src/gallium/drivers/r300/r300_fs.c

   1 /*
   2  * Copyright 2008 Corbin Simpson <MostAwesomeDude@gmail.com>
   3  *                Joakim Sindholt <opensource@zhasha.com>
   4  * Copyright 2009 Marek Olšák <maraeo@gmail.com>
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * on the rights to use, copy, modify, merge, publish, distribute, sub
  10  * license, and/or sell copies of the Software, and to permit persons to whom
  11  * the Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice (including the next
  14  * paragraph) shall be included in all copies or substantial portions of the
  15  * Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  20  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  21  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  22  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  23  * USE OR OTHER DEALINGS IN THE SOFTWARE. */
  24
  25 #include "util/u_math.h"
  26 #include "util/u_memory.h"
  27
  28 #include "tgsi/tgsi_dump.h"
  29 #include "tgsi/tgsi_ureg.h"
  30
  31 #include "r300_cb.h"
  32 #include "r300_context.h"
  33 #include "r300_emit.h"
  34 #include "r300_screen.h"
  35 #include "r300_fs.h"
  36 #include "r300_reg.h"
  37 #include "r300_tgsi_to_rc.h"
  38
  39 #include "radeon_code.h"
  40 #include "radeon_compiler.h"
  41
  42 /* Convert info about FS input semantics to r300_shader_semantics. */
  43 void r300_shader_read_fs_inputs(struct tgsi_shader_info* info,
  44                                 struct r300_shader_semantics* fs_inputs)
  45 {
  46     int i;
  47     unsigned index;
  48
  49     r300_shader_semantics_reset(fs_inputs);
  50
  51     for (i = 0; i < info->num_inputs; i++) {
  52         index = info->input_semantic_index[i];
  53
  54         switch (info->input_semantic_name[i]) {
  55             case TGSI_SEMANTIC_COLOR:
  56                 assert(index < ATTR_COLOR_COUNT);
  57                 fs_inputs->color[index] = i;
  58                 break;
  59
  60             case TGSI_SEMANTIC_GENERIC:
  61                 assert(index < ATTR_GENERIC_COUNT);
  62                 fs_inputs->generic[index] = i;
  63                 break;
  64
  65             case TGSI_SEMANTIC_FOG:
  66                 assert(index == 0);
  67                 fs_inputs->fog = i;
  68                 break;
  69
  70             case TGSI_SEMANTIC_POSITION:
  71                 assert(index == 0);
  72                 fs_inputs->wpos = i;
  73                 break;
  74
  75             default:
  76                 fprintf(stderr, "r300: FP: Unknown input semantic: %i\n",
  77                         info->input_semantic_name[i]);
  78         }
  79     }
  80 }
  81
  82 static void find_output_registers(struct r300_fragment_program_compiler * compiler,
  83                                   struct r300_fragment_shader_code *shader)
  84 {
  85     unsigned i, colorbuf_count = 0;
  86
  87     /* Mark the outputs as not present initially */
  88     compiler->OutputColor[0] = shader->info.num_outputs;
  89     compiler->OutputColor[1] = shader->info.num_outputs;
  90     compiler->OutputColor[2] = shader->info.num_outputs;
  91     compiler->OutputColor[3] = shader->info.num_outputs;
  92     compiler->OutputDepth = shader->info.num_outputs;
  93
  94     /* Now see where they really are. */
  95     for(i = 0; i < shader->info.num_outputs; ++i) {
  96         switch(shader->info.output_semantic_name[i]) {
  97             case TGSI_SEMANTIC_COLOR:
  98                 compiler->OutputColor[colorbuf_count] = i;
  99                 colorbuf_count++;
 100                 break;
 101             case TGSI_SEMANTIC_POSITION:
 102                 compiler->OutputDepth = i;
 103                 break;
 104         }
 105     }
 106 }
 107
 108 static void allocate_hardware_inputs(
 109     struct r300_fragment_program_compiler * c,
 110     void (*allocate)(void * data, unsigned input, unsigned hwreg),
 111     void * mydata)
 112 {
 113     struct r300_shader_semantics* inputs =
 114         (struct r300_shader_semantics*)c->UserData;
 115     int i, reg = 0;
 116
 117     /* Allocate input registers. */
 118     for (i = 0; i < ATTR_COLOR_COUNT; i++) {
 119         if (inputs->color[i] != ATTR_UNUSED) {
 120             allocate(mydata, inputs->color[i], reg++);
 121         }
 122     }
 123     for (i = 0; i < ATTR_GENERIC_COUNT; i++) {
 124         if (inputs->generic[i] != ATTR_UNUSED) {
 125             allocate(mydata, inputs->generic[i], reg++);
 126         }
 127     }
 128     if (inputs->fog != ATTR_UNUSED) {
 129         allocate(mydata, inputs->fog, reg++);
 130     }
 131     if (inputs->wpos != ATTR_UNUSED) {
 132         allocate(mydata, inputs->wpos, reg++);
 133     }
 134 }
 135
 136 static void get_external_state(
 137     struct r300_context* r300,
 138     struct r300_fragment_program_external_state* state)
 139 {
 140     struct r300_textures_state *texstate = r300->textures_state.state;
 141     unsigned i;
 142     unsigned char *swizzle;
 143
 144     for (i = 0; i < texstate->sampler_state_count; i++) {
 145         struct r300_sampler_state* s = texstate->sampler_states[i];
 146
 147         if (!s) {
 148             continue;
 149         }
 150
 151         if (s->state.compare_mode == PIPE_TEX_COMPARE_R_TO_TEXTURE) {
 152             state->unit[i].compare_mode_enabled = 1;
 153
 154             /* Pass depth texture swizzling to the compiler. */
 155             if (texstate->sampler_views[i]) {
 156                 swizzle = texstate->sampler_views[i]->swizzle;
 157
 158                 state->unit[i].depth_texture_swizzle =
 159                     RC_MAKE_SWIZZLE(swizzle[0], swizzle[1],
 160                                     swizzle[2], swizzle[3]);
 161             } else {
 162                 state->unit[i].depth_texture_swizzle = RC_SWIZZLE_XYZW;
 163             }
 164
 165             /* Fortunately, no need to translate this. */
 166             state->unit[i].texture_compare_func = s->state.compare_func;
 167         }
 168
 169         state->unit[i].non_normalized_coords = !s->state.normalized_coords;
 170
 171         if (texstate->sampler_views[i]) {
 172             struct r300_texture *t;
 173             t = (struct r300_texture*)texstate->sampler_views[i]->base.texture;
 174
 175             /* XXX this should probably take into account STR, not just S. */
 176             if (t->desc.is_npot) {
 177                 switch (s->state.wrap_s) {
 178                     case PIPE_TEX_WRAP_REPEAT:
 179                         state->unit[i].wrap_mode = RC_WRAP_REPEAT;
 180                         state->unit[i].fake_npot = TRUE;
 181                         break;
 182
 183                     case PIPE_TEX_WRAP_MIRROR_REPEAT:
 184                         state->unit[i].wrap_mode = RC_WRAP_MIRRORED_REPEAT;
 185                         state->unit[i].fake_npot = TRUE;
 186                         break;
 187
 188                     case PIPE_TEX_WRAP_MIRROR_CLAMP:
 189                     case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_EDGE:
 190                     case PIPE_TEX_WRAP_MIRROR_CLAMP_TO_BORDER:
 191                         state->unit[i].wrap_mode = RC_WRAP_MIRRORED_CLAMP;
 192                         state->unit[i].fake_npot = TRUE;
 193                         break;
 194
 195                     default:
 196                         state->unit[i].wrap_mode = RC_WRAP_NONE;
 197                         break;
 198                 }
 199             }
 200         }
 201     }
 202 }
 203
 204 static void r300_translate_fragment_shader(
 205     struct r300_context* r300,
 206     struct r300_fragment_shader_code* shader,
 207     const struct tgsi_token *tokens);
 208
 209 static void r300_dummy_fragment_shader(
 210     struct r300_context* r300,
 211     struct r300_fragment_shader_code* shader)
 212 {
 213     struct pipe_shader_state state;
 214     struct ureg_program *ureg;
 215     struct ureg_dst out;
 216     struct ureg_src imm;
 217
 218     /* Make a simple fragment shader which outputs (0, 0, 0, 1) */
 219     ureg = ureg_create(TGSI_PROCESSOR_FRAGMENT);
 220     out = ureg_DECL_output(ureg, TGSI_SEMANTIC_COLOR, 0);
 221     imm = ureg_imm4f(ureg, 0, 0, 0, 1);
 222
 223     ureg_MOV(ureg, out, imm);
 224     ureg_END(ureg);
 225
 226     state.tokens = ureg_finalize(ureg);
 227
 228     shader->dummy = TRUE;
 229     r300_translate_fragment_shader(r300, shader, state.tokens);
 230
 231     ureg_destroy(ureg);
 232 }
 233
 234 static void r300_emit_fs_code_to_buffer(
 235     struct r300_context *r300,
 236     struct r300_fragment_shader_code *shader)
 237 {
 238     struct rX00_fragment_program_code *generic_code = &shader->code;
 239     unsigned imm_count = shader->immediates_count;
 240     unsigned imm_first = shader->externals_count;
 241     unsigned imm_end = generic_code->constants.Count;
 242     struct rc_constant *constants = generic_code->constants.Constants;
 243     unsigned i;
 244     CB_LOCALS;
 245
 246     if (r300->screen->caps.is_r500) {
 247         struct r500_fragment_program_code *code = &generic_code->code.r500;
 248
 249         shader->cb_code_size = 19 +
 250                                ((code->inst_end + 1) * 6) +
 251                                imm_count * 7;
 252
 253         NEW_CB(shader->cb_code, shader->cb_code_size);
 254         OUT_CB_REG(R500_US_CONFIG, R500_ZERO_TIMES_ANYTHING_EQUALS_ZERO);
 255         OUT_CB_REG(R500_US_PIXSIZE, code->max_temp_idx);
 256         OUT_CB_REG(R500_US_FC_CTRL, code->us_fc_ctrl);
 257         OUT_CB_REG(R500_US_CODE_RANGE,
 258                    R500_US_CODE_RANGE_ADDR(0) | R500_US_CODE_RANGE_SIZE(code->inst_end));
 259         OUT_CB_REG(R500_US_CODE_OFFSET, 0);
 260         OUT_CB_REG(R500_US_CODE_ADDR,
 261                    R500_US_CODE_START_ADDR(0) | R500_US_CODE_END_ADDR(code->inst_end));
 262
 263         OUT_CB_REG(R500_GA_US_VECTOR_INDEX, R500_GA_US_VECTOR_INDEX_TYPE_INSTR);
 264         OUT_CB_ONE_REG(R500_GA_US_VECTOR_DATA, (code->inst_end + 1) * 6);
 265         for (i = 0; i <= code->inst_end; i++) {
 266             OUT_CB(code->inst[i].inst0);
 267             OUT_CB(code->inst[i].inst1);
 268             OUT_CB(code->inst[i].inst2);
 269             OUT_CB(code->inst[i].inst3);
 270             OUT_CB(code->inst[i].inst4);
 271             OUT_CB(code->inst[i].inst5);
 272         }
 273
 274         /* Emit immediates. */
 275         if (imm_count) {
 276             for(i = imm_first; i < imm_end; ++i) {
 277                 if (constants[i].Type == RC_CONSTANT_IMMEDIATE) {
 278                     const float *data = constants[i].u.Immediate;
 279
 280                     OUT_CB_REG(R500_GA_US_VECTOR_INDEX,
 281                                R500_GA_US_VECTOR_INDEX_TYPE_CONST |
 282                                (i & R500_GA_US_VECTOR_INDEX_MASK));
 283                     OUT_CB_ONE_REG(R500_GA_US_VECTOR_DATA, 4);
 284                     OUT_CB_TABLE(data, 4);
 285                 }
 286             }
 287         }
 288     } else { /* r300 */
 289         struct r300_fragment_program_code *code = &generic_code->code.r300;
 290
 291         shader->cb_code_size = 19 +
 292                                (r300->screen->caps.is_r400 ? 2 : 0) +
 293                                code->alu.length * 4 +
 294                                (code->tex.length ? (1 + code->tex.length) : 0) +
 295                                imm_count * 5;
 296
 297         NEW_CB(shader->cb_code, shader->cb_code_size);
 298
 299         if (r300->screen->caps.is_r400)
 300             OUT_CB_REG(R400_US_CODE_BANK, 0);
 301
 302         OUT_CB_REG(R300_US_CONFIG, code->config);
 303         OUT_CB_REG(R300_US_PIXSIZE, code->pixsize);
 304         OUT_CB_REG(R300_US_CODE_OFFSET, code->code_offset);
 305
 306         OUT_CB_REG_SEQ(R300_US_CODE_ADDR_0, 4);
 307         OUT_CB_TABLE(code->code_addr, 4);
 308
 309         OUT_CB_REG_SEQ(R300_US_ALU_RGB_INST_0, code->alu.length);
 310         for (i = 0; i < code->alu.length; i++)
 311             OUT_CB(code->alu.inst[i].rgb_inst);
 312
 313         OUT_CB_REG_SEQ(R300_US_ALU_RGB_ADDR_0, code->alu.length);
 314         for (i = 0; i < code->alu.length; i++)
 315             OUT_CB(code->alu.inst[i].rgb_addr);
 316
 317         OUT_CB_REG_SEQ(R300_US_ALU_ALPHA_INST_0, code->alu.length);
 318         for (i = 0; i < code->alu.length; i++)
 319             OUT_CB(code->alu.inst[i].alpha_inst);
 320
 321         OUT_CB_REG_SEQ(R300_US_ALU_ALPHA_ADDR_0, code->alu.length);
 322         for (i = 0; i < code->alu.length; i++)
 323             OUT_CB(code->alu.inst[i].alpha_addr);
 324
 325         if (code->tex.length) {
 326             OUT_CB_REG_SEQ(R300_US_TEX_INST_0, code->tex.length);
 327             OUT_CB_TABLE(code->tex.inst, code->tex.length);
 328         }
 329
 330         /* Emit immediates. */
 331         if (imm_count) {
 332             for(i = imm_first; i < imm_end; ++i) {
 333                 if (constants[i].Type == RC_CONSTANT_IMMEDIATE) {
 334                     const float *data = constants[i].u.Immediate;
 335
 336                     OUT_CB_REG_SEQ(R300_PFS_PARAM_0_X + i * 16, 4);
 337                     OUT_CB(pack_float24(data[0]));
 338                     OUT_CB(pack_float24(data[1]));
 339                     OUT_CB(pack_float24(data[2]));
 340                     OUT_CB(pack_float24(data[3]));
 341                 }
 342             }
 343         }
 344     }
 345
 346     OUT_CB_REG(R300_FG_DEPTH_SRC, shader->fg_depth_src);
 347     OUT_CB_REG(R300_US_W_FMT, shader->us_out_w);
 348     END_CB;
 349 }
 350
 351 static void r300_translate_fragment_shader(
 352     struct r300_context* r300,
 353     struct r300_fragment_shader_code* shader,
 354     const struct tgsi_token *tokens)
 355 {
 356     struct r300_fragment_program_compiler compiler;
 357     struct tgsi_to_rc ttr;
 358     int wpos;
 359     unsigned i;
 360
 361     tgsi_scan_shader(tokens, &shader->info);
 362     r300_shader_read_fs_inputs(&shader->info, &shader->inputs);
 363
 364     wpos = shader->inputs.wpos;
 365
 366     /* Setup the compiler. */
 367     memset(&compiler, 0, sizeof(compiler));
 368     rc_init(&compiler.Base);
 369     compiler.Base.Debug = DBG_ON(r300, DBG_FP);
 370
 371     compiler.code = &shader->code;
 372     compiler.state = shader->compare_state;
 373     compiler.Base.is_r500 = r300->screen->caps.is_r500;
 374     compiler.Base.max_temp_regs = compiler.Base.is_r500 ? 128 : 32;
 375     compiler.AllocateHwInputs = &allocate_hardware_inputs;
 376     compiler.UserData = &shader->inputs;
 377
 378     find_output_registers(&compiler, shader);
 379
 380     if (compiler.Base.Debug) {
 381         debug_printf("r300: Initial fragment program\n");
 382         tgsi_dump(tokens, 0);
 383     }
 384
 385     /* Translate TGSI to our internal representation */
 386     ttr.compiler = &compiler.Base;
 387     ttr.info = &shader->info;
 388     ttr.use_half_swizzles = TRUE;
 389
 390     r300_tgsi_to_rc(&ttr, tokens);
 391
 392     /**
 393      * Transform the program to support WPOS.
 394      *
 395      * Introduce a small fragment at the start of the program that will be
 396      * the only code that directly reads the WPOS input.
 397      * All other code pieces that reference that input will be rewritten
 398      * to read from a newly allocated temporary. */
 399     if (wpos != ATTR_UNUSED) {
 400         /* Moving the input to some other reg is not really necessary. */
 401         rc_transform_fragment_wpos(&compiler.Base, wpos, wpos, TRUE);
 402     }
 403
 404     /* Invoke the compiler */
 405     r3xx_compile_fragment_program(&compiler);
 406
 407     /* Shaders with zero instructions are invalid,
 408      * use the dummy shader instead. */
 409     if (shader->code.code.r500.inst_end == -1) {
 410         rc_destroy(&compiler.Base);
 411         r300_dummy_fragment_shader(r300, shader);
 412         return;
 413     }
 414
 415     if (compiler.Base.Error) {
 416         fprintf(stderr, "r300 FP: Compiler Error:\n%sUsing a dummy shader"
 417                 " instead.\nIf there's an 'unknown opcode' message, please"
 418                 " file a bug report and attach this log.\n", compiler.Base.ErrorMsg);
 419
 420         if (shader->dummy) {
 421             fprintf(stderr, "r300 FP: Cannot compile the dummy shader! "
 422                     "Giving up...\n");
 423             abort();
 424         }
 425
 426         rc_destroy(&compiler.Base);
 427         r300_dummy_fragment_shader(r300, shader);
 428         return;
 429     }
 430
 431     /* Initialize numbers of constants for each type. */
 432     shader->externals_count = ttr.immediate_offset;
 433     shader->immediates_count = 0;
 434     shader->rc_state_count = 0;
 435
 436     for (i = shader->externals_count; i < shader->code.constants.Count; i++) {
 437         switch (shader->code.constants.Constants[i].Type) {
 438             case RC_CONSTANT_IMMEDIATE:
 439                 ++shader->immediates_count;
 440                 break;
 441             case RC_CONSTANT_STATE:
 442                 ++shader->rc_state_count;
 443                 break;
 444             default:
 445                 assert(0);
 446         }
 447     }
 448
 449     /* Setup shader depth output. */
 450     if (shader->code.writes_depth) {
 451         shader->fg_depth_src = R300_FG_DEPTH_SRC_SHADER;
 452         shader->us_out_w = R300_W_FMT_W24 | R300_W_SRC_US;
 453     } else {
 454         shader->fg_depth_src = R300_FG_DEPTH_SRC_SCAN;
 455         shader->us_out_w = R300_W_FMT_W0 | R300_W_SRC_US;
 456     }
 457
 458     /* And, finally... */
 459     rc_destroy(&compiler.Base);
 460
 461     /* Build the command buffer. */
 462     r300_emit_fs_code_to_buffer(r300, shader);
 463 }
 464
 465 boolean r300_pick_fragment_shader(struct r300_context* r300)
 466 {
 467     struct r300_fragment_shader* fs = r300_fs(r300);
 468     struct r300_fragment_program_external_state state = {{{ 0 }}};
 469     struct r300_fragment_shader_code* ptr;
 470
 471     get_external_state(r300, &state);
 472
 473     if (!fs->first) {
 474         /* Build the fragment shader for the first time. */
 475         fs->first = fs->shader = CALLOC_STRUCT(r300_fragment_shader_code);
 476
 477         memcpy(&fs->shader->compare_state, &state,
 478             sizeof(struct r300_fragment_program_external_state));
 479         r300_translate_fragment_shader(r300, fs->shader, fs->state.tokens);
 480         return TRUE;
 481
 482     } else {
 483         /* Check if the currently-bound shader has been compiled
 484          * with the texture-compare state we need. */
 485         if (memcmp(&fs->shader->compare_state, &state, sizeof(state)) != 0) {
 486             /* Search for the right shader. */
 487             ptr = fs->first;
 488             while (ptr) {
 489                 if (memcmp(&ptr->compare_state, &state, sizeof(state)) == 0) {
 490                     if (fs->shader != ptr) {
 491                         fs->shader = ptr;
 492                         return TRUE;
 493                     }
 494                     /* The currently-bound one is OK. */
 495                     return FALSE;
 496                 }
 497                 ptr = ptr->next;
 498             }
 499
 500             /* Not found, gotta compile a new one. */
 501             ptr = CALLOC_STRUCT(r300_fragment_shader_code);
 502             ptr->next = fs->first;
 503             fs->first = fs->shader = ptr;
 504
 505             ptr->compare_state = state;
 506             r300_translate_fragment_shader(r300, ptr, fs->state.tokens);
 507             return TRUE;
 508         }
 509     }
 510
 511     return FALSE;
 512 }