VirtualBox

source: vbox/trunk/src/VBox/Devices/Graphics/shaderlib/shader.c@ 82704

Last change on this file since 82704 was 82704, checked in by vboxsync, 5 years ago

3D: Check register index in shader opcodes, bugref:9613
  • Property svn:eol-style set to native
  • Property svn:keywords set to Author Date Id Revision
File size: 87.9 KB
Line 
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/*
26 * Oracle LGPL Disclaimer: For the avoidance of doubt, except that if any license choice
27 * other than GPL or LGPL is available it will apply instead, Oracle elects to use only
28 * the Lesser General Public License version 2.1 (LGPLv2) at this time for any software where
29 * a choice of LGPL license versions is made available with the language indicating
30 * that LGPLv2 or any later version may be used, or where a choice of which version
31 * of the LGPL is applied is otherwise unspecified.
32 */
33
34#include "config.h"
35#include "wine/port.h"
36
37#include <math.h>
38#include <stdio.h>
39#include <string.h>
40
41#include "wined3d_private.h"
42
43WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
44WINE_DECLARE_DEBUG_CHANNEL(d3d);
45
46static const char *shader_opcode_names[] =
47{
48 /* WINED3DSIH_ABS */ "abs",
49 /* WINED3DSIH_ADD */ "add",
50 /* WINED3DSIH_BEM */ "bem",
51 /* WINED3DSIH_BREAK */ "break",
52 /* WINED3DSIH_BREAKC */ "breakc",
53 /* WINED3DSIH_BREAKP */ "breakp",
54 /* WINED3DSIH_CALL */ "call",
55 /* WINED3DSIH_CALLNZ */ "callnz",
56 /* WINED3DSIH_CMP */ "cmp",
57 /* WINED3DSIH_CND */ "cnd",
58 /* WINED3DSIH_CRS */ "crs",
59 /* WINED3DSIH_CUT */ "cut",
60 /* WINED3DSIH_DCL */ "dcl",
61 /* WINED3DSIH_DEF */ "def",
62 /* WINED3DSIH_DEFB */ "defb",
63 /* WINED3DSIH_DEFI */ "defi",
64 /* WINED3DSIH_DP2ADD */ "dp2add",
65 /* WINED3DSIH_DP3 */ "dp3",
66 /* WINED3DSIH_DP4 */ "dp4",
67 /* WINED3DSIH_DST */ "dst",
68 /* WINED3DSIH_DSX */ "dsx",
69 /* WINED3DSIH_DSY */ "dsy",
70 /* WINED3DSIH_ELSE */ "else",
71 /* WINED3DSIH_EMIT */ "emit",
72 /* WINED3DSIH_ENDIF */ "endif",
73 /* WINED3DSIH_ENDLOOP */ "endloop",
74 /* WINED3DSIH_ENDREP */ "endrep",
75 /* WINED3DSIH_EXP */ "exp",
76 /* WINED3DSIH_EXPP */ "expp",
77 /* WINED3DSIH_FRC */ "frc",
78 /* WINED3DSIH_IADD */ "iadd",
79 /* WINED3DSIH_IF */ "if",
80 /* WINED3DSIH_IFC */ "ifc",
81 /* WINED3DSIH_IGE */ "ige",
82 /* WINED3DSIH_LABEL */ "label",
83 /* WINED3DSIH_LIT */ "lit",
84 /* WINED3DSIH_LOG */ "log",
85 /* WINED3DSIH_LOGP */ "logp",
86 /* WINED3DSIH_LOOP */ "loop",
87 /* WINED3DSIH_LRP */ "lrp",
88 /* WINED3DSIH_LT */ "lt",
89 /* WINED3DSIH_M3x2 */ "m3x2",
90 /* WINED3DSIH_M3x3 */ "m3x3",
91 /* WINED3DSIH_M3x4 */ "m3x4",
92 /* WINED3DSIH_M4x3 */ "m4x3",
93 /* WINED3DSIH_M4x4 */ "m4x4",
94 /* WINED3DSIH_MAD */ "mad",
95 /* WINED3DSIH_MAX */ "max",
96 /* WINED3DSIH_MIN */ "min",
97 /* WINED3DSIH_MOV */ "mov",
98 /* WINED3DSIH_MOVA */ "mova",
99 /* WINED3DSIH_MUL */ "mul",
100 /* WINED3DSIH_NOP */ "nop",
101 /* WINED3DSIH_NRM */ "nrm",
102 /* WINED3DSIH_PHASE */ "phase",
103 /* WINED3DSIH_POW */ "pow",
104 /* WINED3DSIH_RCP */ "rcp",
105 /* WINED3DSIH_REP */ "rep",
106 /* WINED3DSIH_RET */ "ret",
107 /* WINED3DSIH_RSQ */ "rsq",
108 /* WINED3DSIH_SETP */ "setp",
109 /* WINED3DSIH_SGE */ "sge",
110 /* WINED3DSIH_SGN */ "sgn",
111 /* WINED3DSIH_SINCOS */ "sincos",
112 /* WINED3DSIH_SLT */ "slt",
113 /* WINED3DSIH_SUB */ "sub",
114 /* WINED3DSIH_TEX */ "texld",
115 /* WINED3DSIH_TEXBEM */ "texbem",
116 /* WINED3DSIH_TEXBEML */ "texbeml",
117 /* WINED3DSIH_TEXCOORD */ "texcrd",
118 /* WINED3DSIH_TEXDEPTH */ "texdepth",
119 /* WINED3DSIH_TEXDP3 */ "texdp3",
120 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
121 /* WINED3DSIH_TEXKILL */ "texkill",
122 /* WINED3DSIH_TEXLDD */ "texldd",
123 /* WINED3DSIH_TEXLDL */ "texldl",
124 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
125 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
126 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
127 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
128 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
129 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
130 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
131 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
132 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
133 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
134 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
135 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
136};
137
138static const char *semantic_names[] =
139{
140 /* WINED3DDECLUSAGE_POSITION */ "SV_POSITION",
141 /* WINED3DDECLUSAGE_BLENDWEIGHT */ "BLENDWEIGHT",
142 /* WINED3DDECLUSAGE_BLENDINDICES */ "BLENDINDICES",
143 /* WINED3DDECLUSAGE_NORMAL */ "NORMAL",
144 /* WINED3DDECLUSAGE_PSIZE */ "PSIZE",
145 /* WINED3DDECLUSAGE_TEXCOORD */ "TEXCOORD",
146 /* WINED3DDECLUSAGE_TANGENT */ "TANGENT",
147 /* WINED3DDECLUSAGE_BINORMAL */ "BINORMAL",
148 /* WINED3DDECLUSAGE_TESSFACTOR */ "TESSFACTOR",
149 /* WINED3DDECLUSAGE_POSITIONT */ "POSITIONT",
150 /* WINED3DDECLUSAGE_COLOR */ "COLOR",
151 /* WINED3DDECLUSAGE_FOG */ "FOG",
152 /* WINED3DDECLUSAGE_DEPTH */ "DEPTH",
153 /* WINED3DDECLUSAGE_SAMPLE */ "SAMPLE",
154};
155
156static const char *shader_semantic_name_from_usage(WINED3DDECLUSAGE usage)
157{
158 if (usage >= sizeof(semantic_names) / sizeof(*semantic_names))
159 {
160 FIXME("Unrecognized usage %#x.\n", usage);
161 return "UNRECOGNIZED";
162 }
163
164 return semantic_names[usage];
165}
166
167static WINED3DDECLUSAGE shader_usage_from_semantic_name(const char *name)
168{
169 unsigned int i;
170
171 for (i = 0; i < sizeof(semantic_names) / sizeof(*semantic_names); ++i)
172 {
173 if (!strcmp(name, semantic_names[i])) return i;
174 }
175
176 return ~0U;
177}
178
179BOOL shader_match_semantic(const char *semantic_name, WINED3DDECLUSAGE usage)
180{
181 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
182}
183
184static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
185 const struct wined3d_shader_semantic *s)
186{
187 e->semantic_name = shader_semantic_name_from_usage(s->usage);
188 e->semantic_idx = s->usage_idx;
189 e->sysval_semantic = 0;
190 e->component_type = 0;
191 e->register_idx = s->reg.reg.idx;
192 e->mask = s->reg.write_mask;
193}
194
195static const struct wined3d_shader_frontend *shader_select_frontend(DWORD version_token)
196{
197 switch (version_token >> 16)
198 {
199 case WINED3D_SM1_VS:
200 case WINED3D_SM1_PS:
201 return &sm1_shader_frontend;
202
203 case WINED3D_SM4_PS:
204 case WINED3D_SM4_VS:
205 case WINED3D_SM4_GS:
206 return &sm4_shader_frontend;
207
208 default:
209 FIXME("Unrecognised version token %#x\n", version_token);
210 return NULL;
211 }
212}
213
214void shader_buffer_clear(struct wined3d_shader_buffer *buffer)
215{
216 buffer->buffer[0] = '\0';
217 buffer->bsize = 0;
218 buffer->lineNo = 0;
219 buffer->newline = TRUE;
220}
221
222BOOL shader_buffer_init(struct wined3d_shader_buffer *buffer)
223{
224 buffer->buffer = HeapAlloc(GetProcessHeap(), 0, SHADER_PGMSIZE);
225 if (!buffer->buffer)
226 {
227 ERR("Failed to allocate shader buffer memory.\n");
228 return FALSE;
229 }
230
231 shader_buffer_clear(buffer);
232 return TRUE;
233}
234
235void shader_buffer_free(struct wined3d_shader_buffer *buffer)
236{
237 HeapFree(GetProcessHeap(), 0, buffer->buffer);
238}
239
240int shader_vaddline(struct wined3d_shader_buffer *buffer, const char *format, va_list args)
241{
242 char *base = buffer->buffer + buffer->bsize;
243 int rc;
244
245 rc = vsnprintf(base, SHADER_PGMSIZE - 1 - buffer->bsize, format, args);
246
247 if (rc < 0 /* C89 */ || (unsigned int)rc > SHADER_PGMSIZE - 1 - buffer->bsize /* C99 */)
248 {
249 ERR("The buffer allocated for the shader program string "
250 "is too small at %d bytes.\n", SHADER_PGMSIZE);
251 buffer->bsize = SHADER_PGMSIZE - 1;
252 return -1;
253 }
254
255 if (buffer->newline)
256 {
257 TRACE("GL HW (%u, %u) : %s", buffer->lineNo + 1, buffer->bsize, base);
258 buffer->newline = FALSE;
259 }
260 else
261 {
262 TRACE("%s", base);
263 }
264
265 buffer->bsize += rc;
266 if (buffer->buffer[buffer->bsize-1] == '\n')
267 {
268 ++buffer->lineNo;
269 buffer->newline = TRUE;
270 }
271
272 return 0;
273}
274
275int shader_addline(struct wined3d_shader_buffer *buffer, const char *format, ...)
276{
277 va_list args;
278 int ret;
279
280 va_start(args, format);
281 ret = shader_vaddline(buffer, format, args);
282 va_end(args);
283
284 return ret;
285}
286
287static void shader_init(struct IWineD3DBaseShaderClass *shader, IWineD3DDeviceImpl *device,
288 IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
289{
290 shader->ref = 1;
291 shader->device = (IWineD3DDevice *)device;
292 shader->parent = parent;
293 shader->parent_ops = parent_ops;
294 list_init(&shader->linked_programs);
295 list_add_head(&device->shaders, &shader->shader_list_entry);
296}
297
298/* Convert floating point offset relative to a register file to an absolute
299 * offset for float constants. */
300static unsigned int shader_get_float_offset(WINED3DSHADER_PARAM_REGISTER_TYPE register_type, UINT register_idx)
301{
302 switch (register_type)
303 {
304 case WINED3DSPR_CONST: return register_idx;
305 case WINED3DSPR_CONST2: return 2048 + register_idx;
306 case WINED3DSPR_CONST3: return 4096 + register_idx;
307 case WINED3DSPR_CONST4: return 6144 + register_idx;
308 default:
309 FIXME("Unsupported register type: %u.\n", register_type);
310 return register_idx;
311 }
312}
313
314static void shader_delete_constant_list(struct list *clist)
315{
316 struct local_constant *constant;
317 struct list *ptr;
318
319 ptr = list_head(clist);
320 while (ptr)
321 {
322 constant = LIST_ENTRY(ptr, struct local_constant, entry);
323 ptr = list_next(clist, ptr);
324 HeapFree(GetProcessHeap(), 0, constant);
325 }
326 list_init(clist);
327}
328
329static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
330{
331 DWORD idx, shift;
332 idx = bit >> 5;
333 shift = bit & 0x1f;
334 bitmap[idx] |= (1 << shift);
335}
336
337static HRESULT shader_record_register_usage(IWineD3DBaseShaderImpl *shader, struct shader_reg_maps *reg_maps,
338 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type)
339{
340 switch (reg->type)
341 {
342 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
343 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
344 {
345 AssertReturn(reg->idx < MAX_REG_TEXCRD, E_INVALIDARG);
346 reg_maps->texcoord |= 1 << reg->idx;
347 }
348 else
349 {
350 AssertReturn(reg->idx < MAX_REG_ADDR, E_INVALIDARG);
351 reg_maps->address |= 1 << reg->idx;
352 }
353 break;
354
355 case WINED3DSPR_TEMP:
356 AssertReturn(reg->idx < MAX_REG_TEMP, E_INVALIDARG);
357 reg_maps->temporary |= 1 << reg->idx;
358 break;
359
360 case WINED3DSPR_INPUT:
361 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
362 {
363 if (reg->rel_addr)
364 {
365 /* If relative addressing is used, we must assume that all registers
366 * are used. Even if it is a construct like v3[aL], we can't assume
367 * that v0, v1 and v2 aren't read because aL can be negative */
368 unsigned int i;
369 for (i = 0; i < MAX_REG_INPUT; ++i)
370 {
371 ((IWineD3DPixelShaderImpl *)shader)->input_reg_used[i] = TRUE;
372 }
373 }
374 else
375 {
376 unsigned int reg_idx = reg->idx;
377 AssertReturn(reg_idx < MAX_REG_INPUT, E_INVALIDARG);
378
379 ((IWineD3DPixelShaderImpl *)shader)->input_reg_used[reg_idx] = TRUE;
380 }
381 }
382 else
383 {
384 AssertReturn(reg->idx < MAX_REG_INPUT, E_INVALIDARG);
385 reg_maps->input_registers |= 1 << reg->idx;
386 }
387 break;
388
389 case WINED3DSPR_RASTOUT:
390 if (reg->idx == 1) reg_maps->fog = 1;
391 break;
392
393 case WINED3DSPR_MISCTYPE:
394 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
395 {
396 if (reg->idx == 0) reg_maps->vpos = 1;
397 else if (reg->idx == 1) reg_maps->usesfacing = 1;
398 }
399 break;
400
401 case WINED3DSPR_CONST:
402 if (reg->rel_addr)
403 {
404 if (shader_type != WINED3D_SHADER_TYPE_PIXEL)
405 {
406 if (reg->idx < ((IWineD3DVertexShaderImpl *)shader)->min_rel_offset)
407 {
408 ((IWineD3DVertexShaderImpl *)shader)->min_rel_offset = reg->idx;
409 }
410 if (reg->idx > ((IWineD3DVertexShaderImpl *)shader)->max_rel_offset)
411 {
412 ((IWineD3DVertexShaderImpl *)shader)->max_rel_offset = reg->idx;
413 }
414 }
415 reg_maps->usesrelconstF = TRUE;
416 }
417 else
418 {
419 AssertReturn(reg->idx < reg_maps->constf_size, E_INVALIDARG);
420 set_bitmap_bit(reg_maps->constf, reg->idx);
421 }
422 break;
423
424 case WINED3DSPR_CONSTINT:
425 AssertReturn(reg->idx < MAX_CONST_I, E_INVALIDARG);
426 reg_maps->integer_constants |= (1 << reg->idx);
427 break;
428
429 case WINED3DSPR_CONSTBOOL:
430 AssertReturn(reg->idx < MAX_CONST_B, E_INVALIDARG);
431 reg_maps->boolean_constants |= (1 << reg->idx);
432 break;
433
434 case WINED3DSPR_COLOROUT:
435 reg_maps->highest_render_target = max(reg_maps->highest_render_target, reg->idx);
436 break;
437
438 default:
439 TRACE("Not recording register of type %#x and idx %u\n", reg->type, reg->idx);
440 return E_INVALIDARG;
441 }
442
443 return S_OK;
444}
445
446static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
447{
448 switch (instr)
449 {
450 case WINED3DSIH_M4x4:
451 case WINED3DSIH_M3x4:
452 return param == 1 ? 3 : 0;
453
454 case WINED3DSIH_M4x3:
455 case WINED3DSIH_M3x3:
456 return param == 1 ? 2 : 0;
457
458 case WINED3DSIH_M3x2:
459 return param == 1 ? 1 : 0;
460
461 default:
462 return 0;
463 }
464}
465
466/* Note that this does not count the loop register as an address register. */
467static HRESULT shader_get_registers_used(IWineD3DBaseShader *iface, const struct wined3d_shader_frontend *fe,
468 struct shader_reg_maps *reg_maps, struct wined3d_shader_signature_element *input_signature,
469 struct wined3d_shader_signature_element *output_signature, const DWORD *byte_code, DWORD constf_size)
470{
471 IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)iface;
472 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
473 void *fe_data = shader->baseShader.frontend_data;
474 struct wined3d_shader_version shader_version;
475 const DWORD *ptr = byte_code;
476
477 memset(reg_maps, 0, sizeof(*reg_maps));
478
479 /* get_registers_used() is called on every compile on some 1.x shaders,
480 * which can result in stacking up a collection of local constants.
481 * Delete the old constants if existing. */
482 shader_delete_constant_list(&shader->baseShader.constantsF);
483 shader_delete_constant_list(&shader->baseShader.constantsB);
484 shader_delete_constant_list(&shader->baseShader.constantsI);
485
486 fe->shader_read_header(fe_data, &ptr, &shader_version);
487 reg_maps->shader_version = shader_version;
488
489 reg_maps->constf = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY,
490 sizeof(*reg_maps->constf) * ((constf_size + 31) / 32));
491 if (!reg_maps->constf)
492 {
493 ERR("Failed to allocate constant map memory.\n");
494 return E_OUTOFMEMORY;
495 }
496
497 reg_maps->constf_size = constf_size;
498
499 while (!fe->shader_is_end(fe_data, &ptr))
500 {
501 struct wined3d_shader_instruction ins;
502 const char *comment;
503 UINT comment_size;
504 UINT param_size;
505
506 /* Skip comments. */
507 fe->shader_read_comment(&ptr, &comment, &comment_size);
508 if (comment) continue;
509
510 /* Fetch opcode. */
511 fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
512
513 /* Unhandled opcode, and its parameters. */
514 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
515 {
516 TRACE("Skipping unrecognized instruction.\n");
517 ptr += param_size;
518 continue;
519 }
520
521 /* Handle declarations. */
522 if (ins.handler_idx == WINED3DSIH_DCL)
523 {
524 struct wined3d_shader_semantic semantic;
525 unsigned int reg_idx;
526
527 fe->shader_read_semantic(&ptr, &semantic);
528 reg_idx = semantic.reg.reg.idx;
529
530 switch (semantic.reg.reg.type)
531 {
532 /* Mark input registers used. */
533 case WINED3DSPR_INPUT:
534 AssertReturn(reg_idx < max(MAX_ATTRIBS, MAX_REG_INPUT), E_INVALIDARG);
535 reg_maps->input_registers |= 1 << reg_idx;
536 shader_signature_from_semantic(&input_signature[reg_idx], &semantic);
537 break;
538
539 /* Vertex shader: mark 3.0 output registers used, save token. */
540 case WINED3DSPR_OUTPUT:
541 AssertReturn(reg_idx < MAX_REG_OUTPUT, E_INVALIDARG);
542 reg_maps->output_registers |= 1 << reg_idx;
543 shader_signature_from_semantic(&output_signature[reg_idx], &semantic);
544 if (semantic.usage == WINED3DDECLUSAGE_FOG) reg_maps->fog = 1;
545 break;
546
547 /* Save sampler usage token. */
548 case WINED3DSPR_SAMPLER:
549 AssertReturn(reg_idx < RT_ELEMENTS(reg_maps->sampler_type), E_INVALIDARG);
550 reg_maps->sampler_type[reg_idx] = semantic.sampler_type;
551 break;
552
553 default:
554 TRACE("Not recording DCL register type %#x.\n", semantic.reg.reg.type);
555 break;
556 }
557 }
558 else if (ins.handler_idx == WINED3DSIH_DEF)
559 {
560 struct wined3d_shader_src_param rel_addr;
561 struct wined3d_shader_dst_param dst;
562 local_constant* lconst;
563
564 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
565
566 AssertReturn(dst.reg.idx < constf_size, E_INVALIDARG);
567
568 lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
569 if (!lconst) return E_OUTOFMEMORY;
570
571 lconst->idx = dst.reg.idx;
572 memcpy(lconst->value, ptr, 4 * sizeof(DWORD));
573 ptr += 4;
574
575 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
576 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
577 {
578 float *value = (float *)lconst->value;
579 if (value[0] < -1.0f) value[0] = -1.0f;
580 else if (value[0] > 1.0f) value[0] = 1.0f;
581 if (value[1] < -1.0f) value[1] = -1.0f;
582 else if (value[1] > 1.0f) value[1] = 1.0f;
583 if (value[2] < -1.0f) value[2] = -1.0f;
584 else if (value[2] > 1.0f) value[2] = 1.0f;
585 if (value[3] < -1.0f) value[3] = -1.0f;
586 else if (value[3] > 1.0f) value[3] = 1.0f;
587 }
588
589 list_add_head(&shader->baseShader.constantsF, &lconst->entry);
590 }
591 else if (ins.handler_idx == WINED3DSIH_DEFI)
592 {
593 struct wined3d_shader_src_param rel_addr;
594 struct wined3d_shader_dst_param dst;
595 local_constant* lconst;
596
597 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
598
599 AssertReturn(dst.reg.idx < MAX_CONST_I, E_INVALIDARG);
600
601 lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
602 if (!lconst) return E_OUTOFMEMORY;
603
604 lconst->idx = dst.reg.idx;
605 memcpy(lconst->value, ptr, 4 * sizeof(DWORD));
606 ptr += 4;
607
608 list_add_head(&shader->baseShader.constantsI, &lconst->entry);
609 reg_maps->local_int_consts |= (1 << dst.reg.idx);
610 }
611 else if (ins.handler_idx == WINED3DSIH_DEFB)
612 {
613 struct wined3d_shader_src_param rel_addr;
614 struct wined3d_shader_dst_param dst;
615 local_constant* lconst;
616
617 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
618
619 AssertReturn(dst.reg.idx < MAX_CONST_B, E_INVALIDARG);
620
621 lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
622 if (!lconst) return E_OUTOFMEMORY;
623
624 lconst->idx = dst.reg.idx;
625 memcpy(lconst->value, ptr, sizeof(DWORD));
626 ++ptr;
627
628 list_add_head(&shader->baseShader.constantsB, &lconst->entry);
629 reg_maps->local_bool_consts |= (1 << dst.reg.idx);
630 }
631 /* If there's a loop in the shader. */
632 else if (ins.handler_idx == WINED3DSIH_LOOP
633 || ins.handler_idx == WINED3DSIH_REP)
634 {
635 struct wined3d_shader_src_param src, rel_addr;
636
637 fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
638
639 /* Rep and Loop always use an integer constant for the control parameters. */
640 if (ins.handler_idx == WINED3DSIH_REP)
641 {
642 reg_maps->integer_constants |= 1 << src.reg.idx;
643 }
644 else
645 {
646 fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
647 reg_maps->integer_constants |= 1 << src.reg.idx;
648 }
649
650 cur_loop_depth++;
651 if (cur_loop_depth > max_loop_depth) max_loop_depth = cur_loop_depth;
652 }
653 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
654 || ins.handler_idx == WINED3DSIH_ENDREP)
655 {
656 cur_loop_depth--;
657 }
658 /* For subroutine prototypes. */
659 else if (ins.handler_idx == WINED3DSIH_LABEL)
660 {
661 struct wined3d_shader_src_param src, rel_addr;
662
663 fe->shader_read_src_param(fe_data, &ptr, &src, &rel_addr);
664 reg_maps->labels |= 1 << src.reg.idx;
665 }
666 /* Set texture, address, temporary registers. */
667 else
668 {
669 BOOL color0_mov = FALSE;
670 int i, limit;
671
672 /* This will loop over all the registers and try to
673 * make a bitmask of the ones we're interested in.
674 *
675 * Relative addressing tokens are ignored, but that's
676 * okay, since we'll catch any address registers when
677 * they are initialized (required by spec). */
678 if (ins.dst_count)
679 {
680 struct wined3d_shader_src_param dst_rel_addr;
681 struct wined3d_shader_dst_param dst_param;
682 HRESULT hr;
683
684 fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
685
686 hr = shader_record_register_usage(shader, reg_maps, &dst_param.reg, shader_version.type);
687 AssertReturn(FAILED(hr), hr);
688
689 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and
690 * is used in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel
691 * shaders because TECRDOUT isn't used in them, but future register types might cause issues */
692 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX && shader_version.major < 3
693 && dst_param.reg.type == WINED3DSPR_TEXCRDOUT)
694 {
695 unsigned int idx = dst_param.reg.idx;
696 AssertReturn(idx < MAX_REG_TEXCRD, E_INVALIDARG);
697
698 reg_maps->texcoord_mask[idx] |= dst_param.write_mask;
699 }
700
701 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
702 {
703 IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)shader;
704
705 if (dst_param.reg.type == WINED3DSPR_COLOROUT && dst_param.reg.idx == 0)
706 {
707 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
708 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
709 * the mov and perform the sRGB write correction from the source register.
710 *
711 * However, if the mov is only partial, we can't do this, and if the write
712 * comes from an instruction other than MOV it is hard to do as well. If
713 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
714
715 ps->color0_mov = FALSE;
716 if (ins.handler_idx == WINED3DSIH_MOV)
717 {
718 /* Used later when the source register is read. */
719 color0_mov = TRUE;
720 }
721 }
722 /* Also drop the MOV marker if the source register is overwritten prior to the shader
723 * end
724 */
725 else if (dst_param.reg.type == WINED3DSPR_TEMP && dst_param.reg.idx == ps->color0_reg)
726 {
727 ps->color0_mov = FALSE;
728 }
729 }
730
731 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
732 if (ins.handler_idx == WINED3DSIH_TEX
733 || ins.handler_idx == WINED3DSIH_TEXBEM
734 || ins.handler_idx == WINED3DSIH_TEXBEML
735 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
736 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
737 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
738 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
739 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
740 || ins.handler_idx == WINED3DSIH_TEXREG2AR
741 || ins.handler_idx == WINED3DSIH_TEXREG2GB
742 || ins.handler_idx == WINED3DSIH_TEXREG2RGB)
743 {
744 unsigned int sampler_idx;
745 AssertReturn(shader_version.major == 1, E_INVALIDARG);
746
747 /* Fake sampler usage, only set reserved bit and type. */
748 sampler_idx = dst_param.reg.idx;
749 AssertReturn(sampler_idx < RT_ELEMENTS(reg_maps->sampler_type), E_INVALIDARG);
750
751 TRACE("Setting fake 2D sampler for 1.x pixelshader.\n");
752 reg_maps->sampler_type[sampler_idx] = WINED3DSTT_2D;
753
754 /* texbem is only valid with < 1.4 pixel shaders */
755 if (ins.handler_idx == WINED3DSIH_TEXBEM
756 || ins.handler_idx == WINED3DSIH_TEXBEML)
757 {
758 reg_maps->bumpmat |= 1 << sampler_idx;
759 if (ins.handler_idx == WINED3DSIH_TEXBEML)
760 {
761 reg_maps->luminanceparams |= 1 << sampler_idx;
762 }
763 }
764 }
765 else if (ins.handler_idx == WINED3DSIH_BEM)
766 {
767 unsigned int sampler_idx = dst_param.reg.idx;
768 AssertReturn(shader_version.major == 1, E_INVALIDARG);
769 AssertReturn(sampler_idx < RT_ELEMENTS(reg_maps->sampler_type), E_INVALIDARG);
770
771 reg_maps->bumpmat |= 1 << sampler_idx;
772 }
773 }
774
775 if (ins.handler_idx == WINED3DSIH_NRM) reg_maps->usesnrm = 1;
776 else if (ins.handler_idx == WINED3DSIH_DSY) reg_maps->usesdsy = 1;
777 else if (ins.handler_idx == WINED3DSIH_DSX) reg_maps->usesdsx = 1;
778 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
779 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
780 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
781 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
782 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
783
784 limit = ins.src_count + (ins.predicate ? 1 : 0);
785 for (i = 0; i < limit; ++i)
786 {
787 struct wined3d_shader_src_param src_param, src_rel_addr;
788 unsigned int count;
789 HRESULT hr;
790
791 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
792 count = get_instr_extra_regcount(ins.handler_idx, i);
793
794 hr = shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
795 AssertReturn(FAILED(hr), hr);
796
797 while (count)
798 {
799 ++src_param.reg.idx;
800 hr = shader_record_register_usage(shader, reg_maps, &src_param.reg, shader_version.type);
801 AssertReturn(FAILED(hr), hr);
802 --count;
803 }
804
805 if (color0_mov)
806 {
807 IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)shader;
808 if (src_param.reg.type == WINED3DSPR_TEMP
809 && src_param.swizzle == WINED3DSP_NOSWIZZLE)
810 {
811 ps->color0_mov = TRUE;
812 ps->color0_reg = src_param.reg.idx;
813 }
814 }
815 }
816 }
817 }
818 reg_maps->loop_depth = max_loop_depth;
819
820 shader->baseShader.functionLength = ((const char *)ptr - (const char *)byte_code);
821
822 return WINED3D_OK;
823}
824
825unsigned int shader_find_free_input_register(const struct shader_reg_maps *reg_maps, unsigned int max)
826{
827 DWORD map = 1 << max;
828 map |= map - 1;
829 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
830
831 return wined3d_log2i(map);
832}
833
834static void shader_dump_decl_usage(const struct wined3d_shader_semantic *semantic,
835 const struct wined3d_shader_version *shader_version)
836{
837 TRACE("dcl");
838
839 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
840 {
841 switch (semantic->sampler_type)
842 {
843 case WINED3DSTT_2D: TRACE("_2d"); break;
844 case WINED3DSTT_CUBE: TRACE("_cube"); break;
845 case WINED3DSTT_VOLUME: TRACE("_volume"); break;
846 default: TRACE("_unknown_ttype(0x%08x)", semantic->sampler_type);
847 }
848 }
849 else
850 {
851 /* Pixel shaders 3.0 don't have usage semantics. */
852 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL) return;
853 else TRACE("_");
854
855 switch (semantic->usage)
856 {
857 case WINED3DDECLUSAGE_POSITION:
858 TRACE("position%u", semantic->usage_idx);
859 break;
860
861 case WINED3DDECLUSAGE_BLENDINDICES:
862 TRACE("blend");
863 break;
864
865 case WINED3DDECLUSAGE_BLENDWEIGHT:
866 TRACE("weight");
867 break;
868
869 case WINED3DDECLUSAGE_NORMAL:
870 TRACE("normal%u", semantic->usage_idx);
871 break;
872
873 case WINED3DDECLUSAGE_PSIZE:
874 TRACE("psize");
875 break;
876
877 case WINED3DDECLUSAGE_COLOR:
878 if (semantic->usage_idx == 0) TRACE("color");
879 else TRACE("specular%u", (semantic->usage_idx - 1));
880 break;
881
882 case WINED3DDECLUSAGE_TEXCOORD:
883 TRACE("texture%u", semantic->usage_idx);
884 break;
885
886 case WINED3DDECLUSAGE_TANGENT:
887 TRACE("tangent");
888 break;
889
890 case WINED3DDECLUSAGE_BINORMAL:
891 TRACE("binormal");
892 break;
893
894 case WINED3DDECLUSAGE_TESSFACTOR:
895 TRACE("tessfactor");
896 break;
897
898 case WINED3DDECLUSAGE_POSITIONT:
899 TRACE("positionT%u", semantic->usage_idx);
900 break;
901
902 case WINED3DDECLUSAGE_FOG:
903 TRACE("fog");
904 break;
905
906 case WINED3DDECLUSAGE_DEPTH:
907 TRACE("depth");
908 break;
909
910 case WINED3DDECLUSAGE_SAMPLE:
911 TRACE("sample");
912 break;
913
914 default:
915 FIXME("unknown_semantics(0x%08x)", semantic->usage);
916 }
917 }
918}
919
920static void shader_dump_register(const struct wined3d_shader_register *reg,
921 const struct wined3d_shader_version *shader_version)
922{
923 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
924 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
925 UINT offset = reg->idx;
926
927 switch (reg->type)
928 {
929 case WINED3DSPR_TEMP:
930 TRACE("r");
931 break;
932
933 case WINED3DSPR_INPUT:
934 TRACE("v");
935 break;
936
937 case WINED3DSPR_CONST:
938 case WINED3DSPR_CONST2:
939 case WINED3DSPR_CONST3:
940 case WINED3DSPR_CONST4:
941 TRACE("c");
942 offset = shader_get_float_offset(reg->type, reg->idx);
943 break;
944
945 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
946 TRACE("%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
947 break;
948
949 case WINED3DSPR_RASTOUT:
950 TRACE("%s", rastout_reg_names[reg->idx]);
951 break;
952
953 case WINED3DSPR_COLOROUT:
954 TRACE("oC");
955 break;
956
957 case WINED3DSPR_DEPTHOUT:
958 TRACE("oDepth");
959 break;
960
961 case WINED3DSPR_ATTROUT:
962 TRACE("oD");
963 break;
964
965 case WINED3DSPR_TEXCRDOUT:
966 /* Vertex shaders >= 3.0 use general purpose output registers
967 * (WINED3DSPR_OUTPUT), which can include an address token. */
968 if (shader_version->major >= 3) TRACE("o");
969 else TRACE("oT");
970 break;
971
972 case WINED3DSPR_CONSTINT:
973 TRACE("i");
974 break;
975
976 case WINED3DSPR_CONSTBOOL:
977 TRACE("b");
978 break;
979
980 case WINED3DSPR_LABEL:
981 TRACE("l");
982 break;
983
984 case WINED3DSPR_LOOP:
985 TRACE("aL");
986 break;
987
988 case WINED3DSPR_SAMPLER:
989 TRACE("s");
990 break;
991
992 case WINED3DSPR_MISCTYPE:
993 if (reg->idx > 1) FIXME("Unhandled misctype register %u.\n", reg->idx);
994 else TRACE("%s", misctype_reg_names[reg->idx]);
995 break;
996
997 case WINED3DSPR_PREDICATE:
998 TRACE("p");
999 break;
1000
1001 case WINED3DSPR_IMMCONST:
1002 TRACE("l");
1003 break;
1004
1005 case WINED3DSPR_CONSTBUFFER:
1006 TRACE("cb");
1007 break;
1008
1009 default:
1010 TRACE("unhandled_rtype(%#x)", reg->type);
1011 break;
1012 }
1013
1014 if (reg->type == WINED3DSPR_IMMCONST)
1015 {
1016 TRACE("(");
1017 switch (reg->immconst_type)
1018 {
1019 case WINED3D_IMMCONST_FLOAT:
1020 TRACE("%.8e", *(const float *)reg->immconst_data);
1021 break;
1022
1023 case WINED3D_IMMCONST_FLOAT4:
1024 TRACE("%.8e, %.8e, %.8e, %.8e",
1025 *(const float *)&reg->immconst_data[0], *(const float *)&reg->immconst_data[1],
1026 *(const float *)&reg->immconst_data[2], *(const float *)&reg->immconst_data[3]);
1027 break;
1028
1029 default:
1030 TRACE("<unhandled immconst_type %#x>", reg->immconst_type);
1031 break;
1032 }
1033 TRACE(")");
1034 }
1035 else if (reg->type != WINED3DSPR_RASTOUT && reg->type != WINED3DSPR_MISCTYPE)
1036 {
1037 if (reg->array_idx != ~0U)
1038 {
1039 TRACE("%u[%u", offset, reg->array_idx);
1040 if (reg->rel_addr)
1041 {
1042 TRACE(" + ");
1043 shader_dump_src_param(reg->rel_addr, shader_version);
1044 }
1045 TRACE("]");
1046 }
1047 else
1048 {
1049 if (reg->rel_addr)
1050 {
1051 TRACE("[");
1052 shader_dump_src_param(reg->rel_addr, shader_version);
1053 TRACE(" + ");
1054 }
1055 TRACE("%u", offset);
1056 if (reg->rel_addr) TRACE("]");
1057 }
1058 }
1059}
1060
1061void shader_dump_dst_param(const struct wined3d_shader_dst_param *param,
1062 const struct wined3d_shader_version *shader_version)
1063{
1064 DWORD write_mask = param->write_mask;
1065
1066 shader_dump_register(&param->reg, shader_version);
1067
1068 if (write_mask != WINED3DSP_WRITEMASK_ALL)
1069 {
1070 static const char *write_mask_chars = "xyzw";
1071
1072 TRACE(".");
1073 if (write_mask & WINED3DSP_WRITEMASK_0) TRACE("%c", write_mask_chars[0]);
1074 if (write_mask & WINED3DSP_WRITEMASK_1) TRACE("%c", write_mask_chars[1]);
1075 if (write_mask & WINED3DSP_WRITEMASK_2) TRACE("%c", write_mask_chars[2]);
1076 if (write_mask & WINED3DSP_WRITEMASK_3) TRACE("%c", write_mask_chars[3]);
1077 }
1078}
1079
1080void shader_dump_src_param(const struct wined3d_shader_src_param *param,
1081 const struct wined3d_shader_version *shader_version)
1082{
1083 DWORD src_modifier = param->modifiers;
1084 DWORD swizzle = param->swizzle;
1085
1086 if (src_modifier == WINED3DSPSM_NEG
1087 || src_modifier == WINED3DSPSM_BIASNEG
1088 || src_modifier == WINED3DSPSM_SIGNNEG
1089 || src_modifier == WINED3DSPSM_X2NEG
1090 || src_modifier == WINED3DSPSM_ABSNEG)
1091 TRACE("-");
1092 else if (src_modifier == WINED3DSPSM_COMP)
1093 TRACE("1-");
1094 else if (src_modifier == WINED3DSPSM_NOT)
1095 TRACE("!");
1096
1097 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
1098 TRACE("abs(");
1099
1100 shader_dump_register(&param->reg, shader_version);
1101
1102 if (src_modifier)
1103 {
1104 switch (src_modifier)
1105 {
1106 case WINED3DSPSM_NONE: break;
1107 case WINED3DSPSM_NEG: break;
1108 case WINED3DSPSM_NOT: break;
1109 case WINED3DSPSM_BIAS: TRACE("_bias"); break;
1110 case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
1111 case WINED3DSPSM_SIGN: TRACE("_bx2"); break;
1112 case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
1113 case WINED3DSPSM_COMP: break;
1114 case WINED3DSPSM_X2: TRACE("_x2"); break;
1115 case WINED3DSPSM_X2NEG: TRACE("_x2"); break;
1116 case WINED3DSPSM_DZ: TRACE("_dz"); break;
1117 case WINED3DSPSM_DW: TRACE("_dw"); break;
1118 case WINED3DSPSM_ABSNEG: TRACE(")"); break;
1119 case WINED3DSPSM_ABS: TRACE(")"); break;
1120 default: TRACE("_unknown_modifier(%#x)", src_modifier);
1121 }
1122 }
1123
1124 if (swizzle != WINED3DSP_NOSWIZZLE)
1125 {
1126 static const char *swizzle_chars = "xyzw";
1127 DWORD swizzle_x = swizzle & 0x03;
1128 DWORD swizzle_y = (swizzle >> 2) & 0x03;
1129 DWORD swizzle_z = (swizzle >> 4) & 0x03;
1130 DWORD swizzle_w = (swizzle >> 6) & 0x03;
1131
1132 if (swizzle_x == swizzle_y
1133 && swizzle_x == swizzle_z
1134 && swizzle_x == swizzle_w)
1135 {
1136 TRACE(".%c", swizzle_chars[swizzle_x]);
1137 }
1138 else
1139 {
1140 TRACE(".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
1141 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
1142 }
1143 }
1144}
1145
1146/* Shared code in order to generate the bulk of the shader string.
1147 * NOTE: A description of how to parse tokens can be found on MSDN. */
1148void shader_generate_main(IWineD3DBaseShader *iface, struct wined3d_shader_buffer *buffer,
1149 const shader_reg_maps *reg_maps, const DWORD *byte_code, void *backend_ctx)
1150{
1151 IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)iface;
1152 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1153 const struct wined3d_shader_frontend *fe = shader->baseShader.frontend;
1154 void *fe_data = shader->baseShader.frontend_data;
1155 struct wined3d_shader_src_param src_rel_addr[4];
1156 struct wined3d_shader_src_param src_param[4];
1157 struct wined3d_shader_version shader_version;
1158 struct wined3d_shader_src_param dst_rel_addr;
1159 struct wined3d_shader_dst_param dst_param;
1160 struct wined3d_shader_instruction ins;
1161 struct wined3d_shader_context ctx;
1162 const DWORD *ptr = byte_code;
1163 DWORD i;
1164
1165 /* Initialize current parsing state. */
1166 ctx.shader = iface;
1167 ctx.gl_info = &device->adapter->gl_info;
1168 ctx.reg_maps = reg_maps;
1169 ctx.buffer = buffer;
1170 ctx.backend_data = backend_ctx;
1171
1172 ins.ctx = &ctx;
1173 ins.dst = &dst_param;
1174 ins.src = src_param;
1175 shader->baseShader.parse_state.current_row = 0;
1176
1177 fe->shader_read_header(fe_data, &ptr, &shader_version);
1178
1179 while (!fe->shader_is_end(fe_data, &ptr))
1180 {
1181 const char *comment;
1182 UINT comment_size;
1183 UINT param_size;
1184
1185 /* Skip comment tokens. */
1186 fe->shader_read_comment(&ptr, &comment, &comment_size);
1187 if (comment) continue;
1188
1189 /* Read opcode. */
1190 fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1191
1192 /* Unknown opcode and its parameters. */
1193 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1194 {
1195 TRACE("Skipping unrecognized instruction.\n");
1196 ptr += param_size;
1197 continue;
1198 }
1199
1200 /* Nothing to do. */
1201 if (ins.handler_idx == WINED3DSIH_DCL
1202 || ins.handler_idx == WINED3DSIH_NOP
1203 || ins.handler_idx == WINED3DSIH_DEF
1204 || ins.handler_idx == WINED3DSIH_DEFI
1205 || ins.handler_idx == WINED3DSIH_DEFB
1206 || ins.handler_idx == WINED3DSIH_PHASE)
1207 {
1208 ptr += param_size;
1209 continue;
1210 }
1211
1212 /* Destination token */
1213 if (ins.dst_count) fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
1214
1215 /* Predication token */
1216 if (ins.predicate)
1217 {
1218#ifndef VBOX_WITH_VMSVGA
1219 FIXME("Predicates not implemented.\n");
1220#endif
1221 ins.predicate = *ptr++;
1222 }
1223
1224 /* Other source tokens */
1225 for (i = 0; i < ins.src_count; ++i)
1226 {
1227 fe->shader_read_src_param(fe_data, &ptr, &src_param[i], &src_rel_addr[i]);
1228 }
1229
1230 /* Call appropriate function for output target */
1231 device->shader_backend->shader_handle_instruction(&ins);
1232 }
1233}
1234
1235static void shader_dump_ins_modifiers(const struct wined3d_shader_dst_param *dst)
1236{
1237 DWORD mmask = dst->modifiers;
1238
1239 switch (dst->shift)
1240 {
1241 case 0: break;
1242 case 13: TRACE("_d8"); break;
1243 case 14: TRACE("_d4"); break;
1244 case 15: TRACE("_d2"); break;
1245 case 1: TRACE("_x2"); break;
1246 case 2: TRACE("_x4"); break;
1247 case 3: TRACE("_x8"); break;
1248 default: TRACE("_unhandled_shift(%d)", dst->shift); break;
1249 }
1250
1251 if (mmask & WINED3DSPDM_SATURATE) TRACE("_sat");
1252 if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
1253 if (mmask & WINED3DSPDM_MSAMPCENTROID) TRACE("_centroid");
1254
1255 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
1256 if (mmask) FIXME("_unrecognized_modifier(%#x)", mmask);
1257}
1258
1259static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data, const DWORD *byte_code)
1260{
1261 struct wined3d_shader_version shader_version;
1262 const DWORD *ptr = byte_code;
1263 const char *type_prefix;
1264 DWORD i;
1265
1266 TRACE("Parsing %p.\n", byte_code);
1267
1268 fe->shader_read_header(fe_data, &ptr, &shader_version);
1269
1270 switch (shader_version.type)
1271 {
1272 case WINED3D_SHADER_TYPE_VERTEX:
1273 type_prefix = "vs";
1274 break;
1275
1276 case WINED3D_SHADER_TYPE_GEOMETRY:
1277 type_prefix = "gs";
1278 break;
1279
1280 case WINED3D_SHADER_TYPE_PIXEL:
1281 type_prefix = "ps";
1282 break;
1283
1284 default:
1285 FIXME("Unhandled shader type %#x.\n", shader_version.type);
1286 type_prefix = "unknown";
1287 break;
1288 }
1289
1290 TRACE("%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
1291
1292 while (!fe->shader_is_end(fe_data, &ptr))
1293 {
1294 struct wined3d_shader_instruction ins;
1295 const char *comment;
1296 UINT comment_size;
1297 UINT param_size;
1298
1299 /* comment */
1300 fe->shader_read_comment(&ptr, &comment, &comment_size);
1301 if (comment)
1302 {
1303 if (comment_size > 4 && *(const DWORD *)comment == WINEMAKEFOURCC('T', 'E', 'X', 'T'))
1304 {
1305 const char *end = comment + comment_size;
1306 const char *cur = comment + 4;
1307 const char *line = cur;
1308
1309 TRACE("// TEXT\n");
1310 while (cur != end)
1311 {
1312 if (*cur == '\n')
1313 {
1314 UINT len = cur - line;
1315 if (len && *(cur - 1) == '\r') --len;
1316 TRACE("// %s\n", debugstr_an(line, len));
1317 line = ++cur;
1318 }
1319 else ++cur;
1320 }
1321 if (line != cur) TRACE("// %s\n", debugstr_an(line, cur - line));
1322 }
1323 else TRACE("// %s\n", debugstr_an(comment, comment_size));
1324 continue;
1325 }
1326
1327 fe->shader_read_opcode(fe_data, &ptr, &ins, &param_size);
1328 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1329 {
1330 TRACE("Skipping unrecognized instruction.\n");
1331 ptr += param_size;
1332 continue;
1333 }
1334
1335 if (ins.handler_idx == WINED3DSIH_DCL)
1336 {
1337 struct wined3d_shader_semantic semantic;
1338
1339 fe->shader_read_semantic(&ptr, &semantic);
1340
1341 shader_dump_decl_usage(&semantic, &shader_version);
1342 shader_dump_ins_modifiers(&semantic.reg);
1343 TRACE(" ");
1344 shader_dump_dst_param(&semantic.reg, &shader_version);
1345 }
1346 else if (ins.handler_idx == WINED3DSIH_DEF)
1347 {
1348 struct wined3d_shader_src_param rel_addr;
1349 struct wined3d_shader_dst_param dst;
1350
1351 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1352
1353 TRACE("def c%u = %f, %f, %f, %f", shader_get_float_offset(dst.reg.type, dst.reg.idx),
1354 *(const float *)(ptr),
1355 *(const float *)(ptr + 1),
1356 *(const float *)(ptr + 2),
1357 *(const float *)(ptr + 3));
1358 ptr += 4;
1359 }
1360 else if (ins.handler_idx == WINED3DSIH_DEFI)
1361 {
1362 struct wined3d_shader_src_param rel_addr;
1363 struct wined3d_shader_dst_param dst;
1364
1365 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1366
1367 TRACE("defi i%u = %d, %d, %d, %d", dst.reg.idx,
1368 *(ptr),
1369 *(ptr + 1),
1370 *(ptr + 2),
1371 *(ptr + 3));
1372 ptr += 4;
1373 }
1374 else if (ins.handler_idx == WINED3DSIH_DEFB)
1375 {
1376 struct wined3d_shader_src_param rel_addr;
1377 struct wined3d_shader_dst_param dst;
1378
1379 fe->shader_read_dst_param(fe_data, &ptr, &dst, &rel_addr);
1380
1381 TRACE("defb b%u = %s", dst.reg.idx, *ptr ? "true" : "false");
1382 ++ptr;
1383 }
1384 else
1385 {
1386 struct wined3d_shader_src_param dst_rel_addr, src_rel_addr;
1387 struct wined3d_shader_dst_param dst_param;
1388 struct wined3d_shader_src_param src_param;
1389
1390 if (ins.dst_count)
1391 {
1392 fe->shader_read_dst_param(fe_data, &ptr, &dst_param, &dst_rel_addr);
1393 }
1394
1395 /* Print out predication source token first - it follows
1396 * the destination token. */
1397 if (ins.predicate)
1398 {
1399 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1400 TRACE("(");
1401 shader_dump_src_param(&src_param, &shader_version);
1402 TRACE(") ");
1403 }
1404
1405 /* PixWin marks instructions with the coissue flag with a '+' */
1406 if (ins.coissue) TRACE("+");
1407
1408 TRACE("%s", shader_opcode_names[ins.handler_idx]);
1409
1410 if (ins.handler_idx == WINED3DSIH_IFC
1411 || ins.handler_idx == WINED3DSIH_BREAKC)
1412 {
1413 switch (ins.flags)
1414 {
1415 case COMPARISON_GT: TRACE("_gt"); break;
1416 case COMPARISON_EQ: TRACE("_eq"); break;
1417 case COMPARISON_GE: TRACE("_ge"); break;
1418 case COMPARISON_LT: TRACE("_lt"); break;
1419 case COMPARISON_NE: TRACE("_ne"); break;
1420 case COMPARISON_LE: TRACE("_le"); break;
1421 default: TRACE("_(%u)", ins.flags);
1422 }
1423 }
1424 else if (ins.handler_idx == WINED3DSIH_TEX
1425 && shader_version.major >= 2
1426 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
1427 {
1428 TRACE("p");
1429 }
1430
1431 /* We already read the destination token, print it. */
1432 if (ins.dst_count)
1433 {
1434 shader_dump_ins_modifiers(&dst_param);
1435 TRACE(" ");
1436 shader_dump_dst_param(&dst_param, &shader_version);
1437 }
1438
1439 /* Other source tokens */
1440 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
1441 {
1442 fe->shader_read_src_param(fe_data, &ptr, &src_param, &src_rel_addr);
1443 TRACE(!i ? " " : ", ");
1444 shader_dump_src_param(&src_param, &shader_version);
1445 }
1446 }
1447 TRACE("\n");
1448 }
1449}
1450
1451static void shader_cleanup(IWineD3DBaseShader *iface)
1452{
1453 IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)iface;
1454
1455 ((IWineD3DDeviceImpl *)shader->baseShader.device)->shader_backend->shader_destroy(iface);
1456 HeapFree(GetProcessHeap(), 0, shader->baseShader.reg_maps.constf);
1457 HeapFree(GetProcessHeap(), 0, shader->baseShader.function);
1458 shader_delete_constant_list(&shader->baseShader.constantsF);
1459 shader_delete_constant_list(&shader->baseShader.constantsB);
1460 shader_delete_constant_list(&shader->baseShader.constantsI);
1461 list_remove(&shader->baseShader.shader_list_entry);
1462
1463 if (shader->baseShader.frontend && shader->baseShader.frontend_data)
1464 {
1465 shader->baseShader.frontend->shader_free(shader->baseShader.frontend_data);
1466 }
1467}
1468
1469static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1470static void shader_none_select(const struct wined3d_context *context, BOOL usePS, BOOL useVS) {}
1471static void shader_none_select_depth_blt(IWineD3DDevice *iface, enum tex_types tex_type) {}
1472static void shader_none_deselect_depth_blt(IWineD3DDevice *iface) {}
1473static void shader_none_update_float_vertex_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1474static void shader_none_update_float_pixel_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1475static void shader_none_load_constants(const struct wined3d_context *context, char usePS, char useVS) {}
1476static void shader_none_load_np2fixup_constants(IWineD3DDevice *iface, char usePS, char useVS) {}
1477static void shader_none_destroy(IWineD3DBaseShader *iface) {}
1478static HRESULT shader_none_alloc(IWineD3DDevice *iface) {return WINED3D_OK;}
1479static void shader_none_free(IWineD3DDevice *iface) {}
1480static BOOL shader_none_dirty_const(IWineD3DDevice *iface) {return FALSE;}
1481
1482static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
1483{
1484 /* Set the shader caps to 0 for the none shader backend */
1485 caps->VertexShaderVersion = 0;
1486 caps->PixelShaderVersion = 0;
1487 caps->PixelShader1xMaxValue = 0.0f;
1488}
1489
1490static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
1491{
1492 if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1493 {
1494 TRACE("Checking support for fixup:\n");
1495 dump_color_fixup_desc(fixup);
1496 }
1497
1498 /* Faked to make some apps happy. */
1499 if (!is_complex_fixup(fixup))
1500 {
1501 TRACE("[OK]\n");
1502 return TRUE;
1503 }
1504
1505 TRACE("[FAILED]\n");
1506 return FALSE;
1507}
1508
1509const shader_backend_t none_shader_backend = {
1510 shader_none_handle_instruction,
1511 shader_none_select,
1512 shader_none_select_depth_blt,
1513 shader_none_deselect_depth_blt,
1514 shader_none_update_float_vertex_constants,
1515 shader_none_update_float_pixel_constants,
1516 shader_none_load_constants,
1517 shader_none_load_np2fixup_constants,
1518 shader_none_destroy,
1519 shader_none_alloc,
1520 shader_none_free,
1521 shader_none_dirty_const,
1522 shader_none_get_caps,
1523 shader_none_color_fixup_supported,
1524};
1525
1526static void shader_get_parent(IWineD3DBaseShaderImpl *shader, IUnknown **parent)
1527{
1528 *parent = shader->baseShader.parent;
1529 IUnknown_AddRef(*parent);
1530 TRACE("shader %p, returning %p.\n", shader, *parent);
1531}
1532
1533static HRESULT shader_get_function(IWineD3DBaseShaderImpl *shader, void *data, UINT *data_size)
1534{
1535 if (!data)
1536 {
1537 *data_size = shader->baseShader.functionLength;
1538 return WINED3D_OK;
1539 }
1540
1541 if (*data_size < shader->baseShader.functionLength)
1542 {
1543 /* MSDN claims (for d3d8 at least) that if *pSizeOfData is smaller
1544 * than the required size we should write the required size and
1545 * return D3DERR_MOREDATA. That's not actually true. */
1546 return WINED3DERR_INVALIDCALL;
1547 }
1548
1549 memcpy(data, shader->baseShader.function, shader->baseShader.functionLength);
1550
1551 return WINED3D_OK;
1552}
1553
1554static HRESULT shader_set_function(IWineD3DBaseShaderImpl *shader, const DWORD *byte_code,
1555 const struct wined3d_shader_signature *output_signature, DWORD float_const_count)
1556{
1557 struct shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
1558 const struct wined3d_shader_frontend *fe;
1559 HRESULT hr;
1560
1561 TRACE("shader %p, byte_code %p, output_signature %p, float_const_count %u.\n",
1562 shader, byte_code, output_signature, float_const_count);
1563
1564 fe = shader_select_frontend(*byte_code);
1565 if (!fe)
1566 {
1567 FIXME("Unable to find frontend for shader.\n");
1568 return WINED3DERR_INVALIDCALL;
1569 }
1570 shader->baseShader.frontend = fe;
1571 shader->baseShader.frontend_data = fe->shader_init(byte_code, output_signature);
1572 if (!shader->baseShader.frontend_data)
1573 {
1574 FIXME("Failed to initialize frontend.\n");
1575 return WINED3DERR_INVALIDCALL;
1576 }
1577
1578 /* First pass: trace shader. */
1579 if (TRACE_ON(d3d_shader)) shader_trace_init(fe, shader->baseShader.frontend_data, byte_code);
1580
1581 /* Initialize immediate constant lists. */
1582 list_init(&shader->baseShader.constantsF);
1583 list_init(&shader->baseShader.constantsB);
1584 list_init(&shader->baseShader.constantsI);
1585
1586 /* Second pass: figure out which registers are used, what the semantics are, etc. */
1587 hr = shader_get_registers_used((IWineD3DBaseShader *)shader, fe,
1588 reg_maps, shader->baseShader.input_signature, shader->baseShader.output_signature,
1589 byte_code, float_const_count);
1590 if (FAILED(hr)) return hr;
1591
1592 shader->baseShader.function = HeapAlloc(GetProcessHeap(), 0, shader->baseShader.functionLength);
1593 if (!shader->baseShader.function) return E_OUTOFMEMORY;
1594 memcpy(shader->baseShader.function, byte_code, shader->baseShader.functionLength);
1595
1596 return WINED3D_OK;
1597}
1598
1599static HRESULT STDMETHODCALLTYPE vertexshader_QueryInterface(IWineD3DVertexShader *iface, REFIID riid, void **object)
1600{
1601 TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1602
1603 if (IsEqualGUID(riid, &IID_IWineD3DVertexShader)
1604 || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1605 || IsEqualGUID(riid, &IID_IWineD3DBase)
1606 || IsEqualGUID(riid, &IID_IUnknown))
1607 {
1608 IUnknown_AddRef(iface);
1609 *object = iface;
1610 return S_OK;
1611 }
1612
1613 WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1614
1615 *object = NULL;
1616 return E_NOINTERFACE;
1617}
1618
1619static ULONG STDMETHODCALLTYPE vertexshader_AddRef(IWineD3DVertexShader *iface)
1620{
1621 IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1622 ULONG refcount = InterlockedIncrement(&shader->baseShader.ref);
1623
1624 TRACE("%p increasing refcount to %u.\n", shader, refcount);
1625
1626 return refcount;
1627}
1628
1629#ifdef VBOX_WINE_WITH_SHADER_CACHE
1630static void shader_chache_put(PVBOXEXT_HASHCACHE pCache, IWineD3DBaseShaderImpl *pShader);
1631#endif
1632
1633static ULONG STDMETHODCALLTYPE vertexshader_Release(IWineD3DVertexShader *iface)
1634{
1635 IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1636 ULONG refcount = InterlockedDecrement(&shader->baseShader.ref);
1637
1638 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1639
1640 if (!refcount)
1641 {
1642#ifdef VBOX_WINE_WITH_SHADER_CACHE
1643 shader_chache_put(&((IWineD3DDeviceImpl*)shader->baseShader.device)->vshaderCache, (IWineD3DBaseShaderImpl*)shader);
1644#else
1645 shader_cleanup((IWineD3DBaseShader *)iface);
1646#ifndef VBOX_WITH_VMSVGA
1647 shader->baseShader.parent_ops->wined3d_object_destroyed(shader->baseShader.parent);
1648#endif
1649 HeapFree(GetProcessHeap(), 0, shader);
1650#endif
1651 }
1652
1653 return refcount;
1654}
1655
1656static HRESULT STDMETHODCALLTYPE vertexshader_GetParent(IWineD3DVertexShader *iface, IUnknown **parent)
1657{
1658 TRACE("iface %p, parent %p.\n", iface, parent);
1659
1660 shader_get_parent((IWineD3DBaseShaderImpl *)iface, parent);
1661
1662 return WINED3D_OK;
1663}
1664
1665static HRESULT STDMETHODCALLTYPE vertexshader_GetFunction(IWineD3DVertexShader *iface, void *data, UINT *data_size)
1666{
1667 TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1668
1669 return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1670}
1671
1672/* Set local constants for d3d8 shaders. */
1673static HRESULT STDMETHODCALLTYPE vertexshader_SetLocalConstantsF(IWineD3DVertexShader *iface,
1674 UINT start_idx, const float *src_data, UINT count)
1675{
1676 IWineD3DVertexShaderImpl *shader =(IWineD3DVertexShaderImpl *)iface;
1677 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1678 UINT i, end_idx;
1679
1680 TRACE("iface %p, start_idx %u, src_data %p, count %u.\n", iface, start_idx, src_data, count);
1681
1682 end_idx = start_idx + count;
1683 if (end_idx > device->d3d_vshader_constantF)
1684 {
1685 WARN("end_idx %u > float constants limit %u.\n", end_idx, device->d3d_vshader_constantF);
1686 end_idx = device->d3d_vshader_constantF;
1687 }
1688
1689 for (i = start_idx; i < end_idx; ++i)
1690 {
1691 local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
1692 if (!lconst) return E_OUTOFMEMORY;
1693
1694 lconst->idx = i;
1695 memcpy(lconst->value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
1696 list_add_head(&shader->baseShader.constantsF, &lconst->entry);
1697 }
1698
1699 return WINED3D_OK;
1700}
1701
1702static const IWineD3DVertexShaderVtbl IWineD3DVertexShader_Vtbl =
1703{
1704 /* IUnknown methods */
1705 vertexshader_QueryInterface,
1706 vertexshader_AddRef,
1707 vertexshader_Release,
1708 /* IWineD3DBase methods */
1709 vertexshader_GetParent,
1710 /* IWineD3DBaseShader methods */
1711 vertexshader_GetFunction,
1712 /* IWineD3DVertexShader methods */
1713 vertexshader_SetLocalConstantsF,
1714};
1715
1716void find_vs_compile_args(IWineD3DVertexShaderImpl *shader,
1717 IWineD3DStateBlockImpl *stateblock, struct vs_compile_args *args)
1718{
1719 args->fog_src = stateblock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
1720 args->clip_enabled = stateblock->renderState[WINED3DRS_CLIPPING]
1721 && stateblock->renderState[WINED3DRS_CLIPPLANEENABLE];
1722 args->swizzle_map = ((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.swizzle_map;
1723}
1724
1725static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
1726{
1727 if (usage_idx1 != usage_idx2) return FALSE;
1728 if (usage1 == usage2) return TRUE;
1729 if (usage1 == WINED3DDECLUSAGE_POSITION && usage2 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1730 if (usage2 == WINED3DDECLUSAGE_POSITION && usage1 == WINED3DDECLUSAGE_POSITIONT) return TRUE;
1731
1732 return FALSE;
1733}
1734
1735BOOL vshader_get_input(IWineD3DVertexShader *iface, BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
1736{
1737 IWineD3DVertexShaderImpl *shader = (IWineD3DVertexShaderImpl *)iface;
1738 WORD map = shader->baseShader.reg_maps.input_registers;
1739 unsigned int i;
1740
1741 for (i = 0; map; map >>= 1, ++i)
1742 {
1743 if (!(map & 1)) continue;
1744
1745 if (match_usage(shader->attributes[i].usage,
1746 shader->attributes[i].usage_idx, usage_req, usage_idx_req))
1747 {
1748 *regnum = i;
1749 return TRUE;
1750 }
1751 }
1752 return FALSE;
1753}
1754
1755static void vertexshader_set_limits(IWineD3DVertexShaderImpl *shader)
1756{
1757 DWORD shader_version = WINED3D_SHADER_VERSION(shader->baseShader.reg_maps.shader_version.major,
1758 shader->baseShader.reg_maps.shader_version.minor);
1759 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)shader->baseShader.device;
1760
1761 shader->baseShader.limits.texcoord = 0;
1762 shader->baseShader.limits.attributes = 16;
1763 shader->baseShader.limits.packed_input = 0;
1764
1765 switch (shader_version)
1766 {
1767 case WINED3D_SHADER_VERSION(1, 0):
1768 case WINED3D_SHADER_VERSION(1, 1):
1769 shader->baseShader.limits.temporary = 12;
1770 shader->baseShader.limits.constant_bool = 0;
1771 shader->baseShader.limits.constant_int = 0;
1772 shader->baseShader.limits.address = 1;
1773 shader->baseShader.limits.packed_output = 0;
1774 shader->baseShader.limits.sampler = 0;
1775 shader->baseShader.limits.label = 0;
1776 /* TODO: vs_1_1 has a minimum of 96 constants. What happens when
1777 * a vs_1_1 shader is used on a vs_3_0 capable card that has 256
1778 * constants? */
1779 shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1780 break;
1781
1782 case WINED3D_SHADER_VERSION(2, 0):
1783 case WINED3D_SHADER_VERSION(2, 1):
1784 shader->baseShader.limits.temporary = 12;
1785 shader->baseShader.limits.constant_bool = 16;
1786 shader->baseShader.limits.constant_int = 16;
1787 shader->baseShader.limits.address = 1;
1788 shader->baseShader.limits.packed_output = 0;
1789 shader->baseShader.limits.sampler = 0;
1790 shader->baseShader.limits.label = 16;
1791 shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1792 break;
1793
1794 case WINED3D_SHADER_VERSION(4, 0):
1795 FIXME("Using 3.0 limits for 4.0 shader.\n");
1796 RT_FALL_THRU();
1797
1798 case WINED3D_SHADER_VERSION(3, 0):
1799 shader->baseShader.limits.temporary = 32;
1800 shader->baseShader.limits.constant_bool = 32;
1801 shader->baseShader.limits.constant_int = 32;
1802 shader->baseShader.limits.address = 1;
1803 shader->baseShader.limits.packed_output = 12;
1804 shader->baseShader.limits.sampler = 4;
1805 shader->baseShader.limits.label = 16; /* FIXME: 2048 */
1806 /* DX10 cards on Windows advertise a d3d9 constant limit of 256
1807 * even though they are capable of supporting much more (GL
1808 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
1809 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
1810 * shaders to 256. */
1811 shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1812 break;
1813
1814 default:
1815 shader->baseShader.limits.temporary = 12;
1816 shader->baseShader.limits.constant_bool = 16;
1817 shader->baseShader.limits.constant_int = 16;
1818 shader->baseShader.limits.address = 1;
1819 shader->baseShader.limits.packed_output = 0;
1820 shader->baseShader.limits.sampler = 0;
1821 shader->baseShader.limits.label = 16;
1822 shader->baseShader.limits.constant_float = min(256, device->d3d_vshader_constantF);
1823 FIXME("Unrecognized vertex shader version \"%u.%u\".\n",
1824 shader->baseShader.reg_maps.shader_version.major,
1825 shader->baseShader.reg_maps.shader_version.minor);
1826 }
1827}
1828
1829HRESULT vertexshader_init(IWineD3DVertexShaderImpl *shader, IWineD3DDeviceImpl *device,
1830 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1831 IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
1832{
1833 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
1834 struct shader_reg_maps *reg_maps = &shader->baseShader.reg_maps;
1835 unsigned int i;
1836 HRESULT hr;
1837 WORD map;
1838
1839 if (!byte_code) return WINED3DERR_INVALIDCALL;
1840
1841 shader->lpVtbl = &IWineD3DVertexShader_Vtbl;
1842 shader_init(&shader->baseShader, device, parent, parent_ops);
1843
1844 hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code,
1845 output_signature, device->d3d_vshader_constantF);
1846 if (FAILED(hr))
1847 {
1848 WARN("Failed to set function, hr %#x.\n", hr);
1849 shader_cleanup((IWineD3DBaseShader *)shader);
1850 return hr;
1851 }
1852
1853 map = shader->baseShader.reg_maps.input_registers;
1854 for (i = 0; map; map >>= 1, ++i)
1855 {
1856 if (!(map & 1) || !shader->baseShader.input_signature[i].semantic_name) continue;
1857
1858 shader->attributes[i].usage =
1859 shader_usage_from_semantic_name(shader->baseShader.input_signature[i].semantic_name);
1860 shader->attributes[i].usage_idx = shader->baseShader.input_signature[i].semantic_idx;
1861 }
1862
1863 if (output_signature)
1864 {
1865 for (i = 0; i < output_signature->element_count; ++i)
1866 {
1867 struct wined3d_shader_signature_element *e = &output_signature->elements[i];
1868 reg_maps->output_registers |= 1 << e->register_idx;
1869 shader->baseShader.output_signature[e->register_idx] = *e;
1870 }
1871 }
1872
1873 vertexshader_set_limits(shader);
1874
1875 if (device->vs_selected_mode == SHADER_ARB
1876 && (gl_info->quirks & WINED3D_QUIRK_ARB_VS_OFFSET_LIMIT)
1877 && shader->min_rel_offset <= shader->max_rel_offset)
1878 {
1879 if (shader->max_rel_offset - shader->min_rel_offset > 127)
1880 {
1881 FIXME("The difference between the minimum and maximum relative offset is > 127.\n");
1882 FIXME("Which this OpenGL implementation does not support. Try using GLSL.\n");
1883 FIXME("Min: %d, Max: %d.\n", shader->min_rel_offset, shader->max_rel_offset);
1884 }
1885 else if (shader->max_rel_offset - shader->min_rel_offset > 63)
1886 {
1887 shader->rel_offset = shader->min_rel_offset + 63;
1888 }
1889 else if (shader->max_rel_offset > 63)
1890 {
1891 shader->rel_offset = shader->min_rel_offset;
1892 }
1893 else
1894 {
1895 shader->rel_offset = 0;
1896 }
1897 }
1898
1899 shader->baseShader.load_local_constsF = shader->baseShader.reg_maps.usesrelconstF
1900 && !list_empty(&shader->baseShader.constantsF);
1901
1902 return WINED3D_OK;
1903}
1904
1905static HRESULT STDMETHODCALLTYPE geometryshader_QueryInterface(IWineD3DGeometryShader *iface,
1906 REFIID riid, void **object)
1907{
1908 TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
1909
1910 if (IsEqualGUID(riid, &IID_IWineD3DGeometryShader)
1911 || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1912 || IsEqualGUID(riid, &IID_IWineD3DBase)
1913 || IsEqualGUID(riid, &IID_IUnknown))
1914 {
1915 IUnknown_AddRef(iface);
1916 *object = iface;
1917 return S_OK;
1918 }
1919
1920 WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
1921
1922 *object = NULL;
1923 return E_NOINTERFACE;
1924}
1925
1926static ULONG STDMETHODCALLTYPE geometryshader_AddRef(IWineD3DGeometryShader *iface)
1927{
1928 struct wined3d_geometryshader *shader = (struct wined3d_geometryshader *)iface;
1929 ULONG refcount = InterlockedIncrement(&shader->base_shader.ref);
1930
1931 TRACE("%p increasing refcount to %u.\n", shader, refcount);
1932
1933 return refcount;
1934}
1935
1936static ULONG STDMETHODCALLTYPE geometryshader_Release(IWineD3DGeometryShader *iface)
1937{
1938 struct wined3d_geometryshader *shader = (struct wined3d_geometryshader *)iface;
1939 ULONG refcount = InterlockedDecrement(&shader->base_shader.ref);
1940
1941 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
1942
1943 if (!refcount)
1944 {
1945 shader_cleanup((IWineD3DBaseShader *)iface);
1946 shader->base_shader.parent_ops->wined3d_object_destroyed(shader->base_shader.parent);
1947 HeapFree(GetProcessHeap(), 0, shader);
1948 }
1949
1950 return refcount;
1951}
1952
1953static HRESULT STDMETHODCALLTYPE geometryshader_GetParent(IWineD3DGeometryShader *iface, IUnknown **parent)
1954{
1955 TRACE("iface %p, parent %p.\n", iface, parent);
1956
1957 shader_get_parent((IWineD3DBaseShaderImpl *)iface, parent);
1958
1959 return WINED3D_OK;
1960}
1961
1962static HRESULT STDMETHODCALLTYPE geometryshader_GetFunction(IWineD3DGeometryShader *iface, void *data, UINT *data_size)
1963{
1964 TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
1965
1966 return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
1967}
1968
1969static const IWineD3DGeometryShaderVtbl wined3d_geometryshader_vtbl =
1970{
1971 /* IUnknown methods */
1972 geometryshader_QueryInterface,
1973 geometryshader_AddRef,
1974 geometryshader_Release,
1975 /* IWineD3DBase methods */
1976 geometryshader_GetParent,
1977 /* IWineD3DBaseShader methods */
1978 geometryshader_GetFunction,
1979};
1980
1981HRESULT geometryshader_init(struct wined3d_geometryshader *shader, IWineD3DDeviceImpl *device,
1982 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
1983 IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
1984{
1985 HRESULT hr;
1986
1987 shader->vtbl = &wined3d_geometryshader_vtbl;
1988 shader_init(&shader->base_shader, device, parent, parent_ops);
1989
1990 hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code, output_signature, 0);
1991 if (FAILED(hr))
1992 {
1993 WARN("Failed to set function, hr %#x.\n", hr);
1994 shader_cleanup((IWineD3DBaseShader *)shader);
1995 return hr;
1996 }
1997
1998 shader->base_shader.load_local_constsF = FALSE;
1999
2000 return WINED3D_OK;
2001}
2002
2003static HRESULT STDMETHODCALLTYPE pixelshader_QueryInterface(IWineD3DPixelShader *iface, REFIID riid, void **object)
2004{
2005 TRACE("iface %p, riid %s, object %p.\n", iface, debugstr_guid(riid), object);
2006
2007 if (IsEqualGUID(riid, &IID_IWineD3DPixelShader)
2008 || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
2009 || IsEqualGUID(riid, &IID_IWineD3DBase)
2010 || IsEqualGUID(riid, &IID_IUnknown))
2011 {
2012 IUnknown_AddRef(iface);
2013 *object = iface;
2014 return S_OK;
2015 }
2016
2017 WARN("%s not implemented, returning E_NOINTERFACE.\n", debugstr_guid(riid));
2018
2019 *object = NULL;
2020 return E_NOINTERFACE;
2021}
2022
2023static ULONG STDMETHODCALLTYPE pixelshader_AddRef(IWineD3DPixelShader *iface)
2024{
2025 IWineD3DPixelShaderImpl *shader = (IWineD3DPixelShaderImpl *)iface;
2026 ULONG refcount = InterlockedIncrement(&shader->baseShader.ref);
2027
2028 TRACE("%p increasing refcount to %u.\n", shader, refcount);
2029
2030 return refcount;
2031}
2032
2033static ULONG STDMETHODCALLTYPE pixelshader_Release(IWineD3DPixelShader *iface)
2034{
2035 IWineD3DPixelShaderImpl *shader = (IWineD3DPixelShaderImpl *)iface;
2036 ULONG refcount = InterlockedDecrement(&shader->baseShader.ref);
2037
2038 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
2039
2040 if (!refcount)
2041 {
2042#ifdef VBOX_WINE_WITH_SHADER_CACHE
2043 shader_chache_put(&((IWineD3DDeviceImpl*)shader->baseShader.device)->pshaderCache, (IWineD3DBaseShaderImpl*)shader);
2044#else
2045 shader_cleanup((IWineD3DBaseShader *)iface);
2046#ifndef VBOX_WITH_VMSVGA
2047 shader->baseShader.parent_ops->wined3d_object_destroyed(shader->baseShader.parent);
2048#endif
2049 HeapFree(GetProcessHeap(), 0, shader);
2050#endif
2051 }
2052
2053 return refcount;
2054}
2055
2056static HRESULT STDMETHODCALLTYPE pixelshader_GetParent(IWineD3DPixelShader *iface, IUnknown **parent)
2057{
2058 TRACE("iface %p, parent %p.\n", iface, parent);
2059
2060 shader_get_parent((IWineD3DBaseShaderImpl *)iface, parent);
2061
2062 return WINED3D_OK;
2063}
2064
2065static HRESULT STDMETHODCALLTYPE pixelshader_GetFunction(IWineD3DPixelShader *iface, void *data, UINT *data_size)
2066{
2067 TRACE("iface %p, data %p, data_size %p.\n", iface, data, data_size);
2068
2069 return shader_get_function((IWineD3DBaseShaderImpl *)iface, data, data_size);
2070}
2071
2072static const IWineD3DPixelShaderVtbl IWineD3DPixelShader_Vtbl =
2073{
2074 /* IUnknown methods */
2075 pixelshader_QueryInterface,
2076 pixelshader_AddRef,
2077 pixelshader_Release,
2078 /* IWineD3DBase methods */
2079 pixelshader_GetParent,
2080 /* IWineD3DBaseShader methods */
2081 pixelshader_GetFunction
2082};
2083
2084void find_ps_compile_args(IWineD3DPixelShaderImpl *shader,
2085 IWineD3DStateBlockImpl *stateblock, struct ps_compile_args *args)
2086{
2087 IWineD3DBaseTextureImpl *texture;
2088 UINT i;
2089
2090 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
2091 args->srgb_correction = stateblock->renderState[WINED3DRS_SRGBWRITEENABLE] ? 1 : 0;
2092 args->np2_fixup = 0;
2093 args->t_mirror = 0;
2094
2095 if (shader->baseShader.reg_maps.shader_version.major==1
2096 && shader->baseShader.reg_maps.shader_version.minor<=3)
2097 {
2098 for (i=0; i<4; ++i)
2099 {
2100 DWORD flags = stateblock->textureState[i][WINED3DTSS_TEXTURETRANSFORMFLAGS];
2101 DWORD tex_transform = flags & ~WINED3DTTFF_PROJECTED;
2102 if (flags & WINED3DTTFF_PROJECTED)
2103 {
2104 tex_transform |= WINED3D_PSARGS_PROJECTED;
2105 }
2106 args->tex_transform |= tex_transform << (i*WINED3D_PSARGS_TEXTRANSFORM_SHIFT);
2107 }
2108 }
2109
2110 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
2111 {
2112 if (!shader->baseShader.reg_maps.sampler_type[i]) continue;
2113 texture = (IWineD3DBaseTextureImpl *)stateblock->textures[i];
2114 if (!texture)
2115 {
2116 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2117 continue;
2118 }
2119 args->color_fixup[i] = texture->resource.format_desc->color_fixup;
2120
2121 /* Flag samplers that need NP2 texcoord fixup. */
2122 if (!texture->baseTexture.pow2Matrix_identity)
2123 {
2124 args->np2_fixup |= (1 << i);
2125 }
2126
2127 if (texture->baseTexture.t_mirror)
2128 {
2129 args->t_mirror |= (1 << i);
2130 }
2131 }
2132 if (shader->baseShader.reg_maps.shader_version.major >= 3)
2133 {
2134 if (((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.position_transformed)
2135 {
2136 args->vp_mode = pretransformed;
2137 }
2138 else if (use_vs(stateblock))
2139 {
2140 args->vp_mode = vertexshader;
2141 }
2142 else
2143 {
2144 args->vp_mode = fixedfunction;
2145 }
2146 args->fog = FOG_OFF;
2147 }
2148 else
2149 {
2150 args->vp_mode = vertexshader;
2151 if (stateblock->renderState[WINED3DRS_FOGENABLE])
2152 {
2153 switch (stateblock->renderState[WINED3DRS_FOGTABLEMODE])
2154 {
2155 case WINED3DFOG_NONE:
2156 if (((IWineD3DDeviceImpl *)shader->baseShader.device)->strided_streams.position_transformed
2157 || use_vs(stateblock))
2158 {
2159 args->fog = FOG_LINEAR;
2160 break;
2161 }
2162
2163 switch (stateblock->renderState[WINED3DRS_FOGVERTEXMODE])
2164 {
2165 case WINED3DFOG_NONE: RT_FALL_THRU();
2166 case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
2167 case WINED3DFOG_EXP: args->fog = FOG_EXP; break;
2168 case WINED3DFOG_EXP2: args->fog = FOG_EXP2; break;
2169 }
2170 break;
2171
2172 case WINED3DFOG_LINEAR: args->fog = FOG_LINEAR; break;
2173 case WINED3DFOG_EXP: args->fog = FOG_EXP; break;
2174 case WINED3DFOG_EXP2: args->fog = FOG_EXP2; break;
2175 }
2176 }
2177 else
2178 {
2179 args->fog = FOG_OFF;
2180 }
2181 }
2182}
2183
2184static void pixelshader_set_limits(IWineD3DPixelShaderImpl *shader)
2185{
2186 DWORD shader_version = WINED3D_SHADER_VERSION(shader->baseShader.reg_maps.shader_version.major,
2187 shader->baseShader.reg_maps.shader_version.minor);
2188
2189 shader->baseShader.limits.attributes = 0;
2190 shader->baseShader.limits.address = 0;
2191 shader->baseShader.limits.packed_output = 0;
2192
2193 switch (shader_version)
2194 {
2195 case WINED3D_SHADER_VERSION(1, 0):
2196 case WINED3D_SHADER_VERSION(1, 1):
2197 case WINED3D_SHADER_VERSION(1, 2):
2198 case WINED3D_SHADER_VERSION(1, 3):
2199 shader->baseShader.limits.temporary = 2;
2200 shader->baseShader.limits.constant_float = 8;
2201 shader->baseShader.limits.constant_int = 0;
2202 shader->baseShader.limits.constant_bool = 0;
2203 shader->baseShader.limits.texcoord = 4;
2204 shader->baseShader.limits.sampler = 4;
2205 shader->baseShader.limits.packed_input = 0;
2206 shader->baseShader.limits.label = 0;
2207 break;
2208
2209 case WINED3D_SHADER_VERSION(1, 4):
2210 shader->baseShader.limits.temporary = 6;
2211 shader->baseShader.limits.constant_float = 8;
2212 shader->baseShader.limits.constant_int = 0;
2213 shader->baseShader.limits.constant_bool = 0;
2214 shader->baseShader.limits.texcoord = 6;
2215 shader->baseShader.limits.sampler = 6;
2216 shader->baseShader.limits.packed_input = 0;
2217 shader->baseShader.limits.label = 0;
2218 break;
2219
2220 /* FIXME: Temporaries must match D3DPSHADERCAPS2_0.NumTemps. */
2221 case WINED3D_SHADER_VERSION(2, 0):
2222 shader->baseShader.limits.temporary = 32;
2223 shader->baseShader.limits.constant_float = 32;
2224 shader->baseShader.limits.constant_int = 16;
2225 shader->baseShader.limits.constant_bool = 16;
2226 shader->baseShader.limits.texcoord = 8;
2227 shader->baseShader.limits.sampler = 16;
2228 shader->baseShader.limits.packed_input = 0;
2229 break;
2230
2231 case WINED3D_SHADER_VERSION(2, 1):
2232 shader->baseShader.limits.temporary = 32;
2233 shader->baseShader.limits.constant_float = 32;
2234 shader->baseShader.limits.constant_int = 16;
2235 shader->baseShader.limits.constant_bool = 16;
2236 shader->baseShader.limits.texcoord = 8;
2237 shader->baseShader.limits.sampler = 16;
2238 shader->baseShader.limits.packed_input = 0;
2239 shader->baseShader.limits.label = 16;
2240 break;
2241
2242 case WINED3D_SHADER_VERSION(4, 0):
2243 FIXME("Using 3.0 limits for 4.0 shader.\n");
2244 RT_FALL_THRU();
2245
2246 case WINED3D_SHADER_VERSION(3, 0):
2247 shader->baseShader.limits.temporary = 32;
2248 shader->baseShader.limits.constant_float = 224;
2249 shader->baseShader.limits.constant_int = 16;
2250 shader->baseShader.limits.constant_bool = 16;
2251 shader->baseShader.limits.texcoord = 0;
2252 shader->baseShader.limits.sampler = 16;
2253 shader->baseShader.limits.packed_input = 12;
2254 shader->baseShader.limits.label = 16; /* FIXME: 2048 */
2255 break;
2256
2257 default:
2258 shader->baseShader.limits.temporary = 32;
2259 shader->baseShader.limits.constant_float = 32;
2260 shader->baseShader.limits.constant_int = 16;
2261 shader->baseShader.limits.constant_bool = 16;
2262 shader->baseShader.limits.texcoord = 8;
2263 shader->baseShader.limits.sampler = 16;
2264 shader->baseShader.limits.packed_input = 0;
2265 shader->baseShader.limits.label = 0;
2266 FIXME("Unrecognized pixel shader version %u.%u\n",
2267 shader->baseShader.reg_maps.shader_version.major,
2268 shader->baseShader.reg_maps.shader_version.minor);
2269 }
2270}
2271
2272HRESULT pixelshader_init(IWineD3DPixelShaderImpl *shader, IWineD3DDeviceImpl *device,
2273 const DWORD *byte_code, const struct wined3d_shader_signature *output_signature,
2274 IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
2275{
2276 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
2277 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
2278 HRESULT hr;
2279
2280 if (!byte_code) return WINED3DERR_INVALIDCALL;
2281
2282 shader->lpVtbl = &IWineD3DPixelShader_Vtbl;
2283 shader_init(&shader->baseShader, device, parent, parent_ops);
2284
2285 hr = shader_set_function((IWineD3DBaseShaderImpl *)shader, byte_code,
2286 output_signature, device->d3d_pshader_constantF);
2287 if (FAILED(hr))
2288 {
2289 WARN("Failed to set function, hr %#x.\n", hr);
2290 shader_cleanup((IWineD3DBaseShader *)shader);
2291 return hr;
2292 }
2293
2294 pixelshader_set_limits(shader);
2295
2296 for (i = 0; i < MAX_REG_INPUT; ++i)
2297 {
2298 if (shader->input_reg_used[i])
2299 {
2300 ++num_regs_used;
2301 highest_reg_used = i;
2302 }
2303 }
2304
2305 /* Don't do any register mapping magic if it is not needed, or if we can't
2306 * achieve anything anyway */
2307 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
2308 || num_regs_used > (gl_info->limits.glsl_varyings / 4))
2309 {
2310 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
2311 {
2312 /* This happens with relative addressing. The input mapper function
2313 * warns about this if the higher registers are declared too, so
2314 * don't write a FIXME here */
2315 WARN("More varying registers used than supported\n");
2316 }
2317
2318 for (i = 0; i < MAX_REG_INPUT; ++i)
2319 {
2320 shader->input_reg_map[i] = i;
2321 }
2322
2323 shader->declared_in_count = highest_reg_used + 1;
2324 }
2325 else
2326 {
2327 shader->declared_in_count = 0;
2328 for (i = 0; i < MAX_REG_INPUT; ++i)
2329 {
2330 if (shader->input_reg_used[i]) shader->input_reg_map[i] = shader->declared_in_count++;
2331 else shader->input_reg_map[i] = ~0U;
2332 }
2333 }
2334
2335 shader->baseShader.load_local_constsF = FALSE;
2336
2337 return WINED3D_OK;
2338}
2339
2340void pixelshader_update_samplers(struct shader_reg_maps *reg_maps, IWineD3DBaseTexture * const *textures)
2341{
2342 WINED3DSAMPLER_TEXTURE_TYPE *sampler_type = reg_maps->sampler_type;
2343 unsigned int i;
2344
2345 if (reg_maps->shader_version.major != 1) return;
2346
2347 for (i = 0; i < max(MAX_FRAGMENT_SAMPLERS, MAX_VERTEX_SAMPLERS); ++i)
2348 {
2349 /* We don't sample from this sampler. */
2350 if (!sampler_type[i]) continue;
2351
2352 if (!textures[i])
2353 {
2354 WARN("No texture bound to sampler %u, using 2D.\n", i);
2355 sampler_type[i] = WINED3DSTT_2D;
2356 continue;
2357 }
2358
2359 switch (IWineD3DBaseTexture_GetTextureDimensions(textures[i]))
2360 {
2361 case GL_TEXTURE_RECTANGLE_ARB:
2362 case GL_TEXTURE_2D:
2363 /* We have to select between texture rectangles and 2D
2364 * textures later because 2.0 and 3.0 shaders only have
2365 * WINED3DSTT_2D as well. */
2366 sampler_type[i] = WINED3DSTT_2D;
2367 break;
2368
2369 case GL_TEXTURE_3D:
2370 sampler_type[i] = WINED3DSTT_VOLUME;
2371 break;
2372
2373 case GL_TEXTURE_CUBE_MAP_ARB:
2374 sampler_type[i] = WINED3DSTT_CUBE;
2375 break;
2376
2377 default:
2378 FIXME("Unrecognized texture type %#x, using 2D.\n",
2379 IWineD3DBaseTexture_GetTextureDimensions(textures[i]));
2380 sampler_type[i] = WINED3DSTT_2D;
2381 }
2382 }
2383}
2384
2385#ifdef VBOX_WINE_WITH_SHADER_CACHE
2386
2387static void shader_reinit(struct IWineD3DBaseShaderClass *shader, IWineD3DDeviceImpl *device,
2388 IUnknown *parent, const struct wined3d_parent_ops *parent_ops)
2389{
2390 shader->ref = 1;
2391 shader->device = (IWineD3DDevice *)device;
2392 shader->parent = parent;
2393 shader->parent_ops = parent_ops;
2394#ifdef DEBUG
2395 {
2396 IWineD3DBaseShaderImpl *tstShader;
2397 LIST_FOR_EACH_ENTRY(tstShader, &device->shaders, IWineD3DBaseShaderImpl, baseShader.shader_list_entry) {
2398 if (&tstShader->baseShader == shader)
2399 return;
2400 }
2401 ERR("shader not in list!");
2402 }
2403#endif
2404}
2405
2406static DECLCALLBACK(uint32_t) shader_cache_hash(void *pvKey)
2407{
2408 uint32_t u32Hash, i;
2409 DWORD *function;
2410 IWineD3DBaseShaderImpl *pShader = (IWineD3DBaseShaderImpl *)pvKey;
2411 if (pShader->baseShader.u32CacheDataInited)
2412 return pShader->baseShader.u32Hash;
2413
2414 u32Hash = pShader->baseShader.functionLength;
2415 function = pShader->baseShader.function;
2416 for (i = 0; i < pShader->baseShader.functionLength / 4; ++i)
2417 {
2418 u32Hash += *function++;
2419 }
2420
2421 for (i = 0; i < pShader->baseShader.functionLength % 4; ++i)
2422 {
2423 u32Hash += *((uint8_t*)function)++;
2424 }
2425
2426 pShader->baseShader.u32Hash = u32Hash;
2427 pShader->baseShader.u32CacheDataInited = TRUE;
2428 return pShader->baseShader.u32Hash;
2429}
2430
2431static DECLCALLBACK(bool) shader_cache_equal(void *pvKey1, void *pvKey2)
2432{
2433 IWineD3DBaseShaderImpl *pShader1 = (IWineD3DBaseShaderImpl *)pvKey1;
2434 IWineD3DBaseShaderImpl *pShader2 = (IWineD3DBaseShaderImpl *)pvKey2;
2435
2436 if (pShader1 == pShader2)
2437 return TRUE;
2438
2439 if (pShader1->baseShader.functionLength != pShader2->baseShader.functionLength)
2440 return FALSE;
2441
2442 if (shader_cache_hash(pvKey1) != shader_cache_hash(pvKey2))
2443 {
2444#ifdef DEBUG_misha
2445 Assert(memcmp(pShader1->baseShader.function, pShader2->baseShader.function, pShader1->baseShader.functionLength));
2446#endif
2447 return FALSE;
2448 }
2449
2450 return !memcmp(pShader1->baseShader.function, pShader2->baseShader.function, pShader1->baseShader.functionLength);
2451}
2452
2453#define VBOX_SHADER_FROM_CACHE_ENTRY(_pEntry) RT_FROM_MEMBER(RT_FROM_MEMBER((_pEntry), IWineD3DBaseShaderClass, CacheEntry), IWineD3DBaseShaderImpl, baseShader)
2454
2455static DECLCALLBACK(void) shader_cache_cleanup_entry(void *pvKey, struct VBOXEXT_HASHCACHE_ENTRY *pEntry)
2456{
2457 IWineD3DBaseShaderImpl *pShader = VBOX_SHADER_FROM_CACHE_ENTRY(pEntry);
2458 shader_cleanup((IWineD3DBaseShader *)pShader);
2459 if (pShader->baseShader.parent)
2460 pShader->baseShader.parent_ops->wined3d_object_destroyed(pShader->baseShader.parent);
2461 HeapFree(GetProcessHeap(), 0, pShader);
2462}
2463
2464static void shader_chache_init(PVBOXEXT_HASHCACHE pCache)
2465{
2466 VBoxExtCacheInit(pCache, 1024, shader_cache_hash, shader_cache_equal, shader_cache_cleanup_entry);
2467}
2468
2469static void shader_chache_term(PVBOXEXT_HASHCACHE pCache)
2470{
2471 VBoxExtCacheTerm(pCache);
2472}
2473
2474static void shader_chache_put(PVBOXEXT_HASHCACHE pCache, IWineD3DBaseShaderImpl *pShader)
2475{
2476 if (pShader->baseShader.parent)
2477 {
2478 pShader->baseShader.parent_ops->wined3d_object_destroyed(pShader->baseShader.parent);
2479 pShader->baseShader.parent = NULL;
2480 }
2481
2482 VBoxExtCachePut(pCache, pShader, &pShader->baseShader.CacheEntry);
2483}
2484
2485static IWineD3DBaseShaderImpl* shader_chache_get(PVBOXEXT_HASHCACHE pCache, IWineD3DBaseShaderImpl *pShader)
2486{
2487 PVBOXEXT_HASHCACHE_ENTRY pEntry = VBoxExtCacheGet(pCache, pShader);
2488 IWineD3DBaseShaderImpl *pCachedShader;
2489
2490 if (!pEntry)
2491 return pShader;
2492
2493 pCachedShader = VBOX_SHADER_FROM_CACHE_ENTRY(pEntry);
2494 shader_reinit(&pCachedShader->baseShader, (IWineD3DDeviceImpl *)pShader->baseShader.device,
2495 pShader->baseShader.parent, pShader->baseShader.parent_ops);
2496 pShader->baseShader.parent = NULL;
2497 /* we can not do a IWineD3DBaseShader_Release here since this would result in putting a shader to the cache */
2498 shader_cleanup((IWineD3DBaseShader *)pShader);
2499 HeapFree(GetProcessHeap(), 0, pShader);
2500 return pCachedShader;
2501}
2502
2503IWineD3DVertexShaderImpl * vertexshader_check_cached(IWineD3DDeviceImpl *device, IWineD3DVertexShaderImpl *object)
2504{
2505 return (IWineD3DVertexShaderImpl*)shader_chache_get(&device->vshaderCache, (IWineD3DBaseShaderImpl *)object);
2506}
2507
2508IWineD3DPixelShaderImpl * pixelshader_check_cached(IWineD3DDeviceImpl *device, IWineD3DPixelShaderImpl *object)
2509{
2510 return (IWineD3DPixelShaderImpl*)shader_chache_get(&device->pshaderCache, (IWineD3DBaseShaderImpl *)object);
2511}
2512
2513void shader_chaches_init(IWineD3DDeviceImpl *device)
2514{
2515 shader_chache_init(&device->vshaderCache);
2516 shader_chache_init(&device->pshaderCache);
2517}
2518
2519void shader_chaches_term(IWineD3DDeviceImpl *device)
2520{
2521 shader_chache_term(&device->vshaderCache);
2522 shader_chache_term(&device->pshaderCache);
2523}
2524#endif
Note: See TracBrowser for help on using the repository browser.

© 2025 Oracle Support Privacy / Do Not Sell My Info Terms of Use Trademark Policy Automated Access Etiquette