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 #define OPTIMIZE_IN_OUT_TEST 0
 208 #if OPTIMIZE_IN_OUT_TEST
 209    struct lp_build_if_state ifctx;
 210    LLVMValueRef not_draw_all;
 211 #endif
 212    struct lp_build_flow_context *flow;
 213    struct lp_type i32_type;
 214    LLVMTypeRef i32vec4_type, mask_type;
 215    LLVMValueRef c0_vec, c1_vec, c2_vec;
 216    LLVMValueRef in_out_mask;
 217
 218    assert(i < 4);
 219
 220    /* int32 vector type */
 221    memset(&i32_type, 0, sizeof i32_type);
 222    i32_type.floating = FALSE; /* values are integers */
 223    i32_type.sign = TRUE;      /* values are signed */
 224    i32_type.norm = FALSE;     /* values are not normalized */
 225    i32_type.width = 32;       /* 32-bit int values */
 226    i32_type.length = 4;       /* 4 elements per vector */
 227
 228    i32vec4_type = lp_build_int32_vec4_type();
 229
 230    mask_type = LLVMIntType(32 * 4);
 231
 232    /*
 233     * Use a conditional here to do detailed pixel in/out testing.
 234     * We only have to do this if c0 != INT_MIN.
 235     */
 236    flow = lp_build_flow_create(builder);
 237    lp_build_flow_scope_begin(flow);
 238
 239    {
 240 #if OPTIMIZE_IN_OUT_TEST
 241       /* not_draw_all = (c0 != INT_MIN) */
 242       not_draw_all = LLVMBuildICmp(builder,
 243                                    LLVMIntNE,
 244                                    c0,
 245                                    LLVMConstInt(LLVMInt32Type(), INT_MIN, 0),
 246                                    "");
 247
 248       in_out_mask = lp_build_int_const_scalar(i32_type, ~0);
 249
 250
 251       lp_build_flow_scope_declare(flow, &in_out_mask);
 252
 253       /* if (not_draw_all) {... */
 254       lp_build_if(&ifctx, flow, builder, not_draw_all);
 255 #endif
 256       {
 257          LLVMValueRef step0_vec, step1_vec, step2_vec;
 258          LLVMValueRef m0_vec, m1_vec, m2_vec;
 259          LLVMValueRef index, m;
 260
 261          /* c0_vec = {c0, c0, c0, c0}
 262           * Note that we emit this code four times but LLVM optimizes away
 263           * three instances of it.
 264           */
 265          c0_vec = lp_build_broadcast(builder, i32vec4_type, c0);
 266          c1_vec = lp_build_broadcast(builder, i32vec4_type, c1);
 267          c2_vec = lp_build_broadcast(builder, i32vec4_type, c2);
 268          lp_build_name(c0_vec, "edgeconst0vec");
 269          lp_build_name(c1_vec, "edgeconst1vec");
 270          lp_build_name(c2_vec, "edgeconst2vec");
 271
 272          /* load step0vec, step1, step2 vec from memory */
 273          index = LLVMConstInt(LLVMInt32Type(), i, 0);
 274          step0_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step0_ptr, &index, 1, ""), "");
 275          step1_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step1_ptr, &index, 1, ""), "");
 276          step2_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step2_ptr, &index, 1, ""), "");
 277          lp_build_name(step0_vec, "step0vec");
 278          lp_build_name(step1_vec, "step1vec");
 279          lp_build_name(step2_vec, "step2vec");
 280
 281          /* m0_vec = step0_ptr[i] > c0_vec */
 282          m0_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step0_vec, c0_vec);
 283          m1_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step1_vec, c1_vec);
 284          m2_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step2_vec, c2_vec);
 285
 286          /* in_out_mask = m0_vec & m1_vec & m2_vec */
 287          m = LLVMBuildAnd(builder, m0_vec, m1_vec, "");
 288          in_out_mask = LLVMBuildAnd(builder, m, m2_vec, "");
 289          lp_build_name(in_out_mask, "inoutmaskvec");
 290       }
 291 #if OPTIMIZE_IN_OUT_TEST
 292       lp_build_endif(&ifctx);
 293 #endif
 294
 295    }
 296    lp_build_flow_scope_end(flow);
 297    lp_build_flow_destroy(flow);
 298
 299    /* This is the initial alive/dead pixel mask for a quad of four pixels.
 300     * It's an int[4] vector with each word set to 0 or ~0.
 301     * Words will get cleared when pixels faile the Z test, etc.
 302     */
 303    *mask = in_out_mask;
 304 }
 305
 306
 307 /**
 308  * Generate the fragment shader, depth/stencil test, and alpha tests.
 309  * \param i  which quad in the tile, in range [0,3]
 310  */
 311 static void
 312 generate_fs(struct llvmpipe_context *lp,
 313             struct lp_fragment_shader *shader,
 314             const struct lp_fragment_shader_variant_key *key,
 315             LLVMBuilderRef builder,
 316             struct lp_type type,
 317             LLVMValueRef context_ptr,
 318             unsigned i,
 319             const struct lp_build_interp_soa_context *interp,
 320             struct lp_build_sampler_soa *sampler,
 321             LLVMValueRef *pmask,
 322             LLVMValueRef (*color)[4],
 323             LLVMValueRef depth_ptr,
 324             LLVMValueRef c0,
 325             LLVMValueRef c1,
 326             LLVMValueRef c2,
 327             LLVMValueRef step0_ptr,
 328             LLVMValueRef step1_ptr,
 329             LLVMValueRef step2_ptr)
 330 {
 331    const struct tgsi_token *tokens = shader->base.tokens;
 332    LLVMTypeRef elem_type;
 333    LLVMTypeRef vec_type;
 334    LLVMTypeRef int_vec_type;
 335    LLVMValueRef consts_ptr;
 336    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
 337    LLVMValueRef z = interp->pos[2];
 338    struct lp_build_flow_context *flow;
 339    struct lp_build_mask_context mask;
 340    boolean early_depth_test;
 341    unsigned attrib;
 342    unsigned chan;
 343    unsigned cbuf;
 344
 345    assert(i < 4);
 346
 347    elem_type = lp_build_elem_type(type);
 348    vec_type = lp_build_vec_type(type);
 349    int_vec_type = lp_build_int_vec_type(type);
 350
 351    consts_ptr = lp_jit_context_constants(builder, context_ptr);
 352
 353    flow = lp_build_flow_create(builder);
 354
 355    memset(outputs, 0, sizeof outputs);
 356
 357    lp_build_flow_scope_begin(flow);
 358
 359    /* Declare the color and z variables */
 360    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
 361       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 362          color[cbuf][chan] = LLVMGetUndef(vec_type);
 363          lp_build_flow_scope_declare(flow, &color[cbuf][chan]);
 364       }
 365    }
 366    lp_build_flow_scope_declare(flow, &z);
 367
 368    /* do triangle edge testing */
 369    generate_tri_edge_mask(builder, i, pmask,
 370                           c0, c1, c2, step0_ptr, step1_ptr, step2_ptr);
 371
 372    /* 'mask' will control execution based on quad's pixel alive/killed state */
 373    lp_build_mask_begin(&mask, flow, type, *pmask);
 374
 375
 376    early_depth_test =
 377       key->depth.enabled &&
 378       !key->alpha.enabled &&
 379       !shader->info.uses_kill &&
 380       !shader->info.writes_z;
 381
 382    if(early_depth_test)
 383       generate_depth(builder, key,
 384                      type, &mask,
 385                      z, depth_ptr);
 386
 387    lp_build_tgsi_soa(builder, tokens, type, &mask,
 388                      consts_ptr, interp->pos, interp->inputs,
 389                      outputs, sampler);
 390
 391    for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
 392       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 393          if(outputs[attrib][chan]) {
 394             lp_build_name(outputs[attrib][chan], "output%u.%u.%c", i, attrib, "xyzw"[chan]);
 395
 396             switch (shader->info.output_semantic_name[attrib]) {
 397             case TGSI_SEMANTIC_COLOR:
 398                {
 399                   unsigned cbuf = shader->info.output_semantic_index[attrib];
 400
 401                   lp_build_name(outputs[attrib][chan], "color%u.%u.%c", i, attrib, "rgba"[chan]);
 402
 403                   /* Alpha test */
 404                   /* XXX: should the alpha reference value be passed separately? */
 405                   /* XXX: should only test the final assignment to alpha */
 406                   if(cbuf == 0 && chan == 3) {
 407                      LLVMValueRef alpha = outputs[attrib][chan];
 408                      LLVMValueRef alpha_ref_value;
 409                      alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
 410                      alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
 411                      lp_build_alpha_test(builder, &key->alpha, type,
 412                                          &mask, alpha, alpha_ref_value);
 413                   }
 414
 415                   color[cbuf][chan] = outputs[attrib][chan];
 416                   break;
 417                }
 418
 419             case TGSI_SEMANTIC_POSITION:
 420                if(chan == 2)
 421                   z = outputs[attrib][chan];
 422                break;
 423             }
 424          }
 425       }
 426    }
 427
 428    if(!early_depth_test)
 429       generate_depth(builder, key,
 430                      type, &mask,
 431                      z, depth_ptr);
 432
 433    lp_build_mask_end(&mask);
 434
 435    lp_build_flow_scope_end(flow);
 436
 437    lp_build_flow_destroy(flow);
 438
 439    *pmask = mask.value;
 440
 441 }
 442
 443
 444 /**
 445  * Generate color blending and color output.
 446  */
 447 static void
 448 generate_blend(const struct pipe_blend_state *blend,
 449                LLVMBuilderRef builder,
 450                struct lp_type type,
 451                LLVMValueRef context_ptr,
 452                LLVMValueRef mask,
 453                LLVMValueRef *src,
 454                LLVMValueRef dst_ptr)
 455 {
 456    struct lp_build_context bld;
 457    struct lp_build_flow_context *flow;
 458    struct lp_build_mask_context mask_ctx;
 459    LLVMTypeRef vec_type;
 460    LLVMTypeRef int_vec_type;
 461    LLVMValueRef const_ptr;
 462    LLVMValueRef con[4];
 463    LLVMValueRef dst[4];
 464    LLVMValueRef res[4];
 465    unsigned chan;
 466
 467    lp_build_context_init(&bld, builder, type);
 468
 469    flow = lp_build_flow_create(builder);
 470
 471    /* we'll use this mask context to skip blending if all pixels are dead */
 472    lp_build_mask_begin(&mask_ctx, flow, type, mask);
 473
 474    vec_type = lp_build_vec_type(type);
 475    int_vec_type = lp_build_int_vec_type(type);
 476
 477    const_ptr = lp_jit_context_blend_color(builder, context_ptr);
 478    const_ptr = LLVMBuildBitCast(builder, const_ptr,
 479                                 LLVMPointerType(vec_type, 0), "");
 480
 481    for(chan = 0; chan < 4; ++chan) {
 482       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 483       con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
 484
 485       dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
 486
 487       lp_build_name(con[chan], "con.%c", "rgba"[chan]);
 488       lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
 489    }
 490
 491    lp_build_blend_soa(builder, blend, type, src, dst, con, res);
 492
 493    for(chan = 0; chan < 4; ++chan) {
 494       if(blend->colormask & (1 << chan)) {
 495          LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 496          lp_build_name(res[chan], "res.%c", "rgba"[chan]);
 497          res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
 498          LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
 499       }
 500    }
 501
 502    lp_build_mask_end(&mask_ctx);
 503    lp_build_flow_destroy(flow);
 504 }
 505
 506
 507 /**
 508  * Generate the runtime callable function for the whole fragment pipeline.
 509  * Note that the function which we generate operates on a block of 16
 510  * pixels at at time.  The block contains 2x2 quads.  Each quad contains
 511  * 2x2 pixels.
 512  */
 513 static void
 514 generate_fragment(struct llvmpipe_context *lp,
 515                   struct lp_fragment_shader *shader,
 516                   struct lp_fragment_shader_variant *variant)
 517 {
 518    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
 519    const struct lp_fragment_shader_variant_key *key = &variant->key;
 520    struct lp_type fs_type;
 521    struct lp_type blend_type;
 522    LLVMTypeRef fs_elem_type;
 523    LLVMTypeRef fs_vec_type;
 524    LLVMTypeRef fs_int_vec_type;
 525    LLVMTypeRef blend_vec_type;
 526    LLVMTypeRef blend_int_vec_type;
 527    LLVMTypeRef arg_types[14];
 528    LLVMTypeRef func_type;
 529    LLVMTypeRef int32_vec4_type = lp_build_int32_vec4_type();
 530    LLVMValueRef context_ptr;
 531    LLVMValueRef x;
 532    LLVMValueRef y;
 533    LLVMValueRef a0_ptr;
 534    LLVMValueRef dadx_ptr;
 535    LLVMValueRef dady_ptr;
 536    LLVMValueRef color_ptr_ptr;
 537    LLVMValueRef depth_ptr;
 538    LLVMValueRef c0, c1, c2, step0_ptr, step1_ptr, step2_ptr;
 539    LLVMBasicBlockRef block;
 540    LLVMBuilderRef builder;
 541    LLVMValueRef x0;
 542    LLVMValueRef y0;
 543    struct lp_build_sampler_soa *sampler;
 544    struct lp_build_interp_soa_context interp;
 545    LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
 546    LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
 547    LLVMValueRef blend_mask;
 548    LLVMValueRef blend_in_color[NUM_CHANNELS];
 549    unsigned num_fs;
 550    unsigned i;
 551    unsigned chan;
 552    unsigned cbuf;
 553
 554
 555    /* TODO: actually pick these based on the fs and color buffer
 556     * characteristics. */
 557
 558    memset(&fs_type, 0, sizeof fs_type);
 559    fs_type.floating = TRUE; /* floating point values */
 560    fs_type.sign = TRUE;     /* values are signed */
 561    fs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
 562    fs_type.width = 32;      /* 32-bit float */
 563    fs_type.length = 4;      /* 4 elements per vector */
 564    num_fs = 4;              /* number of quads per block */
 565
 566    memset(&blend_type, 0, sizeof blend_type);
 567    blend_type.floating = FALSE; /* values are integers */
 568    blend_type.sign = FALSE;     /* values are unsigned */
 569    blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
 570    blend_type.width = 8;        /* 8-bit ubyte values */
 571    blend_type.length = 16;      /* 16 elements per vector */
 572
 573    /*
 574     * Generate the function prototype. Any change here must be reflected in
 575     * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
 576     */
 577
 578    fs_elem_type = lp_build_elem_type(fs_type);
 579    fs_vec_type = lp_build_vec_type(fs_type);
 580    fs_int_vec_type = lp_build_int_vec_type(fs_type);
 581
 582    blend_vec_type = lp_build_vec_type(blend_type);
 583    blend_int_vec_type = lp_build_int_vec_type(blend_type);
 584
 585    arg_types[0] = screen->context_ptr_type;            /* context */
 586    arg_types[1] = LLVMInt32Type();                     /* x */
 587    arg_types[2] = LLVMInt32Type();                     /* y */
 588    arg_types[3] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
 589    arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
 590    arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* dady */
 591    arg_types[6] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0);  /* color */
 592    arg_types[7] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
 593    arg_types[8] = LLVMInt32Type();                     /* c0 */
 594    arg_types[9] = LLVMInt32Type();                    /* c1 */
 595    arg_types[10] = LLVMInt32Type();                    /* c2 */
 596    /* Note: the step arrays are built as int32[16] but we interpret
 597     * them here as int32_vec4[4].
 598     */
 599    arg_types[11] = LLVMPointerType(int32_vec4_type, 0);/* step0 */
 600    arg_types[12] = LLVMPointerType(int32_vec4_type, 0);/* step1 */
 601    arg_types[13] = LLVMPointerType(int32_vec4_type, 0);/* step2 */
 602
 603    func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
 604
 605    variant->function = LLVMAddFunction(screen->module, "shader", func_type);
 606    LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
 607
 608    /* XXX: need to propagate noalias down into color param now we are
 609     * passing a pointer-to-pointer?
 610     */
 611    for(i = 0; i < Elements(arg_types); ++i)
 612       if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
 613          LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute);
 614
 615    context_ptr  = LLVMGetParam(variant->function, 0);
 616    x            = LLVMGetParam(variant->function, 1);
 617    y            = LLVMGetParam(variant->function, 2);
 618    a0_ptr       = LLVMGetParam(variant->function, 3);
 619    dadx_ptr     = LLVMGetParam(variant->function, 4);
 620    dady_ptr     = LLVMGetParam(variant->function, 5);
 621    color_ptr_ptr = LLVMGetParam(variant->function, 6);
 622    depth_ptr    = LLVMGetParam(variant->function, 7);
 623    c0           = LLVMGetParam(variant->function, 8);
 624    c1           = LLVMGetParam(variant->function, 9);
 625    c2           = LLVMGetParam(variant->function, 10);
 626    step0_ptr    = LLVMGetParam(variant->function, 11);
 627    step1_ptr    = LLVMGetParam(variant->function, 12);
 628    step2_ptr    = LLVMGetParam(variant->function, 13);
 629
 630    lp_build_name(context_ptr, "context");
 631    lp_build_name(x, "x");
 632    lp_build_name(y, "y");
 633    lp_build_name(a0_ptr, "a0");
 634    lp_build_name(dadx_ptr, "dadx");
 635    lp_build_name(dady_ptr, "dady");
 636    lp_build_name(color_ptr_ptr, "color_ptr");
 637    lp_build_name(depth_ptr, "depth");
 638    lp_build_name(c0, "c0");
 639    lp_build_name(c1, "c1");
 640    lp_build_name(c2, "c2");
 641    lp_build_name(step0_ptr, "step0");
 642    lp_build_name(step1_ptr, "step1");
 643    lp_build_name(step2_ptr, "step2");
 644
 645    /*
 646     * Function body
 647     */
 648
 649    block = LLVMAppendBasicBlock(variant->function, "entry");
 650    builder = LLVMCreateBuilder();
 651    LLVMPositionBuilderAtEnd(builder, block);
 652
 653    generate_pos0(builder, x, y, &x0, &y0);
 654
 655    lp_build_interp_soa_init(&interp,
 656                             shader->base.tokens,
 657                             key->flatshade,
 658                             builder, fs_type,
 659                             a0_ptr, dadx_ptr, dady_ptr,
 660                             x0, y0);
 661
 662    /* code generated texture sampling */
 663    sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
 664
 665    /* loop over quads in the block */
 666    for(i = 0; i < num_fs; ++i) {
 667       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
 668       LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
 669       LLVMValueRef depth_ptr_i;
 670       int cbuf;
 671
 672       if(i != 0)
 673          lp_build_interp_soa_update(&interp, i);
 674
 675       depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
 676
 677       generate_fs(lp, shader, key,
 678                   builder,
 679                   fs_type,
 680                   context_ptr,
 681                   i,
 682                   &interp,
 683                   sampler,
 684                   &fs_mask[i], /* output */
 685                   out_color,
 686                   depth_ptr_i,
 687                   c0, c1, c2,
 688                   step0_ptr, step1_ptr, step2_ptr);
 689
 690       for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
 691          for(chan = 0; chan < NUM_CHANNELS; ++chan)
 692             fs_out_color[cbuf][chan][i] = out_color[cbuf][chan];
 693    }
 694
 695    sampler->destroy(sampler);
 696
 697    /* Loop over color outputs / color buffers to do blending.
 698     */
 699    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
 700       LLVMValueRef color_ptr;
 701       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0);
 702
 703       /*
 704        * Convert the fs's output color and mask to fit to the blending type.
 705        */
 706       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 707          lp_build_conv(builder, fs_type, blend_type,
 708                        fs_out_color[cbuf][chan], num_fs,
 709                        &blend_in_color[chan], 1);
 710          lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
 711       }
 712
 713       lp_build_conv_mask(builder, fs_type, blend_type,
 714                          fs_mask, num_fs,
 715                          &blend_mask, 1);
 716
 717       color_ptr = LLVMBuildLoad(builder,
 718                                 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
 719                                 "");
 720       lp_build_name(color_ptr, "color_ptr%d", cbuf);
 721
 722       /*
 723        * Blending.
 724        */
 725       generate_blend(&key->blend,
 726                      builder,
 727                      blend_type,
 728                      context_ptr,
 729                      blend_mask,
 730                      blend_in_color,
 731                      color_ptr);
 732    }
 733
 734    LLVMBuildRetVoid(builder);
 735
 736    LLVMDisposeBuilder(builder);
 737
 738
 739    /* Verify the LLVM IR.  If invalid, dump and abort */
 740 #ifdef DEBUG
 741    if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) {
 742       if (1)
 743          LLVMDumpValue(variant->function);
 744       abort();
 745    }
 746 #endif
 747
 748    /* Apply optimizations to LLVM IR */
 749    if (1)
 750       LLVMRunFunctionPassManager(screen->pass, variant->function);
 751
 752    if (LP_DEBUG & DEBUG_JIT) {
 753       /* Print the LLVM IR to stderr */
 754       LLVMDumpValue(variant->function);
 755       debug_printf("\n");
 756    }
 757
 758    /*
 759     * Translate the LLVM IR into machine code.
 760     */
 761    variant->jit_function = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, variant->function);
 762
 763    if (LP_DEBUG & DEBUG_ASM)
 764       lp_disassemble(variant->jit_function);
 765
 766    variant->next = shader->variants;
 767    shader->variants = variant;
 768 }
 769
 770
 771 static struct lp_fragment_shader_variant *
 772 generate_variant(struct llvmpipe_context *lp,
 773                  struct lp_fragment_shader *shader,
 774                  const struct lp_fragment_shader_variant_key *key)
 775 {
 776    struct lp_fragment_shader_variant *variant;
 777
 778    if (LP_DEBUG & DEBUG_JIT) {
 779       unsigned i;
 780
 781       tgsi_dump(shader->base.tokens, 0);
 782       if(key->depth.enabled) {
 783          debug_printf("depth.format = %s\n", pf_name(key->zsbuf_format));
 784          debug_printf("depth.func = %s\n", debug_dump_func(key->depth.func, TRUE));
 785          debug_printf("depth.writemask = %u\n", key->depth.writemask);
 786       }
 787       if(key->alpha.enabled) {
 788          debug_printf("alpha.func = %s\n", debug_dump_func(key->alpha.func, TRUE));
 789          debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
 790       }
 791       if(key->blend.logicop_enable) {
 792          debug_printf("blend.logicop_func = %u\n", key->blend.logicop_func);
 793       }
 794       else if(key->blend.blend_enable) {
 795          debug_printf("blend.rgb_func = %s\n",   debug_dump_blend_func  (key->blend.rgb_func, TRUE));
 796          debug_printf("rgb_src_factor = %s\n",   debug_dump_blend_factor(key->blend.rgb_src_factor, TRUE));
 797          debug_printf("rgb_dst_factor = %s\n",   debug_dump_blend_factor(key->blend.rgb_dst_factor, TRUE));
 798          debug_printf("alpha_func = %s\n",       debug_dump_blend_func  (key->blend.alpha_func, TRUE));
 799          debug_printf("alpha_src_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_src_factor, TRUE));
 800          debug_printf("alpha_dst_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_dst_factor, TRUE));
 801       }
 802       debug_printf("blend.colormask = 0x%x\n", key->blend.colormask);
 803       for(i = 0; i < PIPE_MAX_SAMPLERS; ++i) {
 804          if(key->sampler[i].format) {
 805             debug_printf("sampler[%u] = \n", i);
 806             debug_printf("  .format = %s\n",
 807                          pf_name(key->sampler[i].format));
 808             debug_printf("  .target = %s\n",
 809                          debug_dump_tex_target(key->sampler[i].target, TRUE));
 810             debug_printf("  .pot = %u %u %u\n",
 811                          key->sampler[i].pot_width,
 812                          key->sampler[i].pot_height,
 813                          key->sampler[i].pot_depth);
 814             debug_printf("  .wrap = %s %s %s\n",
 815                          debug_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
 816                          debug_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
 817                          debug_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
 818             debug_printf("  .min_img_filter = %s\n",
 819                          debug_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
 820             debug_printf("  .min_mip_filter = %s\n",
 821                          debug_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
 822             debug_printf("  .mag_img_filter = %s\n",
 823                          debug_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
 824             if(key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE)
 825                debug_printf("  .compare_func = %s\n", debug_dump_func(key->sampler[i].compare_func, TRUE));
 826             debug_printf("  .normalized_coords = %u\n", key->sampler[i].normalized_coords);
 827             debug_printf("  .prefilter = %u\n", key->sampler[i].prefilter);
 828          }
 829       }
 830    }
 831
 832    variant = CALLOC_STRUCT(lp_fragment_shader_variant);
 833    if(!variant)
 834       return NULL;
 835
 836    variant->shader = shader;
 837    memcpy(&variant->key, key, sizeof *key);
 838
 839    generate_fragment(lp, shader, variant);
 840
 841    return variant;
 842 }
 843
 844
 845 void *
 846 llvmpipe_create_fs_state(struct pipe_context *pipe,
 847                          const struct pipe_shader_state *templ)
 848 {
 849    struct lp_fragment_shader *shader;
 850
 851    shader = CALLOC_STRUCT(lp_fragment_shader);
 852    if (!shader)
 853       return NULL;
 854
 855    /* get/save the summary info for this shader */
 856    tgsi_scan_shader(templ->tokens, &shader->info);
 857
 858    /* we need to keep a local copy of the tokens */
 859    shader->base.tokens = tgsi_dup_tokens(templ->tokens);
 860
 861    return shader;
 862 }
 863
 864
 865 void
 866 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
 867 {
 868    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 869
 870    if (llvmpipe->fs == fs)
 871       return;
 872
 873    draw_flush(llvmpipe->draw);
 874
 875    llvmpipe->fs = fs;
 876
 877    llvmpipe->dirty |= LP_NEW_FS;
 878 }
 879
 880
 881 void
 882 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
 883 {
 884    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 885    struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
 886    struct lp_fragment_shader *shader = fs;
 887    struct lp_fragment_shader_variant *variant;
 888
 889    assert(fs != llvmpipe->fs);
 890    (void) llvmpipe;
 891
 892    variant = shader->variants;
 893    while(variant) {
 894       struct lp_fragment_shader_variant *next = variant->next;
 895
 896       if(variant->function) {
 897          if(variant->jit_function)
 898             LLVMFreeMachineCodeForFunction(screen->engine, variant->function);
 899          LLVMDeleteFunction(variant->function);
 900       }
 901
 902       FREE(variant);
 903
 904       variant = next;
 905    }
 906
 907    FREE((void *) shader->base.tokens);
 908    FREE(shader);
 909 }
 910
 911
 912
 913 void
 914 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
 915                              uint shader, uint index,
 916                              const struct pipe_constant_buffer *constants)
 917 {
 918    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 919    struct pipe_buffer *buffer = constants ? constants->buffer : NULL;
 920    unsigned size = buffer ? buffer->size : 0;
 921    const void *data = buffer ? llvmpipe_buffer(buffer)->data : NULL;
 922
 923    assert(shader < PIPE_SHADER_TYPES);
 924    assert(index == 0);
 925
 926    if(llvmpipe->constants[shader].buffer == buffer)
 927       return;
 928
 929    draw_flush(llvmpipe->draw);
 930
 931    /* note: reference counting */
 932    pipe_buffer_reference(&llvmpipe->constants[shader].buffer, buffer);
 933
 934    if(shader == PIPE_SHADER_VERTEX) {
 935       draw_set_mapped_constant_buffer(llvmpipe->draw, PIPE_SHADER_VERTEX,
 936                                       data, size);
 937    }
 938
 939    llvmpipe->dirty |= LP_NEW_CONSTANTS;
 940 }
 941
 942
 943 /**
 944  * We need to generate several variants of the fragment pipeline to match
 945  * all the combinations of the contributing state atoms.
 946  *
 947  * TODO: there is actually no reason to tie this to context state -- the
 948  * generated code could be cached globally in the screen.
 949  */
 950 static void
 951 make_variant_key(struct llvmpipe_context *lp,
 952                  struct lp_fragment_shader *shader,
 953                  struct lp_fragment_shader_variant_key *key)
 954 {
 955    unsigned i;
 956
 957    memset(key, 0, sizeof *key);
 958
 959    if(lp->framebuffer.zsbuf &&
 960       lp->depth_stencil->depth.enabled) {
 961       key->zsbuf_format = lp->framebuffer.zsbuf->format;
 962       memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
 963    }
 964
 965    key->alpha.enabled = lp->depth_stencil->alpha.enabled;
 966    if(key->alpha.enabled)
 967       key->alpha.func = lp->depth_stencil->alpha.func;
 968    /* alpha.ref_value is passed in jit_context */
 969
 970    key->flatshade = lp->rasterizer->flatshade;
 971
 972    if (lp->framebuffer.nr_cbufs) {
 973       memcpy(&key->blend, lp->blend, sizeof key->blend);
 974    }
 975
 976    key->nr_cbufs = lp->framebuffer.nr_cbufs;
 977    for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
 978       const struct util_format_description *format_desc;
 979       unsigned chan;
 980
 981       format_desc = util_format_description(lp->framebuffer.cbufs[i]->format);
 982       assert(format_desc->layout == UTIL_FORMAT_COLORSPACE_RGB ||
 983              format_desc->layout == UTIL_FORMAT_COLORSPACE_SRGB);
 984
 985       /* mask out color channels not present in the color buffer.
 986        * Should be simple to incorporate per-cbuf writemasks:
 987        */
 988       for(chan = 0; chan < 4; ++chan) {
 989          enum util_format_swizzle swizzle = format_desc->swizzle[chan];
 990
 991          if(swizzle <= UTIL_FORMAT_SWIZZLE_W)
 992             key->cbuf_blend[i].colormask |= (1 << chan);
 993       }
 994    }
 995
 996    for(i = 0; i < PIPE_MAX_SAMPLERS; ++i)
 997       if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i))
 998          lp_sampler_static_state(&key->sampler[i], lp->texture[i], lp->sampler[i]);
 999 }
1000
1001
1002 void
1003 llvmpipe_update_fs(struct llvmpipe_context *lp)
1004 {
1005    struct lp_fragment_shader *shader = lp->fs;
1006    struct lp_fragment_shader_variant_key key;
1007    struct lp_fragment_shader_variant *variant;
1008
1009    make_variant_key(lp, shader, &key);
1010
1011    variant = shader->variants;
1012    while(variant) {
1013       if(memcmp(&variant->key, &key, sizeof key) == 0)
1014          break;
1015
1016       variant = variant->next;
1017    }
1018
1019    if(!variant)
1020       variant = generate_variant(lp, shader, &key);
1021
1022    shader->current = variant;
1023
1024    lp_setup_set_fs_function(lp->setup,
1025                             shader->current->jit_function);
1026 }