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