?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
wined3d: Simplify shader_glsl_get_swizzle().
[wine/multimedia.git] / dlls / wined3d / glsl_shader.c
blobf0fe21678d44582bbafe70325d025efe0391e998
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 const char *name;
40 DWORD coord_mask;
41 } glsl_sample_function_t;
42
43 /** Prints the GLSL info log which will contain error messages if they exist */
44 void print_glsl_info_log(WineD3D_GL_Info *gl_info, GLhandleARB obj) {
45
46 int infologLength = 0;
47 char *infoLog;
48
49 GL_EXTCALL(glGetObjectParameterivARB(obj,
50 GL_OBJECT_INFO_LOG_LENGTH_ARB,
51 &infologLength));
52
53 /* A size of 1 is just a null-terminated string, so the log should be bigger than
54 * that if there are errors. */
55 if (infologLength > 1)
56 {
57 infoLog = (char *)HeapAlloc(GetProcessHeap(), 0, infologLength);
58 GL_EXTCALL(glGetInfoLogARB(obj, infologLength, NULL, infoLog));
59 FIXME("Error received from GLSL shader #%u: %s\n", obj, debugstr_a(infoLog));
60 HeapFree(GetProcessHeap(), 0, infoLog);
61 }
62 }
63
64 /**
65 * Loads (pixel shader) samplers
66 */
67 void shader_glsl_load_psamplers(
68 WineD3D_GL_Info *gl_info,
69 IWineD3DStateBlock* iface) {
70
71 IWineD3DStateBlockImpl* stateBlock = (IWineD3DStateBlockImpl*) iface;
72 GLhandleARB programId = stateBlock->glsl_program->programId;
73 GLhandleARB name_loc;
74 int i;
75 char sampler_name[20];
76
77 for (i=0; i< GL_LIMITS(samplers); ++i) {
78 if (stateBlock->textures[i] != NULL) {
79 snprintf(sampler_name, sizeof(sampler_name), "Psampler%d", i);
80 name_loc = GL_EXTCALL(glGetUniformLocationARB(programId, sampler_name));
81 if (name_loc != -1) {
82 TRACE("Loading %s for texture %d\n", sampler_name, i);
83 GL_EXTCALL(glUniform1iARB(name_loc, i));
84 checkGLcall("glUniform1iARB");
85 }
86 }
87 }
88 }
89
90 /**
91 * Loads floating point constants (aka uniforms) into the currently set GLSL program.
92 * When constant_list == NULL, it will load all the constants.
93 */
94 static void shader_glsl_load_constantsF(IWineD3DBaseShaderImpl* This, WineD3D_GL_Info *gl_info,
95 unsigned int max_constants, float* constants, GLhandleARB *constant_locations,
96 struct list *constant_list) {
97 local_constant* lconst;
98 GLhandleARB tmp_loc;
99 int i;
100
101 if (!constant_list) {
102 if (TRACE_ON(d3d_shader)) {
103 for (i = 0; i < max_constants; ++i) {
104 tmp_loc = constant_locations[i];
105 if (tmp_loc != -1) {
106 TRACE("Loading constants %i: %f, %f, %f, %f\n", i,
107 constants[i * 4 + 0], constants[i * 4 + 1],
108 constants[i * 4 + 2], constants[i * 4 + 3]);
109 }
110 }
111 }
112 for (i = 0; i < max_constants; ++i) {
113 tmp_loc = constant_locations[i];
114 if (tmp_loc != -1) {
115 /* We found this uniform name in the program - go ahead and send the data */
116 GL_EXTCALL(glUniform4fvARB(tmp_loc, 1, constants + (i * 4)));
117 }
118 }
119 checkGLcall("glUniform4fvARB()");
120 } else {
121 constant_entry *constant;
122 if (TRACE_ON(d3d_shader)) {
123 LIST_FOR_EACH_ENTRY(constant, constant_list, constant_entry, entry) {
124 i = constant->idx;
125 tmp_loc = constant_locations[i];
126 if (tmp_loc != -1) {
127 TRACE("Loading constants %i: %f, %f, %f, %f\n", i,
128 constants[i * 4 + 0], constants[i * 4 + 1],
129 constants[i * 4 + 2], constants[i * 4 + 3]);
130 }
131 }
132 }
133 LIST_FOR_EACH_ENTRY(constant, constant_list, constant_entry, entry) {
134 i = constant->idx;
135 tmp_loc = constant_locations[i];
136 if (tmp_loc != -1) {
137 /* We found this uniform name in the program - go ahead and send the data */
138 GL_EXTCALL(glUniform4fvARB(tmp_loc, 1, constants + (i * 4)));
139 }
140 }
141 checkGLcall("glUniform4fvARB()");
142 }
143
144 /* Load immediate constants */
145 if (TRACE_ON(d3d_shader)) {
146 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
147 tmp_loc = constant_locations[lconst->idx];
148 if (tmp_loc != -1) {
149 GLfloat* values = (GLfloat*)lconst->value;
150 TRACE("Loading local constants %i: %f, %f, %f, %f\n", lconst->idx,
151 values[0], values[1], values[2], values[3]);
152 }
153 }
154 }
155 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
156 tmp_loc = constant_locations[lconst->idx];
157 if (tmp_loc != -1) {
158 /* We found this uniform name in the program - go ahead and send the data */
159 GL_EXTCALL(glUniform4fvARB(tmp_loc, 1, (GLfloat*)lconst->value));
160 }
161 }
162 checkGLcall("glUniform4fvARB()");
163 }
164
165 /**
166 * Loads integer constants (aka uniforms) into the currently set GLSL program.
167 * When @constants_set == NULL, it will load all the constants.
168 */
169 void shader_glsl_load_constantsI(
170 IWineD3DBaseShaderImpl* This,
171 WineD3D_GL_Info *gl_info,
172 GLhandleARB programId,
173 unsigned max_constants,
174 int* constants,
175 BOOL* constants_set) {
176
177 GLhandleARB tmp_loc;
178 int i;
179 char tmp_name[8];
180 char is_pshader = shader_is_pshader_version(This->baseShader.hex_version);
181 const char* prefix = is_pshader? "PI":"VI";
182 struct list* ptr;
183
184 for (i=0; i<max_constants; ++i) {
185 if (NULL == constants_set || constants_set[i]) {
186
187 TRACE("Loading constants %i: %i, %i, %i, %i\n",
188 i, constants[i*4], constants[i*4+1], constants[i*4+2], constants[i*4+3]);
189
190 /* TODO: Benchmark and see if it would be beneficial to store the
191 * locations of the constants to avoid looking up each time */
192 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, i);
193 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
194 if (tmp_loc != -1) {
195 /* We found this uniform name in the program - go ahead and send the data */
196 GL_EXTCALL(glUniform4ivARB(tmp_loc, 1, &constants[i*4]));
197 checkGLcall("glUniform4ivARB");
198 }
199 }
200 }
201
202 /* Load immediate constants */
203 ptr = list_head(&This->baseShader.constantsI);
204 while (ptr) {
205 local_constant* lconst = LIST_ENTRY(ptr, struct local_constant, entry);
206 unsigned int idx = lconst->idx;
207 GLint* values = (GLint*) lconst->value;
208
209 TRACE("Loading local constants %i: %i, %i, %i, %i\n", idx,
210 values[0], values[1], values[2], values[3]);
211
212 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, idx);
213 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
214 if (tmp_loc != -1) {
215 /* We found this uniform name in the program - go ahead and send the data */
216 GL_EXTCALL(glUniform4ivARB(tmp_loc, 1, values));
217 checkGLcall("glUniform4ivARB");
218 }
219 ptr = list_next(&This->baseShader.constantsI, ptr);
220 }
221 }
222
223 /**
224 * Loads boolean constants (aka uniforms) into the currently set GLSL program.
225 * When @constants_set == NULL, it will load all the constants.
226 */
227 void shader_glsl_load_constantsB(
228 IWineD3DBaseShaderImpl* This,
229 WineD3D_GL_Info *gl_info,
230 GLhandleARB programId,
231 unsigned max_constants,
232 BOOL* constants,
233 BOOL* constants_set) {
234
235 GLhandleARB tmp_loc;
236 int i;
237 char tmp_name[8];
238 char is_pshader = shader_is_pshader_version(This->baseShader.hex_version);
239 const char* prefix = is_pshader? "PB":"VB";
240 struct list* ptr;
241
242 for (i=0; i<max_constants; ++i) {
243 if (NULL == constants_set || constants_set[i]) {
244
245 TRACE("Loading constants %i: %i;\n", i, constants[i*4]);
246
247 /* TODO: Benchmark and see if it would be beneficial to store the
248 * locations of the constants to avoid looking up each time */
249 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, i);
250 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
251 if (tmp_loc != -1) {
252 /* We found this uniform name in the program - go ahead and send the data */
253 GL_EXTCALL(glUniform1ivARB(tmp_loc, 1, &constants[i*4]));
254 checkGLcall("glUniform1ivARB");
255 }
256 }
257 }
258
259 /* Load immediate constants */
260 ptr = list_head(&This->baseShader.constantsB);
261 while (ptr) {
262 local_constant* lconst = LIST_ENTRY(ptr, struct local_constant, entry);
263 unsigned int idx = lconst->idx;
264 GLint* values = (GLint*) lconst->value;
265
266 TRACE("Loading local constants %i: %i\n", idx, values[0]);
267
268 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, idx);
269 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
270 if (tmp_loc != -1) {
271 /* We found this uniform name in the program - go ahead and send the data */
272 GL_EXTCALL(glUniform1ivARB(tmp_loc, 1, values));
273 checkGLcall("glUniform1ivARB");
274 }
275 ptr = list_next(&This->baseShader.constantsB, ptr);
276 }
277 }
278
279
280
281 /**
282 * Loads the app-supplied constants into the currently set GLSL program.
283 */
284 void shader_glsl_load_constants(
285 IWineD3DDevice* device,
286 char usePixelShader,
287 char useVertexShader) {
288
289 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) device;
290 IWineD3DStateBlockImpl* stateBlock = deviceImpl->stateBlock;
291 WineD3D_GL_Info *gl_info = &((IWineD3DImpl*) deviceImpl->wineD3D)->gl_info;
292
293 GLhandleARB *constant_locations;
294 struct list *constant_list;
295 GLhandleARB programId;
296
297 if (!stateBlock->glsl_program) {
298 /* No GLSL program set - nothing to do. */
299 return;
300 }
301 programId = stateBlock->glsl_program->programId;
302
303 if (useVertexShader) {
304 IWineD3DBaseShaderImpl* vshader = (IWineD3DBaseShaderImpl*) stateBlock->vertexShader;
305 IWineD3DVertexShaderImpl* vshader_impl = (IWineD3DVertexShaderImpl*) vshader;
306 GLint pos;
307
308 IWineD3DVertexDeclarationImpl* vertexDeclaration =
309 (IWineD3DVertexDeclarationImpl*) vshader_impl->vertexDeclaration;
310
311 constant_locations = stateBlock->glsl_program->vuniformF_locations;
312 constant_list = &stateBlock->set_vconstantsF;
313
314 if (NULL != vertexDeclaration && NULL != vertexDeclaration->constants) {
315 /* Load DirectX 8 float constants/uniforms for vertex shader */
316 shader_glsl_load_constantsF(vshader, gl_info, GL_LIMITS(vshader_constantsF),
317 vertexDeclaration->constants, constant_locations, NULL);
318 }
319
320 /* Load DirectX 9 float constants/uniforms for vertex shader */
321 shader_glsl_load_constantsF(vshader, gl_info, GL_LIMITS(vshader_constantsF),
322 stateBlock->vertexShaderConstantF, constant_locations, constant_list);
323
324 /* Load DirectX 9 integer constants/uniforms for vertex shader */
325 shader_glsl_load_constantsI(vshader, gl_info, programId, MAX_CONST_I,
326 stateBlock->vertexShaderConstantI,
327 stateBlock->set.vertexShaderConstantsI);
328
329 /* Load DirectX 9 boolean constants/uniforms for vertex shader */
330 shader_glsl_load_constantsB(vshader, gl_info, programId, MAX_CONST_B,
331 stateBlock->vertexShaderConstantB,
332 stateBlock->set.vertexShaderConstantsB);
333
334 /* Upload the position fixup params */
335 pos = GL_EXTCALL(glGetUniformLocationARB(programId, "posFixup"));
336 checkGLcall("glGetUniformLocationARB");
337 GL_EXTCALL(glUniform4fvARB(pos, 1, &deviceImpl->posFixup[0]));
338 checkGLcall("glUniform4fvARB");
339 }
340
341 if (usePixelShader) {
342
343 IWineD3DBaseShaderImpl* pshader = (IWineD3DBaseShaderImpl*) stateBlock->pixelShader;
344
345 constant_locations = stateBlock->glsl_program->puniformF_locations;
346 constant_list = &stateBlock->set_pconstantsF;
347
348 /* Load pixel shader samplers */
349 shader_glsl_load_psamplers(gl_info, (IWineD3DStateBlock*) stateBlock);
350
351 /* Load DirectX 9 float constants/uniforms for pixel shader */
352 shader_glsl_load_constantsF(pshader, gl_info, GL_LIMITS(pshader_constantsF),
353 stateBlock->pixelShaderConstantF, constant_locations, constant_list);
354
355 /* Load DirectX 9 integer constants/uniforms for pixel shader */
356 shader_glsl_load_constantsI(pshader, gl_info, programId, MAX_CONST_I,
357 stateBlock->pixelShaderConstantI,
358 stateBlock->set.pixelShaderConstantsI);
359
360 /* Load DirectX 9 boolean constants/uniforms for pixel shader */
361 shader_glsl_load_constantsB(pshader, gl_info, programId, MAX_CONST_B,
362 stateBlock->pixelShaderConstantB,
363 stateBlock->set.pixelShaderConstantsB);
364 }
365 }
366
367 /** Generate the variable & register declarations for the GLSL output target */
368 void shader_generate_glsl_declarations(
369 IWineD3DBaseShader *iface,
370 shader_reg_maps* reg_maps,
371 SHADER_BUFFER* buffer,
372 WineD3D_GL_Info* gl_info) {
373
374 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
375 int i;
376
377 /* There are some minor differences between pixel and vertex shaders */
378 char pshader = shader_is_pshader_version(This->baseShader.hex_version);
379 char prefix = pshader ? 'P' : 'V';
380
381 /* Prototype the subroutines */
382 for (i = 0; i < This->baseShader.limits.label; i++) {
383 if (reg_maps->labels[i])
384 shader_addline(buffer, "void subroutine%lu();\n", i);
385 }
386
387 /* Declare the constants (aka uniforms) */
388 if (This->baseShader.limits.constant_float > 0) {
389 unsigned max_constantsF = min(This->baseShader.limits.constant_float,
390 (pshader ? GL_LIMITS(pshader_constantsF) : GL_LIMITS(vshader_constantsF)));
391 shader_addline(buffer, "uniform vec4 %cC[%u];\n", prefix, max_constantsF);
392 }
393
394 if (This->baseShader.limits.constant_int > 0)
395 shader_addline(buffer, "uniform ivec4 %cI[%u];\n", prefix, This->baseShader.limits.constant_int);
396
397 if (This->baseShader.limits.constant_bool > 0)
398 shader_addline(buffer, "uniform bool %cB[%u];\n", prefix, This->baseShader.limits.constant_bool);
399
400 if(!pshader)
401 shader_addline(buffer, "uniform vec4 posFixup;\n");
402
403 /* Declare texture samplers */
404 for (i = 0; i < This->baseShader.limits.sampler; i++) {
405 if (reg_maps->samplers[i]) {
406
407 DWORD stype = reg_maps->samplers[i] & WINED3DSP_TEXTURETYPE_MASK;
408 switch (stype) {
409
410 case WINED3DSTT_1D:
411 shader_addline(buffer, "uniform sampler1D %csampler%lu;\n", prefix, i);
412 break;
413 case WINED3DSTT_2D:
414 shader_addline(buffer, "uniform sampler2D %csampler%lu;\n", prefix, i);
415 break;
416 case WINED3DSTT_CUBE:
417 shader_addline(buffer, "uniform samplerCube %csampler%lu;\n", prefix, i);
418 break;
419 case WINED3DSTT_VOLUME:
420 shader_addline(buffer, "uniform sampler3D %csampler%lu;\n", prefix, i);
421 break;
422 default:
423 shader_addline(buffer, "uniform unsupported_sampler %csampler%lu;\n", prefix, i);
424 FIXME("Unrecognized sampler type: %#x\n", stype);
425 break;
426 }
427 }
428 }
429
430 /* Declare address variables */
431 for (i = 0; i < This->baseShader.limits.address; i++) {
432 if (reg_maps->address[i])
433 shader_addline(buffer, "ivec4 A%d;\n", i);
434 }
435
436 /* Declare texture coordinate temporaries and initialize them */
437 for (i = 0; i < This->baseShader.limits.texcoord; i++) {
438 if (reg_maps->texcoord[i])
439 shader_addline(buffer, "vec4 T%lu = gl_TexCoord[%lu];\n", i, i);
440 }
441
442 /* Declare input register temporaries */
443 for (i=0; i < This->baseShader.limits.packed_input; i++) {
444 if (reg_maps->packed_input[i])
445 shader_addline(buffer, "vec4 IN%lu;\n", i);
446 }
447
448 /* Declare output register temporaries */
449 for (i = 0; i < This->baseShader.limits.packed_output; i++) {
450 if (reg_maps->packed_output[i])
451 shader_addline(buffer, "vec4 OUT%lu;\n", i);
452 }
453
454 /* Declare temporary variables */
455 for(i = 0; i < This->baseShader.limits.temporary; i++) {
456 if (reg_maps->temporary[i])
457 shader_addline(buffer, "vec4 R%lu;\n", i);
458 }
459
460 /* Declare attributes */
461 for (i = 0; i < This->baseShader.limits.attributes; i++) {
462 if (reg_maps->attributes[i])
463 shader_addline(buffer, "attribute vec4 attrib%i;\n", i);
464 }
465
466 /* Declare loop register aL */
467 if (reg_maps->loop) {
468 shader_addline(buffer, "int aL;\n");
469 shader_addline(buffer, "int tmpInt;\n");
470 }
471
472 /* Temporary variables for matrix operations */
473 shader_addline(buffer, "vec4 tmp0;\n");
474 shader_addline(buffer, "vec4 tmp1;\n");
475
476 /* Start the main program */
477 shader_addline(buffer, "void main() {\n");
478 }
479
480 /*****************************************************************************
481 * Functions to generate GLSL strings from DirectX Shader bytecode begin here.
482 *
483 * For more information, see http://wiki.winehq.org/DirectX-Shaders
484 ****************************************************************************/
485
486 /* Prototypes */
487 static void shader_glsl_add_src_param(SHADER_OPCODE_ARG* arg, const DWORD param,
488 const DWORD addr_token, DWORD mask, char *reg_name, char *swizzle, char *out_str);
489
490 /** Used for opcode modifiers - They multiply the result by the specified amount */
491 static const char * const shift_glsl_tab[] = {
492 "", /* 0 (none) */
493 "2.0 * ", /* 1 (x2) */
494 "4.0 * ", /* 2 (x4) */
495 "8.0 * ", /* 3 (x8) */
496 "16.0 * ", /* 4 (x16) */
497 "32.0 * ", /* 5 (x32) */
498 "", /* 6 (x64) */
499 "", /* 7 (x128) */
500 "", /* 8 (d256) */
501 "", /* 9 (d128) */
502 "", /* 10 (d64) */
503 "", /* 11 (d32) */
504 "0.0625 * ", /* 12 (d16) */
505 "0.125 * ", /* 13 (d8) */
506 "0.25 * ", /* 14 (d4) */
507 "0.5 * " /* 15 (d2) */
508 };
509
510 /* Generate a GLSL parameter that does the input modifier computation and return the input register/mask to use */
511 static void shader_glsl_gen_modifier (
512 const DWORD instr,
513 const char *in_reg,
514 const char *in_regswizzle,
515 char *out_str) {
516
517 out_str[0] = 0;
518
519 if (instr == WINED3DSIO_TEXKILL)
520 return;
521
522 switch (instr & WINED3DSP_SRCMOD_MASK) {
523 case WINED3DSPSM_DZ: /* Need to handle this in the instructions itself (texld & texcrd). */
524 case WINED3DSPSM_DW:
525 case WINED3DSPSM_NONE:
526 sprintf(out_str, "%s%s", in_reg, in_regswizzle);
527 break;
528 case WINED3DSPSM_NEG:
529 sprintf(out_str, "-%s%s", in_reg, in_regswizzle);
530 break;
531 case WINED3DSPSM_NOT:
532 sprintf(out_str, "!%s%s", in_reg, in_regswizzle);
533 break;
534 case WINED3DSPSM_BIAS:
535 sprintf(out_str, "(%s%s - vec4(0.5)%s)", in_reg, in_regswizzle, in_regswizzle);
536 break;
537 case WINED3DSPSM_BIASNEG:
538 sprintf(out_str, "-(%s%s - vec4(0.5)%s)", in_reg, in_regswizzle, in_regswizzle);
539 break;
540 case WINED3DSPSM_SIGN:
541 sprintf(out_str, "(2.0 * (%s%s - 0.5))", in_reg, in_regswizzle);
542 break;
543 case WINED3DSPSM_SIGNNEG:
544 sprintf(out_str, "-(2.0 * (%s%s - 0.5))", in_reg, in_regswizzle);
545 break;
546 case WINED3DSPSM_COMP:
547 sprintf(out_str, "(1.0 - %s%s)", in_reg, in_regswizzle);
548 break;
549 case WINED3DSPSM_X2:
550 sprintf(out_str, "(2.0 * %s%s)", in_reg, in_regswizzle);
551 break;
552 case WINED3DSPSM_X2NEG:
553 sprintf(out_str, "-(2.0 * %s%s)", in_reg, in_regswizzle);
554 break;
555 case WINED3DSPSM_ABS:
556 sprintf(out_str, "abs(%s%s)", in_reg, in_regswizzle);
557 break;
558 case WINED3DSPSM_ABSNEG:
559 sprintf(out_str, "-abs(%s%s)", in_reg, in_regswizzle);
560 break;
561 default:
562 FIXME("Unhandled modifier %u\n", (instr & WINED3DSP_SRCMOD_MASK));
563 sprintf(out_str, "%s%s", in_reg, in_regswizzle);
564 }
565 }
566
567 /** Writes the GLSL variable name that corresponds to the register that the
568 * DX opcode parameter is trying to access */
569 static void shader_glsl_get_register_name(
570 const DWORD param,
571 const DWORD addr_token,
572 char* regstr,
573 BOOL* is_color,
574 SHADER_OPCODE_ARG* arg) {
575
576 /* oPos, oFog and oPts in D3D */
577 static const char * const hwrastout_reg_names[] = { "gl_Position", "gl_FogFragCoord", "gl_PointSize" };
578
579 DWORD reg = param & WINED3DSP_REGNUM_MASK;
580 DWORD regtype = shader_get_regtype(param);
581 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) arg->shader;
582 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
583 WineD3D_GL_Info* gl_info = &((IWineD3DImpl*)deviceImpl->wineD3D)->gl_info;
584
585 char pshader = shader_is_pshader_version(This->baseShader.hex_version);
586 char tmpStr[50];
587
588 *is_color = FALSE;
589
590 switch (regtype) {
591 case WINED3DSPR_TEMP:
592 sprintf(tmpStr, "R%u", reg);
593 break;
594 case WINED3DSPR_INPUT:
595 if (pshader) {
596 /* Pixel shaders >= 3.0 */
597 if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3)
598 sprintf(tmpStr, "IN%u", reg);
599 else {
600 if (reg==0)
601 strcpy(tmpStr, "gl_Color");
602 else
603 strcpy(tmpStr, "gl_SecondaryColor");
604 }
605 } else {
606 if (vshader_input_is_color((IWineD3DVertexShader*) This, reg))
607 *is_color = TRUE;
608 sprintf(tmpStr, "attrib%u", reg);
609 }
610 break;
611 case WINED3DSPR_CONST:
612 {
613 const char* prefix = pshader? "PC":"VC";
614
615 /* Relative addressing */
616 if (param & WINED3DSHADER_ADDRMODE_RELATIVE) {
617
618 /* Relative addressing on shaders 2.0+ have a relative address token,
619 * prior to that, it was hard-coded as "A0.x" because there's only 1 register */
620 if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 2) {
621 char relStr[100], relReg[50], relMask[6];
622 shader_glsl_add_src_param(arg, addr_token, 0, WINED3DSP_WRITEMASK_0, relReg, relMask, relStr);
623 sprintf(tmpStr, "%s[%s + %u]", prefix, relStr, reg);
624 } else
625 sprintf(tmpStr, "%s[A0.x + %u]", prefix, reg);
626
627 } else
628 sprintf(tmpStr, "%s[%u]", prefix, reg);
629
630 break;
631 }
632 case WINED3DSPR_CONSTINT:
633 if (pshader)
634 sprintf(tmpStr, "PI[%u]", reg);
635 else
636 sprintf(tmpStr, "VI[%u]", reg);
637 break;
638 case WINED3DSPR_CONSTBOOL:
639 if (pshader)
640 sprintf(tmpStr, "PB[%u]", reg);
641 else
642 sprintf(tmpStr, "VB[%u]", reg);
643 break;
644 case WINED3DSPR_TEXTURE: /* case WINED3DSPR_ADDR: */
645 if (pshader) {
646 sprintf(tmpStr, "T%u", reg);
647 } else {
648 sprintf(tmpStr, "A%u", reg);
649 }
650 break;
651 case WINED3DSPR_LOOP:
652 sprintf(tmpStr, "aL");
653 break;
654 case WINED3DSPR_SAMPLER:
655 if (pshader)
656 sprintf(tmpStr, "Psampler%u", reg);
657 else
658 sprintf(tmpStr, "Vsampler%u", reg);
659 break;
660 case WINED3DSPR_COLOROUT:
661 if (reg >= GL_LIMITS(buffers)) {
662 WARN("Write to render target %u, only %d supported\n", reg, 4);
663 }
664 if (GL_SUPPORT(ARB_DRAW_BUFFERS)) {
665 sprintf(tmpStr, "gl_FragData[%u]", reg);
666 } else { /* On older cards with GLSL support like the GeforceFX there's only one buffer. */
667 sprintf(tmpStr, "gl_FragColor");
668 }
669 break;
670 case WINED3DSPR_RASTOUT:
671 sprintf(tmpStr, "%s", hwrastout_reg_names[reg]);
672 break;
673 case WINED3DSPR_DEPTHOUT:
674 sprintf(tmpStr, "gl_FragDepth");
675 break;
676 case WINED3DSPR_ATTROUT:
677 if (reg == 0) {
678 sprintf(tmpStr, "gl_FrontColor");
679 } else {
680 sprintf(tmpStr, "gl_FrontSecondaryColor");
681 }
682 break;
683 case WINED3DSPR_TEXCRDOUT:
684 /* Vertex shaders >= 3.0: WINED3DSPR_OUTPUT */
685 if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3)
686 sprintf(tmpStr, "OUT%u", reg);
687 else
688 sprintf(tmpStr, "gl_TexCoord[%u]", reg);
689 break;
690 default:
691 FIXME("Unhandled register name Type(%d)\n", regtype);
692 sprintf(tmpStr, "unrecognized_register");
693 break;
694 }
695
696 strcat(regstr, tmpStr);
697 }
698
699 /* Get the GLSL write mask for the destination register */
700 static DWORD shader_glsl_get_write_mask(const DWORD param, char *write_mask) {
701 char *ptr = write_mask;
702 DWORD mask = param & WINED3DSP_WRITEMASK_ALL;
703
704 /* gl_FogFragCoord and glPointSize are floats, fixup the write mask. */
705 if ((shader_get_regtype(param) == WINED3DSPR_RASTOUT) && ((param & WINED3DSP_REGNUM_MASK) != 0)) {
706 mask = WINED3DSP_WRITEMASK_0;
707 } else {
708 if (mask != WINED3DSP_WRITEMASK_ALL) {
709 *ptr++ = '.';
710 if (param & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
711 if (param & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
712 if (param & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
713 if (param & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
714 }
715 }
716
717 *ptr = '\0';
718
719 return mask;
720 }
721
722 static size_t shader_glsl_get_write_mask_size(DWORD write_mask) {
723 size_t size = 0;
724
725 if (write_mask & WINED3DSP_WRITEMASK_0) ++size;
726 if (write_mask & WINED3DSP_WRITEMASK_1) ++size;
727 if (write_mask & WINED3DSP_WRITEMASK_2) ++size;
728 if (write_mask & WINED3DSP_WRITEMASK_3) ++size;
729
730 return size;
731 }
732
733 static void shader_glsl_get_swizzle(const DWORD param, BOOL fixup, DWORD mask, char *swizzle_str) {
734 /* For registers of type WINED3DDECLTYPE_D3DCOLOR, data is stored as "bgra",
735 * but addressed as "rgba". To fix this we need to swap the register's x
736 * and z components. */
737 DWORD swizzle = (param & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
738 const char *swizzle_chars = fixup ? "zyxw" : "xyzw";
739 char *ptr = swizzle_str;
740
741 *ptr++ = '.';
742 /* swizzle bits fields: wwzzyyxx */
743 if (mask & WINED3DSP_WRITEMASK_0) *ptr++ = swizzle_chars[swizzle & 0x03];
744 if (mask & WINED3DSP_WRITEMASK_1) *ptr++ = swizzle_chars[(swizzle >> 2) & 0x03];
745 if (mask & WINED3DSP_WRITEMASK_2) *ptr++ = swizzle_chars[(swizzle >> 4) & 0x03];
746 if (mask & WINED3DSP_WRITEMASK_3) *ptr++ = swizzle_chars[(swizzle >> 6) & 0x03];
747
748 *ptr = '\0';
749 }
750
751 /* From a given parameter token, generate the corresponding GLSL string.
752 * Also, return the actual register name and swizzle in case the
753 * caller needs this information as well. */
754 static void shader_glsl_add_src_param(SHADER_OPCODE_ARG* arg, const DWORD param,
755 const DWORD addr_token, DWORD mask, char *reg_name, char *swizzle, char *out_str) {
756 BOOL is_color = FALSE;
757 swizzle[0] = reg_name[0] = out_str[0] = 0;
758
759 shader_glsl_get_register_name(param, addr_token, reg_name, &is_color, arg);
760
761 shader_glsl_get_swizzle(param, is_color, mask, swizzle);
762 shader_glsl_gen_modifier(param, reg_name, swizzle, out_str);
763 }
764
765 /* From a given parameter token, generate the corresponding GLSL string.
766 * Also, return the actual register name and swizzle in case the
767 * caller needs this information as well. */
768 static DWORD shader_glsl_add_dst_param(SHADER_OPCODE_ARG* arg, const DWORD param,
769 const DWORD addr_token, char *reg_name, char *write_mask, char *out_str) {
770 BOOL is_color = FALSE;
771 DWORD mask;
772 write_mask[0] = reg_name[0] = out_str[0] = 0;
773
774 shader_glsl_get_register_name(param, addr_token, reg_name, &is_color, arg);
775
776 mask = shader_glsl_get_write_mask(param, write_mask);
777 sprintf(out_str, "%s%s", reg_name, write_mask);
778
779 return mask;
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 char reg_name[50], write_mask[6], reg_str[100];
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, reg_name, write_mask, reg_str);
790 shader_addline(buffer, "%s = %s(", reg_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 char dst_reg[50];
802 char dst_mask[6];
803 char dst_str[100];
804
805 shader_glsl_add_dst_param(arg, arg->dst, 0, dst_reg, dst_mask, dst_str);
806
807 if (mask & WINED3DSPDM_SATURATE) {
808 /* _SAT means to clamp the value of the register to between 0 and 1 */
809 shader_addline(arg->buffer, "%s%s = clamp(%s%s, 0.0, 1.0);\n", dst_reg, dst_mask, dst_reg, dst_mask);
810 }
811 if (mask & WINED3DSPDM_MSAMPCENTROID) {
812 FIXME("_centroid modifier not handled\n");
813 }
814 if (mask & WINED3DSPDM_PARTIALPRECISION) {
815 /* MSDN says this modifier can be safely ignored, so that's what we'll do. */
816 }
817 }
818 }
819
820 static inline const char* shader_get_comp_op(
821 const DWORD opcode) {
822
823 DWORD op = (opcode & INST_CONTROLS_MASK) >> INST_CONTROLS_SHIFT;
824 switch (op) {
825 case COMPARISON_GT: return ">";
826 case COMPARISON_EQ: return "==";
827 case COMPARISON_GE: return ">=";
828 case COMPARISON_LT: return "<";
829 case COMPARISON_NE: return "!=";
830 case COMPARISON_LE: return "<=";
831 default:
832 FIXME("Unrecognized comparison value: %u\n", op);
833 return "(\?\?)";
834 }
835 }
836
837 static void shader_glsl_get_sample_function(DWORD sampler_type, BOOL projected, glsl_sample_function_t *sample_function) {
838 /* Note that there's no such thing as a projected cube texture. */
839 switch(sampler_type) {
840 case WINED3DSTT_1D:
841 sample_function->name = projected ? "texture1DProj" : "texture1D";
842 sample_function->coord_mask = WINED3DSP_WRITEMASK_0;
843 break;
844 case WINED3DSTT_2D:
845 sample_function->name = projected ? "texture2DProj" : "texture2D";
846 sample_function->coord_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1;
847 break;
848 case WINED3DSTT_CUBE:
849 sample_function->name = "textureCube";
850 sample_function->coord_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
851 break;
852 case WINED3DSTT_VOLUME:
853 sample_function->name = projected ? "texture3DProj" : "texture3D";
854 sample_function->coord_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
855 break;
856 default:
857 sample_function->name = "";
858 FIXME("Unrecognized sampler type: %#x;\n", sampler_type);
859 break;
860 }
861 }
862
863
864 /*****************************************************************************
865 *
866 * Begin processing individual instruction opcodes
867 *
868 ****************************************************************************/
869
870 /* Generate GLSL arithmetic functions (dst = src1 + src2) */
871 void shader_glsl_arith(SHADER_OPCODE_ARG* arg) {
872 CONST SHADER_OPCODE* curOpcode = arg->opcode;
873 SHADER_BUFFER* buffer = arg->buffer;
874 char src0_reg[50], src1_reg[50];
875 char src0_mask[6], src1_mask[6];
876 char src0_str[100], src1_str[100];
877 DWORD write_mask;
878 char op;
879
880 /* Determine the GLSL operator to use based on the opcode */
881 switch (curOpcode->opcode) {
882 case WINED3DSIO_MUL: op = '*'; break;
883 case WINED3DSIO_ADD: op = '+'; break;
884 case WINED3DSIO_SUB: op = '-'; break;
885 default:
886 op = ' ';
887 FIXME("Opcode %s not yet handled in GLSL\n", curOpcode->name);
888 break;
889 }
890
891 write_mask = shader_glsl_append_dst(buffer, arg);
892 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, src0_reg, src0_mask, src0_str);
893 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, src1_reg, src1_mask, src1_str);
894 shader_addline(buffer, "%s %c %s);\n", src0_str, op, src1_str);
895 }
896
897 /* Process the WINED3DSIO_MOV opcode using GLSL (dst = src) */
898 void shader_glsl_mov(SHADER_OPCODE_ARG* arg) {
899 IWineD3DBaseShaderImpl* shader = (IWineD3DBaseShaderImpl*) arg->shader;
900 SHADER_BUFFER* buffer = arg->buffer;
901 char src0_str[100];
902 char src0_reg[50];
903 char src0_mask[6];
904 DWORD write_mask;
905
906 write_mask = shader_glsl_append_dst(buffer, arg);
907 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, src0_reg, src0_mask, src0_str);
908
909 /* In vs_1_1 WINED3DSIO_MOV can write to the adress register. In later
910 * shader versions WINED3DSIO_MOVA is used for this. */
911 if ((WINED3DSHADER_VERSION_MAJOR(shader->baseShader.hex_version) == 1 &&
912 !shader_is_pshader_version(shader->baseShader.hex_version) &&
913 shader_get_regtype(arg->dst) == WINED3DSPR_ADDR) ||
914 arg->opcode->opcode == WINED3DSIO_MOVA) {
915 /* We need to *round* to the nearest int here. */
916 size_t mask_size = shader_glsl_get_write_mask_size(write_mask);
917 if (mask_size > 1) {
918 shader_addline(buffer, "ivec%d(floor(%s + vec%d(0.5))));\n", mask_size, src0_str, mask_size);
919 } else {
920 shader_addline(buffer, "int(floor(%s + 0.5)));\n", src0_str);
921 }
922 } else {
923 shader_addline(buffer, "%s);\n", src0_str);
924 }
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 CONST SHADER_OPCODE* curOpcode = arg->opcode;
930 SHADER_BUFFER* buffer = arg->buffer;
931 char src0_str[100], src1_str[100];
932 char src0_reg[50], src1_reg[50];
933 char src0_mask[6], src1_mask[6];
934 DWORD dst_write_mask, src_write_mask;
935 size_t dst_size = 0;
936
937 dst_write_mask = shader_glsl_append_dst(buffer, arg);
938 dst_size = shader_glsl_get_write_mask_size(dst_write_mask);
939
940 /* dp3 works on vec3, dp4 on vec4 */
941 if (curOpcode->opcode == WINED3DSIO_DP4) {
942 src_write_mask = WINED3DSP_WRITEMASK_ALL;
943 } else {
944 src_write_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
945 }
946
947 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_write_mask, src0_reg, src0_mask, src0_str);
948 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], src_write_mask, src1_reg, src1_mask, src1_str);
949
950 if (dst_size > 1) {
951 shader_addline(buffer, "vec%d(dot(%s, %s)));\n", dst_size, src0_str, src1_str);
952 } else {
953 shader_addline(buffer, "dot(%s, %s));\n", src0_str, src1_str);
954 }
955 }
956
957 /* Note that this instruction has some restrictions. The destination write mask
958 * can't contain the w component, and the source swizzles have to be .xyzw */
959 void shader_glsl_cross(SHADER_OPCODE_ARG *arg) {
960 char src0_reg[50], src0_mask[6], src0_str[100];
961 char src1_reg[50], src1_mask[6], src1_str[100];
962 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
963 char dst_mask[6];
964
965 shader_glsl_get_write_mask(arg->dst, dst_mask);
966 shader_glsl_append_dst(arg->buffer, arg);
967 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, src0_reg, src0_mask, src0_str);
968 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], src_mask, src1_reg, src1_mask, src1_str);
969 shader_addline(arg->buffer, "cross(%s, %s).%s);\n", src0_str, src1_str, dst_mask);
970 }
971
972 /* Map the opcode 1-to-1 to the GL code (arg->dst = instruction(src0, src1, ...) */
973 void shader_glsl_map2gl(SHADER_OPCODE_ARG* arg) {
974 CONST SHADER_OPCODE* curOpcode = arg->opcode;
975 SHADER_BUFFER* buffer = arg->buffer;
976 const char *instruction;
977 char arguments[256];
978 char src_str[100];
979 char src_reg[50];
980 char src_mask[6];
981 DWORD write_mask;
982 unsigned i;
983
984 /* Determine the GLSL function to use based on the opcode */
985 /* TODO: Possibly make this a table for faster lookups */
986 switch (curOpcode->opcode) {
987 case WINED3DSIO_MIN: instruction = "min"; break;
988 case WINED3DSIO_MAX: instruction = "max"; break;
989 case WINED3DSIO_RSQ: instruction = "inversesqrt"; break;
990 case WINED3DSIO_ABS: instruction = "abs"; break;
991 case WINED3DSIO_FRC: instruction = "fract"; break;
992 case WINED3DSIO_POW: instruction = "pow"; break;
993 case WINED3DSIO_NRM: instruction = "normalize"; break;
994 case WINED3DSIO_LOGP:
995 case WINED3DSIO_LOG: instruction = "log2"; break;
996 case WINED3DSIO_EXP: instruction = "exp2"; break;
997 case WINED3DSIO_SGN: instruction = "sign"; break;
998 default: instruction = "";
999 FIXME("Opcode %s not yet handled in GLSL\n", curOpcode->name);
1000 break;
1001 }
1002
1003 write_mask = shader_glsl_append_dst(buffer, arg);
1004
1005 arguments[0] = '\0';
1006 if (curOpcode->num_params > 0) {
1007 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, src_reg, src_mask, src_str);
1008 strcat(arguments, src_str);
1009 for (i = 2; i < curOpcode->num_params; ++i) {
1010 strcat(arguments, ", ");
1011 shader_glsl_add_src_param(arg, arg->src[i-1], arg->src_addr[i-1], write_mask, src_reg, src_mask, src_str);
1012 strcat(arguments, src_str);
1013 }
1014 }
1015
1016 shader_addline(buffer, "%s(%s));\n", instruction, arguments);
1017 }
1018
1019 /** Process the WINED3DSIO_EXPP instruction in GLSL:
1020 * For shader model 1.x, do the following (and honor the writemask, so use a temporary variable):
1021 * dst.x = 2^(floor(src))
1022 * dst.y = src - floor(src)
1023 * dst.z = 2^src (partial precision is allowed, but optional)
1024 * dst.w = 1.0;
1025 * For 2.0 shaders, just do this (honoring writemask and swizzle):
1026 * dst = 2^src; (partial precision is allowed, but optional)
1027 */
1028 void shader_glsl_expp(SHADER_OPCODE_ARG* arg) {
1029 IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)arg->shader;
1030 char src_str[100];
1031 char src_reg[50];
1032 char src_mask[6];
1033
1034 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, src_reg, src_mask, src_str);
1035
1036 if (shader->baseShader.hex_version < WINED3DPS_VERSION(2,0)) {
1037 char dst_mask[6];
1038
1039 shader_addline(arg->buffer, "tmp0.x = exp2(floor(%s));\n", src_str);
1040 shader_addline(arg->buffer, "tmp0.y = %s - floor(%s);\n", src_str, src_str);
1041 shader_addline(arg->buffer, "tmp0.z = exp2(%s);\n", src_str);
1042 shader_addline(arg->buffer, "tmp0.w = 1.0;\n");
1043
1044 shader_glsl_append_dst(arg->buffer, arg);
1045 shader_glsl_get_write_mask(arg->dst, dst_mask);
1046 shader_addline(arg->buffer, "tmp0%s);\n", dst_mask);
1047 } else {
1048 DWORD write_mask;
1049 size_t mask_size;
1050
1051 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1052 mask_size = shader_glsl_get_write_mask_size(write_mask);
1053
1054 if (mask_size > 1) {
1055 shader_addline(arg->buffer, "vec%d(exp2(%s)));\n", mask_size, src_str);
1056 } else {
1057 shader_addline(arg->buffer, "exp2(%s));\n", src_str);
1058 }
1059 }
1060 }
1061
1062 /** Process the RCP (reciprocal or inverse) opcode in GLSL (dst = 1 / src) */
1063 void shader_glsl_rcp(SHADER_OPCODE_ARG* arg) {
1064 char src_str[100];
1065 char src_reg[50];
1066 char src_mask[6];
1067 DWORD write_mask;
1068 size_t mask_size;
1069
1070 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1071 mask_size = shader_glsl_get_write_mask_size(write_mask);
1072 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, src_reg, src_mask, src_str);
1073
1074 if (mask_size > 1) {
1075 shader_addline(arg->buffer, "vec%d(1.0 / %s));\n", mask_size, src_str);
1076 } else {
1077 shader_addline(arg->buffer, "1.0 / %s);\n", src_str);
1078 }
1079 }
1080
1081 /** Process signed comparison opcodes in GLSL. */
1082 void shader_glsl_compare(SHADER_OPCODE_ARG* arg) {
1083 char src0_str[100], src1_str[100];
1084 char src0_reg[50], src1_reg[50];
1085 char src0_mask[6], src1_mask[6];
1086 DWORD write_mask;
1087 size_t mask_size;
1088
1089 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1090 mask_size = shader_glsl_get_write_mask_size(write_mask);
1091 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, src0_reg, src0_mask, src0_str);
1092 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, src1_reg, src1_mask, src1_str);
1093
1094 if (mask_size > 1) {
1095 const char *compare;
1096
1097 switch(arg->opcode->opcode) {
1098 case WINED3DSIO_SLT: compare = "lessThan"; break;
1099 case WINED3DSIO_SGE: compare = "greaterThanEqual"; break;
1100 default: compare = "";
1101 FIXME("Can't handle opcode %s\n", arg->opcode->name);
1102 }
1103
1104 shader_addline(arg->buffer, "vec%d(%s(%s, %s)));\n", mask_size, compare, src0_str, src1_str);
1105 } else {
1106 const char *compare;
1107
1108 switch(arg->opcode->opcode) {
1109 case WINED3DSIO_SLT: compare = "<"; break;
1110 case WINED3DSIO_SGE: compare = ">="; break;
1111 default: compare = "";
1112 FIXME("Can't handle opcode %s\n", arg->opcode->name);
1113 }
1114
1115 shader_addline(arg->buffer, "(%s %s %s) ? 1.0 : 0.0);\n", src0_str, compare, src1_str);
1116 }
1117 }
1118
1119 /** Process CMP instruction in GLSL (dst = src0 >= 0.0 ? src1 : src2), per channel */
1120 void shader_glsl_cmp(SHADER_OPCODE_ARG* arg) {
1121 char src0_str[100], src1_str[100], src2_str[100];
1122 char src0_reg[50], src1_reg[50], src2_reg[50];
1123 char src0_mask[6], src1_mask[6], src2_mask[6];
1124 DWORD write_mask;
1125 size_t mask_size;
1126
1127 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1128 mask_size = shader_glsl_get_write_mask_size(write_mask);
1129
1130 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, src0_reg, src0_mask, src0_str);
1131 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, src1_reg, src1_mask, src1_str);
1132 shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], write_mask, src2_reg, src2_mask, src2_str);
1133
1134 if (mask_size > 1) {
1135 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);
1136 } else {
1137 shader_addline(arg->buffer, "%s >= 0.0 ? %s : %s);\n", src0_str, src1_str, src2_str);
1138 }
1139 }
1140
1141 /** Process the CND opcode in GLSL (dst = (src0 < 0.5) ? src1 : src2) */
1142 /* For ps 1.1-1.3, only a single component of src0 is used. For ps 1.4
1143 * the compare is done per component of src0. */
1144 void shader_glsl_cnd(SHADER_OPCODE_ARG* arg) {
1145 IWineD3DBaseShaderImpl* shader = (IWineD3DBaseShaderImpl*) arg->shader;
1146 char src0_str[100], src1_str[100], src2_str[100];
1147 char src0_reg[50], src1_reg[50], src2_reg[50];
1148 char src0_mask[6], src1_mask[6], src2_mask[6];
1149 DWORD write_mask;
1150 size_t mask_size;
1151
1152 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1153
1154 if (shader->baseShader.hex_version < WINED3DPS_VERSION(1, 4)) {
1155 mask_size = 1;
1156 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, src0_reg, src0_mask, src0_str);
1157 } else {
1158 mask_size = shader_glsl_get_write_mask_size(write_mask);
1159 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, src0_reg, src0_mask, src0_str);
1160 }
1161
1162 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, src1_reg, src1_mask, src1_str);
1163 shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], write_mask, src2_reg, src2_mask, src2_str);
1164
1165 if (mask_size > 1) {
1166 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);
1167 } else {
1168 shader_addline(arg->buffer, "%s < 0.5 ? %s : %s);\n", src0_str, src1_str, src2_str);
1169 }
1170 }
1171
1172 /** GLSL code generation for WINED3DSIO_MAD: Multiply the first 2 opcodes, then add the last */
1173 void shader_glsl_mad(SHADER_OPCODE_ARG* arg) {
1174 char src0_str[100], src1_str[100], src2_str[100];
1175 char src0_reg[50], src1_reg[50], src2_reg[50];
1176 char src0_mask[6], src1_mask[6], src2_mask[6];
1177 DWORD write_mask;
1178
1179 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1180 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, src0_reg, src0_mask, src0_str);
1181 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, src1_reg, src1_mask, src1_str);
1182 shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], write_mask, src2_reg, src2_mask, src2_str);
1183 shader_addline(arg->buffer, "(%s * %s) + %s);\n", src0_str, src1_str, src2_str);
1184 }
1185
1186 /** Handles transforming all WINED3DSIO_M?x? opcodes for
1187 Vertex shaders to GLSL codes */
1188 void shader_glsl_mnxn(SHADER_OPCODE_ARG* arg) {
1189 int i;
1190 int nComponents = 0;
1191 SHADER_OPCODE_ARG tmpArg;
1192
1193 memset(&tmpArg, 0, sizeof(SHADER_OPCODE_ARG));
1194
1195 /* Set constants for the temporary argument */
1196 tmpArg.shader = arg->shader;
1197 tmpArg.buffer = arg->buffer;
1198 tmpArg.src[0] = arg->src[0];
1199 tmpArg.src_addr[0] = arg->src_addr[0];
1200 tmpArg.src_addr[1] = arg->src_addr[1];
1201 tmpArg.reg_maps = arg->reg_maps;
1202
1203 switch(arg->opcode->opcode) {
1204 case WINED3DSIO_M4x4:
1205 nComponents = 4;
1206 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP4);
1207 break;
1208 case WINED3DSIO_M4x3:
1209 nComponents = 3;
1210 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP4);
1211 break;
1212 case WINED3DSIO_M3x4:
1213 nComponents = 4;
1214 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP3);
1215 break;
1216 case WINED3DSIO_M3x3:
1217 nComponents = 3;
1218 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP3);
1219 break;
1220 case WINED3DSIO_M3x2:
1221 nComponents = 2;
1222 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP3);
1223 break;
1224 default:
1225 break;
1226 }
1227
1228 for (i = 0; i < nComponents; i++) {
1229 tmpArg.dst = ((arg->dst) & ~WINED3DSP_WRITEMASK_ALL)|(WINED3DSP_WRITEMASK_0<<i);
1230 tmpArg.src[1] = arg->src[1]+i;
1231 shader_glsl_dot(&tmpArg);
1232 }
1233 }
1234
1235 /**
1236 The LRP instruction performs a component-wise linear interpolation
1237 between the second and third operands using the first operand as the
1238 blend factor. Equation: (dst = src2 + src0 * (src1 - src2))
1239 This is equivalent to mix(src2, src1, src0);
1240 */
1241 void shader_glsl_lrp(SHADER_OPCODE_ARG* arg) {
1242 char src0_str[100], src1_str[100], src2_str[100];
1243 char src0_reg[50], src1_reg[50], src2_reg[50];
1244 char src0_mask[6], src1_mask[6], src2_mask[6];
1245 DWORD write_mask;
1246
1247 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1248
1249 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, src0_reg, src0_mask, src0_str);
1250 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, src1_reg, src1_mask, src1_str);
1251 shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], write_mask, src2_reg, src2_mask, src2_str);
1252
1253 shader_addline(arg->buffer, "mix(%s, %s, %s));\n", src2_str, src1_str, src0_str);
1254 }
1255
1256 /** Process the WINED3DSIO_LIT instruction in GLSL:
1257 * dst.x = dst.w = 1.0
1258 * dst.y = (src0.x > 0) ? src0.x
1259 * dst.z = (src0.x > 0) ? ((src0.y > 0) ? pow(src0.y, src.w) : 0) : 0
1260 * where src.w is clamped at +- 128
1261 */
1262 void shader_glsl_lit(SHADER_OPCODE_ARG* arg) {
1263 char src0_str[100];
1264 char src0_reg[50];
1265 char src0_mask[6];
1266 char src1_str[100];
1267 char src1_reg[50];
1268 char src1_mask[6];
1269 char src3_str[100];
1270 char src3_reg[50];
1271 char src3_mask[6];
1272 char dst_mask[6];
1273
1274 shader_glsl_append_dst(arg->buffer, arg);
1275 shader_glsl_get_write_mask(arg->dst, dst_mask);
1276
1277 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, src0_reg, src0_mask, src0_str);
1278 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_1, src1_reg, src1_mask, src1_str);
1279 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_3, src3_reg, src3_mask, src3_str);
1280
1281 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",
1282 src0_str, src0_str, src0_str, src1_str, src1_str, src3_str, dst_mask);
1283 }
1284
1285 /** Process the WINED3DSIO_DST instruction in GLSL:
1286 * dst.x = 1.0
1287 * dst.y = src0.x * src0.y
1288 * dst.z = src0.z
1289 * dst.w = src1.w
1290 */
1291 void shader_glsl_dst(SHADER_OPCODE_ARG* arg) {
1292 char src0y_str[100], src0z_str[100], src1y_str[100], src1w_str[100];
1293 char src0y_reg[50], src0z_reg[50], src1y_reg[50], src1w_reg[50];
1294 char dst_mask[6], src0y_mask[6], src0z_mask[6], src1y_mask[6], src1w_mask[6];
1295
1296 shader_glsl_append_dst(arg->buffer, arg);
1297 shader_glsl_get_write_mask(arg->dst, dst_mask);
1298
1299 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_1, src0y_reg, src0y_mask, src0y_str);
1300 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_2, src0z_reg, src0z_mask, src0z_str);
1301 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_1, src1y_reg, src1y_mask, src1y_str);
1302 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_3, src1w_reg, src1w_mask, src1w_str);
1303
1304 shader_addline(arg->buffer, "vec4(1.0, %s * %s, %s, %s))%s;\n",
1305 src0y_str, src1y_str, src0z_str, src1w_str, dst_mask);
1306 }
1307
1308 /** Process the WINED3DSIO_SINCOS instruction in GLSL:
1309 * VS 2.0 requires that specific cosine and sine constants be passed to this instruction so the hardware
1310 * can handle it. But, these functions are built-in for GLSL, so we can just ignore the last 2 params.
1311 *
1312 * dst.x = cos(src0.?)
1313 * dst.y = sin(src0.?)
1314 * dst.z = dst.z
1315 * dst.w = dst.w
1316 */
1317 void shader_glsl_sincos(SHADER_OPCODE_ARG* arg) {
1318 char src0_str[100];
1319 char src0_reg[50];
1320 char src0_mask[6];
1321 DWORD write_mask;
1322
1323 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1324 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, src0_reg, src0_mask, src0_str);
1325
1326 switch (write_mask) {
1327 case WINED3DSP_WRITEMASK_0:
1328 shader_addline(arg->buffer, "cos(%s));\n", src0_str);
1329 break;
1330
1331 case WINED3DSP_WRITEMASK_1:
1332 shader_addline(arg->buffer, "sin(%s));\n", src0_str);
1333 break;
1334
1335 case (WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1):
1336 shader_addline(arg->buffer, "vec2(cos(%s), sin(%s)));\n", src0_str, src0_str);
1337 break;
1338
1339 default:
1340 ERR("Write mask should be .x, .y or .xy\n");
1341 break;
1342 }
1343 }
1344
1345 /** Process the WINED3DSIO_LOOP instruction in GLSL:
1346 * Start a for() loop where src1.y is the initial value of aL,
1347 * increment aL by src1.z for a total of src1.x iterations.
1348 * Need to use a temporary variable for this operation.
1349 */
1350 /* FIXME: I don't think nested loops will work correctly this way. */
1351 void shader_glsl_loop(SHADER_OPCODE_ARG* arg) {
1352
1353 char src1_str[100];
1354 char src1_reg[50];
1355 char src1_mask[6];
1356
1357 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_ALL, src1_reg, src1_mask, src1_str);
1358
1359 shader_addline(arg->buffer, "for (tmpInt = 0, aL = %s.y; tmpInt < %s.x; tmpInt++, aL += %s.z) {\n",
1360 src1_reg, src1_reg, src1_reg);
1361 }
1362
1363 void shader_glsl_end(SHADER_OPCODE_ARG* arg) {
1364 shader_addline(arg->buffer, "}\n");
1365 }
1366
1367 void shader_glsl_rep(SHADER_OPCODE_ARG* arg) {
1368
1369 char src0_str[100];
1370 char src0_reg[50];
1371 char src0_mask[6];
1372
1373 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, src0_reg, src0_mask, src0_str);
1374 shader_addline(arg->buffer, "for (tmpInt = 0; tmpInt < %s; tmpInt++) {\n", src0_str);
1375 }
1376
1377 void shader_glsl_if(SHADER_OPCODE_ARG* arg) {
1378
1379 char src0_str[100];
1380 char src0_reg[50];
1381 char src0_mask[6];
1382
1383 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, src0_reg, src0_mask, src0_str);
1384 shader_addline(arg->buffer, "if (%s) {\n", src0_str);
1385 }
1386
1387 void shader_glsl_ifc(SHADER_OPCODE_ARG* arg) {
1388
1389 char src0_str[100], src1_str[100];
1390 char src0_reg[50], src1_reg[50];
1391 char src0_mask[6], src1_mask[6];
1392
1393 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, src0_reg, src0_mask, src0_str);
1394 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_0, src1_reg, src1_mask, src1_str);
1395
1396 shader_addline(arg->buffer, "if (%s %s %s) {\n",
1397 src0_str, shader_get_comp_op(arg->opcode_token), src1_str);
1398 }
1399
1400 void shader_glsl_else(SHADER_OPCODE_ARG* arg) {
1401 shader_addline(arg->buffer, "} else {\n");
1402 }
1403
1404 void shader_glsl_break(SHADER_OPCODE_ARG* arg) {
1405 shader_addline(arg->buffer, "break;\n");
1406 }
1407
1408 /* FIXME: According to MSDN the compare is done per component. */
1409 void shader_glsl_breakc(SHADER_OPCODE_ARG* arg) {
1410
1411 char src0_str[100], src1_str[100];
1412 char src0_reg[50], src1_reg[50];
1413 char src0_mask[6], src1_mask[6];
1414
1415 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, src0_reg, src0_mask, src0_str);
1416 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_0, src1_reg, src1_mask, src1_str);
1417
1418 shader_addline(arg->buffer, "if (%s %s %s) break;\n",
1419 src0_str, shader_get_comp_op(arg->opcode_token), src1_str);
1420 }
1421
1422 void shader_glsl_label(SHADER_OPCODE_ARG* arg) {
1423
1424 DWORD snum = (arg->src[0]) & WINED3DSP_REGNUM_MASK;
1425 shader_addline(arg->buffer, "}\n");
1426 shader_addline(arg->buffer, "void subroutine%lu () {\n", snum);
1427 }
1428
1429 void shader_glsl_call(SHADER_OPCODE_ARG* arg) {
1430 DWORD snum = (arg->src[0]) & WINED3DSP_REGNUM_MASK;
1431 shader_addline(arg->buffer, "subroutine%lu();\n", snum);
1432 }
1433
1434 void shader_glsl_callnz(SHADER_OPCODE_ARG* arg) {
1435
1436 char src1_str[100];
1437 char src1_reg[50];
1438 char src1_mask[6];
1439
1440 DWORD snum = (arg->src[0]) & WINED3DSP_REGNUM_MASK;
1441 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_0, src1_reg, src1_mask, src1_str);
1442 shader_addline(arg->buffer, "if (%s) subroutine%lu();\n", src1_str, snum);
1443 }
1444
1445 /*********************************************
1446 * Pixel Shader Specific Code begins here
1447 ********************************************/
1448 void pshader_glsl_tex(SHADER_OPCODE_ARG* arg) {
1449 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1450 DWORD hex_version = This->baseShader.hex_version;
1451 char coord_str[100], coord_reg[50], coord_mask[6];
1452 char dst_swizzle[6];
1453 glsl_sample_function_t sample_function;
1454 DWORD sampler_type;
1455 DWORD sampler_idx;
1456 BOOL projected;
1457 DWORD mask = 0;
1458
1459 /* All versions have a destination register */
1460 shader_glsl_append_dst(arg->buffer, arg);
1461
1462 /* 1.0-1.4: Use destination register as sampler source.
1463 * 2.0+: Use provided sampler source. */
1464 if (hex_version < WINED3DPS_VERSION(1,4)) {
1465 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1466 DWORD flags;
1467
1468 sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1469 flags = deviceImpl->stateBlock->textureState[sampler_idx][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1470
1471 if (flags & WINED3DTTFF_PROJECTED) {
1472 projected = TRUE;
1473 switch (flags & ~WINED3DTTFF_PROJECTED) {
1474 case WINED3DTTFF_COUNT1: FIXME("WINED3DTTFF_PROJECTED with WINED3DTTFF_COUNT1?\n"); break;
1475 case WINED3DTTFF_COUNT2: mask = WINED3DSP_WRITEMASK_1; break;
1476 case WINED3DTTFF_COUNT3: mask = WINED3DSP_WRITEMASK_2; break;
1477 case WINED3DTTFF_COUNT4:
1478 case WINED3DTTFF_DISABLE: mask = WINED3DSP_WRITEMASK_3; break;
1479 }
1480 } else {
1481 projected = FALSE;
1482 }
1483 } else if (hex_version < WINED3DPS_VERSION(2,0)) {
1484 DWORD src_mod = arg->src[0] & WINED3DSP_SRCMOD_MASK;
1485 sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1486
1487 if (src_mod == WINED3DSPSM_DZ) {
1488 projected = TRUE;
1489 mask = WINED3DSP_WRITEMASK_2;
1490 } else if (src_mod == WINED3DSPSM_DW) {
1491 projected = TRUE;
1492 mask = WINED3DSP_WRITEMASK_3;
1493 } else {
1494 projected = FALSE;
1495 }
1496 } else {
1497 sampler_idx = arg->src[1] & WINED3DSP_REGNUM_MASK;
1498 /* TODO: Handle D3DSI_TEXLD_PROJECTED... */
1499 projected = FALSE;
1500 }
1501
1502 sampler_type = arg->reg_maps->samplers[sampler_idx] & WINED3DSP_TEXTURETYPE_MASK;
1503 shader_glsl_get_sample_function(sampler_type, projected, &sample_function);
1504 mask |= sample_function.coord_mask;
1505
1506 if (hex_version < WINED3DPS_VERSION(2,0)) {
1507 shader_glsl_get_write_mask(arg->dst, dst_swizzle);
1508 } else {
1509 shader_glsl_get_swizzle(arg->src[1], FALSE, arg->dst, dst_swizzle);
1510 }
1511
1512 /* 1.0-1.3: Use destination register as coordinate source.
1513 1.4+: Use provided coordinate source register. */
1514 if (hex_version < WINED3DPS_VERSION(1,4)) {
1515 shader_glsl_get_write_mask(mask, coord_mask);
1516 shader_addline(arg->buffer, "%s(Psampler%u, T%u%s)%s);\n",
1517 sample_function.name, sampler_idx, sampler_idx, coord_mask, dst_swizzle);
1518 } else {
1519 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], mask, coord_reg, coord_mask, coord_str);
1520 shader_addline(arg->buffer, "%s(Psampler%u, %s)%s);\n",
1521 sample_function.name, sampler_idx, coord_str, dst_swizzle);
1522 }
1523 }
1524
1525 void pshader_glsl_texcoord(SHADER_OPCODE_ARG* arg) {
1526
1527 /* FIXME: Make this work for more than just 2D textures */
1528
1529 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1530 SHADER_BUFFER* buffer = arg->buffer;
1531 DWORD hex_version = This->baseShader.hex_version;
1532 DWORD write_mask;
1533 char dst_mask[6];
1534
1535 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1536 shader_glsl_get_write_mask(write_mask, dst_mask);
1537
1538 if (hex_version != WINED3DPS_VERSION(1,4)) {
1539 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1540 shader_addline(buffer, "clamp(gl_TexCoord[%u], 0.0, 1.0)%s);\n", reg, dst_mask);
1541 } else {
1542 DWORD reg = arg->src[0] & WINED3DSP_REGNUM_MASK;
1543 DWORD src_mod = arg->src[0] & WINED3DSP_SRCMOD_MASK;
1544 char dst_swizzle[6];
1545
1546 shader_glsl_get_swizzle(arg->src[0], FALSE, write_mask, dst_swizzle);
1547
1548 if (src_mod == WINED3DSPSM_DZ) {
1549 char div_str[100], div_reg[50], div_mask[6];
1550 size_t mask_size = shader_glsl_get_write_mask_size(write_mask);
1551 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_2, div_reg, div_mask, div_str);
1552
1553 if (mask_size > 1) {
1554 shader_addline(buffer, "gl_TexCoord[%u]%s / vec%d(%s));\n", reg, dst_swizzle, mask_size, div_str);
1555 } else {
1556 shader_addline(buffer, "gl_TexCoord[%u]%s / %s);\n", reg, dst_swizzle, div_str);
1557 }
1558 } else if (src_mod == WINED3DSPSM_DW) {
1559 char div_str[100], div_reg[50], div_mask[6];
1560 size_t mask_size = shader_glsl_get_write_mask_size(write_mask);
1561 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_3, div_reg, div_mask, div_str);
1562
1563 if (mask_size > 1) {
1564 shader_addline(buffer, "gl_TexCoord[%u]%s / vec%d(%s));\n", reg, dst_swizzle, mask_size, div_str);
1565 } else {
1566 shader_addline(buffer, "gl_TexCoord[%u]%s / %s);\n", reg, dst_swizzle, div_str);
1567 }
1568 } else {
1569 shader_addline(buffer, "gl_TexCoord[%u]%s);\n", reg, dst_swizzle);
1570 }
1571 }
1572 }
1573
1574 /** Process the WINED3DSIO_TEXDP3TEX instruction in GLSL:
1575 * Take a 3-component dot product of the TexCoord[dstreg] and src,
1576 * then perform a 1D texture lookup from stage dstregnum, place into dst. */
1577 void pshader_glsl_texdp3tex(SHADER_OPCODE_ARG* arg) {
1578 char src0_str[100];
1579 char src0_name[50];
1580 char src0_mask[6];
1581 char dst_mask[6];
1582 DWORD sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1583 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1584
1585 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, src0_name, src0_mask, src0_str);
1586
1587 shader_glsl_append_dst(arg->buffer, arg);
1588 shader_glsl_get_write_mask(arg->dst, dst_mask);
1589 shader_addline(arg->buffer, "texture1D(Psampler%u, dot(gl_TexCoord[%u].xyz, %s))%s);\n",
1590 sampler_idx, sampler_idx, src0_str, dst_mask);
1591 }
1592
1593 /** Process the WINED3DSIO_TEXDP3 instruction in GLSL:
1594 * Take a 3-component dot product of the TexCoord[dstreg] and src. */
1595 void pshader_glsl_texdp3(SHADER_OPCODE_ARG* arg) {
1596 char src0_str[100];
1597 char src0_name[50];
1598 char src0_mask[6];
1599 DWORD dstreg = arg->dst & WINED3DSP_REGNUM_MASK;
1600 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1601 DWORD dst_mask;
1602 size_t mask_size;
1603
1604 dst_mask = shader_glsl_append_dst(arg->buffer, arg);
1605 mask_size = shader_glsl_get_write_mask_size(dst_mask);
1606 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, src0_name, src0_mask, src0_str);
1607
1608 if (mask_size > 1) {
1609 shader_addline(arg->buffer, "vec%d(dot(T%u.xyz, %s)));\n", mask_size, dstreg, src0_str);
1610 } else {
1611 shader_addline(arg->buffer, "dot(T%u.xyz, %s));\n", dstreg, src0_str);
1612 }
1613 }
1614
1615 /** Process the WINED3DSIO_TEXDEPTH instruction in GLSL:
1616 * Calculate the depth as dst.x / dst.y */
1617 void pshader_glsl_texdepth(SHADER_OPCODE_ARG* arg) {
1618
1619 char dst_str[100];
1620 char dst_reg[50];
1621 char dst_mask[6];
1622
1623 shader_glsl_add_dst_param(arg, arg->dst, 0, dst_reg, dst_mask, dst_str);
1624
1625 shader_addline(arg->buffer, "gl_FragDepth = %s.x / %s.y;\n", dst_reg, dst_reg);
1626 }
1627
1628 /** Process the WINED3DSIO_TEXM3X2DEPTH instruction in GLSL:
1629 * Last row of a 3x2 matrix multiply, use the result to calculate the depth:
1630 * Calculate tmp0.y = TexCoord[dstreg] . src.xyz; (tmp0.x has already been calculated)
1631 * depth = (tmp0.y == 0.0) ? 1.0 : tmp0.x / tmp0.y
1632 */
1633 void pshader_glsl_texm3x2depth(SHADER_OPCODE_ARG* arg) {
1634 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1635 DWORD dstreg = arg->dst & WINED3DSP_REGNUM_MASK;
1636 char src0_str[100], src0_name[50], src0_mask[6];
1637
1638 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, src0_name, src0_mask, src0_str);
1639
1640 shader_addline(arg->buffer, "tmp0.y = dot(T%u.xyz, %s);\n", dstreg, src0_str);
1641 shader_addline(arg->buffer, "gl_FragDepth = (tmp0.y == 0.0) ? 1.0 : clamp(tmp0.x / tmp0.y, 0.0, 1.0);\n");
1642 }
1643
1644 /** Process the WINED3DSIO_TEXM3X2PAD instruction in GLSL
1645 * Calculate the 1st of a 2-row matrix multiplication. */
1646 void pshader_glsl_texm3x2pad(SHADER_OPCODE_ARG* arg) {
1647 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1648 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1649 SHADER_BUFFER* buffer = arg->buffer;
1650 char src0_str[100];
1651 char src0_name[50];
1652 char src0_mask[6];
1653
1654 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, src0_name, src0_mask, src0_str);
1655 shader_addline(buffer, "tmp0.x = dot(T%u.xyz, %s);\n", reg, src0_str);
1656 }
1657
1658 /** Process the WINED3DSIO_TEXM3X3PAD instruction in GLSL
1659 * Calculate the 1st or 2nd row of a 3-row matrix multiplication. */
1660 void pshader_glsl_texm3x3pad(SHADER_OPCODE_ARG* arg) {
1661
1662 IWineD3DPixelShaderImpl* shader = (IWineD3DPixelShaderImpl*) arg->shader;
1663 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1664 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1665 SHADER_BUFFER* buffer = arg->buffer;
1666 SHADER_PARSE_STATE* current_state = &shader->baseShader.parse_state;
1667 char src0_str[100];
1668 char src0_name[50];
1669 char src0_mask[6];
1670
1671 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, src0_name, src0_mask, src0_str);
1672 shader_addline(buffer, "tmp0.%c = dot(T%u.xyz, %s);\n", 'x' + current_state->current_row, reg, src0_str);
1673 current_state->texcoord_w[current_state->current_row++] = reg;
1674 }
1675
1676 void pshader_glsl_texm3x2tex(SHADER_OPCODE_ARG* arg) {
1677 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1678 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1679 SHADER_BUFFER* buffer = arg->buffer;
1680 char src0_str[100];
1681 char src0_name[50];
1682 char src0_mask[6];
1683 char dst_mask[6];
1684
1685 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, src0_name, src0_mask, src0_str);
1686 shader_addline(buffer, "tmp0.y = dot(T%u.xyz, %s);\n", reg, src0_str);
1687
1688 shader_glsl_append_dst(buffer, arg);
1689 shader_glsl_get_write_mask(arg->dst, dst_mask);
1690
1691 /* Sample the texture using the calculated coordinates */
1692 shader_addline(buffer, "texture2D(Psampler%u, tmp0.xy)%s);\n", reg, dst_mask);
1693 }
1694
1695 /** Process the WINED3DSIO_TEXM3X3TEX instruction in GLSL
1696 * Perform the 3rd row of a 3x3 matrix multiply, then sample the texture using the calculated coordinates */
1697 void pshader_glsl_texm3x3tex(SHADER_OPCODE_ARG* arg) {
1698 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1699 char src0_str[100];
1700 char src0_name[50];
1701 char src0_mask[6];
1702 char dst_mask[6];
1703 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1704 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1705 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1706 DWORD sampler_type = arg->reg_maps->samplers[reg] & WINED3DSP_TEXTURETYPE_MASK;
1707 glsl_sample_function_t sample_function;
1708
1709 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, src0_name, src0_mask, src0_str);
1710 shader_addline(arg->buffer, "tmp0.z = dot(T%u.xyz, %s);\n", reg, src0_str);
1711
1712 shader_glsl_append_dst(arg->buffer, arg);
1713 shader_glsl_get_write_mask(arg->dst, dst_mask);
1714 shader_glsl_get_sample_function(sampler_type, FALSE, &sample_function);
1715
1716 /* Sample the texture using the calculated coordinates */
1717 shader_addline(arg->buffer, "%s(Psampler%u, tmp0.xyz)%s);\n", sample_function.name, reg, dst_mask);
1718
1719 current_state->current_row = 0;
1720 }
1721
1722 /** Process the WINED3DSIO_TEXM3X3 instruction in GLSL
1723 * Perform the 3rd row of a 3x3 matrix multiply */
1724 void pshader_glsl_texm3x3(SHADER_OPCODE_ARG* arg) {
1725 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1726 char src0_str[100];
1727 char src0_name[50];
1728 char src0_mask[6];
1729 char dst_mask[6];
1730 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1731 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1732 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1733
1734 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, src0_name, src0_mask, src0_str);
1735
1736 shader_glsl_append_dst(arg->buffer, arg);
1737 shader_glsl_get_write_mask(arg->dst, dst_mask);
1738 shader_addline(arg->buffer, "vec4(tmp.xy, dot(T%u.xyz, %s), 1.0)%s);\n", reg, src0_str, dst_mask);
1739
1740 current_state->current_row = 0;
1741 }
1742
1743 /** Process the WINED3DSIO_TEXM3X3SPEC instruction in GLSL
1744 * Peform the final texture lookup based on the previous 2 3x3 matrix multiplies */
1745 void pshader_glsl_texm3x3spec(SHADER_OPCODE_ARG* arg) {
1746
1747 IWineD3DPixelShaderImpl* shader = (IWineD3DPixelShaderImpl*) arg->shader;
1748 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1749 char src0_str[100], src0_name[50], src0_mask[6];
1750 char src1_str[100], src1_name[50], src1_mask[6];
1751 char dst_mask[6];
1752 SHADER_BUFFER* buffer = arg->buffer;
1753 SHADER_PARSE_STATE* current_state = &shader->baseShader.parse_state;
1754 DWORD stype = arg->reg_maps->samplers[reg] & WINED3DSP_TEXTURETYPE_MASK;
1755 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1756 glsl_sample_function_t sample_function;
1757
1758 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, src0_name, src0_mask, src0_str);
1759 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], src_mask, src1_name, src1_mask, src1_str);
1760
1761 /* Perform the last matrix multiply operation */
1762 shader_addline(buffer, "tmp0.z = dot(T%u.xyz, %s);\n", reg, src0_str);
1763
1764 /* Calculate reflection vector, 2*(tmp0.src1)*tmp0-src1
1765 * This is equavalent to reflect(-src1, tmp0); */
1766 shader_addline(buffer, "tmp0.xyz = reflect(-(%s), tmp0.xyz);\n", src1_str);
1767
1768 shader_glsl_append_dst(buffer, arg);
1769 shader_glsl_get_write_mask(arg->dst, dst_mask);
1770 shader_glsl_get_sample_function(stype, FALSE, &sample_function);
1771
1772 /* Sample the texture */
1773 shader_addline(buffer, "%s(Psampler%u, tmp0.xyz)%s);\n", sample_function.name, reg, dst_mask);
1774
1775 current_state->current_row = 0;
1776 }
1777
1778 /** Process the WINED3DSIO_TEXM3X3VSPEC instruction in GLSL
1779 * Peform the final texture lookup based on the previous 2 3x3 matrix multiplies */
1780 void pshader_glsl_texm3x3vspec(SHADER_OPCODE_ARG* arg) {
1781
1782 IWineD3DPixelShaderImpl* shader = (IWineD3DPixelShaderImpl*) arg->shader;
1783 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1784 SHADER_BUFFER* buffer = arg->buffer;
1785 SHADER_PARSE_STATE* current_state = &shader->baseShader.parse_state;
1786 char src0_str[100], src0_name[50], src0_mask[6], dst_mask[6];
1787 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1788 DWORD sampler_type = arg->reg_maps->samplers[reg] & WINED3DSP_TEXTURETYPE_MASK;
1789 glsl_sample_function_t sample_function;
1790
1791 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, src0_name, src0_mask, src0_str);
1792
1793 /* Perform the last matrix multiply operation */
1794 shader_addline(buffer, "tmp0.z = dot(vec3(T%u), vec3(%s));\n", reg, src0_str);
1795
1796 /* Construct the eye-ray vector from w coordinates */
1797 shader_addline(buffer, "tmp1.xyz = normalize(vec3(gl_TexCoord[%u].w, gl_TexCoord[%u].w, gl_TexCoord[%u].w));\n",
1798 current_state->texcoord_w[0], current_state->texcoord_w[1], reg);
1799
1800 /* Calculate reflection vector (Assume normal is normalized): RF = 2*(tmp0.tmp1)*tmp0-tmp1
1801 * This is equavalent to reflect(-tmp1, tmp0); */
1802 shader_addline(buffer, "tmp0.xyz = reflect(-tmp1.xyz, tmp0.xyz);\n");
1803
1804 shader_glsl_append_dst(buffer, arg);
1805 shader_glsl_get_write_mask(arg->dst, dst_mask);
1806 shader_glsl_get_sample_function(sampler_type, FALSE, &sample_function);
1807
1808 /* Sample the texture using the calculated coordinates */
1809 shader_addline(buffer, "%s(Psampler%u, tmp0.xyz)%s);\n", sample_function.name, reg, dst_mask);
1810
1811 current_state->current_row = 0;
1812 }
1813
1814 /** Process the WINED3DSIO_TEXBEM instruction in GLSL.
1815 * Apply a fake bump map transform.
1816 * FIXME: Should apply the BUMPMAPENV matrix. For now, just sample the texture */
1817 void pshader_glsl_texbem(SHADER_OPCODE_ARG* arg) {
1818
1819 DWORD reg1 = arg->dst & WINED3DSP_REGNUM_MASK;
1820 DWORD reg2 = arg->src[0] & WINED3DSP_REGNUM_MASK;
1821
1822 FIXME("Not applying the BUMPMAPENV matrix for pixel shader instruction texbem.\n");
1823 shader_addline(arg->buffer, "T%u = texture2D(Psampler%u, gl_TexCoord[%u].xy + T%u.xy);\n",
1824 reg1, reg1, reg1, reg2);
1825 }
1826
1827 /** Process the WINED3DSIO_TEXREG2AR instruction in GLSL
1828 * Sample 2D texture at dst using the alpha & red (wx) components of src as texture coordinates */
1829 void pshader_glsl_texreg2ar(SHADER_OPCODE_ARG* arg) {
1830 char src0_str[100];
1831 char src0_reg[50];
1832 char src0_mask[6];
1833 DWORD sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1834 char dst_mask[6];
1835
1836 shader_glsl_append_dst(arg->buffer, arg);
1837 shader_glsl_get_write_mask(arg->dst, dst_mask);
1838 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_ALL, src0_reg, src0_mask, src0_str);
1839
1840 shader_addline(arg->buffer, "texture2D(Psampler%u, %s.wx)%s);\n", sampler_idx, src0_reg, dst_mask);
1841 }
1842
1843 /** Process the WINED3DSIO_TEXREG2GB instruction in GLSL
1844 * Sample 2D texture at dst using the green & blue (yz) components of src as texture coordinates */
1845 void pshader_glsl_texreg2gb(SHADER_OPCODE_ARG* arg) {
1846 char src0_str[100];
1847 char src0_reg[50];
1848 char src0_mask[6];
1849 DWORD sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1850 char dst_mask[6];
1851
1852 shader_glsl_append_dst(arg->buffer, arg);
1853 shader_glsl_get_write_mask(arg->dst, dst_mask);
1854 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_ALL, src0_reg, src0_mask, src0_str);
1855
1856 shader_addline(arg->buffer, "texture2D(Psampler%u, %s.yz)%s);\n", sampler_idx, src0_reg, dst_mask);
1857 }
1858
1859 /** Process the WINED3DSIO_TEXREG2RGB instruction in GLSL
1860 * Sample texture at dst using the rgb (xyz) components of src as texture coordinates */
1861 void pshader_glsl_texreg2rgb(SHADER_OPCODE_ARG* arg) {
1862 char src0_str[100];
1863 char src0_reg[50];
1864 char src0_mask[6];
1865 char dst_mask[6];
1866 DWORD sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1867 DWORD sampler_type = arg->reg_maps->samplers[sampler_idx] & WINED3DSP_TEXTURETYPE_MASK;
1868 glsl_sample_function_t sample_function;
1869
1870 shader_glsl_append_dst(arg->buffer, arg);
1871 shader_glsl_get_write_mask(arg->dst, dst_mask);
1872 shader_glsl_get_sample_function(sampler_type, FALSE, &sample_function);
1873 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], sample_function.coord_mask, src0_reg, src0_mask, src0_str);
1874
1875 shader_addline(arg->buffer, "%s(Psampler%u, %s)%s);\n", sample_function.name, sampler_idx, src0_str, dst_mask);
1876 }
1877
1878 /** Process the WINED3DSIO_TEXKILL instruction in GLSL.
1879 * If any of the first 3 components are < 0, discard this pixel */
1880 void pshader_glsl_texkill(SHADER_OPCODE_ARG* arg) {
1881
1882 char dst_str[100], dst_name[50], dst_mask[6];
1883
1884 shader_glsl_add_dst_param(arg, arg->dst, 0, dst_name, dst_mask, dst_str);
1885 shader_addline(arg->buffer, "if (any(lessThan(%s.xyz, vec3(0.0)))) discard;\n", dst_name);
1886 }
1887
1888 /** Process the WINED3DSIO_DP2ADD instruction in GLSL.
1889 * dst = dot2(src0, src1) + src2 */
1890 void pshader_glsl_dp2add(SHADER_OPCODE_ARG* arg) {
1891 char src0_str[100], src1_str[100], src2_str[100];
1892 char src0_reg[50], src1_reg[50], src2_reg[50];
1893 char src0_mask[6], src1_mask[6], src2_mask[6];
1894 DWORD write_mask;
1895 size_t mask_size;
1896
1897 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1898 mask_size = shader_glsl_get_write_mask_size(write_mask);
1899
1900 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);
1901 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);
1902 shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], WINED3DSP_WRITEMASK_0, src2_reg, src2_mask, src2_str);
1903
1904 shader_addline(arg->buffer, "dot(%s, %s) + %s);\n", src0_str, src1_str, src2_str);
1905 }
1906
1907 void pshader_glsl_input_pack(
1908 SHADER_BUFFER* buffer,
1909 semantic* semantics_in) {
1910
1911 unsigned int i;
1912
1913 for (i = 0; i < MAX_REG_INPUT; i++) {
1914
1915 DWORD usage_token = semantics_in[i].usage;
1916 DWORD register_token = semantics_in[i].reg;
1917 DWORD usage, usage_idx;
1918 char reg_mask[6];
1919
1920 /* Uninitialized */
1921 if (!usage_token) continue;
1922 usage = (usage_token & WINED3DSP_DCL_USAGE_MASK) >> WINED3DSP_DCL_USAGE_SHIFT;
1923 usage_idx = (usage_token & WINED3DSP_DCL_USAGEINDEX_MASK) >> WINED3DSP_DCL_USAGEINDEX_SHIFT;
1924 shader_glsl_get_write_mask(register_token, reg_mask);
1925
1926 switch(usage) {
1927
1928 case D3DDECLUSAGE_COLOR:
1929 if (usage_idx == 0)
1930 shader_addline(buffer, "IN%u%s = vec4(gl_Color)%s;\n",
1931 i, reg_mask, reg_mask);
1932 else if (usage_idx == 1)
1933 shader_addline(buffer, "IN%u%s = vec4(gl_SecondaryColor)%s;\n",
1934 i, reg_mask, reg_mask);
1935 else
1936 shader_addline(buffer, "IN%u%s = vec4(unsupported_color_input)%s;\n",
1937 i, reg_mask, reg_mask);
1938 break;
1939
1940 case D3DDECLUSAGE_TEXCOORD:
1941 shader_addline(buffer, "IN%u%s = vec4(gl_TexCoord[%u])%s;\n",
1942 i, reg_mask, usage_idx, reg_mask );
1943 break;
1944
1945 case D3DDECLUSAGE_FOG:
1946 shader_addline(buffer, "IN%u%s = vec4(gl_FogFragCoord)%s;\n",
1947 i, reg_mask, reg_mask);
1948 break;
1949
1950 default:
1951 shader_addline(buffer, "IN%u%s = vec4(unsupported_input)%s;\n",
1952 i, reg_mask, reg_mask);
1953 }
1954 }
1955 }
1956
1957 /*********************************************
1958 * Vertex Shader Specific Code begins here
1959 ********************************************/
1960
1961 void vshader_glsl_output_unpack(
1962 SHADER_BUFFER* buffer,
1963 semantic* semantics_out) {
1964
1965 unsigned int i;
1966
1967 for (i = 0; i < MAX_REG_OUTPUT; i++) {
1968
1969 DWORD usage_token = semantics_out[i].usage;
1970 DWORD register_token = semantics_out[i].reg;
1971 DWORD usage, usage_idx;
1972 char reg_mask[6];
1973
1974 /* Uninitialized */
1975 if (!usage_token) continue;
1976
1977 usage = (usage_token & WINED3DSP_DCL_USAGE_MASK) >> WINED3DSP_DCL_USAGE_SHIFT;
1978 usage_idx = (usage_token & WINED3DSP_DCL_USAGEINDEX_MASK) >> WINED3DSP_DCL_USAGEINDEX_SHIFT;
1979 shader_glsl_get_write_mask(register_token, reg_mask);
1980
1981 switch(usage) {
1982
1983 case D3DDECLUSAGE_COLOR:
1984 if (usage_idx == 0)
1985 shader_addline(buffer, "gl_FrontColor%s = OUT%u%s;\n", reg_mask, i, reg_mask);
1986 else if (usage_idx == 1)
1987 shader_addline(buffer, "gl_FrontSecondaryColor%s = OUT%u%s;\n", reg_mask, i, reg_mask);
1988 else
1989 shader_addline(buffer, "unsupported_color_output%s = OUT%u%s;\n", reg_mask, i, reg_mask);
1990 break;
1991
1992 case D3DDECLUSAGE_POSITION:
1993 shader_addline(buffer, "gl_Position%s = OUT%u%s;\n", reg_mask, i, reg_mask);
1994 break;
1995
1996 case D3DDECLUSAGE_TEXCOORD:
1997 shader_addline(buffer, "gl_TexCoord[%u]%s = OUT%u%s;\n",
1998 usage_idx, reg_mask, i, reg_mask);
1999 break;
2000
2001 case WINED3DSHADERDECLUSAGE_PSIZE:
2002 shader_addline(buffer, "gl_PointSize = OUT%u.x;\n", i);
2003 break;
2004
2005 case WINED3DSHADERDECLUSAGE_FOG:
2006 shader_addline(buffer, "gl_FogFragCoord%s = OUT%u%s;\n", reg_mask, i, reg_mask);
2007 break;
2008
2009 default:
2010 shader_addline(buffer, "unsupported_output%s = OUT%u%s;\n", reg_mask, i, reg_mask);
2011 }
2012 }
2013 }
2014
2015 /** Attach a GLSL pixel or vertex shader object to the shader program */
2016 static void attach_glsl_shader(IWineD3DDevice *iface, IWineD3DBaseShader* shader) {
2017 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
2018 WineD3D_GL_Info *gl_info = &((IWineD3DImpl *)(This->wineD3D))->gl_info;
2019 GLhandleARB shaderObj = ((IWineD3DBaseShaderImpl*)shader)->baseShader.prgId;
2020 if (This->stateBlock->glsl_program && shaderObj != 0) {
2021 TRACE("Attaching GLSL shader object %u to program %u\n", shaderObj, This->stateBlock->glsl_program->programId);
2022 GL_EXTCALL(glAttachObjectARB(This->stateBlock->glsl_program->programId, shaderObj));
2023 checkGLcall("glAttachObjectARB");
2024 }
2025 }
2026
2027 /** Sets the GLSL program ID for the given pixel and vertex shader combination.
2028 * It sets the programId on the current StateBlock (because it should be called
2029 * inside of the DrawPrimitive() part of the render loop).
2030 *
2031 * If a program for the given combination does not exist, create one, and store
2032 * the program in the list. If it creates a program, it will link the given
2033 * objects, too.
2034 *
2035 * We keep the shader programs around on a list because linking
2036 * shader objects together is an expensive operation. It's much
2037 * faster to loop through a list of pre-compiled & linked programs
2038 * each time that the application sets a new pixel or vertex shader
2039 * than it is to re-link them together at that time.
2040 *
2041 * The list will be deleted in IWineD3DDevice::Release().
2042 */
2043 static void set_glsl_shader_program(IWineD3DDevice *iface) {
2044 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
2045 WineD3D_GL_Info *gl_info = &((IWineD3DImpl *)(This->wineD3D))->gl_info;
2046 IWineD3DPixelShader *pshader = This->stateBlock->pixelShader;
2047 IWineD3DVertexShader *vshader = This->stateBlock->vertexShader;
2048 struct glsl_shader_prog_link *curLink = NULL;
2049 struct glsl_shader_prog_link *newLink = NULL;
2050 struct list *ptr = NULL;
2051 GLhandleARB programId = 0;
2052 int i;
2053 char glsl_name[8];
2054
2055 ptr = list_head( &This->glsl_shader_progs );
2056 while (ptr) {
2057 /* At least one program exists - see if it matches our ps/vs combination */
2058 curLink = LIST_ENTRY( ptr, struct glsl_shader_prog_link, entry );
2059 if (vshader == curLink->vertexShader && pshader == curLink->pixelShader) {
2060 /* Existing Program found, use it */
2061 TRACE("Found existing program (%u) for this vertex/pixel shader combination\n",
2062 curLink->programId);
2063 This->stateBlock->glsl_program = curLink;
2064 return;
2065 }
2066 /* This isn't the entry we need - try the next one */
2067 ptr = list_next( &This->glsl_shader_progs, ptr );
2068 }
2069
2070 /* If we get to this point, then no matching program exists, so we create one */
2071 programId = GL_EXTCALL(glCreateProgramObjectARB());
2072 TRACE("Created new GLSL shader program %u\n", programId);
2073
2074 /* Allocate a new link for the list of programs */
2075 newLink = HeapAlloc(GetProcessHeap(), 0, sizeof(struct glsl_shader_prog_link));
2076 newLink->programId = programId;
2077 This->stateBlock->glsl_program = newLink;
2078
2079 /* Attach GLSL vshader */
2080 if (NULL != vshader && This->vs_selected_mode == SHADER_GLSL) {
2081 int i;
2082 int max_attribs = 16; /* TODO: Will this always be the case? It is at the moment... */
2083 char tmp_name[10];
2084
2085 TRACE("Attaching vertex shader to GLSL program\n");
2086 attach_glsl_shader(iface, (IWineD3DBaseShader*)vshader);
2087
2088 /* Bind vertex attributes to a corresponding index number to match
2089 * the same index numbers as ARB_vertex_programs (makes loading
2090 * vertex attributes simpler). With this method, we can use the
2091 * exact same code to load the attributes later for both ARB and
2092 * GLSL shaders.
2093 *
2094 * We have to do this here because we need to know the Program ID
2095 * in order to make the bindings work, and it has to be done prior
2096 * to linking the GLSL program. */
2097 for (i = 0; i < max_attribs; ++i) {
2098 snprintf(tmp_name, sizeof(tmp_name), "attrib%i", i);
2099 GL_EXTCALL(glBindAttribLocationARB(programId, i, tmp_name));
2100 }
2101 checkGLcall("glBindAttribLocationARB");
2102 newLink->vertexShader = vshader;
2103 }
2104
2105 /* Attach GLSL pshader */
2106 if (NULL != pshader && This->ps_selected_mode == SHADER_GLSL) {
2107 TRACE("Attaching pixel shader to GLSL program\n");
2108 attach_glsl_shader(iface, (IWineD3DBaseShader*)pshader);
2109 newLink->pixelShader = pshader;
2110 }
2111
2112 /* Link the program */
2113 TRACE("Linking GLSL shader program %u\n", programId);
2114 GL_EXTCALL(glLinkProgramARB(programId));
2115 print_glsl_info_log(&GLINFO_LOCATION, programId);
2116 list_add_head( &This->glsl_shader_progs, &newLink->entry);
2117
2118 newLink->vuniformF_locations = HeapAlloc(GetProcessHeap(), 0, sizeof(GLhandleARB) * GL_LIMITS(vshader_constantsF));
2119 for (i = 0; i < GL_LIMITS(vshader_constantsF); ++i) {
2120 snprintf(glsl_name, sizeof(glsl_name), "VC[%i]", i);
2121 newLink->vuniformF_locations[i] = GL_EXTCALL(glGetUniformLocationARB(programId, glsl_name));
2122 }
2123 newLink->puniformF_locations = HeapAlloc(GetProcessHeap(), 0, sizeof(GLhandleARB) * GL_LIMITS(pshader_constantsF));
2124 for (i = 0; i < GL_LIMITS(pshader_constantsF); ++i) {
2125 snprintf(glsl_name, sizeof(glsl_name), "PC[%i]", i);
2126 newLink->puniformF_locations[i] = GL_EXTCALL(glGetUniformLocationARB(programId, glsl_name));
2127 }
2128
2129 return;
2130 }
2131
2132 static GLhandleARB create_glsl_blt_shader(WineD3D_GL_Info *gl_info) {
2133 GLhandleARB program_id;
2134 GLhandleARB vshader_id, pshader_id;
2135 const char *blt_vshader[] = {
2136 "void main(void)\n"
2137 "{\n"
2138 " gl_Position = gl_Vertex;\n"
2139 " gl_FrontColor = vec4(1.0);\n"
2140 " gl_TexCoord[0].x = (gl_Vertex.x * 0.5) + 0.5;\n"
2141 " gl_TexCoord[0].y = (-gl_Vertex.y * 0.5) + 0.5;\n"
2142 "}\n"
2143 };
2144
2145 const char *blt_pshader[] = {
2146 "uniform sampler2D sampler;\n"
2147 "void main(void)\n"
2148 "{\n"
2149 " gl_FragDepth = texture2D(sampler, gl_TexCoord[0].xy).x;\n"
2150 "}\n"
2151 };
2152
2153 vshader_id = GL_EXTCALL(glCreateShaderObjectARB(GL_VERTEX_SHADER_ARB));
2154 GL_EXTCALL(glShaderSourceARB(vshader_id, 1, blt_vshader, NULL));
2155 GL_EXTCALL(glCompileShaderARB(vshader_id));
2156
2157 pshader_id = GL_EXTCALL(glCreateShaderObjectARB(GL_FRAGMENT_SHADER_ARB));
2158 GL_EXTCALL(glShaderSourceARB(pshader_id, 1, blt_pshader, NULL));
2159 GL_EXTCALL(glCompileShaderARB(pshader_id));
2160
2161 program_id = GL_EXTCALL(glCreateProgramObjectARB());
2162 GL_EXTCALL(glAttachObjectARB(program_id, vshader_id));
2163 GL_EXTCALL(glAttachObjectARB(program_id, pshader_id));
2164 GL_EXTCALL(glLinkProgramARB(program_id));
2165
2166 print_glsl_info_log(&GLINFO_LOCATION, program_id);
2167
2168 return program_id;
2169 }
2170
2171 static void shader_glsl_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {
2172 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
2173 WineD3D_GL_Info *gl_info = &((IWineD3DImpl *)(This->wineD3D))->gl_info;
2174 GLhandleARB program_id = 0;
2175
2176 if (useVS || usePS) set_glsl_shader_program(iface);
2177 else This->stateBlock->glsl_program = NULL;
2178
2179 program_id = This->stateBlock->glsl_program ? This->stateBlock->glsl_program->programId : 0;
2180 if (program_id) TRACE("Using GLSL program %u\n", program_id);
2181 GL_EXTCALL(glUseProgramObjectARB(program_id));
2182 checkGLcall("glUseProgramObjectARB");
2183 }
2184
2185 static void shader_glsl_select_depth_blt(IWineD3DDevice *iface) {
2186 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
2187 WineD3D_GL_Info *gl_info = &((IWineD3DImpl *)(This->wineD3D))->gl_info;
2188 static GLhandleARB program_id = 0;
2189 static GLhandleARB loc = -1;
2190
2191 if (!program_id) {
2192 program_id = create_glsl_blt_shader(gl_info);
2193 loc = GL_EXTCALL(glGetUniformLocationARB(program_id, "sampler"));
2194 }
2195
2196 GL_EXTCALL(glUseProgramObjectARB(program_id));
2197 GL_EXTCALL(glUniform1iARB(loc, 0));
2198 }
2199
2200 static void shader_glsl_cleanup(BOOL usePS, BOOL useVS) {
2201 /* Nothing to do */
2202 }
2203
2204 const shader_backend_t glsl_shader_backend = {
2205 &shader_glsl_select,
2206 &shader_glsl_select_depth_blt,
2207 &shader_glsl_load_constants,
2208 &shader_glsl_cleanup
2209 };
Maarten's multimedia branch