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