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  * - triangle edge in/out testing
  35  * - scissor test
  36  * - stipple (TBI)
  37  * - early depth test
  38  * - fragment shader
  39  * - alpha test
  40  * - depth/stencil test
  41  * - blending
  42  *
  43  * This file has only the glue to assemble the fragment pipeline.  The actual
  44  * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
  45  * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
  46  * muster the LLVM JIT execution engine to create a function that follows an
  47  * established binary interface and that can be called from C directly.
  48  *
  49  * A big source of complexity here is that we often want to run different
  50  * stages with different precisions and data types and precisions. For example,
  51  * the fragment shader needs typically to be done in floats, but the
  52  * depth/stencil test and blending is better done in the type that most closely
  53  * matches the depth/stencil and color buffer respectively.
  54  *
  55  * Since the width of a SIMD vector register stays the same regardless of the
  56  * element type, different types imply different number of elements, so we must
  57  * code generate more instances of the stages with larger types to be able to
  58  * feed/consume the stages with smaller types.
  59  *
  60  * @author Jose Fonseca <jfonseca@vmware.com>
  61  */
  62
  63 #include <limits.h>
  64 #include "pipe/p_defines.h"
  65 #include "util/u_inlines.h"
  66 #include "util/u_memory.h"
  67 #include "util/u_pointer.h"
  68 #include "util/u_format.h"
  69 #include "util/u_dump.h"
  70 #include "os/os_time.h"
  71 #include "pipe/p_shader_tokens.h"
  72 #include "draw/draw_context.h"
  73 #include "tgsi/tgsi_dump.h"
  74 #include "tgsi/tgsi_scan.h"
  75 #include "tgsi/tgsi_parse.h"
  76 #include "gallivm/lp_bld_type.h"
  77 #include "gallivm/lp_bld_const.h"
  78 #include "gallivm/lp_bld_conv.h"
  79 #include "gallivm/lp_bld_intr.h"
  80 #include "gallivm/lp_bld_logic.h"
  81 #include "gallivm/lp_bld_tgsi.h"
  82 #include "gallivm/lp_bld_swizzle.h"
  83 #include "gallivm/lp_bld_flow.h"
  84 #include "gallivm/lp_bld_debug.h"
  85
  86 #include "lp_bld_alpha.h"
  87 #include "lp_bld_blend.h"
  88 #include "lp_bld_depth.h"
  89 #include "lp_bld_interp.h"
  90 #include "lp_context.h"
  91 #include "lp_debug.h"
  92 #include "lp_perf.h"
  93 #include "lp_screen.h"
  94 #include "lp_setup.h"
  95 #include "lp_state.h"
  96 #include "lp_tex_sample.h"
  97
  98
  99 #include <llvm-c/Analysis.h>
 100
 101
 102 static const unsigned char quad_offset_x[4] = {0, 1, 0, 1};
 103 static const unsigned char quad_offset_y[4] = {0, 0, 1, 1};
 104
 105
 106 /*
 107  * Derive from the quad's upper left scalar coordinates the coordinates for
 108  * all other quad pixels
 109  */
 110 static void
 111 generate_pos0(LLVMBuilderRef builder,
 112               LLVMValueRef x,
 113               LLVMValueRef y,
 114               LLVMValueRef *x0,
 115               LLVMValueRef *y0)
 116 {
 117    LLVMTypeRef int_elem_type = LLVMInt32Type();
 118    LLVMTypeRef int_vec_type = LLVMVectorType(int_elem_type, QUAD_SIZE);
 119    LLVMTypeRef elem_type = LLVMFloatType();
 120    LLVMTypeRef vec_type = LLVMVectorType(elem_type, QUAD_SIZE);
 121    LLVMValueRef x_offsets[QUAD_SIZE];
 122    LLVMValueRef y_offsets[QUAD_SIZE];
 123    unsigned i;
 124
 125    x = lp_build_broadcast(builder, int_vec_type, x);
 126    y = lp_build_broadcast(builder, int_vec_type, y);
 127
 128    for(i = 0; i < QUAD_SIZE; ++i) {
 129       x_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_x[i], 0);
 130       y_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_y[i], 0);
 131    }
 132
 133    x = LLVMBuildAdd(builder, x, LLVMConstVector(x_offsets, QUAD_SIZE), "");
 134    y = LLVMBuildAdd(builder, y, LLVMConstVector(y_offsets, QUAD_SIZE), "");
 135
 136    *x0 = LLVMBuildSIToFP(builder, x, vec_type, "");
 137    *y0 = LLVMBuildSIToFP(builder, y, vec_type, "");
 138 }
 139
 140
 141 /**
 142  * Generate the depth /stencil test code.
 143  */
 144 static void
 145 generate_depth_stencil(LLVMBuilderRef builder,
 146                        const struct lp_fragment_shader_variant_key *key,
 147                        struct lp_type src_type,
 148                        struct lp_build_mask_context *mask,
 149                        LLVMValueRef stencil_refs[2],
 150                        LLVMValueRef src,
 151                        LLVMValueRef dst_ptr,
 152                        LLVMValueRef facing,
 153                        LLVMValueRef counter)
 154 {
 155    const struct util_format_description *format_desc;
 156    struct lp_type dst_type;
 157
 158    if (!key->depth.enabled && !key->stencil[0].enabled && !key->stencil[1].enabled)
 159       return;
 160
 161    format_desc = util_format_description(key->zsbuf_format);
 162    assert(format_desc);
 163
 164    /*
 165     * Depths are expected to be between 0 and 1, even if they are stored in
 166     * floats. Setting these bits here will ensure that the lp_build_conv() call
 167     * below won't try to unnecessarily clamp the incoming values.
 168     */
 169    if(src_type.floating) {
 170       src_type.sign = FALSE;
 171       src_type.norm = TRUE;
 172    }
 173    else {
 174       assert(!src_type.sign);
 175       assert(src_type.norm);
 176    }
 177
 178    /* Pick the depth type. */
 179    dst_type = lp_depth_type(format_desc, src_type.width*src_type.length);
 180
 181    /* FIXME: Cope with a depth test type with a different bit width. */
 182    assert(dst_type.width == src_type.width);
 183    assert(dst_type.length == src_type.length);
 184
 185    /* Convert fragment Z from float to integer */
 186    lp_build_conv(builder, src_type, dst_type, &src, 1, &src, 1);
 187
 188    dst_ptr = LLVMBuildBitCast(builder,
 189                               dst_ptr,
 190                               LLVMPointerType(lp_build_vec_type(dst_type), 0), "");
 191    lp_build_depth_stencil_test(builder,
 192                                &key->depth,
 193                                key->stencil,
 194                                dst_type,
 195                                format_desc,
 196                                mask,
 197                                stencil_refs,
 198                                src,
 199                                dst_ptr,
 200                                facing,
 201                                counter);
 202 }
 203
 204
 205 /**
 206  * Generate the code to do inside/outside triangle testing for the
 207  * four pixels in a 2x2 quad.  This will set the four elements of the
 208  * quad mask vector to 0 or ~0.
 209  * \param i  which quad of the quad group to test, in [0,3]
 210  */
 211 static void
 212 generate_tri_edge_mask(LLVMBuilderRef builder,
 213                        unsigned i,
 214                        LLVMValueRef *mask,      /* ivec4, out */
 215                        LLVMValueRef c0,         /* int32 */
 216                        LLVMValueRef c1,         /* int32 */
 217                        LLVMValueRef c2,         /* int32 */
 218                        LLVMValueRef step0_ptr,  /* ivec4 */
 219                        LLVMValueRef step1_ptr,  /* ivec4 */
 220                        LLVMValueRef step2_ptr)  /* ivec4 */
 221 {
 222 #define OPTIMIZE_IN_OUT_TEST 0
 223 #if OPTIMIZE_IN_OUT_TEST
 224    struct lp_build_if_state ifctx;
 225    LLVMValueRef not_draw_all;
 226 #endif
 227    struct lp_build_flow_context *flow;
 228    struct lp_type i32_type;
 229    LLVMTypeRef i32vec4_type;
 230    LLVMValueRef c0_vec, c1_vec, c2_vec;
 231    LLVMValueRef in_out_mask;
 232
 233    assert(i < 4);
 234
 235    /* int32 vector type */
 236    memset(&i32_type, 0, sizeof i32_type);
 237    i32_type.floating = FALSE; /* values are integers */
 238    i32_type.sign = TRUE;      /* values are signed */
 239    i32_type.norm = FALSE;     /* values are not normalized */
 240    i32_type.width = 32;       /* 32-bit int values */
 241    i32_type.length = 4;       /* 4 elements per vector */
 242
 243    i32vec4_type = lp_build_int32_vec4_type();
 244
 245    /*
 246     * Use a conditional here to do detailed pixel in/out testing.
 247     * We only have to do this if c0 != INT_MIN.
 248     */
 249    flow = lp_build_flow_create(builder);
 250    lp_build_flow_scope_begin(flow);
 251
 252    {
 253 #if OPTIMIZE_IN_OUT_TEST
 254       /* not_draw_all = (c0 != INT_MIN) */
 255       not_draw_all = LLVMBuildICmp(builder,
 256                                    LLVMIntNE,
 257                                    c0,
 258                                    LLVMConstInt(LLVMInt32Type(), INT_MIN, 0),
 259                                    "");
 260
 261       in_out_mask = lp_build_const_int_vec(i32_type, ~0);
 262
 263
 264       lp_build_flow_scope_declare(flow, &in_out_mask);
 265
 266       /* if (not_draw_all) {... */
 267       lp_build_if(&ifctx, flow, builder, not_draw_all);
 268 #endif
 269       {
 270          LLVMValueRef step0_vec, step1_vec, step2_vec;
 271          LLVMValueRef m0_vec, m1_vec, m2_vec;
 272          LLVMValueRef index, m;
 273
 274          /* c0_vec = {c0, c0, c0, c0}
 275           * Note that we emit this code four times but LLVM optimizes away
 276           * three instances of it.
 277           */
 278          c0_vec = lp_build_broadcast(builder, i32vec4_type, c0);
 279          c1_vec = lp_build_broadcast(builder, i32vec4_type, c1);
 280          c2_vec = lp_build_broadcast(builder, i32vec4_type, c2);
 281          lp_build_name(c0_vec, "edgeconst0vec");
 282          lp_build_name(c1_vec, "edgeconst1vec");
 283          lp_build_name(c2_vec, "edgeconst2vec");
 284
 285          /* load step0vec, step1, step2 vec from memory */
 286          index = LLVMConstInt(LLVMInt32Type(), i, 0);
 287          step0_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step0_ptr, &index, 1, ""), "");
 288          step1_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step1_ptr, &index, 1, ""), "");
 289          step2_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step2_ptr, &index, 1, ""), "");
 290          lp_build_name(step0_vec, "step0vec");
 291          lp_build_name(step1_vec, "step1vec");
 292          lp_build_name(step2_vec, "step2vec");
 293
 294          /* m0_vec = step0_ptr[i] > c0_vec */
 295          m0_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step0_vec, c0_vec);
 296          m1_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step1_vec, c1_vec);
 297          m2_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step2_vec, c2_vec);
 298
 299          /* in_out_mask = m0_vec & m1_vec & m2_vec */
 300          m = LLVMBuildAnd(builder, m0_vec, m1_vec, "");
 301          in_out_mask = LLVMBuildAnd(builder, m, m2_vec, "");
 302          lp_build_name(in_out_mask, "inoutmaskvec");
 303       }
 304 #if OPTIMIZE_IN_OUT_TEST
 305       lp_build_endif(&ifctx);
 306 #endif
 307
 308    }
 309    lp_build_flow_scope_end(flow);
 310    lp_build_flow_destroy(flow);
 311
 312    /* This is the initial alive/dead pixel mask for a quad of four pixels.
 313     * It's an int[4] vector with each word set to 0 or ~0.
 314     * Words will get cleared when pixels faile the Z test, etc.
 315     */
 316    *mask = in_out_mask;
 317 }
 318
 319
 320 static LLVMValueRef
 321 generate_scissor_test(LLVMBuilderRef builder,
 322                       LLVMValueRef context_ptr,
 323                       const struct lp_build_interp_soa_context *interp,
 324                       struct lp_type type)
 325 {
 326    LLVMTypeRef vec_type = lp_build_vec_type(type);
 327    LLVMValueRef xpos = interp->pos[0], ypos = interp->pos[1];
 328    LLVMValueRef xmin, ymin, xmax, ymax;
 329    LLVMValueRef m0, m1, m2, m3, m;
 330
 331    /* xpos, ypos contain the window coords for the four pixels in the quad */
 332    assert(xpos);
 333    assert(ypos);
 334
 335    /* get the current scissor bounds, convert to vectors */
 336    xmin = lp_jit_context_scissor_xmin_value(builder, context_ptr);
 337    xmin = lp_build_broadcast(builder, vec_type, xmin);
 338
 339    ymin = lp_jit_context_scissor_ymin_value(builder, context_ptr);
 340    ymin = lp_build_broadcast(builder, vec_type, ymin);
 341
 342    xmax = lp_jit_context_scissor_xmax_value(builder, context_ptr);
 343    xmax = lp_build_broadcast(builder, vec_type, xmax);
 344
 345    ymax = lp_jit_context_scissor_ymax_value(builder, context_ptr);
 346    ymax = lp_build_broadcast(builder, vec_type, ymax);
 347
 348    /* compare the fragment's position coordinates against the scissor bounds */
 349    m0 = lp_build_compare(builder, type, PIPE_FUNC_GEQUAL, xpos, xmin);
 350    m1 = lp_build_compare(builder, type, PIPE_FUNC_GEQUAL, ypos, ymin);
 351    m2 = lp_build_compare(builder, type, PIPE_FUNC_LESS, xpos, xmax);
 352    m3 = lp_build_compare(builder, type, PIPE_FUNC_LESS, ypos, ymax);
 353
 354    /* AND all the masks together */
 355    m = LLVMBuildAnd(builder, m0, m1, "");
 356    m = LLVMBuildAnd(builder, m, m2, "");
 357    m = LLVMBuildAnd(builder, m, m3, "");
 358
 359    lp_build_name(m, "scissormask");
 360
 361    return m;
 362 }
 363
 364
 365 static LLVMValueRef
 366 build_int32_vec_const(int value)
 367 {
 368    struct lp_type i32_type;
 369
 370    memset(&i32_type, 0, sizeof i32_type);
 371    i32_type.floating = FALSE; /* values are integers */
 372    i32_type.sign = TRUE;      /* values are signed */
 373    i32_type.norm = FALSE;     /* values are not normalized */
 374    i32_type.width = 32;       /* 32-bit int values */
 375    i32_type.length = 4;       /* 4 elements per vector */
 376    return lp_build_const_int_vec(i32_type, value);
 377 }
 378
 379
 380
 381 /**
 382  * Generate the fragment shader, depth/stencil test, and alpha tests.
 383  * \param i  which quad in the tile, in range [0,3]
 384  * \param do_tri_test  if 1, do triangle edge in/out testing
 385  */
 386 static void
 387 generate_fs(struct llvmpipe_context *lp,
 388             struct lp_fragment_shader *shader,
 389             const struct lp_fragment_shader_variant_key *key,
 390             LLVMBuilderRef builder,
 391             struct lp_type type,
 392             LLVMValueRef context_ptr,
 393             unsigned i,
 394             const struct lp_build_interp_soa_context *interp,
 395             struct lp_build_sampler_soa *sampler,
 396             LLVMValueRef *pmask,
 397             LLVMValueRef (*color)[4],
 398             LLVMValueRef depth_ptr,
 399             LLVMValueRef facing,
 400             unsigned do_tri_test,
 401             LLVMValueRef c0,
 402             LLVMValueRef c1,
 403             LLVMValueRef c2,
 404             LLVMValueRef step0_ptr,
 405             LLVMValueRef step1_ptr,
 406             LLVMValueRef step2_ptr,
 407             LLVMValueRef counter)
 408 {
 409    const struct tgsi_token *tokens = shader->base.tokens;
 410    LLVMTypeRef vec_type;
 411    LLVMValueRef consts_ptr;
 412    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
 413    LLVMValueRef z = interp->pos[2];
 414    LLVMValueRef stencil_refs[2];
 415    struct lp_build_flow_context *flow;
 416    struct lp_build_mask_context mask;
 417    boolean early_depth_stencil_test;
 418    unsigned attrib;
 419    unsigned chan;
 420    unsigned cbuf;
 421
 422    assert(i < 4);
 423
 424    stencil_refs[0] = lp_jit_context_stencil_ref_front_value(builder, context_ptr);
 425    stencil_refs[1] = lp_jit_context_stencil_ref_back_value(builder, context_ptr);
 426
 427    vec_type = lp_build_vec_type(type);
 428
 429    consts_ptr = lp_jit_context_constants(builder, context_ptr);
 430
 431    flow = lp_build_flow_create(builder);
 432
 433    memset(outputs, 0, sizeof outputs);
 434
 435    lp_build_flow_scope_begin(flow);
 436
 437    /* Declare the color and z variables */
 438    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
 439       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 440          color[cbuf][chan] = LLVMGetUndef(vec_type);
 441          lp_build_flow_scope_declare(flow, &color[cbuf][chan]);
 442       }
 443    }
 444    lp_build_flow_scope_declare(flow, &z);
 445
 446    /* do triangle edge testing */
 447    if (do_tri_test) {
 448       generate_tri_edge_mask(builder, i, pmask,
 449                              c0, c1, c2, step0_ptr, step1_ptr, step2_ptr);
 450    }
 451    else {
 452       *pmask = build_int32_vec_const(~0);
 453    }
 454
 455    /* 'mask' will control execution based on quad's pixel alive/killed state */
 456    lp_build_mask_begin(&mask, flow, type, *pmask);
 457
 458    if (key->scissor) {
 459       LLVMValueRef smask =
 460          generate_scissor_test(builder, context_ptr, interp, type);
 461       lp_build_mask_update(&mask, smask);
 462    }
 463
 464    early_depth_stencil_test =
 465       (key->depth.enabled || key->stencil[0].enabled) &&
 466       !key->alpha.enabled &&
 467       !shader->info.uses_kill &&
 468       !shader->info.writes_z;
 469
 470    if (early_depth_stencil_test)
 471       generate_depth_stencil(builder, key,
 472                              type, &mask,
 473                              stencil_refs, z, depth_ptr, facing, counter);
 474
 475    lp_build_tgsi_soa(builder, tokens, type, &mask,
 476                      consts_ptr, interp->pos, interp->inputs,
 477                      outputs, sampler, &shader->info);
 478
 479    /* loop over fragment shader outputs/results */
 480    for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
 481       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 482          if(outputs[attrib][chan]) {
 483             LLVMValueRef out = LLVMBuildLoad(builder, outputs[attrib][chan], "");
 484             lp_build_name(out, "output%u.%u.%c", i, attrib, "xyzw"[chan]);
 485
 486             switch (shader->info.output_semantic_name[attrib]) {
 487             case TGSI_SEMANTIC_COLOR:
 488                {
 489                   unsigned cbuf = shader->info.output_semantic_index[attrib];
 490
 491                   lp_build_name(out, "color%u.%u.%c", i, attrib, "rgba"[chan]);
 492
 493                   /* Alpha test */
 494                   /* XXX: should the alpha reference value be passed separately? */
 495                   /* XXX: should only test the final assignment to alpha */
 496                   if(cbuf == 0 && chan == 3) {
 497                      LLVMValueRef alpha = out;
 498                      LLVMValueRef alpha_ref_value;
 499                      alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
 500                      alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
 501                      lp_build_alpha_test(builder, &key->alpha, type,
 502                                          &mask, alpha, alpha_ref_value);
 503                   }
 504
 505                   color[cbuf][chan] = out;
 506                   break;
 507                }
 508
 509             case TGSI_SEMANTIC_POSITION:
 510                if(chan == 2)
 511                   z = out;
 512                break;
 513             }
 514          }
 515       }
 516    }
 517
 518    if (!early_depth_stencil_test)
 519       generate_depth_stencil(builder, key,
 520                              type, &mask,
 521                              stencil_refs, z, depth_ptr, facing, counter);
 522
 523    lp_build_mask_end(&mask);
 524
 525    lp_build_flow_scope_end(flow);
 526
 527    lp_build_flow_destroy(flow);
 528
 529    *pmask = mask.value;
 530
 531 }
 532
 533
 534 /**
 535  * Generate color blending and color output.
 536  * \param rt  the render target index (to index blend, colormask state)
 537  * \param type  the pixel color type
 538  * \param context_ptr  pointer to the runtime JIT context
 539  * \param mask  execution mask (active fragment/pixel mask)
 540  * \param src  colors from the fragment shader
 541  * \param dst_ptr  the destination color buffer pointer
 542  */
 543 static void
 544 generate_blend(const struct pipe_blend_state *blend,
 545                unsigned rt,
 546                LLVMBuilderRef builder,
 547                struct lp_type type,
 548                LLVMValueRef context_ptr,
 549                LLVMValueRef mask,
 550                LLVMValueRef *src,
 551                LLVMValueRef dst_ptr)
 552 {
 553    struct lp_build_context bld;
 554    struct lp_build_flow_context *flow;
 555    struct lp_build_mask_context mask_ctx;
 556    LLVMTypeRef vec_type;
 557    LLVMValueRef const_ptr;
 558    LLVMValueRef con[4];
 559    LLVMValueRef dst[4];
 560    LLVMValueRef res[4];
 561    unsigned chan;
 562
 563    lp_build_context_init(&bld, builder, type);
 564
 565    flow = lp_build_flow_create(builder);
 566
 567    /* we'll use this mask context to skip blending if all pixels are dead */
 568    lp_build_mask_begin(&mask_ctx, flow, type, mask);
 569
 570    vec_type = lp_build_vec_type(type);
 571
 572    const_ptr = lp_jit_context_blend_color(builder, context_ptr);
 573    const_ptr = LLVMBuildBitCast(builder, const_ptr,
 574                                 LLVMPointerType(vec_type, 0), "");
 575
 576    /* load constant blend color and colors from the dest color buffer */
 577    for(chan = 0; chan < 4; ++chan) {
 578       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 579       con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
 580
 581       dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
 582
 583       lp_build_name(con[chan], "con.%c", "rgba"[chan]);
 584       lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
 585    }
 586
 587    /* do blend */
 588    lp_build_blend_soa(builder, blend, type, rt, src, dst, con, res);
 589
 590    /* store results to color buffer */
 591    for(chan = 0; chan < 4; ++chan) {
 592       if(blend->rt[rt].colormask & (1 << chan)) {
 593          LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 594          lp_build_name(res[chan], "res.%c", "rgba"[chan]);
 595          res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
 596          LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
 597       }
 598    }
 599
 600    lp_build_mask_end(&mask_ctx);
 601    lp_build_flow_destroy(flow);
 602 }
 603
 604
 605 /**
 606  * Generate the runtime callable function for the whole fragment pipeline.
 607  * Note that the function which we generate operates on a block of 16
 608  * pixels at at time.  The block contains 2x2 quads.  Each quad contains
 609  * 2x2 pixels.
 610  */
 611 static void
 612 generate_fragment(struct llvmpipe_context *lp,
 613                   struct lp_fragment_shader *shader,
 614                   struct lp_fragment_shader_variant *variant,
 615                   unsigned do_tri_test)
 616 {
 617    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
 618    const struct lp_fragment_shader_variant_key *key = &variant->key;
 619    struct lp_type fs_type;
 620    struct lp_type blend_type;
 621    LLVMTypeRef fs_elem_type;
 622    LLVMTypeRef fs_int_vec_type;
 623    LLVMTypeRef blend_vec_type;
 624    LLVMTypeRef arg_types[16];
 625    LLVMTypeRef func_type;
 626    LLVMTypeRef int32_vec4_type = lp_build_int32_vec4_type();
 627    LLVMValueRef context_ptr;
 628    LLVMValueRef x;
 629    LLVMValueRef y;
 630    LLVMValueRef a0_ptr;
 631    LLVMValueRef dadx_ptr;
 632    LLVMValueRef dady_ptr;
 633    LLVMValueRef color_ptr_ptr;
 634    LLVMValueRef depth_ptr;
 635    LLVMValueRef c0, c1, c2, step0_ptr, step1_ptr, step2_ptr, counter = NULL;
 636    LLVMBasicBlockRef block;
 637    LLVMBuilderRef builder;
 638    LLVMValueRef x0;
 639    LLVMValueRef y0;
 640    struct lp_build_sampler_soa *sampler;
 641    struct lp_build_interp_soa_context interp;
 642    LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
 643    LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
 644    LLVMValueRef blend_mask;
 645    LLVMValueRef function;
 646    LLVMValueRef facing;
 647    unsigned num_fs;
 648    unsigned i;
 649    unsigned chan;
 650    unsigned cbuf;
 651
 652
 653    /* TODO: actually pick these based on the fs and color buffer
 654     * characteristics. */
 655
 656    memset(&fs_type, 0, sizeof fs_type);
 657    fs_type.floating = TRUE; /* floating point values */
 658    fs_type.sign = TRUE;     /* values are signed */
 659    fs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
 660    fs_type.width = 32;      /* 32-bit float */
 661    fs_type.length = 4;      /* 4 elements per vector */
 662    num_fs = 4;              /* number of quads per block */
 663
 664    memset(&blend_type, 0, sizeof blend_type);
 665    blend_type.floating = FALSE; /* values are integers */
 666    blend_type.sign = FALSE;     /* values are unsigned */
 667    blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
 668    blend_type.width = 8;        /* 8-bit ubyte values */
 669    blend_type.length = 16;      /* 16 elements per vector */
 670
 671    /*
 672     * Generate the function prototype. Any change here must be reflected in
 673     * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
 674     */
 675
 676    fs_elem_type = lp_build_elem_type(fs_type);
 677    fs_int_vec_type = lp_build_int_vec_type(fs_type);
 678
 679    blend_vec_type = lp_build_vec_type(blend_type);
 680
 681    arg_types[0] = screen->context_ptr_type;            /* context */
 682    arg_types[1] = LLVMInt32Type();                     /* x */
 683    arg_types[2] = LLVMInt32Type();                     /* y */
 684    arg_types[3] = LLVMFloatType();                     /* facing */
 685    arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
 686    arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
 687    arg_types[6] = LLVMPointerType(fs_elem_type, 0);    /* dady */
 688    arg_types[7] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0);  /* color */
 689    arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
 690    arg_types[9] = LLVMInt32Type();                     /* c0 */
 691    arg_types[10] = LLVMInt32Type();                    /* c1 */
 692    arg_types[11] = LLVMInt32Type();                    /* c2 */
 693    /* Note: the step arrays are built as int32[16] but we interpret
 694     * them here as int32_vec4[4].
 695     */
 696    arg_types[12] = LLVMPointerType(int32_vec4_type, 0);/* step0 */
 697    arg_types[13] = LLVMPointerType(int32_vec4_type, 0);/* step1 */
 698    arg_types[14] = LLVMPointerType(int32_vec4_type, 0);/* step2 */
 699    arg_types[15] = LLVMPointerType(LLVMInt32Type(), 0);/* counter */
 700
 701    func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
 702
 703    function = LLVMAddFunction(screen->module, "shader", func_type);
 704    LLVMSetFunctionCallConv(function, LLVMCCallConv);
 705
 706    variant->function[do_tri_test] = function;
 707
 708
 709    /* XXX: need to propagate noalias down into color param now we are
 710     * passing a pointer-to-pointer?
 711     */
 712    for(i = 0; i < Elements(arg_types); ++i)
 713       if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
 714          LLVMAddAttribute(LLVMGetParam(function, i), LLVMNoAliasAttribute);
 715
 716    context_ptr  = LLVMGetParam(function, 0);
 717    x            = LLVMGetParam(function, 1);
 718    y            = LLVMGetParam(function, 2);
 719    facing       = LLVMGetParam(function, 3);
 720    a0_ptr       = LLVMGetParam(function, 4);
 721    dadx_ptr     = LLVMGetParam(function, 5);
 722    dady_ptr     = LLVMGetParam(function, 6);
 723    color_ptr_ptr = LLVMGetParam(function, 7);
 724    depth_ptr    = LLVMGetParam(function, 8);
 725    c0           = LLVMGetParam(function, 9);
 726    c1           = LLVMGetParam(function, 10);
 727    c2           = LLVMGetParam(function, 11);
 728    step0_ptr    = LLVMGetParam(function, 12);
 729    step1_ptr    = LLVMGetParam(function, 13);
 730    step2_ptr    = LLVMGetParam(function, 14);
 731
 732    lp_build_name(context_ptr, "context");
 733    lp_build_name(x, "x");
 734    lp_build_name(y, "y");
 735    lp_build_name(a0_ptr, "a0");
 736    lp_build_name(dadx_ptr, "dadx");
 737    lp_build_name(dady_ptr, "dady");
 738    lp_build_name(color_ptr_ptr, "color_ptr_ptr");
 739    lp_build_name(depth_ptr, "depth");
 740    lp_build_name(c0, "c0");
 741    lp_build_name(c1, "c1");
 742    lp_build_name(c2, "c2");
 743    lp_build_name(step0_ptr, "step0");
 744    lp_build_name(step1_ptr, "step1");
 745    lp_build_name(step2_ptr, "step2");
 746
 747    if (key->occlusion_count) {
 748       counter = LLVMGetParam(function, 15);
 749       lp_build_name(counter, "counter");
 750    }
 751
 752    /*
 753     * Function body
 754     */
 755
 756    block = LLVMAppendBasicBlock(function, "entry");
 757    builder = LLVMCreateBuilder();
 758    LLVMPositionBuilderAtEnd(builder, block);
 759
 760    generate_pos0(builder, x, y, &x0, &y0);
 761
 762    lp_build_interp_soa_init(&interp,
 763                             &shader->info,
 764                             key->flatshade,
 765                             builder, fs_type,
 766                             a0_ptr, dadx_ptr, dady_ptr,
 767                             x0, y0);
 768
 769    /* code generated texture sampling */
 770    sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
 771
 772    /* loop over quads in the block */
 773    for(i = 0; i < num_fs; ++i) {
 774       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
 775       LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
 776       LLVMValueRef depth_ptr_i;
 777
 778       if(i != 0)
 779          lp_build_interp_soa_update(&interp, i);
 780
 781       depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
 782
 783       generate_fs(lp, shader, key,
 784                   builder,
 785                   fs_type,
 786                   context_ptr,
 787                   i,
 788                   &interp,
 789                   sampler,
 790                   &fs_mask[i], /* output */
 791                   out_color,
 792                   depth_ptr_i,
 793                   facing,
 794                   do_tri_test,
 795                   c0, c1, c2,
 796                   step0_ptr, step1_ptr, step2_ptr, counter);
 797
 798       for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
 799          for(chan = 0; chan < NUM_CHANNELS; ++chan)
 800             fs_out_color[cbuf][chan][i] = out_color[cbuf][chan];
 801    }
 802
 803    sampler->destroy(sampler);
 804
 805    /* Loop over color outputs / color buffers to do blending.
 806     */
 807    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
 808       LLVMValueRef color_ptr;
 809       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0);
 810       LLVMValueRef blend_in_color[NUM_CHANNELS];
 811       unsigned rt;
 812
 813       /*
 814        * Convert the fs's output color and mask to fit to the blending type.
 815        */
 816       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 817          lp_build_conv(builder, fs_type, blend_type,
 818                        fs_out_color[cbuf][chan], num_fs,
 819                        &blend_in_color[chan], 1);
 820          lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
 821       }
 822
 823       lp_build_conv_mask(builder, fs_type, blend_type,
 824                          fs_mask, num_fs,
 825                          &blend_mask, 1);
 826
 827       color_ptr = LLVMBuildLoad(builder,
 828                                 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
 829                                 "");
 830       lp_build_name(color_ptr, "color_ptr%d", cbuf);
 831
 832       /* which blend/colormask state to use */
 833       rt = key->blend.independent_blend_enable ? cbuf : 0;
 834
 835       /*
 836        * Blending.
 837        */
 838       generate_blend(&key->blend,
 839                      rt,
 840                      builder,
 841                      blend_type,
 842                      context_ptr,
 843                      blend_mask,
 844                      blend_in_color,
 845                      color_ptr);
 846    }
 847
 848    LLVMBuildRetVoid(builder);
 849
 850    LLVMDisposeBuilder(builder);
 851
 852
 853    /* Verify the LLVM IR.  If invalid, dump and abort */
 854 #ifdef DEBUG
 855    if(LLVMVerifyFunction(function, LLVMPrintMessageAction)) {
 856       if (1)
 857          lp_debug_dump_value(function);
 858       abort();
 859    }
 860 #endif
 861
 862    /* Apply optimizations to LLVM IR */
 863    if (1)
 864       LLVMRunFunctionPassManager(screen->pass, function);
 865
 866    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 867       /* Print the LLVM IR to stderr */
 868       lp_debug_dump_value(function);
 869       debug_printf("\n");
 870    }
 871
 872    /*
 873     * Translate the LLVM IR into machine code.
 874     */
 875    {
 876       void *f = LLVMGetPointerToGlobal(screen->engine, function);
 877
 878       variant->jit_function[do_tri_test] = (lp_jit_frag_func)pointer_to_func(f);
 879
 880       if (gallivm_debug & GALLIVM_DEBUG_ASM) {
 881          lp_disassemble(f);
 882       }
 883    }
 884 }
 885
 886
 887 static void
 888 dump_fs_variant_key(const struct lp_fragment_shader_variant_key *key)
 889 {
 890    unsigned i;
 891
 892    debug_printf("fs variant %p:\n", (void *) key);
 893
 894    if (key->depth.enabled) {
 895       debug_printf("depth.format = %s\n", util_format_name(key->zsbuf_format));
 896       debug_printf("depth.func = %s\n", util_dump_func(key->depth.func, TRUE));
 897       debug_printf("depth.writemask = %u\n", key->depth.writemask);
 898    }
 899
 900    for (i = 0; i < 2; ++i) {
 901       if (key->stencil[i].enabled) {
 902          debug_printf("stencil[%u].func = %s\n", i, util_dump_func(key->stencil[i].func, TRUE));
 903          debug_printf("stencil[%u].fail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].fail_op, TRUE));
 904          debug_printf("stencil[%u].zpass_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zpass_op, TRUE));
 905          debug_printf("stencil[%u].zfail_op = %s\n", i, util_dump_stencil_op(key->stencil[i].zfail_op, TRUE));
 906          debug_printf("stencil[%u].valuemask = 0x%x\n", i, key->stencil[i].valuemask);
 907          debug_printf("stencil[%u].writemask = 0x%x\n", i, key->stencil[i].writemask);
 908       }
 909    }
 910
 911    if (key->alpha.enabled) {
 912       debug_printf("alpha.func = %s\n", util_dump_func(key->alpha.func, TRUE));
 913       debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
 914    }
 915
 916    if (key->blend.logicop_enable) {
 917       debug_printf("blend.logicop_func = %s\n", util_dump_logicop(key->blend.logicop_func, TRUE));
 918    }
 919    else if (key->blend.rt[0].blend_enable) {
 920       debug_printf("blend.rgb_func = %s\n",   util_dump_blend_func  (key->blend.rt[0].rgb_func, TRUE));
 921       debug_printf("blend.rgb_src_factor = %s\n",   util_dump_blend_factor(key->blend.rt[0].rgb_src_factor, TRUE));
 922       debug_printf("blend.rgb_dst_factor = %s\n",   util_dump_blend_factor(key->blend.rt[0].rgb_dst_factor, TRUE));
 923       debug_printf("blend.alpha_func = %s\n",       util_dump_blend_func  (key->blend.rt[0].alpha_func, TRUE));
 924       debug_printf("blend.alpha_src_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_src_factor, TRUE));
 925       debug_printf("blend.alpha_dst_factor = %s\n", util_dump_blend_factor(key->blend.rt[0].alpha_dst_factor, TRUE));
 926    }
 927    debug_printf("blend.colormask = 0x%x\n", key->blend.rt[0].colormask);
 928    for (i = 0; i < PIPE_MAX_SAMPLERS; ++i) {
 929       if (key->sampler[i].format) {
 930          debug_printf("sampler[%u] = \n", i);
 931          debug_printf("  .format = %s\n",
 932                       util_format_name(key->sampler[i].format));
 933          debug_printf("  .target = %s\n",
 934                       util_dump_tex_target(key->sampler[i].target, TRUE));
 935          debug_printf("  .pot = %u %u %u\n",
 936                       key->sampler[i].pot_width,
 937                       key->sampler[i].pot_height,
 938                       key->sampler[i].pot_depth);
 939          debug_printf("  .wrap = %s %s %s\n",
 940                       util_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
 941                       util_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
 942                       util_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
 943          debug_printf("  .min_img_filter = %s\n",
 944                       util_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
 945          debug_printf("  .min_mip_filter = %s\n",
 946                       util_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
 947          debug_printf("  .mag_img_filter = %s\n",
 948                       util_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
 949          if (key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE)
 950             debug_printf("  .compare_func = %s\n", util_dump_func(key->sampler[i].compare_func, TRUE));
 951          debug_printf("  .normalized_coords = %u\n", key->sampler[i].normalized_coords);
 952       }
 953    }
 954 }
 955
 956
 957
 958 static struct lp_fragment_shader_variant *
 959 generate_variant(struct llvmpipe_context *lp,
 960                  struct lp_fragment_shader *shader,
 961                  const struct lp_fragment_shader_variant_key *key)
 962 {
 963    struct lp_fragment_shader_variant *variant;
 964
 965    if (gallivm_debug & GALLIVM_DEBUG_IR) {
 966       tgsi_dump(shader->base.tokens, 0);
 967       dump_fs_variant_key(key);
 968    }
 969
 970    variant = CALLOC_STRUCT(lp_fragment_shader_variant);
 971    if(!variant)
 972       return NULL;
 973
 974    memcpy(&variant->key, key, sizeof *key);
 975
 976    generate_fragment(lp, shader, variant, RAST_WHOLE);
 977    generate_fragment(lp, shader, variant, RAST_EDGE_TEST);
 978
 979    /* TODO: most of these can be relaxed, in particular the colormask */
 980    variant->opaque =
 981          !key->blend.logicop_enable &&
 982          !key->blend.rt[0].blend_enable &&
 983          key->blend.rt[0].colormask == 0xf &&
 984          !key->stencil[0].enabled &&
 985          !key->alpha.enabled &&
 986          !key->depth.enabled &&
 987          !key->scissor &&
 988          !shader->info.uses_kill
 989          ? TRUE : FALSE;
 990
 991    /* insert new variant into linked list */
 992    variant->next = shader->variants;
 993    shader->variants = variant;
 994
 995    return variant;
 996 }
 997
 998
 999 static void *
1000 llvmpipe_create_fs_state(struct pipe_context *pipe,
1001                          const struct pipe_shader_state *templ)
1002 {
1003    struct lp_fragment_shader *shader;
1004
1005    shader = CALLOC_STRUCT(lp_fragment_shader);
1006    if (!shader)
1007       return NULL;
1008
1009    /* get/save the summary info for this shader */
1010    tgsi_scan_shader(templ->tokens, &shader->info);
1011
1012    /* we need to keep a local copy of the tokens */
1013    shader->base.tokens = tgsi_dup_tokens(templ->tokens);
1014
1015    if (LP_DEBUG & DEBUG_TGSI) {
1016       unsigned attrib;
1017       debug_printf("llvmpipe: Create fragment shader %p:\n", (void *) shader);
1018       tgsi_dump(templ->tokens, 0);
1019       debug_printf("usage masks:\n");
1020       for (attrib = 0; attrib < shader->info.num_inputs; ++attrib) {
1021          unsigned usage_mask = shader->info.input_usage_mask[attrib];
1022          debug_printf("  IN[%u].%s%s%s%s\n",
1023                       attrib,
1024                       usage_mask & TGSI_WRITEMASK_X ? "x" : "",
1025                       usage_mask & TGSI_WRITEMASK_Y ? "y" : "",
1026                       usage_mask & TGSI_WRITEMASK_Z ? "z" : "",
1027                       usage_mask & TGSI_WRITEMASK_W ? "w" : "");
1028       }
1029       debug_printf("\n");
1030    }
1031
1032    return shader;
1033 }
1034
1035
1036 static void
1037 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
1038 {
1039    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1040
1041    if (llvmpipe->fs == fs)
1042       return;
1043
1044    draw_flush(llvmpipe->draw);
1045
1046    llvmpipe->fs = fs;
1047
1048    llvmpipe->dirty |= LP_NEW_FS;
1049 }
1050
1051
1052 static void
1053 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
1054 {
1055    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1056    struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
1057    struct lp_fragment_shader *shader = fs;
1058    struct lp_fragment_shader_variant *variant;
1059
1060    assert(fs != llvmpipe->fs);
1061    (void) llvmpipe;
1062
1063    /*
1064     * XXX: we need to flush the context until we have some sort of reference
1065     * counting in fragment shaders as they may still be binned
1066     */
1067    draw_flush(llvmpipe->draw);
1068    lp_setup_flush(llvmpipe->setup, 0);
1069
1070    variant = shader->variants;
1071    while(variant) {
1072       struct lp_fragment_shader_variant *next = variant->next;
1073       unsigned i;
1074
1075       for (i = 0; i < Elements(variant->function); i++) {
1076          if (variant->function[i]) {
1077             if (variant->jit_function[i])
1078                LLVMFreeMachineCodeForFunction(screen->engine,
1079                                               variant->function[i]);
1080             LLVMDeleteFunction(variant->function[i]);
1081          }
1082       }
1083
1084       FREE(variant);
1085
1086       variant = next;
1087    }
1088
1089    FREE((void *) shader->base.tokens);
1090    FREE(shader);
1091 }
1092
1093
1094
1095 static void
1096 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
1097                              uint shader, uint index,
1098                              struct pipe_resource *constants)
1099 {
1100    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
1101    unsigned size = constants ? constants->width0 : 0;
1102    const void *data = constants ? llvmpipe_resource_data(constants) : NULL;
1103
1104    assert(shader < PIPE_SHADER_TYPES);
1105    assert(index == 0);
1106
1107    if(llvmpipe->constants[shader] == constants)
1108       return;
1109
1110    draw_flush(llvmpipe->draw);
1111
1112    /* note: reference counting */
1113    pipe_resource_reference(&llvmpipe->constants[shader], constants);
1114
1115    if(shader == PIPE_SHADER_VERTEX) {
1116       draw_set_mapped_constant_buffer(llvmpipe->draw, PIPE_SHADER_VERTEX, 0,
1117                                       data, size);
1118    }
1119
1120    llvmpipe->dirty |= LP_NEW_CONSTANTS;
1121 }
1122
1123
1124 /**
1125  * Return the blend factor equivalent to a destination alpha of one.
1126  */
1127 static INLINE unsigned
1128 force_dst_alpha_one(unsigned factor, boolean alpha)
1129 {
1130    switch(factor) {
1131    case PIPE_BLENDFACTOR_DST_ALPHA:
1132       return PIPE_BLENDFACTOR_ONE;
1133    case PIPE_BLENDFACTOR_INV_DST_ALPHA:
1134       return PIPE_BLENDFACTOR_ZERO;
1135    case PIPE_BLENDFACTOR_SRC_ALPHA_SATURATE:
1136       return PIPE_BLENDFACTOR_ZERO;
1137    }
1138
1139    if (alpha) {
1140       switch(factor) {
1141       case PIPE_BLENDFACTOR_DST_COLOR:
1142          return PIPE_BLENDFACTOR_ONE;
1143       case PIPE_BLENDFACTOR_INV_DST_COLOR:
1144          return PIPE_BLENDFACTOR_ZERO;
1145       }
1146    }
1147
1148    return factor;
1149 }
1150
1151
1152 /**
1153  * We need to generate several variants of the fragment pipeline to match
1154  * all the combinations of the contributing state atoms.
1155  *
1156  * TODO: there is actually no reason to tie this to context state -- the
1157  * generated code could be cached globally in the screen.
1158  */
1159 static void
1160 make_variant_key(struct llvmpipe_context *lp,
1161                  struct lp_fragment_shader *shader,
1162                  struct lp_fragment_shader_variant_key *key)
1163 {
1164    unsigned i;
1165
1166    memset(key, 0, sizeof *key);
1167
1168    if (lp->framebuffer.zsbuf) {
1169       if (lp->depth_stencil->depth.enabled) {
1170          key->zsbuf_format = lp->framebuffer.zsbuf->format;
1171          memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
1172       }
1173       if (lp->depth_stencil->stencil[0].enabled) {
1174          key->zsbuf_format = lp->framebuffer.zsbuf->format;
1175          memcpy(&key->stencil, &lp->depth_stencil->stencil, sizeof key->stencil);
1176       }
1177    }
1178
1179    key->alpha.enabled = lp->depth_stencil->alpha.enabled;
1180    if(key->alpha.enabled)
1181       key->alpha.func = lp->depth_stencil->alpha.func;
1182    /* alpha.ref_value is passed in jit_context */
1183
1184    key->flatshade = lp->rasterizer->flatshade;
1185    key->scissor = lp->rasterizer->scissor;
1186    if (lp->active_query_count) {
1187       key->occlusion_count = TRUE;
1188    }
1189
1190    if (lp->framebuffer.nr_cbufs) {
1191       memcpy(&key->blend, lp->blend, sizeof key->blend);
1192    }
1193
1194    key->nr_cbufs = lp->framebuffer.nr_cbufs;
1195    for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
1196       struct pipe_rt_blend_state *blend_rt = &key->blend.rt[i];
1197       const struct util_format_description *format_desc;
1198       unsigned chan;
1199
1200       format_desc = util_format_description(lp->framebuffer.cbufs[i]->format);
1201       assert(format_desc->colorspace == UTIL_FORMAT_COLORSPACE_RGB ||
1202              format_desc->colorspace == UTIL_FORMAT_COLORSPACE_SRGB);
1203
1204       blend_rt->colormask = lp->blend->rt[i].colormask;
1205
1206       /* mask out color channels not present in the color buffer.
1207        * Should be simple to incorporate per-cbuf writemasks:
1208        */
1209       for(chan = 0; chan < 4; ++chan) {
1210          enum util_format_swizzle swizzle = format_desc->swizzle[chan];
1211
1212          if(swizzle > UTIL_FORMAT_SWIZZLE_W)
1213             blend_rt->colormask &= ~(1 << chan);
1214       }
1215
1216       /*
1217        * Our swizzled render tiles always have an alpha channel, but the linear
1218        * render target format often does not, so force here the dst alpha to be
1219        * one.
1220        *
1221        * This is not a mere optimization. Wrong results will be produced if the
1222        * dst alpha is used, the dst format does not have alpha, and the previous
1223        * rendering was not flushed from the swizzled to linear buffer. For
1224        * example, NonPowTwo DCT.
1225        *
1226        * TODO: This should be generalized to all channels for better
1227        * performance, but only alpha causes correctness issues.
1228        */
1229       if (format_desc->swizzle[3] > UTIL_FORMAT_SWIZZLE_W) {
1230          blend_rt->rgb_src_factor = force_dst_alpha_one(blend_rt->rgb_src_factor, FALSE);
1231          blend_rt->rgb_dst_factor = force_dst_alpha_one(blend_rt->rgb_dst_factor, FALSE);
1232          blend_rt->alpha_src_factor = force_dst_alpha_one(blend_rt->alpha_src_factor, TRUE);
1233          blend_rt->alpha_dst_factor = force_dst_alpha_one(blend_rt->alpha_dst_factor, TRUE);
1234       }
1235    }
1236
1237    for(i = 0; i < PIPE_MAX_SAMPLERS; ++i)
1238       if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i))
1239          lp_sampler_static_state(&key->sampler[i], lp->fragment_sampler_views[i], lp->sampler[i]);
1240 }
1241
1242
1243 /**
1244  * Update fragment state.  This is called just prior to drawing
1245  * something when some fragment-related state has changed.
1246  */
1247 void
1248 llvmpipe_update_fs(struct llvmpipe_context *lp)
1249 {
1250    struct lp_fragment_shader *shader = lp->fs;
1251    struct lp_fragment_shader_variant_key key;
1252    struct lp_fragment_shader_variant *variant;
1253
1254    make_variant_key(lp, shader, &key);
1255
1256    variant = shader->variants;
1257    while(variant) {
1258       if(memcmp(&variant->key, &key, sizeof key) == 0)
1259          break;
1260
1261       variant = variant->next;
1262    }
1263
1264    if (!variant) {
1265       int64_t t0, t1;
1266       int64_t dt;
1267       t0 = os_time_get();
1268
1269       variant = generate_variant(lp, shader, &key);
1270
1271       t1 = os_time_get();
1272       dt = t1 - t0;
1273       LP_COUNT_ADD(llvm_compile_time, dt);
1274       LP_COUNT_ADD(nr_llvm_compiles, 2);  /* emit vs. omit in/out test */
1275    }
1276
1277    lp_setup_set_fs_variant(lp->setup, variant);
1278 }
1279
1280
1281
1282 void
1283 llvmpipe_init_fs_funcs(struct llvmpipe_context *llvmpipe)
1284 {
1285    llvmpipe->pipe.create_fs_state = llvmpipe_create_fs_state;
1286    llvmpipe->pipe.bind_fs_state   = llvmpipe_bind_fs_state;
1287    llvmpipe->pipe.delete_fs_state = llvmpipe_delete_fs_state;
1288
1289    llvmpipe->pipe.set_constant_buffer = llvmpipe_set_constant_buffer;
1290 }