intel/eu: Provide desc immediate argument up front to brw_send_indirect_message().
[mesa.git] / src / intel / compiler / brw_eu.h
1 /*
2 Copyright (C) Intel Corp. 2006. All Rights Reserved.
3 Intel funded Tungsten Graphics 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 <keithw@vmware.com>
30 */
31
32
33 #ifndef BRW_EU_H
34 #define BRW_EU_H
35
36 #include <stdbool.h>
37 #include <stdio.h>
38 #include "brw_inst.h"
39 #include "brw_eu_defines.h"
40 #include "brw_reg.h"
41 #include "brw_disasm_info.h"
42
43 #ifdef __cplusplus
44 extern "C" {
45 #endif
46
47 #define BRW_EU_MAX_INSN_STACK 5
48
49 struct brw_insn_state {
50 /* One of BRW_EXECUTE_* */
51 unsigned exec_size:3;
52
53 /* Group in units of channels */
54 unsigned group:5;
55
56 /* Compression control on gen4-5 */
57 bool compressed:1;
58
59 /* One of BRW_MASK_* */
60 unsigned mask_control:1;
61
62 bool saturate:1;
63
64 /* One of BRW_ALIGN_* */
65 unsigned access_mode:1;
66
67 /* One of BRW_PREDICATE_* */
68 enum brw_predicate predicate:4;
69
70 bool pred_inv:1;
71
72 /* Flag subreg. Bottom bit is subreg, top bit is reg */
73 unsigned flag_subreg:2;
74
75 bool acc_wr_control:1;
76 };
77
78
79 /* A helper for accessing the last instruction emitted. This makes it easy
80 * to set various bits on an instruction without having to create temporary
81 * variable and assign the emitted instruction to those.
82 */
83 #define brw_last_inst (&p->store[p->nr_insn - 1])
84
85 struct brw_codegen {
86 brw_inst *store;
87 int store_size;
88 unsigned nr_insn;
89 unsigned int next_insn_offset;
90
91 void *mem_ctx;
92
93 /* Allow clients to push/pop instruction state:
94 */
95 struct brw_insn_state stack[BRW_EU_MAX_INSN_STACK];
96 struct brw_insn_state *current;
97
98 /** Whether or not the user wants automatic exec sizes
99 *
100 * If true, codegen will try to automatically infer the exec size of an
101 * instruction from the width of the destination register. If false, it
102 * will take whatever is set by brw_set_default_exec_size verbatim.
103 *
104 * This is set to true by default in brw_init_codegen.
105 */
106 bool automatic_exec_sizes;
107
108 bool single_program_flow;
109 const struct gen_device_info *devinfo;
110
111 /* Control flow stacks:
112 * - if_stack contains IF and ELSE instructions which must be patched
113 * (and popped) once the matching ENDIF instruction is encountered.
114 *
115 * Just store the instruction pointer(an index).
116 */
117 int *if_stack;
118 int if_stack_depth;
119 int if_stack_array_size;
120
121 /**
122 * loop_stack contains the instruction pointers of the starts of loops which
123 * must be patched (and popped) once the matching WHILE instruction is
124 * encountered.
125 */
126 int *loop_stack;
127 /**
128 * pre-gen6, the BREAK and CONT instructions had to tell how many IF/ENDIF
129 * blocks they were popping out of, to fix up the mask stack. This tracks
130 * the IF/ENDIF nesting in each current nested loop level.
131 */
132 int *if_depth_in_loop;
133 int loop_stack_depth;
134 int loop_stack_array_size;
135 };
136
137 void brw_pop_insn_state( struct brw_codegen *p );
138 void brw_push_insn_state( struct brw_codegen *p );
139 unsigned brw_get_default_exec_size(struct brw_codegen *p);
140 unsigned brw_get_default_group(struct brw_codegen *p);
141 unsigned brw_get_default_access_mode(struct brw_codegen *p);
142 void brw_set_default_exec_size(struct brw_codegen *p, unsigned value);
143 void brw_set_default_mask_control( struct brw_codegen *p, unsigned value );
144 void brw_set_default_saturate( struct brw_codegen *p, bool enable );
145 void brw_set_default_access_mode( struct brw_codegen *p, unsigned access_mode );
146 void brw_inst_set_compression(const struct gen_device_info *devinfo,
147 brw_inst *inst, bool on);
148 void brw_set_default_compression(struct brw_codegen *p, bool on);
149 void brw_inst_set_group(const struct gen_device_info *devinfo,
150 brw_inst *inst, unsigned group);
151 void brw_set_default_group(struct brw_codegen *p, unsigned group);
152 void brw_set_default_compression_control(struct brw_codegen *p, enum brw_compression c);
153 void brw_set_default_predicate_control( struct brw_codegen *p, unsigned pc );
154 void brw_set_default_predicate_inverse(struct brw_codegen *p, bool predicate_inverse);
155 void brw_set_default_flag_reg(struct brw_codegen *p, int reg, int subreg);
156 void brw_set_default_acc_write_control(struct brw_codegen *p, unsigned value);
157
158 void brw_init_codegen(const struct gen_device_info *, struct brw_codegen *p,
159 void *mem_ctx);
160 int brw_disassemble_inst(FILE *file, const struct gen_device_info *devinfo,
161 const struct brw_inst *inst, bool is_compacted);
162 void brw_disassemble(const struct gen_device_info *devinfo,
163 const void *assembly, int start, int end, FILE *out);
164 const unsigned *brw_get_program( struct brw_codegen *p, unsigned *sz );
165
166 brw_inst *brw_next_insn(struct brw_codegen *p, unsigned opcode);
167 void brw_set_dest(struct brw_codegen *p, brw_inst *insn, struct brw_reg dest);
168 void brw_set_src0(struct brw_codegen *p, brw_inst *insn, struct brw_reg reg);
169
170 void gen6_resolve_implied_move(struct brw_codegen *p,
171 struct brw_reg *src,
172 unsigned msg_reg_nr);
173
174 /* Helpers for regular instructions:
175 */
176 #define ALU1(OP) \
177 brw_inst *brw_##OP(struct brw_codegen *p, \
178 struct brw_reg dest, \
179 struct brw_reg src0);
180
181 #define ALU2(OP) \
182 brw_inst *brw_##OP(struct brw_codegen *p, \
183 struct brw_reg dest, \
184 struct brw_reg src0, \
185 struct brw_reg src1);
186
187 #define ALU3(OP) \
188 brw_inst *brw_##OP(struct brw_codegen *p, \
189 struct brw_reg dest, \
190 struct brw_reg src0, \
191 struct brw_reg src1, \
192 struct brw_reg src2);
193
194 #define ROUND(OP) \
195 void brw_##OP(struct brw_codegen *p, struct brw_reg dest, struct brw_reg src0);
196
197 ALU1(MOV)
198 ALU2(SEL)
199 ALU1(NOT)
200 ALU2(AND)
201 ALU2(OR)
202 ALU2(XOR)
203 ALU2(SHR)
204 ALU2(SHL)
205 ALU1(DIM)
206 ALU2(ASR)
207 ALU3(CSEL)
208 ALU1(F32TO16)
209 ALU1(F16TO32)
210 ALU2(ADD)
211 ALU2(AVG)
212 ALU2(MUL)
213 ALU1(FRC)
214 ALU1(RNDD)
215 ALU2(MAC)
216 ALU2(MACH)
217 ALU1(LZD)
218 ALU2(DP4)
219 ALU2(DPH)
220 ALU2(DP3)
221 ALU2(DP2)
222 ALU2(LINE)
223 ALU2(PLN)
224 ALU3(MAD)
225 ALU3(LRP)
226 ALU1(BFREV)
227 ALU3(BFE)
228 ALU2(BFI1)
229 ALU3(BFI2)
230 ALU1(FBH)
231 ALU1(FBL)
232 ALU1(CBIT)
233 ALU2(ADDC)
234 ALU2(SUBB)
235 ALU2(MAC)
236
237 ROUND(RNDZ)
238 ROUND(RNDE)
239
240 #undef ALU1
241 #undef ALU2
242 #undef ALU3
243 #undef ROUND
244
245
246 /* Helpers for SEND instruction:
247 */
248 void brw_set_sampler_message(struct brw_codegen *p,
249 brw_inst *insn,
250 unsigned binding_table_index,
251 unsigned sampler,
252 unsigned msg_type,
253 unsigned response_length,
254 unsigned msg_length,
255 unsigned header_present,
256 unsigned simd_mode,
257 unsigned return_format);
258
259 void brw_set_dp_read_message(struct brw_codegen *p,
260 brw_inst *insn,
261 unsigned binding_table_index,
262 unsigned msg_control,
263 unsigned msg_type,
264 unsigned target_cache,
265 unsigned msg_length,
266 bool header_present,
267 unsigned response_length);
268
269 void brw_set_dp_write_message(struct brw_codegen *p,
270 brw_inst *insn,
271 unsigned binding_table_index,
272 unsigned msg_control,
273 unsigned msg_type,
274 unsigned target_cache,
275 unsigned msg_length,
276 bool header_present,
277 unsigned last_render_target,
278 unsigned response_length,
279 unsigned end_of_thread,
280 unsigned send_commit_msg);
281
282 /**
283 * Construct a message descriptor immediate with the specified common
284 * descriptor controls.
285 */
286 static inline uint32_t
287 brw_message_desc(const struct gen_device_info *devinfo,
288 unsigned msg_length,
289 unsigned response_length,
290 bool header_present)
291 {
292 if (devinfo->gen >= 5) {
293 return (SET_BITS(msg_length, 28, 25) |
294 SET_BITS(response_length, 24, 20) |
295 SET_BITS(header_present, 19, 19));
296 } else {
297 return (SET_BITS(msg_length, 23, 20) |
298 SET_BITS(response_length, 19, 16));
299 }
300 }
301
302
303 void brw_urb_WRITE(struct brw_codegen *p,
304 struct brw_reg dest,
305 unsigned msg_reg_nr,
306 struct brw_reg src0,
307 enum brw_urb_write_flags flags,
308 unsigned msg_length,
309 unsigned response_length,
310 unsigned offset,
311 unsigned swizzle);
312
313 /**
314 * Send message to shared unit \p sfid with a possibly indirect descriptor \p
315 * desc. If \p desc is not an immediate it will be transparently loaded to an
316 * address register using an OR instruction. The returned instruction can be
317 * passed as argument to the usual brw_set_*_message() functions in order to
318 * specify any additional descriptor bits -- If \p desc is an immediate this
319 * will be the SEND instruction itself, otherwise it will be the OR
320 * instruction.
321 */
322 struct brw_inst *
323 brw_send_indirect_message(struct brw_codegen *p,
324 unsigned sfid,
325 struct brw_reg dst,
326 struct brw_reg payload,
327 struct brw_reg desc,
328 unsigned desc_imm);
329
330 void brw_ff_sync(struct brw_codegen *p,
331 struct brw_reg dest,
332 unsigned msg_reg_nr,
333 struct brw_reg src0,
334 bool allocate,
335 unsigned response_length,
336 bool eot);
337
338 void brw_svb_write(struct brw_codegen *p,
339 struct brw_reg dest,
340 unsigned msg_reg_nr,
341 struct brw_reg src0,
342 unsigned binding_table_index,
343 bool send_commit_msg);
344
345 brw_inst *brw_fb_WRITE(struct brw_codegen *p,
346 struct brw_reg payload,
347 struct brw_reg implied_header,
348 unsigned msg_control,
349 unsigned binding_table_index,
350 unsigned msg_length,
351 unsigned response_length,
352 bool eot,
353 bool last_render_target,
354 bool header_present);
355
356 brw_inst *gen9_fb_READ(struct brw_codegen *p,
357 struct brw_reg dst,
358 struct brw_reg payload,
359 unsigned binding_table_index,
360 unsigned msg_length,
361 unsigned response_length,
362 bool per_sample);
363
364 void brw_SAMPLE(struct brw_codegen *p,
365 struct brw_reg dest,
366 unsigned msg_reg_nr,
367 struct brw_reg src0,
368 unsigned binding_table_index,
369 unsigned sampler,
370 unsigned msg_type,
371 unsigned response_length,
372 unsigned msg_length,
373 unsigned header_present,
374 unsigned simd_mode,
375 unsigned return_format);
376
377 void brw_adjust_sampler_state_pointer(struct brw_codegen *p,
378 struct brw_reg header,
379 struct brw_reg sampler_index);
380
381 void gen4_math(struct brw_codegen *p,
382 struct brw_reg dest,
383 unsigned function,
384 unsigned msg_reg_nr,
385 struct brw_reg src,
386 unsigned precision );
387
388 void gen6_math(struct brw_codegen *p,
389 struct brw_reg dest,
390 unsigned function,
391 struct brw_reg src0,
392 struct brw_reg src1);
393
394 void brw_oword_block_read(struct brw_codegen *p,
395 struct brw_reg dest,
396 struct brw_reg mrf,
397 uint32_t offset,
398 uint32_t bind_table_index);
399
400 unsigned brw_scratch_surface_idx(const struct brw_codegen *p);
401
402 void brw_oword_block_read_scratch(struct brw_codegen *p,
403 struct brw_reg dest,
404 struct brw_reg mrf,
405 int num_regs,
406 unsigned offset);
407
408 void brw_oword_block_write_scratch(struct brw_codegen *p,
409 struct brw_reg mrf,
410 int num_regs,
411 unsigned offset);
412
413 void gen7_block_read_scratch(struct brw_codegen *p,
414 struct brw_reg dest,
415 int num_regs,
416 unsigned offset);
417
418 void brw_shader_time_add(struct brw_codegen *p,
419 struct brw_reg payload,
420 uint32_t surf_index);
421
422 /**
423 * Return the generation-specific jump distance scaling factor.
424 *
425 * Given the number of instructions to jump, we need to scale by
426 * some number to obtain the actual jump distance to program in an
427 * instruction.
428 */
429 static inline unsigned
430 brw_jump_scale(const struct gen_device_info *devinfo)
431 {
432 /* Broadwell measures jump targets in bytes. */
433 if (devinfo->gen >= 8)
434 return 16;
435
436 /* Ironlake and later measure jump targets in 64-bit data chunks (in order
437 * (to support compaction), so each 128-bit instruction requires 2 chunks.
438 */
439 if (devinfo->gen >= 5)
440 return 2;
441
442 /* Gen4 simply uses the number of 128-bit instructions. */
443 return 1;
444 }
445
446 void brw_barrier(struct brw_codegen *p, struct brw_reg src);
447
448 /* If/else/endif. Works by manipulating the execution flags on each
449 * channel.
450 */
451 brw_inst *brw_IF(struct brw_codegen *p, unsigned execute_size);
452 brw_inst *gen6_IF(struct brw_codegen *p, enum brw_conditional_mod conditional,
453 struct brw_reg src0, struct brw_reg src1);
454
455 void brw_ELSE(struct brw_codegen *p);
456 void brw_ENDIF(struct brw_codegen *p);
457
458 /* DO/WHILE loops:
459 */
460 brw_inst *brw_DO(struct brw_codegen *p, unsigned execute_size);
461
462 brw_inst *brw_WHILE(struct brw_codegen *p);
463
464 brw_inst *brw_BREAK(struct brw_codegen *p);
465 brw_inst *brw_CONT(struct brw_codegen *p);
466 brw_inst *gen6_HALT(struct brw_codegen *p);
467
468 /* Forward jumps:
469 */
470 void brw_land_fwd_jump(struct brw_codegen *p, int jmp_insn_idx);
471
472 brw_inst *brw_JMPI(struct brw_codegen *p, struct brw_reg index,
473 unsigned predicate_control);
474
475 void brw_NOP(struct brw_codegen *p);
476
477 void brw_WAIT(struct brw_codegen *p);
478
479 /* Special case: there is never a destination, execution size will be
480 * taken from src0:
481 */
482 void brw_CMP(struct brw_codegen *p,
483 struct brw_reg dest,
484 unsigned conditional,
485 struct brw_reg src0,
486 struct brw_reg src1);
487
488 void
489 brw_untyped_atomic(struct brw_codegen *p,
490 struct brw_reg dst,
491 struct brw_reg payload,
492 struct brw_reg surface,
493 unsigned atomic_op,
494 unsigned msg_length,
495 bool response_expected,
496 bool header_present);
497
498 void
499 brw_untyped_surface_read(struct brw_codegen *p,
500 struct brw_reg dst,
501 struct brw_reg payload,
502 struct brw_reg surface,
503 unsigned msg_length,
504 unsigned num_channels);
505
506 void
507 brw_untyped_surface_write(struct brw_codegen *p,
508 struct brw_reg payload,
509 struct brw_reg surface,
510 unsigned msg_length,
511 unsigned num_channels,
512 bool header_present);
513
514 void
515 brw_typed_atomic(struct brw_codegen *p,
516 struct brw_reg dst,
517 struct brw_reg payload,
518 struct brw_reg surface,
519 unsigned atomic_op,
520 unsigned msg_length,
521 bool response_expected,
522 bool header_present);
523
524 void
525 brw_typed_surface_read(struct brw_codegen *p,
526 struct brw_reg dst,
527 struct brw_reg payload,
528 struct brw_reg surface,
529 unsigned msg_length,
530 unsigned num_channels,
531 bool header_present);
532
533 void
534 brw_typed_surface_write(struct brw_codegen *p,
535 struct brw_reg payload,
536 struct brw_reg surface,
537 unsigned msg_length,
538 unsigned num_channels,
539 bool header_present);
540
541 void
542 brw_byte_scattered_read(struct brw_codegen *p,
543 struct brw_reg dst,
544 struct brw_reg payload,
545 struct brw_reg surface,
546 unsigned msg_length,
547 unsigned bit_size);
548
549 void
550 brw_byte_scattered_write(struct brw_codegen *p,
551 struct brw_reg payload,
552 struct brw_reg surface,
553 unsigned msg_length,
554 unsigned bit_size,
555 bool header_present);
556
557 void
558 brw_memory_fence(struct brw_codegen *p,
559 struct brw_reg dst,
560 enum opcode send_op);
561
562 void
563 brw_pixel_interpolator_query(struct brw_codegen *p,
564 struct brw_reg dest,
565 struct brw_reg mrf,
566 bool noperspective,
567 unsigned mode,
568 struct brw_reg data,
569 unsigned msg_length,
570 unsigned response_length);
571
572 void
573 brw_find_live_channel(struct brw_codegen *p,
574 struct brw_reg dst,
575 struct brw_reg mask);
576
577 void
578 brw_broadcast(struct brw_codegen *p,
579 struct brw_reg dst,
580 struct brw_reg src,
581 struct brw_reg idx);
582
583 void
584 brw_rounding_mode(struct brw_codegen *p,
585 enum brw_rnd_mode mode);
586
587 /***********************************************************************
588 * brw_eu_util.c:
589 */
590
591 void brw_copy_indirect_to_indirect(struct brw_codegen *p,
592 struct brw_indirect dst_ptr,
593 struct brw_indirect src_ptr,
594 unsigned count);
595
596 void brw_copy_from_indirect(struct brw_codegen *p,
597 struct brw_reg dst,
598 struct brw_indirect ptr,
599 unsigned count);
600
601 void brw_copy4(struct brw_codegen *p,
602 struct brw_reg dst,
603 struct brw_reg src,
604 unsigned count);
605
606 void brw_copy8(struct brw_codegen *p,
607 struct brw_reg dst,
608 struct brw_reg src,
609 unsigned count);
610
611 void brw_math_invert( struct brw_codegen *p,
612 struct brw_reg dst,
613 struct brw_reg src);
614
615 void brw_set_src1(struct brw_codegen *p, brw_inst *insn, struct brw_reg reg);
616
617 void brw_set_desc_ex(struct brw_codegen *p, brw_inst *insn,
618 unsigned desc, unsigned ex_desc);
619
620 static inline void
621 brw_set_desc(struct brw_codegen *p, brw_inst *insn, unsigned desc)
622 {
623 brw_set_desc_ex(p, insn, desc, 0);
624 }
625
626 void brw_set_uip_jip(struct brw_codegen *p, int start_offset);
627
628 enum brw_conditional_mod brw_negate_cmod(uint32_t cmod);
629 enum brw_conditional_mod brw_swap_cmod(uint32_t cmod);
630
631 /* brw_eu_compact.c */
632 void brw_init_compaction_tables(const struct gen_device_info *devinfo);
633 void brw_compact_instructions(struct brw_codegen *p, int start_offset,
634 struct disasm_info *disasm);
635 void brw_uncompact_instruction(const struct gen_device_info *devinfo,
636 brw_inst *dst, brw_compact_inst *src);
637 bool brw_try_compact_instruction(const struct gen_device_info *devinfo,
638 brw_compact_inst *dst, const brw_inst *src);
639
640 void brw_debug_compact_uncompact(const struct gen_device_info *devinfo,
641 brw_inst *orig, brw_inst *uncompacted);
642
643 /* brw_eu_validate.c */
644 bool brw_validate_instructions(const struct gen_device_info *devinfo,
645 const void *assembly, int start_offset, int end_offset,
646 struct disasm_info *disasm);
647
648 static inline int
649 next_offset(const struct gen_device_info *devinfo, void *store, int offset)
650 {
651 brw_inst *insn = (brw_inst *)((char *)store + offset);
652
653 if (brw_inst_cmpt_control(devinfo, insn))
654 return offset + 8;
655 else
656 return offset + 16;
657 }
658
659 struct opcode_desc {
660 /* The union is an implementation detail used by brw_opcode_desc() to handle
661 * opcodes that have been reused for different instructions across hardware
662 * generations.
663 *
664 * The gens field acts as a tag. If it is non-zero, name points to a string
665 * containing the instruction mnemonic. If it is zero, the table field is
666 * valid and either points to a secondary opcode_desc table with 'size'
667 * elements or is NULL and no such instruction exists for the opcode.
668 */
669 union {
670 struct {
671 char *name;
672 int nsrc;
673 };
674 struct {
675 const struct opcode_desc *table;
676 unsigned size;
677 };
678 };
679 int ndst;
680 int gens;
681 };
682
683 const struct opcode_desc *
684 brw_opcode_desc(const struct gen_device_info *devinfo, enum opcode opcode);
685
686 static inline bool
687 is_3src(const struct gen_device_info *devinfo, enum opcode opcode)
688 {
689 const struct opcode_desc *desc = brw_opcode_desc(devinfo, opcode);
690 return desc && desc->nsrc == 3;
691 }
692
693 /** Maximum SEND message length */
694 #define BRW_MAX_MSG_LENGTH 15
695
696 /** First MRF register used by pull loads */
697 #define FIRST_SPILL_MRF(gen) ((gen) == 6 ? 21 : 13)
698
699 /** First MRF register used by spills */
700 #define FIRST_PULL_LOAD_MRF(gen) ((gen) == 6 ? 16 : 13)
701
702 #ifdef __cplusplus
703 }
704 #endif
705
706 #endif