tgsi: Allow TEMPORARY registers as indirect address into source operands.
[mesa.git] / src / gallium / auxiliary / tgsi / tgsi_ureg.c
1 /**************************************************************************
2 *
3 * Copyright 2009 VMware, Inc.
4 * All Rights Reserved.
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a
7 * 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, sub license, 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 portions
16 * of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
21 * IN NO EVENT SHALL VMWARE, INC AND/OR ITS SUPPLIERS BE LIABLE FOR
22 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
23 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
24 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
25 *
26 **************************************************************************/
27
28
29 #include "pipe/p_context.h"
30 #include "pipe/p_state.h"
31 #include "tgsi/tgsi_ureg.h"
32 #include "tgsi/tgsi_build.h"
33 #include "tgsi/tgsi_info.h"
34 #include "tgsi/tgsi_dump.h"
35 #include "tgsi/tgsi_sanity.h"
36 #include "util/u_memory.h"
37 #include "util/u_math.h"
38
39 union tgsi_any_token {
40 struct tgsi_header header;
41 struct tgsi_processor processor;
42 struct tgsi_token token;
43 struct tgsi_property prop;
44 struct tgsi_property_data prop_data;
45 struct tgsi_declaration decl;
46 struct tgsi_declaration_range decl_range;
47 struct tgsi_declaration_semantic decl_semantic;
48 struct tgsi_immediate imm;
49 union tgsi_immediate_data imm_data;
50 struct tgsi_instruction insn;
51 struct tgsi_instruction_predicate insn_predicate;
52 struct tgsi_instruction_label insn_label;
53 struct tgsi_instruction_texture insn_texture;
54 struct tgsi_src_register src;
55 struct tgsi_dimension dim;
56 struct tgsi_dst_register dst;
57 unsigned value;
58 };
59
60
61 struct ureg_tokens {
62 union tgsi_any_token *tokens;
63 unsigned size;
64 unsigned order;
65 unsigned count;
66 };
67
68 #define UREG_MAX_INPUT PIPE_MAX_ATTRIBS
69 #define UREG_MAX_SYSTEM_VALUE PIPE_MAX_ATTRIBS
70 #define UREG_MAX_OUTPUT PIPE_MAX_ATTRIBS
71 #define UREG_MAX_CONSTANT_RANGE 32
72 #define UREG_MAX_IMMEDIATE 32
73 #define UREG_MAX_TEMP 256
74 #define UREG_MAX_ADDR 2
75 #define UREG_MAX_LOOP 1
76 #define UREG_MAX_PRED 1
77
78 #define DOMAIN_DECL 0
79 #define DOMAIN_INSN 1
80
81 struct ureg_program
82 {
83 unsigned processor;
84 struct pipe_context *pipe;
85
86 struct {
87 unsigned semantic_name;
88 unsigned semantic_index;
89 unsigned interp;
90 } fs_input[UREG_MAX_INPUT];
91 unsigned nr_fs_inputs;
92
93 unsigned vs_inputs[UREG_MAX_INPUT/32];
94
95 struct {
96 unsigned index;
97 } gs_input[UREG_MAX_INPUT];
98 unsigned nr_gs_inputs;
99
100 struct {
101 unsigned index;
102 unsigned semantic_name;
103 unsigned semantic_index;
104 } system_value[UREG_MAX_SYSTEM_VALUE];
105 unsigned nr_system_values;
106
107 struct {
108 unsigned semantic_name;
109 unsigned semantic_index;
110 } output[UREG_MAX_OUTPUT];
111 unsigned nr_outputs;
112
113 struct {
114 union {
115 float f[4];
116 unsigned u[4];
117 int i[4];
118 } value;
119 unsigned nr;
120 unsigned type;
121 } immediate[UREG_MAX_IMMEDIATE];
122 unsigned nr_immediates;
123
124 struct ureg_src sampler[PIPE_MAX_SAMPLERS];
125 unsigned nr_samplers;
126
127 unsigned temps_active[UREG_MAX_TEMP / 32];
128 unsigned nr_temps;
129
130 struct {
131 unsigned first;
132 unsigned last;
133 } constant_range[UREG_MAX_CONSTANT_RANGE];
134 unsigned nr_constant_ranges;
135
136 unsigned property_gs_input_prim;
137
138 unsigned nr_addrs;
139 unsigned nr_preds;
140 unsigned nr_loops;
141 unsigned nr_instructions;
142
143 struct ureg_tokens domain[2];
144 };
145
146 static union tgsi_any_token error_tokens[32];
147
148 static void tokens_error( struct ureg_tokens *tokens )
149 {
150 if (tokens->tokens && tokens->tokens != error_tokens)
151 FREE(tokens->tokens);
152
153 tokens->tokens = error_tokens;
154 tokens->size = Elements(error_tokens);
155 tokens->count = 0;
156 }
157
158
159 static void tokens_expand( struct ureg_tokens *tokens,
160 unsigned count )
161 {
162 unsigned old_size = tokens->size * sizeof(unsigned);
163
164 if (tokens->tokens == error_tokens) {
165 return;
166 }
167
168 while (tokens->count + count > tokens->size) {
169 tokens->size = (1 << ++tokens->order);
170 }
171
172 tokens->tokens = REALLOC(tokens->tokens,
173 old_size,
174 tokens->size * sizeof(unsigned));
175 if (tokens->tokens == NULL) {
176 tokens_error(tokens);
177 }
178 }
179
180 static void set_bad( struct ureg_program *ureg )
181 {
182 tokens_error(&ureg->domain[0]);
183 }
184
185
186
187 static union tgsi_any_token *get_tokens( struct ureg_program *ureg,
188 unsigned domain,
189 unsigned count )
190 {
191 struct ureg_tokens *tokens = &ureg->domain[domain];
192 union tgsi_any_token *result;
193
194 if (tokens->count + count > tokens->size)
195 tokens_expand(tokens, count);
196
197 result = &tokens->tokens[tokens->count];
198 tokens->count += count;
199 return result;
200 }
201
202
203 static union tgsi_any_token *retrieve_token( struct ureg_program *ureg,
204 unsigned domain,
205 unsigned nr )
206 {
207 if (ureg->domain[domain].tokens == error_tokens)
208 return &error_tokens[0];
209
210 return &ureg->domain[domain].tokens[nr];
211 }
212
213
214
215 static INLINE struct ureg_dst
216 ureg_dst_register( unsigned file,
217 unsigned index )
218 {
219 struct ureg_dst dst;
220
221 dst.File = file;
222 dst.WriteMask = TGSI_WRITEMASK_XYZW;
223 dst.Indirect = 0;
224 dst.IndirectIndex = 0;
225 dst.IndirectSwizzle = 0;
226 dst.Saturate = 0;
227 dst.Predicate = 0;
228 dst.PredNegate = 0;
229 dst.PredSwizzleX = TGSI_SWIZZLE_X;
230 dst.PredSwizzleY = TGSI_SWIZZLE_Y;
231 dst.PredSwizzleZ = TGSI_SWIZZLE_Z;
232 dst.PredSwizzleW = TGSI_SWIZZLE_W;
233 dst.Index = index;
234
235 return dst;
236 }
237
238 static INLINE struct ureg_src
239 ureg_src_register( unsigned file,
240 unsigned index )
241 {
242 struct ureg_src src;
243
244 src.File = file;
245 src.SwizzleX = TGSI_SWIZZLE_X;
246 src.SwizzleY = TGSI_SWIZZLE_Y;
247 src.SwizzleZ = TGSI_SWIZZLE_Z;
248 src.SwizzleW = TGSI_SWIZZLE_W;
249 src.Indirect = 0;
250 src.IndirectFile = TGSI_FILE_NULL;
251 src.IndirectIndex = 0;
252 src.IndirectSwizzle = 0;
253 src.Absolute = 0;
254 src.Index = index;
255 src.Negate = 0;
256 src.Dimension = 0;
257 src.DimensionIndex = 0;
258
259 return src;
260 }
261
262
263
264 void
265 ureg_property_gs_input_prim(struct ureg_program *ureg,
266 unsigned gs_input_prim)
267 {
268 ureg->property_gs_input_prim = gs_input_prim;
269 }
270
271
272
273 struct ureg_src
274 ureg_DECL_fs_input( struct ureg_program *ureg,
275 unsigned name,
276 unsigned index,
277 unsigned interp_mode )
278 {
279 unsigned i;
280
281 for (i = 0; i < ureg->nr_fs_inputs; i++) {
282 if (ureg->fs_input[i].semantic_name == name &&
283 ureg->fs_input[i].semantic_index == index)
284 goto out;
285 }
286
287 if (ureg->nr_fs_inputs < UREG_MAX_INPUT) {
288 ureg->fs_input[i].semantic_name = name;
289 ureg->fs_input[i].semantic_index = index;
290 ureg->fs_input[i].interp = interp_mode;
291 ureg->nr_fs_inputs++;
292 }
293 else {
294 set_bad( ureg );
295 }
296
297 out:
298 return ureg_src_register( TGSI_FILE_INPUT, i );
299 }
300
301
302 struct ureg_src
303 ureg_DECL_vs_input( struct ureg_program *ureg,
304 unsigned index )
305 {
306 assert(ureg->processor == TGSI_PROCESSOR_VERTEX);
307
308 ureg->vs_inputs[index/32] |= 1 << (index % 32);
309 return ureg_src_register( TGSI_FILE_INPUT, index );
310 }
311
312
313 struct ureg_src
314 ureg_DECL_gs_input(struct ureg_program *ureg,
315 unsigned index)
316 {
317 if (ureg->nr_gs_inputs < UREG_MAX_INPUT) {
318 ureg->gs_input[ureg->nr_gs_inputs].index = index;
319 ureg->nr_gs_inputs++;
320 } else {
321 set_bad(ureg);
322 }
323
324 /* XXX: Add suport for true 2D input registers. */
325 return ureg_src_register(TGSI_FILE_INPUT, index);
326 }
327
328
329 struct ureg_src
330 ureg_DECL_system_value(struct ureg_program *ureg,
331 unsigned index,
332 unsigned semantic_name,
333 unsigned semantic_index)
334 {
335 if (ureg->nr_system_values < UREG_MAX_SYSTEM_VALUE) {
336 ureg->system_value[ureg->nr_system_values].index = index;
337 ureg->system_value[ureg->nr_system_values].semantic_name = semantic_name;
338 ureg->system_value[ureg->nr_system_values].semantic_index = semantic_index;
339 ureg->nr_system_values++;
340 } else {
341 set_bad(ureg);
342 }
343
344 return ureg_src_register(TGSI_FILE_SYSTEM_VALUE, index);
345 }
346
347
348 struct ureg_dst
349 ureg_DECL_output( struct ureg_program *ureg,
350 unsigned name,
351 unsigned index )
352 {
353 unsigned i;
354
355 for (i = 0; i < ureg->nr_outputs; i++) {
356 if (ureg->output[i].semantic_name == name &&
357 ureg->output[i].semantic_index == index)
358 goto out;
359 }
360
361 if (ureg->nr_outputs < UREG_MAX_OUTPUT) {
362 ureg->output[i].semantic_name = name;
363 ureg->output[i].semantic_index = index;
364 ureg->nr_outputs++;
365 }
366 else {
367 set_bad( ureg );
368 }
369
370 out:
371 return ureg_dst_register( TGSI_FILE_OUTPUT, i );
372 }
373
374
375 /* Returns a new constant register. Keep track of which have been
376 * referred to so that we can emit decls later.
377 *
378 * There is nothing in this code to bind this constant to any tracked
379 * value or manage any constant_buffer contents -- that's the
380 * resposibility of the calling code.
381 */
382 struct ureg_src ureg_DECL_constant(struct ureg_program *ureg,
383 unsigned index )
384 {
385 unsigned minconst = index, maxconst = index;
386 unsigned i;
387
388 /* Inside existing range?
389 */
390 for (i = 0; i < ureg->nr_constant_ranges; i++) {
391 if (ureg->constant_range[i].first <= index &&
392 ureg->constant_range[i].last >= index)
393 goto out;
394 }
395
396 /* Extend existing range?
397 */
398 for (i = 0; i < ureg->nr_constant_ranges; i++) {
399 if (ureg->constant_range[i].last == index - 1) {
400 ureg->constant_range[i].last = index;
401 goto out;
402 }
403
404 if (ureg->constant_range[i].first == index + 1) {
405 ureg->constant_range[i].first = index;
406 goto out;
407 }
408
409 minconst = MIN2(minconst, ureg->constant_range[i].first);
410 maxconst = MAX2(maxconst, ureg->constant_range[i].last);
411 }
412
413 /* Create new range?
414 */
415 if (ureg->nr_constant_ranges < UREG_MAX_CONSTANT_RANGE) {
416 i = ureg->nr_constant_ranges++;
417 ureg->constant_range[i].first = index;
418 ureg->constant_range[i].last = index;
419 goto out;
420 }
421
422 /* Collapse all ranges down to one:
423 */
424 i = 0;
425 ureg->constant_range[0].first = minconst;
426 ureg->constant_range[0].last = maxconst;
427 ureg->nr_constant_ranges = 1;
428
429 out:
430 assert(i < ureg->nr_constant_ranges);
431 assert(ureg->constant_range[i].first <= index);
432 assert(ureg->constant_range[i].last >= index);
433 return ureg_src_register( TGSI_FILE_CONSTANT, index );
434 }
435
436
437 /* Allocate a new temporary. Temporaries greater than UREG_MAX_TEMP
438 * are legal, but will not be released.
439 */
440 struct ureg_dst ureg_DECL_temporary( struct ureg_program *ureg )
441 {
442 unsigned i;
443
444 for (i = 0; i < UREG_MAX_TEMP; i += 32) {
445 int bit = ffs(~ureg->temps_active[i/32]);
446 if (bit != 0) {
447 i += bit - 1;
448 goto out;
449 }
450 }
451
452 /* No reusable temps, so allocate a new one:
453 */
454 i = ureg->nr_temps++;
455
456 out:
457 if (i < UREG_MAX_TEMP)
458 ureg->temps_active[i/32] |= 1 << (i % 32);
459
460 if (i >= ureg->nr_temps)
461 ureg->nr_temps = i + 1;
462
463 return ureg_dst_register( TGSI_FILE_TEMPORARY, i );
464 }
465
466
467 void ureg_release_temporary( struct ureg_program *ureg,
468 struct ureg_dst tmp )
469 {
470 if(tmp.File == TGSI_FILE_TEMPORARY)
471 if (tmp.Index < UREG_MAX_TEMP)
472 ureg->temps_active[tmp.Index/32] &= ~(1 << (tmp.Index % 32));
473 }
474
475
476 /* Allocate a new address register.
477 */
478 struct ureg_dst ureg_DECL_address( struct ureg_program *ureg )
479 {
480 if (ureg->nr_addrs < UREG_MAX_ADDR)
481 return ureg_dst_register( TGSI_FILE_ADDRESS, ureg->nr_addrs++ );
482
483 assert( 0 );
484 return ureg_dst_register( TGSI_FILE_ADDRESS, 0 );
485 }
486
487 /* Allocate a new loop register.
488 */
489 struct ureg_dst
490 ureg_DECL_loop(struct ureg_program *ureg)
491 {
492 if (ureg->nr_loops < UREG_MAX_LOOP) {
493 return ureg_dst_register(TGSI_FILE_LOOP, ureg->nr_loops++);
494 }
495
496 assert(0);
497 return ureg_dst_register(TGSI_FILE_LOOP, 0);
498 }
499
500 /* Allocate a new predicate register.
501 */
502 struct ureg_dst
503 ureg_DECL_predicate(struct ureg_program *ureg)
504 {
505 if (ureg->nr_preds < UREG_MAX_PRED) {
506 return ureg_dst_register(TGSI_FILE_PREDICATE, ureg->nr_preds++);
507 }
508
509 assert(0);
510 return ureg_dst_register(TGSI_FILE_PREDICATE, 0);
511 }
512
513 /* Allocate a new sampler.
514 */
515 struct ureg_src ureg_DECL_sampler( struct ureg_program *ureg,
516 unsigned nr )
517 {
518 unsigned i;
519
520 for (i = 0; i < ureg->nr_samplers; i++)
521 if (ureg->sampler[i].Index == nr)
522 return ureg->sampler[i];
523
524 if (i < PIPE_MAX_SAMPLERS) {
525 ureg->sampler[i] = ureg_src_register( TGSI_FILE_SAMPLER, nr );
526 ureg->nr_samplers++;
527 return ureg->sampler[i];
528 }
529
530 assert( 0 );
531 return ureg->sampler[0];
532 }
533
534
535 static int
536 match_or_expand_immediate( const unsigned *v,
537 unsigned nr,
538 unsigned *v2,
539 unsigned *pnr2,
540 unsigned *swizzle )
541 {
542 unsigned nr2 = *pnr2;
543 unsigned i, j;
544
545 *swizzle = 0;
546
547 for (i = 0; i < nr; i++) {
548 boolean found = FALSE;
549
550 for (j = 0; j < nr2 && !found; j++) {
551 if (v[i] == v2[j]) {
552 *swizzle |= j << (i * 2);
553 found = TRUE;
554 }
555 }
556
557 if (!found) {
558 if (nr2 >= 4) {
559 return FALSE;
560 }
561
562 v2[nr2] = v[i];
563 *swizzle |= nr2 << (i * 2);
564 nr2++;
565 }
566 }
567
568 /* Actually expand immediate only when fully succeeded.
569 */
570 *pnr2 = nr2;
571 return TRUE;
572 }
573
574
575 static struct ureg_src
576 decl_immediate( struct ureg_program *ureg,
577 const unsigned *v,
578 unsigned nr,
579 unsigned type )
580 {
581 unsigned i, j;
582 unsigned swizzle;
583
584 /* Could do a first pass where we examine all existing immediates
585 * without expanding.
586 */
587
588 for (i = 0; i < ureg->nr_immediates; i++) {
589 if (ureg->immediate[i].type != type) {
590 continue;
591 }
592 if (match_or_expand_immediate(v,
593 nr,
594 ureg->immediate[i].value.u,
595 &ureg->immediate[i].nr,
596 &swizzle)) {
597 goto out;
598 }
599 }
600
601 if (ureg->nr_immediates < UREG_MAX_IMMEDIATE) {
602 i = ureg->nr_immediates++;
603 ureg->immediate[i].type = type;
604 if (match_or_expand_immediate(v,
605 nr,
606 ureg->immediate[i].value.u,
607 &ureg->immediate[i].nr,
608 &swizzle)) {
609 goto out;
610 }
611 }
612
613 set_bad(ureg);
614
615 out:
616 /* Make sure that all referenced elements are from this immediate.
617 * Has the effect of making size-one immediates into scalars.
618 */
619 for (j = nr; j < 4; j++) {
620 swizzle |= (swizzle & 0x3) << (j * 2);
621 }
622
623 return ureg_swizzle(ureg_src_register(TGSI_FILE_IMMEDIATE, i),
624 (swizzle >> 0) & 0x3,
625 (swizzle >> 2) & 0x3,
626 (swizzle >> 4) & 0x3,
627 (swizzle >> 6) & 0x3);
628 }
629
630
631 struct ureg_src
632 ureg_DECL_immediate( struct ureg_program *ureg,
633 const float *v,
634 unsigned nr )
635 {
636 union {
637 float f[4];
638 unsigned u[4];
639 } fu;
640 unsigned int i;
641
642 for (i = 0; i < nr; i++) {
643 fu.f[i] = v[i];
644 }
645
646 return decl_immediate(ureg, fu.u, nr, TGSI_IMM_FLOAT32);
647 }
648
649
650 struct ureg_src
651 ureg_DECL_immediate_uint( struct ureg_program *ureg,
652 const unsigned *v,
653 unsigned nr )
654 {
655 return decl_immediate(ureg, v, nr, TGSI_IMM_UINT32);
656 }
657
658
659 struct ureg_src
660 ureg_DECL_immediate_block_uint( struct ureg_program *ureg,
661 const unsigned *v,
662 unsigned nr )
663 {
664 uint index;
665 uint i;
666
667 if (ureg->nr_immediates + (nr + 3) / 4 > UREG_MAX_IMMEDIATE) {
668 set_bad(ureg);
669 return ureg_src_register(TGSI_FILE_IMMEDIATE, 0);
670 }
671
672 index = ureg->nr_immediates;
673 ureg->nr_immediates += (nr + 3) / 4;
674
675 for (i = index; i < ureg->nr_immediates; i++) {
676 ureg->immediate[i].type = TGSI_IMM_UINT32;
677 ureg->immediate[i].nr = nr > 4 ? 4 : nr;
678 memcpy(ureg->immediate[i].value.u,
679 &v[(i - index) * 4],
680 ureg->immediate[i].nr * sizeof(uint));
681 nr -= 4;
682 }
683
684 return ureg_src_register(TGSI_FILE_IMMEDIATE, index);
685 }
686
687
688 struct ureg_src
689 ureg_DECL_immediate_int( struct ureg_program *ureg,
690 const int *v,
691 unsigned nr )
692 {
693 return decl_immediate(ureg, (const unsigned *)v, nr, TGSI_IMM_INT32);
694 }
695
696
697 void
698 ureg_emit_src( struct ureg_program *ureg,
699 struct ureg_src src )
700 {
701 unsigned size = 1 + (src.Indirect ? 1 : 0) + (src.Dimension ? 1 : 0);
702
703 union tgsi_any_token *out = get_tokens( ureg, DOMAIN_INSN, size );
704 unsigned n = 0;
705
706 assert(src.File != TGSI_FILE_NULL);
707 assert(src.File != TGSI_FILE_OUTPUT);
708 assert(src.File < TGSI_FILE_COUNT);
709
710 out[n].value = 0;
711 out[n].src.File = src.File;
712 out[n].src.SwizzleX = src.SwizzleX;
713 out[n].src.SwizzleY = src.SwizzleY;
714 out[n].src.SwizzleZ = src.SwizzleZ;
715 out[n].src.SwizzleW = src.SwizzleW;
716 out[n].src.Index = src.Index;
717 out[n].src.Negate = src.Negate;
718 out[0].src.Absolute = src.Absolute;
719 n++;
720
721 if (src.Indirect) {
722 out[0].src.Indirect = 1;
723 out[n].value = 0;
724 out[n].src.File = src.IndirectFile;
725 out[n].src.SwizzleX = src.IndirectSwizzle;
726 out[n].src.SwizzleY = src.IndirectSwizzle;
727 out[n].src.SwizzleZ = src.IndirectSwizzle;
728 out[n].src.SwizzleW = src.IndirectSwizzle;
729 out[n].src.Index = src.IndirectIndex;
730 n++;
731 }
732
733 if (src.Dimension) {
734 out[0].src.Dimension = 1;
735 out[n].dim.Indirect = 0;
736 out[n].dim.Dimension = 0;
737 out[n].dim.Padding = 0;
738 out[n].dim.Index = src.DimensionIndex;
739 n++;
740 }
741
742 assert(n == size);
743 }
744
745
746 void
747 ureg_emit_dst( struct ureg_program *ureg,
748 struct ureg_dst dst )
749 {
750 unsigned size = (1 +
751 (dst.Indirect ? 1 : 0));
752
753 union tgsi_any_token *out = get_tokens( ureg, DOMAIN_INSN, size );
754 unsigned n = 0;
755
756 assert(dst.File != TGSI_FILE_NULL);
757 assert(dst.File != TGSI_FILE_CONSTANT);
758 assert(dst.File != TGSI_FILE_INPUT);
759 assert(dst.File != TGSI_FILE_SAMPLER);
760 assert(dst.File != TGSI_FILE_IMMEDIATE);
761 assert(dst.File < TGSI_FILE_COUNT);
762
763 out[n].value = 0;
764 out[n].dst.File = dst.File;
765 out[n].dst.WriteMask = dst.WriteMask;
766 out[n].dst.Indirect = dst.Indirect;
767 out[n].dst.Index = dst.Index;
768 n++;
769
770 if (dst.Indirect) {
771 out[n].value = 0;
772 out[n].src.File = TGSI_FILE_ADDRESS;
773 out[n].src.SwizzleX = dst.IndirectSwizzle;
774 out[n].src.SwizzleY = dst.IndirectSwizzle;
775 out[n].src.SwizzleZ = dst.IndirectSwizzle;
776 out[n].src.SwizzleW = dst.IndirectSwizzle;
777 out[n].src.Index = dst.IndirectIndex;
778 n++;
779 }
780
781 assert(n == size);
782 }
783
784
785 static void validate( unsigned opcode,
786 unsigned nr_dst,
787 unsigned nr_src )
788 {
789 #ifdef DEBUG
790 const struct tgsi_opcode_info *info = tgsi_get_opcode_info( opcode );
791 assert(info);
792 if(info) {
793 assert(nr_dst == info->num_dst);
794 assert(nr_src == info->num_src);
795 }
796 #endif
797 }
798
799 struct ureg_emit_insn_result
800 ureg_emit_insn(struct ureg_program *ureg,
801 unsigned opcode,
802 boolean saturate,
803 boolean predicate,
804 boolean pred_negate,
805 unsigned pred_swizzle_x,
806 unsigned pred_swizzle_y,
807 unsigned pred_swizzle_z,
808 unsigned pred_swizzle_w,
809 unsigned num_dst,
810 unsigned num_src )
811 {
812 union tgsi_any_token *out;
813 uint count = predicate ? 2 : 1;
814 struct ureg_emit_insn_result result;
815
816 validate( opcode, num_dst, num_src );
817
818 out = get_tokens( ureg, DOMAIN_INSN, count );
819 out[0].insn = tgsi_default_instruction();
820 out[0].insn.Opcode = opcode;
821 out[0].insn.Saturate = saturate;
822 out[0].insn.NumDstRegs = num_dst;
823 out[0].insn.NumSrcRegs = num_src;
824
825 result.insn_token = ureg->domain[DOMAIN_INSN].count - count;
826 result.extended_token = result.insn_token;
827
828 if (predicate) {
829 out[0].insn.Predicate = 1;
830 out[1].insn_predicate = tgsi_default_instruction_predicate();
831 out[1].insn_predicate.Negate = pred_negate;
832 out[1].insn_predicate.SwizzleX = pred_swizzle_x;
833 out[1].insn_predicate.SwizzleY = pred_swizzle_y;
834 out[1].insn_predicate.SwizzleZ = pred_swizzle_z;
835 out[1].insn_predicate.SwizzleW = pred_swizzle_w;
836 }
837
838 ureg->nr_instructions++;
839
840 return result;
841 }
842
843
844 void
845 ureg_emit_label(struct ureg_program *ureg,
846 unsigned extended_token,
847 unsigned *label_token )
848 {
849 union tgsi_any_token *out, *insn;
850
851 if(!label_token)
852 return;
853
854 out = get_tokens( ureg, DOMAIN_INSN, 1 );
855 out[0].value = 0;
856
857 insn = retrieve_token( ureg, DOMAIN_INSN, extended_token );
858 insn->insn.Label = 1;
859
860 *label_token = ureg->domain[DOMAIN_INSN].count - 1;
861 }
862
863 /* Will return a number which can be used in a label to point to the
864 * next instruction to be emitted.
865 */
866 unsigned
867 ureg_get_instruction_number( struct ureg_program *ureg )
868 {
869 return ureg->nr_instructions;
870 }
871
872 /* Patch a given label (expressed as a token number) to point to a
873 * given instruction (expressed as an instruction number).
874 */
875 void
876 ureg_fixup_label(struct ureg_program *ureg,
877 unsigned label_token,
878 unsigned instruction_number )
879 {
880 union tgsi_any_token *out = retrieve_token( ureg, DOMAIN_INSN, label_token );
881
882 out->insn_label.Label = instruction_number;
883 }
884
885
886 void
887 ureg_emit_texture(struct ureg_program *ureg,
888 unsigned extended_token,
889 unsigned target )
890 {
891 union tgsi_any_token *out, *insn;
892
893 out = get_tokens( ureg, DOMAIN_INSN, 1 );
894 insn = retrieve_token( ureg, DOMAIN_INSN, extended_token );
895
896 insn->insn.Texture = 1;
897
898 out[0].value = 0;
899 out[0].insn_texture.Texture = target;
900 }
901
902
903 void
904 ureg_fixup_insn_size(struct ureg_program *ureg,
905 unsigned insn )
906 {
907 union tgsi_any_token *out = retrieve_token( ureg, DOMAIN_INSN, insn );
908
909 assert(out->insn.Type == TGSI_TOKEN_TYPE_INSTRUCTION);
910 out->insn.NrTokens = ureg->domain[DOMAIN_INSN].count - insn - 1;
911 }
912
913
914 void
915 ureg_insn(struct ureg_program *ureg,
916 unsigned opcode,
917 const struct ureg_dst *dst,
918 unsigned nr_dst,
919 const struct ureg_src *src,
920 unsigned nr_src )
921 {
922 struct ureg_emit_insn_result insn;
923 unsigned i;
924 boolean saturate;
925 boolean predicate;
926 boolean negate = FALSE;
927 unsigned swizzle[4] = { 0 };
928
929 saturate = nr_dst ? dst[0].Saturate : FALSE;
930 predicate = nr_dst ? dst[0].Predicate : FALSE;
931 if (predicate) {
932 negate = dst[0].PredNegate;
933 swizzle[0] = dst[0].PredSwizzleX;
934 swizzle[1] = dst[0].PredSwizzleY;
935 swizzle[2] = dst[0].PredSwizzleZ;
936 swizzle[3] = dst[0].PredSwizzleW;
937 }
938
939 insn = ureg_emit_insn(ureg,
940 opcode,
941 saturate,
942 predicate,
943 negate,
944 swizzle[0],
945 swizzle[1],
946 swizzle[2],
947 swizzle[3],
948 nr_dst,
949 nr_src);
950
951 for (i = 0; i < nr_dst; i++)
952 ureg_emit_dst( ureg, dst[i] );
953
954 for (i = 0; i < nr_src; i++)
955 ureg_emit_src( ureg, src[i] );
956
957 ureg_fixup_insn_size( ureg, insn.insn_token );
958 }
959
960 void
961 ureg_tex_insn(struct ureg_program *ureg,
962 unsigned opcode,
963 const struct ureg_dst *dst,
964 unsigned nr_dst,
965 unsigned target,
966 const struct ureg_src *src,
967 unsigned nr_src )
968 {
969 struct ureg_emit_insn_result insn;
970 unsigned i;
971 boolean saturate;
972 boolean predicate;
973 boolean negate = FALSE;
974 unsigned swizzle[4] = { 0 };
975
976 saturate = nr_dst ? dst[0].Saturate : FALSE;
977 predicate = nr_dst ? dst[0].Predicate : FALSE;
978 if (predicate) {
979 negate = dst[0].PredNegate;
980 swizzle[0] = dst[0].PredSwizzleX;
981 swizzle[1] = dst[0].PredSwizzleY;
982 swizzle[2] = dst[0].PredSwizzleZ;
983 swizzle[3] = dst[0].PredSwizzleW;
984 }
985
986 insn = ureg_emit_insn(ureg,
987 opcode,
988 saturate,
989 predicate,
990 negate,
991 swizzle[0],
992 swizzle[1],
993 swizzle[2],
994 swizzle[3],
995 nr_dst,
996 nr_src);
997
998 ureg_emit_texture( ureg, insn.extended_token, target );
999
1000 for (i = 0; i < nr_dst; i++)
1001 ureg_emit_dst( ureg, dst[i] );
1002
1003 for (i = 0; i < nr_src; i++)
1004 ureg_emit_src( ureg, src[i] );
1005
1006 ureg_fixup_insn_size( ureg, insn.insn_token );
1007 }
1008
1009
1010 void
1011 ureg_label_insn(struct ureg_program *ureg,
1012 unsigned opcode,
1013 const struct ureg_src *src,
1014 unsigned nr_src,
1015 unsigned *label_token )
1016 {
1017 struct ureg_emit_insn_result insn;
1018 unsigned i;
1019
1020 insn = ureg_emit_insn(ureg,
1021 opcode,
1022 FALSE,
1023 FALSE,
1024 FALSE,
1025 TGSI_SWIZZLE_X,
1026 TGSI_SWIZZLE_Y,
1027 TGSI_SWIZZLE_Z,
1028 TGSI_SWIZZLE_W,
1029 0,
1030 nr_src);
1031
1032 ureg_emit_label( ureg, insn.extended_token, label_token );
1033
1034 for (i = 0; i < nr_src; i++)
1035 ureg_emit_src( ureg, src[i] );
1036
1037 ureg_fixup_insn_size( ureg, insn.insn_token );
1038 }
1039
1040
1041
1042 static void emit_decl( struct ureg_program *ureg,
1043 unsigned file,
1044 unsigned index,
1045 unsigned semantic_name,
1046 unsigned semantic_index,
1047 unsigned interp )
1048 {
1049 union tgsi_any_token *out = get_tokens( ureg, DOMAIN_DECL, 3 );
1050
1051 out[0].value = 0;
1052 out[0].decl.Type = TGSI_TOKEN_TYPE_DECLARATION;
1053 out[0].decl.NrTokens = 3;
1054 out[0].decl.File = file;
1055 out[0].decl.UsageMask = TGSI_WRITEMASK_XYZW; /* FIXME! */
1056 out[0].decl.Interpolate = interp;
1057 out[0].decl.Semantic = 1;
1058
1059 out[1].value = 0;
1060 out[1].decl_range.First =
1061 out[1].decl_range.Last = index;
1062
1063 out[2].value = 0;
1064 out[2].decl_semantic.Name = semantic_name;
1065 out[2].decl_semantic.Index = semantic_index;
1066
1067 }
1068
1069
1070 static void emit_decl_range( struct ureg_program *ureg,
1071 unsigned file,
1072 unsigned first,
1073 unsigned count )
1074 {
1075 union tgsi_any_token *out = get_tokens( ureg, DOMAIN_DECL, 2 );
1076
1077 out[0].value = 0;
1078 out[0].decl.Type = TGSI_TOKEN_TYPE_DECLARATION;
1079 out[0].decl.NrTokens = 2;
1080 out[0].decl.File = file;
1081 out[0].decl.UsageMask = 0xf;
1082 out[0].decl.Interpolate = TGSI_INTERPOLATE_CONSTANT;
1083 out[0].decl.Semantic = 0;
1084
1085 out[1].value = 0;
1086 out[1].decl_range.First = first;
1087 out[1].decl_range.Last = first + count - 1;
1088 }
1089
1090 static void
1091 emit_immediate( struct ureg_program *ureg,
1092 const unsigned *v,
1093 unsigned type )
1094 {
1095 union tgsi_any_token *out = get_tokens( ureg, DOMAIN_DECL, 5 );
1096
1097 out[0].value = 0;
1098 out[0].imm.Type = TGSI_TOKEN_TYPE_IMMEDIATE;
1099 out[0].imm.NrTokens = 5;
1100 out[0].imm.DataType = type;
1101 out[0].imm.Padding = 0;
1102
1103 out[1].imm_data.Uint = v[0];
1104 out[2].imm_data.Uint = v[1];
1105 out[3].imm_data.Uint = v[2];
1106 out[4].imm_data.Uint = v[3];
1107 }
1108
1109 static void
1110 emit_property(struct ureg_program *ureg,
1111 unsigned name,
1112 unsigned data)
1113 {
1114 union tgsi_any_token *out = get_tokens(ureg, DOMAIN_DECL, 2);
1115
1116 out[0].value = 0;
1117 out[0].prop.Type = TGSI_TOKEN_TYPE_PROPERTY;
1118 out[0].prop.NrTokens = 2;
1119 out[0].prop.PropertyName = name;
1120
1121 out[1].prop_data.Data = data;
1122 }
1123
1124
1125 static void emit_decls( struct ureg_program *ureg )
1126 {
1127 unsigned i;
1128
1129 if (ureg->property_gs_input_prim != ~0) {
1130 assert(ureg->processor == TGSI_PROCESSOR_GEOMETRY);
1131
1132 emit_property(ureg,
1133 TGSI_PROPERTY_GS_INPUT_PRIM,
1134 ureg->property_gs_input_prim);
1135 }
1136
1137 if (ureg->processor == TGSI_PROCESSOR_VERTEX) {
1138 for (i = 0; i < UREG_MAX_INPUT; i++) {
1139 if (ureg->vs_inputs[i/32] & (1 << (i%32))) {
1140 emit_decl_range( ureg, TGSI_FILE_INPUT, i, 1 );
1141 }
1142 }
1143 } else if (ureg->processor == TGSI_PROCESSOR_FRAGMENT) {
1144 for (i = 0; i < ureg->nr_fs_inputs; i++) {
1145 emit_decl( ureg,
1146 TGSI_FILE_INPUT,
1147 i,
1148 ureg->fs_input[i].semantic_name,
1149 ureg->fs_input[i].semantic_index,
1150 ureg->fs_input[i].interp );
1151 }
1152 } else {
1153 for (i = 0; i < ureg->nr_gs_inputs; i++) {
1154 emit_decl_range(ureg,
1155 TGSI_FILE_INPUT,
1156 ureg->gs_input[i].index,
1157 1);
1158 }
1159 }
1160
1161 for (i = 0; i < ureg->nr_system_values; i++) {
1162 emit_decl(ureg,
1163 TGSI_FILE_SYSTEM_VALUE,
1164 ureg->system_value[i].index,
1165 ureg->system_value[i].semantic_name,
1166 ureg->system_value[i].semantic_index,
1167 TGSI_INTERPOLATE_CONSTANT);
1168 }
1169
1170 for (i = 0; i < ureg->nr_outputs; i++) {
1171 emit_decl( ureg,
1172 TGSI_FILE_OUTPUT,
1173 i,
1174 ureg->output[i].semantic_name,
1175 ureg->output[i].semantic_index,
1176 TGSI_INTERPOLATE_CONSTANT );
1177 }
1178
1179 for (i = 0; i < ureg->nr_samplers; i++) {
1180 emit_decl_range( ureg,
1181 TGSI_FILE_SAMPLER,
1182 ureg->sampler[i].Index, 1 );
1183 }
1184
1185 if (ureg->nr_constant_ranges) {
1186 for (i = 0; i < ureg->nr_constant_ranges; i++)
1187 emit_decl_range( ureg,
1188 TGSI_FILE_CONSTANT,
1189 ureg->constant_range[i].first,
1190 (ureg->constant_range[i].last + 1 -
1191 ureg->constant_range[i].first) );
1192 }
1193
1194 if (ureg->nr_temps) {
1195 emit_decl_range( ureg,
1196 TGSI_FILE_TEMPORARY,
1197 0, ureg->nr_temps );
1198 }
1199
1200 if (ureg->nr_addrs) {
1201 emit_decl_range( ureg,
1202 TGSI_FILE_ADDRESS,
1203 0, ureg->nr_addrs );
1204 }
1205
1206 if (ureg->nr_loops) {
1207 emit_decl_range(ureg,
1208 TGSI_FILE_LOOP,
1209 0,
1210 ureg->nr_loops);
1211 }
1212
1213 if (ureg->nr_preds) {
1214 emit_decl_range(ureg,
1215 TGSI_FILE_PREDICATE,
1216 0,
1217 ureg->nr_preds);
1218 }
1219
1220 for (i = 0; i < ureg->nr_immediates; i++) {
1221 emit_immediate( ureg,
1222 ureg->immediate[i].value.u,
1223 ureg->immediate[i].type );
1224 }
1225 }
1226
1227 /* Append the instruction tokens onto the declarations to build a
1228 * contiguous stream suitable to send to the driver.
1229 */
1230 static void copy_instructions( struct ureg_program *ureg )
1231 {
1232 unsigned nr_tokens = ureg->domain[DOMAIN_INSN].count;
1233 union tgsi_any_token *out = get_tokens( ureg,
1234 DOMAIN_DECL,
1235 nr_tokens );
1236
1237 memcpy(out,
1238 ureg->domain[DOMAIN_INSN].tokens,
1239 nr_tokens * sizeof out[0] );
1240 }
1241
1242
1243 static void
1244 fixup_header_size(struct ureg_program *ureg)
1245 {
1246 union tgsi_any_token *out = retrieve_token( ureg, DOMAIN_DECL, 0 );
1247
1248 out->header.BodySize = ureg->domain[DOMAIN_DECL].count - 2;
1249 }
1250
1251
1252 static void
1253 emit_header( struct ureg_program *ureg )
1254 {
1255 union tgsi_any_token *out = get_tokens( ureg, DOMAIN_DECL, 2 );
1256
1257 out[0].header.HeaderSize = 2;
1258 out[0].header.BodySize = 0;
1259
1260 out[1].processor.Processor = ureg->processor;
1261 out[1].processor.Padding = 0;
1262 }
1263
1264
1265 const struct tgsi_token *ureg_finalize( struct ureg_program *ureg )
1266 {
1267 const struct tgsi_token *tokens;
1268
1269 emit_header( ureg );
1270 emit_decls( ureg );
1271 copy_instructions( ureg );
1272 fixup_header_size( ureg );
1273
1274 if (ureg->domain[0].tokens == error_tokens ||
1275 ureg->domain[1].tokens == error_tokens) {
1276 debug_printf("%s: error in generated shader\n", __FUNCTION__);
1277 assert(0);
1278 return NULL;
1279 }
1280
1281 tokens = &ureg->domain[DOMAIN_DECL].tokens[0].token;
1282
1283 if (0) {
1284 debug_printf("%s: emitted shader %d tokens:\n", __FUNCTION__,
1285 ureg->domain[DOMAIN_DECL].count);
1286 tgsi_dump( tokens, 0 );
1287 }
1288
1289 #if DEBUG
1290 if (tokens && !tgsi_sanity_check(tokens)) {
1291 debug_printf("tgsi_ureg.c, sanity check failed on generated tokens:\n");
1292 tgsi_dump(tokens, 0);
1293 assert(0);
1294 }
1295 #endif
1296
1297
1298 return tokens;
1299 }
1300
1301
1302 void *ureg_create_shader( struct ureg_program *ureg,
1303 struct pipe_context *pipe )
1304 {
1305 struct pipe_shader_state state;
1306
1307 state.tokens = ureg_finalize(ureg);
1308 if(!state.tokens)
1309 return NULL;
1310
1311 if (ureg->processor == TGSI_PROCESSOR_VERTEX)
1312 return pipe->create_vs_state( pipe, &state );
1313 else
1314 return pipe->create_fs_state( pipe, &state );
1315 }
1316
1317
1318 const struct tgsi_token *ureg_get_tokens( struct ureg_program *ureg,
1319 unsigned *nr_tokens )
1320 {
1321 const struct tgsi_token *tokens;
1322
1323 ureg_finalize(ureg);
1324
1325 tokens = &ureg->domain[DOMAIN_DECL].tokens[0].token;
1326
1327 if (nr_tokens)
1328 *nr_tokens = ureg->domain[DOMAIN_DECL].size;
1329
1330 ureg->domain[DOMAIN_DECL].tokens = 0;
1331 ureg->domain[DOMAIN_DECL].size = 0;
1332 ureg->domain[DOMAIN_DECL].order = 0;
1333 ureg->domain[DOMAIN_DECL].count = 0;
1334
1335 return tokens;
1336 }
1337
1338
1339 struct ureg_program *ureg_create( unsigned processor )
1340 {
1341 struct ureg_program *ureg = CALLOC_STRUCT( ureg_program );
1342 if (ureg == NULL)
1343 return NULL;
1344
1345 ureg->processor = processor;
1346 ureg->property_gs_input_prim = ~0;
1347 return ureg;
1348 }
1349
1350
1351 void ureg_destroy( struct ureg_program *ureg )
1352 {
1353 unsigned i;
1354
1355 for (i = 0; i < Elements(ureg->domain); i++) {
1356 if (ureg->domain[i].tokens &&
1357 ureg->domain[i].tokens != error_tokens)
1358 FREE(ureg->domain[i].tokens);
1359 }
1360
1361 FREE(ureg);
1362 }