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 fragment shader, depth/stencil test, and alpha tests.
 180  */
 181 static void
 182 generate_fs(struct llvmpipe_context *lp,
 183             struct lp_fragment_shader *shader,
 184             const struct lp_fragment_shader_variant_key *key,
 185             LLVMBuilderRef builder,
 186             struct lp_type type,
 187             LLVMValueRef context_ptr,
 188             unsigned i,
 189             const struct lp_build_interp_soa_context *interp,
 190             struct lp_build_sampler_soa *sampler,
 191             LLVMValueRef *pmask,
 192             LLVMValueRef *color,
 193             LLVMValueRef depth_ptr)
 194 {
 195    const struct tgsi_token *tokens = shader->base.tokens;
 196    LLVMTypeRef elem_type;
 197    LLVMTypeRef vec_type;
 198    LLVMTypeRef int_vec_type;
 199    LLVMValueRef consts_ptr;
 200    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
 201    LLVMValueRef z = interp->pos[2];
 202    struct lp_build_flow_context *flow;
 203    struct lp_build_mask_context mask;
 204    boolean early_depth_test;
 205    unsigned attrib;
 206    unsigned chan;
 207
 208    elem_type = lp_build_elem_type(type);
 209    vec_type = lp_build_vec_type(type);
 210    int_vec_type = lp_build_int_vec_type(type);
 211
 212    consts_ptr = lp_jit_context_constants(builder, context_ptr);
 213
 214    flow = lp_build_flow_create(builder);
 215
 216    memset(outputs, 0, sizeof outputs);
 217
 218    lp_build_flow_scope_begin(flow);
 219
 220    /* Declare the color and z variables */
 221    for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 222       color[chan] = LLVMGetUndef(vec_type);
 223       lp_build_flow_scope_declare(flow, &color[chan]);
 224    }
 225    lp_build_flow_scope_declare(flow, &z);
 226
 227    lp_build_mask_begin(&mask, flow, type, *pmask);
 228
 229    early_depth_test =
 230       key->depth.enabled &&
 231       !key->alpha.enabled &&
 232       !shader->info.uses_kill &&
 233       !shader->info.writes_z;
 234
 235    if(early_depth_test)
 236       generate_depth(builder, key,
 237                      type, &mask,
 238                      z, depth_ptr);
 239
 240    lp_build_tgsi_soa(builder, tokens, type, &mask,
 241                      consts_ptr, interp->pos, interp->inputs,
 242                      outputs, sampler);
 243
 244    for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
 245       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 246          if(outputs[attrib][chan]) {
 247             lp_build_name(outputs[attrib][chan], "output%u.%u.%c", i, attrib, "xyzw"[chan]);
 248
 249             switch (shader->info.output_semantic_name[attrib]) {
 250             case TGSI_SEMANTIC_COLOR:
 251                {
 252                   unsigned cbuf = shader->info.output_semantic_index[attrib];
 253
 254                   lp_build_name(outputs[attrib][chan], "color%u.%u.%c", i, attrib, "rgba"[chan]);
 255
 256                   /* Alpha test */
 257                   /* XXX: should the alpha reference value be passed separately? */
 258                   if(cbuf == 0 && chan == 3) {
 259                      LLVMValueRef alpha = outputs[attrib][chan];
 260                      LLVMValueRef alpha_ref_value;
 261                      alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
 262                      alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
 263                      lp_build_alpha_test(builder, &key->alpha, type,
 264                                          &mask, alpha, alpha_ref_value);
 265                   }
 266
 267                   if(cbuf == 0)
 268                      color[chan] = outputs[attrib][chan];
 269
 270                   break;
 271                }
 272
 273             case TGSI_SEMANTIC_POSITION:
 274                if(chan == 2)
 275                   z = outputs[attrib][chan];
 276                break;
 277             }
 278          }
 279       }
 280    }
 281
 282    if(!early_depth_test)
 283       generate_depth(builder, key,
 284                      type, &mask,
 285                      z, depth_ptr);
 286
 287    lp_build_mask_end(&mask);
 288
 289    lp_build_flow_scope_end(flow);
 290
 291    lp_build_flow_destroy(flow);
 292
 293    *pmask = mask.value;
 294
 295 }
 296
 297
 298 /**
 299  * Generate color blending and color output.
 300  */
 301 static void
 302 generate_blend(const struct pipe_blend_state *blend,
 303                LLVMBuilderRef builder,
 304                struct lp_type type,
 305                LLVMValueRef context_ptr,
 306                LLVMValueRef mask,
 307                LLVMValueRef *src,
 308                LLVMValueRef dst_ptr)
 309 {
 310    struct lp_build_context bld;
 311    struct lp_build_flow_context *flow;
 312    struct lp_build_mask_context mask_ctx;
 313    LLVMTypeRef vec_type;
 314    LLVMTypeRef int_vec_type;
 315    LLVMValueRef const_ptr;
 316    LLVMValueRef con[4];
 317    LLVMValueRef dst[4];
 318    LLVMValueRef res[4];
 319    unsigned chan;
 320
 321    lp_build_context_init(&bld, builder, type);
 322
 323    flow = lp_build_flow_create(builder);
 324    lp_build_mask_begin(&mask_ctx, flow, type, mask);
 325
 326    vec_type = lp_build_vec_type(type);
 327    int_vec_type = lp_build_int_vec_type(type);
 328
 329    const_ptr = lp_jit_context_blend_color(builder, context_ptr);
 330    const_ptr = LLVMBuildBitCast(builder, const_ptr,
 331                                 LLVMPointerType(vec_type, 0), "");
 332
 333    for(chan = 0; chan < 4; ++chan) {
 334       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 335       con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
 336
 337       dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
 338
 339       lp_build_name(con[chan], "con.%c", "rgba"[chan]);
 340       lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
 341    }
 342
 343    lp_build_blend_soa(builder, blend, type, src, dst, con, res);
 344
 345    for(chan = 0; chan < 4; ++chan) {
 346       if(blend->colormask & (1 << chan)) {
 347          LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 348          lp_build_name(res[chan], "res.%c", "rgba"[chan]);
 349          res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
 350          LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
 351       }
 352    }
 353
 354    lp_build_mask_end(&mask_ctx);
 355    lp_build_flow_destroy(flow);
 356 }
 357
 358
 359 /**
 360  * Generate the runtime callable function for the whole fragment pipeline.
 361  * Note that the function which we generate operates on a block of 16
 362  * pixels at at time.  The block contains 2x2 quads.  Each quad contains
 363  * 2x2 pixels.
 364  */
 365 static struct lp_fragment_shader_variant *
 366 generate_fragment(struct llvmpipe_context *lp,
 367                   struct lp_fragment_shader *shader,
 368                   const struct lp_fragment_shader_variant_key *key)
 369 {
 370    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
 371    struct lp_fragment_shader_variant *variant;
 372    struct lp_type fs_type;
 373    struct lp_type blend_type;
 374    LLVMTypeRef fs_elem_type;
 375    LLVMTypeRef fs_vec_type;
 376    LLVMTypeRef fs_int_vec_type;
 377    LLVMTypeRef blend_vec_type;
 378    LLVMTypeRef blend_int_vec_type;
 379    LLVMTypeRef arg_types[9];
 380    LLVMTypeRef func_type;
 381    LLVMValueRef context_ptr;
 382    LLVMValueRef x;
 383    LLVMValueRef y;
 384    LLVMValueRef a0_ptr;
 385    LLVMValueRef dadx_ptr;
 386    LLVMValueRef dady_ptr;
 387    LLVMValueRef mask_ptr;
 388    LLVMValueRef color_ptr;
 389    LLVMValueRef depth_ptr;
 390    LLVMBasicBlockRef block;
 391    LLVMBuilderRef builder;
 392    LLVMValueRef x0;
 393    LLVMValueRef y0;
 394    struct lp_build_sampler_soa *sampler;
 395    struct lp_build_interp_soa_context interp;
 396    LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
 397    LLVMValueRef fs_out_color[NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
 398    LLVMValueRef blend_mask;
 399    LLVMValueRef blend_in_color[NUM_CHANNELS];
 400    unsigned num_fs;
 401    unsigned i;
 402    unsigned chan;
 403
 404    if (LP_DEBUG & DEBUG_JIT) {
 405       tgsi_dump(shader->base.tokens, 0);
 406       if(key->depth.enabled) {
 407          debug_printf("depth.func = %s\n", debug_dump_func(key->depth.func, TRUE));
 408          debug_printf("depth.writemask = %u\n", key->depth.writemask);
 409       }
 410       if(key->alpha.enabled) {
 411          debug_printf("alpha.func = %s\n", debug_dump_func(key->alpha.func, TRUE));
 412          debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
 413       }
 414       if(key->blend.logicop_enable) {
 415          debug_printf("blend.logicop_func = %u\n", key->blend.logicop_func);
 416       }
 417       else if(key->blend.blend_enable) {
 418          debug_printf("blend.rgb_func = %s\n",   debug_dump_blend_func  (key->blend.rgb_func, TRUE));
 419          debug_printf("blend.rgb_src_factor = %s\n",   debug_dump_blend_factor(key->blend.rgb_src_factor, TRUE));
 420          debug_printf("blend.rgb_dst_factor = %s\n",   debug_dump_blend_factor(key->blend.rgb_dst_factor, TRUE));
 421          debug_printf("blend.alpha_func = %s\n",       debug_dump_blend_func  (key->blend.alpha_func, TRUE));
 422          debug_printf("blend.alpha_src_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_src_factor, TRUE));
 423          debug_printf("blend.alpha_dst_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_dst_factor, TRUE));
 424       }
 425       debug_printf("blend.colormask = 0x%x\n", key->blend.colormask);
 426    }
 427
 428    variant = CALLOC_STRUCT(lp_fragment_shader_variant);
 429    if(!variant)
 430       return NULL;
 431
 432    variant->shader = shader;
 433    memcpy(&variant->key, key, sizeof *key);
 434
 435    /* TODO: actually pick these based on the fs and color buffer
 436     * characteristics. */
 437
 438    memset(&fs_type, 0, sizeof fs_type);
 439    fs_type.floating = TRUE; /* floating point values */
 440    fs_type.sign = TRUE;     /* values are signed */
 441    fs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
 442    fs_type.width = 32;      /* 32-bit float */
 443    fs_type.length = 4;      /* 4 elements per vector */
 444    num_fs = 4;              /* number of quads per block */
 445
 446    memset(&blend_type, 0, sizeof blend_type);
 447    blend_type.floating = FALSE; /* values are integers */
 448    blend_type.sign = FALSE;     /* values are unsigned */
 449    blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
 450    blend_type.width = 8;        /* 8-bit ubyte values */
 451    blend_type.length = 16;      /* 16 elements per vector */
 452
 453    /*
 454     * Generate the function prototype. Any change here must be reflected in
 455     * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
 456     */
 457
 458    fs_elem_type = lp_build_elem_type(fs_type);
 459    fs_vec_type = lp_build_vec_type(fs_type);
 460    fs_int_vec_type = lp_build_int_vec_type(fs_type);
 461
 462    blend_vec_type = lp_build_vec_type(blend_type);
 463    blend_int_vec_type = lp_build_int_vec_type(blend_type);
 464
 465    arg_types[0] = screen->context_ptr_type;            /* context */
 466    arg_types[1] = LLVMInt32Type();                     /* x */
 467    arg_types[2] = LLVMInt32Type();                     /* y */
 468    arg_types[3] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
 469    arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
 470    arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* dady */
 471    arg_types[6] = LLVMPointerType(fs_int_vec_type, 0); /* mask */
 472    arg_types[7] = LLVMPointerType(blend_vec_type, 0);  /* color */
 473    arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
 474
 475    func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
 476
 477    variant->function = LLVMAddFunction(screen->module, "shader", func_type);
 478    LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
 479    for(i = 0; i < Elements(arg_types); ++i)
 480       if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
 481          LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute);
 482
 483    context_ptr  = LLVMGetParam(variant->function, 0);
 484    x            = LLVMGetParam(variant->function, 1);
 485    y            = LLVMGetParam(variant->function, 2);
 486    a0_ptr       = LLVMGetParam(variant->function, 3);
 487    dadx_ptr     = LLVMGetParam(variant->function, 4);
 488    dady_ptr     = LLVMGetParam(variant->function, 5);
 489    mask_ptr     = LLVMGetParam(variant->function, 6);
 490    color_ptr    = LLVMGetParam(variant->function, 7);
 491    depth_ptr    = LLVMGetParam(variant->function, 8);
 492
 493    lp_build_name(context_ptr, "context");
 494    lp_build_name(x, "x");
 495    lp_build_name(y, "y");
 496    lp_build_name(a0_ptr, "a0");
 497    lp_build_name(dadx_ptr, "dadx");
 498    lp_build_name(dady_ptr, "dady");
 499    lp_build_name(mask_ptr, "mask");
 500    lp_build_name(color_ptr, "color");
 501    lp_build_name(depth_ptr, "depth");
 502
 503    /*
 504     * Function body
 505     */
 506
 507    block = LLVMAppendBasicBlock(variant->function, "entry");
 508    builder = LLVMCreateBuilder();
 509    LLVMPositionBuilderAtEnd(builder, block);
 510
 511    generate_pos0(builder, x, y, &x0, &y0);
 512
 513    lp_build_interp_soa_init(&interp, shader->base.tokens, builder, fs_type,
 514                             a0_ptr, dadx_ptr, dady_ptr,
 515                             x0, y0);
 516
 517    /* code generated texture sampling */
 518    sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
 519
 520    /* loop over quads in the block */
 521    for(i = 0; i < num_fs; ++i) {
 522       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
 523       LLVMValueRef out_color[NUM_CHANNELS];
 524       LLVMValueRef depth_ptr_i;
 525
 526       if(i != 0)
 527          lp_build_interp_soa_update(&interp, i);
 528
 529       fs_mask[i] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, mask_ptr, &index, 1, ""), "");
 530       depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
 531
 532       generate_fs(lp, shader, key,
 533                   builder,
 534                   fs_type,
 535                   context_ptr,
 536                   i,
 537                   &interp,
 538                   sampler,
 539                   &fs_mask[i],
 540                   out_color,
 541                   depth_ptr_i);
 542
 543       for(chan = 0; chan < NUM_CHANNELS; ++chan)
 544          fs_out_color[chan][i] = out_color[chan];
 545    }
 546
 547    sampler->destroy(sampler);
 548
 549    /*
 550     * Convert the fs's output color and mask to fit to the blending type.
 551     */
 552
 553    for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 554       lp_build_conv(builder, fs_type, blend_type,
 555                     fs_out_color[chan], num_fs,
 556                     &blend_in_color[chan], 1);
 557       lp_build_name(blend_in_color[chan], "color.%c", "rgba"[chan]);
 558
 559    }
 560
 561    lp_build_conv_mask(builder, fs_type, blend_type,
 562                       fs_mask, num_fs,
 563                       &blend_mask, 1);
 564
 565    /*
 566     * Blending.
 567     */
 568
 569    generate_blend(&key->blend,
 570                   builder,
 571                   blend_type,
 572                   context_ptr,
 573                   blend_mask,
 574                   blend_in_color,
 575                   color_ptr);
 576
 577    LLVMBuildRetVoid(builder);
 578
 579    LLVMDisposeBuilder(builder);
 580
 581    /*
 582     * Translate the LLVM IR into machine code.
 583     */
 584
 585    if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) {
 586       LLVMDumpValue(variant->function);
 587       abort();
 588    }
 589
 590    LLVMRunFunctionPassManager(screen->pass, variant->function);
 591
 592    if (LP_DEBUG & DEBUG_JIT) {
 593       LLVMDumpValue(variant->function);
 594       debug_printf("\n");
 595    }
 596
 597    variant->jit_function = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, variant->function);
 598
 599    if (LP_DEBUG & DEBUG_ASM)
 600       lp_disassemble(variant->jit_function);
 601
 602    variant->next = shader->variants;
 603    shader->variants = variant;
 604
 605    return variant;
 606 }
 607
 608
 609 void *
 610 llvmpipe_create_fs_state(struct pipe_context *pipe,
 611                          const struct pipe_shader_state *templ)
 612 {
 613    struct lp_fragment_shader *shader;
 614
 615    shader = CALLOC_STRUCT(lp_fragment_shader);
 616    if (!shader)
 617       return NULL;
 618
 619    /* get/save the summary info for this shader */
 620    tgsi_scan_shader(templ->tokens, &shader->info);
 621
 622    /* we need to keep a local copy of the tokens */
 623    shader->base.tokens = tgsi_dup_tokens(templ->tokens);
 624
 625    return shader;
 626 }
 627
 628
 629 void
 630 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
 631 {
 632    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 633
 634    llvmpipe->fs = (struct lp_fragment_shader *) fs;
 635
 636    llvmpipe->dirty |= LP_NEW_FS;
 637 }
 638
 639
 640 void
 641 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
 642 {
 643    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 644    struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
 645    struct lp_fragment_shader *shader = fs;
 646    struct lp_fragment_shader_variant *variant;
 647
 648    assert(fs != llvmpipe->fs);
 649
 650    variant = shader->variants;
 651    while(variant) {
 652       struct lp_fragment_shader_variant *next = variant->next;
 653
 654       if(variant->function) {
 655          if(variant->jit_function)
 656             LLVMFreeMachineCodeForFunction(screen->engine, variant->function);
 657          LLVMDeleteFunction(variant->function);
 658       }
 659
 660       FREE(variant);
 661
 662       variant = next;
 663    }
 664
 665    FREE((void *) shader->base.tokens);
 666    FREE(shader);
 667 }
 668
 669
 670
 671 void
 672 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
 673                              uint shader, uint index,
 674                              const struct pipe_constant_buffer *constants)
 675 {
 676    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 677    struct pipe_buffer *buffer = constants ? constants->buffer : NULL;
 678    unsigned size = buffer ? buffer->size : 0;
 679    const void *data = buffer ? llvmpipe_buffer(buffer)->data : NULL;
 680
 681    assert(shader < PIPE_SHADER_TYPES);
 682    assert(index == 0);
 683
 684    if(llvmpipe->constants[shader].buffer == buffer)
 685       return;
 686
 687    if(shader == PIPE_SHADER_VERTEX)
 688       draw_flush(llvmpipe->draw);
 689
 690    /* note: reference counting */
 691    pipe_buffer_reference(&llvmpipe->constants[shader].buffer, buffer);
 692
 693    if(shader == PIPE_SHADER_VERTEX) {
 694       draw_set_mapped_constant_buffer(llvmpipe->draw, data, size);
 695    }
 696
 697    llvmpipe->dirty |= LP_NEW_CONSTANTS;
 698 }
 699
 700
 701 /**
 702  * We need to generate several variants of the fragment pipeline to match
 703  * all the combinations of the contributing state atoms.
 704  *
 705  * TODO: there is actually no reason to tie this to context state -- the
 706  * generated code could be cached globally in the screen.
 707  */
 708 static void
 709 make_variant_key(struct llvmpipe_context *lp,
 710                  struct lp_fragment_shader *shader,
 711                  struct lp_fragment_shader_variant_key *key)
 712 {
 713    unsigned i;
 714
 715    memset(key, 0, sizeof *key);
 716
 717    if(lp->framebuffer.zsbuf &&
 718       lp->depth_stencil->depth.enabled) {
 719       key->zsbuf_format = lp->framebuffer.zsbuf->format;
 720       memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
 721    }
 722
 723    key->alpha.enabled = lp->depth_stencil->alpha.enabled;
 724    if(key->alpha.enabled)
 725       key->alpha.func = lp->depth_stencil->alpha.func;
 726    /* alpha.ref_value is passed in jit_context */
 727
 728    if(lp->framebuffer.cbufs[0]) {
 729       const struct util_format_description *format_desc;
 730       unsigned chan;
 731
 732       memcpy(&key->blend, lp->blend, sizeof key->blend);
 733
 734       format_desc = util_format_description(lp->framebuffer.cbufs[0]->format);
 735       assert(format_desc->layout == UTIL_FORMAT_COLORSPACE_RGB ||
 736              format_desc->layout == UTIL_FORMAT_COLORSPACE_SRGB);
 737
 738       /* mask out color channels not present in the color buffer */
 739       for(chan = 0; chan < 4; ++chan) {
 740          enum util_format_swizzle swizzle = format_desc->swizzle[chan];
 741          if(swizzle > 4)
 742             key->blend.colormask &= ~(1 << chan);
 743       }
 744    }
 745
 746    for(i = 0; i < PIPE_MAX_SAMPLERS; ++i)
 747       if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i))
 748          lp_sampler_static_state(&key->sampler[i], lp->texture[i], lp->sampler[i]);
 749 }
 750
 751
 752 void
 753 llvmpipe_update_fs(struct llvmpipe_context *lp)
 754 {
 755    struct lp_fragment_shader *shader = lp->fs;
 756    struct lp_fragment_shader_variant_key key;
 757    struct lp_fragment_shader_variant *variant;
 758
 759    make_variant_key(lp, shader, &key);
 760
 761    variant = shader->variants;
 762    while(variant) {
 763       if(memcmp(&variant->key, &key, sizeof key) == 0)
 764          break;
 765
 766       variant = variant->next;
 767    }
 768
 769    if(!variant)
 770       variant = generate_fragment(lp, shader, &key);
 771
 772    shader->current = variant;
 773
 774    lp_setup_set_fs(lp->setup, shader);
 775 }