?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
wined3d: Remove dead NV_TEXTURE_SHADER code
[wine/multimedia.git] / dlls / wined3d / glsl_shader.c
blobfd441110d9817520f5581b8a52b546de50b5cf7c
1 /*
2 * GLSL pixel and vertex shader implementation
3 *
4 * Copyright 2006 Jason Green
5 * Copyright 2006 Henri Verbeet
6 *
7 * This library is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
11 *
12 * This library is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
16 *
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with this library; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
20 */
21
22 #include "config.h"
23 #include <stdio.h>
24 #include "wined3d_private.h"
25
26 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
27
28 #define GLINFO_LOCATION (*gl_info)
29
30 /** Prints the GLSL info log which will contain error messages if they exist */
31 void print_glsl_info_log(WineD3D_GL_Info *gl_info, GLhandleARB obj) {
32
33 int infologLength = 0;
34 char *infoLog;
35
36 GL_EXTCALL(glGetObjectParameterivARB(obj,
37 GL_OBJECT_INFO_LOG_LENGTH_ARB,
38 &infologLength));
39
40 /* A size of 1 is just a null-terminated string, so the log should be bigger than
41 * that if there are errors. */
42 if (infologLength > 1)
43 {
44 infoLog = (char *)HeapAlloc(GetProcessHeap(), 0, infologLength);
45 GL_EXTCALL(glGetInfoLogARB(obj, infologLength, NULL, infoLog));
46 FIXME("Error received from GLSL shader #%u: %s\n", obj, debugstr_a(infoLog));
47 HeapFree(GetProcessHeap(), 0, infoLog);
48 }
49 }
50
51 /**
52 * Loads (pixel shader) samplers
53 */
54 void shader_glsl_load_psamplers(
55 WineD3D_GL_Info *gl_info,
56 IWineD3DStateBlock* iface) {
57
58 IWineD3DStateBlockImpl* stateBlock = (IWineD3DStateBlockImpl*) iface;
59 GLhandleARB programId = stateBlock->glsl_program->programId;
60 GLhandleARB name_loc;
61 int i;
62 char sampler_name[20];
63
64 for (i=0; i< GL_LIMITS(samplers); ++i) {
65 if (stateBlock->textures[i] != NULL) {
66 snprintf(sampler_name, sizeof(sampler_name), "Psampler%d", i);
67 name_loc = GL_EXTCALL(glGetUniformLocationARB(programId, sampler_name));
68 if (name_loc != -1) {
69 TRACE("Loading %s for texture %d\n", sampler_name, i);
70 GL_EXTCALL(glUniform1iARB(name_loc, i));
71 checkGLcall("glUniform1iARB");
72 }
73 }
74 }
75 }
76
77 /**
78 * Loads floating point constants (aka uniforms) into the currently set GLSL program.
79 * When constant_list == NULL, it will load all the constants.
80 */
81 static void shader_glsl_load_constantsF(IWineD3DBaseShaderImpl* This, WineD3D_GL_Info *gl_info,
82 unsigned int max_constants, float* constants, GLhandleARB *constant_locations,
83 struct list *constant_list) {
84 local_constant* lconst;
85 GLhandleARB tmp_loc;
86 int i;
87
88 if (!constant_list) {
89 if (TRACE_ON(d3d_shader)) {
90 for (i = 0; i < max_constants; ++i) {
91 tmp_loc = constant_locations[i];
92 if (tmp_loc != -1) {
93 TRACE("Loading constants %i: %f, %f, %f, %f\n", i,
94 constants[i * 4 + 0], constants[i * 4 + 1],
95 constants[i * 4 + 2], constants[i * 4 + 3]);
96 }
97 }
98 }
99 for (i = 0; i < max_constants; ++i) {
100 tmp_loc = constant_locations[i];
101 if (tmp_loc != -1) {
102 /* We found this uniform name in the program - go ahead and send the data */
103 GL_EXTCALL(glUniform4fvARB(tmp_loc, 1, constants + (i * 4)));
104 }
105 }
106 checkGLcall("glUniform4fvARB()");
107 } else {
108 constant_entry *constant;
109 if (TRACE_ON(d3d_shader)) {
110 LIST_FOR_EACH_ENTRY(constant, constant_list, constant_entry, entry) {
111 i = constant->idx;
112 tmp_loc = constant_locations[i];
113 if (tmp_loc != -1) {
114 TRACE("Loading constants %i: %f, %f, %f, %f\n", i,
115 constants[i * 4 + 0], constants[i * 4 + 1],
116 constants[i * 4 + 2], constants[i * 4 + 3]);
117 }
118 }
119 }
120 LIST_FOR_EACH_ENTRY(constant, constant_list, constant_entry, entry) {
121 i = constant->idx;
122 tmp_loc = constant_locations[i];
123 if (tmp_loc != -1) {
124 /* We found this uniform name in the program - go ahead and send the data */
125 GL_EXTCALL(glUniform4fvARB(tmp_loc, 1, constants + (i * 4)));
126 }
127 }
128 checkGLcall("glUniform4fvARB()");
129 }
130
131 /* Load immediate constants */
132 if (TRACE_ON(d3d_shader)) {
133 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
134 tmp_loc = constant_locations[lconst->idx];
135 if (tmp_loc != -1) {
136 GLfloat* values = (GLfloat*)lconst->value;
137 TRACE("Loading local constants %i: %f, %f, %f, %f\n", lconst->idx,
138 values[0], values[1], values[2], values[3]);
139 }
140 }
141 }
142 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
143 tmp_loc = constant_locations[lconst->idx];
144 if (tmp_loc != -1) {
145 /* We found this uniform name in the program - go ahead and send the data */
146 GL_EXTCALL(glUniform4fvARB(tmp_loc, 1, (GLfloat*)lconst->value));
147 }
148 }
149 checkGLcall("glUniform4fvARB()");
150 }
151
152 /**
153 * Loads integer constants (aka uniforms) into the currently set GLSL program.
154 * When @constants_set == NULL, it will load all the constants.
155 */
156 void shader_glsl_load_constantsI(
157 IWineD3DBaseShaderImpl* This,
158 WineD3D_GL_Info *gl_info,
159 GLhandleARB programId,
160 unsigned max_constants,
161 int* constants,
162 BOOL* constants_set) {
163
164 GLhandleARB tmp_loc;
165 int i;
166 char tmp_name[8];
167 char is_pshader = shader_is_pshader_version(This->baseShader.hex_version);
168 const char* prefix = is_pshader? "PI":"VI";
169 struct list* ptr;
170
171 for (i=0; i<max_constants; ++i) {
172 if (NULL == constants_set || constants_set[i]) {
173
174 TRACE("Loading constants %i: %i, %i, %i, %i\n",
175 i, constants[i*4], constants[i*4+1], constants[i*4+2], constants[i*4+3]);
176
177 /* TODO: Benchmark and see if it would be beneficial to store the
178 * locations of the constants to avoid looking up each time */
179 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, i);
180 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
181 if (tmp_loc != -1) {
182 /* We found this uniform name in the program - go ahead and send the data */
183 GL_EXTCALL(glUniform4ivARB(tmp_loc, 1, &constants[i*4]));
184 checkGLcall("glUniform4ivARB");
185 }
186 }
187 }
188
189 /* Load immediate constants */
190 ptr = list_head(&This->baseShader.constantsI);
191 while (ptr) {
192 local_constant* lconst = LIST_ENTRY(ptr, struct local_constant, entry);
193 unsigned int idx = lconst->idx;
194 GLint* values = (GLint*) lconst->value;
195
196 TRACE("Loading local constants %i: %i, %i, %i, %i\n", idx,
197 values[0], values[1], values[2], values[3]);
198
199 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, idx);
200 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
201 if (tmp_loc != -1) {
202 /* We found this uniform name in the program - go ahead and send the data */
203 GL_EXTCALL(glUniform4ivARB(tmp_loc, 1, values));
204 checkGLcall("glUniform4ivARB");
205 }
206 ptr = list_next(&This->baseShader.constantsI, ptr);
207 }
208 }
209
210 /**
211 * Loads boolean constants (aka uniforms) into the currently set GLSL program.
212 * When @constants_set == NULL, it will load all the constants.
213 */
214 void shader_glsl_load_constantsB(
215 IWineD3DBaseShaderImpl* This,
216 WineD3D_GL_Info *gl_info,
217 GLhandleARB programId,
218 unsigned max_constants,
219 BOOL* constants,
220 BOOL* constants_set) {
221
222 GLhandleARB tmp_loc;
223 int i;
224 char tmp_name[8];
225 char is_pshader = shader_is_pshader_version(This->baseShader.hex_version);
226 const char* prefix = is_pshader? "PB":"VB";
227 struct list* ptr;
228
229 for (i=0; i<max_constants; ++i) {
230 if (NULL == constants_set || constants_set[i]) {
231
232 TRACE("Loading constants %i: %i;\n", i, constants[i*4]);
233
234 /* TODO: Benchmark and see if it would be beneficial to store the
235 * locations of the constants to avoid looking up each time */
236 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, i);
237 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
238 if (tmp_loc != -1) {
239 /* We found this uniform name in the program - go ahead and send the data */
240 GL_EXTCALL(glUniform1ivARB(tmp_loc, 1, &constants[i*4]));
241 checkGLcall("glUniform1ivARB");
242 }
243 }
244 }
245
246 /* Load immediate constants */
247 ptr = list_head(&This->baseShader.constantsB);
248 while (ptr) {
249 local_constant* lconst = LIST_ENTRY(ptr, struct local_constant, entry);
250 unsigned int idx = lconst->idx;
251 GLint* values = (GLint*) lconst->value;
252
253 TRACE("Loading local constants %i: %i\n", idx, values[0]);
254
255 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, idx);
256 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
257 if (tmp_loc != -1) {
258 /* We found this uniform name in the program - go ahead and send the data */
259 GL_EXTCALL(glUniform1ivARB(tmp_loc, 1, values));
260 checkGLcall("glUniform1ivARB");
261 }
262 ptr = list_next(&This->baseShader.constantsB, ptr);
263 }
264 }
265
266
267
268 /**
269 * Loads the app-supplied constants into the currently set GLSL program.
270 */
271 void shader_glsl_load_constants(
272 IWineD3DDevice* device,
273 char usePixelShader,
274 char useVertexShader) {
275
276 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) device;
277 IWineD3DStateBlockImpl* stateBlock = deviceImpl->stateBlock;
278 WineD3D_GL_Info *gl_info = &((IWineD3DImpl*) deviceImpl->wineD3D)->gl_info;
279
280 GLhandleARB *constant_locations;
281 struct list *constant_list;
282 GLhandleARB programId;
283
284 if (!stateBlock->glsl_program) {
285 /* No GLSL program set - nothing to do. */
286 return;
287 }
288 programId = stateBlock->glsl_program->programId;
289
290 if (useVertexShader) {
291 IWineD3DBaseShaderImpl* vshader = (IWineD3DBaseShaderImpl*) stateBlock->vertexShader;
292 IWineD3DVertexShaderImpl* vshader_impl = (IWineD3DVertexShaderImpl*) vshader;
293 GLint pos;
294
295 IWineD3DVertexDeclarationImpl* vertexDeclaration =
296 (IWineD3DVertexDeclarationImpl*) vshader_impl->vertexDeclaration;
297
298 constant_locations = stateBlock->glsl_program->vuniformF_locations;
299 constant_list = &stateBlock->set_vconstantsF;
300
301 if (NULL != vertexDeclaration && NULL != vertexDeclaration->constants) {
302 /* Load DirectX 8 float constants/uniforms for vertex shader */
303 shader_glsl_load_constantsF(vshader, gl_info, GL_LIMITS(vshader_constantsF),
304 vertexDeclaration->constants, constant_locations, NULL);
305 }
306
307 /* Load DirectX 9 float constants/uniforms for vertex shader */
308 shader_glsl_load_constantsF(vshader, gl_info, GL_LIMITS(vshader_constantsF),
309 stateBlock->vertexShaderConstantF, constant_locations, constant_list);
310
311 /* Load DirectX 9 integer constants/uniforms for vertex shader */
312 shader_glsl_load_constantsI(vshader, gl_info, programId, MAX_CONST_I,
313 stateBlock->vertexShaderConstantI,
314 stateBlock->set.vertexShaderConstantsI);
315
316 /* Load DirectX 9 boolean constants/uniforms for vertex shader */
317 shader_glsl_load_constantsB(vshader, gl_info, programId, MAX_CONST_B,
318 stateBlock->vertexShaderConstantB,
319 stateBlock->set.vertexShaderConstantsB);
320
321 /* Upload the position fixup params */
322 pos = GL_EXTCALL(glGetUniformLocationARB(programId, "posFixup"));
323 checkGLcall("glGetUniformLocationARB");
324 GL_EXTCALL(glUniform4fvARB(pos, 1, &deviceImpl->posFixup[0]));
325 checkGLcall("glUniform4fvARB");
326 }
327
328 if (usePixelShader) {
329
330 IWineD3DBaseShaderImpl* pshader = (IWineD3DBaseShaderImpl*) stateBlock->pixelShader;
331
332 constant_locations = stateBlock->glsl_program->puniformF_locations;
333 constant_list = &stateBlock->set_pconstantsF;
334
335 /* Load pixel shader samplers */
336 shader_glsl_load_psamplers(gl_info, (IWineD3DStateBlock*) stateBlock);
337
338 /* Load DirectX 9 float constants/uniforms for pixel shader */
339 shader_glsl_load_constantsF(pshader, gl_info, GL_LIMITS(pshader_constantsF),
340 stateBlock->pixelShaderConstantF, constant_locations, constant_list);
341
342 /* Load DirectX 9 integer constants/uniforms for pixel shader */
343 shader_glsl_load_constantsI(pshader, gl_info, programId, MAX_CONST_I,
344 stateBlock->pixelShaderConstantI,
345 stateBlock->set.pixelShaderConstantsI);
346
347 /* Load DirectX 9 boolean constants/uniforms for pixel shader */
348 shader_glsl_load_constantsB(pshader, gl_info, programId, MAX_CONST_B,
349 stateBlock->pixelShaderConstantB,
350 stateBlock->set.pixelShaderConstantsB);
351 }
352 }
353
354 /** Generate the variable & register declarations for the GLSL output target */
355 void shader_generate_glsl_declarations(
356 IWineD3DBaseShader *iface,
357 shader_reg_maps* reg_maps,
358 SHADER_BUFFER* buffer,
359 WineD3D_GL_Info* gl_info) {
360
361 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
362 int i;
363
364 /* There are some minor differences between pixel and vertex shaders */
365 char pshader = shader_is_pshader_version(This->baseShader.hex_version);
366 char prefix = pshader ? 'P' : 'V';
367
368 /* Prototype the subroutines */
369 for (i = 0; i < This->baseShader.limits.label; i++) {
370 if (reg_maps->labels[i])
371 shader_addline(buffer, "void subroutine%lu();\n", i);
372 }
373
374 /* Declare the constants (aka uniforms) */
375 if (This->baseShader.limits.constant_float > 0) {
376 unsigned max_constantsF = min(This->baseShader.limits.constant_float,
377 (pshader ? GL_LIMITS(pshader_constantsF) : GL_LIMITS(vshader_constantsF)));
378 shader_addline(buffer, "uniform vec4 %cC[%u];\n", prefix, max_constantsF);
379 }
380
381 if (This->baseShader.limits.constant_int > 0)
382 shader_addline(buffer, "uniform ivec4 %cI[%u];\n", prefix, This->baseShader.limits.constant_int);
383
384 if (This->baseShader.limits.constant_bool > 0)
385 shader_addline(buffer, "uniform bool %cB[%u];\n", prefix, This->baseShader.limits.constant_bool);
386
387 if(!pshader)
388 shader_addline(buffer, "uniform vec4 posFixup;\n");
389
390 /* Declare texture samplers */
391 for (i = 0; i < This->baseShader.limits.sampler; i++) {
392 if (reg_maps->samplers[i]) {
393
394 DWORD stype = reg_maps->samplers[i] & WINED3DSP_TEXTURETYPE_MASK;
395 switch (stype) {
396
397 case WINED3DSTT_1D:
398 shader_addline(buffer, "uniform sampler1D %csampler%lu;\n", prefix, i);
399 break;
400 case WINED3DSTT_2D:
401 shader_addline(buffer, "uniform sampler2D %csampler%lu;\n", prefix, i);
402 break;
403 case WINED3DSTT_CUBE:
404 shader_addline(buffer, "uniform samplerCube %csampler%lu;\n", prefix, i);
405 break;
406 case WINED3DSTT_VOLUME:
407 shader_addline(buffer, "uniform sampler3D %csampler%lu;\n", prefix, i);
408 break;
409 default:
410 shader_addline(buffer, "uniform unsupported_sampler %csampler%lu;\n", prefix, i);
411 FIXME("Unrecognized sampler type: %#x\n", stype);
412 break;
413 }
414 }
415 }
416
417 /* Declare address variables */
418 for (i = 0; i < This->baseShader.limits.address; i++) {
419 if (reg_maps->address[i])
420 shader_addline(buffer, "ivec4 A%d;\n", i);
421 }
422
423 /* Declare texture coordinate temporaries and initialize them */
424 for (i = 0; i < This->baseShader.limits.texcoord; i++) {
425 if (reg_maps->texcoord[i])
426 shader_addline(buffer, "vec4 T%lu = gl_TexCoord[%lu];\n", i, i);
427 }
428
429 /* Declare input register temporaries */
430 for (i=0; i < This->baseShader.limits.packed_input; i++) {
431 if (reg_maps->packed_input[i])
432 shader_addline(buffer, "vec4 IN%lu;\n", i);
433 }
434
435 /* Declare output register temporaries */
436 for (i = 0; i < This->baseShader.limits.packed_output; i++) {
437 if (reg_maps->packed_output[i])
438 shader_addline(buffer, "vec4 OUT%lu;\n", i);
439 }
440
441 /* Declare temporary variables */
442 for(i = 0; i < This->baseShader.limits.temporary; i++) {
443 if (reg_maps->temporary[i])
444 shader_addline(buffer, "vec4 R%lu;\n", i);
445 }
446
447 /* Declare attributes */
448 for (i = 0; i < This->baseShader.limits.attributes; i++) {
449 if (reg_maps->attributes[i])
450 shader_addline(buffer, "attribute vec4 attrib%i;\n", i);
451 }
452
453 /* Declare loop register aL */
454 if (reg_maps->loop) {
455 shader_addline(buffer, "int aL;\n");
456 shader_addline(buffer, "int tmpInt;\n");
457 }
458
459 /* Temporary variables for matrix operations */
460 shader_addline(buffer, "vec4 tmp0;\n");
461 shader_addline(buffer, "vec4 tmp1;\n");
462
463 /* Start the main program */
464 shader_addline(buffer, "void main() {\n");
465 }
466
467 /*****************************************************************************
468 * Functions to generate GLSL strings from DirectX Shader bytecode begin here.
469 *
470 * For more information, see http://wiki.winehq.org/DirectX-Shaders
471 ****************************************************************************/
472
473 /* Prototypes */
474 static void shader_glsl_add_param(
475 SHADER_OPCODE_ARG* arg,
476 const DWORD param,
477 const DWORD addr_token,
478 BOOL is_input,
479 char *reg_name,
480 char *reg_mask,
481 char *out_str);
482
483 /** Used for opcode modifiers - They multiply the result by the specified amount */
484 static const char * const shift_glsl_tab[] = {
485 "", /* 0 (none) */
486 "2.0 * ", /* 1 (x2) */
487 "4.0 * ", /* 2 (x4) */
488 "8.0 * ", /* 3 (x8) */
489 "16.0 * ", /* 4 (x16) */
490 "32.0 * ", /* 5 (x32) */
491 "", /* 6 (x64) */
492 "", /* 7 (x128) */
493 "", /* 8 (d256) */
494 "", /* 9 (d128) */
495 "", /* 10 (d64) */
496 "", /* 11 (d32) */
497 "0.0625 * ", /* 12 (d16) */
498 "0.125 * ", /* 13 (d8) */
499 "0.25 * ", /* 14 (d4) */
500 "0.5 * " /* 15 (d2) */
501 };
502
503 /** Print the beginning of the generated GLSL string. example: "reg_name.xyzw = vec4("
504 * Will also change the reg_mask if necessary (not all register types are equal in DX vs GL) */
505 static void shader_glsl_add_dst(DWORD param, const char* reg_name, char* reg_mask, char* outStr) {
506
507 int shift = (param & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
508 char cast[6];
509
510 if ((shader_get_regtype(param) == WINED3DSPR_RASTOUT)
511 && ((param & WINED3DSP_REGNUM_MASK) != 0)) {
512 /* gl_FogFragCoord or glPointSize - both floats */
513 strcpy(cast, "float");
514 strcpy(reg_mask, "");
515
516 } else if (reg_name[0] == 'A') {
517 /* Address register for vertex shaders (ivec4) */
518 strcpy(cast, "ivec4");
519
520 } else {
521 /* Everything else should be a 4 component float vector */
522 strcpy(cast, "vec4");
523 }
524
525 sprintf(outStr, "%s%s = %s%s(", reg_name, reg_mask, shift_glsl_tab[shift], cast);
526 }
527
528 /* Generate a GLSL parameter that does the input modifier computation and return the input register/mask to use */
529 static void shader_glsl_gen_modifier (
530 const DWORD instr,
531 const char *in_reg,
532 const char *in_regswizzle,
533 char *out_str) {
534
535 out_str[0] = 0;
536
537 if (instr == WINED3DSIO_TEXKILL)
538 return;
539
540 switch (instr & WINED3DSP_SRCMOD_MASK) {
541 case WINED3DSPSM_NONE:
542 sprintf(out_str, "%s%s", in_reg, in_regswizzle);
543 break;
544 case WINED3DSPSM_NEG:
545 sprintf(out_str, "-%s%s", in_reg, in_regswizzle);
546 break;
547 case WINED3DSPSM_NOT:
548 sprintf(out_str, "!%s%s", in_reg, in_regswizzle);
549 break;
550 case WINED3DSPSM_BIAS:
551 sprintf(out_str, "(%s%s - vec4(0.5)%s)", in_reg, in_regswizzle, in_regswizzle);
552 break;
553 case WINED3DSPSM_BIASNEG:
554 sprintf(out_str, "-(%s%s - vec4(0.5)%s)", in_reg, in_regswizzle, in_regswizzle);
555 break;
556 case WINED3DSPSM_SIGN:
557 sprintf(out_str, "(2.0 * (%s%s - 0.5))", in_reg, in_regswizzle);
558 break;
559 case WINED3DSPSM_SIGNNEG:
560 sprintf(out_str, "-(2.0 * (%s%s - 0.5))", in_reg, in_regswizzle);
561 break;
562 case WINED3DSPSM_COMP:
563 sprintf(out_str, "(1.0 - %s%s)", in_reg, in_regswizzle);
564 break;
565 case WINED3DSPSM_X2:
566 sprintf(out_str, "(2.0 * %s%s)", in_reg, in_regswizzle);
567 break;
568 case WINED3DSPSM_X2NEG:
569 sprintf(out_str, "-(2.0 * %s%s)", in_reg, in_regswizzle);
570 break;
571 case WINED3DSPSM_DZ: /* reg1_db = { reg1.r/b, reg1.g/b, ...} The g & a components are undefined, so we'll leave them alone */
572 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);
573 break;
574 case WINED3DSPSM_DW:
575 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);
576 break;
577 case WINED3DSPSM_ABS:
578 sprintf(out_str, "abs(%s%s)", in_reg, in_regswizzle);
579 break;
580 case WINED3DSPSM_ABSNEG:
581 sprintf(out_str, "-abs(%s%s)", in_reg, in_regswizzle);
582 break;
583 default:
584 FIXME("Unhandled modifier %u\n", (instr & WINED3DSP_SRCMOD_MASK));
585 sprintf(out_str, "%s%s", in_reg, in_regswizzle);
586 }
587 }
588
589 /** Writes the GLSL variable name that corresponds to the register that the
590 * DX opcode parameter is trying to access */
591 static void shader_glsl_get_register_name(
592 const DWORD param,
593 const DWORD addr_token,
594 char* regstr,
595 BOOL* is_color,
596 SHADER_OPCODE_ARG* arg) {
597
598 /* oPos, oFog and oPts in D3D */
599 static const char * const hwrastout_reg_names[] = { "gl_Position", "gl_FogFragCoord", "gl_PointSize" };
600
601 DWORD reg = param & WINED3DSP_REGNUM_MASK;
602 DWORD regtype = shader_get_regtype(param);
603 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) arg->shader;
604 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
605 WineD3D_GL_Info* gl_info = &((IWineD3DImpl*)deviceImpl->wineD3D)->gl_info;
606
607 char pshader = shader_is_pshader_version(This->baseShader.hex_version);
608 char tmpStr[50];
609
610 *is_color = FALSE;
611
612 switch (regtype) {
613 case WINED3DSPR_TEMP:
614 sprintf(tmpStr, "R%u", reg);
615 break;
616 case WINED3DSPR_INPUT:
617 if (pshader) {
618 /* Pixel shaders >= 3.0 */
619 if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3)
620 sprintf(tmpStr, "IN%u", reg);
621 else {
622 if (reg==0)
623 strcpy(tmpStr, "gl_Color");
624 else
625 strcpy(tmpStr, "gl_SecondaryColor");
626 }
627 } else {
628 if (vshader_input_is_color((IWineD3DVertexShader*) This, reg))
629 *is_color = TRUE;
630 sprintf(tmpStr, "attrib%u", reg);
631 }
632 break;
633 case WINED3DSPR_CONST:
634 {
635 const char* prefix = pshader? "PC":"VC";
636
637 /* Relative addressing */
638 if (param & WINED3DSHADER_ADDRMODE_RELATIVE) {
639
640 /* Relative addressing on shaders 2.0+ have a relative address token,
641 * prior to that, it was hard-coded as "A0.x" because there's only 1 register */
642 if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 2) {
643 char relStr[100], relReg[50], relMask[6];
644 shader_glsl_add_param(arg, addr_token, 0, TRUE, relReg, relMask, relStr);
645 sprintf(tmpStr, "%s[%s + %u]", prefix, relStr, reg);
646 } else
647 sprintf(tmpStr, "%s[A0.x + %u]", prefix, reg);
648
649 } else
650 sprintf(tmpStr, "%s[%u]", prefix, reg);
651
652 break;
653 }
654 case WINED3DSPR_CONSTINT:
655 if (pshader)
656 sprintf(tmpStr, "PI[%u]", reg);
657 else
658 sprintf(tmpStr, "VI[%u]", reg);
659 break;
660 case WINED3DSPR_CONSTBOOL:
661 if (pshader)
662 sprintf(tmpStr, "PB[%u]", reg);
663 else
664 sprintf(tmpStr, "VB[%u]", reg);
665 break;
666 case WINED3DSPR_TEXTURE: /* case WINED3DSPR_ADDR: */
667 if (pshader) {
668 sprintf(tmpStr, "T%u", reg);
669 } else {
670 sprintf(tmpStr, "A%u", reg);
671 }
672 break;
673 case WINED3DSPR_LOOP:
674 sprintf(tmpStr, "aL");
675 break;
676 case WINED3DSPR_SAMPLER:
677 if (pshader)
678 sprintf(tmpStr, "Psampler%u", reg);
679 else
680 sprintf(tmpStr, "Vsampler%u", reg);
681 break;
682 case WINED3DSPR_COLOROUT:
683 if (reg >= GL_LIMITS(buffers)) {
684 WARN("Write to render target %u, only %d supported\n", reg, 4);
685 }
686 if (GL_SUPPORT(ARB_DRAW_BUFFERS)) {
687 sprintf(tmpStr, "gl_FragData[%u]", reg);
688 } else { /* On older cards with GLSL support like the GeforceFX there's only one buffer. */
689 sprintf(tmpStr, "gl_FragColor");
690 }
691 break;
692 case WINED3DSPR_RASTOUT:
693 sprintf(tmpStr, "%s", hwrastout_reg_names[reg]);
694 break;
695 case WINED3DSPR_DEPTHOUT:
696 sprintf(tmpStr, "gl_FragDepth");
697 break;
698 case WINED3DSPR_ATTROUT:
699 if (reg == 0) {
700 sprintf(tmpStr, "gl_FrontColor");
701 } else {
702 sprintf(tmpStr, "gl_FrontSecondaryColor");
703 }
704 break;
705 case WINED3DSPR_TEXCRDOUT:
706 /* Vertex shaders >= 3.0: WINED3DSPR_OUTPUT */
707 if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3)
708 sprintf(tmpStr, "OUT%u", reg);
709 else
710 sprintf(tmpStr, "gl_TexCoord[%u]", reg);
711 break;
712 default:
713 FIXME("Unhandled register name Type(%d)\n", regtype);
714 sprintf(tmpStr, "unrecognized_register");
715 break;
716 }
717
718 strcat(regstr, tmpStr);
719 }
720
721 /* Get the GLSL write mask for the destination register */
722 static void shader_glsl_get_write_mask(const DWORD param, char *write_mask) {
723 char *ptr = write_mask;
724
725 if ((param & WINED3DSP_WRITEMASK_ALL) != WINED3DSP_WRITEMASK_ALL) {
726 *ptr++ = '.';
727 if (param & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
728 if (param & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
729 if (param & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
730 if (param & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
731 }
732
733 *ptr = '\0';
734 }
735
736 static void shader_glsl_get_swizzle(const DWORD param, BOOL fixup, char *swizzle_str) {
737 /* For registers of type WINED3DDECLTYPE_D3DCOLOR, data is stored as "bgra",
738 * but addressed as "rgba". To fix this we need to swap the register's x
739 * and z components. */
740 const char *swizzle_chars = fixup ? "zyxw" : "xyzw";
741 char *ptr = swizzle_str;
742
743 /* swizzle bits fields: wwzzyyxx */
744 DWORD swizzle = (param & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
745 DWORD swizzle_x = swizzle & 0x03;
746 DWORD swizzle_y = (swizzle >> 2) & 0x03;
747 DWORD swizzle_z = (swizzle >> 4) & 0x03;
748 DWORD swizzle_w = (swizzle >> 6) & 0x03;
749
750 /* If the swizzle is the default swizzle (ie, "xyzw"), we don't need to
751 * generate a swizzle string. Unless we need to our own swizzling. */
752 if ((WINED3DSP_NOSWIZZLE >> WINED3DSP_SWIZZLE_SHIFT) != swizzle || fixup) {
753 *ptr++ = '.';
754 if (swizzle_x == swizzle_y && swizzle_x == swizzle_z && swizzle_x == swizzle_w) {
755 *ptr++ = swizzle_chars[swizzle_x];
756 } else {
757 *ptr++ = swizzle_chars[swizzle_x];
758 *ptr++ = swizzle_chars[swizzle_y];
759 *ptr++ = swizzle_chars[swizzle_z];
760 *ptr++ = swizzle_chars[swizzle_w];
761 }
762 }
763
764 *ptr = '\0';
765 }
766
767 /** From a given parameter token, generate the corresponding GLSL string.
768 * Also, return the actual register name and swizzle in case the
769 * caller needs this information as well. */
770 static void shader_glsl_add_param(
771 SHADER_OPCODE_ARG* arg,
772 const DWORD param,
773 const DWORD addr_token,
774 BOOL is_input,
775 char *reg_name,
776 char *reg_mask,
777 char *out_str) {
778
779 BOOL is_color = FALSE;
780 reg_mask[0] = reg_name[0] = out_str[0] = 0;
781
782 shader_glsl_get_register_name(param, addr_token, reg_name, &is_color, arg);
783
784 if (is_input) {
785 shader_glsl_get_swizzle(param, is_color, reg_mask);
786 shader_glsl_gen_modifier(param, reg_name, reg_mask, out_str);
787 } else {
788 shader_glsl_get_write_mask(param, reg_mask);
789 sprintf(out_str, "%s%s", reg_name, reg_mask);
790 }
791 }
792
793 /** Process GLSL instruction modifiers */
794 void shader_glsl_add_instruction_modifiers(SHADER_OPCODE_ARG* arg) {
795
796 DWORD mask = arg->dst & WINED3DSP_DSTMOD_MASK;
797
798 if (arg->opcode->dst_token && mask != 0) {
799 char dst_reg[50];
800 char dst_mask[6];
801 char dst_str[100];
802
803 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
804
805 if (mask & WINED3DSPDM_SATURATE) {
806 /* _SAT means to clamp the value of the register to between 0 and 1 */
807 shader_addline(arg->buffer, "%s%s = clamp(%s%s, 0.0, 1.0);\n", dst_reg, dst_mask, dst_reg, dst_mask);
808 }
809 if (mask & WINED3DSPDM_MSAMPCENTROID) {
810 FIXME("_centroid modifier not handled\n");
811 }
812 if (mask & WINED3DSPDM_PARTIALPRECISION) {
813 /* MSDN says this modifier can be safely ignored, so that's what we'll do. */
814 }
815 }
816 }
817
818 static inline const char* shader_get_comp_op(
819 const DWORD opcode) {
820
821 DWORD op = (opcode & INST_CONTROLS_MASK) >> INST_CONTROLS_SHIFT;
822 switch (op) {
823 case COMPARISON_GT: return ">";
824 case COMPARISON_EQ: return "==";
825 case COMPARISON_GE: return ">=";
826 case COMPARISON_LT: return "<";
827 case COMPARISON_NE: return "!=";
828 case COMPARISON_LE: return "<=";
829 default:
830 FIXME("Unrecognized comparison value: %u\n", op);
831 return "(\?\?)";
832 }
833 }
834
835 static void shader_glsl_sample(SHADER_OPCODE_ARG* arg, DWORD sampler_idx, const char *dst_str, const char *coord_reg) {
836 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
837 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
838 DWORD sampler_type = arg->reg_maps->samplers[sampler_idx] & WINED3DSP_TEXTURETYPE_MASK;
839 const char sampler_prefix = shader_is_pshader_version(This->baseShader.hex_version) ? 'P' : 'V';
840 SHADER_BUFFER* buffer = arg->buffer;
841
842 if(deviceImpl->stateBlock->textureState[sampler_idx][WINED3DTSS_TEXTURETRANSFORMFLAGS] & WINED3DTTFF_PROJECTED) {
843 /* Note that there's no such thing as a projected cube texture. */
844 switch(sampler_type) {
845 case WINED3DSTT_2D:
846 shader_addline(buffer, "%s = texture2DProj(%csampler%u, %s);\n", dst_str, sampler_prefix, sampler_idx, coord_reg);
847 break;
848 case WINED3DSTT_VOLUME:
849 shader_addline(buffer, "%s = texture3DProj(%csampler%u, %s);\n", dst_str, sampler_prefix, sampler_idx, coord_reg);
850 break;
851 default:
852 shader_addline(buffer, "%s = unrecognized_stype(%csampler%u, %s);\n", dst_str, sampler_prefix, sampler_idx, coord_reg);
853 FIXME("Unrecognized sampler type: %#x;\n", sampler_type);
854 break;
855 }
856 } else {
857 switch(sampler_type) {
858 case WINED3DSTT_2D:
859 shader_addline(buffer, "%s = texture2D(%csampler%u, %s.xy);\n", dst_str, sampler_prefix, sampler_idx, coord_reg);
860 break;
861 case WINED3DSTT_CUBE:
862 shader_addline(buffer, "%s = textureCube(%csampler%u, %s.xyz);\n", dst_str, sampler_prefix, sampler_idx, coord_reg);
863 break;
864 case WINED3DSTT_VOLUME:
865 shader_addline(buffer, "%s = texture3D(%csampler%u, %s.xyz);\n", dst_str, sampler_prefix, sampler_idx, coord_reg);
866 break;
867 default:
868 shader_addline(buffer, "%s = unrecognized_stype(%csampler%u, %s);\n", dst_str, sampler_prefix, sampler_idx, coord_reg);
869 FIXME("Unrecognized sampler type: %#x;\n", sampler_type);
870 break;
871 }
872 }
873 }
874
875
876 /*****************************************************************************
877 *
878 * Begin processing individual instruction opcodes
879 *
880 ****************************************************************************/
881
882 /* Generate GLSL arithmetic functions (dst = src1 + src2) */
883 void shader_glsl_arith(SHADER_OPCODE_ARG* arg) {
884
885 CONST SHADER_OPCODE* curOpcode = arg->opcode;
886 SHADER_BUFFER* buffer = arg->buffer;
887 char tmpLine[256];
888 char dst_reg[50], src0_reg[50], src1_reg[50];
889 char dst_mask[6], src0_mask[6], src1_mask[6];
890 char dst_str[100], src0_str[100], src1_str[100];
891
892 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
893 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
894 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
895 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
896 strcat(tmpLine, "vec4(");
897 strcat(tmpLine, src0_str);
898 strcat(tmpLine, ")");
899
900 /* Determine the GLSL operator to use based on the opcode */
901 switch (curOpcode->opcode) {
902 case WINED3DSIO_MUL: strcat(tmpLine, " * "); break;
903 case WINED3DSIO_ADD: strcat(tmpLine, " + "); break;
904 case WINED3DSIO_SUB: strcat(tmpLine, " - "); break;
905 default:
906 FIXME("Opcode %s not yet handled in GLSL\n", curOpcode->name);
907 break;
908 }
909 shader_addline(buffer, "%svec4(%s))%s;\n", tmpLine, src1_str, dst_mask);
910 }
911
912 /* Process the WINED3DSIO_MOV opcode using GLSL (dst = src) */
913 void shader_glsl_mov(SHADER_OPCODE_ARG* arg) {
914
915 SHADER_BUFFER* buffer = arg->buffer;
916 char tmpLine[256];
917 char dst_str[100], src0_str[100];
918 char dst_reg[50], src0_reg[50];
919 char dst_mask[6], src0_mask[6];
920
921 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
922 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
923 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
924 shader_addline(buffer, "%s%s)%s;\n", tmpLine, src0_str, dst_mask);
925 }
926
927 /* Process the dot product operators DP3 and DP4 in GLSL (dst = dot(src0, src1)) */
928 void shader_glsl_dot(SHADER_OPCODE_ARG* arg) {
929
930 CONST SHADER_OPCODE* curOpcode = arg->opcode;
931 SHADER_BUFFER* buffer = arg->buffer;
932 char tmpDest[100];
933 char dst_str[100], src0_str[100], src1_str[100];
934 char dst_reg[50], src0_reg[50], src1_reg[50];
935 char dst_mask[6], src0_mask[6], src1_mask[6];
936 char cast[6];
937
938 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
939 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
940 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
941
942 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpDest);
943
944 /* Need to cast the src vectors to vec3 for dp3, and vec4 for dp4 */
945 if (curOpcode->opcode == WINED3DSIO_DP4)
946 strcpy(cast, "vec4(");
947 else
948 strcpy(cast, "vec3(");
949
950 shader_addline(buffer, "%sdot(%s%s), %s%s)))%s;\n",
951 tmpDest, cast, src0_str, cast, src1_str, dst_mask);
952 }
953
954 /* Map the opcode 1-to-1 to the GL code (arg->dst = instruction(src0, src1, ...) */
955 void shader_glsl_map2gl(SHADER_OPCODE_ARG* arg) {
956
957 CONST SHADER_OPCODE* curOpcode = arg->opcode;
958 SHADER_BUFFER* buffer = arg->buffer;
959 char tmpLine[256];
960 char dst_str[100], src_str[100];
961 char dst_reg[50], src_reg[50];
962 char dst_mask[6], src_mask[6];
963 unsigned i;
964
965 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
966
967 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
968
969 /* Determine the GLSL function to use based on the opcode */
970 /* TODO: Possibly make this a table for faster lookups */
971 switch (curOpcode->opcode) {
972 case WINED3DSIO_MIN: strcat(tmpLine, "min"); break;
973 case WINED3DSIO_MAX: strcat(tmpLine, "max"); break;
974 case WINED3DSIO_RSQ: strcat(tmpLine, "inversesqrt"); break;
975 case WINED3DSIO_ABS: strcat(tmpLine, "abs"); break;
976 case WINED3DSIO_FRC: strcat(tmpLine, "fract"); break;
977 case WINED3DSIO_POW: strcat(tmpLine, "pow"); break;
978 case WINED3DSIO_CRS: strcat(tmpLine, "cross"); break;
979 case WINED3DSIO_NRM: strcat(tmpLine, "normalize"); break;
980 case WINED3DSIO_LOGP:
981 case WINED3DSIO_LOG: strcat(tmpLine, "log2"); break;
982 case WINED3DSIO_EXP: strcat(tmpLine, "exp2"); break;
983 case WINED3DSIO_SGE: strcat(tmpLine, "greaterThanEqual"); break;
984 case WINED3DSIO_SLT: strcat(tmpLine, "lessThan"); break;
985 case WINED3DSIO_SGN: strcat(tmpLine, "sign"); break;
986 default:
987 FIXME("Opcode %s not yet handled in GLSL\n", curOpcode->name);
988 break;
989 }
990
991 strcat(tmpLine, "(");
992
993 if (curOpcode->num_params > 0) {
994 strcat(tmpLine, "vec4(");
995 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src_reg, src_mask, src_str);
996 strcat(tmpLine, src_str);
997 strcat(tmpLine, ")");
998 for (i = 2; i < curOpcode->num_params; ++i) {
999 strcat(tmpLine, ", vec4(");
1000 shader_glsl_add_param(arg, arg->src[i-1], arg->src_addr[i-1], TRUE, src_reg, src_mask, src_str);
1001 strcat(tmpLine, src_str);
1002 strcat(tmpLine, ")");
1003 }
1004 }
1005 shader_addline(buffer, "%s))%s;\n", tmpLine, dst_mask);
1006
1007 }
1008
1009 /** Process the WINED3DSIO_EXPP instruction in GLSL:
1010 * For shader model 1.x, do the following (and honor the writemask, so use a temporary variable):
1011 * dst.x = 2^(floor(src))
1012 * dst.y = src - floor(src)
1013 * dst.z = 2^src (partial precision is allowed, but optional)
1014 * dst.w = 1.0;
1015 * For 2.0 shaders, just do this (honoring writemask and swizzle):
1016 * dst = 2^src; (partial precision is allowed, but optional)
1017 */
1018 void shader_glsl_expp(SHADER_OPCODE_ARG* arg) {
1019
1020 char tmpLine[256];
1021 char dst_str[100], src_str[100];
1022 char dst_reg[50], src_reg[50];
1023 char dst_mask[6], src_mask[6];
1024 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1025 DWORD hex_version = This->baseShader.hex_version;
1026
1027 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1028 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src_reg, src_mask, src_str);
1029 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
1030
1031 if (hex_version < WINED3DPS_VERSION(2,0)) {
1032 shader_addline(arg->buffer, "tmp0.x = vec4(exp2(floor(%s))).x;\n", src_str);
1033 shader_addline(arg->buffer, "tmp0.y = vec4(%s - floor(%s)).y;\n", src_str, src_str);
1034 shader_addline(arg->buffer, "tmp0.z = vec4(exp2(%s)).x;\n", src_str);
1035 shader_addline(arg->buffer, "tmp0.w = 1.0;\n");
1036 shader_addline(arg->buffer, "%svec4(tmp0))%s;\n", tmpLine, dst_mask);
1037 } else {
1038 shader_addline(arg->buffer, "%svec4(exp2(%s)))%s;\n", tmpLine, src_str, dst_mask);
1039 }
1040 }
1041
1042 /** Process the RCP (reciprocal or inverse) opcode in GLSL (dst = 1 / src) */
1043 void shader_glsl_rcp(SHADER_OPCODE_ARG* arg) {
1044
1045 char tmpLine[256];
1046 char dst_str[100], src_str[100];
1047 char dst_reg[50], src_reg[50];
1048 char dst_mask[6], src_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, src_reg, src_mask, src_str);
1052 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
1053 strcat(tmpLine, "1.0 / ");
1054 shader_addline(arg->buffer, "%s%s)%s;\n", tmpLine, src_str, dst_mask);
1055 }
1056
1057 /** Process signed comparison opcodes in GLSL. */
1058 void shader_glsl_compare(SHADER_OPCODE_ARG* arg) {
1059
1060 char tmpLine[256];
1061 char dst_str[100], src0_str[100], src1_str[100];
1062 char dst_reg[50], src0_reg[50], src1_reg[50];
1063 char dst_mask[6], src0_mask[6], src1_mask[6];
1064
1065 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1066 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1067 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
1068
1069 /* If we are comparing vectors and not scalars, we should process this through map2gl using the GLSL functions. */
1070 if (strlen(src0_mask) != 2) {
1071 shader_glsl_map2gl(arg);
1072 } else {
1073 char compareStr[3];
1074 compareStr[0] = 0;
1075 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
1076
1077 switch (arg->opcode->opcode) {
1078 case WINED3DSIO_SLT: strcpy(compareStr, "<"); break;
1079 case WINED3DSIO_SGE: strcpy(compareStr, ">="); break;
1080 default:
1081 FIXME("Can't handle opcode %s\n", arg->opcode->name);
1082 }
1083 shader_addline(arg->buffer, "%s(float(%s) %s float(%s)) ? 1.0 : 0.0)%s;\n",
1084 tmpLine, src0_str, compareStr, src1_str, dst_mask);
1085 }
1086 }
1087
1088 /** Process CMP instruction in GLSL (dst = src0.x > 0.0 ? src1.x : src2.x), per channel */
1089 void shader_glsl_cmp(SHADER_OPCODE_ARG* arg) {
1090
1091 char tmpLine[256];
1092 char dst_str[100], src0_str[100], src1_str[100], src2_str[100];
1093 char dst_reg[50], src0_reg[50], src1_reg[50], src2_reg[50];
1094 char dst_mask[6], src0_mask[6], src1_mask[6], src2_mask[6];
1095
1096 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1097 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1098 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
1099 shader_glsl_add_param(arg, arg->src[2], arg->src_addr[2], TRUE, src2_reg, src2_mask, src2_str);
1100
1101 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
1102 shader_addline(arg->buffer, "%smix(vec4(%s), vec4(%s), vec4(lessThan(vec4(%s), vec4(0.0)))))%s;\n",
1103 tmpLine, src1_str, src2_str, src0_str, dst_mask);
1104 }
1105
1106 /** Process the CND opcode in GLSL (dst = (src0 < 0.5) ? src1 : src2) */
1107 void shader_glsl_cnd(SHADER_OPCODE_ARG* arg) {
1108
1109 char tmpLine[256];
1110 char dst_str[100], src0_str[100], src1_str[100], src2_str[100];
1111 char dst_reg[50], src0_reg[50], src1_reg[50], src2_reg[50];
1112 char dst_mask[6], src0_mask[6], src1_mask[6], src2_mask[6];
1113
1114 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1115 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1116 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
1117 shader_glsl_add_param(arg, arg->src[2], arg->src_addr[2], TRUE, src2_reg, src2_mask, src2_str);
1118 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
1119 shader_addline(arg->buffer, "%s(%s < 0.5) ? %s : %s)%s;\n",
1120 tmpLine, src0_str, src1_str, src2_str, dst_mask);
1121 }
1122
1123 /** GLSL code generation for WINED3DSIO_MAD: Multiply the first 2 opcodes, then add the last */
1124 void shader_glsl_mad(SHADER_OPCODE_ARG* arg) {
1125
1126 char tmpLine[256];
1127 char dst_str[100], src0_str[100], src1_str[100], src2_str[100];
1128 char dst_reg[50], src0_reg[50], src1_reg[50], src2_reg[50];
1129 char dst_mask[6], src0_mask[6], src1_mask[6], src2_mask[6];
1130
1131 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1132 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1133 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
1134 shader_glsl_add_param(arg, arg->src[2], arg->src_addr[2], TRUE, src2_reg, src2_mask, src2_str);
1135 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
1136
1137 shader_addline(arg->buffer, "%s(vec4(%s) * vec4(%s)) + vec4(%s))%s;\n",
1138 tmpLine, src0_str, src1_str, src2_str, dst_mask);
1139 }
1140
1141 /** Handles transforming all WINED3DSIO_M?x? opcodes for
1142 Vertex shaders to GLSL codes */
1143 void shader_glsl_mnxn(SHADER_OPCODE_ARG* arg) {
1144 int i;
1145 int nComponents = 0;
1146 SHADER_OPCODE_ARG tmpArg;
1147
1148 memset(&tmpArg, 0, sizeof(SHADER_OPCODE_ARG));
1149
1150 /* Set constants for the temporary argument */
1151 tmpArg.shader = arg->shader;
1152 tmpArg.buffer = arg->buffer;
1153 tmpArg.src[0] = arg->src[0];
1154 tmpArg.src_addr[0] = arg->src_addr[0];
1155 tmpArg.src_addr[1] = arg->src_addr[1];
1156 tmpArg.reg_maps = arg->reg_maps;
1157
1158 switch(arg->opcode->opcode) {
1159 case WINED3DSIO_M4x4:
1160 nComponents = 4;
1161 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP4);
1162 break;
1163 case WINED3DSIO_M4x3:
1164 nComponents = 3;
1165 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP4);
1166 break;
1167 case WINED3DSIO_M3x4:
1168 nComponents = 4;
1169 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP3);
1170 break;
1171 case WINED3DSIO_M3x3:
1172 nComponents = 3;
1173 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP3);
1174 break;
1175 case WINED3DSIO_M3x2:
1176 nComponents = 2;
1177 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP3);
1178 break;
1179 default:
1180 break;
1181 }
1182
1183 for (i = 0; i < nComponents; i++) {
1184 tmpArg.dst = ((arg->dst) & ~WINED3DSP_WRITEMASK_ALL)|(WINED3DSP_WRITEMASK_0<<i);
1185 tmpArg.src[1] = arg->src[1]+i;
1186 shader_glsl_dot(&tmpArg);
1187 }
1188 }
1189
1190 /**
1191 The LRP instruction performs a component-wise linear interpolation
1192 between the second and third operands using the first operand as the
1193 blend factor. Equation: (dst = src2 * (src1 - src0) + src0)
1194 */
1195 void shader_glsl_lrp(SHADER_OPCODE_ARG* arg) {
1196
1197 char tmpLine[256];
1198 char dst_str[100], src0_str[100], src1_str[100], src2_str[100];
1199 char dst_reg[50], src0_reg[50], src1_reg[50], src2_reg[50];
1200 char dst_mask[6], src0_mask[6], src1_mask[6], src2_mask[6];
1201
1202 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1203 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1204 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
1205 shader_glsl_add_param(arg, arg->src[2], arg->src_addr[2], TRUE, src2_reg, src2_mask, src2_str);
1206
1207 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
1208
1209 shader_addline(arg->buffer, "%s%s + %s * (%s - %s))%s;\n",
1210 tmpLine, src2_str, src0_str, src1_str, src2_str, dst_mask);
1211 }
1212
1213 /** Process the WINED3DSIO_LIT instruction in GLSL:
1214 * dst.x = dst.w = 1.0
1215 * dst.y = (src0.x > 0) ? src0.x
1216 * dst.z = (src0.x > 0) ? ((src0.y > 0) ? pow(src0.y, src.w) : 0) : 0
1217 * where src.w is clamped at +- 128
1218 */
1219 void shader_glsl_lit(SHADER_OPCODE_ARG* arg) {
1220
1221 char dst_str[100], src0_str[100];
1222 char dst_reg[50], src0_reg[50];
1223 char dst_mask[6], src0_mask[6];
1224
1225 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1226 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1227
1228 shader_addline(arg->buffer,
1229 "%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",
1230 dst_str, src0_reg, src0_reg, src0_reg, src0_reg, src0_reg, src0_reg, dst_mask);
1231 }
1232
1233 /** Process the WINED3DSIO_DST instruction in GLSL:
1234 * dst.x = 1.0
1235 * dst.y = src0.x * src0.y
1236 * dst.z = src0.z
1237 * dst.w = src1.w
1238 */
1239 void shader_glsl_dst(SHADER_OPCODE_ARG* arg) {
1240
1241 char dst_str[100], src0_str[100], src1_str[100];
1242 char dst_reg[50], src0_reg[50], src1_reg[50];
1243 char dst_mask[6], src0_mask[6], src1_mask[6];
1244
1245 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1246 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1247 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
1248
1249 shader_addline(arg->buffer, "%s = vec4(1.0, %s.x * %s.y, %s.z, %s.w)%s;\n",
1250 dst_str, src0_reg, src1_reg, src0_reg, src1_reg, dst_mask);
1251 }
1252
1253 /** Process the WINED3DSIO_SINCOS instruction in GLSL:
1254 * VS 2.0 requires that specific cosine and sine constants be passed to this instruction so the hardware
1255 * can handle it. But, these functions are built-in for GLSL, so we can just ignore the last 2 params.
1256 *
1257 * dst.x = cos(src0.?)
1258 * dst.y = sin(src0.?)
1259 * dst.z = dst.z
1260 * dst.w = dst.w
1261 */
1262 void shader_glsl_sincos(SHADER_OPCODE_ARG* arg) {
1263
1264 char dst_str[100], src0_str[100];
1265 char dst_reg[50], src0_reg[50];
1266 char dst_mask[6], src0_mask[6];
1267
1268 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1269 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1270
1271 shader_addline(arg->buffer, "%s = vec4(cos(%s), sin(%s), %s.z, %s.w)%s;\n",
1272 dst_str, src0_str, src0_str, dst_reg, dst_reg, dst_mask);
1273 }
1274
1275 /** Process the WINED3DSIO_LOOP instruction in GLSL:
1276 * Start a for() loop where src0.y is the initial value of aL,
1277 * increment aL by src0.z for a total of src0.x iterations.
1278 * Need to use a temporary variable for this operation.
1279 */
1280 void shader_glsl_loop(SHADER_OPCODE_ARG* arg) {
1281
1282 char src1_str[100];
1283 char src1_reg[50];
1284 char src1_mask[6];
1285
1286 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
1287
1288 shader_addline(arg->buffer, "for (tmpInt = 0, aL = %s.y; tmpInt < %s.x; tmpInt++, aL += %s.z) {\n",
1289 src1_reg, src1_reg, src1_reg);
1290 }
1291
1292 void shader_glsl_end(SHADER_OPCODE_ARG* arg) {
1293 shader_addline(arg->buffer, "}\n");
1294 }
1295
1296 void shader_glsl_rep(SHADER_OPCODE_ARG* arg) {
1297
1298 char src0_str[100];
1299 char src0_reg[50];
1300 char src0_mask[6];
1301
1302 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1303 shader_addline(arg->buffer, "for (tmpInt = 0; tmpInt < %s.x; tmpInt++) {\n", src0_reg);
1304 }
1305
1306 void shader_glsl_if(SHADER_OPCODE_ARG* arg) {
1307
1308 char src0_str[100];
1309 char src0_reg[50];
1310 char src0_mask[6];
1311
1312 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1313 shader_addline(arg->buffer, "if (%s) {\n", src0_str);
1314 }
1315
1316 void shader_glsl_ifc(SHADER_OPCODE_ARG* arg) {
1317
1318 char src0_str[100], src1_str[100];
1319 char src0_reg[50], src1_reg[50];
1320 char src0_mask[6], src1_mask[6];
1321
1322 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1323 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
1324
1325 shader_addline(arg->buffer, "if (%s %s %s) {\n",
1326 src0_str, shader_get_comp_op(arg->opcode_token), src1_str);
1327 }
1328
1329 void shader_glsl_else(SHADER_OPCODE_ARG* arg) {
1330 shader_addline(arg->buffer, "} else {\n");
1331 }
1332
1333 void shader_glsl_break(SHADER_OPCODE_ARG* arg) {
1334 shader_addline(arg->buffer, "break;\n");
1335 }
1336
1337 void shader_glsl_breakc(SHADER_OPCODE_ARG* arg) {
1338
1339 char src0_str[100], src1_str[100];
1340 char src0_reg[50], src1_reg[50];
1341 char src0_mask[6], src1_mask[6];
1342
1343 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1344 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
1345
1346 shader_addline(arg->buffer, "if (%s %s %s) break;\n",
1347 src0_str, shader_get_comp_op(arg->opcode_token), src1_str);
1348 }
1349
1350 void shader_glsl_label(SHADER_OPCODE_ARG* arg) {
1351
1352 DWORD snum = (arg->src[0]) & WINED3DSP_REGNUM_MASK;
1353 shader_addline(arg->buffer, "}\n");
1354 shader_addline(arg->buffer, "void subroutine%lu () {\n", snum);
1355 }
1356
1357 void shader_glsl_call(SHADER_OPCODE_ARG* arg) {
1358 DWORD snum = (arg->src[0]) & WINED3DSP_REGNUM_MASK;
1359 shader_addline(arg->buffer, "subroutine%lu();\n", snum);
1360 }
1361
1362 void shader_glsl_callnz(SHADER_OPCODE_ARG* arg) {
1363
1364 char src1_str[100];
1365 char src1_reg[50];
1366 char src1_mask[6];
1367
1368 DWORD snum = (arg->src[0]) & WINED3DSP_REGNUM_MASK;
1369 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
1370 shader_addline(arg->buffer, "if (%s) subroutine%lu();\n", src1_str, snum);
1371 }
1372
1373 /*********************************************
1374 * Pixel Shader Specific Code begins here
1375 ********************************************/
1376 void pshader_glsl_tex(SHADER_OPCODE_ARG* arg) {
1377 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1378
1379 DWORD hex_version = This->baseShader.hex_version;
1380
1381 char dst_str[100], dst_reg[50], dst_mask[6];
1382 char coord_str[100], coord_reg[50], coord_mask[6];
1383
1384 /* All versions have a destination register */
1385 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1386
1387 /* 1.0-1.3: Use destination register as coordinate source.
1388 1.4+: Use provided coordinate source register. */
1389 if (hex_version < WINED3DPS_VERSION(1,4))
1390 strcpy(coord_reg, dst_reg);
1391 else
1392 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, coord_reg, coord_mask, coord_str);
1393
1394 /* 1.0-1.4: Use destination register as sampler source.
1395 * 2.0+: Use provided sampler source. */
1396 if (hex_version < WINED3DPS_VERSION(2,0)) {
1397 shader_glsl_sample(arg, arg->dst & WINED3DSP_REGNUM_MASK, dst_str, coord_reg);
1398 } else {
1399 shader_glsl_sample(arg, arg->src[1] & WINED3DSP_REGNUM_MASK, dst_str, coord_reg);
1400 }
1401
1402 }
1403
1404 void pshader_glsl_texcoord(SHADER_OPCODE_ARG* arg) {
1405
1406 /* FIXME: Make this work for more than just 2D textures */
1407
1408 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1409 SHADER_BUFFER* buffer = arg->buffer;
1410 DWORD hex_version = This->baseShader.hex_version;
1411
1412 char tmpStr[100];
1413 char tmpReg[50];
1414 char tmpMask[6];
1415 tmpReg[0] = 0;
1416
1417 shader_glsl_add_param(arg, arg->dst, 0, FALSE, tmpReg, tmpMask, tmpStr);
1418
1419 if (hex_version != WINED3DPS_VERSION(1,4)) {
1420 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1421 shader_addline(buffer, "%s = clamp(gl_TexCoord[%u], 0.0, 1.0);\n", tmpReg, reg);
1422 } else {
1423 DWORD reg2 = arg->src[0] & WINED3DSP_REGNUM_MASK;
1424 shader_addline(buffer, "%s = gl_TexCoord[%u]%s;\n", tmpStr, reg2, tmpMask);
1425 }
1426 }
1427
1428 /** Process the WINED3DSIO_TEXDP3TEX instruction in GLSL:
1429 * Take a 3-component dot product of the TexCoord[dstreg] and src,
1430 * then perform a 1D texture lookup from stage dstregnum, place into dst. */
1431 void pshader_glsl_texdp3tex(SHADER_OPCODE_ARG* arg) {
1432
1433 DWORD dstreg = arg->dst & WINED3DSP_REGNUM_MASK;
1434 char src0_str[100], dst_str[100];
1435 char src0_name[50], dst_name[50];
1436 char src0_mask[6], dst_mask[6];
1437
1438 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_name, dst_mask, dst_str);
1439 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_name, src0_mask, src0_str);
1440
1441 shader_addline(arg->buffer, "tmp0.x = dot(vec3(gl_TexCoord[%u]), vec3(%s));\n", dstreg, src0_str);
1442 shader_addline(arg->buffer, "%s = vec4(texture1D(Psampler%u, tmp0.x))%s;\n", dst_str, dstreg, dst_mask);
1443 }
1444
1445 /** Process the WINED3DSIO_TEXDP3 instruction in GLSL:
1446 * Take a 3-component dot product of the TexCoord[dstreg] and src. */
1447 void pshader_glsl_texdp3(SHADER_OPCODE_ARG* arg) {
1448
1449 DWORD dstreg = arg->dst & WINED3DSP_REGNUM_MASK;
1450 char src0_str[100], dst_str[100];
1451 char src0_name[50], dst_name[50];
1452 char src0_mask[6], dst_mask[6];
1453
1454 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_name, dst_mask, dst_str);
1455 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_name, src0_mask, src0_str);
1456
1457 shader_addline(arg->buffer, "%s = vec4(dot(vec3(T%u), vec3(%s)))%s;\n",
1458 dst_str, dstreg, src0_str, dst_mask);
1459 }
1460
1461 /** Process the WINED3DSIO_TEXDEPTH instruction in GLSL:
1462 * Calculate the depth as dst.x / dst.y */
1463 void pshader_glsl_texdepth(SHADER_OPCODE_ARG* arg) {
1464
1465 char dst_str[100];
1466 char dst_reg[50];
1467 char dst_mask[6];
1468
1469 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1470
1471 shader_addline(arg->buffer, "gl_FragDepth = %s.x / %s.y;\n", dst_reg, dst_reg);
1472 }
1473
1474 /** Process the WINED3DSIO_TEXM3X2DEPTH instruction in GLSL:
1475 * Last row of a 3x2 matrix multiply, use the result to calculate the depth:
1476 * Calculate tmp0.y = TexCoord[dstreg] . src.xyz; (tmp0.x has already been calculated)
1477 * depth = (tmp0.y == 0.0) ? 1.0 : tmp0.x / tmp0.y
1478 */
1479 void pshader_glsl_texm3x2depth(SHADER_OPCODE_ARG* arg) {
1480
1481 DWORD dstreg = arg->dst & WINED3DSP_REGNUM_MASK;
1482 char src0_str[100], dst_str[100];
1483 char src0_name[50], dst_name[50];
1484 char src0_mask[6], dst_mask[6];
1485
1486 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_name, dst_mask, dst_str);
1487 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_name, src0_mask, src0_str);
1488
1489 shader_addline(arg->buffer, "tmp0.y = dot(vec3(T%u), vec3(%s));\n", dstreg, src0_str);
1490 shader_addline(arg->buffer, "gl_FragDepth = vec4((tmp0.y == 0.0) ? 1.0 : tmp0.x / tmp0.y)%s;\n", dst_str, dst_name);
1491 }
1492
1493 /** Process the WINED3DSIO_TEXM3X2PAD instruction in GLSL
1494 * Calculate the 1st of a 2-row matrix multiplication. */
1495 void pshader_glsl_texm3x2pad(SHADER_OPCODE_ARG* arg) {
1496
1497 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1498 SHADER_BUFFER* buffer = arg->buffer;
1499 char src0_str[100];
1500 char src0_name[50];
1501 char src0_mask[6];
1502
1503 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_name, src0_mask, src0_str);
1504 shader_addline(buffer, "tmp0.x = dot(vec3(T%u), vec3(%s));\n", reg, src0_str);
1505 }
1506
1507 /** Process the WINED3DSIO_TEXM3X3PAD instruction in GLSL
1508 * Calculate the 1st or 2nd row of a 3-row matrix multiplication. */
1509 void pshader_glsl_texm3x3pad(SHADER_OPCODE_ARG* arg) {
1510
1511 IWineD3DPixelShaderImpl* shader = (IWineD3DPixelShaderImpl*) arg->shader;
1512 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1513 SHADER_BUFFER* buffer = arg->buffer;
1514 SHADER_PARSE_STATE* current_state = &shader->baseShader.parse_state;
1515 char src0_str[100];
1516 char src0_name[50];
1517 char src0_mask[6];
1518
1519 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_name, src0_mask, src0_str);
1520 shader_addline(buffer, "tmp0.%c = dot(vec3(T%u), vec3(%s));\n", 'x' + current_state->current_row, reg, src0_str);
1521 current_state->texcoord_w[current_state->current_row++] = reg;
1522 }
1523
1524 void pshader_glsl_texm3x2tex(SHADER_OPCODE_ARG* arg) {
1525 char dst_str[8];
1526 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1527 SHADER_BUFFER* buffer = arg->buffer;
1528 char src0_str[100];
1529 char src0_name[50];
1530 char src0_mask[6];
1531
1532 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_name, src0_mask, src0_str);
1533 shader_addline(buffer, "tmp0.y = dot(vec3(T%u), vec3(%s));\n", reg, src0_str);
1534
1535 /* Sample the texture using the calculated coordinates */
1536 sprintf(dst_str, "T%u", reg);
1537 shader_glsl_sample(arg, reg, dst_str, "tmp0");
1538 }
1539
1540 /** Process the WINED3DSIO_TEXM3X3TEX instruction in GLSL
1541 * Perform the 3rd row of a 3x3 matrix multiply, then sample the texture using the calculated coordinates */
1542 void pshader_glsl_texm3x3tex(SHADER_OPCODE_ARG* arg) {
1543 char dst_str[8];
1544 char src0_str[100];
1545 char src0_name[50];
1546 char src0_mask[6];
1547 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1548 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1549 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1550
1551 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_name, src0_mask, src0_str);
1552 shader_addline(arg->buffer, "tmp0.z = dot(vec3(T%u), vec3(%s));\n", reg, src0_str);
1553
1554 /* Sample the texture using the calculated coordinates */
1555 sprintf(dst_str, "T%u", reg);
1556 shader_glsl_sample(arg, reg, dst_str, "tmp0");
1557 current_state->current_row = 0;
1558 }
1559
1560 /** Process the WINED3DSIO_TEXM3X3 instruction in GLSL
1561 * Perform the 3rd row of a 3x3 matrix multiply */
1562 void pshader_glsl_texm3x3(SHADER_OPCODE_ARG* arg) {
1563
1564 char src0_str[100];
1565 char src0_name[50];
1566 char src0_mask[6];
1567 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1568 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1569 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1570
1571 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_name, src0_mask, src0_str);
1572
1573 shader_addline(arg->buffer, "tmp0.z = dot(vec3(T%u), vec3(%s));\n", reg, src0_str);
1574 shader_addline(arg->buffer, "T%u = vec4(tmp0.x, tmp0.y, tmp0.z, 1.0);\n", reg);
1575 current_state->current_row = 0;
1576 }
1577
1578 /** Process the WINED3DSIO_TEXM3X3SPEC instruction in GLSL
1579 * Peform the final texture lookup based on the previous 2 3x3 matrix multiplies */
1580 void pshader_glsl_texm3x3spec(SHADER_OPCODE_ARG* arg) {
1581
1582 IWineD3DPixelShaderImpl* shader = (IWineD3DPixelShaderImpl*) arg->shader;
1583 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1584 char dimensions[5];
1585 char dst_str[8];
1586 char src0_str[100], src0_name[50], src0_mask[6];
1587 char src1_str[100], src1_name[50], src1_mask[6];
1588 SHADER_BUFFER* buffer = arg->buffer;
1589 SHADER_PARSE_STATE* current_state = &shader->baseShader.parse_state;
1590 DWORD stype = arg->reg_maps->samplers[reg] & WINED3DSP_TEXTURETYPE_MASK;
1591
1592 switch (stype) {
1593 case WINED3DSTT_2D: strcpy(dimensions, "2D"); break;
1594 case WINED3DSTT_CUBE: strcpy(dimensions, "Cube"); break;
1595 case WINED3DSTT_VOLUME: strcpy(dimensions, "3D"); break;
1596 default:
1597 strcpy(dimensions, "");
1598 FIXME("Unrecognized sampler type: %#x\n", stype);
1599 break;
1600 }
1601
1602 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_name, src0_mask, src0_str);
1603 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_name, src1_mask, src1_str);
1604
1605 /* Perform the last matrix multiply operation */
1606 shader_addline(buffer, "tmp0.z = dot(vec3(T%u), vec3(%s));\n", reg, src0_str);
1607
1608 /* Calculate reflection vector */
1609 shader_addline(buffer, "tmp0.xyz = reflect(-vec3(%s), vec3(tmp0));\n", src1_str);
1610
1611 /* Sample the texture */
1612 sprintf(dst_str, "T%u", reg);
1613 shader_glsl_sample(arg, reg, dst_str, "tmp0");
1614 current_state->current_row = 0;
1615 }
1616
1617 /** Process the WINED3DSIO_TEXM3X3VSPEC instruction in GLSL
1618 * Peform the final texture lookup based on the previous 2 3x3 matrix multiplies */
1619 void pshader_glsl_texm3x3vspec(SHADER_OPCODE_ARG* arg) {
1620
1621 IWineD3DPixelShaderImpl* shader = (IWineD3DPixelShaderImpl*) arg->shader;
1622 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1623 SHADER_BUFFER* buffer = arg->buffer;
1624 SHADER_PARSE_STATE* current_state = &shader->baseShader.parse_state;
1625 char dst_str[8];
1626 char src0_str[100], src0_name[50], src0_mask[6];
1627
1628 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_name, src0_mask, src0_str);
1629
1630 /* Perform the last matrix multiply operation */
1631 shader_addline(buffer, "tmp0.z = dot(vec3(T%u), vec3(%s));\n", reg, src0_str);
1632
1633 /* Construct the eye-ray vector from w coordinates */
1634 shader_addline(buffer, "tmp1.x = gl_TexCoord[%u].w;\n", current_state->texcoord_w[0]);
1635 shader_addline(buffer, "tmp1.y = gl_TexCoord[%u].w;\n", current_state->texcoord_w[1]);
1636 shader_addline(buffer, "tmp1.z = gl_TexCoord[%u].w;\n", reg);
1637
1638 /* Calculate reflection vector (Assume normal is normalized): RF = 2*(N.E)*N -E */
1639 shader_addline(buffer, "tmp0.x = dot(vec3(tmp0), vec3(tmp1));\n");
1640 shader_addline(buffer, "tmp0 = tmp0.w * tmp0;\n");
1641 shader_addline(buffer, "tmp0 = (2.0 * tmp0) - tmp1;\n");
1642
1643 /* Sample the texture using the calculated coordinates */
1644 sprintf(dst_str, "T%u", reg);
1645 shader_glsl_sample(arg, reg, dst_str, "tmp0");
1646 current_state->current_row = 0;
1647 }
1648
1649 /** Process the WINED3DSIO_TEXBEM instruction in GLSL.
1650 * Apply a fake bump map transform.
1651 * FIXME: Should apply the BUMPMAPENV matrix. For now, just sample the texture */
1652 void pshader_glsl_texbem(SHADER_OPCODE_ARG* arg) {
1653
1654 DWORD reg1 = arg->dst & WINED3DSP_REGNUM_MASK;
1655 DWORD reg2 = arg->src[0] & WINED3DSP_REGNUM_MASK;
1656
1657 FIXME("Not applying the BUMPMAPENV matrix for pixel shader instruction texbem.\n");
1658 shader_addline(arg->buffer, "T%u = texture2D(Psampler%u, gl_TexCoord[%u].xy + T%u.xy);\n",
1659 reg1, reg1, reg1, reg2);
1660 }
1661
1662 /** Process the WINED3DSIO_TEXREG2AR instruction in GLSL
1663 * Sample 2D texture at dst using the alpha & red (wx) components of src as texture coordinates */
1664 void pshader_glsl_texreg2ar(SHADER_OPCODE_ARG* arg) {
1665
1666 char tmpLine[255];
1667 char dst_str[100], src0_str[100];
1668 char dst_reg[50], src0_reg[50];
1669 char dst_mask[6], src0_mask[6];
1670 DWORD src0_regnum = arg->src[0] & WINED3DSP_REGNUM_MASK;
1671
1672 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1673 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1674
1675 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
1676 shader_addline(arg->buffer, "%stexture2D(Psampler%u, %s.yz))%s;\n",
1677 tmpLine, src0_regnum, dst_reg, dst_mask);
1678 }
1679
1680 /** Process the WINED3DSIO_TEXREG2GB instruction in GLSL
1681 * Sample 2D texture at dst using the green & blue (yz) components of src as texture coordinates */
1682 void pshader_glsl_texreg2gb(SHADER_OPCODE_ARG* arg) {
1683
1684 char tmpLine[255];
1685 char dst_str[100], src0_str[100];
1686 char dst_reg[50], src0_reg[50];
1687 char dst_mask[6], src0_mask[6];
1688 DWORD src0_regnum = arg->src[0] & WINED3DSP_REGNUM_MASK;
1689
1690 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1691 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1692
1693 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
1694 shader_addline(arg->buffer, "%stexture2D(Psampler%u, %s.yz))%s;\n",
1695 tmpLine, src0_regnum, dst_reg, dst_mask);
1696 }
1697
1698 /** Process the WINED3DSIO_TEXREG2RGB instruction in GLSL
1699 * Sample texture at dst using the rgb (xyz) components of src as texture coordinates */
1700 void pshader_glsl_texreg2rgb(SHADER_OPCODE_ARG* arg) {
1701
1702 char tmpLine[255];
1703 char dst_str[100], src0_str[100];
1704 char dst_reg[50], src0_reg[50];
1705 char dst_mask[6], src0_mask[6];
1706 char dimensions[5];
1707 DWORD src0_regnum = arg->src[0] & WINED3DSP_REGNUM_MASK;
1708 DWORD stype = arg->reg_maps->samplers[src0_regnum] & WINED3DSP_TEXTURETYPE_MASK;
1709 switch (stype) {
1710 case WINED3DSTT_2D: strcpy(dimensions, "2D"); break;
1711 case WINED3DSTT_CUBE: strcpy(dimensions, "Cube"); break;
1712 case WINED3DSTT_VOLUME: strcpy(dimensions, "3D"); break;
1713 default:
1714 strcpy(dimensions, "");
1715 FIXME("Unrecognized sampler type: %#x\n", stype);
1716 break;
1717 }
1718
1719 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1720 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1721
1722 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
1723 shader_addline(arg->buffer, "%stexture%s(Psampler%u, %s.%s))%s;\n",
1724 tmpLine, dimensions, src0_regnum, dst_reg, (stype == WINED3DSTT_2D) ? "xy" : "xyz", dst_mask);
1725 }
1726
1727 /** Process the WINED3DSIO_TEXKILL instruction in GLSL.
1728 * If any of the first 3 components are < 0, discard this pixel */
1729 void pshader_glsl_texkill(SHADER_OPCODE_ARG* arg) {
1730
1731 char dst_str[100], dst_name[50], dst_mask[6];
1732
1733 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_name, dst_mask, dst_str);
1734 shader_addline(arg->buffer, "if (any(lessThan(%s.xyz, vec3(0.0)))) discard;\n", dst_name);
1735 }
1736
1737 /** Process the WINED3DSIO_DP2ADD instruction in GLSL.
1738 * dst = dot2(src0, src1) + src2 */
1739 void pshader_glsl_dp2add(SHADER_OPCODE_ARG* arg) {
1740
1741 char tmpLine[256];
1742 char dst_str[100], src0_str[100], src1_str[100], src2_str[100];
1743 char dst_reg[50], src0_reg[50], src1_reg[50], src2_reg[50];
1744 char dst_mask[6], src0_mask[6], src1_mask[6], src2_mask[6];
1745
1746 shader_glsl_add_param(arg, arg->dst, 0, FALSE, dst_reg, dst_mask, dst_str);
1747 shader_glsl_add_param(arg, arg->src[0], arg->src_addr[0], TRUE, src0_reg, src0_mask, src0_str);
1748 shader_glsl_add_param(arg, arg->src[1], arg->src_addr[1], TRUE, src1_reg, src1_mask, src1_str);
1749 shader_glsl_add_param(arg, arg->src[2], arg->src_addr[2], TRUE, src2_reg, src2_mask, src2_str);
1750 shader_glsl_add_dst(arg->dst, dst_reg, dst_mask, tmpLine);
1751 shader_addline(arg->buffer, "%sdot(vec2(%s), vec2(%s)) + %s)%s;\n",
1752 tmpLine, src0_str, src1_str, src2_str, dst_mask);
1753 }
1754
1755 void pshader_glsl_input_pack(
1756 SHADER_BUFFER* buffer,
1757 semantic* semantics_in) {
1758
1759 unsigned int i;
1760
1761 for (i = 0; i < MAX_REG_INPUT; i++) {
1762
1763 DWORD usage_token = semantics_in[i].usage;
1764 DWORD register_token = semantics_in[i].reg;
1765 DWORD usage, usage_idx;
1766 char reg_mask[6];
1767
1768 /* Uninitialized */
1769 if (!usage_token) continue;
1770 usage = (usage_token & WINED3DSP_DCL_USAGE_MASK) >> WINED3DSP_DCL_USAGE_SHIFT;
1771 usage_idx = (usage_token & WINED3DSP_DCL_USAGEINDEX_MASK) >> WINED3DSP_DCL_USAGEINDEX_SHIFT;
1772 shader_glsl_get_write_mask(register_token, reg_mask);
1773
1774 switch(usage) {
1775
1776 case D3DDECLUSAGE_COLOR:
1777 if (usage_idx == 0)
1778 shader_addline(buffer, "IN%u%s = vec4(gl_Color)%s;\n",
1779 i, reg_mask, reg_mask);
1780 else if (usage_idx == 1)
1781 shader_addline(buffer, "IN%u%s = vec4(gl_SecondaryColor)%s;\n",
1782 i, reg_mask, reg_mask);
1783 else
1784 shader_addline(buffer, "IN%u%s = vec4(unsupported_color_input)%s;\n",
1785 i, reg_mask, reg_mask);
1786 break;
1787
1788 case D3DDECLUSAGE_TEXCOORD:
1789 shader_addline(buffer, "IN%u%s = vec4(gl_TexCoord[%u])%s;\n",
1790 i, reg_mask, usage_idx, reg_mask );
1791 break;
1792
1793 case D3DDECLUSAGE_FOG:
1794 shader_addline(buffer, "IN%u%s = vec4(gl_FogFragCoord)%s;\n",
1795 i, reg_mask, reg_mask);
1796 break;
1797
1798 default:
1799 shader_addline(buffer, "IN%u%s = vec4(unsupported_input)%s;\n",
1800 i, reg_mask, reg_mask);
1801 }
1802 }
1803 }
1804
1805 /*********************************************
1806 * Vertex Shader Specific Code begins here
1807 ********************************************/
1808
1809 void vshader_glsl_output_unpack(
1810 SHADER_BUFFER* buffer,
1811 semantic* semantics_out) {
1812
1813 unsigned int i;
1814
1815 for (i = 0; i < MAX_REG_OUTPUT; i++) {
1816
1817 DWORD usage_token = semantics_out[i].usage;
1818 DWORD register_token = semantics_out[i].reg;
1819 DWORD usage, usage_idx;
1820 char reg_mask[6];
1821
1822 /* Uninitialized */
1823 if (!usage_token) continue;
1824
1825 usage = (usage_token & WINED3DSP_DCL_USAGE_MASK) >> WINED3DSP_DCL_USAGE_SHIFT;
1826 usage_idx = (usage_token & WINED3DSP_DCL_USAGEINDEX_MASK) >> WINED3DSP_DCL_USAGEINDEX_SHIFT;
1827 shader_glsl_get_write_mask(register_token, reg_mask);
1828
1829 switch(usage) {
1830
1831 case D3DDECLUSAGE_COLOR:
1832 if (usage_idx == 0)
1833 shader_addline(buffer, "gl_FrontColor%s = OUT%u%s;\n", reg_mask, i, reg_mask);
1834 else if (usage_idx == 1)
1835 shader_addline(buffer, "gl_FrontSecondaryColor%s = OUT%u%s;\n", reg_mask, i, reg_mask);
1836 else
1837 shader_addline(buffer, "unsupported_color_output%s = OUT%u%s;\n", reg_mask, i, reg_mask);
1838 break;
1839
1840 case D3DDECLUSAGE_POSITION:
1841 shader_addline(buffer, "gl_Position%s = OUT%u%s;\n", reg_mask, i, reg_mask);
1842 break;
1843
1844 case D3DDECLUSAGE_TEXCOORD:
1845 shader_addline(buffer, "gl_TexCoord[%u]%s = OUT%u%s;\n",
1846 usage_idx, reg_mask, i, reg_mask);
1847 break;
1848
1849 case WINED3DSHADERDECLUSAGE_PSIZE:
1850 shader_addline(buffer, "gl_PointSize = OUT%u.x;\n", i);
1851 break;
1852
1853 case WINED3DSHADERDECLUSAGE_FOG:
1854 shader_addline(buffer, "gl_FogFragCoord%s = OUT%u%s;\n", reg_mask, i, reg_mask);
1855 break;
1856
1857 default:
1858 shader_addline(buffer, "unsupported_output%s = OUT%u%s;\n", reg_mask, i, reg_mask);
1859 }
1860 }
1861 }
1862
1863 /** Attach a GLSL pixel or vertex shader object to the shader program */
1864 static void attach_glsl_shader(IWineD3DDevice *iface, IWineD3DBaseShader* shader) {
1865 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1866 WineD3D_GL_Info *gl_info = &((IWineD3DImpl *)(This->wineD3D))->gl_info;
1867 GLhandleARB shaderObj = ((IWineD3DBaseShaderImpl*)shader)->baseShader.prgId;
1868 if (This->stateBlock->glsl_program && shaderObj != 0) {
1869 TRACE("Attaching GLSL shader object %u to program %u\n", shaderObj, This->stateBlock->glsl_program->programId);
1870 GL_EXTCALL(glAttachObjectARB(This->stateBlock->glsl_program->programId, shaderObj));
1871 checkGLcall("glAttachObjectARB");
1872 }
1873 }
1874
1875 /** Sets the GLSL program ID for the given pixel and vertex shader combination.
1876 * It sets the programId on the current StateBlock (because it should be called
1877 * inside of the DrawPrimitive() part of the render loop).
1878 *
1879 * If a program for the given combination does not exist, create one, and store
1880 * the program in the list. If it creates a program, it will link the given
1881 * objects, too.
1882 *
1883 * We keep the shader programs around on a list because linking
1884 * shader objects together is an expensive operation. It's much
1885 * faster to loop through a list of pre-compiled & linked programs
1886 * each time that the application sets a new pixel or vertex shader
1887 * than it is to re-link them together at that time.
1888 *
1889 * The list will be deleted in IWineD3DDevice::Release().
1890 */
1891 static void set_glsl_shader_program(IWineD3DDevice *iface) {
1892 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1893 WineD3D_GL_Info *gl_info = &((IWineD3DImpl *)(This->wineD3D))->gl_info;
1894 IWineD3DPixelShader *pshader = This->stateBlock->pixelShader;
1895 IWineD3DVertexShader *vshader = This->stateBlock->vertexShader;
1896 struct glsl_shader_prog_link *curLink = NULL;
1897 struct glsl_shader_prog_link *newLink = NULL;
1898 struct list *ptr = NULL;
1899 GLhandleARB programId = 0;
1900 int i;
1901 char glsl_name[8];
1902
1903 ptr = list_head( &This->glsl_shader_progs );
1904 while (ptr) {
1905 /* At least one program exists - see if it matches our ps/vs combination */
1906 curLink = LIST_ENTRY( ptr, struct glsl_shader_prog_link, entry );
1907 if (vshader == curLink->vertexShader && pshader == curLink->pixelShader) {
1908 /* Existing Program found, use it */
1909 TRACE("Found existing program (%u) for this vertex/pixel shader combination\n",
1910 curLink->programId);
1911 This->stateBlock->glsl_program = curLink;
1912 return;
1913 }
1914 /* This isn't the entry we need - try the next one */
1915 ptr = list_next( &This->glsl_shader_progs, ptr );
1916 }
1917
1918 /* If we get to this point, then no matching program exists, so we create one */
1919 programId = GL_EXTCALL(glCreateProgramObjectARB());
1920 TRACE("Created new GLSL shader program %u\n", programId);
1921
1922 /* Allocate a new link for the list of programs */
1923 newLink = HeapAlloc(GetProcessHeap(), 0, sizeof(struct glsl_shader_prog_link));
1924 newLink->programId = programId;
1925 This->stateBlock->glsl_program = newLink;
1926
1927 /* Attach GLSL vshader */
1928 if (NULL != vshader && This->vs_selected_mode == SHADER_GLSL) {
1929 int i;
1930 int max_attribs = 16; /* TODO: Will this always be the case? It is at the moment... */
1931 char tmp_name[10];
1932
1933 TRACE("Attaching vertex shader to GLSL program\n");
1934 attach_glsl_shader(iface, (IWineD3DBaseShader*)vshader);
1935
1936 /* Bind vertex attributes to a corresponding index number to match
1937 * the same index numbers as ARB_vertex_programs (makes loading
1938 * vertex attributes simpler). With this method, we can use the
1939 * exact same code to load the attributes later for both ARB and
1940 * GLSL shaders.
1941 *
1942 * We have to do this here because we need to know the Program ID
1943 * in order to make the bindings work, and it has to be done prior
1944 * to linking the GLSL program. */
1945 for (i = 0; i < max_attribs; ++i) {
1946 snprintf(tmp_name, sizeof(tmp_name), "attrib%i", i);
1947 GL_EXTCALL(glBindAttribLocationARB(programId, i, tmp_name));
1948 }
1949 checkGLcall("glBindAttribLocationARB");
1950 newLink->vertexShader = vshader;
1951 }
1952
1953 /* Attach GLSL pshader */
1954 if (NULL != pshader && This->ps_selected_mode == SHADER_GLSL) {
1955 TRACE("Attaching pixel shader to GLSL program\n");
1956 attach_glsl_shader(iface, (IWineD3DBaseShader*)pshader);
1957 newLink->pixelShader = pshader;
1958 }
1959
1960 /* Link the program */
1961 TRACE("Linking GLSL shader program %u\n", programId);
1962 GL_EXTCALL(glLinkProgramARB(programId));
1963 print_glsl_info_log(&GLINFO_LOCATION, programId);
1964 list_add_head( &This->glsl_shader_progs, &newLink->entry);
1965
1966 newLink->vuniformF_locations = HeapAlloc(GetProcessHeap(), 0, sizeof(GLhandleARB) * GL_LIMITS(vshader_constantsF));
1967 for (i = 0; i < GL_LIMITS(vshader_constantsF); ++i) {
1968 snprintf(glsl_name, sizeof(glsl_name), "VC[%i]", i);
1969 newLink->vuniformF_locations[i] = GL_EXTCALL(glGetUniformLocationARB(programId, glsl_name));
1970 }
1971 newLink->puniformF_locations = HeapAlloc(GetProcessHeap(), 0, sizeof(GLhandleARB) * GL_LIMITS(pshader_constantsF));
1972 for (i = 0; i < GL_LIMITS(pshader_constantsF); ++i) {
1973 snprintf(glsl_name, sizeof(glsl_name), "PC[%i]", i);
1974 newLink->puniformF_locations[i] = GL_EXTCALL(glGetUniformLocationARB(programId, glsl_name));
1975 }
1976
1977 return;
1978 }
1979
1980 static GLhandleARB create_glsl_blt_shader(WineD3D_GL_Info *gl_info) {
1981 GLhandleARB program_id;
1982 GLhandleARB vshader_id, pshader_id;
1983 const char *blt_vshader[] = {
1984 "void main(void)\n"
1985 "{\n"
1986 " gl_Position = gl_Vertex;\n"
1987 " gl_FrontColor = vec4(1.0);\n"
1988 " gl_TexCoord[0].x = (gl_Vertex.x * 0.5) + 0.5;\n"
1989 " gl_TexCoord[0].y = (-gl_Vertex.y * 0.5) + 0.5;\n"
1990 "}\n"
1991 };
1992
1993 const char *blt_pshader[] = {
1994 "uniform sampler2D sampler;\n"
1995 "void main(void)\n"
1996 "{\n"
1997 " gl_FragDepth = texture2D(sampler, gl_TexCoord[0].xy).x;\n"
1998 "}\n"
1999 };
2000
2001 vshader_id = GL_EXTCALL(glCreateShaderObjectARB(GL_VERTEX_SHADER_ARB));
2002 GL_EXTCALL(glShaderSourceARB(vshader_id, 1, blt_vshader, NULL));
2003 GL_EXTCALL(glCompileShaderARB(vshader_id));
2004
2005 pshader_id = GL_EXTCALL(glCreateShaderObjectARB(GL_FRAGMENT_SHADER_ARB));
2006 GL_EXTCALL(glShaderSourceARB(pshader_id, 1, blt_pshader, NULL));
2007 GL_EXTCALL(glCompileShaderARB(pshader_id));
2008
2009 program_id = GL_EXTCALL(glCreateProgramObjectARB());
2010 GL_EXTCALL(glAttachObjectARB(program_id, vshader_id));
2011 GL_EXTCALL(glAttachObjectARB(program_id, pshader_id));
2012 GL_EXTCALL(glLinkProgramARB(program_id));
2013
2014 print_glsl_info_log(&GLINFO_LOCATION, program_id);
2015
2016 return program_id;
2017 }
2018
2019 static void shader_glsl_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {
2020 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
2021 WineD3D_GL_Info *gl_info = &((IWineD3DImpl *)(This->wineD3D))->gl_info;
2022 GLhandleARB program_id = 0;
2023
2024 if (useVS || usePS) set_glsl_shader_program(iface);
2025 else This->stateBlock->glsl_program = NULL;
2026
2027 program_id = This->stateBlock->glsl_program ? This->stateBlock->glsl_program->programId : 0;
2028 if (program_id) TRACE("Using GLSL program %u\n", program_id);
2029 GL_EXTCALL(glUseProgramObjectARB(program_id));
2030 checkGLcall("glUseProgramObjectARB");
2031 }
2032
2033 static void shader_glsl_select_depth_blt(IWineD3DDevice *iface) {
2034 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
2035 WineD3D_GL_Info *gl_info = &((IWineD3DImpl *)(This->wineD3D))->gl_info;
2036 static GLhandleARB program_id = 0;
2037 static GLhandleARB loc = -1;
2038
2039 if (!program_id) {
2040 program_id = create_glsl_blt_shader(gl_info);
2041 loc = GL_EXTCALL(glGetUniformLocationARB(program_id, "sampler"));
2042 }
2043
2044 GL_EXTCALL(glUseProgramObjectARB(program_id));
2045 GL_EXTCALL(glUniform1iARB(loc, 0));
2046 }
2047
2048 static void shader_glsl_cleanup(BOOL usePS, BOOL useVS) {
2049 /* Nothing to do */
2050 }
2051
2052 const shader_backend_t glsl_shader_backend = {
2053 &shader_glsl_select,
2054 &shader_glsl_select_depth_blt,
2055 &shader_glsl_load_constants,
2056 &shader_glsl_cleanup
2057 };
Maarten's multimedia branch