projects / reactos.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
Sync trunk.
[reactos.git] / 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 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 *shader_opcode_names[] =
37 {
38 /* WINED3DSIH_ABS */ "abs",
39 /* WINED3DSIH_ADD */ "add",
40 /* WINED3DSIH_BEM */ "bem",
41 /* WINED3DSIH_BREAK */ "break",
42 /* WINED3DSIH_BREAKC */ "breakc",
43 /* WINED3DSIH_BREAKP */ "breakp",
44 /* WINED3DSIH_CALL */ "call",
45 /* WINED3DSIH_CALLNZ */ "callnz",
46 /* WINED3DSIH_CMP */ "cmp",
47 /* WINED3DSIH_CND */ "cnd",
48 /* WINED3DSIH_CRS */ "crs",
49 /* WINED3DSIH_CUT */ "cut",
50 /* WINED3DSIH_DCL */ "dcl",
51 /* WINED3DSIH_DEF */ "def",
52 /* WINED3DSIH_DEFB */ "defb",
53 /* WINED3DSIH_DEFI */ "defi",
54 /* WINED3DSIH_DP2ADD */ "dp2add",
55 /* WINED3DSIH_DP3 */ "dp3",
56 /* WINED3DSIH_DP4 */ "dp4",
57 /* WINED3DSIH_DST */ "dst",
58 /* WINED3DSIH_DSX */ "dsx",
59 /* WINED3DSIH_DSY */ "dsy",
60 /* WINED3DSIH_ELSE */ "else",
61 /* WINED3DSIH_EMIT */ "emit",
62 /* WINED3DSIH_ENDIF */ "endif",
63 /* WINED3DSIH_ENDLOOP */ "endloop",
64 /* WINED3DSIH_ENDREP */ "endrep",
65 /* WINED3DSIH_EXP */ "exp",
66 /* WINED3DSIH_EXPP */ "expp",
67 /* WINED3DSIH_FRC */ "frc",
68 /* WINED3DSIH_IADD */ "iadd",
69 /* WINED3DSIH_IF */ "if",
70 /* WINED3DSIH_IFC */ "ifc",
71 /* WINED3DSIH_IGE */ "ige",
72 /* WINED3DSIH_LABEL */ "label",
73 /* WINED3DSIH_LIT */ "lit",
74 /* WINED3DSIH_LOG */ "log",
75 /* WINED3DSIH_LOGP */ "logp",
76 /* WINED3DSIH_LOOP */ "loop",
77 /* WINED3DSIH_LRP */ "lrp",
78 /* WINED3DSIH_LT */ "lt",
79 /* WINED3DSIH_M3x2 */ "m3x2",
80 /* WINED3DSIH_M3x3 */ "m3x3",
81 /* WINED3DSIH_M3x4 */ "m3x4",
82 /* WINED3DSIH_M4x3 */ "m4x3",
83 /* WINED3DSIH_M4x4 */ "m4x4",
84 /* WINED3DSIH_MAD */ "mad",
85 /* WINED3DSIH_MAX */ "max",
86 /* WINED3DSIH_MIN */ "min",
87 /* WINED3DSIH_MOV */ "mov",
88 /* WINED3DSIH_MOVA */ "mova",
89 /* WINED3DSIH_MUL */ "mul",
90 /* WINED3DSIH_NOP */ "nop",
91 /* WINED3DSIH_NRM */ "nrm",
92 /* WINED3DSIH_PHASE */ "phase",
93 /* WINED3DSIH_POW */ "pow",
94 /* WINED3DSIH_RCP */ "rcp",
95 /* WINED3DSIH_REP */ "rep",
96 /* WINED3DSIH_RET */ "ret",
97 /* WINED3DSIH_RSQ */ "rsq",
98 /* WINED3DSIH_SETP */ "setp",
99 /* WINED3DSIH_SGE */ "sge",
100 /* WINED3DSIH_SGN */ "sgn",
101 /* WINED3DSIH_SINCOS */ "sincos",
102 /* WINED3DSIH_SLT */ "slt",
103 /* WINED3DSIH_SUB */ "sub",
104 /* WINED3DSIH_TEX */ "texld",
105 /* WINED3DSIH_TEXBEM */ "texbem",
106 /* WINED3DSIH_TEXBEML */ "texbeml",
107 /* WINED3DSIH_TEXCOORD */ "texcrd",
108 /* WINED3DSIH_TEXDEPTH */ "texdepth",
109 /* WINED3DSIH_TEXDP3 */ "texdp3",
110 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
111 /* WINED3DSIH_TEXKILL */ "texkill",
112 /* WINED3DSIH_TEXLDD */ "texldd",
113 /* WINED3DSIH_TEXLDL */ "texldl",
114 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
115 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
116 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
117 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
118 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
119 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
120 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
121 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
122 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
123 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
124 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
125 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
126 };
127
128 static const char *semantic_names[] =
129 {
130 /* WINED3DDECLUSAGE_POSITION */ "SV_POSITION",
131 /* WINED3DDECLUSAGE_BLENDWEIGHT */ "BLENDWEIGHT",
132 /* WINED3DDECLUSAGE_BLENDINDICES */ "BLENDINDICES",
133 /* WINED3DDECLUSAGE_NORMAL */ "NORMAL",
134 /* WINED3DDECLUSAGE_PSIZE */ "PSIZE",
135 /* WINED3DDECLUSAGE_TEXCOORD */ "TEXCOORD",
136 /* WINED3DDECLUSAGE_TANGENT */ "TANGENT",
137 /* WINED3DDECLUSAGE_BINORMAL */ "BINORMAL",
138 /* WINED3DDECLUSAGE_TESSFACTOR */ "TESSFACTOR",
139 /* WINED3DDECLUSAGE_POSITIONT */ "POSITIONT",
140 /* WINED3DDECLUSAGE_COLOR */ "COLOR",
141 /* WINED3DDECLUSAGE_FOG */ "FOG",
142 /* WINED3DDECLUSAGE_DEPTH */ "DEPTH",
143 /* WINED3DDECLUSAGE_SAMPLE */ "SAMPLE",
144 };
145
146 static const char *shader_semantic_name_from_usage(WINED3DDECLUSAGE usage)
147 {
148 if (usage >= sizeof(semantic_names) / sizeof(*semantic_names))
149 {
150 FIXME("Unrecognized usage %#x.\n", usage);
151 return "UNRECOGNIZED";
152 }
153
154 return semantic_names[usage];
155 }
156
157 static WINED3DDECLUSAGE shader_usage_from_semantic_name(const char *name)
158 {
159 unsigned int i;
160
161 for (i = 0; i < sizeof(semantic_names) / sizeof(*semantic_names); ++i)
162 {
163 if (!strcmp(name, semantic_names[i])) return i;
164 }
165
166 return ~0U;
167 }
168
169 BOOL shader_match_semantic(const char *semantic_name, WINED3DDECLUSAGE usage)
170 {
171 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
172 }
173
174 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
175 const struct wined3d_shader_semantic *s)
176 {
177 e->semantic_name = shader_semantic_name_from_usage(s->usage);
178 e->semantic_idx = s->usage_idx;
179 e->sysval_semantic = 0;
180 e->component_type = 0;
181 e->register_idx = s->reg.reg.idx;
182 e->mask = s->reg.write_mask;
183 }
184
185 static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
186 {
187 switch (version_token >> 16)
188 {
189 case WINED3D_SM1_VS:
190 case WINED3D_SM1_PS:
191 return &sm1_shader_frontend;
192
193 case WINED3D_SM4_PS:
194 case WINED3D_SM4_VS:
195 case WINED3D_SM4_GS:
196 return &sm4_shader_frontend;
197
198 default:
199 FIXME("Unrecognised version token %#x\n", version_token);
200 return NULL;
201 }
202 }
203
204 void shader_buffer_clear(struct wined3d_shader_buffer *buffer)
205 {
206 buffer->buffer[0] = '\0';
207 buffer->bsize = 0;
208 buffer->lineNo = 0;
209 buffer->newline = TRUE;
210 }
211
212 BOOL shader_buffer_init(struct wined3d_shader_buffer *buffer)
213 {
214 buffer->buffer = HeapAlloc(GetProcessHeap(), 0, SHADER_PGMSIZE);
215 if (!buffer->buffer)
216 {
217 ERR("Failed to allocate shader buffer memory.\n");
218 return FALSE;
219 }
220
221 shader_buffer_clear(buffer);
222 return TRUE;
223 }
224
225 void shader_buffer_free(struct wined3d_shader_buffer *buffer)
226 {
227 HeapFree(GetProcessHeap(), 0, buffer->buffer);
228 }
229
230 int shader_vaddline(struct wined3d_shader_buffer *buffer, const char *format, va_list args)
231 {
232 char *base = buffer->buffer + buffer->bsize;
233 int rc;
234
235 rc = vsnprintf(base, SHADER_PGMSIZE - 1 - buffer->bsize, format, args);
236
237 if (rc < 0 /* C89 */ || (unsigned int)rc > SHADER_PGMSIZE - 1 - buffer->bsize /* C99 */)
238 {
239 ERR("The buffer allocated for the shader program string "
240 "is too small at %d bytes.\n", SHADER_PGMSIZE);
241 buffer->bsize = SHADER_PGMSIZE - 1;
242 return -1;
243 }
244
245 if (buffer->newline)
246 {
247 TRACE("GL HW (%u, %u) : %s", buffer->lineNo + 1, buffer->bsize, base);
248 buffer->newline = FALSE;
249 }
250 else
251 {
252 TRACE("%s", base);
253 }
254
255 buffer->bsize += rc;
256 if (buffer->buffer[buffer->bsize-1] == '\n')
257 {
258 ++buffer->lineNo;
259 buffer->newline = TRUE;
260 }
261
262 return 0;
263 }
264
265 int shader_addline(struct wined3d_shader_buffer *buffer, const char *format, ...)
266 {
267 va_list args;
268 int ret;
269
270 va_start(args, format);
271 ret = shader_vaddline(buffer, format, args);
272 va_end(args);
273
274 return ret;
275 }
276
277 static void shader_init(struct IWineD3DBaseShaderClass *shader, IWineD3DDeviceImpl *device,
278 IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
279 {
280 shader->ref = 1;
281 shader->device = (IWineD3DDevice *)device;
282 shader->parent = parent;
283 shader->parent_ops = parent_ops;
284 list_init(&shader->linked_programs);
285 list_add_head(&device->shaders, &shader->shader_list_entry);
286 }
287
288 /* Convert floating point offset relative to a register file to an absolute
289 * offset for float constants. */
290 static unsigned int shader_get_float_offset(WINED3DSHADER_PARAM_REGISTER_TYPE register_type, UINT register_idx)
291 {
292 switch (register_type)
293 {
294 case WINED3DSPR_CONST: return register_idx;
295 case WINED3DSPR_CONST2: return 2048 + register_idx;
296 case WINED3DSPR_CONST3: return 4096 + register_idx;
297 case WINED3DSPR_CONST4: return 6144 + register_idx;
298 default:
299 FIXME("Unsupported register type: %u.\n", register_type);
300 return register_idx;
301 }
302 }
303
304 static void shader_delete_constant_list(struct list *clist)
305 {
306 struct local_constant *constant;
307 struct list *ptr;
308
309 ptr = list_head(clist);
310 while (ptr)
311 {
312 constant = LIST_ENTRY(ptr, struct local_constant, entry);
313 ptr = list_next(clist, ptr);
314 HeapFree(GetProcessHeap(), 0, constant);
315 }
316 list_init(clist);
317 }
318
319 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
320 {
321 DWORD idx, shift;
322 idx = bit >> 5;
323 shift = bit & 0x1f;
324 bitmap[idx] |= (1 << shift);
325 }
326
327 static void shader_record_register_usage(IWineD3DBaseShaderImpl *shader, struct shader_reg_maps *reg_maps,
328 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type)
329 {
330 switch (reg->type)
331 {
332 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
333 if (shader_type == WINED3D_SHADER_TYPE_PIXEL) reg_maps->texcoord |= 1 << reg->idx;
334 else reg_maps->address |= 1 << reg->idx;
335 break;
336
337 case WINED3DSPR_TEMP:
338 reg_maps->temporary |= 1 << reg->idx;
339 break;
340
341 case WINED3DSPR_INPUT:
342 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
343 {
344 if (reg->rel_addr)
345 {
346 /* If relative addressing is used, we must assume that all registers
347 * are used. Even if it is a construct like v3[aL], we can't assume
348 * that v0, v1 and v2 aren't read because aL can be negative */
349 unsigned int i;
350 for (i = 0; i < MAX_REG_INPUT; ++i)
351 {
352 ((IWineD3DPixelShaderImpl *)shader)->input_reg_used[i] = TRUE;
353 }
354 }
355 else
356 {
357 ((IWineD3DPixelShaderImpl *)shader)->input_reg_used[reg->idx] = TRUE;
358 }
359 }
360 else reg_maps->input_registers |= 1 << reg->idx;
361 break;
362
363 case WINED3DSPR_RASTOUT:
364 if (reg->idx == 1) reg_maps->fog = 1;
365 break;
366
367 case WINED3DSPR_MISCTYPE:
368 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
369 {
370 if (reg->idx == 0) reg_maps->vpos = 1;
371 else if (reg->idx == 1) reg_maps->usesfacing = 1;
372 }
373 break;
374
375 case WINED3DSPR_CONST:
376 if (reg->rel_addr)
377 {
378 if (shader_type != WINED3D_SHADER_TYPE_PIXEL)
379 {
380 if (reg->idx < ((IWineD3DVertexShaderImpl *)shader)->min_rel_offset)
381 {
382 ((IWineD3DVertexShaderImpl *)shader)->min_rel_offset = reg->idx;
383 }
384 if (reg->idx > ((IWineD3DVertexShaderImpl *)shader)->max_rel_offset)
385 {
386 ((IWineD3DVertexShaderImpl *)shader)->max_rel_offset = reg->idx;
387 }
388 }
389 reg_maps->usesrelconstF = TRUE;
390 }
391 else
392 {
393 set_bitmap_bit(reg_maps->constf, reg->idx);
394 }
395 break;
396
397 case WINED3DSPR_CONSTINT:
398 reg_maps->integer_constants |= (1 << reg->idx);
399 break;
400
401 case WINED3DSPR_CONSTBOOL:
402 reg_maps->boolean_constants |= (1 << reg->idx);
403 break;
404
405 case WINED3DSPR_COLOROUT:
406 reg_maps->highest_render_target = max(reg_maps->highest_render_target, reg->idx);
407 break;
408
409 default:
410 TRACE("Not recording register of type %#x and idx %u\n", reg->type, reg->idx);
411 break;
412 }
413 }
414
415 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
416 {
417 switch (instr)
418 {
419 case WINED3DSIH_M4x4:
420 case WINED3DSIH_M3x4:
421 return param == 1 ? 3 : 0;
422
423 case WINED3DSIH_M4x3:
424 case WINED3DSIH_M3x3:
425 return param == 1 ? 2 : 0;
426
427 case WINED3DSIH_M3x2:
428 return param == 1 ? 1 : 0;
429
430 default:
431 return 0;
432 }
433 }
434
435 /* Note that this does not count the loop register as an address register. */
436 static HRESULT shader_get_registers_used(IWineD3DBaseShader *iface, const struct wined3d_shader_frontend *fe,
437 struct shader_reg_maps *reg_maps, struct wined3d_shader_signature_element *input_signature,
438 struct wined3d_shader_signature_element *output_signature, const DWORD *byte_code, DWORD constf_size)
439 {
440 IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)iface;
441 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
442 void *fe_data = shader->baseShader.frontend_data;
443 struct wined3d_shader_version shader_version;
444 const DWORD *ptr = byte_code;
445
446 memset(reg_maps, 0, sizeof(*reg_maps));
447
448 /* get_registers_used() is called on every compile on some 1.x shaders,
449 * which can result in stacking up a collection of local constants.
450 * Delete the old constants if existing. */
451 shader_delete_constant_list(&shader->baseShader.constantsF);
452 shader_delete_constant_list(&shader->baseShader.constantsB);
453 shader_delete_constant_list(&shader->baseShader.constantsI);
454
455 fe->shader_read_header(fe_data, &ptr, &shader_version);
456 reg_maps->shader_version = shader_version;
457
458 reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
459 sizeof(*reg_maps->constf) * ((constf_size + 31) / 32));
460 if (!reg_maps->constf)
461 {
462 ERR("Failed to allocate constant map memory.\n");
463 return E_OUTOFMEMORY;
464 }
465
466 while (!fe->shader_is_end(fe_data, &ptr))
467 {
468 struct wined3d_shader_instruction ins;
469 const char *comment;
470 UINT comment_size;
471 UINT param_size;
472
473 /* Skip comments. */
474 fe->shader_read_comment(&ptr, &comment, &comment_size);
475 if (comment) continue;
476
477 /* Fetch opcode. */
478 fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
479
480 /* Unhandled opcode, and its parameters. */
481 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
482 {
483 TRACE("Skipping unrecognized instruction.\n");
484 ptr += param_size;
485 continue;
486 }
487
488 /* Handle declarations. */
489 if (ins.handler_idx == WINED3DSIH_DCL)
490 {
491 struct wined3d_shader_semantic semantic;
492
493 fe->shader_read_semantic(&ptr, &semantic);
494
495 switch (semantic.reg.reg.type)
496 {
497 /* Mark input registers used. */
498 case WINED3DSPR_INPUT:
499 reg_maps->input_registers |= 1 << semantic.reg.reg.idx;
500 shader_signature_from_semantic(&input_signature[semantic.reg.reg.idx], &semantic);
501 break;
502
503 /* Vertex shader: mark 3.0 output registers used, save token. */
504 case WINED3DSPR_OUTPUT:
505 reg_maps->output_registers |= 1 << semantic.reg.reg.idx;
506 shader_signature_from_semantic(&output_signature[semantic.reg.reg.idx], &semantic);
507 if (semantic.usage == WINED3DDECLUSAGE_FOG) reg_maps->fog = 1;
508 break;
509
510 /* Save sampler usage token. */
511 case WINED3DSPR_SAMPLER:
512 reg_maps->sampler_type[semantic.reg.reg.idx] = semantic.sampler_type;
513 break;
514
515 default:
516 TRACE("Not recording DCL register type %#x.\n", semantic.reg.reg.type);
517 break;
518 }
519 }
520 else if (ins.handler_idx == WINED3DSIH_DEF)
521 {
522 struct wined3d_shader_src_param rel_addr;
523 struct wined3d_shader_dst_param dst;
524
525 local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
526 if (!lconst) return E_OUTOFMEMORY;
527
528 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
529 lconst->idx = dst.reg.idx;
530
531 memcpy(lconst->value, ptr, 4 * sizeof(DWORD));
532 ptr += 4;
533
534 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
535 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
536 {
537 float *value = (float *)lconst->value;
538 if (value[0] < -1.0f) value[0] = -1.0f;
539 else if (value[0] > 1.0f) value[0] = 1.0f;
540 if (value[1] < -1.0f) value[1] = -1.0f;
541 else if (value[1] > 1.0f) value[1] = 1.0f;
542 if (value[2] < -1.0f) value[2] = -1.0f;
543 else if (value[2] > 1.0f) value[2] = 1.0f;
544 if (value[3] < -1.0f) value[3] = -1.0f;
545 else if (value[3] > 1.0f) value[3] = 1.0f;
546 }
547
548 list_add_head(&shader->baseShader.constantsF, &lconst->entry);
549 }
550 else if (ins.handler_idx == WINED3DSIH_DEFI)
551 {
552 struct wined3d_shader_src_param rel_addr;
553 struct wined3d_shader_dst_param dst;
554
555 local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
556 if (!lconst) return E_OUTOFMEMORY;
557
558 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
559 lconst->idx = dst.reg.idx;
560
561 memcpy(lconst->value, ptr, 4 * sizeof(DWORD));
562 ptr += 4;
563
564 list_add_head(&shader->baseShader.constantsI, &lconst->entry);
565 reg_maps->local_int_consts |= (1 << dst.reg.idx);
566 }
567 else if (ins.handler_idx == WINED3DSIH_DEFB)
568 {
569 struct wined3d_shader_src_param rel_addr;
570 struct wined3d_shader_dst_param dst;
571
572 local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
573 if (!lconst) return E_OUTOFMEMORY;
574
575 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
576 lconst->idx = dst.reg.idx;
577
578 memcpy(lconst->value, ptr, sizeof(DWORD));
579 ++ptr;
580
581 list_add_head(&shader->baseShader.constantsB, &lconst->entry);
582 reg_maps->local_bool_consts |= (1 << dst.reg.idx);
583 }
584 /* If there's a loop in the shader. */
585 else if (ins.handler_idx == WINED3DSIH_LOOP
586 || ins.handler_idx == WINED3DSIH_REP)
587 {
588 struct wined3d_shader_src_param src, rel_addr;
589
590 fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
591
592 /* Rep and Loop always use an integer constant for the control parameters. */
593 if (ins.handler_idx == WINED3DSIH_REP)
594 {
595 reg_maps->integer_constants |= 1 << src.reg.idx;
596 }
597 else
598 {
599 fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
600 reg_maps->integer_constants |= 1 << src.reg.idx;
601 }
602
603 cur_loop_depth++;
604 if (cur_loop_depth > max_loop_depth) max_loop_depth = cur_loop_depth;
605 }
606 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
607 || ins.handler_idx == WINED3DSIH_ENDREP)
608 {
609 cur_loop_depth--;
610 }
611 /* For subroutine prototypes. */
612 else if (ins.handler_idx == WINED3DSIH_LABEL)
613 {
614 struct wined3d_shader_src_param src, rel_addr;
615
616 fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
617 reg_maps->labels |= 1 << src.reg.idx;
618 }
619 /* Set texture, address, temporary registers. */
620 else
621 {
622 BOOL color0_mov = FALSE;
623 int i, limit;
624
625 /* This will loop over all the registers and try to
626 * make a bitmask of the ones we're interested in.
627 *
628 * Relative addressing tokens are ignored, but that's
629 * okay, since we'll catch any address registers when
630 * they are initialized (required by spec). */
631 if (ins.dst_count)
632 {
633 struct wined3d_shader_src_param dst_rel_addr;
634 struct wined3d_shader_dst_param dst_param;
635
636 fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
637
638 shader_record_register_usage(shader, reg_maps, &dst_param.reg, shader_version.type);
639
640 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and
641 * is used in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel
642 * shaders because TECRDOUT isn't used in them, but future register types might cause issues */
643 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX && shader_version.major < 3
644 && dst_param.reg.type == WINED3DSPR_TEXCRDOUT)
645 {
646 reg_maps->texcoord_mask[dst_param.reg.idx] |= dst_param.write_mask;
647 }
648
649 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
650 {
651 IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)shader;
652
653 if (dst_param.reg.type == WINED3DSPR_COLOROUT && dst_param.reg.idx == 0)
654 {
655 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
656 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
657 * the mov and perform the sRGB write correction from the source register.
658 *
659 * However, if the mov is only partial, we can't do this, and if the write
660 * comes from an instruction other than MOV it is hard to do as well. If
661 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
662
663 ps->color0_mov = FALSE;
664 if (ins.handler_idx == WINED3DSIH_MOV)
665 {
666 /* Used later when the source register is read. */
667 color0_mov = TRUE;
668 }
669 }
670 /* Also drop the MOV marker if the source register is overwritten prior to the shader
671 * end
672 */
673 else if (dst_param.reg.type == WINED3DSPR_TEMP && dst_param.reg.idx == ps->color0_reg)
674 {
675 ps->color0_mov = FALSE;
676 }
677 }
678
679 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
680 if (shader_version.major == 1
681 && (ins.handler_idx == WINED3DSIH_TEX
682 || ins.handler_idx == WINED3DSIH_TEXBEM
683 || ins.handler_idx == WINED3DSIH_TEXBEML
684 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
685 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
686 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
687 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
688 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
689 || ins.handler_idx == WINED3DSIH_TEXREG2AR
690 || ins.handler_idx == WINED3DSIH_TEXREG2GB
691 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
692 {
693 /* Fake sampler usage, only set reserved bit and type. */
694 DWORD sampler_code = dst_param.reg.idx;
695
696 TRACE("Setting fake 2D sampler for 1.x pixelshader.\n");
697 reg_maps->sampler_type[sampler_code] = WINED3DSTT_2D;
698
699 /* texbem is only valid with < 1.4 pixel shaders */
700 if (ins.handler_idx == WINED3DSIH_TEXBEM
701 || ins.handler_idx == WINED3DSIH_TEXBEML)
702 {
703 reg_maps->bumpmat |= 1 << dst_param.reg.idx;
704 if (ins.handler_idx == WINED3DSIH_TEXBEML)
705 {
706 reg_maps->luminanceparams |= 1 << dst_param.reg.idx;
707 }
708 }
709 }
710 else if (ins.handler_idx == WINED3DSIH_BEM)
711 {
712 reg_maps->bumpmat |= 1 << dst_param.reg.idx;
713 }
714 }
715
716 if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
717 else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
718 else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
719 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
720 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
721 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
722 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
723 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
724 else if (ins.handler_idx == WINED3DSIH_RCP) reg_maps->usesrcp = 1;
725
726 limit = ins.src_count + (ins.predicate ? 1 : 0);
727 for (i = 0; i < limit; ++i)
728 {
729 struct wined3d_shader_src_param src_param, src_rel_addr;
730 unsigned int count;
731
732 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
733 count = get_instr_extra_regcount(ins.handler_idx, i);
734
735 shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
736 while (count)
737 {
738 ++src_param.reg.idx;
739 shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
740 --count;
741 }
742
743 if (color0_mov)
744 {
745 IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)shader;
746 if (src_param.reg.type == WINED3DSPR_TEMP
747 && src_param.swizzle == WINED3DSP_NOSWIZZLE)
748 {
749 ps->color0_mov = TRUE;
750 ps->color0_reg = src_param.reg.idx;
751 }
752 }
753 }
754 }
755 }
756 reg_maps->loop_depth = max_loop_depth;
757
758 shader->baseShader.functionLength = ((const char *)ptr - (const char *)byte_code);
759
760 return WINED3D_OK;
761 }
762
763 unsigned int shader_find_free_input_register(const struct shader_reg_maps *reg_maps, unsigned int max)
764 {
765 DWORD map = 1 << max;
766 map |= map - 1;
767 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
768
769 return wined3d_log2i(map);
770 }
771
772 static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
773 const struct wined3d_shader_version *shader_version)
774 {
775 TRACE("dcl");
776
777 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
778 {
779 switch (semantic->sampler_type)
780 {
781 case WINED3DSTT_2D: TRACE("_2d"); break;
782 case WINED3DSTT_CUBE: TRACE("_cube"); break;
783 case WINED3DSTT_VOLUME: TRACE("_volume"); break;
784 default: TRACE("_unknown_ttype(0x%08x)", semantic->sampler_type);
785 }
786 }
787 else
788 {
789 /* Pixel shaders 3.0 don't have usage semantics. */
790 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL) return;
791 else TRACE("_");
792
793 switch (semantic->usage)
794 {
795 case WINED3DDECLUSAGE_POSITION:
796 TRACE("position%u", semantic->usage_idx);
797 break;
798
799 case WINED3DDECLUSAGE_BLENDINDICES:
800 TRACE("blend");
801 break;
802
803 case WINED3DDECLUSAGE_BLENDWEIGHT:
804 TRACE("weight");
805 break;
806
807 case WINED3DDECLUSAGE_NORMAL:
808 TRACE("normal%u", semantic->usage_idx);
809 break;
810
811 case WINED3DDECLUSAGE_PSIZE:
812 TRACE("psize");
813 break;
814
815 case WINED3DDECLUSAGE_COLOR:
816 if (semantic->usage_idx == 0) TRACE("color");
817 else TRACE("specular%u", (semantic->usage_idx - 1));
818 break;
819
820 case WINED3DDECLUSAGE_TEXCOORD:
821 TRACE("texture%u", semantic->usage_idx);
822 break;
823
824 case WINED3DDECLUSAGE_TANGENT:
825 TRACE("tangent");
826 break;
827
828 case WINED3DDECLUSAGE_BINORMAL:
829 TRACE("binormal");
830 break;
831
832 case WINED3DDECLUSAGE_TESSFACTOR:
833 TRACE("tessfactor");
834 break;
835
836 case WINED3DDECLUSAGE_POSITIONT:
837 TRACE("positionT%u", semantic->usage_idx);
838 break;
839
840 case WINED3DDECLUSAGE_FOG:
841 TRACE("fog");
842 break;
843
844 case WINED3DDECLUSAGE_DEPTH:
845 TRACE("depth");
846 break;
847
848 case WINED3DDECLUSAGE_SAMPLE:
849 TRACE("sample");
850 break;
851
852 default:
853 FIXME("unknown_semantics(0x%08x)", semantic->usage);
854 }
855 }
856 }
857
858 static void shader_dump_register(const struct wined3d_shader_register *reg,
859 const struct wined3d_shader_version *shader_version)
860 {
861 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
862 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
863 UINT offset = reg->idx;
864
865 switch (reg->type)
866 {
867 case WINED3DSPR_TEMP:
868 TRACE("r");
869 break;
870
871 case WINED3DSPR_INPUT:
872 TRACE("v");
873 break;
874
875 case WINED3DSPR_CONST:
876 case WINED3DSPR_CONST2:
877 case WINED3DSPR_CONST3:
878 case WINED3DSPR_CONST4:
879 TRACE("c");
880 offset = shader_get_float_offset(reg->type, reg->idx);
881 break;
882
883 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
884 TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
885 break;
886
887 case WINED3DSPR_RASTOUT:
888 TRACE("%s", rastout_reg_names[reg->idx]);
889 break;
890
891 case WINED3DSPR_COLOROUT:
892 TRACE("oC");
893 break;
894
895 case WINED3DSPR_DEPTHOUT:
896 TRACE("oDepth");
897 break;
898
899 case WINED3DSPR_ATTROUT:
900 TRACE("oD");
901 break;
902
903 case WINED3DSPR_TEXCRDOUT:
904 /* Vertex shaders >= 3.0 use general purpose output registers
905 * (WINED3DSPR_OUTPUT), which can include an address token. */
906 if (shader_version->major >= 3) TRACE("o");
907 else TRACE("oT");
908 break;
909
910 case WINED3DSPR_CONSTINT:
911 TRACE("i");
912 break;
913
914 case WINED3DSPR_CONSTBOOL:
915 TRACE("b");
916 break;
917
918 case WINED3DSPR_LABEL:
919 TRACE("l");
920 break;
921
922 case WINED3DSPR_LOOP:
923 TRACE("aL");
924 break;
925
926 case WINED3DSPR_SAMPLER:
927 TRACE("s");
928 break;
929
930 case WINED3DSPR_MISCTYPE:
931 if (reg->idx > 1) FIXME("Unhandled misctype register %u.\n", reg->idx);
932 else TRACE("%s", misctype_reg_names[reg->idx]);
933 break;
934
935 case WINED3DSPR_PREDICATE:
936 TRACE("p");
937 break;
938
939 case WINED3DSPR_IMMCONST:
940 TRACE("l");
941 break;
942
943 case WINED3DSPR_CONSTBUFFER:
944 TRACE("cb");
945 break;
946
947 default:
948 TRACE("unhandled_rtype(%#x)", reg->type);
949 break;
950 }
951
952 if (reg->type == WINED3DSPR_IMMCONST)
953 {
954 TRACE("(");
955 switch (reg->immconst_type)
956 {
957 case WINED3D_IMMCONST_FLOAT:
958 TRACE("%.8e", *(const float *)reg->immconst_data);
959 break;
960
961 case WINED3D_IMMCONST_FLOAT4:
962 TRACE("%.8e, %.8e, %.8e, %.8e",
963 *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
964 *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
965 break;
966
967 default:
968 TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
969 break;
970 }
971 TRACE(")");
972 }
973 else if (reg->type != WINED3DSPR_RASTOUT && reg->type != WINED3DSPR_MISCTYPE)
974 {
975 if (reg->array_idx != ~0U)
976 {
977 TRACE("%u[%u", offset, reg->array_idx);
978 if (reg->rel_addr)
979 {
980 TRACE(" + ");
981 shader_dump_src_param(reg->rel_addr, shader_version);
982 }
983 TRACE("]");
984 }
985 else
986 {
987 if (reg->rel_addr)
988 {
989 TRACE("[");
990 shader_dump_src_param(reg->rel_addr, shader_version);
991 TRACE(" + ");
992 }
993 TRACE("%u", offset);
994 if (reg->rel_addr) TRACE("]");
995 }
996 }
997 }
998
999 void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
1000 const struct wined3d_shader_version *shader_version)
1001 {
1002 DWORD write_mask = param->write_mask;
1003
1004 shader_dump_register(&param->reg, shader_version);
1005
1006 if (write_mask != WINED3DSP_WRITEMASK_ALL)
1007 {
1008 static const char *write_mask_chars = "xyzw";
1009
1010 TRACE(".");
1011 if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
1012 if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
1013 if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
1014 if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
1015 }
1016 }
1017
1018 void shader_dump_src_param(const struct wined3d_shader_src_param *param,
1019 const struct wined3d_shader_version *shader_version)
1020 {
1021 DWORD src_modifier = param->modifiers;
1022 DWORD swizzle = param->swizzle;
1023
1024 if (src_modifier == WINED3DSPSM_NEG
1025 || src_modifier == WINED3DSPSM_BIASNEG
1026 || src_modifier == WINED3DSPSM_SIGNNEG
1027 || src_modifier == WINED3DSPSM_X2NEG
1028 || src_modifier == WINED3DSPSM_ABSNEG)
1029 TRACE("-");
1030 else if (src_modifier == WINED3DSPSM_COMP)
1031 TRACE("1-");
1032 else if (src_modifier == WINED3DSPSM_NOT)
1033 TRACE("!");
1034
1035 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1036 TRACE("abs(");
1037
1038 shader_dump_register(&param->reg, shader_version);
1039
1040 if (src_modifier)
1041 {
1042 switch (src_modifier)
1043 {
1044 case WINED3DSPSM_NONE: break;
1045 case WINED3DSPSM_NEG: break;
1046 case WINED3DSPSM_NOT: break;
1047 case WINED3DSPSM_BIAS: TRACE("_bias"); break;
1048 case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
1049 case WINED3DSPSM_SIGN: TRACE("_bx2"); break;
1050 case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1051 case WINED3DSPSM_COMP: break;
1052 case WINED3DSPSM_X2: TRACE("_x2"); break;
1053 case WINED3DSPSM_X2NEG: TRACE("_x2"); break;
1054 case WINED3DSPSM_DZ: TRACE("_dz"); break;
1055 case WINED3DSPSM_DW: TRACE("_dw"); break;
1056 case WINED3DSPSM_ABSNEG: TRACE(")"); break;
1057 case WINED3DSPSM_ABS: TRACE(")"); break;
1058 default: TRACE("_unknown_modifier(%#x)", src_modifier);
1059 }
1060 }
1061
1062 if (swizzle != WINED3DSP_NOSWIZZLE)
1063 {
1064 static const char *swizzle_chars = "xyzw";
1065 DWORD swizzle_x = swizzle & 0x03;
1066 DWORD swizzle_y = (swizzle >> 2) & 0x03;
1067 DWORD swizzle_z = (swizzle >> 4) & 0x03;
1068 DWORD swizzle_w = (swizzle >> 6) & 0x03;
1069
1070 if (swizzle_x == swizzle_y
1071 && swizzle_x == swizzle_z
1072 && swizzle_x == swizzle_w)
1073 {
1074 TRACE(".%c", swizzle_chars[swizzle_x]);
1075 }
1076 else
1077 {
1078 TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1079 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1080 }
1081 }
1082 }
1083
1084 /* Shared code in order to generate the bulk of the shader string.
1085 * NOTE: A description of how to parse tokens can be found on MSDN. */
1086 void shader_generate_main(IWineD3DBaseShader *iface, struct wined3d_shader_buffer *buffer,
1087 const shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1088 {
1089 IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)iface;
1090 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1091 const struct wined3d_shader_frontend *fe = shader->baseShader.frontend;
1092 void *fe_data = shader->baseShader.frontend_data;
1093 struct wined3d_shader_src_param src_rel_addr[4];
1094 struct wined3d_shader_src_param src_param[4];
1095 struct wined3d_shader_version shader_version;
1096 struct wined3d_shader_src_param dst_rel_addr;
1097 struct wined3d_shader_dst_param dst_param;
1098 struct wined3d_shader_instruction ins;
1099 struct wined3d_shader_context ctx;
1100 const DWORD *ptr = byte_code;
1101 DWORD i;
1102
1103 /* Initialize current parsing state. */
1104 ctx.shader = iface;
1105 ctx.gl_info = &device->adapter->gl_info;
1106 ctx.reg_maps = reg_maps;
1107 ctx.buffer = buffer;
1108 ctx.backend_data = backend_ctx;
1109
1110 ins.ctx = &ctx;
1111 ins.dst = &dst_param;
1112 ins.src = src_param;
1113 shader->baseShader.parse_state.current_row = 0;
1114
1115 fe->shader_read_header(fe_data, &ptr, &shader_version);
1116
1117 while (!fe->shader_is_end(fe_data, &ptr))
1118 {
1119 const char *comment;
1120 UINT comment_size;
1121 UINT param_size;
1122
1123 /* Skip comment tokens. */
1124 fe->shader_read_comment(&ptr, &comment, &comment_size);
1125 if (comment) continue;
1126
1127 /* Read opcode. */
1128 fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1129
1130 /* Unknown opcode and its parameters. */
1131 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1132 {
1133 TRACE("Skipping unrecognized instruction.\n");
1134 ptr += param_size;
1135 continue;
1136 }
1137
1138 /* Nothing to do. */
1139 if (ins.handler_idx == WINED3DSIH_DCL
1140 || ins.handler_idx == WINED3DSIH_NOP
1141 || ins.handler_idx == WINED3DSIH_DEF
1142 || ins.handler_idx == WINED3DSIH_DEFI
1143 || ins.handler_idx == WINED3DSIH_DEFB
1144 || ins.handler_idx == WINED3DSIH_PHASE)
1145 {
1146 ptr += param_size;
1147 continue;
1148 }
1149
1150 /* Destination token */
1151 if (ins.dst_count) fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
1152
1153 /* Predication token */
1154 if (ins.predicate)
1155 {
1156 FIXME("Predicates not implemented.\n");
1157 ins.predicate = *ptr++;
1158 }
1159
1160 /* Other source tokens */
1161 for (i = 0; i < ins.src_count; ++i)
1162 {
1163 fe->shader_read_src_param(fe_data, &ptr, &src_param[i], &src_rel_addr[i]);
1164 }
1165
1166 /* Call appropriate function for output target */
1167 device->shader_backend->shader_handle_instruction(&ins);
1168 }
1169 }
1170
1171 static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1172 {
1173 DWORD mmask = dst->modifiers;
1174
1175 switch (dst->shift)
1176 {
1177 case 0: break;
1178 case 13: TRACE("_d8"); break;
1179 case 14: TRACE("_d4"); break;
1180 case 15: TRACE("_d2"); break;
1181 case 1: TRACE("_x2"); break;
1182 case 2: TRACE("_x4"); break;
1183 case 3: TRACE("_x8"); break;
1184 default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1185 }
1186
1187 if (mmask & WINED3DSPDM_SATURATE) TRACE("_sat");
1188 if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1189 if (mmask & WINED3DSPDM_MSAMPCENTROID) TRACE("_centroid");
1190
1191 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1192 if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1193 }
1194
1195 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1196 {
1197 struct wined3d_shader_version shader_version;
1198 const DWORD *ptr = byte_code;
1199 const char *type_prefix;
1200 DWORD i;
1201
1202 TRACE("Parsing %p.\n", byte_code);
1203
1204 fe->shader_read_header(fe_data, &ptr, &shader_version);
1205
1206 switch (shader_version.type)
1207 {
1208 case WINED3D_SHADER_TYPE_VERTEX:
1209 type_prefix = "vs";
1210 break;
1211
1212 case WINED3D_SHADER_TYPE_GEOMETRY:
1213 type_prefix = "gs";
1214 break;
1215
1216 case WINED3D_SHADER_TYPE_PIXEL:
1217 type_prefix = "ps";
1218 break;
1219
1220 default:
1221 FIXME("Unhandled shader type %#x.\n", shader_version.type);
1222 type_prefix = "unknown";
1223 break;
1224 }
1225
1226 TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1227
1228 while (!fe->shader_is_end(fe_data, &ptr))
1229 {
1230 struct wined3d_shader_instruction ins;
1231 const char *comment;
1232 UINT comment_size;
1233 UINT param_size;
1234
1235 /* comment */
1236 fe->shader_read_comment(&ptr, &comment, &comment_size);
1237 if (comment)
1238 {
1239 if (comment_size > 4 && *(const DWORD *)comment == WINEMAKEFOURCC('T', 'E', 'X', 'T'))
1240 {
1241 const char *end = comment + comment_size;
1242 const char *ptr = comment + 4;
1243 const char *line = ptr;
1244
1245 TRACE("// TEXT\n");
1246 while (ptr != end)
1247 {
1248 if (*ptr == '\n')
1249 {
1250 UINT len = ptr - line;
1251 if (len && *(ptr - 1) == '\r') --len;
1252 TRACE("// %s\n", debugstr_an(line, len));
1253 line = ++ptr;
1254 }
1255 else ++ptr;
1256 }
1257 if (line != ptr) TRACE("// %s\n", debugstr_an(line, ptr - line));
1258 }
1259 else TRACE("// %s\n", debugstr_an(comment, comment_size));
1260 continue;
1261 }
1262
1263 fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1264 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1265 {
1266 TRACE("Skipping unrecognized instruction.\n");
1267 ptr += param_size;
1268 continue;
1269 }
1270
1271 if (ins.handler_idx == WINED3DSIH_DCL)
1272 {
1273 struct wined3d_shader_semantic semantic;
1274
1275 fe->shader_read_semantic(&ptr, &semantic);
1276
1277 shader_dump_decl_usage(&semantic, &shader_version);
1278 shader_dump_ins_modifiers(&semantic.reg);
1279 TRACE(" ");
1280 shader_dump_dst_param(&semantic.reg, &shader_version);
1281 }
1282 else if (ins.handler_idx == WINED3DSIH_DEF)
1283 {
1284 struct wined3d_shader_src_param rel_addr;
1285 struct wined3d_shader_dst_param dst;
1286
1287 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1288
1289 TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(dst.reg.type, dst.reg.idx),
1290 *(const float *)(ptr),
1291 *(const float *)(ptr + 1),
1292 *(const float *)(ptr + 2),
1293 *(const float *)(ptr + 3));
1294 ptr += 4;
1295 }
1296 else if (ins.handler_idx == WINED3DSIH_DEFI)
1297 {
1298 struct wined3d_shader_src_param rel_addr;
1299 struct wined3d_shader_dst_param dst;
1300
1301 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1302
1303 TRACE("defi i%u = %d, %d, %d, %d", dst.reg.idx,
1304 *(ptr),
1305 *(ptr + 1),
1306 *(ptr + 2),
1307 *(ptr + 3));
1308 ptr += 4;
1309 }
1310 else if (ins.handler_idx == WINED3DSIH_DEFB)
1311 {
1312 struct wined3d_shader_src_param rel_addr;
1313 struct wined3d_shader_dst_param dst;
1314
1315 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1316
1317 TRACE("defb b%u = %s", dst.reg.idx, *ptr ? "true" : "false");
1318 ++ptr;
1319 }
1320 else
1321 {
1322 struct wined3d_shader_src_param dst_rel_addr, src_rel_addr;
1323 struct wined3d_shader_dst_param dst_param;
1324 struct wined3d_shader_src_param src_param;
1325
1326 if (ins.dst_count)
1327 {
1328 fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
1329 }
1330
1331 /* Print out predication source token first - it follows
1332 * the destination token. */
1333 if (ins.predicate)
1334 {
1335 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1336 TRACE("(");
1337 shader_dump_src_param(&src_param, &shader_version);
1338 TRACE(") ");
1339 }
1340
1341 /* PixWin marks instructions with the coissue flag with a '+' */
1342 if (ins.coissue) TRACE("+");
1343
1344 TRACE("%s", shader_opcode_names[ins.handler_idx]);
1345
1346 if (ins.handler_idx == WINED3DSIH_IFC
1347 || ins.handler_idx == WINED3DSIH_BREAKC)
1348 {
1349 switch (ins.flags)
1350 {
1351 case COMPARISON_GT: TRACE("_gt"); break;
1352 case COMPARISON_EQ: TRACE("_eq"); break;
1353 case COMPARISON_GE: TRACE("_ge"); break;
1354 case COMPARISON_LT: TRACE("_lt"); break;
1355 case COMPARISON_NE: TRACE("_ne"); break;
1356 case COMPARISON_LE: TRACE("_le"); break;
1357 default: TRACE("_(%u)", ins.flags);
1358 }
1359 }
1360 else if (ins.handler_idx == WINED3DSIH_TEX
1361 && shader_version.major >= 2
1362 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1363 {
1364 TRACE("p");
1365 }
1366
1367 /* We already read the destination token, print it. */
1368 if (ins.dst_count)
1369 {
1370 shader_dump_ins_modifiers(&dst_param);
1371 TRACE(" ");
1372 shader_dump_dst_param(&dst_param, &shader_version);
1373 }
1374
1375 /* Other source tokens */
1376 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1377 {
1378 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1379 TRACE(!i ? " " : ", ");
1380 shader_dump_src_param(&src_param, &shader_version);
1381 }
1382 }
1383 TRACE("\n");
1384 }
1385 }
1386
1387 static void shader_cleanup(IWineD3DBaseShader *iface)
1388 {
1389 IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)iface;
1390
1391 ((IWineD3DDeviceImpl *)shader->baseShader.device)->shader_backend->shader_destroy(iface);
1392 HeapFree(GetProcessHeap(), 0, shader->baseShader.reg_maps.constf);
1393 HeapFree(GetProcessHeap(), 0, shader->baseShader.function);
1394 shader_delete_constant_list(&shader->baseShader.constantsF);
1395 shader_delete_constant_list(&shader->baseShader.constantsB);
1396 shader_delete_constant_list(&shader->baseShader.constantsI);
1397 list_remove(&shader->baseShader.shader_list_entry);
1398
1399 if (shader->baseShader.frontend && shader->baseShader.frontend_data)
1400 {
1401 shader->baseShader.frontend->shader_free(shader->baseShader.frontend_data);
1402 }
1403 }
1404
1405 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1406 static void shader_none_select(const struct wined3d_context *context, BOOL usePS, BOOL useVS) {}
1407 static void shader_none_select_depth_blt(IWineD3DDevice *iface, enum tex_types tex_type, const SIZE *ds_mask_size) {}
1408 static void shader_none_deselect_depth_blt(IWineD3DDevice *iface) {}
1409 static void shader_none_update_float_vertex_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1410 static void shader_none_update_float_pixel_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1411 static void shader_none_load_constants(const struct wined3d_context *context, char usePS, char useVS) {}
1412 static void shader_none_load_np2fixup_constants(IWineD3DDevice *iface, char usePS, char useVS) {}
1413 static void shader_none_destroy(IWineD3DBaseShader *iface) {}
1414 static HRESULT shader_none_alloc(IWineD3DDevice *iface) {return WINED3D_OK;}
1415 static void shader_none_free(IWineD3DDevice *iface) {}
1416 static BOOL shader_none_dirty_const(IWineD3DDevice *iface) {return FALSE;}
1417
1418 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
1419 {
1420 /* Set the shader caps to 0 for the none shader backend */
1421 caps->VertexShaderVersion = 0;
1422 caps->MaxVertexShaderConst = 0;
1423 caps->PixelShaderVersion = 0;
1424 caps->PixelShader1xMaxValue = 0.0f;
1425 caps->MaxPixelShaderConst = 0;
1426 caps->VSClipping = FALSE;
1427 }
1428
1429 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
1430 {
1431 if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1432 {
1433 TRACE("Checking support for fixup:\n");
1434 dump_color_fixup_desc(fixup);
1435 }
1436
1437 /* Faked to make some apps happy. */
1438 if (!is_complex_fixup(fixup))
1439 {
1440 TRACE("[OK]\n");
1441 return TRUE;
1442 }
1443
1444 TRACE("[FAILED]\n");
1445 return FALSE;
1446 }
1447
1448 const shader_backend_t none_shader_backend = {
1449 shader_none_handle_instruction,
1450 shader_none_select,
1451 shader_none_select_depth_blt,
1452 shader_none_deselect_depth_blt,
1453 shader_none_update_float_vertex_constants,
1454 shader_none_update_float_pixel_constants,
1455 shader_none_load_constants,
1456 shader_none_load_np2fixup_constants,
1457 shader_none_destroy,
1458 shader_none_alloc,
1459 shader_none_free,
1460 shader_none_dirty_const,
1461 shader_none_get_caps,
1462 shader_none_color_fixup_supported,
1463 };
1464
1465 static void shader_get_parent(IWineD3DBaseShaderImpl *shader, IUnknown **parent)
1466 {
1467 *parent = shader->baseShader.parent;
1468 IUnknown_AddRef(*parent);
1469 TRACE("shader %p, returning %p.\n", shader, *parent);
1470 }
1471
1472 static HRESULT shader_get_function(IWineD3DBaseShaderImpl *shader, void *data, UINT *data_size)
1473 {
1474 if (!data)
1475 {
1476 *data_size = shader->baseShader.functionLength;
1477 return WINED3D_OK;
1478 }
1479
1480 if (*data_size < shader->baseShader.functionLength)
1481 {
1482 /* MSDN claims (for d3d8 at least) that if *pSizeOfData is smaller
1483 * than the required size we should write the required size and
1484 * return D3DERR_MOREDATA. That's not actually true. */
1485 return WINED3DERR_INVALIDCALL;
1486 }
1487
1488 memcpy(data, shader->baseShader.function, shader->baseShader.functionLength);
1489
1490 return WINED3D_OK;
1491 }
1492
1493 static HRESULT shader_set_function(IWineD3DBaseShaderImpl *shader, const DWORD *byte_code,
1494 const struct wined3d_shader_signature *output_signature, DWORD float_const_count)
1495 {
1496 struct shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
1497 const struct wined3d_shader_frontend *fe;
1498 HRESULT hr;
1499
1500 TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u.\n",
1501 shader, byte_code, output_signature, float_const_count);
1502
1503 fe = shader_select_frontend(*byte_code);
1504 if (!fe)
1505 {
1506 FIXME("Unable to find frontend for shader.\n");
1507 return WINED3DERR_INVALIDCALL;
1508 }
1509 shader->baseShader.frontend = fe;
1510 shader->baseShader.frontend_data = fe->shader_init(byte_code, output_signature);
1511 if (!shader->baseShader.frontend_data)
1512 {
1513 FIXME("Failed to initialize frontend.\n");
1514 return WINED3DERR_INVALIDCALL;
1515 }
1516
1517 /* First pass: trace shader. */
1518 if (TRACE_ON(d3d_shader)) shader_trace_init(fe, shader->baseShader.frontend_data, byte_code);
1519
1520 /* Initialize immediate constant lists. */
1521 list_init(&shader->baseShader.constantsF);
1522 list_init(&shader->baseShader.constantsB);
1523 list_init(&shader->baseShader.constantsI);
1524
1525 /* Second pass: figure out which registers are used, what the semantics are, etc. */
1526 hr = shader_get_registers_used((IWineD3DBaseShader *)shader, fe,
1527 reg_maps, shader->baseShader.input_signature, shader->baseShader.output_signature,
1528 byte_code, float_const_count);
1529 if (FAILED(hr)) return hr;
1530
1531 shader->baseShader.function = HeapAlloc(GetProcessHeap(), 0, shader->baseShader.functionLength);
1532 if (!shader->baseShader.function) return E_OUTOFMEMORY;
1533 memcpy(shader->baseShader.function, byte_code, shader->baseShader.functionLength);
1534
1535 return WINED3D_OK;
1536 }
1537
1538 static HRESULT STDMETHODCALLTYPE vertexshader_QueryInterface(IWineD3DVertexShader *iface, REFIID riid, void **object)
1539 {
1540 TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1541
1542 if (IsEqualGUID(riid, &IID_IWineD3DVertexShader)
1543 || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1544 || IsEqualGUID(riid, &IID_IWineD3DBase)
1545 || IsEqualGUID(riid, &IID_IUnknown))
1546 {
1547 IUnknown_AddRef(iface);
1548 *object = iface;
1549 return S_OK;
1550 }
1551
1552 WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1553
1554 *object = NULL;
1555 return E_NOINTERFACE;
1556 }
1557
1558 static ULONG STDMETHODCALLTYPE vertexshader_AddRef(IWineD3DVertexShader *iface)
1559 {
1560 IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1561 ULONG refcount = InterlockedIncrement(&shader->baseShader.ref);
1562
1563 TRACE("%p increasing refcount to %u.\n", shader, refcount);
1564
1565 return refcount;
1566 }
1567
1568 static ULONG STDMETHODCALLTYPE vertexshader_Release(IWineD3DVertexShader *iface)
1569 {
1570 IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1571 ULONG refcount = InterlockedDecrement(&shader->baseShader.ref);
1572
1573 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1574
1575 if (!refcount)
1576 {
1577 shader_cleanup((IWineD3DBaseShader *)iface);
1578 shader->baseShader.parent_ops->wined3d_object_destroyed(shader->baseShader.parent);
1579 HeapFree(GetProcessHeap(), 0, shader);
1580 }
1581
1582 return refcount;
1583 }
1584
1585 static HRESULT STDMETHODCALLTYPE vertexshader_GetParent(IWineD3DVertexShader *iface, IUnknown **parent)
1586 {
1587 TRACE("iface %p, parent %p.\n", iface, parent);
1588
1589 shader_get_parent((IWineD3DBaseShaderImpl *)iface, parent);
1590
1591 return WINED3D_OK;
1592 }
1593
1594 static HRESULT STDMETHODCALLTYPE vertexshader_GetFunction(IWineD3DVertexShader *iface, void *data, UINT *data_size)
1595 {
1596 TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1597
1598 return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1599 }
1600
1601 /* Set local constants for d3d8 shaders. */
1602 static HRESULT STDMETHODCALLTYPE vertexshader_SetLocalConstantsF(IWineD3DVertexShader *iface,
1603 UINT start_idx, const float *src_data, UINT count)
1604 {
1605 IWineD3DVertexShaderImpl *shader =(IWineD3DVertexShaderImpl *)iface;
1606 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1607 UINT i, end_idx;
1608
1609 TRACE("iface %p, start_idx %u, src_data %p, count %u.\n", iface, start_idx, src_data, count);
1610
1611 end_idx = start_idx + count;
1612 if (end_idx > device->d3d_vshader_constantF)
1613 {
1614 WARN("end_idx %u > float constants limit %u.\n", end_idx, device->d3d_vshader_constantF);
1615 end_idx = device->d3d_vshader_constantF;
1616 }
1617
1618 for (i = start_idx; i < end_idx; ++i)
1619 {
1620 local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
1621 if (!lconst) return E_OUTOFMEMORY;
1622
1623 lconst->idx = i;
1624 memcpy(lconst->value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
1625 list_add_head(&shader->baseShader.constantsF, &lconst->entry);
1626 }
1627
1628 return WINED3D_OK;
1629 }
1630
1631 static const IWineD3DVertexShaderVtbl IWineD3DVertexShader_Vtbl =
1632 {
1633 /* IUnknown methods */
1634 vertexshader_QueryInterface,
1635 vertexshader_AddRef,
1636 vertexshader_Release,
1637 /* IWineD3DBase methods */
1638 vertexshader_GetParent,
1639 /* IWineD3DBaseShader methods */
1640 vertexshader_GetFunction,
1641 /* IWineD3DVertexShader methods */
1642 vertexshader_SetLocalConstantsF,
1643 };
1644
1645 void find_vs_compile_args(IWineD3DVertexShaderImpl *shader,
1646 IWineD3DStateBlockImpl *stateblock, struct vs_compile_args *args)
1647 {
1648 args->fog_src = stateblock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
1649 args->clip_enabled = stateblock->renderState[WINED3DRS_CLIPPING]
1650 && stateblock->renderState[WINED3DRS_CLIPPLANEENABLE];
1651 args->swizzle_map = ((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.swizzle_map;
1652 }
1653
1654 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
1655 {
1656 if (usage_idx1 != usage_idx2) return FALSE;
1657 if (usage1 == usage2) return TRUE;
1658 if (usage1 == WINED3DDECLUSAGE_POSITION && usage2 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1659 if (usage2 == WINED3DDECLUSAGE_POSITION && usage1 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1660
1661 return FALSE;
1662 }
1663
1664 BOOL vshader_get_input(IWineD3DVertexShader *iface, BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
1665 {
1666 IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1667 WORD map = shader->baseShader.reg_maps.input_registers;
1668 unsigned int i;
1669
1670 for (i = 0; map; map >>= 1, ++i)
1671 {
1672 if (!(map & 1)) continue;
1673
1674 if (match_usage(shader->attributes[i].usage,
1675 shader->attributes[i].usage_idx, usage_req, usage_idx_req))
1676 {
1677 *regnum = i;
1678 return TRUE;
1679 }
1680 }
1681 return FALSE;
1682 }
1683
1684 static void vertexshader_set_limits(IWineD3DVertexShaderImpl *shader)
1685 {
1686 DWORD shader_version = WINED3D_SHADER_VERSION(shader->baseShader.reg_maps.shader_version.major,
1687 shader->baseShader.reg_maps.shader_version.minor);
1688 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1689
1690 shader->baseShader.limits.texcoord = 0;
1691 shader->baseShader.limits.attributes = 16;
1692 shader->baseShader.limits.packed_input = 0;
1693
1694 switch (shader_version)
1695 {
1696 case WINED3D_SHADER_VERSION(1, 0):
1697 case WINED3D_SHADER_VERSION(1, 1):
1698 shader->baseShader.limits.temporary = 12;
1699 shader->baseShader.limits.constant_bool = 0;
1700 shader->baseShader.limits.constant_int = 0;
1701 shader->baseShader.limits.address = 1;
1702 shader->baseShader.limits.packed_output = 0;
1703 shader->baseShader.limits.sampler = 0;
1704 shader->baseShader.limits.label = 0;
1705 /* TODO: vs_1_1 has a minimum of 96 constants. What happens when
1706 * a vs_1_1 shader is used on a vs_3_0 capable card that has 256
1707 * constants? */
1708 shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1709 break;
1710
1711 case WINED3D_SHADER_VERSION(2, 0):
1712 case WINED3D_SHADER_VERSION(2, 1):
1713 shader->baseShader.limits.temporary = 12;
1714 shader->baseShader.limits.constant_bool = 16;
1715 shader->baseShader.limits.constant_int = 16;
1716 shader->baseShader.limits.address = 1;
1717 shader->baseShader.limits.packed_output = 0;
1718 shader->baseShader.limits.sampler = 0;
1719 shader->baseShader.limits.label = 16;
1720 shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1721 break;
1722
1723 case WINED3D_SHADER_VERSION(4, 0):
1724 FIXME("Using 3.0 limits for 4.0 shader.\n");
1725 /* Fall through. */
1726
1727 case WINED3D_SHADER_VERSION(3, 0):
1728 shader->baseShader.limits.temporary = 32;
1729 shader->baseShader.limits.constant_bool = 32;
1730 shader->baseShader.limits.constant_int = 32;
1731 shader->baseShader.limits.address = 1;
1732 shader->baseShader.limits.packed_output = 12;
1733 shader->baseShader.limits.sampler = 4;
1734 shader->baseShader.limits.label = 16; /* FIXME: 2048 */
1735 /* DX10 cards on Windows advertise a d3d9 constant limit of 256
1736 * even though they are capable of supporting much more (GL
1737 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
1738 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
1739 * shaders to 256. */
1740 shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1741 break;
1742
1743 default:
1744 shader->baseShader.limits.temporary = 12;
1745 shader->baseShader.limits.constant_bool = 16;
1746 shader->baseShader.limits.constant_int = 16;
1747 shader->baseShader.limits.address = 1;
1748 shader->baseShader.limits.packed_output = 0;
1749 shader->baseShader.limits.sampler = 0;
1750 shader->baseShader.limits.label = 16;
1751 shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1752 FIXME("Unrecognized vertex shader version \"%u.%u\".\n",
1753 shader->baseShader.reg_maps.shader_version.major,
1754 shader->baseShader.reg_maps.shader_version.minor);
1755 }
1756 }
1757
1758 HRESULT vertexshader_init(IWineD3DVertexShaderImpl *shader, IWineD3DDeviceImpl *device,
1759 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1760 IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
1761 {
1762 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
1763 struct shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
1764 unsigned int i;
1765 HRESULT hr;
1766 WORD map;
1767
1768 if (!byte_code) return WINED3DERR_INVALIDCALL;
1769
1770 shader->lpVtbl = &IWineD3DVertexShader_Vtbl;
1771 shader_init(&shader->baseShader, device, parent, parent_ops);
1772
1773 hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code,
1774 output_signature, device->d3d_vshader_constantF);
1775 if (FAILED(hr))
1776 {
1777 WARN("Failed to set function, hr %#x.\n", hr);
1778 shader_cleanup((IWineD3DBaseShader *)shader);
1779 return hr;
1780 }
1781
1782 map = shader->baseShader.reg_maps.input_registers;
1783 for (i = 0; map; map >>= 1, ++i)
1784 {
1785 if (!(map & 1) || !shader->baseShader.input_signature[i].semantic_name) continue;
1786
1787 shader->attributes[i].usage =
1788 shader_usage_from_semantic_name(shader->baseShader.input_signature[i].semantic_name);
1789 shader->attributes[i].usage_idx = shader->baseShader.input_signature[i].semantic_idx;
1790 }
1791
1792 if (output_signature)
1793 {
1794 for (i = 0; i < output_signature->element_count; ++i)
1795 {
1796 struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1797 reg_maps->output_registers |= 1 << e->register_idx;
1798 shader->baseShader.output_signature[e->register_idx] = *e;
1799 }
1800 }
1801
1802 vertexshader_set_limits(shader);
1803
1804 if (device->vs_selected_mode == SHADER_ARB
1805 && (gl_info->quirks & WINED3D_QUIRK_ARB_VS_OFFSET_LIMIT)
1806 && shader->min_rel_offset <= shader->max_rel_offset)
1807 {
1808 if (shader->max_rel_offset - shader->min_rel_offset > 127)
1809 {
1810 FIXME("The difference between the minimum and maximum relative offset is > 127.\n");
1811 FIXME("Which this OpenGL implementation does not support. Try using GLSL.\n");
1812 FIXME("Min: %d, Max: %d.\n", shader->min_rel_offset, shader->max_rel_offset);
1813 }
1814 else if (shader->max_rel_offset - shader->min_rel_offset > 63)
1815 {
1816 shader->rel_offset = shader->min_rel_offset + 63;
1817 }
1818 else if (shader->max_rel_offset > 63)
1819 {
1820 shader->rel_offset = shader->min_rel_offset;
1821 }
1822 else
1823 {
1824 shader->rel_offset = 0;
1825 }
1826 }
1827
1828 shader->baseShader.load_local_constsF = shader->baseShader.reg_maps.usesrelconstF
1829 && !list_empty(&shader->baseShader.constantsF);
1830
1831 return WINED3D_OK;
1832 }
1833
1834 static HRESULT STDMETHODCALLTYPE geometryshader_QueryInterface(IWineD3DGeometryShader *iface,
1835 REFIID riid, void **object)
1836 {
1837 TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1838
1839 if (IsEqualGUID(riid, &IID_IWineD3DGeometryShader)
1840 || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1841 || IsEqualGUID(riid, &IID_IWineD3DBase)
1842 || IsEqualGUID(riid, &IID_IUnknown))
1843 {
1844 IUnknown_AddRef(iface);
1845 *object = iface;
1846 return S_OK;
1847 }
1848
1849 WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1850
1851 *object = NULL;
1852 return E_NOINTERFACE;
1853 }
1854
1855 static ULONG STDMETHODCALLTYPE geometryshader_AddRef(IWineD3DGeometryShader *iface)
1856 {
1857 struct wined3d_geometryshader *shader = (struct wined3d_geometryshader *)iface;
1858 ULONG refcount = InterlockedIncrement(&shader->base_shader.ref);
1859
1860 TRACE("%p increasing refcount to %u.\n", shader, refcount);
1861
1862 return refcount;
1863 }
1864
1865 static ULONG STDMETHODCALLTYPE geometryshader_Release(IWineD3DGeometryShader *iface)
1866 {
1867 struct wined3d_geometryshader *shader = (struct wined3d_geometryshader *)iface;
1868 ULONG refcount = InterlockedDecrement(&shader->base_shader.ref);
1869
1870 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1871
1872 if (!refcount)
1873 {
1874 shader_cleanup((IWineD3DBaseShader *)iface);
1875 shader->base_shader.parent_ops->wined3d_object_destroyed(shader->base_shader.parent);
1876 HeapFree(GetProcessHeap(), 0, shader);
1877 }
1878
1879 return refcount;
1880 }
1881
1882 static HRESULT STDMETHODCALLTYPE geometryshader_GetParent(IWineD3DGeometryShader *iface, IUnknown **parent)
1883 {
1884 TRACE("iface %p, parent %p.\n", iface, parent);
1885
1886 shader_get_parent((IWineD3DBaseShaderImpl *)iface, parent);
1887
1888 return WINED3D_OK;
1889 }
1890
1891 static HRESULT STDMETHODCALLTYPE geometryshader_GetFunction(IWineD3DGeometryShader *iface, void *data, UINT *data_size)
1892 {
1893 TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1894
1895 return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1896 }
1897
1898 static const IWineD3DGeometryShaderVtbl wined3d_geometryshader_vtbl =
1899 {
1900 /* IUnknown methods */
1901 geometryshader_QueryInterface,
1902 geometryshader_AddRef,
1903 geometryshader_Release,
1904 /* IWineD3DBase methods */
1905 geometryshader_GetParent,
1906 /* IWineD3DBaseShader methods */
1907 geometryshader_GetFunction,
1908 };
1909
1910 HRESULT geometryshader_init(struct wined3d_geometryshader *shader, IWineD3DDeviceImpl *device,
1911 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1912 IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
1913 {
1914 HRESULT hr;
1915
1916 shader->vtbl = &wined3d_geometryshader_vtbl;
1917 shader_init(&shader->base_shader, device, parent, parent_ops);
1918
1919 hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code, output_signature, 0);
1920 if (FAILED(hr))
1921 {
1922 WARN("Failed to set function, hr %#x.\n", hr);
1923 shader_cleanup((IWineD3DBaseShader *)shader);
1924 return hr;
1925 }
1926
1927 shader->base_shader.load_local_constsF = FALSE;
1928
1929 return WINED3D_OK;
1930 }
1931
1932 static HRESULT STDMETHODCALLTYPE pixelshader_QueryInterface(IWineD3DPixelShader *iface, REFIID riid, void **object)
1933 {
1934 TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1935
1936 if (IsEqualGUID(riid, &IID_IWineD3DPixelShader)
1937 || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1938 || IsEqualGUID(riid, &IID_IWineD3DBase)
1939 || IsEqualGUID(riid, &IID_IUnknown))
1940 {
1941 IUnknown_AddRef(iface);
1942 *object = iface;
1943 return S_OK;
1944 }
1945
1946 WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1947
1948 *object = NULL;
1949 return E_NOINTERFACE;
1950 }
1951
1952 static ULONG STDMETHODCALLTYPE pixelshader_AddRef(IWineD3DPixelShader *iface)
1953 {
1954 IWineD3DPixelShaderImpl *shader = (IWineD3DPixelShaderImpl *)iface;
1955 ULONG refcount = InterlockedIncrement(&shader->baseShader.ref);
1956
1957 TRACE("%p increasing refcount to %u.\n", shader, refcount);
1958
1959 return refcount;
1960 }
1961
1962 static ULONG STDMETHODCALLTYPE pixelshader_Release(IWineD3DPixelShader *iface)
1963 {
1964 IWineD3DPixelShaderImpl *shader = (IWineD3DPixelShaderImpl *)iface;
1965 ULONG refcount = InterlockedDecrement(&shader->baseShader.ref);
1966
1967 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1968
1969 if (!refcount)
1970 {
1971 shader_cleanup((IWineD3DBaseShader *)iface);
1972 shader->baseShader.parent_ops->wined3d_object_destroyed(shader->baseShader.parent);
1973 HeapFree(GetProcessHeap(), 0, shader);
1974 }
1975
1976 return refcount;
1977 }
1978
1979 static HRESULT STDMETHODCALLTYPE pixelshader_GetParent(IWineD3DPixelShader *iface, IUnknown **parent)
1980 {
1981 TRACE("iface %p, parent %p.\n", iface, parent);
1982
1983 shader_get_parent((IWineD3DBaseShaderImpl *)iface, parent);
1984
1985 return WINED3D_OK;
1986 }
1987
1988 static HRESULT STDMETHODCALLTYPE pixelshader_GetFunction(IWineD3DPixelShader *iface, void *data, UINT *data_size)
1989 {
1990 TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1991
1992 return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1993 }
1994
1995 static const IWineD3DPixelShaderVtbl IWineD3DPixelShader_Vtbl =
1996 {
1997 /* IUnknown methods */
1998 pixelshader_QueryInterface,
1999 pixelshader_AddRef,
2000 pixelshader_Release,
2001 /* IWineD3DBase methods */
2002 pixelshader_GetParent,
2003 /* IWineD3DBaseShader methods */
2004 pixelshader_GetFunction
2005 };
2006
2007 void find_ps_compile_args(IWineD3DPixelShaderImpl *shader,
2008 IWineD3DStateBlockImpl *stateblock, struct ps_compile_args *args)
2009 {
2010 IWineD3DBaseTextureImpl *texture;
2011 UINT i;
2012
2013 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
2014 args->srgb_correction = stateblock->renderState[WINED3DRS_SRGBWRITEENABLE] ? 1 : 0;
2015 args->np2_fixup = 0;
2016
2017 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
2018 {
2019 if (!shader->baseShader.reg_maps.sampler_type[i]) continue;
2020 texture = (IWineD3DBaseTextureImpl *)stateblock->textures[i];
2021 if (!texture)
2022 {
2023 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2024 continue;
2025 }
2026 args->color_fixup[i] = texture->resource.format_desc->color_fixup;
2027
2028 if (texture->resource.format_desc->Flags & WINED3DFMT_FLAG_SHADOW)
2029 args->shadow |= 1 << i;
2030
2031 /* Flag samplers that need NP2 texcoord fixup. */
2032 if (!texture->baseTexture.pow2Matrix_identity)
2033 {
2034 args->np2_fixup |= (1 << i);
2035 }
2036 }
2037 if (shader->baseShader.reg_maps.shader_version.major >= 3)
2038 {
2039 if (((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.position_transformed)
2040 {
2041 args->vp_mode = pretransformed;
2042 }
2043 else if (use_vs(stateblock))
2044 {
2045 args->vp_mode = vertexshader;
2046 }
2047 else
2048 {
2049 args->vp_mode = fixedfunction;
2050 }
2051 args->fog = FOG_OFF;
2052 }
2053 else
2054 {
2055 args->vp_mode = vertexshader;
2056 if (stateblock->renderState[WINED3DRS_FOGENABLE])
2057 {
2058 switch (stateblock->renderState[WINED3DRS_FOGTABLEMODE])
2059 {
2060 case WINED3DFOG_NONE:
2061 if (((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.position_transformed
2062 || use_vs(stateblock))
2063 {
2064 args->fog = FOG_LINEAR;
2065 break;
2066 }
2067
2068 switch (stateblock->renderState[WINED3DRS_FOGVERTEXMODE])
2069 {
2070 case WINED3DFOG_NONE: /* Fall through. */
2071 case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
2072 case WINED3DFOG_EXP: args->fog = FOG_EXP; break;
2073 case WINED3DFOG_EXP2: args->fog = FOG_EXP2; break;
2074 }
2075 break;
2076
2077 case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
2078 case WINED3DFOG_EXP: args->fog = FOG_EXP; break;
2079 case WINED3DFOG_EXP2: args->fog = FOG_EXP2; break;
2080 }
2081 }
2082 else
2083 {
2084 args->fog = FOG_OFF;
2085 }
2086 }
2087 }
2088
2089 static void pixelshader_set_limits(IWineD3DPixelShaderImpl *shader)
2090 {
2091 DWORD shader_version = WINED3D_SHADER_VERSION(shader->baseShader.reg_maps.shader_version.major,
2092 shader->baseShader.reg_maps.shader_version.minor);
2093
2094 shader->baseShader.limits.attributes = 0;
2095 shader->baseShader.limits.address = 0;
2096 shader->baseShader.limits.packed_output = 0;
2097
2098 switch (shader_version)
2099 {
2100 case WINED3D_SHADER_VERSION(1, 0):
2101 case WINED3D_SHADER_VERSION(1, 1):
2102 case WINED3D_SHADER_VERSION(1, 2):
2103 case WINED3D_SHADER_VERSION(1, 3):
2104 shader->baseShader.limits.temporary = 2;
2105 shader->baseShader.limits.constant_float = 8;
2106 shader->baseShader.limits.constant_int = 0;
2107 shader->baseShader.limits.constant_bool = 0;
2108 shader->baseShader.limits.texcoord = 4;
2109 shader->baseShader.limits.sampler = 4;
2110 shader->baseShader.limits.packed_input = 0;
2111 shader->baseShader.limits.label = 0;
2112 break;
2113
2114 case WINED3D_SHADER_VERSION(1, 4):
2115 shader->baseShader.limits.temporary = 6;
2116 shader->baseShader.limits.constant_float = 8;
2117 shader->baseShader.limits.constant_int = 0;
2118 shader->baseShader.limits.constant_bool = 0;
2119 shader->baseShader.limits.texcoord = 6;
2120 shader->baseShader.limits.sampler = 6;
2121 shader->baseShader.limits.packed_input = 0;
2122 shader->baseShader.limits.label = 0;
2123 break;
2124
2125 /* FIXME: Temporaries must match D3DPSHADERCAPS2_0.NumTemps. */
2126 case WINED3D_SHADER_VERSION(2, 0):
2127 shader->baseShader.limits.temporary = 32;
2128 shader->baseShader.limits.constant_float = 32;
2129 shader->baseShader.limits.constant_int = 16;
2130 shader->baseShader.limits.constant_bool = 16;
2131 shader->baseShader.limits.texcoord = 8;
2132 shader->baseShader.limits.sampler = 16;
2133 shader->baseShader.limits.packed_input = 0;
2134 break;
2135
2136 case WINED3D_SHADER_VERSION(2, 1):
2137 shader->baseShader.limits.temporary = 32;
2138 shader->baseShader.limits.constant_float = 32;
2139 shader->baseShader.limits.constant_int = 16;
2140 shader->baseShader.limits.constant_bool = 16;
2141 shader->baseShader.limits.texcoord = 8;
2142 shader->baseShader.limits.sampler = 16;
2143 shader->baseShader.limits.packed_input = 0;
2144 shader->baseShader.limits.label = 16;
2145 break;
2146
2147 case WINED3D_SHADER_VERSION(4, 0):
2148 FIXME("Using 3.0 limits for 4.0 shader.\n");
2149 /* Fall through. */
2150
2151 case WINED3D_SHADER_VERSION(3, 0):
2152 shader->baseShader.limits.temporary = 32;
2153 shader->baseShader.limits.constant_float = 224;
2154 shader->baseShader.limits.constant_int = 16;
2155 shader->baseShader.limits.constant_bool = 16;
2156 shader->baseShader.limits.texcoord = 0;
2157 shader->baseShader.limits.sampler = 16;
2158 shader->baseShader.limits.packed_input = 12;
2159 shader->baseShader.limits.label = 16; /* FIXME: 2048 */
2160 break;
2161
2162 default:
2163 shader->baseShader.limits.temporary = 32;
2164 shader->baseShader.limits.constant_float = 32;
2165 shader->baseShader.limits.constant_int = 16;
2166 shader->baseShader.limits.constant_bool = 16;
2167 shader->baseShader.limits.texcoord = 8;
2168 shader->baseShader.limits.sampler = 16;
2169 shader->baseShader.limits.packed_input = 0;
2170 shader->baseShader.limits.label = 0;
2171 FIXME("Unrecognized pixel shader version %u.%u\n",
2172 shader->baseShader.reg_maps.shader_version.major,
2173 shader->baseShader.reg_maps.shader_version.minor);
2174 }
2175 }
2176
2177 HRESULT pixelshader_init(IWineD3DPixelShaderImpl *shader, IWineD3DDeviceImpl *device,
2178 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2179 IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
2180 {
2181 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2182 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2183 HRESULT hr;
2184
2185 if (!byte_code) return WINED3DERR_INVALIDCALL;
2186
2187 shader->lpVtbl = &IWineD3DPixelShader_Vtbl;
2188 shader_init(&shader->baseShader, device, parent, parent_ops);
2189
2190 hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code,
2191 output_signature, device->d3d_pshader_constantF);
2192 if (FAILED(hr))
2193 {
2194 WARN("Failed to set function, hr %#x.\n", hr);
2195 shader_cleanup((IWineD3DBaseShader *)shader);
2196 return hr;
2197 }
2198
2199 pixelshader_set_limits(shader);
2200
2201 for (i = 0; i < MAX_REG_INPUT; ++i)
2202 {
2203 if (shader->input_reg_used[i])
2204 {
2205 ++num_regs_used;
2206 highest_reg_used = i;
2207 }
2208 }
2209
2210 /* Don't do any register mapping magic if it is not needed, or if we can't
2211 * achieve anything anyway */
2212 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2213 || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2214 {
2215 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2216 {
2217 /* This happens with relative addressing. The input mapper function
2218 * warns about this if the higher registers are declared too, so
2219 * don't write a FIXME here */
2220 WARN("More varying registers used than supported\n");
2221 }
2222
2223 for (i = 0; i < MAX_REG_INPUT; ++i)
2224 {
2225 shader->input_reg_map[i] = i;
2226 }
2227
2228 shader->declared_in_count = highest_reg_used + 1;
2229 }
2230 else
2231 {
2232 shader->declared_in_count = 0;
2233 for (i = 0; i < MAX_REG_INPUT; ++i)
2234 {
2235 if (shader->input_reg_used[i]) shader->input_reg_map[i] = shader->declared_in_count++;
2236 else shader->input_reg_map[i] = ~0U;
2237 }
2238 }
2239
2240 shader->baseShader.load_local_constsF = FALSE;
2241
2242 return WINED3D_OK;
2243 }
2244
2245 void pixelshader_update_samplers(struct shader_reg_maps *reg_maps, IWineD3DBaseTexture * const *textures)
2246 {
2247 WINED3DSAMPLER_TEXTURE_TYPE *sampler_type = reg_maps->sampler_type;
2248 unsigned int i;
2249
2250 if (reg_maps->shader_version.major != 1) return;
2251
2252 for (i = 0; i < max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS); ++i)
2253 {
2254 /* We don't sample from this sampler. */
2255 if (!sampler_type[i]) continue;
2256
2257 if (!textures[i])
2258 {
2259 WARN("No texture bound to sampler %u, using 2D.\n", i);
2260 sampler_type[i] = WINED3DSTT_2D;
2261 continue;
2262 }
2263
2264 switch (IWineD3DBaseTexture_GetTextureDimensions(textures[i]))
2265 {
2266 case GL_TEXTURE_RECTANGLE_ARB:
2267 case GL_TEXTURE_2D:
2268 /* We have to select between texture rectangles and 2D
2269 * textures later because 2.0 and 3.0 shaders only have
2270 * WINED3DSTT_2D as well. */
2271 sampler_type[i] = WINED3DSTT_2D;
2272 break;
2273
2274 case GL_TEXTURE_3D:
2275 sampler_type[i] = WINED3DSTT_VOLUME;
2276 break;
2277
2278 case GL_TEXTURE_CUBE_MAP_ARB:
2279 sampler_type[i] = WINED3DSTT_CUBE;
2280 break;
2281
2282 default:
2283 FIXME("Unrecognized texture type %#x, using 2D.\n",
2284 IWineD3DBaseTexture_GetTextureDimensions(textures[i]));
2285 sampler_type[i] = WINED3DSTT_2D;
2286 }
2287 }
2288 }
A free Windows-compatible Operating System