?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
wined3d: Add a glsl_dst_param_t type, pass it to shader_glsl_add_dst_param() instead...
[wine.git] / dlls / wined3d / glsl_shader.c
blob4c5b48915b67f06071b7ed13a83108b89af9c8ae
1 /*
2 * GLSL pixel and vertex shader implementation
3 *
4 * Copyright 2006 Jason Green
5 * Copyright 2006-2007 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 /*
23 * D3D shader asm has swizzles on source parameters, and write masks for
24 * destination parameters. GLSL uses swizzles for both. The result of this is
25 * that for example "mov dst.xw, src.zyxw" becomes "dst.xw = src.zw" in GLSL.
26 * Ie, to generate a proper GLSL source swizzle, we need to take the D3D write
27 * mask for the destination parameter into account.
28 */
29
30 #include "config.h"
31 #include <stdio.h>
32 #include "wined3d_private.h"
33
34 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
35
36 #define GLINFO_LOCATION (*gl_info)
37
38 typedef struct {
39 char reg_name[50];
40 char mask_str[6];
41 } glsl_dst_param_t;
42
43 typedef struct {
44 const char *name;
45 DWORD coord_mask;
46 } glsl_sample_function_t;
47
48 /** Prints the GLSL info log which will contain error messages if they exist */
49 void print_glsl_info_log(WineD3D_GL_Info *gl_info, GLhandleARB obj) {
50
51 int infologLength = 0;
52 char *infoLog;
53
54 GL_EXTCALL(glGetObjectParameterivARB(obj,
55 GL_OBJECT_INFO_LOG_LENGTH_ARB,
56 &infologLength));
57
58 /* A size of 1 is just a null-terminated string, so the log should be bigger than
59 * that if there are errors. */
60 if (infologLength > 1)
61 {
62 infoLog = (char *)HeapAlloc(GetProcessHeap(), 0, infologLength);
63 GL_EXTCALL(glGetInfoLogARB(obj, infologLength, NULL, infoLog));
64 FIXME("Error received from GLSL shader #%u: %s\n", obj, debugstr_a(infoLog));
65 HeapFree(GetProcessHeap(), 0, infoLog);
66 }
67 }
68
69 /**
70 * Loads (pixel shader) samplers
71 */
72 void shader_glsl_load_psamplers(
73 WineD3D_GL_Info *gl_info,
74 IWineD3DStateBlock* iface) {
75
76 IWineD3DStateBlockImpl* stateBlock = (IWineD3DStateBlockImpl*) iface;
77 GLhandleARB programId = stateBlock->glsl_program->programId;
78 GLhandleARB name_loc;
79 int i;
80 char sampler_name[20];
81
82 for (i=0; i< GL_LIMITS(samplers); ++i) {
83 if (stateBlock->textures[i] != NULL) {
84 snprintf(sampler_name, sizeof(sampler_name), "Psampler%d", i);
85 name_loc = GL_EXTCALL(glGetUniformLocationARB(programId, sampler_name));
86 if (name_loc != -1) {
87 TRACE("Loading %s for texture %d\n", sampler_name, i);
88 GL_EXTCALL(glUniform1iARB(name_loc, i));
89 checkGLcall("glUniform1iARB");
90 }
91 }
92 }
93 }
94
95 /**
96 * Loads floating point constants (aka uniforms) into the currently set GLSL program.
97 * When constant_list == NULL, it will load all the constants.
98 */
99 static void shader_glsl_load_constantsF(IWineD3DBaseShaderImpl* This, WineD3D_GL_Info *gl_info,
100 unsigned int max_constants, float* constants, GLhandleARB *constant_locations,
101 struct list *constant_list) {
102 local_constant* lconst;
103 GLhandleARB tmp_loc;
104 int i;
105
106 if (!constant_list) {
107 if (TRACE_ON(d3d_shader)) {
108 for (i = 0; i < max_constants; ++i) {
109 tmp_loc = constant_locations[i];
110 if (tmp_loc != -1) {
111 TRACE("Loading constants %i: %f, %f, %f, %f\n", i,
112 constants[i * 4 + 0], constants[i * 4 + 1],
113 constants[i * 4 + 2], constants[i * 4 + 3]);
114 }
115 }
116 }
117 for (i = 0; i < max_constants; ++i) {
118 tmp_loc = constant_locations[i];
119 if (tmp_loc != -1) {
120 /* We found this uniform name in the program - go ahead and send the data */
121 GL_EXTCALL(glUniform4fvARB(tmp_loc, 1, constants + (i * 4)));
122 }
123 }
124 checkGLcall("glUniform4fvARB()");
125 } else {
126 constant_entry *constant;
127 if (TRACE_ON(d3d_shader)) {
128 LIST_FOR_EACH_ENTRY(constant, constant_list, constant_entry, entry) {
129 i = constant->idx;
130 tmp_loc = constant_locations[i];
131 if (tmp_loc != -1) {
132 TRACE("Loading constants %i: %f, %f, %f, %f\n", i,
133 constants[i * 4 + 0], constants[i * 4 + 1],
134 constants[i * 4 + 2], constants[i * 4 + 3]);
135 }
136 }
137 }
138 LIST_FOR_EACH_ENTRY(constant, constant_list, constant_entry, entry) {
139 i = constant->idx;
140 tmp_loc = constant_locations[i];
141 if (tmp_loc != -1) {
142 /* We found this uniform name in the program - go ahead and send the data */
143 GL_EXTCALL(glUniform4fvARB(tmp_loc, 1, constants + (i * 4)));
144 }
145 }
146 checkGLcall("glUniform4fvARB()");
147 }
148
149 /* Load immediate constants */
150 if (TRACE_ON(d3d_shader)) {
151 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
152 tmp_loc = constant_locations[lconst->idx];
153 if (tmp_loc != -1) {
154 GLfloat* values = (GLfloat*)lconst->value;
155 TRACE("Loading local constants %i: %f, %f, %f, %f\n", lconst->idx,
156 values[0], values[1], values[2], values[3]);
157 }
158 }
159 }
160 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
161 tmp_loc = constant_locations[lconst->idx];
162 if (tmp_loc != -1) {
163 /* We found this uniform name in the program - go ahead and send the data */
164 GL_EXTCALL(glUniform4fvARB(tmp_loc, 1, (GLfloat*)lconst->value));
165 }
166 }
167 checkGLcall("glUniform4fvARB()");
168 }
169
170 /**
171 * Loads integer constants (aka uniforms) into the currently set GLSL program.
172 * When @constants_set == NULL, it will load all the constants.
173 */
174 void shader_glsl_load_constantsI(
175 IWineD3DBaseShaderImpl* This,
176 WineD3D_GL_Info *gl_info,
177 GLhandleARB programId,
178 unsigned max_constants,
179 int* constants,
180 BOOL* constants_set) {
181
182 GLhandleARB tmp_loc;
183 int i;
184 char tmp_name[8];
185 char is_pshader = shader_is_pshader_version(This->baseShader.hex_version);
186 const char* prefix = is_pshader? "PI":"VI";
187 struct list* ptr;
188
189 for (i=0; i<max_constants; ++i) {
190 if (NULL == constants_set || constants_set[i]) {
191
192 TRACE("Loading constants %i: %i, %i, %i, %i\n",
193 i, constants[i*4], constants[i*4+1], constants[i*4+2], constants[i*4+3]);
194
195 /* TODO: Benchmark and see if it would be beneficial to store the
196 * locations of the constants to avoid looking up each time */
197 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, i);
198 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
199 if (tmp_loc != -1) {
200 /* We found this uniform name in the program - go ahead and send the data */
201 GL_EXTCALL(glUniform4ivARB(tmp_loc, 1, &constants[i*4]));
202 checkGLcall("glUniform4ivARB");
203 }
204 }
205 }
206
207 /* Load immediate constants */
208 ptr = list_head(&This->baseShader.constantsI);
209 while (ptr) {
210 local_constant* lconst = LIST_ENTRY(ptr, struct local_constant, entry);
211 unsigned int idx = lconst->idx;
212 GLint* values = (GLint*) lconst->value;
213
214 TRACE("Loading local constants %i: %i, %i, %i, %i\n", idx,
215 values[0], values[1], values[2], values[3]);
216
217 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, idx);
218 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
219 if (tmp_loc != -1) {
220 /* We found this uniform name in the program - go ahead and send the data */
221 GL_EXTCALL(glUniform4ivARB(tmp_loc, 1, values));
222 checkGLcall("glUniform4ivARB");
223 }
224 ptr = list_next(&This->baseShader.constantsI, ptr);
225 }
226 }
227
228 /**
229 * Loads boolean constants (aka uniforms) into the currently set GLSL program.
230 * When @constants_set == NULL, it will load all the constants.
231 */
232 void shader_glsl_load_constantsB(
233 IWineD3DBaseShaderImpl* This,
234 WineD3D_GL_Info *gl_info,
235 GLhandleARB programId,
236 unsigned max_constants,
237 BOOL* constants,
238 BOOL* constants_set) {
239
240 GLhandleARB tmp_loc;
241 int i;
242 char tmp_name[8];
243 char is_pshader = shader_is_pshader_version(This->baseShader.hex_version);
244 const char* prefix = is_pshader? "PB":"VB";
245 struct list* ptr;
246
247 for (i=0; i<max_constants; ++i) {
248 if (NULL == constants_set || constants_set[i]) {
249
250 TRACE("Loading constants %i: %i;\n", i, constants[i*4]);
251
252 /* TODO: Benchmark and see if it would be beneficial to store the
253 * locations of the constants to avoid looking up each time */
254 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, i);
255 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
256 if (tmp_loc != -1) {
257 /* We found this uniform name in the program - go ahead and send the data */
258 GL_EXTCALL(glUniform1ivARB(tmp_loc, 1, &constants[i*4]));
259 checkGLcall("glUniform1ivARB");
260 }
261 }
262 }
263
264 /* Load immediate constants */
265 ptr = list_head(&This->baseShader.constantsB);
266 while (ptr) {
267 local_constant* lconst = LIST_ENTRY(ptr, struct local_constant, entry);
268 unsigned int idx = lconst->idx;
269 GLint* values = (GLint*) lconst->value;
270
271 TRACE("Loading local constants %i: %i\n", idx, values[0]);
272
273 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, idx);
274 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
275 if (tmp_loc != -1) {
276 /* We found this uniform name in the program - go ahead and send the data */
277 GL_EXTCALL(glUniform1ivARB(tmp_loc, 1, values));
278 checkGLcall("glUniform1ivARB");
279 }
280 ptr = list_next(&This->baseShader.constantsB, ptr);
281 }
282 }
283
284
285
286 /**
287 * Loads the app-supplied constants into the currently set GLSL program.
288 */
289 void shader_glsl_load_constants(
290 IWineD3DDevice* device,
291 char usePixelShader,
292 char useVertexShader) {
293
294 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) device;
295 IWineD3DStateBlockImpl* stateBlock = deviceImpl->stateBlock;
296 WineD3D_GL_Info *gl_info = &((IWineD3DImpl*) deviceImpl->wineD3D)->gl_info;
297
298 GLhandleARB *constant_locations;
299 struct list *constant_list;
300 GLhandleARB programId;
301
302 if (!stateBlock->glsl_program) {
303 /* No GLSL program set - nothing to do. */
304 return;
305 }
306 programId = stateBlock->glsl_program->programId;
307
308 if (useVertexShader) {
309 IWineD3DBaseShaderImpl* vshader = (IWineD3DBaseShaderImpl*) stateBlock->vertexShader;
310 IWineD3DVertexShaderImpl* vshader_impl = (IWineD3DVertexShaderImpl*) vshader;
311 GLint pos;
312
313 IWineD3DVertexDeclarationImpl* vertexDeclaration =
314 (IWineD3DVertexDeclarationImpl*) vshader_impl->vertexDeclaration;
315
316 constant_locations = stateBlock->glsl_program->vuniformF_locations;
317 constant_list = &stateBlock->set_vconstantsF;
318
319 if (NULL != vertexDeclaration && NULL != vertexDeclaration->constants) {
320 /* Load DirectX 8 float constants/uniforms for vertex shader */
321 shader_glsl_load_constantsF(vshader, gl_info, GL_LIMITS(vshader_constantsF),
322 vertexDeclaration->constants, constant_locations, NULL);
323 }
324
325 /* Load DirectX 9 float constants/uniforms for vertex shader */
326 shader_glsl_load_constantsF(vshader, gl_info, GL_LIMITS(vshader_constantsF),
327 stateBlock->vertexShaderConstantF, constant_locations, constant_list);
328
329 /* Load DirectX 9 integer constants/uniforms for vertex shader */
330 shader_glsl_load_constantsI(vshader, gl_info, programId, MAX_CONST_I,
331 stateBlock->vertexShaderConstantI,
332 stateBlock->set.vertexShaderConstantsI);
333
334 /* Load DirectX 9 boolean constants/uniforms for vertex shader */
335 shader_glsl_load_constantsB(vshader, gl_info, programId, MAX_CONST_B,
336 stateBlock->vertexShaderConstantB,
337 stateBlock->set.vertexShaderConstantsB);
338
339 /* Upload the position fixup params */
340 pos = GL_EXTCALL(glGetUniformLocationARB(programId, "posFixup"));
341 checkGLcall("glGetUniformLocationARB");
342 GL_EXTCALL(glUniform4fvARB(pos, 1, &deviceImpl->posFixup[0]));
343 checkGLcall("glUniform4fvARB");
344 }
345
346 if (usePixelShader) {
347
348 IWineD3DBaseShaderImpl* pshader = (IWineD3DBaseShaderImpl*) stateBlock->pixelShader;
349
350 constant_locations = stateBlock->glsl_program->puniformF_locations;
351 constant_list = &stateBlock->set_pconstantsF;
352
353 /* Load pixel shader samplers */
354 shader_glsl_load_psamplers(gl_info, (IWineD3DStateBlock*) stateBlock);
355
356 /* Load DirectX 9 float constants/uniforms for pixel shader */
357 shader_glsl_load_constantsF(pshader, gl_info, GL_LIMITS(pshader_constantsF),
358 stateBlock->pixelShaderConstantF, constant_locations, constant_list);
359
360 /* Load DirectX 9 integer constants/uniforms for pixel shader */
361 shader_glsl_load_constantsI(pshader, gl_info, programId, MAX_CONST_I,
362 stateBlock->pixelShaderConstantI,
363 stateBlock->set.pixelShaderConstantsI);
364
365 /* Load DirectX 9 boolean constants/uniforms for pixel shader */
366 shader_glsl_load_constantsB(pshader, gl_info, programId, MAX_CONST_B,
367 stateBlock->pixelShaderConstantB,
368 stateBlock->set.pixelShaderConstantsB);
369 }
370 }
371
372 /** Generate the variable & register declarations for the GLSL output target */
373 void shader_generate_glsl_declarations(
374 IWineD3DBaseShader *iface,
375 shader_reg_maps* reg_maps,
376 SHADER_BUFFER* buffer,
377 WineD3D_GL_Info* gl_info) {
378
379 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
380 int i;
381
382 /* There are some minor differences between pixel and vertex shaders */
383 char pshader = shader_is_pshader_version(This->baseShader.hex_version);
384 char prefix = pshader ? 'P' : 'V';
385
386 /* Prototype the subroutines */
387 for (i = 0; i < This->baseShader.limits.label; i++) {
388 if (reg_maps->labels[i])
389 shader_addline(buffer, "void subroutine%lu();\n", i);
390 }
391
392 /* Declare the constants (aka uniforms) */
393 if (This->baseShader.limits.constant_float > 0) {
394 unsigned max_constantsF = min(This->baseShader.limits.constant_float,
395 (pshader ? GL_LIMITS(pshader_constantsF) : GL_LIMITS(vshader_constantsF)));
396 shader_addline(buffer, "uniform vec4 %cC[%u];\n", prefix, max_constantsF);
397 }
398
399 if (This->baseShader.limits.constant_int > 0)
400 shader_addline(buffer, "uniform ivec4 %cI[%u];\n", prefix, This->baseShader.limits.constant_int);
401
402 if (This->baseShader.limits.constant_bool > 0)
403 shader_addline(buffer, "uniform bool %cB[%u];\n", prefix, This->baseShader.limits.constant_bool);
404
405 if(!pshader)
406 shader_addline(buffer, "uniform vec4 posFixup;\n");
407
408 /* Declare texture samplers */
409 for (i = 0; i < This->baseShader.limits.sampler; i++) {
410 if (reg_maps->samplers[i]) {
411
412 DWORD stype = reg_maps->samplers[i] & WINED3DSP_TEXTURETYPE_MASK;
413 switch (stype) {
414
415 case WINED3DSTT_1D:
416 shader_addline(buffer, "uniform sampler1D %csampler%lu;\n", prefix, i);
417 break;
418 case WINED3DSTT_2D:
419 shader_addline(buffer, "uniform sampler2D %csampler%lu;\n", prefix, i);
420 break;
421 case WINED3DSTT_CUBE:
422 shader_addline(buffer, "uniform samplerCube %csampler%lu;\n", prefix, i);
423 break;
424 case WINED3DSTT_VOLUME:
425 shader_addline(buffer, "uniform sampler3D %csampler%lu;\n", prefix, i);
426 break;
427 default:
428 shader_addline(buffer, "uniform unsupported_sampler %csampler%lu;\n", prefix, i);
429 FIXME("Unrecognized sampler type: %#x\n", stype);
430 break;
431 }
432 }
433 }
434
435 /* Declare address variables */
436 for (i = 0; i < This->baseShader.limits.address; i++) {
437 if (reg_maps->address[i])
438 shader_addline(buffer, "ivec4 A%d;\n", i);
439 }
440
441 /* Declare texture coordinate temporaries and initialize them */
442 for (i = 0; i < This->baseShader.limits.texcoord; i++) {
443 if (reg_maps->texcoord[i])
444 shader_addline(buffer, "vec4 T%lu = gl_TexCoord[%lu];\n", i, i);
445 }
446
447 /* Declare input register temporaries */
448 for (i=0; i < This->baseShader.limits.packed_input; i++) {
449 if (reg_maps->packed_input[i])
450 shader_addline(buffer, "vec4 IN%lu;\n", i);
451 }
452
453 /* Declare output register temporaries */
454 for (i = 0; i < This->baseShader.limits.packed_output; i++) {
455 if (reg_maps->packed_output[i])
456 shader_addline(buffer, "vec4 OUT%lu;\n", i);
457 }
458
459 /* Declare temporary variables */
460 for(i = 0; i < This->baseShader.limits.temporary; i++) {
461 if (reg_maps->temporary[i])
462 shader_addline(buffer, "vec4 R%lu;\n", i);
463 }
464
465 /* Declare attributes */
466 for (i = 0; i < This->baseShader.limits.attributes; i++) {
467 if (reg_maps->attributes[i])
468 shader_addline(buffer, "attribute vec4 attrib%i;\n", i);
469 }
470
471 /* Declare loop register aL */
472 if (reg_maps->loop) {
473 shader_addline(buffer, "int aL;\n");
474 shader_addline(buffer, "int tmpInt;\n");
475 }
476
477 /* Temporary variables for matrix operations */
478 shader_addline(buffer, "vec4 tmp0;\n");
479 shader_addline(buffer, "vec4 tmp1;\n");
480
481 /* Start the main program */
482 shader_addline(buffer, "void main() {\n");
483 }
484
485 /*****************************************************************************
486 * Functions to generate GLSL strings from DirectX Shader bytecode begin here.
487 *
488 * For more information, see http://wiki.winehq.org/DirectX-Shaders
489 ****************************************************************************/
490
491 /* Prototypes */
492 static void shader_glsl_add_src_param(SHADER_OPCODE_ARG* arg, const DWORD param,
493 const DWORD addr_token, DWORD mask, char *reg_name, char *swizzle, char *out_str);
494
495 /** Used for opcode modifiers - They multiply the result by the specified amount */
496 static const char * const shift_glsl_tab[] = {
497 "", /* 0 (none) */
498 "2.0 * ", /* 1 (x2) */
499 "4.0 * ", /* 2 (x4) */
500 "8.0 * ", /* 3 (x8) */
501 "16.0 * ", /* 4 (x16) */
502 "32.0 * ", /* 5 (x32) */
503 "", /* 6 (x64) */
504 "", /* 7 (x128) */
505 "", /* 8 (d256) */
506 "", /* 9 (d128) */
507 "", /* 10 (d64) */
508 "", /* 11 (d32) */
509 "0.0625 * ", /* 12 (d16) */
510 "0.125 * ", /* 13 (d8) */
511 "0.25 * ", /* 14 (d4) */
512 "0.5 * " /* 15 (d2) */
513 };
514
515 /* Generate a GLSL parameter that does the input modifier computation and return the input register/mask to use */
516 static void shader_glsl_gen_modifier (
517 const DWORD instr,
518 const char *in_reg,
519 const char *in_regswizzle,
520 char *out_str) {
521
522 out_str[0] = 0;
523
524 if (instr == WINED3DSIO_TEXKILL)
525 return;
526
527 switch (instr & WINED3DSP_SRCMOD_MASK) {
528 case WINED3DSPSM_DZ: /* Need to handle this in the instructions itself (texld & texcrd). */
529 case WINED3DSPSM_DW:
530 case WINED3DSPSM_NONE:
531 sprintf(out_str, "%s%s", in_reg, in_regswizzle);
532 break;
533 case WINED3DSPSM_NEG:
534 sprintf(out_str, "-%s%s", in_reg, in_regswizzle);
535 break;
536 case WINED3DSPSM_NOT:
537 sprintf(out_str, "!%s%s", in_reg, in_regswizzle);
538 break;
539 case WINED3DSPSM_BIAS:
540 sprintf(out_str, "(%s%s - vec4(0.5)%s)", in_reg, in_regswizzle, in_regswizzle);
541 break;
542 case WINED3DSPSM_BIASNEG:
543 sprintf(out_str, "-(%s%s - vec4(0.5)%s)", in_reg, in_regswizzle, in_regswizzle);
544 break;
545 case WINED3DSPSM_SIGN:
546 sprintf(out_str, "(2.0 * (%s%s - 0.5))", in_reg, in_regswizzle);
547 break;
548 case WINED3DSPSM_SIGNNEG:
549 sprintf(out_str, "-(2.0 * (%s%s - 0.5))", in_reg, in_regswizzle);
550 break;
551 case WINED3DSPSM_COMP:
552 sprintf(out_str, "(1.0 - %s%s)", in_reg, in_regswizzle);
553 break;
554 case WINED3DSPSM_X2:
555 sprintf(out_str, "(2.0 * %s%s)", in_reg, in_regswizzle);
556 break;
557 case WINED3DSPSM_X2NEG:
558 sprintf(out_str, "-(2.0 * %s%s)", in_reg, in_regswizzle);
559 break;
560 case WINED3DSPSM_ABS:
561 sprintf(out_str, "abs(%s%s)", in_reg, in_regswizzle);
562 break;
563 case WINED3DSPSM_ABSNEG:
564 sprintf(out_str, "-abs(%s%s)", in_reg, in_regswizzle);
565 break;
566 default:
567 FIXME("Unhandled modifier %u\n", (instr & WINED3DSP_SRCMOD_MASK));
568 sprintf(out_str, "%s%s", in_reg, in_regswizzle);
569 }
570 }
571
572 /** Writes the GLSL variable name that corresponds to the register that the
573 * DX opcode parameter is trying to access */
574 static void shader_glsl_get_register_name(
575 const DWORD param,
576 const DWORD addr_token,
577 char* regstr,
578 BOOL* is_color,
579 SHADER_OPCODE_ARG* arg) {
580
581 /* oPos, oFog and oPts in D3D */
582 static const char * const hwrastout_reg_names[] = { "gl_Position", "gl_FogFragCoord", "gl_PointSize" };
583
584 DWORD reg = param & WINED3DSP_REGNUM_MASK;
585 DWORD regtype = shader_get_regtype(param);
586 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) arg->shader;
587 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
588 WineD3D_GL_Info* gl_info = &((IWineD3DImpl*)deviceImpl->wineD3D)->gl_info;
589
590 char pshader = shader_is_pshader_version(This->baseShader.hex_version);
591 char tmpStr[50];
592
593 *is_color = FALSE;
594
595 switch (regtype) {
596 case WINED3DSPR_TEMP:
597 sprintf(tmpStr, "R%u", reg);
598 break;
599 case WINED3DSPR_INPUT:
600 if (pshader) {
601 /* Pixel shaders >= 3.0 */
602 if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3)
603 sprintf(tmpStr, "IN%u", reg);
604 else {
605 if (reg==0)
606 strcpy(tmpStr, "gl_Color");
607 else
608 strcpy(tmpStr, "gl_SecondaryColor");
609 }
610 } else {
611 if (vshader_input_is_color((IWineD3DVertexShader*) This, reg))
612 *is_color = TRUE;
613 sprintf(tmpStr, "attrib%u", reg);
614 }
615 break;
616 case WINED3DSPR_CONST:
617 {
618 const char* prefix = pshader? "PC":"VC";
619
620 /* Relative addressing */
621 if (param & WINED3DSHADER_ADDRMODE_RELATIVE) {
622
623 /* Relative addressing on shaders 2.0+ have a relative address token,
624 * prior to that, it was hard-coded as "A0.x" because there's only 1 register */
625 if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 2) {
626 char relStr[100], relReg[50], relMask[6];
627 shader_glsl_add_src_param(arg, addr_token, 0, WINED3DSP_WRITEMASK_0, relReg, relMask, relStr);
628 sprintf(tmpStr, "%s[%s + %u]", prefix, relStr, reg);
629 } else
630 sprintf(tmpStr, "%s[A0.x + %u]", prefix, reg);
631
632 } else
633 sprintf(tmpStr, "%s[%u]", prefix, reg);
634
635 break;
636 }
637 case WINED3DSPR_CONSTINT:
638 if (pshader)
639 sprintf(tmpStr, "PI[%u]", reg);
640 else
641 sprintf(tmpStr, "VI[%u]", reg);
642 break;
643 case WINED3DSPR_CONSTBOOL:
644 if (pshader)
645 sprintf(tmpStr, "PB[%u]", reg);
646 else
647 sprintf(tmpStr, "VB[%u]", reg);
648 break;
649 case WINED3DSPR_TEXTURE: /* case WINED3DSPR_ADDR: */
650 if (pshader) {
651 sprintf(tmpStr, "T%u", reg);
652 } else {
653 sprintf(tmpStr, "A%u", reg);
654 }
655 break;
656 case WINED3DSPR_LOOP:
657 sprintf(tmpStr, "aL");
658 break;
659 case WINED3DSPR_SAMPLER:
660 if (pshader)
661 sprintf(tmpStr, "Psampler%u", reg);
662 else
663 sprintf(tmpStr, "Vsampler%u", reg);
664 break;
665 case WINED3DSPR_COLOROUT:
666 if (reg >= GL_LIMITS(buffers)) {
667 WARN("Write to render target %u, only %d supported\n", reg, 4);
668 }
669 if (GL_SUPPORT(ARB_DRAW_BUFFERS)) {
670 sprintf(tmpStr, "gl_FragData[%u]", reg);
671 } else { /* On older cards with GLSL support like the GeforceFX there's only one buffer. */
672 sprintf(tmpStr, "gl_FragColor");
673 }
674 break;
675 case WINED3DSPR_RASTOUT:
676 sprintf(tmpStr, "%s", hwrastout_reg_names[reg]);
677 break;
678 case WINED3DSPR_DEPTHOUT:
679 sprintf(tmpStr, "gl_FragDepth");
680 break;
681 case WINED3DSPR_ATTROUT:
682 if (reg == 0) {
683 sprintf(tmpStr, "gl_FrontColor");
684 } else {
685 sprintf(tmpStr, "gl_FrontSecondaryColor");
686 }
687 break;
688 case WINED3DSPR_TEXCRDOUT:
689 /* Vertex shaders >= 3.0: WINED3DSPR_OUTPUT */
690 if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3)
691 sprintf(tmpStr, "OUT%u", reg);
692 else
693 sprintf(tmpStr, "gl_TexCoord[%u]", reg);
694 break;
695 default:
696 FIXME("Unhandled register name Type(%d)\n", regtype);
697 sprintf(tmpStr, "unrecognized_register");
698 break;
699 }
700
701 strcat(regstr, tmpStr);
702 }
703
704 /* Get the GLSL write mask for the destination register */
705 static DWORD shader_glsl_get_write_mask(const DWORD param, char *write_mask) {
706 char *ptr = write_mask;
707 DWORD mask = param & WINED3DSP_WRITEMASK_ALL;
708
709 /* gl_FogFragCoord and glPointSize are floats, fixup the write mask. */
710 if ((shader_get_regtype(param) == WINED3DSPR_RASTOUT) && ((param & WINED3DSP_REGNUM_MASK) != 0)) {
711 mask = WINED3DSP_WRITEMASK_0;
712 } else {
713 *ptr++ = '.';
714 if (param & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
715 if (param & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
716 if (param & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
717 if (param & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
718 }
719
720 *ptr = '\0';
721
722 return mask;
723 }
724
725 static size_t shader_glsl_get_write_mask_size(DWORD write_mask) {
726 size_t size = 0;
727
728 if (write_mask & WINED3DSP_WRITEMASK_0) ++size;
729 if (write_mask & WINED3DSP_WRITEMASK_1) ++size;
730 if (write_mask & WINED3DSP_WRITEMASK_2) ++size;
731 if (write_mask & WINED3DSP_WRITEMASK_3) ++size;
732
733 return size;
734 }
735
736 static void shader_glsl_get_swizzle(const DWORD param, BOOL fixup, DWORD mask, 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 DWORD swizzle = (param & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
741 const char *swizzle_chars = fixup ? "zyxw" : "xyzw";
742 char *ptr = swizzle_str;
743
744 *ptr++ = '.';
745 /* swizzle bits fields: wwzzyyxx */
746 if (mask & WINED3DSP_WRITEMASK_0) *ptr++ = swizzle_chars[swizzle & 0x03];
747 if (mask & WINED3DSP_WRITEMASK_1) *ptr++ = swizzle_chars[(swizzle >> 2) & 0x03];
748 if (mask & WINED3DSP_WRITEMASK_2) *ptr++ = swizzle_chars[(swizzle >> 4) & 0x03];
749 if (mask & WINED3DSP_WRITEMASK_3) *ptr++ = swizzle_chars[(swizzle >> 6) & 0x03];
750
751 *ptr = '\0';
752 }
753
754 /* From a given parameter token, generate the corresponding GLSL string.
755 * Also, return the actual register name and swizzle in case the
756 * caller needs this information as well. */
757 static void shader_glsl_add_src_param(SHADER_OPCODE_ARG* arg, const DWORD param,
758 const DWORD addr_token, DWORD mask, char *reg_name, char *swizzle, char *out_str) {
759 BOOL is_color = FALSE;
760 swizzle[0] = reg_name[0] = out_str[0] = 0;
761
762 shader_glsl_get_register_name(param, addr_token, reg_name, &is_color, arg);
763
764 shader_glsl_get_swizzle(param, is_color, mask, swizzle);
765 shader_glsl_gen_modifier(param, reg_name, swizzle, out_str);
766 }
767
768 /* From a given parameter token, generate the corresponding GLSL string.
769 * Also, return the actual register name and swizzle in case the
770 * caller needs this information as well. */
771 static DWORD shader_glsl_add_dst_param(SHADER_OPCODE_ARG* arg, const DWORD param,
772 const DWORD addr_token, glsl_dst_param_t *dst_param) {
773 BOOL is_color = FALSE;
774
775 dst_param->mask_str[0] = '\0';
776 dst_param->reg_name[0] = '\0';
777
778 shader_glsl_get_register_name(param, addr_token, dst_param->reg_name, &is_color, arg);
779 return shader_glsl_get_write_mask(param, dst_param->mask_str);
780 }
781
782 /* Append the destination part of the instruction to the buffer, return the effective write mask */
783 static DWORD shader_glsl_append_dst(SHADER_BUFFER *buffer, SHADER_OPCODE_ARG *arg) {
784 glsl_dst_param_t dst_param;
785 DWORD mask;
786 int shift;
787
788 shift = (arg->dst & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
789 mask = shader_glsl_add_dst_param(arg, arg->dst, arg->dst_addr, &dst_param);
790 shader_addline(buffer, "%s%s = %s(", dst_param.reg_name, dst_param.mask_str, shift_glsl_tab[shift]);
791
792 return mask;
793 }
794
795 /** Process GLSL instruction modifiers */
796 void shader_glsl_add_instruction_modifiers(SHADER_OPCODE_ARG* arg) {
797
798 DWORD mask = arg->dst & WINED3DSP_DSTMOD_MASK;
799
800 if (arg->opcode->dst_token && mask != 0) {
801 glsl_dst_param_t dst_param;
802
803 shader_glsl_add_dst_param(arg, arg->dst, 0, &dst_param);
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_param.reg_name,
808 dst_param.mask_str, dst_param.reg_name, dst_param.mask_str);
809 }
810 if (mask & WINED3DSPDM_MSAMPCENTROID) {
811 FIXME("_centroid modifier not handled\n");
812 }
813 if (mask & WINED3DSPDM_PARTIALPRECISION) {
814 /* MSDN says this modifier can be safely ignored, so that's what we'll do. */
815 }
816 }
817 }
818
819 static inline const char* shader_get_comp_op(
820 const DWORD opcode) {
821
822 DWORD op = (opcode & INST_CONTROLS_MASK) >> INST_CONTROLS_SHIFT;
823 switch (op) {
824 case COMPARISON_GT: return ">";
825 case COMPARISON_EQ: return "==";
826 case COMPARISON_GE: return ">=";
827 case COMPARISON_LT: return "<";
828 case COMPARISON_NE: return "!=";
829 case COMPARISON_LE: return "<=";
830 default:
831 FIXME("Unrecognized comparison value: %u\n", op);
832 return "(\?\?)";
833 }
834 }
835
836 static void shader_glsl_get_sample_function(DWORD sampler_type, BOOL projected, glsl_sample_function_t *sample_function) {
837 /* Note that there's no such thing as a projected cube texture. */
838 switch(sampler_type) {
839 case WINED3DSTT_1D:
840 sample_function->name = projected ? "texture1DProj" : "texture1D";
841 sample_function->coord_mask = WINED3DSP_WRITEMASK_0;
842 break;
843 case WINED3DSTT_2D:
844 sample_function->name = projected ? "texture2DProj" : "texture2D";
845 sample_function->coord_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1;
846 break;
847 case WINED3DSTT_CUBE:
848 sample_function->name = "textureCube";
849 sample_function->coord_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
850 break;
851 case WINED3DSTT_VOLUME:
852 sample_function->name = projected ? "texture3DProj" : "texture3D";
853 sample_function->coord_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
854 break;
855 default:
856 sample_function->name = "";
857 FIXME("Unrecognized sampler type: %#x;\n", sampler_type);
858 break;
859 }
860 }
861
862
863 /*****************************************************************************
864 *
865 * Begin processing individual instruction opcodes
866 *
867 ****************************************************************************/
868
869 /* Generate GLSL arithmetic functions (dst = src1 + src2) */
870 void shader_glsl_arith(SHADER_OPCODE_ARG* arg) {
871 CONST SHADER_OPCODE* curOpcode = arg->opcode;
872 SHADER_BUFFER* buffer = arg->buffer;
873 char src0_reg[50], src1_reg[50];
874 char src0_mask[6], src1_mask[6];
875 char src0_str[100], src1_str[100];
876 DWORD write_mask;
877 char op;
878
879 /* Determine the GLSL operator to use based on the opcode */
880 switch (curOpcode->opcode) {
881 case WINED3DSIO_MUL: op = '*'; break;
882 case WINED3DSIO_ADD: op = '+'; break;
883 case WINED3DSIO_SUB: op = '-'; break;
884 default:
885 op = ' ';
886 FIXME("Opcode %s not yet handled in GLSL\n", curOpcode->name);
887 break;
888 }
889
890 write_mask = shader_glsl_append_dst(buffer, arg);
891 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, src0_reg, src0_mask, src0_str);
892 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, src1_reg, src1_mask, src1_str);
893 shader_addline(buffer, "%s %c %s);\n", src0_str, op, src1_str);
894 }
895
896 /* Process the WINED3DSIO_MOV opcode using GLSL (dst = src) */
897 void shader_glsl_mov(SHADER_OPCODE_ARG* arg) {
898 IWineD3DBaseShaderImpl* shader = (IWineD3DBaseShaderImpl*) arg->shader;
899 SHADER_BUFFER* buffer = arg->buffer;
900 char src0_str[100];
901 char src0_reg[50];
902 char src0_mask[6];
903 DWORD write_mask;
904
905 write_mask = shader_glsl_append_dst(buffer, arg);
906 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, src0_reg, src0_mask, src0_str);
907
908 /* In vs_1_1 WINED3DSIO_MOV can write to the adress register. In later
909 * shader versions WINED3DSIO_MOVA is used for this. */
910 if ((WINED3DSHADER_VERSION_MAJOR(shader->baseShader.hex_version) == 1 &&
911 !shader_is_pshader_version(shader->baseShader.hex_version) &&
912 shader_get_regtype(arg->dst) == WINED3DSPR_ADDR) ||
913 arg->opcode->opcode == WINED3DSIO_MOVA) {
914 /* We need to *round* to the nearest int here. */
915 size_t mask_size = shader_glsl_get_write_mask_size(write_mask);
916 if (mask_size > 1) {
917 shader_addline(buffer, "ivec%d(floor(%s + vec%d(0.5))));\n", mask_size, src0_str, mask_size);
918 } else {
919 shader_addline(buffer, "int(floor(%s + 0.5)));\n", src0_str);
920 }
921 } else {
922 shader_addline(buffer, "%s);\n", src0_str);
923 }
924 }
925
926 /* Process the dot product operators DP3 and DP4 in GLSL (dst = dot(src0, src1)) */
927 void shader_glsl_dot(SHADER_OPCODE_ARG* arg) {
928 CONST SHADER_OPCODE* curOpcode = arg->opcode;
929 SHADER_BUFFER* buffer = arg->buffer;
930 char src0_str[100], src1_str[100];
931 char src0_reg[50], src1_reg[50];
932 char src0_mask[6], src1_mask[6];
933 DWORD dst_write_mask, src_write_mask;
934 size_t dst_size = 0;
935
936 dst_write_mask = shader_glsl_append_dst(buffer, arg);
937 dst_size = shader_glsl_get_write_mask_size(dst_write_mask);
938
939 /* dp3 works on vec3, dp4 on vec4 */
940 if (curOpcode->opcode == WINED3DSIO_DP4) {
941 src_write_mask = WINED3DSP_WRITEMASK_ALL;
942 } else {
943 src_write_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
944 }
945
946 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_write_mask, src0_reg, src0_mask, src0_str);
947 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], src_write_mask, src1_reg, src1_mask, src1_str);
948
949 if (dst_size > 1) {
950 shader_addline(buffer, "vec%d(dot(%s, %s)));\n", dst_size, src0_str, src1_str);
951 } else {
952 shader_addline(buffer, "dot(%s, %s));\n", src0_str, src1_str);
953 }
954 }
955
956 /* Note that this instruction has some restrictions. The destination write mask
957 * can't contain the w component, and the source swizzles have to be .xyzw */
958 void shader_glsl_cross(SHADER_OPCODE_ARG *arg) {
959 char src0_reg[50], src0_mask[6], src0_str[100];
960 char src1_reg[50], src1_mask[6], src1_str[100];
961 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
962 char dst_mask[6];
963
964 shader_glsl_get_write_mask(arg->dst, dst_mask);
965 shader_glsl_append_dst(arg->buffer, arg);
966 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, src0_reg, src0_mask, src0_str);
967 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], src_mask, src1_reg, src1_mask, src1_str);
968 shader_addline(arg->buffer, "cross(%s, %s).%s);\n", src0_str, src1_str, dst_mask);
969 }
970
971 /* Map the opcode 1-to-1 to the GL code (arg->dst = instruction(src0, src1, ...) */
972 void shader_glsl_map2gl(SHADER_OPCODE_ARG* arg) {
973 CONST SHADER_OPCODE* curOpcode = arg->opcode;
974 SHADER_BUFFER* buffer = arg->buffer;
975 const char *instruction;
976 char arguments[256];
977 char src_str[100];
978 char src_reg[50];
979 char src_mask[6];
980 DWORD write_mask;
981 unsigned i;
982
983 /* Determine the GLSL function to use based on the opcode */
984 /* TODO: Possibly make this a table for faster lookups */
985 switch (curOpcode->opcode) {
986 case WINED3DSIO_MIN: instruction = "min"; break;
987 case WINED3DSIO_MAX: instruction = "max"; break;
988 case WINED3DSIO_RSQ: instruction = "inversesqrt"; break;
989 case WINED3DSIO_ABS: instruction = "abs"; break;
990 case WINED3DSIO_FRC: instruction = "fract"; break;
991 case WINED3DSIO_POW: instruction = "pow"; break;
992 case WINED3DSIO_NRM: instruction = "normalize"; break;
993 case WINED3DSIO_LOGP:
994 case WINED3DSIO_LOG: instruction = "log2"; break;
995 case WINED3DSIO_EXP: instruction = "exp2"; break;
996 case WINED3DSIO_SGN: instruction = "sign"; break;
997 default: instruction = "";
998 FIXME("Opcode %s not yet handled in GLSL\n", curOpcode->name);
999 break;
1000 }
1001
1002 write_mask = shader_glsl_append_dst(buffer, arg);
1003
1004 arguments[0] = '\0';
1005 if (curOpcode->num_params > 0) {
1006 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, src_reg, src_mask, src_str);
1007 strcat(arguments, src_str);
1008 for (i = 2; i < curOpcode->num_params; ++i) {
1009 strcat(arguments, ", ");
1010 shader_glsl_add_src_param(arg, arg->src[i-1], arg->src_addr[i-1], write_mask, src_reg, src_mask, src_str);
1011 strcat(arguments, src_str);
1012 }
1013 }
1014
1015 shader_addline(buffer, "%s(%s));\n", instruction, arguments);
1016 }
1017
1018 /** Process the WINED3DSIO_EXPP instruction in GLSL:
1019 * For shader model 1.x, do the following (and honor the writemask, so use a temporary variable):
1020 * dst.x = 2^(floor(src))
1021 * dst.y = src - floor(src)
1022 * dst.z = 2^src (partial precision is allowed, but optional)
1023 * dst.w = 1.0;
1024 * For 2.0 shaders, just do this (honoring writemask and swizzle):
1025 * dst = 2^src; (partial precision is allowed, but optional)
1026 */
1027 void shader_glsl_expp(SHADER_OPCODE_ARG* arg) {
1028 IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)arg->shader;
1029 char src_str[100];
1030 char src_reg[50];
1031 char src_mask[6];
1032
1033 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, src_reg, src_mask, src_str);
1034
1035 if (shader->baseShader.hex_version < WINED3DPS_VERSION(2,0)) {
1036 char dst_mask[6];
1037
1038 shader_addline(arg->buffer, "tmp0.x = exp2(floor(%s));\n", src_str);
1039 shader_addline(arg->buffer, "tmp0.y = %s - floor(%s);\n", src_str, src_str);
1040 shader_addline(arg->buffer, "tmp0.z = exp2(%s);\n", src_str);
1041 shader_addline(arg->buffer, "tmp0.w = 1.0;\n");
1042
1043 shader_glsl_append_dst(arg->buffer, arg);
1044 shader_glsl_get_write_mask(arg->dst, dst_mask);
1045 shader_addline(arg->buffer, "tmp0%s);\n", dst_mask);
1046 } else {
1047 DWORD write_mask;
1048 size_t mask_size;
1049
1050 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1051 mask_size = shader_glsl_get_write_mask_size(write_mask);
1052
1053 if (mask_size > 1) {
1054 shader_addline(arg->buffer, "vec%d(exp2(%s)));\n", mask_size, src_str);
1055 } else {
1056 shader_addline(arg->buffer, "exp2(%s));\n", src_str);
1057 }
1058 }
1059 }
1060
1061 /** Process the RCP (reciprocal or inverse) opcode in GLSL (dst = 1 / src) */
1062 void shader_glsl_rcp(SHADER_OPCODE_ARG* arg) {
1063 char src_str[100];
1064 char src_reg[50];
1065 char src_mask[6];
1066 DWORD write_mask;
1067 size_t mask_size;
1068
1069 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1070 mask_size = shader_glsl_get_write_mask_size(write_mask);
1071 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, src_reg, src_mask, src_str);
1072
1073 if (mask_size > 1) {
1074 shader_addline(arg->buffer, "vec%d(1.0 / %s));\n", mask_size, src_str);
1075 } else {
1076 shader_addline(arg->buffer, "1.0 / %s);\n", src_str);
1077 }
1078 }
1079
1080 /** Process signed comparison opcodes in GLSL. */
1081 void shader_glsl_compare(SHADER_OPCODE_ARG* arg) {
1082 char src0_str[100], src1_str[100];
1083 char src0_reg[50], src1_reg[50];
1084 char src0_mask[6], src1_mask[6];
1085 DWORD write_mask;
1086 size_t mask_size;
1087
1088 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1089 mask_size = shader_glsl_get_write_mask_size(write_mask);
1090 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, src0_reg, src0_mask, src0_str);
1091 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, src1_reg, src1_mask, src1_str);
1092
1093 if (mask_size > 1) {
1094 const char *compare;
1095
1096 switch(arg->opcode->opcode) {
1097 case WINED3DSIO_SLT: compare = "lessThan"; break;
1098 case WINED3DSIO_SGE: compare = "greaterThanEqual"; break;
1099 default: compare = "";
1100 FIXME("Can't handle opcode %s\n", arg->opcode->name);
1101 }
1102
1103 shader_addline(arg->buffer, "vec%d(%s(%s, %s)));\n", mask_size, compare, src0_str, src1_str);
1104 } else {
1105 const char *compare;
1106
1107 switch(arg->opcode->opcode) {
1108 case WINED3DSIO_SLT: compare = "<"; break;
1109 case WINED3DSIO_SGE: compare = ">="; break;
1110 default: compare = "";
1111 FIXME("Can't handle opcode %s\n", arg->opcode->name);
1112 }
1113
1114 shader_addline(arg->buffer, "(%s %s %s) ? 1.0 : 0.0);\n", src0_str, compare, src1_str);
1115 }
1116 }
1117
1118 /** Process CMP instruction in GLSL (dst = src0 >= 0.0 ? src1 : src2), per channel */
1119 void shader_glsl_cmp(SHADER_OPCODE_ARG* arg) {
1120 char src0_str[100], src1_str[100], src2_str[100];
1121 char src0_reg[50], src1_reg[50], src2_reg[50];
1122 char src0_mask[6], src1_mask[6], src2_mask[6];
1123 DWORD write_mask;
1124 size_t mask_size;
1125
1126 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1127 mask_size = shader_glsl_get_write_mask_size(write_mask);
1128
1129 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, src0_reg, src0_mask, src0_str);
1130 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, src1_reg, src1_mask, src1_str);
1131 shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], write_mask, src2_reg, src2_mask, src2_str);
1132
1133 if (mask_size > 1) {
1134 shader_addline(arg->buffer, "mix(%s, %s, vec%d(lessThan(%s, vec%d(0.0)))));\n", src1_str, src2_str, mask_size, src0_str, mask_size);
1135 } else {
1136 shader_addline(arg->buffer, "%s >= 0.0 ? %s : %s);\n", src0_str, src1_str, src2_str);
1137 }
1138 }
1139
1140 /** Process the CND opcode in GLSL (dst = (src0 < 0.5) ? src1 : src2) */
1141 /* For ps 1.1-1.3, only a single component of src0 is used. For ps 1.4
1142 * the compare is done per component of src0. */
1143 void shader_glsl_cnd(SHADER_OPCODE_ARG* arg) {
1144 IWineD3DBaseShaderImpl* shader = (IWineD3DBaseShaderImpl*) arg->shader;
1145 char src0_str[100], src1_str[100], src2_str[100];
1146 char src0_reg[50], src1_reg[50], src2_reg[50];
1147 char src0_mask[6], src1_mask[6], src2_mask[6];
1148 DWORD write_mask;
1149 size_t mask_size;
1150
1151 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1152
1153 if (shader->baseShader.hex_version < WINED3DPS_VERSION(1, 4)) {
1154 mask_size = 1;
1155 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, src0_reg, src0_mask, src0_str);
1156 } else {
1157 mask_size = shader_glsl_get_write_mask_size(write_mask);
1158 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, src0_reg, src0_mask, src0_str);
1159 }
1160
1161 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, src1_reg, src1_mask, src1_str);
1162 shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], write_mask, src2_reg, src2_mask, src2_str);
1163
1164 if (mask_size > 1) {
1165 shader_addline(arg->buffer, "mix(%s, %s, vec%d(lessThan(%s, vec%d(0.5)))));\n", src2_str, src1_str, mask_size, src0_str, mask_size);
1166 } else {
1167 shader_addline(arg->buffer, "%s < 0.5 ? %s : %s);\n", src0_str, src1_str, src2_str);
1168 }
1169 }
1170
1171 /** GLSL code generation for WINED3DSIO_MAD: Multiply the first 2 opcodes, then add the last */
1172 void shader_glsl_mad(SHADER_OPCODE_ARG* arg) {
1173 char src0_str[100], src1_str[100], src2_str[100];
1174 char src0_reg[50], src1_reg[50], src2_reg[50];
1175 char src0_mask[6], src1_mask[6], src2_mask[6];
1176 DWORD write_mask;
1177
1178 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1179 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, src0_reg, src0_mask, src0_str);
1180 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, src1_reg, src1_mask, src1_str);
1181 shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], write_mask, src2_reg, src2_mask, src2_str);
1182 shader_addline(arg->buffer, "(%s * %s) + %s);\n", src0_str, src1_str, src2_str);
1183 }
1184
1185 /** Handles transforming all WINED3DSIO_M?x? opcodes for
1186 Vertex shaders to GLSL codes */
1187 void shader_glsl_mnxn(SHADER_OPCODE_ARG* arg) {
1188 int i;
1189 int nComponents = 0;
1190 SHADER_OPCODE_ARG tmpArg;
1191
1192 memset(&tmpArg, 0, sizeof(SHADER_OPCODE_ARG));
1193
1194 /* Set constants for the temporary argument */
1195 tmpArg.shader = arg->shader;
1196 tmpArg.buffer = arg->buffer;
1197 tmpArg.src[0] = arg->src[0];
1198 tmpArg.src_addr[0] = arg->src_addr[0];
1199 tmpArg.src_addr[1] = arg->src_addr[1];
1200 tmpArg.reg_maps = arg->reg_maps;
1201
1202 switch(arg->opcode->opcode) {
1203 case WINED3DSIO_M4x4:
1204 nComponents = 4;
1205 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP4);
1206 break;
1207 case WINED3DSIO_M4x3:
1208 nComponents = 3;
1209 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP4);
1210 break;
1211 case WINED3DSIO_M3x4:
1212 nComponents = 4;
1213 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP3);
1214 break;
1215 case WINED3DSIO_M3x3:
1216 nComponents = 3;
1217 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP3);
1218 break;
1219 case WINED3DSIO_M3x2:
1220 nComponents = 2;
1221 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP3);
1222 break;
1223 default:
1224 break;
1225 }
1226
1227 for (i = 0; i < nComponents; i++) {
1228 tmpArg.dst = ((arg->dst) & ~WINED3DSP_WRITEMASK_ALL)|(WINED3DSP_WRITEMASK_0<<i);
1229 tmpArg.src[1] = arg->src[1]+i;
1230 shader_glsl_dot(&tmpArg);
1231 }
1232 }
1233
1234 /**
1235 The LRP instruction performs a component-wise linear interpolation
1236 between the second and third operands using the first operand as the
1237 blend factor. Equation: (dst = src2 + src0 * (src1 - src2))
1238 This is equivalent to mix(src2, src1, src0);
1239 */
1240 void shader_glsl_lrp(SHADER_OPCODE_ARG* arg) {
1241 char src0_str[100], src1_str[100], src2_str[100];
1242 char src0_reg[50], src1_reg[50], src2_reg[50];
1243 char src0_mask[6], src1_mask[6], src2_mask[6];
1244 DWORD write_mask;
1245
1246 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1247
1248 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, src0_reg, src0_mask, src0_str);
1249 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, src1_reg, src1_mask, src1_str);
1250 shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], write_mask, src2_reg, src2_mask, src2_str);
1251
1252 shader_addline(arg->buffer, "mix(%s, %s, %s));\n", src2_str, src1_str, src0_str);
1253 }
1254
1255 /** Process the WINED3DSIO_LIT instruction in GLSL:
1256 * dst.x = dst.w = 1.0
1257 * dst.y = (src0.x > 0) ? src0.x
1258 * dst.z = (src0.x > 0) ? ((src0.y > 0) ? pow(src0.y, src.w) : 0) : 0
1259 * where src.w is clamped at +- 128
1260 */
1261 void shader_glsl_lit(SHADER_OPCODE_ARG* arg) {
1262 char src0_str[100];
1263 char src0_reg[50];
1264 char src0_mask[6];
1265 char src1_str[100];
1266 char src1_reg[50];
1267 char src1_mask[6];
1268 char src3_str[100];
1269 char src3_reg[50];
1270 char src3_mask[6];
1271 char dst_mask[6];
1272
1273 shader_glsl_append_dst(arg->buffer, arg);
1274 shader_glsl_get_write_mask(arg->dst, dst_mask);
1275
1276 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, src0_reg, src0_mask, src0_str);
1277 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_1, src1_reg, src1_mask, src1_str);
1278 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_3, src3_reg, src3_mask, src3_str);
1279
1280 shader_addline(arg->buffer, "vec4(1.0, (%s > 0.0 ? %s : 0.0), (%s > 0.0 ? ((%s > 0.0) ? pow(%s, clamp(%s, -128.0, 128.0)) : 0.0) : 0.0), 1.0)%s);\n",
1281 src0_str, src0_str, src0_str, src1_str, src1_str, src3_str, dst_mask);
1282 }
1283
1284 /** Process the WINED3DSIO_DST instruction in GLSL:
1285 * dst.x = 1.0
1286 * dst.y = src0.x * src0.y
1287 * dst.z = src0.z
1288 * dst.w = src1.w
1289 */
1290 void shader_glsl_dst(SHADER_OPCODE_ARG* arg) {
1291 char src0y_str[100], src0z_str[100], src1y_str[100], src1w_str[100];
1292 char src0y_reg[50], src0z_reg[50], src1y_reg[50], src1w_reg[50];
1293 char dst_mask[6], src0y_mask[6], src0z_mask[6], src1y_mask[6], src1w_mask[6];
1294
1295 shader_glsl_append_dst(arg->buffer, arg);
1296 shader_glsl_get_write_mask(arg->dst, dst_mask);
1297
1298 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_1, src0y_reg, src0y_mask, src0y_str);
1299 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_2, src0z_reg, src0z_mask, src0z_str);
1300 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_1, src1y_reg, src1y_mask, src1y_str);
1301 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_3, src1w_reg, src1w_mask, src1w_str);
1302
1303 shader_addline(arg->buffer, "vec4(1.0, %s * %s, %s, %s))%s;\n",
1304 src0y_str, src1y_str, src0z_str, src1w_str, dst_mask);
1305 }
1306
1307 /** Process the WINED3DSIO_SINCOS instruction in GLSL:
1308 * VS 2.0 requires that specific cosine and sine constants be passed to this instruction so the hardware
1309 * can handle it. But, these functions are built-in for GLSL, so we can just ignore the last 2 params.
1310 *
1311 * dst.x = cos(src0.?)
1312 * dst.y = sin(src0.?)
1313 * dst.z = dst.z
1314 * dst.w = dst.w
1315 */
1316 void shader_glsl_sincos(SHADER_OPCODE_ARG* arg) {
1317 char src0_str[100];
1318 char src0_reg[50];
1319 char src0_mask[6];
1320 DWORD write_mask;
1321
1322 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1323 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, src0_reg, src0_mask, src0_str);
1324
1325 switch (write_mask) {
1326 case WINED3DSP_WRITEMASK_0:
1327 shader_addline(arg->buffer, "cos(%s));\n", src0_str);
1328 break;
1329
1330 case WINED3DSP_WRITEMASK_1:
1331 shader_addline(arg->buffer, "sin(%s));\n", src0_str);
1332 break;
1333
1334 case (WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1):
1335 shader_addline(arg->buffer, "vec2(cos(%s), sin(%s)));\n", src0_str, src0_str);
1336 break;
1337
1338 default:
1339 ERR("Write mask should be .x, .y or .xy\n");
1340 break;
1341 }
1342 }
1343
1344 /** Process the WINED3DSIO_LOOP instruction in GLSL:
1345 * Start a for() loop where src1.y is the initial value of aL,
1346 * increment aL by src1.z for a total of src1.x iterations.
1347 * Need to use a temporary variable for this operation.
1348 */
1349 /* FIXME: I don't think nested loops will work correctly this way. */
1350 void shader_glsl_loop(SHADER_OPCODE_ARG* arg) {
1351
1352 char src1_str[100];
1353 char src1_reg[50];
1354 char src1_mask[6];
1355
1356 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_ALL, src1_reg, src1_mask, src1_str);
1357
1358 shader_addline(arg->buffer, "for (tmpInt = 0, aL = %s.y; tmpInt < %s.x; tmpInt++, aL += %s.z) {\n",
1359 src1_reg, src1_reg, src1_reg);
1360 }
1361
1362 void shader_glsl_end(SHADER_OPCODE_ARG* arg) {
1363 shader_addline(arg->buffer, "}\n");
1364 }
1365
1366 void shader_glsl_rep(SHADER_OPCODE_ARG* arg) {
1367
1368 char src0_str[100];
1369 char src0_reg[50];
1370 char src0_mask[6];
1371
1372 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, src0_reg, src0_mask, src0_str);
1373 shader_addline(arg->buffer, "for (tmpInt = 0; tmpInt < %s; tmpInt++) {\n", src0_str);
1374 }
1375
1376 void shader_glsl_if(SHADER_OPCODE_ARG* arg) {
1377
1378 char src0_str[100];
1379 char src0_reg[50];
1380 char src0_mask[6];
1381
1382 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, src0_reg, src0_mask, src0_str);
1383 shader_addline(arg->buffer, "if (%s) {\n", src0_str);
1384 }
1385
1386 void shader_glsl_ifc(SHADER_OPCODE_ARG* arg) {
1387
1388 char src0_str[100], src1_str[100];
1389 char src0_reg[50], src1_reg[50];
1390 char src0_mask[6], src1_mask[6];
1391
1392 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, src0_reg, src0_mask, src0_str);
1393 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_0, src1_reg, src1_mask, src1_str);
1394
1395 shader_addline(arg->buffer, "if (%s %s %s) {\n",
1396 src0_str, shader_get_comp_op(arg->opcode_token), src1_str);
1397 }
1398
1399 void shader_glsl_else(SHADER_OPCODE_ARG* arg) {
1400 shader_addline(arg->buffer, "} else {\n");
1401 }
1402
1403 void shader_glsl_break(SHADER_OPCODE_ARG* arg) {
1404 shader_addline(arg->buffer, "break;\n");
1405 }
1406
1407 /* FIXME: According to MSDN the compare is done per component. */
1408 void shader_glsl_breakc(SHADER_OPCODE_ARG* arg) {
1409
1410 char src0_str[100], src1_str[100];
1411 char src0_reg[50], src1_reg[50];
1412 char src0_mask[6], src1_mask[6];
1413
1414 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, src0_reg, src0_mask, src0_str);
1415 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_0, src1_reg, src1_mask, src1_str);
1416
1417 shader_addline(arg->buffer, "if (%s %s %s) break;\n",
1418 src0_str, shader_get_comp_op(arg->opcode_token), src1_str);
1419 }
1420
1421 void shader_glsl_label(SHADER_OPCODE_ARG* arg) {
1422
1423 DWORD snum = (arg->src[0]) & WINED3DSP_REGNUM_MASK;
1424 shader_addline(arg->buffer, "}\n");
1425 shader_addline(arg->buffer, "void subroutine%lu () {\n", snum);
1426 }
1427
1428 void shader_glsl_call(SHADER_OPCODE_ARG* arg) {
1429 DWORD snum = (arg->src[0]) & WINED3DSP_REGNUM_MASK;
1430 shader_addline(arg->buffer, "subroutine%lu();\n", snum);
1431 }
1432
1433 void shader_glsl_callnz(SHADER_OPCODE_ARG* arg) {
1434
1435 char src1_str[100];
1436 char src1_reg[50];
1437 char src1_mask[6];
1438
1439 DWORD snum = (arg->src[0]) & WINED3DSP_REGNUM_MASK;
1440 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_0, src1_reg, src1_mask, src1_str);
1441 shader_addline(arg->buffer, "if (%s) subroutine%lu();\n", src1_str, snum);
1442 }
1443
1444 /*********************************************
1445 * Pixel Shader Specific Code begins here
1446 ********************************************/
1447 void pshader_glsl_tex(SHADER_OPCODE_ARG* arg) {
1448 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1449 DWORD hex_version = This->baseShader.hex_version;
1450 char coord_str[100], coord_reg[50], coord_mask[6];
1451 char dst_swizzle[6];
1452 glsl_sample_function_t sample_function;
1453 DWORD sampler_type;
1454 DWORD sampler_idx;
1455 BOOL projected;
1456 DWORD mask = 0;
1457
1458 /* All versions have a destination register */
1459 shader_glsl_append_dst(arg->buffer, arg);
1460
1461 /* 1.0-1.4: Use destination register as sampler source.
1462 * 2.0+: Use provided sampler source. */
1463 if (hex_version < WINED3DPS_VERSION(1,4)) {
1464 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1465 DWORD flags;
1466
1467 sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1468 flags = deviceImpl->stateBlock->textureState[sampler_idx][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1469
1470 if (flags & WINED3DTTFF_PROJECTED) {
1471 projected = TRUE;
1472 switch (flags & ~WINED3DTTFF_PROJECTED) {
1473 case WINED3DTTFF_COUNT1: FIXME("WINED3DTTFF_PROJECTED with WINED3DTTFF_COUNT1?\n"); break;
1474 case WINED3DTTFF_COUNT2: mask = WINED3DSP_WRITEMASK_1; break;
1475 case WINED3DTTFF_COUNT3: mask = WINED3DSP_WRITEMASK_2; break;
1476 case WINED3DTTFF_COUNT4:
1477 case WINED3DTTFF_DISABLE: mask = WINED3DSP_WRITEMASK_3; break;
1478 }
1479 } else {
1480 projected = FALSE;
1481 }
1482 } else if (hex_version < WINED3DPS_VERSION(2,0)) {
1483 DWORD src_mod = arg->src[0] & WINED3DSP_SRCMOD_MASK;
1484 sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1485
1486 if (src_mod == WINED3DSPSM_DZ) {
1487 projected = TRUE;
1488 mask = WINED3DSP_WRITEMASK_2;
1489 } else if (src_mod == WINED3DSPSM_DW) {
1490 projected = TRUE;
1491 mask = WINED3DSP_WRITEMASK_3;
1492 } else {
1493 projected = FALSE;
1494 }
1495 } else {
1496 sampler_idx = arg->src[1] & WINED3DSP_REGNUM_MASK;
1497 /* TODO: Handle D3DSI_TEXLD_PROJECTED... */
1498 projected = FALSE;
1499 }
1500
1501 sampler_type = arg->reg_maps->samplers[sampler_idx] & WINED3DSP_TEXTURETYPE_MASK;
1502 shader_glsl_get_sample_function(sampler_type, projected, &sample_function);
1503 mask |= sample_function.coord_mask;
1504
1505 if (hex_version < WINED3DPS_VERSION(2,0)) {
1506 shader_glsl_get_write_mask(arg->dst, dst_swizzle);
1507 } else {
1508 shader_glsl_get_swizzle(arg->src[1], FALSE, arg->dst, dst_swizzle);
1509 }
1510
1511 /* 1.0-1.3: Use destination register as coordinate source.
1512 1.4+: Use provided coordinate source register. */
1513 if (hex_version < WINED3DPS_VERSION(1,4)) {
1514 shader_glsl_get_write_mask(mask, coord_mask);
1515 shader_addline(arg->buffer, "%s(Psampler%u, T%u%s)%s);\n",
1516 sample_function.name, sampler_idx, sampler_idx, coord_mask, dst_swizzle);
1517 } else {
1518 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], mask, coord_reg, coord_mask, coord_str);
1519 shader_addline(arg->buffer, "%s(Psampler%u, %s)%s);\n",
1520 sample_function.name, sampler_idx, coord_str, dst_swizzle);
1521 }
1522 }
1523
1524 void pshader_glsl_texcoord(SHADER_OPCODE_ARG* arg) {
1525
1526 /* FIXME: Make this work for more than just 2D textures */
1527
1528 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1529 SHADER_BUFFER* buffer = arg->buffer;
1530 DWORD hex_version = This->baseShader.hex_version;
1531 DWORD write_mask;
1532 char dst_mask[6];
1533
1534 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1535 shader_glsl_get_write_mask(write_mask, dst_mask);
1536
1537 if (hex_version != WINED3DPS_VERSION(1,4)) {
1538 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1539 shader_addline(buffer, "clamp(gl_TexCoord[%u], 0.0, 1.0)%s);\n", reg, dst_mask);
1540 } else {
1541 DWORD reg = arg->src[0] & WINED3DSP_REGNUM_MASK;
1542 DWORD src_mod = arg->src[0] & WINED3DSP_SRCMOD_MASK;
1543 char dst_swizzle[6];
1544
1545 shader_glsl_get_swizzle(arg->src[0], FALSE, write_mask, dst_swizzle);
1546
1547 if (src_mod == WINED3DSPSM_DZ) {
1548 char div_str[100], div_reg[50], div_mask[6];
1549 size_t mask_size = shader_glsl_get_write_mask_size(write_mask);
1550 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_2, div_reg, div_mask, div_str);
1551
1552 if (mask_size > 1) {
1553 shader_addline(buffer, "gl_TexCoord[%u]%s / vec%d(%s));\n", reg, dst_swizzle, mask_size, div_str);
1554 } else {
1555 shader_addline(buffer, "gl_TexCoord[%u]%s / %s);\n", reg, dst_swizzle, div_str);
1556 }
1557 } else if (src_mod == WINED3DSPSM_DW) {
1558 char div_str[100], div_reg[50], div_mask[6];
1559 size_t mask_size = shader_glsl_get_write_mask_size(write_mask);
1560 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_3, div_reg, div_mask, div_str);
1561
1562 if (mask_size > 1) {
1563 shader_addline(buffer, "gl_TexCoord[%u]%s / vec%d(%s));\n", reg, dst_swizzle, mask_size, div_str);
1564 } else {
1565 shader_addline(buffer, "gl_TexCoord[%u]%s / %s);\n", reg, dst_swizzle, div_str);
1566 }
1567 } else {
1568 shader_addline(buffer, "gl_TexCoord[%u]%s);\n", reg, dst_swizzle);
1569 }
1570 }
1571 }
1572
1573 /** Process the WINED3DSIO_TEXDP3TEX instruction in GLSL:
1574 * Take a 3-component dot product of the TexCoord[dstreg] and src,
1575 * then perform a 1D texture lookup from stage dstregnum, place into dst. */
1576 void pshader_glsl_texdp3tex(SHADER_OPCODE_ARG* arg) {
1577 char src0_str[100];
1578 char src0_name[50];
1579 char src0_mask[6];
1580 char dst_mask[6];
1581 DWORD sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1582 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1583
1584 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, src0_name, src0_mask, src0_str);
1585
1586 shader_glsl_append_dst(arg->buffer, arg);
1587 shader_glsl_get_write_mask(arg->dst, dst_mask);
1588 shader_addline(arg->buffer, "texture1D(Psampler%u, dot(gl_TexCoord[%u].xyz, %s))%s);\n",
1589 sampler_idx, sampler_idx, src0_str, dst_mask);
1590 }
1591
1592 /** Process the WINED3DSIO_TEXDP3 instruction in GLSL:
1593 * Take a 3-component dot product of the TexCoord[dstreg] and src. */
1594 void pshader_glsl_texdp3(SHADER_OPCODE_ARG* arg) {
1595 char src0_str[100];
1596 char src0_name[50];
1597 char src0_mask[6];
1598 DWORD dstreg = arg->dst & WINED3DSP_REGNUM_MASK;
1599 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1600 DWORD dst_mask;
1601 size_t mask_size;
1602
1603 dst_mask = shader_glsl_append_dst(arg->buffer, arg);
1604 mask_size = shader_glsl_get_write_mask_size(dst_mask);
1605 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, src0_name, src0_mask, src0_str);
1606
1607 if (mask_size > 1) {
1608 shader_addline(arg->buffer, "vec%d(dot(T%u.xyz, %s)));\n", mask_size, dstreg, src0_str);
1609 } else {
1610 shader_addline(arg->buffer, "dot(T%u.xyz, %s));\n", dstreg, src0_str);
1611 }
1612 }
1613
1614 /** Process the WINED3DSIO_TEXDEPTH instruction in GLSL:
1615 * Calculate the depth as dst.x / dst.y */
1616 void pshader_glsl_texdepth(SHADER_OPCODE_ARG* arg) {
1617 glsl_dst_param_t dst_param;
1618
1619 shader_glsl_add_dst_param(arg, arg->dst, 0, &dst_param);
1620
1621 shader_addline(arg->buffer, "gl_FragDepth = %s.x / %s.y;\n", dst_param.reg_name, dst_param.reg_name);
1622 }
1623
1624 /** Process the WINED3DSIO_TEXM3X2DEPTH instruction in GLSL:
1625 * Last row of a 3x2 matrix multiply, use the result to calculate the depth:
1626 * Calculate tmp0.y = TexCoord[dstreg] . src.xyz; (tmp0.x has already been calculated)
1627 * depth = (tmp0.y == 0.0) ? 1.0 : tmp0.x / tmp0.y
1628 */
1629 void pshader_glsl_texm3x2depth(SHADER_OPCODE_ARG* arg) {
1630 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1631 DWORD dstreg = arg->dst & WINED3DSP_REGNUM_MASK;
1632 char src0_str[100], src0_name[50], src0_mask[6];
1633
1634 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, src0_name, src0_mask, src0_str);
1635
1636 shader_addline(arg->buffer, "tmp0.y = dot(T%u.xyz, %s);\n", dstreg, src0_str);
1637 shader_addline(arg->buffer, "gl_FragDepth = (tmp0.y == 0.0) ? 1.0 : clamp(tmp0.x / tmp0.y, 0.0, 1.0);\n");
1638 }
1639
1640 /** Process the WINED3DSIO_TEXM3X2PAD instruction in GLSL
1641 * Calculate the 1st of a 2-row matrix multiplication. */
1642 void pshader_glsl_texm3x2pad(SHADER_OPCODE_ARG* arg) {
1643 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1644 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1645 SHADER_BUFFER* buffer = arg->buffer;
1646 char src0_str[100];
1647 char src0_name[50];
1648 char src0_mask[6];
1649
1650 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, src0_name, src0_mask, src0_str);
1651 shader_addline(buffer, "tmp0.x = dot(T%u.xyz, %s);\n", reg, src0_str);
1652 }
1653
1654 /** Process the WINED3DSIO_TEXM3X3PAD instruction in GLSL
1655 * Calculate the 1st or 2nd row of a 3-row matrix multiplication. */
1656 void pshader_glsl_texm3x3pad(SHADER_OPCODE_ARG* arg) {
1657
1658 IWineD3DPixelShaderImpl* shader = (IWineD3DPixelShaderImpl*) arg->shader;
1659 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1660 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1661 SHADER_BUFFER* buffer = arg->buffer;
1662 SHADER_PARSE_STATE* current_state = &shader->baseShader.parse_state;
1663 char src0_str[100];
1664 char src0_name[50];
1665 char src0_mask[6];
1666
1667 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, src0_name, src0_mask, src0_str);
1668 shader_addline(buffer, "tmp0.%c = dot(T%u.xyz, %s);\n", 'x' + current_state->current_row, reg, src0_str);
1669 current_state->texcoord_w[current_state->current_row++] = reg;
1670 }
1671
1672 void pshader_glsl_texm3x2tex(SHADER_OPCODE_ARG* arg) {
1673 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1674 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1675 SHADER_BUFFER* buffer = arg->buffer;
1676 char src0_str[100];
1677 char src0_name[50];
1678 char src0_mask[6];
1679 char dst_mask[6];
1680
1681 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, src0_name, src0_mask, src0_str);
1682 shader_addline(buffer, "tmp0.y = dot(T%u.xyz, %s);\n", reg, src0_str);
1683
1684 shader_glsl_append_dst(buffer, arg);
1685 shader_glsl_get_write_mask(arg->dst, dst_mask);
1686
1687 /* Sample the texture using the calculated coordinates */
1688 shader_addline(buffer, "texture2D(Psampler%u, tmp0.xy)%s);\n", reg, dst_mask);
1689 }
1690
1691 /** Process the WINED3DSIO_TEXM3X3TEX instruction in GLSL
1692 * Perform the 3rd row of a 3x3 matrix multiply, then sample the texture using the calculated coordinates */
1693 void pshader_glsl_texm3x3tex(SHADER_OPCODE_ARG* arg) {
1694 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1695 char src0_str[100];
1696 char src0_name[50];
1697 char src0_mask[6];
1698 char dst_mask[6];
1699 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1700 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1701 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1702 DWORD sampler_type = arg->reg_maps->samplers[reg] & WINED3DSP_TEXTURETYPE_MASK;
1703 glsl_sample_function_t sample_function;
1704
1705 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, src0_name, src0_mask, src0_str);
1706 shader_addline(arg->buffer, "tmp0.z = dot(T%u.xyz, %s);\n", reg, src0_str);
1707
1708 shader_glsl_append_dst(arg->buffer, arg);
1709 shader_glsl_get_write_mask(arg->dst, dst_mask);
1710 shader_glsl_get_sample_function(sampler_type, FALSE, &sample_function);
1711
1712 /* Sample the texture using the calculated coordinates */
1713 shader_addline(arg->buffer, "%s(Psampler%u, tmp0.xyz)%s);\n", sample_function.name, reg, dst_mask);
1714
1715 current_state->current_row = 0;
1716 }
1717
1718 /** Process the WINED3DSIO_TEXM3X3 instruction in GLSL
1719 * Perform the 3rd row of a 3x3 matrix multiply */
1720 void pshader_glsl_texm3x3(SHADER_OPCODE_ARG* arg) {
1721 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1722 char src0_str[100];
1723 char src0_name[50];
1724 char src0_mask[6];
1725 char dst_mask[6];
1726 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1727 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1728 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1729
1730 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, src0_name, src0_mask, src0_str);
1731
1732 shader_glsl_append_dst(arg->buffer, arg);
1733 shader_glsl_get_write_mask(arg->dst, dst_mask);
1734 shader_addline(arg->buffer, "vec4(tmp.xy, dot(T%u.xyz, %s), 1.0)%s);\n", reg, src0_str, dst_mask);
1735
1736 current_state->current_row = 0;
1737 }
1738
1739 /** Process the WINED3DSIO_TEXM3X3SPEC instruction in GLSL
1740 * Peform the final texture lookup based on the previous 2 3x3 matrix multiplies */
1741 void pshader_glsl_texm3x3spec(SHADER_OPCODE_ARG* arg) {
1742
1743 IWineD3DPixelShaderImpl* shader = (IWineD3DPixelShaderImpl*) arg->shader;
1744 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1745 char src0_str[100], src0_name[50], src0_mask[6];
1746 char src1_str[100], src1_name[50], src1_mask[6];
1747 char dst_mask[6];
1748 SHADER_BUFFER* buffer = arg->buffer;
1749 SHADER_PARSE_STATE* current_state = &shader->baseShader.parse_state;
1750 DWORD stype = arg->reg_maps->samplers[reg] & WINED3DSP_TEXTURETYPE_MASK;
1751 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1752 glsl_sample_function_t sample_function;
1753
1754 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, src0_name, src0_mask, src0_str);
1755 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], src_mask, src1_name, src1_mask, src1_str);
1756
1757 /* Perform the last matrix multiply operation */
1758 shader_addline(buffer, "tmp0.z = dot(T%u.xyz, %s);\n", reg, src0_str);
1759
1760 /* Calculate reflection vector, 2*(tmp0.src1)*tmp0-src1
1761 * This is equavalent to reflect(-src1, tmp0); */
1762 shader_addline(buffer, "tmp0.xyz = reflect(-(%s), tmp0.xyz);\n", src1_str);
1763
1764 shader_glsl_append_dst(buffer, arg);
1765 shader_glsl_get_write_mask(arg->dst, dst_mask);
1766 shader_glsl_get_sample_function(stype, FALSE, &sample_function);
1767
1768 /* Sample the texture */
1769 shader_addline(buffer, "%s(Psampler%u, tmp0.xyz)%s);\n", sample_function.name, reg, dst_mask);
1770
1771 current_state->current_row = 0;
1772 }
1773
1774 /** Process the WINED3DSIO_TEXM3X3VSPEC instruction in GLSL
1775 * Peform the final texture lookup based on the previous 2 3x3 matrix multiplies */
1776 void pshader_glsl_texm3x3vspec(SHADER_OPCODE_ARG* arg) {
1777
1778 IWineD3DPixelShaderImpl* shader = (IWineD3DPixelShaderImpl*) arg->shader;
1779 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1780 SHADER_BUFFER* buffer = arg->buffer;
1781 SHADER_PARSE_STATE* current_state = &shader->baseShader.parse_state;
1782 char src0_str[100], src0_name[50], src0_mask[6], dst_mask[6];
1783 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1784 DWORD sampler_type = arg->reg_maps->samplers[reg] & WINED3DSP_TEXTURETYPE_MASK;
1785 glsl_sample_function_t sample_function;
1786
1787 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, src0_name, src0_mask, src0_str);
1788
1789 /* Perform the last matrix multiply operation */
1790 shader_addline(buffer, "tmp0.z = dot(vec3(T%u), vec3(%s));\n", reg, src0_str);
1791
1792 /* Construct the eye-ray vector from w coordinates */
1793 shader_addline(buffer, "tmp1.xyz = normalize(vec3(gl_TexCoord[%u].w, gl_TexCoord[%u].w, gl_TexCoord[%u].w));\n",
1794 current_state->texcoord_w[0], current_state->texcoord_w[1], reg);
1795
1796 /* Calculate reflection vector (Assume normal is normalized): RF = 2*(tmp0.tmp1)*tmp0-tmp1
1797 * This is equavalent to reflect(-tmp1, tmp0); */
1798 shader_addline(buffer, "tmp0.xyz = reflect(-tmp1.xyz, tmp0.xyz);\n");
1799
1800 shader_glsl_append_dst(buffer, arg);
1801 shader_glsl_get_write_mask(arg->dst, dst_mask);
1802 shader_glsl_get_sample_function(sampler_type, FALSE, &sample_function);
1803
1804 /* Sample the texture using the calculated coordinates */
1805 shader_addline(buffer, "%s(Psampler%u, tmp0.xyz)%s);\n", sample_function.name, reg, dst_mask);
1806
1807 current_state->current_row = 0;
1808 }
1809
1810 /** Process the WINED3DSIO_TEXBEM instruction in GLSL.
1811 * Apply a fake bump map transform.
1812 * FIXME: Should apply the BUMPMAPENV matrix. For now, just sample the texture */
1813 void pshader_glsl_texbem(SHADER_OPCODE_ARG* arg) {
1814
1815 DWORD reg1 = arg->dst & WINED3DSP_REGNUM_MASK;
1816 DWORD reg2 = arg->src[0] & WINED3DSP_REGNUM_MASK;
1817
1818 FIXME("Not applying the BUMPMAPENV matrix for pixel shader instruction texbem.\n");
1819 shader_addline(arg->buffer, "T%u = texture2D(Psampler%u, gl_TexCoord[%u].xy + T%u.xy);\n",
1820 reg1, reg1, reg1, reg2);
1821 }
1822
1823 /** Process the WINED3DSIO_TEXREG2AR instruction in GLSL
1824 * Sample 2D texture at dst using the alpha & red (wx) components of src as texture coordinates */
1825 void pshader_glsl_texreg2ar(SHADER_OPCODE_ARG* arg) {
1826 char src0_str[100];
1827 char src0_reg[50];
1828 char src0_mask[6];
1829 DWORD sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1830 char dst_mask[6];
1831
1832 shader_glsl_append_dst(arg->buffer, arg);
1833 shader_glsl_get_write_mask(arg->dst, dst_mask);
1834 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_ALL, src0_reg, src0_mask, src0_str);
1835
1836 shader_addline(arg->buffer, "texture2D(Psampler%u, %s.wx)%s);\n", sampler_idx, src0_reg, dst_mask);
1837 }
1838
1839 /** Process the WINED3DSIO_TEXREG2GB instruction in GLSL
1840 * Sample 2D texture at dst using the green & blue (yz) components of src as texture coordinates */
1841 void pshader_glsl_texreg2gb(SHADER_OPCODE_ARG* arg) {
1842 char src0_str[100];
1843 char src0_reg[50];
1844 char src0_mask[6];
1845 DWORD sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1846 char dst_mask[6];
1847
1848 shader_glsl_append_dst(arg->buffer, arg);
1849 shader_glsl_get_write_mask(arg->dst, dst_mask);
1850 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_ALL, src0_reg, src0_mask, src0_str);
1851
1852 shader_addline(arg->buffer, "texture2D(Psampler%u, %s.yz)%s);\n", sampler_idx, src0_reg, dst_mask);
1853 }
1854
1855 /** Process the WINED3DSIO_TEXREG2RGB instruction in GLSL
1856 * Sample texture at dst using the rgb (xyz) components of src as texture coordinates */
1857 void pshader_glsl_texreg2rgb(SHADER_OPCODE_ARG* arg) {
1858 char src0_str[100];
1859 char src0_reg[50];
1860 char src0_mask[6];
1861 char dst_mask[6];
1862 DWORD sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1863 DWORD sampler_type = arg->reg_maps->samplers[sampler_idx] & WINED3DSP_TEXTURETYPE_MASK;
1864 glsl_sample_function_t sample_function;
1865
1866 shader_glsl_append_dst(arg->buffer, arg);
1867 shader_glsl_get_write_mask(arg->dst, dst_mask);
1868 shader_glsl_get_sample_function(sampler_type, FALSE, &sample_function);
1869 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], sample_function.coord_mask, src0_reg, src0_mask, src0_str);
1870
1871 shader_addline(arg->buffer, "%s(Psampler%u, %s)%s);\n", sample_function.name, sampler_idx, src0_str, dst_mask);
1872 }
1873
1874 /** Process the WINED3DSIO_TEXKILL instruction in GLSL.
1875 * If any of the first 3 components are < 0, discard this pixel */
1876 void pshader_glsl_texkill(SHADER_OPCODE_ARG* arg) {
1877 glsl_dst_param_t dst_param;
1878
1879 shader_glsl_add_dst_param(arg, arg->dst, 0, &dst_param);
1880 shader_addline(arg->buffer, "if (any(lessThan(%s.xyz, vec3(0.0)))) discard;\n", dst_param.reg_name);
1881 }
1882
1883 /** Process the WINED3DSIO_DP2ADD instruction in GLSL.
1884 * dst = dot2(src0, src1) + src2 */
1885 void pshader_glsl_dp2add(SHADER_OPCODE_ARG* arg) {
1886 char src0_str[100], src1_str[100], src2_str[100];
1887 char src0_reg[50], src1_reg[50], src2_reg[50];
1888 char src0_mask[6], src1_mask[6], src2_mask[6];
1889 DWORD write_mask;
1890 size_t mask_size;
1891
1892 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1893 mask_size = shader_glsl_get_write_mask_size(write_mask);
1894
1895 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1, src0_reg, src0_mask, src0_str);
1896 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1, src1_reg, src1_mask, src1_str);
1897 shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], WINED3DSP_WRITEMASK_0, src2_reg, src2_mask, src2_str);
1898
1899 shader_addline(arg->buffer, "dot(%s, %s) + %s);\n", src0_str, src1_str, src2_str);
1900 }
1901
1902 void pshader_glsl_input_pack(
1903 SHADER_BUFFER* buffer,
1904 semantic* semantics_in) {
1905
1906 unsigned int i;
1907
1908 for (i = 0; i < MAX_REG_INPUT; i++) {
1909
1910 DWORD usage_token = semantics_in[i].usage;
1911 DWORD register_token = semantics_in[i].reg;
1912 DWORD usage, usage_idx;
1913 char reg_mask[6];
1914
1915 /* Uninitialized */
1916 if (!usage_token) continue;
1917 usage = (usage_token & WINED3DSP_DCL_USAGE_MASK) >> WINED3DSP_DCL_USAGE_SHIFT;
1918 usage_idx = (usage_token & WINED3DSP_DCL_USAGEINDEX_MASK) >> WINED3DSP_DCL_USAGEINDEX_SHIFT;
1919 shader_glsl_get_write_mask(register_token, reg_mask);
1920
1921 switch(usage) {
1922
1923 case D3DDECLUSAGE_COLOR:
1924 if (usage_idx == 0)
1925 shader_addline(buffer, "IN%u%s = vec4(gl_Color)%s;\n",
1926 i, reg_mask, reg_mask);
1927 else if (usage_idx == 1)
1928 shader_addline(buffer, "IN%u%s = vec4(gl_SecondaryColor)%s;\n",
1929 i, reg_mask, reg_mask);
1930 else
1931 shader_addline(buffer, "IN%u%s = vec4(unsupported_color_input)%s;\n",
1932 i, reg_mask, reg_mask);
1933 break;
1934
1935 case D3DDECLUSAGE_TEXCOORD:
1936 shader_addline(buffer, "IN%u%s = vec4(gl_TexCoord[%u])%s;\n",
1937 i, reg_mask, usage_idx, reg_mask );
1938 break;
1939
1940 case D3DDECLUSAGE_FOG:
1941 shader_addline(buffer, "IN%u%s = vec4(gl_FogFragCoord)%s;\n",
1942 i, reg_mask, reg_mask);
1943 break;
1944
1945 default:
1946 shader_addline(buffer, "IN%u%s = vec4(unsupported_input)%s;\n",
1947 i, reg_mask, reg_mask);
1948 }
1949 }
1950 }
1951
1952 /*********************************************
1953 * Vertex Shader Specific Code begins here
1954 ********************************************/
1955
1956 void vshader_glsl_output_unpack(
1957 SHADER_BUFFER* buffer,
1958 semantic* semantics_out) {
1959
1960 unsigned int i;
1961
1962 for (i = 0; i < MAX_REG_OUTPUT; i++) {
1963
1964 DWORD usage_token = semantics_out[i].usage;
1965 DWORD register_token = semantics_out[i].reg;
1966 DWORD usage, usage_idx;
1967 char reg_mask[6];
1968
1969 /* Uninitialized */
1970 if (!usage_token) continue;
1971
1972 usage = (usage_token & WINED3DSP_DCL_USAGE_MASK) >> WINED3DSP_DCL_USAGE_SHIFT;
1973 usage_idx = (usage_token & WINED3DSP_DCL_USAGEINDEX_MASK) >> WINED3DSP_DCL_USAGEINDEX_SHIFT;
1974 shader_glsl_get_write_mask(register_token, reg_mask);
1975
1976 switch(usage) {
1977
1978 case D3DDECLUSAGE_COLOR:
1979 if (usage_idx == 0)
1980 shader_addline(buffer, "gl_FrontColor%s = OUT%u%s;\n", reg_mask, i, reg_mask);
1981 else if (usage_idx == 1)
1982 shader_addline(buffer, "gl_FrontSecondaryColor%s = OUT%u%s;\n", reg_mask, i, reg_mask);
1983 else
1984 shader_addline(buffer, "unsupported_color_output%s = OUT%u%s;\n", reg_mask, i, reg_mask);
1985 break;
1986
1987 case D3DDECLUSAGE_POSITION:
1988 shader_addline(buffer, "gl_Position%s = OUT%u%s;\n", reg_mask, i, reg_mask);
1989 break;
1990
1991 case D3DDECLUSAGE_TEXCOORD:
1992 shader_addline(buffer, "gl_TexCoord[%u]%s = OUT%u%s;\n",
1993 usage_idx, reg_mask, i, reg_mask);
1994 break;
1995
1996 case WINED3DSHADERDECLUSAGE_PSIZE:
1997 shader_addline(buffer, "gl_PointSize = OUT%u.x;\n", i);
1998 break;
1999
2000 case WINED3DSHADERDECLUSAGE_FOG:
2001 shader_addline(buffer, "gl_FogFragCoord%s = OUT%u%s;\n", reg_mask, i, reg_mask);
2002 break;
2003
2004 default:
2005 shader_addline(buffer, "unsupported_output%s = OUT%u%s;\n", reg_mask, i, reg_mask);
2006 }
2007 }
2008 }
2009
2010 /** Attach a GLSL pixel or vertex shader object to the shader program */
2011 static void attach_glsl_shader(IWineD3DDevice *iface, IWineD3DBaseShader* shader) {
2012 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
2013 WineD3D_GL_Info *gl_info = &((IWineD3DImpl *)(This->wineD3D))->gl_info;
2014 GLhandleARB shaderObj = ((IWineD3DBaseShaderImpl*)shader)->baseShader.prgId;
2015 if (This->stateBlock->glsl_program && shaderObj != 0) {
2016 TRACE("Attaching GLSL shader object %u to program %u\n", shaderObj, This->stateBlock->glsl_program->programId);
2017 GL_EXTCALL(glAttachObjectARB(This->stateBlock->glsl_program->programId, shaderObj));
2018 checkGLcall("glAttachObjectARB");
2019 }
2020 }
2021
2022 /** Sets the GLSL program ID for the given pixel and vertex shader combination.
2023 * It sets the programId on the current StateBlock (because it should be called
2024 * inside of the DrawPrimitive() part of the render loop).
2025 *
2026 * If a program for the given combination does not exist, create one, and store
2027 * the program in the list. If it creates a program, it will link the given
2028 * objects, too.
2029 *
2030 * We keep the shader programs around on a list because linking
2031 * shader objects together is an expensive operation. It's much
2032 * faster to loop through a list of pre-compiled & linked programs
2033 * each time that the application sets a new pixel or vertex shader
2034 * than it is to re-link them together at that time.
2035 *
2036 * The list will be deleted in IWineD3DDevice::Release().
2037 */
2038 static void set_glsl_shader_program(IWineD3DDevice *iface) {
2039 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
2040 WineD3D_GL_Info *gl_info = &((IWineD3DImpl *)(This->wineD3D))->gl_info;
2041 IWineD3DPixelShader *pshader = This->stateBlock->pixelShader;
2042 IWineD3DVertexShader *vshader = This->stateBlock->vertexShader;
2043 struct glsl_shader_prog_link *curLink = NULL;
2044 struct glsl_shader_prog_link *newLink = NULL;
2045 struct list *ptr = NULL;
2046 GLhandleARB programId = 0;
2047 int i;
2048 char glsl_name[8];
2049
2050 ptr = list_head( &This->glsl_shader_progs );
2051 while (ptr) {
2052 /* At least one program exists - see if it matches our ps/vs combination */
2053 curLink = LIST_ENTRY( ptr, struct glsl_shader_prog_link, entry );
2054 if (vshader == curLink->vertexShader && pshader == curLink->pixelShader) {
2055 /* Existing Program found, use it */
2056 TRACE("Found existing program (%u) for this vertex/pixel shader combination\n",
2057 curLink->programId);
2058 This->stateBlock->glsl_program = curLink;
2059 return;
2060 }
2061 /* This isn't the entry we need - try the next one */
2062 ptr = list_next( &This->glsl_shader_progs, ptr );
2063 }
2064
2065 /* If we get to this point, then no matching program exists, so we create one */
2066 programId = GL_EXTCALL(glCreateProgramObjectARB());
2067 TRACE("Created new GLSL shader program %u\n", programId);
2068
2069 /* Allocate a new link for the list of programs */
2070 newLink = HeapAlloc(GetProcessHeap(), 0, sizeof(struct glsl_shader_prog_link));
2071 newLink->programId = programId;
2072 This->stateBlock->glsl_program = newLink;
2073
2074 /* Attach GLSL vshader */
2075 if (NULL != vshader && This->vs_selected_mode == SHADER_GLSL) {
2076 int i;
2077 int max_attribs = 16; /* TODO: Will this always be the case? It is at the moment... */
2078 char tmp_name[10];
2079
2080 TRACE("Attaching vertex shader to GLSL program\n");
2081 attach_glsl_shader(iface, (IWineD3DBaseShader*)vshader);
2082
2083 /* Bind vertex attributes to a corresponding index number to match
2084 * the same index numbers as ARB_vertex_programs (makes loading
2085 * vertex attributes simpler). With this method, we can use the
2086 * exact same code to load the attributes later for both ARB and
2087 * GLSL shaders.
2088 *
2089 * We have to do this here because we need to know the Program ID
2090 * in order to make the bindings work, and it has to be done prior
2091 * to linking the GLSL program. */
2092 for (i = 0; i < max_attribs; ++i) {
2093 snprintf(tmp_name, sizeof(tmp_name), "attrib%i", i);
2094 GL_EXTCALL(glBindAttribLocationARB(programId, i, tmp_name));
2095 }
2096 checkGLcall("glBindAttribLocationARB");
2097 newLink->vertexShader = vshader;
2098 }
2099
2100 /* Attach GLSL pshader */
2101 if (NULL != pshader && This->ps_selected_mode == SHADER_GLSL) {
2102 TRACE("Attaching pixel shader to GLSL program\n");
2103 attach_glsl_shader(iface, (IWineD3DBaseShader*)pshader);
2104 newLink->pixelShader = pshader;
2105 }
2106
2107 /* Link the program */
2108 TRACE("Linking GLSL shader program %u\n", programId);
2109 GL_EXTCALL(glLinkProgramARB(programId));
2110 print_glsl_info_log(&GLINFO_LOCATION, programId);
2111 list_add_head( &This->glsl_shader_progs, &newLink->entry);
2112
2113 newLink->vuniformF_locations = HeapAlloc(GetProcessHeap(), 0, sizeof(GLhandleARB) * GL_LIMITS(vshader_constantsF));
2114 for (i = 0; i < GL_LIMITS(vshader_constantsF); ++i) {
2115 snprintf(glsl_name, sizeof(glsl_name), "VC[%i]", i);
2116 newLink->vuniformF_locations[i] = GL_EXTCALL(glGetUniformLocationARB(programId, glsl_name));
2117 }
2118 newLink->puniformF_locations = HeapAlloc(GetProcessHeap(), 0, sizeof(GLhandleARB) * GL_LIMITS(pshader_constantsF));
2119 for (i = 0; i < GL_LIMITS(pshader_constantsF); ++i) {
2120 snprintf(glsl_name, sizeof(glsl_name), "PC[%i]", i);
2121 newLink->puniformF_locations[i] = GL_EXTCALL(glGetUniformLocationARB(programId, glsl_name));
2122 }
2123
2124 return;
2125 }
2126
2127 static GLhandleARB create_glsl_blt_shader(WineD3D_GL_Info *gl_info) {
2128 GLhandleARB program_id;
2129 GLhandleARB vshader_id, pshader_id;
2130 const char *blt_vshader[] = {
2131 "void main(void)\n"
2132 "{\n"
2133 " gl_Position = gl_Vertex;\n"
2134 " gl_FrontColor = vec4(1.0);\n"
2135 " gl_TexCoord[0].x = (gl_Vertex.x * 0.5) + 0.5;\n"
2136 " gl_TexCoord[0].y = (-gl_Vertex.y * 0.5) + 0.5;\n"
2137 "}\n"
2138 };
2139
2140 const char *blt_pshader[] = {
2141 "uniform sampler2D sampler;\n"
2142 "void main(void)\n"
2143 "{\n"
2144 " gl_FragDepth = texture2D(sampler, gl_TexCoord[0].xy).x;\n"
2145 "}\n"
2146 };
2147
2148 vshader_id = GL_EXTCALL(glCreateShaderObjectARB(GL_VERTEX_SHADER_ARB));
2149 GL_EXTCALL(glShaderSourceARB(vshader_id, 1, blt_vshader, NULL));
2150 GL_EXTCALL(glCompileShaderARB(vshader_id));
2151
2152 pshader_id = GL_EXTCALL(glCreateShaderObjectARB(GL_FRAGMENT_SHADER_ARB));
2153 GL_EXTCALL(glShaderSourceARB(pshader_id, 1, blt_pshader, NULL));
2154 GL_EXTCALL(glCompileShaderARB(pshader_id));
2155
2156 program_id = GL_EXTCALL(glCreateProgramObjectARB());
2157 GL_EXTCALL(glAttachObjectARB(program_id, vshader_id));
2158 GL_EXTCALL(glAttachObjectARB(program_id, pshader_id));
2159 GL_EXTCALL(glLinkProgramARB(program_id));
2160
2161 print_glsl_info_log(&GLINFO_LOCATION, program_id);
2162
2163 return program_id;
2164 }
2165
2166 static void shader_glsl_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {
2167 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
2168 WineD3D_GL_Info *gl_info = &((IWineD3DImpl *)(This->wineD3D))->gl_info;
2169 GLhandleARB program_id = 0;
2170
2171 if (useVS || usePS) set_glsl_shader_program(iface);
2172 else This->stateBlock->glsl_program = NULL;
2173
2174 program_id = This->stateBlock->glsl_program ? This->stateBlock->glsl_program->programId : 0;
2175 if (program_id) TRACE("Using GLSL program %u\n", program_id);
2176 GL_EXTCALL(glUseProgramObjectARB(program_id));
2177 checkGLcall("glUseProgramObjectARB");
2178 }
2179
2180 static void shader_glsl_select_depth_blt(IWineD3DDevice *iface) {
2181 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
2182 WineD3D_GL_Info *gl_info = &((IWineD3DImpl *)(This->wineD3D))->gl_info;
2183 static GLhandleARB program_id = 0;
2184 static GLhandleARB loc = -1;
2185
2186 if (!program_id) {
2187 program_id = create_glsl_blt_shader(gl_info);
2188 loc = GL_EXTCALL(glGetUniformLocationARB(program_id, "sampler"));
2189 }
2190
2191 GL_EXTCALL(glUseProgramObjectARB(program_id));
2192 GL_EXTCALL(glUniform1iARB(loc, 0));
2193 }
2194
2195 static void shader_glsl_cleanup(BOOL usePS, BOOL useVS) {
2196 /* Nothing to do */
2197 }
2198
2199 const shader_backend_t glsl_shader_backend = {
2200 &shader_glsl_select,
2201 &shader_glsl_select_depth_blt,
2202 &shader_glsl_load_constants,
2203 &shader_glsl_cleanup
2204 };
Windows API implementation