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 static LLVMValueRef
 308 generate_scissor_test(LLVMBuilderRef builder,
 309                       LLVMValueRef context_ptr,
 310                       const struct lp_build_interp_soa_context *interp,
 311                       struct lp_type type)
 312 {
 313    LLVMTypeRef vec_type = lp_build_vec_type(type);
 314    LLVMValueRef xpos = interp->pos[0], ypos = interp->pos[1];
 315    LLVMValueRef xmin, ymin, xmax, ymax;
 316    LLVMValueRef m0, m1, m2, m3, m;
 317
 318    /* xpos, ypos contain the window coords for the four pixels in the quad */
 319    assert(xpos);
 320    assert(ypos);
 321
 322    /* get the current scissor bounds, convert to vectors */
 323    xmin = lp_jit_context_scissor_xmin_value(builder, context_ptr);
 324    xmin = lp_build_broadcast(builder, vec_type, xmin);
 325
 326    ymin = lp_jit_context_scissor_ymin_value(builder, context_ptr);
 327    ymin = lp_build_broadcast(builder, vec_type, ymin);
 328
 329    xmax = lp_jit_context_scissor_xmax_value(builder, context_ptr);
 330    xmax = lp_build_broadcast(builder, vec_type, xmax);
 331
 332    ymax = lp_jit_context_scissor_ymax_value(builder, context_ptr);
 333    ymax = lp_build_broadcast(builder, vec_type, ymax);
 334
 335    /* compare the fragment's position coordinates against the scissor bounds */
 336    m0 = lp_build_compare(builder, type, PIPE_FUNC_GEQUAL, xpos, xmin);
 337    m1 = lp_build_compare(builder, type, PIPE_FUNC_GEQUAL, ypos, ymin);
 338    m2 = lp_build_compare(builder, type, PIPE_FUNC_LESS, xpos, xmax);
 339    m3 = lp_build_compare(builder, type, PIPE_FUNC_LESS, ypos, ymax);
 340
 341    /* AND all the masks together */
 342    m = LLVMBuildAnd(builder, m0, m1, "");
 343    m = LLVMBuildAnd(builder, m, m2, "");
 344    m = LLVMBuildAnd(builder, m, m3, "");
 345
 346    lp_build_name(m, "scissormask");
 347
 348    return m;
 349 }
 350
 351
 352 /**
 353  * Generate the fragment shader, depth/stencil test, and alpha tests.
 354  * \param i  which quad in the tile, in range [0,3]
 355  */
 356 static void
 357 generate_fs(struct llvmpipe_context *lp,
 358             struct lp_fragment_shader *shader,
 359             const struct lp_fragment_shader_variant_key *key,
 360             LLVMBuilderRef builder,
 361             struct lp_type type,
 362             LLVMValueRef context_ptr,
 363             unsigned i,
 364             const struct lp_build_interp_soa_context *interp,
 365             struct lp_build_sampler_soa *sampler,
 366             LLVMValueRef *pmask,
 367             LLVMValueRef (*color)[4],
 368             LLVMValueRef depth_ptr,
 369             LLVMValueRef c0,
 370             LLVMValueRef c1,
 371             LLVMValueRef c2,
 372             LLVMValueRef step0_ptr,
 373             LLVMValueRef step1_ptr,
 374             LLVMValueRef step2_ptr)
 375 {
 376    const struct tgsi_token *tokens = shader->base.tokens;
 377    LLVMTypeRef elem_type;
 378    LLVMTypeRef vec_type;
 379    LLVMTypeRef int_vec_type;
 380    LLVMValueRef consts_ptr;
 381    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
 382    LLVMValueRef z = interp->pos[2];
 383    struct lp_build_flow_context *flow;
 384    struct lp_build_mask_context mask;
 385    boolean early_depth_test;
 386    unsigned attrib;
 387    unsigned chan;
 388    unsigned cbuf;
 389
 390    assert(i < 4);
 391
 392    elem_type = lp_build_elem_type(type);
 393    vec_type = lp_build_vec_type(type);
 394    int_vec_type = lp_build_int_vec_type(type);
 395
 396    consts_ptr = lp_jit_context_constants(builder, context_ptr);
 397
 398    flow = lp_build_flow_create(builder);
 399
 400    memset(outputs, 0, sizeof outputs);
 401
 402    lp_build_flow_scope_begin(flow);
 403
 404    /* Declare the color and z variables */
 405    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
 406       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 407          color[cbuf][chan] = LLVMGetUndef(vec_type);
 408          lp_build_flow_scope_declare(flow, &color[cbuf][chan]);
 409       }
 410    }
 411    lp_build_flow_scope_declare(flow, &z);
 412
 413    /* do triangle edge testing */
 414    generate_tri_edge_mask(builder, i, pmask,
 415                           c0, c1, c2, step0_ptr, step1_ptr, step2_ptr);
 416
 417    /* 'mask' will control execution based on quad's pixel alive/killed state */
 418    lp_build_mask_begin(&mask, flow, type, *pmask);
 419
 420    if (key->scissor) {
 421       LLVMValueRef smask =
 422          generate_scissor_test(builder, context_ptr, interp, type);
 423       lp_build_mask_update(&mask, smask);
 424    }
 425
 426    early_depth_test =
 427       key->depth.enabled &&
 428       !key->alpha.enabled &&
 429       !shader->info.uses_kill &&
 430       !shader->info.writes_z;
 431
 432    if(early_depth_test)
 433       generate_depth(builder, key,
 434                      type, &mask,
 435                      z, depth_ptr);
 436
 437    lp_build_tgsi_soa(builder, tokens, type, &mask,
 438                      consts_ptr, interp->pos, interp->inputs,
 439                      outputs, sampler);
 440
 441    for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
 442       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 443          if(outputs[attrib][chan]) {
 444             lp_build_name(outputs[attrib][chan], "output%u.%u.%c", i, attrib, "xyzw"[chan]);
 445
 446             switch (shader->info.output_semantic_name[attrib]) {
 447             case TGSI_SEMANTIC_COLOR:
 448                {
 449                   unsigned cbuf = shader->info.output_semantic_index[attrib];
 450
 451                   lp_build_name(outputs[attrib][chan], "color%u.%u.%c", i, attrib, "rgba"[chan]);
 452
 453                   /* Alpha test */
 454                   /* XXX: should the alpha reference value be passed separately? */
 455                   /* XXX: should only test the final assignment to alpha */
 456                   if(cbuf == 0 && chan == 3) {
 457                      LLVMValueRef alpha = outputs[attrib][chan];
 458                      LLVMValueRef alpha_ref_value;
 459                      alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
 460                      alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
 461                      lp_build_alpha_test(builder, &key->alpha, type,
 462                                          &mask, alpha, alpha_ref_value);
 463                   }
 464
 465                   color[cbuf][chan] = outputs[attrib][chan];
 466                   break;
 467                }
 468
 469             case TGSI_SEMANTIC_POSITION:
 470                if(chan == 2)
 471                   z = outputs[attrib][chan];
 472                break;
 473             }
 474          }
 475       }
 476    }
 477
 478    if(!early_depth_test)
 479       generate_depth(builder, key,
 480                      type, &mask,
 481                      z, depth_ptr);
 482
 483    lp_build_mask_end(&mask);
 484
 485    lp_build_flow_scope_end(flow);
 486
 487    lp_build_flow_destroy(flow);
 488
 489    *pmask = mask.value;
 490
 491 }
 492
 493
 494 /**
 495  * Generate color blending and color output.
 496  */
 497 static void
 498 generate_blend(const struct pipe_blend_state *blend,
 499                LLVMBuilderRef builder,
 500                struct lp_type type,
 501                LLVMValueRef context_ptr,
 502                LLVMValueRef mask,
 503                LLVMValueRef *src,
 504                LLVMValueRef dst_ptr)
 505 {
 506    struct lp_build_context bld;
 507    struct lp_build_flow_context *flow;
 508    struct lp_build_mask_context mask_ctx;
 509    LLVMTypeRef vec_type;
 510    LLVMTypeRef int_vec_type;
 511    LLVMValueRef const_ptr;
 512    LLVMValueRef con[4];
 513    LLVMValueRef dst[4];
 514    LLVMValueRef res[4];
 515    unsigned chan;
 516
 517    lp_build_context_init(&bld, builder, type);
 518
 519    flow = lp_build_flow_create(builder);
 520
 521    /* we'll use this mask context to skip blending if all pixels are dead */
 522    lp_build_mask_begin(&mask_ctx, flow, type, mask);
 523
 524    vec_type = lp_build_vec_type(type);
 525    int_vec_type = lp_build_int_vec_type(type);
 526
 527    const_ptr = lp_jit_context_blend_color(builder, context_ptr);
 528    const_ptr = LLVMBuildBitCast(builder, const_ptr,
 529                                 LLVMPointerType(vec_type, 0), "");
 530
 531    for(chan = 0; chan < 4; ++chan) {
 532       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 533       con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
 534
 535       dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
 536
 537       lp_build_name(con[chan], "con.%c", "rgba"[chan]);
 538       lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
 539    }
 540
 541    lp_build_blend_soa(builder, blend, type, src, dst, con, res);
 542
 543    for(chan = 0; chan < 4; ++chan) {
 544       if(blend->colormask & (1 << chan)) {
 545          LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 546          lp_build_name(res[chan], "res.%c", "rgba"[chan]);
 547          res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
 548          LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
 549       }
 550    }
 551
 552    lp_build_mask_end(&mask_ctx);
 553    lp_build_flow_destroy(flow);
 554 }
 555
 556
 557 /**
 558  * Generate the runtime callable function for the whole fragment pipeline.
 559  * Note that the function which we generate operates on a block of 16
 560  * pixels at at time.  The block contains 2x2 quads.  Each quad contains
 561  * 2x2 pixels.
 562  */
 563 static void
 564 generate_fragment(struct llvmpipe_context *lp,
 565                   struct lp_fragment_shader *shader,
 566                   struct lp_fragment_shader_variant *variant)
 567 {
 568    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
 569    const struct lp_fragment_shader_variant_key *key = &variant->key;
 570    struct lp_type fs_type;
 571    struct lp_type blend_type;
 572    LLVMTypeRef fs_elem_type;
 573    LLVMTypeRef fs_vec_type;
 574    LLVMTypeRef fs_int_vec_type;
 575    LLVMTypeRef blend_vec_type;
 576    LLVMTypeRef blend_int_vec_type;
 577    LLVMTypeRef arg_types[14];
 578    LLVMTypeRef func_type;
 579    LLVMTypeRef int32_vec4_type = lp_build_int32_vec4_type();
 580    LLVMValueRef context_ptr;
 581    LLVMValueRef x;
 582    LLVMValueRef y;
 583    LLVMValueRef a0_ptr;
 584    LLVMValueRef dadx_ptr;
 585    LLVMValueRef dady_ptr;
 586    LLVMValueRef color_ptr_ptr;
 587    LLVMValueRef depth_ptr;
 588    LLVMValueRef c0, c1, c2, step0_ptr, step1_ptr, step2_ptr;
 589    LLVMBasicBlockRef block;
 590    LLVMBuilderRef builder;
 591    LLVMValueRef x0;
 592    LLVMValueRef y0;
 593    struct lp_build_sampler_soa *sampler;
 594    struct lp_build_interp_soa_context interp;
 595    LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
 596    LLVMValueRef fs_out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
 597    LLVMValueRef blend_mask;
 598    LLVMValueRef blend_in_color[NUM_CHANNELS];
 599    unsigned num_fs;
 600    unsigned i;
 601    unsigned chan;
 602    unsigned cbuf;
 603
 604
 605    /* TODO: actually pick these based on the fs and color buffer
 606     * characteristics. */
 607
 608    memset(&fs_type, 0, sizeof fs_type);
 609    fs_type.floating = TRUE; /* floating point values */
 610    fs_type.sign = TRUE;     /* values are signed */
 611    fs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
 612    fs_type.width = 32;      /* 32-bit float */
 613    fs_type.length = 4;      /* 4 elements per vector */
 614    num_fs = 4;              /* number of quads per block */
 615
 616    memset(&blend_type, 0, sizeof blend_type);
 617    blend_type.floating = FALSE; /* values are integers */
 618    blend_type.sign = FALSE;     /* values are unsigned */
 619    blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
 620    blend_type.width = 8;        /* 8-bit ubyte values */
 621    blend_type.length = 16;      /* 16 elements per vector */
 622
 623    /*
 624     * Generate the function prototype. Any change here must be reflected in
 625     * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
 626     */
 627
 628    fs_elem_type = lp_build_elem_type(fs_type);
 629    fs_vec_type = lp_build_vec_type(fs_type);
 630    fs_int_vec_type = lp_build_int_vec_type(fs_type);
 631
 632    blend_vec_type = lp_build_vec_type(blend_type);
 633    blend_int_vec_type = lp_build_int_vec_type(blend_type);
 634
 635    arg_types[0] = screen->context_ptr_type;            /* context */
 636    arg_types[1] = LLVMInt32Type();                     /* x */
 637    arg_types[2] = LLVMInt32Type();                     /* y */
 638    arg_types[3] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
 639    arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
 640    arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* dady */
 641    arg_types[6] = LLVMPointerType(LLVMPointerType(blend_vec_type, 0), 0);  /* color */
 642    arg_types[7] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
 643    arg_types[8] = LLVMInt32Type();                     /* c0 */
 644    arg_types[9] = LLVMInt32Type();                    /* c1 */
 645    arg_types[10] = LLVMInt32Type();                    /* c2 */
 646    /* Note: the step arrays are built as int32[16] but we interpret
 647     * them here as int32_vec4[4].
 648     */
 649    arg_types[11] = LLVMPointerType(int32_vec4_type, 0);/* step0 */
 650    arg_types[12] = LLVMPointerType(int32_vec4_type, 0);/* step1 */
 651    arg_types[13] = LLVMPointerType(int32_vec4_type, 0);/* step2 */
 652
 653    func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
 654
 655    variant->function = LLVMAddFunction(screen->module, "shader", func_type);
 656    LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
 657
 658    /* XXX: need to propagate noalias down into color param now we are
 659     * passing a pointer-to-pointer?
 660     */
 661    for(i = 0; i < Elements(arg_types); ++i)
 662       if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
 663          LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute);
 664
 665    context_ptr  = LLVMGetParam(variant->function, 0);
 666    x            = LLVMGetParam(variant->function, 1);
 667    y            = LLVMGetParam(variant->function, 2);
 668    a0_ptr       = LLVMGetParam(variant->function, 3);
 669    dadx_ptr     = LLVMGetParam(variant->function, 4);
 670    dady_ptr     = LLVMGetParam(variant->function, 5);
 671    color_ptr_ptr = LLVMGetParam(variant->function, 6);
 672    depth_ptr    = LLVMGetParam(variant->function, 7);
 673    c0           = LLVMGetParam(variant->function, 8);
 674    c1           = LLVMGetParam(variant->function, 9);
 675    c2           = LLVMGetParam(variant->function, 10);
 676    step0_ptr    = LLVMGetParam(variant->function, 11);
 677    step1_ptr    = LLVMGetParam(variant->function, 12);
 678    step2_ptr    = LLVMGetParam(variant->function, 13);
 679
 680    lp_build_name(context_ptr, "context");
 681    lp_build_name(x, "x");
 682    lp_build_name(y, "y");
 683    lp_build_name(a0_ptr, "a0");
 684    lp_build_name(dadx_ptr, "dadx");
 685    lp_build_name(dady_ptr, "dady");
 686    lp_build_name(color_ptr_ptr, "color_ptr");
 687    lp_build_name(depth_ptr, "depth");
 688    lp_build_name(c0, "c0");
 689    lp_build_name(c1, "c1");
 690    lp_build_name(c2, "c2");
 691    lp_build_name(step0_ptr, "step0");
 692    lp_build_name(step1_ptr, "step1");
 693    lp_build_name(step2_ptr, "step2");
 694
 695    /*
 696     * Function body
 697     */
 698
 699    block = LLVMAppendBasicBlock(variant->function, "entry");
 700    builder = LLVMCreateBuilder();
 701    LLVMPositionBuilderAtEnd(builder, block);
 702
 703    generate_pos0(builder, x, y, &x0, &y0);
 704
 705    lp_build_interp_soa_init(&interp,
 706                             shader->base.tokens,
 707                             key->flatshade,
 708                             builder, fs_type,
 709                             a0_ptr, dadx_ptr, dady_ptr,
 710                             x0, y0);
 711
 712    /* code generated texture sampling */
 713    sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
 714
 715    /* loop over quads in the block */
 716    for(i = 0; i < num_fs; ++i) {
 717       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
 718       LLVMValueRef out_color[PIPE_MAX_COLOR_BUFS][NUM_CHANNELS];
 719       LLVMValueRef depth_ptr_i;
 720       int cbuf;
 721
 722       if(i != 0)
 723          lp_build_interp_soa_update(&interp, i);
 724
 725       depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
 726
 727       generate_fs(lp, shader, key,
 728                   builder,
 729                   fs_type,
 730                   context_ptr,
 731                   i,
 732                   &interp,
 733                   sampler,
 734                   &fs_mask[i], /* output */
 735                   out_color,
 736                   depth_ptr_i,
 737                   c0, c1, c2,
 738                   step0_ptr, step1_ptr, step2_ptr);
 739
 740       for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++)
 741          for(chan = 0; chan < NUM_CHANNELS; ++chan)
 742             fs_out_color[cbuf][chan][i] = out_color[cbuf][chan];
 743    }
 744
 745    sampler->destroy(sampler);
 746
 747    /* Loop over color outputs / color buffers to do blending.
 748     */
 749    for(cbuf = 0; cbuf < key->nr_cbufs; cbuf++) {
 750       LLVMValueRef color_ptr;
 751       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), cbuf, 0);
 752
 753       /*
 754        * Convert the fs's output color and mask to fit to the blending type.
 755        */
 756       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 757          lp_build_conv(builder, fs_type, blend_type,
 758                        fs_out_color[cbuf][chan], num_fs,
 759                        &blend_in_color[chan], 1);
 760          lp_build_name(blend_in_color[chan], "color%d.%c", cbuf, "rgba"[chan]);
 761       }
 762
 763       lp_build_conv_mask(builder, fs_type, blend_type,
 764                          fs_mask, num_fs,
 765                          &blend_mask, 1);
 766
 767       color_ptr = LLVMBuildLoad(builder,
 768                                 LLVMBuildGEP(builder, color_ptr_ptr, &index, 1, ""),
 769                                 "");
 770       lp_build_name(color_ptr, "color_ptr%d", cbuf);
 771
 772       /*
 773        * Blending.
 774        */
 775       generate_blend(&key->blend,
 776                      builder,
 777                      blend_type,
 778                      context_ptr,
 779                      blend_mask,
 780                      blend_in_color,
 781                      color_ptr);
 782    }
 783
 784    LLVMBuildRetVoid(builder);
 785
 786    LLVMDisposeBuilder(builder);
 787
 788
 789    /* Verify the LLVM IR.  If invalid, dump and abort */
 790 #ifdef DEBUG
 791    if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) {
 792       if (1)
 793          LLVMDumpValue(variant->function);
 794       abort();
 795    }
 796 #endif
 797
 798    /* Apply optimizations to LLVM IR */
 799    if (1)
 800       LLVMRunFunctionPassManager(screen->pass, variant->function);
 801
 802    if (LP_DEBUG & DEBUG_JIT) {
 803       /* Print the LLVM IR to stderr */
 804       LLVMDumpValue(variant->function);
 805       debug_printf("\n");
 806    }
 807
 808    /*
 809     * Translate the LLVM IR into machine code.
 810     */
 811    variant->jit_function = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, variant->function);
 812
 813    if (LP_DEBUG & DEBUG_ASM)
 814       lp_disassemble(variant->jit_function);
 815
 816    variant->next = shader->variants;
 817    shader->variants = variant;
 818 }
 819
 820
 821 static struct lp_fragment_shader_variant *
 822 generate_variant(struct llvmpipe_context *lp,
 823                  struct lp_fragment_shader *shader,
 824                  const struct lp_fragment_shader_variant_key *key)
 825 {
 826    struct lp_fragment_shader_variant *variant;
 827
 828    if (LP_DEBUG & DEBUG_JIT) {
 829       unsigned i;
 830
 831       tgsi_dump(shader->base.tokens, 0);
 832       if(key->depth.enabled) {
 833          debug_printf("depth.format = %s\n", pf_name(key->zsbuf_format));
 834          debug_printf("depth.func = %s\n", debug_dump_func(key->depth.func, TRUE));
 835          debug_printf("depth.writemask = %u\n", key->depth.writemask);
 836       }
 837       if(key->alpha.enabled) {
 838          debug_printf("alpha.func = %s\n", debug_dump_func(key->alpha.func, TRUE));
 839          debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
 840       }
 841       if(key->blend.logicop_enable) {
 842          debug_printf("blend.logicop_func = %u\n", key->blend.logicop_func);
 843       }
 844       else if(key->blend.blend_enable) {
 845          debug_printf("blend.rgb_func = %s\n",   debug_dump_blend_func  (key->blend.rgb_func, TRUE));
 846          debug_printf("rgb_src_factor = %s\n",   debug_dump_blend_factor(key->blend.rgb_src_factor, TRUE));
 847          debug_printf("rgb_dst_factor = %s\n",   debug_dump_blend_factor(key->blend.rgb_dst_factor, TRUE));
 848          debug_printf("alpha_func = %s\n",       debug_dump_blend_func  (key->blend.alpha_func, TRUE));
 849          debug_printf("alpha_src_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_src_factor, TRUE));
 850          debug_printf("alpha_dst_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_dst_factor, TRUE));
 851       }
 852       debug_printf("blend.colormask = 0x%x\n", key->blend.colormask);
 853       for(i = 0; i < PIPE_MAX_SAMPLERS; ++i) {
 854          if(key->sampler[i].format) {
 855             debug_printf("sampler[%u] = \n", i);
 856             debug_printf("  .format = %s\n",
 857                          pf_name(key->sampler[i].format));
 858             debug_printf("  .target = %s\n",
 859                          debug_dump_tex_target(key->sampler[i].target, TRUE));
 860             debug_printf("  .pot = %u %u %u\n",
 861                          key->sampler[i].pot_width,
 862                          key->sampler[i].pot_height,
 863                          key->sampler[i].pot_depth);
 864             debug_printf("  .wrap = %s %s %s\n",
 865                          debug_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
 866                          debug_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
 867                          debug_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
 868             debug_printf("  .min_img_filter = %s\n",
 869                          debug_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
 870             debug_printf("  .min_mip_filter = %s\n",
 871                          debug_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
 872             debug_printf("  .mag_img_filter = %s\n",
 873                          debug_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
 874             if(key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE)
 875                debug_printf("  .compare_func = %s\n", debug_dump_func(key->sampler[i].compare_func, TRUE));
 876             debug_printf("  .normalized_coords = %u\n", key->sampler[i].normalized_coords);
 877             debug_printf("  .prefilter = %u\n", key->sampler[i].prefilter);
 878          }
 879       }
 880    }
 881
 882    variant = CALLOC_STRUCT(lp_fragment_shader_variant);
 883    if(!variant)
 884       return NULL;
 885
 886    variant->shader = shader;
 887    memcpy(&variant->key, key, sizeof *key);
 888
 889    generate_fragment(lp, shader, variant);
 890
 891    return variant;
 892 }
 893
 894
 895 void *
 896 llvmpipe_create_fs_state(struct pipe_context *pipe,
 897                          const struct pipe_shader_state *templ)
 898 {
 899    struct lp_fragment_shader *shader;
 900
 901    shader = CALLOC_STRUCT(lp_fragment_shader);
 902    if (!shader)
 903       return NULL;
 904
 905    /* get/save the summary info for this shader */
 906    tgsi_scan_shader(templ->tokens, &shader->info);
 907
 908    /* we need to keep a local copy of the tokens */
 909    shader->base.tokens = tgsi_dup_tokens(templ->tokens);
 910
 911    return shader;
 912 }
 913
 914
 915 void
 916 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
 917 {
 918    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 919
 920    if (llvmpipe->fs == fs)
 921       return;
 922
 923    draw_flush(llvmpipe->draw);
 924
 925    llvmpipe->fs = fs;
 926
 927    llvmpipe->dirty |= LP_NEW_FS;
 928 }
 929
 930
 931 void
 932 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
 933 {
 934    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 935    struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
 936    struct lp_fragment_shader *shader = fs;
 937    struct lp_fragment_shader_variant *variant;
 938
 939    assert(fs != llvmpipe->fs);
 940    (void) llvmpipe;
 941
 942    variant = shader->variants;
 943    while(variant) {
 944       struct lp_fragment_shader_variant *next = variant->next;
 945
 946       if(variant->function) {
 947          if(variant->jit_function)
 948             LLVMFreeMachineCodeForFunction(screen->engine, variant->function);
 949          LLVMDeleteFunction(variant->function);
 950       }
 951
 952       FREE(variant);
 953
 954       variant = next;
 955    }
 956
 957    FREE((void *) shader->base.tokens);
 958    FREE(shader);
 959 }
 960
 961
 962
 963 void
 964 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
 965                              uint shader, uint index,
 966                              const struct pipe_constant_buffer *constants)
 967 {
 968    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 969    struct pipe_buffer *buffer = constants ? constants->buffer : NULL;
 970    unsigned size = buffer ? buffer->size : 0;
 971    const void *data = buffer ? llvmpipe_buffer(buffer)->data : NULL;
 972
 973    assert(shader < PIPE_SHADER_TYPES);
 974    assert(index == 0);
 975
 976    if(llvmpipe->constants[shader].buffer == buffer)
 977       return;
 978
 979    draw_flush(llvmpipe->draw);
 980
 981    /* note: reference counting */
 982    pipe_buffer_reference(&llvmpipe->constants[shader].buffer, buffer);
 983
 984    if(shader == PIPE_SHADER_VERTEX) {
 985       draw_set_mapped_constant_buffer(llvmpipe->draw, PIPE_SHADER_VERTEX,
 986                                       data, size);
 987    }
 988
 989    llvmpipe->dirty |= LP_NEW_CONSTANTS;
 990 }
 991
 992
 993 /**
 994  * We need to generate several variants of the fragment pipeline to match
 995  * all the combinations of the contributing state atoms.
 996  *
 997  * TODO: there is actually no reason to tie this to context state -- the
 998  * generated code could be cached globally in the screen.
 999  */
1000 static void
1001 make_variant_key(struct llvmpipe_context *lp,
1002                  struct lp_fragment_shader *shader,
1003                  struct lp_fragment_shader_variant_key *key)
1004 {
1005    unsigned i;
1006
1007    memset(key, 0, sizeof *key);
1008
1009    if(lp->framebuffer.zsbuf &&
1010       lp->depth_stencil->depth.enabled) {
1011       key->zsbuf_format = lp->framebuffer.zsbuf->format;
1012       memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
1013    }
1014
1015    key->alpha.enabled = lp->depth_stencil->alpha.enabled;
1016    if(key->alpha.enabled)
1017       key->alpha.func = lp->depth_stencil->alpha.func;
1018    /* alpha.ref_value is passed in jit_context */
1019
1020    key->flatshade = lp->rasterizer->flatshade;
1021    key->scissor = lp->rasterizer->scissor;
1022
1023    if (lp->framebuffer.nr_cbufs) {
1024       memcpy(&key->blend, lp->blend, sizeof key->blend);
1025    }
1026
1027    key->nr_cbufs = lp->framebuffer.nr_cbufs;
1028    for (i = 0; i < lp->framebuffer.nr_cbufs; i++) {
1029       const struct util_format_description *format_desc;
1030       unsigned chan;
1031
1032       format_desc = util_format_description(lp->framebuffer.cbufs[i]->format);
1033       assert(format_desc->layout == UTIL_FORMAT_COLORSPACE_RGB ||
1034              format_desc->layout == UTIL_FORMAT_COLORSPACE_SRGB);
1035
1036       /* mask out color channels not present in the color buffer.
1037        * Should be simple to incorporate per-cbuf writemasks:
1038        */
1039       for(chan = 0; chan < 4; ++chan) {
1040          enum util_format_swizzle swizzle = format_desc->swizzle[chan];
1041
1042          if(swizzle <= UTIL_FORMAT_SWIZZLE_W)
1043             key->cbuf_blend[i].colormask |= (1 << chan);
1044       }
1045    }
1046
1047    for(i = 0; i < PIPE_MAX_SAMPLERS; ++i)
1048       if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i))
1049          lp_sampler_static_state(&key->sampler[i], lp->texture[i], lp->sampler[i]);
1050 }
1051
1052
1053 /**
1054  * Update fragment state.  This is called just prior to drawing
1055  * something when some fragment-related state has changed.
1056  */
1057 void
1058 llvmpipe_update_fs(struct llvmpipe_context *lp)
1059 {
1060    struct lp_fragment_shader *shader = lp->fs;
1061    struct lp_fragment_shader_variant_key key;
1062    struct lp_fragment_shader_variant *variant;
1063    boolean opaque;
1064
1065    make_variant_key(lp, shader, &key);
1066
1067    variant = shader->variants;
1068    while(variant) {
1069       if(memcmp(&variant->key, &key, sizeof key) == 0)
1070          break;
1071
1072       variant = variant->next;
1073    }
1074
1075    if(!variant)
1076       variant = generate_variant(lp, shader, &key);
1077
1078    shader->current = variant;
1079
1080    /* TODO: put this in the variant */
1081    /* TODO: most of these can be relaxed, in particular the colormask */
1082    opaque = !key.blend.logicop_enable &&
1083             !key.blend.blend_enable &&
1084             key.blend.colormask == 0xf &&
1085             !key.alpha.enabled &&
1086             !key.depth.enabled &&
1087             !key.scissor &&
1088             !shader->info.uses_kill
1089             ? TRUE : FALSE;
1090
1091    lp_setup_set_fs_function(lp->setup,
1092                             shader->current->jit_function,
1093                             opaque);
1094 }