?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
push 014043c4937c940c54cd1214c96e33a3b3c8cf7d
[wine/hacks.git] / dlls / wined3d / baseshader.c
blobed292ce329f330ae7d9f80137c7b589e435a3608
1 /*
2 * shaders implementation
3 *
4 * Copyright 2002-2003 Jason Edmeades
5 * Copyright 2002-2003 Raphael Junqueira
6 * Copyright 2004 Christian Costa
7 * Copyright 2005 Oliver Stieber
8 * Copyright 2006 Ivan Gyurdiev
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 <string.h>
27 #include <stdio.h>
28 #include "wined3d_private.h"
29
30 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
31
32 #define GLNAME_REQUIRE_GLSL ((const char *)1)
33
34 static inline BOOL shader_is_version_token(DWORD token) {
35 return shader_is_pshader_version(token) ||
36 shader_is_vshader_version(token);
37 }
38
39 int shader_addline(
40 SHADER_BUFFER* buffer,
41 const char *format, ...) {
42
43 char* base = buffer->buffer + buffer->bsize;
44 int rc;
45
46 va_list args;
47 va_start(args, format);
48 rc = vsnprintf(base, SHADER_PGMSIZE - 1 - buffer->bsize, format, args);
49 va_end(args);
50
51 if (rc < 0 || /* C89 */
52 rc > SHADER_PGMSIZE - 1 - buffer->bsize) { /* C99 */
53
54 ERR("The buffer allocated for the shader program string "
55 "is too small at %d bytes.\n", SHADER_PGMSIZE);
56 buffer->bsize = SHADER_PGMSIZE - 1;
57 return -1;
58 }
59
60 if (buffer->newline) {
61 TRACE("GL HW (%u, %u) : %s", buffer->lineNo + 1, buffer->bsize, base);
62 buffer->newline = FALSE;
63 } else {
64 TRACE("%s", base);
65 }
66
67 buffer->bsize += rc;
68 if (buffer->buffer[buffer->bsize-1] == '\n') {
69 buffer->lineNo++;
70 buffer->newline = TRUE;
71 }
72 return 0;
73 }
74
75 const SHADER_OPCODE* shader_get_opcode(
76 IWineD3DBaseShader *iface, const DWORD code) {
77
78 IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl*) iface;
79
80 DWORD i = 0;
81 DWORD hex_version = This->baseShader.hex_version;
82 const SHADER_OPCODE *shader_ins = This->baseShader.shader_ins;
83
84 /** TODO: use dichotomic search */
85 while (NULL != shader_ins[i].name) {
86 if (((code & WINED3DSI_OPCODE_MASK) == shader_ins[i].opcode) &&
87 (((hex_version >= shader_ins[i].min_version) && (hex_version <= shader_ins[i].max_version)) ||
88 ((shader_ins[i].min_version == 0) && (shader_ins[i].max_version == 0)))) {
89 return &shader_ins[i];
90 }
91 ++i;
92 }
93 FIXME("Unsupported opcode %#x(%d) masked %#x, shader version %#x\n",
94 code, code, code & WINED3DSI_OPCODE_MASK, hex_version);
95 return NULL;
96 }
97
98 /* Read a parameter opcode from the input stream,
99 * and possibly a relative addressing token.
100 * Return the number of tokens read */
101 int shader_get_param(
102 IWineD3DBaseShader* iface,
103 const DWORD* pToken,
104 DWORD* param,
105 DWORD* addr_token) {
106
107 /* PS >= 3.0 have relative addressing (with token)
108 * VS >= 2.0 have relative addressing (with token)
109 * VS >= 1.0 < 2.0 have relative addressing (without token)
110 * The version check below should work in general */
111
112 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
113 char rel_token = WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 2 &&
114 ((*pToken & WINED3DSHADER_ADDRESSMODE_MASK) == WINED3DSHADER_ADDRMODE_RELATIVE);
115
116 *param = *pToken;
117 *addr_token = rel_token? *(pToken + 1): 0;
118 return rel_token? 2:1;
119 }
120
121 /* Return the number of parameters to skip for an opcode */
122 static inline int shader_skip_opcode(
123 IWineD3DBaseShaderImpl* This,
124 const SHADER_OPCODE* curOpcode,
125 DWORD opcode_token) {
126
127 /* Shaders >= 2.0 may contain address tokens, but fortunately they
128 * have a useful length mask - use it here. Shaders 1.0 contain no such tokens */
129
130 return (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 2)?
131 ((opcode_token & WINED3DSI_INSTLENGTH_MASK) >> WINED3DSI_INSTLENGTH_SHIFT):
132 curOpcode->num_params;
133 }
134
135 /* Read the parameters of an unrecognized opcode from the input stream
136 * Return the number of tokens read.
137 *
138 * Note: This function assumes source or destination token format.
139 * It will not work with specially-formatted tokens like DEF or DCL,
140 * but hopefully those would be recognized */
141
142 int shader_skip_unrecognized(
143 IWineD3DBaseShader* iface,
144 const DWORD* pToken) {
145
146 int tokens_read = 0;
147 int i = 0;
148
149 /* TODO: Think of a good name for 0x80000000 and replace it with a constant */
150 while (*pToken & 0x80000000) {
151
152 DWORD param, addr_token;
153 tokens_read += shader_get_param(iface, pToken, &param, &addr_token);
154 pToken += tokens_read;
155
156 FIXME("Unrecognized opcode param: token=%08x "
157 "addr_token=%08x name=", param, addr_token);
158 shader_dump_param(iface, param, addr_token, i);
159 FIXME("\n");
160 ++i;
161 }
162 return tokens_read;
163 }
164
165 /* Convert floating point offset relative
166 * to a register file to an absolute offset for float constants */
167
168 unsigned int shader_get_float_offset(const DWORD reg) {
169
170 unsigned int regnum = reg & WINED3DSP_REGNUM_MASK;
171 int regtype = shader_get_regtype(reg);
172
173 switch (regtype) {
174 case WINED3DSPR_CONST: return regnum;
175 case WINED3DSPR_CONST2: return 2048 + regnum;
176 case WINED3DSPR_CONST3: return 4096 + regnum;
177 case WINED3DSPR_CONST4: return 6144 + regnum;
178 default:
179 FIXME("Unsupported register type: %d\n", regtype);
180 return regnum;
181 }
182 }
183
184 static void shader_delete_constant_list(struct list* clist) {
185
186 struct list *ptr;
187 struct local_constant* constant;
188
189 ptr = list_head(clist);
190 while (ptr) {
191 constant = LIST_ENTRY(ptr, struct local_constant, entry);
192 ptr = list_next(clist, ptr);
193 HeapFree(GetProcessHeap(), 0, constant);
194 }
195 list_init(clist);
196 }
197
198 /* Note that this does not count the loop register
199 * as an address register. */
200
201 HRESULT shader_get_registers_used(
202 IWineD3DBaseShader *iface,
203 shader_reg_maps* reg_maps,
204 semantic* semantics_in,
205 semantic* semantics_out,
206 CONST DWORD* pToken,
207 IWineD3DStateBlockImpl *stateBlock) {
208
209 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
210 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
211
212 /* There are some minor differences between pixel and vertex shaders */
213 char pshader = shader_is_pshader_version(This->baseShader.hex_version);
214
215 reg_maps->bumpmat = -1;
216 reg_maps->luminanceparams = -1;
217
218 if (pToken == NULL)
219 return WINED3D_OK;
220
221 /* get_registers_used is called on every compile on some 1.x shaders, which can result
222 * in stacking up a collection of local constants. Delete the old constants if existing
223 */
224 shader_delete_constant_list(&This->baseShader.constantsF);
225 shader_delete_constant_list(&This->baseShader.constantsB);
226 shader_delete_constant_list(&This->baseShader.constantsI);
227
228 while (WINED3DVS_END() != *pToken) {
229 CONST SHADER_OPCODE* curOpcode;
230 DWORD opcode_token;
231
232 /* Skip version */
233 if (shader_is_version_token(*pToken)) {
234 ++pToken;
235 continue;
236
237 /* Skip comments */
238 } else if (shader_is_comment(*pToken)) {
239 DWORD comment_len = (*pToken & WINED3DSI_COMMENTSIZE_MASK) >> WINED3DSI_COMMENTSIZE_SHIFT;
240 ++pToken;
241 pToken += comment_len;
242 continue;
243 }
244
245 /* Fetch opcode */
246 opcode_token = *pToken++;
247 curOpcode = shader_get_opcode(iface, opcode_token);
248
249 /* Unhandled opcode, and its parameters */
250 if (NULL == curOpcode) {
251 while (*pToken & 0x80000000)
252 ++pToken;
253
254 /* Handle declarations */
255 } else if (WINED3DSIO_DCL == curOpcode->opcode) {
256
257 DWORD usage = *pToken++;
258 DWORD param = *pToken++;
259 DWORD regtype = shader_get_regtype(param);
260 unsigned int regnum = param & WINED3DSP_REGNUM_MASK;
261
262 /* Vshader: mark attributes used
263 Pshader: mark 3.0 input registers used, save token */
264 if (WINED3DSPR_INPUT == regtype) {
265
266 if (!pshader)
267 reg_maps->attributes[regnum] = 1;
268 else
269 reg_maps->packed_input[regnum] = 1;
270
271 semantics_in[regnum].usage = usage;
272 semantics_in[regnum].reg = param;
273
274 /* Vshader: mark 3.0 output registers used, save token */
275 } else if (WINED3DSPR_OUTPUT == regtype) {
276 reg_maps->packed_output[regnum] = 1;
277 semantics_out[regnum].usage = usage;
278 semantics_out[regnum].reg = param;
279 if (usage & (WINED3DDECLUSAGE_FOG << WINED3DSP_DCL_USAGE_SHIFT))
280 reg_maps->fog = 1;
281
282 /* Save sampler usage token */
283 } else if (WINED3DSPR_SAMPLER == regtype)
284 reg_maps->samplers[regnum] = usage;
285
286 } else if (WINED3DSIO_DEF == curOpcode->opcode) {
287
288 local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
289 if (!lconst) return E_OUTOFMEMORY;
290 lconst->idx = *pToken & WINED3DSP_REGNUM_MASK;
291 memcpy(&lconst->value, pToken + 1, 4 * sizeof(DWORD));
292
293 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
294 if(WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) == 1 && pshader) {
295 float *value = (float *) lconst->value;
296 if(value[0] < -1.0) value[0] = -1.0;
297 else if(value[0] > 1.0) value[0] = 1.0;
298 if(value[1] < -1.0) value[1] = -1.0;
299 else if(value[1] > 1.0) value[1] = 1.0;
300 if(value[2] < -1.0) value[2] = -1.0;
301 else if(value[2] > 1.0) value[2] = 1.0;
302 if(value[3] < -1.0) value[3] = -1.0;
303 else if(value[3] > 1.0) value[3] = 1.0;
304 }
305
306 list_add_head(&This->baseShader.constantsF, &lconst->entry);
307 pToken += curOpcode->num_params;
308
309 } else if (WINED3DSIO_DEFI == curOpcode->opcode) {
310
311 local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
312 if (!lconst) return E_OUTOFMEMORY;
313 lconst->idx = *pToken & WINED3DSP_REGNUM_MASK;
314 memcpy(&lconst->value, pToken + 1, 4 * sizeof(DWORD));
315 list_add_head(&This->baseShader.constantsI, &lconst->entry);
316 pToken += curOpcode->num_params;
317
318 } else if (WINED3DSIO_DEFB == curOpcode->opcode) {
319
320 local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
321 if (!lconst) return E_OUTOFMEMORY;
322 lconst->idx = *pToken & WINED3DSP_REGNUM_MASK;
323 memcpy(&lconst->value, pToken + 1, 1 * sizeof(DWORD));
324 list_add_head(&This->baseShader.constantsB, &lconst->entry);
325 pToken += curOpcode->num_params;
326
327 /* If there's a loop in the shader */
328 } else if (WINED3DSIO_LOOP == curOpcode->opcode ||
329 WINED3DSIO_REP == curOpcode->opcode) {
330 cur_loop_depth++;
331 if(cur_loop_depth > max_loop_depth)
332 max_loop_depth = cur_loop_depth;
333 pToken += curOpcode->num_params;
334
335 } else if (WINED3DSIO_ENDLOOP == curOpcode->opcode ||
336 WINED3DSIO_ENDREP == curOpcode->opcode) {
337 cur_loop_depth--;
338
339 /* For subroutine prototypes */
340 } else if (WINED3DSIO_LABEL == curOpcode->opcode) {
341
342 DWORD snum = *pToken & WINED3DSP_REGNUM_MASK;
343 reg_maps->labels[snum] = 1;
344 pToken += curOpcode->num_params;
345
346 /* Set texture, address, temporary registers */
347 } else {
348 int i, limit;
349
350 /* Declare 1.X samplers implicitly, based on the destination reg. number */
351 if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) == 1 &&
352 (WINED3DSIO_TEX == curOpcode->opcode ||
353 WINED3DSIO_TEXBEM == curOpcode->opcode ||
354 WINED3DSIO_TEXBEML == curOpcode->opcode ||
355 WINED3DSIO_TEXDP3TEX == curOpcode->opcode ||
356 WINED3DSIO_TEXM3x2TEX == curOpcode->opcode ||
357 WINED3DSIO_TEXM3x3SPEC == curOpcode->opcode ||
358 WINED3DSIO_TEXM3x3TEX == curOpcode->opcode ||
359 WINED3DSIO_TEXM3x3VSPEC == curOpcode->opcode ||
360 WINED3DSIO_TEXREG2AR == curOpcode->opcode ||
361 WINED3DSIO_TEXREG2GB == curOpcode->opcode ||
362 WINED3DSIO_TEXREG2RGB == curOpcode->opcode)) {
363
364 /* Fake sampler usage, only set reserved bit and ttype */
365 DWORD sampler_code = *pToken & WINED3DSP_REGNUM_MASK;
366
367 if(!stateBlock->textures[sampler_code]) {
368 ERR("No texture bound to sampler %d\n", sampler_code);
369 reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
370 } else {
371 int texType = IWineD3DBaseTexture_GetTextureDimensions(stateBlock->textures[sampler_code]);
372 switch(texType) {
373 /* We have to select between texture rectangles and 2D textures later because 2.0 and
374 * 3.0 shaders only have WINED3DSTT_2D as well
375 */
376 case GL_TEXTURE_RECTANGLE_ARB:
377 case GL_TEXTURE_2D:
378 reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
379 break;
380
381 case GL_TEXTURE_3D:
382 reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_VOLUME;
383 break;
384
385 case GL_TEXTURE_CUBE_MAP_ARB:
386 reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_CUBE;
387 break;
388
389 default:
390 ERR("Unexpected gl texture type found: %d\n", texType);
391 reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
392 }
393 }
394
395 /* texbem is only valid with < 1.4 pixel shaders */
396 if(WINED3DSIO_TEXBEM == curOpcode->opcode ||
397 WINED3DSIO_TEXBEML == curOpcode->opcode) {
398 if(reg_maps->bumpmat != -1 && reg_maps->bumpmat != sampler_code) {
399 FIXME("Pixel shader uses texbem instruction on more than 1 sampler\n");
400 } else {
401 reg_maps->bumpmat = sampler_code;
402 if(WINED3DSIO_TEXBEML == curOpcode->opcode) {
403 reg_maps->luminanceparams = sampler_code;
404 }
405 }
406 }
407 }
408 if(WINED3DSIO_NRM == curOpcode->opcode) {
409 reg_maps->usesnrm = 1;
410 } else if(WINED3DSIO_BEM == curOpcode->opcode) {
411 DWORD regnum = *pToken & WINED3DSP_REGNUM_MASK;
412 if(reg_maps->bumpmat != -1 && reg_maps->bumpmat != regnum) {
413 FIXME("Pixel shader uses bem or texbem instruction on more than 1 sampler\n");
414 } else {
415 reg_maps->bumpmat = regnum;
416 }
417 } else if(WINED3DSIO_DSY == curOpcode->opcode) {
418 reg_maps->usesdsy = 1;
419 }
420
421 /* This will loop over all the registers and try to
422 * make a bitmask of the ones we're interested in.
423 *
424 * Relative addressing tokens are ignored, but that's
425 * okay, since we'll catch any address registers when
426 * they are initialized (required by spec) */
427
428 limit = (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED)?
429 curOpcode->num_params + 1: curOpcode->num_params;
430
431 for (i = 0; i < limit; ++i) {
432
433 DWORD param, addr_token, reg, regtype;
434 pToken += shader_get_param(iface, pToken, &param, &addr_token);
435
436 regtype = shader_get_regtype(param);
437 reg = param & WINED3DSP_REGNUM_MASK;
438
439 if (WINED3DSPR_TEXTURE == regtype) { /* vs: WINED3DSPR_ADDR */
440
441 if (pshader)
442 reg_maps->texcoord[reg] = 1;
443 else
444 reg_maps->address[reg] = 1;
445 }
446
447 else if (WINED3DSPR_TEMP == regtype)
448 reg_maps->temporary[reg] = 1;
449
450 else if (WINED3DSPR_INPUT == regtype) {
451 if( !pshader)
452 reg_maps->attributes[reg] = 1;
453 else {
454 if(param & WINED3DSHADER_ADDRMODE_RELATIVE) {
455 /* If relative addressing is used, we must assume that all registers
456 * are used. Even if it is a construct like v3[aL], we can't assume
457 * that v0, v1 and v2 aren't read because aL can be negative
458 */
459 unsigned int i;
460 for(i = 0; i < MAX_REG_INPUT; i++) {
461 ((IWineD3DPixelShaderImpl *) This)->input_reg_used[i] = TRUE;
462 }
463 } else {
464 ((IWineD3DPixelShaderImpl *) This)->input_reg_used[reg] = TRUE;
465 }
466 }
467 }
468
469 else if (WINED3DSPR_RASTOUT == regtype && reg == 1)
470 reg_maps->fog = 1;
471
472 else if (WINED3DSPR_MISCTYPE == regtype && reg == 0 && pshader)
473 reg_maps->vpos = 1;
474
475 else if(WINED3DSPR_CONST == regtype) {
476 if(param & WINED3DSHADER_ADDRMODE_RELATIVE) {
477 if(!pshader) {
478 if(reg <= ((IWineD3DVertexShaderImpl *) This)->min_rel_offset) {
479 ((IWineD3DVertexShaderImpl *) This)->min_rel_offset = reg;
480 } else if(reg >= ((IWineD3DVertexShaderImpl *) This)->max_rel_offset) {
481 ((IWineD3DVertexShaderImpl *) This)->max_rel_offset = reg;
482 }
483 }
484 reg_maps->usesrelconstF = TRUE;
485 }
486 }
487
488 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and is used
489 * in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel shaders because TECRDOUT
490 * isn't used in them, but future register types might cause issues
491 */
492 else if(WINED3DSPR_TEXCRDOUT == regtype && i == 0 /* Only look at writes */ &&
493 !pshader && WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) < 3) {
494 reg_maps->texcoord_mask[reg] |= shader_get_writemask(param);
495 }
496 }
497 }
498 }
499 reg_maps->loop_depth = max_loop_depth;
500
501 return WINED3D_OK;
502 }
503
504 static void shader_dump_decl_usage(
505 IWineD3DBaseShaderImpl* This,
506 DWORD decl,
507 DWORD param) {
508
509 DWORD regtype = shader_get_regtype(param);
510
511 TRACE("dcl");
512
513 if (regtype == WINED3DSPR_SAMPLER) {
514 DWORD ttype = decl & WINED3DSP_TEXTURETYPE_MASK;
515
516 switch (ttype) {
517 case WINED3DSTT_2D: TRACE("_2d"); break;
518 case WINED3DSTT_CUBE: TRACE("_cube"); break;
519 case WINED3DSTT_VOLUME: TRACE("_volume"); break;
520 default: TRACE("_unknown_ttype(%08x)", ttype);
521 }
522
523 } else {
524
525 DWORD usage = decl & WINED3DSP_DCL_USAGE_MASK;
526 DWORD idx = (decl & WINED3DSP_DCL_USAGEINDEX_MASK) >> WINED3DSP_DCL_USAGEINDEX_SHIFT;
527
528 /* Pixel shaders 3.0 don't have usage semantics */
529 char pshader = shader_is_pshader_version(This->baseShader.hex_version);
530 if (pshader && This->baseShader.hex_version < WINED3DPS_VERSION(3,0))
531 return;
532 else
533 TRACE("_");
534
535 switch(usage) {
536 case WINED3DDECLUSAGE_POSITION:
537 TRACE("position%d", idx);
538 break;
539 case WINED3DDECLUSAGE_BLENDINDICES:
540 TRACE("blend");
541 break;
542 case WINED3DDECLUSAGE_BLENDWEIGHT:
543 TRACE("weight");
544 break;
545 case WINED3DDECLUSAGE_NORMAL:
546 TRACE("normal%d", idx);
547 break;
548 case WINED3DDECLUSAGE_PSIZE:
549 TRACE("psize");
550 break;
551 case WINED3DDECLUSAGE_COLOR:
552 if(idx == 0) {
553 TRACE("color");
554 } else {
555 TRACE("specular%d", (idx - 1));
556 }
557 break;
558 case WINED3DDECLUSAGE_TEXCOORD:
559 TRACE("texture%d", idx);
560 break;
561 case WINED3DDECLUSAGE_TANGENT:
562 TRACE("tangent");
563 break;
564 case WINED3DDECLUSAGE_BINORMAL:
565 TRACE("binormal");
566 break;
567 case WINED3DDECLUSAGE_TESSFACTOR:
568 TRACE("tessfactor");
569 break;
570 case WINED3DDECLUSAGE_POSITIONT:
571 TRACE("positionT%d", idx);
572 break;
573 case WINED3DDECLUSAGE_FOG:
574 TRACE("fog");
575 break;
576 case WINED3DDECLUSAGE_DEPTH:
577 TRACE("depth");
578 break;
579 case WINED3DDECLUSAGE_SAMPLE:
580 TRACE("sample");
581 break;
582 default:
583 FIXME("unknown_semantics(%08x)", usage);
584 }
585 }
586 }
587
588 static void shader_dump_arr_entry(
589 IWineD3DBaseShader *iface,
590 const DWORD param,
591 const DWORD addr_token,
592 unsigned int reg,
593 int input) {
594
595 char relative =
596 ((param & WINED3DSHADER_ADDRESSMODE_MASK) == WINED3DSHADER_ADDRMODE_RELATIVE);
597
598 if (relative) {
599 TRACE("[");
600 if (addr_token)
601 shader_dump_param(iface, addr_token, 0, input);
602 else
603 TRACE("a0.x");
604 TRACE(" + ");
605 }
606 TRACE("%u", reg);
607 if (relative)
608 TRACE("]");
609 }
610
611 void shader_dump_param(
612 IWineD3DBaseShader *iface,
613 const DWORD param,
614 const DWORD addr_token,
615 int input) {
616
617 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
618 static const char * const rastout_reg_names[] = { "oPos", "oFog", "oPts" };
619 static const char * const misctype_reg_names[] = { "vPos", "vFace"};
620 char swizzle_reg_chars[4];
621
622 DWORD reg = param & WINED3DSP_REGNUM_MASK;
623 DWORD regtype = shader_get_regtype(param);
624 DWORD modifier = param & WINED3DSP_SRCMOD_MASK;
625
626 /* There are some minor differences between pixel and vertex shaders */
627 char pshader = shader_is_pshader_version(This->baseShader.hex_version);
628
629 /* For one, we'd prefer color components to be shown for pshaders.
630 * FIXME: use the swizzle function for this */
631
632 swizzle_reg_chars[0] = pshader? 'r': 'x';
633 swizzle_reg_chars[1] = pshader? 'g': 'y';
634 swizzle_reg_chars[2] = pshader? 'b': 'z';
635 swizzle_reg_chars[3] = pshader? 'a': 'w';
636
637 if (input) {
638 if ( (modifier == WINED3DSPSM_NEG) ||
639 (modifier == WINED3DSPSM_BIASNEG) ||
640 (modifier == WINED3DSPSM_SIGNNEG) ||
641 (modifier == WINED3DSPSM_X2NEG) ||
642 (modifier == WINED3DSPSM_ABSNEG) )
643 TRACE("-");
644 else if (modifier == WINED3DSPSM_COMP)
645 TRACE("1-");
646 else if (modifier == WINED3DSPSM_NOT)
647 TRACE("!");
648
649 if (modifier == WINED3DSPSM_ABS || modifier == WINED3DSPSM_ABSNEG)
650 TRACE("abs(");
651 }
652
653 switch (regtype) {
654 case WINED3DSPR_TEMP:
655 TRACE("r%u", reg);
656 break;
657 case WINED3DSPR_INPUT:
658 TRACE("v");
659 shader_dump_arr_entry(iface, param, addr_token, reg, input);
660 break;
661 case WINED3DSPR_CONST:
662 case WINED3DSPR_CONST2:
663 case WINED3DSPR_CONST3:
664 case WINED3DSPR_CONST4:
665 TRACE("c");
666 shader_dump_arr_entry(iface, param, addr_token, shader_get_float_offset(param), input);
667 break;
668 case WINED3DSPR_TEXTURE: /* vs: case D3DSPR_ADDR */
669 TRACE("%c%u", (pshader? 't':'a'), reg);
670 break;
671 case WINED3DSPR_RASTOUT:
672 TRACE("%s", rastout_reg_names[reg]);
673 break;
674 case WINED3DSPR_COLOROUT:
675 TRACE("oC%u", reg);
676 break;
677 case WINED3DSPR_DEPTHOUT:
678 TRACE("oDepth");
679 break;
680 case WINED3DSPR_ATTROUT:
681 TRACE("oD%u", reg);
682 break;
683 case WINED3DSPR_TEXCRDOUT:
684
685 /* Vertex shaders >= 3.0 use general purpose output registers
686 * (WINED3DSPR_OUTPUT), which can include an address token */
687
688 if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3) {
689 TRACE("o");
690 shader_dump_arr_entry(iface, param, addr_token, reg, input);
691 }
692 else
693 TRACE("oT%u", reg);
694 break;
695 case WINED3DSPR_CONSTINT:
696 TRACE("i");
697 shader_dump_arr_entry(iface, param, addr_token, reg, input);
698 break;
699 case WINED3DSPR_CONSTBOOL:
700 TRACE("b");
701 shader_dump_arr_entry(iface, param, addr_token, reg, input);
702 break;
703 case WINED3DSPR_LABEL:
704 TRACE("l%u", reg);
705 break;
706 case WINED3DSPR_LOOP:
707 TRACE("aL");
708 break;
709 case WINED3DSPR_SAMPLER:
710 TRACE("s%u", reg);
711 break;
712 case WINED3DSPR_MISCTYPE:
713 if (reg > 1) {
714 FIXME("Unhandled misctype register %d\n", reg);
715 } else {
716 TRACE("%s", misctype_reg_names[reg]);
717 }
718 break;
719 case WINED3DSPR_PREDICATE:
720 TRACE("p%u", reg);
721 break;
722 default:
723 TRACE("unhandled_rtype(%#x)", regtype);
724 break;
725 }
726
727 if (!input) {
728 /* operand output (for modifiers and shift, see dump_ins_modifiers) */
729
730 if ((param & WINED3DSP_WRITEMASK_ALL) != WINED3DSP_WRITEMASK_ALL) {
731 TRACE(".");
732 if (param & WINED3DSP_WRITEMASK_0) TRACE("%c", swizzle_reg_chars[0]);
733 if (param & WINED3DSP_WRITEMASK_1) TRACE("%c", swizzle_reg_chars[1]);
734 if (param & WINED3DSP_WRITEMASK_2) TRACE("%c", swizzle_reg_chars[2]);
735 if (param & WINED3DSP_WRITEMASK_3) TRACE("%c", swizzle_reg_chars[3]);
736 }
737
738 } else {
739 /** operand input */
740 DWORD swizzle = (param & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
741 DWORD swizzle_r = swizzle & 0x03;
742 DWORD swizzle_g = (swizzle >> 2) & 0x03;
743 DWORD swizzle_b = (swizzle >> 4) & 0x03;
744 DWORD swizzle_a = (swizzle >> 6) & 0x03;
745
746 if (0 != modifier) {
747 switch (modifier) {
748 case WINED3DSPSM_NONE: break;
749 case WINED3DSPSM_NEG: break;
750 case WINED3DSPSM_NOT: break;
751 case WINED3DSPSM_BIAS: TRACE("_bias"); break;
752 case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
753 case WINED3DSPSM_SIGN: TRACE("_bx2"); break;
754 case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
755 case WINED3DSPSM_COMP: break;
756 case WINED3DSPSM_X2: TRACE("_x2"); break;
757 case WINED3DSPSM_X2NEG: TRACE("_x2"); break;
758 case WINED3DSPSM_DZ: TRACE("_dz"); break;
759 case WINED3DSPSM_DW: TRACE("_dw"); break;
760 case WINED3DSPSM_ABSNEG: TRACE(")"); break;
761 case WINED3DSPSM_ABS: TRACE(")"); break;
762 default:
763 TRACE("_unknown_modifier(%#x)", modifier >> WINED3DSP_SRCMOD_SHIFT);
764 }
765 }
766
767 /**
768 * swizzle bits fields:
769 * RRGGBBAA
770 */
771 if ((WINED3DVS_NOSWIZZLE >> WINED3DVS_SWIZZLE_SHIFT) != swizzle) {
772 if (swizzle_r == swizzle_g &&
773 swizzle_r == swizzle_b &&
774 swizzle_r == swizzle_a) {
775 TRACE(".%c", swizzle_reg_chars[swizzle_r]);
776 } else {
777 TRACE(".%c%c%c%c",
778 swizzle_reg_chars[swizzle_r],
779 swizzle_reg_chars[swizzle_g],
780 swizzle_reg_chars[swizzle_b],
781 swizzle_reg_chars[swizzle_a]);
782 }
783 }
784 }
785 }
786
787 /** Shared code in order to generate the bulk of the shader string.
788 Use the shader_header_fct & shader_footer_fct to add strings
789 that are specific to pixel or vertex functions
790 NOTE: A description of how to parse tokens can be found on msdn */
791 void shader_generate_main(
792 IWineD3DBaseShader *iface,
793 SHADER_BUFFER* buffer,
794 shader_reg_maps* reg_maps,
795 CONST DWORD* pFunction) {
796
797 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
798 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device; /* To access shader backend callbacks */
799 const DWORD *pToken = pFunction;
800 const SHADER_OPCODE *curOpcode = NULL;
801 SHADER_HANDLER hw_fct = NULL;
802 DWORD i;
803 SHADER_OPCODE_ARG hw_arg;
804
805 /* Initialize current parsing state */
806 hw_arg.shader = iface;
807 hw_arg.buffer = buffer;
808 hw_arg.reg_maps = reg_maps;
809 This->baseShader.parse_state.current_row = 0;
810
811 /* Second pass, process opcodes */
812 if (NULL != pToken) {
813 while (WINED3DPS_END() != *pToken) {
814
815 /* Skip version token */
816 if (shader_is_version_token(*pToken)) {
817 ++pToken;
818 continue;
819 }
820
821 /* Skip comment tokens */
822 if (shader_is_comment(*pToken)) {
823 DWORD comment_len = (*pToken & WINED3DSI_COMMENTSIZE_MASK) >> WINED3DSI_COMMENTSIZE_SHIFT;
824 ++pToken;
825 TRACE("#%s\n", (const char*)pToken);
826 pToken += comment_len;
827 continue;
828 }
829
830 /* Read opcode */
831 hw_arg.opcode_token = *pToken++;
832 curOpcode = shader_get_opcode(iface, hw_arg.opcode_token);
833
834 /* Select handler */
835 if (curOpcode == NULL)
836 hw_fct = NULL;
837 else if (This->baseShader.shader_mode == SHADER_GLSL)
838 hw_fct = curOpcode->hw_glsl_fct;
839 else if (This->baseShader.shader_mode == SHADER_ARB)
840 hw_fct = curOpcode->hw_fct;
841
842 /* Unknown opcode and its parameters */
843 if (NULL == curOpcode) {
844 FIXME("Unrecognized opcode: token=%08x\n", hw_arg.opcode_token);
845 pToken += shader_skip_unrecognized(iface, pToken);
846
847 /* Nothing to do */
848 } else if (WINED3DSIO_DCL == curOpcode->opcode ||
849 WINED3DSIO_NOP == curOpcode->opcode ||
850 WINED3DSIO_DEF == curOpcode->opcode ||
851 WINED3DSIO_DEFI == curOpcode->opcode ||
852 WINED3DSIO_DEFB == curOpcode->opcode ||
853 WINED3DSIO_PHASE == curOpcode->opcode ||
854 WINED3DSIO_RET == curOpcode->opcode) {
855
856 pToken += shader_skip_opcode(This, curOpcode, hw_arg.opcode_token);
857
858 /* If a generator function is set for current shader target, use it */
859 } else if (hw_fct != NULL) {
860
861 hw_arg.opcode = curOpcode;
862
863 /* Destination token */
864 if (curOpcode->dst_token) {
865
866 DWORD param, addr_token = 0;
867 pToken += shader_get_param(iface, pToken, &param, &addr_token);
868 hw_arg.dst = param;
869 hw_arg.dst_addr = addr_token;
870 }
871
872 /* Predication token */
873 if (hw_arg.opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED)
874 hw_arg.predicate = *pToken++;
875
876 /* Other source tokens */
877 for (i = 0; i < (curOpcode->num_params - curOpcode->dst_token); i++) {
878
879 DWORD param, addr_token = 0;
880 pToken += shader_get_param(iface, pToken, &param, &addr_token);
881 hw_arg.src[i] = param;
882 hw_arg.src_addr[i] = addr_token;
883 }
884
885 /* Call appropriate function for output target */
886 hw_fct(&hw_arg);
887
888 /* Add color correction if needed */
889 device->shader_backend->shader_color_correction(&hw_arg);
890
891 /* Process instruction modifiers for GLSL apps ( _sat, etc. ) */
892 if (This->baseShader.shader_mode == SHADER_GLSL)
893 shader_glsl_add_instruction_modifiers(&hw_arg);
894
895 /* Unhandled opcode */
896 } else {
897
898 FIXME("Can't handle opcode %s in hwShader\n", curOpcode->name);
899 pToken += shader_skip_opcode(This, curOpcode, hw_arg.opcode_token);
900 }
901 }
902 /* TODO: What about result.depth? */
903
904 }
905 }
906
907 void shader_dump_ins_modifiers(const DWORD output) {
908
909 DWORD shift = (output & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
910 DWORD mmask = output & WINED3DSP_DSTMOD_MASK;
911
912 switch (shift) {
913 case 0: break;
914 case 13: TRACE("_d8"); break;
915 case 14: TRACE("_d4"); break;
916 case 15: TRACE("_d2"); break;
917 case 1: TRACE("_x2"); break;
918 case 2: TRACE("_x4"); break;
919 case 3: TRACE("_x8"); break;
920 default: TRACE("_unhandled_shift(%d)", shift); break;
921 }
922
923 if (mmask & WINED3DSPDM_SATURATE) TRACE("_sat");
924 if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
925 if (mmask & WINED3DSPDM_MSAMPCENTROID) TRACE("_centroid");
926
927 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
928 if (mmask)
929 FIXME("_unrecognized_modifier(%#x)", mmask >> WINED3DSP_DSTMOD_SHIFT);
930 }
931
932 /* First pass: trace shader, initialize length and version */
933 void shader_trace_init(
934 IWineD3DBaseShader *iface,
935 const DWORD* pFunction) {
936
937 IWineD3DBaseShaderImpl *This =(IWineD3DBaseShaderImpl *)iface;
938
939 const DWORD* pToken = pFunction;
940 const SHADER_OPCODE* curOpcode = NULL;
941 DWORD opcode_token;
942 unsigned int len = 0;
943 DWORD i;
944
945 TRACE("(%p) : Parsing programme\n", This);
946
947 if (NULL != pToken) {
948 while (WINED3DVS_END() != *pToken) {
949 if (shader_is_version_token(*pToken)) { /** version */
950 This->baseShader.hex_version = *pToken;
951 TRACE("%s_%u_%u\n", shader_is_pshader_version(This->baseShader.hex_version)? "ps": "vs",
952 WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version),
953 WINED3DSHADER_VERSION_MINOR(This->baseShader.hex_version));
954 ++pToken;
955 ++len;
956 continue;
957 }
958 if (shader_is_comment(*pToken)) { /** comment */
959 DWORD comment_len = (*pToken & WINED3DSI_COMMENTSIZE_MASK) >> WINED3DSI_COMMENTSIZE_SHIFT;
960 ++pToken;
961 TRACE("//%s\n", (const char*)pToken);
962 pToken += comment_len;
963 len += comment_len + 1;
964 continue;
965 }
966 opcode_token = *pToken++;
967 curOpcode = shader_get_opcode(iface, opcode_token);
968 len++;
969
970 if (NULL == curOpcode) {
971 int tokens_read;
972 FIXME("Unrecognized opcode: token=%08x\n", opcode_token);
973 tokens_read = shader_skip_unrecognized(iface, pToken);
974 pToken += tokens_read;
975 len += tokens_read;
976
977 } else {
978 if (curOpcode->opcode == WINED3DSIO_DCL) {
979
980 DWORD usage = *pToken;
981 DWORD param = *(pToken + 1);
982
983 shader_dump_decl_usage(This, usage, param);
984 shader_dump_ins_modifiers(param);
985 TRACE(" ");
986 shader_dump_param(iface, param, 0, 0);
987 pToken += 2;
988 len += 2;
989
990 } else if (curOpcode->opcode == WINED3DSIO_DEF) {
991
992 unsigned int offset = shader_get_float_offset(*pToken);
993
994 TRACE("def c%u = %f, %f, %f, %f", offset,
995 *(const float *)(pToken + 1),
996 *(const float *)(pToken + 2),
997 *(const float *)(pToken + 3),
998 *(const float *)(pToken + 4));
999
1000 pToken += 5;
1001 len += 5;
1002 } else if (curOpcode->opcode == WINED3DSIO_DEFI) {
1003
1004 TRACE("defi i%u = %d, %d, %d, %d", *pToken & WINED3DSP_REGNUM_MASK,
1005 *(pToken + 1),
1006 *(pToken + 2),
1007 *(pToken + 3),
1008 *(pToken + 4));
1009
1010 pToken += 5;
1011 len += 5;
1012
1013 } else if (curOpcode->opcode == WINED3DSIO_DEFB) {
1014
1015 TRACE("defb b%u = %s", *pToken & WINED3DSP_REGNUM_MASK,
1016 *(pToken + 1)? "true": "false");
1017
1018 pToken += 2;
1019 len += 2;
1020
1021 } else {
1022
1023 DWORD param, addr_token;
1024 int tokens_read;
1025
1026 /* Print out predication source token first - it follows
1027 * the destination token. */
1028 if (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED) {
1029 TRACE("(");
1030 shader_dump_param(iface, *(pToken + 2), 0, 1);
1031 TRACE(") ");
1032 }
1033 if (opcode_token & WINED3DSI_COISSUE) {
1034 /* PixWin marks instructions with the coissue flag with a '+' */
1035 TRACE("+");
1036 }
1037
1038 TRACE("%s", curOpcode->name);
1039
1040 if (curOpcode->opcode == WINED3DSIO_IFC ||
1041 curOpcode->opcode == WINED3DSIO_BREAKC) {
1042
1043 DWORD op = (opcode_token & INST_CONTROLS_MASK) >> INST_CONTROLS_SHIFT;
1044 switch (op) {
1045 case COMPARISON_GT: TRACE("_gt"); break;
1046 case COMPARISON_EQ: TRACE("_eq"); break;
1047 case COMPARISON_GE: TRACE("_ge"); break;
1048 case COMPARISON_LT: TRACE("_lt"); break;
1049 case COMPARISON_NE: TRACE("_ne"); break;
1050 case COMPARISON_LE: TRACE("_le"); break;
1051 default:
1052 TRACE("_(%u)", op);
1053 }
1054 } else if (curOpcode->opcode == WINED3DSIO_TEX &&
1055 This->baseShader.hex_version >= WINED3DPS_VERSION(2,0)) {
1056 if(opcode_token & WINED3DSI_TEXLD_PROJECT) TRACE("p");
1057 }
1058
1059 /* Destination token */
1060 if (curOpcode->dst_token) {
1061
1062 /* Destination token */
1063 tokens_read = shader_get_param(iface, pToken, &param, &addr_token);
1064 pToken += tokens_read;
1065 len += tokens_read;
1066
1067 shader_dump_ins_modifiers(param);
1068 TRACE(" ");
1069 shader_dump_param(iface, param, addr_token, 0);
1070 }
1071
1072 /* Predication token - already printed out, just skip it */
1073 if (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED) {
1074 pToken++;
1075 len++;
1076 }
1077
1078 /* Other source tokens */
1079 for (i = curOpcode->dst_token; i < curOpcode->num_params; ++i) {
1080
1081 tokens_read = shader_get_param(iface, pToken, &param, &addr_token);
1082 pToken += tokens_read;
1083 len += tokens_read;
1084
1085 TRACE((i == 0)? " " : ", ");
1086 shader_dump_param(iface, param, addr_token, 1);
1087 }
1088 }
1089 TRACE("\n");
1090 }
1091 }
1092 This->baseShader.functionLength = (len + 1) * sizeof(DWORD);
1093 } else {
1094 This->baseShader.functionLength = 1; /* no Function defined use fixed function vertex processing */
1095 }
1096 }
1097
1098 static void shader_none_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {}
1099 static void shader_none_select_depth_blt(IWineD3DDevice *iface) {}
1100 static void shader_none_destroy_depth_blt(IWineD3DDevice *iface) {}
1101 static void shader_none_load_constants(IWineD3DDevice *iface, char usePS, char useVS) {}
1102 static void shader_none_cleanup(IWineD3DDevice *iface) {}
1103 static void shader_none_color_correction(SHADER_OPCODE_ARG* arg) {}
1104 static void shader_none_destroy(IWineD3DBaseShader *iface) {}
1105 static HRESULT shader_none_alloc(IWineD3DDevice *iface) {return WINED3D_OK;}
1106 static void shader_none_free(IWineD3DDevice *iface) {}
1107
1108 const shader_backend_t none_shader_backend = {
1109 &shader_none_select,
1110 &shader_none_select_depth_blt,
1111 &shader_none_destroy_depth_blt,
1112 &shader_none_load_constants,
1113 &shader_none_cleanup,
1114 &shader_none_color_correction,
1115 &shader_none_destroy,
1116 &shader_none_alloc,
1117 &shader_none_free
1118 };
1119
1120 /* *******************************************
1121 IWineD3DPixelShader IUnknown parts follow
1122 ******************************************* */
1123 HRESULT WINAPI IWineD3DBaseShaderImpl_QueryInterface(IWineD3DBaseShader *iface, REFIID riid, LPVOID *ppobj)
1124 {
1125 IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)iface;
1126 TRACE("(%p)->(%s,%p)\n",This,debugstr_guid(riid),ppobj);
1127 if (IsEqualGUID(riid, &IID_IUnknown)
1128 || IsEqualGUID(riid, &IID_IWineD3DBase)
1129 || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1130 || IsEqualGUID(riid, &IID_IWineD3DPixelShader)) {
1131 IUnknown_AddRef(iface);
1132 *ppobj = This;
1133 return S_OK;
1134 }
1135 *ppobj = NULL;
1136 return E_NOINTERFACE;
1137 }
1138
1139 ULONG WINAPI IWineD3DBaseShaderImpl_AddRef(IWineD3DBaseShader *iface) {
1140 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
1141 TRACE("(%p) : AddRef increasing from %d\n", This, This->baseShader.ref);
1142 return InterlockedIncrement(&This->baseShader.ref);
1143 }
1144
1145 ULONG WINAPI IWineD3DBaseShaderImpl_Release(IWineD3DBaseShader *iface) {
1146 IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)iface;
1147 IWineD3DDeviceImpl *deviceImpl = (IWineD3DDeviceImpl *) This->baseShader.device;
1148 ULONG ref;
1149 TRACE("(%p) : Releasing from %d\n", This, This->baseShader.ref);
1150 ref = InterlockedDecrement(&This->baseShader.ref);
1151 if (ref == 0) {
1152 deviceImpl->shader_backend->shader_destroy(iface);
1153 HeapFree(GetProcessHeap(), 0, This->baseShader.function);
1154 shader_delete_constant_list(&This->baseShader.constantsF);
1155 shader_delete_constant_list(&This->baseShader.constantsB);
1156 shader_delete_constant_list(&This->baseShader.constantsI);
1157 list_remove(&This->baseShader.shader_list_entry);
1158 HeapFree(GetProcessHeap(), 0, This);
1159 }
1160 return ref;
1161 }
useful hacks and other patches not (yet) suitable for mainline