src/mesa/drivers/dri/i965/brw_eu.h

   1 /*
   2  Copyright (C) Intel Corp.  2006.  All Rights Reserved.
   3  Intel funded Tungsten Graphics (http://www.tungstengraphics.com) to
   4  develop this 3D driver.
   5
   6  Permission is hereby granted, free of charge, to any person obtaining
   7  a copy of this software and associated documentation files (the
   8  "Software"), to deal in the Software without restriction, including
   9  without limitation the rights to use, copy, modify, merge, publish,
  10  distribute, sublicense, and/or sell copies of the Software, and to
  11  permit persons to whom the Software is furnished to do so, subject to
  12  the following conditions:
  13
  14  The above copyright notice and this permission notice (including the
  15  next paragraph) shall be included in all copies or substantial
  16  portions of the Software.
  17
  18  THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  19  EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
  21  IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
  22  LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
  23  OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
  24  WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25
  26  **********************************************************************/
  27  /*
  28   * Authors:
  29   *   Keith Whitwell <keith@tungstengraphics.com>
  30   */
  31
  32
  33 #ifndef BRW_EU_H
  34 #define BRW_EU_H
  35
  36 #include <stdbool.h>
  37 #include "brw_structs.h"
  38 #include "brw_defines.h"
  39 #include "brw_reg.h"
  40 #include "program/prog_instruction.h"
  41
  42 #ifdef __cplusplus
  43 extern "C" {
  44 #endif
  45
  46 #define BRW_EU_MAX_INSN_STACK 5
  47
  48 struct brw_compile {
  49    struct brw_instruction *store;
  50    int store_size;
  51    GLuint nr_insn;
  52    unsigned int next_insn_offset;
  53
  54    void *mem_ctx;
  55
  56    /* Allow clients to push/pop instruction state:
  57     */
  58    struct brw_instruction stack[BRW_EU_MAX_INSN_STACK];
  59    bool compressed_stack[BRW_EU_MAX_INSN_STACK];
  60    struct brw_instruction *current;
  61
  62    GLuint flag_value;
  63    bool single_program_flow;
  64    bool compressed;
  65    struct brw_context *brw;
  66
  67    /* Control flow stacks:
  68     * - if_stack contains IF and ELSE instructions which must be patched
  69     *   (and popped) once the matching ENDIF instruction is encountered.
  70     *
  71     *   Just store the instruction pointer(an index).
  72     */
  73    int *if_stack;
  74    int if_stack_depth;
  75    int if_stack_array_size;
  76
  77    /**
  78     * loop_stack contains the instruction pointers of the starts of loops which
  79     * must be patched (and popped) once the matching WHILE instruction is
  80     * encountered.
  81     */
  82    int *loop_stack;
  83    /**
  84     * pre-gen6, the BREAK and CONT instructions had to tell how many IF/ENDIF
  85     * blocks they were popping out of, to fix up the mask stack.  This tracks
  86     * the IF/ENDIF nesting in each current nested loop level.
  87     */
  88    int *if_depth_in_loop;
  89    int loop_stack_depth;
  90    int loop_stack_array_size;
  91 };
  92
  93 static INLINE struct brw_instruction *current_insn( struct brw_compile *p)
  94 {
  95    return &p->store[p->nr_insn];
  96 }
  97
  98 void brw_pop_insn_state( struct brw_compile *p );
  99 void brw_push_insn_state( struct brw_compile *p );
 100 void brw_set_mask_control( struct brw_compile *p, GLuint value );
 101 void brw_set_saturate( struct brw_compile *p, bool enable );
 102 void brw_set_access_mode( struct brw_compile *p, GLuint access_mode );
 103 void brw_set_compression_control(struct brw_compile *p, enum brw_compression c);
 104 void brw_set_predicate_control_flag_value( struct brw_compile *p, GLuint value );
 105 void brw_set_predicate_control( struct brw_compile *p, GLuint pc );
 106 void brw_set_predicate_inverse(struct brw_compile *p, bool predicate_inverse);
 107 void brw_set_conditionalmod( struct brw_compile *p, GLuint conditional );
 108 void brw_set_flag_reg(struct brw_compile *p, int reg, int subreg);
 109 void brw_set_acc_write_control(struct brw_compile *p, GLuint value);
 110
 111 void brw_init_compile(struct brw_context *, struct brw_compile *p,
 112                       void *mem_ctx);
 113 void brw_dump_compile(struct brw_compile *p, FILE *out, int start, int end);
 114 const GLuint *brw_get_program( struct brw_compile *p, GLuint *sz );
 115
 116 struct brw_instruction *brw_next_insn(struct brw_compile *p, GLuint opcode);
 117 void brw_set_dest(struct brw_compile *p, struct brw_instruction *insn,
 118                   struct brw_reg dest);
 119 void brw_set_src0(struct brw_compile *p, struct brw_instruction *insn,
 120                   struct brw_reg reg);
 121
 122 void gen6_resolve_implied_move(struct brw_compile *p,
 123                                struct brw_reg *src,
 124                                GLuint msg_reg_nr);
 125
 126 /* Helpers for regular instructions:
 127  */
 128 #define ALU1(OP)                                        \
 129 struct brw_instruction *brw_##OP(struct brw_compile *p, \
 130               struct brw_reg dest,                      \
 131               struct brw_reg src0);
 132
 133 #define ALU2(OP)                                        \
 134 struct brw_instruction *brw_##OP(struct brw_compile *p, \
 135               struct brw_reg dest,                      \
 136               struct brw_reg src0,                      \
 137               struct brw_reg src1);
 138
 139 #define ALU3(OP)                                        \
 140 struct brw_instruction *brw_##OP(struct brw_compile *p, \
 141               struct brw_reg dest,                      \
 142               struct brw_reg src0,                      \
 143               struct brw_reg src1,                      \
 144               struct brw_reg src2);
 145
 146 #define ROUND(OP) \
 147 void brw_##OP(struct brw_compile *p, struct brw_reg dest, struct brw_reg src0);
 148
 149 ALU1(MOV)
 150 ALU2(SEL)
 151 ALU1(NOT)
 152 ALU2(AND)
 153 ALU2(OR)
 154 ALU2(XOR)
 155 ALU2(SHR)
 156 ALU2(SHL)
 157 ALU2(ASR)
 158 ALU1(F32TO16)
 159 ALU1(F16TO32)
 160 ALU2(JMPI)
 161 ALU2(ADD)
 162 ALU2(AVG)
 163 ALU2(MUL)
 164 ALU1(FRC)
 165 ALU1(RNDD)
 166 ALU2(MAC)
 167 ALU2(MACH)
 168 ALU1(LZD)
 169 ALU2(DP4)
 170 ALU2(DPH)
 171 ALU2(DP3)
 172 ALU2(DP2)
 173 ALU2(LINE)
 174 ALU2(PLN)
 175 ALU3(MAD)
 176 ALU3(LRP)
 177 ALU1(BFREV)
 178 ALU3(BFE)
 179 ALU2(BFI1)
 180 ALU3(BFI2)
 181 ALU1(FBH)
 182 ALU1(FBL)
 183 ALU1(CBIT)
 184 ALU2(ADDC)
 185 ALU2(SUBB)
 186
 187 ROUND(RNDZ)
 188 ROUND(RNDE)
 189
 190 #undef ALU1
 191 #undef ALU2
 192 #undef ALU3
 193 #undef ROUND
 194
 195
 196 /* Helpers for SEND instruction:
 197  */
 198 void brw_set_sampler_message(struct brw_compile *p,
 199                              struct brw_instruction *insn,
 200                              GLuint binding_table_index,
 201                              GLuint sampler,
 202                              GLuint msg_type,
 203                              GLuint response_length,
 204                              GLuint msg_length,
 205                              GLuint header_present,
 206                              GLuint simd_mode,
 207                              GLuint return_format);
 208
 209 void brw_set_dp_read_message(struct brw_compile *p,
 210                              struct brw_instruction *insn,
 211                              GLuint binding_table_index,
 212                              GLuint msg_control,
 213                              GLuint msg_type,
 214                              GLuint target_cache,
 215                              GLuint msg_length,
 216                              bool header_present,
 217                              GLuint response_length);
 218
 219 void brw_set_dp_write_message(struct brw_compile *p,
 220                               struct brw_instruction *insn,
 221                               GLuint binding_table_index,
 222                               GLuint msg_control,
 223                               GLuint msg_type,
 224                               GLuint msg_length,
 225                               bool header_present,
 226                               GLuint last_render_target,
 227                               GLuint response_length,
 228                               GLuint end_of_thread,
 229                               GLuint send_commit_msg);
 230
 231 enum brw_urb_write_flags {
 232    BRW_URB_WRITE_NO_FLAGS = 0,
 233
 234    /**
 235     * Causes a new URB entry to be allocated, and its address stored in the
 236     * destination register (gen < 7).
 237     */
 238    BRW_URB_WRITE_ALLOCATE = 0x1,
 239
 240    /**
 241     * Causes the current URB entry to be deallocated (gen < 7).
 242     */
 243    BRW_URB_WRITE_UNUSED = 0x2,
 244
 245    /**
 246     * Causes the thread to terminate.
 247     */
 248    BRW_URB_WRITE_EOT = 0x4,
 249
 250    /**
 251     * Indicates that the given URB entry is complete, and may be sent further
 252     * down the 3D pipeline (gen < 7).
 253     */
 254    BRW_URB_WRITE_COMPLETE = 0x8,
 255
 256    /**
 257     * Indicates that an additional offset (which may be different for the two
 258     * vec4 slots) is stored in the message header (gen == 7).
 259     */
 260    BRW_URB_WRITE_PER_SLOT_OFFSET = 0x10,
 261
 262    /**
 263     * Indicates that the channel masks in the URB_WRITE message header should
 264     * not be overridden to 0xff (gen == 7).
 265     */
 266    BRW_URB_WRITE_USE_CHANNEL_MASKS = 0x20,
 267
 268    /**
 269     * Indicates that the data should be sent to the URB using the
 270     * URB_WRITE_OWORD message rather than URB_WRITE_HWORD (gen == 7).  This
 271     * causes offsets to be interpreted as multiples of an OWORD instead of an
 272     * HWORD, and only allows one OWORD to be written.
 273     */
 274    BRW_URB_WRITE_OWORD = 0x40,
 275
 276    /**
 277     * Convenient combination of flags: end the thread while simultaneously
 278     * marking the given URB entry as complete.
 279     */
 280    BRW_URB_WRITE_EOT_COMPLETE = BRW_URB_WRITE_EOT | BRW_URB_WRITE_COMPLETE,
 281
 282    /**
 283     * Convenient combination of flags: mark the given URB entry as complete
 284     * and simultaneously allocate a new one.
 285     */
 286    BRW_URB_WRITE_ALLOCATE_COMPLETE =
 287       BRW_URB_WRITE_ALLOCATE | BRW_URB_WRITE_COMPLETE,
 288 };
 289
 290 #ifdef __cplusplus
 291 /**
 292  * Allow brw_urb_write_flags enums to be ORed together.
 293  */
 294 inline brw_urb_write_flags
 295 operator|(brw_urb_write_flags x, brw_urb_write_flags y)
 296 {
 297    return static_cast<brw_urb_write_flags>(static_cast<int>(x) |
 298                                            static_cast<int>(y));
 299 }
 300 #endif
 301
 302 void brw_urb_WRITE(struct brw_compile *p,
 303                    struct brw_reg dest,
 304                    GLuint msg_reg_nr,
 305                    struct brw_reg src0,
 306                    enum brw_urb_write_flags flags,
 307                    GLuint msg_length,
 308                    GLuint response_length,
 309                    GLuint offset,
 310                    GLuint swizzle);
 311
 312 void brw_ff_sync(struct brw_compile *p,
 313                    struct brw_reg dest,
 314                    GLuint msg_reg_nr,
 315                    struct brw_reg src0,
 316                    bool allocate,
 317                    GLuint response_length,
 318                    bool eot);
 319
 320 void brw_svb_write(struct brw_compile *p,
 321                    struct brw_reg dest,
 322                    GLuint msg_reg_nr,
 323                    struct brw_reg src0,
 324                    GLuint binding_table_index,
 325                    bool   send_commit_msg);
 326
 327 void brw_fb_WRITE(struct brw_compile *p,
 328                   int dispatch_width,
 329                    GLuint msg_reg_nr,
 330                    struct brw_reg src0,
 331                    GLuint msg_control,
 332                    GLuint binding_table_index,
 333                    GLuint msg_length,
 334                    GLuint response_length,
 335                    bool eot,
 336                    bool header_present);
 337
 338 void brw_SAMPLE(struct brw_compile *p,
 339                 struct brw_reg dest,
 340                 GLuint msg_reg_nr,
 341                 struct brw_reg src0,
 342                 GLuint binding_table_index,
 343                 GLuint sampler,
 344                 GLuint msg_type,
 345                 GLuint response_length,
 346                 GLuint msg_length,
 347                 GLuint header_present,
 348                 GLuint simd_mode,
 349                 GLuint return_format);
 350
 351 void brw_math( struct brw_compile *p,
 352                struct brw_reg dest,
 353                GLuint function,
 354                GLuint msg_reg_nr,
 355                struct brw_reg src,
 356                GLuint data_type,
 357                GLuint precision );
 358
 359 void brw_math2(struct brw_compile *p,
 360                struct brw_reg dest,
 361                GLuint function,
 362                struct brw_reg src0,
 363                struct brw_reg src1);
 364
 365 void brw_oword_block_read(struct brw_compile *p,
 366                           struct brw_reg dest,
 367                           struct brw_reg mrf,
 368                           uint32_t offset,
 369                           uint32_t bind_table_index);
 370
 371 void brw_oword_block_read_scratch(struct brw_compile *p,
 372                                   struct brw_reg dest,
 373                                   struct brw_reg mrf,
 374                                   int num_regs,
 375                                   GLuint offset);
 376
 377 void brw_oword_block_write_scratch(struct brw_compile *p,
 378                                    struct brw_reg mrf,
 379                                    int num_regs,
 380                                    GLuint offset);
 381
 382 void brw_shader_time_add(struct brw_compile *p,
 383                          struct brw_reg payload,
 384                          uint32_t surf_index);
 385
 386 /* If/else/endif.  Works by manipulating the execution flags on each
 387  * channel.
 388  */
 389 struct brw_instruction *brw_IF(struct brw_compile *p,
 390                                GLuint execute_size);
 391 struct brw_instruction *gen6_IF(struct brw_compile *p, uint32_t conditional,
 392                                 struct brw_reg src0, struct brw_reg src1);
 393
 394 void brw_ELSE(struct brw_compile *p);
 395 void brw_ENDIF(struct brw_compile *p);
 396
 397 /* DO/WHILE loops:
 398  */
 399 struct brw_instruction *brw_DO(struct brw_compile *p,
 400                                GLuint execute_size);
 401
 402 struct brw_instruction *brw_WHILE(struct brw_compile *p);
 403
 404 struct brw_instruction *brw_BREAK(struct brw_compile *p);
 405 struct brw_instruction *brw_CONT(struct brw_compile *p);
 406 struct brw_instruction *gen6_CONT(struct brw_compile *p);
 407 struct brw_instruction *gen6_HALT(struct brw_compile *p);
 408 /* Forward jumps:
 409  */
 410 void brw_land_fwd_jump(struct brw_compile *p, int jmp_insn_idx);
 411
 412
 413
 414 void brw_NOP(struct brw_compile *p);
 415
 416 void brw_WAIT(struct brw_compile *p);
 417
 418 /* Special case: there is never a destination, execution size will be
 419  * taken from src0:
 420  */
 421 void brw_CMP(struct brw_compile *p,
 422              struct brw_reg dest,
 423              GLuint conditional,
 424              struct brw_reg src0,
 425              struct brw_reg src1);
 426
 427 void
 428 brw_untyped_atomic(struct brw_compile *p,
 429                    struct brw_reg dest,
 430                    struct brw_reg mrf,
 431                    GLuint atomic_op,
 432                    GLuint bind_table_index,
 433                    GLuint msg_length,
 434                    GLuint response_length);
 435
 436 void
 437 brw_untyped_surface_read(struct brw_compile *p,
 438                          struct brw_reg dest,
 439                          struct brw_reg mrf,
 440                          GLuint bind_table_index,
 441                          GLuint msg_length,
 442                          GLuint response_length);
 443
 444 /***********************************************************************
 445  * brw_eu_util.c:
 446  */
 447
 448 void brw_copy_indirect_to_indirect(struct brw_compile *p,
 449                                    struct brw_indirect dst_ptr,
 450                                    struct brw_indirect src_ptr,
 451                                    GLuint count);
 452
 453 void brw_copy_from_indirect(struct brw_compile *p,
 454                             struct brw_reg dst,
 455                             struct brw_indirect ptr,
 456                             GLuint count);
 457
 458 void brw_copy4(struct brw_compile *p,
 459                struct brw_reg dst,
 460                struct brw_reg src,
 461                GLuint count);
 462
 463 void brw_copy8(struct brw_compile *p,
 464                struct brw_reg dst,
 465                struct brw_reg src,
 466                GLuint count);
 467
 468 void brw_math_invert( struct brw_compile *p,
 469                       struct brw_reg dst,
 470                       struct brw_reg src);
 471
 472 void brw_set_src1(struct brw_compile *p,
 473                   struct brw_instruction *insn,
 474                   struct brw_reg reg);
 475
 476 void brw_set_uip_jip(struct brw_compile *p);
 477
 478 uint32_t brw_swap_cmod(uint32_t cmod);
 479
 480 /* brw_eu_compact.c */
 481 void brw_init_compaction_tables(struct brw_context *brw);
 482 void brw_compact_instructions(struct brw_compile *p);
 483 void brw_uncompact_instruction(struct brw_context *brw,
 484                                struct brw_instruction *dst,
 485                                struct brw_compact_instruction *src);
 486 bool brw_try_compact_instruction(struct brw_compile *p,
 487                                  struct brw_compact_instruction *dst,
 488                                  struct brw_instruction *src);
 489
 490 void brw_debug_compact_uncompact(struct brw_context *brw,
 491                                  struct brw_instruction *orig,
 492                                  struct brw_instruction *uncompacted);
 493
 494 #ifdef __cplusplus
 495 }
 496 #endif
 497
 498 #endif