?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
msi: Change an ERR to a WARN.
[wine.git] / dlls / wined3d / glsl_shader.c
blob62f27e3314bf151a5d1a6736621bdb4d2769396e
1 /*
2 * GLSL pixel and vertex shader implementation
3 *
4 * Copyright 2006 Jason Green
5 *
6 * This library is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2.1 of the License, or (at your option) any later version.
10 *
11 * This library is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
15 *
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with this library; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
19 */
20
21 #include "config.h"
22 #include <stdio.h>
23 #include "wined3d_private.h"
24
25 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
26
27 #define GLINFO_LOCATION (*gl_info)
28
29 /** Prints the GLSL info log which will contain error messages if they exist */
30 void print_glsl_info_log(WineD3D_GL_Info *gl_info, GLhandleARB obj) {
31
32 int infologLength = 0;
33 char *infoLog;
34
35 GL_EXTCALL(glGetObjectParameterivARB(obj,
36 GL_OBJECT_INFO_LOG_LENGTH_ARB,
37 &infologLength));
38
39 /* A size of 1 is just a null-terminated string, so the log should be bigger than
40 * that if there are errors. */
41 if (infologLength > 1)
42 {
43 infoLog = (char *)HeapAlloc(GetProcessHeap(), 0, infologLength);
44 GL_EXTCALL(glGetInfoLogARB(obj, infologLength, NULL, infoLog));
45 FIXME("Error received from GLSL shader #%u: %s\n", obj, debugstr_a(infoLog));
46 HeapFree(GetProcessHeap(), 0, infoLog);
47 }
48 }
49
50 /**
51 * Loads (pixel shader) samplers
52 */
53 void shader_glsl_load_psamplers(
54 WineD3D_GL_Info *gl_info,
55 IWineD3DStateBlock* iface) {
56
57 IWineD3DStateBlockImpl* stateBlock = (IWineD3DStateBlockImpl*) iface;
58 GLhandleARB programId = stateBlock->shaderPrgId;
59 GLhandleARB name_loc;
60 int i;
61 char sampler_name[20];
62
63 for (i=0; i< GL_LIMITS(samplers); ++i) {
64 if (stateBlock->textures[i] != NULL) {
65 snprintf(sampler_name, sizeof(sampler_name), "Psampler%d", i);
66 name_loc = GL_EXTCALL(glGetUniformLocationARB(programId, sampler_name));
67 if (name_loc != -1) {
68 TRACE_(d3d_shader)("Loading %s for texture %d\n", sampler_name, i);
69 GL_EXTCALL(glUniform1iARB(name_loc, i));
70 checkGLcall("glUniform1iARB");
71 }
72 }
73 }
74 }
75
76 /**
77 * Loads floating point constants (aka uniforms) into the currently set GLSL program.
78 * When @constants_set == NULL, it will load all the constants.
79 */
80 void shader_glsl_load_constantsF(
81 IWineD3DBaseShaderImpl* This,
82 WineD3D_GL_Info *gl_info,
83 GLhandleARB programId,
84 unsigned max_constants,
85 float* constants,
86 BOOL* constants_set) {
87
88 GLhandleARB tmp_loc;
89 int i;
90 char tmp_name[7];
91 char is_pshader = shader_is_pshader_version(This->baseShader.hex_version);
92 const char* prefix = is_pshader? "PC":"VC";
93 struct list* ptr;
94
95 for (i=0; i<max_constants; ++i) {
96 if (NULL == constants_set || constants_set[i]) {
97
98 TRACE("Loading constants %i: %f, %f, %f, %f\n", i,
99 constants[i * sizeof(float) + 0], constants[i * sizeof(float) + 1],
100 constants[i * sizeof(float) + 2], constants[i * sizeof(float) + 3]);
101
102 /* TODO: Benchmark and see if it would be beneficial to store the
103 * locations of the constants to avoid looking up each time */
104 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, i);
105 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
106 if (tmp_loc != -1) {
107 /* We found this uniform name in the program - go ahead and send the data */
108 GL_EXTCALL(glUniform4fvARB(tmp_loc, 1, &constants[i * sizeof(float)]));
109 checkGLcall("glUniform4fvARB");
110 }
111 }
112 }
113
114 /* Load immediate constants */
115 ptr = list_head(&This->baseShader.constantsF);
116 while (ptr) {
117 local_constant* lconst = LIST_ENTRY(ptr, struct local_constant, entry);
118 unsigned int idx = lconst->idx;
119 GLfloat* values = (GLfloat*) lconst->value;
120
121 TRACE("Loading local constants %i: %f, %f, %f, %f\n", idx,
122 values[0], values[1], values[2], values[3]);
123
124 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, idx);
125 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
126 if (tmp_loc != -1) {
127 /* We found this uniform name in the program - go ahead and send the data */
128 GL_EXTCALL(glUniform4fvARB(tmp_loc, 1, values));
129 checkGLcall("glUniform4fvARB");
130 }
131 ptr = list_next(&This->baseShader.constantsF, ptr);
132 }
133 }
134
135 /**
136 * Loads integer constants (aka uniforms) into the currently set GLSL program.
137 * When @constants_set == NULL, it will load all the constants.
138 */
139 void shader_glsl_load_constantsI(
140 IWineD3DBaseShaderImpl* This,
141 WineD3D_GL_Info *gl_info,
142 GLhandleARB programId,
143 unsigned max_constants,
144 int* constants,
145 BOOL* constants_set) {
146
147 GLhandleARB tmp_loc;
148 int i;
149 char tmp_name[7];
150 char is_pshader = shader_is_pshader_version(This->baseShader.hex_version);
151 const char* prefix = is_pshader? "PI":"VI";
152 struct list* ptr;
153
154 for (i=0; i<max_constants; ++i) {
155 if (NULL == constants_set || constants_set[i]) {
156
157 TRACE("Loading constants %i: %i, %i, %i, %i\n",
158 i, constants[i*4], constants[i*4+1], constants[i*4+2], constants[i*4+3]);
159
160 /* TODO: Benchmark and see if it would be beneficial to store the
161 * locations of the constants to avoid looking up each time */
162 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, i);
163 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
164 if (tmp_loc != -1) {
165 /* We found this uniform name in the program - go ahead and send the data */
166 GL_EXTCALL(glUniform4ivARB(tmp_loc, 1, &constants[i*4]));
167 checkGLcall("glUniform4ivARB");
168 }
169 }
170 }
171
172 /* Load immediate constants */
173 ptr = list_head(&This->baseShader.constantsI);
174 while (ptr) {
175 local_constant* lconst = LIST_ENTRY(ptr, struct local_constant, entry);
176 unsigned int idx = lconst->idx;
177 GLint* values = (GLint*) lconst->value;
178
179 TRACE("Loading local constants %i: %i, %i, %i, %i\n", idx,
180 values[0], values[1], values[2], values[3]);
181
182 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, idx);
183 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
184 if (tmp_loc != -1) {
185 /* We found this uniform name in the program - go ahead and send the data */
186 GL_EXTCALL(glUniform4ivARB(tmp_loc, 1, values));
187 checkGLcall("glUniform4ivARB");
188 }
189 ptr = list_next(&This->baseShader.constantsI, ptr);
190 }
191 }
192
193 /**
194 * Loads boolean constants (aka uniforms) into the currently set GLSL program.
195 * When @constants_set == NULL, it will load all the constants.
196 */
197 void shader_glsl_load_constantsB(
198 IWineD3DBaseShaderImpl* This,
199 WineD3D_GL_Info *gl_info,
200 GLhandleARB programId,
201 unsigned max_constants,
202 BOOL* constants,
203 BOOL* constants_set) {
204
205 GLhandleARB tmp_loc;
206 int i;
207 char tmp_name[7];
208 char is_pshader = shader_is_pshader_version(This->baseShader.hex_version);
209 const char* prefix = is_pshader? "PB":"VB";
210 struct list* ptr;
211
212 for (i=0; i<max_constants; ++i) {
213 if (NULL == constants_set || constants_set[i]) {
214
215 TRACE("Loading constants %i: %i;\n", i, constants[i*4]);
216
217 /* TODO: Benchmark and see if it would be beneficial to store the
218 * locations of the constants to avoid looking up each time */
219 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, i);
220 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
221 if (tmp_loc != -1) {
222 /* We found this uniform name in the program - go ahead and send the data */
223 GL_EXTCALL(glUniform1ivARB(tmp_loc, 1, &constants[i*4]));
224 checkGLcall("glUniform1ivARB");
225 }
226 }
227 }
228
229 /* Load immediate constants */
230 ptr = list_head(&This->baseShader.constantsB);
231 while (ptr) {
232 local_constant* lconst = LIST_ENTRY(ptr, struct local_constant, entry);
233 unsigned int idx = lconst->idx;
234 GLint* values = (GLint*) lconst->value;
235
236 TRACE("Loading local constants %i: %i\n", idx, values[0]);
237
238 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, idx);
239 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
240 if (tmp_loc != -1) {
241 /* We found this uniform name in the program - go ahead and send the data */
242 GL_EXTCALL(glUniform1ivARB(tmp_loc, 1, values));
243 checkGLcall("glUniform1ivARB");
244 }
245 ptr = list_next(&This->baseShader.constantsB, ptr);
246 }
247 }
248
249
250
251 /**
252 * Loads the app-supplied constants into the currently set GLSL program.
253 */
254 void shader_glsl_load_constants(
255 IWineD3DStateBlock* iface,
256 char usePixelShader,
257 char useVertexShader) {
258
259 IWineD3DStateBlockImpl* stateBlock = (IWineD3DStateBlockImpl*) iface;
260 WineD3D_GL_Info *gl_info = &((IWineD3DImpl*)stateBlock->wineD3DDevice->wineD3D)->gl_info;
261 GLhandleARB programId = stateBlock->shaderPrgId;
262
263 if (programId == 0) {
264 /* No GLSL program set - nothing to do. */
265 return;
266 }
267
268 if (useVertexShader) {
269 IWineD3DBaseShaderImpl* vshader = (IWineD3DBaseShaderImpl*) stateBlock->vertexShader;
270 IWineD3DVertexShaderImpl* vshader_impl = (IWineD3DVertexShaderImpl*) vshader;
271
272 IWineD3DVertexDeclarationImpl* vertexDeclaration =
273 (IWineD3DVertexDeclarationImpl*) vshader_impl->vertexDeclaration;
274
275 if (NULL != vertexDeclaration && NULL != vertexDeclaration->constants) {
276 /* Load DirectX 8 float constants/uniforms for vertex shader */
277 shader_glsl_load_constantsF(vshader, gl_info, programId, GL_LIMITS(vshader_constantsF),
278 vertexDeclaration->constants, NULL);
279 }
280
281 /* Load DirectX 9 float constants/uniforms for vertex shader */
282 shader_glsl_load_constantsF(vshader, gl_info, programId, GL_LIMITS(vshader_constantsF),
283 stateBlock->vertexShaderConstantF,
284 stateBlock->set.vertexShaderConstantsF);
285
286 /* Load DirectX 9 integer constants/uniforms for vertex shader */
287 shader_glsl_load_constantsI(vshader, gl_info, programId, MAX_CONST_I,
288 stateBlock->vertexShaderConstantI,
289 stateBlock->set.vertexShaderConstantsI);
290
291 /* Load DirectX 9 boolean constants/uniforms for vertex shader */
292 shader_glsl_load_constantsB(vshader, gl_info, programId, MAX_CONST_B,
293 stateBlock->vertexShaderConstantB,
294 stateBlock->set.vertexShaderConstantsB);
295 }
296
297 if (usePixelShader) {
298
299 IWineD3DBaseShaderImpl* pshader = (IWineD3DBaseShaderImpl*) stateBlock->pixelShader;
300
301 /* Load pixel shader samplers */
302 shader_glsl_load_psamplers(gl_info, iface);
303
304 /* Load DirectX 9 float constants/uniforms for pixel shader */
305 shader_glsl_load_constantsF(pshader, gl_info, programId, GL_LIMITS(pshader_constantsF),
306 stateBlock->pixelShaderConstantF,
307 stateBlock->set.pixelShaderConstantsF);
308
309 /* Load DirectX 9 integer constants/uniforms for pixel shader */
310 shader_glsl_load_constantsI(pshader, gl_info, programId, MAX_CONST_I,
311 stateBlock->pixelShaderConstantI,
312 stateBlock->set.pixelShaderConstantsI);
313
314 /* Load DirectX 9 boolean constants/uniforms for pixel shader */
315 shader_glsl_load_constantsB(pshader, gl_info, programId, MAX_CONST_B,
316 stateBlock->pixelShaderConstantB,
317 stateBlock->set.pixelShaderConstantsB);
318 }
319 }
320
321 /** Generate the variable & register declarations for the GLSL output target */
322 void shader_generate_glsl_declarations(
323 IWineD3DBaseShader *iface,
324 shader_reg_maps* reg_maps,
325 SHADER_BUFFER* buffer,
326 WineD3D_GL_Info* gl_info) {
327
328 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
329 int i;
330
331 /* There are some minor differences between pixel and vertex shaders */
332 char pshader = shader_is_pshader_version(This->baseShader.hex_version);
333 char prefix = pshader ? 'P' : 'V';
334
335 /* Prototype the subroutines */
336 for (i = 0; i < This->baseShader.limits.label; i++) {
337 if (reg_maps->labels[i])
338 shader_addline(buffer, "void subroutine%lu();\n", i);
339 }
340
341 /* Declare the constants (aka uniforms) */
342 if (This->baseShader.limits.constant_float > 0) {
343 unsigned max_constantsF = min(This->baseShader.limits.constant_float,
344 (pshader ? GL_LIMITS(pshader_constantsF) : GL_LIMITS(vshader_constantsF)));
345 shader_addline(buffer, "uniform vec4 %cC[%u];\n", prefix, max_constantsF);
346 }
347
348 if (This->baseShader.limits.constant_int > 0)
349 shader_addline(buffer, "uniform ivec4 %cI[%u];\n", prefix, This->baseShader.limits.constant_int);
350
351 if (This->baseShader.limits.constant_bool > 0)
352 shader_addline(buffer, "uniform bool %cB[%u];\n", prefix, This->baseShader.limits.constant_bool);
353
354 /* Declare texture samplers */
355 for (i = 0; i < This->baseShader.limits.sampler; i++) {
356 if (reg_maps->samplers[i]) {
357
358 DWORD stype = reg_maps->samplers[i] & WINED3DSP_TEXTURETYPE_MASK;
359 switch (stype) {
360
361 case WINED3DSTT_1D:
362 shader_addline(buffer, "uniform sampler1D %csampler%lu;\n", prefix, i);
363 break;
364 case WINED3DSTT_2D:
365 shader_addline(buffer, "uniform sampler2D %csampler%lu;\n", prefix, i);
366 break;
367 case WINED3DSTT_CUBE:
368 shader_addline(buffer, "uniform samplerCube %csampler%lu;\n", prefix, i);
369 break;
370 case WINED3DSTT_VOLUME:
371 shader_addline(buffer, "uniform sampler3D %csampler%lu;\n", prefix, i);
372 break;
373 default:
374 shader_addline(buffer, "uniform unsupported_sampler %csampler%lu;\n", prefix, i);
375 FIXME("Unrecognized sampler type: %#lx\n", stype);
376 break;
377 }
378 }
379 }
380
381 /* Declare address variables */
382 for (i = 0; i < This->baseShader.limits.address; i++) {
383 if (reg_maps->address[i])
384 shader_addline(buffer, "ivec4 A%ld;\n", i);
385 }
386
387 /* Declare texture coordinate temporaries and initialize them */
388 for (i = 0; i < This->baseShader.limits.texcoord; i++) {
389 if (reg_maps->texcoord[i])
390 shader_addline(buffer, "vec4 T%lu = gl_TexCoord[%lu];\n", i, i);
391 }
392
393 /* Declare input register temporaries */
394 for (i=0; i < This->baseShader.limits.packed_input; i++) {
395 if (reg_maps->packed_input[i])
396 shader_addline(buffer, "vec4 IN%lu;\n", i);
397 }
398
399 /* Declare output register temporaries */
400 for (i = 0; i < This->baseShader.limits.packed_output; i++) {
401 if (reg_maps->packed_output[i])
402 shader_addline(buffer, "vec4 OUT%lu;\n", i);
403 }
404
405 /* Declare temporary variables */
406 for(i = 0; i < This->baseShader.limits.temporary; i++) {
407 if (reg_maps->temporary[i])
408 shader_addline(buffer, "vec4 R%lu;\n", i);
409 }
410
411 /* Declare attributes */
412 for (i = 0; i < This->baseShader.limits.attributes; i++) {
413 if (reg_maps->attributes[i])
414 shader_addline(buffer, "attribute vec4 attrib%i;\n", i);
415 }
416
417 /* Declare loop register aL */
418 if (reg_maps->loop) {
419 shader_addline(buffer, "int aL;\n");
420 shader_addline(buffer, "int tmpInt;\n");
421 }
422
423 /* Temporary variables for matrix operations */
424 shader_addline(buffer, "vec4 tmp0;\n");
425 shader_addline(buffer, "vec4 tmp1;\n");
426
427 /* Start the main program */
428 shader_addline(buffer, "void main() {\n");
429 }
430
431 /*****************************************************************************
432 * Functions to generate GLSL strings from DirectX Shader bytecode begin here.
433 *
434 * For more information, see http://wiki.winehq.org/DirectX-Shaders
435 ****************************************************************************/
436
437 /* Prototypes */
438 static void shader_glsl_add_param(
439 SHADER_OPCODE_ARG* arg,
440 const DWORD param,
441 const DWORD addr_token,
442 BOOL is_input,
443 char *reg_name,
444 char *reg_mask,
445 char *out_str);
446
447 /** Used for opcode modifiers - They multiply the result by the specified amount */
448 static const char* shift_glsl_tab[] = {
449 "", /* 0 (none) */
450 "2.0 * ", /* 1 (x2) */
451 "4.0 * ", /* 2 (x4) */
452 "8.0 * ", /* 3 (x8) */
453 "16.0 * ", /* 4 (x16) */
454 "32.0 * ", /* 5 (x32) */
455 "", /* 6 (x64) */
456 "", /* 7 (x128) */
457 "", /* 8 (d256) */
458 "", /* 9 (d128) */
459 "", /* 10 (d64) */
460 "", /* 11 (d32) */
461 "0.0625 * ", /* 12 (d16) */
462 "0.125 * ", /* 13 (d8) */
463 "0.25 * ", /* 14 (d4) */
464 "0.5 * " /* 15 (d2) */
465 };
466
467 /** Print the beginning of the generated GLSL string. example: "reg_name.xyzw = vec4("
468 * Will also change the reg_mask if necessary (not all register types are equal in DX vs GL) */
469 static void shader_glsl_add_dst(DWORD param, const char* reg_name, char* reg_mask, char* outStr) {
470
471 int shift = (param & D3DSP_DSTSHIFT_MASK) >> D3DSP_DSTSHIFT_SHIFT;
472 char cast[6];
473
474 if ((shader_get_regtype(param) == D3DSPR_RASTOUT)
475 && ((param & D3DSP_REGNUM_MASK) != 0)) {
476 /* gl_FogFragCoord or glPointSize - both floats */
477 strcpy(cast, "float");
478 strcpy(reg_mask, "");
479
480 } else if (reg_name[0] == 'A') {
481 /* Address register for vertex shaders (ivec4) */
482 strcpy(cast, "ivec4");
483
484 } else {
485 /* Everything else should be a 4 component float vector */
486 strcpy(cast, "vec4");
487 }
488
489 sprintf(outStr, "%s%s = %s%s(", reg_name, reg_mask, shift_glsl_tab[shift], cast);
490 }
491
492 /* Generate a GLSL parameter that does the input modifier computation and return the input register/mask to use */
493 static void shader_glsl_gen_modifier (
494 const DWORD instr,
495 const char *in_reg,
496 const char *in_regswizzle,
497 char *out_str) {
498
499 out_str[0] = 0;
500
501 if (instr == D3DSIO_TEXKILL)
502 return;
503
504 switch (instr & D3DSP_SRCMOD_MASK) {
505 case D3DSPSM_NONE:
506 sprintf(out_str, "%s%s", in_reg, in_regswizzle);
507 break;
508 case D3DSPSM_NEG:
509 sprintf(out_str, "-%s%s", in_reg, in_regswizzle);
510 break;
511 case D3DSPSM_NOT:
512 sprintf(out_str, "!%s%s", in_reg, in_regswizzle);
513 break;
514 case D3DSPSM_BIAS:
515 sprintf(out_str, "(%s%s - vec4(0.5)%s)", in_reg, in_regswizzle, in_regswizzle);
516 break;
517 case D3DSPSM_BIASNEG:
518 sprintf(out_str, "-(%s%s - vec4(0.5)%s)", in_reg, in_regswizzle, in_regswizzle);
519 break;
520 case D3DSPSM_SIGN:
521 sprintf(out_str, "(2.0 * (%s%s - 0.5))", in_reg, in_regswizzle);
522 break;
523 case D3DSPSM_SIGNNEG:
524 sprintf(out_str, "-(2.0 * (%s%s - 0.5))", in_reg, in_regswizzle);
525 break;
526 case D3DSPSM_COMP:
527 sprintf(out_str, "(1.0 - %s%s)", in_reg, in_regswizzle);
528 break;
529 case D3DSPSM_X2:
530 sprintf(out_str, "(2.0 * %s%s)", in_reg, in_regswizzle);
531 break;
532 case D3DSPSM_X2NEG:
533 sprintf(out_str, "-(2.0 * %s%s)", in_reg, in_regswizzle);
534 break;
535 case D3DSPSM_DZ: /* reg1_db = { reg1.r/b, reg1.g/b, ...} The g & a components are undefined, so we'll leave them alone */
536 sprintf(out_str, "vec4(%s.r / %s.b, %s.g / %s.b, %s.b, %s.a)", in_reg, in_reg, in_reg, in_reg, in_reg, in_reg);
537 break;
538 case D3DSPSM_DW:
539 sprintf(out_str, "vec4(%s.r / %s.a, %s.g / %s.a, %s.b, %s.a)", in_reg, in_reg, in_reg, in_reg, in_reg, in_reg);
540 break;
541 case D3DSPSM_ABS:
542 sprintf(out_str, "abs(%s%s)", in_reg, in_regswizzle);
543 break;
544 case D3DSPSM_ABSNEG:
545 sprintf(out_str, "-abs(%s%s)", in_reg, in_regswizzle);
546 break;
547 default:
548 FIXME("Unhandled modifier %lu\n", (instr & D3DSP_SRCMOD_MASK));
549 sprintf(out_str, "%s%s", in_reg, in_regswizzle);
550 }
551 }
552
553 /** Writes the GLSL variable name that corresponds to the register that the
554 * DX opcode parameter is trying to access */
555 static void shader_glsl_get_register_name(
556 const DWORD param,
557 const DWORD addr_token,
558 char* regstr,
559 BOOL* is_color,
560 SHADER_OPCODE_ARG* arg) {
561
562 /* oPos, oFog and oPts in D3D */
563 const char* hwrastout_reg_names[] = { "gl_Position", "gl_FogFragCoord", "gl_PointSize" };
564
565 DWORD reg = param & D3DSP_REGNUM_MASK;
566 DWORD regtype = shader_get_regtype(param);
567 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) arg->shader;
568 char pshader = shader_is_pshader_version(This->baseShader.hex_version);
569 char tmpStr[50];
570
571 *is_color = FALSE;
572
573 switch (regtype) {
574 case D3DSPR_TEMP:
575 sprintf(tmpStr, "R%lu", reg);
576 break;
577 case D3DSPR_INPUT:
578 if (pshader) {
579 /* Pixel shaders >= 3.0 */
580 if (D3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3)
581 sprintf(tmpStr, "IN%lu", reg);
582 else {
583 if (reg==0)
584 strcpy(tmpStr, "gl_Color");
585 else
586 strcpy(tmpStr, "gl_SecondaryColor");
587 }
588 } else {
589 if (vshader_input_is_color((IWineD3DVertexShader*) This, reg))
590 *is_color = TRUE;
591 sprintf(tmpStr, "attrib%lu", reg);
592 }
593 break;
594 case D3DSPR_CONST:
595 {
596 const char* prefix = pshader? "PC":"VC";
597
598 /* Relative addressing */
599 if (param & D3DVS_ADDRMODE_RELATIVE) {
600
601 /* Relative addressing on shaders 2.0+ have a relative address token,
602 * prior to that, it was hard-coded as "A0.x" because there's only 1 register */
603 if (D3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 2) {
604 char relStr[100], relReg[50], relMask[6];
605 shader_glsl_add_param(arg, addr_token, 0, TRUE, relReg, relMask, relStr);
606 sprintf(tmpStr, "%s[%s + %lu]", prefix, relStr, reg);
607 } else
608 sprintf(tmpStr, "%s[A0.x + %lu]", prefix, reg);
609
610 } else
611 sprintf(tmpStr, "%s[%lu]", prefix, reg);
612
613 break;
614 }
615 case D3DSPR_CONSTINT:
616 if (pshader)
617 sprintf(tmpStr, "PI[%lu]", reg);
618 else
619 sprintf(tmpStr, "VI[%lu]", reg);
620 break;
621 case D3DSPR_CONSTBOOL:
622 if (pshader)
623 sprintf(tmpStr, "PB[%lu]", reg);
624 else
625 sprintf(tmpStr, "VB[%lu]", reg);
626 break;
627 case D3DSPR_TEXTURE: /* case D3DSPR_ADDR: */
628 if (pshader) {
629 sprintf(tmpStr, "T%lu", reg);
630 } else {
631 sprintf(tmpStr, "A%lu", reg);
632 }
633 break;
634 case D3DSPR_LOOP:
635 sprintf(tmpStr, "aL");
636 break;
637 case D3DSPR_SAMPLER:
638 if (pshader)
639 sprintf(tmpStr, "Psampler%lu", reg);
640 else
641 sprintf(tmpStr, "Vsampler%lu", reg);
642 break;
643 case D3DSPR_COLOROUT:
644 sprintf(tmpStr, "gl_FragData[%lu]", reg);
645 if (reg > 0) {
646 /* TODO: See GL_ARB_draw_buffers */
647 FIXME("Unsupported write to render target %lu\n", reg);
648 }
649 break;
650 case D3DSPR_RASTOUT:
651 sprintf(tmpStr, "%s", hwrastout_reg_names[reg]);
652 break;
653 case D3DSPR_DEPTHOUT:
654 sprintf(tmpStr, "gl_FragDepth");
655 break;
656 case D3DSPR_ATTROUT:
657 if (reg == 0) {
658 sprintf(tmpStr, "gl_FrontColor");
659 } else {
660 sprintf(tmpStr, "gl_FrontSecondaryColor");
661 }
662 break;
663 case D3DSPR_TEXCRDOUT:
664 /* Vertex shaders >= 3.0: D3DSPR_OUTPUT */
665 if (D3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3)
666 sprintf(tmpStr, "OUT%lu", reg);
667 else
668 sprintf(tmpStr, "gl_TexCoord[%lu]", reg);
669 break;
670 default:
671 FIXME("Unhandled register name Type(%ld)\n", regtype);
672 sprintf(tmpStr, "unrecognized_register");
673 break;
674 }
675
676 strcat(regstr, tmpStr);
677 }
678
679 /* Writes the GLSL writemask for the destination register */
680 static void shader_glsl_get_output_register_swizzle(
681 const DWORD param,
682 char *write_mask) {
683
684 *write_mask = 0;
685 if ((param & D3DSP_WRITEMASK_ALL) != D3DSP_WRITEMASK_ALL) {
686 strcat(write_mask, ".");
687 if (param & D3DSP_WRITEMASK_0) strcat(write_mask, "x");
688 if (param & D3DSP_WRITEMASK_1) strcat(write_mask, "y");
689 if (param & D3DSP_WRITEMASK_2) strcat(write_mask, "z");
690 if (param & D3DSP_WRITEMASK_3) strcat(write_mask, "w");
691 }
692 }
693
694 static void shader_glsl_get_input_register_swizzle(
695 const DWORD param,
696 BOOL is_color,
697 char *reg_mask) {
698
699 const char swizzle_reg_chars_color_fix[] = "zyxw";
700 const char swizzle_reg_chars[] = "xyzw";
701 const char* swizzle_regs = NULL;
702
703 /** operand input */
704 DWORD swizzle = (param & D3DVS_SWIZZLE_MASK) >> D3DVS_SWIZZLE_SHIFT;
705 DWORD swizzle_x = swizzle & 0x03;
706 DWORD swizzle_y = (swizzle >> 2) & 0x03;
707 DWORD swizzle_z = (swizzle >> 4) & 0x03;
708 DWORD swizzle_w = (swizzle >> 6) & 0x03;
709
710 if (is_color) {
711 swizzle_regs = swizzle_reg_chars_color_fix;
712 } else {
713 swizzle_regs = swizzle_reg_chars;
714 }
715
716 /**
717 * swizzle bits fields:
718 * WWZZYYXX
719 */
720 if ((D3DVS_NOSWIZZLE >> D3DVS_SWIZZLE_SHIFT) == swizzle) { /* D3DVS_NOSWIZZLE == 0xE4 << D3DVS_SWIZZLE_SHIFT */
721 if (is_color) {
722 sprintf(reg_mask, ".%c%c%c%c",
723 swizzle_regs[swizzle_x],
724 swizzle_regs[swizzle_y],
725 swizzle_regs[swizzle_z],
726 swizzle_regs[swizzle_w]);
727 }
728 return ;
729 }
730 if (swizzle_x == swizzle_y &&
731 swizzle_x == swizzle_z &&
732 swizzle_x == swizzle_w)
733 {
734 sprintf(reg_mask, ".%c", swizzle_regs[swizzle_x]);
735 } else {
736 sprintf(reg_mask, ".%c%c%c%c",
737 swizzle_regs[swizzle_x],
738 swizzle_regs[swizzle_y],
739 swizzle_regs[swizzle_z],
740 swizzle_regs[swizzle_w]);
741 }
742 }
743
744 /** From a given parameter token, generate the corresponding GLSL string.
745 * Also, return the actual register name and swizzle in case the
746 * caller needs this information as well. */
747 static void shader_glsl_add_param(
748 SHADER_OPCODE_ARG* arg,
749 const DWORD param,
750 const DWORD addr_token,
751 BOOL is_input,
752 char *reg_name,
753 char *reg_mask,
754 char *out_str) {
755
756 BOOL is_color = FALSE;
757 reg_mask[0] = reg_name[0] = out_str[0] = 0;
758
759 shader_glsl_get_register_name(param, addr_token, reg_name, &is_color, arg);
760
761 if (is_input) {
762 shader_glsl_get_input_register_swizzle(param, is_color, reg_mask);
763 shader_glsl_gen_modifier(param, reg_name, reg_mask, out_str);
764 } else {
765 shader_glsl_get_output_register_swizzle(param, reg_mask);
766 sprintf(out_str, "%s%s", reg_name, reg_mask);
767 }
768 }
769
770 /** Process GLSL instruction modifiers */
771 void shader_glsl_add_instruction_modifiers(SHADER_OPCODE_ARG* arg) {
772
773 DWORD mask = arg->dst & D3DSP_DSTMOD_MASK;
774
775 if (arg->opcode->dst_token && mask != 0) {
776 char dst_reg[50];
777 char dst_mask[6];
778 char dst_str[100];
779
780 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
781
782 if (mask & D3DSPDM_SATURATE) {
783 /* _SAT means to clamp the value of the register to between 0 and 1 */
784 shader_addline(arg->buffer, "%s%s = clamp(%s%s, 0.0, 1.0);\n", dst_reg, dst_mask, dst_reg, dst_mask);
785 }
786 if (mask & D3DSPDM_MSAMPCENTROID) {
787 FIXME("_centroid modifier not handled\n");
788 }
789 if (mask & D3DSPDM_PARTIALPRECISION) {
790 /* MSDN says this modifier can be safely ignored, so that's what we'll do. */
791 }
792 }
793 }
794
795 static inline const char* shader_get_comp_op(
796 const DWORD opcode) {
797
798 DWORD op = (opcode & INST_CONTROLS_MASK) >> INST_CONTROLS_SHIFT;
799 switch (op) {
800 case COMPARISON_GT: return ">";
801 case COMPARISON_EQ: return "==";
802 case COMPARISON_GE: return ">=";
803 case COMPARISON_LT: return "<";
804 case COMPARISON_NE: return "!=";
805 case COMPARISON_LE: return "<=";
806 default:
807 FIXME("Unrecognized comparison value: %lu\n", op);
808 return "(\?\?)";
809 }
810 }
811
812 /*****************************************************************************
813 *
814 * Begin processing individual instruction opcodes
815 *
816 ****************************************************************************/
817
818 /* Generate GLSL arithmetic functions (dst = src1 + src2) */
819 void shader_glsl_arith(SHADER_OPCODE_ARG* arg) {
820
821 CONST SHADER_OPCODE* curOpcode = arg->opcode;
822 SHADER_BUFFER* buffer = arg->buffer;
823 char tmpLine[256];
824 char dst_reg[50], src0_reg[50], src1_reg[50];
825 char dst_mask[6], src0_mask[6], src1_mask[6];
826 char dst_str[100], src0_str[100], src1_str[100];
827
828 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
829 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
830 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
831 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
832 strcat(tmpLine, "vec4(");
833 strcat(tmpLine, src0_str);
834 strcat(tmpLine, ")");
835
836 /* Determine the GLSL operator to use based on the opcode */
837 switch (curOpcode->opcode) {
838 case D3DSIO_MUL: strcat(tmpLine, " * "); break;
839 case D3DSIO_ADD: strcat(tmpLine, " + "); break;
840 case D3DSIO_SUB: strcat(tmpLine, " - "); break;
841 default:
842 FIXME("Opcode %s not yet handled in GLSL\n", curOpcode->name);
843 break;
844 }
845 shader_addline(buffer, "%svec4(%s))%s;\n", tmpLine, src1_str, dst_mask);
846 }
847
848 /* Process the D3DSIO_MOV opcode using GLSL (dst = src) */
849 void shader_glsl_mov(SHADER_OPCODE_ARG* arg) {
850
851 SHADER_BUFFER* buffer = arg->buffer;
852 char tmpLine[256];
853 char dst_str[100], src0_str[100];
854 char dst_reg[50], src0_reg[50];
855 char dst_mask[6], src0_mask[6];
856
857 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
858 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
859 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
860 shader_addline(buffer, "%s%s)%s;\n", tmpLine, src0_str, dst_mask);
861 }
862
863 /* Process the dot product operators DP3 and DP4 in GLSL (dst = dot(src0, src1)) */
864 void shader_glsl_dot(SHADER_OPCODE_ARG* arg) {
865
866 CONST SHADER_OPCODE* curOpcode = arg->opcode;
867 SHADER_BUFFER* buffer = arg->buffer;
868 char tmpDest[100];
869 char dst_str[100], src0_str[100], src1_str[100];
870 char dst_reg[50], src0_reg[50], src1_reg[50];
871 char dst_mask[6], src0_mask[6], src1_mask[6];
872 char cast[6];
873
874 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
875 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
876 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
877
878 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpDest);
879
880 /* Need to cast the src vectors to vec3 for dp3, and vec4 for dp4 */
881 if (curOpcode->opcode == D3DSIO_DP4)
882 strcpy(cast, "vec4(");
883 else
884 strcpy(cast, "vec3(");
885
886 shader_addline(buffer, "%sdot(%s%s), %s%s)))%s;\n",
887 tmpDest, cast, src0_str, cast, src1_str, dst_mask);
888 }
889
890 /* Map the opcode 1-to-1 to the GL code (arg->dst = instruction(src0, src1, ...) */
891 void shader_glsl_map2gl(SHADER_OPCODE_ARG* arg) {
892
893 CONST SHADER_OPCODE* curOpcode = arg->opcode;
894 SHADER_BUFFER* buffer = arg->buffer;
895 char tmpLine[256];
896 char dst_str[100], src_str[100];
897 char dst_reg[50], src_reg[50];
898 char dst_mask[6], src_mask[6];
899 unsigned i;
900
901 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
902
903 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
904
905 /* Determine the GLSL function to use based on the opcode */
906 /* TODO: Possibly make this a table for faster lookups */
907 switch (curOpcode->opcode) {
908 case D3DSIO_MIN: strcat(tmpLine, "min"); break;
909 case D3DSIO_MAX: strcat(tmpLine, "max"); break;
910 case D3DSIO_RSQ: strcat(tmpLine, "inversesqrt"); break;
911 case D3DSIO_ABS: strcat(tmpLine, "abs"); break;
912 case D3DSIO_FRC: strcat(tmpLine, "fract"); break;
913 case D3DSIO_POW: strcat(tmpLine, "pow"); break;
914 case D3DSIO_CRS: strcat(tmpLine, "cross"); break;
915 case D3DSIO_NRM: strcat(tmpLine, "normalize"); break;
916 case D3DSIO_LOGP:
917 case D3DSIO_LOG: strcat(tmpLine, "log2"); break;
918 case D3DSIO_EXP: strcat(tmpLine, "exp2"); break;
919 case D3DSIO_SGE: strcat(tmpLine, "greaterThanEqual"); break;
920 case D3DSIO_SLT: strcat(tmpLine, "lessThan"); break;
921 case D3DSIO_SGN: strcat(tmpLine, "sign"); break;
922 default:
923 FIXME("Opcode %s not yet handled in GLSL\n", curOpcode->name);
924 break;
925 }
926
927 strcat(tmpLine, "(");
928
929 if (curOpcode->num_params > 0) {
930 strcat(tmpLine, "vec4(");
931 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src_reg, src_mask, src_str);
932 strcat(tmpLine, src_str);
933 strcat(tmpLine, ")");
934 for (i = 2; i < curOpcode->num_params; ++i) {
935 strcat(tmpLine, ", vec4(");
936 shader_glsl_add_param(arg, arg->src[i-1], arg->src_addr[i-1], TRUE, src_reg, src_mask, src_str);
937 strcat(tmpLine, src_str);
938 strcat(tmpLine, ")");
939 }
940 }
941 shader_addline(buffer, "%s))%s;\n", tmpLine, dst_mask);
942
943 }
944
945 /** Process the D3DSIO_EXPP instruction in GLSL:
946 * For shader model 1.x, do the following (and honor the writemask, so use a temporary variable):
947 * dst.x = 2^(floor(src))
948 * dst.y = src - floor(src)
949 * dst.z = 2^src (partial precision is allowed, but optional)
950 * dst.w = 1.0;
951 * For 2.0 shaders, just do this (honoring writemask and swizzle):
952 * dst = 2^src; (partial precision is allowed, but optional)
953 */
954 void shader_glsl_expp(SHADER_OPCODE_ARG* arg) {
955
956 char tmpLine[256];
957 char dst_str[100], src_str[100];
958 char dst_reg[50], src_reg[50];
959 char dst_mask[6], src_mask[6];
960 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
961 DWORD hex_version = This->baseShader.hex_version;
962
963 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
964 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src_reg, src_mask, src_str);
965 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
966
967 if (hex_version < D3DPS_VERSION(2,0)) {
968 shader_addline(arg->buffer, "tmp0.x = vec4(exp2(floor(%s))).x;\n", src_str);
969 shader_addline(arg->buffer, "tmp0.y = vec4(%s - floor(%s)).y;\n", src_str, src_str);
970 shader_addline(arg->buffer, "tmp0.z = vec4(exp2(%s)).x;\n", src_str);
971 shader_addline(arg->buffer, "tmp0.w = 1.0;\n");
972 shader_addline(arg->buffer, "%svec4(tmp0))%s;\n", tmpLine, dst_mask);
973 } else {
974 shader_addline(arg->buffer, "%svec4(exp2(%s)))%s;\n", tmpLine, src_str, dst_mask);
975 }
976 }
977
978 /** Process the RCP (reciprocal or inverse) opcode in GLSL (dst = 1 / src) */
979 void shader_glsl_rcp(SHADER_OPCODE_ARG* arg) {
980
981 char tmpLine[256];
982 char dst_str[100], src_str[100];
983 char dst_reg[50], src_reg[50];
984 char dst_mask[6], src_mask[6];
985
986 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
987 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src_reg, src_mask, src_str);
988 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
989 strcat(tmpLine, "1.0 / ");
990 shader_addline(arg->buffer, "%s%s)%s;\n", tmpLine, src_str, dst_mask);
991 }
992
993 /** Process signed comparison opcodes in GLSL. */
994 void shader_glsl_compare(SHADER_OPCODE_ARG* arg) {
995
996 char tmpLine[256];
997 char dst_str[100], src0_str[100], src1_str[100];
998 char dst_reg[50], src0_reg[50], src1_reg[50];
999 char dst_mask[6], src0_mask[6], src1_mask[6];
1000
1001 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1002 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1003 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
1004
1005 /* If we are comparing vectors and not scalars, we should process this through map2gl using the GLSL functions. */
1006 if (strlen(src0_mask) != 2) {
1007 shader_glsl_map2gl(arg);
1008 } else {
1009 char compareStr[3];
1010 compareStr[0] = 0;
1011 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
1012
1013 switch (arg->opcode->opcode) {
1014 case D3DSIO_SLT: strcpy(compareStr, "<"); break;
1015 case D3DSIO_SGE: strcpy(compareStr, ">="); break;
1016 default:
1017 FIXME("Can't handle opcode %s\n", arg->opcode->name);
1018 }
1019 shader_addline(arg->buffer, "%s(float(%s) %s float(%s)) ? 1.0 : 0.0)%s;\n",
1020 tmpLine, src0_str, compareStr, src1_str, dst_mask);
1021 }
1022 }
1023
1024 /** Process CMP instruction in GLSL (dst = src0.x > 0.0 ? src1.x : src2.x), per channel */
1025 void shader_glsl_cmp(SHADER_OPCODE_ARG* arg) {
1026
1027 char tmpLine[256];
1028 char dst_str[100], src0_str[100], src1_str[100], src2_str[100];
1029 char dst_reg[50], src0_reg[50], src1_reg[50], src2_reg[50];
1030 char dst_mask[6], src0_mask[6], src1_mask[6], src2_mask[6];
1031
1032 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1033 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1034 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
1035 shader_glsl_add_param(arg, arg->src[2], arg->src_addr[2], TRUE, src2_reg, src2_mask, src2_str);
1036
1037 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
1038 shader_addline(arg->buffer, "%smix(vec4(%s), vec4(%s), vec4(lessThan(vec4(%s), vec4(0.0)))))%s;\n",
1039 tmpLine, src1_str, src2_str, src0_str, dst_mask);
1040 }
1041
1042 /** Process the CND opcode in GLSL (dst = (src0 < 0.5) ? src1 : src2) */
1043 void shader_glsl_cnd(SHADER_OPCODE_ARG* arg) {
1044
1045 char tmpLine[256];
1046 char dst_str[100], src0_str[100], src1_str[100], src2_str[100];
1047 char dst_reg[50], src0_reg[50], src1_reg[50], src2_reg[50];
1048 char dst_mask[6], src0_mask[6], src1_mask[6], src2_mask[6];
1049
1050 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1051 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1052 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
1053 shader_glsl_add_param(arg, arg->src[2], arg->src_addr[2], TRUE, src2_reg, src2_mask, src2_str);
1054 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
1055 shader_addline(arg->buffer, "%s(%s < 0.5) ? %s : %s)%s;\n",
1056 tmpLine, src0_str, src1_str, src2_str, dst_mask);
1057 }
1058
1059 /** GLSL code generation for D3DSIO_MAD: Multiply the first 2 opcodes, then add the last */
1060 void shader_glsl_mad(SHADER_OPCODE_ARG* arg) {
1061
1062 char tmpLine[256];
1063 char dst_str[100], src0_str[100], src1_str[100], src2_str[100];
1064 char dst_reg[50], src0_reg[50], src1_reg[50], src2_reg[50];
1065 char dst_mask[6], src0_mask[6], src1_mask[6], src2_mask[6];
1066
1067 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1068 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1069 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
1070 shader_glsl_add_param(arg, arg->src[2], arg->src_addr[2], TRUE, src2_reg, src2_mask, src2_str);
1071 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
1072
1073 shader_addline(arg->buffer, "%s(vec4(%s) * vec4(%s)) + vec4(%s))%s;\n",
1074 tmpLine, src0_str, src1_str, src2_str, dst_mask);
1075 }
1076
1077 /** Handles transforming all D3DSIO_M?x? opcodes for
1078 Vertex shaders to GLSL codes */
1079 void shader_glsl_mnxn(SHADER_OPCODE_ARG* arg) {
1080 int i;
1081 int nComponents = 0;
1082 SHADER_OPCODE_ARG tmpArg;
1083
1084 memset(&tmpArg, 0, sizeof(SHADER_OPCODE_ARG));
1085
1086 /* Set constants for the temporary argument */
1087 tmpArg.shader = arg->shader;
1088 tmpArg.buffer = arg->buffer;
1089 tmpArg.src[0] = arg->src[0];
1090 tmpArg.src_addr[0] = arg->src_addr[0];
1091 tmpArg.src_addr[1] = arg->src_addr[1];
1092 tmpArg.reg_maps = arg->reg_maps;
1093
1094 switch(arg->opcode->opcode) {
1095 case D3DSIO_M4x4:
1096 nComponents = 4;
1097 tmpArg.opcode = &IWineD3DVertexShaderImpl_shader_ins[D3DSIO_DP4];
1098 break;
1099 case D3DSIO_M4x3:
1100 nComponents = 3;
1101 tmpArg.opcode = &IWineD3DVertexShaderImpl_shader_ins[D3DSIO_DP4];
1102 break;
1103 case D3DSIO_M3x4:
1104 nComponents = 4;
1105 tmpArg.opcode = &IWineD3DVertexShaderImpl_shader_ins[D3DSIO_DP3];
1106 break;
1107 case D3DSIO_M3x3:
1108 nComponents = 3;
1109 tmpArg.opcode = &IWineD3DVertexShaderImpl_shader_ins[D3DSIO_DP3];
1110 break;
1111 case D3DSIO_M3x2:
1112 nComponents = 2;
1113 tmpArg.opcode = &IWineD3DVertexShaderImpl_shader_ins[D3DSIO_DP3];
1114 break;
1115 default:
1116 break;
1117 }
1118
1119 for (i = 0; i < nComponents; i++) {
1120 tmpArg.dst = ((arg->dst) & ~D3DSP_WRITEMASK_ALL)|(D3DSP_WRITEMASK_0<<i);
1121 tmpArg.src[1] = arg->src[1]+i;
1122 shader_glsl_dot(&tmpArg);
1123 }
1124 }
1125
1126 /**
1127 The LRP instruction performs a component-wise linear interpolation
1128 between the second and third operands using the first operand as the
1129 blend factor. Equation: (dst = src2 * (src1 - src0) + src0)
1130 */
1131 void shader_glsl_lrp(SHADER_OPCODE_ARG* arg) {
1132
1133 char tmpLine[256];
1134 char dst_str[100], src0_str[100], src1_str[100], src2_str[100];
1135 char dst_reg[50], src0_reg[50], src1_reg[50], src2_reg[50];
1136 char dst_mask[6], src0_mask[6], src1_mask[6], src2_mask[6];
1137
1138 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1139 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1140 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
1141 shader_glsl_add_param(arg, arg->src[2], arg->src_addr[2], TRUE, src2_reg, src2_mask, src2_str);
1142
1143 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
1144
1145 shader_addline(arg->buffer, "%s%s + %s * (%s - %s))%s;\n",
1146 tmpLine, src2_str, src0_str, src1_str, src2_str, dst_mask);
1147 }
1148
1149 /** Process the D3DSIO_LIT instruction in GLSL:
1150 * dst.x = dst.w = 1.0
1151 * dst.y = (src0.x > 0) ? src0.x
1152 * dst.z = (src0.x > 0) ? ((src0.y > 0) ? pow(src0.y, src.w) : 0) : 0
1153 * where src.w is clamped at +- 128
1154 */
1155 void shader_glsl_lit(SHADER_OPCODE_ARG* arg) {
1156
1157 char dst_str[100], src0_str[100];
1158 char dst_reg[50], src0_reg[50];
1159 char dst_mask[6], src0_mask[6];
1160
1161 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1162 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1163
1164 shader_addline(arg->buffer,
1165 "%s = vec4(1.0, (%s.x > 0.0 ? %s.x : 0.0), (%s.x > 0.0 ? ((%s.y > 0.0) ? pow(%s.y, clamp(%s.w, -128.0, 128.0)) : 0.0) : 0.0), 1.0)%s;\n",
1166 dst_str, src0_reg, src0_reg, src0_reg, src0_reg, src0_reg, src0_reg, dst_mask);
1167 }
1168
1169 /** Process the D3DSIO_DST instruction in GLSL:
1170 * dst.x = 1.0
1171 * dst.y = src0.x * src0.y
1172 * dst.z = src0.z
1173 * dst.w = src1.w
1174 */
1175 void shader_glsl_dst(SHADER_OPCODE_ARG* arg) {
1176
1177 char dst_str[100], src0_str[100], src1_str[100];
1178 char dst_reg[50], src0_reg[50], src1_reg[50];
1179 char dst_mask[6], src0_mask[6], src1_mask[6];
1180
1181 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1182 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1183 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
1184
1185 shader_addline(arg->buffer, "%s = vec4(1.0, %s.x * %s.y, %s.z, %s.w)%s;\n",
1186 dst_str, src0_reg, src1_reg, src0_reg, src1_reg, dst_mask);
1187 }
1188
1189 /** Process the D3DSIO_SINCOS instruction in GLSL:
1190 * VS 2.0 requires that specific cosine and sine constants be passed to this instruction so the hardware
1191 * can handle it. But, these functions are built-in for GLSL, so we can just ignore the last 2 params.
1192 *
1193 * dst.x = cos(src0.?)
1194 * dst.y = sin(src0.?)
1195 * dst.z = dst.z
1196 * dst.w = dst.w
1197 */
1198 void shader_glsl_sincos(SHADER_OPCODE_ARG* arg) {
1199
1200 char dst_str[100], src0_str[100];
1201 char dst_reg[50], src0_reg[50];
1202 char dst_mask[6], src0_mask[6];
1203
1204 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1205 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1206
1207 shader_addline(arg->buffer, "%s = vec4(cos(%s), sin(%s), %s.z, %s.w)%s;\n",
1208 dst_str, src0_str, src0_str, dst_reg, dst_reg, dst_mask);
1209 }
1210
1211 /** Process the D3DSIO_LOOP instruction in GLSL:
1212 * Start a for() loop where src0.y is the initial value of aL,
1213 * increment aL by src0.z for a total of src0.x iterations.
1214 * Need to use a temporary variable for this operation.
1215 */
1216 void shader_glsl_loop(SHADER_OPCODE_ARG* arg) {
1217
1218 char src1_str[100];
1219 char src1_reg[50];
1220 char src1_mask[6];
1221
1222 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
1223
1224 shader_addline(arg->buffer, "for (tmpInt = 0, aL = %s.y; tmpInt < %s.x; tmpInt++, aL += %s.z) {\n",
1225 src1_reg, src1_reg, src1_reg);
1226 }
1227
1228 void shader_glsl_end(SHADER_OPCODE_ARG* arg) {
1229 shader_addline(arg->buffer, "}\n");
1230 }
1231
1232 void shader_glsl_rep(SHADER_OPCODE_ARG* arg) {
1233
1234 char src0_str[100];
1235 char src0_reg[50];
1236 char src0_mask[6];
1237
1238 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1239 shader_addline(arg->buffer, "for (tmpInt = 0; tmpInt < %s.x; tmpInt++) {\n", src0_reg);
1240 }
1241
1242 void shader_glsl_if(SHADER_OPCODE_ARG* arg) {
1243
1244 char src0_str[100];
1245 char src0_reg[50];
1246 char src0_mask[6];
1247
1248 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1249 shader_addline(arg->buffer, "if (%s) {\n", src0_str);
1250 }
1251
1252 void shader_glsl_ifc(SHADER_OPCODE_ARG* arg) {
1253
1254 char src0_str[100], src1_str[100];
1255 char src0_reg[50], src1_reg[50];
1256 char src0_mask[6], src1_mask[6];
1257
1258 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1259 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
1260
1261 shader_addline(arg->buffer, "if (%s %s %s) {\n",
1262 src0_str, shader_get_comp_op(arg->opcode_token), src1_str);
1263 }
1264
1265 void shader_glsl_else(SHADER_OPCODE_ARG* arg) {
1266 shader_addline(arg->buffer, "} else {\n");
1267 }
1268
1269 void shader_glsl_break(SHADER_OPCODE_ARG* arg) {
1270 shader_addline(arg->buffer, "break;\n");
1271 }
1272
1273 void shader_glsl_breakc(SHADER_OPCODE_ARG* arg) {
1274
1275 char src0_str[100], src1_str[100];
1276 char src0_reg[50], src1_reg[50];
1277 char src0_mask[6], src1_mask[6];
1278
1279 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1280 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
1281
1282 shader_addline(arg->buffer, "if (%s %s %s) break;\n",
1283 src0_str, shader_get_comp_op(arg->opcode_token), src1_str);
1284 }
1285
1286 void shader_glsl_label(SHADER_OPCODE_ARG* arg) {
1287
1288 DWORD snum = (arg->src[0]) & D3DSP_REGNUM_MASK;
1289 shader_addline(arg->buffer, "}\n");
1290 shader_addline(arg->buffer, "void subroutine%lu () {\n", snum);
1291 }
1292
1293 void shader_glsl_call(SHADER_OPCODE_ARG* arg) {
1294 DWORD snum = (arg->src[0]) & D3DSP_REGNUM_MASK;
1295 shader_addline(arg->buffer, "subroutine%lu();\n", snum);
1296 }
1297
1298 void shader_glsl_callnz(SHADER_OPCODE_ARG* arg) {
1299
1300 char src1_str[100];
1301 char src1_reg[50];
1302 char src1_mask[6];
1303
1304 DWORD snum = (arg->src[0]) & D3DSP_REGNUM_MASK;
1305 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
1306 shader_addline(arg->buffer, "if (%s) subroutine%lu();\n", src1_str, snum);
1307 }
1308
1309 /*********************************************
1310 * Pixel Shader Specific Code begins here
1311 ********************************************/
1312 void pshader_glsl_tex(SHADER_OPCODE_ARG* arg) {
1313
1314 /* FIXME: Make this work for more than just 2D textures */
1315
1316 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1317 SHADER_BUFFER* buffer = arg->buffer;
1318 DWORD hex_version = This->baseShader.hex_version;
1319
1320 char dst_str[100], dst_reg[50], dst_mask[6];
1321 char coord_str[100], coord_reg[50], coord_mask[6];
1322 char sampler_str[100], sampler_reg[50], sampler_mask[6];
1323 DWORD reg_dest_code = arg->dst & D3DSP_REGNUM_MASK;
1324 DWORD sampler_code, sampler_type;
1325
1326 /* All versions have a destination register */
1327 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1328
1329 /* 1.0-1.3: Use destination register as coordinate source.
1330 1.4+: Use provided coordinate source register. */
1331 if (hex_version < D3DPS_VERSION(1,4))
1332 strcpy(coord_reg, dst_reg);
1333 else
1334 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, coord_reg, coord_mask, coord_str);
1335
1336 /* 1.0-1.4: Use destination register as coordinate source.
1337 * 2.0+: Use provided coordinate source register. */
1338 if (hex_version < D3DPS_VERSION(2,0)) {
1339 sprintf(sampler_str, "Psampler%lu", reg_dest_code);
1340 sampler_code = reg_dest_code;
1341 }
1342 else {
1343 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, sampler_reg, sampler_mask, sampler_str);
1344 sampler_code = arg->src[1] & D3DSP_REGNUM_MASK;
1345 }
1346
1347 sampler_type = arg->reg_maps->samplers[sampler_code] & WINED3DSP_TEXTURETYPE_MASK;
1348 switch(sampler_type) {
1349
1350 case WINED3DSTT_2D:
1351 shader_addline(buffer, "%s = texture2D(%s, %s.st);\n", dst_str, sampler_str, coord_reg);
1352 break;
1353 case WINED3DSTT_CUBE:
1354 shader_addline(buffer, "%s = textureCube(%s, %s.stp);\n", dst_str, sampler_str, coord_reg);
1355 break;
1356 case WINED3DSTT_VOLUME:
1357 shader_addline(buffer, "%s = texture3D(%s, %s.stp);\n", dst_str, sampler_str, coord_reg);
1358 break;
1359 default:
1360 shader_addline(buffer, "%s = unrecognized_stype(%s, %s.stp);\n", dst_str, sampler_str, coord_reg);
1361 FIXME("Unrecognized sampler type: %#lx;\n", sampler_type);
1362 break;
1363 }
1364 }
1365
1366 void pshader_glsl_texcoord(SHADER_OPCODE_ARG* arg) {
1367
1368 /* FIXME: Make this work for more than just 2D textures */
1369
1370 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1371 SHADER_BUFFER* buffer = arg->buffer;
1372 DWORD hex_version = This->baseShader.hex_version;
1373
1374 char tmpStr[100];
1375 char tmpReg[50];
1376 char tmpMask[6];
1377 tmpReg[0] = 0;
1378
1379 shader_glsl_add_param(arg, arg->dst, 0, FALSE, tmpReg, tmpMask, tmpStr);
1380
1381 if (hex_version != D3DPS_VERSION(1,4)) {
1382 DWORD reg = arg->dst & D3DSP_REGNUM_MASK;
1383 shader_addline(buffer, "%s = clamp(gl_TexCoord[%lu], 0.0, 1.0);\n", tmpReg, reg);
1384 } else {
1385 DWORD reg2 = arg->src[0] & D3DSP_REGNUM_MASK;
1386 shader_addline(buffer, "%s = gl_TexCoord[%lu]%s;\n", tmpStr, reg2, tmpMask);
1387 }
1388 }
1389
1390 /** Process the D3DSIO_TEXDP3TEX instruction in GLSL:
1391 * Take a 3-component dot product of the TexCoord[dstreg] and src,
1392 * then perform a 1D texture lookup from stage dstregnum, place into dst. */
1393 void pshader_glsl_texdp3tex(SHADER_OPCODE_ARG* arg) {
1394
1395 DWORD dstreg = arg->dst & D3DSP_REGNUM_MASK;
1396 char src0_str[100], dst_str[100];
1397 char src0_name[50], dst_name[50];
1398 char src0_mask[6], dst_mask[6];
1399
1400 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_name, dst_mask, dst_str);
1401 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_name, src0_mask, src0_str);
1402
1403 shader_addline(arg->buffer, "tmp0.x = dot(vec3(gl_TexCoord[%lu]), vec3(%s));\n", dstreg, src0_str);
1404 shader_addline(arg->buffer, "%s = vec4(texture1D(Psampler%lu, tmp0.x))%s;\n", dst_str, dstreg, dst_mask);
1405 }
1406
1407 /** Process the D3DSIO_TEXDP3 instruction in GLSL:
1408 * Take a 3-component dot product of the TexCoord[dstreg] and src. */
1409 void pshader_glsl_texdp3(SHADER_OPCODE_ARG* arg) {
1410
1411 DWORD dstreg = arg->dst & D3DSP_REGNUM_MASK;
1412 char src0_str[100], dst_str[100];
1413 char src0_name[50], dst_name[50];
1414 char src0_mask[6], dst_mask[6];
1415
1416 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_name, dst_mask, dst_str);
1417 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_name, src0_mask, src0_str);
1418
1419 shader_addline(arg->buffer, "%s = vec4(dot(vec3(T%lu), vec3(%s)))%s;\n",
1420 dst_str, dstreg, src0_str, dst_mask);
1421 }
1422
1423 /** Process the D3DSIO_TEXDEPTH instruction in GLSL:
1424 * Calculate the depth as dst.x / dst.y */
1425 void pshader_glsl_texdepth(SHADER_OPCODE_ARG* arg) {
1426
1427 char dst_str[100];
1428 char dst_reg[50];
1429 char dst_mask[6];
1430
1431 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1432
1433 shader_addline(arg->buffer, "gl_FragDepth = %s.x / %s.y;\n", dst_reg, dst_reg);
1434 }
1435
1436 /** Process the D3DSIO_TEXM3X2DEPTH instruction in GLSL:
1437 * Last row of a 3x2 matrix multiply, use the result to calculate the depth:
1438 * Calculate tmp0.y = TexCoord[dstreg] . src.xyz; (tmp0.x has already been calculated)
1439 * depth = (tmp0.y == 0.0) ? 1.0 : tmp0.x / tmp0.y
1440 */
1441 void pshader_glsl_texm3x2depth(SHADER_OPCODE_ARG* arg) {
1442
1443 DWORD dstreg = arg->dst & D3DSP_REGNUM_MASK;
1444 char src0_str[100], dst_str[100];
1445 char src0_name[50], dst_name[50];
1446 char src0_mask[6], dst_mask[6];
1447
1448 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_name, dst_mask, dst_str);
1449 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_name, src0_mask, src0_str);
1450
1451 shader_addline(arg->buffer, "tmp0.y = dot(vec3(T%lu), vec3(%s));\n", dstreg, src0_str);
1452 shader_addline(arg->buffer, "gl_FragDepth = vec4((tmp0.y == 0.0) ? 1.0 : tmp0.x / tmp0.y)%s;\n", dst_str, dst_name);
1453 }
1454
1455 /** Process the D3DSIO_TEXM3X2PAD instruction in GLSL
1456 * Calculate the 1st of a 2-row matrix multiplication. */
1457 void pshader_glsl_texm3x2pad(SHADER_OPCODE_ARG* arg) {
1458
1459 DWORD reg = arg->dst & D3DSP_REGNUM_MASK;
1460 SHADER_BUFFER* buffer = arg->buffer;
1461 char src0_str[100];
1462 char src0_name[50];
1463 char src0_mask[6];
1464
1465 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_name, src0_mask, src0_str);
1466 shader_addline(buffer, "tmp0.x = dot(vec3(T%lu), vec3(%s));\n", reg, src0_str);
1467 }
1468
1469 /** Process the D3DSIO_TEXM3X3PAD instruction in GLSL
1470 * Calculate the 1st or 2nd row of a 3-row matrix multiplication. */
1471 void pshader_glsl_texm3x3pad(SHADER_OPCODE_ARG* arg) {
1472
1473 IWineD3DPixelShaderImpl* shader = (IWineD3DPixelShaderImpl*) arg->shader;
1474 DWORD reg = arg->dst & D3DSP_REGNUM_MASK;
1475 SHADER_BUFFER* buffer = arg->buffer;
1476 SHADER_PARSE_STATE* current_state = &shader->baseShader.parse_state;
1477 char src0_str[100];
1478 char src0_name[50];
1479 char src0_mask[6];
1480
1481 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_name, src0_mask, src0_str);
1482 shader_addline(buffer, "tmp0.%c = dot(vec3(T%lu), vec3(%s));\n", 'x' + current_state->current_row, reg, src0_str);
1483 current_state->texcoord_w[current_state->current_row++] = reg;
1484 }
1485
1486 void pshader_glsl_texm3x2tex(SHADER_OPCODE_ARG* arg) {
1487
1488 /* FIXME: Make this work for more than just 2D textures */
1489
1490 DWORD reg = arg->dst & D3DSP_REGNUM_MASK;
1491 SHADER_BUFFER* buffer = arg->buffer;
1492 char src0_str[100];
1493 char src0_name[50];
1494 char src0_mask[6];
1495
1496 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_name, src0_mask, src0_str);
1497 shader_addline(buffer, "tmp0.y = dot(vec3(T%lu), vec3(%s));\n", reg, src0_str);
1498 shader_addline(buffer, "T%lu = texture2D(Psampler%lu, tmp0.st);\n", reg, reg);
1499 }
1500
1501 /** Process the D3DSIO_TEXM3X3TEX instruction in GLSL
1502 * Perform the 3rd row of a 3x3 matrix multiply, then sample the texture using the calculate coordinates */
1503 void pshader_glsl_texm3x3tex(SHADER_OPCODE_ARG* arg) {
1504
1505 char src0_str[100];
1506 char src0_name[50];
1507 char src0_mask[6];
1508 char dimensions[5];
1509 DWORD reg = arg->dst & D3DSP_REGNUM_MASK;
1510 DWORD src0_regnum = arg->src[0] & D3DSP_REGNUM_MASK;
1511 DWORD stype = arg->reg_maps->samplers[src0_regnum] & WINED3DSP_TEXTURETYPE_MASK;
1512 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1513 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1514
1515 switch (stype) {
1516 case WINED3DSTT_2D: strcpy(dimensions, "2D"); break;
1517 case WINED3DSTT_CUBE: strcpy(dimensions, "Cube"); break;
1518 case WINED3DSTT_VOLUME: strcpy(dimensions, "3D"); break;
1519 default:
1520 strcpy(dimensions, "");
1521 FIXME("Unrecognized sampler type: %#lx\n", stype);
1522 break;
1523 }
1524
1525 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_name, src0_mask, src0_str);
1526 shader_addline(arg->buffer, "tmp0.z = dot(vec3(T%lu), vec3(%s));\n", reg, src0_str);
1527 shader_addline(arg->buffer, "T%lu = texture%s(Psampler%lu, tmp0.%s);\n",
1528 reg, dimensions, reg, (stype == WINED3DSTT_2D) ? "xy" : "xyz");
1529 current_state->current_row = 0;
1530 }
1531
1532 /** Process the D3DSIO_TEXM3X3 instruction in GLSL
1533 * Perform the 3rd row of a 3x3 matrix multiply */
1534 void pshader_glsl_texm3x3(SHADER_OPCODE_ARG* arg) {
1535
1536 char src0_str[100];
1537 char src0_name[50];
1538 char src0_mask[6];
1539 DWORD reg = arg->dst & D3DSP_REGNUM_MASK;
1540 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1541 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1542
1543 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_name, src0_mask, src0_str);
1544
1545 shader_addline(arg->buffer, "tmp0.z = dot(vec3(T%lu), vec3(%s));\n", reg, src0_str);
1546 shader_addline(arg->buffer, "T%lu = vec4(tmp0.x, tmp0.y, tmp0.z, 1.0);\n", reg);
1547 current_state->current_row = 0;
1548 }
1549
1550 /** Process the D3DSIO_TEXM3X3SPEC instruction in GLSL
1551 * Peform the final texture lookup based on the previous 2 3x3 matrix multiplies */
1552 void pshader_glsl_texm3x3spec(SHADER_OPCODE_ARG* arg) {
1553
1554 IWineD3DPixelShaderImpl* shader = (IWineD3DPixelShaderImpl*) arg->shader;
1555 DWORD reg = arg->dst & D3DSP_REGNUM_MASK;
1556 char dimensions[5];
1557 char src0_str[100], src0_name[50], src0_mask[6];
1558 char src1_str[100], src1_name[50], src1_mask[6];
1559 SHADER_BUFFER* buffer = arg->buffer;
1560 SHADER_PARSE_STATE* current_state = &shader->baseShader.parse_state;
1561 DWORD stype = arg->reg_maps->samplers[reg] & WINED3DSP_TEXTURETYPE_MASK;
1562
1563 switch (stype) {
1564 case WINED3DSTT_2D: strcpy(dimensions, "2D"); break;
1565 case WINED3DSTT_CUBE: strcpy(dimensions, "Cube"); break;
1566 case WINED3DSTT_VOLUME: strcpy(dimensions, "3D"); break;
1567 default:
1568 strcpy(dimensions, "");
1569 FIXME("Unrecognized sampler type: %#lx\n", stype);
1570 break;
1571 }
1572
1573 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_name, src0_mask, src0_str);
1574 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_name, src1_mask, src1_str);
1575
1576 /* Perform the last matrix multiply operation */
1577 shader_addline(buffer, "tmp0.z = dot(vec3(T%lu), vec3(%s));\n", reg, src0_str);
1578
1579 /* Calculate reflection vector */
1580 shader_addline(buffer, "tmp0.xyz = reflect(-vec3(%s), vec3(tmp0));\n", src1_str);
1581
1582 /* Sample the texture */
1583 shader_addline(buffer, "T%lu = texture%s(Psampler%lu, tmp0.%s);\n",
1584 reg, dimensions, reg, (stype == WINED3DSTT_2D) ? "xy" : "xyz");
1585 current_state->current_row = 0;
1586 }
1587
1588 /** Process the D3DSIO_TEXM3X3VSPEC instruction in GLSL
1589 * Peform the final texture lookup based on the previous 2 3x3 matrix multiplies */
1590 void pshader_glsl_texm3x3vspec(SHADER_OPCODE_ARG* arg) {
1591
1592 IWineD3DPixelShaderImpl* shader = (IWineD3DPixelShaderImpl*) arg->shader;
1593 DWORD reg = arg->dst & D3DSP_REGNUM_MASK;
1594 SHADER_BUFFER* buffer = arg->buffer;
1595 SHADER_PARSE_STATE* current_state = &shader->baseShader.parse_state;
1596 char src0_str[100], src0_name[50], src0_mask[6];
1597
1598 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_name, src0_mask, src0_str);
1599
1600 /* Perform the last matrix multiply operation */
1601 shader_addline(buffer, "tmp0.z = dot(vec3(T%lu), vec3(%s));\n", reg, src0_str);
1602
1603 /* Construct the eye-ray vector from w coordinates */
1604 shader_addline(buffer, "tmp1.x = gl_TexCoord[%lu].w;\n", current_state->texcoord_w[0]);
1605 shader_addline(buffer, "tmp1.y = gl_TexCoord[%lu].w;\n", current_state->texcoord_w[1]);
1606 shader_addline(buffer, "tmp1.z = gl_TexCoord[%lu].w;\n", reg);
1607
1608 /* Calculate reflection vector (Assume normal is normalized): RF = 2*(N.E)*N -E */
1609 shader_addline(buffer, "tmp0.x = dot(vec3(tmp0), vec3(tmp1));\n");
1610 shader_addline(buffer, "tmp0 = tmp0.w * tmp0;\n");
1611 shader_addline(buffer, "tmp0 = (2.0 * tmp0) - tmp1;\n");
1612
1613 /* FIXME:
1614 * We don't really know if a Cube or a Volume texture is being sampled, but since Cube textures
1615 * are used more commonly, we'll default to that.
1616 * We probably need to push back the pixel shader generation code until drawPrimitive() for
1617 * shader versions < 2.0, since that's the only time we can guarantee that we're sampling
1618 * the correct type of texture because we can lookup what textures are bound at that point.
1619 */
1620 shader_addline(buffer, "T%lu = textureCube(Psampler%lu, tmp0.xyz);\n", reg, reg);
1621 current_state->current_row = 0;
1622 }
1623
1624 /** Process the D3DSIO_TEXBEM instruction in GLSL.
1625 * Apply a fake bump map transform.
1626 * FIXME: Should apply the BUMPMAPENV matrix. For now, just sample the texture */
1627 void pshader_glsl_texbem(SHADER_OPCODE_ARG* arg) {
1628
1629 DWORD reg1 = arg->dst & D3DSP_REGNUM_MASK;
1630 DWORD reg2 = arg->src[0] & D3DSP_REGNUM_MASK;
1631
1632 FIXME("Not applying the BUMPMAPENV matrix for pixel shader instruction texbem.\n");
1633 shader_addline(arg->buffer, "T%lu = texture2D(Psampler%lu, gl_TexCoord[%lu].xy + T%lu.xy);\n",
1634 reg1, reg1, reg1, reg2);
1635 }
1636
1637 /** Process the D3DSIO_TEXREG2AR instruction in GLSL
1638 * Sample 2D texture at dst using the alpha & red (wx) components of src as texture coordinates */
1639 void pshader_glsl_texreg2ar(SHADER_OPCODE_ARG* arg) {
1640
1641 char tmpLine[255];
1642 char dst_str[100], src0_str[100];
1643 char dst_reg[50], src0_reg[50];
1644 char dst_mask[6], src0_mask[6];
1645 DWORD src0_regnum = arg->src[0] & D3DSP_REGNUM_MASK;
1646
1647 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1648 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1649
1650 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
1651 shader_addline(arg->buffer, "%stexture2D(Psampler%lu, %s.yz))%s;\n",
1652 tmpLine, src0_regnum, dst_reg, dst_mask);
1653 }
1654
1655 /** Process the D3DSIO_TEXREG2GB instruction in GLSL
1656 * Sample 2D texture at dst using the green & blue (yz) components of src as texture coordinates */
1657 void pshader_glsl_texreg2gb(SHADER_OPCODE_ARG* arg) {
1658
1659 char tmpLine[255];
1660 char dst_str[100], src0_str[100];
1661 char dst_reg[50], src0_reg[50];
1662 char dst_mask[6], src0_mask[6];
1663 DWORD src0_regnum = arg->src[0] & D3DSP_REGNUM_MASK;
1664
1665 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1666 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1667
1668 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
1669 shader_addline(arg->buffer, "%stexture2D(Psampler%lu, %s.yz))%s;\n",
1670 tmpLine, src0_regnum, dst_reg, dst_mask);
1671 }
1672
1673 /** Process the D3DSIO_TEXREG2RGB instruction in GLSL
1674 * Sample texture at dst using the rgb (xyz) components of src as texture coordinates */
1675 void pshader_glsl_texreg2rgb(SHADER_OPCODE_ARG* arg) {
1676
1677 char tmpLine[255];
1678 char dst_str[100], src0_str[100];
1679 char dst_reg[50], src0_reg[50];
1680 char dst_mask[6], src0_mask[6];
1681 char dimensions[5];
1682 DWORD src0_regnum = arg->src[0] & D3DSP_REGNUM_MASK;
1683 DWORD stype = arg->reg_maps->samplers[src0_regnum] & WINED3DSP_TEXTURETYPE_MASK;
1684 switch (stype) {
1685 case WINED3DSTT_2D: strcpy(dimensions, "2D"); break;
1686 case WINED3DSTT_CUBE: strcpy(dimensions, "Cube"); break;
1687 case WINED3DSTT_VOLUME: strcpy(dimensions, "3D"); break;
1688 default:
1689 strcpy(dimensions, "");
1690 FIXME("Unrecognized sampler type: %#lx\n", stype);
1691 break;
1692 }
1693
1694 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1695 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1696
1697 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
1698 shader_addline(arg->buffer, "%stexture%s(Psampler%lu, %s.%s))%s;\n",
1699 tmpLine, dimensions, src0_regnum, dst_reg, (stype == WINED3DSTT_2D) ? "xy" : "xyz", dst_mask);
1700 }
1701
1702 /** Process the D3DSIO_TEXKILL instruction in GLSL.
1703 * If any of the first 3 components are < 0, discard this pixel */
1704 void pshader_glsl_texkill(SHADER_OPCODE_ARG* arg) {
1705
1706 char dst_str[100], dst_name[50], dst_mask[6];
1707
1708 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_name, dst_mask, dst_str);
1709 shader_addline(arg->buffer, "if (any(lessThan(%s.xyz, vec3(0.0)))) discard;\n", dst_name);
1710 }
1711
1712 /** Process the D3DSIO_DP2ADD instruction in GLSL.
1713 * dst = dot2(src0, src1) + src2 */
1714 void pshader_glsl_dp2add(SHADER_OPCODE_ARG* arg) {
1715
1716 char tmpLine[256];
1717 char dst_str[100], src0_str[100], src1_str[100], src2_str[100];
1718 char dst_reg[50], src0_reg[50], src1_reg[50], src2_reg[50];
1719 char dst_mask[6], src0_mask[6], src1_mask[6], src2_mask[6];
1720
1721 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1722 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1723 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
1724 shader_glsl_add_param(arg, arg->src[2], arg->src_addr[2], TRUE, src2_reg, src2_mask, src2_str);
1725 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
1726 shader_addline(arg->buffer, "%sdot(vec2(%s), vec2(%s)) + %s)%s;\n",
1727 tmpLine, src0_str, src1_str, src2_str, dst_mask);
1728 }
1729
1730 void pshader_glsl_input_pack(
1731 SHADER_BUFFER* buffer,
1732 semantic* semantics_in) {
1733
1734 unsigned int i;
1735
1736 for (i = 0; i < MAX_REG_INPUT; i++) {
1737
1738 DWORD usage_token = semantics_in[i].usage;
1739 DWORD register_token = semantics_in[i].reg;
1740 DWORD usage, usage_idx;
1741 char reg_mask[6];
1742
1743 /* Uninitialized */
1744 if (!usage_token) continue;
1745 usage = (usage_token & D3DSP_DCL_USAGE_MASK) >> D3DSP_DCL_USAGE_SHIFT;
1746 usage_idx = (usage_token & D3DSP_DCL_USAGEINDEX_MASK) >> D3DSP_DCL_USAGEINDEX_SHIFT;
1747 shader_glsl_get_output_register_swizzle(register_token, reg_mask);
1748
1749 switch(usage) {
1750
1751 case D3DDECLUSAGE_COLOR:
1752 if (usage_idx == 0)
1753 shader_addline(buffer, "IN%lu%s = vec4(gl_Color)%s;\n",
1754 i, reg_mask, reg_mask);
1755 else if (usage_idx == 1)
1756 shader_addline(buffer, "IN%lu%s = vec4(gl_SecondaryColor)%s;\n",
1757 i, reg_mask, reg_mask);
1758 else
1759 shader_addline(buffer, "IN%lu%s = vec4(unsupported_color_input)%s;\n",
1760 i, reg_mask, reg_mask);
1761 break;
1762
1763 case D3DDECLUSAGE_TEXCOORD:
1764 shader_addline(buffer, "IN%lu%s = vec4(gl_TexCoord[%lu])%s;\n",
1765 i, reg_mask, usage_idx, reg_mask );
1766 break;
1767
1768 case D3DDECLUSAGE_FOG:
1769 shader_addline(buffer, "IN%lu%s = vec4(gl_FogFragCoord)%s;\n",
1770 i, reg_mask, reg_mask);
1771 break;
1772
1773 default:
1774 shader_addline(buffer, "IN%lu%s = vec4(unsupported_input)%s;\n",
1775 i, reg_mask, reg_mask);
1776 }
1777 }
1778 }
1779
1780 /*********************************************
1781 * Vertex Shader Specific Code begins here
1782 ********************************************/
1783
1784 void vshader_glsl_output_unpack(
1785 SHADER_BUFFER* buffer,
1786 semantic* semantics_out) {
1787
1788 unsigned int i;
1789
1790 for (i = 0; i < MAX_REG_OUTPUT; i++) {
1791
1792 DWORD usage_token = semantics_out[i].usage;
1793 DWORD register_token = semantics_out[i].reg;
1794 DWORD usage, usage_idx;
1795 char reg_mask[6];
1796
1797 /* Uninitialized */
1798 if (!usage_token) continue;
1799
1800 usage = (usage_token & D3DSP_DCL_USAGE_MASK) >> D3DSP_DCL_USAGE_SHIFT;
1801 usage_idx = (usage_token & D3DSP_DCL_USAGEINDEX_MASK) >> D3DSP_DCL_USAGEINDEX_SHIFT;
1802 shader_glsl_get_output_register_swizzle(register_token, reg_mask);
1803
1804 switch(usage) {
1805
1806 case D3DDECLUSAGE_COLOR:
1807 if (usage_idx == 0)
1808 shader_addline(buffer, "gl_FrontColor%s = OUT%lu%s;\n", reg_mask, i, reg_mask);
1809 else if (usage_idx == 1)
1810 shader_addline(buffer, "gl_FrontSecondaryColor%s = OUT%lu%s;\n", reg_mask, i, reg_mask);
1811 else
1812 shader_addline(buffer, "unsupported_color_output%s = OUT%lu%s;\n", reg_mask, i, reg_mask);
1813 break;
1814
1815 case D3DDECLUSAGE_POSITION:
1816 shader_addline(buffer, "gl_Position%s = OUT%lu%s;\n", reg_mask, i, reg_mask);
1817 break;
1818
1819 case D3DDECLUSAGE_TEXCOORD:
1820 shader_addline(buffer, "gl_TexCoord[%lu]%s = OUT%lu%s;\n",
1821 usage_idx, reg_mask, i, reg_mask);
1822 break;
1823
1824 case WINED3DSHADERDECLUSAGE_PSIZE:
1825 shader_addline(buffer, "gl_PointSize = OUT%lu.x;\n", i);
1826 break;
1827
1828 case WINED3DSHADERDECLUSAGE_FOG:
1829 shader_addline(buffer, "gl_FogFragCoord%s = OUT%lu%s;\n", reg_mask, i, reg_mask);
1830 break;
1831
1832 default:
1833 shader_addline(buffer, "unsupported_output%s = OUT%lu%s;\n", reg_mask, i, reg_mask);
1834 }
1835 }
1836 }
Windows API implementation