src/gallium/drivers/llvmpipe/lp_state_fs.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2009 VMware, Inc.
   4  * Copyright 2007 Tungsten Graphics, Inc., Cedar Park, Texas.
   5  * All Rights Reserved.
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a
   8  * copy of this software and associated documentation files (the
   9  * "Software"), to deal in the Software without restriction, including
  10  * without limitation the rights to use, copy, modify, merge, publish,
  11  * distribute, sub license, and/or sell copies of the Software, and to
  12  * permit persons to whom the Software is furnished to do so, subject to
  13  * the following conditions:
  14  *
  15  * The above copyright notice and this permission notice (including the
  16  * next paragraph) shall be included in all copies or substantial portions
  17  * of the Software.
  18  *
  19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  20  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  21  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  22  * IN NO EVENT SHALL TUNGSTEN GRAPHICS AND/OR ITS SUPPLIERS BE LIABLE FOR
  23  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  24  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  25  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  26  *
  27  **************************************************************************/
  28
  29 /**
  30  * @file
  31  * Code generate the whole fragment pipeline.
  32  *
  33  * The fragment pipeline consists of the following stages:
  34  * - stipple (TBI)
  35  * - early depth test
  36  * - fragment shader
  37  * - alpha test
  38  * - depth/stencil test (stencil TBI)
  39  * - blending
  40  *
  41  * This file has only the glue to assembly the fragment pipeline.  The actual
  42  * plumbing of converting Gallium state into LLVM IR is done elsewhere, in the
  43  * lp_bld_*.[ch] files, and in a complete generic and reusable way. Here we
  44  * muster the LLVM JIT execution engine to create a function that follows an
  45  * established binary interface and that can be called from C directly.
  46  *
  47  * A big source of complexity here is that we often want to run different
  48  * stages with different precisions and data types and precisions. For example,
  49  * the fragment shader needs typically to be done in floats, but the
  50  * depth/stencil test and blending is better done in the type that most closely
  51  * matches the depth/stencil and color buffer respectively.
  52  *
  53  * Since the width of a SIMD vector register stays the same regardless of the
  54  * element type, different types imply different number of elements, so we must
  55  * code generate more instances of the stages with larger types to be able to
  56  * feed/consume the stages with smaller types.
  57  *
  58  * @author Jose Fonseca <jfonseca@vmware.com>
  59  */
  60
  61 #include "pipe/p_defines.h"
  62 #include "util/u_memory.h"
  63 #include "util/u_format.h"
  64 #include "util/u_debug_dump.h"
  65 #include "pipe/p_shader_tokens.h"
  66 #include "draw/draw_context.h"
  67 #include "tgsi/tgsi_dump.h"
  68 #include "tgsi/tgsi_scan.h"
  69 #include "tgsi/tgsi_parse.h"
  70 #include "lp_bld_type.h"
  71 #include "lp_bld_const.h"
  72 #include "lp_bld_conv.h"
  73 #include "lp_bld_intr.h"
  74 #include "lp_bld_logic.h"
  75 #include "lp_bld_depth.h"
  76 #include "lp_bld_interp.h"
  77 #include "lp_bld_tgsi.h"
  78 #include "lp_bld_alpha.h"
  79 #include "lp_bld_blend.h"
  80 #include "lp_bld_swizzle.h"
  81 #include "lp_bld_flow.h"
  82 #include "lp_bld_debug.h"
  83 #include "lp_screen.h"
  84 #include "lp_context.h"
  85 #include "lp_buffer.h"
  86 #include "lp_state.h"
  87 #include "lp_tex_sample.h"
  88
  89
  90 static const unsigned char quad_offset_x[4] = {0, 1, 0, 1};
  91 static const unsigned char quad_offset_y[4] = {0, 0, 1, 1};
  92
  93
  94 /*
  95  * Derive from the quad's upper left scalar coordinates the coordinates for
  96  * all other quad pixels
  97  */
  98 static void
  99 generate_pos0(LLVMBuilderRef builder,
 100               LLVMValueRef x,
 101               LLVMValueRef y,
 102               LLVMValueRef *x0,
 103               LLVMValueRef *y0)
 104 {
 105    LLVMTypeRef int_elem_type = LLVMInt32Type();
 106    LLVMTypeRef int_vec_type = LLVMVectorType(int_elem_type, QUAD_SIZE);
 107    LLVMTypeRef elem_type = LLVMFloatType();
 108    LLVMTypeRef vec_type = LLVMVectorType(elem_type, QUAD_SIZE);
 109    LLVMValueRef x_offsets[QUAD_SIZE];
 110    LLVMValueRef y_offsets[QUAD_SIZE];
 111    unsigned i;
 112
 113    x = lp_build_broadcast(builder, int_vec_type, x);
 114    y = lp_build_broadcast(builder, int_vec_type, y);
 115
 116    for(i = 0; i < QUAD_SIZE; ++i) {
 117       x_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_x[i], 0);
 118       y_offsets[i] = LLVMConstInt(int_elem_type, quad_offset_y[i], 0);
 119    }
 120
 121    x = LLVMBuildAdd(builder, x, LLVMConstVector(x_offsets, QUAD_SIZE), "");
 122    y = LLVMBuildAdd(builder, y, LLVMConstVector(y_offsets, QUAD_SIZE), "");
 123
 124    *x0 = LLVMBuildSIToFP(builder, x, vec_type, "");
 125    *y0 = LLVMBuildSIToFP(builder, y, vec_type, "");
 126 }
 127
 128
 129 /**
 130  * Generate the depth test.
 131  */
 132 static void
 133 generate_depth(LLVMBuilderRef builder,
 134                const struct lp_fragment_shader_variant_key *key,
 135                struct lp_type src_type,
 136                struct lp_build_mask_context *mask,
 137                LLVMValueRef src,
 138                LLVMValueRef dst_ptr)
 139 {
 140    const struct util_format_description *format_desc;
 141    struct lp_type dst_type;
 142
 143    if(!key->depth.enabled)
 144       return;
 145
 146    format_desc = util_format_description(key->zsbuf_format);
 147    assert(format_desc);
 148
 149    /* Pick the depth type. */
 150    dst_type = lp_depth_type(format_desc, src_type.width*src_type.length);
 151
 152    /* FIXME: Cope with a depth test type with a different bit width. */
 153    assert(dst_type.width == src_type.width);
 154    assert(dst_type.length == src_type.length);
 155
 156 #if 1
 157    src = lp_build_clamped_float_to_unsigned_norm(builder,
 158                                                  src_type,
 159                                                  dst_type.width,
 160                                                  src);
 161 #else
 162    lp_build_conv(builder, src_type, dst_type, &src, 1, &src, 1);
 163 #endif
 164
 165    lp_build_depth_test(builder,
 166                        &key->depth,
 167                        dst_type,
 168                        format_desc,
 169                        mask,
 170                        src,
 171                        dst_ptr);
 172 }
 173
 174
 175 /**
 176  * Generate the fragment shader, depth/stencil test, and alpha tests.
 177  */
 178 static void
 179 generate_fs(struct llvmpipe_context *lp,
 180             struct lp_fragment_shader *shader,
 181             const struct lp_fragment_shader_variant_key *key,
 182             LLVMBuilderRef builder,
 183             struct lp_type type,
 184             LLVMValueRef context_ptr,
 185             unsigned i,
 186             const struct lp_build_interp_soa_context *interp,
 187             struct lp_build_sampler_soa *sampler,
 188             LLVMValueRef *pmask,
 189             LLVMValueRef *color,
 190             LLVMValueRef depth_ptr)
 191 {
 192    const struct tgsi_token *tokens = shader->base.tokens;
 193    LLVMTypeRef elem_type;
 194    LLVMTypeRef vec_type;
 195    LLVMTypeRef int_vec_type;
 196    LLVMValueRef consts_ptr;
 197    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][NUM_CHANNELS];
 198    LLVMValueRef z = interp->pos[2];
 199    struct lp_build_flow_context *flow;
 200    struct lp_build_mask_context mask;
 201    boolean early_depth_test;
 202    unsigned attrib;
 203    unsigned chan;
 204
 205    elem_type = lp_build_elem_type(type);
 206    vec_type = lp_build_vec_type(type);
 207    int_vec_type = lp_build_int_vec_type(type);
 208
 209    consts_ptr = lp_jit_context_constants(builder, context_ptr);
 210
 211    flow = lp_build_flow_create(builder);
 212
 213    memset(outputs, 0, sizeof outputs);
 214
 215    lp_build_flow_scope_begin(flow);
 216
 217    /* Declare the color and z variables */
 218    for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 219       color[chan] = LLVMGetUndef(vec_type);
 220       lp_build_flow_scope_declare(flow, &color[chan]);
 221    }
 222    lp_build_flow_scope_declare(flow, &z);
 223
 224    lp_build_mask_begin(&mask, flow, type, *pmask);
 225
 226    early_depth_test =
 227       key->depth.enabled &&
 228       !key->alpha.enabled &&
 229       !shader->info.uses_kill &&
 230       !shader->info.writes_z;
 231
 232    if(early_depth_test)
 233       generate_depth(builder, key,
 234                      type, &mask,
 235                      z, depth_ptr);
 236
 237    lp_build_tgsi_soa(builder, tokens, type, &mask,
 238                      consts_ptr, interp->pos, interp->inputs,
 239                      outputs, sampler);
 240
 241    for (attrib = 0; attrib < shader->info.num_outputs; ++attrib) {
 242       for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 243          if(outputs[attrib][chan]) {
 244             lp_build_name(outputs[attrib][chan], "output%u.%u.%c", i, attrib, "xyzw"[chan]);
 245
 246             switch (shader->info.output_semantic_name[attrib]) {
 247             case TGSI_SEMANTIC_COLOR:
 248                {
 249                   unsigned cbuf = shader->info.output_semantic_index[attrib];
 250
 251                   lp_build_name(outputs[attrib][chan], "color%u.%u.%c", i, attrib, "rgba"[chan]);
 252
 253                   /* Alpha test */
 254                   /* XXX: should the alpha reference value be passed separately? */
 255                   if(cbuf == 0 && chan == 3) {
 256                      LLVMValueRef alpha = outputs[attrib][chan];
 257                      LLVMValueRef alpha_ref_value;
 258                      alpha_ref_value = lp_jit_context_alpha_ref_value(builder, context_ptr);
 259                      alpha_ref_value = lp_build_broadcast(builder, vec_type, alpha_ref_value);
 260                      lp_build_alpha_test(builder, &key->alpha, type,
 261                                          &mask, alpha, alpha_ref_value);
 262                   }
 263
 264                   if(cbuf == 0)
 265                      color[chan] = outputs[attrib][chan];
 266
 267                   break;
 268                }
 269
 270             case TGSI_SEMANTIC_POSITION:
 271                if(chan == 2)
 272                   z = outputs[attrib][chan];
 273                break;
 274             }
 275          }
 276       }
 277    }
 278
 279    if(!early_depth_test)
 280       generate_depth(builder, key,
 281                      type, &mask,
 282                      z, depth_ptr);
 283
 284    lp_build_mask_end(&mask);
 285
 286    lp_build_flow_scope_end(flow);
 287
 288    lp_build_flow_destroy(flow);
 289
 290    *pmask = mask.value;
 291
 292 }
 293
 294
 295 /**
 296  * Generate color blending and color output.
 297  */
 298 static void
 299 generate_blend(const struct pipe_blend_state *blend,
 300                LLVMBuilderRef builder,
 301                struct lp_type type,
 302                LLVMValueRef context_ptr,
 303                LLVMValueRef mask,
 304                LLVMValueRef *src,
 305                LLVMValueRef dst_ptr)
 306 {
 307    struct lp_build_context bld;
 308    struct lp_build_flow_context *flow;
 309    struct lp_build_mask_context mask_ctx;
 310    LLVMTypeRef vec_type;
 311    LLVMTypeRef int_vec_type;
 312    LLVMValueRef const_ptr;
 313    LLVMValueRef con[4];
 314    LLVMValueRef dst[4];
 315    LLVMValueRef res[4];
 316    unsigned chan;
 317
 318    lp_build_context_init(&bld, builder, type);
 319
 320    flow = lp_build_flow_create(builder);
 321    lp_build_mask_begin(&mask_ctx, flow, type, mask);
 322
 323    vec_type = lp_build_vec_type(type);
 324    int_vec_type = lp_build_int_vec_type(type);
 325
 326    const_ptr = lp_jit_context_blend_color(builder, context_ptr);
 327    const_ptr = LLVMBuildBitCast(builder, const_ptr,
 328                                 LLVMPointerType(vec_type, 0), "");
 329
 330    for(chan = 0; chan < 4; ++chan) {
 331       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 332       con[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, const_ptr, &index, 1, ""), "");
 333
 334       dst[chan] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, dst_ptr, &index, 1, ""), "");
 335
 336       lp_build_name(con[chan], "con.%c", "rgba"[chan]);
 337       lp_build_name(dst[chan], "dst.%c", "rgba"[chan]);
 338    }
 339
 340    lp_build_blend_soa(builder, blend, type, src, dst, con, res);
 341
 342    for(chan = 0; chan < 4; ++chan) {
 343       if(blend->colormask & (1 << chan)) {
 344          LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), chan, 0);
 345          lp_build_name(res[chan], "res.%c", "rgba"[chan]);
 346          res[chan] = lp_build_select(&bld, mask, res[chan], dst[chan]);
 347          LLVMBuildStore(builder, res[chan], LLVMBuildGEP(builder, dst_ptr, &index, 1, ""));
 348       }
 349    }
 350
 351    lp_build_mask_end(&mask_ctx);
 352    lp_build_flow_destroy(flow);
 353 }
 354
 355
 356 /**
 357  * Generate the runtime callable function for the whole fragment pipeline.
 358  */
 359 static struct lp_fragment_shader_variant *
 360 generate_fragment(struct llvmpipe_context *lp,
 361                   struct lp_fragment_shader *shader,
 362                   const struct lp_fragment_shader_variant_key *key)
 363 {
 364    struct llvmpipe_screen *screen = llvmpipe_screen(lp->pipe.screen);
 365    struct lp_fragment_shader_variant *variant;
 366    struct lp_type fs_type;
 367    struct lp_type blend_type;
 368    LLVMTypeRef fs_elem_type;
 369    LLVMTypeRef fs_vec_type;
 370    LLVMTypeRef fs_int_vec_type;
 371    LLVMTypeRef blend_vec_type;
 372    LLVMTypeRef blend_int_vec_type;
 373    LLVMTypeRef arg_types[9];
 374    LLVMTypeRef func_type;
 375    LLVMValueRef context_ptr;
 376    LLVMValueRef x;
 377    LLVMValueRef y;
 378    LLVMValueRef a0_ptr;
 379    LLVMValueRef dadx_ptr;
 380    LLVMValueRef dady_ptr;
 381    LLVMValueRef mask_ptr;
 382    LLVMValueRef color_ptr;
 383    LLVMValueRef depth_ptr;
 384    LLVMBasicBlockRef block;
 385    LLVMBuilderRef builder;
 386    LLVMValueRef x0;
 387    LLVMValueRef y0;
 388    struct lp_build_sampler_soa *sampler;
 389    struct lp_build_interp_soa_context interp;
 390    LLVMValueRef fs_mask[LP_MAX_VECTOR_LENGTH];
 391    LLVMValueRef fs_out_color[NUM_CHANNELS][LP_MAX_VECTOR_LENGTH];
 392    LLVMValueRef blend_mask;
 393    LLVMValueRef blend_in_color[NUM_CHANNELS];
 394    unsigned num_fs;
 395    unsigned i;
 396    unsigned chan;
 397
 398 #ifdef DEBUG
 399    tgsi_dump(shader->base.tokens, 0);
 400    if(key->depth.enabled) {
 401       debug_printf("depth.format = %s\n", pf_name(key->zsbuf_format));
 402       debug_printf("depth.func = %s\n", debug_dump_func(key->depth.func, TRUE));
 403       debug_printf("depth.writemask = %u\n", key->depth.writemask);
 404    }
 405    if(key->alpha.enabled) {
 406       debug_printf("alpha.func = %s\n", debug_dump_func(key->alpha.func, TRUE));
 407       debug_printf("alpha.ref_value = %f\n", key->alpha.ref_value);
 408    }
 409    if(key->blend.logicop_enable) {
 410       debug_printf("blend.logicop_func = %u\n", key->blend.logicop_func);
 411    }
 412    else if(key->blend.blend_enable) {
 413       debug_printf("blend.rgb_func = %s\n",   debug_dump_blend_func  (key->blend.rgb_func, TRUE));
 414       debug_printf("rgb_src_factor = %s\n",   debug_dump_blend_factor(key->blend.rgb_src_factor, TRUE));
 415       debug_printf("rgb_dst_factor = %s\n",   debug_dump_blend_factor(key->blend.rgb_dst_factor, TRUE));
 416       debug_printf("alpha_func = %s\n",       debug_dump_blend_func  (key->blend.alpha_func, TRUE));
 417       debug_printf("alpha_src_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_src_factor, TRUE));
 418       debug_printf("alpha_dst_factor = %s\n", debug_dump_blend_factor(key->blend.alpha_dst_factor, TRUE));
 419    }
 420    debug_printf("blend.colormask = 0x%x\n", key->blend.colormask);
 421    for(i = 0; i < PIPE_MAX_SAMPLERS; ++i) {
 422       if(key->sampler[i].format) {
 423          debug_printf("sampler[%u] = \n", i);
 424          debug_printf("  .format = %s\n",
 425                       pf_name(key->sampler[i].format));
 426          debug_printf("  .target = %s\n",
 427                       debug_dump_tex_target(key->sampler[i].target, TRUE));
 428          debug_printf("  .pot = %u %u %u\n",
 429                       key->sampler[i].pot_width,
 430                       key->sampler[i].pot_height,
 431                       key->sampler[i].pot_depth);
 432          debug_printf("  .wrap = %s %s %s\n",
 433                       debug_dump_tex_wrap(key->sampler[i].wrap_s, TRUE),
 434                       debug_dump_tex_wrap(key->sampler[i].wrap_t, TRUE),
 435                       debug_dump_tex_wrap(key->sampler[i].wrap_r, TRUE));
 436          debug_printf("  .min_img_filter = %s\n",
 437                       debug_dump_tex_filter(key->sampler[i].min_img_filter, TRUE));
 438          debug_printf("  .min_mip_filter = %s\n",
 439                       debug_dump_tex_mipfilter(key->sampler[i].min_mip_filter, TRUE));
 440          debug_printf("  .mag_img_filter = %s\n",
 441                       debug_dump_tex_filter(key->sampler[i].mag_img_filter, TRUE));
 442          if(key->sampler[i].compare_mode)
 443             debug_printf("  .compare_mode = %s\n", debug_dump_func(key->sampler[i].compare_func, TRUE));
 444          debug_printf("  .normalized_coords = %u\n", key->sampler[i].normalized_coords);
 445          debug_printf("  .prefilter = %u\n", key->sampler[i].prefilter);
 446       }
 447    }
 448
 449 #endif
 450
 451    variant = CALLOC_STRUCT(lp_fragment_shader_variant);
 452    if(!variant)
 453       return NULL;
 454
 455    variant->shader = shader;
 456    memcpy(&variant->key, key, sizeof *key);
 457
 458    /* TODO: actually pick these based on the fs and color buffer
 459     * characteristics. */
 460
 461    memset(&fs_type, 0, sizeof fs_type);
 462    fs_type.floating = TRUE; /* floating point values */
 463    fs_type.sign = TRUE;     /* values are signed */
 464    fs_type.norm = FALSE;    /* values are not limited to [0,1] or [-1,1] */
 465    fs_type.width = 32;      /* 32-bit float */
 466    fs_type.length = 4;      /* 4 element per vector */
 467    num_fs = 4;
 468
 469    memset(&blend_type, 0, sizeof blend_type);
 470    blend_type.floating = FALSE; /* values are integers */
 471    blend_type.sign = FALSE;     /* values are unsigned */
 472    blend_type.norm = TRUE;      /* values are in [0,1] or [-1,1] */
 473    blend_type.width = 8;        /* 8-bit ubyte values */
 474    blend_type.length = 16;      /* 16 elements per vector */
 475
 476    /*
 477     * Generate the function prototype. Any change here must be reflected in
 478     * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
 479     */
 480
 481    fs_elem_type = lp_build_elem_type(fs_type);
 482    fs_vec_type = lp_build_vec_type(fs_type);
 483    fs_int_vec_type = lp_build_int_vec_type(fs_type);
 484
 485    blend_vec_type = lp_build_vec_type(blend_type);
 486    blend_int_vec_type = lp_build_int_vec_type(blend_type);
 487
 488    arg_types[0] = screen->context_ptr_type;            /* context */
 489    arg_types[1] = LLVMInt32Type();                     /* x */
 490    arg_types[2] = LLVMInt32Type();                     /* y */
 491    arg_types[3] = LLVMPointerType(fs_elem_type, 0);    /* a0 */
 492    arg_types[4] = LLVMPointerType(fs_elem_type, 0);    /* dadx */
 493    arg_types[5] = LLVMPointerType(fs_elem_type, 0);    /* dady */
 494    arg_types[6] = LLVMPointerType(fs_int_vec_type, 0); /* mask */
 495    arg_types[7] = LLVMPointerType(blend_vec_type, 0);  /* color */
 496    arg_types[8] = LLVMPointerType(fs_int_vec_type, 0); /* depth */
 497
 498    func_type = LLVMFunctionType(LLVMVoidType(), arg_types, Elements(arg_types), 0);
 499
 500    variant->function = LLVMAddFunction(screen->module, "shader", func_type);
 501    LLVMSetFunctionCallConv(variant->function, LLVMCCallConv);
 502    for(i = 0; i < Elements(arg_types); ++i)
 503       if(LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind)
 504          LLVMAddAttribute(LLVMGetParam(variant->function, i), LLVMNoAliasAttribute);
 505
 506    context_ptr  = LLVMGetParam(variant->function, 0);
 507    x            = LLVMGetParam(variant->function, 1);
 508    y            = LLVMGetParam(variant->function, 2);
 509    a0_ptr       = LLVMGetParam(variant->function, 3);
 510    dadx_ptr     = LLVMGetParam(variant->function, 4);
 511    dady_ptr     = LLVMGetParam(variant->function, 5);
 512    mask_ptr     = LLVMGetParam(variant->function, 6);
 513    color_ptr    = LLVMGetParam(variant->function, 7);
 514    depth_ptr    = LLVMGetParam(variant->function, 8);
 515
 516    lp_build_name(context_ptr, "context");
 517    lp_build_name(x, "x");
 518    lp_build_name(y, "y");
 519    lp_build_name(a0_ptr, "a0");
 520    lp_build_name(dadx_ptr, "dadx");
 521    lp_build_name(dady_ptr, "dady");
 522    lp_build_name(mask_ptr, "mask");
 523    lp_build_name(color_ptr, "color");
 524    lp_build_name(depth_ptr, "depth");
 525
 526    /*
 527     * Function body
 528     */
 529
 530    block = LLVMAppendBasicBlock(variant->function, "entry");
 531    builder = LLVMCreateBuilder();
 532    LLVMPositionBuilderAtEnd(builder, block);
 533
 534    generate_pos0(builder, x, y, &x0, &y0);
 535
 536    lp_build_interp_soa_init(&interp, shader->base.tokens, builder, fs_type,
 537                             a0_ptr, dadx_ptr, dady_ptr,
 538                             x0, y0, 2, 0);
 539
 540 #if 0
 541    /* C texture sampling */
 542    sampler = lp_c_sampler_soa_create(context_ptr);
 543 #else
 544    /* code generated texture sampling */
 545    sampler = lp_llvm_sampler_soa_create(key->sampler, context_ptr);
 546 #endif
 547
 548    for(i = 0; i < num_fs; ++i) {
 549       LLVMValueRef index = LLVMConstInt(LLVMInt32Type(), i, 0);
 550       LLVMValueRef out_color[NUM_CHANNELS];
 551       LLVMValueRef depth_ptr_i;
 552
 553       if(i != 0)
 554          lp_build_interp_soa_update(&interp);
 555
 556       fs_mask[i] = LLVMBuildLoad(builder, LLVMBuildGEP(builder, mask_ptr, &index, 1, ""), "");
 557       depth_ptr_i = LLVMBuildGEP(builder, depth_ptr, &index, 1, "");
 558
 559       generate_fs(lp, shader, key,
 560                   builder,
 561                   fs_type,
 562                   context_ptr,
 563                   i,
 564                   &interp,
 565                   sampler,
 566                   &fs_mask[i],
 567                   out_color,
 568                   depth_ptr_i);
 569
 570       for(chan = 0; chan < NUM_CHANNELS; ++chan)
 571          fs_out_color[chan][i] = out_color[chan];
 572    }
 573
 574    sampler->destroy(sampler);
 575
 576    /*
 577     * Convert the fs's output color and mask to fit to the blending type.
 578     */
 579
 580    for(chan = 0; chan < NUM_CHANNELS; ++chan) {
 581       lp_build_conv(builder, fs_type, blend_type,
 582                     fs_out_color[chan], num_fs,
 583                     &blend_in_color[chan], 1);
 584       lp_build_name(blend_in_color[chan], "color.%c", "rgba"[chan]);
 585
 586    }
 587
 588    lp_build_conv_mask(builder, fs_type, blend_type,
 589                                fs_mask, num_fs,
 590                                &blend_mask, 1);
 591
 592    /*
 593     * Blending.
 594     */
 595
 596    generate_blend(&key->blend,
 597                   builder,
 598                   blend_type,
 599                   context_ptr,
 600                   blend_mask,
 601                   blend_in_color,
 602                   color_ptr);
 603
 604    LLVMBuildRetVoid(builder);
 605
 606    LLVMDisposeBuilder(builder);
 607
 608    /*
 609     * Translate the LLVM IR into machine code.
 610     */
 611
 612    if(LLVMVerifyFunction(variant->function, LLVMPrintMessageAction)) {
 613       LLVMDumpValue(variant->function);
 614       abort();
 615    }
 616
 617    LLVMRunFunctionPassManager(screen->pass, variant->function);
 618
 619 #ifdef DEBUG
 620    LLVMDumpValue(variant->function);
 621    debug_printf("\n");
 622 #endif
 623
 624    variant->jit_function = (lp_jit_frag_func)LLVMGetPointerToGlobal(screen->engine, variant->function);
 625
 626 #ifdef DEBUG
 627    lp_disassemble(variant->jit_function);
 628 #endif
 629
 630    variant->next = shader->variants;
 631    shader->variants = variant;
 632
 633    return variant;
 634 }
 635
 636
 637 void *
 638 llvmpipe_create_fs_state(struct pipe_context *pipe,
 639                          const struct pipe_shader_state *templ)
 640 {
 641    struct lp_fragment_shader *shader;
 642
 643    shader = CALLOC_STRUCT(lp_fragment_shader);
 644    if (!shader)
 645       return NULL;
 646
 647    /* get/save the summary info for this shader */
 648    tgsi_scan_shader(templ->tokens, &shader->info);
 649
 650    /* we need to keep a local copy of the tokens */
 651    shader->base.tokens = tgsi_dup_tokens(templ->tokens);
 652
 653    return shader;
 654 }
 655
 656
 657 void
 658 llvmpipe_bind_fs_state(struct pipe_context *pipe, void *fs)
 659 {
 660    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 661
 662    llvmpipe->fs = (struct lp_fragment_shader *) fs;
 663
 664    llvmpipe->dirty |= LP_NEW_FS;
 665 }
 666
 667
 668 void
 669 llvmpipe_delete_fs_state(struct pipe_context *pipe, void *fs)
 670 {
 671    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 672    struct llvmpipe_screen *screen = llvmpipe_screen(pipe->screen);
 673    struct lp_fragment_shader *shader = fs;
 674    struct lp_fragment_shader_variant *variant;
 675
 676    assert(fs != llvmpipe->fs);
 677    (void) llvmpipe;
 678
 679    variant = shader->variants;
 680    while(variant) {
 681       struct lp_fragment_shader_variant *next = variant->next;
 682
 683       if(variant->function) {
 684          if(variant->jit_function)
 685             LLVMFreeMachineCodeForFunction(screen->engine, variant->function);
 686          LLVMDeleteFunction(variant->function);
 687       }
 688
 689       FREE(variant);
 690
 691       variant = next;
 692    }
 693
 694    FREE((void *) shader->base.tokens);
 695    FREE(shader);
 696 }
 697
 698
 699
 700 void
 701 llvmpipe_set_constant_buffer(struct pipe_context *pipe,
 702                              uint shader, uint index,
 703                              const struct pipe_constant_buffer *constants)
 704 {
 705    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 706    struct pipe_buffer *buffer = constants ? constants->buffer : NULL;
 707    unsigned size = buffer ? buffer->size : 0;
 708    const void *data = buffer ? llvmpipe_buffer(buffer)->data : NULL;
 709
 710    assert(shader < PIPE_SHADER_TYPES);
 711    assert(index == 0);
 712
 713    if(shader == PIPE_SHADER_VERTEX)
 714       draw_flush(llvmpipe->draw);
 715
 716    /* note: reference counting */
 717    pipe_buffer_reference(&llvmpipe->constants[shader].buffer, buffer);
 718
 719    if(shader == PIPE_SHADER_FRAGMENT) {
 720       llvmpipe->jit_context.constants = data;
 721    }
 722
 723    if(shader == PIPE_SHADER_VERTEX) {
 724       draw_set_mapped_constant_buffer(llvmpipe->draw, data, size);
 725    }
 726
 727    llvmpipe->dirty |= LP_NEW_CONSTANTS;
 728 }
 729
 730
 731 /**
 732  * We need to generate several variants of the fragment pipeline to match
 733  * all the combinations of the contributing state atoms.
 734  *
 735  * TODO: there is actually no reason to tie this to context state -- the
 736  * generated code could be cached globally in the screen.
 737  */
 738 static void
 739 make_variant_key(struct llvmpipe_context *lp,
 740                  struct lp_fragment_shader *shader,
 741                  struct lp_fragment_shader_variant_key *key)
 742 {
 743    unsigned i;
 744
 745    memset(key, 0, sizeof *key);
 746
 747    if(lp->framebuffer.zsbuf &&
 748       lp->depth_stencil->depth.enabled) {
 749       key->zsbuf_format = lp->framebuffer.zsbuf->format;
 750       memcpy(&key->depth, &lp->depth_stencil->depth, sizeof key->depth);
 751    }
 752
 753    key->alpha.enabled = lp->depth_stencil->alpha.enabled;
 754    if(key->alpha.enabled)
 755       key->alpha.func = lp->depth_stencil->alpha.func;
 756    /* alpha.ref_value is passed in jit_context */
 757
 758    if(lp->framebuffer.cbufs[0]) {
 759       const struct util_format_description *format_desc;
 760       unsigned chan;
 761
 762       memcpy(&key->blend, lp->blend, sizeof key->blend);
 763
 764       format_desc = util_format_description(lp->framebuffer.cbufs[0]->format);
 765       assert(format_desc->layout == UTIL_FORMAT_COLORSPACE_RGB ||
 766              format_desc->layout == UTIL_FORMAT_COLORSPACE_SRGB);
 767
 768       /* mask out color channels not present in the color buffer */
 769       for(chan = 0; chan < 4; ++chan) {
 770          enum util_format_swizzle swizzle = format_desc->swizzle[chan];
 771          if(swizzle > 4)
 772             key->blend.colormask &= ~(1 << chan);
 773       }
 774    }
 775
 776    for(i = 0; i < PIPE_MAX_SAMPLERS; ++i)
 777       if(shader->info.file_mask[TGSI_FILE_SAMPLER] & (1 << i))
 778          lp_sampler_static_state(&key->sampler[i], lp->texture[i], lp->sampler[i]);
 779 }
 780
 781
 782 void
 783 llvmpipe_update_fs(struct llvmpipe_context *lp)
 784 {
 785    struct lp_fragment_shader *shader = lp->fs;
 786    struct lp_fragment_shader_variant_key key;
 787    struct lp_fragment_shader_variant *variant;
 788
 789    make_variant_key(lp, shader, &key);
 790
 791    variant = shader->variants;
 792    while(variant) {
 793       if(memcmp(&variant->key, &key, sizeof key) == 0)
 794          break;
 795
 796       variant = variant->next;
 797    }
 798
 799    if(!variant)
 800       variant = generate_fragment(lp, shader, &key);
 801
 802    shader->current = variant;
 803 }