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 "pipe/p_defines.h"
  62 #include "util/u_memory.h"
  63 #include "util/u_format.h"
  64 #include "util/u_debug_dump.h"
  65 #include "pipe/internal/p_winsys_screen.h"
  66 #include "pipe/p_shader_tokens.h"
  67 #include "draw/draw_context.h"
  68 #include "tgsi/tgsi_dump.h"
  69 #include "tgsi/tgsi_scan.h"
  70 #include "tgsi/tgsi_parse.h"
  71 #include "lp_bld_type.h"
  72 #include "lp_bld_const.h"
  73 #include "lp_bld_conv.h"
  74 #include "lp_bld_intr.h"
  75 #include "lp_bld_logic.h"
  76 #include "lp_bld_depth.h"
  77 #include "lp_bld_interp.h"
  78 #include "lp_bld_tgsi.h"
  79 #include "lp_bld_alpha.h"
  80 #include "lp_bld_blend.h"
  81 #include "lp_bld_swizzle.h"
  82 #include "lp_bld_flow.h"
  83 #include "lp_bld_debug.h"
  84 #include "lp_screen.h"
  85 #include "lp_context.h"
  86 #include "lp_buffer.h"
  87 #include "lp_setup.h"
  88 #include "lp_state.h"
  89 #include "lp_tex_sample.h"
  90 #include "lp_debug.h"
  91
  92
  93 static const unsigned char quad_offset_x[4] = {0, 1, 0, 1};
  94 static const unsigned char quad_offset_y[4] = {0, 0, 1, 1};
  95
  96
  97 /*
  98  * Derive from the quad's upper left scalar coordinates the coordinates for
  99  * all other quad pixels
 100  */
 101 static void
 102 generate_pos0(LLVMBuilderRef builder,
 103               LLVMValueRef x,
 104               LLVMValueRef y,
 105               LLVMValueRef *x0,
 106               LLVMValueRef *y0)
 107 {
 108    LLVMTypeRef int_elem_type = LLVMInt32Type();
 109    LLVMTypeRef int_vec_type = LLVMVectorType(int_elem_type, QUAD_SIZE);
 110    LLVMTypeRef elem_type = LLVMFloatType();
 111    LLVMTypeRef vec_type = LLVMVectorType(elem_type, QUAD_SIZE);
 112    LLVMValueRef x_offsets[QUAD_SIZE];
 113    LLVMValueRef y_offsets[QUAD_SIZE];
 114    unsigned i;
 115
 116    x = lp_build_broadcast(builder, int_vec_type, x);
 117    y = lp_build_broadcast(builder, int_vec_type, y);
 118
 119    for(i = 0; i < QUAD_SIZE; ++i) {
 120       x_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_x[i], 0);
 121       y_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_y[i], 0);
 122    }
 123
 124    x = LLVMBuildAdd(builder, x, LLVMConstVector(x_offsets, QUAD_SIZE), "");
 125    y = LLVMBuildAdd(builder, y, LLVMConstVector(y_offsets, QUAD_SIZE), "");
 126
 127    *x0 = LLVMBuildSIToFP(builder, x, vec_type, "");
 128    *y0 = LLVMBuildSIToFP(builder, y, vec_type, "");
 129 }
 130
 131
 132 /**
 133  * Generate the depth test.
 134  */
 135 static void
 136 generate_depth(LLVMBuilderRef builder,
 137                const struct lp_fragment_shader_variant_key *key,
 138                struct lp_type src_type,
 139                struct lp_build_mask_context *mask,
 140                LLVMValueRef src,
 141                LLVMValueRef dst_ptr)
 142 {
 143    const struct util_format_description *format_desc;
 144    struct lp_type dst_type;
 145
 146    if(!key->depth.enabled)
 147       return;
 148
 149    format_desc = util_format_description(key->zsbuf_format);
 150    assert(format_desc);
 151
 152    /* Pick the depth type. */
 153    dst_type = lp_depth_type(format_desc, src_type.width*src_type.length);
 154
 155    /* FIXME: Cope with a depth test type with a different bit width. */
 156    assert(dst_type.width == src_type.width);
 157    assert(dst_type.length == src_type.length);
 158
 159 #if 1
 160    src = lp_build_clamped_float_to_unsigned_norm(builder,
 161                                                  src_type,
 162                                                  dst_type.width,
 163                                                  src);
 164 #else
 165    lp_build_conv(builder, src_type, dst_type, &src, 1, &src, 1);
 166 #endif
 167
 168    lp_build_depth_test(builder,
 169                        &key->depth,
 170                        dst_type,
 171                        format_desc,
 172                        mask,
 173                        src,
 174                        dst_ptr);
 175 }
 176
 177
 178 /**
 179  * Generate the code to do inside/outside triangle testing for the
 180  * four pixels in a 2x2 quad.  This will set the four elements of the
 181  * quad mask vector to 0 or ~0.
 182  * \param i  which quad of the quad group to test, in [0,3]
 183  */
 184 static void
 185 generate_tri_edge_mask(LLVMBuilderRef builder,
 186                        unsigned i,
 187                        LLVMValueRef *mask,      /* ivec4, out */
 188                        LLVMValueRef c0,         /* int32 */
 189                        LLVMValueRef c1,         /* int32 */
 190                        LLVMValueRef c2,         /* int32 */
 191                        LLVMValueRef step0_ptr,  /* ivec4 */
 192                        LLVMValueRef step1_ptr,  /* ivec4 */
 193                        LLVMValueRef step2_ptr)  /* ivec4 */
 194 {
 195    /*
 196      c0_vec = splat(c0)
 197      c1_vec = splat(c1)
 198      c2_vec = splat(c2)
 199      m0_vec = step0_ptr[i] > c0_vec
 200      m1_vec = step1_ptr[i] > c1_vec
 201      m2_vec = step2_ptr[i] > c2_vec
 202      mask = m0_vec & m1_vec & m2_vec
 203     */
 204    struct lp_type i32_type;
 205    LLVMTypeRef i32vec4_type;
 206
 207    LLVMValueRef index;
 208    LLVMValueRef c0_vec, c1_vec, c2_vec;
 209    LLVMValueRef step0_vec, step1_vec, step2_vec;
 210    LLVMValueRef m0_vec, m1_vec, m2_vec;
 211    LLVMValueRef m;
 212
 213    assert(i < 4);
 214
 215    /* int32 vector type */
 216    memset(&i32_type, 0, sizeof i32_type);
 217    i32_type.floating = FALSE; /* values are integers */
 218    i32_type.sign = TRUE;      /* values are signed */
 219    i32_type.norm = FALSE;     /* values are not normalized */
 220    i32_type.width = 32;       /* 32-bit int values */
 221    i32_type.length = 4;       /* 4 elements per vector */
 222
 223    i32vec4_type = lp_build_int32_vec4_type();
 224
 225    /* c0_vec = {c0, c0, c0, c0}
 226     * Note that we emit this code four times but LLVM optimizes away
 227     * three instances of it.
 228     */
 229    c0_vec = lp_build_broadcast(builder, i32vec4_type, c0);
 230    c1_vec = lp_build_broadcast(builder, i32vec4_type, c1);
 231    c2_vec = lp_build_broadcast(builder, i32vec4_type, c2);
 232
 233    lp_build_name(c0_vec, "edgeconst0vec");
 234    lp_build_name(c1_vec, "edgeconst1vec");
 235    lp_build_name(c2_vec, "edgeconst2vec");
 236
 237    index = LLVMConstInt(LLVMInt32Type(), i, 0);
 238    step0_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step0_ptr, &index, 1, ""), "");
 239    step1_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step1_ptr, &index, 1, ""), "");
 240    step2_vec = LLVMBuildLoad(builder, LLVMBuildGEP(builder, step2_ptr, &index, 1, ""), "");
 241
 242    lp_build_name(step0_vec, "step0vec");
 243    lp_build_name(step1_vec, "step1vec");
 244    lp_build_name(step2_vec, "step2vec");
 245
 246    m0_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step0_vec, c0_vec);
 247    m1_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step1_vec, c1_vec);
 248    m2_vec = lp_build_compare(builder, i32_type, PIPE_FUNC_GREATER, step2_vec, c2_vec);
 249
 250    m = LLVMBuildAnd(builder, m0_vec, m1_vec, "");
 251    m = LLVMBuildAnd(builder, m, m2_vec, "");
 252
 253    lp_build_name(m, "inoutmaskvec");
 254
 255    *mask = m;
 256
 257    /*
 258     * if mask = {0,0,0,0} skip quad
 259     */
 260 }
 261
 262
 263 /**
 264  * Generate the fragment shader, depth/stencil test, and alpha tests.
 265  * \param i  which quad in the tile, in range [0,3]
 266  */
 267 static void
 268 generate_fs(struct llvmpipe_context *lp,
 269             struct lp_fragment_shader *shader,
 270             const struct lp_fragment_shader_variant_key *key,
 271             LLVMBuilderRef builder,
 272             struct lp_type type,
 273             LLVMValueRef context_ptr,
 274             unsigned i,
 275             const struct lp_build_interp_soa_context *interp,
 276             struct lp_build_sampler_soa *sampler,
 277             LLVMValueRef *pmask,
 278             LLVMValueRef *color,
 279             LLVMValueRef depth_ptr,
 280             LLVMValueRef c0,
 281             LLVMValueRef c1,
 282             LLVMValueRef c2,
 283             LLVMValueRef step0_ptr,
 284             LLVMValueRef step1_ptr,
 285             LLVMValueRef step2_ptr)
 286 {
 287    const struct tgsi_token *tokens = shader->base.tokens;
 288    LLVMTypeRef elem_type;
 289    LLVMTypeRef vec_type;
 290    LLVMTypeRef int_vec_type;
 291    LLVMValueRef consts_ptr;
 292    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
 293    LLVMValueRef z = interp->pos[2];
 294    struct lp_build_flow_context *flow;
 295    struct lp_build_mask_context mask;
 296    boolean early_depth_test;
 297    unsigned attrib;
 298    unsigned chan;
 299
 300    assert(i < 4);
 301
 302    elem_type = lp_build_elem_type(type);
 303    vec_type = lp_build_vec_type(type);
 304    int_vec_type = lp_build_int_vec_type(type);
 305
 306    consts_ptr = lp_jit_context_constants(builder, context_ptr);
 307
 308    flow = lp_build_flow_create(builder);
 309
 310    memset(outputs, 0, sizeof outputs);
 311
 312    lp_build_flow_scope_begin(flow);
 313
 314    /* Declare the color and z variables */
 315    for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 316       color[chan] = LLVMGetUndef(vec_type);
 317       lp_build_flow_scope_declare(flow, &color[chan]);
 318    }
 319    lp_build_flow_scope_declare(flow, &z);
 320
 321    /* do triangle edge testing */
 322    generate_tri_edge_mask(builder, i, pmask,
 323                           c0, c1, c2, step0_ptr, step1_ptr, step2_ptr);
 324
 325    /* 'mask' will control execution based on quad's pixel alive/killed state */
 326    lp_build_mask_begin(&mask, flow, type, *pmask);
 327
 328
 329    early_depth_test =
 330       key->depth.enabled &&
 331       !key->alpha.enabled &&
 332       !shader->info.uses_kill &&
 333       !shader->info.writes_z;
 334
 335    if(early_depth_test)
 336       generate_depth(builder, key,
 337                      type, &mask,
 338                      z, depth_ptr);
 339
 340    lp_build_tgsi_soa(builder, tokens, type, &mask,
 341                      consts_ptr, interp->pos, interp->inputs,
 342                      outputs, sampler);
 343
 344    for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
 345       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 346          if(outputs[attrib][chan]) {
 347             lp_build_name(outputs[attrib][chan], "output%u.%u.%c", i, attrib, "xyzw"[chan]);
 348
 349             switch (shader->info.output_semantic_name[attrib]) {
 350             case TGSI_SEMANTIC_COLOR:
 351                {
 352                   unsigned cbuf = shader->info.output_semantic_index[attrib];
 353
 354                   lp_build_name(outputs[attrib][chan], "color%u.%u.%c", i, attrib, "rgba"[chan]);
 355
 356                   /* Alpha test */
 357                   /* XXX: should the alpha reference value be passed separately? */
 358                   if(cbuf == 0 && chan == 3) {
 359                      LLVMValueRef alpha = outputs[attrib][chan];
 360                      LLVMValueRef alpha_ref_value;
 361                      alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
 362                      alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
 363                      lp_build_alpha_test(builder, &key->alpha, type,
 364                                          &mask, alpha, alpha_ref_value);
 365                   }
 366
 367                   if(cbuf == 0)
 368                      color[chan] = outputs[attrib][chan];
 369
 370                   break;
 371                }
 372
 373             case TGSI_SEMANTIC_POSITION:
 374                if(chan == 2)
 375                   z = outputs[attrib][chan];
 376                break;
 377             }
 378          }
 379       }
 380    }
 381
 382    if(!early_depth_test)
 383       generate_depth(builder, key,
 384                      type, &mask,
 385                      z, depth_ptr);
 386
 387    lp_build_mask_end(&mask);
 388
 389    lp_build_flow_scope_end(flow);
 390
 391    lp_build_flow_destroy(flow);
 392
 393    *pmask = mask.value;
 394
 395 }
 396
 397
 398 /**
 399  * Generate color blending and color output.
 400  */
 401 static void
 402 generate_blend(const struct pipe_blend_state *blend,
 403                LLVMBuilderRef builder,
 404                struct lp_type type,
 405                LLVMValueRef context_ptr,
 406                LLVMValueRef mask,
 407                LLVMValueRef *src,
 408                LLVMValueRef dst_ptr)
 409 {
 410    struct lp_build_context bld;
 411    struct lp_build_flow_context *flow;
 412    struct lp_build_mask_context mask_ctx;
 413    LLVMTypeRef vec_type;
 414    LLVMTypeRef int_vec_type;
 415    LLVMValueRef const_ptr;
 416    LLVMValueRef con[4];
 417    LLVMValueRef dst[4];
 418    LLVMValueRef res[4];
 419    unsigned chan;
 420
 421    lp_build_context_init(&bld, builder, type);
 422
 423    flow = lp_build_flow_create(builder);
 424    lp_build_mask_begin(&mask_ctx, flow, type, mask);
 425
 426    vec_type = lp_build_vec_type(type);
 427    int_vec_type = lp_build_int_vec_type(type);
 428
 429    const_ptr = lp_jit_context_blend_color(builder, context_ptr);
 430    const_ptr = LLVMBuildBitCast(builder, const_ptr,
 431                                 LLVMPointerType(vec_type, 0), "");
 432
 433    for(chan = 0; chan < 4; ++chan) {
 434       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 435       con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
 436
 437       dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
 438
 439       lp_build_name(con[chan], "con.%c", "rgba"[chan]);
 440       lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
 441    }
 442
 443    lp_build_blend_soa(builder, blend, type, src, dst, con, res);
 444
 445    for(chan = 0; chan < 4; ++chan) {
 446       if(blend->colormask & (1 << chan)) {
 447          LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 448          lp_build_name(res[chan], "res.%c", "rgba"[chan]);
 449          res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
 450          LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
 451       }
 452    }
 453
 454    lp_build_mask_end(&mask_ctx);
 455    lp_build_flow_destroy(flow);
 456 }
 457
 458
 459 /**
 460  * Generate the runtime callable function for the whole fragment pipeline.
 461  * Note that the function which we generate operates on a block of 16
 462  * pixels at at time.  The block contains 2x2 quads.  Each quad contains
 463  * 2x2 pixels.
 464  */
 465 static struct lp_fragment_shader_variant *
 466 generate_fragment(struct llvmpipe_context *lp,
 467                   struct lp_fragment_shader *shader,
 468                   const struct lp_fragment_shader_variant_key *key)
 469 {
 470    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
 471    struct lp_fragment_shader_variant *variant;
 472    struct lp_type fs_type;
 473    struct lp_type blend_type;
 474    LLVMTypeRef fs_elem_type;
 475    LLVMTypeRef fs_vec_type;
 476    LLVMTypeRef fs_int_vec_type;
 477    LLVMTypeRef blend_vec_type;
 478    LLVMTypeRef blend_int_vec_type;
 479    LLVMTypeRef arg_types[14];
 480    LLVMTypeRef func_type;
 481    LLVMTypeRef int32_vec4_type = lp_build_int32_vec4_type();
 482    LLVMValueRef context_ptr;
 483    LLVMValueRef x;
 484    LLVMValueRef y;
 485    LLVMValueRef a0_ptr;
 486    LLVMValueRef dadx_ptr;
 487    LLVMValueRef dady_ptr;
 488    LLVMValueRef color_ptr;
 489    LLVMValueRef depth_ptr;
 490    LLVMValueRef c0, c1, c2, step0_ptr, step1_ptr, step2_ptr;
 491    LLVMBasicBlockRef block;
 492    LLVMBuilderRef builder;
 493    LLVMValueRef x0;
 494    LLVMValueRef y0;
 495    struct lp_build_sampler_soa *sampler;
 496    struct lp_build_interp_soa_context interp;
 497    LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
 498    LLVMValueRef fs_out_color[NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
 499    LLVMValueRef blend_mask;
 500    LLVMValueRef blend_in_color[NUM_CHANNELS];
 501    unsigned num_fs;
 502    unsigned i;
 503    unsigned chan;
 504
 505    if (LP_DEBUG & DEBUG_JIT) {
 506       tgsi_dump(shader->base.tokens, 0);
 507       if(key->depth.enabled) {
 508          debug_printf("depth.func = %s\n", debug_dump_func(key->depth.func, TRUE));
 509          debug_printf("depth.writemask = %u\n", key->depth.writemask);
 510       }
 511       if(key->alpha.enabled) {
 512          debug_printf("alpha.func = %s\n", debug_dump_func(key->alpha.func, TRUE));
 513          debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
 514       }
 515       if(key->blend.logicop_enable) {
 516          debug_printf("blend.logicop_func = %u\n", key->blend.logicop_func);
 517       }
 518       else if(key->blend.blend_enable) {
 519          debug_printf("blend.rgb_func = %s\n",   debug_dump_blend_func  (key->blend.rgb_func, TRUE));
 520          debug_printf("blend.rgb_src_factor = %s\n",   debug_dump_blend_factor(key->blend.rgb_src_factor, TRUE));
 521          debug_printf("blend.rgb_dst_factor = %s\n",   debug_dump_blend_factor(key->blend.rgb_dst_factor, TRUE));
 522          debug_printf("blend.alpha_func = %s\n",       debug_dump_blend_func  (key->blend.alpha_func, TRUE));
 523          debug_printf("blend.alpha_src_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_src_factor, TRUE));
 524          debug_printf("blend.alpha_dst_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_dst_factor, TRUE));
 525       }
 526       debug_printf("blend.colormask = 0x%x\n", key->blend.colormask);
 527    }
 528
 529    variant = CALLOC_STRUCT(lp_fragment_shader_variant);
 530    if(!variant)
 531       return NULL;
 532
 533    variant->shader = shader;
 534    memcpy(&variant->key, key, sizeof *key);
 535
 536    /* TODO: actually pick these based on the fs and color buffer
 537     * characteristics. */
 538
 539    memset(&fs_type, 0, sizeof fs_type);
 540    fs_type.floating = TRUE; /* floating point values */
 541    fs_type.sign = TRUE;     /* values are signed */
 542    fs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
 543    fs_type.width = 32;      /* 32-bit float */
 544    fs_type.length = 4;      /* 4 elements per vector */
 545    num_fs = 4;              /* number of quads per block */
 546
 547    memset(&blend_type, 0, sizeof blend_type);
 548    blend_type.floating = FALSE; /* values are integers */
 549    blend_type.sign = FALSE;     /* values are unsigned */
 550    blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
 551    blend_type.width = 8;        /* 8-bit ubyte values */
 552    blend_type.length = 16;      /* 16 elements per vector */
 553
 554    /*
 555     * Generate the function prototype. Any change here must be reflected in
 556     * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
 557     */
 558
 559    fs_elem_type = lp_build_elem_type(fs_type);
 560    fs_vec_type = lp_build_vec_type(fs_type);
 561    fs_int_vec_type = lp_build_int_vec_type(fs_type);
 562
 563    blend_vec_type = lp_build_vec_type(blend_type);
 564    blend_int_vec_type = lp_build_int_vec_type(blend_type);
 565
 566    arg_types[0] = screen->context_ptr_type;            /* context */
 567    arg_types[1] = LLVMInt32Type();                     /* x */
 568    arg_types[2] = LLVMInt32Type();                     /* y */
 569    arg_types[3] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
 570    arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
 571    arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* dady */
 572    arg_types[6] = LLVMPointerType(blend_vec_type, 0);  /* color */
 573    arg_types[7] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
 574    arg_types[8] = LLVMInt32Type();                     /* c0 */
 575    arg_types[9] = LLVMInt32Type();                    /* c1 */
 576    arg_types[10] = LLVMInt32Type();                    /* c2 */
 577    /* Note: the step arrays are built as int32[16] but we interpret
 578     * them here as int32_vec4[4].
 579     */
 580    arg_types[11] = LLVMPointerType(int32_vec4_type, 0);/* step0 */
 581    arg_types[12] = LLVMPointerType(int32_vec4_type, 0);/* step1 */
 582    arg_types[13] = LLVMPointerType(int32_vec4_type, 0);/* step2 */
 583
 584    func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
 585
 586    variant->function = LLVMAddFunction(screen->module, "shader", func_type);
 587    LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
 588    for(i = 0; i < Elements(arg_types); ++i)
 589       if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
 590          LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute);
 591
 592    context_ptr  = LLVMGetParam(variant->function, 0);
 593    x            = LLVMGetParam(variant->function, 1);
 594    y            = LLVMGetParam(variant->function, 2);
 595    a0_ptr       = LLVMGetParam(variant->function, 3);
 596    dadx_ptr     = LLVMGetParam(variant->function, 4);
 597    dady_ptr     = LLVMGetParam(variant->function, 5);
 598    color_ptr    = LLVMGetParam(variant->function, 6);
 599    depth_ptr    = LLVMGetParam(variant->function, 7);
 600    c0           = LLVMGetParam(variant->function, 8);
 601    c1           = LLVMGetParam(variant->function, 9);
 602    c2           = LLVMGetParam(variant->function, 10);
 603    step0_ptr    = LLVMGetParam(variant->function, 11);
 604    step1_ptr    = LLVMGetParam(variant->function, 12);
 605    step2_ptr    = LLVMGetParam(variant->function, 13);
 606
 607    lp_build_name(context_ptr, "context");
 608    lp_build_name(x, "x");
 609    lp_build_name(y, "y");
 610    lp_build_name(a0_ptr, "a0");
 611    lp_build_name(dadx_ptr, "dadx");
 612    lp_build_name(dady_ptr, "dady");
 613    lp_build_name(color_ptr, "color");
 614    lp_build_name(depth_ptr, "depth");
 615    lp_build_name(c0, "c0");
 616    lp_build_name(c1, "c1");
 617    lp_build_name(c2, "c2");
 618    lp_build_name(step0_ptr, "step0");
 619    lp_build_name(step1_ptr, "step1");
 620    lp_build_name(step2_ptr, "step2");
 621
 622    /*
 623     * Function body
 624     */
 625
 626    block = LLVMAppendBasicBlock(variant->function, "entry");
 627    builder = LLVMCreateBuilder();
 628    LLVMPositionBuilderAtEnd(builder, block);
 629
 630    generate_pos0(builder, x, y, &x0, &y0);
 631
 632    lp_build_interp_soa_init(&interp, shader->base.tokens, builder, fs_type,
 633                             a0_ptr, dadx_ptr, dady_ptr,
 634                             x0, y0);
 635
 636    /* code generated texture sampling */
 637    sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
 638
 639    /* loop over quads in the block */
 640    for(i = 0; i < num_fs; ++i) {
 641       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
 642       LLVMValueRef out_color[NUM_CHANNELS];
 643       LLVMValueRef depth_ptr_i;
 644
 645       if(i != 0)
 646          lp_build_interp_soa_update(&interp, i);
 647
 648       depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
 649
 650       generate_fs(lp, shader, key,
 651                   builder,
 652                   fs_type,
 653                   context_ptr,
 654                   i,
 655                   &interp,
 656                   sampler,
 657                   &fs_mask[i], /* output */
 658                   out_color,
 659                   depth_ptr_i,
 660                   c0, c1, c2,
 661                   step0_ptr, step1_ptr, step2_ptr);
 662
 663       for(chan = 0; chan < NUM_CHANNELS; ++chan)
 664          fs_out_color[chan][i] = out_color[chan];
 665    }
 666
 667    sampler->destroy(sampler);
 668
 669    /*
 670     * Convert the fs's output color and mask to fit to the blending type.
 671     */
 672
 673    for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 674       lp_build_conv(builder, fs_type, blend_type,
 675                     fs_out_color[chan], num_fs,
 676                     &blend_in_color[chan], 1);
 677       lp_build_name(blend_in_color[chan], "color.%c", "rgba"[chan]);
 678
 679    }
 680
 681    lp_build_conv_mask(builder, fs_type, blend_type,
 682                       fs_mask, num_fs,
 683                       &blend_mask, 1);
 684
 685    /*
 686     * Blending.
 687     */
 688
 689    generate_blend(&key->blend,
 690                   builder,
 691                   blend_type,
 692                   context_ptr,
 693                   blend_mask,
 694                   blend_in_color,
 695                   color_ptr);
 696
 697    LLVMBuildRetVoid(builder);
 698
 699    LLVMDisposeBuilder(builder);
 700
 701    /*
 702     * Translate the LLVM IR into machine code.
 703     */
 704
 705    if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) {
 706       LLVMDumpValue(variant->function);
 707       abort();
 708    }
 709
 710    LLVMRunFunctionPassManager(screen->pass, variant->function);
 711
 712    if (LP_DEBUG & DEBUG_JIT) {
 713       LLVMDumpValue(variant->function);
 714       debug_printf("\n");
 715    }
 716
 717    variant->jit_function = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, variant->function);
 718
 719    if (LP_DEBUG & DEBUG_ASM)
 720       lp_disassemble(variant->jit_function);
 721
 722    variant->next = shader->variants;
 723    shader->variants = variant;
 724
 725    return variant;
 726 }
 727
 728
 729 void *
 730 llvmpipe_create_fs_state(struct pipe_context *pipe,
 731                          const struct pipe_shader_state *templ)
 732 {
 733    struct lp_fragment_shader *shader;
 734
 735    shader = CALLOC_STRUCT(lp_fragment_shader);
 736    if (!shader)
 737       return NULL;
 738
 739    /* get/save the summary info for this shader */
 740    tgsi_scan_shader(templ->tokens, &shader->info);
 741
 742    /* we need to keep a local copy of the tokens */
 743    shader->base.tokens = tgsi_dup_tokens(templ->tokens);
 744
 745    return shader;
 746 }
 747
 748
 749 void
 750 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
 751 {
 752    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 753
 754    llvmpipe->fs = (struct lp_fragment_shader *) fs;
 755
 756    llvmpipe->dirty |= LP_NEW_FS;
 757 }
 758
 759
 760 void
 761 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
 762 {
 763    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 764    struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
 765    struct lp_fragment_shader *shader = fs;
 766    struct lp_fragment_shader_variant *variant;
 767
 768    assert(fs != llvmpipe->fs);
 769
 770    variant = shader->variants;
 771    while(variant) {
 772       struct lp_fragment_shader_variant *next = variant->next;
 773
 774       if(variant->function) {
 775          if(variant->jit_function)
 776             LLVMFreeMachineCodeForFunction(screen->engine, variant->function);
 777          LLVMDeleteFunction(variant->function);
 778       }
 779
 780       FREE(variant);
 781
 782       variant = next;
 783    }
 784
 785    FREE((void *) shader->base.tokens);
 786    FREE(shader);
 787 }
 788
 789
 790
 791 void
 792 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
 793                              uint shader, uint index,
 794                              const struct pipe_constant_buffer *constants)
 795 {
 796    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 797    struct pipe_buffer *buffer = constants ? constants->buffer : NULL;
 798    unsigned size = buffer ? buffer->size : 0;
 799    const void *data = buffer ? llvmpipe_buffer(buffer)->data : NULL;
 800
 801    assert(shader < PIPE_SHADER_TYPES);
 802    assert(index == 0);
 803
 804    if(llvmpipe->constants[shader].buffer == buffer)
 805       return;
 806
 807    if(shader == PIPE_SHADER_VERTEX)
 808       draw_flush(llvmpipe->draw);
 809
 810    /* note: reference counting */
 811    pipe_buffer_reference(&llvmpipe->constants[shader].buffer, buffer);
 812
 813    if(shader == PIPE_SHADER_VERTEX) {
 814       draw_set_mapped_constant_buffer(llvmpipe->draw, data, size);
 815    }
 816
 817    llvmpipe->dirty |= LP_NEW_CONSTANTS;
 818 }
 819
 820
 821 /**
 822  * We need to generate several variants of the fragment pipeline to match
 823  * all the combinations of the contributing state atoms.
 824  *
 825  * TODO: there is actually no reason to tie this to context state -- the
 826  * generated code could be cached globally in the screen.
 827  */
 828 static void
 829 make_variant_key(struct llvmpipe_context *lp,
 830                  struct lp_fragment_shader *shader,
 831                  struct lp_fragment_shader_variant_key *key)
 832 {
 833    unsigned i;
 834
 835    memset(key, 0, sizeof *key);
 836
 837    if(lp->framebuffer.zsbuf &&
 838       lp->depth_stencil->depth.enabled) {
 839       key->zsbuf_format = lp->framebuffer.zsbuf->format;
 840       memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
 841    }
 842
 843    key->alpha.enabled = lp->depth_stencil->alpha.enabled;
 844    if(key->alpha.enabled)
 845       key->alpha.func = lp->depth_stencil->alpha.func;
 846    /* alpha.ref_value is passed in jit_context */
 847
 848    if(lp->framebuffer.cbufs[0]) {
 849       const struct util_format_description *format_desc;
 850       unsigned chan;
 851
 852       memcpy(&key->blend, lp->blend, sizeof key->blend);
 853
 854       format_desc = util_format_description(lp->framebuffer.cbufs[0]->format);
 855       assert(format_desc->layout == UTIL_FORMAT_COLORSPACE_RGB ||
 856              format_desc->layout == UTIL_FORMAT_COLORSPACE_SRGB);
 857
 858       /* mask out color channels not present in the color buffer */
 859       for(chan = 0; chan < 4; ++chan) {
 860          enum util_format_swizzle swizzle = format_desc->swizzle[chan];
 861          if(swizzle > 4)
 862             key->blend.colormask &= ~(1 << chan);
 863       }
 864    }
 865
 866    for(i = 0; i < PIPE_MAX_SAMPLERS; ++i)
 867       if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i))
 868          lp_sampler_static_state(&key->sampler[i], lp->texture[i], lp->sampler[i]);
 869 }
 870
 871
 872 void
 873 llvmpipe_update_fs(struct llvmpipe_context *lp)
 874 {
 875    struct lp_fragment_shader *shader = lp->fs;
 876    struct lp_fragment_shader_variant_key key;
 877    struct lp_fragment_shader_variant *variant;
 878
 879    make_variant_key(lp, shader, &key);
 880
 881    variant = shader->variants;
 882    while(variant) {
 883       if(memcmp(&variant->key, &key, sizeof key) == 0)
 884          break;
 885
 886       variant = variant->next;
 887    }
 888
 889    if(!variant)
 890       variant = generate_fragment(lp, shader, &key);
 891
 892    shader->current = variant;
 893
 894    lp_setup_set_fs_function(lp->setup,
 895                             shader->current->jit_function);
 896 }