?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
regedit: Allow importing strings with escaped NULL.
[wine.git] / dlls / wined3d / shader.c
blob5ae4abc0657bae663bc2bcea39538fa3a66262f4
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, 2013 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
36 /* pow, mul_high, sub_high, mul_low */
37 const float wined3d_srgb_const0[] = {0.41666f, 1.055f, 0.055f, 12.92f};
38 /* cmp */
39 const float wined3d_srgb_const1[] = {0.0031308f, 0.0f, 0.0f, 0.0f};
40
41 static const char * const shader_opcode_names[] =
42 {
43 /* WINED3DSIH_ABS */ "abs",
44 /* WINED3DSIH_ADD */ "add",
45 /* WINED3DSIH_AND */ "and",
46 /* WINED3DSIH_BEM */ "bem",
47 /* WINED3DSIH_BREAK */ "break",
48 /* WINED3DSIH_BREAKC */ "breakc",
49 /* WINED3DSIH_BREAKP */ "breakp",
50 /* WINED3DSIH_CALL */ "call",
51 /* WINED3DSIH_CALLNZ */ "callnz",
52 /* WINED3DSIH_CMP */ "cmp",
53 /* WINED3DSIH_CND */ "cnd",
54 /* WINED3DSIH_CRS */ "crs",
55 /* WINED3DSIH_CUT */ "cut",
56 /* WINED3DSIH_DCL */ "dcl",
57 /* WINED3DSIH_DCL_CONSTANT_BUFFER */ "dcl_constantBuffer",
58 /* WINED3DSIH_DCL_INPUT_PRIMITIVE */ "dcl_inputPrimitive",
59 /* WINED3DSIH_DCL_OUTPUT_TOPOLOGY */ "dcl_outputTopology",
60 /* WINED3DSIH_DCL_VERTICES_OUT */ "dcl_maxOutputVertexCount",
61 /* WINED3DSIH_DEF */ "def",
62 /* WINED3DSIH_DEFB */ "defb",
63 /* WINED3DSIH_DEFI */ "defi",
64 /* WINED3DSIH_DIV */ "div",
65 /* WINED3DSIH_DP2 */ "dp2",
66 /* WINED3DSIH_DP2ADD */ "dp2add",
67 /* WINED3DSIH_DP3 */ "dp3",
68 /* WINED3DSIH_DP4 */ "dp4",
69 /* WINED3DSIH_DST */ "dst",
70 /* WINED3DSIH_DSX */ "dsx",
71 /* WINED3DSIH_DSY */ "dsy",
72 /* WINED3DSIH_ELSE */ "else",
73 /* WINED3DSIH_EMIT */ "emit",
74 /* WINED3DSIH_ENDIF */ "endif",
75 /* WINED3DSIH_ENDLOOP */ "endloop",
76 /* WINED3DSIH_ENDREP */ "endrep",
77 /* WINED3DSIH_EQ */ "eq",
78 /* WINED3DSIH_EXP */ "exp",
79 /* WINED3DSIH_EXPP */ "expp",
80 /* WINED3DSIH_FRC */ "frc",
81 /* WINED3DSIH_FTOI */ "ftoi",
82 /* WINED3DSIH_GE */ "ge",
83 /* WINED3DSIH_IADD */ "iadd",
84 /* WINED3DSIH_IEQ */ "ieq",
85 /* WINED3DSIH_IF */ "if",
86 /* WINED3DSIH_IFC */ "ifc",
87 /* WINED3DSIH_IGE */ "ige",
88 /* WINED3DSIH_IMUL */ "imul",
89 /* WINED3DSIH_ISHL */ "ishl",
90 /* WINED3DSIH_ITOF */ "itof",
91 /* WINED3DSIH_LABEL */ "label",
92 /* WINED3DSIH_LD */ "ld",
93 /* WINED3DSIH_LIT */ "lit",
94 /* WINED3DSIH_LOG */ "log",
95 /* WINED3DSIH_LOGP */ "logp",
96 /* WINED3DSIH_LOOP */ "loop",
97 /* WINED3DSIH_LRP */ "lrp",
98 /* WINED3DSIH_LT */ "lt",
99 /* WINED3DSIH_M3x2 */ "m3x2",
100 /* WINED3DSIH_M3x3 */ "m3x3",
101 /* WINED3DSIH_M3x4 */ "m3x4",
102 /* WINED3DSIH_M4x3 */ "m4x3",
103 /* WINED3DSIH_M4x4 */ "m4x4",
104 /* WINED3DSIH_MAD */ "mad",
105 /* WINED3DSIH_MAX */ "max",
106 /* WINED3DSIH_MIN */ "min",
107 /* WINED3DSIH_MOV */ "mov",
108 /* WINED3DSIH_MOVA */ "mova",
109 /* WINED3DSIH_MOVC */ "movc",
110 /* WINED3DSIH_MUL */ "mul",
111 /* WINED3DSIH_NOP */ "nop",
112 /* WINED3DSIH_NRM */ "nrm",
113 /* WINED3DSIH_PHASE */ "phase",
114 /* WINED3DSIH_POW */ "pow",
115 /* WINED3DSIH_RCP */ "rcp",
116 /* WINED3DSIH_REP */ "rep",
117 /* WINED3DSIH_RET */ "ret",
118 /* WINED3DSIH_ROUND_NI */ "round_ni",
119 /* WINED3DSIH_RSQ */ "rsq",
120 /* WINED3DSIH_SAMPLE */ "sample",
121 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
122 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
123 /* WINED3DSIH_SETP */ "setp",
124 /* WINED3DSIH_SGE */ "sge",
125 /* WINED3DSIH_SGN */ "sgn",
126 /* WINED3DSIH_SINCOS */ "sincos",
127 /* WINED3DSIH_SLT */ "slt",
128 /* WINED3DSIH_SQRT */ "sqrt",
129 /* WINED3DSIH_SUB */ "sub",
130 /* WINED3DSIH_TEX */ "texld",
131 /* WINED3DSIH_TEXBEM */ "texbem",
132 /* WINED3DSIH_TEXBEML */ "texbeml",
133 /* WINED3DSIH_TEXCOORD */ "texcrd",
134 /* WINED3DSIH_TEXDEPTH */ "texdepth",
135 /* WINED3DSIH_TEXDP3 */ "texdp3",
136 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
137 /* WINED3DSIH_TEXKILL */ "texkill",
138 /* WINED3DSIH_TEXLDD */ "texldd",
139 /* WINED3DSIH_TEXLDL */ "texldl",
140 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
141 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
142 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
143 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
144 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
145 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
146 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
147 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
148 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
149 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
150 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
151 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
152 /* WINED3DSIH_UDIV */ "udiv",
153 /* WINED3DSIH_USHR */ "ushr",
154 /* WINED3DSIH_UTOF */ "utof",
155 /* WINED3DSIH_XOR */ "xor",
156 };
157
158 static const char * const semantic_names[] =
159 {
160 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
161 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
162 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
163 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
164 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
165 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
166 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
167 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
168 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
169 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
170 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
171 /* WINED3D_DECL_USAGE_FOG */ "FOG",
172 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
173 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
174 };
175
176 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
177 {
178 if (usage >= sizeof(semantic_names) / sizeof(*semantic_names))
179 {
180 FIXME("Unrecognized usage %#x.\n", usage);
181 return "UNRECOGNIZED";
182 }
183
184 return semantic_names[usage];
185 }
186
187 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
188 {
189 unsigned int i;
190
191 for (i = 0; i < sizeof(semantic_names) / sizeof(*semantic_names); ++i)
192 {
193 if (!strcmp(name, semantic_names[i])) return i;
194 }
195
196 return ~0U;
197 }
198
199 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
200 {
201 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
202 }
203
204 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
205 const struct wined3d_shader_semantic *s)
206 {
207 e->semantic_name = shader_semantic_name_from_usage(s->usage);
208 e->semantic_idx = s->usage_idx;
209 e->sysval_semantic = 0;
210 e->component_type = 0;
211 e->register_idx = s->reg.reg.idx[0].offset;
212 e->mask = s->reg.write_mask;
213 }
214
215 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
216 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
217 {
218 e->semantic_name = shader_semantic_name_from_usage(usage);
219 e->semantic_idx = usage_idx;
220 e->sysval_semantic = 0;
221 e->component_type = 0;
222 e->register_idx = reg_idx;
223 e->mask = write_mask;
224 }
225
226 static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
227 {
228 switch (version_token >> 16)
229 {
230 case WINED3D_SM1_VS:
231 case WINED3D_SM1_PS:
232 return &sm1_shader_frontend;
233
234 case WINED3D_SM4_PS:
235 case WINED3D_SM4_VS:
236 case WINED3D_SM4_GS:
237 return &sm4_shader_frontend;
238
239 default:
240 FIXME("Unrecognised version token %#x\n", version_token);
241 return NULL;
242 }
243 }
244
245 void shader_buffer_clear(struct wined3d_shader_buffer *buffer)
246 {
247 buffer->buffer[0] = '\0';
248 buffer->content_size = 0;
249 buffer->lineNo = 0;
250 buffer->newline = TRUE;
251 }
252
253 BOOL shader_buffer_init(struct wined3d_shader_buffer *buffer)
254 {
255 buffer->buffer_size = 16384;
256 if (!(buffer->buffer = HeapAlloc(GetProcessHeap(), 0, buffer->buffer_size)))
257 {
258 ERR("Failed to allocate shader buffer memory.\n");
259 return FALSE;
260 }
261
262 shader_buffer_clear(buffer);
263 return TRUE;
264 }
265
266 void shader_buffer_free(struct wined3d_shader_buffer *buffer)
267 {
268 HeapFree(GetProcessHeap(), 0, buffer->buffer);
269 }
270
271 int shader_vaddline(struct wined3d_shader_buffer *buffer, const char *format, va_list args)
272 {
273 char *base = buffer->buffer + buffer->content_size;
274 int rc;
275 char *new_buffer;
276
277 while(1)
278 {
279 rc = vsnprintf(base, buffer->buffer_size - buffer->content_size, format, args);
280 if (rc < 0 /* C89 */ || (unsigned int)rc >= buffer->buffer_size - buffer->content_size /* C99 */)
281 {
282 new_buffer = HeapReAlloc(GetProcessHeap(), 0, buffer->buffer, buffer->buffer_size * 2);
283 if (!new_buffer)
284 {
285 ERR("The buffer allocated for the shader program string is too small at %d bytes.\n", buffer->buffer_size);
286 buffer->content_size = buffer->buffer_size - 1;
287 return -1;
288 }
289 buffer->buffer = new_buffer;
290 buffer->buffer_size = buffer->buffer_size * 2;
291 base = buffer->buffer + buffer->content_size;
292 }
293 else
294 {
295 break;
296 }
297 }
298
299 if (buffer->newline)
300 {
301 TRACE("GL HW (%u, %u) : %s", buffer->lineNo + 1, buffer->content_size, base);
302 buffer->newline = FALSE;
303 }
304 else
305 {
306 TRACE("%s", base);
307 }
308
309 buffer->content_size += rc;
310 if (buffer->buffer[buffer->content_size-1] == '\n')
311 {
312 ++buffer->lineNo;
313 buffer->newline = TRUE;
314 }
315
316 return 0;
317 }
318
319 int shader_addline(struct wined3d_shader_buffer *buffer, const char *format, ...)
320 {
321 va_list args;
322 int ret;
323
324 va_start(args, format);
325 ret = shader_vaddline(buffer, format, args);
326 va_end(args);
327
328 return ret;
329 }
330
331 static void shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
332 void *parent, const struct wined3d_parent_ops *parent_ops)
333 {
334 shader->ref = 1;
335 shader->device = device;
336 shader->parent = parent;
337 shader->parent_ops = parent_ops;
338 list_init(&shader->linked_programs);
339 list_add_head(&device->shaders, &shader->shader_list_entry);
340 }
341
342 /* Convert floating point offset relative to a register file to an absolute
343 * offset for float constants. */
344 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
345 {
346 switch (register_type)
347 {
348 case WINED3DSPR_CONST: return register_idx;
349 case WINED3DSPR_CONST2: return 2048 + register_idx;
350 case WINED3DSPR_CONST3: return 4096 + register_idx;
351 case WINED3DSPR_CONST4: return 6144 + register_idx;
352 default:
353 FIXME("Unsupported register type: %u.\n", register_type);
354 return register_idx;
355 }
356 }
357
358 static void shader_delete_constant_list(struct list *clist)
359 {
360 struct wined3d_shader_lconst *constant;
361 struct list *ptr;
362
363 ptr = list_head(clist);
364 while (ptr)
365 {
366 constant = LIST_ENTRY(ptr, struct wined3d_shader_lconst, entry);
367 ptr = list_next(clist, ptr);
368 HeapFree(GetProcessHeap(), 0, constant);
369 }
370 list_init(clist);
371 }
372
373 static void shader_set_limits(struct wined3d_shader *shader)
374 {
375 static const struct limits_entry
376 {
377 unsigned int min_version;
378 unsigned int max_version;
379 struct wined3d_shader_limits limits;
380 }
381 vs_limits[] =
382 {
383 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
384 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 1), { 0, 0, 256, 0, 12, 0}},
385 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 1), { 0, 16, 256, 16, 12, 0}},
386 /* DX10 cards on Windows advertise a D3D9 constant limit of 256
387 * even though they are capable of supporting much more (GL
388 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
389 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
390 * shaders to 256. */
391 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), { 4, 16, 256, 16, 12, 0}},
392 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 16, 0}},
393 {0}
394 },
395 gs_limits[] =
396 {
397 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
398 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 32, 16}},
399 {0}
400 },
401 ps_limits[] =
402 {
403 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
404 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 3), { 4, 0, 8, 0, 0, 0}},
405 {WINED3D_SHADER_VERSION(1, 4), WINED3D_SHADER_VERSION(1, 4), { 6, 0, 8, 0, 0, 0}},
406 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 0), {16, 0, 32, 0, 0, 0}},
407 {WINED3D_SHADER_VERSION(2, 1), WINED3D_SHADER_VERSION(2, 1), {16, 16, 32, 16, 0, 0}},
408 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), {16, 16, 224, 16, 0, 12}},
409 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 0, 32}},
410 {0}
411 };
412 const struct limits_entry *limits_array;
413 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
414 shader->reg_maps.shader_version.minor);
415 int i = 0;
416
417 switch (shader->reg_maps.shader_version.type)
418 {
419 default:
420 FIXME("Unexpected shader type %u found.\n", shader->reg_maps.shader_version.type);
421 /* Fall-through. */
422 case WINED3D_SHADER_TYPE_VERTEX:
423 limits_array = vs_limits;
424 break;
425 case WINED3D_SHADER_TYPE_GEOMETRY:
426 limits_array = gs_limits;
427 break;
428 case WINED3D_SHADER_TYPE_PIXEL:
429 limits_array = ps_limits;
430 break;
431 }
432
433 while (limits_array[i].min_version && limits_array[i].min_version <= shader_version)
434 {
435 if (shader_version <= limits_array[i].max_version)
436 {
437 shader->limits = &limits_array[i].limits;
438 break;
439 }
440 ++i;
441 }
442 if (!shader->limits)
443 {
444 FIXME("Unexpected shader version \"%u.%u\".\n",
445 shader->reg_maps.shader_version.major,
446 shader->reg_maps.shader_version.minor);
447 shader->limits = &limits_array[max(0, i - 1)].limits;
448 }
449 }
450
451 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
452 {
453 DWORD idx, shift;
454 idx = bit >> 5;
455 shift = bit & 0x1f;
456 bitmap[idx] |= (1 << shift);
457 }
458
459 static BOOL shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
460 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type, unsigned int constf_size)
461 {
462 switch (reg->type)
463 {
464 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
465 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
466 reg_maps->texcoord |= 1 << reg->idx[0].offset;
467 else
468 reg_maps->address |= 1 << reg->idx[0].offset;
469 break;
470
471 case WINED3DSPR_TEMP:
472 reg_maps->temporary |= 1 << reg->idx[0].offset;
473 break;
474
475 case WINED3DSPR_INPUT:
476 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
477 {
478 if (reg->idx[0].rel_addr)
479 {
480 /* If relative addressing is used, we must assume that all registers
481 * are used. Even if it is a construct like v3[aL], we can't assume
482 * that v0, v1 and v2 aren't read because aL can be negative */
483 unsigned int i;
484 for (i = 0; i < MAX_REG_INPUT; ++i)
485 {
486 shader->u.ps.input_reg_used[i] = TRUE;
487 }
488 }
489 else
490 {
491 shader->u.ps.input_reg_used[reg->idx[0].offset] = TRUE;
492 }
493 }
494 else
495 reg_maps->input_registers |= 1 << reg->idx[0].offset;
496 break;
497
498 case WINED3DSPR_RASTOUT:
499 if (reg->idx[0].offset == 1)
500 reg_maps->fog = 1;
501 break;
502
503 case WINED3DSPR_MISCTYPE:
504 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
505 {
506 if (!reg->idx[0].offset)
507 reg_maps->vpos = 1;
508 else if (reg->idx[0].offset == 1)
509 reg_maps->usesfacing = 1;
510 }
511 break;
512
513 case WINED3DSPR_CONST:
514 if (reg->idx[0].rel_addr)
515 {
516 if (reg->idx[0].offset < reg_maps->min_rel_offset)
517 reg_maps->min_rel_offset = reg->idx[0].offset;
518 if (reg->idx[0].offset > reg_maps->max_rel_offset)
519 reg_maps->max_rel_offset = reg->idx[0].offset;
520 reg_maps->usesrelconstF = TRUE;
521 }
522 else
523 {
524 if (reg->idx[0].offset >= min(shader->limits->constant_float, constf_size))
525 {
526 WARN("Shader using float constant %u which is not supported.\n", reg->idx[0].offset);
527 return FALSE;
528 }
529 else
530 {
531 set_bitmap_bit(reg_maps->constf, reg->idx[0].offset);
532 }
533 }
534 break;
535
536 case WINED3DSPR_CONSTINT:
537 if (reg->idx[0].offset >= shader->limits->constant_int)
538 {
539 WARN("Shader using integer constant %u which is not supported.\n", reg->idx[0].offset);
540 return FALSE;
541 }
542 else
543 {
544 reg_maps->integer_constants |= (1 << reg->idx[0].offset);
545 }
546 break;
547
548 case WINED3DSPR_CONSTBOOL:
549 if (reg->idx[0].offset >= shader->limits->constant_bool)
550 {
551 WARN("Shader using bool constant %u which is not supported.\n", reg->idx[0].offset);
552 return FALSE;
553 }
554 else
555 {
556 reg_maps->boolean_constants |= (1 << reg->idx[0].offset);
557 }
558 break;
559
560 case WINED3DSPR_COLOROUT:
561 reg_maps->rt_mask |= (1 << reg->idx[0].offset);
562 break;
563
564 default:
565 TRACE("Not recording register of type %#x and [%#x][%#x].\n",
566 reg->type, reg->idx[0].offset, reg->idx[1].offset);
567 break;
568 }
569 return TRUE;
570 }
571
572 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
573 {
574 switch (instr)
575 {
576 case WINED3DSIH_M4x4:
577 case WINED3DSIH_M3x4:
578 return param == 1 ? 3 : 0;
579
580 case WINED3DSIH_M4x3:
581 case WINED3DSIH_M3x3:
582 return param == 1 ? 2 : 0;
583
584 case WINED3DSIH_M3x2:
585 return param == 1 ? 1 : 0;
586
587 default:
588 return 0;
589 }
590 }
591
592 /* Note that this does not count the loop register as an address register. */
593 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
594 struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature_element *input_signature,
595 struct wined3d_shader_signature_element *output_signature, const DWORD *byte_code, DWORD constf_size)
596 {
597 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
598 void *fe_data = shader->frontend_data;
599 struct wined3d_shader_version shader_version;
600 const DWORD *ptr = byte_code;
601
602 memset(reg_maps, 0, sizeof(*reg_maps));
603 reg_maps->min_rel_offset = ~0U;
604
605 fe->shader_read_header(fe_data, &ptr, &shader_version);
606 reg_maps->shader_version = shader_version;
607
608 shader_set_limits(shader);
609
610 reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
611 sizeof(*reg_maps->constf) * ((min(shader->limits->constant_float, constf_size) + 31) / 32));
612 if (!reg_maps->constf)
613 {
614 ERR("Failed to allocate constant map memory.\n");
615 return E_OUTOFMEMORY;
616 }
617
618 while (!fe->shader_is_end(fe_data, &ptr))
619 {
620 struct wined3d_shader_instruction ins;
621
622 /* Fetch opcode. */
623 fe->shader_read_instruction(fe_data, &ptr, &ins);
624
625 /* Unhandled opcode, and its parameters. */
626 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
627 {
628 TRACE("Skipping unrecognized instruction.\n");
629 continue;
630 }
631
632 /* Handle declarations. */
633 if (ins.handler_idx == WINED3DSIH_DCL)
634 {
635 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
636
637 switch (semantic->reg.reg.type)
638 {
639 /* Mark input registers used. */
640 case WINED3DSPR_INPUT:
641 reg_maps->input_registers |= 1 << semantic->reg.reg.idx[0].offset;
642 shader_signature_from_semantic(&input_signature[semantic->reg.reg.idx[0].offset], semantic);
643 break;
644
645 /* Vertex shader: mark 3.0 output registers used, save token. */
646 case WINED3DSPR_OUTPUT:
647 reg_maps->output_registers |= 1 << semantic->reg.reg.idx[0].offset;
648 shader_signature_from_semantic(&output_signature[semantic->reg.reg.idx[0].offset], semantic);
649 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
650 reg_maps->fog = 1;
651 break;
652
653 /* Save sampler usage token. */
654 case WINED3DSPR_SAMPLER:
655 reg_maps->sampler_type[semantic->reg.reg.idx[0].offset] = semantic->sampler_type;
656 break;
657
658 default:
659 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
660 break;
661 }
662 }
663 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
664 {
665 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
666 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
667 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
668 else
669 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
670 }
671 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
672 {
673 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
674 shader->u.gs.input_type = ins.declaration.primitive_type;
675 else
676 FIXME("Invalid instruction %#x for shader type %#x.\n",
677 ins.handler_idx, shader_version.type);
678 }
679 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
680 {
681 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
682 shader->u.gs.output_type = ins.declaration.primitive_type;
683 else
684 FIXME("Invalid instruction %#x for shader type %#x.\n",
685 ins.handler_idx, shader_version.type);
686 }
687 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
688 {
689 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
690 shader->u.gs.vertices_out = ins.declaration.count;
691 else
692 FIXME("Invalid instruction %#x for shader type %#x.\n",
693 ins.handler_idx, shader_version.type);
694 }
695 else if (ins.handler_idx == WINED3DSIH_DEF)
696 {
697 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
698 float *value;
699 if (!lconst) return E_OUTOFMEMORY;
700
701 lconst->idx = ins.dst[0].reg.idx[0].offset;
702 memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
703 value = (float *)lconst->value;
704
705 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
706 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
707 {
708 if (value[0] < -1.0f) value[0] = -1.0f;
709 else if (value[0] > 1.0f) value[0] = 1.0f;
710 if (value[1] < -1.0f) value[1] = -1.0f;
711 else if (value[1] > 1.0f) value[1] = 1.0f;
712 if (value[2] < -1.0f) value[2] = -1.0f;
713 else if (value[2] > 1.0f) value[2] = 1.0f;
714 if (value[3] < -1.0f) value[3] = -1.0f;
715 else if (value[3] > 1.0f) value[3] = 1.0f;
716 }
717
718 list_add_head(&shader->constantsF, &lconst->entry);
719
720 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
721 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
722 {
723 shader->lconst_inf_or_nan = TRUE;
724 }
725 }
726 else if (ins.handler_idx == WINED3DSIH_DEFI)
727 {
728 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
729 if (!lconst) return E_OUTOFMEMORY;
730
731 lconst->idx = ins.dst[0].reg.idx[0].offset;
732 memcpy(lconst->value, ins.src[0].reg.immconst_data, 4 * sizeof(DWORD));
733
734 list_add_head(&shader->constantsI, &lconst->entry);
735 reg_maps->local_int_consts |= (1 << lconst->idx);
736 }
737 else if (ins.handler_idx == WINED3DSIH_DEFB)
738 {
739 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
740 if (!lconst) return E_OUTOFMEMORY;
741
742 lconst->idx = ins.dst[0].reg.idx[0].offset;
743 memcpy(lconst->value, ins.src[0].reg.immconst_data, sizeof(DWORD));
744
745 list_add_head(&shader->constantsB, &lconst->entry);
746 reg_maps->local_bool_consts |= (1 << lconst->idx);
747 }
748 /* For subroutine prototypes. */
749 else if (ins.handler_idx == WINED3DSIH_LABEL)
750 {
751 reg_maps->labels |= 1 << ins.src[0].reg.idx[0].offset;
752 }
753 /* Set texture, address, temporary registers. */
754 else
755 {
756 BOOL color0_mov = FALSE;
757 unsigned int i;
758
759 /* This will loop over all the registers and try to
760 * make a bitmask of the ones we're interested in.
761 *
762 * Relative addressing tokens are ignored, but that's
763 * okay, since we'll catch any address registers when
764 * they are initialized (required by spec). */
765 for (i = 0; i < ins.dst_count; ++i)
766 {
767 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
768 shader_version.type, constf_size))
769 return WINED3DERR_INVALIDCALL;
770
771 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and
772 * is used in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel
773 * shaders because TECRDOUT isn't used in them, but future register types might cause issues */
774 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX && shader_version.major < 3)
775 {
776 UINT idx = ins.dst[i].reg.idx[0].offset;
777
778 switch (ins.dst[i].reg.type)
779 {
780 case WINED3DSPR_RASTOUT:
781 switch (idx)
782 {
783 case 0: /* oPos */
784 reg_maps->output_registers |= 1 << 10;
785 shader_signature_from_usage(&output_signature[10],
786 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
787 break;
788
789 case 1: /* oFog */
790 reg_maps->output_registers |= 1 << 11;
791 shader_signature_from_usage(&output_signature[11],
792 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
793 break;
794
795 case 2: /* oPts */
796 reg_maps->output_registers |= 1 << 11;
797 shader_signature_from_usage(&output_signature[11],
798 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
799 break;
800 }
801 break;
802
803 case WINED3DSPR_ATTROUT:
804 if (idx < 2)
805 {
806 idx += 8;
807 if (reg_maps->output_registers & (1 << idx))
808 {
809 output_signature[idx].mask |= ins.dst[i].write_mask;
810 }
811 else
812 {
813 reg_maps->output_registers |= 1 << idx;
814 shader_signature_from_usage(&output_signature[idx],
815 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
816 }
817 }
818 break;
819
820 case WINED3DSPR_TEXCRDOUT:
821
822 reg_maps->texcoord_mask[idx] |= ins.dst[i].write_mask;
823 if (reg_maps->output_registers & (1 << idx))
824 {
825 output_signature[idx].mask |= ins.dst[i].write_mask;
826 }
827 else
828 {
829 reg_maps->output_registers |= 1 << idx;
830 shader_signature_from_usage(&output_signature[idx],
831 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
832 }
833 break;
834
835 default:
836 break;
837 }
838 }
839
840 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
841 {
842 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
843 {
844 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
845 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
846 * the mov and perform the sRGB write correction from the source register.
847 *
848 * However, if the mov is only partial, we can't do this, and if the write
849 * comes from an instruction other than MOV it is hard to do as well. If
850 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
851 shader->u.ps.color0_mov = FALSE;
852 if (ins.handler_idx == WINED3DSIH_MOV
853 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
854 {
855 /* Used later when the source register is read. */
856 color0_mov = TRUE;
857 }
858 }
859 /* Also drop the MOV marker if the source register is overwritten prior to the shader
860 * end
861 */
862 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
863 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
864 {
865 shader->u.ps.color0_mov = FALSE;
866 }
867 }
868
869 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
870 if (shader_version.major == 1
871 && (ins.handler_idx == WINED3DSIH_TEX
872 || ins.handler_idx == WINED3DSIH_TEXBEM
873 || ins.handler_idx == WINED3DSIH_TEXBEML
874 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
875 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
876 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
877 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
878 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
879 || ins.handler_idx == WINED3DSIH_TEXREG2AR
880 || ins.handler_idx == WINED3DSIH_TEXREG2GB
881 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
882 {
883 /* Fake sampler usage, only set reserved bit and type. */
884 DWORD sampler_code = ins.dst[i].reg.idx[0].offset;
885
886 TRACE("Setting fake 2D sampler for 1.x pixelshader.\n");
887 reg_maps->sampler_type[sampler_code] = WINED3DSTT_2D;
888
889 /* texbem is only valid with < 1.4 pixel shaders */
890 if (ins.handler_idx == WINED3DSIH_TEXBEM
891 || ins.handler_idx == WINED3DSIH_TEXBEML)
892 {
893 reg_maps->bumpmat |= 1 << ins.dst[i].reg.idx[0].offset;
894 if (ins.handler_idx == WINED3DSIH_TEXBEML)
895 {
896 reg_maps->luminanceparams |= 1 << ins.dst[i].reg.idx[0].offset;
897 }
898 }
899 }
900 else if (ins.handler_idx == WINED3DSIH_BEM)
901 {
902 reg_maps->bumpmat |= 1 << ins.dst[i].reg.idx[0].offset;
903 }
904 }
905
906 if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
907 else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
908 else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
909 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
910 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
911 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
912 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
913 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
914 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
915 else if (ins.handler_idx == WINED3DSIH_LOOP
916 || ins.handler_idx == WINED3DSIH_REP)
917 {
918 ++cur_loop_depth;
919 if (cur_loop_depth > max_loop_depth)
920 max_loop_depth = cur_loop_depth;
921 }
922 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
923 || ins.handler_idx == WINED3DSIH_ENDREP)
924 --cur_loop_depth;
925
926 if (ins.predicate)
927 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
928 shader_version.type, constf_size))
929 return WINED3DERR_INVALIDCALL;
930
931 for (i = 0; i < ins.src_count; ++i)
932 {
933 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
934 struct wined3d_shader_register reg = ins.src[i].reg;
935
936 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
937 shader_version.type, constf_size))
938 return WINED3DERR_INVALIDCALL;
939 while (count)
940 {
941 ++reg.idx[0].offset;
942 if (!shader_record_register_usage(shader, reg_maps, &reg,
943 shader_version.type, constf_size))
944 return WINED3DERR_INVALIDCALL;
945 --count;
946 }
947
948 if (color0_mov)
949 {
950 if (ins.src[i].reg.type == WINED3DSPR_TEMP
951 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
952 {
953 shader->u.ps.color0_mov = TRUE;
954 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
955 }
956 }
957 }
958 }
959 }
960 reg_maps->loop_depth = max_loop_depth;
961
962 /* PS before 2.0 don't have explicit color outputs. Instead the value of
963 * R0 is written to the render target. */
964 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
965 reg_maps->rt_mask |= (1 << 0);
966
967 shader->functionLength = ((const char *)ptr - (const char *)byte_code);
968
969 return WINED3D_OK;
970 }
971
972 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
973 {
974 DWORD map = 1 << max;
975 map |= map - 1;
976 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
977
978 return wined3d_log2i(map);
979 }
980
981 static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
982 const struct wined3d_shader_version *shader_version)
983 {
984 TRACE("dcl");
985
986 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
987 {
988 switch (semantic->sampler_type)
989 {
990 case WINED3DSTT_2D: TRACE("_2d"); break;
991 case WINED3DSTT_CUBE: TRACE("_cube"); break;
992 case WINED3DSTT_VOLUME: TRACE("_volume"); break;
993 default: TRACE("_unknown_ttype(0x%08x)", semantic->sampler_type);
994 }
995 }
996 else
997 {
998 /* Pixel shaders 3.0 don't have usage semantics. */
999 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL) return;
1000 else TRACE("_");
1001
1002 switch (semantic->usage)
1003 {
1004 case WINED3D_DECL_USAGE_POSITION:
1005 TRACE("position%u", semantic->usage_idx);
1006 break;
1007
1008 case WINED3D_DECL_USAGE_BLEND_INDICES:
1009 TRACE("blend");
1010 break;
1011
1012 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
1013 TRACE("weight");
1014 break;
1015
1016 case WINED3D_DECL_USAGE_NORMAL:
1017 TRACE("normal%u", semantic->usage_idx);
1018 break;
1019
1020 case WINED3D_DECL_USAGE_PSIZE:
1021 TRACE("psize");
1022 break;
1023
1024 case WINED3D_DECL_USAGE_COLOR:
1025 if (!semantic->usage_idx) TRACE("color");
1026 else TRACE("specular%u", (semantic->usage_idx - 1));
1027 break;
1028
1029 case WINED3D_DECL_USAGE_TEXCOORD:
1030 TRACE("texture%u", semantic->usage_idx);
1031 break;
1032
1033 case WINED3D_DECL_USAGE_TANGENT:
1034 TRACE("tangent");
1035 break;
1036
1037 case WINED3D_DECL_USAGE_BINORMAL:
1038 TRACE("binormal");
1039 break;
1040
1041 case WINED3D_DECL_USAGE_TESS_FACTOR:
1042 TRACE("tessfactor");
1043 break;
1044
1045 case WINED3D_DECL_USAGE_POSITIONT:
1046 TRACE("positionT%u", semantic->usage_idx);
1047 break;
1048
1049 case WINED3D_DECL_USAGE_FOG:
1050 TRACE("fog");
1051 break;
1052
1053 case WINED3D_DECL_USAGE_DEPTH:
1054 TRACE("depth");
1055 break;
1056
1057 case WINED3D_DECL_USAGE_SAMPLE:
1058 TRACE("sample");
1059 break;
1060
1061 default:
1062 FIXME("unknown_semantics(0x%08x)", semantic->usage);
1063 }
1064 }
1065 }
1066
1067 static void shader_dump_register(const struct wined3d_shader_register *reg,
1068 const struct wined3d_shader_version *shader_version)
1069 {
1070 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
1071 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
1072 UINT offset = reg->idx[0].offset;
1073
1074 switch (reg->type)
1075 {
1076 case WINED3DSPR_TEMP:
1077 TRACE("r");
1078 break;
1079
1080 case WINED3DSPR_INPUT:
1081 TRACE("v");
1082 break;
1083
1084 case WINED3DSPR_CONST:
1085 case WINED3DSPR_CONST2:
1086 case WINED3DSPR_CONST3:
1087 case WINED3DSPR_CONST4:
1088 TRACE("c");
1089 offset = shader_get_float_offset(reg->type, offset);
1090 break;
1091
1092 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
1093 TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
1094 break;
1095
1096 case WINED3DSPR_RASTOUT:
1097 TRACE("%s", rastout_reg_names[offset]);
1098 break;
1099
1100 case WINED3DSPR_COLOROUT:
1101 TRACE("oC");
1102 break;
1103
1104 case WINED3DSPR_DEPTHOUT:
1105 TRACE("oDepth");
1106 break;
1107
1108 case WINED3DSPR_ATTROUT:
1109 TRACE("oD");
1110 break;
1111
1112 case WINED3DSPR_TEXCRDOUT:
1113 /* Vertex shaders >= 3.0 use general purpose output registers
1114 * (WINED3DSPR_OUTPUT), which can include an address token. */
1115 if (shader_version->major >= 3) TRACE("o");
1116 else TRACE("oT");
1117 break;
1118
1119 case WINED3DSPR_CONSTINT:
1120 TRACE("i");
1121 break;
1122
1123 case WINED3DSPR_CONSTBOOL:
1124 TRACE("b");
1125 break;
1126
1127 case WINED3DSPR_LABEL:
1128 TRACE("l");
1129 break;
1130
1131 case WINED3DSPR_LOOP:
1132 TRACE("aL");
1133 break;
1134
1135 case WINED3DSPR_SAMPLER:
1136 TRACE("s");
1137 break;
1138
1139 case WINED3DSPR_MISCTYPE:
1140 if (offset > 1)
1141 FIXME("Unhandled misctype register %u.\n", offset);
1142 else
1143 TRACE("%s", misctype_reg_names[offset]);
1144 break;
1145
1146 case WINED3DSPR_PREDICATE:
1147 TRACE("p");
1148 break;
1149
1150 case WINED3DSPR_IMMCONST:
1151 TRACE("l");
1152 break;
1153
1154 case WINED3DSPR_CONSTBUFFER:
1155 TRACE("cb");
1156 break;
1157
1158 case WINED3DSPR_PRIMID:
1159 TRACE("primID");
1160 break;
1161
1162 case WINED3DSPR_NULL:
1163 TRACE("null");
1164 break;
1165
1166 case WINED3DSPR_RESOURCE:
1167 TRACE("t");
1168 break;
1169
1170 default:
1171 TRACE("unhandled_rtype(%#x)", reg->type);
1172 break;
1173 }
1174
1175 if (reg->type == WINED3DSPR_IMMCONST)
1176 {
1177 TRACE("(");
1178 switch (reg->immconst_type)
1179 {
1180 case WINED3D_IMMCONST_SCALAR:
1181 switch (reg->data_type)
1182 {
1183 case WINED3D_DATA_FLOAT:
1184 TRACE("%.8e", *(const float *)reg->immconst_data);
1185 break;
1186 case WINED3D_DATA_INT:
1187 TRACE("%d", reg->immconst_data[0]);
1188 break;
1189 case WINED3D_DATA_RESOURCE:
1190 case WINED3D_DATA_SAMPLER:
1191 case WINED3D_DATA_UINT:
1192 TRACE("%u", reg->immconst_data[0]);
1193 break;
1194 default:
1195 TRACE("<unhandled data type %#x>", reg->data_type);
1196 break;
1197 }
1198 break;
1199
1200 case WINED3D_IMMCONST_VEC4:
1201 switch (reg->data_type)
1202 {
1203 case WINED3D_DATA_FLOAT:
1204 TRACE("%.8e, %.8e, %.8e, %.8e",
1205 *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
1206 *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
1207 break;
1208 case WINED3D_DATA_INT:
1209 TRACE("%d, %d, %d, %d",
1210 reg->immconst_data[0], reg->immconst_data[1],
1211 reg->immconst_data[2], reg->immconst_data[3]);
1212 break;
1213 case WINED3D_DATA_RESOURCE:
1214 case WINED3D_DATA_SAMPLER:
1215 case WINED3D_DATA_UINT:
1216 TRACE("%u, %u, %u, %u",
1217 reg->immconst_data[0], reg->immconst_data[1],
1218 reg->immconst_data[2], reg->immconst_data[3]);
1219 break;
1220 default:
1221 TRACE("<unhandled data type %#x>", reg->data_type);
1222 break;
1223 }
1224 break;
1225
1226 default:
1227 TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
1228 break;
1229 }
1230 TRACE(")");
1231 }
1232 else if (reg->type != WINED3DSPR_RASTOUT
1233 && reg->type != WINED3DSPR_MISCTYPE
1234 && reg->type != WINED3DSPR_NULL)
1235 {
1236 if (offset != ~0U)
1237 {
1238 TRACE("[");
1239 if (reg->idx[0].rel_addr)
1240 {
1241 shader_dump_src_param(reg->idx[0].rel_addr, shader_version);
1242 TRACE(" + ");
1243 }
1244 TRACE("%u]", offset);
1245
1246 if (reg->idx[1].offset != ~0U)
1247 {
1248 TRACE("[");
1249 if (reg->idx[1].rel_addr)
1250 {
1251 shader_dump_src_param(reg->idx[1].rel_addr, shader_version);
1252 TRACE(" + ");
1253 }
1254 TRACE("%u]", reg->idx[1].offset);
1255 }
1256 }
1257 }
1258 }
1259
1260 void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
1261 const struct wined3d_shader_version *shader_version)
1262 {
1263 DWORD write_mask = param->write_mask;
1264
1265 shader_dump_register(&param->reg, shader_version);
1266
1267 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
1268 {
1269 static const char write_mask_chars[] = "xyzw";
1270
1271 TRACE(".");
1272 if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
1273 if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
1274 if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
1275 if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
1276 }
1277 }
1278
1279 void shader_dump_src_param(const struct wined3d_shader_src_param *param,
1280 const struct wined3d_shader_version *shader_version)
1281 {
1282 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
1283 DWORD swizzle = param->swizzle;
1284
1285 if (src_modifier == WINED3DSPSM_NEG
1286 || src_modifier == WINED3DSPSM_BIASNEG
1287 || src_modifier == WINED3DSPSM_SIGNNEG
1288 || src_modifier == WINED3DSPSM_X2NEG
1289 || src_modifier == WINED3DSPSM_ABSNEG)
1290 TRACE("-");
1291 else if (src_modifier == WINED3DSPSM_COMP)
1292 TRACE("1-");
1293 else if (src_modifier == WINED3DSPSM_NOT)
1294 TRACE("!");
1295
1296 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1297 TRACE("abs(");
1298
1299 shader_dump_register(&param->reg, shader_version);
1300
1301 if (src_modifier)
1302 {
1303 switch (src_modifier)
1304 {
1305 case WINED3DSPSM_NONE: break;
1306 case WINED3DSPSM_NEG: break;
1307 case WINED3DSPSM_NOT: break;
1308 case WINED3DSPSM_BIAS: TRACE("_bias"); break;
1309 case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
1310 case WINED3DSPSM_SIGN: TRACE("_bx2"); break;
1311 case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1312 case WINED3DSPSM_COMP: break;
1313 case WINED3DSPSM_X2: TRACE("_x2"); break;
1314 case WINED3DSPSM_X2NEG: TRACE("_x2"); break;
1315 case WINED3DSPSM_DZ: TRACE("_dz"); break;
1316 case WINED3DSPSM_DW: TRACE("_dw"); break;
1317 case WINED3DSPSM_ABSNEG: TRACE(")"); break;
1318 case WINED3DSPSM_ABS: TRACE(")"); break;
1319 default: TRACE("_unknown_modifier(%#x)", src_modifier);
1320 }
1321 }
1322
1323 if (swizzle != WINED3DSP_NOSWIZZLE)
1324 {
1325 static const char swizzle_chars[] = "xyzw";
1326 DWORD swizzle_x = swizzle & 0x03;
1327 DWORD swizzle_y = (swizzle >> 2) & 0x03;
1328 DWORD swizzle_z = (swizzle >> 4) & 0x03;
1329 DWORD swizzle_w = (swizzle >> 6) & 0x03;
1330
1331 if (swizzle_x == swizzle_y
1332 && swizzle_x == swizzle_z
1333 && swizzle_x == swizzle_w)
1334 {
1335 TRACE(".%c", swizzle_chars[swizzle_x]);
1336 }
1337 else
1338 {
1339 TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1340 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1341 }
1342 }
1343 }
1344
1345 /* Shared code in order to generate the bulk of the shader string.
1346 * NOTE: A description of how to parse tokens can be found on MSDN. */
1347 void shader_generate_main(const struct wined3d_shader *shader, struct wined3d_shader_buffer *buffer,
1348 const struct wined3d_shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1349 {
1350 struct wined3d_device *device = shader->device;
1351 const struct wined3d_shader_frontend *fe = shader->frontend;
1352 void *fe_data = shader->frontend_data;
1353 struct wined3d_shader_version shader_version;
1354 struct wined3d_shader_loop_state loop_state;
1355 struct wined3d_shader_instruction ins;
1356 struct wined3d_shader_tex_mx tex_mx;
1357 struct wined3d_shader_context ctx;
1358 const DWORD *ptr = byte_code;
1359
1360 /* Initialize current parsing state. */
1361 tex_mx.current_row = 0;
1362 loop_state.current_depth = 0;
1363 loop_state.current_reg = 0;
1364
1365 ctx.shader = shader;
1366 ctx.gl_info = &device->adapter->gl_info;
1367 ctx.reg_maps = reg_maps;
1368 ctx.buffer = buffer;
1369 ctx.tex_mx = &tex_mx;
1370 ctx.loop_state = &loop_state;
1371 ctx.backend_data = backend_ctx;
1372 ins.ctx = &ctx;
1373
1374 fe->shader_read_header(fe_data, &ptr, &shader_version);
1375
1376 while (!fe->shader_is_end(fe_data, &ptr))
1377 {
1378 /* Read opcode. */
1379 fe->shader_read_instruction(fe_data, &ptr, &ins);
1380
1381 /* Unknown opcode and its parameters. */
1382 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1383 {
1384 TRACE("Skipping unrecognized instruction.\n");
1385 continue;
1386 }
1387
1388 if (ins.predicate)
1389 FIXME("Predicates not implemented.\n");
1390
1391 /* Call appropriate function for output target */
1392 device->shader_backend->shader_handle_instruction(&ins);
1393 }
1394 }
1395
1396 static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1397 {
1398 DWORD mmask = dst->modifiers;
1399
1400 switch (dst->shift)
1401 {
1402 case 0: break;
1403 case 13: TRACE("_d8"); break;
1404 case 14: TRACE("_d4"); break;
1405 case 15: TRACE("_d2"); break;
1406 case 1: TRACE("_x2"); break;
1407 case 2: TRACE("_x4"); break;
1408 case 3: TRACE("_x8"); break;
1409 default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1410 }
1411
1412 if (mmask & WINED3DSPDM_SATURATE) TRACE("_sat");
1413 if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1414 if (mmask & WINED3DSPDM_MSAMPCENTROID) TRACE("_centroid");
1415
1416 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1417 if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1418 }
1419
1420 static void shader_dump_primitive_type(enum wined3d_primitive_type primitive_type)
1421 {
1422 switch (primitive_type)
1423 {
1424 case WINED3D_PT_UNDEFINED:
1425 TRACE("undefined");
1426 break;
1427 case WINED3D_PT_POINTLIST:
1428 TRACE("pointlist");
1429 break;
1430 case WINED3D_PT_LINELIST:
1431 TRACE("linelist");
1432 break;
1433 case WINED3D_PT_LINESTRIP:
1434 TRACE("linestrip");
1435 break;
1436 case WINED3D_PT_TRIANGLELIST:
1437 TRACE("trianglelist");
1438 break;
1439 case WINED3D_PT_TRIANGLESTRIP:
1440 TRACE("trianglestrip");
1441 break;
1442 case WINED3D_PT_TRIANGLEFAN:
1443 TRACE("trianglefan");
1444 break;
1445 case WINED3D_PT_LINELIST_ADJ:
1446 TRACE("linelist_adj");
1447 break;
1448 case WINED3D_PT_LINESTRIP_ADJ:
1449 TRACE("linestrip_adj");
1450 break;
1451 case WINED3D_PT_TRIANGLELIST_ADJ:
1452 TRACE("trianglelist_adj");
1453 break;
1454 case WINED3D_PT_TRIANGLESTRIP_ADJ:
1455 TRACE("trianglestrip_adj");
1456 break;
1457 default:
1458 TRACE("<unrecognized_primitive_type %#x>", primitive_type);
1459 break;
1460 }
1461 }
1462
1463 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1464 {
1465 struct wined3d_shader_version shader_version;
1466 const DWORD *ptr = byte_code;
1467 const char *type_prefix;
1468 DWORD i;
1469
1470 TRACE("Parsing %p.\n", byte_code);
1471
1472 fe->shader_read_header(fe_data, &ptr, &shader_version);
1473
1474 switch (shader_version.type)
1475 {
1476 case WINED3D_SHADER_TYPE_VERTEX:
1477 type_prefix = "vs";
1478 break;
1479
1480 case WINED3D_SHADER_TYPE_GEOMETRY:
1481 type_prefix = "gs";
1482 break;
1483
1484 case WINED3D_SHADER_TYPE_PIXEL:
1485 type_prefix = "ps";
1486 break;
1487
1488 default:
1489 FIXME("Unhandled shader type %#x.\n", shader_version.type);
1490 type_prefix = "unknown";
1491 break;
1492 }
1493
1494 TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1495
1496 while (!fe->shader_is_end(fe_data, &ptr))
1497 {
1498 struct wined3d_shader_instruction ins;
1499
1500 fe->shader_read_instruction(fe_data, &ptr, &ins);
1501 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1502 {
1503 TRACE("Skipping unrecognized instruction.\n");
1504 continue;
1505 }
1506
1507 if (ins.handler_idx == WINED3DSIH_DCL)
1508 {
1509 shader_dump_decl_usage(&ins.declaration.semantic, &shader_version);
1510 shader_dump_ins_modifiers(&ins.declaration.semantic.reg);
1511 TRACE(" ");
1512 shader_dump_dst_param(&ins.declaration.semantic.reg, &shader_version);
1513 }
1514 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
1515 {
1516 TRACE("%s ", shader_opcode_names[ins.handler_idx]);
1517 shader_dump_src_param(&ins.declaration.src, &shader_version);
1518 TRACE(", %s", ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
1519 }
1520 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
1521 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
1522 {
1523 TRACE("%s ", shader_opcode_names[ins.handler_idx]);
1524 shader_dump_primitive_type(ins.declaration.primitive_type);
1525 }
1526 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1527 {
1528 TRACE("%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
1529 }
1530 else if (ins.handler_idx == WINED3DSIH_DEF)
1531 {
1532 TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(ins.dst[0].reg.type,
1533 ins.dst[0].reg.idx[0].offset),
1534 *(const float *)&ins.src[0].reg.immconst_data[0],
1535 *(const float *)&ins.src[0].reg.immconst_data[1],
1536 *(const float *)&ins.src[0].reg.immconst_data[2],
1537 *(const float *)&ins.src[0].reg.immconst_data[3]);
1538 }
1539 else if (ins.handler_idx == WINED3DSIH_DEFI)
1540 {
1541 TRACE("defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
1542 ins.src[0].reg.immconst_data[0],
1543 ins.src[0].reg.immconst_data[1],
1544 ins.src[0].reg.immconst_data[2],
1545 ins.src[0].reg.immconst_data[3]);
1546 }
1547 else if (ins.handler_idx == WINED3DSIH_DEFB)
1548 {
1549 TRACE("defb b%u = %s", ins.dst[0].reg.idx[0].offset, ins.src[0].reg.immconst_data[0] ? "true" : "false");
1550 }
1551 else
1552 {
1553 if (ins.predicate)
1554 {
1555 TRACE("(");
1556 shader_dump_src_param(ins.predicate, &shader_version);
1557 TRACE(") ");
1558 }
1559
1560 /* PixWin marks instructions with the coissue flag with a '+' */
1561 if (ins.coissue) TRACE("+");
1562
1563 TRACE("%s", shader_opcode_names[ins.handler_idx]);
1564
1565 if (ins.handler_idx == WINED3DSIH_IFC
1566 || ins.handler_idx == WINED3DSIH_BREAKC)
1567 {
1568 switch (ins.flags)
1569 {
1570 case WINED3D_SHADER_REL_OP_GT: TRACE("_gt"); break;
1571 case WINED3D_SHADER_REL_OP_EQ: TRACE("_eq"); break;
1572 case WINED3D_SHADER_REL_OP_GE: TRACE("_ge"); break;
1573 case WINED3D_SHADER_REL_OP_LT: TRACE("_lt"); break;
1574 case WINED3D_SHADER_REL_OP_NE: TRACE("_ne"); break;
1575 case WINED3D_SHADER_REL_OP_LE: TRACE("_le"); break;
1576 default: TRACE("_(%u)", ins.flags);
1577 }
1578 }
1579 else if (ins.handler_idx == WINED3DSIH_TEX
1580 && shader_version.major >= 2
1581 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1582 {
1583 TRACE("p");
1584 }
1585
1586 for (i = 0; i < ins.dst_count; ++i)
1587 {
1588 shader_dump_ins_modifiers(&ins.dst[i]);
1589 TRACE(!i ? " " : ", ");
1590 shader_dump_dst_param(&ins.dst[i], &shader_version);
1591 }
1592
1593 /* Other source tokens */
1594 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1595 {
1596 TRACE(!i ? " " : ", ");
1597 shader_dump_src_param(&ins.src[i - ins.dst_count], &shader_version);
1598 }
1599 }
1600 TRACE("\n");
1601 }
1602 }
1603
1604 static void shader_cleanup(struct wined3d_shader *shader)
1605 {
1606 shader->device->shader_backend->shader_destroy(shader);
1607 HeapFree(GetProcessHeap(), 0, shader->reg_maps.constf);
1608 HeapFree(GetProcessHeap(), 0, shader->function);
1609 shader_delete_constant_list(&shader->constantsF);
1610 shader_delete_constant_list(&shader->constantsB);
1611 shader_delete_constant_list(&shader->constantsI);
1612 list_remove(&shader->shader_list_entry);
1613
1614 if (shader->frontend && shader->frontend_data)
1615 shader->frontend->shader_free(shader->frontend_data);
1616 }
1617
1618 struct shader_none_priv
1619 {
1620 const struct wined3d_vertex_pipe_ops *vertex_pipe;
1621 const struct fragment_pipeline *fragment_pipe;
1622 BOOL ffp_proj_control;
1623 };
1624
1625 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1626 static void shader_none_select_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info,
1627 enum tex_types tex_type, const SIZE *ds_mask_size) {}
1628 static void shader_none_deselect_depth_blt(void *shader_priv, const struct wined3d_gl_info *gl_info) {}
1629 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
1630 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
1631 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
1632 const struct wined3d_state *state) {}
1633 static void shader_none_destroy(struct wined3d_shader *shader) {}
1634 static void shader_none_free_context_data(struct wined3d_context *context) {}
1635
1636 /* Context activation is done by the caller. */
1637 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
1638 const struct wined3d_state *state)
1639 {
1640 const struct wined3d_gl_info *gl_info = context->gl_info;
1641 struct shader_none_priv *priv = shader_priv;
1642
1643 priv->vertex_pipe->vp_enable(gl_info, !use_vs(state));
1644 priv->fragment_pipe->enable_extension(gl_info, !use_ps(state));
1645 }
1646
1647 /* Context activation is done by the caller. */
1648 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
1649 {
1650 struct shader_none_priv *priv = shader_priv;
1651 const struct wined3d_gl_info *gl_info = context->gl_info;
1652
1653 priv->vertex_pipe->vp_enable(gl_info, FALSE);
1654 priv->fragment_pipe->enable_extension(gl_info, FALSE);
1655
1656 context->shader_update_mask = (1 << WINED3D_SHADER_TYPE_PIXEL)
1657 | (1 << WINED3D_SHADER_TYPE_VERTEX)
1658 | (1 << WINED3D_SHADER_TYPE_GEOMETRY);
1659 }
1660
1661 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
1662 const struct fragment_pipeline *fragment_pipe)
1663 {
1664 struct fragment_caps fragment_caps;
1665 void *vertex_priv, *fragment_priv;
1666 struct shader_none_priv *priv;
1667
1668 if (!(priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv))))
1669 return E_OUTOFMEMORY;
1670
1671 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
1672 {
1673 ERR("Failed to initialize vertex pipe.\n");
1674 HeapFree(GetProcessHeap(), 0, priv);
1675 return E_FAIL;
1676 }
1677
1678 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
1679 {
1680 ERR("Failed to initialize fragment pipe.\n");
1681 vertex_pipe->vp_free(device);
1682 HeapFree(GetProcessHeap(), 0, priv);
1683 return E_FAIL;
1684 }
1685
1686 priv->vertex_pipe = vertex_pipe;
1687 priv->fragment_pipe = fragment_pipe;
1688 fragment_pipe->get_caps(&device->adapter->gl_info, &fragment_caps);
1689 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
1690
1691 device->vertex_priv = vertex_priv;
1692 device->fragment_priv = fragment_priv;
1693 device->shader_priv = priv;
1694
1695 return WINED3D_OK;
1696 }
1697
1698 static void shader_none_free(struct wined3d_device *device)
1699 {
1700 struct shader_none_priv *priv = device->shader_priv;
1701
1702 priv->fragment_pipe->free_private(device);
1703 priv->vertex_pipe->vp_free(device);
1704 HeapFree(GetProcessHeap(), 0, priv);
1705 }
1706
1707 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
1708 {
1709 return TRUE;
1710 }
1711
1712 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
1713 {
1714 /* Set the shader caps to 0 for the none shader backend */
1715 caps->vs_version = 0;
1716 caps->gs_version = 0;
1717 caps->ps_version = 0;
1718 caps->vs_uniform_count = 0;
1719 caps->ps_uniform_count = 0;
1720 caps->ps_1x_max_value = 0.0f;
1721 caps->wined3d_caps = 0;
1722 }
1723
1724 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
1725 {
1726 /* We "support" every possible fixup, since we don't support any shader
1727 * model, and will never have to actually sample a texture. */
1728 return TRUE;
1729 }
1730
1731 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
1732 {
1733 struct shader_none_priv *priv = shader_priv;
1734
1735 return priv->ffp_proj_control;
1736 }
1737
1738 const struct wined3d_shader_backend_ops none_shader_backend =
1739 {
1740 shader_none_handle_instruction,
1741 shader_none_select,
1742 shader_none_disable,
1743 shader_none_select_depth_blt,
1744 shader_none_deselect_depth_blt,
1745 shader_none_update_float_vertex_constants,
1746 shader_none_update_float_pixel_constants,
1747 shader_none_load_constants,
1748 shader_none_destroy,
1749 shader_none_alloc,
1750 shader_none_free,
1751 shader_none_allocate_context_data,
1752 shader_none_free_context_data,
1753 shader_none_get_caps,
1754 shader_none_color_fixup_supported,
1755 shader_none_has_ffp_proj_control,
1756 };
1757
1758 static HRESULT shader_set_function(struct wined3d_shader *shader, const DWORD *byte_code,
1759 const struct wined3d_shader_signature *output_signature, DWORD float_const_count,
1760 enum wined3d_shader_type type, unsigned int max_version)
1761 {
1762 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1763 const struct wined3d_shader_frontend *fe;
1764 HRESULT hr;
1765 unsigned int backend_version;
1766 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
1767
1768 TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u.\n",
1769 shader, byte_code, output_signature, float_const_count);
1770
1771 fe = shader_select_frontend(*byte_code);
1772 if (!fe)
1773 {
1774 FIXME("Unable to find frontend for shader.\n");
1775 return WINED3DERR_INVALIDCALL;
1776 }
1777 shader->frontend = fe;
1778 shader->frontend_data = fe->shader_init(byte_code, output_signature);
1779 if (!shader->frontend_data)
1780 {
1781 FIXME("Failed to initialize frontend.\n");
1782 return WINED3DERR_INVALIDCALL;
1783 }
1784
1785 /* First pass: trace shader. */
1786 if (TRACE_ON(d3d_shader))
1787 shader_trace_init(fe, shader->frontend_data, byte_code);
1788
1789 /* Initialize immediate constant lists. */
1790 list_init(&shader->constantsF);
1791 list_init(&shader->constantsB);
1792 list_init(&shader->constantsI);
1793 shader->lconst_inf_or_nan = FALSE;
1794
1795 /* Second pass: figure out which registers are used, what the semantics are, etc. */
1796 hr = shader_get_registers_used(shader, fe,
1797 reg_maps, shader->input_signature, shader->output_signature,
1798 byte_code, float_const_count);
1799 if (FAILED(hr)) return hr;
1800
1801 if (reg_maps->shader_version.type != type)
1802 {
1803 WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
1804 return WINED3DERR_INVALIDCALL;
1805 }
1806 if (reg_maps->shader_version.major > max_version)
1807 {
1808 WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
1809 return WINED3DERR_INVALIDCALL;
1810 }
1811 switch (type)
1812 {
1813 case WINED3D_SHADER_TYPE_VERTEX:
1814 backend_version = d3d_info->limits.vs_version;
1815 break;
1816 case WINED3D_SHADER_TYPE_GEOMETRY:
1817 backend_version = d3d_info->limits.gs_version;
1818 break;
1819 case WINED3D_SHADER_TYPE_PIXEL:
1820 backend_version = d3d_info->limits.ps_version;
1821 break;
1822 default:
1823 FIXME("No backend version-checking for this shader type\n");
1824 backend_version = 0;
1825 }
1826 if (reg_maps->shader_version.major > backend_version)
1827 {
1828 WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
1829 reg_maps->shader_version.major, reg_maps->shader_version.minor);
1830 return WINED3DERR_INVALIDCALL;
1831 }
1832
1833 shader->function = HeapAlloc(GetProcessHeap(), 0, shader->functionLength);
1834 if (!shader->function)
1835 return E_OUTOFMEMORY;
1836 memcpy(shader->function, byte_code, shader->functionLength);
1837
1838 return WINED3D_OK;
1839 }
1840
1841 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
1842 {
1843 ULONG refcount = InterlockedIncrement(&shader->ref);
1844
1845 TRACE("%p increasing refcount to %u.\n", shader, refcount);
1846
1847 return refcount;
1848 }
1849
1850 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
1851 {
1852 ULONG refcount = InterlockedDecrement(&shader->ref);
1853
1854 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1855
1856 if (!refcount)
1857 {
1858 shader_cleanup(shader);
1859 shader->parent_ops->wined3d_object_destroyed(shader->parent);
1860 HeapFree(GetProcessHeap(), 0, shader);
1861 }
1862
1863 return refcount;
1864 }
1865
1866 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
1867 {
1868 TRACE("shader %p.\n", shader);
1869
1870 return shader->parent;
1871 }
1872
1873 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
1874 void *byte_code, UINT *byte_code_size)
1875 {
1876 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
1877
1878 if (!byte_code)
1879 {
1880 *byte_code_size = shader->functionLength;
1881 return WINED3D_OK;
1882 }
1883
1884 if (*byte_code_size < shader->functionLength)
1885 {
1886 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
1887 * than the required size we should write the required size and
1888 * return D3DERR_MOREDATA. That's not actually true. */
1889 return WINED3DERR_INVALIDCALL;
1890 }
1891
1892 memcpy(byte_code, shader->function, shader->functionLength);
1893
1894 return WINED3D_OK;
1895 }
1896
1897 /* Set local constants for d3d8 shaders. */
1898 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
1899 UINT start_idx, const float *src_data, UINT count)
1900 {
1901 UINT end_idx = start_idx + count;
1902 UINT i;
1903
1904 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
1905
1906 if (end_idx > shader->limits->constant_float)
1907 {
1908 WARN("end_idx %u > float constants limit %u.\n",
1909 end_idx, shader->limits->constant_float);
1910 end_idx = shader->limits->constant_float;
1911 }
1912
1913 for (i = start_idx; i < end_idx; ++i)
1914 {
1915 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1916 float *value;
1917 if (!lconst)
1918 return E_OUTOFMEMORY;
1919
1920 lconst->idx = i;
1921 value = (float *)lconst->value;
1922 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
1923 list_add_head(&shader->constantsF, &lconst->entry);
1924
1925 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
1926 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
1927 {
1928 shader->lconst_inf_or_nan = TRUE;
1929 }
1930 }
1931
1932 return WINED3D_OK;
1933 }
1934
1935 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
1936 WORD swizzle_map, struct vs_compile_args *args)
1937 {
1938 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
1939 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
1940 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
1941 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
1942 args->swizzle_map = swizzle_map;
1943 }
1944
1945 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
1946 {
1947 if (usage_idx1 != usage_idx2)
1948 return FALSE;
1949 if (usage1 == usage2)
1950 return TRUE;
1951 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
1952 return TRUE;
1953 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
1954 return TRUE;
1955
1956 return FALSE;
1957 }
1958
1959 BOOL vshader_get_input(const struct wined3d_shader *shader,
1960 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
1961 {
1962 WORD map = shader->reg_maps.input_registers;
1963 unsigned int i;
1964
1965 for (i = 0; map; map >>= 1, ++i)
1966 {
1967 if (!(map & 1)) continue;
1968
1969 if (match_usage(shader->u.vs.attributes[i].usage,
1970 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
1971 {
1972 *regnum = i;
1973 return TRUE;
1974 }
1975 }
1976 return FALSE;
1977 }
1978
1979 static HRESULT vertexshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
1980 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1981 void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
1982 {
1983 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1984 unsigned int i;
1985 HRESULT hr;
1986 WORD map;
1987 const DWORD vs_uniform_count = device->adapter->d3d_info.limits.vs_uniform_count;
1988
1989 if (!byte_code) return WINED3DERR_INVALIDCALL;
1990
1991 shader_init(shader, device, parent, parent_ops);
1992 hr = shader_set_function(shader, byte_code, output_signature, vs_uniform_count,
1993 WINED3D_SHADER_TYPE_VERTEX, max_version);
1994 if (FAILED(hr))
1995 {
1996 WARN("Failed to set function, hr %#x.\n", hr);
1997 shader_cleanup(shader);
1998 return hr;
1999 }
2000
2001 map = reg_maps->input_registers;
2002 for (i = 0; map; map >>= 1, ++i)
2003 {
2004 if (!(map & 1) || !shader->input_signature[i].semantic_name)
2005 continue;
2006
2007 shader->u.vs.attributes[i].usage =
2008 shader_usage_from_semantic_name(shader->input_signature[i].semantic_name);
2009 shader->u.vs.attributes[i].usage_idx = shader->input_signature[i].semantic_idx;
2010 }
2011
2012 if (output_signature)
2013 {
2014 for (i = 0; i < output_signature->element_count; ++i)
2015 {
2016 struct wined3d_shader_signature_element *e = &output_signature->elements[i];
2017 reg_maps->output_registers |= 1 << e->register_idx;
2018 shader->output_signature[e->register_idx] = *e;
2019 }
2020 }
2021
2022 shader->load_local_constsF = (reg_maps->usesrelconstF && !list_empty(&shader->constantsF)) ||
2023 shader->lconst_inf_or_nan;
2024
2025 return WINED3D_OK;
2026 }
2027
2028 static HRESULT geometryshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2029 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2030 void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
2031 {
2032 HRESULT hr;
2033
2034 shader_init(shader, device, parent, parent_ops);
2035 hr = shader_set_function(shader, byte_code, output_signature, 0,
2036 WINED3D_SHADER_TYPE_GEOMETRY, max_version);
2037 if (FAILED(hr))
2038 {
2039 WARN("Failed to set function, hr %#x.\n", hr);
2040 shader_cleanup(shader);
2041 return hr;
2042 }
2043
2044 shader->load_local_constsF = shader->lconst_inf_or_nan;
2045
2046 return WINED3D_OK;
2047 }
2048
2049 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2050 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_gl_info *gl_info)
2051 {
2052 const struct wined3d_texture *texture;
2053 UINT i;
2054
2055 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
2056 if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && state->render_states[WINED3D_RS_SRGBWRITEENABLE])
2057 {
2058 const struct wined3d_format *rt_format = state->fb->render_targets[0]->format;
2059 if (rt_format->flags & WINED3DFMT_FLAG_SRGB_WRITE)
2060 {
2061 static unsigned int warned = 0;
2062
2063 args->srgb_correction = 1;
2064 if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
2065 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
2066 "support, expect rendering artifacts.\n");
2067 }
2068 }
2069
2070 if (shader->reg_maps.shader_version.major == 1
2071 && shader->reg_maps.shader_version.minor <= 3)
2072 {
2073 for (i = 0; i < shader->limits->sampler; ++i)
2074 {
2075 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
2076
2077 if (flags & WINED3D_TTFF_PROJECTED)
2078 {
2079 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
2080
2081 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
2082 {
2083 unsigned int j;
2084 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
2085 DWORD max_valid = WINED3D_TTFF_COUNT4;
2086 enum wined3d_sampler_texture_type sampler_type = shader->reg_maps.sampler_type[i];
2087
2088 for (j = 0; j < state->vertex_declaration->element_count; ++j)
2089 {
2090 struct wined3d_vertex_declaration_element *element =
2091 &state->vertex_declaration->elements[j];
2092
2093 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
2094 && element->usage_idx == index)
2095 {
2096 max_valid = element->format->component_count;
2097 break;
2098 }
2099 }
2100 if (!tex_transform || tex_transform > max_valid)
2101 {
2102 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
2103 tex_transform, max_valid);
2104 tex_transform = max_valid;
2105 }
2106 if ((sampler_type == WINED3DSTT_1D && tex_transform > WINED3D_TTFF_COUNT1)
2107 || (sampler_type == WINED3DSTT_2D && tex_transform > WINED3D_TTFF_COUNT2)
2108 || (sampler_type == WINED3DSTT_VOLUME && tex_transform > WINED3D_TTFF_COUNT3))
2109 tex_transform |= WINED3D_PSARGS_PROJECTED;
2110 else
2111 {
2112 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
2113 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
2114 i, tex_transform, sampler_type);
2115 }
2116 }
2117 else
2118 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
2119
2120 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
2121 }
2122 }
2123 }
2124 if (shader->reg_maps.shader_version.major == 1
2125 && shader->reg_maps.shader_version.minor <= 4)
2126 {
2127 for (i = 0; i < shader->limits->sampler; ++i)
2128 {
2129 const struct wined3d_texture *texture = state->textures[i];
2130
2131 if (!shader->reg_maps.sampler_type[i])
2132 continue;
2133
2134 /* Treat unbound textures as 2D. The dummy texture will provide
2135 * the proper sample value. The tex_types bitmap defaults to
2136 * 2D because of the memset. */
2137 if (!texture)
2138 continue;
2139
2140 switch (texture->target)
2141 {
2142 /* RECT textures are distinguished from 2D textures via np2_fixup */
2143 case GL_TEXTURE_RECTANGLE_ARB:
2144 case GL_TEXTURE_2D:
2145 break;
2146
2147 case GL_TEXTURE_3D:
2148 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
2149 break;
2150
2151 case GL_TEXTURE_CUBE_MAP_ARB:
2152 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
2153 break;
2154 }
2155 }
2156 }
2157
2158 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
2159 {
2160 if (!shader->reg_maps.sampler_type[i])
2161 continue;
2162
2163 texture = state->textures[i];
2164 if (!texture)
2165 {
2166 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2167 continue;
2168 }
2169 args->color_fixup[i] = texture->resource.format->color_fixup;
2170
2171 if (texture->resource.format->flags & WINED3DFMT_FLAG_SHADOW)
2172 args->shadow |= 1 << i;
2173
2174 /* Flag samplers that need NP2 texcoord fixup. */
2175 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
2176 args->np2_fixup |= (1 << i);
2177 }
2178 if (shader->reg_maps.shader_version.major >= 3)
2179 {
2180 if (position_transformed)
2181 args->vp_mode = pretransformed;
2182 else if (use_vs(state))
2183 args->vp_mode = vertexshader;
2184 else
2185 args->vp_mode = fixedfunction;
2186 args->fog = WINED3D_FFP_PS_FOG_OFF;
2187 }
2188 else
2189 {
2190 args->vp_mode = vertexshader;
2191 if (state->render_states[WINED3D_RS_FOGENABLE])
2192 {
2193 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
2194 {
2195 case WINED3D_FOG_NONE:
2196 if (position_transformed || use_vs(state))
2197 {
2198 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
2199 break;
2200 }
2201
2202 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
2203 {
2204 case WINED3D_FOG_NONE: /* Fall through. */
2205 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
2206 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
2207 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
2208 }
2209 break;
2210
2211 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
2212 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
2213 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
2214 }
2215 }
2216 else
2217 {
2218 args->fog = WINED3D_FFP_PS_FOG_OFF;
2219 }
2220 }
2221 }
2222
2223 static HRESULT pixelshader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2224 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2225 void *parent, const struct wined3d_parent_ops *parent_ops, unsigned int max_version)
2226 {
2227 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2228 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2229 HRESULT hr;
2230 const DWORD ps_uniform_count = device->adapter->d3d_info.limits.ps_uniform_count;
2231
2232 if (!byte_code) return WINED3DERR_INVALIDCALL;
2233
2234 shader_init(shader, device, parent, parent_ops);
2235 hr = shader_set_function(shader, byte_code, output_signature, ps_uniform_count,
2236 WINED3D_SHADER_TYPE_PIXEL, max_version);
2237 if (FAILED(hr))
2238 {
2239 WARN("Failed to set function, hr %#x.\n", hr);
2240 shader_cleanup(shader);
2241 return hr;
2242 }
2243
2244 for (i = 0; i < MAX_REG_INPUT; ++i)
2245 {
2246 if (shader->u.ps.input_reg_used[i])
2247 {
2248 ++num_regs_used;
2249 highest_reg_used = i;
2250 }
2251 }
2252
2253 /* Don't do any register mapping magic if it is not needed, or if we can't
2254 * achieve anything anyway */
2255 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2256 || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2257 {
2258 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2259 {
2260 /* This happens with relative addressing. The input mapper function
2261 * warns about this if the higher registers are declared too, so
2262 * don't write a FIXME here */
2263 WARN("More varying registers used than supported\n");
2264 }
2265
2266 for (i = 0; i < MAX_REG_INPUT; ++i)
2267 {
2268 shader->u.ps.input_reg_map[i] = i;
2269 }
2270
2271 shader->u.ps.declared_in_count = highest_reg_used + 1;
2272 }
2273 else
2274 {
2275 shader->u.ps.declared_in_count = 0;
2276 for (i = 0; i < MAX_REG_INPUT; ++i)
2277 {
2278 if (shader->u.ps.input_reg_used[i])
2279 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
2280 else shader->u.ps.input_reg_map[i] = ~0U;
2281 }
2282 }
2283
2284 shader->load_local_constsF = shader->lconst_inf_or_nan;
2285
2286 return WINED3D_OK;
2287 }
2288
2289 void pixelshader_update_samplers(struct wined3d_shader *shader, WORD tex_types)
2290 {
2291 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2292 enum wined3d_sampler_texture_type *sampler_type = reg_maps->sampler_type;
2293 unsigned int i;
2294
2295 if (reg_maps->shader_version.major != 1) return;
2296
2297 for (i = 0; i < shader->limits->sampler; ++i)
2298 {
2299 /* We don't sample from this sampler. */
2300 if (!sampler_type[i]) continue;
2301
2302 switch ((tex_types >> i * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK)
2303 {
2304 case WINED3D_SHADER_TEX_2D:
2305 sampler_type[i] = WINED3DSTT_2D;
2306 break;
2307
2308 case WINED3D_SHADER_TEX_3D:
2309 sampler_type[i] = WINED3DSTT_VOLUME;
2310 break;
2311
2312 case WINED3D_SHADER_TEX_CUBE:
2313 sampler_type[i] = WINED3DSTT_CUBE;
2314 break;
2315 }
2316 }
2317 }
2318
2319 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const DWORD *byte_code,
2320 const struct wined3d_shader_signature *output_signature, void *parent,
2321 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2322 {
2323 struct wined3d_shader *object;
2324 HRESULT hr;
2325
2326 TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2327 device, byte_code, output_signature, parent, parent_ops, shader);
2328
2329 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2330 if (!object)
2331 return E_OUTOFMEMORY;
2332
2333 hr = geometryshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2334 if (FAILED(hr))
2335 {
2336 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
2337 HeapFree(GetProcessHeap(), 0, object);
2338 return hr;
2339 }
2340
2341 TRACE("Created geometry shader %p.\n", object);
2342 *shader = object;
2343
2344 return WINED3D_OK;
2345 }
2346
2347 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const DWORD *byte_code,
2348 const struct wined3d_shader_signature *output_signature, void *parent,
2349 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2350 {
2351 struct wined3d_shader *object;
2352 HRESULT hr;
2353
2354 TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2355 device, byte_code, output_signature, parent, parent_ops, shader);
2356
2357 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2358 if (!object)
2359 return E_OUTOFMEMORY;
2360
2361 hr = pixelshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2362 if (FAILED(hr))
2363 {
2364 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
2365 HeapFree(GetProcessHeap(), 0, object);
2366 return hr;
2367 }
2368
2369 TRACE("Created pixel shader %p.\n", object);
2370 *shader = object;
2371
2372 return WINED3D_OK;
2373 }
2374
2375 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const DWORD *byte_code,
2376 const struct wined3d_shader_signature *output_signature, void *parent,
2377 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader, unsigned int max_version)
2378 {
2379 struct wined3d_shader *object;
2380 HRESULT hr;
2381
2382 TRACE("device %p, byte_code %p, output_signature %p, parent %p, parent_ops %p, shader %p.\n",
2383 device, byte_code, output_signature, parent, parent_ops, shader);
2384
2385 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
2386 if (!object)
2387 return E_OUTOFMEMORY;
2388
2389 hr = vertexshader_init(object, device, byte_code, output_signature, parent, parent_ops, max_version);
2390 if (FAILED(hr))
2391 {
2392 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
2393 HeapFree(GetProcessHeap(), 0, object);
2394 return hr;
2395 }
2396
2397 TRACE("Created vertex shader %p.\n", object);
2398 *shader = object;
2399
2400 return WINED3D_OK;
2401 }
Windows API implementation