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