src/gallium/drivers/llvmpipe/lp_state_fs.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2009 VMware, Inc.
   4  * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
   5  * All Rights Reserved.
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * copy of this software and associated documentation files (the
   9  * "Software"), to deal in the Software without restriction, including
  10  * without limitation the rights to use, copy, modify, merge, publish,
  11  * distribute, sub license, and/or sell copies of the Software, and to
  12  * permit persons to whom the Software is furnished to do so, subject to
  13  * the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the
  16  * next paragraph) shall be included in all copies or substantial portions
  17  * of the Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  20  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  21  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  22  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
  23  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  24  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  25  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  26  *
  27  **************************************************************************/
  28
  29 /**
  30  * @file
  31  * Code generate the whole fragment pipeline.
  32  *
  33  * The fragment pipeline consists of the following stages:
  34  * - stipple (TBI)
  35  * - early depth test
  36  * - fragment shader
  37  * - alpha test
  38  * - depth/stencil test (stencil TBI)
  39  * - blending
  40  *
  41  * This file has only the glue to assembly the fragment pipeline.  The actual
  42  * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
  43  * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
  44  * muster the LLVM JIT execution engine to create a function that follows an
  45  * established binary interface and that can be called from C directly.
  46  *
  47  * A big source of complexity here is that we often want to run different
  48  * stages with different precisions and data types and precisions. For example,
  49  * the fragment shader needs typically to be done in floats, but the
  50  * depth/stencil test and blending is better done in the type that most closely
  51  * matches the depth/stencil and color buffer respectively.
  52  *
  53  * Since the width of a SIMD vector register stays the same regardless of the
  54  * element type, different types imply different number of elements, so we must
  55  * code generate more instances of the stages with larger types to be able to
  56  * feed/consume the stages with smaller types.
  57  *
  58  * @author Jose Fonseca <jfonseca@vmware.com>
  59  */
  60
  61 #include <limits.h>
  62 #include "pipe/p_defines.h"
  63 #include "util/u_memory.h"
  64 #include "util/u_format.h"
  65 #include "util/u_debug_dump.h"
  66 #include "pipe/internal/p_winsys_screen.h"
  67 #include "pipe/p_shader_tokens.h"
  68 #include "draw/draw_context.h"
  69 #include "tgsi/tgsi_dump.h"
  70 #include "tgsi/tgsi_scan.h"
  71 #include "tgsi/tgsi_parse.h"
  72 #include "lp_bld_type.h"
  73 #include "lp_bld_const.h"
  74 #include "lp_bld_conv.h"
  75 #include "lp_bld_intr.h"
  76 #include "lp_bld_logic.h"
  77 #include "lp_bld_depth.h"
  78 #include "lp_bld_interp.h"
  79 #include "lp_bld_tgsi.h"
  80 #include "lp_bld_alpha.h"
  81 #include "lp_bld_blend.h"
  82 #include "lp_bld_swizzle.h"
  83 #include "lp_bld_flow.h"
  84 #include "lp_bld_debug.h"
  85 #include "lp_screen.h"
  86 #include "lp_context.h"
  87 #include "lp_buffer.h"
  88 #include "lp_setup.h"
  89 #include "lp_state.h"
  90 #include "lp_tex_sample.h"
  91 #include "lp_debug.h"
  92
  93
  94 static const unsigned char quad_offset_x[4] = {0, 1, 0, 1};
  95 static const unsigned char quad_offset_y[4] = {0, 0, 1, 1};
  96
  97
  98 /*
  99  * Derive from the quad's upper left scalar coordinates the coordinates for
 100  * all other quad pixels
 101  */
 102 static void
 103 generate_pos0(LLVMBuilderRef builder,
 104               LLVMValueRef x,
 105               LLVMValueRef y,
 106               LLVMValueRef *x0,
 107               LLVMValueRef *y0)
 108 {
 109    LLVMTypeRef int_elem_type = LLVMInt32Type();
 110    LLVMTypeRef int_vec_type = LLVMVectorType(int_elem_type, QUAD_SIZE);
 111    LLVMTypeRef elem_type = LLVMFloatType();
 112    LLVMTypeRef vec_type = LLVMVectorType(elem_type, QUAD_SIZE);
 113    LLVMValueRef x_offsets[QUAD_SIZE];
 114    LLVMValueRef y_offsets[QUAD_SIZE];
 115    unsigned i;
 116
 117    x = lp_build_broadcast(builder, int_vec_type, x);
 118    y = lp_build_broadcast(builder, int_vec_type, y);
 119
 120    for(i = 0; i < QUAD_SIZE; ++i) {
 121       x_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_x[i], 0);
 122       y_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_y[i], 0);
 123    }
 124
 125    x = LLVMBuildAdd(builder, x, LLVMConstVector(x_offsets, QUAD_SIZE), "");
 126    y = LLVMBuildAdd(builder, y, LLVMConstVector(y_offsets, QUAD_SIZE), "");
 127
 128    *x0 = LLVMBuildSIToFP(builder, x, vec_type, "");
 129    *y0 = LLVMBuildSIToFP(builder, y, vec_type, "");
 130 }
 131
 132
 133 /**
 134  * Generate the depth test.
 135  */
 136 static void
 137 generate_depth(LLVMBuilderRef builder,
 138                const struct lp_fragment_shader_variant_key *key,
 139                struct lp_type src_type,
 140                struct lp_build_mask_context *mask,
 141                LLVMValueRef src,
 142                LLVMValueRef dst_ptr)
 143 {
 144    const struct util_format_description *format_desc;
 145    struct lp_type dst_type;
 146
 147    if(!key->depth.enabled)
 148       return;
 149
 150    format_desc = util_format_description(key->zsbuf_format);
 151    assert(format_desc);
 152
 153    /*
 154     * Depths are expected to be between 0 and 1, even if they are stored in
 155     * floats. Setting these bits here will ensure that the lp_build_conv() call
 156     * below won't try to unnecessarily clamp the incoming values.
 157     */
 158    if(src_type.floating) {
 159       src_type.sign = FALSE;
 160       src_type.norm = TRUE;
 161    }
 162    else {
 163       assert(!src_type.sign);
 164       assert(src_type.norm);
 165    }
 166
 167    /* Pick the depth type. */
 168    dst_type = lp_depth_type(format_desc, src_type.width*src_type.length);
 169
 170    /* FIXME: Cope with a depth test type with a different bit width. */
 171    assert(dst_type.width == src_type.width);
 172    assert(dst_type.length == src_type.length);
 173
 174    lp_build_conv(builder, src_type, dst_type, &src, 1, &src, 1);
 175
 176    dst_ptr = LLVMBuildBitCast(builder,
 177                               dst_ptr,
 178                               LLVMPointerType(lp_build_vec_type(dst_type), 0), "");
 179
 180    lp_build_depth_test(builder,
 181                        &key->depth,
 182                        dst_type,
 183                        format_desc,
 184                        mask,
 185                        src,
 186                        dst_ptr);
 187 }
 188
 189
 190 /**
 191  * Generate the code to do inside/outside triangle testing for the
 192  * four pixels in a 2x2 quad.  This will set the four elements of the
 193  * quad mask vector to 0 or ~0.
 194  * \param i  which quad of the quad group to test, in [0,3]
 195  */
 196 static void
 197 generate_tri_edge_mask(LLVMBuilderRef builder,
 198                        unsigned i,
 199                        LLVMValueRef *mask,      /* ivec4, out */
 200                        LLVMValueRef c0,         /* int32 */
 201                        LLVMValueRef c1,         /* int32 */
 202                        LLVMValueRef c2,         /* int32 */
 203                        LLVMValueRef step0_ptr,  /* ivec4 */
 204                        LLVMValueRef step1_ptr,  /* ivec4 */
 205                        LLVMValueRef step2_ptr)  /* ivec4 */
 206 {
 207    /*
 208      c0_vec = splat(c0)
 209      c1_vec = splat(c1)
 210      c2_vec = splat(c2)
 211      m0_vec = step0_ptr[i] > c0_vec
 212      m1_vec = step1_ptr[i] > c1_vec
 213      m2_vec = step2_ptr[i] > c2_vec
 214      mask = m0_vec & m1_vec & m2_vec
 215     */
 216    struct lp_build_flow_context *flow;
 217    struct lp_build_if_state ifctx;
 218    struct lp_type i32_type;
 219    LLVMTypeRef i32vec4_type, mask_type;
 220    LLVMValueRef c0_vec, c1_vec, c2_vec;
 221    LLVMValueRef not_draw_all;
 222    LLVMValueRef in_out_mask;
 223
 224    assert(i < 4);
 225
 226    /* int32 vector type */
 227    memset(&i32_type, 0, sizeof i32_type);
 228    i32_type.floating = FALSE; /* values are integers */
 229    i32_type.sign = TRUE;      /* values are signed */
 230    i32_type.norm = FALSE;     /* values are not normalized */
 231    i32_type.width = 32;       /* 32-bit int values */
 232    i32_type.length = 4;       /* 4 elements per vector */
 233
 234    i32vec4_type = lp_build_int32_vec4_type();
 235
 236    mask_type = LLVMIntType(32 * 4);
 237
 238    /*
 239     * Use a conditional here to do detailed pixel in/out testing.
 240     * We only have to do this if c0 != {INT_MIN, INT_MIN, INT_MIN, INT_MIN}
 241     */
 242    flow = lp_build_flow_create(builder);
 243    lp_build_flow_scope_begin(flow);
 244
 245    {
 246 #define OPTIMIZE_IN_OUT_TEST 1
 247 #if OPTIMIZE_IN_OUT_TEST
 248
 249       not_draw_all = LLVMBuildICmp(builder,
 250                                    LLVMIntNE,
 251                                    c0,
 252                                    LLVMConstInt(LLVMInt32Type(), INT_MIN, 0),
 253                                    "");
 254
 255       in_out_mask = lp_build_int_const_scalar(i32_type, ~0);
 256
 257
 258       lp_build_flow_scope_declare(flow, &in_out_mask);
 259
 260       lp_build_if(&ifctx, flow, builder, not_draw_all);
 261 #endif
 262       {
 263          LLVMValueRef step0_vec, step1_vec, step2_vec;
 264          LLVMValueRef m0_vec, m1_vec, m2_vec;
 265          LLVMValueRef index, m;
 266
 267          /* c0_vec = {c0, c0, c0, c0}
 268           * Note that we emit this code four times but LLVM optimizes away
 269           * three instances of it.
 270           */
 271          c0_vec = lp_build_broadcast(builder, i32vec4_type, c0);
 272          c1_vec = lp_build_broadcast(builder, i32vec4_type, c1);
 273          c2_vec = lp_build_broadcast(builder, i32vec4_type, c2);
 274          lp_build_name(c0_vec, "edgeconst0vec");
 275          lp_build_name(c1_vec, "edgeconst1vec");
 276          lp_build_name(c2_vec, "edgeconst2vec");
 277
 278
 279          index = LLVMConstInt(LLVMInt32Type(), i, 0);
 280          step0_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step0_ptr, &index, 1, ""), "");
 281          step1_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step1_ptr, &index, 1, ""), "");
 282          step2_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step2_ptr, &index, 1, ""), "");
 283
 284          lp_build_name(step0_vec, "step0vec");
 285          lp_build_name(step1_vec, "step1vec");
 286          lp_build_name(step2_vec, "step2vec");
 287
 288          m0_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step0_vec, c0_vec);
 289          m1_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step1_vec, c1_vec);
 290          m2_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step2_vec, c2_vec);
 291
 292          m = LLVMBuildAnd(builder, m0_vec, m1_vec, "");
 293          in_out_mask = LLVMBuildAnd(builder, m, m2_vec, "");
 294          lp_build_name(in_out_mask, "inoutmaskvec");
 295
 296          /* This is the initial alive/dead pixel mask.  Additional bits will get cleared
 297           * when the Z test fails, etc.
 298           */
 299       }
 300 #if OPTIMIZE_IN_OUT_TEST
 301       lp_build_endif(&ifctx);
 302 #endif
 303
 304    }
 305    lp_build_flow_scope_end(flow);
 306    lp_build_flow_destroy(flow);
 307
 308    *mask = in_out_mask;
 309 }
 310
 311
 312 /**
 313  * Generate the fragment shader, depth/stencil test, and alpha tests.
 314  * \param i  which quad in the tile, in range [0,3]
 315  */
 316 static void
 317 generate_fs(struct llvmpipe_context *lp,
 318             struct lp_fragment_shader *shader,
 319             const struct lp_fragment_shader_variant_key *key,
 320             LLVMBuilderRef builder,
 321             struct lp_type type,
 322             LLVMValueRef context_ptr,
 323             unsigned i,
 324             const struct lp_build_interp_soa_context *interp,
 325             struct lp_build_sampler_soa *sampler,
 326             LLVMValueRef *pmask,
 327             LLVMValueRef (*color)[4],
 328             LLVMValueRef depth_ptr,
 329             LLVMValueRef c0,
 330             LLVMValueRef c1,
 331             LLVMValueRef c2,
 332             LLVMValueRef step0_ptr,
 333             LLVMValueRef step1_ptr,
 334             LLVMValueRef step2_ptr)
 335 {
 336    const struct tgsi_token *tokens = shader->base.tokens;
 337    LLVMTypeRef elem_type;
 338    LLVMTypeRef vec_type;
 339    LLVMTypeRef int_vec_type;
 340    LLVMValueRef consts_ptr;
 341    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
 342    LLVMValueRef z = interp->pos[2];
 343    struct lp_build_flow_context *flow;
 344    struct lp_build_mask_context mask;
 345    boolean early_depth_test;
 346    unsigned attrib;
 347    unsigned chan;
 348    unsigned cbuf;
 349
 350    assert(i < 4);
 351
 352    elem_type = lp_build_elem_type(type);
 353    vec_type = lp_build_vec_type(type);
 354    int_vec_type = lp_build_int_vec_type(type);
 355
 356    consts_ptr = lp_jit_context_constants(builder, context_ptr);
 357
 358    flow = lp_build_flow_create(builder);
 359
 360    memset(outputs, 0, sizeof outputs);
 361
 362    lp_build_flow_scope_begin(flow);
 363
 364    /* Declare the color and z variables */
 365    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
 366       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 367          color[cbuf][chan] = LLVMGetUndef(vec_type);
 368          lp_build_flow_scope_declare(flow, &color[cbuf][chan]);
 369       }
 370    }
 371    lp_build_flow_scope_declare(flow, &z);
 372
 373    /* do triangle edge testing */
 374    generate_tri_edge_mask(builder, i, pmask,
 375                           c0, c1, c2, step0_ptr, step1_ptr, step2_ptr);
 376
 377    /* 'mask' will control execution based on quad's pixel alive/killed state */
 378    lp_build_mask_begin(&mask, flow, type, *pmask);
 379
 380
 381    early_depth_test =
 382       key->depth.enabled &&
 383       !key->alpha.enabled &&
 384       !shader->info.uses_kill &&
 385       !shader->info.writes_z;
 386
 387    if(early_depth_test)
 388       generate_depth(builder, key,
 389                      type, &mask,
 390                      z, depth_ptr);
 391
 392    lp_build_tgsi_soa(builder, tokens, type, &mask,
 393                      consts_ptr, interp->pos, interp->inputs,
 394                      outputs, sampler);
 395
 396    for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
 397       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 398          if(outputs[attrib][chan]) {
 399             lp_build_name(outputs[attrib][chan], "output%u.%u.%c", i, attrib, "xyzw"[chan]);
 400
 401             switch (shader->info.output_semantic_name[attrib]) {
 402             case TGSI_SEMANTIC_COLOR:
 403                {
 404                   unsigned cbuf = shader->info.output_semantic_index[attrib];
 405
 406                   lp_build_name(outputs[attrib][chan], "color%u.%u.%c", i, attrib, "rgba"[chan]);
 407
 408                   /* Alpha test */
 409                   /* XXX: should the alpha reference value be passed separately? */
 410                   /* XXX: should only test the final assignment to alpha */
 411                   if(cbuf == 0 && chan == 3) {
 412                      LLVMValueRef alpha = outputs[attrib][chan];
 413                      LLVMValueRef alpha_ref_value;
 414                      alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
 415                      alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
 416                      lp_build_alpha_test(builder, &key->alpha, type,
 417                                          &mask, alpha, alpha_ref_value);
 418                   }
 419
 420                   color[cbuf][chan] = outputs[attrib][chan];
 421                   break;
 422                }
 423
 424             case TGSI_SEMANTIC_POSITION:
 425                if(chan == 2)
 426                   z = outputs[attrib][chan];
 427                break;
 428             }
 429          }
 430       }
 431    }
 432
 433    if(!early_depth_test)
 434       generate_depth(builder, key,
 435                      type, &mask,
 436                      z, depth_ptr);
 437
 438    lp_build_mask_end(&mask);
 439
 440    lp_build_flow_scope_end(flow);
 441
 442    lp_build_flow_destroy(flow);
 443
 444    *pmask = mask.value;
 445
 446 }
 447
 448
 449 /**
 450  * Generate color blending and color output.
 451  */
 452 static void
 453 generate_blend(const struct pipe_blend_state *blend,
 454                LLVMBuilderRef builder,
 455                struct lp_type type,
 456                LLVMValueRef context_ptr,
 457                LLVMValueRef mask,
 458                LLVMValueRef *src,
 459                LLVMValueRef dst_ptr)
 460 {
 461    struct lp_build_context bld;
 462    struct lp_build_flow_context *flow;
 463    struct lp_build_mask_context mask_ctx;
 464    LLVMTypeRef vec_type;
 465    LLVMTypeRef int_vec_type;
 466    LLVMValueRef const_ptr;
 467    LLVMValueRef con[4];
 468    LLVMValueRef dst[4];
 469    LLVMValueRef res[4];
 470    unsigned chan;
 471
 472    lp_build_context_init(&bld, builder, type);
 473
 474    flow = lp_build_flow_create(builder);
 475
 476    /* we'll use this mask context to skip blending if all pixels are dead */
 477    lp_build_mask_begin(&mask_ctx, flow, type, mask);
 478
 479    vec_type = lp_build_vec_type(type);
 480    int_vec_type = lp_build_int_vec_type(type);
 481
 482    const_ptr = lp_jit_context_blend_color(builder, context_ptr);
 483    const_ptr = LLVMBuildBitCast(builder, const_ptr,
 484                                 LLVMPointerType(vec_type, 0), "");
 485
 486    for(chan = 0; chan < 4; ++chan) {
 487       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 488       con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
 489
 490       dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
 491
 492       lp_build_name(con[chan], "con.%c", "rgba"[chan]);
 493       lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
 494    }
 495
 496    lp_build_blend_soa(builder, blend, type, src, dst, con, res);
 497
 498    for(chan = 0; chan < 4; ++chan) {
 499       if(blend->colormask & (1 << chan)) {
 500          LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 501          lp_build_name(res[chan], "res.%c", "rgba"[chan]);
 502          res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
 503          LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
 504       }
 505    }
 506
 507    lp_build_mask_end(&mask_ctx);
 508    lp_build_flow_destroy(flow);
 509 }
 510
 511
 512 /**
 513  * Generate the runtime callable function for the whole fragment pipeline.
 514  * Note that the function which we generate operates on a block of 16
 515  * pixels at at time.  The block contains 2x2 quads.  Each quad contains
 516  * 2x2 pixels.
 517  */
 518 static struct lp_fragment_shader_variant *
 519 generate_fragment(struct llvmpipe_context *lp,
 520                   struct lp_fragment_shader *shader,
 521                   const struct lp_fragment_shader_variant_key *key)
 522 {
 523    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
 524    struct lp_fragment_shader_variant *variant;
 525    struct lp_type fs_type;
 526    struct lp_type blend_type;
 527    LLVMTypeRef fs_elem_type;
 528    LLVMTypeRef fs_vec_type;
 529    LLVMTypeRef fs_int_vec_type;
 530    LLVMTypeRef blend_vec_type;
 531    LLVMTypeRef blend_int_vec_type;
 532    LLVMTypeRef arg_types[14];
 533    LLVMTypeRef func_type;
 534    LLVMTypeRef int32_vec4_type = lp_build_int32_vec4_type();
 535    LLVMValueRef context_ptr;
 536    LLVMValueRef x;
 537    LLVMValueRef y;
 538    LLVMValueRef a0_ptr;
 539    LLVMValueRef dadx_ptr;
 540    LLVMValueRef dady_ptr;
 541    LLVMValueRef color_ptr_ptr;
 542    LLVMValueRef depth_ptr;
 543    LLVMValueRef c0, c1, c2, step0_ptr, step1_ptr, step2_ptr;
 544    LLVMBasicBlockRef block;
 545    LLVMBuilderRef builder;
 546    LLVMValueRef x0;
 547    LLVMValueRef y0;
 548    struct lp_build_sampler_soa *sampler;
 549    struct lp_build_interp_soa_context interp;
 550    LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
 551    LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
 552    LLVMValueRef blend_mask;
 553    LLVMValueRef blend_in_color[NUM_CHANNELS];
 554    unsigned num_fs;
 555    unsigned i;
 556    unsigned chan;
 557    unsigned cbuf;
 558
 559    if (LP_DEBUG & DEBUG_JIT) {
 560       tgsi_dump(shader->base.tokens, 0);
 561       if(key->depth.enabled) {
 562          debug_printf("depth.format = %s\n", pf_name(key->zsbuf_format));
 563          debug_printf("depth.func = %s\n", debug_dump_func(key->depth.func, TRUE));
 564          debug_printf("depth.writemask = %u\n", key->depth.writemask);
 565       }
 566       if(key->alpha.enabled) {
 567          debug_printf("alpha.func = %s\n", debug_dump_func(key->alpha.func, TRUE));
 568          debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
 569       }
 570       if(key->blend.logicop_enable) {
 571          debug_printf("blend.logicop_func = %u\n", key->blend.logicop_func);
 572       }
 573       else if(key->blend.blend_enable) {
 574          debug_printf("blend.rgb_func = %s\n",   debug_dump_blend_func  (key->blend.rgb_func, TRUE));
 575          debug_printf("rgb_src_factor = %s\n",   debug_dump_blend_factor(key->blend.rgb_src_factor, TRUE));
 576          debug_printf("rgb_dst_factor = %s\n",   debug_dump_blend_factor(key->blend.rgb_dst_factor, TRUE));
 577          debug_printf("alpha_func = %s\n",       debug_dump_blend_func  (key->blend.alpha_func, TRUE));
 578          debug_printf("alpha_src_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_src_factor, TRUE));
 579          debug_printf("alpha_dst_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_dst_factor, TRUE));
 580       }
 581       debug_printf("blend.colormask = 0x%x\n", key->blend.colormask);
 582       for(i = 0; i < PIPE_MAX_SAMPLERS; ++i) {
 583          if(key->sampler[i].format) {
 584             debug_printf("sampler[%u] = \n", i);
 585             debug_printf("  .format = %s\n",
 586                          pf_name(key->sampler[i].format));
 587             debug_printf("  .target = %s\n",
 588                          debug_dump_tex_target(key->sampler[i].target, TRUE));
 589             debug_printf("  .pot = %u %u %u\n",
 590                          key->sampler[i].pot_width,
 591                          key->sampler[i].pot_height,
 592                          key->sampler[i].pot_depth);
 593             debug_printf("  .wrap = %s %s %s\n",
 594                          debug_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
 595                          debug_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
 596                          debug_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
 597             debug_printf("  .min_img_filter = %s\n",
 598                          debug_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
 599             debug_printf("  .min_mip_filter = %s\n",
 600                          debug_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
 601             debug_printf("  .mag_img_filter = %s\n",
 602                          debug_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
 603             if(key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE)
 604                debug_printf("  .compare_func = %s\n", debug_dump_func(key->sampler[i].compare_func, TRUE));
 605             debug_printf("  .normalized_coords = %u\n", key->sampler[i].normalized_coords);
 606             debug_printf("  .prefilter = %u\n", key->sampler[i].prefilter);
 607          }
 608       }
 609    }
 610
 611    variant = CALLOC_STRUCT(lp_fragment_shader_variant);
 612    if(!variant)
 613       return NULL;
 614
 615    variant->shader = shader;
 616    memcpy(&variant->key, key, sizeof *key);
 617
 618    /* TODO: actually pick these based on the fs and color buffer
 619     * characteristics. */
 620
 621    memset(&fs_type, 0, sizeof fs_type);
 622    fs_type.floating = TRUE; /* floating point values */
 623    fs_type.sign = TRUE;     /* values are signed */
 624    fs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
 625    fs_type.width = 32;      /* 32-bit float */
 626    fs_type.length = 4;      /* 4 elements per vector */
 627    num_fs = 4;              /* number of quads per block */
 628
 629    memset(&blend_type, 0, sizeof blend_type);
 630    blend_type.floating = FALSE; /* values are integers */
 631    blend_type.sign = FALSE;     /* values are unsigned */
 632    blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
 633    blend_type.width = 8;        /* 8-bit ubyte values */
 634    blend_type.length = 16;      /* 16 elements per vector */
 635
 636    /*
 637     * Generate the function prototype. Any change here must be reflected in
 638     * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
 639     */
 640
 641    fs_elem_type = lp_build_elem_type(fs_type);
 642    fs_vec_type = lp_build_vec_type(fs_type);
 643    fs_int_vec_type = lp_build_int_vec_type(fs_type);
 644
 645    blend_vec_type = lp_build_vec_type(blend_type);
 646    blend_int_vec_type = lp_build_int_vec_type(blend_type);
 647
 648    arg_types[0] = screen->context_ptr_type;            /* context */
 649    arg_types[1] = LLVMInt32Type();                     /* x */
 650    arg_types[2] = LLVMInt32Type();                     /* y */
 651    arg_types[3] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
 652    arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
 653    arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* dady */
 654    arg_types[6] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0);  /* color */
 655    arg_types[7] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
 656    arg_types[8] = LLVMInt32Type();                     /* c0 */
 657    arg_types[9] = LLVMInt32Type();                    /* c1 */
 658    arg_types[10] = LLVMInt32Type();                    /* c2 */
 659    /* Note: the step arrays are built as int32[16] but we interpret
 660     * them here as int32_vec4[4].
 661     */
 662    arg_types[11] = LLVMPointerType(int32_vec4_type, 0);/* step0 */
 663    arg_types[12] = LLVMPointerType(int32_vec4_type, 0);/* step1 */
 664    arg_types[13] = LLVMPointerType(int32_vec4_type, 0);/* step2 */
 665
 666    func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
 667
 668    variant->function = LLVMAddFunction(screen->module, "shader", func_type);
 669    LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
 670
 671    /* XXX: need to propagate noalias down into color param now we are
 672     * passing a pointer-to-pointer?
 673     */
 674    for(i = 0; i < Elements(arg_types); ++i)
 675       if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
 676          LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute);
 677
 678    context_ptr  = LLVMGetParam(variant->function, 0);
 679    x            = LLVMGetParam(variant->function, 1);
 680    y            = LLVMGetParam(variant->function, 2);
 681    a0_ptr       = LLVMGetParam(variant->function, 3);
 682    dadx_ptr     = LLVMGetParam(variant->function, 4);
 683    dady_ptr     = LLVMGetParam(variant->function, 5);
 684    color_ptr_ptr = LLVMGetParam(variant->function, 6);
 685    depth_ptr    = LLVMGetParam(variant->function, 7);
 686    c0           = LLVMGetParam(variant->function, 8);
 687    c1           = LLVMGetParam(variant->function, 9);
 688    c2           = LLVMGetParam(variant->function, 10);
 689    step0_ptr    = LLVMGetParam(variant->function, 11);
 690    step1_ptr    = LLVMGetParam(variant->function, 12);
 691    step2_ptr    = LLVMGetParam(variant->function, 13);
 692
 693    lp_build_name(context_ptr, "context");
 694    lp_build_name(x, "x");
 695    lp_build_name(y, "y");
 696    lp_build_name(a0_ptr, "a0");
 697    lp_build_name(dadx_ptr, "dadx");
 698    lp_build_name(dady_ptr, "dady");
 699    lp_build_name(color_ptr_ptr, "color_ptr");
 700    lp_build_name(depth_ptr, "depth");
 701    lp_build_name(c0, "c0");
 702    lp_build_name(c1, "c1");
 703    lp_build_name(c2, "c2");
 704    lp_build_name(step0_ptr, "step0");
 705    lp_build_name(step1_ptr, "step1");
 706    lp_build_name(step2_ptr, "step2");
 707
 708    /*
 709     * Function body
 710     */
 711
 712    block = LLVMAppendBasicBlock(variant->function, "entry");
 713    builder = LLVMCreateBuilder();
 714    LLVMPositionBuilderAtEnd(builder, block);
 715
 716    generate_pos0(builder, x, y, &x0, &y0);
 717
 718    lp_build_interp_soa_init(&interp,
 719                             shader->base.tokens,
 720                             key->flatshade,
 721                             builder, fs_type,
 722                             a0_ptr, dadx_ptr, dady_ptr,
 723                             x0, y0);
 724
 725    /* code generated texture sampling */
 726    sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
 727
 728    /* loop over quads in the block */
 729    for(i = 0; i < num_fs; ++i) {
 730       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
 731       LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
 732       LLVMValueRef depth_ptr_i;
 733       int cbuf;
 734
 735       if(i != 0)
 736          lp_build_interp_soa_update(&interp, i);
 737
 738       depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
 739
 740       generate_fs(lp, shader, key,
 741                   builder,
 742                   fs_type,
 743                   context_ptr,
 744                   i,
 745                   &interp,
 746                   sampler,
 747                   &fs_mask[i], /* output */
 748                   out_color,
 749                   depth_ptr_i,
 750                   c0, c1, c2,
 751                   step0_ptr, step1_ptr, step2_ptr);
 752
 753       for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
 754          for(chan = 0; chan < NUM_CHANNELS; ++chan)
 755             fs_out_color[cbuf][chan][i] = out_color[cbuf][chan];
 756    }
 757
 758    sampler->destroy(sampler);
 759
 760    /* Loop over color outputs / color buffers to do blending.
 761     */
 762    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
 763       LLVMValueRef color_ptr;
 764       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0);
 765
 766       /*
 767        * Convert the fs's output color and mask to fit to the blending type.
 768        */
 769       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 770          lp_build_conv(builder, fs_type, blend_type,
 771                        fs_out_color[cbuf][chan], num_fs,
 772                        &blend_in_color[chan], 1);
 773          lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
 774       }
 775
 776       lp_build_conv_mask(builder, fs_type, blend_type,
 777                          fs_mask, num_fs,
 778                          &blend_mask, 1);
 779
 780       color_ptr = LLVMBuildLoad(builder,
 781                                 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
 782                                 "");
 783       lp_build_name(color_ptr, "color_ptr%d", cbuf);
 784
 785       /*
 786        * Blending.
 787        */
 788       generate_blend(&key->blend,
 789                      builder,
 790                      blend_type,
 791                      context_ptr,
 792                      blend_mask,
 793                      blend_in_color,
 794                      color_ptr);
 795    }
 796
 797    LLVMBuildRetVoid(builder);
 798
 799    LLVMDisposeBuilder(builder);
 800
 801
 802    /* Verify the LLVM IR.  If invalid, dump and abort */
 803 #ifdef DEBUG
 804    if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) {
 805       if (1)
 806          LLVMDumpValue(variant->function);
 807       abort();
 808    }
 809 #endif
 810
 811    /* Apply optimizations to LLVM IR */
 812    if (1)
 813       LLVMRunFunctionPassManager(screen->pass, variant->function);
 814
 815    if (LP_DEBUG & DEBUG_JIT) {
 816       /* Print the LLVM IR to stderr */
 817       LLVMDumpValue(variant->function);
 818       debug_printf("\n");
 819    }
 820
 821    /*
 822     * Translate the LLVM IR into machine code.
 823     */
 824    variant->jit_function = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, variant->function);
 825
 826    if (LP_DEBUG & DEBUG_ASM)
 827       lp_disassemble(variant->jit_function);
 828
 829    variant->next = shader->variants;
 830    shader->variants = variant;
 831
 832    return variant;
 833 }
 834
 835
 836 void *
 837 llvmpipe_create_fs_state(struct pipe_context *pipe,
 838                          const struct pipe_shader_state *templ)
 839 {
 840    struct lp_fragment_shader *shader;
 841
 842    shader = CALLOC_STRUCT(lp_fragment_shader);
 843    if (!shader)
 844       return NULL;
 845
 846    /* get/save the summary info for this shader */
 847    tgsi_scan_shader(templ->tokens, &shader->info);
 848
 849    /* we need to keep a local copy of the tokens */
 850    shader->base.tokens = tgsi_dup_tokens(templ->tokens);
 851
 852    return shader;
 853 }
 854
 855
 856 void
 857 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
 858 {
 859    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 860
 861    if (llvmpipe->fs == fs)
 862       return;
 863
 864    draw_flush(llvmpipe->draw);
 865
 866    llvmpipe->fs = fs;
 867
 868    llvmpipe->dirty |= LP_NEW_FS;
 869 }
 870
 871
 872 void
 873 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
 874 {
 875    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 876    struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
 877    struct lp_fragment_shader *shader = fs;
 878    struct lp_fragment_shader_variant *variant;
 879
 880    assert(fs != llvmpipe->fs);
 881    (void) llvmpipe;
 882
 883    variant = shader->variants;
 884    while(variant) {
 885       struct lp_fragment_shader_variant *next = variant->next;
 886
 887       if(variant->function) {
 888          if(variant->jit_function)
 889             LLVMFreeMachineCodeForFunction(screen->engine, variant->function);
 890          LLVMDeleteFunction(variant->function);
 891       }
 892
 893       FREE(variant);
 894
 895       variant = next;
 896    }
 897
 898    FREE((void *) shader->base.tokens);
 899    FREE(shader);
 900 }
 901
 902
 903
 904 void
 905 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
 906                              uint shader, uint index,
 907                              const struct pipe_constant_buffer *constants)
 908 {
 909    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 910    struct pipe_buffer *buffer = constants ? constants->buffer : NULL;
 911    unsigned size = buffer ? buffer->size : 0;
 912    const void *data = buffer ? llvmpipe_buffer(buffer)->data : NULL;
 913
 914    assert(shader < PIPE_SHADER_TYPES);
 915    assert(index == 0);
 916
 917    if(llvmpipe->constants[shader].buffer == buffer)
 918       return;
 919
 920    draw_flush(llvmpipe->draw);
 921
 922    /* note: reference counting */
 923    pipe_buffer_reference(&llvmpipe->constants[shader].buffer, buffer);
 924
 925    if(shader == PIPE_SHADER_VERTEX) {
 926       draw_set_mapped_constant_buffer(llvmpipe->draw, PIPE_SHADER_VERTEX,
 927                                       data, size);
 928    }
 929
 930    llvmpipe->dirty |= LP_NEW_CONSTANTS;
 931 }
 932
 933
 934 /**
 935  * We need to generate several variants of the fragment pipeline to match
 936  * all the combinations of the contributing state atoms.
 937  *
 938  * TODO: there is actually no reason to tie this to context state -- the
 939  * generated code could be cached globally in the screen.
 940  */
 941 static void
 942 make_variant_key(struct llvmpipe_context *lp,
 943                  struct lp_fragment_shader *shader,
 944                  struct lp_fragment_shader_variant_key *key)
 945 {
 946    unsigned i;
 947
 948    memset(key, 0, sizeof *key);
 949
 950    if(lp->framebuffer.zsbuf &&
 951       lp->depth_stencil->depth.enabled) {
 952       key->zsbuf_format = lp->framebuffer.zsbuf->format;
 953       memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
 954    }
 955
 956    key->alpha.enabled = lp->depth_stencil->alpha.enabled;
 957    if(key->alpha.enabled)
 958       key->alpha.func = lp->depth_stencil->alpha.func;
 959    /* alpha.ref_value is passed in jit_context */
 960
 961    key->flatshade = lp->rasterizer->flatshade;
 962
 963    if (lp->framebuffer.nr_cbufs) {
 964       memcpy(&key->blend, lp->blend, sizeof key->blend);
 965    }
 966
 967    key->nr_cbufs = lp->framebuffer.nr_cbufs;
 968    for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
 969       const struct util_format_description *format_desc;
 970       unsigned chan;
 971
 972       format_desc = util_format_description(lp->framebuffer.cbufs[i]->format);
 973       assert(format_desc->layout == UTIL_FORMAT_COLORSPACE_RGB ||
 974              format_desc->layout == UTIL_FORMAT_COLORSPACE_SRGB);
 975
 976       /* mask out color channels not present in the color buffer.
 977        * Should be simple to incorporate per-cbuf writemasks:
 978        */
 979       for(chan = 0; chan < 4; ++chan) {
 980          enum util_format_swizzle swizzle = format_desc->swizzle[chan];
 981
 982          if(swizzle <= UTIL_FORMAT_SWIZZLE_W)
 983             key->cbuf_blend[i].colormask |= (1 << chan);
 984       }
 985    }
 986
 987    for(i = 0; i < PIPE_MAX_SAMPLERS; ++i)
 988       if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i))
 989          lp_sampler_static_state(&key->sampler[i], lp->texture[i], lp->sampler[i]);
 990 }
 991
 992
 993 void
 994 llvmpipe_update_fs(struct llvmpipe_context *lp)
 995 {
 996    struct lp_fragment_shader *shader = lp->fs;
 997    struct lp_fragment_shader_variant_key key;
 998    struct lp_fragment_shader_variant *variant;
 999
1000    make_variant_key(lp, shader, &key);
1001
1002    variant = shader->variants;
1003    while(variant) {
1004       if(memcmp(&variant->key, &key, sizeof key) == 0)
1005          break;
1006
1007       variant = variant->next;
1008    }
1009
1010    if(!variant)
1011       variant = generate_fragment(lp, shader, &key);
1012
1013    shader->current = variant;
1014
1015    lp_setup_set_fs_function(lp->setup,
1016                             shader->current->jit_function);
1017 }