nv50_program.c revision 3a68fcfb6b406cf864afbf200e436fc384fd0865 
1/*
2 * Copyright 2010 Chrsitoph Bumiller
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17 * THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
18 * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF
19 * OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
20 * SOFTWARE.
21 */
22
23#include "nv50_program.h"
24#include "nv50_pc.h"
25#include "nv50_context.h"
26
27#include "pipe/p_shader_tokens.h"
28#include "tgsi/tgsi_parse.h"
29#include "tgsi/tgsi_util.h"
30#include "tgsi/tgsi_dump.h"
31
32static INLINE unsigned
33bitcount4(const uint32_t val)
34{
35   static const unsigned cnt[16]
36   = { 0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4 };
37   return cnt[val & 0xf];
38}
39
40static unsigned
41nv50_tgsi_src_mask(const struct tgsi_full_instruction *inst, int c)
42{
43   unsigned mask = inst->Dst[0].Register.WriteMask;
44
45   switch (inst->Instruction.Opcode) {
46   case TGSI_OPCODE_COS:
47   case TGSI_OPCODE_SIN:
48      return (mask & 0x8) | ((mask & 0x7) ? 0x1 : 0x0);
49   case TGSI_OPCODE_DP3:
50      return 0x7;
51   case TGSI_OPCODE_DP4:
52   case TGSI_OPCODE_DPH:
53   case TGSI_OPCODE_KIL: /* WriteMask ignored */
54      return 0xf;
55   case TGSI_OPCODE_DST:
56      return mask & (c ? 0xa : 0x6);
57   case TGSI_OPCODE_EX2:
58   case TGSI_OPCODE_EXP:
59   case TGSI_OPCODE_LG2:
60   case TGSI_OPCODE_LOG:
61   case TGSI_OPCODE_POW:
62   case TGSI_OPCODE_RCP:
63   case TGSI_OPCODE_RSQ:
64   case TGSI_OPCODE_SCS:
65      return 0x1;
66   case TGSI_OPCODE_IF:
67      return 0x1;
68   case TGSI_OPCODE_LIT:
69      return 0xb;
70   case TGSI_OPCODE_TEX:
71   case TGSI_OPCODE_TXB:
72   case TGSI_OPCODE_TXL:
73   case TGSI_OPCODE_TXP:
74   {
75      const struct tgsi_instruction_texture *tex;
76
77      assert(inst->Instruction.Texture);
78      tex = &inst->Texture;
79
80      mask = 0x7;
81      if (inst->Instruction.Opcode != TGSI_OPCODE_TEX &&
82          inst->Instruction.Opcode != TGSI_OPCODE_TXD)
83         mask |= 0x8; /* bias, lod or proj */
84
85      switch (tex->Texture) {
86      case TGSI_TEXTURE_1D:
87         mask &= 0x9;
88         break;
89      case TGSI_TEXTURE_SHADOW1D:
90         mask &= 0x5;
91         break;
92      case TGSI_TEXTURE_2D:
93         mask &= 0xb;
94         break;
95      default:
96         break;
97      }
98   }
99  	   return mask;
100   case TGSI_OPCODE_XPD:
101   {
102      unsigned x = 0;
103      if (mask & 1) x |= 0x6;
104      if (mask & 2) x |= 0x5;
105      if (mask & 4) x |= 0x3;
106      return x;
107   }
108   default:
109      break;
110   }
111
112   return mask;
113}
114
115static void
116nv50_indirect_inputs(struct nv50_translation_info *ti, int id)
117{
118   int i, c;
119
120   for (i = 0; i < PIPE_MAX_SHADER_INPUTS; ++i)
121      for (c = 0; c < 4; ++c)
122         ti->input_access[i][c] = id;
123
124   ti->indirect_inputs = TRUE;
125}
126
127static void
128nv50_indirect_outputs(struct nv50_translation_info *ti, int id)
129{
130   int i, c;
131
132   for (i = 0; i < PIPE_MAX_SHADER_OUTPUTS; ++i)
133      for (c = 0; c < 4; ++c)
134         ti->output_access[i][c] = id;
135
136   ti->indirect_outputs = TRUE;
137}
138
139static void
140prog_inst(struct nv50_translation_info *ti,
141          const struct tgsi_full_instruction *inst, int id)
142{
143   const struct tgsi_dst_register *dst;
144   const struct tgsi_src_register *src;
145   int s, c, k;
146   unsigned mask;
147
148   if (inst->Dst[0].Register.File == TGSI_FILE_OUTPUT) {
149      for (c = 0; c < 4; ++c) {
150         dst = &inst->Dst[0].Register;
151         if (inst->Dst[0].Register.Indirect)
152            nv50_indirect_outputs(ti, id);
153         if (!(dst->WriteMask & (1 << c)))
154            continue;
155         ti->output_access[dst->Index][c] = id;
156      }
157
158      if (inst->Instruction.Opcode == TGSI_OPCODE_MOV &&
159          inst->Src[0].Register.File == TGSI_FILE_INPUT &&
160          dst->Index == ti->edgeflag_out)
161         ti->p->vp.edgeflag = inst->Src[0].Register.Index;
162   }
163
164   for (s = 0; s < inst->Instruction.NumSrcRegs; ++s) {
165      src = &inst->Src[s].Register;
166      if (src->File != TGSI_FILE_INPUT)
167         continue;
168      mask = nv50_tgsi_src_mask(inst, s);
169
170      if (inst->Src[s].Register.Indirect)
171         nv50_indirect_inputs(ti, id);
172
173      for (c = 0; c < 4; ++c) {
174         if (!(mask & (1 << c)))
175            continue;
176         k = tgsi_util_get_full_src_register_swizzle(&inst->Src[s], c);
177         if (k <= TGSI_SWIZZLE_W)
178            ti->input_access[src->Index][k] = id;
179      }
180   }
181}
182
183static void
184prog_immediate(struct nv50_translation_info *ti,
185               const struct tgsi_full_immediate *imm)
186{
187   int c;
188   unsigned n = ++ti->immd32_nr;
189
190   tgsi_dump_immediate(imm);
191
192   if (n == (1 << (ffs(n) - 1)))
193      ti->immd32 = REALLOC(ti->immd32, (n / 2) * 16, (n * 2) * 16);
194
195   for (c = 0; c < 4; ++c)
196      ti->immd32[(n - 1) * 4 + c] = imm->u[c].Uint;
197}
198
199static INLINE unsigned
200translate_interpolate(const struct tgsi_full_declaration *decl)
201{
202   unsigned mode;
203
204   if (decl->Declaration.Interpolate == TGSI_INTERPOLATE_CONSTANT)
205      mode = NV50_INTERP_FLAT;
206   else
207   if (decl->Declaration.Interpolate == TGSI_INTERPOLATE_PERSPECTIVE)
208      mode = 0;
209   else
210      mode = NV50_INTERP_LINEAR;
211
212   if (decl->Declaration.Centroid)
213      mode |= NV50_INTERP_CENTROID;
214
215   return mode;
216}
217
218static void
219prog_decl(struct nv50_translation_info *ti,
220          const struct tgsi_full_declaration *decl)
221{
222   unsigned i, first, last, sn = 0, si = 0;
223
224   first = decl->Range.First;
225   last = decl->Range.Last;
226
227   if (decl->Declaration.Semantic) {
228      sn = decl->Semantic.Name;
229      si = decl->Semantic.Index;
230   }
231   tgsi_dump_declaration(decl);
232
233   switch (decl->Declaration.File) {
234   case TGSI_FILE_INPUT:
235      for (i = first; i <= last; ++i)
236         ti->interp_mode[i] = translate_interpolate(decl);
237
238      if (!decl->Declaration.Semantic)
239         break;
240
241      for (i = first; i <= last; ++i) {
242         ti->p->in[i].sn = sn;
243         ti->p->in[i].si = si;
244      }
245
246      switch (sn) {
247      case TGSI_SEMANTIC_FACE:
248         break;
249      case TGSI_SEMANTIC_COLOR:
250         if (ti->p->type == PIPE_SHADER_FRAGMENT)
251            ti->p->vp.bfc[si] = first;
252         break;
253      }
254      break;
255   case TGSI_FILE_OUTPUT:
256      if (!decl->Declaration.Semantic)
257         break;
258
259      for (i = first; i <= last; ++i) {
260         ti->p->out[i].sn = sn;
261         ti->p->out[i].si = si;
262      }
263
264      switch (sn) {
265      case TGSI_SEMANTIC_BCOLOR:
266         ti->p->vp.bfc[si] = first;
267         break;
268      case TGSI_SEMANTIC_PSIZE:
269         ti->p->vp.psiz = first;
270         break;
271      case TGSI_SEMANTIC_EDGEFLAG:
272         ti->edgeflag_out = first;
273         break;
274      default:
275         break;
276      }
277      break;
278   case TGSI_FILE_SYSTEM_VALUE:
279      switch (decl->Semantic.Name) {
280      case TGSI_SEMANTIC_FACE:
281         break;
282      case TGSI_SEMANTIC_INSTANCEID:
283         break;
284      case TGSI_SEMANTIC_PRIMID:
285         break;
286         /*
287      case TGSI_SEMANTIC_PRIMIDIN:
288         break;
289      case TGSI_SEMANTIC_VERTEXID:
290         break;
291         */
292      default:
293         break;
294      }
295      break;
296   case TGSI_FILE_CONSTANT:
297      ti->p->parm_size = MAX2(ti->p->parm_size, (last + 1) * 16);
298      break;
299   case TGSI_FILE_ADDRESS:
300   case TGSI_FILE_SAMPLER:
301   case TGSI_FILE_TEMPORARY:
302      break;
303   default:
304      assert(0);
305      break;
306   }
307}
308
309static int
310nv50_vertprog_prepare(struct nv50_translation_info *ti)
311{
312   struct nv50_program *p = ti->p;
313   int i, c;
314   unsigned num_inputs = 0;
315
316   ti->input_file = NV_FILE_MEM_S;
317   ti->output_file = NV_FILE_OUT;
318
319   for (i = 0; i <= ti->scan.file_max[TGSI_FILE_INPUT]; ++i) {
320      p->in[i].id = i;
321      p->in[i].hw = num_inputs;
322
323      for (c = 0; c < 4; ++c) {
324         if (!ti->input_access[i][c])
325            continue;
326         ti->input_map[i][c] = num_inputs++;
327         p->vp.attrs[(4 * i + c) / 32] |= 1 << ((i * 4 + c) % 32);
328      }
329   }
330
331   for (i = 0; i <= ti->scan.file_max[TGSI_FILE_OUTPUT]; ++i) {
332      p->out[i].id = i;
333      p->out[i].hw = p->max_out;
334
335      for (c = 0; c < 4; ++c) {
336         if (!ti->output_access[i][c])
337            continue;
338         ti->output_map[i][c] = p->max_out++;
339         p->out[i].mask |= 1 << c;
340      }
341   }
342
343   if (p->vp.psiz < 0x40)
344      p->vp.psiz = p->out[p->vp.psiz].hw;
345
346   return 0;
347}
348
349static int
350nv50_fragprog_prepare(struct nv50_translation_info *ti)
351{
352   struct nv50_program *p = ti->p;
353   int i, j, c;
354   unsigned nvary, nintp, depr;
355   unsigned n = 0, m = 0, skip = 0;
356   ubyte sn[16], si[16];
357
358   /* FP flags */
359
360   if (ti->scan.writes_z) {
361      p->fp.flags[1] = 0x11;
362      p->fp.flags[0] |= NV50TCL_FP_CONTROL_EXPORTS_Z;
363   }
364
365   if (ti->scan.uses_kill)
366      p->fp.flags[0] |= NV50TCL_FP_CONTROL_USES_KIL;
367
368   /* FP inputs */
369
370   ti->input_file = NV_FILE_MEM_V;
371   ti->output_file = NV_FILE_GPR;
372
373   /* count non-flat inputs, save semantic info */
374   for (i = 0; i < p->in_nr; ++i) {
375      m += (ti->interp_mode[i] & NV50_INTERP_FLAT) ? 0 : 1;
376      sn[i] = p->in[i].sn;
377      si[i] = p->in[i].si;
378   }
379
380   /* reorder p->in[] so that non-flat inputs are first and
381    * kick out special inputs that don't use VP/GP_RESULT_MAP
382    */
383   nintp = 0;
384   for (i = 0; i < p->in_nr; ++i) {
385      if (sn[i] == TGSI_SEMANTIC_POSITION) {
386         for (c = 0; c < 4; ++c) {
387            ti->input_map[i][c] = nintp;
388            if (ti->input_access[i][c]) {
389               p->fp.interp |= 1 << (24 + c);
390               ++nintp;
391            }
392         }
393         skip++;
394         continue;
395      } else
396      if (sn[i] == TGSI_SEMANTIC_FACE) {
397         ti->input_map[i][0] = 255;
398         skip++;
399         continue;
400      }
401
402      j = (ti->interp_mode[i] & NV50_INTERP_FLAT) ? m++ : n++;
403
404      if (sn[i] == TGSI_SEMANTIC_COLOR)
405         p->vp.bfc[si[i]] = j;
406
407      p->in[j].linear = (ti->interp_mode[i] & NV50_INTERP_LINEAR) ? 1 : 0;
408      p->in[j].id = i;
409      p->in[j].sn = sn[i];
410      p->in[j].si = si[i];
411   }
412   assert(n <= m);
413   p->in_nr -= skip;
414
415   if (!(p->fp.interp & (8 << 24))) {
416      p->fp.interp |= (8 << 24);
417      ++nintp;
418   }
419
420   p->fp.colors = (1 << 24) | 4; /* CLAMP, FFC0_ID = 4 */
421
422   for (i = 0; i < p->in_nr; ++i) {
423      int j = p->in[i].id;
424      p->in[i].hw = nintp;
425
426      for (c = 0; c < 4; ++c) {
427         if (!ti->input_access[j][c])
428            continue;
429         p->in[i].mask |= 1 << c;
430         ti->input_map[j][c] = nintp++;
431      }
432      /* count color inputs */
433      if (i == p->vp.bfc[0] || i == p->vp.bfc[1])
434         p->fp.colors += bitcount4(p->in[i].mask) << 16;
435   }
436   nintp -= bitcount4(p->fp.interp >> 24); /* subtract position inputs */
437   nvary = nintp;
438   if (n < m)
439      nvary -= p->in[n].hw;
440
441   p->fp.interp |= nvary << NV50TCL_FP_INTERPOLANT_CTRL_COUNT_NONFLAT_SHIFT;
442   p->fp.interp |= nintp << NV50TCL_FP_INTERPOLANT_CTRL_COUNT_SHIFT;
443
444   /* FP outputs */
445
446   if (p->out_nr > (1 + (ti->scan.writes_z ? 1 : 0)))
447      p->fp.flags[0] |= NV50TCL_FP_CONTROL_MULTIPLE_RESULTS;
448
449   depr = p->out_nr;
450   for (i = 0; i < p->out_nr; ++i) {
451      p->out[i].id = i;
452      if (p->out[i].sn == TGSI_SEMANTIC_POSITION) {
453         depr = i;
454         continue;
455      }
456      p->out[i].hw = p->max_out;
457      p->out[i].mask = 0xf;
458
459      for (c = 0; c < 4; ++c)
460         ti->output_map[i][c] = p->max_out++;
461   }
462   if (depr < p->out_nr) {
463      p->out[depr].mask = 0x4;
464      p->out[depr].hw = p->max_out++;
465   }
466
467   return 0;
468}
469
470static int
471nv50_geomprog_prepare(struct nv50_translation_info *ti)
472{
473   ti->input_file = NV_FILE_MEM_S;
474   ti->output_file = NV_FILE_OUT;
475
476   assert(0);
477   return 1;
478}
479
480static int
481nv50_prog_scan(struct nv50_translation_info *ti)
482{
483   struct nv50_program *p = ti->p;
484   struct tgsi_parse_context parse;
485   int ret;
486
487   p->vp.psiz = 0x40;
488   p->vp.bfc[0] = 0x40;
489   p->vp.bfc[1] = 0x40;
490   p->gp.primid = 0x80;
491
492   tgsi_scan_shader(p->pipe.tokens, &ti->scan);
493
494   tgsi_parse_init(&parse, p->pipe.tokens);
495   while (!tgsi_parse_end_of_tokens(&parse)) {
496      tgsi_parse_token(&parse);
497
498      switch (parse.FullToken.Token.Type) {
499      case TGSI_TOKEN_TYPE_IMMEDIATE:
500         prog_immediate(ti, &parse.FullToken.FullImmediate);
501         break;
502      case TGSI_TOKEN_TYPE_DECLARATION:
503         prog_decl(ti, &parse.FullToken.FullDeclaration);
504         break;
505      case TGSI_TOKEN_TYPE_INSTRUCTION:
506         prog_inst(ti, &parse.FullToken.FullInstruction, ++ti->inst_nr);
507         break;
508      }
509   }
510
511   p->in_nr = ti->scan.file_max[TGSI_FILE_INPUT] + 1;
512   p->out_nr = ti->scan.file_max[TGSI_FILE_OUTPUT] + 1;
513
514   switch (p->type) {
515   case PIPE_SHADER_VERTEX:
516      ret = nv50_vertprog_prepare(ti);
517      break;
518   case PIPE_SHADER_FRAGMENT:
519      ret = nv50_fragprog_prepare(ti);
520      break;
521   case PIPE_SHADER_GEOMETRY:
522      ret = nv50_geomprog_prepare(ti);
523      break;
524   default:
525      assert(!"unsupported program type");
526      ret = -1;
527      break;
528   }
529
530   assert(!ret);
531   return ret;
532}
533
534boolean
535nv50_program_tx(struct nv50_program *p)
536{
537   struct nv50_translation_info *ti;
538   int ret;
539
540   ti = CALLOC_STRUCT(nv50_translation_info);
541   ti->p = p;
542
543   ti->edgeflag_out = PIPE_MAX_SHADER_OUTPUTS;
544
545   ret = nv50_prog_scan(ti);
546   if (ret) {
547      NOUVEAU_ERR("unsupported shader program\n");
548      goto out;
549   }
550
551   ret = nv50_generate_code(ti);
552   if (ret) {
553      NOUVEAU_ERR("error during shader translation\n");
554      goto out;
555   }
556
557out:
558   if (ti->immd32)
559      FREE(ti->immd32);
560   FREE(ti);
561   return ret ? FALSE : TRUE;
562}
563
564void
565nv50_program_destroy(struct nv50_context *nv50, struct nv50_program *p)
566{
567   nouveau_bo_ref(NULL, &p->bo);
568
569   so_ref(NULL, &p->so);
570
571   if (p->code)
572      FREE(p->code);
573
574   p->translated = FALSE;
575}
576