?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
wined3d: Move constant loading into target-specific files.
[wine.git] / dlls / wined3d / baseshader.c
blobf19a03bd9f0a5043163cef09b9a10449a01b3d80
1 /*
2 * shaders implementation
3 *
4 * Copyright 2002-2003 Jason Edmeades
5 * Copyright 2002-2003 Raphael Junqueira
6 * Copyright 2005 Oliver Stieber
7 * Copyright 2006 Ivan Gyurdiev
8 *
9 * This library is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
13 *
14 * This library is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
18 *
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with this library; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
22 */
23
24 #include "config.h"
25 #include <string.h>
26 #include <stdio.h>
27 #include "wined3d_private.h"
28
29 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
30
31 #define GLNAME_REQUIRE_GLSL ((const char *)1)
32
33 inline static BOOL shader_is_version_token(DWORD token) {
34 return shader_is_pshader_version(token) ||
35 shader_is_vshader_version(token);
36 }
37
38 int shader_addline(
39 SHADER_BUFFER* buffer,
40 const char *format, ...) {
41
42 char* base = buffer->buffer + buffer->bsize;
43 int rc;
44
45 va_list args;
46 va_start(args, format);
47 rc = vsnprintf(base, SHADER_PGMSIZE - 1 - buffer->bsize, format, args);
48 va_end(args);
49
50 if (rc < 0 || /* C89 */
51 rc > SHADER_PGMSIZE - 1 - buffer->bsize) { /* C99 */
52
53 ERR("The buffer allocated for the shader program string "
54 "is too small at %d bytes.\n", SHADER_PGMSIZE);
55 buffer->bsize = SHADER_PGMSIZE - 1;
56 return -1;
57 }
58
59 buffer->bsize += rc;
60 buffer->lineNo++;
61 TRACE("GL HW (%u, %u) : %s", buffer->lineNo, buffer->bsize, base);
62 return 0;
63 }
64
65 const SHADER_OPCODE* shader_get_opcode(
66 IWineD3DBaseShader *iface, const DWORD code) {
67
68 IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl*) iface;
69
70 DWORD i = 0;
71 DWORD hex_version = This->baseShader.hex_version;
72 const SHADER_OPCODE *shader_ins = This->baseShader.shader_ins;
73
74 /** TODO: use dichotomic search */
75 while (NULL != shader_ins[i].name) {
76 if (((code & D3DSI_OPCODE_MASK) == shader_ins[i].opcode) &&
77 (((hex_version >= shader_ins[i].min_version) && (hex_version <= shader_ins[i].max_version)) ||
78 ((shader_ins[i].min_version == 0) && (shader_ins[i].max_version == 0)))) {
79 return &shader_ins[i];
80 }
81 ++i;
82 }
83 FIXME("Unsupported opcode %#lx(%ld) masked %#lx, shader version %#lx\n",
84 code, code, code & D3DSI_OPCODE_MASK, hex_version);
85 return NULL;
86 }
87
88 /* Read a parameter opcode from the input stream,
89 * and possibly a relative addressing token.
90 * Return the number of tokens read */
91 int shader_get_param(
92 IWineD3DBaseShader* iface,
93 const DWORD* pToken,
94 DWORD* param,
95 DWORD* addr_token) {
96
97 /* PS >= 3.0 have relative addressing (with token)
98 * VS >= 2.0 have relative addressing (with token)
99 * VS >= 1.0 < 2.0 have relative addressing (without token)
100 * The version check below should work in general */
101
102 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
103 char rel_token = D3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 2 &&
104 ((*pToken & D3DSHADER_ADDRESSMODE_MASK) == D3DSHADER_ADDRMODE_RELATIVE);
105
106 *param = *pToken;
107 *addr_token = rel_token? *(pToken + 1): 0;
108 return rel_token? 2:1;
109 }
110
111 /* Return the number of parameters to skip for an opcode */
112 static inline int shader_skip_opcode(
113 IWineD3DBaseShaderImpl* This,
114 const SHADER_OPCODE* curOpcode,
115 DWORD opcode_token) {
116
117 /* Shaders >= 2.0 may contain address tokens, but fortunately they
118 * have a useful legnth mask - use it here. Shaders 1.0 contain no such tokens */
119
120 return (D3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 2)?
121 ((opcode_token & D3DSI_INSTLENGTH_MASK) >> D3DSI_INSTLENGTH_SHIFT):
122 curOpcode->num_params;
123 }
124
125 /* Read the parameters of an unrecognized opcode from the input stream
126 * Return the number of tokens read.
127 *
128 * Note: This function assumes source or destination token format.
129 * It will not work with specially-formatted tokens like DEF or DCL,
130 * but hopefully those would be recognized */
131
132 int shader_skip_unrecognized(
133 IWineD3DBaseShader* iface,
134 const DWORD* pToken) {
135
136 int tokens_read = 0;
137 int i = 0;
138
139 /* TODO: Think of a good name for 0x80000000 and replace it with a constant */
140 while (*pToken & 0x80000000) {
141
142 DWORD param, addr_token;
143 tokens_read += shader_get_param(iface, pToken, &param, &addr_token);
144 pToken += tokens_read;
145
146 FIXME("Unrecognized opcode param: token=%08lX "
147 "addr_token=%08lX name=", param, addr_token);
148 shader_dump_param(iface, param, addr_token, i);
149 FIXME("\n");
150 ++i;
151 }
152 return tokens_read;
153 }
154
155 /* Convert floating point offset relative
156 * to a register file to an absolute offset for float constants */
157
158 unsigned int shader_get_float_offset(const DWORD reg) {
159
160 unsigned int regnum = reg & D3DSP_REGNUM_MASK;
161 int regtype = shader_get_regtype(reg);
162
163 switch (regtype) {
164 case D3DSPR_CONST: return regnum;
165 case D3DSPR_CONST2: return 2048 + regnum;
166 case D3DSPR_CONST3: return 4096 + regnum;
167 case D3DSPR_CONST4: return 6144 + regnum;
168 default:
169 FIXME("Unsupported register type: %d\n", regtype);
170 return regnum;
171 }
172 }
173
174 static void shader_parse_decl_usage(
175 DWORD *semantics_map,
176 DWORD usage_token, DWORD param) {
177
178 unsigned int usage = (usage_token & D3DSP_DCL_USAGE_MASK) >> D3DSP_DCL_USAGE_SHIFT;
179 unsigned int usage_idx = (usage_token & D3DSP_DCL_USAGEINDEX_MASK) >> D3DSP_DCL_USAGEINDEX_SHIFT;
180 unsigned int regnum = param & D3DSP_REGNUM_MASK;
181
182 switch(usage) {
183 case D3DDECLUSAGE_POSITION:
184 if (usage_idx == 0) { /* tween data */
185 TRACE("Setting position to %d\n", regnum);
186 semantics_map[WINED3DSHADERDECLUSAGE_POSITION] = param;
187 } else {
188 /* TODO: position indexes go from 0-8!!*/
189 TRACE("Setting position 2 to %d because usage_idx = %d\n", regnum, usage_idx);
190 /* robots uses positions up to 8, the position arrays are just packed.*/
191 if (usage_idx > 1) {
192 TRACE("Loaded for position %d (greater than 2)\n", usage_idx);
193 }
194 semantics_map[WINED3DSHADERDECLUSAGE_POSITION2 + usage_idx-1] = param;
195 }
196 break;
197
198 case D3DDECLUSAGE_BLENDINDICES:
199 TRACE("Setting BLENDINDICES to %d\n", regnum);
200 semantics_map[WINED3DSHADERDECLUSAGE_BLENDINDICES] = param;
201 if (usage_idx != 0) FIXME("Extended BLENDINDICES\n");
202 break;
203
204 case D3DDECLUSAGE_BLENDWEIGHT:
205 TRACE("Setting BLENDWEIGHT to %d\n", regnum);
206 semantics_map[WINED3DSHADERDECLUSAGE_BLENDWEIGHT] = param;
207 if (usage_idx != 0) FIXME("Extended blend weights\n");
208 break;
209
210 case D3DDECLUSAGE_NORMAL:
211 if (usage_idx == 0) { /* tween data */
212 TRACE("Setting normal to %d\n", regnum);
213 semantics_map[WINED3DSHADERDECLUSAGE_NORMAL] = param;
214 } else {
215 TRACE("Setting normal 2 to %d because usage = %d\n", regnum, usage_idx);
216 semantics_map[WINED3DSHADERDECLUSAGE_NORMAL2] = param;
217 }
218 break;
219
220 case D3DDECLUSAGE_PSIZE:
221 TRACE("Setting PSIZE to %d\n", regnum);
222 semantics_map[WINED3DSHADERDECLUSAGE_PSIZE] = param;
223 if (usage_idx != 0) FIXME("Extended PSIZE\n");
224 break;
225
226 case D3DDECLUSAGE_COLOR:
227 if (usage_idx == 0) {
228 TRACE("Setting DIFFUSE to %d\n", regnum);
229 semantics_map[WINED3DSHADERDECLUSAGE_DIFFUSE] = param;
230 } else {
231 TRACE("Setting SPECULAR to %d\n", regnum);
232 semantics_map[WINED3DSHADERDECLUSAGE_SPECULAR] = param;
233 }
234 break;
235
236 case D3DDECLUSAGE_TEXCOORD:
237 if (usage_idx > 7) {
238 FIXME("Program uses texture coordinate %d but only 0-7 have been "
239 "implemented\n", usage_idx);
240 } else {
241 TRACE("Setting TEXCOORD %d to %d\n", usage_idx, regnum);
242 semantics_map[WINED3DSHADERDECLUSAGE_TEXCOORD0 + usage_idx] = param;
243 }
244 break;
245
246 case D3DDECLUSAGE_TANGENT:
247 TRACE("Setting TANGENT to %d\n", regnum);
248 semantics_map[WINED3DSHADERDECLUSAGE_TANGENT] = param;
249 break;
250
251 case D3DDECLUSAGE_BINORMAL:
252 TRACE("Setting BINORMAL to %d\n", regnum);
253 semantics_map[WINED3DSHADERDECLUSAGE_BINORMAL] = param;
254 break;
255
256 case D3DDECLUSAGE_TESSFACTOR:
257 TRACE("Setting TESSFACTOR to %d\n", regnum);
258 semantics_map[WINED3DSHADERDECLUSAGE_TESSFACTOR] = param;
259 break;
260
261 case D3DDECLUSAGE_POSITIONT:
262 if (usage_idx == 0) { /* tween data */
263 FIXME("Setting positiont to %d\n", regnum);
264 semantics_map[WINED3DSHADERDECLUSAGE_POSITIONT] = param;
265 } else {
266 FIXME("Setting positiont 2 to %d because usage = %d\n", regnum, usage_idx);
267 semantics_map[WINED3DSHADERDECLUSAGE_POSITIONT2] = param;
268 if (usage_idx != 0) FIXME("Extended positiont\n");
269 }
270 break;
271
272 case D3DDECLUSAGE_FOG:
273 TRACE("Setting FOG to %d\n", regnum);
274 semantics_map[WINED3DSHADERDECLUSAGE_FOG] = param;
275 break;
276
277 case D3DDECLUSAGE_DEPTH:
278 TRACE("Setting DEPTH to %d\n", regnum);
279 semantics_map[WINED3DSHADERDECLUSAGE_DEPTH] = param;
280 break;
281
282 case D3DDECLUSAGE_SAMPLE:
283 TRACE("Setting SAMPLE to %d\n", regnum);
284 semantics_map[WINED3DSHADERDECLUSAGE_SAMPLE] = param;
285 break;
286
287 default:
288 FIXME("Unrecognised dcl %#x", usage);
289 }
290 }
291
292 /* Note that this does not count the loop register
293 * as an address register. */
294
295 void shader_get_registers_used(
296 IWineD3DBaseShader *iface,
297 shader_reg_maps* reg_maps,
298 CONST DWORD* pToken) {
299
300 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
301
302 /* There are some minor differences between pixel and vertex shaders */
303 char pshader = shader_is_pshader_version(This->baseShader.hex_version);
304
305 if (pToken == NULL)
306 return;
307
308 while (D3DVS_END() != *pToken) {
309 CONST SHADER_OPCODE* curOpcode;
310 DWORD opcode_token;
311
312 /* Skip version */
313 if (shader_is_version_token(*pToken)) {
314 ++pToken;
315 continue;
316
317 /* Skip comments */
318 } else if (shader_is_comment(*pToken)) {
319 DWORD comment_len = (*pToken & D3DSI_COMMENTSIZE_MASK) >> D3DSI_COMMENTSIZE_SHIFT;
320 ++pToken;
321 pToken += comment_len;
322 continue;
323 }
324
325 /* Fetch opcode */
326 opcode_token = *pToken++;
327 curOpcode = shader_get_opcode(iface, opcode_token);
328
329 /* Unhandled opcode, and its parameters */
330 if (NULL == curOpcode) {
331 while (*pToken & 0x80000000)
332 ++pToken;
333
334 /* Handle declarations */
335 } else if (D3DSIO_DCL == curOpcode->opcode) {
336
337 DWORD usage = *pToken++;
338 DWORD param = *pToken++;
339 DWORD regtype = shader_get_regtype(param);
340 unsigned int regnum = param & D3DSP_REGNUM_MASK;
341
342 /* Vshader: mark attributes used
343 Pshader: mark 3.0 input registers used, save token */
344 if (D3DSPR_INPUT == regtype) {
345
346 if (!pshader)
347 reg_maps->attributes[regnum] = 1;
348 else
349 reg_maps->packed_input[regnum] = 1;
350
351 shader_parse_decl_usage(reg_maps->semantics_in, usage, param);
352
353 /* Vshader: mark 3.0 output registers used, save token */
354 } else if (D3DSPR_OUTPUT == regtype) {
355 reg_maps->packed_output[regnum] = 1;
356 shader_parse_decl_usage(reg_maps->semantics_out, usage, param);
357
358 /* Save sampler usage token */
359 } else if (D3DSPR_SAMPLER == regtype)
360 reg_maps->samplers[regnum] = usage;
361
362 /* Skip definitions (for now) */
363 } else if (D3DSIO_DEF == curOpcode->opcode) {
364 pToken += curOpcode->num_params;
365
366 /* Set texture, address, temporary registers */
367 } else {
368 int i, limit;
369
370 /* Declare 1.X samplers implicitly, based on the destination reg. number */
371 if (D3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) == 1 &&
372 (D3DSIO_TEX == curOpcode->opcode ||
373 D3DSIO_TEXM3x2TEX == curOpcode->opcode ||
374 D3DSIO_TEXM3x3TEX == curOpcode->opcode ||
375 D3DSIO_TEXM3x3SPEC == curOpcode->opcode ||
376 D3DSIO_TEXM3x3VSPEC == curOpcode->opcode)) {
377
378 /* Fake sampler usage, only set reserved bit and ttype */
379 DWORD sampler_code = *pToken & D3DSP_REGNUM_MASK;
380 reg_maps->samplers[sampler_code] = (0x1 << 31) | D3DSTT_2D;
381 }
382
383 /* This will loop over all the registers and try to
384 * make a bitmask of the ones we're interested in.
385 *
386 * Relative addressing tokens are ignored, but that's
387 * okay, since we'll catch any address registers when
388 * they are initialized (required by spec) */
389
390 limit = (opcode_token & D3DSHADER_INSTRUCTION_PREDICATED)?
391 curOpcode->num_params + 1: curOpcode->num_params;
392
393 for (i = 0; i < limit; ++i) {
394
395 DWORD param, addr_token, reg, regtype;
396 pToken += shader_get_param(iface, pToken, &param, &addr_token);
397
398 regtype = (param & D3DSP_REGTYPE_MASK) >> D3DSP_REGTYPE_SHIFT;
399 reg = param & D3DSP_REGNUM_MASK;
400
401 if (D3DSPR_TEXTURE == regtype) { /* vs: D3DSPR_ADDR */
402
403 if (pshader)
404 reg_maps->texcoord[reg] = 1;
405 else
406 reg_maps->address[reg] = 1;
407 }
408
409 else if (D3DSPR_TEMP == regtype)
410 reg_maps->temporary[reg] = 1;
411
412 else if (D3DSPR_INPUT == regtype && !pshader)
413 reg_maps->attributes[reg] = 1;
414
415 else if (D3DSPR_LOOP == regtype)
416 reg_maps->loop = 1;
417 }
418 }
419 }
420 }
421
422 static void shader_dump_decl_usage(
423 DWORD decl,
424 DWORD param) {
425
426 DWORD regtype = shader_get_regtype(param);
427 TRACE("dcl_");
428
429 if (regtype == D3DSPR_SAMPLER) {
430 DWORD ttype = decl & D3DSP_TEXTURETYPE_MASK;
431
432 switch (ttype) {
433 case D3DSTT_2D: TRACE("2d"); break;
434 case D3DSTT_CUBE: TRACE("cube"); break;
435 case D3DSTT_VOLUME: TRACE("volume"); break;
436 default: TRACE("unknown_ttype(%08lx)", ttype);
437 }
438
439 } else {
440
441 DWORD usage = decl & D3DSP_DCL_USAGE_MASK;
442 DWORD idx = (decl & D3DSP_DCL_USAGEINDEX_MASK) >> D3DSP_DCL_USAGEINDEX_SHIFT;
443
444 switch(usage) {
445 case D3DDECLUSAGE_POSITION:
446 TRACE("%s%ld", "position", idx);
447 break;
448 case D3DDECLUSAGE_BLENDINDICES:
449 TRACE("%s", "blend");
450 break;
451 case D3DDECLUSAGE_BLENDWEIGHT:
452 TRACE("%s", "weight");
453 break;
454 case D3DDECLUSAGE_NORMAL:
455 TRACE("%s%ld", "normal", idx);
456 break;
457 case D3DDECLUSAGE_PSIZE:
458 TRACE("%s", "psize");
459 break;
460 case D3DDECLUSAGE_COLOR:
461 if(idx == 0) {
462 TRACE("%s", "color");
463 } else {
464 TRACE("%s%ld", "specular", (idx - 1));
465 }
466 break;
467 case D3DDECLUSAGE_TEXCOORD:
468 TRACE("%s%ld", "texture", idx);
469 break;
470 case D3DDECLUSAGE_TANGENT:
471 TRACE("%s", "tangent");
472 break;
473 case D3DDECLUSAGE_BINORMAL:
474 TRACE("%s", "binormal");
475 break;
476 case D3DDECLUSAGE_TESSFACTOR:
477 TRACE("%s", "tessfactor");
478 break;
479 case D3DDECLUSAGE_POSITIONT:
480 TRACE("%s%ld", "positionT", idx);
481 break;
482 case D3DDECLUSAGE_FOG:
483 TRACE("%s", "fog");
484 break;
485 case D3DDECLUSAGE_DEPTH:
486 TRACE("%s", "depth");
487 break;
488 case D3DDECLUSAGE_SAMPLE:
489 TRACE("%s", "sample");
490 break;
491 default:
492 FIXME("unknown_semantics(%08lx)", usage);
493 }
494 }
495 }
496
497 static void shader_dump_arr_entry(
498 IWineD3DBaseShader *iface,
499 const DWORD param,
500 const DWORD addr_token,
501 unsigned int reg,
502 int input) {
503
504 char relative =
505 ((param & D3DSHADER_ADDRESSMODE_MASK) == D3DSHADER_ADDRMODE_RELATIVE);
506
507 if (relative) {
508 TRACE("[");
509 if (addr_token)
510 shader_dump_param(iface, addr_token, 0, input);
511 else
512 TRACE("a0.x");
513 TRACE(" + ");
514 }
515 TRACE("%u", reg);
516 if (relative)
517 TRACE("]");
518 }
519
520 void shader_dump_param(
521 IWineD3DBaseShader *iface,
522 const DWORD param,
523 const DWORD addr_token,
524 int input) {
525
526 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
527 static const char* rastout_reg_names[] = { "oPos", "oFog", "oPts" };
528 char swizzle_reg_chars[4];
529
530 DWORD reg = param & D3DSP_REGNUM_MASK;
531 DWORD regtype = shader_get_regtype(param);
532
533 /* There are some minor differences between pixel and vertex shaders */
534 char pshader = shader_is_pshader_version(This->baseShader.hex_version);
535
536 /* For one, we'd prefer color components to be shown for pshaders.
537 * FIXME: use the swizzle function for this */
538
539 swizzle_reg_chars[0] = pshader? 'r': 'x';
540 swizzle_reg_chars[1] = pshader? 'g': 'y';
541 swizzle_reg_chars[2] = pshader? 'b': 'z';
542 swizzle_reg_chars[3] = pshader? 'a': 'w';
543
544 if (input) {
545 if ( ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_NEG) ||
546 ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_BIASNEG) ||
547 ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_SIGNNEG) ||
548 ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_X2NEG) )
549 TRACE("-");
550 else if ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_COMP)
551 TRACE("1-");
552 else if ((param & D3DSP_SRCMOD_MASK) == D3DSPSM_NOT)
553 TRACE("!");
554 }
555
556 switch (regtype) {
557 case D3DSPR_TEMP:
558 TRACE("r%lu", reg);
559 break;
560 case D3DSPR_INPUT:
561 TRACE("v");
562 shader_dump_arr_entry(iface, param, addr_token, reg, input);
563 break;
564 case D3DSPR_CONST:
565 case D3DSPR_CONST2:
566 case D3DSPR_CONST3:
567 case D3DSPR_CONST4:
568 TRACE("c");
569 shader_dump_arr_entry(iface, param, addr_token, shader_get_float_offset(param), input);
570 break;
571 case D3DSPR_TEXTURE: /* vs: case D3DSPR_ADDR */
572 TRACE("%c%lu", (pshader? 't':'a'), reg);
573 break;
574 case D3DSPR_RASTOUT:
575 TRACE("%s", rastout_reg_names[reg]);
576 break;
577 case D3DSPR_COLOROUT:
578 TRACE("oC%lu", reg);
579 break;
580 case D3DSPR_DEPTHOUT:
581 TRACE("oDepth");
582 break;
583 case D3DSPR_ATTROUT:
584 TRACE("oD%lu", reg);
585 break;
586 case D3DSPR_TEXCRDOUT:
587
588 /* Vertex shaders >= 3.0 use general purpose output registers
589 * (D3DSPR_OUTPUT), which can include an address token */
590
591 if (D3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3) {
592 TRACE("o");
593 shader_dump_arr_entry(iface, param, addr_token, reg, input);
594 }
595 else
596 TRACE("oT%lu", reg);
597 break;
598 case D3DSPR_CONSTINT:
599 TRACE("i");
600 shader_dump_arr_entry(iface, param, addr_token, reg, input);
601 break;
602 case D3DSPR_CONSTBOOL:
603 TRACE("b");
604 shader_dump_arr_entry(iface, param, addr_token, reg, input);
605 break;
606 case D3DSPR_LABEL:
607 TRACE("l%lu", reg);
608 break;
609 case D3DSPR_LOOP:
610 TRACE("aL");
611 break;
612 case D3DSPR_SAMPLER:
613 TRACE("s%lu", reg);
614 break;
615 case D3DSPR_PREDICATE:
616 TRACE("p%lu", reg);
617 break;
618 default:
619 TRACE("unhandled_rtype(%#lx)", regtype);
620 break;
621 }
622
623 if (!input) {
624 /* operand output (for modifiers and shift, see dump_ins_modifiers) */
625
626 if ((param & D3DSP_WRITEMASK_ALL) != D3DSP_WRITEMASK_ALL) {
627 TRACE(".");
628 if (param & D3DSP_WRITEMASK_0) TRACE("%c", swizzle_reg_chars[0]);
629 if (param & D3DSP_WRITEMASK_1) TRACE("%c", swizzle_reg_chars[1]);
630 if (param & D3DSP_WRITEMASK_2) TRACE("%c", swizzle_reg_chars[2]);
631 if (param & D3DSP_WRITEMASK_3) TRACE("%c", swizzle_reg_chars[3]);
632 }
633
634 } else {
635 /** operand input */
636 DWORD swizzle = (param & D3DSP_SWIZZLE_MASK) >> D3DSP_SWIZZLE_SHIFT;
637 DWORD swizzle_r = swizzle & 0x03;
638 DWORD swizzle_g = (swizzle >> 2) & 0x03;
639 DWORD swizzle_b = (swizzle >> 4) & 0x03;
640 DWORD swizzle_a = (swizzle >> 6) & 0x03;
641
642 if (0 != (param & D3DSP_SRCMOD_MASK)) {
643 DWORD mask = param & D3DSP_SRCMOD_MASK;
644 switch (mask) {
645 case D3DSPSM_NONE: break;
646 case D3DSPSM_NEG: break;
647 case D3DSPSM_NOT: break;
648 case D3DSPSM_BIAS: TRACE("_bias"); break;
649 case D3DSPSM_BIASNEG: TRACE("_bias"); break;
650 case D3DSPSM_SIGN: TRACE("_bx2"); break;
651 case D3DSPSM_SIGNNEG: TRACE("_bx2"); break;
652 case D3DSPSM_COMP: break;
653 case D3DSPSM_X2: TRACE("_x2"); break;
654 case D3DSPSM_X2NEG: TRACE("_x2"); break;
655 case D3DSPSM_DZ: TRACE("_dz"); break;
656 case D3DSPSM_DW: TRACE("_dw"); break;
657 default:
658 TRACE("_unknown_modifier(%#lx)", mask >> D3DSP_SRCMOD_SHIFT);
659 }
660 }
661
662 /**
663 * swizzle bits fields:
664 * RRGGBBAA
665 */
666 if ((D3DVS_NOSWIZZLE >> D3DVS_SWIZZLE_SHIFT) != swizzle) { /* ! D3DVS_NOSWIZZLE == 0xE4 << D3DVS_SWIZZLE_SHIFT */
667 if (swizzle_r == swizzle_g &&
668 swizzle_r == swizzle_b &&
669 swizzle_r == swizzle_a) {
670 TRACE(".%c", swizzle_reg_chars[swizzle_r]);
671 } else {
672 TRACE(".%c%c%c%c",
673 swizzle_reg_chars[swizzle_r],
674 swizzle_reg_chars[swizzle_g],
675 swizzle_reg_chars[swizzle_b],
676 swizzle_reg_chars[swizzle_a]);
677 }
678 }
679 }
680 }
681
682 /* Generate the variable & register declarations for the ARB_vertex_program output target */
683 void shader_generate_arb_declarations(
684 IWineD3DBaseShader *iface,
685 shader_reg_maps* reg_maps,
686 SHADER_BUFFER* buffer) {
687
688 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
689 DWORD i;
690
691 for(i = 0; i < This->baseShader.limits.temporary; i++) {
692 if (reg_maps->temporary[i])
693 shader_addline(buffer, "TEMP R%lu;\n", i);
694 }
695
696 for (i = 0; i < This->baseShader.limits.address; i++) {
697 if (reg_maps->address[i])
698 shader_addline(buffer, "ADDRESS A%ld;\n", i);
699 }
700
701 for(i = 0; i < This->baseShader.limits.texcoord; i++) {
702 if (reg_maps->texcoord[i])
703 shader_addline(buffer,"TEMP T%lu;\n", i);
704 }
705
706 /* Texture coordinate registers must be pre-loaded */
707 for (i = 0; i < This->baseShader.limits.texcoord; i++) {
708 if (reg_maps->texcoord[i])
709 shader_addline(buffer, "MOV T%lu, fragment.texcoord[%lu];\n", i, i);
710 }
711
712 /* Need to PARAM the environment parameters (constants) so we can use relative addressing */
713 shader_addline(buffer, "PARAM C[%d] = { program.env[0..%d] };\n",
714 This->baseShader.limits.constant_float,
715 This->baseShader.limits.constant_float - 1);
716 }
717
718 /** Generate the variable & register declarations for the GLSL output target */
719 void shader_generate_glsl_declarations(
720 IWineD3DBaseShader *iface,
721 shader_reg_maps* reg_maps,
722 SHADER_BUFFER* buffer) {
723
724 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
725 int i;
726
727 /* There are some minor differences between pixel and vertex shaders */
728 char pshader = shader_is_pshader_version(This->baseShader.hex_version);
729
730 /* Declare the constants (aka uniforms) */
731 if (pshader)
732 shader_addline(buffer, "uniform vec4 PC[%u];\n", This->baseShader.limits.constant_float);
733 else
734 shader_addline(buffer, "uniform vec4 VC[%u];\n", This->baseShader.limits.constant_float);
735
736 /* Declare texture samplers */
737 for (i = 0; i < This->baseShader.limits.sampler; i++) {
738 if (reg_maps->samplers[i]) {
739
740 const char* prefix = pshader? "psampler": "vsampler";
741
742 DWORD stype = reg_maps->samplers[i] & D3DSP_TEXTURETYPE_MASK;
743 switch (stype) {
744
745 case D3DSTT_2D:
746 shader_addline(buffer, "uniform sampler2D %s%lu;\n", prefix, i);
747 break;
748 case D3DSTT_CUBE:
749 shader_addline(buffer, "uniform samplerCube %s%lu;\n", prefix, i);
750 break;
751 case D3DSTT_VOLUME:
752 shader_addline(buffer, "uniform sampler3D %s%lu;\n", prefix, i);
753 break;
754 default:
755 shader_addline(buffer, "uniform unsupported_sampler %s%lu;\n", prefix, i);
756 FIXME("Unrecognized sampler type: %#lx\n", stype);
757 break;
758 }
759 }
760 }
761
762 /* Declare address variables */
763 for (i = 0; i < This->baseShader.limits.address; i++) {
764 if (reg_maps->address[i])
765 shader_addline(buffer, "ivec4 A%ld;\n", i);
766 }
767
768 /* Declare texture coordinate temporaries and initialize them */
769 for (i = 0; i < This->baseShader.limits.texcoord; i++) {
770 if (reg_maps->texcoord[i])
771 shader_addline(buffer, "vec4 T%lu = gl_TexCoord[%lu];\n", i, i);
772 }
773
774 /* Declare input register temporaries */
775 for (i=0; i < This->baseShader.limits.packed_input; i++) {
776 if (reg_maps->packed_input[i])
777 shader_addline(buffer, "vec4 IN%lu;\n", i);
778 }
779
780 /* Declare output register temporaries */
781 for (i = 0; i < This->baseShader.limits.packed_output; i++) {
782 if (reg_maps->packed_output[i])
783 shader_addline(buffer, "vec4 OUT%lu;\n", i);
784 }
785
786 /* Declare temporary variables */
787 for(i = 0; i < This->baseShader.limits.temporary; i++) {
788 if (reg_maps->temporary[i])
789 shader_addline(buffer, "vec4 R%lu;\n", i);
790 }
791
792 /* Declare attributes */
793 for (i = 0; i < This->baseShader.limits.attributes; i++) {
794 if (reg_maps->attributes[i])
795 shader_addline(buffer, "attribute vec4 attrib%i;\n", i);
796 }
797
798 /* Declare loop register aL */
799 if (reg_maps->loop)
800 shader_addline(buffer, "int aL;\n");
801
802 /* Temporary variables for matrix operations */
803 shader_addline(buffer, "vec4 tmp0;\n");
804 shader_addline(buffer, "vec4 tmp1;\n");
805
806 /* Start the main program */
807 shader_addline(buffer, "void main() {\n");
808 }
809
810 /** Shared code in order to generate the bulk of the shader string.
811 Use the shader_header_fct & shader_footer_fct to add strings
812 that are specific to pixel or vertex functions
813 NOTE: A description of how to parse tokens can be found at:
814 http://msdn.microsoft.com/library/default.asp?url=/library/en-us/graphics/hh/graphics/usermodedisplaydriver_shader_cc8e4e05-f5c3-4ec0-8853-8ce07c1551b2.xml.asp */
815 void shader_generate_main(
816 IWineD3DBaseShader *iface,
817 SHADER_BUFFER* buffer,
818 shader_reg_maps* reg_maps,
819 CONST DWORD* pFunction) {
820
821 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
822 const DWORD *pToken = pFunction;
823 const SHADER_OPCODE *curOpcode = NULL;
824 SHADER_HANDLER hw_fct = NULL;
825 DWORD opcode_token;
826 DWORD i;
827 SHADER_OPCODE_ARG hw_arg;
828
829 /* Initialize current parsing state */
830 hw_arg.shader = iface;
831 hw_arg.buffer = buffer;
832 hw_arg.reg_maps = reg_maps;
833 This->baseShader.parse_state.current_row = 0;
834
835 /* Second pass, process opcodes */
836 if (NULL != pToken) {
837 while (D3DPS_END() != *pToken) {
838
839 /* Skip version token */
840 if (shader_is_version_token(*pToken)) {
841 ++pToken;
842 continue;
843 }
844
845 /* Skip comment tokens */
846 if (shader_is_comment(*pToken)) {
847 DWORD comment_len = (*pToken & D3DSI_COMMENTSIZE_MASK) >> D3DSI_COMMENTSIZE_SHIFT;
848 ++pToken;
849 TRACE("#%s\n", (char*)pToken);
850 pToken += comment_len;
851 continue;
852 }
853
854 /* Read opcode */
855 opcode_token = *pToken++;
856 curOpcode = shader_get_opcode(iface, opcode_token);
857 hw_fct = (curOpcode == NULL)? NULL:
858 (wined3d_settings.shader_mode == SHADER_GLSL)?
859 curOpcode->hw_glsl_fct : curOpcode->hw_fct;
860
861 /* Unknown opcode and its parameters */
862 if (NULL == curOpcode) {
863 FIXME("Unrecognized opcode: token=%08lX\n", opcode_token);
864 pToken += shader_skip_unrecognized(iface, pToken);
865
866 /* Nothing to do */
867 } else if (D3DSIO_DCL == curOpcode->opcode ||
868 D3DSIO_NOP == curOpcode->opcode) {
869
870 pToken += shader_skip_opcode(This, curOpcode, opcode_token);
871
872 /* If a generator function is set for current shader target, use it */
873 } else if (hw_fct != NULL) {
874
875 hw_arg.opcode = curOpcode;
876
877 /* Destination token */
878 if (curOpcode->dst_token) {
879
880 DWORD param, addr_token = 0;
881 pToken += shader_get_param(iface, pToken, &param, &addr_token);
882 hw_arg.dst = param;
883 hw_arg.dst_addr = addr_token;
884 }
885
886 /* Predication token */
887 if (opcode_token & D3DSHADER_INSTRUCTION_PREDICATED)
888 hw_arg.predicate = *pToken++;
889
890 /* Other source tokens */
891 for (i = curOpcode->dst_token; i < curOpcode->num_params; i++) {
892
893 DWORD param, addr_token = 0;
894
895 /* DEF* instructions have constant src parameters, not registers */
896 if (curOpcode->opcode == D3DSIO_DEF ||
897 curOpcode->opcode == D3DSIO_DEFI ||
898 curOpcode->opcode == D3DSIO_DEFB) {
899 param = *pToken++;
900
901 } else
902 pToken += shader_get_param(iface, pToken, &param, &addr_token);
903
904 hw_arg.src[i-1] = param;
905 hw_arg.src_addr[i-1] = addr_token;
906 }
907
908 /* Call appropriate function for output target */
909 hw_fct(&hw_arg);
910
911 /* Process instruction modifiers for GLSL apps ( _sat, etc. ) */
912 if (wined3d_settings.shader_mode == SHADER_GLSL)
913 shader_glsl_add_instruction_modifiers(&hw_arg);
914
915 /* Unhandled opcode */
916 } else {
917
918 FIXME("Can't handle opcode %s in hwShader\n", curOpcode->name);
919 pToken += shader_skip_opcode(This, curOpcode, opcode_token);
920 }
921 }
922 /* TODO: What about result.depth? */
923
924 }
925 }
926
927 void shader_dump_ins_modifiers(const DWORD output) {
928
929 DWORD shift = (output & D3DSP_DSTSHIFT_MASK) >> D3DSP_DSTSHIFT_SHIFT;
930 DWORD mmask = output & D3DSP_DSTMOD_MASK;
931
932 switch (shift) {
933 case 0: break;
934 case 13: TRACE("_d8"); break;
935 case 14: TRACE("_d4"); break;
936 case 15: TRACE("_d2"); break;
937 case 1: TRACE("_x2"); break;
938 case 2: TRACE("_x4"); break;
939 case 3: TRACE("_x8"); break;
940 default: TRACE("_unhandled_shift(%ld)", shift); break;
941 }
942
943 if (mmask & D3DSPDM_SATURATE) TRACE("_sat");
944 if (mmask & D3DSPDM_PARTIALPRECISION) TRACE("_pp");
945 if (mmask & D3DSPDM_MSAMPCENTROID) TRACE("_centroid");
946
947 mmask &= ~(D3DSPDM_SATURATE | D3DSPDM_PARTIALPRECISION | D3DSPDM_MSAMPCENTROID);
948 if (mmask)
949 FIXME("_unrecognized_modifier(%#lx)", mmask >> D3DSP_DSTMOD_SHIFT);
950 }
951
952 /** Process the D3DSIO_DEF opcode into an ARB string - creates a local vec4
953 * float constant, and stores it's usage on the regmaps. */
954 void shader_hw_def(SHADER_OPCODE_ARG* arg) {
955
956 DWORD reg = arg->dst & D3DSP_REGNUM_MASK;
957
958 shader_addline(arg->buffer,
959 "PARAM C%lu = { %f, %f, %f, %f };\n", reg,
960 *((const float *)(arg->src + 0)),
961 *((const float *)(arg->src + 1)),
962 *((const float *)(arg->src + 2)),
963 *((const float *)(arg->src + 3)) );
964
965 arg->reg_maps->constantsF[reg] = 1;
966 }
967
968 /* First pass: trace shader, initialize length and version */
969 void shader_trace_init(
970 IWineD3DBaseShader *iface,
971 const DWORD* pFunction) {
972
973 IWineD3DBaseShaderImpl *This =(IWineD3DBaseShaderImpl *)iface;
974
975 const DWORD* pToken = pFunction;
976 const SHADER_OPCODE* curOpcode = NULL;
977 DWORD opcode_token;
978 unsigned int len = 0;
979 DWORD i;
980
981 TRACE("(%p) : Parsing programme\n", This);
982
983 if (NULL != pToken) {
984 while (D3DVS_END() != *pToken) {
985 if (shader_is_version_token(*pToken)) { /** version */
986 This->baseShader.hex_version = *pToken;
987 TRACE("%s_%lu_%lu\n", shader_is_pshader_version(This->baseShader.hex_version)? "ps": "vs",
988 D3DSHADER_VERSION_MAJOR(This->baseShader.hex_version),
989 D3DSHADER_VERSION_MINOR(This->baseShader.hex_version));
990 ++pToken;
991 ++len;
992 continue;
993 }
994 if (shader_is_comment(*pToken)) { /** comment */
995 DWORD comment_len = (*pToken & D3DSI_COMMENTSIZE_MASK) >> D3DSI_COMMENTSIZE_SHIFT;
996 ++pToken;
997 TRACE("//%s\n", (char*)pToken);
998 pToken += comment_len;
999 len += comment_len + 1;
1000 continue;
1001 }
1002 opcode_token = *pToken++;
1003 curOpcode = shader_get_opcode(iface, opcode_token);
1004 len++;
1005
1006 if (NULL == curOpcode) {
1007 int tokens_read;
1008 FIXME("Unrecognized opcode: token=%08lX\n", opcode_token);
1009 tokens_read = shader_skip_unrecognized(iface, pToken);
1010 pToken += tokens_read;
1011 len += tokens_read;
1012
1013 } else {
1014 if (curOpcode->opcode == D3DSIO_DCL) {
1015
1016 DWORD usage = *pToken;
1017 DWORD param = *(pToken + 1);
1018
1019 shader_dump_decl_usage(usage, param);
1020 shader_dump_ins_modifiers(param);
1021 TRACE(" ");
1022 shader_dump_param(iface, param, 0, 0);
1023 pToken += 2;
1024 len += 2;
1025
1026 } else if (curOpcode->opcode == D3DSIO_DEF) {
1027
1028 unsigned int offset = shader_get_float_offset(*pToken);
1029
1030 TRACE("def c%u = %f, %f, %f, %f", offset,
1031 *(float *)(pToken + 1),
1032 *(float *)(pToken + 2),
1033 *(float *)(pToken + 3),
1034 *(float *)(pToken + 4));
1035
1036 pToken += 5;
1037 len += 5;
1038 } else if (curOpcode->opcode == D3DSIO_DEFI) {
1039
1040 TRACE("defi i%lu = %ld, %ld, %ld, %ld", *pToken & D3DSP_REGNUM_MASK,
1041 (long) *(pToken + 1),
1042 (long) *(pToken + 2),
1043 (long) *(pToken + 3),
1044 (long) *(pToken + 4));
1045
1046 pToken += 5;
1047 len += 5;
1048
1049 } else if (curOpcode->opcode == D3DSIO_DEFB) {
1050
1051 TRACE("defb b%lu = %s", *pToken & D3DSP_REGNUM_MASK,
1052 *(pToken + 1)? "true": "false");
1053
1054 pToken += 2;
1055 len += 2;
1056
1057 } else {
1058
1059 DWORD param, addr_token;
1060 int tokens_read;
1061
1062 /* Print out predication source token first - it follows
1063 * the destination token. */
1064 if (opcode_token & D3DSHADER_INSTRUCTION_PREDICATED) {
1065 TRACE("(");
1066 shader_dump_param(iface, *(pToken + 2), 0, 1);
1067 TRACE(") ");
1068 }
1069
1070 TRACE("%s", curOpcode->name);
1071
1072 /* Destination token */
1073 if (curOpcode->dst_token) {
1074
1075 /* Destination token */
1076 tokens_read = shader_get_param(iface, pToken, &param, &addr_token);
1077 pToken += tokens_read;
1078 len += tokens_read;
1079
1080 shader_dump_ins_modifiers(param);
1081 TRACE(" ");
1082 shader_dump_param(iface, param, addr_token, 0);
1083 }
1084
1085 /* Predication token - already printed out, just skip it */
1086 if (opcode_token & D3DSHADER_INSTRUCTION_PREDICATED) {
1087 pToken++;
1088 len++;
1089 }
1090
1091 /* Other source tokens */
1092 for (i = curOpcode->dst_token; i < curOpcode->num_params; ++i) {
1093
1094 tokens_read = shader_get_param(iface, pToken, &param, &addr_token);
1095 pToken += tokens_read;
1096 len += tokens_read;
1097
1098 TRACE((i == 0)? " " : ", ");
1099 shader_dump_param(iface, param, addr_token, 1);
1100 }
1101 }
1102 TRACE("\n");
1103 }
1104 }
1105 This->baseShader.functionLength = (len + 1) * sizeof(DWORD);
1106 } else {
1107 This->baseShader.functionLength = 1; /* no Function defined use fixed function vertex processing */
1108 }
1109 }
1110
1111 /* TODO: Move other shared code here */
Windows API implementation