tgsi_ureg.c revision 6d8dbd3d1ec888300fb0e9ac3cf61808ba8ecc2b 
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_info.h"
33#include "tgsi/tgsi_dump.h"
34#include "tgsi/tgsi_sanity.h"
35#include "util/u_memory.h"
36#include "util/u_math.h"
37
38union tgsi_any_token {
39   struct tgsi_version version;
40   struct tgsi_header header;
41   struct tgsi_processor processor;
42   struct tgsi_token token;
43   struct tgsi_declaration decl;
44   struct tgsi_declaration_range decl_range;
45   struct tgsi_declaration_semantic decl_semantic;
46   struct tgsi_immediate imm;
47   union  tgsi_immediate_data imm_data;
48   struct tgsi_instruction insn;
49   struct tgsi_instruction_ext_nv insn_ext_nv;
50   struct tgsi_instruction_ext_label insn_ext_label;
51   struct tgsi_instruction_ext_texture insn_ext_texture;
52   struct tgsi_instruction_ext_predicate insn_ext_predicate;
53   struct tgsi_src_register src;
54   struct tgsi_src_register_ext_swz src_ext_swz;
55   struct tgsi_src_register_ext_mod src_ext_mod;
56   struct tgsi_dimension dim;
57   struct tgsi_dst_register dst;
58   struct tgsi_dst_register_ext_concode dst_ext_code;
59   struct tgsi_dst_register_ext_modulate dst_ext_mod;
60   struct tgsi_dst_register_ext_predicate dst_ext_pred;
61   unsigned value;
62};
63
64
65struct ureg_tokens {
66   union tgsi_any_token *tokens;
67   unsigned size;
68   unsigned order;
69   unsigned count;
70};
71
72#define UREG_MAX_INPUT PIPE_MAX_ATTRIBS
73#define UREG_MAX_OUTPUT PIPE_MAX_ATTRIBS
74#define UREG_MAX_IMMEDIATE 32
75#define UREG_MAX_TEMP 256
76#define UREG_MAX_ADDR 2
77
78#define DOMAIN_DECL 0
79#define DOMAIN_INSN 1
80
81struct 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 semantic_name;
97      unsigned semantic_index;
98   } output[UREG_MAX_OUTPUT];
99   unsigned nr_outputs;
100
101   struct {
102      float v[4];
103      unsigned nr;
104   } immediate[UREG_MAX_IMMEDIATE];
105   unsigned nr_immediates;
106
107   struct ureg_src sampler[PIPE_MAX_SAMPLERS];
108   unsigned nr_samplers;
109
110   unsigned temps_active[UREG_MAX_TEMP / 32];
111   unsigned nr_temps;
112
113   unsigned nr_addrs;
114
115   unsigned nr_constants;
116   unsigned nr_instructions;
117
118   struct ureg_tokens domain[2];
119};
120
121static union tgsi_any_token error_tokens[32];
122
123static void tokens_error( struct ureg_tokens *tokens )
124{
125   if (tokens->tokens && tokens->tokens != error_tokens)
126      FREE(tokens->tokens);
127
128   tokens->tokens = error_tokens;
129   tokens->size = Elements(error_tokens);
130   tokens->count = 0;
131}
132
133
134static void tokens_expand( struct ureg_tokens *tokens,
135                           unsigned count )
136{
137   unsigned old_size = tokens->size * sizeof(unsigned);
138
139   if (tokens->tokens == error_tokens)
140      goto fail;
141
142   while (tokens->count + count > tokens->size) {
143      tokens->size = (1 << ++tokens->order);
144   }
145
146   tokens->tokens = REALLOC(tokens->tokens,
147                            old_size,
148                            tokens->size * sizeof(unsigned));
149   if (tokens->tokens == NULL)
150      goto fail;
151
152   return;
153
154fail:
155   tokens_error(tokens);
156}
157
158static void set_bad( struct ureg_program *ureg )
159{
160   tokens_error(&ureg->domain[0]);
161}
162
163
164
165static union tgsi_any_token *get_tokens( struct ureg_program *ureg,
166                                         unsigned domain,
167                                         unsigned count )
168{
169   struct ureg_tokens *tokens = &ureg->domain[domain];
170   union tgsi_any_token *result;
171
172   if (tokens->count + count > tokens->size)
173      tokens_expand(tokens, count);
174
175   result = &tokens->tokens[tokens->count];
176   tokens->count += count;
177   return result;
178}
179
180
181static union tgsi_any_token *retrieve_token( struct ureg_program *ureg,
182                                            unsigned domain,
183                                            unsigned nr )
184{
185   if (ureg->domain[domain].tokens == error_tokens)
186      return &error_tokens[0];
187
188   return &ureg->domain[domain].tokens[nr];
189}
190
191
192
193static INLINE struct ureg_dst
194ureg_dst_register( unsigned file,
195                   unsigned index )
196{
197   struct ureg_dst dst;
198
199   dst.File      = file;
200   dst.WriteMask = TGSI_WRITEMASK_XYZW;
201   dst.Indirect  = 0;
202   dst.IndirectIndex = 0;
203   dst.IndirectSwizzle = 0;
204   dst.Saturate  = 0;
205   dst.Index     = index;
206   dst.Pad1      = 0;
207   dst.Pad2      = 0;
208
209   return dst;
210}
211
212static INLINE struct ureg_src
213ureg_src_register( unsigned file,
214                   unsigned index )
215{
216   struct ureg_src src;
217
218   src.File     = file;
219   src.SwizzleX = TGSI_SWIZZLE_X;
220   src.SwizzleY = TGSI_SWIZZLE_Y;
221   src.SwizzleZ = TGSI_SWIZZLE_Z;
222   src.SwizzleW = TGSI_SWIZZLE_W;
223   src.Pad      = 0;
224   src.Indirect = 0;
225   src.IndirectIndex = 0;
226   src.IndirectSwizzle = 0;
227   src.Absolute = 0;
228   src.Index    = index;
229   src.Negate   = 0;
230
231   return src;
232}
233
234
235
236
237struct ureg_src
238ureg_DECL_fs_input( struct ureg_program *ureg,
239                    unsigned name,
240                    unsigned index,
241                    unsigned interp_mode )
242{
243   unsigned i;
244
245   for (i = 0; i < ureg->nr_fs_inputs; i++) {
246      if (ureg->fs_input[i].semantic_name == name &&
247          ureg->fs_input[i].semantic_index == index)
248         goto out;
249   }
250
251   if (ureg->nr_fs_inputs < UREG_MAX_INPUT) {
252      ureg->fs_input[i].semantic_name = name;
253      ureg->fs_input[i].semantic_index = index;
254      ureg->fs_input[i].interp = interp_mode;
255      ureg->nr_fs_inputs++;
256   }
257   else {
258      set_bad( ureg );
259   }
260
261out:
262   return ureg_src_register( TGSI_FILE_INPUT, i );
263}
264
265
266struct ureg_src
267ureg_DECL_vs_input( struct ureg_program *ureg,
268                    unsigned index )
269{
270   assert(ureg->processor == TGSI_PROCESSOR_VERTEX);
271
272   ureg->vs_inputs[index/32] |= 1 << (index % 32);
273   return ureg_src_register( TGSI_FILE_INPUT, index );
274}
275
276
277struct ureg_dst
278ureg_DECL_output( struct ureg_program *ureg,
279                  unsigned name,
280                  unsigned index )
281{
282   unsigned i;
283
284   for (i = 0; i < ureg->nr_outputs; i++) {
285      if (ureg->output[i].semantic_name == name &&
286          ureg->output[i].semantic_index == index)
287         goto out;
288   }
289
290   if (ureg->nr_outputs < UREG_MAX_OUTPUT) {
291      ureg->output[i].semantic_name = name;
292      ureg->output[i].semantic_index = index;
293      ureg->nr_outputs++;
294   }
295   else {
296      set_bad( ureg );
297   }
298
299out:
300   return ureg_dst_register( TGSI_FILE_OUTPUT, i );
301}
302
303
304/* Returns a new constant register.  Keep track of which have been
305 * referred to so that we can emit decls later.
306 *
307 * There is nothing in this code to bind this constant to any tracked
308 * value or manage any constant_buffer contents -- that's the
309 * resposibility of the calling code.
310 */
311struct ureg_src ureg_DECL_constant(struct ureg_program *ureg )
312{
313   return ureg_src_register( TGSI_FILE_CONSTANT, ureg->nr_constants++ );
314}
315
316
317/* Allocate a new temporary.  Temporaries greater than UREG_MAX_TEMP
318 * are legal, but will not be released.
319 */
320struct ureg_dst ureg_DECL_temporary( struct ureg_program *ureg )
321{
322   unsigned i;
323
324   for (i = 0; i < UREG_MAX_TEMP; i += 32) {
325      int bit = ffs(~ureg->temps_active[i/32]);
326      if (bit != 0) {
327         i += bit - 1;
328         goto out;
329      }
330   }
331
332   /* No reusable temps, so allocate a new one:
333    */
334   i = ureg->nr_temps++;
335
336out:
337   if (i < UREG_MAX_TEMP)
338      ureg->temps_active[i/32] |= 1 << (i % 32);
339
340   if (i >= ureg->nr_temps)
341      ureg->nr_temps = i + 1;
342
343   return ureg_dst_register( TGSI_FILE_TEMPORARY, i );
344}
345
346
347void ureg_release_temporary( struct ureg_program *ureg,
348                             struct ureg_dst tmp )
349{
350   if(tmp.File == TGSI_FILE_TEMPORARY)
351      if (tmp.Index < UREG_MAX_TEMP)
352         ureg->temps_active[tmp.Index/32] &= ~(1 << (tmp.Index % 32));
353}
354
355
356/* Allocate a new address register.
357 */
358struct ureg_dst ureg_DECL_address( struct ureg_program *ureg )
359{
360   if (ureg->nr_addrs < UREG_MAX_ADDR)
361      return ureg_dst_register( TGSI_FILE_ADDRESS, ureg->nr_addrs++ );
362
363   assert( 0 );
364   return ureg_dst_register( TGSI_FILE_ADDRESS, 0 );
365}
366
367/* Allocate a new sampler.
368 */
369struct ureg_src ureg_DECL_sampler( struct ureg_program *ureg,
370                                   unsigned nr )
371{
372   unsigned i;
373
374   for (i = 0; i < ureg->nr_samplers; i++)
375      if (ureg->sampler[i].Index == nr)
376         return ureg->sampler[i];
377
378   if (i < PIPE_MAX_SAMPLERS) {
379      ureg->sampler[i] = ureg_src_register( TGSI_FILE_SAMPLER, nr );
380      ureg->nr_samplers++;
381      return ureg->sampler[i];
382   }
383
384   assert( 0 );
385   return ureg->sampler[0];
386}
387
388
389
390
391static int match_or_expand_immediate( const float *v,
392                                      unsigned nr,
393                                      float *v2,
394                                      unsigned *nr2,
395                                      unsigned *swizzle )
396{
397   unsigned i, j;
398
399   *swizzle = 0;
400
401   for (i = 0; i < nr; i++) {
402      boolean found = FALSE;
403
404      for (j = 0; j < *nr2 && !found; j++) {
405         if (v[i] == v2[j]) {
406            *swizzle |= j << (i * 2);
407            found = TRUE;
408         }
409      }
410
411      if (!found) {
412         if (*nr2 >= 4)
413            return FALSE;
414
415         v2[*nr2] = v[i];
416         *swizzle |= *nr2 << (i * 2);
417         (*nr2)++;
418      }
419   }
420
421   return TRUE;
422}
423
424
425
426
427struct ureg_src ureg_DECL_immediate( struct ureg_program *ureg,
428                                     const float *v,
429                                     unsigned nr )
430{
431   unsigned i, j;
432   unsigned swizzle;
433
434   /* Could do a first pass where we examine all existing immediates
435    * without expanding.
436    */
437
438   for (i = 0; i < ureg->nr_immediates; i++) {
439      if (match_or_expand_immediate( v,
440                                     nr,
441                                     ureg->immediate[i].v,
442                                     &ureg->immediate[i].nr,
443                                     &swizzle ))
444         goto out;
445   }
446
447   if (ureg->nr_immediates < UREG_MAX_IMMEDIATE) {
448      i = ureg->nr_immediates++;
449      if (match_or_expand_immediate( v,
450                                     nr,
451                                     ureg->immediate[i].v,
452                                     &ureg->immediate[i].nr,
453                                     &swizzle ))
454         goto out;
455   }
456
457   set_bad( ureg );
458
459out:
460   /* Make sure that all referenced elements are from this immediate.
461    * Has the effect of making size-one immediates into scalars.
462    */
463   for (j = nr; j < 4; j++)
464      swizzle |= (swizzle & 0x3) << (j * 2);
465
466   return ureg_swizzle( ureg_src_register( TGSI_FILE_IMMEDIATE, i ),
467                        (swizzle >> 0) & 0x3,
468                        (swizzle >> 2) & 0x3,
469                        (swizzle >> 4) & 0x3,
470                        (swizzle >> 6) & 0x3);
471}
472
473
474void
475ureg_emit_src( struct ureg_program *ureg,
476               struct ureg_src src )
477{
478   unsigned size = (1 +
479                    (src.Absolute ? 1 : 0) +
480                    (src.Indirect ? 1 : 0));
481
482   union tgsi_any_token *out = get_tokens( ureg, DOMAIN_INSN, size );
483   unsigned n = 0;
484
485   assert(src.File != TGSI_FILE_NULL);
486   assert(src.File != TGSI_FILE_OUTPUT);
487   assert(src.File < TGSI_FILE_COUNT);
488
489   out[n].value = 0;
490   out[n].src.File = src.File;
491   out[n].src.SwizzleX = src.SwizzleX;
492   out[n].src.SwizzleY = src.SwizzleY;
493   out[n].src.SwizzleZ = src.SwizzleZ;
494   out[n].src.SwizzleW = src.SwizzleW;
495   out[n].src.Index = src.Index;
496   out[n].src.Negate = src.Negate;
497   n++;
498
499   if (src.Absolute) {
500      out[0].src.Extended = 1;
501      out[0].src.Negate = 0;
502      out[n].value = 0;
503      out[n].src_ext_mod.Type = TGSI_SRC_REGISTER_EXT_TYPE_MOD;
504      out[n].src_ext_mod.Absolute = 1;
505      out[n].src_ext_mod.Negate = src.Negate;
506      n++;
507   }
508
509   if (src.Indirect) {
510      out[0].src.Indirect = 1;
511      out[n].value = 0;
512      out[n].src.File = TGSI_FILE_ADDRESS;
513      out[n].src.SwizzleX = src.IndirectSwizzle;
514      out[n].src.SwizzleY = src.IndirectSwizzle;
515      out[n].src.SwizzleZ = src.IndirectSwizzle;
516      out[n].src.SwizzleW = src.IndirectSwizzle;
517      out[n].src.Index = src.IndirectIndex;
518      n++;
519   }
520
521   assert(n == size);
522}
523
524
525void
526ureg_emit_dst( struct ureg_program *ureg,
527               struct ureg_dst dst )
528{
529   unsigned size = (1 +
530                    (dst.Indirect ? 1 : 0));
531
532   union tgsi_any_token *out = get_tokens( ureg, DOMAIN_INSN, size );
533   unsigned n = 0;
534
535   assert(dst.File != TGSI_FILE_NULL);
536   assert(dst.File != TGSI_FILE_CONSTANT);
537   assert(dst.File != TGSI_FILE_INPUT);
538   assert(dst.File != TGSI_FILE_SAMPLER);
539   assert(dst.File != TGSI_FILE_IMMEDIATE);
540   assert(dst.File < TGSI_FILE_COUNT);
541
542   out[n].value = 0;
543   out[n].dst.File = dst.File;
544   out[n].dst.WriteMask = dst.WriteMask;
545   out[n].dst.Indirect = dst.Indirect;
546   out[n].dst.Index = dst.Index;
547   n++;
548
549   if (dst.Indirect) {
550      out[n].value = 0;
551      out[n].src.File = TGSI_FILE_ADDRESS;
552      out[n].src.SwizzleX = dst.IndirectSwizzle;
553      out[n].src.SwizzleY = dst.IndirectSwizzle;
554      out[n].src.SwizzleZ = dst.IndirectSwizzle;
555      out[n].src.SwizzleW = dst.IndirectSwizzle;
556      out[n].src.Index = dst.IndirectIndex;
557      n++;
558   }
559
560   assert(n == size);
561}
562
563
564
565unsigned
566ureg_emit_insn(struct ureg_program *ureg,
567               unsigned opcode,
568               boolean saturate,
569               unsigned num_dst,
570               unsigned num_src )
571{
572   union tgsi_any_token *out;
573
574   out = get_tokens( ureg, DOMAIN_INSN, 1 );
575   out[0].value = 0;
576   out[0].insn.Type = TGSI_TOKEN_TYPE_INSTRUCTION;
577   out[0].insn.NrTokens = 0;
578   out[0].insn.Opcode = opcode;
579   out[0].insn.Saturate = saturate;
580   out[0].insn.NumDstRegs = num_dst;
581   out[0].insn.NumSrcRegs = num_src;
582   out[0].insn.Padding = 0;
583   out[0].insn.Extended = 0;
584
585   ureg->nr_instructions++;
586
587   return ureg->domain[DOMAIN_INSN].count - 1;
588}
589
590
591void
592ureg_emit_label(struct ureg_program *ureg,
593                unsigned insn_token,
594                unsigned *label_token )
595{
596   union tgsi_any_token *out, *insn;
597
598   if(!label_token)
599      return;
600
601   out = get_tokens( ureg, DOMAIN_INSN, 1 );
602   insn = retrieve_token( ureg, DOMAIN_INSN, insn_token );
603
604   insn->insn.Extended = 1;
605
606   out[0].value = 0;
607   out[0].insn_ext_label.Type = TGSI_INSTRUCTION_EXT_TYPE_LABEL;
608
609   *label_token = ureg->domain[DOMAIN_INSN].count - 1;
610}
611
612/* Will return a number which can be used in a label to point to the
613 * next instruction to be emitted.
614 */
615unsigned
616ureg_get_instruction_number( struct ureg_program *ureg )
617{
618   return ureg->nr_instructions;
619}
620
621/* Patch a given label (expressed as a token number) to point to a
622 * given instruction (expressed as an instruction number).
623 */
624void
625ureg_fixup_label(struct ureg_program *ureg,
626                 unsigned label_token,
627                 unsigned instruction_number )
628{
629   union tgsi_any_token *out = retrieve_token( ureg, DOMAIN_INSN, label_token );
630
631   assert(out->insn_ext_label.Type == TGSI_INSTRUCTION_EXT_TYPE_LABEL);
632   out->insn_ext_label.Label = instruction_number;
633}
634
635
636void
637ureg_emit_texture(struct ureg_program *ureg,
638                  unsigned insn_token,
639                  unsigned target )
640{
641   union tgsi_any_token *out, *insn;
642
643   out = get_tokens( ureg, DOMAIN_INSN, 1 );
644   insn = retrieve_token( ureg, DOMAIN_INSN, insn_token );
645
646   insn->insn.Extended = 1;
647
648   out[0].value = 0;
649   out[0].insn_ext_texture.Type = TGSI_INSTRUCTION_EXT_TYPE_TEXTURE;
650   out[0].insn_ext_texture.Texture = target;
651}
652
653
654void
655ureg_fixup_insn_size(struct ureg_program *ureg,
656                     unsigned insn )
657{
658   union tgsi_any_token *out = retrieve_token( ureg, DOMAIN_INSN, insn );
659
660   assert(out->insn.Type == TGSI_TOKEN_TYPE_INSTRUCTION);
661   out->insn.NrTokens = ureg->domain[DOMAIN_INSN].count - insn - 1;
662}
663
664
665void
666ureg_insn(struct ureg_program *ureg,
667          unsigned opcode,
668          const struct ureg_dst *dst,
669          unsigned nr_dst,
670          const struct ureg_src *src,
671          unsigned nr_src )
672{
673   unsigned insn, i;
674   boolean saturate;
675
676#ifdef DEBUG
677   {
678      const struct tgsi_opcode_info *info = tgsi_get_opcode_info( opcode );
679      assert(info);
680      if(info) {
681         assert(nr_dst == info->num_dst);
682         assert(nr_src == info->num_src);
683      }
684   }
685#endif
686
687   saturate = nr_dst ? dst[0].Saturate : FALSE;
688
689   insn = ureg_emit_insn( ureg, opcode, saturate, nr_dst, nr_src );
690
691   for (i = 0; i < nr_dst; i++)
692      ureg_emit_dst( ureg, dst[i] );
693
694   for (i = 0; i < nr_src; i++)
695      ureg_emit_src( ureg, src[i] );
696
697   ureg_fixup_insn_size( ureg, insn );
698}
699
700
701
702static void emit_decl( struct ureg_program *ureg,
703                       unsigned file,
704                       unsigned index,
705                       unsigned semantic_name,
706                       unsigned semantic_index,
707                       unsigned interp )
708{
709   union tgsi_any_token *out = get_tokens( ureg, DOMAIN_DECL, 3 );
710
711   out[0].value = 0;
712   out[0].decl.Type = TGSI_TOKEN_TYPE_DECLARATION;
713   out[0].decl.NrTokens = 3;
714   out[0].decl.File = file;
715   out[0].decl.UsageMask = TGSI_WRITEMASK_XYZW; /* FIXME! */
716   out[0].decl.Interpolate = interp;
717   out[0].decl.Semantic = 1;
718
719   out[1].value = 0;
720   out[1].decl_range.First =
721      out[1].decl_range.Last = index;
722
723   out[2].value = 0;
724   out[2].decl_semantic.SemanticName = semantic_name;
725   out[2].decl_semantic.SemanticIndex = semantic_index;
726
727}
728
729
730static void emit_decl_range( struct ureg_program *ureg,
731                             unsigned file,
732                             unsigned first,
733                             unsigned count )
734{
735   union tgsi_any_token *out = get_tokens( ureg, DOMAIN_DECL, 2 );
736
737   out[0].value = 0;
738   out[0].decl.Type = TGSI_TOKEN_TYPE_DECLARATION;
739   out[0].decl.NrTokens = 2;
740   out[0].decl.File = file;
741   out[0].decl.UsageMask = 0xf;
742   out[0].decl.Interpolate = TGSI_INTERPOLATE_CONSTANT;
743   out[0].decl.Semantic = 0;
744
745   out[1].value = 0;
746   out[1].decl_range.First = first;
747   out[1].decl_range.Last = first + count - 1;
748}
749
750static void emit_immediate( struct ureg_program *ureg,
751                            const float *v )
752{
753   union tgsi_any_token *out = get_tokens( ureg, DOMAIN_DECL, 5 );
754
755   out[0].value = 0;
756   out[0].imm.Type = TGSI_TOKEN_TYPE_IMMEDIATE;
757   out[0].imm.NrTokens = 5;
758   out[0].imm.DataType = TGSI_IMM_FLOAT32;
759   out[0].imm.Padding = 0;
760   out[0].imm.Extended = 0;
761
762   out[1].imm_data.Float = v[0];
763   out[2].imm_data.Float = v[1];
764   out[3].imm_data.Float = v[2];
765   out[4].imm_data.Float = v[3];
766}
767
768
769
770
771static void emit_decls( struct ureg_program *ureg )
772{
773   unsigned i;
774
775   if (ureg->processor == TGSI_PROCESSOR_VERTEX) {
776      for (i = 0; i < UREG_MAX_INPUT; i++) {
777         if (ureg->vs_inputs[i/32] & (1 << (i%32))) {
778            emit_decl_range( ureg, TGSI_FILE_INPUT, i, 1 );
779         }
780      }
781   }
782   else {
783      for (i = 0; i < ureg->nr_fs_inputs; i++) {
784         emit_decl( ureg,
785                    TGSI_FILE_INPUT,
786                    i,
787                    ureg->fs_input[i].semantic_name,
788                    ureg->fs_input[i].semantic_index,
789                    ureg->fs_input[i].interp );
790      }
791   }
792
793   for (i = 0; i < ureg->nr_outputs; i++) {
794      emit_decl( ureg,
795                 TGSI_FILE_OUTPUT,
796                 i,
797                 ureg->output[i].semantic_name,
798                 ureg->output[i].semantic_index,
799                 TGSI_INTERPOLATE_CONSTANT );
800   }
801
802   for (i = 0; i < ureg->nr_samplers; i++) {
803      emit_decl_range( ureg,
804                       TGSI_FILE_SAMPLER,
805                       ureg->sampler[i].Index, 1 );
806   }
807
808   if (ureg->nr_constants) {
809      emit_decl_range( ureg,
810                       TGSI_FILE_CONSTANT,
811                       0, ureg->nr_constants );
812   }
813
814   if (ureg->nr_temps) {
815      emit_decl_range( ureg,
816                       TGSI_FILE_TEMPORARY,
817                       0, ureg->nr_temps );
818   }
819
820   if (ureg->nr_addrs) {
821      emit_decl_range( ureg,
822                       TGSI_FILE_ADDRESS,
823                       0, ureg->nr_addrs );
824   }
825
826   for (i = 0; i < ureg->nr_immediates; i++) {
827      emit_immediate( ureg,
828                      ureg->immediate[i].v );
829   }
830}
831
832/* Append the instruction tokens onto the declarations to build a
833 * contiguous stream suitable to send to the driver.
834 */
835static void copy_instructions( struct ureg_program *ureg )
836{
837   unsigned nr_tokens = ureg->domain[DOMAIN_INSN].count;
838   union tgsi_any_token *out = get_tokens( ureg,
839                                           DOMAIN_DECL,
840                                           nr_tokens );
841
842   memcpy(out,
843          ureg->domain[DOMAIN_INSN].tokens,
844          nr_tokens * sizeof out[0] );
845}
846
847
848static void
849fixup_header_size(struct ureg_program *ureg)
850{
851   union tgsi_any_token *out = retrieve_token( ureg, DOMAIN_DECL, 1 );
852
853   out->header.BodySize = ureg->domain[DOMAIN_DECL].count - 3;
854}
855
856
857static void
858emit_header( struct ureg_program *ureg )
859{
860   union tgsi_any_token *out = get_tokens( ureg, DOMAIN_DECL, 3 );
861
862   out[0].version.MajorVersion = 1;
863   out[0].version.MinorVersion = 1;
864   out[0].version.Padding = 0;
865
866   out[1].header.HeaderSize = 2;
867   out[1].header.BodySize = 0;
868
869   out[2].processor.Processor = ureg->processor;
870   out[2].processor.Padding = 0;
871}
872
873
874const struct tgsi_token *ureg_finalize( struct ureg_program *ureg )
875{
876   const struct tgsi_token *tokens;
877
878   emit_header( ureg );
879   emit_decls( ureg );
880   copy_instructions( ureg );
881   fixup_header_size( ureg );
882
883   if (ureg->domain[0].tokens == error_tokens ||
884       ureg->domain[1].tokens == error_tokens) {
885      debug_printf("%s: error in generated shader\n", __FUNCTION__);
886      assert(0);
887      return NULL;
888   }
889
890   tokens = &ureg->domain[DOMAIN_DECL].tokens[0].token;
891
892   if (0) {
893      debug_printf("%s: emitted shader %d tokens:\n", __FUNCTION__,
894                   ureg->domain[DOMAIN_DECL].count);
895      tgsi_dump( tokens, 0 );
896   }
897
898#if DEBUG
899   if (tokens && !tgsi_sanity_check(tokens)) {
900      debug_printf("tgsi_ureg.c, sanity check failed on generated tokens:\n");
901      tgsi_dump(tokens, 0);
902      assert(0);
903   }
904#endif
905
906
907   return tokens;
908}
909
910
911void *ureg_create_shader( struct ureg_program *ureg,
912                          struct pipe_context *pipe )
913{
914   struct pipe_shader_state state;
915
916   state.tokens = ureg_finalize(ureg);
917   if(!state.tokens)
918      return NULL;
919
920   if (ureg->processor == TGSI_PROCESSOR_VERTEX)
921      return pipe->create_vs_state( pipe, &state );
922   else
923      return pipe->create_fs_state( pipe, &state );
924}
925
926
927
928
929struct ureg_program *ureg_create( unsigned processor )
930{
931   struct ureg_program *ureg = CALLOC_STRUCT( ureg_program );
932   if (ureg == NULL)
933      return NULL;
934
935   ureg->processor = processor;
936   return ureg;
937}
938
939
940void ureg_destroy( struct ureg_program *ureg )
941{
942   unsigned i;
943
944   for (i = 0; i < Elements(ureg->domain); i++) {
945      if (ureg->domain[i].tokens &&
946          ureg->domain[i].tokens != error_tokens)
947         FREE(ureg->domain[i].tokens);
948   }
949
950   FREE(ureg);
951}
952