?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
wined3d: Use the absolute source value in shader_glsl_rsq().
[wine/multimedia.git] / dlls / wined3d / shader.c
blob23de4a747978afc0be1da130cb0d53c9626100e6
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 param_size;
471
472 /* Skip comments. */
473 fe->shader_read_comment(&ptr, &comment);
474 if (comment) continue;
475
476 /* Fetch opcode. */
477 fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
478
479 /* Unhandled opcode, and its parameters. */
480 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
481 {
482 TRACE("Skipping unrecognized instruction.\n");
483 ptr += param_size;
484 continue;
485 }
486
487 /* Handle declarations. */
488 if (ins.handler_idx == WINED3DSIH_DCL)
489 {
490 struct wined3d_shader_semantic semantic;
491
492 fe->shader_read_semantic(&ptr, &semantic);
493
494 switch (semantic.reg.reg.type)
495 {
496 /* Mark input registers used. */
497 case WINED3DSPR_INPUT:
498 reg_maps->input_registers |= 1 << semantic.reg.reg.idx;
499 shader_signature_from_semantic(&input_signature[semantic.reg.reg.idx], &semantic);
500 break;
501
502 /* Vertex shader: mark 3.0 output registers used, save token. */
503 case WINED3DSPR_OUTPUT:
504 reg_maps->output_registers |= 1 << semantic.reg.reg.idx;
505 shader_signature_from_semantic(&output_signature[semantic.reg.reg.idx], &semantic);
506 if (semantic.usage == WINED3DDECLUSAGE_FOG) reg_maps->fog = 1;
507 break;
508
509 /* Save sampler usage token. */
510 case WINED3DSPR_SAMPLER:
511 reg_maps->sampler_type[semantic.reg.reg.idx] = semantic.sampler_type;
512 break;
513
514 default:
515 TRACE("Not recording DCL register type %#x.\n", semantic.reg.reg.type);
516 break;
517 }
518 }
519 else if (ins.handler_idx == WINED3DSIH_DEF)
520 {
521 struct wined3d_shader_src_param rel_addr;
522 struct wined3d_shader_dst_param dst;
523
524 local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
525 if (!lconst) return E_OUTOFMEMORY;
526
527 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
528 lconst->idx = dst.reg.idx;
529
530 memcpy(lconst->value, ptr, 4 * sizeof(DWORD));
531 ptr += 4;
532
533 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
534 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
535 {
536 float *value = (float *)lconst->value;
537 if (value[0] < -1.0f) value[0] = -1.0f;
538 else if (value[0] > 1.0f) value[0] = 1.0f;
539 if (value[1] < -1.0f) value[1] = -1.0f;
540 else if (value[1] > 1.0f) value[1] = 1.0f;
541 if (value[2] < -1.0f) value[2] = -1.0f;
542 else if (value[2] > 1.0f) value[2] = 1.0f;
543 if (value[3] < -1.0f) value[3] = -1.0f;
544 else if (value[3] > 1.0f) value[3] = 1.0f;
545 }
546
547 list_add_head(&shader->baseShader.constantsF, &lconst->entry);
548 }
549 else if (ins.handler_idx == WINED3DSIH_DEFI)
550 {
551 struct wined3d_shader_src_param rel_addr;
552 struct wined3d_shader_dst_param dst;
553
554 local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
555 if (!lconst) return E_OUTOFMEMORY;
556
557 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
558 lconst->idx = dst.reg.idx;
559
560 memcpy(lconst->value, ptr, 4 * sizeof(DWORD));
561 ptr += 4;
562
563 list_add_head(&shader->baseShader.constantsI, &lconst->entry);
564 reg_maps->local_int_consts |= (1 << dst.reg.idx);
565 }
566 else if (ins.handler_idx == WINED3DSIH_DEFB)
567 {
568 struct wined3d_shader_src_param rel_addr;
569 struct wined3d_shader_dst_param dst;
570
571 local_constant *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
572 if (!lconst) return E_OUTOFMEMORY;
573
574 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
575 lconst->idx = dst.reg.idx;
576
577 memcpy(lconst->value, ptr, sizeof(DWORD));
578 ++ptr;
579
580 list_add_head(&shader->baseShader.constantsB, &lconst->entry);
581 reg_maps->local_bool_consts |= (1 << dst.reg.idx);
582 }
583 /* If there's a loop in the shader. */
584 else if (ins.handler_idx == WINED3DSIH_LOOP
585 || ins.handler_idx == WINED3DSIH_REP)
586 {
587 struct wined3d_shader_src_param src, rel_addr;
588
589 fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
590
591 /* Rep and Loop always use an integer constant for the control parameters. */
592 if (ins.handler_idx == WINED3DSIH_REP)
593 {
594 reg_maps->integer_constants |= 1 << src.reg.idx;
595 }
596 else
597 {
598 fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
599 reg_maps->integer_constants |= 1 << src.reg.idx;
600 }
601
602 cur_loop_depth++;
603 if (cur_loop_depth > max_loop_depth) max_loop_depth = cur_loop_depth;
604 }
605 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
606 || ins.handler_idx == WINED3DSIH_ENDREP)
607 {
608 cur_loop_depth--;
609 }
610 /* For subroutine prototypes. */
611 else if (ins.handler_idx == WINED3DSIH_LABEL)
612 {
613 struct wined3d_shader_src_param src, rel_addr;
614
615 fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
616 reg_maps->labels |= 1 << src.reg.idx;
617 }
618 /* Set texture, address, temporary registers. */
619 else
620 {
621 BOOL color0_mov = FALSE;
622 int i, limit;
623
624 /* This will loop over all the registers and try to
625 * make a bitmask of the ones we're interested in.
626 *
627 * Relative addressing tokens are ignored, but that's
628 * okay, since we'll catch any address registers when
629 * they are initialized (required by spec). */
630 if (ins.dst_count)
631 {
632 struct wined3d_shader_src_param dst_rel_addr;
633 struct wined3d_shader_dst_param dst_param;
634
635 fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
636
637 shader_record_register_usage(shader, reg_maps, &dst_param.reg, shader_version.type);
638
639 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and
640 * is used in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel
641 * shaders because TECRDOUT isn't used in them, but future register types might cause issues */
642 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX && shader_version.major < 3
643 && dst_param.reg.type == WINED3DSPR_TEXCRDOUT)
644 {
645 reg_maps->texcoord_mask[dst_param.reg.idx] |= dst_param.write_mask;
646 }
647
648 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
649 {
650 IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)shader;
651
652 if (dst_param.reg.type == WINED3DSPR_COLOROUT && dst_param.reg.idx == 0)
653 {
654 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
655 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
656 * the mov and perform the sRGB write correction from the source register.
657 *
658 * However, if the mov is only partial, we can't do this, and if the write
659 * comes from an instruction other than MOV it is hard to do as well. If
660 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
661
662 ps->color0_mov = FALSE;
663 if (ins.handler_idx == WINED3DSIH_MOV)
664 {
665 /* Used later when the source register is read. */
666 color0_mov = TRUE;
667 }
668 }
669 /* Also drop the MOV marker if the source register is overwritten prior to the shader
670 * end
671 */
672 else if (dst_param.reg.type == WINED3DSPR_TEMP && dst_param.reg.idx == ps->color0_reg)
673 {
674 ps->color0_mov = FALSE;
675 }
676 }
677
678 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
679 if (shader_version.major == 1
680 && (ins.handler_idx == WINED3DSIH_TEX
681 || ins.handler_idx == WINED3DSIH_TEXBEM
682 || ins.handler_idx == WINED3DSIH_TEXBEML
683 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
684 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
685 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
686 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
687 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
688 || ins.handler_idx == WINED3DSIH_TEXREG2AR
689 || ins.handler_idx == WINED3DSIH_TEXREG2GB
690 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
691 {
692 /* Fake sampler usage, only set reserved bit and type. */
693 DWORD sampler_code = dst_param.reg.idx;
694
695 TRACE("Setting fake 2D sampler for 1.x pixelshader.\n");
696 reg_maps->sampler_type[sampler_code] = WINED3DSTT_2D;
697
698 /* texbem is only valid with < 1.4 pixel shaders */
699 if (ins.handler_idx == WINED3DSIH_TEXBEM
700 || ins.handler_idx == WINED3DSIH_TEXBEML)
701 {
702 reg_maps->bumpmat |= 1 << dst_param.reg.idx;
703 if (ins.handler_idx == WINED3DSIH_TEXBEML)
704 {
705 reg_maps->luminanceparams |= 1 << dst_param.reg.idx;
706 }
707 }
708 }
709 else if (ins.handler_idx == WINED3DSIH_BEM)
710 {
711 reg_maps->bumpmat |= 1 << dst_param.reg.idx;
712 }
713 }
714
715 if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
716 else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
717 else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
718 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
719 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
720 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
721 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
722 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
723
724 limit = ins.src_count + (ins.predicate ? 1 : 0);
725 for (i = 0; i < limit; ++i)
726 {
727 struct wined3d_shader_src_param src_param, src_rel_addr;
728 unsigned int count;
729
730 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
731 count = get_instr_extra_regcount(ins.handler_idx, i);
732
733 shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
734 while (count)
735 {
736 ++src_param.reg.idx;
737 shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
738 --count;
739 }
740
741 if (color0_mov)
742 {
743 IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)shader;
744 if (src_param.reg.type == WINED3DSPR_TEMP
745 && src_param.swizzle == WINED3DSP_NOSWIZZLE)
746 {
747 ps->color0_mov = TRUE;
748 ps->color0_reg = src_param.reg.idx;
749 }
750 }
751 }
752 }
753 }
754 reg_maps->loop_depth = max_loop_depth;
755
756 shader->baseShader.functionLength = ((const char *)ptr - (const char *)byte_code);
757
758 return WINED3D_OK;
759 }
760
761 unsigned int shader_find_free_input_register(const struct shader_reg_maps *reg_maps, unsigned int max)
762 {
763 DWORD map = 1 << max;
764 map |= map - 1;
765 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
766
767 return wined3d_log2i(map);
768 }
769
770 static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
771 const struct wined3d_shader_version *shader_version)
772 {
773 TRACE("dcl");
774
775 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
776 {
777 switch (semantic->sampler_type)
778 {
779 case WINED3DSTT_2D: TRACE("_2d"); break;
780 case WINED3DSTT_CUBE: TRACE("_cube"); break;
781 case WINED3DSTT_VOLUME: TRACE("_volume"); break;
782 default: TRACE("_unknown_ttype(0x%08x)", semantic->sampler_type);
783 }
784 }
785 else
786 {
787 /* Pixel shaders 3.0 don't have usage semantics. */
788 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL) return;
789 else TRACE("_");
790
791 switch (semantic->usage)
792 {
793 case WINED3DDECLUSAGE_POSITION:
794 TRACE("position%u", semantic->usage_idx);
795 break;
796
797 case WINED3DDECLUSAGE_BLENDINDICES:
798 TRACE("blend");
799 break;
800
801 case WINED3DDECLUSAGE_BLENDWEIGHT:
802 TRACE("weight");
803 break;
804
805 case WINED3DDECLUSAGE_NORMAL:
806 TRACE("normal%u", semantic->usage_idx);
807 break;
808
809 case WINED3DDECLUSAGE_PSIZE:
810 TRACE("psize");
811 break;
812
813 case WINED3DDECLUSAGE_COLOR:
814 if (semantic->usage_idx == 0) TRACE("color");
815 else TRACE("specular%u", (semantic->usage_idx - 1));
816 break;
817
818 case WINED3DDECLUSAGE_TEXCOORD:
819 TRACE("texture%u", semantic->usage_idx);
820 break;
821
822 case WINED3DDECLUSAGE_TANGENT:
823 TRACE("tangent");
824 break;
825
826 case WINED3DDECLUSAGE_BINORMAL:
827 TRACE("binormal");
828 break;
829
830 case WINED3DDECLUSAGE_TESSFACTOR:
831 TRACE("tessfactor");
832 break;
833
834 case WINED3DDECLUSAGE_POSITIONT:
835 TRACE("positionT%u", semantic->usage_idx);
836 break;
837
838 case WINED3DDECLUSAGE_FOG:
839 TRACE("fog");
840 break;
841
842 case WINED3DDECLUSAGE_DEPTH:
843 TRACE("depth");
844 break;
845
846 case WINED3DDECLUSAGE_SAMPLE:
847 TRACE("sample");
848 break;
849
850 default:
851 FIXME("unknown_semantics(0x%08x)", semantic->usage);
852 }
853 }
854 }
855
856 static void shader_dump_register(const struct wined3d_shader_register *reg,
857 const struct wined3d_shader_version *shader_version)
858 {
859 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
860 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
861 UINT offset = reg->idx;
862
863 switch (reg->type)
864 {
865 case WINED3DSPR_TEMP:
866 TRACE("r");
867 break;
868
869 case WINED3DSPR_INPUT:
870 TRACE("v");
871 break;
872
873 case WINED3DSPR_CONST:
874 case WINED3DSPR_CONST2:
875 case WINED3DSPR_CONST3:
876 case WINED3DSPR_CONST4:
877 TRACE("c");
878 offset = shader_get_float_offset(reg->type, reg->idx);
879 break;
880
881 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
882 TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
883 break;
884
885 case WINED3DSPR_RASTOUT:
886 TRACE("%s", rastout_reg_names[reg->idx]);
887 break;
888
889 case WINED3DSPR_COLOROUT:
890 TRACE("oC");
891 break;
892
893 case WINED3DSPR_DEPTHOUT:
894 TRACE("oDepth");
895 break;
896
897 case WINED3DSPR_ATTROUT:
898 TRACE("oD");
899 break;
900
901 case WINED3DSPR_TEXCRDOUT:
902 /* Vertex shaders >= 3.0 use general purpose output registers
903 * (WINED3DSPR_OUTPUT), which can include an address token. */
904 if (shader_version->major >= 3) TRACE("o");
905 else TRACE("oT");
906 break;
907
908 case WINED3DSPR_CONSTINT:
909 TRACE("i");
910 break;
911
912 case WINED3DSPR_CONSTBOOL:
913 TRACE("b");
914 break;
915
916 case WINED3DSPR_LABEL:
917 TRACE("l");
918 break;
919
920 case WINED3DSPR_LOOP:
921 TRACE("aL");
922 break;
923
924 case WINED3DSPR_SAMPLER:
925 TRACE("s");
926 break;
927
928 case WINED3DSPR_MISCTYPE:
929 if (reg->idx > 1) FIXME("Unhandled misctype register %u.\n", reg->idx);
930 else TRACE("%s", misctype_reg_names[reg->idx]);
931 break;
932
933 case WINED3DSPR_PREDICATE:
934 TRACE("p");
935 break;
936
937 case WINED3DSPR_IMMCONST:
938 TRACE("l");
939 break;
940
941 case WINED3DSPR_CONSTBUFFER:
942 TRACE("cb");
943 break;
944
945 default:
946 TRACE("unhandled_rtype(%#x)", reg->type);
947 break;
948 }
949
950 if (reg->type == WINED3DSPR_IMMCONST)
951 {
952 TRACE("(");
953 switch (reg->immconst_type)
954 {
955 case WINED3D_IMMCONST_FLOAT:
956 TRACE("%.8e", *(const float *)reg->immconst_data);
957 break;
958
959 case WINED3D_IMMCONST_FLOAT4:
960 TRACE("%.8e, %.8e, %.8e, %.8e",
961 *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
962 *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
963 break;
964
965 default:
966 TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
967 break;
968 }
969 TRACE(")");
970 }
971 else if (reg->type != WINED3DSPR_RASTOUT && reg->type != WINED3DSPR_MISCTYPE)
972 {
973 if (reg->array_idx != ~0U)
974 {
975 TRACE("%u[%u", offset, reg->array_idx);
976 if (reg->rel_addr)
977 {
978 TRACE(" + ");
979 shader_dump_src_param(reg->rel_addr, shader_version);
980 }
981 TRACE("]");
982 }
983 else
984 {
985 if (reg->rel_addr)
986 {
987 TRACE("[");
988 shader_dump_src_param(reg->rel_addr, shader_version);
989 TRACE(" + ");
990 }
991 TRACE("%u", offset);
992 if (reg->rel_addr) TRACE("]");
993 }
994 }
995 }
996
997 void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
998 const struct wined3d_shader_version *shader_version)
999 {
1000 DWORD write_mask = param->write_mask;
1001
1002 shader_dump_register(&param->reg, shader_version);
1003
1004 if (write_mask != WINED3DSP_WRITEMASK_ALL)
1005 {
1006 static const char *write_mask_chars = "xyzw";
1007
1008 TRACE(".");
1009 if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
1010 if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
1011 if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
1012 if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
1013 }
1014 }
1015
1016 void shader_dump_src_param(const struct wined3d_shader_src_param *param,
1017 const struct wined3d_shader_version *shader_version)
1018 {
1019 DWORD src_modifier = param->modifiers;
1020 DWORD swizzle = param->swizzle;
1021
1022 if (src_modifier == WINED3DSPSM_NEG
1023 || src_modifier == WINED3DSPSM_BIASNEG
1024 || src_modifier == WINED3DSPSM_SIGNNEG
1025 || src_modifier == WINED3DSPSM_X2NEG
1026 || src_modifier == WINED3DSPSM_ABSNEG)
1027 TRACE("-");
1028 else if (src_modifier == WINED3DSPSM_COMP)
1029 TRACE("1-");
1030 else if (src_modifier == WINED3DSPSM_NOT)
1031 TRACE("!");
1032
1033 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1034 TRACE("abs(");
1035
1036 shader_dump_register(&param->reg, shader_version);
1037
1038 if (src_modifier)
1039 {
1040 switch (src_modifier)
1041 {
1042 case WINED3DSPSM_NONE: break;
1043 case WINED3DSPSM_NEG: break;
1044 case WINED3DSPSM_NOT: break;
1045 case WINED3DSPSM_BIAS: TRACE("_bias"); break;
1046 case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
1047 case WINED3DSPSM_SIGN: TRACE("_bx2"); break;
1048 case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1049 case WINED3DSPSM_COMP: break;
1050 case WINED3DSPSM_X2: TRACE("_x2"); break;
1051 case WINED3DSPSM_X2NEG: TRACE("_x2"); break;
1052 case WINED3DSPSM_DZ: TRACE("_dz"); break;
1053 case WINED3DSPSM_DW: TRACE("_dw"); break;
1054 case WINED3DSPSM_ABSNEG: TRACE(")"); break;
1055 case WINED3DSPSM_ABS: TRACE(")"); break;
1056 default: TRACE("_unknown_modifier(%#x)", src_modifier);
1057 }
1058 }
1059
1060 if (swizzle != WINED3DSP_NOSWIZZLE)
1061 {
1062 static const char *swizzle_chars = "xyzw";
1063 DWORD swizzle_x = swizzle & 0x03;
1064 DWORD swizzle_y = (swizzle >> 2) & 0x03;
1065 DWORD swizzle_z = (swizzle >> 4) & 0x03;
1066 DWORD swizzle_w = (swizzle >> 6) & 0x03;
1067
1068 if (swizzle_x == swizzle_y
1069 && swizzle_x == swizzle_z
1070 && swizzle_x == swizzle_w)
1071 {
1072 TRACE(".%c", swizzle_chars[swizzle_x]);
1073 }
1074 else
1075 {
1076 TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1077 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1078 }
1079 }
1080 }
1081
1082 /* Shared code in order to generate the bulk of the shader string.
1083 * NOTE: A description of how to parse tokens can be found on MSDN. */
1084 void shader_generate_main(IWineD3DBaseShader *iface, struct wined3d_shader_buffer *buffer,
1085 const shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1086 {
1087 IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)iface;
1088 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1089 const struct wined3d_shader_frontend *fe = shader->baseShader.frontend;
1090 void *fe_data = shader->baseShader.frontend_data;
1091 struct wined3d_shader_src_param src_rel_addr[4];
1092 struct wined3d_shader_src_param src_param[4];
1093 struct wined3d_shader_version shader_version;
1094 struct wined3d_shader_src_param dst_rel_addr;
1095 struct wined3d_shader_dst_param dst_param;
1096 struct wined3d_shader_instruction ins;
1097 struct wined3d_shader_context ctx;
1098 const DWORD *ptr = byte_code;
1099 DWORD i;
1100
1101 /* Initialize current parsing state. */
1102 ctx.shader = iface;
1103 ctx.gl_info = &device->adapter->gl_info;
1104 ctx.reg_maps = reg_maps;
1105 ctx.buffer = buffer;
1106 ctx.backend_data = backend_ctx;
1107
1108 ins.ctx = &ctx;
1109 ins.dst = &dst_param;
1110 ins.src = src_param;
1111 shader->baseShader.parse_state.current_row = 0;
1112
1113 fe->shader_read_header(fe_data, &ptr, &shader_version);
1114
1115 while (!fe->shader_is_end(fe_data, &ptr))
1116 {
1117 const char *comment;
1118 UINT param_size;
1119
1120 /* Skip comment tokens. */
1121 fe->shader_read_comment(&ptr, &comment);
1122 if (comment) continue;
1123
1124 /* Read opcode. */
1125 fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1126
1127 /* Unknown opcode and its parameters. */
1128 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1129 {
1130 TRACE("Skipping unrecognized instruction.\n");
1131 ptr += param_size;
1132 continue;
1133 }
1134
1135 /* Nothing to do. */
1136 if (ins.handler_idx == WINED3DSIH_DCL
1137 || ins.handler_idx == WINED3DSIH_NOP
1138 || ins.handler_idx == WINED3DSIH_DEF
1139 || ins.handler_idx == WINED3DSIH_DEFI
1140 || ins.handler_idx == WINED3DSIH_DEFB
1141 || ins.handler_idx == WINED3DSIH_PHASE)
1142 {
1143 ptr += param_size;
1144 continue;
1145 }
1146
1147 /* Destination token */
1148 if (ins.dst_count) fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
1149
1150 /* Predication token */
1151 if (ins.predicate) ins.predicate = *ptr++;
1152
1153 /* Other source tokens */
1154 for (i = 0; i < ins.src_count; ++i)
1155 {
1156 fe->shader_read_src_param(fe_data, &ptr, &src_param[i], &src_rel_addr[i]);
1157 }
1158
1159 /* Call appropriate function for output target */
1160 device->shader_backend->shader_handle_instruction(&ins);
1161 }
1162 }
1163
1164 static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1165 {
1166 DWORD mmask = dst->modifiers;
1167
1168 switch (dst->shift)
1169 {
1170 case 0: break;
1171 case 13: TRACE("_d8"); break;
1172 case 14: TRACE("_d4"); break;
1173 case 15: TRACE("_d2"); break;
1174 case 1: TRACE("_x2"); break;
1175 case 2: TRACE("_x4"); break;
1176 case 3: TRACE("_x8"); break;
1177 default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1178 }
1179
1180 if (mmask & WINED3DSPDM_SATURATE) TRACE("_sat");
1181 if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1182 if (mmask & WINED3DSPDM_MSAMPCENTROID) TRACE("_centroid");
1183
1184 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1185 if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1186 }
1187
1188 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1189 {
1190 struct wined3d_shader_version shader_version;
1191 const DWORD *ptr = byte_code;
1192 const char *type_prefix;
1193 DWORD i;
1194
1195 TRACE("Parsing %p.\n", byte_code);
1196
1197 fe->shader_read_header(fe_data, &ptr, &shader_version);
1198
1199 switch (shader_version.type)
1200 {
1201 case WINED3D_SHADER_TYPE_VERTEX:
1202 type_prefix = "vs";
1203 break;
1204
1205 case WINED3D_SHADER_TYPE_GEOMETRY:
1206 type_prefix = "gs";
1207 break;
1208
1209 case WINED3D_SHADER_TYPE_PIXEL:
1210 type_prefix = "ps";
1211 break;
1212
1213 default:
1214 FIXME("Unhandled shader type %#x.\n", shader_version.type);
1215 type_prefix = "unknown";
1216 break;
1217 }
1218
1219 TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1220
1221 while (!fe->shader_is_end(fe_data, &ptr))
1222 {
1223 struct wined3d_shader_instruction ins;
1224 const char *comment;
1225 UINT param_size;
1226
1227 /* comment */
1228 fe->shader_read_comment(&ptr, &comment);
1229 if (comment)
1230 {
1231 TRACE("//%s\n", comment);
1232 continue;
1233 }
1234
1235 fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1236 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1237 {
1238 TRACE("Skipping unrecognized instruction.\n");
1239 ptr += param_size;
1240 continue;
1241 }
1242
1243 if (ins.handler_idx == WINED3DSIH_DCL)
1244 {
1245 struct wined3d_shader_semantic semantic;
1246
1247 fe->shader_read_semantic(&ptr, &semantic);
1248
1249 shader_dump_decl_usage(&semantic, &shader_version);
1250 shader_dump_ins_modifiers(&semantic.reg);
1251 TRACE(" ");
1252 shader_dump_dst_param(&semantic.reg, &shader_version);
1253 }
1254 else if (ins.handler_idx == WINED3DSIH_DEF)
1255 {
1256 struct wined3d_shader_src_param rel_addr;
1257 struct wined3d_shader_dst_param dst;
1258
1259 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1260
1261 TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(dst.reg.type, dst.reg.idx),
1262 *(const float *)(ptr),
1263 *(const float *)(ptr + 1),
1264 *(const float *)(ptr + 2),
1265 *(const float *)(ptr + 3));
1266 ptr += 4;
1267 }
1268 else if (ins.handler_idx == WINED3DSIH_DEFI)
1269 {
1270 struct wined3d_shader_src_param rel_addr;
1271 struct wined3d_shader_dst_param dst;
1272
1273 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1274
1275 TRACE("defi i%u = %d, %d, %d, %d", dst.reg.idx,
1276 *(ptr),
1277 *(ptr + 1),
1278 *(ptr + 2),
1279 *(ptr + 3));
1280 ptr += 4;
1281 }
1282 else if (ins.handler_idx == WINED3DSIH_DEFB)
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("defb b%u = %s", dst.reg.idx, *ptr ? "true" : "false");
1290 ++ptr;
1291 }
1292 else
1293 {
1294 struct wined3d_shader_src_param dst_rel_addr, src_rel_addr;
1295 struct wined3d_shader_dst_param dst_param;
1296 struct wined3d_shader_src_param src_param;
1297
1298 if (ins.dst_count)
1299 {
1300 fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
1301 }
1302
1303 /* Print out predication source token first - it follows
1304 * the destination token. */
1305 if (ins.predicate)
1306 {
1307 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1308 TRACE("(");
1309 shader_dump_src_param(&src_param, &shader_version);
1310 TRACE(") ");
1311 }
1312
1313 /* PixWin marks instructions with the coissue flag with a '+' */
1314 if (ins.coissue) TRACE("+");
1315
1316 TRACE("%s", shader_opcode_names[ins.handler_idx]);
1317
1318 if (ins.handler_idx == WINED3DSIH_IFC
1319 || ins.handler_idx == WINED3DSIH_BREAKC)
1320 {
1321 switch (ins.flags)
1322 {
1323 case COMPARISON_GT: TRACE("_gt"); break;
1324 case COMPARISON_EQ: TRACE("_eq"); break;
1325 case COMPARISON_GE: TRACE("_ge"); break;
1326 case COMPARISON_LT: TRACE("_lt"); break;
1327 case COMPARISON_NE: TRACE("_ne"); break;
1328 case COMPARISON_LE: TRACE("_le"); break;
1329 default: TRACE("_(%u)", ins.flags);
1330 }
1331 }
1332 else if (ins.handler_idx == WINED3DSIH_TEX
1333 && shader_version.major >= 2
1334 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1335 {
1336 TRACE("p");
1337 }
1338
1339 /* We already read the destination token, print it. */
1340 if (ins.dst_count)
1341 {
1342 shader_dump_ins_modifiers(&dst_param);
1343 TRACE(" ");
1344 shader_dump_dst_param(&dst_param, &shader_version);
1345 }
1346
1347 /* Other source tokens */
1348 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1349 {
1350 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1351 TRACE(!i ? " " : ", ");
1352 shader_dump_src_param(&src_param, &shader_version);
1353 }
1354 }
1355 TRACE("\n");
1356 }
1357 }
1358
1359 static void shader_cleanup(IWineD3DBaseShader *iface)
1360 {
1361 IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)iface;
1362
1363 ((IWineD3DDeviceImpl *)shader->baseShader.device)->shader_backend->shader_destroy(iface);
1364 HeapFree(GetProcessHeap(), 0, shader->baseShader.reg_maps.constf);
1365 HeapFree(GetProcessHeap(), 0, shader->baseShader.function);
1366 shader_delete_constant_list(&shader->baseShader.constantsF);
1367 shader_delete_constant_list(&shader->baseShader.constantsB);
1368 shader_delete_constant_list(&shader->baseShader.constantsI);
1369 list_remove(&shader->baseShader.shader_list_entry);
1370
1371 if (shader->baseShader.frontend && shader->baseShader.frontend_data)
1372 {
1373 shader->baseShader.frontend->shader_free(shader->baseShader.frontend_data);
1374 }
1375 }
1376
1377 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1378 static void shader_none_select(const struct wined3d_context *context, BOOL usePS, BOOL useVS) {}
1379 static void shader_none_select_depth_blt(IWineD3DDevice *iface, enum tex_types tex_type) {}
1380 static void shader_none_deselect_depth_blt(IWineD3DDevice *iface) {}
1381 static void shader_none_update_float_vertex_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1382 static void shader_none_update_float_pixel_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1383 static void shader_none_load_constants(const struct wined3d_context *context, char usePS, char useVS) {}
1384 static void shader_none_load_np2fixup_constants(IWineD3DDevice *iface, char usePS, char useVS) {}
1385 static void shader_none_destroy(IWineD3DBaseShader *iface) {}
1386 static HRESULT shader_none_alloc(IWineD3DDevice *iface) {return WINED3D_OK;}
1387 static void shader_none_free(IWineD3DDevice *iface) {}
1388 static BOOL shader_none_dirty_const(IWineD3DDevice *iface) {return FALSE;}
1389
1390 static void shader_none_get_caps(WINED3DDEVTYPE devtype,
1391 const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
1392 {
1393 /* Set the shader caps to 0 for the none shader backend */
1394 caps->VertexShaderVersion = 0;
1395 caps->PixelShaderVersion = 0;
1396 caps->PixelShader1xMaxValue = 0.0f;
1397 }
1398
1399 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
1400 {
1401 if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1402 {
1403 TRACE("Checking support for fixup:\n");
1404 dump_color_fixup_desc(fixup);
1405 }
1406
1407 /* Faked to make some apps happy. */
1408 if (!is_yuv_fixup(fixup))
1409 {
1410 TRACE("[OK]\n");
1411 return TRUE;
1412 }
1413
1414 TRACE("[FAILED]\n");
1415 return FALSE;
1416 }
1417
1418 const shader_backend_t none_shader_backend = {
1419 shader_none_handle_instruction,
1420 shader_none_select,
1421 shader_none_select_depth_blt,
1422 shader_none_deselect_depth_blt,
1423 shader_none_update_float_vertex_constants,
1424 shader_none_update_float_pixel_constants,
1425 shader_none_load_constants,
1426 shader_none_load_np2fixup_constants,
1427 shader_none_destroy,
1428 shader_none_alloc,
1429 shader_none_free,
1430 shader_none_dirty_const,
1431 shader_none_get_caps,
1432 shader_none_color_fixup_supported,
1433 };
1434
1435 static void shader_get_parent(IWineD3DBaseShaderImpl *shader, IUnknown **parent)
1436 {
1437 *parent = shader->baseShader.parent;
1438 IUnknown_AddRef(*parent);
1439 TRACE("shader %p, returning %p.\n", shader, *parent);
1440 }
1441
1442 static HRESULT shader_get_function(IWineD3DBaseShaderImpl *shader, void *data, UINT *data_size)
1443 {
1444 if (!data)
1445 {
1446 *data_size = shader->baseShader.functionLength;
1447 return WINED3D_OK;
1448 }
1449
1450 if (*data_size < shader->baseShader.functionLength)
1451 {
1452 /* MSDN claims (for d3d8 at least) that if *pSizeOfData is smaller
1453 * than the required size we should write the required size and
1454 * return D3DERR_MOREDATA. That's not actually true. */
1455 return WINED3DERR_INVALIDCALL;
1456 }
1457
1458 memcpy(data, shader->baseShader.function, shader->baseShader.functionLength);
1459
1460 return WINED3D_OK;
1461 }
1462
1463 static HRESULT shader_set_function(IWineD3DBaseShaderImpl *shader, const DWORD *byte_code,
1464 const struct wined3d_shader_signature *output_signature, DWORD float_const_count)
1465 {
1466 struct shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
1467 const struct wined3d_shader_frontend *fe;
1468 HRESULT hr;
1469
1470 TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u.\n",
1471 shader, byte_code, output_signature, float_const_count);
1472
1473 fe = shader_select_frontend(*byte_code);
1474 if (!fe)
1475 {
1476 FIXME("Unable to find frontend for shader.\n");
1477 return WINED3DERR_INVALIDCALL;
1478 }
1479 shader->baseShader.frontend = fe;
1480 shader->baseShader.frontend_data = fe->shader_init(byte_code, output_signature);
1481 if (!shader->baseShader.frontend_data)
1482 {
1483 FIXME("Failed to initialize frontend.\n");
1484 return WINED3DERR_INVALIDCALL;
1485 }
1486
1487 /* First pass: trace shader. */
1488 if (TRACE_ON(d3d_shader)) shader_trace_init(fe, shader->baseShader.frontend_data, byte_code);
1489
1490 /* Initialize immediate constant lists. */
1491 list_init(&shader->baseShader.constantsF);
1492 list_init(&shader->baseShader.constantsB);
1493 list_init(&shader->baseShader.constantsI);
1494
1495 /* Second pass: figure out which registers are used, what the semantics are, etc. */
1496 hr = shader_get_registers_used((IWineD3DBaseShader *)shader, fe,
1497 reg_maps, shader->baseShader.input_signature, shader->baseShader.output_signature,
1498 byte_code, float_const_count);
1499 if (FAILED(hr)) return hr;
1500
1501 shader->baseShader.function = HeapAlloc(GetProcessHeap(), 0, shader->baseShader.functionLength);
1502 if (!shader->baseShader.function) return E_OUTOFMEMORY;
1503 memcpy(shader->baseShader.function, byte_code, shader->baseShader.functionLength);
1504
1505 return WINED3D_OK;
1506 }
1507
1508 static HRESULT STDMETHODCALLTYPE vertexshader_QueryInterface(IWineD3DVertexShader *iface, REFIID riid, void **object)
1509 {
1510 TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1511
1512 if (IsEqualGUID(riid, &IID_IWineD3DVertexShader)
1513 || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1514 || IsEqualGUID(riid, &IID_IWineD3DBase)
1515 || IsEqualGUID(riid, &IID_IUnknown))
1516 {
1517 IUnknown_AddRef(iface);
1518 *object = iface;
1519 return S_OK;
1520 }
1521
1522 WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1523
1524 *object = NULL;
1525 return E_NOINTERFACE;
1526 }
1527
1528 static ULONG STDMETHODCALLTYPE vertexshader_AddRef(IWineD3DVertexShader *iface)
1529 {
1530 IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1531 ULONG refcount = InterlockedIncrement(&shader->baseShader.ref);
1532
1533 TRACE("%p increasing refcount to %u.\n", shader, refcount);
1534
1535 return refcount;
1536 }
1537
1538 static ULONG STDMETHODCALLTYPE vertexshader_Release(IWineD3DVertexShader *iface)
1539 {
1540 IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1541 ULONG refcount = InterlockedDecrement(&shader->baseShader.ref);
1542
1543 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1544
1545 if (!refcount)
1546 {
1547 shader_cleanup((IWineD3DBaseShader *)iface);
1548 shader->baseShader.parent_ops->wined3d_object_destroyed(shader->baseShader.parent);
1549 HeapFree(GetProcessHeap(), 0, shader);
1550 }
1551
1552 return refcount;
1553 }
1554
1555 static HRESULT STDMETHODCALLTYPE vertexshader_GetParent(IWineD3DVertexShader *iface, IUnknown **parent)
1556 {
1557 TRACE("iface %p, parent %p.\n", iface, parent);
1558
1559 shader_get_parent((IWineD3DBaseShaderImpl *)iface, parent);
1560
1561 return WINED3D_OK;
1562 }
1563
1564 static HRESULT STDMETHODCALLTYPE vertexshader_GetFunction(IWineD3DVertexShader *iface, void *data, UINT *data_size)
1565 {
1566 TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1567
1568 return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1569 }
1570
1571 /* Set local constants for d3d8 shaders. */
1572 static HRESULT STDMETHODCALLTYPE vertexshader_SetLocalConstantsF(IWineD3DVertexShader *iface,
1573 UINT start_idx, const float *src_data, UINT count)
1574 {
1575 IWineD3DVertexShaderImpl *shader =(IWineD3DVertexShaderImpl *)iface;
1576 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1577 UINT i, end_idx;
1578
1579 TRACE("iface %p, start_idx %u, src_data %p, count %u.\n", iface, start_idx, src_data, count);
1580
1581 end_idx = start_idx + count;
1582 if (end_idx > device->d3d_vshader_constantF)
1583 {
1584 WARN("end_idx %u > float constants limit %u.\n", end_idx, device->d3d_vshader_constantF);
1585 end_idx = device->d3d_vshader_constantF;
1586 }
1587
1588 for (i = start_idx; i < end_idx; ++i)
1589 {
1590 local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
1591 if (!lconst) return E_OUTOFMEMORY;
1592
1593 lconst->idx = i;
1594 memcpy(lconst->value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
1595 list_add_head(&shader->baseShader.constantsF, &lconst->entry);
1596 }
1597
1598 return WINED3D_OK;
1599 }
1600
1601 static const IWineD3DVertexShaderVtbl IWineD3DVertexShader_Vtbl =
1602 {
1603 /* IUnknown methods */
1604 vertexshader_QueryInterface,
1605 vertexshader_AddRef,
1606 vertexshader_Release,
1607 /* IWineD3DBase methods */
1608 vertexshader_GetParent,
1609 /* IWineD3DBaseShader methods */
1610 vertexshader_GetFunction,
1611 /* IWineD3DVertexShader methods */
1612 vertexshader_SetLocalConstantsF,
1613 };
1614
1615 void find_vs_compile_args(IWineD3DVertexShaderImpl *shader,
1616 IWineD3DStateBlockImpl *stateblock, struct vs_compile_args *args)
1617 {
1618 args->fog_src = stateblock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
1619 args->clip_enabled = stateblock->renderState[WINED3DRS_CLIPPING]
1620 && stateblock->renderState[WINED3DRS_CLIPPLANEENABLE];
1621 args->swizzle_map = ((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.swizzle_map;
1622 }
1623
1624 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
1625 {
1626 if (usage_idx1 != usage_idx2) return FALSE;
1627 if (usage1 == usage2) return TRUE;
1628 if (usage1 == WINED3DDECLUSAGE_POSITION && usage2 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1629 if (usage2 == WINED3DDECLUSAGE_POSITION && usage1 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1630
1631 return FALSE;
1632 }
1633
1634 BOOL vshader_get_input(IWineD3DVertexShader *iface, BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
1635 {
1636 IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1637 WORD map = shader->baseShader.reg_maps.input_registers;
1638 unsigned int i;
1639
1640 for (i = 0; map; map >>= 1, ++i)
1641 {
1642 if (!(map & 1)) continue;
1643
1644 if (match_usage(shader->attributes[i].usage,
1645 shader->attributes[i].usage_idx, usage_req, usage_idx_req))
1646 {
1647 *regnum = i;
1648 return TRUE;
1649 }
1650 }
1651 return FALSE;
1652 }
1653
1654 static void vertexshader_set_limits(IWineD3DVertexShaderImpl *shader)
1655 {
1656 DWORD shader_version = WINED3D_SHADER_VERSION(shader->baseShader.reg_maps.shader_version.major,
1657 shader->baseShader.reg_maps.shader_version.minor);
1658 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1659
1660 shader->baseShader.limits.texcoord = 0;
1661 shader->baseShader.limits.attributes = 16;
1662 shader->baseShader.limits.packed_input = 0;
1663
1664 switch (shader_version)
1665 {
1666 case WINED3D_SHADER_VERSION(1, 0):
1667 case WINED3D_SHADER_VERSION(1, 1):
1668 shader->baseShader.limits.temporary = 12;
1669 shader->baseShader.limits.constant_bool = 0;
1670 shader->baseShader.limits.constant_int = 0;
1671 shader->baseShader.limits.address = 1;
1672 shader->baseShader.limits.packed_output = 0;
1673 shader->baseShader.limits.sampler = 0;
1674 shader->baseShader.limits.label = 0;
1675 /* TODO: vs_1_1 has a minimum of 96 constants. What happens when
1676 * a vs_1_1 shader is used on a vs_3_0 capable card that has 256
1677 * constants? */
1678 shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1679 break;
1680
1681 case WINED3D_SHADER_VERSION(2, 0):
1682 case WINED3D_SHADER_VERSION(2, 1):
1683 shader->baseShader.limits.temporary = 12;
1684 shader->baseShader.limits.constant_bool = 16;
1685 shader->baseShader.limits.constant_int = 16;
1686 shader->baseShader.limits.address = 1;
1687 shader->baseShader.limits.packed_output = 0;
1688 shader->baseShader.limits.sampler = 0;
1689 shader->baseShader.limits.label = 16;
1690 shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1691 break;
1692
1693 case WINED3D_SHADER_VERSION(4, 0):
1694 FIXME("Using 3.0 limits for 4.0 shader.\n");
1695 /* Fall through. */
1696
1697 case WINED3D_SHADER_VERSION(3, 0):
1698 shader->baseShader.limits.temporary = 32;
1699 shader->baseShader.limits.constant_bool = 32;
1700 shader->baseShader.limits.constant_int = 32;
1701 shader->baseShader.limits.address = 1;
1702 shader->baseShader.limits.packed_output = 12;
1703 shader->baseShader.limits.sampler = 4;
1704 shader->baseShader.limits.label = 16; /* FIXME: 2048 */
1705 /* DX10 cards on Windows advertise a d3d9 constant limit of 256
1706 * even though they are capable of supporting much more (GL
1707 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
1708 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
1709 * shaders to 256. */
1710 shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1711 break;
1712
1713 default:
1714 shader->baseShader.limits.temporary = 12;
1715 shader->baseShader.limits.constant_bool = 16;
1716 shader->baseShader.limits.constant_int = 16;
1717 shader->baseShader.limits.address = 1;
1718 shader->baseShader.limits.packed_output = 0;
1719 shader->baseShader.limits.sampler = 0;
1720 shader->baseShader.limits.label = 16;
1721 shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1722 FIXME("Unrecognized vertex shader version \"%u.%u\".\n",
1723 shader->baseShader.reg_maps.shader_version.major,
1724 shader->baseShader.reg_maps.shader_version.minor);
1725 }
1726 }
1727
1728 HRESULT vertexshader_init(IWineD3DVertexShaderImpl *shader, IWineD3DDeviceImpl *device,
1729 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1730 IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
1731 {
1732 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
1733 struct shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
1734 unsigned int i;
1735 HRESULT hr;
1736 WORD map;
1737
1738 if (!byte_code) return WINED3DERR_INVALIDCALL;
1739
1740 shader->lpVtbl = &IWineD3DVertexShader_Vtbl;
1741 shader_init(&shader->baseShader, device, parent, parent_ops);
1742
1743 hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code,
1744 output_signature, device->d3d_vshader_constantF);
1745 if (FAILED(hr))
1746 {
1747 WARN("Failed to set function, hr %#x.\n", hr);
1748 shader_cleanup((IWineD3DBaseShader *)shader);
1749 return hr;
1750 }
1751
1752 map = shader->baseShader.reg_maps.input_registers;
1753 for (i = 0; map; map >>= 1, ++i)
1754 {
1755 if (!(map & 1) || !shader->baseShader.input_signature[i].semantic_name) continue;
1756
1757 shader->attributes[i].usage =
1758 shader_usage_from_semantic_name(shader->baseShader.input_signature[i].semantic_name);
1759 shader->attributes[i].usage_idx = shader->baseShader.input_signature[i].semantic_idx;
1760 }
1761
1762 if (output_signature)
1763 {
1764 for (i = 0; i < output_signature->element_count; ++i)
1765 {
1766 struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1767 reg_maps->output_registers |= 1 << e->register_idx;
1768 shader->baseShader.output_signature[e->register_idx] = *e;
1769 }
1770 }
1771
1772 vertexshader_set_limits(shader);
1773
1774 if (device->vs_selected_mode == SHADER_ARB
1775 && (gl_info->quirks & WINED3D_QUIRK_ARB_VS_OFFSET_LIMIT)
1776 && shader->min_rel_offset <= shader->max_rel_offset)
1777 {
1778 if (shader->max_rel_offset - shader->min_rel_offset > 127)
1779 {
1780 FIXME("The difference between the minimum and maximum relative offset is > 127.\n");
1781 FIXME("Which this OpenGL implementation does not support. Try using GLSL.\n");
1782 FIXME("Min: %d, Max: %d.\n", shader->min_rel_offset, shader->max_rel_offset);
1783 }
1784 else if (shader->max_rel_offset - shader->min_rel_offset > 63)
1785 {
1786 shader->rel_offset = shader->min_rel_offset + 63;
1787 }
1788 else if (shader->max_rel_offset > 63)
1789 {
1790 shader->rel_offset = shader->min_rel_offset;
1791 }
1792 else
1793 {
1794 shader->rel_offset = 0;
1795 }
1796 }
1797
1798 shader->baseShader.load_local_constsF = shader->baseShader.reg_maps.usesrelconstF
1799 && !list_empty(&shader->baseShader.constantsF);
1800
1801 return WINED3D_OK;
1802 }
1803
1804 static HRESULT STDMETHODCALLTYPE geometryshader_QueryInterface(IWineD3DGeometryShader *iface,
1805 REFIID riid, void **object)
1806 {
1807 TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1808
1809 if (IsEqualGUID(riid, &IID_IWineD3DGeometryShader)
1810 || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1811 || IsEqualGUID(riid, &IID_IWineD3DBase)
1812 || IsEqualGUID(riid, &IID_IUnknown))
1813 {
1814 IUnknown_AddRef(iface);
1815 *object = iface;
1816 return S_OK;
1817 }
1818
1819 WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1820
1821 *object = NULL;
1822 return E_NOINTERFACE;
1823 }
1824
1825 static ULONG STDMETHODCALLTYPE geometryshader_AddRef(IWineD3DGeometryShader *iface)
1826 {
1827 struct wined3d_geometryshader *shader = (struct wined3d_geometryshader *)iface;
1828 ULONG refcount = InterlockedIncrement(&shader->base_shader.ref);
1829
1830 TRACE("%p increasing refcount to %u.\n", shader, refcount);
1831
1832 return refcount;
1833 }
1834
1835 static ULONG STDMETHODCALLTYPE geometryshader_Release(IWineD3DGeometryShader *iface)
1836 {
1837 struct wined3d_geometryshader *shader = (struct wined3d_geometryshader *)iface;
1838 ULONG refcount = InterlockedDecrement(&shader->base_shader.ref);
1839
1840 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1841
1842 if (!refcount)
1843 {
1844 shader_cleanup((IWineD3DBaseShader *)iface);
1845 shader->base_shader.parent_ops->wined3d_object_destroyed(shader->base_shader.parent);
1846 HeapFree(GetProcessHeap(), 0, shader);
1847 }
1848
1849 return refcount;
1850 }
1851
1852 static HRESULT STDMETHODCALLTYPE geometryshader_GetParent(IWineD3DGeometryShader *iface, IUnknown **parent)
1853 {
1854 TRACE("iface %p, parent %p.\n", iface, parent);
1855
1856 shader_get_parent((IWineD3DBaseShaderImpl *)iface, parent);
1857
1858 return WINED3D_OK;
1859 }
1860
1861 static HRESULT STDMETHODCALLTYPE geometryshader_GetFunction(IWineD3DGeometryShader *iface, void *data, UINT *data_size)
1862 {
1863 TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1864
1865 return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1866 }
1867
1868 static const IWineD3DGeometryShaderVtbl wined3d_geometryshader_vtbl =
1869 {
1870 /* IUnknown methods */
1871 geometryshader_QueryInterface,
1872 geometryshader_AddRef,
1873 geometryshader_Release,
1874 /* IWineD3DBase methods */
1875 geometryshader_GetParent,
1876 /* IWineD3DBaseShader methods */
1877 geometryshader_GetFunction,
1878 };
1879
1880 HRESULT geometryshader_init(struct wined3d_geometryshader *shader, IWineD3DDeviceImpl *device,
1881 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1882 IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
1883 {
1884 HRESULT hr;
1885
1886 shader->vtbl = &wined3d_geometryshader_vtbl;
1887 shader_init(&shader->base_shader, device, parent, parent_ops);
1888
1889 hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code, output_signature, 0);
1890 if (FAILED(hr))
1891 {
1892 WARN("Failed to set function, hr %#x.\n", hr);
1893 shader_cleanup((IWineD3DBaseShader *)shader);
1894 return hr;
1895 }
1896
1897 shader->base_shader.load_local_constsF = FALSE;
1898
1899 return WINED3D_OK;
1900 }
1901
1902 static HRESULT STDMETHODCALLTYPE pixelshader_QueryInterface(IWineD3DPixelShader *iface, REFIID riid, void **object)
1903 {
1904 TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1905
1906 if (IsEqualGUID(riid, &IID_IWineD3DPixelShader)
1907 || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1908 || IsEqualGUID(riid, &IID_IWineD3DBase)
1909 || IsEqualGUID(riid, &IID_IUnknown))
1910 {
1911 IUnknown_AddRef(iface);
1912 *object = iface;
1913 return S_OK;
1914 }
1915
1916 WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1917
1918 *object = NULL;
1919 return E_NOINTERFACE;
1920 }
1921
1922 static ULONG STDMETHODCALLTYPE pixelshader_AddRef(IWineD3DPixelShader *iface)
1923 {
1924 IWineD3DPixelShaderImpl *shader = (IWineD3DPixelShaderImpl *)iface;
1925 ULONG refcount = InterlockedIncrement(&shader->baseShader.ref);
1926
1927 TRACE("%p increasing refcount to %u.\n", shader, refcount);
1928
1929 return refcount;
1930 }
1931
1932 static ULONG STDMETHODCALLTYPE pixelshader_Release(IWineD3DPixelShader *iface)
1933 {
1934 IWineD3DPixelShaderImpl *shader = (IWineD3DPixelShaderImpl *)iface;
1935 ULONG refcount = InterlockedDecrement(&shader->baseShader.ref);
1936
1937 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1938
1939 if (!refcount)
1940 {
1941 shader_cleanup((IWineD3DBaseShader *)iface);
1942 shader->baseShader.parent_ops->wined3d_object_destroyed(shader->baseShader.parent);
1943 HeapFree(GetProcessHeap(), 0, shader);
1944 }
1945
1946 return refcount;
1947 }
1948
1949 static HRESULT STDMETHODCALLTYPE pixelshader_GetParent(IWineD3DPixelShader *iface, IUnknown **parent)
1950 {
1951 TRACE("iface %p, parent %p.\n", iface, parent);
1952
1953 shader_get_parent((IWineD3DBaseShaderImpl *)iface, parent);
1954
1955 return WINED3D_OK;
1956 }
1957
1958 static HRESULT STDMETHODCALLTYPE pixelshader_GetFunction(IWineD3DPixelShader *iface, void *data, UINT *data_size)
1959 {
1960 TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1961
1962 return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1963 }
1964
1965 static const IWineD3DPixelShaderVtbl IWineD3DPixelShader_Vtbl =
1966 {
1967 /* IUnknown methods */
1968 pixelshader_QueryInterface,
1969 pixelshader_AddRef,
1970 pixelshader_Release,
1971 /* IWineD3DBase methods */
1972 pixelshader_GetParent,
1973 /* IWineD3DBaseShader methods */
1974 pixelshader_GetFunction
1975 };
1976
1977 void find_ps_compile_args(IWineD3DPixelShaderImpl *shader,
1978 IWineD3DStateBlockImpl *stateblock, struct ps_compile_args *args)
1979 {
1980 IWineD3DBaseTextureImpl *texture;
1981 UINT i;
1982
1983 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
1984 args->srgb_correction = stateblock->renderState[WINED3DRS_SRGBWRITEENABLE] ? 1 : 0;
1985 args->np2_fixup = 0;
1986
1987 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
1988 {
1989 if (!shader->baseShader.reg_maps.sampler_type[i]) continue;
1990 texture = (IWineD3DBaseTextureImpl *)stateblock->textures[i];
1991 if (!texture)
1992 {
1993 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
1994 continue;
1995 }
1996 args->color_fixup[i] = texture->resource.format_desc->color_fixup;
1997
1998 /* Flag samplers that need NP2 texcoord fixup. */
1999 if (!texture->baseTexture.pow2Matrix_identity)
2000 {
2001 args->np2_fixup |= (1 << i);
2002 }
2003 }
2004 if (shader->baseShader.reg_maps.shader_version.major >= 3)
2005 {
2006 if (((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.position_transformed)
2007 {
2008 args->vp_mode = pretransformed;
2009 }
2010 else if (use_vs(stateblock))
2011 {
2012 args->vp_mode = vertexshader;
2013 }
2014 else
2015 {
2016 args->vp_mode = fixedfunction;
2017 }
2018 args->fog = FOG_OFF;
2019 }
2020 else
2021 {
2022 args->vp_mode = vertexshader;
2023 if (stateblock->renderState[WINED3DRS_FOGENABLE])
2024 {
2025 switch (stateblock->renderState[WINED3DRS_FOGTABLEMODE])
2026 {
2027 case WINED3DFOG_NONE:
2028 if (((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.position_transformed
2029 || use_vs(stateblock))
2030 {
2031 args->fog = FOG_LINEAR;
2032 break;
2033 }
2034
2035 switch (stateblock->renderState[WINED3DRS_FOGVERTEXMODE])
2036 {
2037 case WINED3DFOG_NONE: /* Fall through. */
2038 case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
2039 case WINED3DFOG_EXP: args->fog = FOG_EXP; break;
2040 case WINED3DFOG_EXP2: args->fog = FOG_EXP2; break;
2041 }
2042 break;
2043
2044 case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
2045 case WINED3DFOG_EXP: args->fog = FOG_EXP; break;
2046 case WINED3DFOG_EXP2: args->fog = FOG_EXP2; break;
2047 }
2048 }
2049 else
2050 {
2051 args->fog = FOG_OFF;
2052 }
2053 }
2054 }
2055
2056 static void pixelshader_set_limits(IWineD3DPixelShaderImpl *shader)
2057 {
2058 DWORD shader_version = WINED3D_SHADER_VERSION(shader->baseShader.reg_maps.shader_version.major,
2059 shader->baseShader.reg_maps.shader_version.minor);
2060
2061 shader->baseShader.limits.attributes = 0;
2062 shader->baseShader.limits.address = 0;
2063 shader->baseShader.limits.packed_output = 0;
2064
2065 switch (shader_version)
2066 {
2067 case WINED3D_SHADER_VERSION(1, 0):
2068 case WINED3D_SHADER_VERSION(1, 1):
2069 case WINED3D_SHADER_VERSION(1, 2):
2070 case WINED3D_SHADER_VERSION(1, 3):
2071 shader->baseShader.limits.temporary = 2;
2072 shader->baseShader.limits.constant_float = 8;
2073 shader->baseShader.limits.constant_int = 0;
2074 shader->baseShader.limits.constant_bool = 0;
2075 shader->baseShader.limits.texcoord = 4;
2076 shader->baseShader.limits.sampler = 4;
2077 shader->baseShader.limits.packed_input = 0;
2078 shader->baseShader.limits.label = 0;
2079 break;
2080
2081 case WINED3D_SHADER_VERSION(1, 4):
2082 shader->baseShader.limits.temporary = 6;
2083 shader->baseShader.limits.constant_float = 8;
2084 shader->baseShader.limits.constant_int = 0;
2085 shader->baseShader.limits.constant_bool = 0;
2086 shader->baseShader.limits.texcoord = 6;
2087 shader->baseShader.limits.sampler = 6;
2088 shader->baseShader.limits.packed_input = 0;
2089 shader->baseShader.limits.label = 0;
2090 break;
2091
2092 /* FIXME: Temporaries must match D3DPSHADERCAPS2_0.NumTemps. */
2093 case WINED3D_SHADER_VERSION(2, 0):
2094 shader->baseShader.limits.temporary = 32;
2095 shader->baseShader.limits.constant_float = 32;
2096 shader->baseShader.limits.constant_int = 16;
2097 shader->baseShader.limits.constant_bool = 16;
2098 shader->baseShader.limits.texcoord = 8;
2099 shader->baseShader.limits.sampler = 16;
2100 shader->baseShader.limits.packed_input = 0;
2101 break;
2102
2103 case WINED3D_SHADER_VERSION(2, 1):
2104 shader->baseShader.limits.temporary = 32;
2105 shader->baseShader.limits.constant_float = 32;
2106 shader->baseShader.limits.constant_int = 16;
2107 shader->baseShader.limits.constant_bool = 16;
2108 shader->baseShader.limits.texcoord = 8;
2109 shader->baseShader.limits.sampler = 16;
2110 shader->baseShader.limits.packed_input = 0;
2111 shader->baseShader.limits.label = 16;
2112 break;
2113
2114 case WINED3D_SHADER_VERSION(4, 0):
2115 FIXME("Using 3.0 limits for 4.0 shader.\n");
2116 /* Fall through. */
2117
2118 case WINED3D_SHADER_VERSION(3, 0):
2119 shader->baseShader.limits.temporary = 32;
2120 shader->baseShader.limits.constant_float = 224;
2121 shader->baseShader.limits.constant_int = 16;
2122 shader->baseShader.limits.constant_bool = 16;
2123 shader->baseShader.limits.texcoord = 0;
2124 shader->baseShader.limits.sampler = 16;
2125 shader->baseShader.limits.packed_input = 12;
2126 shader->baseShader.limits.label = 16; /* FIXME: 2048 */
2127 break;
2128
2129 default:
2130 shader->baseShader.limits.temporary = 32;
2131 shader->baseShader.limits.constant_float = 32;
2132 shader->baseShader.limits.constant_int = 16;
2133 shader->baseShader.limits.constant_bool = 16;
2134 shader->baseShader.limits.texcoord = 8;
2135 shader->baseShader.limits.sampler = 16;
2136 shader->baseShader.limits.packed_input = 0;
2137 shader->baseShader.limits.label = 0;
2138 FIXME("Unrecognized pixel shader version %u.%u\n",
2139 shader->baseShader.reg_maps.shader_version.major,
2140 shader->baseShader.reg_maps.shader_version.minor);
2141 }
2142 }
2143
2144 HRESULT pixelshader_init(IWineD3DPixelShaderImpl *shader, IWineD3DDeviceImpl *device,
2145 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2146 IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
2147 {
2148 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2149 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2150 HRESULT hr;
2151
2152 if (!byte_code) return WINED3DERR_INVALIDCALL;
2153
2154 shader->lpVtbl = &IWineD3DPixelShader_Vtbl;
2155 shader_init(&shader->baseShader, device, parent, parent_ops);
2156
2157 hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code,
2158 output_signature, device->d3d_pshader_constantF);
2159 if (FAILED(hr))
2160 {
2161 WARN("Failed to set function, hr %#x.\n", hr);
2162 shader_cleanup((IWineD3DBaseShader *)shader);
2163 return hr;
2164 }
2165
2166 pixelshader_set_limits(shader);
2167
2168 for (i = 0; i < MAX_REG_INPUT; ++i)
2169 {
2170 if (shader->input_reg_used[i])
2171 {
2172 ++num_regs_used;
2173 highest_reg_used = i;
2174 }
2175 }
2176
2177 /* Don't do any register mapping magic if it is not needed, or if we can't
2178 * achieve anything anyway */
2179 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2180 || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2181 {
2182 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2183 {
2184 /* This happens with relative addressing. The input mapper function
2185 * warns about this if the higher registers are declared too, so
2186 * don't write a FIXME here */
2187 WARN("More varying registers used than supported\n");
2188 }
2189
2190 for (i = 0; i < MAX_REG_INPUT; ++i)
2191 {
2192 shader->input_reg_map[i] = i;
2193 }
2194
2195 shader->declared_in_count = highest_reg_used + 1;
2196 }
2197 else
2198 {
2199 shader->declared_in_count = 0;
2200 for (i = 0; i < MAX_REG_INPUT; ++i)
2201 {
2202 if (shader->input_reg_used[i]) shader->input_reg_map[i] = shader->declared_in_count++;
2203 else shader->input_reg_map[i] = ~0U;
2204 }
2205 }
2206
2207 shader->baseShader.load_local_constsF = FALSE;
2208
2209 return WINED3D_OK;
2210 }
2211
2212 void pixelshader_update_samplers(struct shader_reg_maps *reg_maps, IWineD3DBaseTexture * const *textures)
2213 {
2214 WINED3DSAMPLER_TEXTURE_TYPE *sampler_type = reg_maps->sampler_type;
2215 unsigned int i;
2216
2217 if (reg_maps->shader_version.major != 1) return;
2218
2219 for (i = 0; i < max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS); ++i)
2220 {
2221 /* We don't sample from this sampler. */
2222 if (!sampler_type[i]) continue;
2223
2224 if (!textures[i])
2225 {
2226 WARN("No texture bound to sampler %u, using 2D.\n", i);
2227 sampler_type[i] = WINED3DSTT_2D;
2228 continue;
2229 }
2230
2231 switch (IWineD3DBaseTexture_GetTextureDimensions(textures[i]))
2232 {
2233 case GL_TEXTURE_RECTANGLE_ARB:
2234 case GL_TEXTURE_2D:
2235 /* We have to select between texture rectangles and 2D
2236 * textures later because 2.0 and 3.0 shaders only have
2237 * WINED3DSTT_2D as well. */
2238 sampler_type[i] = WINED3DSTT_2D;
2239 break;
2240
2241 case GL_TEXTURE_3D:
2242 sampler_type[i] = WINED3DSTT_VOLUME;
2243 break;
2244
2245 case GL_TEXTURE_CUBE_MAP_ARB:
2246 sampler_type[i] = WINED3DSTT_CUBE;
2247 break;
2248
2249 default:
2250 FIXME("Unrecognized texture type %#x, using 2D.\n",
2251 IWineD3DBaseTexture_GetTextureDimensions(textures[i]));
2252 sampler_type[i] = WINED3DSTT_2D;
2253 }
2254 }
2255 }
Maarten's multimedia branch