push a9ee6bdc492dbf6a3d2de76ff889abeabd04e8d0
[wine/hacks.git] / dlls / wined3d / baseshader.c
blobc7931eec4f5d46c5581399c728cd2cb40c5b630f
1 /*
2 * shaders implementation
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 * Copyright 2007-2008 Stefan Dösinger for CodeWeavers
11 * This library is free software; you can redistribute it and/or
12 * modify it under the terms of the GNU Lesser General Public
13 * License as published by the Free Software Foundation; either
14 * version 2.1 of the License, or (at your option) any later version.
16 * This library is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * Lesser General Public License for more details.
21 * You should have received a copy of the GNU Lesser General Public
22 * License along with this library; if not, write to the Free Software
23 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
26 #include "config.h"
27 #include <string.h>
28 #include <stdio.h>
29 #include "wined3d_private.h"
31 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
32 WINE_DECLARE_DEBUG_CHANNEL(d3d);
34 static void shader_dump_param(const DWORD param, const DWORD addr_token, int input, DWORD shader_version);
36 static inline BOOL shader_is_version_token(DWORD token) {
37 return shader_is_pshader_version(token) ||
38 shader_is_vshader_version(token);
41 void shader_buffer_init(struct SHADER_BUFFER *buffer)
43 buffer->buffer = HeapAlloc(GetProcessHeap(), 0, SHADER_PGMSIZE);
44 buffer->buffer[0] = '\0';
45 buffer->bsize = 0;
46 buffer->lineNo = 0;
47 buffer->newline = TRUE;
50 void shader_buffer_free(struct SHADER_BUFFER *buffer)
52 HeapFree(GetProcessHeap(), 0, buffer->buffer);
55 int shader_vaddline(SHADER_BUFFER* buffer, const char *format, va_list args)
57 char* base = buffer->buffer + buffer->bsize;
58 int rc;
60 rc = vsnprintf(base, SHADER_PGMSIZE - 1 - buffer->bsize, format, args);
62 if (rc < 0 || /* C89 */
63 rc > SHADER_PGMSIZE - 1 - buffer->bsize) { /* C99 */
65 ERR("The buffer allocated for the shader program string "
66 "is too small at %d bytes.\n", SHADER_PGMSIZE);
67 buffer->bsize = SHADER_PGMSIZE - 1;
68 return -1;
71 if (buffer->newline) {
72 TRACE("GL HW (%u, %u) : %s", buffer->lineNo + 1, buffer->bsize, base);
73 buffer->newline = FALSE;
74 } else {
75 TRACE("%s", base);
78 buffer->bsize += rc;
79 if (buffer->buffer[buffer->bsize-1] == '\n') {
80 buffer->lineNo++;
81 buffer->newline = TRUE;
83 return 0;
86 int shader_addline(SHADER_BUFFER* buffer, const char *format, ...)
88 int ret;
89 va_list args;
91 va_start(args, format);
92 ret = shader_vaddline(buffer, format, args);
93 va_end(args);
95 return ret;
98 void shader_init(struct IWineD3DBaseShaderClass *shader,
99 IWineD3DDevice *device, const SHADER_OPCODE *instruction_table)
101 shader->ref = 1;
102 shader->device = device;
103 shader->shader_ins = instruction_table;
104 list_init(&shader->linked_programs);
107 const SHADER_OPCODE *shader_get_opcode(const SHADER_OPCODE *opcode_table, DWORD shader_version, DWORD code)
109 DWORD i = 0;
111 /** TODO: use dichotomic search */
112 while (opcode_table[i].name)
114 if ((code & WINED3DSI_OPCODE_MASK) == opcode_table[i].opcode
115 && shader_version >= opcode_table[i].min_version
116 && (!opcode_table[i].max_version || shader_version <= opcode_table[i].max_version))
118 return &opcode_table[i];
120 ++i;
123 FIXME("Unsupported opcode %#x(%d) masked %#x, shader version %#x\n",
124 code, code, code & WINED3DSI_OPCODE_MASK, shader_version);
126 return NULL;
129 /* Read a parameter opcode from the input stream,
130 * and possibly a relative addressing token.
131 * Return the number of tokens read */
132 static int shader_get_param(const DWORD *pToken, DWORD shader_version, DWORD *param, DWORD *addr_token)
134 /* PS >= 3.0 have relative addressing (with token)
135 * VS >= 2.0 have relative addressing (with token)
136 * VS >= 1.0 < 2.0 have relative addressing (without token)
137 * The version check below should work in general */
139 char rel_token = WINED3DSHADER_VERSION_MAJOR(shader_version) >= 2 &&
140 ((*pToken & WINED3DSHADER_ADDRESSMODE_MASK) == WINED3DSHADER_ADDRMODE_RELATIVE);
142 *param = *pToken;
143 *addr_token = rel_token? *(pToken + 1): 0;
144 return rel_token? 2:1;
147 /* Return the number of parameters to skip for an opcode */
148 static inline int shader_skip_opcode(const SHADER_OPCODE *curOpcode, DWORD opcode_token, DWORD shader_version)
150 /* Shaders >= 2.0 may contain address tokens, but fortunately they
151 * have a useful length mask - use it here. Shaders 1.0 contain no such tokens */
152 return (WINED3DSHADER_VERSION_MAJOR(shader_version) >= 2)
153 ? ((opcode_token & WINED3DSI_INSTLENGTH_MASK) >> WINED3DSI_INSTLENGTH_SHIFT) : curOpcode->num_params;
156 /* Read the parameters of an unrecognized opcode from the input stream
157 * Return the number of tokens read.
159 * Note: This function assumes source or destination token format.
160 * It will not work with specially-formatted tokens like DEF or DCL,
161 * but hopefully those would be recognized */
162 static int shader_skip_unrecognized(const DWORD *pToken, DWORD shader_version)
164 int tokens_read = 0;
165 int i = 0;
167 /* TODO: Think of a good name for 0x80000000 and replace it with a constant */
168 while (*pToken & 0x80000000) {
170 DWORD param, addr_token;
171 tokens_read += shader_get_param(pToken, shader_version, &param, &addr_token);
172 pToken += tokens_read;
174 FIXME("Unrecognized opcode param: token=0x%08x "
175 "addr_token=0x%08x name=", param, addr_token);
176 shader_dump_param(param, addr_token, i, shader_version);
177 FIXME("\n");
178 ++i;
180 return tokens_read;
183 /* Convert floating point offset relative
184 * to a register file to an absolute offset for float constants */
185 static unsigned int shader_get_float_offset(const DWORD reg)
187 unsigned int regnum = reg & WINED3DSP_REGNUM_MASK;
188 int regtype = shader_get_regtype(reg);
190 switch (regtype) {
191 case WINED3DSPR_CONST: return regnum;
192 case WINED3DSPR_CONST2: return 2048 + regnum;
193 case WINED3DSPR_CONST3: return 4096 + regnum;
194 case WINED3DSPR_CONST4: return 6144 + regnum;
195 default:
196 FIXME("Unsupported register type: %d\n", regtype);
197 return regnum;
201 static void shader_delete_constant_list(struct list* clist) {
203 struct list *ptr;
204 struct local_constant* constant;
206 ptr = list_head(clist);
207 while (ptr) {
208 constant = LIST_ENTRY(ptr, struct local_constant, entry);
209 ptr = list_next(clist, ptr);
210 HeapFree(GetProcessHeap(), 0, constant);
212 list_init(clist);
215 static void shader_parse_dst_param(DWORD param, DWORD addr_param, struct wined3d_shader_dst_param *dst)
217 dst->register_type = ((param & WINED3DSP_REGTYPE_MASK) >> WINED3DSP_REGTYPE_SHIFT)
218 | ((param & WINED3DSP_REGTYPE_MASK2) >> WINED3DSP_REGTYPE_SHIFT2);
219 dst->register_idx = param & WINED3DSP_REGNUM_MASK;
220 dst->write_mask = param & WINED3DSP_WRITEMASK_ALL;
221 dst->modifiers = param & WINED3DSP_DSTMOD_MASK;
222 dst->token = param;
223 dst->addr_token = addr_param;
226 /* Note that this does not count the loop register
227 * as an address register. */
229 HRESULT shader_get_registers_used(IWineD3DBaseShader *iface, struct shader_reg_maps *reg_maps,
230 struct wined3d_shader_semantic *semantics_in, struct wined3d_shader_semantic *semantics_out,
231 const DWORD *byte_code)
233 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
234 const SHADER_OPCODE *shader_ins = This->baseShader.shader_ins;
235 DWORD shader_version;
236 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
237 const DWORD* pToken = byte_code;
238 char pshader;
240 /* There are some minor differences between pixel and vertex shaders */
242 memset(reg_maps->bumpmat, 0, sizeof(reg_maps->bumpmat));
243 memset(reg_maps->luminanceparams, 0, sizeof(reg_maps->luminanceparams));
245 /* get_registers_used is called on every compile on some 1.x shaders, which can result
246 * in stacking up a collection of local constants. Delete the old constants if existing
248 shader_delete_constant_list(&This->baseShader.constantsF);
249 shader_delete_constant_list(&This->baseShader.constantsB);
250 shader_delete_constant_list(&This->baseShader.constantsI);
252 /* The version token is supposed to be the first token */
253 if (!shader_is_version_token(*pToken))
255 FIXME("First token is not a version token, invalid shader.\n");
256 return WINED3DERR_INVALIDCALL;
258 reg_maps->shader_version = shader_version = *pToken++;
259 pshader = shader_is_pshader_version(shader_version);
261 while (WINED3DVS_END() != *pToken) {
262 CONST SHADER_OPCODE* curOpcode;
263 DWORD opcode_token;
265 /* Skip comments */
266 if (shader_is_comment(*pToken))
268 DWORD comment_len = (*pToken & WINED3DSI_COMMENTSIZE_MASK) >> WINED3DSI_COMMENTSIZE_SHIFT;
269 ++pToken;
270 pToken += comment_len;
271 continue;
274 /* Fetch opcode */
275 opcode_token = *pToken++;
276 curOpcode = shader_get_opcode(shader_ins, shader_version, opcode_token);
278 /* Unhandled opcode, and its parameters */
279 if (NULL == curOpcode) {
280 while (*pToken & 0x80000000)
281 ++pToken;
283 /* Handle declarations */
284 } else if (WINED3DSIO_DCL == curOpcode->opcode) {
286 DWORD usage = *pToken++;
287 DWORD param = *pToken++;
288 DWORD regtype = shader_get_regtype(param);
289 unsigned int regnum = param & WINED3DSP_REGNUM_MASK;
291 /* Vshader: mark attributes used
292 Pshader: mark 3.0 input registers used, save token */
293 if (WINED3DSPR_INPUT == regtype) {
295 if (!pshader)
296 reg_maps->attributes[regnum] = 1;
297 else
298 reg_maps->packed_input[regnum] = 1;
300 semantics_in[regnum].usage = (usage & WINED3DSP_DCL_USAGE_MASK) >> WINED3DSP_DCL_USAGE_SHIFT;
301 semantics_in[regnum].usage_idx =
302 (usage & WINED3DSP_DCL_USAGEINDEX_MASK) >> WINED3DSP_DCL_USAGEINDEX_SHIFT;
303 shader_parse_dst_param(param, 0, &semantics_in[regnum].reg);
305 /* Vshader: mark 3.0 output registers used, save token */
306 } else if (WINED3DSPR_OUTPUT == regtype) {
307 reg_maps->packed_output[regnum] = 1;
308 semantics_out[regnum].usage = (usage & WINED3DSP_DCL_USAGE_MASK) >> WINED3DSP_DCL_USAGE_SHIFT;
309 semantics_out[regnum].usage_idx =
310 (usage & WINED3DSP_DCL_USAGEINDEX_MASK) >> WINED3DSP_DCL_USAGEINDEX_SHIFT;
311 shader_parse_dst_param(param, 0, &semantics_out[regnum].reg);
313 if (usage & (WINED3DDECLUSAGE_FOG << WINED3DSP_DCL_USAGE_SHIFT))
314 reg_maps->fog = 1;
316 /* Save sampler usage token */
317 } else if (WINED3DSPR_SAMPLER == regtype)
318 reg_maps->samplers[regnum] = usage;
320 } else if (WINED3DSIO_DEF == curOpcode->opcode) {
322 local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
323 if (!lconst) return E_OUTOFMEMORY;
324 lconst->idx = *pToken & WINED3DSP_REGNUM_MASK;
325 memcpy(lconst->value, pToken + 1, 4 * sizeof(DWORD));
327 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
328 if (WINED3DSHADER_VERSION_MAJOR(shader_version) == 1 && pshader)
330 float *value = (float *) lconst->value;
331 if(value[0] < -1.0) value[0] = -1.0;
332 else if(value[0] > 1.0) value[0] = 1.0;
333 if(value[1] < -1.0) value[1] = -1.0;
334 else if(value[1] > 1.0) value[1] = 1.0;
335 if(value[2] < -1.0) value[2] = -1.0;
336 else if(value[2] > 1.0) value[2] = 1.0;
337 if(value[3] < -1.0) value[3] = -1.0;
338 else if(value[3] > 1.0) value[3] = 1.0;
341 list_add_head(&This->baseShader.constantsF, &lconst->entry);
342 pToken += curOpcode->num_params;
344 } else if (WINED3DSIO_DEFI == curOpcode->opcode) {
346 local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
347 if (!lconst) return E_OUTOFMEMORY;
348 lconst->idx = *pToken & WINED3DSP_REGNUM_MASK;
349 memcpy(lconst->value, pToken + 1, 4 * sizeof(DWORD));
350 list_add_head(&This->baseShader.constantsI, &lconst->entry);
351 pToken += curOpcode->num_params;
353 } else if (WINED3DSIO_DEFB == curOpcode->opcode) {
355 local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
356 if (!lconst) return E_OUTOFMEMORY;
357 lconst->idx = *pToken & WINED3DSP_REGNUM_MASK;
358 memcpy(lconst->value, pToken + 1, 1 * sizeof(DWORD));
359 list_add_head(&This->baseShader.constantsB, &lconst->entry);
360 pToken += curOpcode->num_params;
362 /* If there's a loop in the shader */
363 } else if (WINED3DSIO_LOOP == curOpcode->opcode ||
364 WINED3DSIO_REP == curOpcode->opcode) {
365 cur_loop_depth++;
366 if(cur_loop_depth > max_loop_depth)
367 max_loop_depth = cur_loop_depth;
368 pToken += curOpcode->num_params;
370 /* Rep and Loop always use an integer constant for the control parameters */
371 This->baseShader.uses_int_consts = TRUE;
372 } else if (WINED3DSIO_ENDLOOP == curOpcode->opcode ||
373 WINED3DSIO_ENDREP == curOpcode->opcode) {
374 cur_loop_depth--;
376 /* For subroutine prototypes */
377 } else if (WINED3DSIO_LABEL == curOpcode->opcode) {
379 DWORD snum = *pToken & WINED3DSP_REGNUM_MASK;
380 reg_maps->labels[snum] = 1;
381 pToken += curOpcode->num_params;
383 /* Set texture, address, temporary registers */
384 } else {
385 int i, limit;
387 /* Declare 1.X samplers implicitly, based on the destination reg. number */
388 if (WINED3DSHADER_VERSION_MAJOR(shader_version) == 1
389 && pshader /* Filter different instructions with the same enum values in VS */
390 && (WINED3DSIO_TEX == curOpcode->opcode
391 || WINED3DSIO_TEXBEM == curOpcode->opcode
392 || WINED3DSIO_TEXBEML == curOpcode->opcode
393 || WINED3DSIO_TEXDP3TEX == curOpcode->opcode
394 || WINED3DSIO_TEXM3x2TEX == curOpcode->opcode
395 || WINED3DSIO_TEXM3x3SPEC == curOpcode->opcode
396 || WINED3DSIO_TEXM3x3TEX == curOpcode->opcode
397 || WINED3DSIO_TEXM3x3VSPEC == curOpcode->opcode
398 || WINED3DSIO_TEXREG2AR == curOpcode->opcode
399 || WINED3DSIO_TEXREG2GB == curOpcode->opcode
400 || WINED3DSIO_TEXREG2RGB == curOpcode->opcode))
402 /* Fake sampler usage, only set reserved bit and ttype */
403 DWORD sampler_code = *pToken & WINED3DSP_REGNUM_MASK;
405 TRACE("Setting fake 2D sampler for 1.x pixelshader\n");
406 reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
408 /* texbem is only valid with < 1.4 pixel shaders */
409 if(WINED3DSIO_TEXBEM == curOpcode->opcode ||
410 WINED3DSIO_TEXBEML == curOpcode->opcode) {
411 reg_maps->bumpmat[sampler_code] = TRUE;
412 if(WINED3DSIO_TEXBEML == curOpcode->opcode) {
413 reg_maps->luminanceparams[sampler_code] = TRUE;
417 if(WINED3DSIO_NRM == curOpcode->opcode) {
418 reg_maps->usesnrm = 1;
419 } else if(WINED3DSIO_BEM == curOpcode->opcode && pshader) {
420 DWORD regnum = *pToken & WINED3DSP_REGNUM_MASK;
421 reg_maps->bumpmat[regnum] = TRUE;
422 } else if(WINED3DSIO_DSY == curOpcode->opcode) {
423 reg_maps->usesdsy = 1;
426 /* This will loop over all the registers and try to
427 * make a bitmask of the ones we're interested in.
429 * Relative addressing tokens are ignored, but that's
430 * okay, since we'll catch any address registers when
431 * they are initialized (required by spec) */
433 limit = (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED)?
434 curOpcode->num_params + 1: curOpcode->num_params;
436 for (i = 0; i < limit; ++i) {
438 DWORD param, addr_token, reg, regtype;
439 pToken += shader_get_param(pToken, shader_version, &param, &addr_token);
441 regtype = shader_get_regtype(param);
442 reg = param & WINED3DSP_REGNUM_MASK;
444 if (WINED3DSPR_TEXTURE == regtype) { /* vs: WINED3DSPR_ADDR */
446 if (pshader)
447 reg_maps->texcoord[reg] = 1;
448 else
449 reg_maps->address[reg] = 1;
452 else if (WINED3DSPR_TEMP == regtype)
453 reg_maps->temporary[reg] = 1;
455 else if (WINED3DSPR_INPUT == regtype) {
456 if( !pshader)
457 reg_maps->attributes[reg] = 1;
458 else {
459 if(param & WINED3DSHADER_ADDRMODE_RELATIVE) {
460 /* If relative addressing is used, we must assume that all registers
461 * are used. Even if it is a construct like v3[aL], we can't assume
462 * that v0, v1 and v2 aren't read because aL can be negative
464 unsigned int i;
465 for(i = 0; i < MAX_REG_INPUT; i++) {
466 ((IWineD3DPixelShaderImpl *) This)->input_reg_used[i] = TRUE;
468 } else {
469 ((IWineD3DPixelShaderImpl *) This)->input_reg_used[reg] = TRUE;
474 else if (WINED3DSPR_RASTOUT == regtype && reg == 1)
475 reg_maps->fog = 1;
477 else if (WINED3DSPR_MISCTYPE == regtype && reg == 0 && pshader)
478 reg_maps->vpos = 1;
480 else if(WINED3DSPR_CONST == regtype) {
481 if(param & WINED3DSHADER_ADDRMODE_RELATIVE) {
482 if(!pshader) {
483 if(reg <= ((IWineD3DVertexShaderImpl *) This)->min_rel_offset) {
484 ((IWineD3DVertexShaderImpl *) This)->min_rel_offset = reg;
485 } else if(reg >= ((IWineD3DVertexShaderImpl *) This)->max_rel_offset) {
486 ((IWineD3DVertexShaderImpl *) This)->max_rel_offset = reg;
489 reg_maps->usesrelconstF = TRUE;
492 else if(WINED3DSPR_CONSTINT == regtype) {
493 This->baseShader.uses_int_consts = TRUE;
495 else if(WINED3DSPR_CONSTBOOL == regtype) {
496 This->baseShader.uses_bool_consts = TRUE;
499 /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and is used
500 * in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel shaders because TECRDOUT
501 * isn't used in them, but future register types might cause issues
503 else if (WINED3DSPR_TEXCRDOUT == regtype && i == 0 /* Only look at writes */
504 && !pshader && WINED3DSHADER_VERSION_MAJOR(shader_version) < 3)
506 reg_maps->texcoord_mask[reg] |= shader_get_writemask(param);
511 ++pToken;
512 reg_maps->loop_depth = max_loop_depth;
514 This->baseShader.functionLength = ((char *)pToken - (char *)byte_code);
516 return WINED3D_OK;
519 static void shader_dump_decl_usage(DWORD decl, DWORD param, DWORD shader_version)
521 DWORD regtype = shader_get_regtype(param);
523 TRACE("dcl");
525 if (regtype == WINED3DSPR_SAMPLER) {
526 DWORD ttype = decl & WINED3DSP_TEXTURETYPE_MASK;
528 switch (ttype) {
529 case WINED3DSTT_2D: TRACE("_2d"); break;
530 case WINED3DSTT_CUBE: TRACE("_cube"); break;
531 case WINED3DSTT_VOLUME: TRACE("_volume"); break;
532 default: TRACE("_unknown_ttype(0x%08x)", ttype);
535 } else {
537 DWORD usage = decl & WINED3DSP_DCL_USAGE_MASK;
538 DWORD idx = (decl & WINED3DSP_DCL_USAGEINDEX_MASK) >> WINED3DSP_DCL_USAGEINDEX_SHIFT;
540 /* Pixel shaders 3.0 don't have usage semantics */
541 if (shader_is_pshader_version(shader_version) && shader_version < WINED3DPS_VERSION(3,0))
542 return;
543 else
544 TRACE("_");
546 switch(usage) {
547 case WINED3DDECLUSAGE_POSITION:
548 TRACE("position%d", idx);
549 break;
550 case WINED3DDECLUSAGE_BLENDINDICES:
551 TRACE("blend");
552 break;
553 case WINED3DDECLUSAGE_BLENDWEIGHT:
554 TRACE("weight");
555 break;
556 case WINED3DDECLUSAGE_NORMAL:
557 TRACE("normal%d", idx);
558 break;
559 case WINED3DDECLUSAGE_PSIZE:
560 TRACE("psize");
561 break;
562 case WINED3DDECLUSAGE_COLOR:
563 if(idx == 0) {
564 TRACE("color");
565 } else {
566 TRACE("specular%d", (idx - 1));
568 break;
569 case WINED3DDECLUSAGE_TEXCOORD:
570 TRACE("texture%d", idx);
571 break;
572 case WINED3DDECLUSAGE_TANGENT:
573 TRACE("tangent");
574 break;
575 case WINED3DDECLUSAGE_BINORMAL:
576 TRACE("binormal");
577 break;
578 case WINED3DDECLUSAGE_TESSFACTOR:
579 TRACE("tessfactor");
580 break;
581 case WINED3DDECLUSAGE_POSITIONT:
582 TRACE("positionT%d", idx);
583 break;
584 case WINED3DDECLUSAGE_FOG:
585 TRACE("fog");
586 break;
587 case WINED3DDECLUSAGE_DEPTH:
588 TRACE("depth");
589 break;
590 case WINED3DDECLUSAGE_SAMPLE:
591 TRACE("sample");
592 break;
593 default:
594 FIXME("unknown_semantics(0x%08x)", usage);
599 static void shader_dump_arr_entry(const DWORD param, const DWORD addr_token,
600 unsigned int reg, int input, DWORD shader_version)
602 char relative =
603 ((param & WINED3DSHADER_ADDRESSMODE_MASK) == WINED3DSHADER_ADDRMODE_RELATIVE);
605 if (relative) {
606 TRACE("[");
607 if (addr_token)
608 shader_dump_param(addr_token, 0, input, shader_version);
609 else
610 TRACE("a0.x");
611 TRACE(" + ");
613 TRACE("%u", reg);
614 if (relative)
615 TRACE("]");
618 static void shader_dump_param(const DWORD param, const DWORD addr_token, int input, DWORD shader_version)
620 static const char * const rastout_reg_names[] = { "oPos", "oFog", "oPts" };
621 static const char * const misctype_reg_names[] = { "vPos", "vFace"};
622 const char *swizzle_reg_chars = "xyzw";
624 DWORD reg = param & WINED3DSP_REGNUM_MASK;
625 DWORD regtype = shader_get_regtype(param);
626 DWORD modifier = param & WINED3DSP_SRCMOD_MASK;
628 /* There are some minor differences between pixel and vertex shaders */
629 char pshader = shader_is_pshader_version(shader_version);
631 if (input) {
632 if ( (modifier == WINED3DSPSM_NEG) ||
633 (modifier == WINED3DSPSM_BIASNEG) ||
634 (modifier == WINED3DSPSM_SIGNNEG) ||
635 (modifier == WINED3DSPSM_X2NEG) ||
636 (modifier == WINED3DSPSM_ABSNEG) )
637 TRACE("-");
638 else if (modifier == WINED3DSPSM_COMP)
639 TRACE("1-");
640 else if (modifier == WINED3DSPSM_NOT)
641 TRACE("!");
643 if (modifier == WINED3DSPSM_ABS || modifier == WINED3DSPSM_ABSNEG)
644 TRACE("abs(");
647 switch (regtype) {
648 case WINED3DSPR_TEMP:
649 TRACE("r%u", reg);
650 break;
651 case WINED3DSPR_INPUT:
652 TRACE("v");
653 shader_dump_arr_entry(param, addr_token, reg, input, shader_version);
654 break;
655 case WINED3DSPR_CONST:
656 case WINED3DSPR_CONST2:
657 case WINED3DSPR_CONST3:
658 case WINED3DSPR_CONST4:
659 TRACE("c");
660 shader_dump_arr_entry(param, addr_token, shader_get_float_offset(param), input, shader_version);
661 break;
662 case WINED3DSPR_TEXTURE: /* vs: case D3DSPR_ADDR */
663 TRACE("%c%u", (pshader? 't':'a'), reg);
664 break;
665 case WINED3DSPR_RASTOUT:
666 TRACE("%s", rastout_reg_names[reg]);
667 break;
668 case WINED3DSPR_COLOROUT:
669 TRACE("oC%u", reg);
670 break;
671 case WINED3DSPR_DEPTHOUT:
672 TRACE("oDepth");
673 break;
674 case WINED3DSPR_ATTROUT:
675 TRACE("oD%u", reg);
676 break;
677 case WINED3DSPR_TEXCRDOUT:
679 /* Vertex shaders >= 3.0 use general purpose output registers
680 * (WINED3DSPR_OUTPUT), which can include an address token */
682 if (WINED3DSHADER_VERSION_MAJOR(shader_version) >= 3) {
683 TRACE("o");
684 shader_dump_arr_entry(param, addr_token, reg, input, shader_version);
686 else
687 TRACE("oT%u", reg);
688 break;
689 case WINED3DSPR_CONSTINT:
690 TRACE("i");
691 shader_dump_arr_entry(param, addr_token, reg, input, shader_version);
692 break;
693 case WINED3DSPR_CONSTBOOL:
694 TRACE("b");
695 shader_dump_arr_entry(param, addr_token, reg, input, shader_version);
696 break;
697 case WINED3DSPR_LABEL:
698 TRACE("l%u", reg);
699 break;
700 case WINED3DSPR_LOOP:
701 TRACE("aL");
702 break;
703 case WINED3DSPR_SAMPLER:
704 TRACE("s%u", reg);
705 break;
706 case WINED3DSPR_MISCTYPE:
707 if (reg > 1) {
708 FIXME("Unhandled misctype register %d\n", reg);
709 } else {
710 TRACE("%s", misctype_reg_names[reg]);
712 break;
713 case WINED3DSPR_PREDICATE:
714 TRACE("p%u", reg);
715 break;
716 default:
717 TRACE("unhandled_rtype(%#x)", regtype);
718 break;
721 if (!input) {
722 /* operand output (for modifiers and shift, see dump_ins_modifiers) */
724 if ((param & WINED3DSP_WRITEMASK_ALL) != WINED3DSP_WRITEMASK_ALL) {
725 TRACE(".");
726 if (param & WINED3DSP_WRITEMASK_0) TRACE("%c", swizzle_reg_chars[0]);
727 if (param & WINED3DSP_WRITEMASK_1) TRACE("%c", swizzle_reg_chars[1]);
728 if (param & WINED3DSP_WRITEMASK_2) TRACE("%c", swizzle_reg_chars[2]);
729 if (param & WINED3DSP_WRITEMASK_3) TRACE("%c", swizzle_reg_chars[3]);
732 } else {
733 /** operand input */
734 DWORD swizzle = (param & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
735 DWORD swizzle_x = swizzle & 0x03;
736 DWORD swizzle_y = (swizzle >> 2) & 0x03;
737 DWORD swizzle_z = (swizzle >> 4) & 0x03;
738 DWORD swizzle_w = (swizzle >> 6) & 0x03;
740 if (0 != modifier) {
741 switch (modifier) {
742 case WINED3DSPSM_NONE: break;
743 case WINED3DSPSM_NEG: break;
744 case WINED3DSPSM_NOT: break;
745 case WINED3DSPSM_BIAS: TRACE("_bias"); break;
746 case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
747 case WINED3DSPSM_SIGN: TRACE("_bx2"); break;
748 case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
749 case WINED3DSPSM_COMP: break;
750 case WINED3DSPSM_X2: TRACE("_x2"); break;
751 case WINED3DSPSM_X2NEG: TRACE("_x2"); break;
752 case WINED3DSPSM_DZ: TRACE("_dz"); break;
753 case WINED3DSPSM_DW: TRACE("_dw"); break;
754 case WINED3DSPSM_ABSNEG: TRACE(")"); break;
755 case WINED3DSPSM_ABS: TRACE(")"); break;
756 default:
757 TRACE("_unknown_modifier(%#x)", modifier >> WINED3DSP_SRCMOD_SHIFT);
762 * swizzle bits fields:
763 * RRGGBBAA
765 if ((WINED3DVS_NOSWIZZLE >> WINED3DVS_SWIZZLE_SHIFT) != swizzle) {
766 if (swizzle_x == swizzle_y &&
767 swizzle_x == swizzle_z &&
768 swizzle_x == swizzle_w) {
769 TRACE(".%c", swizzle_reg_chars[swizzle_x]);
770 } else {
771 TRACE(".%c%c%c%c",
772 swizzle_reg_chars[swizzle_x],
773 swizzle_reg_chars[swizzle_y],
774 swizzle_reg_chars[swizzle_z],
775 swizzle_reg_chars[swizzle_w]);
781 /* Shared code in order to generate the bulk of the shader string.
782 * NOTE: A description of how to parse tokens can be found on msdn */
783 void shader_generate_main(IWineD3DBaseShader *iface, SHADER_BUFFER* buffer,
784 const shader_reg_maps* reg_maps, CONST DWORD* pFunction)
786 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
787 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device; /* To access shader backend callbacks */
788 const SHADER_OPCODE *opcode_table = This->baseShader.shader_ins;
789 const SHADER_HANDLER *handler_table = device->shader_backend->shader_instruction_handler_table;
790 DWORD shader_version = reg_maps->shader_version;
791 struct wined3d_shader_dst_param dst_param;
792 struct wined3d_shader_instruction ins;
793 const DWORD *pToken = pFunction;
794 const SHADER_OPCODE *curOpcode;
795 SHADER_HANDLER hw_fct;
796 DWORD i;
798 /* Initialize current parsing state */
799 ins.shader = iface;
800 ins.buffer = buffer;
801 ins.reg_maps = reg_maps;
802 ins.dst = &dst_param;
803 This->baseShader.parse_state.current_row = 0;
805 while (WINED3DPS_END() != *pToken)
807 DWORD opcode_token;
809 /* Skip version token */
810 if (shader_is_version_token(*pToken))
812 ++pToken;
813 continue;
816 /* Skip comment tokens */
817 if (shader_is_comment(*pToken))
819 pToken += (*pToken & WINED3DSI_COMMENTSIZE_MASK) >> WINED3DSI_COMMENTSIZE_SHIFT;
820 ++pToken;
821 continue;
824 /* Read opcode */
825 opcode_token = *pToken++;
826 curOpcode = shader_get_opcode(opcode_table, shader_version, opcode_token);
828 /* Unknown opcode and its parameters */
829 if (!curOpcode)
831 FIXME("Unrecognized opcode: token=0x%08x\n", opcode_token);
832 pToken += shader_skip_unrecognized(pToken, shader_version);
833 continue;
836 /* Nothing to do */
837 if (WINED3DSIO_DCL == curOpcode->opcode
838 || WINED3DSIO_NOP == curOpcode->opcode
839 || WINED3DSIO_DEF == curOpcode->opcode
840 || WINED3DSIO_DEFI == curOpcode->opcode
841 || WINED3DSIO_DEFB == curOpcode->opcode
842 || WINED3DSIO_PHASE == curOpcode->opcode
843 || WINED3DSIO_RET == curOpcode->opcode)
845 pToken += shader_skip_opcode(curOpcode, opcode_token, shader_version);
846 continue;
849 /* Select handler */
850 hw_fct = handler_table[curOpcode->handler_idx];
852 /* Unhandled opcode */
853 if (!hw_fct)
855 FIXME("Can't handle opcode %s in hwShader\n", curOpcode->name);
856 pToken += shader_skip_opcode(curOpcode, opcode_token, shader_version);
857 continue;
860 ins.handler_idx = curOpcode->handler_idx;
861 ins.flags = opcode_token & WINED3D_OPCODESPECIFICCONTROL_MASK;
862 ins.coissue = opcode_token & WINED3DSI_COISSUE;
864 /* Destination token */
865 ins.dst_count = curOpcode->dst_token ? 1 : 0;
866 if (ins.dst_count)
868 DWORD param, addr_param = 0;
869 pToken += shader_get_param(pToken, shader_version, &param, &addr_param);
870 shader_parse_dst_param(param, addr_param, &dst_param);
873 /* Predication token */
874 if (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED) ins.predicate = *pToken++;
876 /* Other source tokens */
877 ins.src_count = curOpcode->num_params - curOpcode->dst_token;
878 for (i = 0; i < ins.src_count; ++i)
880 DWORD param, addr_token = 0;
881 pToken += shader_get_param(pToken, shader_version, &param, &addr_token);
882 ins.src[i] = param;
883 ins.src_addr[i] = addr_token;
886 /* Call appropriate function for output target */
887 hw_fct(&ins);
889 /* Process instruction modifiers for GLSL apps ( _sat, etc. ) */
890 /* FIXME: This should be internal to the shader backend.
891 * Also, right now this is the only reason "shader_mode" exists. */
892 if (This->baseShader.shader_mode == SHADER_GLSL) shader_glsl_add_instruction_modifiers(&ins);
896 static void shader_dump_ins_modifiers(const DWORD output)
898 DWORD shift = (output & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
899 DWORD mmask = output & WINED3DSP_DSTMOD_MASK;
901 switch (shift) {
902 case 0: break;
903 case 13: TRACE("_d8"); break;
904 case 14: TRACE("_d4"); break;
905 case 15: TRACE("_d2"); break;
906 case 1: TRACE("_x2"); break;
907 case 2: TRACE("_x4"); break;
908 case 3: TRACE("_x8"); break;
909 default: TRACE("_unhandled_shift(%d)", shift); break;
912 if (mmask & WINED3DSPDM_SATURATE) TRACE("_sat");
913 if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
914 if (mmask & WINED3DSPDM_MSAMPCENTROID) TRACE("_centroid");
916 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
917 if (mmask)
918 FIXME("_unrecognized_modifier(%#x)", mmask >> WINED3DSP_DSTMOD_SHIFT);
921 void shader_trace_init(const DWORD *pFunction, const SHADER_OPCODE *opcode_table)
923 const DWORD* pToken = pFunction;
924 const SHADER_OPCODE* curOpcode = NULL;
925 DWORD shader_version;
926 DWORD opcode_token;
927 DWORD i;
929 TRACE("Parsing %p\n", pFunction);
931 /* The version token is supposed to be the first token */
932 if (!shader_is_version_token(*pToken))
934 FIXME("First token is not a version token, invalid shader.\n");
935 return;
937 shader_version = *pToken++;
938 TRACE("%s_%u_%u\n", shader_is_pshader_version(shader_version) ? "ps": "vs",
939 WINED3DSHADER_VERSION_MAJOR(shader_version), WINED3DSHADER_VERSION_MINOR(shader_version));
941 while (WINED3DVS_END() != *pToken)
943 if (shader_is_comment(*pToken)) /* comment */
945 DWORD comment_len = (*pToken & WINED3DSI_COMMENTSIZE_MASK) >> WINED3DSI_COMMENTSIZE_SHIFT;
946 ++pToken;
947 TRACE("//%s\n", (const char*)pToken);
948 pToken += comment_len;
949 continue;
951 opcode_token = *pToken++;
952 curOpcode = shader_get_opcode(opcode_table, shader_version, opcode_token);
954 if (!curOpcode)
956 int tokens_read;
957 FIXME("Unrecognized opcode: token=0x%08x\n", opcode_token);
958 tokens_read = shader_skip_unrecognized(pToken, shader_version);
959 pToken += tokens_read;
961 else
963 if (curOpcode->opcode == WINED3DSIO_DCL)
965 DWORD usage = *pToken;
966 DWORD param = *(pToken + 1);
968 shader_dump_decl_usage(usage, param, shader_version);
969 shader_dump_ins_modifiers(param);
970 TRACE(" ");
971 shader_dump_param(param, 0, 0, shader_version);
972 pToken += 2;
974 else if (curOpcode->opcode == WINED3DSIO_DEF)
976 unsigned int offset = shader_get_float_offset(*pToken);
978 TRACE("def c%u = %f, %f, %f, %f", offset,
979 *(const float *)(pToken + 1),
980 *(const float *)(pToken + 2),
981 *(const float *)(pToken + 3),
982 *(const float *)(pToken + 4));
983 pToken += 5;
985 else if (curOpcode->opcode == WINED3DSIO_DEFI)
987 TRACE("defi i%u = %d, %d, %d, %d", *pToken & WINED3DSP_REGNUM_MASK,
988 *(pToken + 1),
989 *(pToken + 2),
990 *(pToken + 3),
991 *(pToken + 4));
992 pToken += 5;
994 else if (curOpcode->opcode == WINED3DSIO_DEFB)
996 TRACE("defb b%u = %s", *pToken & WINED3DSP_REGNUM_MASK,
997 *(pToken + 1)? "true": "false");
998 pToken += 2;
1000 else
1002 DWORD param, addr_token;
1003 int tokens_read;
1005 /* Print out predication source token first - it follows
1006 * the destination token. */
1007 if (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED)
1009 TRACE("(");
1010 shader_dump_param(*(pToken + 2), 0, 1, shader_version);
1011 TRACE(") ");
1013 if (opcode_token & WINED3DSI_COISSUE)
1015 /* PixWin marks instructions with the coissue flag with a '+' */
1016 TRACE("+");
1019 TRACE("%s", curOpcode->name);
1021 if (curOpcode->opcode == WINED3DSIO_IFC
1022 || curOpcode->opcode == WINED3DSIO_BREAKC)
1024 DWORD op = (opcode_token & INST_CONTROLS_MASK) >> INST_CONTROLS_SHIFT;
1026 switch (op)
1028 case COMPARISON_GT: TRACE("_gt"); break;
1029 case COMPARISON_EQ: TRACE("_eq"); break;
1030 case COMPARISON_GE: TRACE("_ge"); break;
1031 case COMPARISON_LT: TRACE("_lt"); break;
1032 case COMPARISON_NE: TRACE("_ne"); break;
1033 case COMPARISON_LE: TRACE("_le"); break;
1034 default: TRACE("_(%u)", op);
1037 else if (curOpcode->opcode == WINED3DSIO_TEX
1038 && shader_version >= WINED3DPS_VERSION(2,0)
1039 && (opcode_token & WINED3DSI_TEXLD_PROJECT))
1041 TRACE("p");
1044 /* Destination token */
1045 if (curOpcode->dst_token)
1047 tokens_read = shader_get_param(pToken, shader_version, &param, &addr_token);
1048 pToken += tokens_read;
1050 shader_dump_ins_modifiers(param);
1051 TRACE(" ");
1052 shader_dump_param(param, addr_token, 0, shader_version);
1055 /* Predication token - already printed out, just skip it */
1056 if (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED)
1058 pToken++;
1061 /* Other source tokens */
1062 for (i = curOpcode->dst_token; i < curOpcode->num_params; ++i)
1064 tokens_read = shader_get_param(pToken, shader_version, &param, &addr_token);
1065 pToken += tokens_read;
1067 TRACE((i == 0)? " " : ", ");
1068 shader_dump_param(param, addr_token, 1, shader_version);
1071 TRACE("\n");
1076 void shader_cleanup(IWineD3DBaseShader *iface)
1078 IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)iface;
1080 ((IWineD3DDeviceImpl *)This->baseShader.device)->shader_backend->shader_destroy(iface);
1081 HeapFree(GetProcessHeap(), 0, This->baseShader.function);
1082 shader_delete_constant_list(&This->baseShader.constantsF);
1083 shader_delete_constant_list(&This->baseShader.constantsB);
1084 shader_delete_constant_list(&This->baseShader.constantsI);
1085 list_remove(&This->baseShader.shader_list_entry);
1088 static const SHADER_HANDLER shader_none_instruction_handler_table[WINED3DSIH_TABLE_SIZE] = {0};
1089 static void shader_none_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {}
1090 static void shader_none_select_depth_blt(IWineD3DDevice *iface, enum tex_types tex_type) {}
1091 static void shader_none_deselect_depth_blt(IWineD3DDevice *iface) {}
1092 static void shader_none_update_float_vertex_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1093 static void shader_none_update_float_pixel_constants(IWineD3DDevice *iface, UINT start, UINT count) {}
1094 static void shader_none_load_constants(IWineD3DDevice *iface, char usePS, char useVS) {}
1095 static void shader_none_destroy(IWineD3DBaseShader *iface) {}
1096 static HRESULT shader_none_alloc(IWineD3DDevice *iface) {return WINED3D_OK;}
1097 static void shader_none_free(IWineD3DDevice *iface) {}
1098 static BOOL shader_none_dirty_const(IWineD3DDevice *iface) {return FALSE;}
1099 static GLuint shader_none_generate_pshader(IWineD3DPixelShader *iface, SHADER_BUFFER *buffer, const struct ps_compile_args *args) {
1100 FIXME("NONE shader backend asked to generate a pixel shader\n");
1101 return 0;
1103 static GLuint shader_none_generate_vshader(IWineD3DVertexShader *iface, SHADER_BUFFER *buffer, const struct vs_compile_args *args) {
1104 FIXME("NONE shader backend asked to generate a vertex shader\n");
1105 return 0;
1108 #define GLINFO_LOCATION (*gl_info)
1109 static void shader_none_get_caps(WINED3DDEVTYPE devtype, const WineD3D_GL_Info *gl_info, struct shader_caps *pCaps)
1111 /* Set the shader caps to 0 for the none shader backend */
1112 pCaps->VertexShaderVersion = 0;
1113 pCaps->PixelShaderVersion = 0;
1114 pCaps->PixelShader1xMaxValue = 0.0;
1116 #undef GLINFO_LOCATION
1117 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
1119 if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
1121 TRACE("Checking support for fixup:\n");
1122 dump_color_fixup_desc(fixup);
1125 /* Faked to make some apps happy. */
1126 if (!is_yuv_fixup(fixup))
1128 TRACE("[OK]\n");
1129 return TRUE;
1132 TRACE("[FAILED]\n");
1133 return FALSE;
1136 const shader_backend_t none_shader_backend = {
1137 shader_none_instruction_handler_table,
1138 shader_none_select,
1139 shader_none_select_depth_blt,
1140 shader_none_deselect_depth_blt,
1141 shader_none_update_float_vertex_constants,
1142 shader_none_update_float_pixel_constants,
1143 shader_none_load_constants,
1144 shader_none_destroy,
1145 shader_none_alloc,
1146 shader_none_free,
1147 shader_none_dirty_const,
1148 shader_none_generate_pshader,
1149 shader_none_generate_vshader,
1150 shader_none_get_caps,
1151 shader_none_color_fixup_supported,