?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
wined3d: Make the state table a property of the shader backend.
[wine/multimedia.git] / dlls / wined3d / baseshader.c
blob11b4861da1123221cd839e2ea5a9893af0caab67
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 memset(reg_maps->bumpmat, 0, sizeof(reg_maps->bumpmat));
216 memset(reg_maps->luminanceparams, 0, sizeof(reg_maps->luminanceparams));
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 pshader /* Filter different instructions with the same enum values in VS */ &&
353 (WINED3DSIO_TEX == curOpcode->opcode ||
354 WINED3DSIO_TEXBEM == curOpcode->opcode ||
355 WINED3DSIO_TEXBEML == curOpcode->opcode ||
356 WINED3DSIO_TEXDP3TEX == curOpcode->opcode ||
357 WINED3DSIO_TEXM3x2TEX == curOpcode->opcode ||
358 WINED3DSIO_TEXM3x3SPEC == curOpcode->opcode ||
359 WINED3DSIO_TEXM3x3TEX == curOpcode->opcode ||
360 WINED3DSIO_TEXM3x3VSPEC == curOpcode->opcode ||
361 WINED3DSIO_TEXREG2AR == curOpcode->opcode ||
362 WINED3DSIO_TEXREG2GB == curOpcode->opcode ||
363 WINED3DSIO_TEXREG2RGB == curOpcode->opcode)) {
364
365 /* Fake sampler usage, only set reserved bit and ttype */
366 DWORD sampler_code = *pToken & WINED3DSP_REGNUM_MASK;
367
368 if(!stateBlock->textures[sampler_code]) {
369 ERR("No texture bound to sampler %d\n", sampler_code);
370 reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
371 } else {
372 int texType = IWineD3DBaseTexture_GetTextureDimensions(stateBlock->textures[sampler_code]);
373 switch(texType) {
374 /* We have to select between texture rectangles and 2D textures later because 2.0 and
375 * 3.0 shaders only have WINED3DSTT_2D as well
376 */
377 case GL_TEXTURE_RECTANGLE_ARB:
378 case GL_TEXTURE_2D:
379 reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
380 break;
381
382 case GL_TEXTURE_3D:
383 reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_VOLUME;
384 break;
385
386 case GL_TEXTURE_CUBE_MAP_ARB:
387 reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_CUBE;
388 break;
389
390 default:
391 ERR("Unexpected gl texture type found: %d\n", texType);
392 reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
393 }
394 }
395
396 /* texbem is only valid with < 1.4 pixel shaders */
397 if(WINED3DSIO_TEXBEM == curOpcode->opcode ||
398 WINED3DSIO_TEXBEML == curOpcode->opcode) {
399 reg_maps->bumpmat[sampler_code] = TRUE;
400 if(WINED3DSIO_TEXBEML == curOpcode->opcode) {
401 reg_maps->luminanceparams[sampler_code] = TRUE;
402 }
403 }
404 }
405 if(WINED3DSIO_NRM == curOpcode->opcode) {
406 reg_maps->usesnrm = 1;
407 } else if(WINED3DSIO_BEM == curOpcode->opcode && pshader) {
408 DWORD regnum = *pToken & WINED3DSP_REGNUM_MASK;
409 reg_maps->bumpmat[regnum] = TRUE;
410 } else if(WINED3DSIO_DSY == curOpcode->opcode) {
411 reg_maps->usesdsy = 1;
412 }
413
414 /* This will loop over all the registers and try to
415 * make a bitmask of the ones we're interested in.
416 *
417 * Relative addressing tokens are ignored, but that's
418 * okay, since we'll catch any address registers when
419 * they are initialized (required by spec) */
420
421 limit = (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED)?
422 curOpcode->num_params + 1: curOpcode->num_params;
423
424 for (i = 0; i < limit; ++i) {
425
426 DWORD param, addr_token, reg, regtype;
427 pToken += shader_get_param(iface, pToken, &param, &addr_token);
428
429 regtype = shader_get_regtype(param);
430 reg = param & WINED3DSP_REGNUM_MASK;
431
432 if (WINED3DSPR_TEXTURE == regtype) { /* vs: WINED3DSPR_ADDR */
433
434 if (pshader)
435 reg_maps->texcoord[reg] = 1;
436 else
437 reg_maps->address[reg] = 1;
438 }
439
440 else if (WINED3DSPR_TEMP == regtype)
441 reg_maps->temporary[reg] = 1;
442
443 else if (WINED3DSPR_INPUT == regtype) {
444 if( !pshader)
445 reg_maps->attributes[reg] = 1;
446 else {
447 if(param & WINED3DSHADER_ADDRMODE_RELATIVE) {
448 /* If relative addressing is used, we must assume that all registers
449 * are used. Even if it is a construct like v3[aL], we can't assume
450 * that v0, v1 and v2 aren't read because aL can be negative
451 */
452 unsigned int i;
453 for(i = 0; i < MAX_REG_INPUT; i++) {
454 ((IWineD3DPixelShaderImpl *) This)->input_reg_used[i] = TRUE;
455 }
456 } else {
457 ((IWineD3DPixelShaderImpl *) This)->input_reg_used[reg] = TRUE;
458 }
459 }
460 }
461
462 else if (WINED3DSPR_RASTOUT == regtype && reg == 1)
463 reg_maps->fog = 1;
464
465 else if (WINED3DSPR_MISCTYPE == regtype && reg == 0 && pshader)
466 reg_maps->vpos = 1;
467
468 else if(WINED3DSPR_CONST == regtype) {
469 if(param & WINED3DSHADER_ADDRMODE_RELATIVE) {
470 if(!pshader) {
471 if(reg <= ((IWineD3DVertexShaderImpl *) This)->min_rel_offset) {
472 ((IWineD3DVertexShaderImpl *) This)->min_rel_offset = reg;
473 } else if(reg >= ((IWineD3DVertexShaderImpl *) This)->max_rel_offset) {
474 ((IWineD3DVertexShaderImpl *) This)->max_rel_offset = reg;
475 }
476 }
477 reg_maps->usesrelconstF = TRUE;
478 }
479 }
480
481 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and is used
482 * in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel shaders because TECRDOUT
483 * isn't used in them, but future register types might cause issues
484 */
485 else if(WINED3DSPR_TEXCRDOUT == regtype && i == 0 /* Only look at writes */ &&
486 !pshader && WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) < 3) {
487 reg_maps->texcoord_mask[reg] |= shader_get_writemask(param);
488 }
489 }
490 }
491 }
492 reg_maps->loop_depth = max_loop_depth;
493
494 return WINED3D_OK;
495 }
496
497 static void shader_dump_decl_usage(
498 IWineD3DBaseShaderImpl* This,
499 DWORD decl,
500 DWORD param) {
501
502 DWORD regtype = shader_get_regtype(param);
503
504 TRACE("dcl");
505
506 if (regtype == WINED3DSPR_SAMPLER) {
507 DWORD ttype = decl & WINED3DSP_TEXTURETYPE_MASK;
508
509 switch (ttype) {
510 case WINED3DSTT_2D: TRACE("_2d"); break;
511 case WINED3DSTT_CUBE: TRACE("_cube"); break;
512 case WINED3DSTT_VOLUME: TRACE("_volume"); break;
513 default: TRACE("_unknown_ttype(%08x)", ttype);
514 }
515
516 } else {
517
518 DWORD usage = decl & WINED3DSP_DCL_USAGE_MASK;
519 DWORD idx = (decl & WINED3DSP_DCL_USAGEINDEX_MASK) >> WINED3DSP_DCL_USAGEINDEX_SHIFT;
520
521 /* Pixel shaders 3.0 don't have usage semantics */
522 char pshader = shader_is_pshader_version(This->baseShader.hex_version);
523 if (pshader && This->baseShader.hex_version < WINED3DPS_VERSION(3,0))
524 return;
525 else
526 TRACE("_");
527
528 switch(usage) {
529 case WINED3DDECLUSAGE_POSITION:
530 TRACE("position%d", idx);
531 break;
532 case WINED3DDECLUSAGE_BLENDINDICES:
533 TRACE("blend");
534 break;
535 case WINED3DDECLUSAGE_BLENDWEIGHT:
536 TRACE("weight");
537 break;
538 case WINED3DDECLUSAGE_NORMAL:
539 TRACE("normal%d", idx);
540 break;
541 case WINED3DDECLUSAGE_PSIZE:
542 TRACE("psize");
543 break;
544 case WINED3DDECLUSAGE_COLOR:
545 if(idx == 0) {
546 TRACE("color");
547 } else {
548 TRACE("specular%d", (idx - 1));
549 }
550 break;
551 case WINED3DDECLUSAGE_TEXCOORD:
552 TRACE("texture%d", idx);
553 break;
554 case WINED3DDECLUSAGE_TANGENT:
555 TRACE("tangent");
556 break;
557 case WINED3DDECLUSAGE_BINORMAL:
558 TRACE("binormal");
559 break;
560 case WINED3DDECLUSAGE_TESSFACTOR:
561 TRACE("tessfactor");
562 break;
563 case WINED3DDECLUSAGE_POSITIONT:
564 TRACE("positionT%d", idx);
565 break;
566 case WINED3DDECLUSAGE_FOG:
567 TRACE("fog");
568 break;
569 case WINED3DDECLUSAGE_DEPTH:
570 TRACE("depth");
571 break;
572 case WINED3DDECLUSAGE_SAMPLE:
573 TRACE("sample");
574 break;
575 default:
576 FIXME("unknown_semantics(%08x)", usage);
577 }
578 }
579 }
580
581 static void shader_dump_arr_entry(
582 IWineD3DBaseShader *iface,
583 const DWORD param,
584 const DWORD addr_token,
585 unsigned int reg,
586 int input) {
587
588 char relative =
589 ((param & WINED3DSHADER_ADDRESSMODE_MASK) == WINED3DSHADER_ADDRMODE_RELATIVE);
590
591 if (relative) {
592 TRACE("[");
593 if (addr_token)
594 shader_dump_param(iface, addr_token, 0, input);
595 else
596 TRACE("a0.x");
597 TRACE(" + ");
598 }
599 TRACE("%u", reg);
600 if (relative)
601 TRACE("]");
602 }
603
604 void shader_dump_param(
605 IWineD3DBaseShader *iface,
606 const DWORD param,
607 const DWORD addr_token,
608 int input) {
609
610 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
611 static const char * const rastout_reg_names[] = { "oPos", "oFog", "oPts" };
612 static const char * const misctype_reg_names[] = { "vPos", "vFace"};
613 char swizzle_reg_chars[4];
614
615 DWORD reg = param & WINED3DSP_REGNUM_MASK;
616 DWORD regtype = shader_get_regtype(param);
617 DWORD modifier = param & WINED3DSP_SRCMOD_MASK;
618
619 /* There are some minor differences between pixel and vertex shaders */
620 char pshader = shader_is_pshader_version(This->baseShader.hex_version);
621
622 /* For one, we'd prefer color components to be shown for pshaders.
623 * FIXME: use the swizzle function for this */
624
625 swizzle_reg_chars[0] = pshader? 'r': 'x';
626 swizzle_reg_chars[1] = pshader? 'g': 'y';
627 swizzle_reg_chars[2] = pshader? 'b': 'z';
628 swizzle_reg_chars[3] = pshader? 'a': 'w';
629
630 if (input) {
631 if ( (modifier == WINED3DSPSM_NEG) ||
632 (modifier == WINED3DSPSM_BIASNEG) ||
633 (modifier == WINED3DSPSM_SIGNNEG) ||
634 (modifier == WINED3DSPSM_X2NEG) ||
635 (modifier == WINED3DSPSM_ABSNEG) )
636 TRACE("-");
637 else if (modifier == WINED3DSPSM_COMP)
638 TRACE("1-");
639 else if (modifier == WINED3DSPSM_NOT)
640 TRACE("!");
641
642 if (modifier == WINED3DSPSM_ABS || modifier == WINED3DSPSM_ABSNEG)
643 TRACE("abs(");
644 }
645
646 switch (regtype) {
647 case WINED3DSPR_TEMP:
648 TRACE("r%u", reg);
649 break;
650 case WINED3DSPR_INPUT:
651 TRACE("v");
652 shader_dump_arr_entry(iface, param, addr_token, reg, input);
653 break;
654 case WINED3DSPR_CONST:
655 case WINED3DSPR_CONST2:
656 case WINED3DSPR_CONST3:
657 case WINED3DSPR_CONST4:
658 TRACE("c");
659 shader_dump_arr_entry(iface, param, addr_token, shader_get_float_offset(param), input);
660 break;
661 case WINED3DSPR_TEXTURE: /* vs: case D3DSPR_ADDR */
662 TRACE("%c%u", (pshader? 't':'a'), reg);
663 break;
664 case WINED3DSPR_RASTOUT:
665 TRACE("%s", rastout_reg_names[reg]);
666 break;
667 case WINED3DSPR_COLOROUT:
668 TRACE("oC%u", reg);
669 break;
670 case WINED3DSPR_DEPTHOUT:
671 TRACE("oDepth");
672 break;
673 case WINED3DSPR_ATTROUT:
674 TRACE("oD%u", reg);
675 break;
676 case WINED3DSPR_TEXCRDOUT:
677
678 /* Vertex shaders >= 3.0 use general purpose output registers
679 * (WINED3DSPR_OUTPUT), which can include an address token */
680
681 if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3) {
682 TRACE("o");
683 shader_dump_arr_entry(iface, param, addr_token, reg, input);
684 }
685 else
686 TRACE("oT%u", reg);
687 break;
688 case WINED3DSPR_CONSTINT:
689 TRACE("i");
690 shader_dump_arr_entry(iface, param, addr_token, reg, input);
691 break;
692 case WINED3DSPR_CONSTBOOL:
693 TRACE("b");
694 shader_dump_arr_entry(iface, param, addr_token, reg, input);
695 break;
696 case WINED3DSPR_LABEL:
697 TRACE("l%u", reg);
698 break;
699 case WINED3DSPR_LOOP:
700 TRACE("aL");
701 break;
702 case WINED3DSPR_SAMPLER:
703 TRACE("s%u", reg);
704 break;
705 case WINED3DSPR_MISCTYPE:
706 if (reg > 1) {
707 FIXME("Unhandled misctype register %d\n", reg);
708 } else {
709 TRACE("%s", misctype_reg_names[reg]);
710 }
711 break;
712 case WINED3DSPR_PREDICATE:
713 TRACE("p%u", reg);
714 break;
715 default:
716 TRACE("unhandled_rtype(%#x)", regtype);
717 break;
718 }
719
720 if (!input) {
721 /* operand output (for modifiers and shift, see dump_ins_modifiers) */
722
723 if ((param & WINED3DSP_WRITEMASK_ALL) != WINED3DSP_WRITEMASK_ALL) {
724 TRACE(".");
725 if (param & WINED3DSP_WRITEMASK_0) TRACE("%c", swizzle_reg_chars[0]);
726 if (param & WINED3DSP_WRITEMASK_1) TRACE("%c", swizzle_reg_chars[1]);
727 if (param & WINED3DSP_WRITEMASK_2) TRACE("%c", swizzle_reg_chars[2]);
728 if (param & WINED3DSP_WRITEMASK_3) TRACE("%c", swizzle_reg_chars[3]);
729 }
730
731 } else {
732 /** operand input */
733 DWORD swizzle = (param & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
734 DWORD swizzle_r = swizzle & 0x03;
735 DWORD swizzle_g = (swizzle >> 2) & 0x03;
736 DWORD swizzle_b = (swizzle >> 4) & 0x03;
737 DWORD swizzle_a = (swizzle >> 6) & 0x03;
738
739 if (0 != modifier) {
740 switch (modifier) {
741 case WINED3DSPSM_NONE: break;
742 case WINED3DSPSM_NEG: break;
743 case WINED3DSPSM_NOT: break;
744 case WINED3DSPSM_BIAS: TRACE("_bias"); break;
745 case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
746 case WINED3DSPSM_SIGN: TRACE("_bx2"); break;
747 case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
748 case WINED3DSPSM_COMP: break;
749 case WINED3DSPSM_X2: TRACE("_x2"); break;
750 case WINED3DSPSM_X2NEG: TRACE("_x2"); break;
751 case WINED3DSPSM_DZ: TRACE("_dz"); break;
752 case WINED3DSPSM_DW: TRACE("_dw"); break;
753 case WINED3DSPSM_ABSNEG: TRACE(")"); break;
754 case WINED3DSPSM_ABS: TRACE(")"); break;
755 default:
756 TRACE("_unknown_modifier(%#x)", modifier >> WINED3DSP_SRCMOD_SHIFT);
757 }
758 }
759
760 /**
761 * swizzle bits fields:
762 * RRGGBBAA
763 */
764 if ((WINED3DVS_NOSWIZZLE >> WINED3DVS_SWIZZLE_SHIFT) != swizzle) {
765 if (swizzle_r == swizzle_g &&
766 swizzle_r == swizzle_b &&
767 swizzle_r == swizzle_a) {
768 TRACE(".%c", swizzle_reg_chars[swizzle_r]);
769 } else {
770 TRACE(".%c%c%c%c",
771 swizzle_reg_chars[swizzle_r],
772 swizzle_reg_chars[swizzle_g],
773 swizzle_reg_chars[swizzle_b],
774 swizzle_reg_chars[swizzle_a]);
775 }
776 }
777 }
778 }
779
780 /** Shared code in order to generate the bulk of the shader string.
781 Use the shader_header_fct & shader_footer_fct to add strings
782 that are specific to pixel or vertex functions
783 NOTE: A description of how to parse tokens can be found on msdn */
784 void shader_generate_main(
785 IWineD3DBaseShader *iface,
786 SHADER_BUFFER* buffer,
787 shader_reg_maps* reg_maps,
788 CONST DWORD* pFunction) {
789
790 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
791 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device; /* To access shader backend callbacks */
792 const DWORD *pToken = pFunction;
793 const SHADER_OPCODE *curOpcode = NULL;
794 SHADER_HANDLER hw_fct = NULL;
795 DWORD i;
796 SHADER_OPCODE_ARG hw_arg;
797
798 /* Initialize current parsing state */
799 hw_arg.shader = iface;
800 hw_arg.buffer = buffer;
801 hw_arg.reg_maps = reg_maps;
802 This->baseShader.parse_state.current_row = 0;
803
804 /* Second pass, process opcodes */
805 if (NULL != pToken) {
806 while (WINED3DPS_END() != *pToken) {
807
808 /* Skip version token */
809 if (shader_is_version_token(*pToken)) {
810 ++pToken;
811 continue;
812 }
813
814 /* Skip comment tokens */
815 if (shader_is_comment(*pToken)) {
816 DWORD comment_len = (*pToken & WINED3DSI_COMMENTSIZE_MASK) >> WINED3DSI_COMMENTSIZE_SHIFT;
817 ++pToken;
818 TRACE("#%s\n", (const char*)pToken);
819 pToken += comment_len;
820 continue;
821 }
822
823 /* Read opcode */
824 hw_arg.opcode_token = *pToken++;
825 curOpcode = shader_get_opcode(iface, hw_arg.opcode_token);
826
827 /* Select handler */
828 if (curOpcode == NULL)
829 hw_fct = NULL;
830 else if (This->baseShader.shader_mode == SHADER_GLSL)
831 hw_fct = curOpcode->hw_glsl_fct;
832 else if (This->baseShader.shader_mode == SHADER_ARB)
833 hw_fct = curOpcode->hw_fct;
834
835 /* Unknown opcode and its parameters */
836 if (NULL == curOpcode) {
837 FIXME("Unrecognized opcode: token=%08x\n", hw_arg.opcode_token);
838 pToken += shader_skip_unrecognized(iface, pToken);
839
840 /* Nothing to do */
841 } else if (WINED3DSIO_DCL == curOpcode->opcode ||
842 WINED3DSIO_NOP == curOpcode->opcode ||
843 WINED3DSIO_DEF == curOpcode->opcode ||
844 WINED3DSIO_DEFI == curOpcode->opcode ||
845 WINED3DSIO_DEFB == curOpcode->opcode ||
846 WINED3DSIO_PHASE == curOpcode->opcode ||
847 WINED3DSIO_RET == curOpcode->opcode) {
848
849 pToken += shader_skip_opcode(This, curOpcode, hw_arg.opcode_token);
850
851 /* If a generator function is set for current shader target, use it */
852 } else if (hw_fct != NULL) {
853
854 hw_arg.opcode = curOpcode;
855
856 /* Destination token */
857 if (curOpcode->dst_token) {
858
859 DWORD param, addr_token = 0;
860 pToken += shader_get_param(iface, pToken, &param, &addr_token);
861 hw_arg.dst = param;
862 hw_arg.dst_addr = addr_token;
863 }
864
865 /* Predication token */
866 if (hw_arg.opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED)
867 hw_arg.predicate = *pToken++;
868
869 /* Other source tokens */
870 for (i = 0; i < (curOpcode->num_params - curOpcode->dst_token); i++) {
871
872 DWORD param, addr_token = 0;
873 pToken += shader_get_param(iface, pToken, &param, &addr_token);
874 hw_arg.src[i] = param;
875 hw_arg.src_addr[i] = addr_token;
876 }
877
878 /* Call appropriate function for output target */
879 hw_fct(&hw_arg);
880
881 /* Add color correction if needed */
882 device->shader_backend->shader_color_correction(&hw_arg);
883
884 /* Process instruction modifiers for GLSL apps ( _sat, etc. ) */
885 if (This->baseShader.shader_mode == SHADER_GLSL)
886 shader_glsl_add_instruction_modifiers(&hw_arg);
887
888 /* Unhandled opcode */
889 } else {
890
891 FIXME("Can't handle opcode %s in hwShader\n", curOpcode->name);
892 pToken += shader_skip_opcode(This, curOpcode, hw_arg.opcode_token);
893 }
894 }
895 /* TODO: What about result.depth? */
896
897 }
898 }
899
900 void shader_dump_ins_modifiers(const DWORD output) {
901
902 DWORD shift = (output & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
903 DWORD mmask = output & WINED3DSP_DSTMOD_MASK;
904
905 switch (shift) {
906 case 0: break;
907 case 13: TRACE("_d8"); break;
908 case 14: TRACE("_d4"); break;
909 case 15: TRACE("_d2"); break;
910 case 1: TRACE("_x2"); break;
911 case 2: TRACE("_x4"); break;
912 case 3: TRACE("_x8"); break;
913 default: TRACE("_unhandled_shift(%d)", shift); break;
914 }
915
916 if (mmask & WINED3DSPDM_SATURATE) TRACE("_sat");
917 if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
918 if (mmask & WINED3DSPDM_MSAMPCENTROID) TRACE("_centroid");
919
920 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
921 if (mmask)
922 FIXME("_unrecognized_modifier(%#x)", mmask >> WINED3DSP_DSTMOD_SHIFT);
923 }
924
925 /* First pass: trace shader, initialize length and version */
926 void shader_trace_init(
927 IWineD3DBaseShader *iface,
928 const DWORD* pFunction) {
929
930 IWineD3DBaseShaderImpl *This =(IWineD3DBaseShaderImpl *)iface;
931
932 const DWORD* pToken = pFunction;
933 const SHADER_OPCODE* curOpcode = NULL;
934 DWORD opcode_token;
935 unsigned int len = 0;
936 DWORD i;
937
938 TRACE("(%p) : Parsing programme\n", This);
939
940 if (NULL != pToken) {
941 while (WINED3DVS_END() != *pToken) {
942 if (shader_is_version_token(*pToken)) { /** version */
943 This->baseShader.hex_version = *pToken;
944 TRACE("%s_%u_%u\n", shader_is_pshader_version(This->baseShader.hex_version)? "ps": "vs",
945 WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version),
946 WINED3DSHADER_VERSION_MINOR(This->baseShader.hex_version));
947 ++pToken;
948 ++len;
949 continue;
950 }
951 if (shader_is_comment(*pToken)) { /** comment */
952 DWORD comment_len = (*pToken & WINED3DSI_COMMENTSIZE_MASK) >> WINED3DSI_COMMENTSIZE_SHIFT;
953 ++pToken;
954 TRACE("//%s\n", (const char*)pToken);
955 pToken += comment_len;
956 len += comment_len + 1;
957 continue;
958 }
959 opcode_token = *pToken++;
960 curOpcode = shader_get_opcode(iface, opcode_token);
961 len++;
962
963 if (NULL == curOpcode) {
964 int tokens_read;
965 FIXME("Unrecognized opcode: token=%08x\n", opcode_token);
966 tokens_read = shader_skip_unrecognized(iface, pToken);
967 pToken += tokens_read;
968 len += tokens_read;
969
970 } else {
971 if (curOpcode->opcode == WINED3DSIO_DCL) {
972
973 DWORD usage = *pToken;
974 DWORD param = *(pToken + 1);
975
976 shader_dump_decl_usage(This, usage, param);
977 shader_dump_ins_modifiers(param);
978 TRACE(" ");
979 shader_dump_param(iface, param, 0, 0);
980 pToken += 2;
981 len += 2;
982
983 } else if (curOpcode->opcode == WINED3DSIO_DEF) {
984
985 unsigned int offset = shader_get_float_offset(*pToken);
986
987 TRACE("def c%u = %f, %f, %f, %f", offset,
988 *(const float *)(pToken + 1),
989 *(const float *)(pToken + 2),
990 *(const float *)(pToken + 3),
991 *(const float *)(pToken + 4));
992
993 pToken += 5;
994 len += 5;
995 } else if (curOpcode->opcode == WINED3DSIO_DEFI) {
996
997 TRACE("defi i%u = %d, %d, %d, %d", *pToken & WINED3DSP_REGNUM_MASK,
998 *(pToken + 1),
999 *(pToken + 2),
1000 *(pToken + 3),
1001 *(pToken + 4));
1002
1003 pToken += 5;
1004 len += 5;
1005
1006 } else if (curOpcode->opcode == WINED3DSIO_DEFB) {
1007
1008 TRACE("defb b%u = %s", *pToken & WINED3DSP_REGNUM_MASK,
1009 *(pToken + 1)? "true": "false");
1010
1011 pToken += 2;
1012 len += 2;
1013
1014 } else {
1015
1016 DWORD param, addr_token;
1017 int tokens_read;
1018
1019 /* Print out predication source token first - it follows
1020 * the destination token. */
1021 if (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED) {
1022 TRACE("(");
1023 shader_dump_param(iface, *(pToken + 2), 0, 1);
1024 TRACE(") ");
1025 }
1026 if (opcode_token & WINED3DSI_COISSUE) {
1027 /* PixWin marks instructions with the coissue flag with a '+' */
1028 TRACE("+");
1029 }
1030
1031 TRACE("%s", curOpcode->name);
1032
1033 if (curOpcode->opcode == WINED3DSIO_IFC ||
1034 curOpcode->opcode == WINED3DSIO_BREAKC) {
1035
1036 DWORD op = (opcode_token & INST_CONTROLS_MASK) >> INST_CONTROLS_SHIFT;
1037 switch (op) {
1038 case COMPARISON_GT: TRACE("_gt"); break;
1039 case COMPARISON_EQ: TRACE("_eq"); break;
1040 case COMPARISON_GE: TRACE("_ge"); break;
1041 case COMPARISON_LT: TRACE("_lt"); break;
1042 case COMPARISON_NE: TRACE("_ne"); break;
1043 case COMPARISON_LE: TRACE("_le"); break;
1044 default:
1045 TRACE("_(%u)", op);
1046 }
1047 } else if (curOpcode->opcode == WINED3DSIO_TEX &&
1048 This->baseShader.hex_version >= WINED3DPS_VERSION(2,0)) {
1049 if(opcode_token & WINED3DSI_TEXLD_PROJECT) TRACE("p");
1050 }
1051
1052 /* Destination token */
1053 if (curOpcode->dst_token) {
1054
1055 /* Destination token */
1056 tokens_read = shader_get_param(iface, pToken, &param, &addr_token);
1057 pToken += tokens_read;
1058 len += tokens_read;
1059
1060 shader_dump_ins_modifiers(param);
1061 TRACE(" ");
1062 shader_dump_param(iface, param, addr_token, 0);
1063 }
1064
1065 /* Predication token - already printed out, just skip it */
1066 if (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED) {
1067 pToken++;
1068 len++;
1069 }
1070
1071 /* Other source tokens */
1072 for (i = curOpcode->dst_token; i < curOpcode->num_params; ++i) {
1073
1074 tokens_read = shader_get_param(iface, pToken, &param, &addr_token);
1075 pToken += tokens_read;
1076 len += tokens_read;
1077
1078 TRACE((i == 0)? " " : ", ");
1079 shader_dump_param(iface, param, addr_token, 1);
1080 }
1081 }
1082 TRACE("\n");
1083 }
1084 }
1085 This->baseShader.functionLength = (len + 1) * sizeof(DWORD);
1086 } else {
1087 This->baseShader.functionLength = 1; /* no Function defined use fixed function vertex processing */
1088 }
1089 }
1090
1091 static void shader_none_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {}
1092 static void shader_none_select_depth_blt(IWineD3DDevice *iface) {}
1093 static void shader_none_destroy_depth_blt(IWineD3DDevice *iface) {}
1094 static void shader_none_load_constants(IWineD3DDevice *iface, char usePS, char useVS) {}
1095 static void shader_none_cleanup(IWineD3DDevice *iface) {}
1096 static void shader_none_color_correction(SHADER_OPCODE_ARG* arg) {}
1097 static void shader_none_destroy(IWineD3DBaseShader *iface) {}
1098 static HRESULT shader_none_alloc(IWineD3DDevice *iface) {return WINED3D_OK;}
1099 static void shader_none_free(IWineD3DDevice *iface) {}
1100 static BOOL shader_none_dirty_const(IWineD3DDevice *iface) {return FALSE;}
1101
1102 const shader_backend_t none_shader_backend = {
1103 &shader_none_select,
1104 &shader_none_select_depth_blt,
1105 &shader_none_destroy_depth_blt,
1106 &shader_none_load_constants,
1107 &shader_none_cleanup,
1108 &shader_none_color_correction,
1109 &shader_none_destroy,
1110 &shader_none_alloc,
1111 &shader_none_free,
1112 &shader_none_dirty_const,
1113 FFPStateTable
1114 };
1115
1116 /* *******************************************
1117 IWineD3DPixelShader IUnknown parts follow
1118 ******************************************* */
1119 HRESULT WINAPI IWineD3DBaseShaderImpl_QueryInterface(IWineD3DBaseShader *iface, REFIID riid, LPVOID *ppobj)
1120 {
1121 IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)iface;
1122 TRACE("(%p)->(%s,%p)\n",This,debugstr_guid(riid),ppobj);
1123 if (IsEqualGUID(riid, &IID_IUnknown)
1124 || IsEqualGUID(riid, &IID_IWineD3DBase)
1125 || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
1126 || IsEqualGUID(riid, &IID_IWineD3DPixelShader)) {
1127 IUnknown_AddRef(iface);
1128 *ppobj = This;
1129 return S_OK;
1130 }
1131 *ppobj = NULL;
1132 return E_NOINTERFACE;
1133 }
1134
1135 ULONG WINAPI IWineD3DBaseShaderImpl_AddRef(IWineD3DBaseShader *iface) {
1136 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
1137 TRACE("(%p) : AddRef increasing from %d\n", This, This->baseShader.ref);
1138 return InterlockedIncrement(&This->baseShader.ref);
1139 }
1140
1141 ULONG WINAPI IWineD3DBaseShaderImpl_Release(IWineD3DBaseShader *iface) {
1142 IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)iface;
1143 IWineD3DDeviceImpl *deviceImpl = (IWineD3DDeviceImpl *) This->baseShader.device;
1144 ULONG ref;
1145 TRACE("(%p) : Releasing from %d\n", This, This->baseShader.ref);
1146 ref = InterlockedDecrement(&This->baseShader.ref);
1147 if (ref == 0) {
1148 deviceImpl->shader_backend->shader_destroy(iface);
1149 HeapFree(GetProcessHeap(), 0, This->baseShader.function);
1150 shader_delete_constant_list(&This->baseShader.constantsF);
1151 shader_delete_constant_list(&This->baseShader.constantsB);
1152 shader_delete_constant_list(&This->baseShader.constantsI);
1153 list_remove(&This->baseShader.shader_list_entry);
1154 HeapFree(GetProcessHeap(), 0, This);
1155 }
1156 return ref;
1157 }
Maarten's multimedia branch