intel/eu: Return new instruction to caller from brw_fb_WRITE().
[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_message_descriptor(struct brw_codegen *p,
260 brw_inst *inst,
261 enum brw_message_target sfid,
262 unsigned msg_length,
263 unsigned response_length,
264 bool header_present,
265 bool end_of_thread);
266
267 void brw_set_dp_read_message(struct brw_codegen *p,
268 brw_inst *insn,
269 unsigned binding_table_index,
270 unsigned msg_control,
271 unsigned msg_type,
272 unsigned target_cache,
273 unsigned msg_length,
274 bool header_present,
275 unsigned response_length);
276
277 void brw_set_dp_write_message(struct brw_codegen *p,
278 brw_inst *insn,
279 unsigned binding_table_index,
280 unsigned msg_control,
281 unsigned msg_type,
282 unsigned target_cache,
283 unsigned msg_length,
284 bool header_present,
285 unsigned last_render_target,
286 unsigned response_length,
287 unsigned end_of_thread,
288 unsigned send_commit_msg);
289
290 void brw_urb_WRITE(struct brw_codegen *p,
291 struct brw_reg dest,
292 unsigned msg_reg_nr,
293 struct brw_reg src0,
294 enum brw_urb_write_flags flags,
295 unsigned msg_length,
296 unsigned response_length,
297 unsigned offset,
298 unsigned swizzle);
299
300 /**
301 * Send message to shared unit \p sfid with a possibly indirect descriptor \p
302 * desc. If \p desc is not an immediate it will be transparently loaded to an
303 * address register using an OR instruction. The returned instruction can be
304 * passed as argument to the usual brw_set_*_message() functions in order to
305 * specify any additional descriptor bits -- If \p desc is an immediate this
306 * will be the SEND instruction itself, otherwise it will be the OR
307 * instruction.
308 */
309 struct brw_inst *
310 brw_send_indirect_message(struct brw_codegen *p,
311 unsigned sfid,
312 struct brw_reg dst,
313 struct brw_reg payload,
314 struct brw_reg desc);
315
316 void brw_ff_sync(struct brw_codegen *p,
317 struct brw_reg dest,
318 unsigned msg_reg_nr,
319 struct brw_reg src0,
320 bool allocate,
321 unsigned response_length,
322 bool eot);
323
324 void brw_svb_write(struct brw_codegen *p,
325 struct brw_reg dest,
326 unsigned msg_reg_nr,
327 struct brw_reg src0,
328 unsigned binding_table_index,
329 bool send_commit_msg);
330
331 brw_inst *brw_fb_WRITE(struct brw_codegen *p,
332 struct brw_reg payload,
333 struct brw_reg implied_header,
334 unsigned msg_control,
335 unsigned binding_table_index,
336 unsigned msg_length,
337 unsigned response_length,
338 bool eot,
339 bool last_render_target,
340 bool header_present);
341
342 brw_inst *gen9_fb_READ(struct brw_codegen *p,
343 struct brw_reg dst,
344 struct brw_reg payload,
345 unsigned binding_table_index,
346 unsigned msg_length,
347 unsigned response_length,
348 bool per_sample);
349
350 void brw_SAMPLE(struct brw_codegen *p,
351 struct brw_reg dest,
352 unsigned msg_reg_nr,
353 struct brw_reg src0,
354 unsigned binding_table_index,
355 unsigned sampler,
356 unsigned msg_type,
357 unsigned response_length,
358 unsigned msg_length,
359 unsigned header_present,
360 unsigned simd_mode,
361 unsigned return_format);
362
363 void brw_adjust_sampler_state_pointer(struct brw_codegen *p,
364 struct brw_reg header,
365 struct brw_reg sampler_index);
366
367 void gen4_math(struct brw_codegen *p,
368 struct brw_reg dest,
369 unsigned function,
370 unsigned msg_reg_nr,
371 struct brw_reg src,
372 unsigned precision );
373
374 void gen6_math(struct brw_codegen *p,
375 struct brw_reg dest,
376 unsigned function,
377 struct brw_reg src0,
378 struct brw_reg src1);
379
380 void brw_oword_block_read(struct brw_codegen *p,
381 struct brw_reg dest,
382 struct brw_reg mrf,
383 uint32_t offset,
384 uint32_t bind_table_index);
385
386 unsigned brw_scratch_surface_idx(const struct brw_codegen *p);
387
388 void brw_oword_block_read_scratch(struct brw_codegen *p,
389 struct brw_reg dest,
390 struct brw_reg mrf,
391 int num_regs,
392 unsigned offset);
393
394 void brw_oword_block_write_scratch(struct brw_codegen *p,
395 struct brw_reg mrf,
396 int num_regs,
397 unsigned offset);
398
399 void gen7_block_read_scratch(struct brw_codegen *p,
400 struct brw_reg dest,
401 int num_regs,
402 unsigned offset);
403
404 void brw_shader_time_add(struct brw_codegen *p,
405 struct brw_reg payload,
406 uint32_t surf_index);
407
408 /**
409 * Return the generation-specific jump distance scaling factor.
410 *
411 * Given the number of instructions to jump, we need to scale by
412 * some number to obtain the actual jump distance to program in an
413 * instruction.
414 */
415 static inline unsigned
416 brw_jump_scale(const struct gen_device_info *devinfo)
417 {
418 /* Broadwell measures jump targets in bytes. */
419 if (devinfo->gen >= 8)
420 return 16;
421
422 /* Ironlake and later measure jump targets in 64-bit data chunks (in order
423 * (to support compaction), so each 128-bit instruction requires 2 chunks.
424 */
425 if (devinfo->gen >= 5)
426 return 2;
427
428 /* Gen4 simply uses the number of 128-bit instructions. */
429 return 1;
430 }
431
432 void brw_barrier(struct brw_codegen *p, struct brw_reg src);
433
434 /* If/else/endif. Works by manipulating the execution flags on each
435 * channel.
436 */
437 brw_inst *brw_IF(struct brw_codegen *p, unsigned execute_size);
438 brw_inst *gen6_IF(struct brw_codegen *p, enum brw_conditional_mod conditional,
439 struct brw_reg src0, struct brw_reg src1);
440
441 void brw_ELSE(struct brw_codegen *p);
442 void brw_ENDIF(struct brw_codegen *p);
443
444 /* DO/WHILE loops:
445 */
446 brw_inst *brw_DO(struct brw_codegen *p, unsigned execute_size);
447
448 brw_inst *brw_WHILE(struct brw_codegen *p);
449
450 brw_inst *brw_BREAK(struct brw_codegen *p);
451 brw_inst *brw_CONT(struct brw_codegen *p);
452 brw_inst *gen6_HALT(struct brw_codegen *p);
453
454 /* Forward jumps:
455 */
456 void brw_land_fwd_jump(struct brw_codegen *p, int jmp_insn_idx);
457
458 brw_inst *brw_JMPI(struct brw_codegen *p, struct brw_reg index,
459 unsigned predicate_control);
460
461 void brw_NOP(struct brw_codegen *p);
462
463 void brw_WAIT(struct brw_codegen *p);
464
465 /* Special case: there is never a destination, execution size will be
466 * taken from src0:
467 */
468 void brw_CMP(struct brw_codegen *p,
469 struct brw_reg dest,
470 unsigned conditional,
471 struct brw_reg src0,
472 struct brw_reg src1);
473
474 void
475 brw_untyped_atomic(struct brw_codegen *p,
476 struct brw_reg dst,
477 struct brw_reg payload,
478 struct brw_reg surface,
479 unsigned atomic_op,
480 unsigned msg_length,
481 bool response_expected,
482 bool header_present);
483
484 void
485 brw_untyped_surface_read(struct brw_codegen *p,
486 struct brw_reg dst,
487 struct brw_reg payload,
488 struct brw_reg surface,
489 unsigned msg_length,
490 unsigned num_channels);
491
492 void
493 brw_untyped_surface_write(struct brw_codegen *p,
494 struct brw_reg payload,
495 struct brw_reg surface,
496 unsigned msg_length,
497 unsigned num_channels,
498 bool header_present);
499
500 void
501 brw_typed_atomic(struct brw_codegen *p,
502 struct brw_reg dst,
503 struct brw_reg payload,
504 struct brw_reg surface,
505 unsigned atomic_op,
506 unsigned msg_length,
507 bool response_expected,
508 bool header_present);
509
510 void
511 brw_typed_surface_read(struct brw_codegen *p,
512 struct brw_reg dst,
513 struct brw_reg payload,
514 struct brw_reg surface,
515 unsigned msg_length,
516 unsigned num_channels,
517 bool header_present);
518
519 void
520 brw_typed_surface_write(struct brw_codegen *p,
521 struct brw_reg payload,
522 struct brw_reg surface,
523 unsigned msg_length,
524 unsigned num_channels,
525 bool header_present);
526
527 void
528 brw_byte_scattered_read(struct brw_codegen *p,
529 struct brw_reg dst,
530 struct brw_reg payload,
531 struct brw_reg surface,
532 unsigned msg_length,
533 unsigned bit_size);
534
535 void
536 brw_byte_scattered_write(struct brw_codegen *p,
537 struct brw_reg payload,
538 struct brw_reg surface,
539 unsigned msg_length,
540 unsigned bit_size,
541 bool header_present);
542
543 void
544 brw_memory_fence(struct brw_codegen *p,
545 struct brw_reg dst,
546 enum opcode send_op);
547
548 void
549 brw_pixel_interpolator_query(struct brw_codegen *p,
550 struct brw_reg dest,
551 struct brw_reg mrf,
552 bool noperspective,
553 unsigned mode,
554 struct brw_reg data,
555 unsigned msg_length,
556 unsigned response_length);
557
558 void
559 brw_find_live_channel(struct brw_codegen *p,
560 struct brw_reg dst,
561 struct brw_reg mask);
562
563 void
564 brw_broadcast(struct brw_codegen *p,
565 struct brw_reg dst,
566 struct brw_reg src,
567 struct brw_reg idx);
568
569 void
570 brw_rounding_mode(struct brw_codegen *p,
571 enum brw_rnd_mode mode);
572
573 /***********************************************************************
574 * brw_eu_util.c:
575 */
576
577 void brw_copy_indirect_to_indirect(struct brw_codegen *p,
578 struct brw_indirect dst_ptr,
579 struct brw_indirect src_ptr,
580 unsigned count);
581
582 void brw_copy_from_indirect(struct brw_codegen *p,
583 struct brw_reg dst,
584 struct brw_indirect ptr,
585 unsigned count);
586
587 void brw_copy4(struct brw_codegen *p,
588 struct brw_reg dst,
589 struct brw_reg src,
590 unsigned count);
591
592 void brw_copy8(struct brw_codegen *p,
593 struct brw_reg dst,
594 struct brw_reg src,
595 unsigned count);
596
597 void brw_math_invert( struct brw_codegen *p,
598 struct brw_reg dst,
599 struct brw_reg src);
600
601 void brw_set_src1(struct brw_codegen *p, brw_inst *insn, struct brw_reg reg);
602
603 void brw_set_uip_jip(struct brw_codegen *p, int start_offset);
604
605 enum brw_conditional_mod brw_negate_cmod(uint32_t cmod);
606 enum brw_conditional_mod brw_swap_cmod(uint32_t cmod);
607
608 /* brw_eu_compact.c */
609 void brw_init_compaction_tables(const struct gen_device_info *devinfo);
610 void brw_compact_instructions(struct brw_codegen *p, int start_offset,
611 struct disasm_info *disasm);
612 void brw_uncompact_instruction(const struct gen_device_info *devinfo,
613 brw_inst *dst, brw_compact_inst *src);
614 bool brw_try_compact_instruction(const struct gen_device_info *devinfo,
615 brw_compact_inst *dst, const brw_inst *src);
616
617 void brw_debug_compact_uncompact(const struct gen_device_info *devinfo,
618 brw_inst *orig, brw_inst *uncompacted);
619
620 /* brw_eu_validate.c */
621 bool brw_validate_instructions(const struct gen_device_info *devinfo,
622 const void *assembly, int start_offset, int end_offset,
623 struct disasm_info *disasm);
624
625 static inline int
626 next_offset(const struct gen_device_info *devinfo, void *store, int offset)
627 {
628 brw_inst *insn = (brw_inst *)((char *)store + offset);
629
630 if (brw_inst_cmpt_control(devinfo, insn))
631 return offset + 8;
632 else
633 return offset + 16;
634 }
635
636 struct opcode_desc {
637 /* The union is an implementation detail used by brw_opcode_desc() to handle
638 * opcodes that have been reused for different instructions across hardware
639 * generations.
640 *
641 * The gens field acts as a tag. If it is non-zero, name points to a string
642 * containing the instruction mnemonic. If it is zero, the table field is
643 * valid and either points to a secondary opcode_desc table with 'size'
644 * elements or is NULL and no such instruction exists for the opcode.
645 */
646 union {
647 struct {
648 char *name;
649 int nsrc;
650 };
651 struct {
652 const struct opcode_desc *table;
653 unsigned size;
654 };
655 };
656 int ndst;
657 int gens;
658 };
659
660 const struct opcode_desc *
661 brw_opcode_desc(const struct gen_device_info *devinfo, enum opcode opcode);
662
663 static inline bool
664 is_3src(const struct gen_device_info *devinfo, enum opcode opcode)
665 {
666 const struct opcode_desc *desc = brw_opcode_desc(devinfo, opcode);
667 return desc && desc->nsrc == 3;
668 }
669
670 /** Maximum SEND message length */
671 #define BRW_MAX_MSG_LENGTH 15
672
673 /** First MRF register used by pull loads */
674 #define FIRST_SPILL_MRF(gen) ((gen) == 6 ? 21 : 13)
675
676 /** First MRF register used by spills */
677 #define FIRST_PULL_LOAD_MRF(gen) ((gen) == 6 ? 16 : 13)
678
679 #ifdef __cplusplus
680 }
681 #endif
682
683 #endif