug_dump_blend_factor(key->blend.rgb_dst_factor, TRUE));
 493       debug_printf("alpha_func = %s\n",       debug_dump_blend_func  (key->blend.alpha_func, TRUE));
 494       debug_printf("alpha_src_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_src_factor, TRUE));
 495       debug_printf("alpha_dst_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_dst_factor, TRUE));
 496    }
 497    debug_printf("blend.colormask = 0x%x\n", key->blend.colormask);
 498 #endif
 499
 500    variant = CALLOC_STRUCT(lp_fragment_shader_variant);
 501    if(!variant)
 502       return NULL;
 503
 504    variant->shader = shader;
 505    memcpy(&variant->key, key, sizeof *key);
 506
 507    /* TODO: actually pick these based on the fs and color buffer
 508     * characteristics. */
 509
 510    fs_type.value = 0;
 511    fs_type.floating = TRUE; /* floating point values */
 512    fs_type.sign = TRUE;     /* values are signed */
 513    fs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
 514    fs_type.width = 32;      /* 32-bit float */
 515    fs_type.length = 4;      /* 4 element per vector */
 516    num_fs = 4;
 517
 518    blend_type.value = 0;
 519    blend_type.floating = FALSE; /* values are integers */
 520    blend_type.sign = FALSE;     /* values are unsigned */
 521    blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
 522    blend_type.width = 8;        /* 8-bit ubyte values */
 523    blend_type.length = 16;      /* 16 elements per vector */
 524
 525    /*
 526     * Generate the function prototype. Any change here must be reflected in
 527     * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
 528     */
 529
 530    fs_elem_type = lp_build_elem_type(fs_type);
 531    fs_vec_type = lp_build_vec_type(fs_type);
 532    fs_int_vec_type = lp_build_int_vec_type(fs_type);
 533
 534    blend_vec_type = lp_build_vec_type(blend_type);
 535    blend_int_vec_type = lp_build_int_vec_type(blend_type);
 536
 537    arg_types[0] = screen->context_ptr_type;            /* context */
 538    arg_types[1] = LLVMInt32Type();                     /* x */
 539    arg_types[2] = LLVMInt32Type();                     /* y */
 540    arg_types[3] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
 541    arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
 542    arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* dady */
 543    arg_types[6] = LLVMPointerType(fs_int_vec_type, 0); /* mask */
 544    arg_types[7] = LLVMPointerType(blend_vec_type, 0);  /* color */
 545    arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
 546
 547    func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
 548
 549    variant->function = LLVMAddFunction(screen->module, "shader", func_type);
 550    LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
 551    for(i = 0; i < Elements(arg_types); ++i)
 552       if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
 553          LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute);
 554
 555    context_ptr  = LLVMGetParam(variant->function, 0);
 556    x            = LLVMGetParam(variant->function, 1);
 557    y            = LLVMGetParam(variant->function, 2);
 558    a0_ptr       = LLVMGetParam(variant->function, 3);
 559    dadx_ptr     = LLVMGetParam(variant->function, 4);
 560    dady_ptr     = LLVMGetParam(variant->function, 5);
 561    mask_ptr     = LLVMGetParam(variant->function, 6);
 562    color_ptr    = LLVMGetParam(variant->function, 7);
 563    depth_ptr    = LLVMGetParam(variant->function, 8);
 564
 565    lp_build_name(context_ptr, "context");
 566    lp_build_name(x, "x");
 567    lp_build_name(y, "y");
 568    lp_build_name(a0_ptr, "a0");
 569    lp_build_name(dadx_ptr, "dadx");
 570    lp_build_name(dady_ptr, "dady");
 571    lp_build_name(mask_ptr, "mask");
 572    lp_build_name(color_ptr, "color");
 573    lp_build_name(depth_ptr, "depth");
 574
 575    /*
 576     * Function body
 577     */
 578
 579    block = LLVMAppendBasicBlock(variant->function, "entry");
 580    builder = LLVMCreateBuilder();
 581    LLVMPositionBuilderAtEnd(builder, block);
 582
 583    generate_pos0(builder, x, y, &x0, &y0);
 584
 585    lp_build_interp_soa_init(&interp, shader->base.tokens, builder, fs_type,
 586                             a0_ptr, dadx_ptr, dady_ptr,
 587                             x0, y0, 2, 0);
 588
 589    memset(&sampler, 0, sizeof sampler);
 590    sampler.context_ptr = context_ptr;
 591
 592    for(i = 0; i < num_fs; ++i) {
 593       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
 594       LLVMValueRef out_color[NUM_CHANNELS];
 595       LLVMValueRef depth_ptr_i;
 596
 597       if(i != 0)
 598          lp_build_interp_soa_update(&interp);
 599
 600       fs_mask[i] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, mask_ptr, &index, 1, ""), "");
 601       depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
 602
 603       generate_fs(lp, shader, key,
 604                   builder,
 605                   fs_type,
 606                   context_ptr,
 607                   i,
 608                   &interp,
 609                   &sampler,
 610                   &fs_mask[i],
 611                   out_color,
 612                   depth_ptr_i);
 613
 614       for(chan = 0; chan < NUM_CHANNELS; ++chan)
 615          fs_out_color[chan][i] = out_color[chan];
 616    }
 617
 618    /*
 619     * Convert the fs's output color and mask to fit to the blending type.
 620     */
 621
 622    for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 623       lp_build_conv(builder, fs_type, blend_type,
 624                     fs_out_color[chan], num_fs,
 625                     &blend_in_color[chan], 1);
 626       lp_build_name(blend_in_color[chan], "color.%c", "rgba"[chan]);
 627
 628    }
 629
 630    lp_build_conv_mask(builder, fs_type, blend_type,
 631                                fs_mask, num_fs,
 632                                &blend_mask, 1);
 633
 634    /*
 635     * Blending.
 636     */
 637
 638    generate_blend(&key->blend,
 639                   builder,
 640                   blend_type,
 641                   context_ptr,
 642                   blend_mask,
 643                   blend_in_color,
 644                   color_ptr);
 645
 646    LLVMBuildRetVoid(builder);
 647
 648    LLVMDisposeBuilder(builder);
 649
 650    /*
 651     * Translate the LLVM IR into machine code.
 652     */
 653
 654    LLVMRunFunctionPassManager(screen->pass, variant->function);
 655
 656 #ifdef DEBUG
 657    LLVMDumpValue(variant->function);
 658    debug_printf("\n");
 659 #endif
 660
 661    if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) {
 662       LLVMDumpValue(variant->function);
 663       abort();
 664    }
 665
 666    variant->jit_function = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, variant->function);
 667
 668 #ifdef DEBUG
 669    lp_disassemble(variant->jit_function);
 670 #endif
 671
 672    variant->next = shader->variants;
 673    shader->variants = variant;
 674
 675    return variant;
 676 }
 677
 678
 679 void *
 680 llvmpipe_create_fs_state(struct pipe_context *pipe,
 681                          const struct pipe_shader_state *templ)
 682 {
 683    struct lp_fragment_shader *shader;
 684
 685    shader = CALLOC_STRUCT(lp_fragment_shader);
 686    if (!shader)
 687       return NULL;
 688
 689    /* get/save the summary info for this shader */
 690    tgsi_scan_shader(templ->tokens, &shader->info);
 691
 692    /* we need to keep a local copy of the tokens */
 693    shader->base.tokens = tgsi_dup_tokens(templ->tokens);
 694
 695    return shader;
 696 }
 697
 698
 699 void
 700 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
 701 {
 702    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 703
 704    llvmpipe->fs = (struct lp_fragment_shader *) fs;
 705
 706    llvmpipe->dirty |= LP_NEW_FS;
 707 }
 708
 709
 710 void
 711 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
 712 {
 713    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 714    struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
 715    struct lp_fragment_shader *shader = fs;
 716    struct lp_fragment_shader_variant *variant;
 717
 718    assert(fs != llvmpipe->fs);
 719
 720    variant = shader->variants;
 721    while(variant) {
 722       struct lp_fragment_shader_variant *next = variant->next;
 723
 724       if(variant->function) {
 725          if(variant->jit_function)
 726             LLVMFreeMachineCodeForFunction(screen->engine, variant->function);
 727          LLVMDeleteFunction(variant->function);
 728       }
 729
 730       FREE(variant);
 731
 732       variant = next;
 733    }
 734
 735    FREE((void *) shader->base.tokens);
 736    FREE(shader);
 737 }
 738
 739
 740
 741 void
 742 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
 743                              uint shader, uint index,
 744                              const struct pipe_constant_buffer *buf)
 745 {
 746    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 747
 748    assert(shader < PIPE_SHADER_TYPES);
 749    assert(index == 0);
 750
 751    /* note: reference counting */
 752    pipe_buffer_reference(&llvmpipe->constants[shader].buffer,
 753                          buf ? buf->buffer : NULL);
 754
 755    llvmpipe->dirty |= LP_NEW_CONSTANTS;
 756 }
 757
 758
 759 /**
 760  * We need to generate several variants of the fragment pipeline to match
 761  * all the combinations of the contributing state atoms.
 762  *
 763  * TODO: there is actually no reason to tie this to context state -- the
 764  * generated code could be cached globally in the screen.
 765  */
 766 static void
 767 make_variant_key(struct llvmpipe_context *lp,
 768                  struct lp_fragment_shader_variant_key *key)
 769 {
 770    memset(key, 0, sizeof *key);
 771
 772    memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
 773
 774    key->alpha.enabled = lp->depth_stencil->alpha.enabled;
 775    if(key->alpha.enabled)
 776       key->alpha.func = lp->depth_stencil->alpha.func;
 777    /* alpha.ref_value is passed in jit_context */
 778
 779    memcpy(&key->blend, lp->blend, sizeof key->blend);
 780 }
 781
 782
 783 void
 784 llvmpipe_update_fs(struct llvmpipe_context *lp)
 785 {
 786    struct lp_fragment_shader *shader = lp->fs;
 787    struct lp_fragment_shader_variant_key key;
 788    struct lp_fragment_shader_variant *variant;
 789
 790    make_variant_key(lp, &key);
 791
 792    variant = shader->variants;
 793    while(variant) {
 794       if(memcmp(&variant->key, &key, sizeof key) == 0)
 795          break;
 796
 797       variant = variant->next;
 798    }
 799
 800    if(!variant)
 801       variant = generate_fragment(lp, shader, &key);
 802
 803    shader->current = variant;
 804 }