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_state.h"
  88 #include "lp_quad.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    /*
 153     * Depths are expected to be between 0 and 1, even if they are stored in
 154     * floats. Setting these bits here will ensure that the lp_build_conv() call
 155     * below won't try to unnecessarily clamp the incoming values.
 156     */
 157    if(src_type.floating) {
 158       src_type.sign = FALSE;
 159       src_type.norm = TRUE;
 160    }
 161    else {
 162       assert(!src_type.sign);
 163       assert(src_type.norm);
 164    }
 165
 166    /* Pick the depth type. */
 167    dst_type = lp_depth_type(format_desc, src_type.width*src_type.length);
 168
 169    /* FIXME: Cope with a depth test type with a different bit width. */
 170    assert(dst_type.width == src_type.width);
 171    assert(dst_type.length == src_type.length);
 172
 173    lp_build_conv(builder, src_type, dst_type, &src, 1, &src, 1);
 174
 175    dst_ptr = LLVMBuildBitCast(builder,
 176                               dst_ptr,
 177                               LLVMPointerType(lp_build_vec_type(dst_type), 0), "");
 178
 179    lp_build_depth_test(builder,
 180                        &key->depth,
 181                        dst_type,
 182                        format_desc,
 183                        mask,
 184                        src,
 185                        dst_ptr);
 186 }
 187
 188
 189 /**
 190  * Generate the fragment shader, depth/stencil test, and alpha tests.
 191  */
 192 static void
 193 generate_fs(struct llvmpipe_context *lp,
 194             struct lp_fragment_shader *shader,
 195             const struct lp_fragment_shader_variant_key *key,
 196             LLVMBuilderRef builder,
 197             struct lp_type type,
 198             LLVMValueRef context_ptr,
 199             unsigned i,
 200             const struct lp_build_interp_soa_context *interp,
 201             struct lp_build_sampler_soa *sampler,
 202             LLVMValueRef *pmask,
 203             LLVMValueRef *color,
 204             LLVMValueRef depth_ptr)
 205 {
 206    const struct tgsi_token *tokens = shader->base.tokens;
 207    LLVMTypeRef elem_type;
 208    LLVMTypeRef vec_type;
 209    LLVMTypeRef int_vec_type;
 210    LLVMValueRef consts_ptr;
 211    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
 212    LLVMValueRef z = interp->pos[2];
 213    struct lp_build_flow_context *flow;
 214    struct lp_build_mask_context mask;
 215    boolean early_depth_test;
 216    unsigned attrib;
 217    unsigned chan;
 218
 219    elem_type = lp_build_elem_type(type);
 220    vec_type = lp_build_vec_type(type);
 221    int_vec_type = lp_build_int_vec_type(type);
 222
 223    consts_ptr = lp_jit_context_constants(builder, context_ptr);
 224
 225    flow = lp_build_flow_create(builder);
 226
 227    memset(outputs, 0, sizeof outputs);
 228
 229    lp_build_flow_scope_begin(flow);
 230
 231    /* Declare the color and z variables */
 232    for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 233       color[chan] = LLVMGetUndef(vec_type);
 234       lp_build_flow_scope_declare(flow, &color[chan]);
 235    }
 236    lp_build_flow_scope_declare(flow, &z);
 237
 238    lp_build_mask_begin(&mask, flow, type, *pmask);
 239
 240    early_depth_test =
 241       key->depth.enabled &&
 242       !key->alpha.enabled &&
 243       !shader->info.uses_kill &&
 244       !shader->info.writes_z;
 245
 246    if(early_depth_test)
 247       generate_depth(builder, key,
 248                      type, &mask,
 249                      z, depth_ptr);
 250
 251    lp_build_tgsi_soa(builder, tokens, type, &mask,
 252                      consts_ptr, interp->pos, interp->inputs,
 253                      outputs, sampler);
 254
 255    for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
 256       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 257          if(outputs[attrib][chan]) {
 258             lp_build_name(outputs[attrib][chan], "output%u.%u.%c", i, attrib, "xyzw"[chan]);
 259
 260             switch (shader->info.output_semantic_name[attrib]) {
 261             case TGSI_SEMANTIC_COLOR:
 262                {
 263                   unsigned cbuf = shader->info.output_semantic_index[attrib];
 264
 265                   lp_build_name(outputs[attrib][chan], "color%u.%u.%c", i, attrib, "rgba"[chan]);
 266
 267                   /* Alpha test */
 268                   /* XXX: should the alpha reference value be passed separately? */
 269                   if(cbuf == 0 && chan == 3) {
 270                      LLVMValueRef alpha = outputs[attrib][chan];
 271                      LLVMValueRef alpha_ref_value;
 272                      alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
 273                      alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
 274                      lp_build_alpha_test(builder, &key->alpha, type,
 275                                          &mask, alpha, alpha_ref_value);
 276                   }
 277
 278                   if(cbuf == 0)
 279                      color[chan] = outputs[attrib][chan];
 280
 281                   break;
 282                }
 283
 284             case TGSI_SEMANTIC_POSITION:
 285                if(chan == 2)
 286                   z = outputs[attrib][chan];
 287                break;
 288             }
 289          }
 290       }
 291    }
 292
 293    if(!early_depth_test)
 294       generate_depth(builder, key,
 295                      type, &mask,
 296                      z, depth_ptr);
 297
 298    lp_build_mask_end(&mask);
 299
 300    lp_build_flow_scope_end(flow);
 301
 302    lp_build_flow_destroy(flow);
 303
 304    *pmask = mask.value;
 305
 306 }
 307
 308
 309 /**
 310  * Generate color blending and color output.
 311  */
 312 static void
 313 generate_blend(const struct pipe_blend_state *blend,
 314                LLVMBuilderRef builder,
 315                struct lp_type type,
 316                LLVMValueRef context_ptr,
 317                LLVMValueRef mask,
 318                LLVMValueRef *src,
 319                LLVMValueRef dst_ptr)
 320 {
 321    struct lp_build_context bld;
 322    struct lp_build_flow_context *flow;
 323    struct lp_build_mask_context mask_ctx;
 324    LLVMTypeRef vec_type;
 325    LLVMTypeRef int_vec_type;
 326    LLVMValueRef const_ptr;
 327    LLVMValueRef con[4];
 328    LLVMValueRef dst[4];
 329    LLVMValueRef res[4];
 330    unsigned chan;
 331
 332    lp_build_context_init(&bld, builder, type);
 333
 334    flow = lp_build_flow_create(builder);
 335    lp_build_mask_begin(&mask_ctx, flow, type, mask);
 336
 337    vec_type = lp_build_vec_type(type);
 338    int_vec_type = lp_build_int_vec_type(type);
 339
 340    const_ptr = lp_jit_context_blend_color(builder, context_ptr);
 341    const_ptr = LLVMBuildBitCast(builder, const_ptr,
 342                                 LLVMPointerType(vec_type, 0), "");
 343
 344    for(chan = 0; chan < 4; ++chan) {
 345       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 346       con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
 347
 348       dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
 349
 350       lp_build_name(con[chan], "con.%c", "rgba"[chan]);
 351       lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
 352    }
 353
 354    lp_build_blend_soa(builder, blend, type, src, dst, con, res);
 355
 356    for(chan = 0; chan < 4; ++chan) {
 357       if(blend->colormask & (1 << chan)) {
 358          LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 359          lp_build_name(res[chan], "res.%c", "rgba"[chan]);
 360          res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
 361          LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
 362       }
 363    }
 364
 365    lp_build_mask_end(&mask_ctx);
 366    lp_build_flow_destroy(flow);
 367 }
 368
 369
 370 /**
 371  * Generate the runtime callable function for the whole fragment pipeline.
 372  */
 373 static struct lp_fragment_shader_variant *
 374 generate_fragment(struct llvmpipe_context *lp,
 375                   struct lp_fragment_shader *shader,
 376                   const struct lp_fragment_shader_variant_key *key)
 377 {
 378    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
 379    struct lp_fragment_shader_variant *variant;
 380    struct lp_type fs_type;
 381    struct lp_type blend_type;
 382    LLVMTypeRef fs_elem_type;
 383    LLVMTypeRef fs_vec_type;
 384    LLVMTypeRef fs_int_vec_type;
 385    LLVMTypeRef blend_vec_type;
 386    LLVMTypeRef blend_int_vec_type;
 387    LLVMTypeRef arg_types[9];
 388    LLVMTypeRef func_type;
 389    LLVMValueRef context_ptr;
 390    LLVMValueRef x;
 391    LLVMValueRef y;
 392    LLVMValueRef a0_ptr;
 393    LLVMValueRef dadx_ptr;
 394    LLVMValueRef dady_ptr;
 395    LLVMValueRef mask_ptr;
 396    LLVMValueRef color_ptr;
 397    LLVMValueRef depth_ptr;
 398    LLVMBasicBlockRef block;
 399    LLVMBuilderRef builder;
 400    LLVMValueRef x0;
 401    LLVMValueRef y0;
 402    struct lp_build_sampler_soa *sampler;
 403    struct lp_build_interp_soa_context interp;
 404    LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
 405    LLVMValueRef fs_out_color[NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
 406    LLVMValueRef blend_mask;
 407    LLVMValueRef blend_in_color[NUM_CHANNELS];
 408    unsigned num_fs;
 409    unsigned i;
 410    unsigned chan;
 411
 412    if (LP_DEBUG & DEBUG_JIT) {
 413       tgsi_dump(shader->base.tokens, 0);
 414       if(key->depth.enabled) {
 415          debug_printf("depth.format = %s\n", pf_name(key->zsbuf_format));
 416          debug_printf("depth.func = %s\n", debug_dump_func(key->depth.func, TRUE));
 417          debug_printf("depth.writemask = %u\n", key->depth.writemask);
 418       }
 419       if(key->alpha.enabled) {
 420          debug_printf("alpha.func = %s\n", debug_dump_func(key->alpha.func, TRUE));
 421          debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
 422       }
 423       if(key->blend.logicop_enable) {
 424          debug_printf("blend.logicop_func = %u\n", key->blend.logicop_func);
 425       }
 426       else if(key->blend.blend_enable) {
 427          debug_printf("blend.rgb_func = %s\n",   debug_dump_blend_func  (key->blend.rgb_func, TRUE));
 428          debug_printf("rgb_src_factor = %s\n",   debug_dump_blend_factor(key->blend.rgb_src_factor, TRUE));
 429          debug_printf("rgb_dst_factor = %s\n",   debug_dump_blend_factor(key->blend.rgb_dst_factor, TRUE));
 430          debug_printf("alpha_func = %s\n",       debug_dump_blend_func  (key->blend.alpha_func, TRUE));
 431          debug_printf("alpha_src_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_src_factor, TRUE));
 432          debug_printf("alpha_dst_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_dst_factor, TRUE));
 433       }
 434       debug_printf("blend.colormask = 0x%x\n", key->blend.colormask);
 435       for(i = 0; i < PIPE_MAX_SAMPLERS; ++i) {
 436          if(key->sampler[i].format) {
 437             debug_printf("sampler[%u] = \n", i);
 438             debug_printf("  .format = %s\n",
 439                          pf_name(key->sampler[i].format));
 440             debug_printf("  .target = %s\n",
 441                          debug_dump_tex_target(key->sampler[i].target, TRUE));
 442             debug_printf("  .pot = %u %u %u\n",
 443                          key->sampler[i].pot_width,
 444                          key->sampler[i].pot_height,
 445                          key->sampler[i].pot_depth);
 446             debug_printf("  .wrap = %s %s %s\n",
 447                          debug_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
 448                          debug_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
 449                          debug_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
 450             debug_printf("  .min_img_filter = %s\n",
 451                          debug_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
 452             debug_printf("  .min_mip_filter = %s\n",
 453                          debug_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
 454             debug_printf("  .mag_img_filter = %s\n",
 455                          debug_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
 456             if(key->sampler[i].compare_mode != PIPE_TEX_COMPARE_NONE)
 457                debug_printf("  .compare_func = %s\n", debug_dump_func(key->sampler[i].compare_func, TRUE));
 458             debug_printf("  .normalized_coords = %u\n", key->sampler[i].normalized_coords);
 459             debug_printf("  .prefilter = %u\n", key->sampler[i].prefilter);
 460          }
 461       }
 462    }
 463
 464    variant = CALLOC_STRUCT(lp_fragment_shader_variant);
 465    if(!variant)
 466       return NULL;
 467
 468    variant->shader = shader;
 469    memcpy(&variant->key, key, sizeof *key);
 470
 471    /* TODO: actually pick these based on the fs and color buffer
 472     * characteristics. */
 473
 474    memset(&fs_type, 0, sizeof fs_type);
 475    fs_type.floating = TRUE; /* floating point values */
 476    fs_type.sign = TRUE;     /* values are signed */
 477    fs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
 478    fs_type.width = 32;      /* 32-bit float */
 479    fs_type.length = 4;      /* 4 element per vector */
 480    num_fs = 4;
 481
 482    memset(&blend_type, 0, sizeof blend_type);
 483    blend_type.floating = FALSE; /* values are integers */
 484    blend_type.sign = FALSE;     /* values are unsigned */
 485    blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
 486    blend_type.width = 8;        /* 8-bit ubyte values */
 487    blend_type.length = 16;      /* 16 elements per vector */
 488
 489    /*
 490     * Generate the function prototype. Any change here must be reflected in
 491     * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
 492     */
 493
 494    fs_elem_type = lp_build_elem_type(fs_type);
 495    fs_vec_type = lp_build_vec_type(fs_type);
 496    fs_int_vec_type = lp_build_int_vec_type(fs_type);
 497
 498    blend_vec_type = lp_build_vec_type(blend_type);
 499    blend_int_vec_type = lp_build_int_vec_type(blend_type);
 500
 501    arg_types[0] = screen->context_ptr_type;            /* context */
 502    arg_types[1] = LLVMInt32Type();                     /* x */
 503    arg_types[2] = LLVMInt32Type();                     /* y */
 504    arg_types[3] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
 505    arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
 506    arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* dady */
 507    arg_types[6] = LLVMPointerType(fs_int_vec_type, 0); /* mask */
 508    arg_types[7] = LLVMPointerType(blend_vec_type, 0);  /* color */
 509    arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
 510
 511    func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
 512
 513    variant->function = LLVMAddFunction(screen->module, "shader", func_type);
 514    LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
 515    for(i = 0; i < Elements(arg_types); ++i)
 516       if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
 517          LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute);
 518
 519    context_ptr  = LLVMGetParam(variant->function, 0);
 520    x            = LLVMGetParam(variant->function, 1);
 521    y            = LLVMGetParam(variant->function, 2);
 522    a0_ptr       = LLVMGetParam(variant->function, 3);
 523    dadx_ptr     = LLVMGetParam(variant->function, 4);
 524    dady_ptr     = LLVMGetParam(variant->function, 5);
 525    mask_ptr     = LLVMGetParam(variant->function, 6);
 526    color_ptr    = LLVMGetParam(variant->function, 7);
 527    depth_ptr    = LLVMGetParam(variant->function, 8);
 528
 529    lp_build_name(context_ptr, "context");
 530    lp_build_name(x, "x");
 531    lp_build_name(y, "y");
 532    lp_build_name(a0_ptr, "a0");
 533    lp_build_name(dadx_ptr, "dadx");
 534    lp_build_name(dady_ptr, "dady");
 535    lp_build_name(mask_ptr, "mask");
 536    lp_build_name(color_ptr, "color");
 537    lp_build_name(depth_ptr, "depth");
 538
 539    /*
 540     * Function body
 541     */
 542
 543    block = LLVMAppendBasicBlock(variant->function, "entry");
 544    builder = LLVMCreateBuilder();
 545    LLVMPositionBuilderAtEnd(builder, block);
 546
 547    generate_pos0(builder, x, y, &x0, &y0);
 548
 549    lp_build_interp_soa_init(&interp, shader->base.tokens, builder, fs_type,
 550                             a0_ptr, dadx_ptr, dady_ptr,
 551                             x0, y0, 2, 0);
 552
 553    /* code generated texture sampling */
 554    sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
 555
 556    for(i = 0; i < num_fs; ++i) {
 557       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
 558       LLVMValueRef out_color[NUM_CHANNELS];
 559       LLVMValueRef depth_ptr_i;
 560
 561       if(i != 0)
 562          lp_build_interp_soa_update(&interp);
 563
 564       fs_mask[i] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, mask_ptr, &index, 1, ""), "");
 565       depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
 566
 567       generate_fs(lp, shader, key,
 568                   builder,
 569                   fs_type,
 570                   context_ptr,
 571                   i,
 572                   &interp,
 573                   sampler,
 574                   &fs_mask[i],
 575                   out_color,
 576                   depth_ptr_i);
 577
 578       for(chan = 0; chan < NUM_CHANNELS; ++chan)
 579          fs_out_color[chan][i] = out_color[chan];
 580    }
 581
 582    sampler->destroy(sampler);
 583
 584    /*
 585     * Convert the fs's output color and mask to fit to the blending type.
 586     */
 587
 588    for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 589       lp_build_conv(builder, fs_type, blend_type,
 590                     fs_out_color[chan], num_fs,
 591                     &blend_in_color[chan], 1);
 592       lp_build_name(blend_in_color[chan], "color.%c", "rgba"[chan]);
 593
 594    }
 595
 596    lp_build_conv_mask(builder, fs_type, blend_type,
 597                       fs_mask, num_fs,
 598                       &blend_mask, 1);
 599
 600    /*
 601     * Blending.
 602     */
 603
 604    generate_blend(&key->blend,
 605                   builder,
 606                   blend_type,
 607                   context_ptr,
 608                   blend_mask,
 609                   blend_in_color,
 610                   color_ptr);
 611
 612    LLVMBuildRetVoid(builder);
 613
 614    LLVMDisposeBuilder(builder);
 615
 616    /*
 617     * Translate the LLVM IR into machine code.
 618     */
 619
 620 #ifdef DEBUG
 621    if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) {
 622       LLVMDumpValue(variant->function);
 623       assert(0);
 624    }
 625 #endif
 626
 627    LLVMRunFunctionPassManager(screen->pass, variant->function);
 628
 629    if (LP_DEBUG & DEBUG_JIT) {
 630       LLVMDumpValue(variant->function);
 631       debug_printf("\n");
 632    }
 633
 634    variant->jit_function = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, variant->function);
 635
 636    if (LP_DEBUG & DEBUG_ASM)
 637       lp_disassemble(variant->jit_function);
 638
 639    variant->next = shader->variants;
 640    shader->variants = variant;
 641
 642    return variant;
 643 }
 644
 645
 646 void *
 647 llvmpipe_create_fs_state(struct pipe_context *pipe,
 648                          const struct pipe_shader_state *templ)
 649 {
 650    struct lp_fragment_shader *shader;
 651
 652    shader = CALLOC_STRUCT(lp_fragment_shader);
 653    if (!shader)
 654       return NULL;
 655
 656    /* get/save the summary info for this shader */
 657    tgsi_scan_shader(templ->tokens, &shader->info);
 658
 659    /* we need to keep a local copy of the tokens */
 660    shader->base.tokens = tgsi_dup_tokens(templ->tokens);
 661
 662    return shader;
 663 }
 664
 665
 666 void
 667 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
 668 {
 669    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 670
 671    if (llvmpipe->fs == fs)
 672       return;
 673
 674    draw_flush(llvmpipe->draw);
 675
 676    llvmpipe->fs = fs;
 677
 678    llvmpipe->dirty |= LP_NEW_FS;
 679 }
 680
 681
 682 void
 683 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
 684 {
 685    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 686    struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
 687    struct lp_fragment_shader *shader = fs;
 688    struct lp_fragment_shader_variant *variant;
 689
 690    assert(fs != llvmpipe->fs);
 691    (void) llvmpipe;
 692
 693    variant = shader->variants;
 694    while(variant) {
 695       struct lp_fragment_shader_variant *next = variant->next;
 696
 697       if(variant->function) {
 698          if(variant->jit_function)
 699             LLVMFreeMachineCodeForFunction(screen->engine, variant->function);
 700          LLVMDeleteFunction(variant->function);
 701       }
 702
 703       FREE(variant);
 704
 705       variant = next;
 706    }
 707
 708    FREE((void *) shader->base.tokens);
 709    FREE(shader);
 710 }
 711
 712
 713
 714 void
 715 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
 716                              uint shader, uint index,
 717                              struct pipe_buffer *constants)
 718 {
 719    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 720    unsigned size = constants ? constants->size : 0;
 721    const void *data = constants ? llvmpipe_buffer(constants)->data : NULL;
 722
 723    assert(shader < PIPE_SHADER_TYPES);
 724    assert(index == 0);
 725
 726    draw_flush(llvmpipe->draw);
 727
 728    /* note: reference counting */
 729    pipe_buffer_reference(&llvmpipe->constants[shader], constants);
 730
 731    if(shader == PIPE_SHADER_FRAGMENT) {
 732       llvmpipe->jit_context.constants = data;
 733    }
 734
 735    if(shader == PIPE_SHADER_VERTEX) {
 736       draw_set_mapped_constant_buffer(llvmpipe->draw, PIPE_SHADER_VERTEX,
 737                                       data, size);
 738    }
 739
 740    llvmpipe->dirty |= LP_NEW_CONSTANTS;
 741 }
 742
 743
 744 /**
 745  * We need to generate several variants of the fragment pipeline to match
 746  * all the combinations of the contributing state atoms.
 747  *
 748  * TODO: there is actually no reason to tie this to context state -- the
 749  * generated code could be cached globally in the screen.
 750  */
 751 static void
 752 make_variant_key(struct llvmpipe_context *lp,
 753                  struct lp_fragment_shader *shader,
 754                  struct lp_fragment_shader_variant_key *key)
 755 {
 756    unsigned i;
 757
 758    memset(key, 0, sizeof *key);
 759
 760    if(lp->framebuffer.zsbuf &&
 761       lp->depth_stencil->depth.enabled) {
 762       key->zsbuf_format = lp->framebuffer.zsbuf->format;
 763       memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
 764    }
 765
 766    key->alpha.enabled = lp->depth_stencil->alpha.enabled;
 767    if(key->alpha.enabled)
 768       key->alpha.func = lp->depth_stencil->alpha.func;
 769    /* alpha.ref_value is passed in jit_context */
 770
 771    if(lp->framebuffer.cbufs[0]) {
 772       const struct util_format_description *format_desc;
 773       unsigned chan;
 774
 775       memcpy(&key->blend, lp->blend, sizeof key->blend);
 776
 777       format_desc = util_format_description(lp->framebuffer.cbufs[0]->format);
 778       assert(format_desc->layout == UTIL_FORMAT_COLORSPACE_RGB ||
 779              format_desc->layout == UTIL_FORMAT_COLORSPACE_SRGB);
 780
 781       /* mask out color channels not present in the color buffer */
 782       for(chan = 0; chan < 4; ++chan) {
 783          enum util_format_swizzle swizzle = format_desc->swizzle[chan];
 784          if(swizzle > 4)
 785             key->blend.colormask &= ~(1 << chan);
 786       }
 787    }
 788
 789    for(i = 0; i < PIPE_MAX_SAMPLERS; ++i)
 790       if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i))
 791          lp_sampler_static_state(&key->sampler[i], lp->texture[i], lp->sampler[i]);
 792 }
 793
 794
 795 void
 796 llvmpipe_update_fs(struct llvmpipe_context *lp)
 797 {
 798    struct lp_fragment_shader *shader = lp->fs;
 799    struct lp_fragment_shader_variant_key key;
 800    struct lp_fragment_shader_variant *variant;
 801
 802    make_variant_key(lp, shader, &key);
 803
 804    variant = shader->variants;
 805    while(variant) {
 806       if(memcmp(&variant->key, &key, sizeof key) == 0)
 807          break;
 808
 809       variant = variant->next;
 810    }
 811
 812    if(!variant)
 813       variant = generate_fragment(lp, shader, &key);
 814
 815    shader->current = variant;
 816 }