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