?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
gphoto2: Convert dialogs to po files.
[wine/multimedia.git] / dlls / wined3d / shader.c
blobd7ba8188de896ba9edbd8ea2a74d0f9147115bb4
1 /*
2 * Copyright 2002-2003 Jason Edmeades
3 * Copyright 2002-2003 Raphael Junqueira
4 * Copyright 2004 Christian Costa
5 * Copyright 2005 Oliver Stieber
6 * Copyright 2006 Ivan Gyurdiev
7 * Copyright 2007-2008 Stefan Dösinger for CodeWeavers
8 * Copyright 2009-2011 Henri Verbeet for CodeWeavers
9 *
10 * This library is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public
12 * License as published by the Free Software Foundation; either
13 * version 2.1 of the License, or (at your option) any later version.
14 *
15 * This library is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
19 *
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with this library; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
23 */
24
25 #include "config.h"
26
27 #include <math.h>
28 #include <stdio.h>
29 #include <string.h>
30
31 #include "wined3d_private.h"
32
33 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
34 WINE_DECLARE_DEBUG_CHANNEL(d3d);
35
36 static const char * const shader_opcode_names[] =
37 {
38 /* WINED3DSIH_ABS */ "abs",
39 /* WINED3DSIH_ADD */ "add",
40 /* WINED3DSIH_AND */ "and",
41 /* WINED3DSIH_BEM */ "bem",
42 /* WINED3DSIH_BREAK */ "break",
43 /* WINED3DSIH_BREAKC */ "breakc",
44 /* WINED3DSIH_BREAKP */ "breakp",
45 /* WINED3DSIH_CALL */ "call",
46 /* WINED3DSIH_CALLNZ */ "callnz",
47 /* WINED3DSIH_CMP */ "cmp",
48 /* WINED3DSIH_CND */ "cnd",
49 /* WINED3DSIH_CRS */ "crs",
50 /* WINED3DSIH_CUT */ "cut",
51 /* WINED3DSIH_DCL */ "dcl",
52 /* WINED3DSIH_DEF */ "def",
53 /* WINED3DSIH_DEFB */ "defb",
54 /* WINED3DSIH_DEFI */ "defi",
55 /* WINED3DSIH_DIV */ "div",
56 /* WINED3DSIH_DP2ADD */ "dp2add",
57 /* WINED3DSIH_DP3 */ "dp3",
58 /* WINED3DSIH_DP4 */ "dp4",
59 /* WINED3DSIH_DST */ "dst",
60 /* WINED3DSIH_DSX */ "dsx",
61 /* WINED3DSIH_DSY */ "dsy",
62 /* WINED3DSIH_ELSE */ "else",
63 /* WINED3DSIH_EMIT */ "emit",
64 /* WINED3DSIH_ENDIF */ "endif",
65 /* WINED3DSIH_ENDLOOP */ "endloop",
66 /* WINED3DSIH_ENDREP */ "endrep",
67 /* WINED3DSIH_EQ */ "eq",
68 /* WINED3DSIH_EXP */ "exp",
69 /* WINED3DSIH_EXPP */ "expp",
70 /* WINED3DSIH_FRC */ "frc",
71 /* WINED3DSIH_FTOI */ "ftoi",
72 /* WINED3DSIH_GE */ "ge",
73 /* WINED3DSIH_IADD */ "iadd",
74 /* WINED3DSIH_IEQ */ "ieq",
75 /* WINED3DSIH_IF */ "if",
76 /* WINED3DSIH_IFC */ "ifc",
77 /* WINED3DSIH_IGE */ "ige",
78 /* WINED3DSIH_IMUL */ "imul",
79 /* WINED3DSIH_ITOF */ "itof",
80 /* WINED3DSIH_LABEL */ "label",
81 /* WINED3DSIH_LD */ "ld",
82 /* WINED3DSIH_LIT */ "lit",
83 /* WINED3DSIH_LOG */ "log",
84 /* WINED3DSIH_LOGP */ "logp",
85 /* WINED3DSIH_LOOP */ "loop",
86 /* WINED3DSIH_LRP */ "lrp",
87 /* WINED3DSIH_LT */ "lt",
88 /* WINED3DSIH_M3x2 */ "m3x2",
89 /* WINED3DSIH_M3x3 */ "m3x3",
90 /* WINED3DSIH_M3x4 */ "m3x4",
91 /* WINED3DSIH_M4x3 */ "m4x3",
92 /* WINED3DSIH_M4x4 */ "m4x4",
93 /* WINED3DSIH_MAD */ "mad",
94 /* WINED3DSIH_MAX */ "max",
95 /* WINED3DSIH_MIN */ "min",
96 /* WINED3DSIH_MOV */ "mov",
97 /* WINED3DSIH_MOVA */ "mova",
98 /* WINED3DSIH_MOVC */ "movc",
99 /* WINED3DSIH_MUL */ "mul",
100 /* WINED3DSIH_NOP */ "nop",
101 /* WINED3DSIH_NRM */ "nrm",
102 /* WINED3DSIH_PHASE */ "phase",
103 /* WINED3DSIH_POW */ "pow",
104 /* WINED3DSIH_RCP */ "rcp",
105 /* WINED3DSIH_REP */ "rep",
106 /* WINED3DSIH_RET */ "ret",
107 /* WINED3DSIH_ROUND_NI */ "round_ni",
108 /* WINED3DSIH_RSQ */ "rsq",
109 /* WINED3DSIH_SAMPLE */ "sample",
110 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
111 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
112 /* WINED3DSIH_SETP */ "setp",
113 /* WINED3DSIH_SGE */ "sge",
114 /* WINED3DSIH_SGN */ "sgn",
115 /* WINED3DSIH_SINCOS */ "sincos",
116 /* WINED3DSIH_SLT */ "slt",
117 /* WINED3DSIH_SQRT */ "sqrt",
118 /* WINED3DSIH_SUB */ "sub",
119 /* WINED3DSIH_TEX */ "texld",
120 /* WINED3DSIH_TEXBEM */ "texbem",
121 /* WINED3DSIH_TEXBEML */ "texbeml",
122 /* WINED3DSIH_TEXCOORD */ "texcrd",
123 /* WINED3DSIH_TEXDEPTH */ "texdepth",
124 /* WINED3DSIH_TEXDP3 */ "texdp3",
125 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
126 /* WINED3DSIH_TEXKILL */ "texkill",
127 /* WINED3DSIH_TEXLDD */ "texldd",
128 /* WINED3DSIH_TEXLDL */ "texldl",
129 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
130 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
131 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
132 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
133 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
134 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
135 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
136 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
137 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
138 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
139 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
140 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
141 /* WINED3DSIH_UDIV */ "udiv",
142 /* WINED3DSIH_USHR */ "ushr",
143 /* WINED3DSIH_UTOF */ "utof",
144 /* WINED3DSIH_XOR */ "xor",
145 };
146
147 static const char * const semantic_names[] =
148 {
149 /* WINED3DDECLUSAGE_POSITION */ "SV_POSITION",
150 /* WINED3DDECLUSAGE_BLENDWEIGHT */ "BLENDWEIGHT",
151 /* WINED3DDECLUSAGE_BLENDINDICES */ "BLENDINDICES",
152 /* WINED3DDECLUSAGE_NORMAL */ "NORMAL",
153 /* WINED3DDECLUSAGE_PSIZE */ "PSIZE",
154 /* WINED3DDECLUSAGE_TEXCOORD */ "TEXCOORD",
155 /* WINED3DDECLUSAGE_TANGENT */ "TANGENT",
156 /* WINED3DDECLUSAGE_BINORMAL */ "BINORMAL",
157 /* WINED3DDECLUSAGE_TESSFACTOR */ "TESSFACTOR",
158 /* WINED3DDECLUSAGE_POSITIONT */ "POSITIONT",
159 /* WINED3DDECLUSAGE_COLOR */ "COLOR",
160 /* WINED3DDECLUSAGE_FOG */ "FOG",
161 /* WINED3DDECLUSAGE_DEPTH */ "DEPTH",
162 /* WINED3DDECLUSAGE_SAMPLE */ "SAMPLE",
163 };
164
165 static const char *shader_semantic_name_from_usage(WINED3DDECLUSAGE usage)
166 {
167 if (usage >= sizeof(semantic_names) / sizeof(*semantic_names))
168 {
169 FIXME("Unrecognized usage %#x.\n", usage);
170 return "UNRECOGNIZED";
171 }
172
173 return semantic_names[usage];
174 }
175
176 static WINED3DDECLUSAGE shader_usage_from_semantic_name(const char *name)
177 {
178 unsigned int i;
179
180 for (i = 0; i < sizeof(semantic_names) / sizeof(*semantic_names); ++i)
181 {
182 if (!strcmp(name, semantic_names[i])) return i;
183 }
184
185 return ~0U;
186 }
187
188 BOOL shader_match_semantic(const char *semantic_name, WINED3DDECLUSAGE usage)
189 {
190 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
191 }
192
193 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
194 const struct wined3d_shader_semantic *s)
195 {
196 e->semantic_name = shader_semantic_name_from_usage(s->usage);
197 e->semantic_idx = s->usage_idx;
198 e->sysval_semantic = 0;
199 e->component_type = 0;
200 e->register_idx = s->reg.reg.idx;
201 e->mask = s->reg.write_mask;
202 }
203
204 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
205 WINED3DDECLUSAGE usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
206 {
207 e->semantic_name = shader_semantic_name_from_usage(usage);
208 e->semantic_idx = usage_idx;
209 e->sysval_semantic = 0;
210 e->component_type = 0;
211 e->register_idx = reg_idx;
212 e->mask = write_mask;
213 }
214
215 static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
216 {
217 switch (version_token >> 16)
218 {
219 case WINED3D_SM1_VS:
220 case WINED3D_SM1_PS:
221 return &sm1_shader_frontend;
222
223 case WINED3D_SM4_PS:
224 case WINED3D_SM4_VS:
225 case WINED3D_SM4_GS:
226 return &sm4_shader_frontend;
227
228 default:
229 FIXME("Unrecognised version token %#x\n", version_token);
230 return NULL;
231 }
232 }
233
234 void shader_buffer_clear(struct wined3d_shader_buffer *buffer)
235 {
236 buffer->buffer[0] = '\0';
237 buffer->bsize = 0;
238 buffer->lineNo = 0;
239 buffer->newline = TRUE;
240 }
241
242 BOOL shader_buffer_init(struct wined3d_shader_buffer *buffer)
243 {
244 buffer->buffer = HeapAlloc(GetProcessHeap(), 0, SHADER_PGMSIZE);
245 if (!buffer->buffer)
246 {
247 ERR("Failed to allocate shader buffer memory.\n");
248 return FALSE;
249 }
250
251 shader_buffer_clear(buffer);
252 return TRUE;
253 }
254
255 void shader_buffer_free(struct wined3d_shader_buffer *buffer)
256 {
257 HeapFree(GetProcessHeap(), 0, buffer->buffer);
258 }
259
260 int shader_vaddline(struct wined3d_shader_buffer *buffer, const char *format, va_list args)
261 {
262 char *base = buffer->buffer + buffer->bsize;
263 int rc;
264
265 rc = vsnprintf(base, SHADER_PGMSIZE - 1 - buffer->bsize, format, args);
266
267 if (rc < 0 /* C89 */ || (unsigned int)rc > SHADER_PGMSIZE - 1 - buffer->bsize /* C99 */)
268 {
269 ERR("The buffer allocated for the shader program string "
270 "is too small at %d bytes.\n", SHADER_PGMSIZE);
271 buffer->bsize = SHADER_PGMSIZE - 1;
272 return -1;
273 }
274
275 if (buffer->newline)
276 {
277 TRACE("GL HW (%u, %u) : %s", buffer->lineNo + 1, buffer->bsize, base);
278 buffer->newline = FALSE;
279 }
280 else
281 {
282 TRACE("%s", base);
283 }
284
285 buffer->bsize += rc;
286 if (buffer->buffer[buffer->bsize-1] == '\n')
287 {
288 ++buffer->lineNo;
289 buffer->newline = TRUE;
290 }
291
292 return 0;
293 }
294
295 int shader_addline(struct wined3d_shader_buffer *buffer, const char *format, ...)
296 {
297 va_list args;
298 int ret;
299
300 va_start(args, format);
301 ret = shader_vaddline(buffer, format, args);
302 va_end(args);
303
304 return ret;
305 }
306
307 static void shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
308 void *parent, const struct wined3d_parent_ops *parent_ops)
309 {
310 shader->ref = 1;
311 shader->device = device;
312 shader->parent = parent;
313 shader->parent_ops = parent_ops;
314 list_init(&shader->linked_programs);
315 list_add_head(&device->shaders, &shader->shader_list_entry);
316 }
317
318 /* Convert floating point offset relative to a register file to an absolute
319 * offset for float constants. */
320 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
321 {
322 switch (register_type)
323 {
324 case WINED3DSPR_CONST: return register_idx;
325 case WINED3DSPR_CONST2: return 2048 + register_idx;
326 case WINED3DSPR_CONST3: return 4096 + register_idx;
327 case WINED3DSPR_CONST4: return 6144 + register_idx;
328 default:
329 FIXME("Unsupported register type: %u.\n", register_type);
330 return register_idx;
331 }
332 }
333
334 static void shader_delete_constant_list(struct list *clist)
335 {
336 struct wined3d_shader_lconst *constant;
337 struct list *ptr;
338
339 ptr = list_head(clist);
340 while (ptr)
341 {
342 constant = LIST_ENTRY(ptr, struct wined3d_shader_lconst, entry);
343 ptr = list_next(clist, ptr);
344 HeapFree(GetProcessHeap(), 0, constant);
345 }
346 list_init(clist);
347 }
348
349 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
350 {
351 DWORD idx, shift;
352 idx = bit >> 5;
353 shift = bit & 0x1f;
354 bitmap[idx] |= (1 << shift);
355 }
356
357 static void shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
358 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type)
359 {
360 switch (reg->type)
361 {
362 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
363 if (shader_type == WINED3D_SHADER_TYPE_PIXEL) reg_maps->texcoord |= 1 << reg->idx;
364 else reg_maps->address |= 1 << reg->idx;
365 break;
366
367 case WINED3DSPR_TEMP:
368 reg_maps->temporary |= 1 << reg->idx;
369 break;
370
371 case WINED3DSPR_INPUT:
372 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
373 {
374 if (reg->rel_addr)
375 {
376 /* If relative addressing is used, we must assume that all registers
377 * are used. Even if it is a construct like v3[aL], we can't assume
378 * that v0, v1 and v2 aren't read because aL can be negative */
379 unsigned int i;
380 for (i = 0; i < MAX_REG_INPUT; ++i)
381 {
382 shader->u.ps.input_reg_used[i] = TRUE;
383 }
384 }
385 else
386 {
387 shader->u.ps.input_reg_used[reg->idx] = TRUE;
388 }
389 }
390 else reg_maps->input_registers |= 1 << reg->idx;
391 break;
392
393 case WINED3DSPR_RASTOUT:
394 if (reg->idx == 1) reg_maps->fog = 1;
395 break;
396
397 case WINED3DSPR_MISCTYPE:
398 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
399 {
400 if (!reg->idx) reg_maps->vpos = 1;
401 else if (reg->idx == 1) reg_maps->usesfacing = 1;
402 }
403 break;
404
405 case WINED3DSPR_CONST:
406 if (reg->rel_addr)
407 {
408 if (reg->idx < reg_maps->min_rel_offset) reg_maps->min_rel_offset = reg->idx;
409 if (reg->idx > reg_maps->max_rel_offset) reg_maps->max_rel_offset = reg->idx;
410 reg_maps->usesrelconstF = TRUE;
411 }
412 else
413 {
414 set_bitmap_bit(reg_maps->constf, reg->idx);
415 }
416 break;
417
418 case WINED3DSPR_CONSTINT:
419 reg_maps->integer_constants |= (1 << reg->idx);
420 break;
421
422 case WINED3DSPR_CONSTBOOL:
423 reg_maps->boolean_constants |= (1 << reg->idx);
424 break;
425
426 case WINED3DSPR_COLOROUT:
427 reg_maps->rt_mask |= (1 << reg->idx);
428 break;
429
430 default:
431 TRACE("Not recording register of type %#x and idx %u\n", reg->type, reg->idx);
432 break;
433 }
434 }
435
436 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
437 {
438 switch (instr)
439 {
440 case WINED3DSIH_M4x4:
441 case WINED3DSIH_M3x4:
442 return param == 1 ? 3 : 0;
443
444 case WINED3DSIH_M4x3:
445 case WINED3DSIH_M3x3:
446 return param == 1 ? 2 : 0;
447
448 case WINED3DSIH_M3x2:
449 return param == 1 ? 1 : 0;
450
451 default:
452 return 0;
453 }
454 }
455
456 /* Note that this does not count the loop register as an address register. */
457 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
458 struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature_element *input_signature,
459 struct wined3d_shader_signature_element *output_signature, const DWORD *byte_code, DWORD constf_size)
460 {
461 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
462 void *fe_data = shader->frontend_data;
463 struct wined3d_shader_version shader_version;
464 const DWORD *ptr = byte_code;
465
466 memset(reg_maps, 0, sizeof(*reg_maps));
467 reg_maps->min_rel_offset = ~0U;
468
469 fe->shader_read_header(fe_data, &ptr, &shader_version);
470 reg_maps->shader_version = shader_version;
471
472 reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
473 sizeof(*reg_maps->constf) * ((constf_size + 31) / 32));
474 if (!reg_maps->constf)
475 {
476 ERR("Failed to allocate constant map memory.\n");
477 return E_OUTOFMEMORY;
478 }
479
480 while (!fe->shader_is_end(fe_data, &ptr))
481 {
482 struct wined3d_shader_instruction ins;
483 const char *comment;
484 UINT comment_size;
485 UINT param_size;
486
487 /* Skip comments. */
488 fe->shader_read_comment(&ptr, &comment, &comment_size);
489 if (comment) continue;
490
491 /* Fetch opcode. */
492 fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
493
494 /* Unhandled opcode, and its parameters. */
495 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
496 {
497 TRACE("Skipping unrecognized instruction.\n");
498 ptr += param_size;
499 continue;
500 }
501
502 /* Handle declarations. */
503 if (ins.handler_idx == WINED3DSIH_DCL)
504 {
505 struct wined3d_shader_semantic semantic;
506
507 fe->shader_read_semantic(&ptr, &semantic);
508
509 switch (semantic.reg.reg.type)
510 {
511 /* Mark input registers used. */
512 case WINED3DSPR_INPUT:
513 reg_maps->input_registers |= 1 << semantic.reg.reg.idx;
514 shader_signature_from_semantic(&input_signature[semantic.reg.reg.idx], &semantic);
515 break;
516
517 /* Vertex shader: mark 3.0 output registers used, save token. */
518 case WINED3DSPR_OUTPUT:
519 reg_maps->output_registers |= 1 << semantic.reg.reg.idx;
520 shader_signature_from_semantic(&output_signature[semantic.reg.reg.idx], &semantic);
521 if (semantic.usage == WINED3DDECLUSAGE_FOG) reg_maps->fog = 1;
522 break;
523
524 /* Save sampler usage token. */
525 case WINED3DSPR_SAMPLER:
526 reg_maps->sampler_type[semantic.reg.reg.idx] = semantic.sampler_type;
527 break;
528
529 default:
530 TRACE("Not recording DCL register type %#x.\n", semantic.reg.reg.type);
531 break;
532 }
533 }
534 else if (ins.handler_idx == WINED3DSIH_DEF)
535 {
536 struct wined3d_shader_src_param rel_addr;
537 struct wined3d_shader_dst_param dst;
538
539 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
540 if (!lconst) return E_OUTOFMEMORY;
541
542 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
543 lconst->idx = dst.reg.idx;
544
545 memcpy(lconst->value, ptr, 4 * sizeof(DWORD));
546 ptr += 4;
547
548 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
549 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
550 {
551 float *value = (float *)lconst->value;
552 if (value[0] < -1.0f) value[0] = -1.0f;
553 else if (value[0] > 1.0f) value[0] = 1.0f;
554 if (value[1] < -1.0f) value[1] = -1.0f;
555 else if (value[1] > 1.0f) value[1] = 1.0f;
556 if (value[2] < -1.0f) value[2] = -1.0f;
557 else if (value[2] > 1.0f) value[2] = 1.0f;
558 if (value[3] < -1.0f) value[3] = -1.0f;
559 else if (value[3] > 1.0f) value[3] = 1.0f;
560 }
561
562 list_add_head(&shader->constantsF, &lconst->entry);
563 }
564 else if (ins.handler_idx == WINED3DSIH_DEFI)
565 {
566 struct wined3d_shader_src_param rel_addr;
567 struct wined3d_shader_dst_param dst;
568
569 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
570 if (!lconst) return E_OUTOFMEMORY;
571
572 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
573 lconst->idx = dst.reg.idx;
574
575 memcpy(lconst->value, ptr, 4 * sizeof(DWORD));
576 ptr += 4;
577
578 list_add_head(&shader->constantsI, &lconst->entry);
579 reg_maps->local_int_consts |= (1 << dst.reg.idx);
580 }
581 else if (ins.handler_idx == WINED3DSIH_DEFB)
582 {
583 struct wined3d_shader_src_param rel_addr;
584 struct wined3d_shader_dst_param dst;
585
586 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
587 if (!lconst) return E_OUTOFMEMORY;
588
589 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
590 lconst->idx = dst.reg.idx;
591
592 memcpy(lconst->value, ptr, sizeof(DWORD));
593 ++ptr;
594
595 list_add_head(&shader->constantsB, &lconst->entry);
596 reg_maps->local_bool_consts |= (1 << dst.reg.idx);
597 }
598 /* If there's a loop in the shader. */
599 else if (ins.handler_idx == WINED3DSIH_LOOP
600 || ins.handler_idx == WINED3DSIH_REP)
601 {
602 struct wined3d_shader_src_param src, rel_addr;
603
604 fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
605
606 /* Rep and Loop always use an integer constant for the control parameters. */
607 if (ins.handler_idx == WINED3DSIH_REP)
608 {
609 reg_maps->integer_constants |= 1 << src.reg.idx;
610 }
611 else
612 {
613 fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
614 reg_maps->integer_constants |= 1 << src.reg.idx;
615 }
616
617 cur_loop_depth++;
618 if (cur_loop_depth > max_loop_depth) max_loop_depth = cur_loop_depth;
619 }
620 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
621 || ins.handler_idx == WINED3DSIH_ENDREP)
622 {
623 cur_loop_depth--;
624 }
625 /* For subroutine prototypes. */
626 else if (ins.handler_idx == WINED3DSIH_LABEL)
627 {
628 struct wined3d_shader_src_param src, rel_addr;
629
630 fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
631 reg_maps->labels |= 1 << src.reg.idx;
632 }
633 /* Set texture, address, temporary registers. */
634 else
635 {
636 BOOL color0_mov = FALSE;
637 unsigned int i, limit;
638
639 /* This will loop over all the registers and try to
640 * make a bitmask of the ones we're interested in.
641 *
642 * Relative addressing tokens are ignored, but that's
643 * okay, since we'll catch any address registers when
644 * they are initialized (required by spec). */
645 for (i = 0; i < ins.dst_count; ++i)
646 {
647 struct wined3d_shader_src_param dst_rel_addr;
648 struct wined3d_shader_dst_param dst_param;
649
650 fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
651
652 shader_record_register_usage(shader, reg_maps, &dst_param.reg, shader_version.type);
653
654 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and
655 * is used in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel
656 * shaders because TECRDOUT isn't used in them, but future register types might cause issues */
657 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX && shader_version.major < 3)
658 {
659 UINT idx = dst_param.reg.idx;
660
661 switch (dst_param.reg.type)
662 {
663 case WINED3DSPR_RASTOUT:
664 switch (idx)
665 {
666 case 0: /* oPos */
667 reg_maps->output_registers |= 1 << 10;
668 shader_signature_from_usage(&output_signature[10],
669 WINED3DDECLUSAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
670 break;
671
672 case 1: /* oFog */
673 reg_maps->output_registers |= 1 << 11;
674 shader_signature_from_usage(&output_signature[11],
675 WINED3DDECLUSAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
676 break;
677
678 case 2: /* oPts */
679 reg_maps->output_registers |= 1 << 11;
680 shader_signature_from_usage(&output_signature[11],
681 WINED3DDECLUSAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
682 break;
683 }
684 break;
685
686 case WINED3DSPR_ATTROUT:
687 if (idx < 2)
688 {
689 idx += 8;
690 if (reg_maps->output_registers & (1 << idx))
691 {
692 output_signature[idx].mask |= dst_param.write_mask;
693 }
694 else
695 {
696 reg_maps->output_registers |= 1 << idx;
697 shader_signature_from_usage(&output_signature[idx],
698 WINED3DDECLUSAGE_COLOR, idx - 8, idx, dst_param.write_mask);
699 }
700 }
701 break;
702
703 case WINED3DSPR_TEXCRDOUT:
704
705 reg_maps->texcoord_mask[idx] |= dst_param.write_mask;
706 if (reg_maps->output_registers & (1 << idx))
707 {
708 output_signature[idx].mask |= dst_param.write_mask;
709 }
710 else
711 {
712 reg_maps->output_registers |= 1 << idx;
713 shader_signature_from_usage(&output_signature[idx],
714 WINED3DDECLUSAGE_TEXCOORD, idx, idx, dst_param.write_mask);
715 }
716 break;
717
718 default:
719 break;
720 }
721 }
722
723 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
724 {
725 if (dst_param.reg.type == WINED3DSPR_COLOROUT && !dst_param.reg.idx)
726 {
727 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
728 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
729 * the mov and perform the sRGB write correction from the source register.
730 *
731 * However, if the mov is only partial, we can't do this, and if the write
732 * comes from an instruction other than MOV it is hard to do as well. If
733 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
734 shader->u.ps.color0_mov = FALSE;
735 if (ins.handler_idx == WINED3DSIH_MOV
736 && dst_param.write_mask == WINED3DSP_WRITEMASK_ALL)
737 {
738 /* Used later when the source register is read. */
739 color0_mov = TRUE;
740 }
741 }
742 /* Also drop the MOV marker if the source register is overwritten prior to the shader
743 * end
744 */
745 else if (dst_param.reg.type == WINED3DSPR_TEMP
746 && dst_param.reg.idx == shader->u.ps.color0_reg)
747 {
748 shader->u.ps.color0_mov = FALSE;
749 }
750 }
751
752 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
753 if (shader_version.major == 1
754 && (ins.handler_idx == WINED3DSIH_TEX
755 || ins.handler_idx == WINED3DSIH_TEXBEM
756 || ins.handler_idx == WINED3DSIH_TEXBEML
757 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
758 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
759 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
760 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
761 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
762 || ins.handler_idx == WINED3DSIH_TEXREG2AR
763 || ins.handler_idx == WINED3DSIH_TEXREG2GB
764 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
765 {
766 /* Fake sampler usage, only set reserved bit and type. */
767 DWORD sampler_code = dst_param.reg.idx;
768
769 TRACE("Setting fake 2D sampler for 1.x pixelshader.\n");
770 reg_maps->sampler_type[sampler_code] = WINED3DSTT_2D;
771
772 /* texbem is only valid with < 1.4 pixel shaders */
773 if (ins.handler_idx == WINED3DSIH_TEXBEM
774 || ins.handler_idx == WINED3DSIH_TEXBEML)
775 {
776 reg_maps->bumpmat |= 1 << dst_param.reg.idx;
777 if (ins.handler_idx == WINED3DSIH_TEXBEML)
778 {
779 reg_maps->luminanceparams |= 1 << dst_param.reg.idx;
780 }
781 }
782 }
783 else if (ins.handler_idx == WINED3DSIH_BEM)
784 {
785 reg_maps->bumpmat |= 1 << dst_param.reg.idx;
786 }
787 }
788
789 if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
790 else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
791 else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
792 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
793 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
794 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
795 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
796 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
797 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
798
799 limit = ins.src_count + (ins.predicate ? 1 : 0);
800 for (i = 0; i < limit; ++i)
801 {
802 struct wined3d_shader_src_param src_param, src_rel_addr;
803 unsigned int count;
804
805 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
806 count = get_instr_extra_regcount(ins.handler_idx, i);
807
808 shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
809 while (count)
810 {
811 ++src_param.reg.idx;
812 shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
813 --count;
814 }
815
816 if (color0_mov)
817 {
818 if (src_param.reg.type == WINED3DSPR_TEMP
819 && src_param.swizzle == WINED3DSP_NOSWIZZLE)
820 {
821 shader->u.ps.color0_mov = TRUE;
822 shader->u.ps.color0_reg = src_param.reg.idx;
823 }
824 }
825 }
826 }
827 }
828 reg_maps->loop_depth = max_loop_depth;
829
830 /* PS before 2.0 don't have explicit color outputs. Instead the value of
831 * R0 is written to the render target. */
832 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
833 reg_maps->rt_mask |= (1 << 0);
834
835 shader->functionLength = ((const char *)ptr - (const char *)byte_code);
836
837 return WINED3D_OK;
838 }
839
840 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
841 {
842 DWORD map = 1 << max;
843 map |= map - 1;
844 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
845
846 return wined3d_log2i(map);
847 }
848
849 static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
850 const struct wined3d_shader_version *shader_version)
851 {
852 TRACE("dcl");
853
854 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
855 {
856 switch (semantic->sampler_type)
857 {
858 case WINED3DSTT_2D: TRACE("_2d"); break;
859 case WINED3DSTT_CUBE: TRACE("_cube"); break;
860 case WINED3DSTT_VOLUME: TRACE("_volume"); break;
861 default: TRACE("_unknown_ttype(0x%08x)", semantic->sampler_type);
862 }
863 }
864 else
865 {
866 /* Pixel shaders 3.0 don't have usage semantics. */
867 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL) return;
868 else TRACE("_");
869
870 switch (semantic->usage)
871 {
872 case WINED3DDECLUSAGE_POSITION:
873 TRACE("position%u", semantic->usage_idx);
874 break;
875
876 case WINED3DDECLUSAGE_BLENDINDICES:
877 TRACE("blend");
878 break;
879
880 case WINED3DDECLUSAGE_BLENDWEIGHT:
881 TRACE("weight");
882 break;
883
884 case WINED3DDECLUSAGE_NORMAL:
885 TRACE("normal%u", semantic->usage_idx);
886 break;
887
888 case WINED3DDECLUSAGE_PSIZE:
889 TRACE("psize");
890 break;
891
892 case WINED3DDECLUSAGE_COLOR:
893 if (!semantic->usage_idx) TRACE("color");
894 else TRACE("specular%u", (semantic->usage_idx - 1));
895 break;
896
897 case WINED3DDECLUSAGE_TEXCOORD:
898 TRACE("texture%u", semantic->usage_idx);
899 break;
900
901 case WINED3DDECLUSAGE_TANGENT:
902 TRACE("tangent");
903 break;
904
905 case WINED3DDECLUSAGE_BINORMAL:
906 TRACE("binormal");
907 break;
908
909 case WINED3DDECLUSAGE_TESSFACTOR:
910 TRACE("tessfactor");
911 break;
912
913 case WINED3DDECLUSAGE_POSITIONT:
914 TRACE("positionT%u", semantic->usage_idx);
915 break;
916
917 case WINED3DDECLUSAGE_FOG:
918 TRACE("fog");
919 break;
920
921 case WINED3DDECLUSAGE_DEPTH:
922 TRACE("depth");
923 break;
924
925 case WINED3DDECLUSAGE_SAMPLE:
926 TRACE("sample");
927 break;
928
929 default:
930 FIXME("unknown_semantics(0x%08x)", semantic->usage);
931 }
932 }
933 }
934
935 static void shader_dump_register(const struct wined3d_shader_register *reg,
936 const struct wined3d_shader_version *shader_version)
937 {
938 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
939 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
940 UINT offset = reg->idx;
941
942 switch (reg->type)
943 {
944 case WINED3DSPR_TEMP:
945 TRACE("r");
946 break;
947
948 case WINED3DSPR_INPUT:
949 TRACE("v");
950 break;
951
952 case WINED3DSPR_CONST:
953 case WINED3DSPR_CONST2:
954 case WINED3DSPR_CONST3:
955 case WINED3DSPR_CONST4:
956 TRACE("c");
957 offset = shader_get_float_offset(reg->type, reg->idx);
958 break;
959
960 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
961 TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
962 break;
963
964 case WINED3DSPR_RASTOUT:
965 TRACE("%s", rastout_reg_names[reg->idx]);
966 break;
967
968 case WINED3DSPR_COLOROUT:
969 TRACE("oC");
970 break;
971
972 case WINED3DSPR_DEPTHOUT:
973 TRACE("oDepth");
974 break;
975
976 case WINED3DSPR_ATTROUT:
977 TRACE("oD");
978 break;
979
980 case WINED3DSPR_TEXCRDOUT:
981 /* Vertex shaders >= 3.0 use general purpose output registers
982 * (WINED3DSPR_OUTPUT), which can include an address token. */
983 if (shader_version->major >= 3) TRACE("o");
984 else TRACE("oT");
985 break;
986
987 case WINED3DSPR_CONSTINT:
988 TRACE("i");
989 break;
990
991 case WINED3DSPR_CONSTBOOL:
992 TRACE("b");
993 break;
994
995 case WINED3DSPR_LABEL:
996 TRACE("l");
997 break;
998
999 case WINED3DSPR_LOOP:
1000 TRACE("aL");
1001 break;
1002
1003 case WINED3DSPR_SAMPLER:
1004 TRACE("s");
1005 break;
1006
1007 case WINED3DSPR_MISCTYPE:
1008 if (reg->idx > 1) FIXME("Unhandled misctype register %u.\n", reg->idx);
1009 else TRACE("%s", misctype_reg_names[reg->idx]);
1010 break;
1011
1012 case WINED3DSPR_PREDICATE:
1013 TRACE("p");
1014 break;
1015
1016 case WINED3DSPR_IMMCONST:
1017 TRACE("l");
1018 break;
1019
1020 case WINED3DSPR_CONSTBUFFER:
1021 TRACE("cb");
1022 break;
1023
1024 case WINED3DSPR_NULL:
1025 TRACE("null");
1026 break;
1027
1028 case WINED3DSPR_RESOURCE:
1029 TRACE("t");
1030 break;
1031
1032 default:
1033 TRACE("unhandled_rtype(%#x)", reg->type);
1034 break;
1035 }
1036
1037 if (reg->type == WINED3DSPR_IMMCONST)
1038 {
1039 TRACE("(");
1040 switch (reg->immconst_type)
1041 {
1042 case WINED3D_IMMCONST_SCALAR:
1043 TRACE("%.8e", *(const float *)reg->immconst_data);
1044 break;
1045
1046 case WINED3D_IMMCONST_VEC4:
1047 TRACE("%.8e, %.8e, %.8e, %.8e",
1048 *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
1049 *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
1050 break;
1051
1052 default:
1053 TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
1054 break;
1055 }
1056 TRACE(")");
1057 }
1058 else if (reg->type != WINED3DSPR_RASTOUT
1059 && reg->type != WINED3DSPR_MISCTYPE
1060 && reg->type != WINED3DSPR_NULL)
1061 {
1062 if (reg->array_idx != ~0U)
1063 {
1064 TRACE("%u[%u", offset, reg->array_idx);
1065 if (reg->rel_addr)
1066 {
1067 TRACE(" + ");
1068 shader_dump_src_param(reg->rel_addr, shader_version);
1069 }
1070 TRACE("]");
1071 }
1072 else
1073 {
1074 if (reg->rel_addr)
1075 {
1076 TRACE("[");
1077 shader_dump_src_param(reg->rel_addr, shader_version);
1078 TRACE(" + ");
1079 }
1080 TRACE("%u", offset);
1081 if (reg->rel_addr) TRACE("]");
1082 }
1083 }
1084 }
1085
1086 void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
1087 const struct wined3d_shader_version *shader_version)
1088 {
1089 DWORD write_mask = param->write_mask;
1090
1091 shader_dump_register(&param->reg, shader_version);
1092
1093 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
1094 {
1095 static const char *write_mask_chars = "xyzw";
1096
1097 TRACE(".");
1098 if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
1099 if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
1100 if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
1101 if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
1102 }
1103 }
1104
1105 void shader_dump_src_param(const struct wined3d_shader_src_param *param,
1106 const struct wined3d_shader_version *shader_version)
1107 {
1108 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
1109 DWORD swizzle = param->swizzle;
1110
1111 if (src_modifier == WINED3DSPSM_NEG
1112 || src_modifier == WINED3DSPSM_BIASNEG
1113 || src_modifier == WINED3DSPSM_SIGNNEG
1114 || src_modifier == WINED3DSPSM_X2NEG
1115 || src_modifier == WINED3DSPSM_ABSNEG)
1116 TRACE("-");
1117 else if (src_modifier == WINED3DSPSM_COMP)
1118 TRACE("1-");
1119 else if (src_modifier == WINED3DSPSM_NOT)
1120 TRACE("!");
1121
1122 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1123 TRACE("abs(");
1124
1125 shader_dump_register(&param->reg, shader_version);
1126
1127 if (src_modifier)
1128 {
1129 switch (src_modifier)
1130 {
1131 case WINED3DSPSM_NONE: break;
1132 case WINED3DSPSM_NEG: break;
1133 case WINED3DSPSM_NOT: break;
1134 case WINED3DSPSM_BIAS: TRACE("_bias"); break;
1135 case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
1136 case WINED3DSPSM_SIGN: TRACE("_bx2"); break;
1137 case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1138 case WINED3DSPSM_COMP: break;
1139 case WINED3DSPSM_X2: TRACE("_x2"); break;
1140 case WINED3DSPSM_X2NEG: TRACE("_x2"); break;
1141 case WINED3DSPSM_DZ: TRACE("_dz"); break;
1142 case WINED3DSPSM_DW: TRACE("_dw"); break;
1143 case WINED3DSPSM_ABSNEG: TRACE(")"); break;
1144 case WINED3DSPSM_ABS: TRACE(")"); break;
1145 default: TRACE("_unknown_modifier(%#x)", src_modifier);
1146 }
1147 }
1148
1149 if (swizzle != WINED3DSP_NOSWIZZLE)
1150 {
1151 static const char *swizzle_chars = "xyzw";
1152 DWORD swizzle_x = swizzle & 0x03;
1153 DWORD swizzle_y = (swizzle >> 2) & 0x03;
1154 DWORD swizzle_z = (swizzle >> 4) & 0x03;
1155 DWORD swizzle_w = (swizzle >> 6) & 0x03;
1156
1157 if (swizzle_x == swizzle_y
1158 && swizzle_x == swizzle_z
1159 && swizzle_x == swizzle_w)
1160 {
1161 TRACE(".%c", swizzle_chars[swizzle_x]);
1162 }
1163 else
1164 {
1165 TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1166 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1167 }
1168 }
1169 }
1170
1171 /* Shared code in order to generate the bulk of the shader string.
1172 * NOTE: A description of how to parse tokens can be found on MSDN. */
1173 void shader_generate_main(const struct wined3d_shader *shader, struct wined3d_shader_buffer *buffer,
1174 const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1175 {
1176 struct wined3d_device *device = shader->device;
1177 const struct wined3d_shader_frontend *fe = shader->frontend;
1178 void *fe_data = shader->frontend_data;
1179 struct wined3d_shader_src_param dst_rel_addr[2];
1180 struct wined3d_shader_src_param src_rel_addr[4];
1181 struct wined3d_shader_dst_param dst_param[2];
1182 struct wined3d_shader_src_param src_param[4];
1183 struct wined3d_shader_version shader_version;
1184 struct wined3d_shader_loop_state loop_state;
1185 struct wined3d_shader_instruction ins;
1186 struct wined3d_shader_tex_mx tex_mx;
1187 struct wined3d_shader_context ctx;
1188 const DWORD *ptr = byte_code;
1189 DWORD i;
1190
1191 /* Initialize current parsing state. */
1192 tex_mx.current_row = 0;
1193 loop_state.current_depth = 0;
1194 loop_state.current_reg = 0;
1195
1196 ctx.shader = shader;
1197 ctx.gl_info = &device->adapter->gl_info;
1198 ctx.reg_maps = reg_maps;
1199 ctx.buffer = buffer;
1200 ctx.tex_mx = &tex_mx;
1201 ctx.loop_state = &loop_state;
1202 ctx.backend_data = backend_ctx;
1203
1204 ins.ctx = &ctx;
1205 ins.dst = dst_param;
1206 ins.src = src_param;
1207
1208 fe->shader_read_header(fe_data, &ptr, &shader_version);
1209
1210 while (!fe->shader_is_end(fe_data, &ptr))
1211 {
1212 const char *comment;
1213 UINT comment_size;
1214 UINT param_size;
1215
1216 /* Skip comment tokens. */
1217 fe->shader_read_comment(&ptr, &comment, &comment_size);
1218 if (comment) continue;
1219
1220 /* Read opcode. */
1221 fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1222
1223 /* Unknown opcode and its parameters. */
1224 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1225 {
1226 TRACE("Skipping unrecognized instruction.\n");
1227 ptr += param_size;
1228 continue;
1229 }
1230
1231 /* Nothing to do. */
1232 if (ins.handler_idx == WINED3DSIH_DCL
1233 || ins.handler_idx == WINED3DSIH_NOP
1234 || ins.handler_idx == WINED3DSIH_DEF
1235 || ins.handler_idx == WINED3DSIH_DEFI
1236 || ins.handler_idx == WINED3DSIH_DEFB
1237 || ins.handler_idx == WINED3DSIH_PHASE)
1238 {
1239 ptr += param_size;
1240 continue;
1241 }
1242
1243 /* Destination tokens */
1244 for (i = 0; i < ins.dst_count; ++i)
1245 {
1246 fe->shader_read_dst_param(fe_data, &ptr, &dst_param[i], &dst_rel_addr[i]);
1247 }
1248
1249 /* Predication token */
1250 if (ins.predicate)
1251 {
1252 FIXME("Predicates not implemented.\n");
1253 ins.predicate = *ptr++;
1254 }
1255
1256 /* Other source tokens */
1257 for (i = 0; i < ins.src_count; ++i)
1258 {
1259 fe->shader_read_src_param(fe_data, &ptr, &src_param[i], &src_rel_addr[i]);
1260 }
1261
1262 /* Call appropriate function for output target */
1263 device->shader_backend->shader_handle_instruction(&ins);
1264 }
1265 }
1266
1267 static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1268 {
1269 DWORD mmask = dst->modifiers;
1270
1271 switch (dst->shift)
1272 {
1273 case 0: break;
1274 case 13: TRACE("_d8"); break;
1275 case 14: TRACE("_d4"); break;
1276 case 15: TRACE("_d2"); break;
1277 case 1: TRACE("_x2"); break;
1278 case 2: TRACE("_x4"); break;
1279 case 3: TRACE("_x8"); break;
1280 default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1281 }
1282
1283 if (mmask & WINED3DSPDM_SATURATE) TRACE("_sat");
1284 if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1285 if (mmask & WINED3DSPDM_MSAMPCENTROID) TRACE("_centroid");
1286
1287 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1288 if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1289 }
1290
1291 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1292 {
1293 struct wined3d_shader_version shader_version;
1294 const DWORD *ptr = byte_code;
1295 const char *type_prefix;
1296 DWORD i;
1297
1298 TRACE("Parsing %p.\n", byte_code);
1299
1300 fe->shader_read_header(fe_data, &ptr, &shader_version);
1301
1302 switch (shader_version.type)
1303 {
1304 case WINED3D_SHADER_TYPE_VERTEX:
1305 type_prefix = "vs";
1306 break;
1307
1308 case WINED3D_SHADER_TYPE_GEOMETRY:
1309 type_prefix = "gs";
1310 break;
1311
1312 case WINED3D_SHADER_TYPE_PIXEL:
1313 type_prefix = "ps";
1314 break;
1315
1316 default:
1317 FIXME("Unhandled shader type %#x.\n", shader_version.type);
1318 type_prefix = "unknown";
1319 break;
1320 }
1321
1322 TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1323
1324 while (!fe->shader_is_end(fe_data, &ptr))
1325 {
1326 struct wined3d_shader_instruction ins;
1327 const char *comment;
1328 UINT comment_size;
1329 UINT param_size;
1330
1331 /* comment */
1332 fe->shader_read_comment(&ptr, &comment, &comment_size);
1333 if (comment)
1334 {
1335 if (comment_size > 4 && *(const DWORD *)comment == WINEMAKEFOURCC('T', 'E', 'X', 'T'))
1336 {
1337 const char *end = comment + comment_size;
1338 const char *ptr = comment + 4;
1339 const char *line = ptr;
1340
1341 TRACE("// TEXT\n");
1342 while (ptr != end)
1343 {
1344 if (*ptr == '\n')
1345 {
1346 UINT len = ptr - line;
1347 if (len && *(ptr - 1) == '\r') --len;
1348 TRACE("// %s\n", debugstr_an(line, len));
1349 line = ++ptr;
1350 }
1351 else ++ptr;
1352 }
1353 if (line != ptr) TRACE("// %s\n", debugstr_an(line, ptr - line));
1354 }
1355 else TRACE("// %s\n", debugstr_an(comment, comment_size));
1356 continue;
1357 }
1358
1359 fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1360 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1361 {
1362 TRACE("Skipping unrecognized instruction.\n");
1363 ptr += param_size;
1364 continue;
1365 }
1366
1367 if (ins.handler_idx == WINED3DSIH_DCL)
1368 {
1369 struct wined3d_shader_semantic semantic;
1370
1371 fe->shader_read_semantic(&ptr, &semantic);
1372
1373 shader_dump_decl_usage(&semantic, &shader_version);
1374 shader_dump_ins_modifiers(&semantic.reg);
1375 TRACE(" ");
1376 shader_dump_dst_param(&semantic.reg, &shader_version);
1377 }
1378 else if (ins.handler_idx == WINED3DSIH_DEF)
1379 {
1380 struct wined3d_shader_src_param rel_addr;
1381 struct wined3d_shader_dst_param dst;
1382
1383 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1384
1385 TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(dst.reg.type, dst.reg.idx),
1386 *(const float *)(ptr),
1387 *(const float *)(ptr + 1),
1388 *(const float *)(ptr + 2),
1389 *(const float *)(ptr + 3));
1390 ptr += 4;
1391 }
1392 else if (ins.handler_idx == WINED3DSIH_DEFI)
1393 {
1394 struct wined3d_shader_src_param rel_addr;
1395 struct wined3d_shader_dst_param dst;
1396
1397 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1398
1399 TRACE("defi i%u = %d, %d, %d, %d", dst.reg.idx,
1400 *(ptr),
1401 *(ptr + 1),
1402 *(ptr + 2),
1403 *(ptr + 3));
1404 ptr += 4;
1405 }
1406 else if (ins.handler_idx == WINED3DSIH_DEFB)
1407 {
1408 struct wined3d_shader_src_param rel_addr;
1409 struct wined3d_shader_dst_param dst;
1410
1411 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1412
1413 TRACE("defb b%u = %s", dst.reg.idx, *ptr ? "true" : "false");
1414 ++ptr;
1415 }
1416 else
1417 {
1418 struct wined3d_shader_src_param dst_rel_addr[2];
1419 struct wined3d_shader_src_param src_rel_addr;
1420 struct wined3d_shader_dst_param dst_param[2];
1421 struct wined3d_shader_src_param src_param;
1422
1423 for (i = 0; i < ins.dst_count; ++i)
1424 {
1425 fe->shader_read_dst_param(fe_data, &ptr, &dst_param[i], &dst_rel_addr[i]);
1426 }
1427
1428 /* Print out predication source token first - it follows
1429 * the destination token. */
1430 if (ins.predicate)
1431 {
1432 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1433 TRACE("(");
1434 shader_dump_src_param(&src_param, &shader_version);
1435 TRACE(") ");
1436 }
1437
1438 /* PixWin marks instructions with the coissue flag with a '+' */
1439 if (ins.coissue) TRACE("+");
1440
1441 TRACE("%s", shader_opcode_names[ins.handler_idx]);
1442
1443 if (ins.handler_idx == WINED3DSIH_IFC
1444 || ins.handler_idx == WINED3DSIH_BREAKC)
1445 {
1446 switch (ins.flags)
1447 {
1448 case WINED3D_SHADER_REL_OP_GT: TRACE("_gt"); break;
1449 case WINED3D_SHADER_REL_OP_EQ: TRACE("_eq"); break;
1450 case WINED3D_SHADER_REL_OP_GE: TRACE("_ge"); break;
1451 case WINED3D_SHADER_REL_OP_LT: TRACE("_lt"); break;
1452 case WINED3D_SHADER_REL_OP_NE: TRACE("_ne"); break;
1453 case WINED3D_SHADER_REL_OP_LE: TRACE("_le"); break;
1454 default: TRACE("_(%u)", ins.flags);
1455 }
1456 }
1457 else if (ins.handler_idx == WINED3DSIH_TEX
1458 && shader_version.major >= 2
1459 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1460 {
1461 TRACE("p");
1462 }
1463
1464 /* We already read the destination tokens, print them. */
1465 for (i = 0; i < ins.dst_count; ++i)
1466 {
1467 shader_dump_ins_modifiers(&dst_param[i]);
1468 TRACE(!i ? " " : ", ");
1469 shader_dump_dst_param(&dst_param[i], &shader_version);
1470 }
1471
1472 /* Other source tokens */
1473 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1474 {
1475 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1476 TRACE(!i ? " " : ", ");
1477 shader_dump_src_param(&src_param, &shader_version);
1478 }
1479 }
1480 TRACE("\n");
1481 }
1482 }
1483
1484 static void shader_cleanup(struct wined3d_shader *shader)
1485 {
1486 shader->device->shader_backend->shader_destroy(shader);
1487 HeapFree(GetProcessHeap(), 0, shader->reg_maps.constf);
1488 HeapFree(GetProcessHeap(), 0, shader->function);
1489 shader_delete_constant_list(&shader->constantsF);
1490 shader_delete_constant_list(&shader->constantsB);
1491 shader_delete_constant_list(&shader->constantsI);
1492 list_remove(&shader->shader_list_entry);
1493
1494 if (shader->frontend && shader->frontend_data)
1495 shader->frontend->shader_free(shader->frontend_data);
1496 }
1497
1498 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1499 static void shader_none_select(const struct wined3d_context *context, BOOL usePS, BOOL useVS) {}
1500 static void shader_none_select_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info,
1501 enum tex_types tex_type, const SIZE *ds_mask_size) {}
1502 static void shader_none_deselect_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info) {}
1503 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
1504 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
1505 static void shader_none_load_constants(const struct wined3d_context *context, char usePS, char useVS) {}
1506 static void shader_none_load_np2fixup_constants(void *shader_priv,
1507 const struct wined3d_gl_info *gl_info, const struct wined3d_state *state) {}
1508 static void shader_none_destroy(struct wined3d_shader *shader) {}
1509 static HRESULT shader_none_alloc(struct wined3d_device *device) {return WINED3D_OK;}
1510 static void shader_none_free(struct wined3d_device *device) {}
1511 static void shader_none_context_destroyed(void *shader_priv, const struct wined3d_context *context) {}
1512
1513 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
1514 {
1515 /* Set the shader caps to 0 for the none shader backend */
1516 caps->VertexShaderVersion = 0;
1517 caps->MaxVertexShaderConst = 0;
1518 caps->PixelShaderVersion = 0;
1519 caps->PixelShader1xMaxValue = 0.0f;
1520 caps->MaxPixelShaderConst = 0;
1521 caps->VSClipping = FALSE;
1522 }
1523
1524 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
1525 {
1526 if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1527 {
1528 TRACE("Checking support for fixup:\n");
1529 dump_color_fixup_desc(fixup);
1530 }
1531
1532 /* Faked to make some apps happy. */
1533 if (!is_complex_fixup(fixup))
1534 {
1535 TRACE("[OK]\n");
1536 return TRUE;
1537 }
1538
1539 TRACE("[FAILED]\n");
1540 return FALSE;
1541 }
1542
1543 const struct wined3d_shader_backend_ops none_shader_backend =
1544 {
1545 shader_none_handle_instruction,
1546 shader_none_select,
1547 shader_none_select_depth_blt,
1548 shader_none_deselect_depth_blt,
1549 shader_none_update_float_vertex_constants,
1550 shader_none_update_float_pixel_constants,
1551 shader_none_load_constants,
1552 shader_none_load_np2fixup_constants,
1553 shader_none_destroy,
1554 shader_none_alloc,
1555 shader_none_free,
1556 shader_none_context_destroyed,
1557 shader_none_get_caps,
1558 shader_none_color_fixup_supported,
1559 };
1560
1561 static HRESULT shader_set_function(struct wined3d_shader *shader, const DWORD *byte_code,
1562 const struct wined3d_shader_signature *output_signature, DWORD float_const_count,
1563 enum wined3d_shader_type type, unsigned int max_version)
1564 {
1565 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1566 const struct wined3d_shader_frontend *fe;
1567 HRESULT hr;
1568 unsigned int backend_version;
1569
1570 TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u.\n",
1571 shader, byte_code, output_signature, float_const_count);
1572
1573 fe = shader_select_frontend(*byte_code);
1574 if (!fe)
1575 {
1576 FIXME("Unable to find frontend for shader.\n");
1577 return WINED3DERR_INVALIDCALL;
1578 }
1579 shader->frontend = fe;
1580 shader->frontend_data = fe->shader_init(byte_code, output_signature);
1581 if (!shader->frontend_data)
1582 {
1583 FIXME("Failed to initialize frontend.\n");
1584 return WINED3DERR_INVALIDCALL;
1585 }
1586
1587 /* First pass: trace shader. */
1588 if (TRACE_ON(d3d_shader))
1589 shader_trace_init(fe, shader->frontend_data, byte_code);
1590
1591 /* Initialize immediate constant lists. */
1592 list_init(&shader->constantsF);
1593 list_init(&shader->constantsB);
1594 list_init(&shader->constantsI);
1595
1596 /* Second pass: figure out which registers are used, what the semantics are, etc. */
1597 hr = shader_get_registers_used(shader, fe,
1598 reg_maps, shader->input_signature, shader->output_signature,
1599 byte_code, float_const_count);
1600 if (FAILED(hr)) return hr;
1601
1602 if (reg_maps->shader_version.type != type)
1603 {
1604 WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
1605 return WINED3DERR_INVALIDCALL;
1606 }
1607 if (reg_maps->shader_version.major > max_version)
1608 {
1609 WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
1610 return WINED3DERR_INVALIDCALL;
1611 }
1612 switch (type)
1613 {
1614 case WINED3D_SHADER_TYPE_VERTEX:
1615 backend_version = shader->device->vshader_version;
1616 break;
1617 case WINED3D_SHADER_TYPE_PIXEL:
1618 backend_version = shader->device->pshader_version;
1619 break;
1620 default:
1621 FIXME("No backend version-checking for this shader type\n");
1622 backend_version = 0;
1623 }
1624 if (reg_maps->shader_version.major > backend_version)
1625 {
1626 WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
1627 reg_maps->shader_version.major, reg_maps->shader_version.minor);
1628 return WINED3DERR_INVALIDCALL;
1629 }
1630
1631 shader->function = HeapAlloc(GetProcessHeap(), 0, shader->functionLength);
1632 if (!shader->function)
1633 return E_OUTOFMEMORY;
1634 memcpy(shader->function, byte_code, shader->functionLength);
1635
1636 return WINED3D_OK;
1637 }
1638
1639 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
1640 {
1641 ULONG refcount = InterlockedIncrement(&shader->ref);
1642
1643 TRACE("%p increasing refcount to %u.\n", shader, refcount);
1644
1645 return refcount;
1646 }
1647
1648 /* Do not call while under the GL lock. */
1649 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
1650 {
1651 ULONG refcount = InterlockedDecrement(&shader->ref);
1652
1653 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1654
1655 if (!refcount)
1656 {
1657 shader_cleanup(shader);
1658 shader->parent_ops->wined3d_object_destroyed(shader->parent);
1659 HeapFree(GetProcessHeap(), 0, shader);
1660 }
1661
1662 return refcount;
1663 }
1664
1665 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
1666 {
1667 TRACE("shader %p.\n", shader);
1668
1669 return shader->parent;
1670 }
1671
1672 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
1673 void *byte_code, UINT *byte_code_size)
1674 {
1675 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
1676
1677 if (!byte_code)
1678 {
1679 *byte_code_size = shader->functionLength;
1680 return WINED3D_OK;
1681 }
1682
1683 if (*byte_code_size < shader->functionLength)
1684 {
1685 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
1686 * than the required size we should write the required size and
1687 * return D3DERR_MOREDATA. That's not actually true. */
1688 return WINED3DERR_INVALIDCALL;
1689 }
1690
1691 memcpy(byte_code, shader->function, shader->functionLength);
1692
1693 return WINED3D_OK;
1694 }
1695
1696 /* Set local constants for d3d8 shaders. */
1697 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
1698 UINT start_idx, const float *src_data, UINT count)
1699 {
1700 UINT end_idx = start_idx + count;
1701 UINT i;
1702
1703 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
1704
1705 if (end_idx > shader->limits.constant_float)
1706 {
1707 WARN("end_idx %u > float constants limit %u.\n",
1708 end_idx, shader->limits.constant_float);
1709 end_idx = shader->limits.constant_float;
1710 }
1711
1712 for (i = start_idx; i < end_idx; ++i)
1713 {
1714 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1715 if (!lconst)
1716 return E_OUTOFMEMORY;
1717
1718 lconst->idx = i;
1719 memcpy(lconst->value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
1720 list_add_head(&shader->constantsF, &lconst->entry);
1721 }
1722
1723 return WINED3D_OK;
1724 }
1725
1726 void find_vs_compile_args(const struct wined3d_state *state,
1727 const struct wined3d_shader *shader, struct vs_compile_args *args)
1728 {
1729 args->fog_src = state->render_states[WINED3DRS_FOGTABLEMODE]
1730 == WINED3DFOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
1731 args->clip_enabled = state->render_states[WINED3DRS_CLIPPING]
1732 && state->render_states[WINED3DRS_CLIPPLANEENABLE];
1733 args->swizzle_map = shader->device->strided_streams.swizzle_map;
1734 }
1735
1736 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
1737 {
1738 if (usage_idx1 != usage_idx2) return FALSE;
1739 if (usage1 == usage2) return TRUE;
1740 if (usage1 == WINED3DDECLUSAGE_POSITION && usage2 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1741 if (usage2 == WINED3DDECLUSAGE_POSITION && usage1 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1742
1743 return FALSE;
1744 }
1745
1746 BOOL vshader_get_input(const struct wined3d_shader *shader,
1747 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
1748 {
1749 WORD map = shader->reg_maps.input_registers;
1750 unsigned int i;
1751
1752 for (i = 0; map; map >>= 1, ++i)
1753 {
1754 if (!(map & 1)) continue;
1755
1756 if (match_usage(shader->u.vs.attributes[i].usage,
1757 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
1758 {
1759 *regnum = i;
1760 return TRUE;
1761 }
1762 }
1763 return FALSE;
1764 }
1765
1766 static void vertexshader_set_limits(struct wined3d_shader *shader)
1767 {
1768 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
1769 shader->reg_maps.shader_version.minor);
1770 struct wined3d_device *device = shader->device;
1771
1772 shader->limits.texcoord = 0;
1773 shader->limits.attributes = 16;
1774 shader->limits.packed_input = 0;
1775
1776 switch (shader_version)
1777 {
1778 case WINED3D_SHADER_VERSION(1, 0):
1779 case WINED3D_SHADER_VERSION(1, 1):
1780 shader->limits.temporary = 12;
1781 shader->limits.constant_bool = 0;
1782 shader->limits.constant_int = 0;
1783 shader->limits.address = 1;
1784 shader->limits.packed_output = 12;
1785 shader->limits.sampler = 0;
1786 shader->limits.label = 0;
1787 /* TODO: vs_1_1 has a minimum of 96 constants. What happens when
1788 * a vs_1_1 shader is used on a vs_3_0 capable card that has 256
1789 * constants? */
1790 shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1791 break;
1792
1793 case WINED3D_SHADER_VERSION(2, 0):
1794 case WINED3D_SHADER_VERSION(2, 1):
1795 shader->limits.temporary = 12;
1796 shader->limits.constant_bool = 16;
1797 shader->limits.constant_int = 16;
1798 shader->limits.address = 1;
1799 shader->limits.packed_output = 12;
1800 shader->limits.sampler = 0;
1801 shader->limits.label = 16;
1802 shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1803 break;
1804
1805 case WINED3D_SHADER_VERSION(4, 0):
1806 FIXME("Using 3.0 limits for 4.0 shader.\n");
1807 /* Fall through. */
1808
1809 case WINED3D_SHADER_VERSION(3, 0):
1810 shader->limits.temporary = 32;
1811 shader->limits.constant_bool = 32;
1812 shader->limits.constant_int = 32;
1813 shader->limits.address = 1;
1814 shader->limits.packed_output = 12;
1815 shader->limits.sampler = 4;
1816 shader->limits.label = 16; /* FIXME: 2048 */
1817 /* DX10 cards on Windows advertise a d3d9 constant limit of 256
1818 * even though they are capable of supporting much more (GL
1819 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
1820 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
1821 * shaders to 256. */
1822 shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1823 break;
1824
1825 default:
1826 shader->limits.temporary = 12;
1827 shader->limits.constant_bool = 16;
1828 shader->limits.constant_int = 16;
1829 shader->limits.address = 1;
1830 shader->limits.packed_output = 12;
1831 shader->limits.sampler = 0;
1832 shader->limits.label = 16;
1833 shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1834 FIXME("Unrecognized vertex shader version \"%u.%u\".\n",
1835 shader->reg_maps.shader_version.major,
1836 shader->reg_maps.shader_version.minor);
1837 }
1838 }
1839
1840 static HRESULT vertexshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
1841 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1842 void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
1843 {
1844 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1845 unsigned int i;
1846 HRESULT hr;
1847 WORD map;
1848
1849 if (!byte_code) return WINED3DERR_INVALIDCALL;
1850
1851 shader_init(shader, device, parent, parent_ops);
1852 hr = shader_set_function(shader, byte_code, output_signature, device->d3d_vshader_constantF,
1853 WINED3D_SHADER_TYPE_VERTEX, max_version);
1854 if (FAILED(hr))
1855 {
1856 WARN("Failed to set function, hr %#x.\n", hr);
1857 shader_cleanup(shader);
1858 return hr;
1859 }
1860
1861 map = reg_maps->input_registers;
1862 for (i = 0; map; map >>= 1, ++i)
1863 {
1864 if (!(map & 1) || !shader->input_signature[i].semantic_name)
1865 continue;
1866
1867 shader->u.vs.attributes[i].usage =
1868 shader_usage_from_semantic_name(shader->input_signature[i].semantic_name);
1869 shader->u.vs.attributes[i].usage_idx = shader->input_signature[i].semantic_idx;
1870 }
1871
1872 if (output_signature)
1873 {
1874 for (i = 0; i < output_signature->element_count; ++i)
1875 {
1876 struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1877 reg_maps->output_registers |= 1 << e->register_idx;
1878 shader->output_signature[e->register_idx] = *e;
1879 }
1880 }
1881
1882 vertexshader_set_limits(shader);
1883
1884 shader->load_local_constsF = reg_maps->usesrelconstF
1885 && !list_empty(&shader->constantsF);
1886
1887 return WINED3D_OK;
1888 }
1889
1890 static HRESULT geometryshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
1891 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1892 void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
1893 {
1894 HRESULT hr;
1895
1896 shader_init(shader, device, parent, parent_ops);
1897 hr = shader_set_function(shader, byte_code, output_signature, 0,
1898 WINED3D_SHADER_TYPE_GEOMETRY, max_version);
1899 if (FAILED(hr))
1900 {
1901 WARN("Failed to set function, hr %#x.\n", hr);
1902 shader_cleanup(shader);
1903 return hr;
1904 }
1905
1906 shader->load_local_constsF = FALSE;
1907
1908 return WINED3D_OK;
1909 }
1910
1911 void find_ps_compile_args(const struct wined3d_state *state,
1912 const struct wined3d_shader *shader, struct ps_compile_args *args)
1913 {
1914 struct wined3d_device *device = shader->device;
1915 const struct wined3d_texture *texture;
1916 UINT i;
1917
1918 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
1919 if (state->render_states[WINED3DRS_SRGBWRITEENABLE])
1920 {
1921 const struct wined3d_surface *rt = state->fb->render_targets[0];
1922 if (rt->resource.format->flags & WINED3DFMT_FLAG_SRGB_WRITE) args->srgb_correction = 1;
1923 }
1924
1925 if (shader->reg_maps.shader_version.major == 1
1926 && shader->reg_maps.shader_version.minor <= 3)
1927 {
1928 for (i = 0; i < 4; ++i)
1929 {
1930 DWORD flags = state->texture_states[i][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1931 DWORD tex_transform = flags & ~WINED3DTTFF_PROJECTED;
1932 if (flags & WINED3DTTFF_PROJECTED)
1933 tex_transform |= WINED3D_PSARGS_PROJECTED;
1934 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
1935 }
1936 }
1937
1938 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
1939 {
1940 if (!shader->reg_maps.sampler_type[i])
1941 continue;
1942
1943 texture = state->textures[i];
1944 if (!texture)
1945 {
1946 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
1947 continue;
1948 }
1949 args->color_fixup[i] = texture->resource.format->color_fixup;
1950
1951 if (texture->resource.format->flags & WINED3DFMT_FLAG_SHADOW)
1952 args->shadow |= 1 << i;
1953
1954 /* Flag samplers that need NP2 texcoord fixup. */
1955 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
1956 args->np2_fixup |= (1 << i);
1957 }
1958 if (shader->reg_maps.shader_version.major >= 3)
1959 {
1960 if (device->strided_streams.position_transformed)
1961 {
1962 args->vp_mode = pretransformed;
1963 }
1964 else if (use_vs(state))
1965 {
1966 args->vp_mode = vertexshader;
1967 }
1968 else
1969 {
1970 args->vp_mode = fixedfunction;
1971 }
1972 args->fog = FOG_OFF;
1973 }
1974 else
1975 {
1976 args->vp_mode = vertexshader;
1977 if (state->render_states[WINED3DRS_FOGENABLE])
1978 {
1979 switch (state->render_states[WINED3DRS_FOGTABLEMODE])
1980 {
1981 case WINED3DFOG_NONE:
1982 if (device->strided_streams.position_transformed || use_vs(state))
1983 {
1984 args->fog = FOG_LINEAR;
1985 break;
1986 }
1987
1988 switch (state->render_states[WINED3DRS_FOGVERTEXMODE])
1989 {
1990 case WINED3DFOG_NONE: /* Fall through. */
1991 case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
1992 case WINED3DFOG_EXP: args->fog = FOG_EXP; break;
1993 case WINED3DFOG_EXP2: args->fog = FOG_EXP2; break;
1994 }
1995 break;
1996
1997 case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
1998 case WINED3DFOG_EXP: args->fog = FOG_EXP; break;
1999 case WINED3DFOG_EXP2: args->fog = FOG_EXP2; break;
2000 }
2001 }
2002 else
2003 {
2004 args->fog = FOG_OFF;
2005 }
2006 }
2007 }
2008
2009 static void pixelshader_set_limits(struct wined3d_shader *shader)
2010 {
2011 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
2012 shader->reg_maps.shader_version.minor);
2013
2014 shader->limits.attributes = 0;
2015 shader->limits.address = 0;
2016 shader->limits.packed_output = 0;
2017
2018 switch (shader_version)
2019 {
2020 case WINED3D_SHADER_VERSION(1, 0):
2021 case WINED3D_SHADER_VERSION(1, 1):
2022 case WINED3D_SHADER_VERSION(1, 2):
2023 case WINED3D_SHADER_VERSION(1, 3):
2024 shader->limits.temporary = 2;
2025 shader->limits.constant_float = 8;
2026 shader->limits.constant_int = 0;
2027 shader->limits.constant_bool = 0;
2028 shader->limits.texcoord = 4;
2029 shader->limits.sampler = 4;
2030 shader->limits.packed_input = 0;
2031 shader->limits.label = 0;
2032 break;
2033
2034 case WINED3D_SHADER_VERSION(1, 4):
2035 shader->limits.temporary = 6;
2036 shader->limits.constant_float = 8;
2037 shader->limits.constant_int = 0;
2038 shader->limits.constant_bool = 0;
2039 shader->limits.texcoord = 6;
2040 shader->limits.sampler = 6;
2041 shader->limits.packed_input = 0;
2042 shader->limits.label = 0;
2043 break;
2044
2045 /* FIXME: Temporaries must match D3DPSHADERCAPS2_0.NumTemps. */
2046 case WINED3D_SHADER_VERSION(2, 0):
2047 shader->limits.temporary = 32;
2048 shader->limits.constant_float = 32;
2049 shader->limits.constant_int = 16;
2050 shader->limits.constant_bool = 16;
2051 shader->limits.texcoord = 8;
2052 shader->limits.sampler = 16;
2053 shader->limits.packed_input = 0;
2054 break;
2055
2056 case WINED3D_SHADER_VERSION(2, 1):
2057 shader->limits.temporary = 32;
2058 shader->limits.constant_float = 32;
2059 shader->limits.constant_int = 16;
2060 shader->limits.constant_bool = 16;
2061 shader->limits.texcoord = 8;
2062 shader->limits.sampler = 16;
2063 shader->limits.packed_input = 0;
2064 shader->limits.label = 16;
2065 break;
2066
2067 case WINED3D_SHADER_VERSION(4, 0):
2068 FIXME("Using 3.0 limits for 4.0 shader.\n");
2069 /* Fall through. */
2070
2071 case WINED3D_SHADER_VERSION(3, 0):
2072 shader->limits.temporary = 32;
2073 shader->limits.constant_float = 224;
2074 shader->limits.constant_int = 16;
2075 shader->limits.constant_bool = 16;
2076 shader->limits.texcoord = 0;
2077 shader->limits.sampler = 16;
2078 shader->limits.packed_input = 12;
2079 shader->limits.label = 16; /* FIXME: 2048 */
2080 break;
2081
2082 default:
2083 shader->limits.temporary = 32;
2084 shader->limits.constant_float = 32;
2085 shader->limits.constant_int = 16;
2086 shader->limits.constant_bool = 16;
2087 shader->limits.texcoord = 8;
2088 shader->limits.sampler = 16;
2089 shader->limits.packed_input = 0;
2090 shader->limits.label = 0;
2091 FIXME("Unrecognized pixel shader version %u.%u\n",
2092 shader->reg_maps.shader_version.major,
2093 shader->reg_maps.shader_version.minor);
2094 }
2095 }
2096
2097 static HRESULT pixelshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2098 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2099 void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
2100 {
2101 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2102 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2103 HRESULT hr;
2104
2105 if (!byte_code) return WINED3DERR_INVALIDCALL;
2106
2107 shader_init(shader, device, parent, parent_ops);
2108 hr = shader_set_function(shader, byte_code, output_signature, device->d3d_pshader_constantF,
2109 WINED3D_SHADER_TYPE_PIXEL, max_version);
2110 if (FAILED(hr))
2111 {
2112 WARN("Failed to set function, hr %#x.\n", hr);
2113 shader_cleanup(shader);
2114 return hr;
2115 }
2116
2117 pixelshader_set_limits(shader);
2118
2119 for (i = 0; i < MAX_REG_INPUT; ++i)
2120 {
2121 if (shader->u.ps.input_reg_used[i])
2122 {
2123 ++num_regs_used;
2124 highest_reg_used = i;
2125 }
2126 }
2127
2128 /* Don't do any register mapping magic if it is not needed, or if we can't
2129 * achieve anything anyway */
2130 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2131 || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2132 {
2133 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2134 {
2135 /* This happens with relative addressing. The input mapper function
2136 * warns about this if the higher registers are declared too, so
2137 * don't write a FIXME here */
2138 WARN("More varying registers used than supported\n");
2139 }
2140
2141 for (i = 0; i < MAX_REG_INPUT; ++i)
2142 {
2143 shader->u.ps.input_reg_map[i] = i;
2144 }
2145
2146 shader->u.ps.declared_in_count = highest_reg_used + 1;
2147 }
2148 else
2149 {
2150 shader->u.ps.declared_in_count = 0;
2151 for (i = 0; i < MAX_REG_INPUT; ++i)
2152 {
2153 if (shader->u.ps.input_reg_used[i])
2154 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
2155 else shader->u.ps.input_reg_map[i] = ~0U;
2156 }
2157 }
2158
2159 shader->load_local_constsF = FALSE;
2160
2161 return WINED3D_OK;
2162 }
2163
2164 void pixelshader_update_samplers(struct wined3d_shader_reg_maps *reg_maps, struct wined3d_texture * const *textures)
2165 {
2166 enum wined3d_sampler_texture_type *sampler_type = reg_maps->sampler_type;
2167 unsigned int i;
2168
2169 if (reg_maps->shader_version.major != 1) return;
2170
2171 for (i = 0; i < max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS); ++i)
2172 {
2173 /* We don't sample from this sampler. */
2174 if (!sampler_type[i]) continue;
2175
2176 if (!textures[i])
2177 {
2178 WARN("No texture bound to sampler %u, using 2D.\n", i);
2179 sampler_type[i] = WINED3DSTT_2D;
2180 continue;
2181 }
2182
2183 switch (textures[i]->target)
2184 {
2185 case GL_TEXTURE_RECTANGLE_ARB:
2186 case GL_TEXTURE_2D:
2187 /* We have to select between texture rectangles and 2D
2188 * textures later because 2.0 and 3.0 shaders only have
2189 * WINED3DSTT_2D as well. */
2190 sampler_type[i] = WINED3DSTT_2D;
2191 break;
2192
2193 case GL_TEXTURE_3D:
2194 sampler_type[i] = WINED3DSTT_VOLUME;
2195 break;
2196
2197 case GL_TEXTURE_CUBE_MAP_ARB:
2198 sampler_type[i] = WINED3DSTT_CUBE;
2199 break;
2200
2201 default:
2202 FIXME("Unrecognized texture type %#x, using 2D.\n", textures[i]->target);
2203 sampler_type[i] = WINED3DSTT_2D;
2204 }
2205 }
2206 }
2207
2208 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const DWORD *byte_code,
2209 const struct wined3d_shader_signature *output_signature, void *parent,
2210 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2211 {
2212 struct wined3d_shader *object;
2213 HRESULT hr;
2214
2215 TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2216 device, byte_code, output_signature, parent, parent_ops, shader);
2217
2218 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2219 if (!object)
2220 {
2221 ERR("Failed to allocate shader memory.\n");
2222 return E_OUTOFMEMORY;
2223 }
2224
2225 hr = geometryshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2226 if (FAILED(hr))
2227 {
2228 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
2229 HeapFree(GetProcessHeap(), 0, object);
2230 return hr;
2231 }
2232
2233 TRACE("Created geometry shader %p.\n", object);
2234 *shader = object;
2235
2236 return WINED3D_OK;
2237 }
2238
2239 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const DWORD *byte_code,
2240 const struct wined3d_shader_signature *output_signature, void *parent,
2241 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2242 {
2243 struct wined3d_shader *object;
2244 HRESULT hr;
2245
2246 TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2247 device, byte_code, output_signature, parent, parent_ops, shader);
2248
2249 if (device->ps_selected_mode == SHADER_NONE)
2250 return WINED3DERR_INVALIDCALL;
2251
2252 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2253 if (!object)
2254 {
2255 ERR("Failed to allocate shader memory.\n");
2256 return E_OUTOFMEMORY;
2257 }
2258
2259 hr = pixelshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2260 if (FAILED(hr))
2261 {
2262 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
2263 HeapFree(GetProcessHeap(), 0, object);
2264 return hr;
2265 }
2266
2267 TRACE("Created pixel shader %p.\n", object);
2268 *shader = object;
2269
2270 return WINED3D_OK;
2271 }
2272
2273 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const DWORD *byte_code,
2274 const struct wined3d_shader_signature *output_signature, void *parent,
2275 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2276 {
2277 struct wined3d_shader *object;
2278 HRESULT hr;
2279
2280 TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2281 device, byte_code, output_signature, parent, parent_ops, shader);
2282
2283 if (device->vs_selected_mode == SHADER_NONE)
2284 return WINED3DERR_INVALIDCALL;
2285
2286 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2287 if (!object)
2288 {
2289 ERR("Failed to allocate shader memory.\n");
2290 return E_OUTOFMEMORY;
2291 }
2292
2293 hr = vertexshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2294 if (FAILED(hr))
2295 {
2296 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
2297 HeapFree(GetProcessHeap(), 0, object);
2298 return hr;
2299 }
2300
2301 TRACE("Created vertex shader %p.\n", object);
2302 *shader = object;
2303
2304 return WINED3D_OK;
2305 }
Maarten's multimedia branch