?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
wined3d: Remove stateblock::set.*.
[wine.git] / dlls / wined3d / glsl_shader.c
blob9233d4bfbe44746fe9f90894a1d9d29de93ee222
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 WINE_DECLARE_DEBUG_CHANNEL(d3d_constants);
36
37 #define GLINFO_LOCATION (*gl_info)
38
39 typedef struct {
40 char reg_name[50];
41 char mask_str[6];
42 } glsl_dst_param_t;
43
44 typedef struct {
45 char reg_name[50];
46 char param_str[100];
47 } glsl_src_param_t;
48
49 typedef struct {
50 const char *name;
51 DWORD coord_mask;
52 } glsl_sample_function_t;
53
54 /** Prints the GLSL info log which will contain error messages if they exist */
55 void print_glsl_info_log(WineD3D_GL_Info *gl_info, GLhandleARB obj) {
56
57 int infologLength = 0;
58 char *infoLog;
59
60 GL_EXTCALL(glGetObjectParameterivARB(obj,
61 GL_OBJECT_INFO_LOG_LENGTH_ARB,
62 &infologLength));
63
64 /* A size of 1 is just a null-terminated string, so the log should be bigger than
65 * that if there are errors. */
66 if (infologLength > 1)
67 {
68 infoLog = HeapAlloc(GetProcessHeap(), 0, infologLength);
69 GL_EXTCALL(glGetInfoLogARB(obj, infologLength, NULL, infoLog));
70 FIXME("Error received from GLSL shader #%u: %s\n", obj, debugstr_a(infoLog));
71 HeapFree(GetProcessHeap(), 0, infoLog);
72 }
73 }
74
75 /**
76 * Loads (pixel shader) samplers
77 */
78 static void shader_glsl_load_psamplers(
79 WineD3D_GL_Info *gl_info,
80 IWineD3DStateBlock* iface) {
81
82 IWineD3DStateBlockImpl* stateBlock = (IWineD3DStateBlockImpl*) iface;
83 GLhandleARB programId = stateBlock->glsl_program->programId;
84 GLhandleARB name_loc;
85 int i;
86 char sampler_name[20];
87
88 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i) {
89 snprintf(sampler_name, sizeof(sampler_name), "Psampler%d", i);
90 name_loc = GL_EXTCALL(glGetUniformLocationARB(programId, sampler_name));
91 if (name_loc != -1) {
92 int mapped_unit = stateBlock->wineD3DDevice->texUnitMap[i];
93 if (mapped_unit != -1 && mapped_unit < GL_LIMITS(fragment_samplers)) {
94 TRACE("Loading %s for texture %d\n", sampler_name, mapped_unit);
95 GL_EXTCALL(glUniform1iARB(name_loc, mapped_unit));
96 checkGLcall("glUniform1iARB");
97 } else {
98 ERR("Trying to load sampler %s on unsupported unit %d\n", sampler_name, mapped_unit);
99 }
100 }
101 }
102 }
103
104 static void shader_glsl_load_vsamplers(WineD3D_GL_Info *gl_info, IWineD3DStateBlock* iface) {
105 IWineD3DStateBlockImpl* stateBlock = (IWineD3DStateBlockImpl*) iface;
106 GLhandleARB programId = stateBlock->glsl_program->programId;
107 GLhandleARB name_loc;
108 char sampler_name[20];
109 int i;
110
111 for (i = 0; i < MAX_VERTEX_SAMPLERS; ++i) {
112 snprintf(sampler_name, sizeof(sampler_name), "Vsampler%d", i);
113 name_loc = GL_EXTCALL(glGetUniformLocationARB(programId, sampler_name));
114 if (name_loc != -1) {
115 int mapped_unit = stateBlock->wineD3DDevice->texUnitMap[MAX_FRAGMENT_SAMPLERS + i];
116 if (mapped_unit != -1 && mapped_unit < GL_LIMITS(combined_samplers)) {
117 TRACE("Loading %s for texture %d\n", sampler_name, mapped_unit);
118 GL_EXTCALL(glUniform1iARB(name_loc, mapped_unit));
119 checkGLcall("glUniform1iARB");
120 } else {
121 ERR("Trying to load sampler %s on unsupported unit %d\n", sampler_name, mapped_unit);
122 }
123 }
124 }
125 }
126
127 /**
128 * Loads floating point constants (aka uniforms) into the currently set GLSL program.
129 * When constant_list == NULL, it will load all the constants.
130 */
131 static void shader_glsl_load_constantsF(IWineD3DBaseShaderImpl* This, WineD3D_GL_Info *gl_info,
132 unsigned int max_constants, float* constants, GLhandleARB *constant_locations,
133 struct list *constant_list) {
134 constants_entry *constant;
135 local_constant* lconst;
136 GLhandleARB tmp_loc;
137 DWORD i, j;
138 DWORD *idx;
139
140 if (TRACE_ON(d3d_shader)) {
141 LIST_FOR_EACH_ENTRY(constant, constant_list, constants_entry, entry) {
142 idx = constant->idx;
143 j = constant->count;
144 while (j--) {
145 i = *idx++;
146 tmp_loc = constant_locations[i];
147 if (tmp_loc != -1) {
148 TRACE_(d3d_constants)("Loading constants %i: %f, %f, %f, %f\n", i,
149 constants[i * 4 + 0], constants[i * 4 + 1],
150 constants[i * 4 + 2], constants[i * 4 + 3]);
151 }
152 }
153 }
154 }
155 LIST_FOR_EACH_ENTRY(constant, constant_list, constants_entry, entry) {
156 idx = constant->idx;
157 j = constant->count;
158 while (j--) {
159 i = *idx++;
160 tmp_loc = constant_locations[i];
161 if (tmp_loc != -1) {
162 /* We found this uniform name in the program - go ahead and send the data */
163 GL_EXTCALL(glUniform4fvARB(tmp_loc, 1, constants + (i * 4)));
164 }
165 }
166 }
167 checkGLcall("glUniform4fvARB()");
168
169 /* Load immediate constants */
170 if (TRACE_ON(d3d_shader)) {
171 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
172 tmp_loc = constant_locations[lconst->idx];
173 if (tmp_loc != -1) {
174 GLfloat* values = (GLfloat*)lconst->value;
175 TRACE_(d3d_constants)("Loading local constants %i: %f, %f, %f, %f\n", lconst->idx,
176 values[0], values[1], values[2], values[3]);
177 }
178 }
179 }
180 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
181 tmp_loc = constant_locations[lconst->idx];
182 if (tmp_loc != -1) {
183 /* We found this uniform name in the program - go ahead and send the data */
184 GL_EXTCALL(glUniform4fvARB(tmp_loc, 1, (GLfloat*)lconst->value));
185 }
186 }
187 checkGLcall("glUniform4fvARB()");
188 }
189
190 /**
191 * Loads integer constants (aka uniforms) into the currently set GLSL program.
192 * When @constants_set == NULL, it will load all the constants.
193 */
194 static void shader_glsl_load_constantsI(
195 IWineD3DBaseShaderImpl* This,
196 WineD3D_GL_Info *gl_info,
197 GLhandleARB programId,
198 unsigned max_constants,
199 int* constants,
200 BOOL* constants_set) {
201
202 GLhandleARB tmp_loc;
203 int i;
204 char tmp_name[8];
205 char is_pshader = shader_is_pshader_version(This->baseShader.hex_version);
206 const char* prefix = is_pshader? "PI":"VI";
207 struct list* ptr;
208
209 for (i=0; i<max_constants; ++i) {
210 if (NULL == constants_set || constants_set[i]) {
211
212 TRACE_(d3d_constants)("Loading constants %i: %i, %i, %i, %i\n",
213 i, constants[i*4], constants[i*4+1], constants[i*4+2], constants[i*4+3]);
214
215 /* TODO: Benchmark and see if it would be beneficial to store the
216 * locations of the constants to avoid looking up each time */
217 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, i);
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, &constants[i*4]));
222 checkGLcall("glUniform4ivARB");
223 }
224 }
225 }
226
227 /* Load immediate constants */
228 ptr = list_head(&This->baseShader.constantsI);
229 while (ptr) {
230 local_constant* lconst = LIST_ENTRY(ptr, struct local_constant, entry);
231 unsigned int idx = lconst->idx;
232 GLint* values = (GLint*) lconst->value;
233
234 TRACE_(d3d_constants)("Loading local constants %i: %i, %i, %i, %i\n", idx,
235 values[0], values[1], values[2], values[3]);
236
237 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, idx);
238 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
239 if (tmp_loc != -1) {
240 /* We found this uniform name in the program - go ahead and send the data */
241 GL_EXTCALL(glUniform4ivARB(tmp_loc, 1, values));
242 checkGLcall("glUniform4ivARB");
243 }
244 ptr = list_next(&This->baseShader.constantsI, ptr);
245 }
246 }
247
248 /**
249 * Loads boolean constants (aka uniforms) into the currently set GLSL program.
250 * When @constants_set == NULL, it will load all the constants.
251 */
252 static void shader_glsl_load_constantsB(
253 IWineD3DBaseShaderImpl* This,
254 WineD3D_GL_Info *gl_info,
255 GLhandleARB programId,
256 unsigned max_constants,
257 BOOL* constants,
258 BOOL* constants_set) {
259
260 GLhandleARB tmp_loc;
261 int i;
262 char tmp_name[8];
263 char is_pshader = shader_is_pshader_version(This->baseShader.hex_version);
264 const char* prefix = is_pshader? "PB":"VB";
265 struct list* ptr;
266
267 for (i=0; i<max_constants; ++i) {
268 if (NULL == constants_set || constants_set[i]) {
269
270 TRACE_(d3d_constants)("Loading constants %i: %i;\n", i, constants[i*4]);
271
272 /* TODO: Benchmark and see if it would be beneficial to store the
273 * locations of the constants to avoid looking up each time */
274 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, i);
275 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
276 if (tmp_loc != -1) {
277 /* We found this uniform name in the program - go ahead and send the data */
278 GL_EXTCALL(glUniform1ivARB(tmp_loc, 1, &constants[i*4]));
279 checkGLcall("glUniform1ivARB");
280 }
281 }
282 }
283
284 /* Load immediate constants */
285 ptr = list_head(&This->baseShader.constantsB);
286 while (ptr) {
287 local_constant* lconst = LIST_ENTRY(ptr, struct local_constant, entry);
288 unsigned int idx = lconst->idx;
289 GLint* values = (GLint*) lconst->value;
290
291 TRACE_(d3d_constants)("Loading local constants %i: %i\n", idx, values[0]);
292
293 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, idx);
294 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
295 if (tmp_loc != -1) {
296 /* We found this uniform name in the program - go ahead and send the data */
297 GL_EXTCALL(glUniform1ivARB(tmp_loc, 1, values));
298 checkGLcall("glUniform1ivARB");
299 }
300 ptr = list_next(&This->baseShader.constantsB, ptr);
301 }
302 }
303
304
305
306 /**
307 * Loads the app-supplied constants into the currently set GLSL program.
308 */
309 void shader_glsl_load_constants(
310 IWineD3DDevice* device,
311 char usePixelShader,
312 char useVertexShader) {
313
314 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) device;
315 IWineD3DStateBlockImpl* stateBlock = deviceImpl->stateBlock;
316 WineD3D_GL_Info *gl_info = &deviceImpl->adapter->gl_info;
317
318 GLhandleARB *constant_locations;
319 struct list *constant_list;
320 GLhandleARB programId;
321 GLint pos;
322
323 if (!stateBlock->glsl_program) {
324 /* No GLSL program set - nothing to do. */
325 return;
326 }
327 programId = stateBlock->glsl_program->programId;
328
329 if (useVertexShader) {
330 IWineD3DBaseShaderImpl* vshader = (IWineD3DBaseShaderImpl*) stateBlock->vertexShader;
331 GLint pos;
332
333 constant_locations = stateBlock->glsl_program->vuniformF_locations;
334 constant_list = &stateBlock->set_vconstantsF;
335
336 /* Load vertex shader samplers */
337 shader_glsl_load_vsamplers(gl_info, (IWineD3DStateBlock*)stateBlock);
338
339 /* Load DirectX 9 float constants/uniforms for vertex shader */
340 shader_glsl_load_constantsF(vshader, gl_info, GL_LIMITS(vshader_constantsF),
341 stateBlock->vertexShaderConstantF, constant_locations, constant_list);
342
343 /* Load DirectX 9 integer constants/uniforms for vertex shader */
344 shader_glsl_load_constantsI(vshader, gl_info, programId, MAX_CONST_I,
345 stateBlock->vertexShaderConstantI,
346 stateBlock->changed.vertexShaderConstantsI);
347
348 /* Load DirectX 9 boolean constants/uniforms for vertex shader */
349 shader_glsl_load_constantsB(vshader, gl_info, programId, MAX_CONST_B,
350 stateBlock->vertexShaderConstantB,
351 stateBlock->changed.vertexShaderConstantsB);
352
353 /* Upload the position fixup params */
354 pos = GL_EXTCALL(glGetUniformLocationARB(programId, "posFixup"));
355 checkGLcall("glGetUniformLocationARB");
356 GL_EXTCALL(glUniform4fvARB(pos, 1, &deviceImpl->posFixup[0]));
357 checkGLcall("glUniform4fvARB");
358 }
359
360 if (usePixelShader) {
361
362 IWineD3DBaseShaderImpl* pshader = (IWineD3DBaseShaderImpl*) stateBlock->pixelShader;
363
364 constant_locations = stateBlock->glsl_program->puniformF_locations;
365 constant_list = &stateBlock->set_pconstantsF;
366
367 /* Load pixel shader samplers */
368 shader_glsl_load_psamplers(gl_info, (IWineD3DStateBlock*) stateBlock);
369
370 /* Load DirectX 9 float constants/uniforms for pixel shader */
371 shader_glsl_load_constantsF(pshader, gl_info, GL_LIMITS(pshader_constantsF),
372 stateBlock->pixelShaderConstantF, constant_locations, constant_list);
373
374 /* Load DirectX 9 integer constants/uniforms for pixel shader */
375 shader_glsl_load_constantsI(pshader, gl_info, programId, MAX_CONST_I,
376 stateBlock->pixelShaderConstantI,
377 stateBlock->changed.pixelShaderConstantsI);
378
379 /* Load DirectX 9 boolean constants/uniforms for pixel shader */
380 shader_glsl_load_constantsB(pshader, gl_info, programId, MAX_CONST_B,
381 stateBlock->pixelShaderConstantB,
382 stateBlock->changed.pixelShaderConstantsB);
383
384 /* Upload the environment bump map matrix if needed. The needsbumpmat member specifies the texture stage to load the matrix from.
385 * It can't be 0 for a valid texbem instruction.
386 */
387 if(((IWineD3DPixelShaderImpl *) pshader)->needsbumpmat != -1) {
388 float *data = (float *) &stateBlock->textureState[(int) ((IWineD3DPixelShaderImpl *) pshader)->needsbumpmat][WINED3DTSS_BUMPENVMAT00];
389 pos = GL_EXTCALL(glGetUniformLocationARB(programId, "bumpenvmat"));
390 checkGLcall("glGetUniformLocationARB");
391 GL_EXTCALL(glUniformMatrix2fvARB(pos, 1, 0, data));
392 checkGLcall("glUniform4fvARB");
393 }
394 }
395 }
396
397 /** Generate the variable & register declarations for the GLSL output target */
398 void shader_generate_glsl_declarations(
399 IWineD3DBaseShader *iface,
400 shader_reg_maps* reg_maps,
401 SHADER_BUFFER* buffer,
402 WineD3D_GL_Info* gl_info) {
403
404 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
405 int i;
406
407 /* There are some minor differences between pixel and vertex shaders */
408 char pshader = shader_is_pshader_version(This->baseShader.hex_version);
409 char prefix = pshader ? 'P' : 'V';
410
411 /* Prototype the subroutines */
412 for (i = 0; i < This->baseShader.limits.label; i++) {
413 if (reg_maps->labels[i])
414 shader_addline(buffer, "void subroutine%lu();\n", i);
415 }
416
417 /* Declare the constants (aka uniforms) */
418 if (This->baseShader.limits.constant_float > 0) {
419 unsigned max_constantsF = min(This->baseShader.limits.constant_float,
420 (pshader ? GL_LIMITS(pshader_constantsF) : GL_LIMITS(vshader_constantsF)));
421 shader_addline(buffer, "uniform vec4 %cC[%u];\n", prefix, max_constantsF);
422 }
423
424 if (This->baseShader.limits.constant_int > 0)
425 shader_addline(buffer, "uniform ivec4 %cI[%u];\n", prefix, This->baseShader.limits.constant_int);
426
427 if (This->baseShader.limits.constant_bool > 0)
428 shader_addline(buffer, "uniform bool %cB[%u];\n", prefix, This->baseShader.limits.constant_bool);
429
430 if(!pshader)
431 shader_addline(buffer, "uniform vec4 posFixup;\n");
432 else if(reg_maps->bumpmat != -1)
433 shader_addline(buffer, "uniform mat2 bumpenvmat;\n");
434
435 /* Declare texture samplers */
436 for (i = 0; i < This->baseShader.limits.sampler; i++) {
437 if (reg_maps->samplers[i]) {
438
439 DWORD stype = reg_maps->samplers[i] & WINED3DSP_TEXTURETYPE_MASK;
440 switch (stype) {
441
442 case WINED3DSTT_1D:
443 shader_addline(buffer, "uniform sampler1D %csampler%lu;\n", prefix, i);
444 break;
445 case WINED3DSTT_2D:
446 shader_addline(buffer, "uniform sampler2D %csampler%lu;\n", prefix, i);
447 break;
448 case WINED3DSTT_CUBE:
449 shader_addline(buffer, "uniform samplerCube %csampler%lu;\n", prefix, i);
450 break;
451 case WINED3DSTT_VOLUME:
452 shader_addline(buffer, "uniform sampler3D %csampler%lu;\n", prefix, i);
453 break;
454 default:
455 shader_addline(buffer, "uniform unsupported_sampler %csampler%lu;\n", prefix, i);
456 FIXME("Unrecognized sampler type: %#x\n", stype);
457 break;
458 }
459 }
460 }
461
462 /* Declare address variables */
463 for (i = 0; i < This->baseShader.limits.address; i++) {
464 if (reg_maps->address[i])
465 shader_addline(buffer, "ivec4 A%d;\n", i);
466 }
467
468 /* Declare texture coordinate temporaries and initialize them */
469 for (i = 0; i < This->baseShader.limits.texcoord; i++) {
470 if (reg_maps->texcoord[i])
471 shader_addline(buffer, "vec4 T%lu = gl_TexCoord[%lu];\n", i, i);
472 }
473
474 /* Declare input register temporaries */
475 for (i=0; i < This->baseShader.limits.packed_input; i++) {
476 if (reg_maps->packed_input[i])
477 shader_addline(buffer, "vec4 IN%lu;\n", i);
478 }
479
480 /* Declare output register temporaries */
481 for (i = 0; i < This->baseShader.limits.packed_output; i++) {
482 if (reg_maps->packed_output[i])
483 shader_addline(buffer, "vec4 OUT%lu;\n", i);
484 }
485
486 /* Declare temporary variables */
487 for(i = 0; i < This->baseShader.limits.temporary; i++) {
488 if (reg_maps->temporary[i])
489 shader_addline(buffer, "vec4 R%lu;\n", i);
490 }
491
492 /* Declare attributes */
493 for (i = 0; i < This->baseShader.limits.attributes; i++) {
494 if (reg_maps->attributes[i])
495 shader_addline(buffer, "attribute vec4 attrib%i;\n", i);
496 }
497
498 /* Declare loop register aL */
499 if (reg_maps->loop) {
500 shader_addline(buffer, "int aL;\n");
501 shader_addline(buffer, "int tmpInt;\n");
502 }
503
504 /* Temporary variables for matrix operations */
505 shader_addline(buffer, "vec4 tmp0;\n");
506 shader_addline(buffer, "vec4 tmp1;\n");
507
508 /* Start the main program */
509 shader_addline(buffer, "void main() {\n");
510 }
511
512 /*****************************************************************************
513 * Functions to generate GLSL strings from DirectX Shader bytecode begin here.
514 *
515 * For more information, see http://wiki.winehq.org/DirectX-Shaders
516 ****************************************************************************/
517
518 /* Prototypes */
519 static void shader_glsl_add_src_param(SHADER_OPCODE_ARG* arg, const DWORD param,
520 const DWORD addr_token, DWORD mask, glsl_src_param_t *src_param);
521
522 /** Used for opcode modifiers - They multiply the result by the specified amount */
523 static const char * const shift_glsl_tab[] = {
524 "", /* 0 (none) */
525 "2.0 * ", /* 1 (x2) */
526 "4.0 * ", /* 2 (x4) */
527 "8.0 * ", /* 3 (x8) */
528 "16.0 * ", /* 4 (x16) */
529 "32.0 * ", /* 5 (x32) */
530 "", /* 6 (x64) */
531 "", /* 7 (x128) */
532 "", /* 8 (d256) */
533 "", /* 9 (d128) */
534 "", /* 10 (d64) */
535 "", /* 11 (d32) */
536 "0.0625 * ", /* 12 (d16) */
537 "0.125 * ", /* 13 (d8) */
538 "0.25 * ", /* 14 (d4) */
539 "0.5 * " /* 15 (d2) */
540 };
541
542 /* Generate a GLSL parameter that does the input modifier computation and return the input register/mask to use */
543 static void shader_glsl_gen_modifier (
544 const DWORD instr,
545 const char *in_reg,
546 const char *in_regswizzle,
547 char *out_str) {
548
549 out_str[0] = 0;
550
551 if (instr == WINED3DSIO_TEXKILL)
552 return;
553
554 switch (instr & WINED3DSP_SRCMOD_MASK) {
555 case WINED3DSPSM_DZ: /* Need to handle this in the instructions itself (texld & texcrd). */
556 case WINED3DSPSM_DW:
557 case WINED3DSPSM_NONE:
558 sprintf(out_str, "%s%s", in_reg, in_regswizzle);
559 break;
560 case WINED3DSPSM_NEG:
561 sprintf(out_str, "-%s%s", in_reg, in_regswizzle);
562 break;
563 case WINED3DSPSM_NOT:
564 sprintf(out_str, "!%s%s", in_reg, in_regswizzle);
565 break;
566 case WINED3DSPSM_BIAS:
567 sprintf(out_str, "(%s%s - vec4(0.5)%s)", in_reg, in_regswizzle, in_regswizzle);
568 break;
569 case WINED3DSPSM_BIASNEG:
570 sprintf(out_str, "-(%s%s - vec4(0.5)%s)", in_reg, in_regswizzle, in_regswizzle);
571 break;
572 case WINED3DSPSM_SIGN:
573 sprintf(out_str, "(2.0 * (%s%s - 0.5))", in_reg, in_regswizzle);
574 break;
575 case WINED3DSPSM_SIGNNEG:
576 sprintf(out_str, "-(2.0 * (%s%s - 0.5))", in_reg, in_regswizzle);
577 break;
578 case WINED3DSPSM_COMP:
579 sprintf(out_str, "(1.0 - %s%s)", in_reg, in_regswizzle);
580 break;
581 case WINED3DSPSM_X2:
582 sprintf(out_str, "(2.0 * %s%s)", in_reg, in_regswizzle);
583 break;
584 case WINED3DSPSM_X2NEG:
585 sprintf(out_str, "-(2.0 * %s%s)", in_reg, in_regswizzle);
586 break;
587 case WINED3DSPSM_ABS:
588 sprintf(out_str, "abs(%s%s)", in_reg, in_regswizzle);
589 break;
590 case WINED3DSPSM_ABSNEG:
591 sprintf(out_str, "-abs(%s%s)", in_reg, in_regswizzle);
592 break;
593 default:
594 FIXME("Unhandled modifier %u\n", (instr & WINED3DSP_SRCMOD_MASK));
595 sprintf(out_str, "%s%s", in_reg, in_regswizzle);
596 }
597 }
598
599 /** Writes the GLSL variable name that corresponds to the register that the
600 * DX opcode parameter is trying to access */
601 static void shader_glsl_get_register_name(
602 const DWORD param,
603 const DWORD addr_token,
604 char* regstr,
605 BOOL* is_color,
606 SHADER_OPCODE_ARG* arg) {
607
608 /* oPos, oFog and oPts in D3D */
609 static const char * const hwrastout_reg_names[] = { "gl_Position", "gl_FogFragCoord", "gl_PointSize" };
610
611 DWORD reg = param & WINED3DSP_REGNUM_MASK;
612 DWORD regtype = shader_get_regtype(param);
613 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) arg->shader;
614 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
615 WineD3D_GL_Info* gl_info = &deviceImpl->adapter->gl_info;
616
617 char pshader = shader_is_pshader_version(This->baseShader.hex_version);
618 char tmpStr[50];
619
620 *is_color = FALSE;
621
622 switch (regtype) {
623 case WINED3DSPR_TEMP:
624 sprintf(tmpStr, "R%u", reg);
625 break;
626 case WINED3DSPR_INPUT:
627 if (pshader) {
628 /* Pixel shaders >= 3.0 */
629 if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3)
630 sprintf(tmpStr, "IN%u", reg);
631 else {
632 if (reg==0)
633 strcpy(tmpStr, "gl_Color");
634 else
635 strcpy(tmpStr, "gl_SecondaryColor");
636 }
637 } else {
638 if (vshader_input_is_color((IWineD3DVertexShader*) This, reg))
639 *is_color = TRUE;
640 sprintf(tmpStr, "attrib%u", reg);
641 }
642 break;
643 case WINED3DSPR_CONST:
644 {
645 const char* prefix = pshader? "PC":"VC";
646
647 /* Relative addressing */
648 if (param & WINED3DSHADER_ADDRMODE_RELATIVE) {
649
650 /* Relative addressing on shaders 2.0+ have a relative address token,
651 * prior to that, it was hard-coded as "A0.x" because there's only 1 register */
652 if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 2) {
653 glsl_src_param_t rel_param;
654 shader_glsl_add_src_param(arg, addr_token, 0, WINED3DSP_WRITEMASK_0, &rel_param);
655 sprintf(tmpStr, "%s[%s + %u]", prefix, rel_param.param_str, reg);
656 } else
657 sprintf(tmpStr, "%s[A0.x + %u]", prefix, reg);
658
659 } else
660 sprintf(tmpStr, "%s[%u]", prefix, reg);
661
662 break;
663 }
664 case WINED3DSPR_CONSTINT:
665 if (pshader)
666 sprintf(tmpStr, "PI[%u]", reg);
667 else
668 sprintf(tmpStr, "VI[%u]", reg);
669 break;
670 case WINED3DSPR_CONSTBOOL:
671 if (pshader)
672 sprintf(tmpStr, "PB[%u]", reg);
673 else
674 sprintf(tmpStr, "VB[%u]", reg);
675 break;
676 case WINED3DSPR_TEXTURE: /* case WINED3DSPR_ADDR: */
677 if (pshader) {
678 sprintf(tmpStr, "T%u", reg);
679 } else {
680 sprintf(tmpStr, "A%u", reg);
681 }
682 break;
683 case WINED3DSPR_LOOP:
684 sprintf(tmpStr, "aL");
685 break;
686 case WINED3DSPR_SAMPLER:
687 if (pshader)
688 sprintf(tmpStr, "Psampler%u", reg);
689 else
690 sprintf(tmpStr, "Vsampler%u", reg);
691 break;
692 case WINED3DSPR_COLOROUT:
693 if (reg >= GL_LIMITS(buffers)) {
694 WARN("Write to render target %u, only %d supported\n", reg, 4);
695 }
696 if (GL_SUPPORT(ARB_DRAW_BUFFERS)) {
697 sprintf(tmpStr, "gl_FragData[%u]", reg);
698 } else { /* On older cards with GLSL support like the GeforceFX there's only one buffer. */
699 sprintf(tmpStr, "gl_FragColor");
700 }
701 break;
702 case WINED3DSPR_RASTOUT:
703 sprintf(tmpStr, "%s", hwrastout_reg_names[reg]);
704 break;
705 case WINED3DSPR_DEPTHOUT:
706 sprintf(tmpStr, "gl_FragDepth");
707 break;
708 case WINED3DSPR_ATTROUT:
709 if (reg == 0) {
710 sprintf(tmpStr, "gl_FrontColor");
711 } else {
712 sprintf(tmpStr, "gl_FrontSecondaryColor");
713 }
714 break;
715 case WINED3DSPR_TEXCRDOUT:
716 /* Vertex shaders >= 3.0: WINED3DSPR_OUTPUT */
717 if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3)
718 sprintf(tmpStr, "OUT%u", reg);
719 else
720 sprintf(tmpStr, "gl_TexCoord[%u]", reg);
721 break;
722 case WINED3DSPR_MISCTYPE:
723 if (reg == 0) {
724 /* vPos */
725 sprintf(tmpStr, "gl_FragCoord");
726 } else {
727 /* gl_FrontFacing could be used for vFace, but note that
728 * gl_FrontFacing is a bool, while vFace is a float for
729 * which the sign determines front/back */
730 FIXME("Unhandled misctype register %d\n", reg);
731 sprintf(tmpStr, "unrecognized_register");
732 }
733 break;
734 default:
735 FIXME("Unhandled register name Type(%d)\n", regtype);
736 sprintf(tmpStr, "unrecognized_register");
737 break;
738 }
739
740 strcat(regstr, tmpStr);
741 }
742
743 /* Get the GLSL write mask for the destination register */
744 static DWORD shader_glsl_get_write_mask(const DWORD param, char *write_mask) {
745 char *ptr = write_mask;
746 DWORD mask = param & WINED3DSP_WRITEMASK_ALL;
747
748 if (shader_is_scalar(param)) {
749 mask = WINED3DSP_WRITEMASK_0;
750 } else {
751 *ptr++ = '.';
752 if (param & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
753 if (param & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
754 if (param & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
755 if (param & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
756 }
757
758 *ptr = '\0';
759
760 return mask;
761 }
762
763 static size_t shader_glsl_get_write_mask_size(DWORD write_mask) {
764 size_t size = 0;
765
766 if (write_mask & WINED3DSP_WRITEMASK_0) ++size;
767 if (write_mask & WINED3DSP_WRITEMASK_1) ++size;
768 if (write_mask & WINED3DSP_WRITEMASK_2) ++size;
769 if (write_mask & WINED3DSP_WRITEMASK_3) ++size;
770
771 return size;
772 }
773
774 static void shader_glsl_get_swizzle(const DWORD param, BOOL fixup, DWORD mask, char *swizzle_str) {
775 /* For registers of type WINED3DDECLTYPE_D3DCOLOR, data is stored as "bgra",
776 * but addressed as "rgba". To fix this we need to swap the register's x
777 * and z components. */
778 DWORD swizzle = (param & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
779 const char *swizzle_chars = fixup ? "zyxw" : "xyzw";
780 char *ptr = swizzle_str;
781
782 if (!shader_is_scalar(param)) {
783 *ptr++ = '.';
784 /* swizzle bits fields: wwzzyyxx */
785 if (mask & WINED3DSP_WRITEMASK_0) *ptr++ = swizzle_chars[swizzle & 0x03];
786 if (mask & WINED3DSP_WRITEMASK_1) *ptr++ = swizzle_chars[(swizzle >> 2) & 0x03];
787 if (mask & WINED3DSP_WRITEMASK_2) *ptr++ = swizzle_chars[(swizzle >> 4) & 0x03];
788 if (mask & WINED3DSP_WRITEMASK_3) *ptr++ = swizzle_chars[(swizzle >> 6) & 0x03];
789 }
790
791 *ptr = '\0';
792 }
793
794 /* From a given parameter token, generate the corresponding GLSL string.
795 * Also, return the actual register name and swizzle in case the
796 * caller needs this information as well. */
797 static void shader_glsl_add_src_param(SHADER_OPCODE_ARG* arg, const DWORD param,
798 const DWORD addr_token, DWORD mask, glsl_src_param_t *src_param) {
799 BOOL is_color = FALSE;
800 char swizzle_str[6];
801
802 src_param->reg_name[0] = '\0';
803 src_param->param_str[0] = '\0';
804 swizzle_str[0] = '\0';
805
806 shader_glsl_get_register_name(param, addr_token, src_param->reg_name, &is_color, arg);
807
808 shader_glsl_get_swizzle(param, is_color, mask, swizzle_str);
809 shader_glsl_gen_modifier(param, src_param->reg_name, swizzle_str, src_param->param_str);
810 }
811
812 /* From a given parameter token, generate the corresponding GLSL string.
813 * Also, return the actual register name and swizzle in case the
814 * caller needs this information as well. */
815 static DWORD shader_glsl_add_dst_param(SHADER_OPCODE_ARG* arg, const DWORD param,
816 const DWORD addr_token, glsl_dst_param_t *dst_param) {
817 BOOL is_color = FALSE;
818
819 dst_param->mask_str[0] = '\0';
820 dst_param->reg_name[0] = '\0';
821
822 shader_glsl_get_register_name(param, addr_token, dst_param->reg_name, &is_color, arg);
823 return shader_glsl_get_write_mask(param, dst_param->mask_str);
824 }
825
826 /* Append the destination part of the instruction to the buffer, return the effective write mask */
827 static DWORD shader_glsl_append_dst_ext(SHADER_BUFFER *buffer, SHADER_OPCODE_ARG *arg, const DWORD param) {
828 glsl_dst_param_t dst_param;
829 DWORD mask;
830 int shift;
831
832 mask = shader_glsl_add_dst_param(arg, param, arg->dst_addr, &dst_param);
833
834 if(mask) {
835 shift = (param & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
836 shader_addline(buffer, "%s%s = %s(", dst_param.reg_name, dst_param.mask_str, shift_glsl_tab[shift]);
837 }
838
839 return mask;
840 }
841
842 /* Append the destination part of the instruction to the buffer, return the effective write mask */
843 static DWORD shader_glsl_append_dst(SHADER_BUFFER *buffer, SHADER_OPCODE_ARG *arg) {
844 return shader_glsl_append_dst_ext(buffer, arg, arg->dst);
845 }
846
847 /** Process GLSL instruction modifiers */
848 void shader_glsl_add_instruction_modifiers(SHADER_OPCODE_ARG* arg) {
849
850 DWORD mask = arg->dst & WINED3DSP_DSTMOD_MASK;
851
852 if (arg->opcode->dst_token && mask != 0) {
853 glsl_dst_param_t dst_param;
854
855 shader_glsl_add_dst_param(arg, arg->dst, 0, &dst_param);
856
857 if (mask & WINED3DSPDM_SATURATE) {
858 /* _SAT means to clamp the value of the register to between 0 and 1 */
859 shader_addline(arg->buffer, "%s%s = clamp(%s%s, 0.0, 1.0);\n", dst_param.reg_name,
860 dst_param.mask_str, dst_param.reg_name, dst_param.mask_str);
861 }
862 if (mask & WINED3DSPDM_MSAMPCENTROID) {
863 FIXME("_centroid modifier not handled\n");
864 }
865 if (mask & WINED3DSPDM_PARTIALPRECISION) {
866 /* MSDN says this modifier can be safely ignored, so that's what we'll do. */
867 }
868 }
869 }
870
871 static inline const char* shader_get_comp_op(
872 const DWORD opcode) {
873
874 DWORD op = (opcode & INST_CONTROLS_MASK) >> INST_CONTROLS_SHIFT;
875 switch (op) {
876 case COMPARISON_GT: return ">";
877 case COMPARISON_EQ: return "==";
878 case COMPARISON_GE: return ">=";
879 case COMPARISON_LT: return "<";
880 case COMPARISON_NE: return "!=";
881 case COMPARISON_LE: return "<=";
882 default:
883 FIXME("Unrecognized comparison value: %u\n", op);
884 return "(\?\?)";
885 }
886 }
887
888 static void shader_glsl_get_sample_function(DWORD sampler_type, BOOL projected, glsl_sample_function_t *sample_function) {
889 /* Note that there's no such thing as a projected cube texture. */
890 switch(sampler_type) {
891 case WINED3DSTT_1D:
892 sample_function->name = projected ? "texture1DProj" : "texture1D";
893 sample_function->coord_mask = WINED3DSP_WRITEMASK_0;
894 break;
895 case WINED3DSTT_2D:
896 sample_function->name = projected ? "texture2DProj" : "texture2D";
897 sample_function->coord_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1;
898 break;
899 case WINED3DSTT_CUBE:
900 sample_function->name = "textureCube";
901 sample_function->coord_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
902 break;
903 case WINED3DSTT_VOLUME:
904 sample_function->name = projected ? "texture3DProj" : "texture3D";
905 sample_function->coord_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
906 break;
907 default:
908 sample_function->name = "";
909 FIXME("Unrecognized sampler type: %#x;\n", sampler_type);
910 break;
911 }
912 }
913
914
915 /*****************************************************************************
916 *
917 * Begin processing individual instruction opcodes
918 *
919 ****************************************************************************/
920
921 /* Generate GLSL arithmetic functions (dst = src1 + src2) */
922 void shader_glsl_arith(SHADER_OPCODE_ARG* arg) {
923 CONST SHADER_OPCODE* curOpcode = arg->opcode;
924 SHADER_BUFFER* buffer = arg->buffer;
925 glsl_src_param_t src0_param;
926 glsl_src_param_t src1_param;
927 DWORD write_mask;
928 char op;
929
930 /* Determine the GLSL operator to use based on the opcode */
931 switch (curOpcode->opcode) {
932 case WINED3DSIO_MUL: op = '*'; break;
933 case WINED3DSIO_ADD: op = '+'; break;
934 case WINED3DSIO_SUB: op = '-'; break;
935 default:
936 op = ' ';
937 FIXME("Opcode %s not yet handled in GLSL\n", curOpcode->name);
938 break;
939 }
940
941 write_mask = shader_glsl_append_dst(buffer, arg);
942 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, &src0_param);
943 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, &src1_param);
944 shader_addline(buffer, "%s %c %s);\n", src0_param.param_str, op, src1_param.param_str);
945 }
946
947 /* Process the WINED3DSIO_MOV opcode using GLSL (dst = src) */
948 void shader_glsl_mov(SHADER_OPCODE_ARG* arg) {
949 IWineD3DBaseShaderImpl* shader = (IWineD3DBaseShaderImpl*) arg->shader;
950 SHADER_BUFFER* buffer = arg->buffer;
951 glsl_src_param_t src0_param;
952 DWORD write_mask;
953
954 write_mask = shader_glsl_append_dst(buffer, arg);
955 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, &src0_param);
956
957 /* In vs_1_1 WINED3DSIO_MOV can write to the address register. In later
958 * shader versions WINED3DSIO_MOVA is used for this. */
959 if ((WINED3DSHADER_VERSION_MAJOR(shader->baseShader.hex_version) == 1 &&
960 !shader_is_pshader_version(shader->baseShader.hex_version) &&
961 shader_get_regtype(arg->dst) == WINED3DSPR_ADDR) ||
962 arg->opcode->opcode == WINED3DSIO_MOVA) {
963 /* We need to *round* to the nearest int here. */
964 size_t mask_size = shader_glsl_get_write_mask_size(write_mask);
965 if (mask_size > 1) {
966 shader_addline(buffer, "ivec%d(floor(abs(%s) + vec%d(0.5)) * sign(%s)));\n", mask_size, src0_param.param_str, mask_size, src0_param.param_str);
967 } else {
968 shader_addline(buffer, "int(floor(abs(%s) + 0.5) * sign(%s)));\n", src0_param.param_str, src0_param.param_str);
969 }
970 } else {
971 shader_addline(buffer, "%s);\n", src0_param.param_str);
972 }
973 }
974
975 /* Process the dot product operators DP3 and DP4 in GLSL (dst = dot(src0, src1)) */
976 void shader_glsl_dot(SHADER_OPCODE_ARG* arg) {
977 CONST SHADER_OPCODE* curOpcode = arg->opcode;
978 SHADER_BUFFER* buffer = arg->buffer;
979 glsl_src_param_t src0_param;
980 glsl_src_param_t src1_param;
981 DWORD dst_write_mask, src_write_mask;
982 size_t dst_size = 0;
983
984 dst_write_mask = shader_glsl_append_dst(buffer, arg);
985 dst_size = shader_glsl_get_write_mask_size(dst_write_mask);
986
987 /* dp3 works on vec3, dp4 on vec4 */
988 if (curOpcode->opcode == WINED3DSIO_DP4) {
989 src_write_mask = WINED3DSP_WRITEMASK_ALL;
990 } else {
991 src_write_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
992 }
993
994 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_write_mask, &src0_param);
995 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], src_write_mask, &src1_param);
996
997 if (dst_size > 1) {
998 shader_addline(buffer, "vec%d(dot(%s, %s)));\n", dst_size, src0_param.param_str, src1_param.param_str);
999 } else {
1000 shader_addline(buffer, "dot(%s, %s));\n", src0_param.param_str, src1_param.param_str);
1001 }
1002 }
1003
1004 /* Note that this instruction has some restrictions. The destination write mask
1005 * can't contain the w component, and the source swizzles have to be .xyzw */
1006 void shader_glsl_cross(SHADER_OPCODE_ARG *arg) {
1007 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1008 glsl_src_param_t src0_param;
1009 glsl_src_param_t src1_param;
1010 char dst_mask[6];
1011
1012 shader_glsl_get_write_mask(arg->dst, dst_mask);
1013 shader_glsl_append_dst(arg->buffer, arg);
1014 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1015 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], src_mask, &src1_param);
1016 shader_addline(arg->buffer, "cross(%s, %s).%s);\n", src0_param.param_str, src1_param.param_str, dst_mask);
1017 }
1018
1019 /* Process the WINED3DSIO_POW instruction in GLSL (dst = |src0|^src1)
1020 * Src0 and src1 are scalars. Note that D3D uses the absolute of src0, while
1021 * GLSL uses the value as-is. */
1022 void shader_glsl_pow(SHADER_OPCODE_ARG *arg) {
1023 SHADER_BUFFER *buffer = arg->buffer;
1024 glsl_src_param_t src0_param;
1025 glsl_src_param_t src1_param;
1026 DWORD dst_write_mask;
1027 size_t dst_size;
1028
1029 dst_write_mask = shader_glsl_append_dst(buffer, arg);
1030 dst_size = shader_glsl_get_write_mask_size(dst_write_mask);
1031
1032 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src0_param);
1033 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_0, &src1_param);
1034
1035 if (dst_size > 1) {
1036 shader_addline(buffer, "vec%d(pow(abs(%s), %s)));\n", dst_size, src0_param.param_str, src1_param.param_str);
1037 } else {
1038 shader_addline(buffer, "pow(abs(%s), %s));\n", src0_param.param_str, src1_param.param_str);
1039 }
1040 }
1041
1042 /* Map the opcode 1-to-1 to the GL code (arg->dst = instruction(src0, src1, ...) */
1043 void shader_glsl_map2gl(SHADER_OPCODE_ARG* arg) {
1044 CONST SHADER_OPCODE* curOpcode = arg->opcode;
1045 SHADER_BUFFER* buffer = arg->buffer;
1046 glsl_src_param_t src_param;
1047 const char *instruction;
1048 char arguments[256];
1049 DWORD write_mask;
1050 unsigned i;
1051
1052 /* Determine the GLSL function to use based on the opcode */
1053 /* TODO: Possibly make this a table for faster lookups */
1054 switch (curOpcode->opcode) {
1055 case WINED3DSIO_MIN: instruction = "min"; break;
1056 case WINED3DSIO_MAX: instruction = "max"; break;
1057 case WINED3DSIO_ABS: instruction = "abs"; break;
1058 case WINED3DSIO_FRC: instruction = "fract"; break;
1059 case WINED3DSIO_NRM: instruction = "normalize"; break;
1060 case WINED3DSIO_LOGP:
1061 case WINED3DSIO_LOG: instruction = "log2"; break;
1062 case WINED3DSIO_EXP: instruction = "exp2"; break;
1063 case WINED3DSIO_SGN: instruction = "sign"; break;
1064 case WINED3DSIO_DSX: instruction = "dFdx"; break;
1065 case WINED3DSIO_DSY: instruction = "dFdy"; break;
1066 default: instruction = "";
1067 FIXME("Opcode %s not yet handled in GLSL\n", curOpcode->name);
1068 break;
1069 }
1070
1071 write_mask = shader_glsl_append_dst(buffer, arg);
1072
1073 arguments[0] = '\0';
1074 if (curOpcode->num_params > 0) {
1075 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, &src_param);
1076 strcat(arguments, src_param.param_str);
1077 for (i = 2; i < curOpcode->num_params; ++i) {
1078 strcat(arguments, ", ");
1079 shader_glsl_add_src_param(arg, arg->src[i-1], arg->src_addr[i-1], write_mask, &src_param);
1080 strcat(arguments, src_param.param_str);
1081 }
1082 }
1083
1084 shader_addline(buffer, "%s(%s));\n", instruction, arguments);
1085 }
1086
1087 /** Process the WINED3DSIO_EXPP instruction in GLSL:
1088 * For shader model 1.x, do the following (and honor the writemask, so use a temporary variable):
1089 * dst.x = 2^(floor(src))
1090 * dst.y = src - floor(src)
1091 * dst.z = 2^src (partial precision is allowed, but optional)
1092 * dst.w = 1.0;
1093 * For 2.0 shaders, just do this (honoring writemask and swizzle):
1094 * dst = 2^src; (partial precision is allowed, but optional)
1095 */
1096 void shader_glsl_expp(SHADER_OPCODE_ARG* arg) {
1097 IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)arg->shader;
1098 glsl_src_param_t src_param;
1099
1100 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src_param);
1101
1102 if (shader->baseShader.hex_version < WINED3DPS_VERSION(2,0)) {
1103 char dst_mask[6];
1104
1105 shader_addline(arg->buffer, "tmp0.x = exp2(floor(%s));\n", src_param.param_str);
1106 shader_addline(arg->buffer, "tmp0.y = %s - floor(%s);\n", src_param.param_str, src_param.param_str);
1107 shader_addline(arg->buffer, "tmp0.z = exp2(%s);\n", src_param.param_str);
1108 shader_addline(arg->buffer, "tmp0.w = 1.0;\n");
1109
1110 shader_glsl_append_dst(arg->buffer, arg);
1111 shader_glsl_get_write_mask(arg->dst, dst_mask);
1112 shader_addline(arg->buffer, "tmp0%s);\n", dst_mask);
1113 } else {
1114 DWORD write_mask;
1115 size_t mask_size;
1116
1117 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1118 mask_size = shader_glsl_get_write_mask_size(write_mask);
1119
1120 if (mask_size > 1) {
1121 shader_addline(arg->buffer, "vec%d(exp2(%s)));\n", mask_size, src_param.param_str);
1122 } else {
1123 shader_addline(arg->buffer, "exp2(%s));\n", src_param.param_str);
1124 }
1125 }
1126 }
1127
1128 /** Process the RCP (reciprocal or inverse) opcode in GLSL (dst = 1 / src) */
1129 void shader_glsl_rcp(SHADER_OPCODE_ARG* arg) {
1130 glsl_src_param_t src_param;
1131 DWORD write_mask;
1132 size_t mask_size;
1133
1134 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1135 mask_size = shader_glsl_get_write_mask_size(write_mask);
1136 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_3, &src_param);
1137
1138 if (mask_size > 1) {
1139 shader_addline(arg->buffer, "vec%d(1.0 / %s));\n", mask_size, src_param.param_str);
1140 } else {
1141 shader_addline(arg->buffer, "1.0 / %s);\n", src_param.param_str);
1142 }
1143 }
1144
1145 void shader_glsl_rsq(SHADER_OPCODE_ARG* arg) {
1146 SHADER_BUFFER* buffer = arg->buffer;
1147 glsl_src_param_t src_param;
1148 DWORD write_mask;
1149 size_t mask_size;
1150
1151 write_mask = shader_glsl_append_dst(buffer, arg);
1152 mask_size = shader_glsl_get_write_mask_size(write_mask);
1153
1154 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_3, &src_param);
1155
1156 if (mask_size > 1) {
1157 shader_addline(buffer, "vec%d(inversesqrt(%s)));\n", mask_size, src_param.param_str);
1158 } else {
1159 shader_addline(buffer, "inversesqrt(%s));\n", src_param.param_str);
1160 }
1161 }
1162
1163 /** Process signed comparison opcodes in GLSL. */
1164 void shader_glsl_compare(SHADER_OPCODE_ARG* arg) {
1165 glsl_src_param_t src0_param;
1166 glsl_src_param_t src1_param;
1167 DWORD write_mask;
1168 size_t mask_size;
1169
1170 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1171 mask_size = shader_glsl_get_write_mask_size(write_mask);
1172 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, &src0_param);
1173 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, &src1_param);
1174
1175 if (mask_size > 1) {
1176 const char *compare;
1177
1178 switch(arg->opcode->opcode) {
1179 case WINED3DSIO_SLT: compare = "lessThan"; break;
1180 case WINED3DSIO_SGE: compare = "greaterThanEqual"; break;
1181 default: compare = "";
1182 FIXME("Can't handle opcode %s\n", arg->opcode->name);
1183 }
1184
1185 shader_addline(arg->buffer, "vec%d(%s(%s, %s)));\n", mask_size, compare,
1186 src0_param.param_str, src1_param.param_str);
1187 } else {
1188 const char *compare;
1189
1190 switch(arg->opcode->opcode) {
1191 case WINED3DSIO_SLT: compare = "<"; break;
1192 case WINED3DSIO_SGE: compare = ">="; break;
1193 default: compare = "";
1194 FIXME("Can't handle opcode %s\n", arg->opcode->name);
1195 }
1196
1197 shader_addline(arg->buffer, "(%s %s %s) ? 1.0 : 0.0);\n",
1198 src0_param.param_str, compare, src1_param.param_str);
1199 }
1200 }
1201
1202 /** Process CMP instruction in GLSL (dst = src0 >= 0.0 ? src1 : src2), per channel */
1203 void shader_glsl_cmp(SHADER_OPCODE_ARG* arg) {
1204 glsl_src_param_t src0_param;
1205 glsl_src_param_t src1_param;
1206 glsl_src_param_t src2_param;
1207 DWORD write_mask, cmp_channel = 0;
1208 unsigned int i, j;
1209
1210 /* Cycle through all source0 channels */
1211 for (i=0; i<4; i++) {
1212 write_mask = 0;
1213 /* Find the destination channels which use the current source0 channel */
1214 for (j=0; j<4; j++) {
1215 if ( ((arg->src[0] >> (WINED3DSP_SWIZZLE_SHIFT + 2*j)) & 0x3) == i ) {
1216 write_mask |= WINED3DSP_WRITEMASK_0 << j;
1217 cmp_channel = WINED3DSP_WRITEMASK_0 << j;
1218 }
1219 }
1220 write_mask = shader_glsl_append_dst_ext(arg->buffer, arg, arg->dst & (~WINED3DSP_SWIZZLE_MASK | write_mask));
1221 if (!write_mask) continue;
1222
1223 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], cmp_channel, &src0_param);
1224 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, &src1_param);
1225 shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], write_mask, &src2_param);
1226
1227 shader_addline(arg->buffer, "%s >= 0.0 ? %s : %s);\n",
1228 src0_param.param_str, src1_param.param_str, src2_param.param_str);
1229 }
1230 }
1231
1232 /** Process the CND opcode in GLSL (dst = (src0 > 0.5) ? src1 : src2) */
1233 /* For ps 1.1-1.3, only a single component of src0 is used. For ps 1.4
1234 * the compare is done per component of src0. */
1235 void shader_glsl_cnd(SHADER_OPCODE_ARG* arg) {
1236 IWineD3DBaseShaderImpl* shader = (IWineD3DBaseShaderImpl*) arg->shader;
1237 glsl_src_param_t src0_param;
1238 glsl_src_param_t src1_param;
1239 glsl_src_param_t src2_param;
1240 DWORD write_mask, cmp_channel = 0;
1241 unsigned int i, j;
1242
1243 if (shader->baseShader.hex_version < WINED3DPS_VERSION(1, 4)) {
1244 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1245 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src0_param);
1246 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, &src1_param);
1247 shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], write_mask, &src2_param);
1248 shader_addline(arg->buffer, "%s > 0.5 ? %s : %s);\n",
1249 src0_param.param_str, src1_param.param_str, src2_param.param_str);
1250 return;
1251 }
1252 /* Cycle through all source0 channels */
1253 for (i=0; i<4; i++) {
1254 write_mask = 0;
1255 /* Find the destination channels which use the current source0 channel */
1256 for (j=0; j<4; j++) {
1257 if ( ((arg->src[0] >> (WINED3DSP_SWIZZLE_SHIFT + 2*j)) & 0x3) == i ) {
1258 write_mask |= WINED3DSP_WRITEMASK_0 << j;
1259 cmp_channel = WINED3DSP_WRITEMASK_0 << j;
1260 }
1261 }
1262 write_mask = shader_glsl_append_dst_ext(arg->buffer, arg, arg->dst & (~WINED3DSP_SWIZZLE_MASK | write_mask));
1263 if (!write_mask) continue;
1264
1265 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], cmp_channel, &src0_param);
1266 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, &src1_param);
1267 shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], write_mask, &src2_param);
1268
1269 shader_addline(arg->buffer, "%s > 0.5 ? %s : %s);\n",
1270 src0_param.param_str, src1_param.param_str, src2_param.param_str);
1271 }
1272 }
1273
1274 /** GLSL code generation for WINED3DSIO_MAD: Multiply the first 2 opcodes, then add the last */
1275 void shader_glsl_mad(SHADER_OPCODE_ARG* arg) {
1276 glsl_src_param_t src0_param;
1277 glsl_src_param_t src1_param;
1278 glsl_src_param_t src2_param;
1279 DWORD write_mask;
1280
1281 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1282 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, &src0_param);
1283 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, &src1_param);
1284 shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], write_mask, &src2_param);
1285 shader_addline(arg->buffer, "(%s * %s) + %s);\n",
1286 src0_param.param_str, src1_param.param_str, src2_param.param_str);
1287 }
1288
1289 /** Handles transforming all WINED3DSIO_M?x? opcodes for
1290 Vertex shaders to GLSL codes */
1291 void shader_glsl_mnxn(SHADER_OPCODE_ARG* arg) {
1292 int i;
1293 int nComponents = 0;
1294 SHADER_OPCODE_ARG tmpArg;
1295
1296 memset(&tmpArg, 0, sizeof(SHADER_OPCODE_ARG));
1297
1298 /* Set constants for the temporary argument */
1299 tmpArg.shader = arg->shader;
1300 tmpArg.buffer = arg->buffer;
1301 tmpArg.src[0] = arg->src[0];
1302 tmpArg.src_addr[0] = arg->src_addr[0];
1303 tmpArg.src_addr[1] = arg->src_addr[1];
1304 tmpArg.reg_maps = arg->reg_maps;
1305
1306 switch(arg->opcode->opcode) {
1307 case WINED3DSIO_M4x4:
1308 nComponents = 4;
1309 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP4);
1310 break;
1311 case WINED3DSIO_M4x3:
1312 nComponents = 3;
1313 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP4);
1314 break;
1315 case WINED3DSIO_M3x4:
1316 nComponents = 4;
1317 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP3);
1318 break;
1319 case WINED3DSIO_M3x3:
1320 nComponents = 3;
1321 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP3);
1322 break;
1323 case WINED3DSIO_M3x2:
1324 nComponents = 2;
1325 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP3);
1326 break;
1327 default:
1328 break;
1329 }
1330
1331 for (i = 0; i < nComponents; i++) {
1332 tmpArg.dst = ((arg->dst) & ~WINED3DSP_WRITEMASK_ALL)|(WINED3DSP_WRITEMASK_0<<i);
1333 tmpArg.src[1] = arg->src[1]+i;
1334 shader_glsl_dot(&tmpArg);
1335 }
1336 }
1337
1338 /**
1339 The LRP instruction performs a component-wise linear interpolation
1340 between the second and third operands using the first operand as the
1341 blend factor. Equation: (dst = src2 + src0 * (src1 - src2))
1342 This is equivalent to mix(src2, src1, src0);
1343 */
1344 void shader_glsl_lrp(SHADER_OPCODE_ARG* arg) {
1345 glsl_src_param_t src0_param;
1346 glsl_src_param_t src1_param;
1347 glsl_src_param_t src2_param;
1348 DWORD write_mask;
1349
1350 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1351
1352 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, &src0_param);
1353 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, &src1_param);
1354 shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], write_mask, &src2_param);
1355
1356 shader_addline(arg->buffer, "mix(%s, %s, %s));\n",
1357 src2_param.param_str, src1_param.param_str, src0_param.param_str);
1358 }
1359
1360 /** Process the WINED3DSIO_LIT instruction in GLSL:
1361 * dst.x = dst.w = 1.0
1362 * dst.y = (src0.x > 0) ? src0.x
1363 * dst.z = (src0.x > 0) ? ((src0.y > 0) ? pow(src0.y, src.w) : 0) : 0
1364 * where src.w is clamped at +- 128
1365 */
1366 void shader_glsl_lit(SHADER_OPCODE_ARG* arg) {
1367 glsl_src_param_t src0_param;
1368 glsl_src_param_t src1_param;
1369 glsl_src_param_t src3_param;
1370 char dst_mask[6];
1371
1372 shader_glsl_append_dst(arg->buffer, arg);
1373 shader_glsl_get_write_mask(arg->dst, dst_mask);
1374
1375 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src0_param);
1376 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_1, &src1_param);
1377 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_3, &src3_param);
1378
1379 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",
1380 src0_param.param_str, src0_param.param_str, src0_param.param_str, src1_param.param_str, src1_param.param_str, src3_param.param_str, dst_mask);
1381 }
1382
1383 /** Process the WINED3DSIO_DST instruction in GLSL:
1384 * dst.x = 1.0
1385 * dst.y = src0.x * src0.y
1386 * dst.z = src0.z
1387 * dst.w = src1.w
1388 */
1389 void shader_glsl_dst(SHADER_OPCODE_ARG* arg) {
1390 glsl_src_param_t src0y_param;
1391 glsl_src_param_t src0z_param;
1392 glsl_src_param_t src1y_param;
1393 glsl_src_param_t src1w_param;
1394 char dst_mask[6];
1395
1396 shader_glsl_append_dst(arg->buffer, arg);
1397 shader_glsl_get_write_mask(arg->dst, dst_mask);
1398
1399 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_1, &src0y_param);
1400 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_2, &src0z_param);
1401 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_1, &src1y_param);
1402 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_3, &src1w_param);
1403
1404 shader_addline(arg->buffer, "vec4(1.0, %s * %s, %s, %s))%s;\n",
1405 src0y_param.param_str, src1y_param.param_str, src0z_param.param_str, src1w_param.param_str, dst_mask);
1406 }
1407
1408 /** Process the WINED3DSIO_SINCOS instruction in GLSL:
1409 * VS 2.0 requires that specific cosine and sine constants be passed to this instruction so the hardware
1410 * can handle it. But, these functions are built-in for GLSL, so we can just ignore the last 2 params.
1411 *
1412 * dst.x = cos(src0.?)
1413 * dst.y = sin(src0.?)
1414 * dst.z = dst.z
1415 * dst.w = dst.w
1416 */
1417 void shader_glsl_sincos(SHADER_OPCODE_ARG* arg) {
1418 glsl_src_param_t src0_param;
1419 DWORD write_mask;
1420
1421 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1422 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src0_param);
1423
1424 switch (write_mask) {
1425 case WINED3DSP_WRITEMASK_0:
1426 shader_addline(arg->buffer, "cos(%s));\n", src0_param.param_str);
1427 break;
1428
1429 case WINED3DSP_WRITEMASK_1:
1430 shader_addline(arg->buffer, "sin(%s));\n", src0_param.param_str);
1431 break;
1432
1433 case (WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1):
1434 shader_addline(arg->buffer, "vec2(cos(%s), sin(%s)));\n", src0_param.param_str, src0_param.param_str);
1435 break;
1436
1437 default:
1438 ERR("Write mask should be .x, .y or .xy\n");
1439 break;
1440 }
1441 }
1442
1443 /** Process the WINED3DSIO_LOOP instruction in GLSL:
1444 * Start a for() loop where src1.y is the initial value of aL,
1445 * increment aL by src1.z for a total of src1.x iterations.
1446 * Need to use a temporary variable for this operation.
1447 */
1448 /* FIXME: I don't think nested loops will work correctly this way. */
1449 void shader_glsl_loop(SHADER_OPCODE_ARG* arg) {
1450 glsl_src_param_t src1_param;
1451
1452 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_ALL, &src1_param);
1453
1454 shader_addline(arg->buffer, "for (tmpInt = 0, aL = %s.y; tmpInt < %s.x; tmpInt++, aL += %s.z) {\n",
1455 src1_param.reg_name, src1_param.reg_name, src1_param.reg_name);
1456 }
1457
1458 void shader_glsl_end(SHADER_OPCODE_ARG* arg) {
1459 shader_addline(arg->buffer, "}\n");
1460 }
1461
1462 void shader_glsl_rep(SHADER_OPCODE_ARG* arg) {
1463 glsl_src_param_t src0_param;
1464
1465 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src0_param);
1466 shader_addline(arg->buffer, "for (tmpInt = 0; tmpInt < %s; tmpInt++) {\n", src0_param.param_str);
1467 }
1468
1469 void shader_glsl_if(SHADER_OPCODE_ARG* arg) {
1470 glsl_src_param_t src0_param;
1471
1472 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src0_param);
1473 shader_addline(arg->buffer, "if (%s) {\n", src0_param.param_str);
1474 }
1475
1476 void shader_glsl_ifc(SHADER_OPCODE_ARG* arg) {
1477 glsl_src_param_t src0_param;
1478 glsl_src_param_t src1_param;
1479
1480 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src0_param);
1481 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_0, &src1_param);
1482
1483 shader_addline(arg->buffer, "if (%s %s %s) {\n",
1484 src0_param.param_str, shader_get_comp_op(arg->opcode_token), src1_param.param_str);
1485 }
1486
1487 void shader_glsl_else(SHADER_OPCODE_ARG* arg) {
1488 shader_addline(arg->buffer, "} else {\n");
1489 }
1490
1491 void shader_glsl_break(SHADER_OPCODE_ARG* arg) {
1492 shader_addline(arg->buffer, "break;\n");
1493 }
1494
1495 /* FIXME: According to MSDN the compare is done per component. */
1496 void shader_glsl_breakc(SHADER_OPCODE_ARG* arg) {
1497 glsl_src_param_t src0_param;
1498 glsl_src_param_t src1_param;
1499
1500 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src0_param);
1501 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_0, &src1_param);
1502
1503 shader_addline(arg->buffer, "if (%s %s %s) break;\n",
1504 src0_param.param_str, shader_get_comp_op(arg->opcode_token), src1_param.param_str);
1505 }
1506
1507 void shader_glsl_label(SHADER_OPCODE_ARG* arg) {
1508
1509 DWORD snum = (arg->src[0]) & WINED3DSP_REGNUM_MASK;
1510 shader_addline(arg->buffer, "}\n");
1511 shader_addline(arg->buffer, "void subroutine%lu () {\n", snum);
1512 }
1513
1514 void shader_glsl_call(SHADER_OPCODE_ARG* arg) {
1515 DWORD snum = (arg->src[0]) & WINED3DSP_REGNUM_MASK;
1516 shader_addline(arg->buffer, "subroutine%lu();\n", snum);
1517 }
1518
1519 void shader_glsl_callnz(SHADER_OPCODE_ARG* arg) {
1520 glsl_src_param_t src1_param;
1521
1522 DWORD snum = (arg->src[0]) & WINED3DSP_REGNUM_MASK;
1523 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_0, &src1_param);
1524 shader_addline(arg->buffer, "if (%s) subroutine%lu();\n", src1_param.param_str, snum);
1525 }
1526
1527 /*********************************************
1528 * Pixel Shader Specific Code begins here
1529 ********************************************/
1530 void pshader_glsl_tex(SHADER_OPCODE_ARG* arg) {
1531 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1532 DWORD hex_version = This->baseShader.hex_version;
1533 char dst_swizzle[6];
1534 glsl_sample_function_t sample_function;
1535 DWORD sampler_type;
1536 DWORD sampler_idx;
1537 BOOL projected;
1538 DWORD mask = 0;
1539
1540 /* All versions have a destination register */
1541 shader_glsl_append_dst(arg->buffer, arg);
1542
1543 /* 1.0-1.4: Use destination register as sampler source.
1544 * 2.0+: Use provided sampler source. */
1545 if (hex_version < WINED3DPS_VERSION(1,4)) {
1546 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1547 DWORD flags;
1548
1549 sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1550 flags = deviceImpl->stateBlock->textureState[sampler_idx][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1551
1552 if (flags & WINED3DTTFF_PROJECTED) {
1553 projected = TRUE;
1554 switch (flags & ~WINED3DTTFF_PROJECTED) {
1555 case WINED3DTTFF_COUNT1: FIXME("WINED3DTTFF_PROJECTED with WINED3DTTFF_COUNT1?\n"); break;
1556 case WINED3DTTFF_COUNT2: mask = WINED3DSP_WRITEMASK_1; break;
1557 case WINED3DTTFF_COUNT3: mask = WINED3DSP_WRITEMASK_2; break;
1558 case WINED3DTTFF_COUNT4:
1559 case WINED3DTTFF_DISABLE: mask = WINED3DSP_WRITEMASK_3; break;
1560 }
1561 } else {
1562 projected = FALSE;
1563 }
1564 } else if (hex_version < WINED3DPS_VERSION(2,0)) {
1565 DWORD src_mod = arg->src[0] & WINED3DSP_SRCMOD_MASK;
1566 sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1567
1568 if (src_mod == WINED3DSPSM_DZ) {
1569 projected = TRUE;
1570 mask = WINED3DSP_WRITEMASK_2;
1571 } else if (src_mod == WINED3DSPSM_DW) {
1572 projected = TRUE;
1573 mask = WINED3DSP_WRITEMASK_3;
1574 } else {
1575 projected = FALSE;
1576 }
1577 } else {
1578 sampler_idx = arg->src[1] & WINED3DSP_REGNUM_MASK;
1579 if(arg->opcode_token & WINED3DSI_TEXLD_PROJECT) {
1580 /* ps 2.0 texldp instruction always divides by the fourth component. */
1581 projected = TRUE;
1582 mask = WINED3DSP_WRITEMASK_3;
1583 } else {
1584 projected = FALSE;
1585 }
1586 }
1587
1588 sampler_type = arg->reg_maps->samplers[sampler_idx] & WINED3DSP_TEXTURETYPE_MASK;
1589 shader_glsl_get_sample_function(sampler_type, projected, &sample_function);
1590 mask |= sample_function.coord_mask;
1591
1592 if (hex_version < WINED3DPS_VERSION(2,0)) {
1593 shader_glsl_get_write_mask(arg->dst, dst_swizzle);
1594 } else {
1595 shader_glsl_get_swizzle(arg->src[1], FALSE, arg->dst, dst_swizzle);
1596 }
1597
1598 /* 1.0-1.3: Use destination register as coordinate source.
1599 1.4+: Use provided coordinate source register. */
1600 if (hex_version < WINED3DPS_VERSION(1,4)) {
1601 char coord_mask[6];
1602 shader_glsl_get_write_mask(mask, coord_mask);
1603 shader_addline(arg->buffer, "%s(Psampler%u, T%u%s)%s);\n",
1604 sample_function.name, sampler_idx, sampler_idx, coord_mask, dst_swizzle);
1605 } else {
1606 glsl_src_param_t coord_param;
1607 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], mask, &coord_param);
1608 shader_addline(arg->buffer, "%s(Psampler%u, %s)%s);\n",
1609 sample_function.name, sampler_idx, coord_param.param_str, dst_swizzle);
1610 }
1611 }
1612
1613 void shader_glsl_texldl(SHADER_OPCODE_ARG* arg) {
1614 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*)arg->shader;
1615 glsl_sample_function_t sample_function;
1616 glsl_src_param_t coord_param, lod_param;
1617 char dst_swizzle[6];
1618 DWORD sampler_type;
1619 DWORD sampler_idx;
1620
1621 shader_glsl_append_dst(arg->buffer, arg);
1622 shader_glsl_get_swizzle(arg->src[1], FALSE, arg->dst, dst_swizzle);
1623
1624 sampler_idx = arg->src[1] & WINED3DSP_REGNUM_MASK;
1625 sampler_type = arg->reg_maps->samplers[sampler_idx] & WINED3DSP_TEXTURETYPE_MASK;
1626 shader_glsl_get_sample_function(sampler_type, FALSE, &sample_function);
1627 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], sample_function.coord_mask, &coord_param);
1628
1629 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_3, &lod_param);
1630
1631 if (shader_is_pshader_version(This->baseShader.hex_version)) {
1632 /* The GLSL spec claims the Lod sampling functions are only supported in vertex shaders.
1633 * However, they seem to work just fine in fragment shaders as well. */
1634 WARN("Using %sLod in fragment shader.\n", sample_function.name);
1635 shader_addline(arg->buffer, "%sLod(Psampler%u, %s, %s)%s);\n",
1636 sample_function.name, sampler_idx, coord_param.param_str, lod_param.param_str, dst_swizzle);
1637 } else {
1638 shader_addline(arg->buffer, "%sLod(Vsampler%u, %s, %s)%s);\n",
1639 sample_function.name, sampler_idx, coord_param.param_str, lod_param.param_str, dst_swizzle);
1640 }
1641 }
1642
1643 void pshader_glsl_texcoord(SHADER_OPCODE_ARG* arg) {
1644
1645 /* FIXME: Make this work for more than just 2D textures */
1646
1647 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1648 SHADER_BUFFER* buffer = arg->buffer;
1649 DWORD hex_version = This->baseShader.hex_version;
1650 DWORD write_mask;
1651 char dst_mask[6];
1652
1653 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1654 shader_glsl_get_write_mask(write_mask, dst_mask);
1655
1656 if (hex_version != WINED3DPS_VERSION(1,4)) {
1657 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1658 shader_addline(buffer, "clamp(gl_TexCoord[%u], 0.0, 1.0)%s);\n", reg, dst_mask);
1659 } else {
1660 DWORD reg = arg->src[0] & WINED3DSP_REGNUM_MASK;
1661 DWORD src_mod = arg->src[0] & WINED3DSP_SRCMOD_MASK;
1662 char dst_swizzle[6];
1663
1664 shader_glsl_get_swizzle(arg->src[0], FALSE, write_mask, dst_swizzle);
1665
1666 if (src_mod == WINED3DSPSM_DZ) {
1667 glsl_src_param_t div_param;
1668 size_t mask_size = shader_glsl_get_write_mask_size(write_mask);
1669 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_2, &div_param);
1670
1671 if (mask_size > 1) {
1672 shader_addline(buffer, "gl_TexCoord[%u]%s / vec%d(%s));\n", reg, dst_swizzle, mask_size, div_param.param_str);
1673 } else {
1674 shader_addline(buffer, "gl_TexCoord[%u]%s / %s);\n", reg, dst_swizzle, div_param.param_str);
1675 }
1676 } else if (src_mod == WINED3DSPSM_DW) {
1677 glsl_src_param_t div_param;
1678 size_t mask_size = shader_glsl_get_write_mask_size(write_mask);
1679 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_3, &div_param);
1680
1681 if (mask_size > 1) {
1682 shader_addline(buffer, "gl_TexCoord[%u]%s / vec%d(%s));\n", reg, dst_swizzle, mask_size, div_param.param_str);
1683 } else {
1684 shader_addline(buffer, "gl_TexCoord[%u]%s / %s);\n", reg, dst_swizzle, div_param.param_str);
1685 }
1686 } else {
1687 shader_addline(buffer, "gl_TexCoord[%u]%s);\n", reg, dst_swizzle);
1688 }
1689 }
1690 }
1691
1692 /** Process the WINED3DSIO_TEXDP3TEX instruction in GLSL:
1693 * Take a 3-component dot product of the TexCoord[dstreg] and src,
1694 * then perform a 1D texture lookup from stage dstregnum, place into dst. */
1695 void pshader_glsl_texdp3tex(SHADER_OPCODE_ARG* arg) {
1696 glsl_src_param_t src0_param;
1697 char dst_mask[6];
1698 DWORD sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1699 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1700
1701 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1702
1703 shader_glsl_append_dst(arg->buffer, arg);
1704 shader_glsl_get_write_mask(arg->dst, dst_mask);
1705 shader_addline(arg->buffer, "texture2D(Psampler%u, vec2(dot(gl_TexCoord[%u].xyz, %s), 0.5))%s);\n",
1706 sampler_idx, sampler_idx, src0_param.param_str, dst_mask);
1707 }
1708
1709 /** Process the WINED3DSIO_TEXDP3 instruction in GLSL:
1710 * Take a 3-component dot product of the TexCoord[dstreg] and src. */
1711 void pshader_glsl_texdp3(SHADER_OPCODE_ARG* arg) {
1712 glsl_src_param_t src0_param;
1713 DWORD dstreg = arg->dst & WINED3DSP_REGNUM_MASK;
1714 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1715 DWORD dst_mask;
1716 size_t mask_size;
1717
1718 dst_mask = shader_glsl_append_dst(arg->buffer, arg);
1719 mask_size = shader_glsl_get_write_mask_size(dst_mask);
1720 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1721
1722 if (mask_size > 1) {
1723 shader_addline(arg->buffer, "vec%d(dot(T%u.xyz, %s)));\n", mask_size, dstreg, src0_param.param_str);
1724 } else {
1725 shader_addline(arg->buffer, "dot(T%u.xyz, %s));\n", dstreg, src0_param.param_str);
1726 }
1727 }
1728
1729 /** Process the WINED3DSIO_TEXDEPTH instruction in GLSL:
1730 * Calculate the depth as dst.x / dst.y */
1731 void pshader_glsl_texdepth(SHADER_OPCODE_ARG* arg) {
1732 glsl_dst_param_t dst_param;
1733
1734 shader_glsl_add_dst_param(arg, arg->dst, 0, &dst_param);
1735
1736 shader_addline(arg->buffer, "gl_FragDepth = %s.x / %s.y;\n", dst_param.reg_name, dst_param.reg_name);
1737 }
1738
1739 /** Process the WINED3DSIO_TEXM3X2DEPTH instruction in GLSL:
1740 * Last row of a 3x2 matrix multiply, use the result to calculate the depth:
1741 * Calculate tmp0.y = TexCoord[dstreg] . src.xyz; (tmp0.x has already been calculated)
1742 * depth = (tmp0.y == 0.0) ? 1.0 : tmp0.x / tmp0.y
1743 */
1744 void pshader_glsl_texm3x2depth(SHADER_OPCODE_ARG* arg) {
1745 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1746 DWORD dstreg = arg->dst & WINED3DSP_REGNUM_MASK;
1747 glsl_src_param_t src0_param;
1748
1749 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1750
1751 shader_addline(arg->buffer, "tmp0.y = dot(T%u.xyz, %s);\n", dstreg, src0_param.param_str);
1752 shader_addline(arg->buffer, "gl_FragDepth = (tmp0.y == 0.0) ? 1.0 : clamp(tmp0.x / tmp0.y, 0.0, 1.0);\n");
1753 }
1754
1755 /** Process the WINED3DSIO_TEXM3X2PAD instruction in GLSL
1756 * Calculate the 1st of a 2-row matrix multiplication. */
1757 void pshader_glsl_texm3x2pad(SHADER_OPCODE_ARG* arg) {
1758 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1759 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1760 SHADER_BUFFER* buffer = arg->buffer;
1761 glsl_src_param_t src0_param;
1762
1763 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1764 shader_addline(buffer, "tmp0.x = dot(T%u.xyz, %s);\n", reg, src0_param.param_str);
1765 }
1766
1767 /** Process the WINED3DSIO_TEXM3X3PAD instruction in GLSL
1768 * Calculate the 1st or 2nd row of a 3-row matrix multiplication. */
1769 void pshader_glsl_texm3x3pad(SHADER_OPCODE_ARG* arg) {
1770
1771 IWineD3DPixelShaderImpl* shader = (IWineD3DPixelShaderImpl*) arg->shader;
1772 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1773 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1774 SHADER_BUFFER* buffer = arg->buffer;
1775 SHADER_PARSE_STATE* current_state = &shader->baseShader.parse_state;
1776 glsl_src_param_t src0_param;
1777
1778 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1779 shader_addline(buffer, "tmp0.%c = dot(T%u.xyz, %s);\n", 'x' + current_state->current_row, reg, src0_param.param_str);
1780 current_state->texcoord_w[current_state->current_row++] = reg;
1781 }
1782
1783 void pshader_glsl_texm3x2tex(SHADER_OPCODE_ARG* arg) {
1784 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1785 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1786 SHADER_BUFFER* buffer = arg->buffer;
1787 glsl_src_param_t src0_param;
1788 char dst_mask[6];
1789
1790 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1791 shader_addline(buffer, "tmp0.y = dot(T%u.xyz, %s);\n", reg, src0_param.param_str);
1792
1793 shader_glsl_append_dst(buffer, arg);
1794 shader_glsl_get_write_mask(arg->dst, dst_mask);
1795
1796 /* Sample the texture using the calculated coordinates */
1797 shader_addline(buffer, "texture2D(Psampler%u, tmp0.xy)%s);\n", reg, dst_mask);
1798 }
1799
1800 /** Process the WINED3DSIO_TEXM3X3TEX instruction in GLSL
1801 * Perform the 3rd row of a 3x3 matrix multiply, then sample the texture using the calculated coordinates */
1802 void pshader_glsl_texm3x3tex(SHADER_OPCODE_ARG* arg) {
1803 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1804 glsl_src_param_t src0_param;
1805 char dst_mask[6];
1806 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1807 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1808 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1809 DWORD sampler_type = arg->reg_maps->samplers[reg] & WINED3DSP_TEXTURETYPE_MASK;
1810 glsl_sample_function_t sample_function;
1811
1812 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1813 shader_addline(arg->buffer, "tmp0.z = dot(T%u.xyz, %s);\n", reg, src0_param.param_str);
1814
1815 shader_glsl_append_dst(arg->buffer, arg);
1816 shader_glsl_get_write_mask(arg->dst, dst_mask);
1817 shader_glsl_get_sample_function(sampler_type, FALSE, &sample_function);
1818
1819 /* Sample the texture using the calculated coordinates */
1820 shader_addline(arg->buffer, "%s(Psampler%u, tmp0.xyz)%s);\n", sample_function.name, reg, dst_mask);
1821
1822 current_state->current_row = 0;
1823 }
1824
1825 /** Process the WINED3DSIO_TEXM3X3 instruction in GLSL
1826 * Perform the 3rd row of a 3x3 matrix multiply */
1827 void pshader_glsl_texm3x3(SHADER_OPCODE_ARG* arg) {
1828 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1829 glsl_src_param_t src0_param;
1830 char dst_mask[6];
1831 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1832 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1833 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1834
1835 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1836
1837 shader_glsl_append_dst(arg->buffer, arg);
1838 shader_glsl_get_write_mask(arg->dst, dst_mask);
1839 shader_addline(arg->buffer, "vec4(tmp.xy, dot(T%u.xyz, %s), 1.0)%s);\n", reg, src0_param.param_str, dst_mask);
1840
1841 current_state->current_row = 0;
1842 }
1843
1844 /** Process the WINED3DSIO_TEXM3X3SPEC instruction in GLSL
1845 * Peform the final texture lookup based on the previous 2 3x3 matrix multiplies */
1846 void pshader_glsl_texm3x3spec(SHADER_OPCODE_ARG* arg) {
1847
1848 IWineD3DPixelShaderImpl* shader = (IWineD3DPixelShaderImpl*) arg->shader;
1849 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1850 glsl_src_param_t src0_param;
1851 glsl_src_param_t src1_param;
1852 char dst_mask[6];
1853 SHADER_BUFFER* buffer = arg->buffer;
1854 SHADER_PARSE_STATE* current_state = &shader->baseShader.parse_state;
1855 DWORD stype = arg->reg_maps->samplers[reg] & WINED3DSP_TEXTURETYPE_MASK;
1856 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1857 glsl_sample_function_t sample_function;
1858
1859 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1860 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], src_mask, &src1_param);
1861
1862 /* Perform the last matrix multiply operation */
1863 shader_addline(buffer, "tmp0.z = dot(T%u.xyz, %s);\n", reg, src0_param.param_str);
1864
1865 /* Calculate reflection vector, 2*(tmp0.src1)*tmp0-src1
1866 * This is equivalent to reflect(-src1, tmp0); */
1867 shader_addline(buffer, "tmp0.xyz = reflect(-(%s), tmp0.xyz);\n", src1_param.param_str);
1868
1869 shader_glsl_append_dst(buffer, arg);
1870 shader_glsl_get_write_mask(arg->dst, dst_mask);
1871 shader_glsl_get_sample_function(stype, FALSE, &sample_function);
1872
1873 /* Sample the texture */
1874 shader_addline(buffer, "%s(Psampler%u, tmp0.xyz)%s);\n", sample_function.name, reg, dst_mask);
1875
1876 current_state->current_row = 0;
1877 }
1878
1879 /** Process the WINED3DSIO_TEXM3X3VSPEC instruction in GLSL
1880 * Peform the final texture lookup based on the previous 2 3x3 matrix multiplies */
1881 void pshader_glsl_texm3x3vspec(SHADER_OPCODE_ARG* arg) {
1882
1883 IWineD3DPixelShaderImpl* shader = (IWineD3DPixelShaderImpl*) arg->shader;
1884 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1885 SHADER_BUFFER* buffer = arg->buffer;
1886 SHADER_PARSE_STATE* current_state = &shader->baseShader.parse_state;
1887 glsl_src_param_t src0_param;
1888 char dst_mask[6];
1889 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1890 DWORD sampler_type = arg->reg_maps->samplers[reg] & WINED3DSP_TEXTURETYPE_MASK;
1891 glsl_sample_function_t sample_function;
1892
1893 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1894
1895 /* Perform the last matrix multiply operation */
1896 shader_addline(buffer, "tmp0.z = dot(vec3(T%u), vec3(%s));\n", reg, src0_param.param_str);
1897
1898 /* Construct the eye-ray vector from w coordinates */
1899 shader_addline(buffer, "tmp1.xyz = normalize(vec3(gl_TexCoord[%u].w, gl_TexCoord[%u].w, gl_TexCoord[%u].w));\n",
1900 current_state->texcoord_w[0], current_state->texcoord_w[1], reg);
1901
1902 /* Calculate reflection vector (Assume normal is normalized): RF = 2*(tmp0.tmp1)*tmp0-tmp1
1903 * This is equivalent to reflect(-tmp1, tmp0); */
1904 shader_addline(buffer, "tmp0.xyz = reflect(-tmp1.xyz, tmp0.xyz);\n");
1905
1906 shader_glsl_append_dst(buffer, arg);
1907 shader_glsl_get_write_mask(arg->dst, dst_mask);
1908 shader_glsl_get_sample_function(sampler_type, FALSE, &sample_function);
1909
1910 /* Sample the texture using the calculated coordinates */
1911 shader_addline(buffer, "%s(Psampler%u, tmp0.xyz)%s);\n", sample_function.name, reg, dst_mask);
1912
1913 current_state->current_row = 0;
1914 }
1915
1916 /** Process the WINED3DSIO_TEXBEM instruction in GLSL.
1917 * Apply a fake bump map transform.
1918 * texbem is pshader <= 1.3 only, this saves a few version checks
1919 */
1920 void pshader_glsl_texbem(SHADER_OPCODE_ARG* arg) {
1921 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1922 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1923 char dst_swizzle[6];
1924 glsl_sample_function_t sample_function;
1925 glsl_src_param_t coord_param;
1926 DWORD sampler_type;
1927 DWORD sampler_idx;
1928 DWORD mask;
1929 DWORD flags;
1930 char coord_mask[6];
1931
1932 sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1933 flags = deviceImpl->stateBlock->textureState[sampler_idx][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1934
1935 sampler_type = arg->reg_maps->samplers[sampler_idx] & WINED3DSP_TEXTURETYPE_MASK;
1936 shader_glsl_get_sample_function(sampler_type, FALSE, &sample_function);
1937 mask = sample_function.coord_mask;
1938
1939 shader_glsl_get_write_mask(arg->dst, dst_swizzle);
1940
1941 shader_glsl_get_write_mask(mask, coord_mask);
1942
1943 /* with projective textures, texbem only divides the static texture coord, not the displacement,
1944 * so we can't let the GL handle this.
1945 */
1946 if (flags & WINED3DTTFF_PROJECTED) {
1947 DWORD div_mask=0;
1948 char coord_div_mask[3];
1949 switch (flags & ~WINED3DTTFF_PROJECTED) {
1950 case WINED3DTTFF_COUNT1: FIXME("WINED3DTTFF_PROJECTED with WINED3DTTFF_COUNT1?\n"); break;
1951 case WINED3DTTFF_COUNT2: div_mask = WINED3DSP_WRITEMASK_1; break;
1952 case WINED3DTTFF_COUNT3: div_mask = WINED3DSP_WRITEMASK_2; break;
1953 case WINED3DTTFF_COUNT4:
1954 case WINED3DTTFF_DISABLE: div_mask = WINED3DSP_WRITEMASK_3; break;
1955 }
1956 shader_glsl_get_write_mask(div_mask, coord_div_mask);
1957 shader_addline(arg->buffer, "T%u%s /= T%u%s;\n", sampler_idx, coord_mask, sampler_idx, coord_div_mask);
1958 }
1959
1960 shader_glsl_append_dst(arg->buffer, arg);
1961 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0|WINED3DSP_WRITEMASK_1, &coord_param);
1962 shader_addline(arg->buffer, "%s(Psampler%u, T%u%s + vec4(bumpenvmat * %s, 0.0, 0.0)%s )%s);\n",
1963 sample_function.name, sampler_idx, sampler_idx, coord_mask, coord_param.param_str, coord_mask, dst_swizzle);
1964 }
1965
1966 void pshader_glsl_bem(SHADER_OPCODE_ARG* arg) {
1967 glsl_src_param_t src0_param, src1_param;
1968
1969 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0|WINED3DSP_WRITEMASK_1, &src0_param);
1970 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_0|WINED3DSP_WRITEMASK_1, &src1_param);
1971
1972 shader_glsl_append_dst(arg->buffer, arg);
1973 shader_addline(arg->buffer, "%s + bumpenvmat * %s);\n",
1974 src0_param.param_str, src1_param.param_str);
1975 }
1976
1977 /** Process the WINED3DSIO_TEXREG2AR instruction in GLSL
1978 * Sample 2D texture at dst using the alpha & red (wx) components of src as texture coordinates */
1979 void pshader_glsl_texreg2ar(SHADER_OPCODE_ARG* arg) {
1980 glsl_src_param_t src0_param;
1981 DWORD sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1982 char dst_mask[6];
1983
1984 shader_glsl_append_dst(arg->buffer, arg);
1985 shader_glsl_get_write_mask(arg->dst, dst_mask);
1986 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_ALL, &src0_param);
1987
1988 shader_addline(arg->buffer, "texture2D(Psampler%u, %s.wx)%s);\n", sampler_idx, src0_param.reg_name, dst_mask);
1989 }
1990
1991 /** Process the WINED3DSIO_TEXREG2GB instruction in GLSL
1992 * Sample 2D texture at dst using the green & blue (yz) components of src as texture coordinates */
1993 void pshader_glsl_texreg2gb(SHADER_OPCODE_ARG* arg) {
1994 glsl_src_param_t src0_param;
1995 DWORD sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1996 char dst_mask[6];
1997
1998 shader_glsl_append_dst(arg->buffer, arg);
1999 shader_glsl_get_write_mask(arg->dst, dst_mask);
2000 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_ALL, &src0_param);
2001
2002 shader_addline(arg->buffer, "texture2D(Psampler%u, %s.yz)%s);\n", sampler_idx, src0_param.reg_name, dst_mask);
2003 }
2004
2005 /** Process the WINED3DSIO_TEXREG2RGB instruction in GLSL
2006 * Sample texture at dst using the rgb (xyz) components of src as texture coordinates */
2007 void pshader_glsl_texreg2rgb(SHADER_OPCODE_ARG* arg) {
2008 glsl_src_param_t src0_param;
2009 char dst_mask[6];
2010 DWORD sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
2011 DWORD sampler_type = arg->reg_maps->samplers[sampler_idx] & WINED3DSP_TEXTURETYPE_MASK;
2012 glsl_sample_function_t sample_function;
2013
2014 shader_glsl_append_dst(arg->buffer, arg);
2015 shader_glsl_get_write_mask(arg->dst, dst_mask);
2016 shader_glsl_get_sample_function(sampler_type, FALSE, &sample_function);
2017 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], sample_function.coord_mask, &src0_param);
2018
2019 shader_addline(arg->buffer, "%s(Psampler%u, %s)%s);\n", sample_function.name, sampler_idx, src0_param.param_str, dst_mask);
2020 }
2021
2022 /** Process the WINED3DSIO_TEXKILL instruction in GLSL.
2023 * If any of the first 3 components are < 0, discard this pixel */
2024 void pshader_glsl_texkill(SHADER_OPCODE_ARG* arg) {
2025 glsl_dst_param_t dst_param;
2026
2027 shader_glsl_add_dst_param(arg, arg->dst, 0, &dst_param);
2028 shader_addline(arg->buffer, "if (any(lessThan(%s.xyz, vec3(0.0)))) discard;\n", dst_param.reg_name);
2029 }
2030
2031 /** Process the WINED3DSIO_DP2ADD instruction in GLSL.
2032 * dst = dot2(src0, src1) + src2 */
2033 void pshader_glsl_dp2add(SHADER_OPCODE_ARG* arg) {
2034 glsl_src_param_t src0_param;
2035 glsl_src_param_t src1_param;
2036 glsl_src_param_t src2_param;
2037 DWORD write_mask;
2038 size_t mask_size;
2039
2040 write_mask = shader_glsl_append_dst(arg->buffer, arg);
2041 mask_size = shader_glsl_get_write_mask_size(write_mask);
2042
2043 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1, &src0_param);
2044 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1, &src1_param);
2045 shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], WINED3DSP_WRITEMASK_0, &src2_param);
2046
2047 shader_addline(arg->buffer, "dot(%s, %s) + %s);\n", src0_param.param_str, src1_param.param_str, src2_param.param_str);
2048 }
2049
2050 void pshader_glsl_input_pack(
2051 SHADER_BUFFER* buffer,
2052 semantic* semantics_in) {
2053
2054 unsigned int i;
2055
2056 for (i = 0; i < MAX_REG_INPUT; i++) {
2057
2058 DWORD usage_token = semantics_in[i].usage;
2059 DWORD register_token = semantics_in[i].reg;
2060 DWORD usage, usage_idx;
2061 char reg_mask[6];
2062
2063 /* Uninitialized */
2064 if (!usage_token) continue;
2065 usage = (usage_token & WINED3DSP_DCL_USAGE_MASK) >> WINED3DSP_DCL_USAGE_SHIFT;
2066 usage_idx = (usage_token & WINED3DSP_DCL_USAGEINDEX_MASK) >> WINED3DSP_DCL_USAGEINDEX_SHIFT;
2067 shader_glsl_get_write_mask(register_token, reg_mask);
2068
2069 switch(usage) {
2070
2071 case WINED3DDECLUSAGE_COLOR:
2072 if (usage_idx == 0)
2073 shader_addline(buffer, "IN%u%s = vec4(gl_Color)%s;\n",
2074 i, reg_mask, reg_mask);
2075 else if (usage_idx == 1)
2076 shader_addline(buffer, "IN%u%s = vec4(gl_SecondaryColor)%s;\n",
2077 i, reg_mask, reg_mask);
2078 else
2079 shader_addline(buffer, "IN%u%s = vec4(unsupported_color_input)%s;\n",
2080 i, reg_mask, reg_mask);
2081 break;
2082
2083 case WINED3DDECLUSAGE_TEXCOORD:
2084 shader_addline(buffer, "IN%u%s = vec4(gl_TexCoord[%u])%s;\n",
2085 i, reg_mask, usage_idx, reg_mask );
2086 break;
2087
2088 case WINED3DDECLUSAGE_FOG:
2089 shader_addline(buffer, "IN%u%s = vec4(gl_FogFragCoord)%s;\n",
2090 i, reg_mask, reg_mask);
2091 break;
2092
2093 default:
2094 shader_addline(buffer, "IN%u%s = vec4(unsupported_input)%s;\n",
2095 i, reg_mask, reg_mask);
2096 }
2097 }
2098 }
2099
2100 /*********************************************
2101 * Vertex Shader Specific Code begins here
2102 ********************************************/
2103
2104 void vshader_glsl_output_unpack(
2105 SHADER_BUFFER* buffer,
2106 semantic* semantics_out) {
2107
2108 unsigned int i;
2109
2110 for (i = 0; i < MAX_REG_OUTPUT; i++) {
2111
2112 DWORD usage_token = semantics_out[i].usage;
2113 DWORD register_token = semantics_out[i].reg;
2114 DWORD usage, usage_idx;
2115 char reg_mask[6];
2116
2117 /* Uninitialized */
2118 if (!usage_token) continue;
2119
2120 usage = (usage_token & WINED3DSP_DCL_USAGE_MASK) >> WINED3DSP_DCL_USAGE_SHIFT;
2121 usage_idx = (usage_token & WINED3DSP_DCL_USAGEINDEX_MASK) >> WINED3DSP_DCL_USAGEINDEX_SHIFT;
2122 shader_glsl_get_write_mask(register_token, reg_mask);
2123
2124 switch(usage) {
2125
2126 case WINED3DDECLUSAGE_COLOR:
2127 if (usage_idx == 0)
2128 shader_addline(buffer, "gl_FrontColor%s = OUT%u%s;\n", reg_mask, i, reg_mask);
2129 else if (usage_idx == 1)
2130 shader_addline(buffer, "gl_FrontSecondaryColor%s = OUT%u%s;\n", reg_mask, i, reg_mask);
2131 else
2132 shader_addline(buffer, "unsupported_color_output%s = OUT%u%s;\n", reg_mask, i, reg_mask);
2133 break;
2134
2135 case WINED3DDECLUSAGE_POSITION:
2136 shader_addline(buffer, "gl_Position%s = OUT%u%s;\n", reg_mask, i, reg_mask);
2137 break;
2138
2139 case WINED3DDECLUSAGE_TEXCOORD:
2140 shader_addline(buffer, "gl_TexCoord[%u]%s = OUT%u%s;\n",
2141 usage_idx, reg_mask, i, reg_mask);
2142 break;
2143
2144 case WINED3DDECLUSAGE_PSIZE:
2145 shader_addline(buffer, "gl_PointSize = OUT%u.x;\n", i);
2146 break;
2147
2148 case WINED3DDECLUSAGE_FOG:
2149 shader_addline(buffer, "gl_FogFragCoord = OUT%u%s;\n", i, reg_mask);
2150 break;
2151
2152 default:
2153 shader_addline(buffer, "unsupported_output%s = OUT%u%s;\n", reg_mask, i, reg_mask);
2154 }
2155 }
2156 }
2157
2158 static void add_glsl_program_entry(IWineD3DDeviceImpl *device, struct glsl_shader_prog_link *entry) {
2159 glsl_program_key_t *key;
2160
2161 key = HeapAlloc(GetProcessHeap(), 0, sizeof(glsl_program_key_t));
2162 key->vshader = entry->vshader;
2163 key->pshader = entry->pshader;
2164
2165 hash_table_put(device->glsl_program_lookup, key, entry);
2166 }
2167
2168 static struct glsl_shader_prog_link *get_glsl_program_entry(IWineD3DDeviceImpl *device,
2169 GLhandleARB vshader, GLhandleARB pshader) {
2170 glsl_program_key_t key;
2171
2172 key.vshader = vshader;
2173 key.pshader = pshader;
2174
2175 return (struct glsl_shader_prog_link *)hash_table_get(device->glsl_program_lookup, &key);
2176 }
2177
2178 void delete_glsl_program_entry(IWineD3DDevice *iface, struct glsl_shader_prog_link *entry) {
2179 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
2180 WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
2181 glsl_program_key_t *key;
2182
2183 key = HeapAlloc(GetProcessHeap(), 0, sizeof(glsl_program_key_t));
2184 key->vshader = entry->vshader;
2185 key->pshader = entry->pshader;
2186 hash_table_remove(This->glsl_program_lookup, key);
2187
2188 GL_EXTCALL(glDeleteObjectARB(entry->programId));
2189 if (entry->vshader) list_remove(&entry->vshader_entry);
2190 if (entry->pshader) list_remove(&entry->pshader_entry);
2191 HeapFree(GetProcessHeap(), 0, entry->vuniformF_locations);
2192 HeapFree(GetProcessHeap(), 0, entry->puniformF_locations);
2193 HeapFree(GetProcessHeap(), 0, entry);
2194 }
2195
2196 /** Sets the GLSL program ID for the given pixel and vertex shader combination.
2197 * It sets the programId on the current StateBlock (because it should be called
2198 * inside of the DrawPrimitive() part of the render loop).
2199 *
2200 * If a program for the given combination does not exist, create one, and store
2201 * the program in the hash table. If it creates a program, it will link the
2202 * given objects, too.
2203 */
2204 static void set_glsl_shader_program(IWineD3DDevice *iface, BOOL use_ps, BOOL use_vs) {
2205 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
2206 WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
2207 IWineD3DPixelShader *pshader = This->stateBlock->pixelShader;
2208 IWineD3DVertexShader *vshader = This->stateBlock->vertexShader;
2209 struct glsl_shader_prog_link *entry = NULL;
2210 GLhandleARB programId = 0;
2211 int i;
2212 char glsl_name[8];
2213
2214 GLhandleARB vshader_id = use_vs ? ((IWineD3DBaseShaderImpl*)vshader)->baseShader.prgId : 0;
2215 GLhandleARB pshader_id = use_ps ? ((IWineD3DBaseShaderImpl*)pshader)->baseShader.prgId : 0;
2216 entry = get_glsl_program_entry(This, vshader_id, pshader_id);
2217 if (entry) {
2218 This->stateBlock->glsl_program = entry;
2219 return;
2220 }
2221
2222 /* If we get to this point, then no matching program exists, so we create one */
2223 programId = GL_EXTCALL(glCreateProgramObjectARB());
2224 TRACE("Created new GLSL shader program %u\n", programId);
2225
2226 /* Create the entry */
2227 entry = HeapAlloc(GetProcessHeap(), 0, sizeof(struct glsl_shader_prog_link));
2228 entry->programId = programId;
2229 entry->vshader = vshader_id;
2230 entry->pshader = pshader_id;
2231 /* Add the hash table entry */
2232 add_glsl_program_entry(This, entry);
2233
2234 /* Set the current program */
2235 This->stateBlock->glsl_program = entry;
2236
2237 /* Attach GLSL vshader */
2238 if (vshader_id) {
2239 int max_attribs = 16; /* TODO: Will this always be the case? It is at the moment... */
2240 char tmp_name[10];
2241
2242 TRACE("Attaching GLSL shader object %u to program %u\n", vshader_id, programId);
2243 GL_EXTCALL(glAttachObjectARB(programId, vshader_id));
2244 checkGLcall("glAttachObjectARB");
2245
2246 /* Bind vertex attributes to a corresponding index number to match
2247 * the same index numbers as ARB_vertex_programs (makes loading
2248 * vertex attributes simpler). With this method, we can use the
2249 * exact same code to load the attributes later for both ARB and
2250 * GLSL shaders.
2251 *
2252 * We have to do this here because we need to know the Program ID
2253 * in order to make the bindings work, and it has to be done prior
2254 * to linking the GLSL program. */
2255 for (i = 0; i < max_attribs; ++i) {
2256 snprintf(tmp_name, sizeof(tmp_name), "attrib%i", i);
2257 GL_EXTCALL(glBindAttribLocationARB(programId, i, tmp_name));
2258 }
2259 checkGLcall("glBindAttribLocationARB");
2260
2261 list_add_head(&((IWineD3DBaseShaderImpl *)vshader)->baseShader.linked_programs, &entry->vshader_entry);
2262 }
2263
2264 /* Attach GLSL pshader */
2265 if (pshader_id) {
2266 TRACE("Attaching GLSL shader object %u to program %u\n", pshader_id, programId);
2267 GL_EXTCALL(glAttachObjectARB(programId, pshader_id));
2268 checkGLcall("glAttachObjectARB");
2269
2270 list_add_head(&((IWineD3DBaseShaderImpl *)pshader)->baseShader.linked_programs, &entry->pshader_entry);
2271 }
2272
2273 /* Link the program */
2274 TRACE("Linking GLSL shader program %u\n", programId);
2275 GL_EXTCALL(glLinkProgramARB(programId));
2276 print_glsl_info_log(&GLINFO_LOCATION, programId);
2277
2278 entry->vuniformF_locations = HeapAlloc(GetProcessHeap(), 0, sizeof(GLhandleARB) * GL_LIMITS(vshader_constantsF));
2279 for (i = 0; i < GL_LIMITS(vshader_constantsF); ++i) {
2280 snprintf(glsl_name, sizeof(glsl_name), "VC[%i]", i);
2281 entry->vuniformF_locations[i] = GL_EXTCALL(glGetUniformLocationARB(programId, glsl_name));
2282 }
2283 entry->puniformF_locations = HeapAlloc(GetProcessHeap(), 0, sizeof(GLhandleARB) * GL_LIMITS(pshader_constantsF));
2284 for (i = 0; i < GL_LIMITS(pshader_constantsF); ++i) {
2285 snprintf(glsl_name, sizeof(glsl_name), "PC[%i]", i);
2286 entry->puniformF_locations[i] = GL_EXTCALL(glGetUniformLocationARB(programId, glsl_name));
2287 }
2288 }
2289
2290 static GLhandleARB create_glsl_blt_shader(WineD3D_GL_Info *gl_info) {
2291 GLhandleARB program_id;
2292 GLhandleARB vshader_id, pshader_id;
2293 const char *blt_vshader[] = {
2294 "void main(void)\n"
2295 "{\n"
2296 " gl_Position = gl_Vertex;\n"
2297 " gl_FrontColor = vec4(1.0);\n"
2298 " gl_TexCoord[0].x = (gl_Vertex.x * 0.5) + 0.5;\n"
2299 " gl_TexCoord[0].y = (-gl_Vertex.y * 0.5) + 0.5;\n"
2300 "}\n"
2301 };
2302
2303 const char *blt_pshader[] = {
2304 "uniform sampler2D sampler;\n"
2305 "void main(void)\n"
2306 "{\n"
2307 " gl_FragDepth = texture2D(sampler, gl_TexCoord[0].xy).x;\n"
2308 "}\n"
2309 };
2310
2311 vshader_id = GL_EXTCALL(glCreateShaderObjectARB(GL_VERTEX_SHADER_ARB));
2312 GL_EXTCALL(glShaderSourceARB(vshader_id, 1, blt_vshader, NULL));
2313 GL_EXTCALL(glCompileShaderARB(vshader_id));
2314
2315 pshader_id = GL_EXTCALL(glCreateShaderObjectARB(GL_FRAGMENT_SHADER_ARB));
2316 GL_EXTCALL(glShaderSourceARB(pshader_id, 1, blt_pshader, NULL));
2317 GL_EXTCALL(glCompileShaderARB(pshader_id));
2318
2319 program_id = GL_EXTCALL(glCreateProgramObjectARB());
2320 GL_EXTCALL(glAttachObjectARB(program_id, vshader_id));
2321 GL_EXTCALL(glAttachObjectARB(program_id, pshader_id));
2322 GL_EXTCALL(glLinkProgramARB(program_id));
2323
2324 print_glsl_info_log(&GLINFO_LOCATION, program_id);
2325
2326 return program_id;
2327 }
2328
2329 static void shader_glsl_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {
2330 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
2331 WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
2332 GLhandleARB program_id = 0;
2333
2334 if (useVS || usePS) set_glsl_shader_program(iface, usePS, useVS);
2335 else This->stateBlock->glsl_program = NULL;
2336
2337 program_id = This->stateBlock->glsl_program ? This->stateBlock->glsl_program->programId : 0;
2338 if (program_id) TRACE("Using GLSL program %u\n", program_id);
2339 GL_EXTCALL(glUseProgramObjectARB(program_id));
2340 checkGLcall("glUseProgramObjectARB");
2341 }
2342
2343 static void shader_glsl_select_depth_blt(IWineD3DDevice *iface) {
2344 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
2345 WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
2346 static GLhandleARB program_id = 0;
2347 static GLhandleARB loc = -1;
2348
2349 if (!program_id) {
2350 program_id = create_glsl_blt_shader(gl_info);
2351 loc = GL_EXTCALL(glGetUniformLocationARB(program_id, "sampler"));
2352 }
2353
2354 GL_EXTCALL(glUseProgramObjectARB(program_id));
2355 GL_EXTCALL(glUniform1iARB(loc, 0));
2356 }
2357
2358 static void shader_glsl_cleanup(IWineD3DDevice *iface) {
2359 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
2360 WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
2361 GL_EXTCALL(glUseProgramObjectARB(0));
2362 }
2363
2364 const shader_backend_t glsl_shader_backend = {
2365 &shader_glsl_select,
2366 &shader_glsl_select_depth_blt,
2367 &shader_glsl_load_constants,
2368 &shader_glsl_cleanup
2369 };
Windows API implementation