projects / reactos.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
* Slap *some* sense into our header inclusions.
[reactos.git] / reactos / dll / directx / wine / wined3d / shader.c
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 #include <wine/port.h>
27
28 //#include <math.h>
29 #include <stdio.h>
30 //#include <string.h>
31
32 #include "wined3d_private.h"
33
34 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
35 WINE_DECLARE_DEBUG_CHANNEL(d3d);
36
37 static const char * const shader_opcode_names[] =
38 {
39 /* WINED3DSIH_ABS */ "abs",
40 /* WINED3DSIH_ADD */ "add",
41 /* WINED3DSIH_AND */ "and",
42 /* WINED3DSIH_BEM */ "bem",
43 /* WINED3DSIH_BREAK */ "break",
44 /* WINED3DSIH_BREAKC */ "breakc",
45 /* WINED3DSIH_BREAKP */ "breakp",
46 /* WINED3DSIH_CALL */ "call",
47 /* WINED3DSIH_CALLNZ */ "callnz",
48 /* WINED3DSIH_CMP */ "cmp",
49 /* WINED3DSIH_CND */ "cnd",
50 /* WINED3DSIH_CRS */ "crs",
51 /* WINED3DSIH_CUT */ "cut",
52 /* WINED3DSIH_DCL */ "dcl",
53 /* WINED3DSIH_DEF */ "def",
54 /* WINED3DSIH_DEFB */ "defb",
55 /* WINED3DSIH_DEFI */ "defi",
56 /* WINED3DSIH_DIV */ "div",
57 /* WINED3DSIH_DP2ADD */ "dp2add",
58 /* WINED3DSIH_DP3 */ "dp3",
59 /* WINED3DSIH_DP4 */ "dp4",
60 /* WINED3DSIH_DST */ "dst",
61 /* WINED3DSIH_DSX */ "dsx",
62 /* WINED3DSIH_DSY */ "dsy",
63 /* WINED3DSIH_ELSE */ "else",
64 /* WINED3DSIH_EMIT */ "emit",
65 /* WINED3DSIH_ENDIF */ "endif",
66 /* WINED3DSIH_ENDLOOP */ "endloop",
67 /* WINED3DSIH_ENDREP */ "endrep",
68 /* WINED3DSIH_EQ */ "eq",
69 /* WINED3DSIH_EXP */ "exp",
70 /* WINED3DSIH_EXPP */ "expp",
71 /* WINED3DSIH_FRC */ "frc",
72 /* WINED3DSIH_FTOI */ "ftoi",
73 /* WINED3DSIH_GE */ "ge",
74 /* WINED3DSIH_IADD */ "iadd",
75 /* WINED3DSIH_IEQ */ "ieq",
76 /* WINED3DSIH_IF */ "if",
77 /* WINED3DSIH_IFC */ "ifc",
78 /* WINED3DSIH_IGE */ "ige",
79 /* WINED3DSIH_IMUL */ "imul",
80 /* WINED3DSIH_ITOF */ "itof",
81 /* WINED3DSIH_LABEL */ "label",
82 /* WINED3DSIH_LD */ "ld",
83 /* WINED3DSIH_LIT */ "lit",
84 /* WINED3DSIH_LOG */ "log",
85 /* WINED3DSIH_LOGP */ "logp",
86 /* WINED3DSIH_LOOP */ "loop",
87 /* WINED3DSIH_LRP */ "lrp",
88 /* WINED3DSIH_LT */ "lt",
89 /* WINED3DSIH_M3x2 */ "m3x2",
90 /* WINED3DSIH_M3x3 */ "m3x3",
91 /* WINED3DSIH_M3x4 */ "m3x4",
92 /* WINED3DSIH_M4x3 */ "m4x3",
93 /* WINED3DSIH_M4x4 */ "m4x4",
94 /* WINED3DSIH_MAD */ "mad",
95 /* WINED3DSIH_MAX */ "max",
96 /* WINED3DSIH_MIN */ "min",
97 /* WINED3DSIH_MOV */ "mov",
98 /* WINED3DSIH_MOVA */ "mova",
99 /* WINED3DSIH_MOVC */ "movc",
100 /* WINED3DSIH_MUL */ "mul",
101 /* WINED3DSIH_NOP */ "nop",
102 /* WINED3DSIH_NRM */ "nrm",
103 /* WINED3DSIH_PHASE */ "phase",
104 /* WINED3DSIH_POW */ "pow",
105 /* WINED3DSIH_RCP */ "rcp",
106 /* WINED3DSIH_REP */ "rep",
107 /* WINED3DSIH_RET */ "ret",
108 /* WINED3DSIH_ROUND_NI */ "round_ni",
109 /* WINED3DSIH_RSQ */ "rsq",
110 /* WINED3DSIH_SAMPLE */ "sample",
111 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
112 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
113 /* WINED3DSIH_SETP */ "setp",
114 /* WINED3DSIH_SGE */ "sge",
115 /* WINED3DSIH_SGN */ "sgn",
116 /* WINED3DSIH_SINCOS */ "sincos",
117 /* WINED3DSIH_SLT */ "slt",
118 /* WINED3DSIH_SQRT */ "sqrt",
119 /* WINED3DSIH_SUB */ "sub",
120 /* WINED3DSIH_TEX */ "texld",
121 /* WINED3DSIH_TEXBEM */ "texbem",
122 /* WINED3DSIH_TEXBEML */ "texbeml",
123 /* WINED3DSIH_TEXCOORD */ "texcrd",
124 /* WINED3DSIH_TEXDEPTH */ "texdepth",
125 /* WINED3DSIH_TEXDP3 */ "texdp3",
126 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
127 /* WINED3DSIH_TEXKILL */ "texkill",
128 /* WINED3DSIH_TEXLDD */ "texldd",
129 /* WINED3DSIH_TEXLDL */ "texldl",
130 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
131 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
132 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
133 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
134 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
135 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
136 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
137 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
138 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
139 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
140 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
141 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
142 /* WINED3DSIH_UDIV */ "udiv",
143 /* WINED3DSIH_USHR */ "ushr",
144 /* WINED3DSIH_UTOF */ "utof",
145 /* WINED3DSIH_XOR */ "xor",
146 };
147
148 static const char * const semantic_names[] =
149 {
150 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
151 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
152 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
153 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
154 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
155 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
156 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
157 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
158 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
159 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
160 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
161 /* WINED3D_DECL_USAGE_FOG */ "FOG",
162 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
163 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
164 };
165
166 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
167 {
168 if (usage >= sizeof(semantic_names) / sizeof(*semantic_names))
169 {
170 FIXME("Unrecognized usage %#x.\n", usage);
171 return "UNRECOGNIZED";
172 }
173
174 return semantic_names[usage];
175 }
176
177 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
178 {
179 unsigned int i;
180
181 for (i = 0; i < sizeof(semantic_names) / sizeof(*semantic_names); ++i)
182 {
183 if (!strcmp(name, semantic_names[i])) return i;
184 }
185
186 return ~0U;
187 }
188
189 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
190 {
191 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
192 }
193
194 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
195 const struct wined3d_shader_semantic *s)
196 {
197 e->semantic_name = shader_semantic_name_from_usage(s->usage);
198 e->semantic_idx = s->usage_idx;
199 e->sysval_semantic = 0;
200 e->component_type = 0;
201 e->register_idx = s->reg.reg.idx;
202 e->mask = s->reg.write_mask;
203 }
204
205 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
206 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
207 {
208 e->semantic_name = shader_semantic_name_from_usage(usage);
209 e->semantic_idx = usage_idx;
210 e->sysval_semantic = 0;
211 e->component_type = 0;
212 e->register_idx = reg_idx;
213 e->mask = write_mask;
214 }
215
216 static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
217 {
218 switch (version_token >> 16)
219 {
220 case WINED3D_SM1_VS:
221 case WINED3D_SM1_PS:
222 return &sm1_shader_frontend;
223
224 case WINED3D_SM4_PS:
225 case WINED3D_SM4_VS:
226 case WINED3D_SM4_GS:
227 return &sm4_shader_frontend;
228
229 default:
230 FIXME("Unrecognised version token %#x\n", version_token);
231 return NULL;
232 }
233 }
234
235 void shader_buffer_clear(struct wined3d_shader_buffer *buffer)
236 {
237 buffer->buffer[0] = '\0';
238 buffer->bsize = 0;
239 buffer->lineNo = 0;
240 buffer->newline = TRUE;
241 }
242
243 BOOL shader_buffer_init(struct wined3d_shader_buffer *buffer)
244 {
245 buffer->buffer = HeapAlloc(GetProcessHeap(), 0, SHADER_PGMSIZE);
246 if (!buffer->buffer)
247 {
248 ERR("Failed to allocate shader buffer memory.\n");
249 return FALSE;
250 }
251
252 shader_buffer_clear(buffer);
253 return TRUE;
254 }
255
256 void shader_buffer_free(struct wined3d_shader_buffer *buffer)
257 {
258 HeapFree(GetProcessHeap(), 0, buffer->buffer);
259 }
260
261 int shader_vaddline(struct wined3d_shader_buffer *buffer, const char *format, va_list args)
262 {
263 char *base = buffer->buffer + buffer->bsize;
264 int rc;
265
266 rc = vsnprintf(base, SHADER_PGMSIZE - 1 - buffer->bsize, format, args);
267
268 if (rc < 0 /* C89 */ || (unsigned int)rc > SHADER_PGMSIZE - 1 - buffer->bsize /* C99 */)
269 {
270 ERR("The buffer allocated for the shader program string "
271 "is too small at %d bytes.\n", SHADER_PGMSIZE);
272 buffer->bsize = SHADER_PGMSIZE - 1;
273 return -1;
274 }
275
276 if (buffer->newline)
277 {
278 TRACE("GL HW (%u, %u) : %s", buffer->lineNo + 1, buffer->bsize, base);
279 buffer->newline = FALSE;
280 }
281 else
282 {
283 TRACE("%s", base);
284 }
285
286 buffer->bsize += rc;
287 if (buffer->buffer[buffer->bsize-1] == '\n')
288 {
289 ++buffer->lineNo;
290 buffer->newline = TRUE;
291 }
292
293 return 0;
294 }
295
296 int shader_addline(struct wined3d_shader_buffer *buffer, const char *format, ...)
297 {
298 va_list args;
299 int ret;
300
301 va_start(args, format);
302 ret = shader_vaddline(buffer, format, args);
303 va_end(args);
304
305 return ret;
306 }
307
308 static void shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
309 void *parent, const struct wined3d_parent_ops *parent_ops)
310 {
311 shader->ref = 1;
312 shader->device = device;
313 shader->parent = parent;
314 shader->parent_ops = parent_ops;
315 list_init(&shader->linked_programs);
316 list_add_head(&device->shaders, &shader->shader_list_entry);
317 }
318
319 /* Convert floating point offset relative to a register file to an absolute
320 * offset for float constants. */
321 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
322 {
323 switch (register_type)
324 {
325 case WINED3DSPR_CONST: return register_idx;
326 case WINED3DSPR_CONST2: return 2048 + register_idx;
327 case WINED3DSPR_CONST3: return 4096 + register_idx;
328 case WINED3DSPR_CONST4: return 6144 + register_idx;
329 default:
330 FIXME("Unsupported register type: %u.\n", register_type);
331 return register_idx;
332 }
333 }
334
335 static void shader_delete_constant_list(struct list *clist)
336 {
337 struct wined3d_shader_lconst *constant;
338 struct list *ptr;
339
340 ptr = list_head(clist);
341 while (ptr)
342 {
343 constant = LIST_ENTRY(ptr, struct wined3d_shader_lconst, entry);
344 ptr = list_next(clist, ptr);
345 HeapFree(GetProcessHeap(), 0, constant);
346 }
347 list_init(clist);
348 }
349
350 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
351 {
352 DWORD idx, shift;
353 idx = bit >> 5;
354 shift = bit & 0x1f;
355 bitmap[idx] |= (1 << shift);
356 }
357
358 static void shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
359 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type)
360 {
361 switch (reg->type)
362 {
363 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
364 if (shader_type == WINED3D_SHADER_TYPE_PIXEL) reg_maps->texcoord |= 1 << reg->idx;
365 else reg_maps->address |= 1 << reg->idx;
366 break;
367
368 case WINED3DSPR_TEMP:
369 reg_maps->temporary |= 1 << reg->idx;
370 break;
371
372 case WINED3DSPR_INPUT:
373 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
374 {
375 if (reg->rel_addr)
376 {
377 /* If relative addressing is used, we must assume that all registers
378 * are used. Even if it is a construct like v3[aL], we can't assume
379 * that v0, v1 and v2 aren't read because aL can be negative */
380 unsigned int i;
381 for (i = 0; i < MAX_REG_INPUT; ++i)
382 {
383 shader->u.ps.input_reg_used[i] = TRUE;
384 }
385 }
386 else
387 {
388 shader->u.ps.input_reg_used[reg->idx] = TRUE;
389 }
390 }
391 else reg_maps->input_registers |= 1 << reg->idx;
392 break;
393
394 case WINED3DSPR_RASTOUT:
395 if (reg->idx == 1) reg_maps->fog = 1;
396 break;
397
398 case WINED3DSPR_MISCTYPE:
399 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
400 {
401 if (!reg->idx) reg_maps->vpos = 1;
402 else if (reg->idx == 1) reg_maps->usesfacing = 1;
403 }
404 break;
405
406 case WINED3DSPR_CONST:
407 if (reg->rel_addr)
408 {
409 if (reg->idx < reg_maps->min_rel_offset) reg_maps->min_rel_offset = reg->idx;
410 if (reg->idx > reg_maps->max_rel_offset) reg_maps->max_rel_offset = reg->idx;
411 reg_maps->usesrelconstF = TRUE;
412 }
413 else
414 {
415 set_bitmap_bit(reg_maps->constf, reg->idx);
416 }
417 break;
418
419 case WINED3DSPR_CONSTINT:
420 reg_maps->integer_constants |= (1 << reg->idx);
421 break;
422
423 case WINED3DSPR_CONSTBOOL:
424 reg_maps->boolean_constants |= (1 << reg->idx);
425 break;
426
427 case WINED3DSPR_COLOROUT:
428 reg_maps->rt_mask |= (1 << reg->idx);
429 break;
430
431 default:
432 TRACE("Not recording register of type %#x and idx %u\n", reg->type, reg->idx);
433 break;
434 }
435 }
436
437 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
438 {
439 switch (instr)
440 {
441 case WINED3DSIH_M4x4:
442 case WINED3DSIH_M3x4:
443 return param == 1 ? 3 : 0;
444
445 case WINED3DSIH_M4x3:
446 case WINED3DSIH_M3x3:
447 return param == 1 ? 2 : 0;
448
449 case WINED3DSIH_M3x2:
450 return param == 1 ? 1 : 0;
451
452 default:
453 return 0;
454 }
455 }
456
457 /* Note that this does not count the loop register as an address register. */
458 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
459 struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature_element *input_signature,
460 struct wined3d_shader_signature_element *output_signature, const DWORD *byte_code, DWORD constf_size)
461 {
462 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
463 void *fe_data = shader->frontend_data;
464 struct wined3d_shader_version shader_version;
465 const DWORD *ptr = byte_code;
466
467 memset(reg_maps, 0, sizeof(*reg_maps));
468 reg_maps->min_rel_offset = ~0U;
469
470 fe->shader_read_header(fe_data, &ptr, &shader_version);
471 reg_maps->shader_version = shader_version;
472
473 reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
474 sizeof(*reg_maps->constf) * ((constf_size + 31) / 32));
475 if (!reg_maps->constf)
476 {
477 ERR("Failed to allocate constant map memory.\n");
478 return E_OUTOFMEMORY;
479 }
480
481 while (!fe->shader_is_end(fe_data, &ptr))
482 {
483 struct wined3d_shader_instruction ins;
484 const char *comment;
485 UINT comment_size;
486 UINT param_size;
487
488 /* Skip comments. */
489 fe->shader_read_comment(&ptr, &comment, &comment_size);
490 if (comment) continue;
491
492 /* Fetch opcode. */
493 fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
494
495 /* Unhandled opcode, and its parameters. */
496 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
497 {
498 TRACE("Skipping unrecognized instruction.\n");
499 ptr += param_size;
500 continue;
501 }
502
503 /* Handle declarations. */
504 if (ins.handler_idx == WINED3DSIH_DCL)
505 {
506 struct wined3d_shader_semantic semantic;
507
508 fe->shader_read_semantic(&ptr, &semantic);
509
510 switch (semantic.reg.reg.type)
511 {
512 /* Mark input registers used. */
513 case WINED3DSPR_INPUT:
514 reg_maps->input_registers |= 1 << semantic.reg.reg.idx;
515 shader_signature_from_semantic(&input_signature[semantic.reg.reg.idx], &semantic);
516 break;
517
518 /* Vertex shader: mark 3.0 output registers used, save token. */
519 case WINED3DSPR_OUTPUT:
520 reg_maps->output_registers |= 1 << semantic.reg.reg.idx;
521 shader_signature_from_semantic(&output_signature[semantic.reg.reg.idx], &semantic);
522 if (semantic.usage == WINED3D_DECL_USAGE_FOG)
523 reg_maps->fog = 1;
524 break;
525
526 /* Save sampler usage token. */
527 case WINED3DSPR_SAMPLER:
528 reg_maps->sampler_type[semantic.reg.reg.idx] = semantic.sampler_type;
529 break;
530
531 default:
532 TRACE("Not recording DCL register type %#x.\n", semantic.reg.reg.type);
533 break;
534 }
535 }
536 else if (ins.handler_idx == WINED3DSIH_DEF)
537 {
538 struct wined3d_shader_src_param rel_addr;
539 struct wined3d_shader_dst_param dst;
540
541 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
542 if (!lconst) return E_OUTOFMEMORY;
543
544 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
545 lconst->idx = dst.reg.idx;
546
547 memcpy(lconst->value, ptr, 4 * sizeof(DWORD));
548 ptr += 4;
549
550 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
551 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
552 {
553 float *value = (float *)lconst->value;
554 if (value[0] < -1.0f) value[0] = -1.0f;
555 else if (value[0] > 1.0f) value[0] = 1.0f;
556 if (value[1] < -1.0f) value[1] = -1.0f;
557 else if (value[1] > 1.0f) value[1] = 1.0f;
558 if (value[2] < -1.0f) value[2] = -1.0f;
559 else if (value[2] > 1.0f) value[2] = 1.0f;
560 if (value[3] < -1.0f) value[3] = -1.0f;
561 else if (value[3] > 1.0f) value[3] = 1.0f;
562 }
563
564 list_add_head(&shader->constantsF, &lconst->entry);
565 }
566 else if (ins.handler_idx == WINED3DSIH_DEFI)
567 {
568 struct wined3d_shader_src_param rel_addr;
569 struct wined3d_shader_dst_param dst;
570
571 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
572 if (!lconst) return E_OUTOFMEMORY;
573
574 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
575 lconst->idx = dst.reg.idx;
576
577 memcpy(lconst->value, ptr, 4 * sizeof(DWORD));
578 ptr += 4;
579
580 list_add_head(&shader->constantsI, &lconst->entry);
581 reg_maps->local_int_consts |= (1 << dst.reg.idx);
582 }
583 else if (ins.handler_idx == WINED3DSIH_DEFB)
584 {
585 struct wined3d_shader_src_param rel_addr;
586 struct wined3d_shader_dst_param dst;
587
588 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
589 if (!lconst) return E_OUTOFMEMORY;
590
591 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
592 lconst->idx = dst.reg.idx;
593
594 memcpy(lconst->value, ptr, sizeof(DWORD));
595 ++ptr;
596
597 list_add_head(&shader->constantsB, &lconst->entry);
598 reg_maps->local_bool_consts |= (1 << dst.reg.idx);
599 }
600 /* For subroutine prototypes. */
601 else if (ins.handler_idx == WINED3DSIH_LABEL)
602 {
603 struct wined3d_shader_src_param src, rel_addr;
604
605 fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
606 reg_maps->labels |= 1 << src.reg.idx;
607 }
608 /* Set texture, address, temporary registers. */
609 else
610 {
611 BOOL color0_mov = FALSE;
612 unsigned int i, limit;
613
614 /* This will loop over all the registers and try to
615 * make a bitmask of the ones we're interested in.
616 *
617 * Relative addressing tokens are ignored, but that's
618 * okay, since we'll catch any address registers when
619 * they are initialized (required by spec). */
620 for (i = 0; i < ins.dst_count; ++i)
621 {
622 struct wined3d_shader_src_param dst_rel_addr;
623 struct wined3d_shader_dst_param dst_param;
624
625 fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
626
627 shader_record_register_usage(shader, reg_maps, &dst_param.reg, shader_version.type);
628
629 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and
630 * is used in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel
631 * shaders because TECRDOUT isn't used in them, but future register types might cause issues */
632 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX && shader_version.major < 3)
633 {
634 UINT idx = dst_param.reg.idx;
635
636 switch (dst_param.reg.type)
637 {
638 case WINED3DSPR_RASTOUT:
639 switch (idx)
640 {
641 case 0: /* oPos */
642 reg_maps->output_registers |= 1 << 10;
643 shader_signature_from_usage(&output_signature[10],
644 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
645 break;
646
647 case 1: /* oFog */
648 reg_maps->output_registers |= 1 << 11;
649 shader_signature_from_usage(&output_signature[11],
650 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
651 break;
652
653 case 2: /* oPts */
654 reg_maps->output_registers |= 1 << 11;
655 shader_signature_from_usage(&output_signature[11],
656 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
657 break;
658 }
659 break;
660
661 case WINED3DSPR_ATTROUT:
662 if (idx < 2)
663 {
664 idx += 8;
665 if (reg_maps->output_registers & (1 << idx))
666 {
667 output_signature[idx].mask |= dst_param.write_mask;
668 }
669 else
670 {
671 reg_maps->output_registers |= 1 << idx;
672 shader_signature_from_usage(&output_signature[idx],
673 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, dst_param.write_mask);
674 }
675 }
676 break;
677
678 case WINED3DSPR_TEXCRDOUT:
679
680 reg_maps->texcoord_mask[idx] |= dst_param.write_mask;
681 if (reg_maps->output_registers & (1 << idx))
682 {
683 output_signature[idx].mask |= dst_param.write_mask;
684 }
685 else
686 {
687 reg_maps->output_registers |= 1 << idx;
688 shader_signature_from_usage(&output_signature[idx],
689 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, dst_param.write_mask);
690 }
691 break;
692
693 default:
694 break;
695 }
696 }
697
698 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
699 {
700 if (dst_param.reg.type == WINED3DSPR_COLOROUT && !dst_param.reg.idx)
701 {
702 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
703 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
704 * the mov and perform the sRGB write correction from the source register.
705 *
706 * However, if the mov is only partial, we can't do this, and if the write
707 * comes from an instruction other than MOV it is hard to do as well. If
708 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
709 shader->u.ps.color0_mov = FALSE;
710 if (ins.handler_idx == WINED3DSIH_MOV
711 && dst_param.write_mask == WINED3DSP_WRITEMASK_ALL)
712 {
713 /* Used later when the source register is read. */
714 color0_mov = TRUE;
715 }
716 }
717 /* Also drop the MOV marker if the source register is overwritten prior to the shader
718 * end
719 */
720 else if (dst_param.reg.type == WINED3DSPR_TEMP
721 && dst_param.reg.idx == shader->u.ps.color0_reg)
722 {
723 shader->u.ps.color0_mov = FALSE;
724 }
725 }
726
727 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
728 if (shader_version.major == 1
729 && (ins.handler_idx == WINED3DSIH_TEX
730 || ins.handler_idx == WINED3DSIH_TEXBEM
731 || ins.handler_idx == WINED3DSIH_TEXBEML
732 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
733 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
734 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
735 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
736 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
737 || ins.handler_idx == WINED3DSIH_TEXREG2AR
738 || ins.handler_idx == WINED3DSIH_TEXREG2GB
739 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
740 {
741 /* Fake sampler usage, only set reserved bit and type. */
742 DWORD sampler_code = dst_param.reg.idx;
743
744 TRACE("Setting fake 2D sampler for 1.x pixelshader.\n");
745 reg_maps->sampler_type[sampler_code] = WINED3DSTT_2D;
746
747 /* texbem is only valid with < 1.4 pixel shaders */
748 if (ins.handler_idx == WINED3DSIH_TEXBEM
749 || ins.handler_idx == WINED3DSIH_TEXBEML)
750 {
751 reg_maps->bumpmat |= 1 << dst_param.reg.idx;
752 if (ins.handler_idx == WINED3DSIH_TEXBEML)
753 {
754 reg_maps->luminanceparams |= 1 << dst_param.reg.idx;
755 }
756 }
757 }
758 else if (ins.handler_idx == WINED3DSIH_BEM)
759 {
760 reg_maps->bumpmat |= 1 << dst_param.reg.idx;
761 }
762 }
763
764 if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
765 else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
766 else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
767 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
768 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
769 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
770 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
771 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
772 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
773 else if (ins.handler_idx == WINED3DSIH_LOOP
774 || ins.handler_idx == WINED3DSIH_REP)
775 {
776 ++cur_loop_depth;
777 if (cur_loop_depth > max_loop_depth)
778 max_loop_depth = cur_loop_depth;
779 }
780 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
781 || ins.handler_idx == WINED3DSIH_ENDREP)
782 --cur_loop_depth;
783
784 limit = ins.src_count + (ins.predicate ? 1 : 0);
785 for (i = 0; i < limit; ++i)
786 {
787 struct wined3d_shader_src_param src_param, src_rel_addr;
788 unsigned int count;
789
790 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
791 count = get_instr_extra_regcount(ins.handler_idx, i);
792
793 shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
794 while (count)
795 {
796 ++src_param.reg.idx;
797 shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
798 --count;
799 }
800
801 if (color0_mov)
802 {
803 if (src_param.reg.type == WINED3DSPR_TEMP
804 && src_param.swizzle == WINED3DSP_NOSWIZZLE)
805 {
806 shader->u.ps.color0_mov = TRUE;
807 shader->u.ps.color0_reg = src_param.reg.idx;
808 }
809 }
810 }
811 }
812 }
813 reg_maps->loop_depth = max_loop_depth;
814
815 /* PS before 2.0 don't have explicit color outputs. Instead the value of
816 * R0 is written to the render target. */
817 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
818 reg_maps->rt_mask |= (1 << 0);
819
820 shader->functionLength = ((const char *)ptr - (const char *)byte_code);
821
822 return WINED3D_OK;
823 }
824
825 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
826 {
827 DWORD map = 1 << max;
828 map |= map - 1;
829 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
830
831 return wined3d_log2i(map);
832 }
833
834 static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
835 const struct wined3d_shader_version *shader_version)
836 {
837 TRACE("dcl");
838
839 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
840 {
841 switch (semantic->sampler_type)
842 {
843 case WINED3DSTT_2D: TRACE("_2d"); break;
844 case WINED3DSTT_CUBE: TRACE("_cube"); break;
845 case WINED3DSTT_VOLUME: TRACE("_volume"); break;
846 default: TRACE("_unknown_ttype(0x%08x)", semantic->sampler_type);
847 }
848 }
849 else
850 {
851 /* Pixel shaders 3.0 don't have usage semantics. */
852 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL) return;
853 else TRACE("_");
854
855 switch (semantic->usage)
856 {
857 case WINED3D_DECL_USAGE_POSITION:
858 TRACE("position%u", semantic->usage_idx);
859 break;
860
861 case WINED3D_DECL_USAGE_BLEND_INDICES:
862 TRACE("blend");
863 break;
864
865 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
866 TRACE("weight");
867 break;
868
869 case WINED3D_DECL_USAGE_NORMAL:
870 TRACE("normal%u", semantic->usage_idx);
871 break;
872
873 case WINED3D_DECL_USAGE_PSIZE:
874 TRACE("psize");
875 break;
876
877 case WINED3D_DECL_USAGE_COLOR:
878 if (!semantic->usage_idx) TRACE("color");
879 else TRACE("specular%u", (semantic->usage_idx - 1));
880 break;
881
882 case WINED3D_DECL_USAGE_TEXCOORD:
883 TRACE("texture%u", semantic->usage_idx);
884 break;
885
886 case WINED3D_DECL_USAGE_TANGENT:
887 TRACE("tangent");
888 break;
889
890 case WINED3D_DECL_USAGE_BINORMAL:
891 TRACE("binormal");
892 break;
893
894 case WINED3D_DECL_USAGE_TESS_FACTOR:
895 TRACE("tessfactor");
896 break;
897
898 case WINED3D_DECL_USAGE_POSITIONT:
899 TRACE("positionT%u", semantic->usage_idx);
900 break;
901
902 case WINED3D_DECL_USAGE_FOG:
903 TRACE("fog");
904 break;
905
906 case WINED3D_DECL_USAGE_DEPTH:
907 TRACE("depth");
908 break;
909
910 case WINED3D_DECL_USAGE_SAMPLE:
911 TRACE("sample");
912 break;
913
914 default:
915 FIXME("unknown_semantics(0x%08x)", semantic->usage);
916 }
917 }
918 }
919
920 static void shader_dump_register(const struct wined3d_shader_register *reg,
921 const struct wined3d_shader_version *shader_version)
922 {
923 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
924 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
925 UINT offset = reg->idx;
926
927 switch (reg->type)
928 {
929 case WINED3DSPR_TEMP:
930 TRACE("r");
931 break;
932
933 case WINED3DSPR_INPUT:
934 TRACE("v");
935 break;
936
937 case WINED3DSPR_CONST:
938 case WINED3DSPR_CONST2:
939 case WINED3DSPR_CONST3:
940 case WINED3DSPR_CONST4:
941 TRACE("c");
942 offset = shader_get_float_offset(reg->type, reg->idx);
943 break;
944
945 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
946 TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
947 break;
948
949 case WINED3DSPR_RASTOUT:
950 TRACE("%s", rastout_reg_names[reg->idx]);
951 break;
952
953 case WINED3DSPR_COLOROUT:
954 TRACE("oC");
955 break;
956
957 case WINED3DSPR_DEPTHOUT:
958 TRACE("oDepth");
959 break;
960
961 case WINED3DSPR_ATTROUT:
962 TRACE("oD");
963 break;
964
965 case WINED3DSPR_TEXCRDOUT:
966 /* Vertex shaders >= 3.0 use general purpose output registers
967 * (WINED3DSPR_OUTPUT), which can include an address token. */
968 if (shader_version->major >= 3) TRACE("o");
969 else TRACE("oT");
970 break;
971
972 case WINED3DSPR_CONSTINT:
973 TRACE("i");
974 break;
975
976 case WINED3DSPR_CONSTBOOL:
977 TRACE("b");
978 break;
979
980 case WINED3DSPR_LABEL:
981 TRACE("l");
982 break;
983
984 case WINED3DSPR_LOOP:
985 TRACE("aL");
986 break;
987
988 case WINED3DSPR_SAMPLER:
989 TRACE("s");
990 break;
991
992 case WINED3DSPR_MISCTYPE:
993 if (reg->idx > 1) FIXME("Unhandled misctype register %u.\n", reg->idx);
994 else TRACE("%s", misctype_reg_names[reg->idx]);
995 break;
996
997 case WINED3DSPR_PREDICATE:
998 TRACE("p");
999 break;
1000
1001 case WINED3DSPR_IMMCONST:
1002 TRACE("l");
1003 break;
1004
1005 case WINED3DSPR_CONSTBUFFER:
1006 TRACE("cb");
1007 break;
1008
1009 case WINED3DSPR_NULL:
1010 TRACE("null");
1011 break;
1012
1013 case WINED3DSPR_RESOURCE:
1014 TRACE("t");
1015 break;
1016
1017 default:
1018 TRACE("unhandled_rtype(%#x)", reg->type);
1019 break;
1020 }
1021
1022 if (reg->type == WINED3DSPR_IMMCONST)
1023 {
1024 TRACE("(");
1025 switch (reg->immconst_type)
1026 {
1027 case WINED3D_IMMCONST_SCALAR:
1028 TRACE("%.8e", *(const float *)reg->immconst_data);
1029 break;
1030
1031 case WINED3D_IMMCONST_VEC4:
1032 TRACE("%.8e, %.8e, %.8e, %.8e",
1033 *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
1034 *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
1035 break;
1036
1037 default:
1038 TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
1039 break;
1040 }
1041 TRACE(")");
1042 }
1043 else if (reg->type != WINED3DSPR_RASTOUT
1044 && reg->type != WINED3DSPR_MISCTYPE
1045 && reg->type != WINED3DSPR_NULL)
1046 {
1047 if (reg->array_idx != ~0U)
1048 {
1049 TRACE("%u[%u", offset, reg->array_idx);
1050 if (reg->rel_addr)
1051 {
1052 TRACE(" + ");
1053 shader_dump_src_param(reg->rel_addr, shader_version);
1054 }
1055 TRACE("]");
1056 }
1057 else
1058 {
1059 if (reg->rel_addr)
1060 {
1061 TRACE("[");
1062 shader_dump_src_param(reg->rel_addr, shader_version);
1063 TRACE(" + ");
1064 }
1065 TRACE("%u", offset);
1066 if (reg->rel_addr) TRACE("]");
1067 }
1068 }
1069 }
1070
1071 void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
1072 const struct wined3d_shader_version *shader_version)
1073 {
1074 DWORD write_mask = param->write_mask;
1075
1076 shader_dump_register(&param->reg, shader_version);
1077
1078 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
1079 {
1080 static const char *write_mask_chars = "xyzw";
1081
1082 TRACE(".");
1083 if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
1084 if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
1085 if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
1086 if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
1087 }
1088 }
1089
1090 void shader_dump_src_param(const struct wined3d_shader_src_param *param,
1091 const struct wined3d_shader_version *shader_version)
1092 {
1093 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
1094 DWORD swizzle = param->swizzle;
1095
1096 if (src_modifier == WINED3DSPSM_NEG
1097 || src_modifier == WINED3DSPSM_BIASNEG
1098 || src_modifier == WINED3DSPSM_SIGNNEG
1099 || src_modifier == WINED3DSPSM_X2NEG
1100 || src_modifier == WINED3DSPSM_ABSNEG)
1101 TRACE("-");
1102 else if (src_modifier == WINED3DSPSM_COMP)
1103 TRACE("1-");
1104 else if (src_modifier == WINED3DSPSM_NOT)
1105 TRACE("!");
1106
1107 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1108 TRACE("abs(");
1109
1110 shader_dump_register(&param->reg, shader_version);
1111
1112 if (src_modifier)
1113 {
1114 switch (src_modifier)
1115 {
1116 case WINED3DSPSM_NONE: break;
1117 case WINED3DSPSM_NEG: break;
1118 case WINED3DSPSM_NOT: break;
1119 case WINED3DSPSM_BIAS: TRACE("_bias"); break;
1120 case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
1121 case WINED3DSPSM_SIGN: TRACE("_bx2"); break;
1122 case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1123 case WINED3DSPSM_COMP: break;
1124 case WINED3DSPSM_X2: TRACE("_x2"); break;
1125 case WINED3DSPSM_X2NEG: TRACE("_x2"); break;
1126 case WINED3DSPSM_DZ: TRACE("_dz"); break;
1127 case WINED3DSPSM_DW: TRACE("_dw"); break;
1128 case WINED3DSPSM_ABSNEG: TRACE(")"); break;
1129 case WINED3DSPSM_ABS: TRACE(")"); break;
1130 default: TRACE("_unknown_modifier(%#x)", src_modifier);
1131 }
1132 }
1133
1134 if (swizzle != WINED3DSP_NOSWIZZLE)
1135 {
1136 static const char *swizzle_chars = "xyzw";
1137 DWORD swizzle_x = swizzle & 0x03;
1138 DWORD swizzle_y = (swizzle >> 2) & 0x03;
1139 DWORD swizzle_z = (swizzle >> 4) & 0x03;
1140 DWORD swizzle_w = (swizzle >> 6) & 0x03;
1141
1142 if (swizzle_x == swizzle_y
1143 && swizzle_x == swizzle_z
1144 && swizzle_x == swizzle_w)
1145 {
1146 TRACE(".%c", swizzle_chars[swizzle_x]);
1147 }
1148 else
1149 {
1150 TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1151 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1152 }
1153 }
1154 }
1155
1156 /* Shared code in order to generate the bulk of the shader string.
1157 * NOTE: A description of how to parse tokens can be found on MSDN. */
1158 void shader_generate_main(const struct wined3d_shader *shader, struct wined3d_shader_buffer *buffer,
1159 const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1160 {
1161 struct wined3d_device *device = shader->device;
1162 const struct wined3d_shader_frontend *fe = shader->frontend;
1163 void *fe_data = shader->frontend_data;
1164 struct wined3d_shader_src_param dst_rel_addr[2];
1165 struct wined3d_shader_src_param src_rel_addr[4];
1166 struct wined3d_shader_dst_param dst_param[2];
1167 struct wined3d_shader_src_param src_param[4];
1168 struct wined3d_shader_version shader_version;
1169 struct wined3d_shader_loop_state loop_state;
1170 struct wined3d_shader_instruction ins;
1171 struct wined3d_shader_tex_mx tex_mx;
1172 struct wined3d_shader_context ctx;
1173 const DWORD *ptr = byte_code;
1174 DWORD i;
1175
1176 /* Initialize current parsing state. */
1177 tex_mx.current_row = 0;
1178 loop_state.current_depth = 0;
1179 loop_state.current_reg = 0;
1180
1181 ctx.shader = shader;
1182 ctx.gl_info = &device->adapter->gl_info;
1183 ctx.reg_maps = reg_maps;
1184 ctx.buffer = buffer;
1185 ctx.tex_mx = &tex_mx;
1186 ctx.loop_state = &loop_state;
1187 ctx.backend_data = backend_ctx;
1188
1189 ins.ctx = &ctx;
1190 ins.dst = dst_param;
1191 ins.src = src_param;
1192
1193 fe->shader_read_header(fe_data, &ptr, &shader_version);
1194
1195 while (!fe->shader_is_end(fe_data, &ptr))
1196 {
1197 const char *comment;
1198 UINT comment_size;
1199 UINT param_size;
1200
1201 /* Skip comment tokens. */
1202 fe->shader_read_comment(&ptr, &comment, &comment_size);
1203 if (comment) continue;
1204
1205 /* Read opcode. */
1206 fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1207
1208 /* Unknown opcode and its parameters. */
1209 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1210 {
1211 TRACE("Skipping unrecognized instruction.\n");
1212 ptr += param_size;
1213 continue;
1214 }
1215
1216 /* Nothing to do. */
1217 if (ins.handler_idx == WINED3DSIH_DCL
1218 || ins.handler_idx == WINED3DSIH_NOP
1219 || ins.handler_idx == WINED3DSIH_DEF
1220 || ins.handler_idx == WINED3DSIH_DEFI
1221 || ins.handler_idx == WINED3DSIH_DEFB
1222 || ins.handler_idx == WINED3DSIH_PHASE)
1223 {
1224 ptr += param_size;
1225 continue;
1226 }
1227
1228 /* Destination tokens */
1229 for (i = 0; i < ins.dst_count; ++i)
1230 {
1231 fe->shader_read_dst_param(fe_data, &ptr, &dst_param[i], &dst_rel_addr[i]);
1232 }
1233
1234 /* Predication token */
1235 if (ins.predicate)
1236 {
1237 FIXME("Predicates not implemented.\n");
1238 ins.predicate = *ptr++;
1239 }
1240
1241 /* Other source tokens */
1242 for (i = 0; i < ins.src_count; ++i)
1243 {
1244 fe->shader_read_src_param(fe_data, &ptr, &src_param[i], &src_rel_addr[i]);
1245 }
1246
1247 /* Call appropriate function for output target */
1248 device->shader_backend->shader_handle_instruction(&ins);
1249 }
1250 }
1251
1252 static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1253 {
1254 DWORD mmask = dst->modifiers;
1255
1256 switch (dst->shift)
1257 {
1258 case 0: break;
1259 case 13: TRACE("_d8"); break;
1260 case 14: TRACE("_d4"); break;
1261 case 15: TRACE("_d2"); break;
1262 case 1: TRACE("_x2"); break;
1263 case 2: TRACE("_x4"); break;
1264 case 3: TRACE("_x8"); break;
1265 default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1266 }
1267
1268 if (mmask & WINED3DSPDM_SATURATE) TRACE("_sat");
1269 if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1270 if (mmask & WINED3DSPDM_MSAMPCENTROID) TRACE("_centroid");
1271
1272 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1273 if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1274 }
1275
1276 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1277 {
1278 struct wined3d_shader_version shader_version;
1279 const DWORD *ptr = byte_code;
1280 const char *type_prefix;
1281 DWORD i;
1282
1283 TRACE("Parsing %p.\n", byte_code);
1284
1285 fe->shader_read_header(fe_data, &ptr, &shader_version);
1286
1287 switch (shader_version.type)
1288 {
1289 case WINED3D_SHADER_TYPE_VERTEX:
1290 type_prefix = "vs";
1291 break;
1292
1293 case WINED3D_SHADER_TYPE_GEOMETRY:
1294 type_prefix = "gs";
1295 break;
1296
1297 case WINED3D_SHADER_TYPE_PIXEL:
1298 type_prefix = "ps";
1299 break;
1300
1301 default:
1302 FIXME("Unhandled shader type %#x.\n", shader_version.type);
1303 type_prefix = "unknown";
1304 break;
1305 }
1306
1307 TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1308
1309 while (!fe->shader_is_end(fe_data, &ptr))
1310 {
1311 struct wined3d_shader_instruction ins;
1312 const char *comment;
1313 UINT comment_size;
1314 UINT param_size;
1315
1316 /* comment */
1317 fe->shader_read_comment(&ptr, &comment, &comment_size);
1318 if (comment)
1319 {
1320 if (comment_size > 4 && *(const DWORD *)comment == WINEMAKEFOURCC('T', 'E', 'X', 'T'))
1321 {
1322 const char *end = comment + comment_size;
1323 const char *ptr = comment + 4;
1324 const char *line = ptr;
1325
1326 TRACE("// TEXT\n");
1327 while (ptr != end)
1328 {
1329 if (*ptr == '\n')
1330 {
1331 UINT len = ptr - line;
1332 if (len && *(ptr - 1) == '\r') --len;
1333 TRACE("// %s\n", debugstr_an(line, len));
1334 line = ++ptr;
1335 }
1336 else ++ptr;
1337 }
1338 if (line != ptr) TRACE("// %s\n", debugstr_an(line, ptr - line));
1339 }
1340 else TRACE("// %s\n", debugstr_an(comment, comment_size));
1341 continue;
1342 }
1343
1344 fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1345 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1346 {
1347 TRACE("Skipping unrecognized instruction.\n");
1348 ptr += param_size;
1349 continue;
1350 }
1351
1352 if (ins.handler_idx == WINED3DSIH_DCL)
1353 {
1354 struct wined3d_shader_semantic semantic;
1355
1356 fe->shader_read_semantic(&ptr, &semantic);
1357
1358 shader_dump_decl_usage(&semantic, &shader_version);
1359 shader_dump_ins_modifiers(&semantic.reg);
1360 TRACE(" ");
1361 shader_dump_dst_param(&semantic.reg, &shader_version);
1362 }
1363 else if (ins.handler_idx == WINED3DSIH_DEF)
1364 {
1365 struct wined3d_shader_src_param rel_addr;
1366 struct wined3d_shader_dst_param dst;
1367
1368 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1369
1370 TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(dst.reg.type, dst.reg.idx),
1371 *(const float *)(ptr),
1372 *(const float *)(ptr + 1),
1373 *(const float *)(ptr + 2),
1374 *(const float *)(ptr + 3));
1375 ptr += 4;
1376 }
1377 else if (ins.handler_idx == WINED3DSIH_DEFI)
1378 {
1379 struct wined3d_shader_src_param rel_addr;
1380 struct wined3d_shader_dst_param dst;
1381
1382 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1383
1384 TRACE("defi i%u = %d, %d, %d, %d", dst.reg.idx,
1385 *(ptr),
1386 *(ptr + 1),
1387 *(ptr + 2),
1388 *(ptr + 3));
1389 ptr += 4;
1390 }
1391 else if (ins.handler_idx == WINED3DSIH_DEFB)
1392 {
1393 struct wined3d_shader_src_param rel_addr;
1394 struct wined3d_shader_dst_param dst;
1395
1396 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1397
1398 TRACE("defb b%u = %s", dst.reg.idx, *ptr ? "true" : "false");
1399 ++ptr;
1400 }
1401 else
1402 {
1403 struct wined3d_shader_src_param dst_rel_addr[2];
1404 struct wined3d_shader_src_param src_rel_addr;
1405 struct wined3d_shader_dst_param dst_param[2];
1406 struct wined3d_shader_src_param src_param;
1407
1408 for (i = 0; i < ins.dst_count; ++i)
1409 {
1410 fe->shader_read_dst_param(fe_data, &ptr, &dst_param[i], &dst_rel_addr[i]);
1411 }
1412
1413 /* Print out predication source token first - it follows
1414 * the destination token. */
1415 if (ins.predicate)
1416 {
1417 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1418 TRACE("(");
1419 shader_dump_src_param(&src_param, &shader_version);
1420 TRACE(") ");
1421 }
1422
1423 /* PixWin marks instructions with the coissue flag with a '+' */
1424 if (ins.coissue) TRACE("+");
1425
1426 TRACE("%s", shader_opcode_names[ins.handler_idx]);
1427
1428 if (ins.handler_idx == WINED3DSIH_IFC
1429 || ins.handler_idx == WINED3DSIH_BREAKC)
1430 {
1431 switch (ins.flags)
1432 {
1433 case WINED3D_SHADER_REL_OP_GT: TRACE("_gt"); break;
1434 case WINED3D_SHADER_REL_OP_EQ: TRACE("_eq"); break;
1435 case WINED3D_SHADER_REL_OP_GE: TRACE("_ge"); break;
1436 case WINED3D_SHADER_REL_OP_LT: TRACE("_lt"); break;
1437 case WINED3D_SHADER_REL_OP_NE: TRACE("_ne"); break;
1438 case WINED3D_SHADER_REL_OP_LE: TRACE("_le"); break;
1439 default: TRACE("_(%u)", ins.flags);
1440 }
1441 }
1442 else if (ins.handler_idx == WINED3DSIH_TEX
1443 && shader_version.major >= 2
1444 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1445 {
1446 TRACE("p");
1447 }
1448
1449 /* We already read the destination tokens, print them. */
1450 for (i = 0; i < ins.dst_count; ++i)
1451 {
1452 shader_dump_ins_modifiers(&dst_param[i]);
1453 TRACE(!i ? " " : ", ");
1454 shader_dump_dst_param(&dst_param[i], &shader_version);
1455 }
1456
1457 /* Other source tokens */
1458 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1459 {
1460 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1461 TRACE(!i ? " " : ", ");
1462 shader_dump_src_param(&src_param, &shader_version);
1463 }
1464 }
1465 TRACE("\n");
1466 }
1467 }
1468
1469 static void shader_cleanup(struct wined3d_shader *shader)
1470 {
1471 shader->device->shader_backend->shader_destroy(shader);
1472 HeapFree(GetProcessHeap(), 0, shader->reg_maps.constf);
1473 HeapFree(GetProcessHeap(), 0, shader->function);
1474 shader_delete_constant_list(&shader->constantsF);
1475 shader_delete_constant_list(&shader->constantsB);
1476 shader_delete_constant_list(&shader->constantsI);
1477 list_remove(&shader->shader_list_entry);
1478
1479 if (shader->frontend && shader->frontend_data)
1480 shader->frontend->shader_free(shader->frontend_data);
1481 }
1482
1483 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1484 static void shader_none_select(const struct wined3d_context *context, BOOL usePS, BOOL useVS) {}
1485 static void shader_none_select_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info,
1486 enum tex_types tex_type, const SIZE *ds_mask_size) {}
1487 static void shader_none_deselect_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info) {}
1488 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
1489 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
1490 static void shader_none_load_constants(const struct wined3d_context *context, char usePS, char useVS) {}
1491 static void shader_none_load_np2fixup_constants(void *shader_priv,
1492 const struct wined3d_gl_info *gl_info, const struct wined3d_state *state) {}
1493 static void shader_none_destroy(struct wined3d_shader *shader) {}
1494 static HRESULT shader_none_alloc(struct wined3d_device *device) {return WINED3D_OK;}
1495 static void shader_none_free(struct wined3d_device *device) {}
1496 static void shader_none_context_destroyed(void *shader_priv, const struct wined3d_context *context) {}
1497
1498 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
1499 {
1500 /* Set the shader caps to 0 for the none shader backend */
1501 caps->VertexShaderVersion = 0;
1502 caps->MaxVertexShaderConst = 0;
1503 caps->PixelShaderVersion = 0;
1504 caps->PixelShader1xMaxValue = 0.0f;
1505 caps->MaxPixelShaderConst = 0;
1506 caps->VSClipping = FALSE;
1507 }
1508
1509 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
1510 {
1511 if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1512 {
1513 TRACE("Checking support for fixup:\n");
1514 dump_color_fixup_desc(fixup);
1515 }
1516
1517 /* Faked to make some apps happy. */
1518 if (!is_complex_fixup(fixup))
1519 {
1520 TRACE("[OK]\n");
1521 return TRUE;
1522 }
1523
1524 TRACE("[FAILED]\n");
1525 return FALSE;
1526 }
1527
1528 const struct wined3d_shader_backend_ops none_shader_backend =
1529 {
1530 shader_none_handle_instruction,
1531 shader_none_select,
1532 shader_none_select_depth_blt,
1533 shader_none_deselect_depth_blt,
1534 shader_none_update_float_vertex_constants,
1535 shader_none_update_float_pixel_constants,
1536 shader_none_load_constants,
1537 shader_none_load_np2fixup_constants,
1538 shader_none_destroy,
1539 shader_none_alloc,
1540 shader_none_free,
1541 shader_none_context_destroyed,
1542 shader_none_get_caps,
1543 shader_none_color_fixup_supported,
1544 };
1545
1546 static HRESULT shader_set_function(struct wined3d_shader *shader, const DWORD *byte_code,
1547 const struct wined3d_shader_signature *output_signature, DWORD float_const_count,
1548 enum wined3d_shader_type type, unsigned int max_version)
1549 {
1550 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1551 const struct wined3d_shader_frontend *fe;
1552 HRESULT hr;
1553 unsigned int backend_version;
1554
1555 TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u.\n",
1556 shader, byte_code, output_signature, float_const_count);
1557
1558 fe = shader_select_frontend(*byte_code);
1559 if (!fe)
1560 {
1561 FIXME("Unable to find frontend for shader.\n");
1562 return WINED3DERR_INVALIDCALL;
1563 }
1564 shader->frontend = fe;
1565 shader->frontend_data = fe->shader_init(byte_code, output_signature);
1566 if (!shader->frontend_data)
1567 {
1568 FIXME("Failed to initialize frontend.\n");
1569 return WINED3DERR_INVALIDCALL;
1570 }
1571
1572 /* First pass: trace shader. */
1573 if (TRACE_ON(d3d_shader))
1574 shader_trace_init(fe, shader->frontend_data, byte_code);
1575
1576 /* Initialize immediate constant lists. */
1577 list_init(&shader->constantsF);
1578 list_init(&shader->constantsB);
1579 list_init(&shader->constantsI);
1580
1581 /* Second pass: figure out which registers are used, what the semantics are, etc. */
1582 hr = shader_get_registers_used(shader, fe,
1583 reg_maps, shader->input_signature, shader->output_signature,
1584 byte_code, float_const_count);
1585 if (FAILED(hr)) return hr;
1586
1587 if (reg_maps->shader_version.type != type)
1588 {
1589 WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
1590 return WINED3DERR_INVALIDCALL;
1591 }
1592 if (reg_maps->shader_version.major > max_version)
1593 {
1594 WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
1595 return WINED3DERR_INVALIDCALL;
1596 }
1597 switch (type)
1598 {
1599 case WINED3D_SHADER_TYPE_VERTEX:
1600 backend_version = shader->device->vshader_version;
1601 break;
1602 case WINED3D_SHADER_TYPE_PIXEL:
1603 backend_version = shader->device->pshader_version;
1604 break;
1605 default:
1606 FIXME("No backend version-checking for this shader type\n");
1607 backend_version = 0;
1608 }
1609 if (reg_maps->shader_version.major > backend_version)
1610 {
1611 WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
1612 reg_maps->shader_version.major, reg_maps->shader_version.minor);
1613 return WINED3DERR_INVALIDCALL;
1614 }
1615
1616 shader->function = HeapAlloc(GetProcessHeap(), 0, shader->functionLength);
1617 if (!shader->function)
1618 return E_OUTOFMEMORY;
1619 memcpy(shader->function, byte_code, shader->functionLength);
1620
1621 return WINED3D_OK;
1622 }
1623
1624 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
1625 {
1626 ULONG refcount = InterlockedIncrement(&shader->ref);
1627
1628 TRACE("%p increasing refcount to %u.\n", shader, refcount);
1629
1630 return refcount;
1631 }
1632
1633 /* Do not call while under the GL lock. */
1634 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
1635 {
1636 ULONG refcount = InterlockedDecrement(&shader->ref);
1637
1638 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1639
1640 if (!refcount)
1641 {
1642 shader_cleanup(shader);
1643 shader->parent_ops->wined3d_object_destroyed(shader->parent);
1644 HeapFree(GetProcessHeap(), 0, shader);
1645 }
1646
1647 return refcount;
1648 }
1649
1650 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
1651 {
1652 TRACE("shader %p.\n", shader);
1653
1654 return shader->parent;
1655 }
1656
1657 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
1658 void *byte_code, UINT *byte_code_size)
1659 {
1660 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
1661
1662 if (!byte_code)
1663 {
1664 *byte_code_size = shader->functionLength;
1665 return WINED3D_OK;
1666 }
1667
1668 if (*byte_code_size < shader->functionLength)
1669 {
1670 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
1671 * than the required size we should write the required size and
1672 * return D3DERR_MOREDATA. That's not actually true. */
1673 return WINED3DERR_INVALIDCALL;
1674 }
1675
1676 memcpy(byte_code, shader->function, shader->functionLength);
1677
1678 return WINED3D_OK;
1679 }
1680
1681 /* Set local constants for d3d8 shaders. */
1682 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
1683 UINT start_idx, const float *src_data, UINT count)
1684 {
1685 UINT end_idx = start_idx + count;
1686 UINT i;
1687
1688 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
1689
1690 if (end_idx > shader->limits.constant_float)
1691 {
1692 WARN("end_idx %u > float constants limit %u.\n",
1693 end_idx, shader->limits.constant_float);
1694 end_idx = shader->limits.constant_float;
1695 }
1696
1697 for (i = start_idx; i < end_idx; ++i)
1698 {
1699 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1700 if (!lconst)
1701 return E_OUTOFMEMORY;
1702
1703 lconst->idx = i;
1704 memcpy(lconst->value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
1705 list_add_head(&shader->constantsF, &lconst->entry);
1706 }
1707
1708 return WINED3D_OK;
1709 }
1710
1711 void find_vs_compile_args(const struct wined3d_state *state,
1712 const struct wined3d_shader *shader, struct vs_compile_args *args)
1713 {
1714 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
1715 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
1716 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
1717 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
1718 args->swizzle_map = shader->device->strided_streams.swizzle_map;
1719 }
1720
1721 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
1722 {
1723 if (usage_idx1 != usage_idx2)
1724 return FALSE;
1725 if (usage1 == usage2)
1726 return TRUE;
1727 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
1728 return TRUE;
1729 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
1730 return TRUE;
1731
1732 return FALSE;
1733 }
1734
1735 BOOL vshader_get_input(const struct wined3d_shader *shader,
1736 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
1737 {
1738 WORD map = shader->reg_maps.input_registers;
1739 unsigned int i;
1740
1741 for (i = 0; map; map >>= 1, ++i)
1742 {
1743 if (!(map & 1)) continue;
1744
1745 if (match_usage(shader->u.vs.attributes[i].usage,
1746 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
1747 {
1748 *regnum = i;
1749 return TRUE;
1750 }
1751 }
1752 return FALSE;
1753 }
1754
1755 static void vertexshader_set_limits(struct wined3d_shader *shader)
1756 {
1757 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
1758 shader->reg_maps.shader_version.minor);
1759 struct wined3d_device *device = shader->device;
1760
1761 shader->limits.texcoord = 0;
1762 shader->limits.attributes = 16;
1763 shader->limits.packed_input = 0;
1764
1765 switch (shader_version)
1766 {
1767 case WINED3D_SHADER_VERSION(1, 0):
1768 case WINED3D_SHADER_VERSION(1, 1):
1769 shader->limits.temporary = 12;
1770 shader->limits.constant_bool = 0;
1771 shader->limits.constant_int = 0;
1772 shader->limits.address = 1;
1773 shader->limits.packed_output = 12;
1774 shader->limits.sampler = 0;
1775 shader->limits.label = 0;
1776 /* TODO: vs_1_1 has a minimum of 96 constants. What happens when
1777 * a vs_1_1 shader is used on a vs_3_0 capable card that has 256
1778 * constants? */
1779 shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1780 break;
1781
1782 case WINED3D_SHADER_VERSION(2, 0):
1783 case WINED3D_SHADER_VERSION(2, 1):
1784 shader->limits.temporary = 12;
1785 shader->limits.constant_bool = 16;
1786 shader->limits.constant_int = 16;
1787 shader->limits.address = 1;
1788 shader->limits.packed_output = 12;
1789 shader->limits.sampler = 0;
1790 shader->limits.label = 16;
1791 shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1792 break;
1793
1794 case WINED3D_SHADER_VERSION(4, 0):
1795 FIXME("Using 3.0 limits for 4.0 shader.\n");
1796 /* Fall through. */
1797
1798 case WINED3D_SHADER_VERSION(3, 0):
1799 shader->limits.temporary = 32;
1800 shader->limits.constant_bool = 32;
1801 shader->limits.constant_int = 32;
1802 shader->limits.address = 1;
1803 shader->limits.packed_output = 12;
1804 shader->limits.sampler = 4;
1805 shader->limits.label = 16; /* FIXME: 2048 */
1806 /* DX10 cards on Windows advertise a d3d9 constant limit of 256
1807 * even though they are capable of supporting much more (GL
1808 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
1809 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
1810 * shaders to 256. */
1811 shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1812 break;
1813
1814 default:
1815 shader->limits.temporary = 12;
1816 shader->limits.constant_bool = 16;
1817 shader->limits.constant_int = 16;
1818 shader->limits.address = 1;
1819 shader->limits.packed_output = 12;
1820 shader->limits.sampler = 0;
1821 shader->limits.label = 16;
1822 shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1823 FIXME("Unrecognized vertex shader version \"%u.%u\".\n",
1824 shader->reg_maps.shader_version.major,
1825 shader->reg_maps.shader_version.minor);
1826 }
1827 }
1828
1829 static HRESULT vertexshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
1830 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1831 void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
1832 {
1833 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1834 unsigned int i;
1835 HRESULT hr;
1836 WORD map;
1837
1838 if (!byte_code) return WINED3DERR_INVALIDCALL;
1839
1840 shader_init(shader, device, parent, parent_ops);
1841 hr = shader_set_function(shader, byte_code, output_signature, device->d3d_vshader_constantF,
1842 WINED3D_SHADER_TYPE_VERTEX, max_version);
1843 if (FAILED(hr))
1844 {
1845 WARN("Failed to set function, hr %#x.\n", hr);
1846 shader_cleanup(shader);
1847 return hr;
1848 }
1849
1850 map = reg_maps->input_registers;
1851 for (i = 0; map; map >>= 1, ++i)
1852 {
1853 if (!(map & 1) || !shader->input_signature[i].semantic_name)
1854 continue;
1855
1856 shader->u.vs.attributes[i].usage =
1857 shader_usage_from_semantic_name(shader->input_signature[i].semantic_name);
1858 shader->u.vs.attributes[i].usage_idx = shader->input_signature[i].semantic_idx;
1859 }
1860
1861 if (output_signature)
1862 {
1863 for (i = 0; i < output_signature->element_count; ++i)
1864 {
1865 struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1866 reg_maps->output_registers |= 1 << e->register_idx;
1867 shader->output_signature[e->register_idx] = *e;
1868 }
1869 }
1870
1871 vertexshader_set_limits(shader);
1872
1873 shader->load_local_constsF = reg_maps->usesrelconstF
1874 && !list_empty(&shader->constantsF);
1875
1876 return WINED3D_OK;
1877 }
1878
1879 static HRESULT geometryshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
1880 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1881 void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
1882 {
1883 HRESULT hr;
1884
1885 shader_init(shader, device, parent, parent_ops);
1886 hr = shader_set_function(shader, byte_code, output_signature, 0,
1887 WINED3D_SHADER_TYPE_GEOMETRY, max_version);
1888 if (FAILED(hr))
1889 {
1890 WARN("Failed to set function, hr %#x.\n", hr);
1891 shader_cleanup(shader);
1892 return hr;
1893 }
1894
1895 shader->load_local_constsF = FALSE;
1896
1897 return WINED3D_OK;
1898 }
1899
1900 void find_ps_compile_args(const struct wined3d_state *state,
1901 const struct wined3d_shader *shader, struct ps_compile_args *args)
1902 {
1903 struct wined3d_device *device = shader->device;
1904 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
1905 const struct wined3d_texture *texture;
1906 UINT i;
1907
1908 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
1909 if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && state->render_states[WINED3D_RS_SRGBWRITEENABLE])
1910 {
1911 const struct wined3d_surface *rt = state->fb->render_targets[0];
1912 if (rt->resource.format->flags & WINED3DFMT_FLAG_SRGB_WRITE)
1913 {
1914 static unsigned int warned = 0;
1915
1916 args->srgb_correction = 1;
1917 if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
1918 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
1919 "support, expect rendering artifacts.\n");
1920 }
1921 }
1922
1923 if (shader->reg_maps.shader_version.major == 1
1924 && shader->reg_maps.shader_version.minor <= 3)
1925 {
1926 for (i = 0; i < 4; ++i)
1927 {
1928 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
1929
1930 if (flags & WINED3D_TTFF_PROJECTED)
1931 {
1932 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
1933
1934 if (!state->vertex_shader)
1935 {
1936 unsigned int j;
1937 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
1938 DWORD max_valid = WINED3D_TTFF_COUNT4;
1939 enum wined3d_sampler_texture_type sampler_type = shader->reg_maps.sampler_type[i];
1940
1941 for (j = 0; j < state->vertex_declaration->element_count; ++j)
1942 {
1943 struct wined3d_vertex_declaration_element *element =
1944 &state->vertex_declaration->elements[j];
1945
1946 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
1947 && element->usage_idx == index)
1948 {
1949 max_valid = element->format->component_count;
1950 break;
1951 }
1952 }
1953 if (!tex_transform || tex_transform > max_valid)
1954 {
1955 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
1956 tex_transform, max_valid);
1957 tex_transform = max_valid;
1958 }
1959 if ((sampler_type == WINED3DSTT_1D && tex_transform > WINED3D_TTFF_COUNT1)
1960 || (sampler_type == WINED3DSTT_2D && tex_transform > WINED3D_TTFF_COUNT2)
1961 || (sampler_type == WINED3DSTT_VOLUME && tex_transform > WINED3D_TTFF_COUNT3))
1962 tex_transform |= WINED3D_PSARGS_PROJECTED;
1963 else
1964 {
1965 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
1966 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
1967 i, tex_transform, sampler_type);
1968 }
1969 }
1970 else
1971 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
1972
1973 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
1974 }
1975 }
1976 }
1977
1978 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
1979 {
1980 if (!shader->reg_maps.sampler_type[i])
1981 continue;
1982
1983 texture = state->textures[i];
1984 if (!texture)
1985 {
1986 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
1987 continue;
1988 }
1989 args->color_fixup[i] = texture->resource.format->color_fixup;
1990
1991 if (texture->resource.format->flags & WINED3DFMT_FLAG_SHADOW)
1992 args->shadow |= 1 << i;
1993
1994 /* Flag samplers that need NP2 texcoord fixup. */
1995 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
1996 args->np2_fixup |= (1 << i);
1997 }
1998 if (shader->reg_maps.shader_version.major >= 3)
1999 {
2000 if (device->strided_streams.position_transformed)
2001 {
2002 args->vp_mode = pretransformed;
2003 }
2004 else if (use_vs(state))
2005 {
2006 args->vp_mode = vertexshader;
2007 }
2008 else
2009 {
2010 args->vp_mode = fixedfunction;
2011 }
2012 args->fog = FOG_OFF;
2013 }
2014 else
2015 {
2016 args->vp_mode = vertexshader;
2017 if (state->render_states[WINED3D_RS_FOGENABLE])
2018 {
2019 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
2020 {
2021 case WINED3D_FOG_NONE:
2022 if (device->strided_streams.position_transformed || use_vs(state))
2023 {
2024 args->fog = FOG_LINEAR;
2025 break;
2026 }
2027
2028 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
2029 {
2030 case WINED3D_FOG_NONE: /* Fall through. */
2031 case WINED3D_FOG_LINEAR: args->fog = FOG_LINEAR; break;
2032 case WINED3D_FOG_EXP: args->fog = FOG_EXP; break;
2033 case WINED3D_FOG_EXP2: args->fog = FOG_EXP2; break;
2034 }
2035 break;
2036
2037 case WINED3D_FOG_LINEAR: args->fog = FOG_LINEAR; break;
2038 case WINED3D_FOG_EXP: args->fog = FOG_EXP; break;
2039 case WINED3D_FOG_EXP2: args->fog = FOG_EXP2; break;
2040 }
2041 }
2042 else
2043 {
2044 args->fog = FOG_OFF;
2045 }
2046 }
2047 }
2048
2049 static void pixelshader_set_limits(struct wined3d_shader *shader)
2050 {
2051 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
2052 shader->reg_maps.shader_version.minor);
2053
2054 shader->limits.attributes = 0;
2055 shader->limits.address = 0;
2056 shader->limits.packed_output = 0;
2057
2058 switch (shader_version)
2059 {
2060 case WINED3D_SHADER_VERSION(1, 0):
2061 case WINED3D_SHADER_VERSION(1, 1):
2062 case WINED3D_SHADER_VERSION(1, 2):
2063 case WINED3D_SHADER_VERSION(1, 3):
2064 shader->limits.temporary = 2;
2065 shader->limits.constant_float = 8;
2066 shader->limits.constant_int = 0;
2067 shader->limits.constant_bool = 0;
2068 shader->limits.texcoord = 4;
2069 shader->limits.sampler = 4;
2070 shader->limits.packed_input = 0;
2071 shader->limits.label = 0;
2072 break;
2073
2074 case WINED3D_SHADER_VERSION(1, 4):
2075 shader->limits.temporary = 6;
2076 shader->limits.constant_float = 8;
2077 shader->limits.constant_int = 0;
2078 shader->limits.constant_bool = 0;
2079 shader->limits.texcoord = 6;
2080 shader->limits.sampler = 6;
2081 shader->limits.packed_input = 0;
2082 shader->limits.label = 0;
2083 break;
2084
2085 /* FIXME: Temporaries must match D3DPSHADERCAPS2_0.NumTemps. */
2086 case WINED3D_SHADER_VERSION(2, 0):
2087 shader->limits.temporary = 32;
2088 shader->limits.constant_float = 32;
2089 shader->limits.constant_int = 16;
2090 shader->limits.constant_bool = 16;
2091 shader->limits.texcoord = 8;
2092 shader->limits.sampler = 16;
2093 shader->limits.packed_input = 0;
2094 break;
2095
2096 case WINED3D_SHADER_VERSION(2, 1):
2097 shader->limits.temporary = 32;
2098 shader->limits.constant_float = 32;
2099 shader->limits.constant_int = 16;
2100 shader->limits.constant_bool = 16;
2101 shader->limits.texcoord = 8;
2102 shader->limits.sampler = 16;
2103 shader->limits.packed_input = 0;
2104 shader->limits.label = 16;
2105 break;
2106
2107 case WINED3D_SHADER_VERSION(4, 0):
2108 FIXME("Using 3.0 limits for 4.0 shader.\n");
2109 /* Fall through. */
2110
2111 case WINED3D_SHADER_VERSION(3, 0):
2112 shader->limits.temporary = 32;
2113 shader->limits.constant_float = 224;
2114 shader->limits.constant_int = 16;
2115 shader->limits.constant_bool = 16;
2116 shader->limits.texcoord = 0;
2117 shader->limits.sampler = 16;
2118 shader->limits.packed_input = 12;
2119 shader->limits.label = 16; /* FIXME: 2048 */
2120 break;
2121
2122 default:
2123 shader->limits.temporary = 32;
2124 shader->limits.constant_float = 32;
2125 shader->limits.constant_int = 16;
2126 shader->limits.constant_bool = 16;
2127 shader->limits.texcoord = 8;
2128 shader->limits.sampler = 16;
2129 shader->limits.packed_input = 0;
2130 shader->limits.label = 0;
2131 FIXME("Unrecognized pixel shader version %u.%u\n",
2132 shader->reg_maps.shader_version.major,
2133 shader->reg_maps.shader_version.minor);
2134 }
2135 }
2136
2137 static HRESULT pixelshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2138 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2139 void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
2140 {
2141 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2142 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2143 HRESULT hr;
2144
2145 if (!byte_code) return WINED3DERR_INVALIDCALL;
2146
2147 shader_init(shader, device, parent, parent_ops);
2148 hr = shader_set_function(shader, byte_code, output_signature, device->d3d_pshader_constantF,
2149 WINED3D_SHADER_TYPE_PIXEL, max_version);
2150 if (FAILED(hr))
2151 {
2152 WARN("Failed to set function, hr %#x.\n", hr);
2153 shader_cleanup(shader);
2154 return hr;
2155 }
2156
2157 pixelshader_set_limits(shader);
2158
2159 for (i = 0; i < MAX_REG_INPUT; ++i)
2160 {
2161 if (shader->u.ps.input_reg_used[i])
2162 {
2163 ++num_regs_used;
2164 highest_reg_used = i;
2165 }
2166 }
2167
2168 /* Don't do any register mapping magic if it is not needed, or if we can't
2169 * achieve anything anyway */
2170 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2171 || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2172 {
2173 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2174 {
2175 /* This happens with relative addressing. The input mapper function
2176 * warns about this if the higher registers are declared too, so
2177 * don't write a FIXME here */
2178 WARN("More varying registers used than supported\n");
2179 }
2180
2181 for (i = 0; i < MAX_REG_INPUT; ++i)
2182 {
2183 shader->u.ps.input_reg_map[i] = i;
2184 }
2185
2186 shader->u.ps.declared_in_count = highest_reg_used + 1;
2187 }
2188 else
2189 {
2190 shader->u.ps.declared_in_count = 0;
2191 for (i = 0; i < MAX_REG_INPUT; ++i)
2192 {
2193 if (shader->u.ps.input_reg_used[i])
2194 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
2195 else shader->u.ps.input_reg_map[i] = ~0U;
2196 }
2197 }
2198
2199 shader->load_local_constsF = FALSE;
2200
2201 return WINED3D_OK;
2202 }
2203
2204 void pixelshader_update_samplers(struct wined3d_shader_reg_maps *reg_maps, struct wined3d_texture * const *textures)
2205 {
2206 enum wined3d_sampler_texture_type *sampler_type = reg_maps->sampler_type;
2207 unsigned int i;
2208
2209 if (reg_maps->shader_version.major != 1) return;
2210
2211 for (i = 0; i < max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS); ++i)
2212 {
2213 /* We don't sample from this sampler. */
2214 if (!sampler_type[i]) continue;
2215
2216 if (!textures[i])
2217 {
2218 WARN("No texture bound to sampler %u, using 2D.\n", i);
2219 sampler_type[i] = WINED3DSTT_2D;
2220 continue;
2221 }
2222
2223 switch (textures[i]->target)
2224 {
2225 case GL_TEXTURE_RECTANGLE_ARB:
2226 case GL_TEXTURE_2D:
2227 /* We have to select between texture rectangles and 2D
2228 * textures later because 2.0 and 3.0 shaders only have
2229 * WINED3DSTT_2D as well. */
2230 sampler_type[i] = WINED3DSTT_2D;
2231 break;
2232
2233 case GL_TEXTURE_3D:
2234 sampler_type[i] = WINED3DSTT_VOLUME;
2235 break;
2236
2237 case GL_TEXTURE_CUBE_MAP_ARB:
2238 sampler_type[i] = WINED3DSTT_CUBE;
2239 break;
2240
2241 default:
2242 FIXME("Unrecognized texture type %#x, using 2D.\n", textures[i]->target);
2243 sampler_type[i] = WINED3DSTT_2D;
2244 }
2245 }
2246 }
2247
2248 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const DWORD *byte_code,
2249 const struct wined3d_shader_signature *output_signature, void *parent,
2250 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2251 {
2252 struct wined3d_shader *object;
2253 HRESULT hr;
2254
2255 TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2256 device, byte_code, output_signature, parent, parent_ops, shader);
2257
2258 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2259 if (!object)
2260 {
2261 ERR("Failed to allocate shader memory.\n");
2262 return E_OUTOFMEMORY;
2263 }
2264
2265 hr = geometryshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2266 if (FAILED(hr))
2267 {
2268 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
2269 HeapFree(GetProcessHeap(), 0, object);
2270 return hr;
2271 }
2272
2273 TRACE("Created geometry shader %p.\n", object);
2274 *shader = object;
2275
2276 return WINED3D_OK;
2277 }
2278
2279 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const DWORD *byte_code,
2280 const struct wined3d_shader_signature *output_signature, void *parent,
2281 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2282 {
2283 struct wined3d_shader *object;
2284 HRESULT hr;
2285
2286 TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2287 device, byte_code, output_signature, parent, parent_ops, shader);
2288
2289 if (device->ps_selected_mode == SHADER_NONE)
2290 return WINED3DERR_INVALIDCALL;
2291
2292 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2293 if (!object)
2294 {
2295 ERR("Failed to allocate shader memory.\n");
2296 return E_OUTOFMEMORY;
2297 }
2298
2299 hr = pixelshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2300 if (FAILED(hr))
2301 {
2302 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
2303 HeapFree(GetProcessHeap(), 0, object);
2304 return hr;
2305 }
2306
2307 TRACE("Created pixel shader %p.\n", object);
2308 *shader = object;
2309
2310 return WINED3D_OK;
2311 }
2312
2313 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const DWORD *byte_code,
2314 const struct wined3d_shader_signature *output_signature, void *parent,
2315 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2316 {
2317 struct wined3d_shader *object;
2318 HRESULT hr;
2319
2320 TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2321 device, byte_code, output_signature, parent, parent_ops, shader);
2322
2323 if (device->vs_selected_mode == SHADER_NONE)
2324 return WINED3DERR_INVALIDCALL;
2325
2326 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2327 if (!object)
2328 {
2329 ERR("Failed to allocate shader memory.\n");
2330 return E_OUTOFMEMORY;
2331 }
2332
2333 hr = vertexshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2334 if (FAILED(hr))
2335 {
2336 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
2337 HeapFree(GetProcessHeap(), 0, object);
2338 return hr;
2339 }
2340
2341 TRACE("Created vertex shader %p.\n", object);
2342 *shader = object;
2343
2344 return WINED3D_OK;
2345 }
A free Windows-compatible Operating System