intel/eu: Encode and decode native instruction opcodes from/to IR opcodes.
[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, enum brw_predicate 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 bool brw_try_override_assembly(struct brw_codegen *p, int start_offset,
167 const char *identifier);
168
169 brw_inst *brw_next_insn(struct brw_codegen *p, unsigned opcode);
170 void brw_set_dest(struct brw_codegen *p, brw_inst *insn, struct brw_reg dest);
171 void brw_set_src0(struct brw_codegen *p, brw_inst *insn, struct brw_reg reg);
172
173 void gen6_resolve_implied_move(struct brw_codegen *p,
174 struct brw_reg *src,
175 unsigned msg_reg_nr);
176
177 /* Helpers for regular instructions:
178 */
179 #define ALU1(OP) \
180 brw_inst *brw_##OP(struct brw_codegen *p, \
181 struct brw_reg dest, \
182 struct brw_reg src0);
183
184 #define ALU2(OP) \
185 brw_inst *brw_##OP(struct brw_codegen *p, \
186 struct brw_reg dest, \
187 struct brw_reg src0, \
188 struct brw_reg src1);
189
190 #define ALU3(OP) \
191 brw_inst *brw_##OP(struct brw_codegen *p, \
192 struct brw_reg dest, \
193 struct brw_reg src0, \
194 struct brw_reg src1, \
195 struct brw_reg src2);
196
197 #define ROUND(OP) \
198 void brw_##OP(struct brw_codegen *p, struct brw_reg dest, struct brw_reg src0);
199
200 ALU1(MOV)
201 ALU2(SEL)
202 ALU1(NOT)
203 ALU2(AND)
204 ALU2(OR)
205 ALU2(XOR)
206 ALU2(SHR)
207 ALU2(SHL)
208 ALU1(DIM)
209 ALU2(ASR)
210 ALU2(ROL)
211 ALU2(ROR)
212 ALU3(CSEL)
213 ALU1(F32TO16)
214 ALU1(F16TO32)
215 ALU2(ADD)
216 ALU2(AVG)
217 ALU2(MUL)
218 ALU1(FRC)
219 ALU1(RNDD)
220 ALU2(MAC)
221 ALU2(MACH)
222 ALU1(LZD)
223 ALU2(DP4)
224 ALU2(DPH)
225 ALU2(DP3)
226 ALU2(DP2)
227 ALU2(LINE)
228 ALU2(PLN)
229 ALU3(MAD)
230 ALU3(LRP)
231 ALU1(BFREV)
232 ALU3(BFE)
233 ALU2(BFI1)
234 ALU3(BFI2)
235 ALU1(FBH)
236 ALU1(FBL)
237 ALU1(CBIT)
238 ALU2(ADDC)
239 ALU2(SUBB)
240 ALU2(MAC)
241
242 ROUND(RNDZ)
243 ROUND(RNDE)
244
245 #undef ALU1
246 #undef ALU2
247 #undef ALU3
248 #undef ROUND
249
250
251 /* Helpers for SEND instruction:
252 */
253
254 /**
255 * Construct a message descriptor immediate with the specified common
256 * descriptor controls.
257 */
258 static inline uint32_t
259 brw_message_desc(const struct gen_device_info *devinfo,
260 unsigned msg_length,
261 unsigned response_length,
262 bool header_present)
263 {
264 if (devinfo->gen >= 5) {
265 return (SET_BITS(msg_length, 28, 25) |
266 SET_BITS(response_length, 24, 20) |
267 SET_BITS(header_present, 19, 19));
268 } else {
269 return (SET_BITS(msg_length, 23, 20) |
270 SET_BITS(response_length, 19, 16));
271 }
272 }
273
274 static inline unsigned
275 brw_message_desc_mlen(const struct gen_device_info *devinfo, uint32_t desc)
276 {
277 if (devinfo->gen >= 5)
278 return GET_BITS(desc, 28, 25);
279 else
280 return GET_BITS(desc, 23, 20);
281 }
282
283 static inline unsigned
284 brw_message_desc_rlen(const struct gen_device_info *devinfo, uint32_t desc)
285 {
286 if (devinfo->gen >= 5)
287 return GET_BITS(desc, 24, 20);
288 else
289 return GET_BITS(desc, 19, 16);
290 }
291
292 static inline bool
293 brw_message_desc_header_present(ASSERTED const struct gen_device_info *devinfo,
294 uint32_t desc)
295 {
296 assert(devinfo->gen >= 5);
297 return GET_BITS(desc, 19, 19);
298 }
299
300 static inline unsigned
301 brw_message_ex_desc(UNUSED const struct gen_device_info *devinfo,
302 unsigned ex_msg_length)
303 {
304 return SET_BITS(ex_msg_length, 9, 6);
305 }
306
307 static inline unsigned
308 brw_message_ex_desc_ex_mlen(UNUSED const struct gen_device_info *devinfo,
309 uint32_t ex_desc)
310 {
311 return GET_BITS(ex_desc, 9, 6);
312 }
313
314 /**
315 * Construct a message descriptor immediate with the specified sampler
316 * function controls.
317 */
318 static inline uint32_t
319 brw_sampler_desc(const struct gen_device_info *devinfo,
320 unsigned binding_table_index,
321 unsigned sampler,
322 unsigned msg_type,
323 unsigned simd_mode,
324 unsigned return_format)
325 {
326 const unsigned desc = (SET_BITS(binding_table_index, 7, 0) |
327 SET_BITS(sampler, 11, 8));
328 if (devinfo->gen >= 7)
329 return (desc | SET_BITS(msg_type, 16, 12) |
330 SET_BITS(simd_mode, 18, 17));
331 else if (devinfo->gen >= 5)
332 return (desc | SET_BITS(msg_type, 15, 12) |
333 SET_BITS(simd_mode, 17, 16));
334 else if (devinfo->is_g4x)
335 return desc | SET_BITS(msg_type, 15, 12);
336 else
337 return (desc | SET_BITS(return_format, 13, 12) |
338 SET_BITS(msg_type, 15, 14));
339 }
340
341 static inline unsigned
342 brw_sampler_desc_binding_table_index(UNUSED const struct gen_device_info *devinfo,
343 uint32_t desc)
344 {
345 return GET_BITS(desc, 7, 0);
346 }
347
348 static inline unsigned
349 brw_sampler_desc_sampler(UNUSED const struct gen_device_info *devinfo, uint32_t desc)
350 {
351 return GET_BITS(desc, 11, 8);
352 }
353
354 static inline unsigned
355 brw_sampler_desc_msg_type(const struct gen_device_info *devinfo, uint32_t desc)
356 {
357 if (devinfo->gen >= 7)
358 return GET_BITS(desc, 16, 12);
359 else if (devinfo->gen >= 5 || devinfo->is_g4x)
360 return GET_BITS(desc, 15, 12);
361 else
362 return GET_BITS(desc, 15, 14);
363 }
364
365 static inline unsigned
366 brw_sampler_desc_simd_mode(const struct gen_device_info *devinfo, uint32_t desc)
367 {
368 assert(devinfo->gen >= 5);
369 if (devinfo->gen >= 7)
370 return GET_BITS(desc, 18, 17);
371 else
372 return GET_BITS(desc, 17, 16);
373 }
374
375 static inline unsigned
376 brw_sampler_desc_return_format(ASSERTED const struct gen_device_info *devinfo,
377 uint32_t desc)
378 {
379 assert(devinfo->gen == 4 && !devinfo->is_g4x);
380 return GET_BITS(desc, 13, 12);
381 }
382
383 /**
384 * Construct a message descriptor for the dataport
385 */
386 static inline uint32_t
387 brw_dp_desc(const struct gen_device_info *devinfo,
388 unsigned binding_table_index,
389 unsigned msg_type,
390 unsigned msg_control)
391 {
392 /* Prior to gen6, things are too inconsistent; use the dp_read/write_desc
393 * helpers instead.
394 */
395 assert(devinfo->gen >= 6);
396 const unsigned desc = SET_BITS(binding_table_index, 7, 0);
397 if (devinfo->gen >= 8) {
398 return (desc | SET_BITS(msg_control, 13, 8) |
399 SET_BITS(msg_type, 18, 14));
400 } else if (devinfo->gen >= 7) {
401 return (desc | SET_BITS(msg_control, 13, 8) |
402 SET_BITS(msg_type, 17, 14));
403 } else {
404 return (desc | SET_BITS(msg_control, 12, 8) |
405 SET_BITS(msg_type, 16, 13));
406 }
407 }
408
409 static inline unsigned
410 brw_dp_desc_binding_table_index(UNUSED const struct gen_device_info *devinfo,
411 uint32_t desc)
412 {
413 return GET_BITS(desc, 7, 0);
414 }
415
416 static inline unsigned
417 brw_dp_desc_msg_type(const struct gen_device_info *devinfo, uint32_t desc)
418 {
419 assert(devinfo->gen >= 6);
420 if (devinfo->gen >= 8)
421 return GET_BITS(desc, 18, 14);
422 else if (devinfo->gen >= 7)
423 return GET_BITS(desc, 17, 14);
424 else
425 return GET_BITS(desc, 16, 13);
426 }
427
428 static inline unsigned
429 brw_dp_desc_msg_control(const struct gen_device_info *devinfo, uint32_t desc)
430 {
431 assert(devinfo->gen >= 6);
432 if (devinfo->gen >= 7)
433 return GET_BITS(desc, 13, 8);
434 else
435 return GET_BITS(desc, 12, 8);
436 }
437
438 /**
439 * Construct a message descriptor immediate with the specified dataport read
440 * function controls.
441 */
442 static inline uint32_t
443 brw_dp_read_desc(const struct gen_device_info *devinfo,
444 unsigned binding_table_index,
445 unsigned msg_control,
446 unsigned msg_type,
447 unsigned target_cache)
448 {
449 if (devinfo->gen >= 6)
450 return brw_dp_desc(devinfo, binding_table_index, msg_type, msg_control);
451 else if (devinfo->gen >= 5 || devinfo->is_g4x)
452 return (SET_BITS(binding_table_index, 7, 0) |
453 SET_BITS(msg_control, 10, 8) |
454 SET_BITS(msg_type, 13, 11) |
455 SET_BITS(target_cache, 15, 14));
456 else
457 return (SET_BITS(binding_table_index, 7, 0) |
458 SET_BITS(msg_control, 11, 8) |
459 SET_BITS(msg_type, 13, 12) |
460 SET_BITS(target_cache, 15, 14));
461 }
462
463 static inline unsigned
464 brw_dp_read_desc_msg_type(const struct gen_device_info *devinfo, uint32_t desc)
465 {
466 if (devinfo->gen >= 6)
467 return brw_dp_desc_msg_type(devinfo, desc);
468 else if (devinfo->gen >= 5 || devinfo->is_g4x)
469 return GET_BITS(desc, 13, 11);
470 else
471 return GET_BITS(desc, 13, 12);
472 }
473
474 static inline unsigned
475 brw_dp_read_desc_msg_control(const struct gen_device_info *devinfo,
476 uint32_t desc)
477 {
478 if (devinfo->gen >= 6)
479 return brw_dp_desc_msg_control(devinfo, desc);
480 else if (devinfo->gen >= 5 || devinfo->is_g4x)
481 return GET_BITS(desc, 10, 8);
482 else
483 return GET_BITS(desc, 11, 8);
484 }
485
486 /**
487 * Construct a message descriptor immediate with the specified dataport write
488 * function controls.
489 */
490 static inline uint32_t
491 brw_dp_write_desc(const struct gen_device_info *devinfo,
492 unsigned binding_table_index,
493 unsigned msg_control,
494 unsigned msg_type,
495 unsigned last_render_target,
496 unsigned send_commit_msg)
497 {
498 assert(devinfo->gen <= 6 || !send_commit_msg);
499 if (devinfo->gen >= 6)
500 return brw_dp_desc(devinfo, binding_table_index, msg_type, msg_control) |
501 SET_BITS(last_render_target, 12, 12) |
502 SET_BITS(send_commit_msg, 17, 17);
503 else
504 return (SET_BITS(binding_table_index, 7, 0) |
505 SET_BITS(msg_control, 11, 8) |
506 SET_BITS(last_render_target, 11, 11) |
507 SET_BITS(msg_type, 14, 12) |
508 SET_BITS(send_commit_msg, 15, 15));
509 }
510
511 static inline unsigned
512 brw_dp_write_desc_msg_type(const struct gen_device_info *devinfo,
513 uint32_t desc)
514 {
515 if (devinfo->gen >= 6)
516 return brw_dp_desc_msg_type(devinfo, desc);
517 else
518 return GET_BITS(desc, 14, 12);
519 }
520
521 static inline unsigned
522 brw_dp_write_desc_msg_control(const struct gen_device_info *devinfo,
523 uint32_t desc)
524 {
525 if (devinfo->gen >= 6)
526 return brw_dp_desc_msg_control(devinfo, desc);
527 else
528 return GET_BITS(desc, 11, 8);
529 }
530
531 static inline bool
532 brw_dp_write_desc_last_render_target(const struct gen_device_info *devinfo,
533 uint32_t desc)
534 {
535 if (devinfo->gen >= 6)
536 return GET_BITS(desc, 12, 12);
537 else
538 return GET_BITS(desc, 11, 11);
539 }
540
541 static inline bool
542 brw_dp_write_desc_write_commit(const struct gen_device_info *devinfo,
543 uint32_t desc)
544 {
545 assert(devinfo->gen <= 6);
546 if (devinfo->gen >= 6)
547 return GET_BITS(desc, 17, 17);
548 else
549 return GET_BITS(desc, 15, 15);
550 }
551
552 /**
553 * Construct a message descriptor immediate with the specified dataport
554 * surface function controls.
555 */
556 static inline uint32_t
557 brw_dp_surface_desc(const struct gen_device_info *devinfo,
558 unsigned msg_type,
559 unsigned msg_control)
560 {
561 assert(devinfo->gen >= 7);
562 /* We'll OR in the binding table index later */
563 return brw_dp_desc(devinfo, 0, msg_type, msg_control);
564 }
565
566 static inline uint32_t
567 brw_dp_untyped_atomic_desc(const struct gen_device_info *devinfo,
568 unsigned exec_size, /**< 0 for SIMD4x2 */
569 unsigned atomic_op,
570 bool response_expected)
571 {
572 assert(exec_size <= 8 || exec_size == 16);
573
574 unsigned msg_type;
575 if (devinfo->gen >= 8 || devinfo->is_haswell) {
576 if (exec_size > 0) {
577 msg_type = HSW_DATAPORT_DC_PORT1_UNTYPED_ATOMIC_OP;
578 } else {
579 msg_type = HSW_DATAPORT_DC_PORT1_UNTYPED_ATOMIC_OP_SIMD4X2;
580 }
581 } else {
582 msg_type = GEN7_DATAPORT_DC_UNTYPED_ATOMIC_OP;
583 }
584
585 const unsigned msg_control =
586 SET_BITS(atomic_op, 3, 0) |
587 SET_BITS(0 < exec_size && exec_size <= 8, 4, 4) |
588 SET_BITS(response_expected, 5, 5);
589
590 return brw_dp_surface_desc(devinfo, msg_type, msg_control);
591 }
592
593 static inline uint32_t
594 brw_dp_untyped_atomic_float_desc(const struct gen_device_info *devinfo,
595 unsigned exec_size,
596 unsigned atomic_op,
597 bool response_expected)
598 {
599 assert(exec_size <= 8 || exec_size == 16);
600 assert(devinfo->gen >= 9);
601
602 assert(exec_size > 0);
603 const unsigned msg_type = GEN9_DATAPORT_DC_PORT1_UNTYPED_ATOMIC_FLOAT_OP;
604
605 const unsigned msg_control =
606 SET_BITS(atomic_op, 1, 0) |
607 SET_BITS(exec_size <= 8, 4, 4) |
608 SET_BITS(response_expected, 5, 5);
609
610 return brw_dp_surface_desc(devinfo, msg_type, msg_control);
611 }
612
613 static inline unsigned
614 brw_mdc_cmask(unsigned num_channels)
615 {
616 /* See also MDC_CMASK in the SKL PRM Vol 2d. */
617 return 0xf & (0xf << num_channels);
618 }
619
620 static inline uint32_t
621 brw_dp_untyped_surface_rw_desc(const struct gen_device_info *devinfo,
622 unsigned exec_size, /**< 0 for SIMD4x2 */
623 unsigned num_channels,
624 bool write)
625 {
626 assert(exec_size <= 8 || exec_size == 16);
627
628 unsigned msg_type;
629 if (write) {
630 if (devinfo->gen >= 8 || devinfo->is_haswell) {
631 msg_type = HSW_DATAPORT_DC_PORT1_UNTYPED_SURFACE_WRITE;
632 } else {
633 msg_type = GEN7_DATAPORT_DC_UNTYPED_SURFACE_WRITE;
634 }
635 } else {
636 /* Read */
637 if (devinfo->gen >= 8 || devinfo->is_haswell) {
638 msg_type = HSW_DATAPORT_DC_PORT1_UNTYPED_SURFACE_READ;
639 } else {
640 msg_type = GEN7_DATAPORT_DC_UNTYPED_SURFACE_READ;
641 }
642 }
643
644 /* SIMD4x2 is only valid for read messages on IVB; use SIMD8 instead */
645 if (write && devinfo->gen == 7 && !devinfo->is_haswell && exec_size == 0)
646 exec_size = 8;
647
648 /* See also MDC_SM3 in the SKL PRM Vol 2d. */
649 const unsigned simd_mode = exec_size == 0 ? 0 : /* SIMD4x2 */
650 exec_size <= 8 ? 2 : 1;
651
652 const unsigned msg_control =
653 SET_BITS(brw_mdc_cmask(num_channels), 3, 0) |
654 SET_BITS(simd_mode, 5, 4);
655
656 return brw_dp_surface_desc(devinfo, msg_type, msg_control);
657 }
658
659 static inline unsigned
660 brw_mdc_ds(unsigned bit_size)
661 {
662 switch (bit_size) {
663 case 8:
664 return GEN7_BYTE_SCATTERED_DATA_ELEMENT_BYTE;
665 case 16:
666 return GEN7_BYTE_SCATTERED_DATA_ELEMENT_WORD;
667 case 32:
668 return GEN7_BYTE_SCATTERED_DATA_ELEMENT_DWORD;
669 default:
670 unreachable("Unsupported bit_size for byte scattered messages");
671 }
672 }
673
674 static inline uint32_t
675 brw_dp_byte_scattered_rw_desc(const struct gen_device_info *devinfo,
676 unsigned exec_size,
677 unsigned bit_size,
678 bool write)
679 {
680 assert(exec_size <= 8 || exec_size == 16);
681
682 assert(devinfo->gen > 7 || devinfo->is_haswell);
683 const unsigned msg_type =
684 write ? HSW_DATAPORT_DC_PORT0_BYTE_SCATTERED_WRITE :
685 HSW_DATAPORT_DC_PORT0_BYTE_SCATTERED_READ;
686
687 assert(exec_size > 0);
688 const unsigned msg_control =
689 SET_BITS(exec_size == 16, 0, 0) |
690 SET_BITS(brw_mdc_ds(bit_size), 3, 2);
691
692 return brw_dp_surface_desc(devinfo, msg_type, msg_control);
693 }
694
695 static inline uint32_t
696 brw_dp_a64_untyped_surface_rw_desc(const struct gen_device_info *devinfo,
697 unsigned exec_size, /**< 0 for SIMD4x2 */
698 unsigned num_channels,
699 bool write)
700 {
701 assert(exec_size <= 8 || exec_size == 16);
702 assert(devinfo->gen >= 8);
703
704 unsigned msg_type =
705 write ? GEN8_DATAPORT_DC_PORT1_A64_UNTYPED_SURFACE_WRITE :
706 GEN8_DATAPORT_DC_PORT1_A64_UNTYPED_SURFACE_READ;
707
708 /* See also MDC_SM3 in the SKL PRM Vol 2d. */
709 const unsigned simd_mode = exec_size == 0 ? 0 : /* SIMD4x2 */
710 exec_size <= 8 ? 2 : 1;
711
712 const unsigned msg_control =
713 SET_BITS(brw_mdc_cmask(num_channels), 3, 0) |
714 SET_BITS(simd_mode, 5, 4);
715
716 return brw_dp_desc(devinfo, BRW_BTI_STATELESS, msg_type, msg_control);
717 }
718
719 /**
720 * Calculate the data size (see MDC_A64_DS in the "Structures" volume of the
721 * Skylake PRM).
722 */
723 static inline uint32_t
724 brw_mdc_a64_ds(unsigned elems)
725 {
726 switch (elems) {
727 case 1: return 0;
728 case 2: return 1;
729 case 4: return 2;
730 case 8: return 3;
731 default:
732 unreachable("Unsupported elmeent count for A64 scattered message");
733 }
734 }
735
736 static inline uint32_t
737 brw_dp_a64_byte_scattered_rw_desc(const struct gen_device_info *devinfo,
738 unsigned exec_size, /**< 0 for SIMD4x2 */
739 unsigned bit_size,
740 bool write)
741 {
742 assert(exec_size <= 8 || exec_size == 16);
743 assert(devinfo->gen >= 8);
744
745 unsigned msg_type =
746 write ? GEN8_DATAPORT_DC_PORT1_A64_SCATTERED_WRITE :
747 GEN9_DATAPORT_DC_PORT1_A64_SCATTERED_READ;
748
749 const unsigned msg_control =
750 SET_BITS(GEN8_A64_SCATTERED_SUBTYPE_BYTE, 1, 0) |
751 SET_BITS(brw_mdc_a64_ds(bit_size / 8), 3, 2) |
752 SET_BITS(exec_size == 16, 4, 4);
753
754 return brw_dp_desc(devinfo, BRW_BTI_STATELESS, msg_type, msg_control);
755 }
756
757 static inline uint32_t
758 brw_dp_a64_untyped_atomic_desc(const struct gen_device_info *devinfo,
759 ASSERTED unsigned exec_size, /**< 0 for SIMD4x2 */
760 unsigned bit_size,
761 unsigned atomic_op,
762 bool response_expected)
763 {
764 assert(exec_size == 8);
765 assert(devinfo->gen >= 8);
766 assert(bit_size == 32 || bit_size == 64);
767
768 const unsigned msg_type = GEN8_DATAPORT_DC_PORT1_A64_UNTYPED_ATOMIC_OP;
769
770 const unsigned msg_control =
771 SET_BITS(atomic_op, 3, 0) |
772 SET_BITS(bit_size == 64, 4, 4) |
773 SET_BITS(response_expected, 5, 5);
774
775 return brw_dp_desc(devinfo, BRW_BTI_STATELESS, msg_type, msg_control);
776 }
777
778 static inline uint32_t
779 brw_dp_a64_untyped_atomic_float_desc(const struct gen_device_info *devinfo,
780 ASSERTED unsigned exec_size,
781 unsigned atomic_op,
782 bool response_expected)
783 {
784 assert(exec_size == 8);
785 assert(devinfo->gen >= 9);
786
787 assert(exec_size > 0);
788 const unsigned msg_type = GEN9_DATAPORT_DC_PORT1_A64_UNTYPED_ATOMIC_FLOAT_OP;
789
790 const unsigned msg_control =
791 SET_BITS(atomic_op, 1, 0) |
792 SET_BITS(response_expected, 5, 5);
793
794 return brw_dp_desc(devinfo, BRW_BTI_STATELESS, msg_type, msg_control);
795 }
796
797 static inline uint32_t
798 brw_dp_typed_atomic_desc(const struct gen_device_info *devinfo,
799 unsigned exec_size,
800 unsigned exec_group,
801 unsigned atomic_op,
802 bool response_expected)
803 {
804 assert(exec_size > 0 || exec_group == 0);
805 assert(exec_group % 8 == 0);
806
807 unsigned msg_type;
808 if (devinfo->gen >= 8 || devinfo->is_haswell) {
809 if (exec_size == 0) {
810 msg_type = HSW_DATAPORT_DC_PORT1_TYPED_ATOMIC_OP_SIMD4X2;
811 } else {
812 msg_type = HSW_DATAPORT_DC_PORT1_TYPED_ATOMIC_OP;
813 }
814 } else {
815 /* SIMD4x2 typed surface R/W messages only exist on HSW+ */
816 assert(exec_size > 0);
817 msg_type = GEN7_DATAPORT_RC_TYPED_ATOMIC_OP;
818 }
819
820 const bool high_sample_mask = (exec_group / 8) % 2 == 1;
821
822 const unsigned msg_control =
823 SET_BITS(atomic_op, 3, 0) |
824 SET_BITS(high_sample_mask, 4, 4) |
825 SET_BITS(response_expected, 5, 5);
826
827 return brw_dp_surface_desc(devinfo, msg_type, msg_control);
828 }
829
830 static inline uint32_t
831 brw_dp_typed_surface_rw_desc(const struct gen_device_info *devinfo,
832 unsigned exec_size,
833 unsigned exec_group,
834 unsigned num_channels,
835 bool write)
836 {
837 assert(exec_size > 0 || exec_group == 0);
838 assert(exec_group % 8 == 0);
839
840 /* Typed surface reads and writes don't support SIMD16 */
841 assert(exec_size <= 8);
842
843 unsigned msg_type;
844 if (write) {
845 if (devinfo->gen >= 8 || devinfo->is_haswell) {
846 msg_type = HSW_DATAPORT_DC_PORT1_TYPED_SURFACE_WRITE;
847 } else {
848 msg_type = GEN7_DATAPORT_RC_TYPED_SURFACE_WRITE;
849 }
850 } else {
851 if (devinfo->gen >= 8 || devinfo->is_haswell) {
852 msg_type = HSW_DATAPORT_DC_PORT1_TYPED_SURFACE_READ;
853 } else {
854 msg_type = GEN7_DATAPORT_RC_TYPED_SURFACE_READ;
855 }
856 }
857
858 /* See also MDC_SG3 in the SKL PRM Vol 2d. */
859 unsigned msg_control;
860 if (devinfo->gen >= 8 || devinfo->is_haswell) {
861 /* See also MDC_SG3 in the SKL PRM Vol 2d. */
862 const unsigned slot_group = exec_size == 0 ? 0 : /* SIMD4x2 */
863 1 + ((exec_group / 8) % 2);
864
865 msg_control =
866 SET_BITS(brw_mdc_cmask(num_channels), 3, 0) |
867 SET_BITS(slot_group, 5, 4);
868 } else {
869 /* SIMD4x2 typed surface R/W messages only exist on HSW+ */
870 assert(exec_size > 0);
871 const unsigned slot_group = ((exec_group / 8) % 2);
872
873 msg_control =
874 SET_BITS(brw_mdc_cmask(num_channels), 3, 0) |
875 SET_BITS(slot_group, 5, 5);
876 }
877
878 return brw_dp_surface_desc(devinfo, msg_type, msg_control);
879 }
880
881 /**
882 * Construct a message descriptor immediate with the specified pixel
883 * interpolator function controls.
884 */
885 static inline uint32_t
886 brw_pixel_interp_desc(UNUSED const struct gen_device_info *devinfo,
887 unsigned msg_type,
888 bool noperspective,
889 unsigned simd_mode,
890 unsigned slot_group)
891 {
892 return (SET_BITS(slot_group, 11, 11) |
893 SET_BITS(msg_type, 13, 12) |
894 SET_BITS(!!noperspective, 14, 14) |
895 SET_BITS(simd_mode, 16, 16));
896 }
897
898 void brw_urb_WRITE(struct brw_codegen *p,
899 struct brw_reg dest,
900 unsigned msg_reg_nr,
901 struct brw_reg src0,
902 enum brw_urb_write_flags flags,
903 unsigned msg_length,
904 unsigned response_length,
905 unsigned offset,
906 unsigned swizzle);
907
908 /**
909 * Send message to shared unit \p sfid with a possibly indirect descriptor \p
910 * desc. If \p desc is not an immediate it will be transparently loaded to an
911 * address register using an OR instruction.
912 */
913 void
914 brw_send_indirect_message(struct brw_codegen *p,
915 unsigned sfid,
916 struct brw_reg dst,
917 struct brw_reg payload,
918 struct brw_reg desc,
919 unsigned desc_imm,
920 bool eot);
921
922 void
923 brw_send_indirect_split_message(struct brw_codegen *p,
924 unsigned sfid,
925 struct brw_reg dst,
926 struct brw_reg payload0,
927 struct brw_reg payload1,
928 struct brw_reg desc,
929 unsigned desc_imm,
930 struct brw_reg ex_desc,
931 unsigned ex_desc_imm,
932 bool eot);
933
934 void brw_ff_sync(struct brw_codegen *p,
935 struct brw_reg dest,
936 unsigned msg_reg_nr,
937 struct brw_reg src0,
938 bool allocate,
939 unsigned response_length,
940 bool eot);
941
942 void brw_svb_write(struct brw_codegen *p,
943 struct brw_reg dest,
944 unsigned msg_reg_nr,
945 struct brw_reg src0,
946 unsigned binding_table_index,
947 bool send_commit_msg);
948
949 brw_inst *brw_fb_WRITE(struct brw_codegen *p,
950 struct brw_reg payload,
951 struct brw_reg implied_header,
952 unsigned msg_control,
953 unsigned binding_table_index,
954 unsigned msg_length,
955 unsigned response_length,
956 bool eot,
957 bool last_render_target,
958 bool header_present);
959
960 brw_inst *gen9_fb_READ(struct brw_codegen *p,
961 struct brw_reg dst,
962 struct brw_reg payload,
963 unsigned binding_table_index,
964 unsigned msg_length,
965 unsigned response_length,
966 bool per_sample);
967
968 void brw_SAMPLE(struct brw_codegen *p,
969 struct brw_reg dest,
970 unsigned msg_reg_nr,
971 struct brw_reg src0,
972 unsigned binding_table_index,
973 unsigned sampler,
974 unsigned msg_type,
975 unsigned response_length,
976 unsigned msg_length,
977 unsigned header_present,
978 unsigned simd_mode,
979 unsigned return_format);
980
981 void brw_adjust_sampler_state_pointer(struct brw_codegen *p,
982 struct brw_reg header,
983 struct brw_reg sampler_index);
984
985 void gen4_math(struct brw_codegen *p,
986 struct brw_reg dest,
987 unsigned function,
988 unsigned msg_reg_nr,
989 struct brw_reg src,
990 unsigned precision );
991
992 void gen6_math(struct brw_codegen *p,
993 struct brw_reg dest,
994 unsigned function,
995 struct brw_reg src0,
996 struct brw_reg src1);
997
998 void brw_oword_block_read(struct brw_codegen *p,
999 struct brw_reg dest,
1000 struct brw_reg mrf,
1001 uint32_t offset,
1002 uint32_t bind_table_index);
1003
1004 unsigned brw_scratch_surface_idx(const struct brw_codegen *p);
1005
1006 void brw_oword_block_read_scratch(struct brw_codegen *p,
1007 struct brw_reg dest,
1008 struct brw_reg mrf,
1009 int num_regs,
1010 unsigned offset);
1011
1012 void brw_oword_block_write_scratch(struct brw_codegen *p,
1013 struct brw_reg mrf,
1014 int num_regs,
1015 unsigned offset);
1016
1017 void gen7_block_read_scratch(struct brw_codegen *p,
1018 struct brw_reg dest,
1019 int num_regs,
1020 unsigned offset);
1021
1022 void brw_shader_time_add(struct brw_codegen *p,
1023 struct brw_reg payload,
1024 uint32_t surf_index);
1025
1026 /**
1027 * Return the generation-specific jump distance scaling factor.
1028 *
1029 * Given the number of instructions to jump, we need to scale by
1030 * some number to obtain the actual jump distance to program in an
1031 * instruction.
1032 */
1033 static inline unsigned
1034 brw_jump_scale(const struct gen_device_info *devinfo)
1035 {
1036 /* Broadwell measures jump targets in bytes. */
1037 if (devinfo->gen >= 8)
1038 return 16;
1039
1040 /* Ironlake and later measure jump targets in 64-bit data chunks (in order
1041 * (to support compaction), so each 128-bit instruction requires 2 chunks.
1042 */
1043 if (devinfo->gen >= 5)
1044 return 2;
1045
1046 /* Gen4 simply uses the number of 128-bit instructions. */
1047 return 1;
1048 }
1049
1050 void brw_barrier(struct brw_codegen *p, struct brw_reg src);
1051
1052 /* If/else/endif. Works by manipulating the execution flags on each
1053 * channel.
1054 */
1055 brw_inst *brw_IF(struct brw_codegen *p, unsigned execute_size);
1056 brw_inst *gen6_IF(struct brw_codegen *p, enum brw_conditional_mod conditional,
1057 struct brw_reg src0, struct brw_reg src1);
1058
1059 void brw_ELSE(struct brw_codegen *p);
1060 void brw_ENDIF(struct brw_codegen *p);
1061
1062 /* DO/WHILE loops:
1063 */
1064 brw_inst *brw_DO(struct brw_codegen *p, unsigned execute_size);
1065
1066 brw_inst *brw_WHILE(struct brw_codegen *p);
1067
1068 brw_inst *brw_BREAK(struct brw_codegen *p);
1069 brw_inst *brw_CONT(struct brw_codegen *p);
1070 brw_inst *gen6_HALT(struct brw_codegen *p);
1071
1072 /* Forward jumps:
1073 */
1074 void brw_land_fwd_jump(struct brw_codegen *p, int jmp_insn_idx);
1075
1076 brw_inst *brw_JMPI(struct brw_codegen *p, struct brw_reg index,
1077 unsigned predicate_control);
1078
1079 void brw_NOP(struct brw_codegen *p);
1080
1081 void brw_WAIT(struct brw_codegen *p);
1082
1083 /* Special case: there is never a destination, execution size will be
1084 * taken from src0:
1085 */
1086 void brw_CMP(struct brw_codegen *p,
1087 struct brw_reg dest,
1088 unsigned conditional,
1089 struct brw_reg src0,
1090 struct brw_reg src1);
1091
1092 void
1093 brw_untyped_atomic(struct brw_codegen *p,
1094 struct brw_reg dst,
1095 struct brw_reg payload,
1096 struct brw_reg surface,
1097 unsigned atomic_op,
1098 unsigned msg_length,
1099 bool response_expected,
1100 bool header_present);
1101
1102 void
1103 brw_untyped_surface_read(struct brw_codegen *p,
1104 struct brw_reg dst,
1105 struct brw_reg payload,
1106 struct brw_reg surface,
1107 unsigned msg_length,
1108 unsigned num_channels);
1109
1110 void
1111 brw_untyped_surface_write(struct brw_codegen *p,
1112 struct brw_reg payload,
1113 struct brw_reg surface,
1114 unsigned msg_length,
1115 unsigned num_channels,
1116 bool header_present);
1117
1118 void
1119 brw_memory_fence(struct brw_codegen *p,
1120 struct brw_reg dst,
1121 struct brw_reg src,
1122 enum opcode send_op,
1123 bool stall,
1124 unsigned bti);
1125
1126 void
1127 brw_pixel_interpolator_query(struct brw_codegen *p,
1128 struct brw_reg dest,
1129 struct brw_reg mrf,
1130 bool noperspective,
1131 unsigned mode,
1132 struct brw_reg data,
1133 unsigned msg_length,
1134 unsigned response_length);
1135
1136 void
1137 brw_find_live_channel(struct brw_codegen *p,
1138 struct brw_reg dst,
1139 struct brw_reg mask);
1140
1141 void
1142 brw_broadcast(struct brw_codegen *p,
1143 struct brw_reg dst,
1144 struct brw_reg src,
1145 struct brw_reg idx);
1146
1147 void
1148 brw_float_controls_mode(struct brw_codegen *p,
1149 unsigned mode, unsigned mask);
1150
1151 /***********************************************************************
1152 * brw_eu_util.c:
1153 */
1154
1155 void brw_copy_indirect_to_indirect(struct brw_codegen *p,
1156 struct brw_indirect dst_ptr,
1157 struct brw_indirect src_ptr,
1158 unsigned count);
1159
1160 void brw_copy_from_indirect(struct brw_codegen *p,
1161 struct brw_reg dst,
1162 struct brw_indirect ptr,
1163 unsigned count);
1164
1165 void brw_copy4(struct brw_codegen *p,
1166 struct brw_reg dst,
1167 struct brw_reg src,
1168 unsigned count);
1169
1170 void brw_copy8(struct brw_codegen *p,
1171 struct brw_reg dst,
1172 struct brw_reg src,
1173 unsigned count);
1174
1175 void brw_math_invert( struct brw_codegen *p,
1176 struct brw_reg dst,
1177 struct brw_reg src);
1178
1179 void brw_set_src1(struct brw_codegen *p, brw_inst *insn, struct brw_reg reg);
1180
1181 void brw_set_desc_ex(struct brw_codegen *p, brw_inst *insn,
1182 unsigned desc, unsigned ex_desc);
1183
1184 static inline void
1185 brw_set_desc(struct brw_codegen *p, brw_inst *insn, unsigned desc)
1186 {
1187 brw_set_desc_ex(p, insn, desc, 0);
1188 }
1189
1190 void brw_set_uip_jip(struct brw_codegen *p, int start_offset);
1191
1192 enum brw_conditional_mod brw_negate_cmod(enum brw_conditional_mod cmod);
1193 enum brw_conditional_mod brw_swap_cmod(enum brw_conditional_mod cmod);
1194
1195 /* brw_eu_compact.c */
1196 void brw_init_compaction_tables(const struct gen_device_info *devinfo);
1197 void brw_compact_instructions(struct brw_codegen *p, int start_offset,
1198 struct disasm_info *disasm);
1199 void brw_uncompact_instruction(const struct gen_device_info *devinfo,
1200 brw_inst *dst, brw_compact_inst *src);
1201 bool brw_try_compact_instruction(const struct gen_device_info *devinfo,
1202 brw_compact_inst *dst, const brw_inst *src);
1203
1204 void brw_debug_compact_uncompact(const struct gen_device_info *devinfo,
1205 brw_inst *orig, brw_inst *uncompacted);
1206
1207 /* brw_eu_validate.c */
1208 bool brw_validate_instructions(const struct gen_device_info *devinfo,
1209 const void *assembly, int start_offset, int end_offset,
1210 struct disasm_info *disasm);
1211
1212 static inline int
1213 next_offset(const struct gen_device_info *devinfo, void *store, int offset)
1214 {
1215 brw_inst *insn = (brw_inst *)((char *)store + offset);
1216
1217 if (brw_inst_cmpt_control(devinfo, insn))
1218 return offset + 8;
1219 else
1220 return offset + 16;
1221 }
1222
1223 struct opcode_desc {
1224 unsigned ir;
1225 unsigned hw;
1226 const char *name;
1227 int nsrc;
1228 int ndst;
1229 int gens;
1230 };
1231
1232 const struct opcode_desc *
1233 brw_opcode_desc(const struct gen_device_info *devinfo, enum opcode opcode);
1234
1235 const struct opcode_desc *
1236 brw_opcode_desc_from_hw(const struct gen_device_info *devinfo, unsigned hw);
1237
1238 static inline unsigned
1239 brw_opcode_encode(const struct gen_device_info *devinfo, enum opcode opcode)
1240 {
1241 return brw_opcode_desc(devinfo, opcode)->hw;
1242 }
1243
1244 static inline enum opcode
1245 brw_opcode_decode(const struct gen_device_info *devinfo, unsigned hw)
1246 {
1247 const struct opcode_desc *desc = brw_opcode_desc_from_hw(devinfo, hw);
1248 return desc ? (enum opcode)desc->ir : BRW_OPCODE_ILLEGAL;
1249 }
1250
1251 static inline void
1252 brw_inst_set_opcode(const struct gen_device_info *devinfo,
1253 brw_inst *inst, enum opcode opcode)
1254 {
1255 brw_inst_set_hw_opcode(devinfo, inst, brw_opcode_encode(devinfo, opcode));
1256 }
1257
1258 static inline enum opcode
1259 brw_inst_opcode(const struct gen_device_info *devinfo, const brw_inst *inst)
1260 {
1261 return brw_opcode_decode(devinfo, brw_inst_hw_opcode(devinfo, inst));
1262 }
1263
1264 static inline bool
1265 is_3src(const struct gen_device_info *devinfo, enum opcode opcode)
1266 {
1267 const struct opcode_desc *desc = brw_opcode_desc(devinfo, opcode);
1268 return desc && desc->nsrc == 3;
1269 }
1270
1271 /** Maximum SEND message length */
1272 #define BRW_MAX_MSG_LENGTH 15
1273
1274 /** First MRF register used by pull loads */
1275 #define FIRST_SPILL_MRF(gen) ((gen) == 6 ? 21 : 13)
1276
1277 /** First MRF register used by spills */
1278 #define FIRST_PULL_LOAD_MRF(gen) ((gen) == 6 ? 16 : 13)
1279
1280 #ifdef __cplusplus
1281 }
1282 #endif
1283
1284 #endif