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