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