?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
wined3d: Add support for SM4 relative addressing.
[wine.git] / dlls / wined3d / shader.c
blobd7e905853f5c9c292ed781daf5a1839e1caca2e0
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_DCL_CONSTANT_BUFFER */ "dcl_constantBuffer",
54 /* WINED3DSIH_DCL_INPUT_PRIMITIVE */ "dcl_inputPrimitive",
55 /* WINED3DSIH_DCL_OUTPUT_TOPOLOGY */ "dcl_outputTopology",
56 /* WINED3DSIH_DCL_VERTICES_OUT */ "dcl_maxOutputVertexCount",
57 /* WINED3DSIH_DEF */ "def",
58 /* WINED3DSIH_DEFB */ "defb",
59 /* WINED3DSIH_DEFI */ "defi",
60 /* WINED3DSIH_DIV */ "div",
61 /* WINED3DSIH_DP2ADD */ "dp2add",
62 /* WINED3DSIH_DP3 */ "dp3",
63 /* WINED3DSIH_DP4 */ "dp4",
64 /* WINED3DSIH_DST */ "dst",
65 /* WINED3DSIH_DSX */ "dsx",
66 /* WINED3DSIH_DSY */ "dsy",
67 /* WINED3DSIH_ELSE */ "else",
68 /* WINED3DSIH_EMIT */ "emit",
69 /* WINED3DSIH_ENDIF */ "endif",
70 /* WINED3DSIH_ENDLOOP */ "endloop",
71 /* WINED3DSIH_ENDREP */ "endrep",
72 /* WINED3DSIH_EQ */ "eq",
73 /* WINED3DSIH_EXP */ "exp",
74 /* WINED3DSIH_EXPP */ "expp",
75 /* WINED3DSIH_FRC */ "frc",
76 /* WINED3DSIH_FTOI */ "ftoi",
77 /* WINED3DSIH_GE */ "ge",
78 /* WINED3DSIH_IADD */ "iadd",
79 /* WINED3DSIH_IEQ */ "ieq",
80 /* WINED3DSIH_IF */ "if",
81 /* WINED3DSIH_IFC */ "ifc",
82 /* WINED3DSIH_IGE */ "ige",
83 /* WINED3DSIH_IMUL */ "imul",
84 /* WINED3DSIH_ITOF */ "itof",
85 /* WINED3DSIH_LABEL */ "label",
86 /* WINED3DSIH_LD */ "ld",
87 /* WINED3DSIH_LIT */ "lit",
88 /* WINED3DSIH_LOG */ "log",
89 /* WINED3DSIH_LOGP */ "logp",
90 /* WINED3DSIH_LOOP */ "loop",
91 /* WINED3DSIH_LRP */ "lrp",
92 /* WINED3DSIH_LT */ "lt",
93 /* WINED3DSIH_M3x2 */ "m3x2",
94 /* WINED3DSIH_M3x3 */ "m3x3",
95 /* WINED3DSIH_M3x4 */ "m3x4",
96 /* WINED3DSIH_M4x3 */ "m4x3",
97 /* WINED3DSIH_M4x4 */ "m4x4",
98 /* WINED3DSIH_MAD */ "mad",
99 /* WINED3DSIH_MAX */ "max",
100 /* WINED3DSIH_MIN */ "min",
101 /* WINED3DSIH_MOV */ "mov",
102 /* WINED3DSIH_MOVA */ "mova",
103 /* WINED3DSIH_MOVC */ "movc",
104 /* WINED3DSIH_MUL */ "mul",
105 /* WINED3DSIH_NOP */ "nop",
106 /* WINED3DSIH_NRM */ "nrm",
107 /* WINED3DSIH_PHASE */ "phase",
108 /* WINED3DSIH_POW */ "pow",
109 /* WINED3DSIH_RCP */ "rcp",
110 /* WINED3DSIH_REP */ "rep",
111 /* WINED3DSIH_RET */ "ret",
112 /* WINED3DSIH_ROUND_NI */ "round_ni",
113 /* WINED3DSIH_RSQ */ "rsq",
114 /* WINED3DSIH_SAMPLE */ "sample",
115 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
116 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
117 /* WINED3DSIH_SETP */ "setp",
118 /* WINED3DSIH_SGE */ "sge",
119 /* WINED3DSIH_SGN */ "sgn",
120 /* WINED3DSIH_SINCOS */ "sincos",
121 /* WINED3DSIH_SLT */ "slt",
122 /* WINED3DSIH_SQRT */ "sqrt",
123 /* WINED3DSIH_SUB */ "sub",
124 /* WINED3DSIH_TEX */ "texld",
125 /* WINED3DSIH_TEXBEM */ "texbem",
126 /* WINED3DSIH_TEXBEML */ "texbeml",
127 /* WINED3DSIH_TEXCOORD */ "texcrd",
128 /* WINED3DSIH_TEXDEPTH */ "texdepth",
129 /* WINED3DSIH_TEXDP3 */ "texdp3",
130 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
131 /* WINED3DSIH_TEXKILL */ "texkill",
132 /* WINED3DSIH_TEXLDD */ "texldd",
133 /* WINED3DSIH_TEXLDL */ "texldl",
134 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
135 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
136 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
137 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
138 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
139 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
140 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
141 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
142 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
143 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
144 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
145 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
146 /* WINED3DSIH_UDIV */ "udiv",
147 /* WINED3DSIH_USHR */ "ushr",
148 /* WINED3DSIH_UTOF */ "utof",
149 /* WINED3DSIH_XOR */ "xor",
150 };
151
152 static const char * const semantic_names[] =
153 {
154 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
155 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
156 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
157 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
158 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
159 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
160 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
161 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
162 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
163 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
164 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
165 /* WINED3D_DECL_USAGE_FOG */ "FOG",
166 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
167 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
168 };
169
170 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
171 {
172 if (usage >= sizeof(semantic_names) / sizeof(*semantic_names))
173 {
174 FIXME("Unrecognized usage %#x.\n", usage);
175 return "UNRECOGNIZED";
176 }
177
178 return semantic_names[usage];
179 }
180
181 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
182 {
183 unsigned int i;
184
185 for (i = 0; i < sizeof(semantic_names) / sizeof(*semantic_names); ++i)
186 {
187 if (!strcmp(name, semantic_names[i])) return i;
188 }
189
190 return ~0U;
191 }
192
193 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
194 {
195 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
196 }
197
198 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
199 const struct wined3d_shader_semantic *s)
200 {
201 e->semantic_name = shader_semantic_name_from_usage(s->usage);
202 e->semantic_idx = s->usage_idx;
203 e->sysval_semantic = 0;
204 e->component_type = 0;
205 e->register_idx = s->reg.reg.idx;
206 e->mask = s->reg.write_mask;
207 }
208
209 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
210 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
211 {
212 e->semantic_name = shader_semantic_name_from_usage(usage);
213 e->semantic_idx = usage_idx;
214 e->sysval_semantic = 0;
215 e->component_type = 0;
216 e->register_idx = reg_idx;
217 e->mask = write_mask;
218 }
219
220 static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
221 {
222 switch (version_token >> 16)
223 {
224 case WINED3D_SM1_VS:
225 case WINED3D_SM1_PS:
226 return &sm1_shader_frontend;
227
228 case WINED3D_SM4_PS:
229 case WINED3D_SM4_VS:
230 case WINED3D_SM4_GS:
231 return &sm4_shader_frontend;
232
233 default:
234 FIXME("Unrecognised version token %#x\n", version_token);
235 return NULL;
236 }
237 }
238
239 void shader_buffer_clear(struct wined3d_shader_buffer *buffer)
240 {
241 buffer->buffer[0] = '\0';
242 buffer->bsize = 0;
243 buffer->lineNo = 0;
244 buffer->newline = TRUE;
245 }
246
247 BOOL shader_buffer_init(struct wined3d_shader_buffer *buffer)
248 {
249 buffer->buffer = HeapAlloc(GetProcessHeap(), 0, SHADER_PGMSIZE);
250 if (!buffer->buffer)
251 {
252 ERR("Failed to allocate shader buffer memory.\n");
253 return FALSE;
254 }
255
256 shader_buffer_clear(buffer);
257 return TRUE;
258 }
259
260 void shader_buffer_free(struct wined3d_shader_buffer *buffer)
261 {
262 HeapFree(GetProcessHeap(), 0, buffer->buffer);
263 }
264
265 int shader_vaddline(struct wined3d_shader_buffer *buffer, const char *format, va_list args)
266 {
267 char *base = buffer->buffer + buffer->bsize;
268 int rc;
269
270 rc = vsnprintf(base, SHADER_PGMSIZE - 1 - buffer->bsize, format, args);
271
272 if (rc < 0 /* C89 */ || (unsigned int)rc > SHADER_PGMSIZE - 1 - buffer->bsize /* C99 */)
273 {
274 ERR("The buffer allocated for the shader program string "
275 "is too small at %d bytes.\n", SHADER_PGMSIZE);
276 buffer->bsize = SHADER_PGMSIZE - 1;
277 return -1;
278 }
279
280 if (buffer->newline)
281 {
282 TRACE("GL HW (%u, %u) : %s", buffer->lineNo + 1, buffer->bsize, base);
283 buffer->newline = FALSE;
284 }
285 else
286 {
287 TRACE("%s", base);
288 }
289
290 buffer->bsize += rc;
291 if (buffer->buffer[buffer->bsize-1] == '\n')
292 {
293 ++buffer->lineNo;
294 buffer->newline = TRUE;
295 }
296
297 return 0;
298 }
299
300 int shader_addline(struct wined3d_shader_buffer *buffer, const char *format, ...)
301 {
302 va_list args;
303 int ret;
304
305 va_start(args, format);
306 ret = shader_vaddline(buffer, format, args);
307 va_end(args);
308
309 return ret;
310 }
311
312 static void shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
313 void *parent, const struct wined3d_parent_ops *parent_ops)
314 {
315 shader->ref = 1;
316 shader->device = device;
317 shader->parent = parent;
318 shader->parent_ops = parent_ops;
319 list_init(&shader->linked_programs);
320 list_add_head(&device->shaders, &shader->shader_list_entry);
321 }
322
323 /* Convert floating point offset relative to a register file to an absolute
324 * offset for float constants. */
325 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
326 {
327 switch (register_type)
328 {
329 case WINED3DSPR_CONST: return register_idx;
330 case WINED3DSPR_CONST2: return 2048 + register_idx;
331 case WINED3DSPR_CONST3: return 4096 + register_idx;
332 case WINED3DSPR_CONST4: return 6144 + register_idx;
333 default:
334 FIXME("Unsupported register type: %u.\n", register_type);
335 return register_idx;
336 }
337 }
338
339 static void shader_delete_constant_list(struct list *clist)
340 {
341 struct wined3d_shader_lconst *constant;
342 struct list *ptr;
343
344 ptr = list_head(clist);
345 while (ptr)
346 {
347 constant = LIST_ENTRY(ptr, struct wined3d_shader_lconst, entry);
348 ptr = list_next(clist, ptr);
349 HeapFree(GetProcessHeap(), 0, constant);
350 }
351 list_init(clist);
352 }
353
354 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
355 {
356 DWORD idx, shift;
357 idx = bit >> 5;
358 shift = bit & 0x1f;
359 bitmap[idx] |= (1 << shift);
360 }
361
362 static void shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
363 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type)
364 {
365 switch (reg->type)
366 {
367 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
368 if (shader_type == WINED3D_SHADER_TYPE_PIXEL) reg_maps->texcoord |= 1 << reg->idx;
369 else reg_maps->address |= 1 << reg->idx;
370 break;
371
372 case WINED3DSPR_TEMP:
373 reg_maps->temporary |= 1 << reg->idx;
374 break;
375
376 case WINED3DSPR_INPUT:
377 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
378 {
379 if (reg->rel_addr)
380 {
381 /* If relative addressing is used, we must assume that all registers
382 * are used. Even if it is a construct like v3[aL], we can't assume
383 * that v0, v1 and v2 aren't read because aL can be negative */
384 unsigned int i;
385 for (i = 0; i < MAX_REG_INPUT; ++i)
386 {
387 shader->u.ps.input_reg_used[i] = TRUE;
388 }
389 }
390 else
391 {
392 shader->u.ps.input_reg_used[reg->idx] = TRUE;
393 }
394 }
395 else reg_maps->input_registers |= 1 << reg->idx;
396 break;
397
398 case WINED3DSPR_RASTOUT:
399 if (reg->idx == 1) reg_maps->fog = 1;
400 break;
401
402 case WINED3DSPR_MISCTYPE:
403 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
404 {
405 if (!reg->idx) reg_maps->vpos = 1;
406 else if (reg->idx == 1) reg_maps->usesfacing = 1;
407 }
408 break;
409
410 case WINED3DSPR_CONST:
411 if (reg->rel_addr)
412 {
413 if (reg->idx < reg_maps->min_rel_offset) reg_maps->min_rel_offset = reg->idx;
414 if (reg->idx > reg_maps->max_rel_offset) reg_maps->max_rel_offset = reg->idx;
415 reg_maps->usesrelconstF = TRUE;
416 }
417 else
418 {
419 set_bitmap_bit(reg_maps->constf, reg->idx);
420 }
421 break;
422
423 case WINED3DSPR_CONSTINT:
424 reg_maps->integer_constants |= (1 << reg->idx);
425 break;
426
427 case WINED3DSPR_CONSTBOOL:
428 reg_maps->boolean_constants |= (1 << reg->idx);
429 break;
430
431 case WINED3DSPR_COLOROUT:
432 reg_maps->rt_mask |= (1 << reg->idx);
433 break;
434
435 default:
436 TRACE("Not recording register of type %#x and idx %u\n", reg->type, reg->idx);
437 break;
438 }
439 }
440
441 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
442 {
443 switch (instr)
444 {
445 case WINED3DSIH_M4x4:
446 case WINED3DSIH_M3x4:
447 return param == 1 ? 3 : 0;
448
449 case WINED3DSIH_M4x3:
450 case WINED3DSIH_M3x3:
451 return param == 1 ? 2 : 0;
452
453 case WINED3DSIH_M3x2:
454 return param == 1 ? 1 : 0;
455
456 default:
457 return 0;
458 }
459 }
460
461 /* Note that this does not count the loop register as an address register. */
462 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
463 struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature_element *input_signature,
464 struct wined3d_shader_signature_element *output_signature, const DWORD *byte_code, DWORD constf_size)
465 {
466 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
467 void *fe_data = shader->frontend_data;
468 struct wined3d_shader_version shader_version;
469 const DWORD *ptr = byte_code;
470
471 memset(reg_maps, 0, sizeof(*reg_maps));
472 reg_maps->min_rel_offset = ~0U;
473
474 fe->shader_read_header(fe_data, &ptr, &shader_version);
475 reg_maps->shader_version = shader_version;
476
477 reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
478 sizeof(*reg_maps->constf) * ((constf_size + 31) / 32));
479 if (!reg_maps->constf)
480 {
481 ERR("Failed to allocate constant map memory.\n");
482 return E_OUTOFMEMORY;
483 }
484
485 while (!fe->shader_is_end(fe_data, &ptr))
486 {
487 struct wined3d_shader_instruction ins;
488 const char *comment;
489 UINT comment_size;
490
491 /* Skip comments. */
492 fe->shader_read_comment(&ptr, &comment, &comment_size);
493 if (comment) continue;
494
495 /* Fetch opcode. */
496 fe->shader_read_instruction(fe_data, &ptr, &ins);
497
498 /* Unhandled opcode, and its parameters. */
499 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
500 {
501 TRACE("Skipping unrecognized instruction.\n");
502 continue;
503 }
504
505 /* Handle declarations. */
506 if (ins.handler_idx == WINED3DSIH_DCL)
507 {
508 struct wined3d_shader_semantic *semantic = &ins.declaration.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_DCL_CONSTANT_BUFFER)
537 {
538 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
539 if (reg->idx >= WINED3D_MAX_CBS)
540 ERR("Invalid CB index %u.\n", reg->idx);
541 else
542 reg_maps->cb_sizes[reg->idx] = reg->array_idx;
543 }
544 else if (ins.handler_idx == WINED3DSIH_DEF)
545 {
546 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
547 if (!lconst) return E_OUTOFMEMORY;
548
549 lconst->idx = ins.dst[0].reg.idx;
550 memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
551
552 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
553 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
554 {
555 float *value = (float *)lconst->value;
556 if (value[0] < -1.0f) value[0] = -1.0f;
557 else if (value[0] > 1.0f) value[0] = 1.0f;
558 if (value[1] < -1.0f) value[1] = -1.0f;
559 else if (value[1] > 1.0f) value[1] = 1.0f;
560 if (value[2] < -1.0f) value[2] = -1.0f;
561 else if (value[2] > 1.0f) value[2] = 1.0f;
562 if (value[3] < -1.0f) value[3] = -1.0f;
563 else if (value[3] > 1.0f) value[3] = 1.0f;
564 }
565
566 list_add_head(&shader->constantsF, &lconst->entry);
567 }
568 else if (ins.handler_idx == WINED3DSIH_DEFI)
569 {
570 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
571 if (!lconst) return E_OUTOFMEMORY;
572
573 lconst->idx = ins.dst[0].reg.idx;
574 memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
575
576 list_add_head(&shader->constantsI, &lconst->entry);
577 reg_maps->local_int_consts |= (1 << lconst->idx);
578 }
579 else if (ins.handler_idx == WINED3DSIH_DEFB)
580 {
581 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
582 if (!lconst) return E_OUTOFMEMORY;
583
584 lconst->idx = ins.dst[0].reg.idx;
585 memcpy(lconst->value, ins.src[0].reg.immconst_data, sizeof(DWORD));
586
587 list_add_head(&shader->constantsB, &lconst->entry);
588 reg_maps->local_bool_consts |= (1 << lconst->idx);
589 }
590 /* For subroutine prototypes. */
591 else if (ins.handler_idx == WINED3DSIH_LABEL)
592 {
593 reg_maps->labels |= 1 << ins.src[0].reg.idx;
594 }
595 /* Set texture, address, temporary registers. */
596 else
597 {
598 BOOL color0_mov = FALSE;
599 unsigned int i;
600
601 /* This will loop over all the registers and try to
602 * make a bitmask of the ones we're interested in.
603 *
604 * Relative addressing tokens are ignored, but that's
605 * okay, since we'll catch any address registers when
606 * they are initialized (required by spec). */
607 for (i = 0; i < ins.dst_count; ++i)
608 {
609 shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg, shader_version.type);
610
611 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and
612 * is used in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel
613 * shaders because TECRDOUT isn't used in them, but future register types might cause issues */
614 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX && shader_version.major < 3)
615 {
616 UINT idx = ins.dst[i].reg.idx;
617
618 switch (ins.dst[i].reg.type)
619 {
620 case WINED3DSPR_RASTOUT:
621 switch (idx)
622 {
623 case 0: /* oPos */
624 reg_maps->output_registers |= 1 << 10;
625 shader_signature_from_usage(&output_signature[10],
626 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
627 break;
628
629 case 1: /* oFog */
630 reg_maps->output_registers |= 1 << 11;
631 shader_signature_from_usage(&output_signature[11],
632 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
633 break;
634
635 case 2: /* oPts */
636 reg_maps->output_registers |= 1 << 11;
637 shader_signature_from_usage(&output_signature[11],
638 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
639 break;
640 }
641 break;
642
643 case WINED3DSPR_ATTROUT:
644 if (idx < 2)
645 {
646 idx += 8;
647 if (reg_maps->output_registers & (1 << idx))
648 {
649 output_signature[idx].mask |= ins.dst[i].write_mask;
650 }
651 else
652 {
653 reg_maps->output_registers |= 1 << idx;
654 shader_signature_from_usage(&output_signature[idx],
655 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
656 }
657 }
658 break;
659
660 case WINED3DSPR_TEXCRDOUT:
661
662 reg_maps->texcoord_mask[idx] |= ins.dst[i].write_mask;
663 if (reg_maps->output_registers & (1 << idx))
664 {
665 output_signature[idx].mask |= ins.dst[i].write_mask;
666 }
667 else
668 {
669 reg_maps->output_registers |= 1 << idx;
670 shader_signature_from_usage(&output_signature[idx],
671 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
672 }
673 break;
674
675 default:
676 break;
677 }
678 }
679
680 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
681 {
682 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx)
683 {
684 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
685 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
686 * the mov and perform the sRGB write correction from the source register.
687 *
688 * However, if the mov is only partial, we can't do this, and if the write
689 * comes from an instruction other than MOV it is hard to do as well. If
690 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
691 shader->u.ps.color0_mov = FALSE;
692 if (ins.handler_idx == WINED3DSIH_MOV
693 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
694 {
695 /* Used later when the source register is read. */
696 color0_mov = TRUE;
697 }
698 }
699 /* Also drop the MOV marker if the source register is overwritten prior to the shader
700 * end
701 */
702 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
703 && ins.dst[i].reg.idx == shader->u.ps.color0_reg)
704 {
705 shader->u.ps.color0_mov = FALSE;
706 }
707 }
708
709 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
710 if (shader_version.major == 1
711 && (ins.handler_idx == WINED3DSIH_TEX
712 || ins.handler_idx == WINED3DSIH_TEXBEM
713 || ins.handler_idx == WINED3DSIH_TEXBEML
714 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
715 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
716 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
717 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
718 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
719 || ins.handler_idx == WINED3DSIH_TEXREG2AR
720 || ins.handler_idx == WINED3DSIH_TEXREG2GB
721 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
722 {
723 /* Fake sampler usage, only set reserved bit and type. */
724 DWORD sampler_code = ins.dst[i].reg.idx;
725
726 TRACE("Setting fake 2D sampler for 1.x pixelshader.\n");
727 reg_maps->sampler_type[sampler_code] = WINED3DSTT_2D;
728
729 /* texbem is only valid with < 1.4 pixel shaders */
730 if (ins.handler_idx == WINED3DSIH_TEXBEM
731 || ins.handler_idx == WINED3DSIH_TEXBEML)
732 {
733 reg_maps->bumpmat |= 1 << ins.dst[i].reg.idx;
734 if (ins.handler_idx == WINED3DSIH_TEXBEML)
735 {
736 reg_maps->luminanceparams |= 1 << ins.dst[i].reg.idx;
737 }
738 }
739 }
740 else if (ins.handler_idx == WINED3DSIH_BEM)
741 {
742 reg_maps->bumpmat |= 1 << ins.dst[i].reg.idx;
743 }
744 }
745
746 if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
747 else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
748 else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
749 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
750 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
751 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
752 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
753 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
754 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
755 else if (ins.handler_idx == WINED3DSIH_LOOP
756 || ins.handler_idx == WINED3DSIH_REP)
757 {
758 ++cur_loop_depth;
759 if (cur_loop_depth > max_loop_depth)
760 max_loop_depth = cur_loop_depth;
761 }
762 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
763 || ins.handler_idx == WINED3DSIH_ENDREP)
764 --cur_loop_depth;
765
766 if (ins.predicate)
767 shader_record_register_usage(shader, reg_maps, &ins.predicate->reg, shader_version.type);
768
769 for (i = 0; i < ins.src_count; ++i)
770 {
771 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
772 struct wined3d_shader_register reg = ins.src[i].reg;
773
774 shader_record_register_usage(shader, reg_maps, &ins.src[i].reg, shader_version.type);
775 while (count)
776 {
777 ++reg.idx;
778 shader_record_register_usage(shader, reg_maps, &reg, shader_version.type);
779 --count;
780 }
781
782 if (color0_mov)
783 {
784 if (ins.src[i].reg.type == WINED3DSPR_TEMP
785 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
786 {
787 shader->u.ps.color0_mov = TRUE;
788 shader->u.ps.color0_reg = ins.src[i].reg.idx;
789 }
790 }
791 }
792 }
793 }
794 reg_maps->loop_depth = max_loop_depth;
795
796 /* PS before 2.0 don't have explicit color outputs. Instead the value of
797 * R0 is written to the render target. */
798 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
799 reg_maps->rt_mask |= (1 << 0);
800
801 shader->functionLength = ((const char *)ptr - (const char *)byte_code);
802
803 return WINED3D_OK;
804 }
805
806 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
807 {
808 DWORD map = 1 << max;
809 map |= map - 1;
810 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
811
812 return wined3d_log2i(map);
813 }
814
815 static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
816 const struct wined3d_shader_version *shader_version)
817 {
818 TRACE("dcl");
819
820 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
821 {
822 switch (semantic->sampler_type)
823 {
824 case WINED3DSTT_2D: TRACE("_2d"); break;
825 case WINED3DSTT_CUBE: TRACE("_cube"); break;
826 case WINED3DSTT_VOLUME: TRACE("_volume"); break;
827 default: TRACE("_unknown_ttype(0x%08x)", semantic->sampler_type);
828 }
829 }
830 else
831 {
832 /* Pixel shaders 3.0 don't have usage semantics. */
833 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL) return;
834 else TRACE("_");
835
836 switch (semantic->usage)
837 {
838 case WINED3D_DECL_USAGE_POSITION:
839 TRACE("position%u", semantic->usage_idx);
840 break;
841
842 case WINED3D_DECL_USAGE_BLEND_INDICES:
843 TRACE("blend");
844 break;
845
846 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
847 TRACE("weight");
848 break;
849
850 case WINED3D_DECL_USAGE_NORMAL:
851 TRACE("normal%u", semantic->usage_idx);
852 break;
853
854 case WINED3D_DECL_USAGE_PSIZE:
855 TRACE("psize");
856 break;
857
858 case WINED3D_DECL_USAGE_COLOR:
859 if (!semantic->usage_idx) TRACE("color");
860 else TRACE("specular%u", (semantic->usage_idx - 1));
861 break;
862
863 case WINED3D_DECL_USAGE_TEXCOORD:
864 TRACE("texture%u", semantic->usage_idx);
865 break;
866
867 case WINED3D_DECL_USAGE_TANGENT:
868 TRACE("tangent");
869 break;
870
871 case WINED3D_DECL_USAGE_BINORMAL:
872 TRACE("binormal");
873 break;
874
875 case WINED3D_DECL_USAGE_TESS_FACTOR:
876 TRACE("tessfactor");
877 break;
878
879 case WINED3D_DECL_USAGE_POSITIONT:
880 TRACE("positionT%u", semantic->usage_idx);
881 break;
882
883 case WINED3D_DECL_USAGE_FOG:
884 TRACE("fog");
885 break;
886
887 case WINED3D_DECL_USAGE_DEPTH:
888 TRACE("depth");
889 break;
890
891 case WINED3D_DECL_USAGE_SAMPLE:
892 TRACE("sample");
893 break;
894
895 default:
896 FIXME("unknown_semantics(0x%08x)", semantic->usage);
897 }
898 }
899 }
900
901 static void shader_dump_register(const struct wined3d_shader_register *reg,
902 const struct wined3d_shader_version *shader_version)
903 {
904 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
905 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
906 UINT offset = reg->idx;
907
908 switch (reg->type)
909 {
910 case WINED3DSPR_TEMP:
911 TRACE("r");
912 break;
913
914 case WINED3DSPR_INPUT:
915 TRACE("v");
916 break;
917
918 case WINED3DSPR_CONST:
919 case WINED3DSPR_CONST2:
920 case WINED3DSPR_CONST3:
921 case WINED3DSPR_CONST4:
922 TRACE("c");
923 offset = shader_get_float_offset(reg->type, reg->idx);
924 break;
925
926 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
927 TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
928 break;
929
930 case WINED3DSPR_RASTOUT:
931 TRACE("%s", rastout_reg_names[reg->idx]);
932 break;
933
934 case WINED3DSPR_COLOROUT:
935 TRACE("oC");
936 break;
937
938 case WINED3DSPR_DEPTHOUT:
939 TRACE("oDepth");
940 break;
941
942 case WINED3DSPR_ATTROUT:
943 TRACE("oD");
944 break;
945
946 case WINED3DSPR_TEXCRDOUT:
947 /* Vertex shaders >= 3.0 use general purpose output registers
948 * (WINED3DSPR_OUTPUT), which can include an address token. */
949 if (shader_version->major >= 3) TRACE("o");
950 else TRACE("oT");
951 break;
952
953 case WINED3DSPR_CONSTINT:
954 TRACE("i");
955 break;
956
957 case WINED3DSPR_CONSTBOOL:
958 TRACE("b");
959 break;
960
961 case WINED3DSPR_LABEL:
962 TRACE("l");
963 break;
964
965 case WINED3DSPR_LOOP:
966 TRACE("aL");
967 break;
968
969 case WINED3DSPR_SAMPLER:
970 TRACE("s");
971 break;
972
973 case WINED3DSPR_MISCTYPE:
974 if (reg->idx > 1) FIXME("Unhandled misctype register %u.\n", reg->idx);
975 else TRACE("%s", misctype_reg_names[reg->idx]);
976 break;
977
978 case WINED3DSPR_PREDICATE:
979 TRACE("p");
980 break;
981
982 case WINED3DSPR_IMMCONST:
983 TRACE("l");
984 break;
985
986 case WINED3DSPR_CONSTBUFFER:
987 TRACE("cb");
988 break;
989
990 case WINED3DSPR_NULL:
991 TRACE("null");
992 break;
993
994 case WINED3DSPR_RESOURCE:
995 TRACE("t");
996 break;
997
998 default:
999 TRACE("unhandled_rtype(%#x)", reg->type);
1000 break;
1001 }
1002
1003 if (reg->type == WINED3DSPR_IMMCONST)
1004 {
1005 TRACE("(");
1006 switch (reg->immconst_type)
1007 {
1008 case WINED3D_IMMCONST_SCALAR:
1009 switch (reg->data_type)
1010 {
1011 case WINED3D_DATA_FLOAT:
1012 TRACE("%.8e", *(const float *)reg->immconst_data);
1013 break;
1014 case WINED3D_DATA_INT:
1015 TRACE("%d", reg->immconst_data[0]);
1016 break;
1017 case WINED3D_DATA_RESOURCE:
1018 case WINED3D_DATA_SAMPLER:
1019 case WINED3D_DATA_UINT:
1020 TRACE("%u", reg->immconst_data[0]);
1021 break;
1022 default:
1023 TRACE("<unhandled data type %#x>", reg->data_type);
1024 break;
1025 }
1026 break;
1027
1028 case WINED3D_IMMCONST_VEC4:
1029 switch (reg->data_type)
1030 {
1031 case WINED3D_DATA_FLOAT:
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 case WINED3D_DATA_INT:
1037 TRACE("%d, %d, %d, %d",
1038 reg->immconst_data[0], reg->immconst_data[1],
1039 reg->immconst_data[2], reg->immconst_data[3]);
1040 break;
1041 case WINED3D_DATA_RESOURCE:
1042 case WINED3D_DATA_SAMPLER:
1043 case WINED3D_DATA_UINT:
1044 TRACE("%u, %u, %u, %u",
1045 reg->immconst_data[0], reg->immconst_data[1],
1046 reg->immconst_data[2], reg->immconst_data[3]);
1047 break;
1048 default:
1049 TRACE("<unhandled data type %#x>", reg->data_type);
1050 break;
1051 }
1052 break;
1053
1054 default:
1055 TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
1056 break;
1057 }
1058 TRACE(")");
1059 }
1060 else if (reg->type != WINED3DSPR_RASTOUT
1061 && reg->type != WINED3DSPR_MISCTYPE
1062 && reg->type != WINED3DSPR_NULL)
1063 {
1064 if (offset != ~0U)
1065 {
1066 TRACE("[");
1067 if (reg->rel_addr)
1068 {
1069 shader_dump_src_param(reg->rel_addr, shader_version);
1070 TRACE(" + ");
1071 }
1072 TRACE("%u]", offset);
1073
1074 if (reg->array_idx != ~0U)
1075 {
1076 TRACE("[");
1077 if (reg->array_rel_addr)
1078 {
1079 shader_dump_src_param(reg->array_rel_addr, shader_version);
1080 TRACE(" + ");
1081 }
1082 TRACE("%u]", reg->array_idx);
1083 }
1084 }
1085 }
1086 }
1087
1088 void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
1089 const struct wined3d_shader_version *shader_version)
1090 {
1091 DWORD write_mask = param->write_mask;
1092
1093 shader_dump_register(&param->reg, shader_version);
1094
1095 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
1096 {
1097 static const char *write_mask_chars = "xyzw";
1098
1099 TRACE(".");
1100 if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
1101 if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
1102 if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
1103 if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
1104 }
1105 }
1106
1107 void shader_dump_src_param(const struct wined3d_shader_src_param *param,
1108 const struct wined3d_shader_version *shader_version)
1109 {
1110 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
1111 DWORD swizzle = param->swizzle;
1112
1113 if (src_modifier == WINED3DSPSM_NEG
1114 || src_modifier == WINED3DSPSM_BIASNEG
1115 || src_modifier == WINED3DSPSM_SIGNNEG
1116 || src_modifier == WINED3DSPSM_X2NEG
1117 || src_modifier == WINED3DSPSM_ABSNEG)
1118 TRACE("-");
1119 else if (src_modifier == WINED3DSPSM_COMP)
1120 TRACE("1-");
1121 else if (src_modifier == WINED3DSPSM_NOT)
1122 TRACE("!");
1123
1124 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1125 TRACE("abs(");
1126
1127 shader_dump_register(&param->reg, shader_version);
1128
1129 if (src_modifier)
1130 {
1131 switch (src_modifier)
1132 {
1133 case WINED3DSPSM_NONE: break;
1134 case WINED3DSPSM_NEG: break;
1135 case WINED3DSPSM_NOT: break;
1136 case WINED3DSPSM_BIAS: TRACE("_bias"); break;
1137 case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
1138 case WINED3DSPSM_SIGN: TRACE("_bx2"); break;
1139 case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1140 case WINED3DSPSM_COMP: break;
1141 case WINED3DSPSM_X2: TRACE("_x2"); break;
1142 case WINED3DSPSM_X2NEG: TRACE("_x2"); break;
1143 case WINED3DSPSM_DZ: TRACE("_dz"); break;
1144 case WINED3DSPSM_DW: TRACE("_dw"); break;
1145 case WINED3DSPSM_ABSNEG: TRACE(")"); break;
1146 case WINED3DSPSM_ABS: TRACE(")"); break;
1147 default: TRACE("_unknown_modifier(%#x)", src_modifier);
1148 }
1149 }
1150
1151 if (swizzle != WINED3DSP_NOSWIZZLE)
1152 {
1153 static const char *swizzle_chars = "xyzw";
1154 DWORD swizzle_x = swizzle & 0x03;
1155 DWORD swizzle_y = (swizzle >> 2) & 0x03;
1156 DWORD swizzle_z = (swizzle >> 4) & 0x03;
1157 DWORD swizzle_w = (swizzle >> 6) & 0x03;
1158
1159 if (swizzle_x == swizzle_y
1160 && swizzle_x == swizzle_z
1161 && swizzle_x == swizzle_w)
1162 {
1163 TRACE(".%c", swizzle_chars[swizzle_x]);
1164 }
1165 else
1166 {
1167 TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1168 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1169 }
1170 }
1171 }
1172
1173 /* Shared code in order to generate the bulk of the shader string.
1174 * NOTE: A description of how to parse tokens can be found on MSDN. */
1175 void shader_generate_main(const struct wined3d_shader *shader, struct wined3d_shader_buffer *buffer,
1176 const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1177 {
1178 struct wined3d_device *device = shader->device;
1179 const struct wined3d_shader_frontend *fe = shader->frontend;
1180 void *fe_data = shader->frontend_data;
1181 struct wined3d_shader_version shader_version;
1182 struct wined3d_shader_loop_state loop_state;
1183 struct wined3d_shader_instruction ins;
1184 struct wined3d_shader_tex_mx tex_mx;
1185 struct wined3d_shader_context ctx;
1186 const DWORD *ptr = byte_code;
1187
1188 /* Initialize current parsing state. */
1189 tex_mx.current_row = 0;
1190 loop_state.current_depth = 0;
1191 loop_state.current_reg = 0;
1192
1193 ctx.shader = shader;
1194 ctx.gl_info = &device->adapter->gl_info;
1195 ctx.reg_maps = reg_maps;
1196 ctx.buffer = buffer;
1197 ctx.tex_mx = &tex_mx;
1198 ctx.loop_state = &loop_state;
1199 ctx.backend_data = backend_ctx;
1200 ins.ctx = &ctx;
1201
1202 fe->shader_read_header(fe_data, &ptr, &shader_version);
1203
1204 while (!fe->shader_is_end(fe_data, &ptr))
1205 {
1206 const char *comment;
1207 UINT comment_size;
1208
1209 /* Skip comment tokens. */
1210 fe->shader_read_comment(&ptr, &comment, &comment_size);
1211 if (comment) continue;
1212
1213 /* Read opcode. */
1214 fe->shader_read_instruction(fe_data, &ptr, &ins);
1215
1216 /* Unknown opcode and its parameters. */
1217 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1218 {
1219 TRACE("Skipping unrecognized instruction.\n");
1220 continue;
1221 }
1222
1223 if (ins.predicate)
1224 FIXME("Predicates not implemented.\n");
1225
1226 /* Call appropriate function for output target */
1227 device->shader_backend->shader_handle_instruction(&ins);
1228 }
1229 }
1230
1231 static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1232 {
1233 DWORD mmask = dst->modifiers;
1234
1235 switch (dst->shift)
1236 {
1237 case 0: break;
1238 case 13: TRACE("_d8"); break;
1239 case 14: TRACE("_d4"); break;
1240 case 15: TRACE("_d2"); break;
1241 case 1: TRACE("_x2"); break;
1242 case 2: TRACE("_x4"); break;
1243 case 3: TRACE("_x8"); break;
1244 default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1245 }
1246
1247 if (mmask & WINED3DSPDM_SATURATE) TRACE("_sat");
1248 if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1249 if (mmask & WINED3DSPDM_MSAMPCENTROID) TRACE("_centroid");
1250
1251 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1252 if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1253 }
1254
1255 static void shader_dump_primitive_type(enum wined3d_primitive_type primitive_type)
1256 {
1257 switch (primitive_type)
1258 {
1259 case WINED3D_PT_UNDEFINED:
1260 TRACE("undefined");
1261 break;
1262 case WINED3D_PT_POINTLIST:
1263 TRACE("pointlist");
1264 break;
1265 case WINED3D_PT_LINELIST:
1266 TRACE("linelist");
1267 break;
1268 case WINED3D_PT_LINESTRIP:
1269 TRACE("linestrip");
1270 break;
1271 case WINED3D_PT_TRIANGLELIST:
1272 TRACE("trianglelist");
1273 break;
1274 case WINED3D_PT_TRIANGLESTRIP:
1275 TRACE("trianglestrip");
1276 break;
1277 case WINED3D_PT_TRIANGLEFAN:
1278 TRACE("trianglefan");
1279 break;
1280 case WINED3D_PT_LINELIST_ADJ:
1281 TRACE("linelist_adj");
1282 break;
1283 case WINED3D_PT_LINESTRIP_ADJ:
1284 TRACE("linestrip_adj");
1285 break;
1286 case WINED3D_PT_TRIANGLELIST_ADJ:
1287 TRACE("trianglelist_adj");
1288 break;
1289 case WINED3D_PT_TRIANGLESTRIP_ADJ:
1290 TRACE("trianglestrip_adj");
1291 break;
1292 default:
1293 TRACE("<unrecognized_primitive_type %#x>", primitive_type);
1294 break;
1295 }
1296 }
1297
1298 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1299 {
1300 struct wined3d_shader_version shader_version;
1301 const DWORD *ptr = byte_code;
1302 const char *type_prefix;
1303 DWORD i;
1304
1305 TRACE("Parsing %p.\n", byte_code);
1306
1307 fe->shader_read_header(fe_data, &ptr, &shader_version);
1308
1309 switch (shader_version.type)
1310 {
1311 case WINED3D_SHADER_TYPE_VERTEX:
1312 type_prefix = "vs";
1313 break;
1314
1315 case WINED3D_SHADER_TYPE_GEOMETRY:
1316 type_prefix = "gs";
1317 break;
1318
1319 case WINED3D_SHADER_TYPE_PIXEL:
1320 type_prefix = "ps";
1321 break;
1322
1323 default:
1324 FIXME("Unhandled shader type %#x.\n", shader_version.type);
1325 type_prefix = "unknown";
1326 break;
1327 }
1328
1329 TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1330
1331 while (!fe->shader_is_end(fe_data, &ptr))
1332 {
1333 struct wined3d_shader_instruction ins;
1334 const char *comment;
1335 UINT comment_size;
1336
1337 /* comment */
1338 fe->shader_read_comment(&ptr, &comment, &comment_size);
1339 if (comment)
1340 {
1341 if (comment_size > 4 && *(const DWORD *)comment == WINEMAKEFOURCC('T', 'E', 'X', 'T'))
1342 {
1343 const char *end = comment + comment_size;
1344 const char *ptr = comment + 4;
1345 const char *line = ptr;
1346
1347 TRACE("// TEXT\n");
1348 while (ptr != end)
1349 {
1350 if (*ptr == '\n')
1351 {
1352 UINT len = ptr - line;
1353 if (len && *(ptr - 1) == '\r') --len;
1354 TRACE("// %s\n", debugstr_an(line, len));
1355 line = ++ptr;
1356 }
1357 else ++ptr;
1358 }
1359 if (line != ptr) TRACE("// %s\n", debugstr_an(line, ptr - line));
1360 }
1361 else TRACE("// %s\n", debugstr_an(comment, comment_size));
1362 continue;
1363 }
1364
1365 fe->shader_read_instruction(fe_data, &ptr, &ins);
1366 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1367 {
1368 TRACE("Skipping unrecognized instruction.\n");
1369 continue;
1370 }
1371
1372 if (ins.handler_idx == WINED3DSIH_DCL)
1373 {
1374 shader_dump_decl_usage(&ins.declaration.semantic, &shader_version);
1375 shader_dump_ins_modifiers(&ins.declaration.semantic.reg);
1376 TRACE(" ");
1377 shader_dump_dst_param(&ins.declaration.semantic.reg, &shader_version);
1378 }
1379 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
1380 {
1381 TRACE("%s ", shader_opcode_names[ins.handler_idx]);
1382 shader_dump_src_param(&ins.declaration.src, &shader_version);
1383 TRACE(", %s", ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
1384 }
1385 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
1386 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
1387 {
1388 TRACE("%s ", shader_opcode_names[ins.handler_idx]);
1389 shader_dump_primitive_type(ins.declaration.primitive_type);
1390 }
1391 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1392 {
1393 TRACE("%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
1394 }
1395 else if (ins.handler_idx == WINED3DSIH_DEF)
1396 {
1397 TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(ins.dst[0].reg.type, ins.dst[0].reg.idx),
1398 *(const float *)&ins.src[0].reg.immconst_data[0],
1399 *(const float *)&ins.src[0].reg.immconst_data[1],
1400 *(const float *)&ins.src[0].reg.immconst_data[2],
1401 *(const float *)&ins.src[0].reg.immconst_data[3]);
1402 }
1403 else if (ins.handler_idx == WINED3DSIH_DEFI)
1404 {
1405 TRACE("defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx,
1406 ins.src[0].reg.immconst_data[0],
1407 ins.src[0].reg.immconst_data[1],
1408 ins.src[0].reg.immconst_data[2],
1409 ins.src[0].reg.immconst_data[3]);
1410 }
1411 else if (ins.handler_idx == WINED3DSIH_DEFB)
1412 {
1413 TRACE("defb b%u = %s", ins.dst[0].reg.idx, ins.src[0].reg.immconst_data[0] ? "true" : "false");
1414 }
1415 else
1416 {
1417 if (ins.predicate)
1418 {
1419 TRACE("(");
1420 shader_dump_src_param(ins.predicate, &shader_version);
1421 TRACE(") ");
1422 }
1423
1424 /* PixWin marks instructions with the coissue flag with a '+' */
1425 if (ins.coissue) TRACE("+");
1426
1427 TRACE("%s", shader_opcode_names[ins.handler_idx]);
1428
1429 if (ins.handler_idx == WINED3DSIH_IFC
1430 || ins.handler_idx == WINED3DSIH_BREAKC)
1431 {
1432 switch (ins.flags)
1433 {
1434 case WINED3D_SHADER_REL_OP_GT: TRACE("_gt"); break;
1435 case WINED3D_SHADER_REL_OP_EQ: TRACE("_eq"); break;
1436 case WINED3D_SHADER_REL_OP_GE: TRACE("_ge"); break;
1437 case WINED3D_SHADER_REL_OP_LT: TRACE("_lt"); break;
1438 case WINED3D_SHADER_REL_OP_NE: TRACE("_ne"); break;
1439 case WINED3D_SHADER_REL_OP_LE: TRACE("_le"); break;
1440 default: TRACE("_(%u)", ins.flags);
1441 }
1442 }
1443 else if (ins.handler_idx == WINED3DSIH_TEX
1444 && shader_version.major >= 2
1445 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1446 {
1447 TRACE("p");
1448 }
1449
1450 for (i = 0; i < ins.dst_count; ++i)
1451 {
1452 shader_dump_ins_modifiers(&ins.dst[i]);
1453 TRACE(!i ? " " : ", ");
1454 shader_dump_dst_param(&ins.dst[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 TRACE(!i ? " " : ", ");
1461 shader_dump_src_param(&ins.src[i - ins.dst_count], &shader_version);
1462 }
1463 }
1464 TRACE("\n");
1465 }
1466 }
1467
1468 static void shader_cleanup(struct wined3d_shader *shader)
1469 {
1470 shader->device->shader_backend->shader_destroy(shader);
1471 HeapFree(GetProcessHeap(), 0, shader->reg_maps.constf);
1472 HeapFree(GetProcessHeap(), 0, shader->function);
1473 shader_delete_constant_list(&shader->constantsF);
1474 shader_delete_constant_list(&shader->constantsB);
1475 shader_delete_constant_list(&shader->constantsI);
1476 list_remove(&shader->shader_list_entry);
1477
1478 if (shader->frontend && shader->frontend_data)
1479 shader->frontend->shader_free(shader->frontend_data);
1480 }
1481
1482 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1483 static void shader_none_select(const struct wined3d_context *context, BOOL usePS, BOOL useVS) {}
1484 static void shader_none_select_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info,
1485 enum tex_types tex_type, const SIZE *ds_mask_size) {}
1486 static void shader_none_deselect_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info) {}
1487 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
1488 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
1489 static void shader_none_load_constants(const struct wined3d_context *context, BOOL usePS, BOOL useVS) {}
1490 static void shader_none_load_np2fixup_constants(void *shader_priv,
1491 const struct wined3d_gl_info *gl_info, const struct wined3d_state *state) {}
1492 static void shader_none_destroy(struct wined3d_shader *shader) {}
1493 static HRESULT shader_none_alloc(struct wined3d_device *device) {return WINED3D_OK;}
1494 static void shader_none_free(struct wined3d_device *device) {}
1495 static void shader_none_context_destroyed(void *shader_priv, const struct wined3d_context *context) {}
1496
1497 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
1498 {
1499 /* Set the shader caps to 0 for the none shader backend */
1500 caps->vs_version = 0;
1501 caps->gs_version = 0;
1502 caps->ps_version = 0;
1503 caps->vs_uniform_count = 0;
1504 caps->ps_uniform_count = 0;
1505 caps->ps_1x_max_value = 0.0f;
1506 caps->vs_clipping = 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->vs_version;
1601 break;
1602 case WINED3D_SHADER_TYPE_GEOMETRY:
1603 backend_version = shader->device->gs_version;
1604 break;
1605 case WINED3D_SHADER_TYPE_PIXEL:
1606 backend_version = shader->device->ps_version;
1607 break;
1608 default:
1609 FIXME("No backend version-checking for this shader type\n");
1610 backend_version = 0;
1611 }
1612 if (reg_maps->shader_version.major > backend_version)
1613 {
1614 WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
1615 reg_maps->shader_version.major, reg_maps->shader_version.minor);
1616 return WINED3DERR_INVALIDCALL;
1617 }
1618
1619 shader->function = HeapAlloc(GetProcessHeap(), 0, shader->functionLength);
1620 if (!shader->function)
1621 return E_OUTOFMEMORY;
1622 memcpy(shader->function, byte_code, shader->functionLength);
1623
1624 return WINED3D_OK;
1625 }
1626
1627 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
1628 {
1629 ULONG refcount = InterlockedIncrement(&shader->ref);
1630
1631 TRACE("%p increasing refcount to %u.\n", shader, refcount);
1632
1633 return refcount;
1634 }
1635
1636 /* Do not call while under the GL lock. */
1637 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
1638 {
1639 ULONG refcount = InterlockedDecrement(&shader->ref);
1640
1641 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1642
1643 if (!refcount)
1644 {
1645 shader_cleanup(shader);
1646 shader->parent_ops->wined3d_object_destroyed(shader->parent);
1647 HeapFree(GetProcessHeap(), 0, shader);
1648 }
1649
1650 return refcount;
1651 }
1652
1653 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
1654 {
1655 TRACE("shader %p.\n", shader);
1656
1657 return shader->parent;
1658 }
1659
1660 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
1661 void *byte_code, UINT *byte_code_size)
1662 {
1663 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
1664
1665 if (!byte_code)
1666 {
1667 *byte_code_size = shader->functionLength;
1668 return WINED3D_OK;
1669 }
1670
1671 if (*byte_code_size < shader->functionLength)
1672 {
1673 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
1674 * than the required size we should write the required size and
1675 * return D3DERR_MOREDATA. That's not actually true. */
1676 return WINED3DERR_INVALIDCALL;
1677 }
1678
1679 memcpy(byte_code, shader->function, shader->functionLength);
1680
1681 return WINED3D_OK;
1682 }
1683
1684 /* Set local constants for d3d8 shaders. */
1685 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
1686 UINT start_idx, const float *src_data, UINT count)
1687 {
1688 UINT end_idx = start_idx + count;
1689 UINT i;
1690
1691 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
1692
1693 if (end_idx > shader->limits.constant_float)
1694 {
1695 WARN("end_idx %u > float constants limit %u.\n",
1696 end_idx, shader->limits.constant_float);
1697 end_idx = shader->limits.constant_float;
1698 }
1699
1700 for (i = start_idx; i < end_idx; ++i)
1701 {
1702 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1703 if (!lconst)
1704 return E_OUTOFMEMORY;
1705
1706 lconst->idx = i;
1707 memcpy(lconst->value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
1708 list_add_head(&shader->constantsF, &lconst->entry);
1709 }
1710
1711 return WINED3D_OK;
1712 }
1713
1714 void find_vs_compile_args(const struct wined3d_state *state,
1715 const struct wined3d_shader *shader, struct vs_compile_args *args)
1716 {
1717 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
1718 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
1719 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
1720 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
1721 args->swizzle_map = shader->device->strided_streams.swizzle_map;
1722 }
1723
1724 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
1725 {
1726 if (usage_idx1 != usage_idx2)
1727 return FALSE;
1728 if (usage1 == usage2)
1729 return TRUE;
1730 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
1731 return TRUE;
1732 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
1733 return TRUE;
1734
1735 return FALSE;
1736 }
1737
1738 BOOL vshader_get_input(const struct wined3d_shader *shader,
1739 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
1740 {
1741 WORD map = shader->reg_maps.input_registers;
1742 unsigned int i;
1743
1744 for (i = 0; map; map >>= 1, ++i)
1745 {
1746 if (!(map & 1)) continue;
1747
1748 if (match_usage(shader->u.vs.attributes[i].usage,
1749 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
1750 {
1751 *regnum = i;
1752 return TRUE;
1753 }
1754 }
1755 return FALSE;
1756 }
1757
1758 static void vertexshader_set_limits(struct wined3d_shader *shader)
1759 {
1760 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
1761 shader->reg_maps.shader_version.minor);
1762 struct wined3d_device *device = shader->device;
1763
1764 shader->limits.packed_input = 0;
1765
1766 switch (shader_version)
1767 {
1768 case WINED3D_SHADER_VERSION(1, 0):
1769 case WINED3D_SHADER_VERSION(1, 1):
1770 shader->limits.constant_bool = 0;
1771 shader->limits.constant_int = 0;
1772 shader->limits.packed_output = 12;
1773 shader->limits.sampler = 0;
1774 /* TODO: vs_1_1 has a minimum of 96 constants. What happens when
1775 * a vs_1_1 shader is used on a vs_3_0 capable card that has 256
1776 * constants? */
1777 shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1778 break;
1779
1780 case WINED3D_SHADER_VERSION(2, 0):
1781 case WINED3D_SHADER_VERSION(2, 1):
1782 shader->limits.constant_bool = 16;
1783 shader->limits.constant_int = 16;
1784 shader->limits.packed_output = 12;
1785 shader->limits.sampler = 0;
1786 shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1787 break;
1788
1789 case WINED3D_SHADER_VERSION(3, 0):
1790 shader->limits.constant_bool = 16;
1791 shader->limits.constant_int = 16;
1792 shader->limits.packed_output = 12;
1793 shader->limits.sampler = 4;
1794 /* DX10 cards on Windows advertise a d3d9 constant limit of 256
1795 * even though they are capable of supporting much more (GL
1796 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
1797 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
1798 * shaders to 256. */
1799 shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1800 break;
1801
1802 case WINED3D_SHADER_VERSION(4, 0):
1803 shader->limits.sampler = 16; /* FIXME: 128 resources, 16 sampler states */
1804 shader->limits.constant_int = 0;
1805 shader->limits.constant_float = 0;
1806 shader->limits.constant_bool = 0;
1807 shader->limits.packed_output = 16;
1808 shader->limits.packed_input = 0;
1809 break;
1810
1811 default:
1812 shader->limits.constant_bool = 16;
1813 shader->limits.constant_int = 16;
1814 shader->limits.packed_output = 12;
1815 shader->limits.sampler = 0;
1816 shader->limits.constant_float = min(256, device->d3d_vshader_constantF);
1817 FIXME("Unrecognized vertex shader version \"%u.%u\".\n",
1818 shader->reg_maps.shader_version.major,
1819 shader->reg_maps.shader_version.minor);
1820 }
1821 }
1822
1823 static HRESULT vertexshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
1824 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1825 void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
1826 {
1827 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1828 unsigned int i;
1829 HRESULT hr;
1830 WORD map;
1831
1832 if (!byte_code) return WINED3DERR_INVALIDCALL;
1833
1834 shader_init(shader, device, parent, parent_ops);
1835 hr = shader_set_function(shader, byte_code, output_signature, device->d3d_vshader_constantF,
1836 WINED3D_SHADER_TYPE_VERTEX, max_version);
1837 if (FAILED(hr))
1838 {
1839 WARN("Failed to set function, hr %#x.\n", hr);
1840 shader_cleanup(shader);
1841 return hr;
1842 }
1843
1844 map = reg_maps->input_registers;
1845 for (i = 0; map; map >>= 1, ++i)
1846 {
1847 if (!(map & 1) || !shader->input_signature[i].semantic_name)
1848 continue;
1849
1850 shader->u.vs.attributes[i].usage =
1851 shader_usage_from_semantic_name(shader->input_signature[i].semantic_name);
1852 shader->u.vs.attributes[i].usage_idx = shader->input_signature[i].semantic_idx;
1853 }
1854
1855 if (output_signature)
1856 {
1857 for (i = 0; i < output_signature->element_count; ++i)
1858 {
1859 struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1860 reg_maps->output_registers |= 1 << e->register_idx;
1861 shader->output_signature[e->register_idx] = *e;
1862 }
1863 }
1864
1865 vertexshader_set_limits(shader);
1866
1867 shader->load_local_constsF = reg_maps->usesrelconstF
1868 && !list_empty(&shader->constantsF);
1869
1870 return WINED3D_OK;
1871 }
1872
1873 static void geometryshader_set_limits(struct wined3d_shader *shader)
1874 {
1875 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
1876 shader->reg_maps.shader_version.minor);
1877
1878 switch (shader_version)
1879 {
1880 case WINED3D_SHADER_VERSION(4, 0):
1881 shader->limits.sampler = 16; /* FIXME: 128 resources, 16 sampler states */
1882 shader->limits.constant_int = 0;
1883 shader->limits.constant_float = 0;
1884 shader->limits.constant_bool = 0;
1885 shader->limits.packed_output = 32;
1886 shader->limits.packed_input = 16;
1887 break;
1888
1889 default:
1890 memset(&shader->limits, 0, sizeof(shader->limits));
1891 FIXME("Unhandled geometry shader version \"%u.%u\".\n",
1892 shader->reg_maps.shader_version.major,
1893 shader->reg_maps.shader_version.minor);
1894 }
1895 }
1896
1897 static HRESULT geometryshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
1898 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1899 void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
1900 {
1901 HRESULT hr;
1902
1903 shader_init(shader, device, parent, parent_ops);
1904 hr = shader_set_function(shader, byte_code, output_signature, 0,
1905 WINED3D_SHADER_TYPE_GEOMETRY, max_version);
1906 if (FAILED(hr))
1907 {
1908 WARN("Failed to set function, hr %#x.\n", hr);
1909 shader_cleanup(shader);
1910 return hr;
1911 }
1912
1913 geometryshader_set_limits(shader);
1914
1915 shader->load_local_constsF = FALSE;
1916
1917 return WINED3D_OK;
1918 }
1919
1920 void find_ps_compile_args(const struct wined3d_state *state,
1921 const struct wined3d_shader *shader, struct ps_compile_args *args)
1922 {
1923 struct wined3d_device *device = shader->device;
1924 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
1925 const struct wined3d_texture *texture;
1926 UINT i;
1927
1928 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
1929 if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && state->render_states[WINED3D_RS_SRGBWRITEENABLE])
1930 {
1931 const struct wined3d_surface *rt = state->fb->render_targets[0];
1932 if (rt->resource.format->flags & WINED3DFMT_FLAG_SRGB_WRITE)
1933 {
1934 static unsigned int warned = 0;
1935
1936 args->srgb_correction = 1;
1937 if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
1938 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
1939 "support, expect rendering artifacts.\n");
1940 }
1941 }
1942
1943 if (shader->reg_maps.shader_version.major == 1
1944 && shader->reg_maps.shader_version.minor <= 3)
1945 {
1946 for (i = 0; i < 4; ++i)
1947 {
1948 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
1949
1950 if (flags & WINED3D_TTFF_PROJECTED)
1951 {
1952 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
1953
1954 if (!state->vertex_shader)
1955 {
1956 unsigned int j;
1957 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
1958 DWORD max_valid = WINED3D_TTFF_COUNT4;
1959 enum wined3d_sampler_texture_type sampler_type = shader->reg_maps.sampler_type[i];
1960
1961 for (j = 0; j < state->vertex_declaration->element_count; ++j)
1962 {
1963 struct wined3d_vertex_declaration_element *element =
1964 &state->vertex_declaration->elements[j];
1965
1966 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
1967 && element->usage_idx == index)
1968 {
1969 max_valid = element->format->component_count;
1970 break;
1971 }
1972 }
1973 if (!tex_transform || tex_transform > max_valid)
1974 {
1975 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
1976 tex_transform, max_valid);
1977 tex_transform = max_valid;
1978 }
1979 if ((sampler_type == WINED3DSTT_1D && tex_transform > WINED3D_TTFF_COUNT1)
1980 || (sampler_type == WINED3DSTT_2D && tex_transform > WINED3D_TTFF_COUNT2)
1981 || (sampler_type == WINED3DSTT_VOLUME && tex_transform > WINED3D_TTFF_COUNT3))
1982 tex_transform |= WINED3D_PSARGS_PROJECTED;
1983 else
1984 {
1985 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
1986 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
1987 i, tex_transform, sampler_type);
1988 }
1989 }
1990 else
1991 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
1992
1993 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
1994 }
1995 }
1996 }
1997
1998 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
1999 {
2000 if (!shader->reg_maps.sampler_type[i])
2001 continue;
2002
2003 texture = state->textures[i];
2004 if (!texture)
2005 {
2006 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2007 continue;
2008 }
2009 args->color_fixup[i] = texture->resource.format->color_fixup;
2010
2011 if (texture->resource.format->flags & WINED3DFMT_FLAG_SHADOW)
2012 args->shadow |= 1 << i;
2013
2014 /* Flag samplers that need NP2 texcoord fixup. */
2015 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
2016 args->np2_fixup |= (1 << i);
2017 }
2018 if (shader->reg_maps.shader_version.major >= 3)
2019 {
2020 if (device->strided_streams.position_transformed)
2021 {
2022 args->vp_mode = pretransformed;
2023 }
2024 else if (use_vs(state))
2025 {
2026 args->vp_mode = vertexshader;
2027 }
2028 else
2029 {
2030 args->vp_mode = fixedfunction;
2031 }
2032 args->fog = FOG_OFF;
2033 }
2034 else
2035 {
2036 args->vp_mode = vertexshader;
2037 if (state->render_states[WINED3D_RS_FOGENABLE])
2038 {
2039 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
2040 {
2041 case WINED3D_FOG_NONE:
2042 if (device->strided_streams.position_transformed || use_vs(state))
2043 {
2044 args->fog = FOG_LINEAR;
2045 break;
2046 }
2047
2048 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
2049 {
2050 case WINED3D_FOG_NONE: /* Fall through. */
2051 case WINED3D_FOG_LINEAR: args->fog = FOG_LINEAR; break;
2052 case WINED3D_FOG_EXP: args->fog = FOG_EXP; break;
2053 case WINED3D_FOG_EXP2: args->fog = FOG_EXP2; break;
2054 }
2055 break;
2056
2057 case WINED3D_FOG_LINEAR: args->fog = FOG_LINEAR; break;
2058 case WINED3D_FOG_EXP: args->fog = FOG_EXP; break;
2059 case WINED3D_FOG_EXP2: args->fog = FOG_EXP2; break;
2060 }
2061 }
2062 else
2063 {
2064 args->fog = FOG_OFF;
2065 }
2066 }
2067 }
2068
2069 static void pixelshader_set_limits(struct wined3d_shader *shader)
2070 {
2071 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
2072 shader->reg_maps.shader_version.minor);
2073
2074 shader->limits.packed_output = 0;
2075
2076 switch (shader_version)
2077 {
2078 case WINED3D_SHADER_VERSION(1, 0):
2079 case WINED3D_SHADER_VERSION(1, 1):
2080 case WINED3D_SHADER_VERSION(1, 2):
2081 case WINED3D_SHADER_VERSION(1, 3):
2082 shader->limits.constant_float = 8;
2083 shader->limits.constant_int = 0;
2084 shader->limits.constant_bool = 0;
2085 shader->limits.sampler = 4;
2086 shader->limits.packed_input = 0;
2087 break;
2088
2089 case WINED3D_SHADER_VERSION(1, 4):
2090 shader->limits.constant_float = 8;
2091 shader->limits.constant_int = 0;
2092 shader->limits.constant_bool = 0;
2093 shader->limits.sampler = 6;
2094 shader->limits.packed_input = 0;
2095 break;
2096
2097 /* FIXME: Temporaries must match D3DPSHADERCAPS2_0.NumTemps. */
2098 case WINED3D_SHADER_VERSION(2, 0):
2099 shader->limits.constant_float = 32;
2100 shader->limits.constant_int = 16;
2101 shader->limits.constant_bool = 16;
2102 shader->limits.sampler = 16;
2103 shader->limits.packed_input = 0;
2104 break;
2105
2106 case WINED3D_SHADER_VERSION(2, 1):
2107 shader->limits.constant_float = 32;
2108 shader->limits.constant_int = 16;
2109 shader->limits.constant_bool = 16;
2110 shader->limits.sampler = 16;
2111 shader->limits.packed_input = 0;
2112 break;
2113
2114 case WINED3D_SHADER_VERSION(3, 0):
2115 shader->limits.constant_float = 224;
2116 shader->limits.constant_int = 16;
2117 shader->limits.constant_bool = 16;
2118 shader->limits.sampler = 16;
2119 shader->limits.packed_input = 12;
2120 break;
2121
2122 case WINED3D_SHADER_VERSION(4, 0):
2123 shader->limits.sampler = 16; /* FIXME: 128 resources, 16 sampler states */
2124 shader->limits.constant_int = 0;
2125 shader->limits.constant_float = 0;
2126 shader->limits.constant_bool = 0;
2127 shader->limits.packed_input = 32;
2128 break;
2129
2130 default:
2131 shader->limits.constant_float = 32;
2132 shader->limits.constant_int = 16;
2133 shader->limits.constant_bool = 16;
2134 shader->limits.sampler = 16;
2135 shader->limits.packed_input = 0;
2136 FIXME("Unrecognized pixel shader version %u.%u\n",
2137 shader->reg_maps.shader_version.major,
2138 shader->reg_maps.shader_version.minor);
2139 }
2140 }
2141
2142 static HRESULT pixelshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2143 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2144 void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
2145 {
2146 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2147 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2148 HRESULT hr;
2149
2150 if (!byte_code) return WINED3DERR_INVALIDCALL;
2151
2152 shader_init(shader, device, parent, parent_ops);
2153 hr = shader_set_function(shader, byte_code, output_signature, device->d3d_pshader_constantF,
2154 WINED3D_SHADER_TYPE_PIXEL, max_version);
2155 if (FAILED(hr))
2156 {
2157 WARN("Failed to set function, hr %#x.\n", hr);
2158 shader_cleanup(shader);
2159 return hr;
2160 }
2161
2162 pixelshader_set_limits(shader);
2163
2164 for (i = 0; i < MAX_REG_INPUT; ++i)
2165 {
2166 if (shader->u.ps.input_reg_used[i])
2167 {
2168 ++num_regs_used;
2169 highest_reg_used = i;
2170 }
2171 }
2172
2173 /* Don't do any register mapping magic if it is not needed, or if we can't
2174 * achieve anything anyway */
2175 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2176 || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2177 {
2178 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2179 {
2180 /* This happens with relative addressing. The input mapper function
2181 * warns about this if the higher registers are declared too, so
2182 * don't write a FIXME here */
2183 WARN("More varying registers used than supported\n");
2184 }
2185
2186 for (i = 0; i < MAX_REG_INPUT; ++i)
2187 {
2188 shader->u.ps.input_reg_map[i] = i;
2189 }
2190
2191 shader->u.ps.declared_in_count = highest_reg_used + 1;
2192 }
2193 else
2194 {
2195 shader->u.ps.declared_in_count = 0;
2196 for (i = 0; i < MAX_REG_INPUT; ++i)
2197 {
2198 if (shader->u.ps.input_reg_used[i])
2199 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
2200 else shader->u.ps.input_reg_map[i] = ~0U;
2201 }
2202 }
2203
2204 shader->load_local_constsF = FALSE;
2205
2206 return WINED3D_OK;
2207 }
2208
2209 void pixelshader_update_samplers(struct wined3d_shader_reg_maps *reg_maps, struct wined3d_texture * const *textures)
2210 {
2211 enum wined3d_sampler_texture_type *sampler_type = reg_maps->sampler_type;
2212 unsigned int i;
2213
2214 if (reg_maps->shader_version.major != 1) return;
2215
2216 for (i = 0; i < max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS); ++i)
2217 {
2218 /* We don't sample from this sampler. */
2219 if (!sampler_type[i]) continue;
2220
2221 if (!textures[i])
2222 {
2223 WARN("No texture bound to sampler %u, using 2D.\n", i);
2224 sampler_type[i] = WINED3DSTT_2D;
2225 continue;
2226 }
2227
2228 switch (textures[i]->target)
2229 {
2230 case GL_TEXTURE_RECTANGLE_ARB:
2231 case GL_TEXTURE_2D:
2232 /* We have to select between texture rectangles and 2D
2233 * textures later because 2.0 and 3.0 shaders only have
2234 * WINED3DSTT_2D as well. */
2235 sampler_type[i] = WINED3DSTT_2D;
2236 break;
2237
2238 case GL_TEXTURE_3D:
2239 sampler_type[i] = WINED3DSTT_VOLUME;
2240 break;
2241
2242 case GL_TEXTURE_CUBE_MAP_ARB:
2243 sampler_type[i] = WINED3DSTT_CUBE;
2244 break;
2245
2246 default:
2247 FIXME("Unrecognized texture type %#x, using 2D.\n", textures[i]->target);
2248 sampler_type[i] = WINED3DSTT_2D;
2249 }
2250 }
2251 }
2252
2253 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const DWORD *byte_code,
2254 const struct wined3d_shader_signature *output_signature, void *parent,
2255 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2256 {
2257 struct wined3d_shader *object;
2258 HRESULT hr;
2259
2260 TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2261 device, byte_code, output_signature, parent, parent_ops, shader);
2262
2263 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2264 if (!object)
2265 {
2266 ERR("Failed to allocate shader memory.\n");
2267 return E_OUTOFMEMORY;
2268 }
2269
2270 hr = geometryshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2271 if (FAILED(hr))
2272 {
2273 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
2274 HeapFree(GetProcessHeap(), 0, object);
2275 return hr;
2276 }
2277
2278 TRACE("Created geometry shader %p.\n", object);
2279 *shader = object;
2280
2281 return WINED3D_OK;
2282 }
2283
2284 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const DWORD *byte_code,
2285 const struct wined3d_shader_signature *output_signature, void *parent,
2286 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2287 {
2288 struct wined3d_shader *object;
2289 HRESULT hr;
2290
2291 TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2292 device, byte_code, output_signature, parent, parent_ops, shader);
2293
2294 if (device->ps_selected_mode == SHADER_NONE)
2295 return WINED3DERR_INVALIDCALL;
2296
2297 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2298 if (!object)
2299 {
2300 ERR("Failed to allocate shader memory.\n");
2301 return E_OUTOFMEMORY;
2302 }
2303
2304 hr = pixelshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2305 if (FAILED(hr))
2306 {
2307 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
2308 HeapFree(GetProcessHeap(), 0, object);
2309 return hr;
2310 }
2311
2312 TRACE("Created pixel shader %p.\n", object);
2313 *shader = object;
2314
2315 return WINED3D_OK;
2316 }
2317
2318 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const DWORD *byte_code,
2319 const struct wined3d_shader_signature *output_signature, void *parent,
2320 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2321 {
2322 struct wined3d_shader *object;
2323 HRESULT hr;
2324
2325 TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2326 device, byte_code, output_signature, parent, parent_ops, shader);
2327
2328 if (device->vs_selected_mode == SHADER_NONE)
2329 return WINED3DERR_INVALIDCALL;
2330
2331 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2332 if (!object)
2333 {
2334 ERR("Failed to allocate shader memory.\n");
2335 return E_OUTOFMEMORY;
2336 }
2337
2338 hr = vertexshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2339 if (FAILED(hr))
2340 {
2341 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
2342 HeapFree(GetProcessHeap(), 0, object);
2343 return hr;
2344 }
2345
2346 TRACE("Created vertex shader %p.\n", object);
2347 *shader = object;
2348
2349 return WINED3D_OK;
2350 }
Windows API implementation