?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
push aea352fc3df615e3f4b48daf6f897ea93ad1fffd
[wine/hacks.git] / dlls / wined3d / glsl_shader.c
blob1e77b96dc72fbbf56a0e26205a8954345f05d1e4
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< GL_LIMITS(fragment_samplers); ++i) {
89 if (stateBlock->textures[i] != NULL) {
90 snprintf(sampler_name, sizeof(sampler_name), "Psampler%d", i);
91 name_loc = GL_EXTCALL(glGetUniformLocationARB(programId, sampler_name));
92 if (name_loc != -1) {
93 int mapped_unit = stateBlock->wineD3DDevice->texUnitMap[i];
94 TRACE("Loading %s for texture %d\n", sampler_name, mapped_unit);
95 GL_EXTCALL(glUniform1iARB(name_loc, mapped_unit));
96 checkGLcall("glUniform1iARB");
97 }
98 }
99 }
100 }
101
102 static void shader_glsl_load_vsamplers(WineD3D_GL_Info *gl_info, IWineD3DStateBlock* iface) {
103 IWineD3DStateBlockImpl* stateBlock = (IWineD3DStateBlockImpl*) iface;
104 GLhandleARB programId = stateBlock->glsl_program->programId;
105 GLhandleARB name_loc;
106 char sampler_name[20];
107 int i;
108
109 for (i = 0; i < MAX_VERTEX_SAMPLERS; ++i) {
110 snprintf(sampler_name, sizeof(sampler_name), "Vsampler%d", i);
111 name_loc = GL_EXTCALL(glGetUniformLocationARB(programId, sampler_name));
112 if (name_loc != -1) {
113 int mapped_unit = stateBlock->wineD3DDevice->texUnitMap[MAX_FRAGMENT_SAMPLERS + i];
114 if (mapped_unit != -1 && mapped_unit < GL_LIMITS(combined_samplers)) {
115 TRACE("Loading %s for texture %d\n", sampler_name, mapped_unit);
116 GL_EXTCALL(glUniform1iARB(name_loc, mapped_unit));
117 checkGLcall("glUniform1iARB");
118 } else {
119 ERR("Trying to load sampler %s on unsupported unit %d\n", sampler_name, mapped_unit);
120 }
121 }
122 }
123 }
124
125 /**
126 * Loads floating point constants (aka uniforms) into the currently set GLSL program.
127 * When constant_list == NULL, it will load all the constants.
128 */
129 static void shader_glsl_load_constantsF(IWineD3DBaseShaderImpl* This, WineD3D_GL_Info *gl_info,
130 unsigned int max_constants, float* constants, GLhandleARB *constant_locations,
131 struct list *constant_list) {
132 constants_entry *constant;
133 local_constant* lconst;
134 GLhandleARB tmp_loc;
135 DWORD i, j;
136 DWORD *idx;
137
138 if (TRACE_ON(d3d_shader)) {
139 LIST_FOR_EACH_ENTRY(constant, constant_list, constants_entry, entry) {
140 idx = constant->idx;
141 j = constant->count;
142 while (j--) {
143 i = *idx++;
144 tmp_loc = constant_locations[i];
145 if (tmp_loc != -1) {
146 TRACE_(d3d_constants)("Loading constants %i: %f, %f, %f, %f\n", i,
147 constants[i * 4 + 0], constants[i * 4 + 1],
148 constants[i * 4 + 2], constants[i * 4 + 3]);
149 }
150 }
151 }
152 }
153 LIST_FOR_EACH_ENTRY(constant, constant_list, constants_entry, entry) {
154 idx = constant->idx;
155 j = constant->count;
156 while (j--) {
157 i = *idx++;
158 tmp_loc = constant_locations[i];
159 if (tmp_loc != -1) {
160 /* We found this uniform name in the program - go ahead and send the data */
161 GL_EXTCALL(glUniform4fvARB(tmp_loc, 1, constants + (i * 4)));
162 }
163 }
164 }
165 checkGLcall("glUniform4fvARB()");
166
167 /* Load immediate constants */
168 if (TRACE_ON(d3d_shader)) {
169 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
170 tmp_loc = constant_locations[lconst->idx];
171 if (tmp_loc != -1) {
172 GLfloat* values = (GLfloat*)lconst->value;
173 TRACE_(d3d_constants)("Loading local constants %i: %f, %f, %f, %f\n", lconst->idx,
174 values[0], values[1], values[2], values[3]);
175 }
176 }
177 }
178 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
179 tmp_loc = constant_locations[lconst->idx];
180 if (tmp_loc != -1) {
181 /* We found this uniform name in the program - go ahead and send the data */
182 GL_EXTCALL(glUniform4fvARB(tmp_loc, 1, (GLfloat*)lconst->value));
183 }
184 }
185 checkGLcall("glUniform4fvARB()");
186 }
187
188 /**
189 * Loads integer constants (aka uniforms) into the currently set GLSL program.
190 * When @constants_set == NULL, it will load all the constants.
191 */
192 static void shader_glsl_load_constantsI(
193 IWineD3DBaseShaderImpl* This,
194 WineD3D_GL_Info *gl_info,
195 GLhandleARB programId,
196 unsigned max_constants,
197 int* constants,
198 BOOL* constants_set) {
199
200 GLhandleARB tmp_loc;
201 int i;
202 char tmp_name[8];
203 char is_pshader = shader_is_pshader_version(This->baseShader.hex_version);
204 const char* prefix = is_pshader? "PI":"VI";
205 struct list* ptr;
206
207 for (i=0; i<max_constants; ++i) {
208 if (NULL == constants_set || constants_set[i]) {
209
210 TRACE_(d3d_constants)("Loading constants %i: %i, %i, %i, %i\n",
211 i, constants[i*4], constants[i*4+1], constants[i*4+2], constants[i*4+3]);
212
213 /* TODO: Benchmark and see if it would be beneficial to store the
214 * locations of the constants to avoid looking up each time */
215 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, i);
216 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
217 if (tmp_loc != -1) {
218 /* We found this uniform name in the program - go ahead and send the data */
219 GL_EXTCALL(glUniform4ivARB(tmp_loc, 1, &constants[i*4]));
220 checkGLcall("glUniform4ivARB");
221 }
222 }
223 }
224
225 /* Load immediate constants */
226 ptr = list_head(&This->baseShader.constantsI);
227 while (ptr) {
228 local_constant* lconst = LIST_ENTRY(ptr, struct local_constant, entry);
229 unsigned int idx = lconst->idx;
230 GLint* values = (GLint*) lconst->value;
231
232 TRACE_(d3d_constants)("Loading local constants %i: %i, %i, %i, %i\n", idx,
233 values[0], values[1], values[2], values[3]);
234
235 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, idx);
236 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
237 if (tmp_loc != -1) {
238 /* We found this uniform name in the program - go ahead and send the data */
239 GL_EXTCALL(glUniform4ivARB(tmp_loc, 1, values));
240 checkGLcall("glUniform4ivARB");
241 }
242 ptr = list_next(&This->baseShader.constantsI, ptr);
243 }
244 }
245
246 /**
247 * Loads boolean constants (aka uniforms) into the currently set GLSL program.
248 * When @constants_set == NULL, it will load all the constants.
249 */
250 static void shader_glsl_load_constantsB(
251 IWineD3DBaseShaderImpl* This,
252 WineD3D_GL_Info *gl_info,
253 GLhandleARB programId,
254 unsigned max_constants,
255 BOOL* constants,
256 BOOL* constants_set) {
257
258 GLhandleARB tmp_loc;
259 int i;
260 char tmp_name[8];
261 char is_pshader = shader_is_pshader_version(This->baseShader.hex_version);
262 const char* prefix = is_pshader? "PB":"VB";
263 struct list* ptr;
264
265 for (i=0; i<max_constants; ++i) {
266 if (NULL == constants_set || constants_set[i]) {
267
268 TRACE_(d3d_constants)("Loading constants %i: %i;\n", i, constants[i*4]);
269
270 /* TODO: Benchmark and see if it would be beneficial to store the
271 * locations of the constants to avoid looking up each time */
272 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, i);
273 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
274 if (tmp_loc != -1) {
275 /* We found this uniform name in the program - go ahead and send the data */
276 GL_EXTCALL(glUniform1ivARB(tmp_loc, 1, &constants[i*4]));
277 checkGLcall("glUniform1ivARB");
278 }
279 }
280 }
281
282 /* Load immediate constants */
283 ptr = list_head(&This->baseShader.constantsB);
284 while (ptr) {
285 local_constant* lconst = LIST_ENTRY(ptr, struct local_constant, entry);
286 unsigned int idx = lconst->idx;
287 GLint* values = (GLint*) lconst->value;
288
289 TRACE_(d3d_constants)("Loading local constants %i: %i\n", idx, values[0]);
290
291 snprintf(tmp_name, sizeof(tmp_name), "%s[%i]", prefix, idx);
292 tmp_loc = GL_EXTCALL(glGetUniformLocationARB(programId, tmp_name));
293 if (tmp_loc != -1) {
294 /* We found this uniform name in the program - go ahead and send the data */
295 GL_EXTCALL(glUniform1ivARB(tmp_loc, 1, values));
296 checkGLcall("glUniform1ivARB");
297 }
298 ptr = list_next(&This->baseShader.constantsB, ptr);
299 }
300 }
301
302
303
304 /**
305 * Loads the app-supplied constants into the currently set GLSL program.
306 */
307 void shader_glsl_load_constants(
308 IWineD3DDevice* device,
309 char usePixelShader,
310 char useVertexShader) {
311
312 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) device;
313 IWineD3DStateBlockImpl* stateBlock = deviceImpl->stateBlock;
314 WineD3D_GL_Info *gl_info = &deviceImpl->adapter->gl_info;
315
316 GLhandleARB *constant_locations;
317 struct list *constant_list;
318 GLhandleARB programId;
319 GLint pos;
320
321 if (!stateBlock->glsl_program) {
322 /* No GLSL program set - nothing to do. */
323 return;
324 }
325 programId = stateBlock->glsl_program->programId;
326
327 if (useVertexShader) {
328 IWineD3DBaseShaderImpl* vshader = (IWineD3DBaseShaderImpl*) stateBlock->vertexShader;
329 GLint pos;
330
331 constant_locations = stateBlock->glsl_program->vuniformF_locations;
332 constant_list = &stateBlock->set_vconstantsF;
333
334 /* Load vertex shader samplers */
335 shader_glsl_load_vsamplers(gl_info, (IWineD3DStateBlock*)stateBlock);
336
337 /* Load DirectX 9 float constants/uniforms for vertex shader */
338 shader_glsl_load_constantsF(vshader, gl_info, GL_LIMITS(vshader_constantsF),
339 stateBlock->vertexShaderConstantF, constant_locations, constant_list);
340
341 /* Load DirectX 9 integer constants/uniforms for vertex shader */
342 shader_glsl_load_constantsI(vshader, gl_info, programId, MAX_CONST_I,
343 stateBlock->vertexShaderConstantI,
344 stateBlock->set.vertexShaderConstantsI);
345
346 /* Load DirectX 9 boolean constants/uniforms for vertex shader */
347 shader_glsl_load_constantsB(vshader, gl_info, programId, MAX_CONST_B,
348 stateBlock->vertexShaderConstantB,
349 stateBlock->set.vertexShaderConstantsB);
350
351 /* Upload the position fixup params */
352 pos = GL_EXTCALL(glGetUniformLocationARB(programId, "posFixup"));
353 checkGLcall("glGetUniformLocationARB");
354 GL_EXTCALL(glUniform4fvARB(pos, 1, &deviceImpl->posFixup[0]));
355 checkGLcall("glUniform4fvARB");
356 }
357
358 if (usePixelShader) {
359
360 IWineD3DBaseShaderImpl* pshader = (IWineD3DBaseShaderImpl*) stateBlock->pixelShader;
361
362 constant_locations = stateBlock->glsl_program->puniformF_locations;
363 constant_list = &stateBlock->set_pconstantsF;
364
365 /* Load pixel shader samplers */
366 shader_glsl_load_psamplers(gl_info, (IWineD3DStateBlock*) stateBlock);
367
368 /* Load DirectX 9 float constants/uniforms for pixel shader */
369 shader_glsl_load_constantsF(pshader, gl_info, GL_LIMITS(pshader_constantsF),
370 stateBlock->pixelShaderConstantF, constant_locations, constant_list);
371
372 /* Load DirectX 9 integer constants/uniforms for pixel shader */
373 shader_glsl_load_constantsI(pshader, gl_info, programId, MAX_CONST_I,
374 stateBlock->pixelShaderConstantI,
375 stateBlock->set.pixelShaderConstantsI);
376
377 /* Load DirectX 9 boolean constants/uniforms for pixel shader */
378 shader_glsl_load_constantsB(pshader, gl_info, programId, MAX_CONST_B,
379 stateBlock->pixelShaderConstantB,
380 stateBlock->set.pixelShaderConstantsB);
381
382 /* Upload the environment bump map matrix if needed. The needsbumpmat member specifies the texture stage to load the matrix from.
383 * It can't be 0 for a valid texbem instruction.
384 */
385 if(((IWineD3DPixelShaderImpl *) pshader)->needsbumpmat != -1) {
386 float *data = (float *) &stateBlock->textureState[(int) ((IWineD3DPixelShaderImpl *) pshader)->needsbumpmat][WINED3DTSS_BUMPENVMAT00];
387 pos = GL_EXTCALL(glGetUniformLocationARB(programId, "bumpenvmat"));
388 checkGLcall("glGetUniformLocationARB");
389 GL_EXTCALL(glUniformMatrix2fvARB(pos, 1, 0, data));
390 checkGLcall("glUniform4fvARB");
391 }
392 }
393 }
394
395 /** Generate the variable & register declarations for the GLSL output target */
396 void shader_generate_glsl_declarations(
397 IWineD3DBaseShader *iface,
398 shader_reg_maps* reg_maps,
399 SHADER_BUFFER* buffer,
400 WineD3D_GL_Info* gl_info) {
401
402 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
403 int i;
404
405 /* There are some minor differences between pixel and vertex shaders */
406 char pshader = shader_is_pshader_version(This->baseShader.hex_version);
407 char prefix = pshader ? 'P' : 'V';
408
409 /* Prototype the subroutines */
410 for (i = 0; i < This->baseShader.limits.label; i++) {
411 if (reg_maps->labels[i])
412 shader_addline(buffer, "void subroutine%lu();\n", i);
413 }
414
415 /* Declare the constants (aka uniforms) */
416 if (This->baseShader.limits.constant_float > 0) {
417 unsigned max_constantsF = min(This->baseShader.limits.constant_float,
418 (pshader ? GL_LIMITS(pshader_constantsF) : GL_LIMITS(vshader_constantsF)));
419 shader_addline(buffer, "uniform vec4 %cC[%u];\n", prefix, max_constantsF);
420 }
421
422 if (This->baseShader.limits.constant_int > 0)
423 shader_addline(buffer, "uniform ivec4 %cI[%u];\n", prefix, This->baseShader.limits.constant_int);
424
425 if (This->baseShader.limits.constant_bool > 0)
426 shader_addline(buffer, "uniform bool %cB[%u];\n", prefix, This->baseShader.limits.constant_bool);
427
428 if(!pshader)
429 shader_addline(buffer, "uniform vec4 posFixup;\n");
430 else if(reg_maps->bumpmat != -1)
431 shader_addline(buffer, "uniform mat2 bumpenvmat;\n");
432
433 /* Declare texture samplers */
434 for (i = 0; i < This->baseShader.limits.sampler; i++) {
435 if (reg_maps->samplers[i]) {
436
437 DWORD stype = reg_maps->samplers[i] & WINED3DSP_TEXTURETYPE_MASK;
438 switch (stype) {
439
440 case WINED3DSTT_1D:
441 shader_addline(buffer, "uniform sampler1D %csampler%lu;\n", prefix, i);
442 break;
443 case WINED3DSTT_2D:
444 shader_addline(buffer, "uniform sampler2D %csampler%lu;\n", prefix, i);
445 break;
446 case WINED3DSTT_CUBE:
447 shader_addline(buffer, "uniform samplerCube %csampler%lu;\n", prefix, i);
448 break;
449 case WINED3DSTT_VOLUME:
450 shader_addline(buffer, "uniform sampler3D %csampler%lu;\n", prefix, i);
451 break;
452 default:
453 shader_addline(buffer, "uniform unsupported_sampler %csampler%lu;\n", prefix, i);
454 FIXME("Unrecognized sampler type: %#x\n", stype);
455 break;
456 }
457 }
458 }
459
460 /* Declare address variables */
461 for (i = 0; i < This->baseShader.limits.address; i++) {
462 if (reg_maps->address[i])
463 shader_addline(buffer, "ivec4 A%d;\n", i);
464 }
465
466 /* Declare texture coordinate temporaries and initialize them */
467 for (i = 0; i < This->baseShader.limits.texcoord; i++) {
468 if (reg_maps->texcoord[i])
469 shader_addline(buffer, "vec4 T%lu = gl_TexCoord[%lu];\n", i, i);
470 }
471
472 /* Declare input register temporaries */
473 for (i=0; i < This->baseShader.limits.packed_input; i++) {
474 if (reg_maps->packed_input[i])
475 shader_addline(buffer, "vec4 IN%lu;\n", i);
476 }
477
478 /* Declare output register temporaries */
479 for (i = 0; i < This->baseShader.limits.packed_output; i++) {
480 if (reg_maps->packed_output[i])
481 shader_addline(buffer, "vec4 OUT%lu;\n", i);
482 }
483
484 /* Declare temporary variables */
485 for(i = 0; i < This->baseShader.limits.temporary; i++) {
486 if (reg_maps->temporary[i])
487 shader_addline(buffer, "vec4 R%lu;\n", i);
488 }
489
490 /* Declare attributes */
491 for (i = 0; i < This->baseShader.limits.attributes; i++) {
492 if (reg_maps->attributes[i])
493 shader_addline(buffer, "attribute vec4 attrib%i;\n", i);
494 }
495
496 /* Declare loop register aL */
497 if (reg_maps->loop) {
498 shader_addline(buffer, "int aL;\n");
499 shader_addline(buffer, "int tmpInt;\n");
500 }
501
502 /* Temporary variables for matrix operations */
503 shader_addline(buffer, "vec4 tmp0;\n");
504 shader_addline(buffer, "vec4 tmp1;\n");
505
506 /* Start the main program */
507 shader_addline(buffer, "void main() {\n");
508 }
509
510 /*****************************************************************************
511 * Functions to generate GLSL strings from DirectX Shader bytecode begin here.
512 *
513 * For more information, see http://wiki.winehq.org/DirectX-Shaders
514 ****************************************************************************/
515
516 /* Prototypes */
517 static void shader_glsl_add_src_param(SHADER_OPCODE_ARG* arg, const DWORD param,
518 const DWORD addr_token, DWORD mask, glsl_src_param_t *src_param);
519
520 /** Used for opcode modifiers - They multiply the result by the specified amount */
521 static const char * const shift_glsl_tab[] = {
522 "", /* 0 (none) */
523 "2.0 * ", /* 1 (x2) */
524 "4.0 * ", /* 2 (x4) */
525 "8.0 * ", /* 3 (x8) */
526 "16.0 * ", /* 4 (x16) */
527 "32.0 * ", /* 5 (x32) */
528 "", /* 6 (x64) */
529 "", /* 7 (x128) */
530 "", /* 8 (d256) */
531 "", /* 9 (d128) */
532 "", /* 10 (d64) */
533 "", /* 11 (d32) */
534 "0.0625 * ", /* 12 (d16) */
535 "0.125 * ", /* 13 (d8) */
536 "0.25 * ", /* 14 (d4) */
537 "0.5 * " /* 15 (d2) */
538 };
539
540 /* Generate a GLSL parameter that does the input modifier computation and return the input register/mask to use */
541 static void shader_glsl_gen_modifier (
542 const DWORD instr,
543 const char *in_reg,
544 const char *in_regswizzle,
545 char *out_str) {
546
547 out_str[0] = 0;
548
549 if (instr == WINED3DSIO_TEXKILL)
550 return;
551
552 switch (instr & WINED3DSP_SRCMOD_MASK) {
553 case WINED3DSPSM_DZ: /* Need to handle this in the instructions itself (texld & texcrd). */
554 case WINED3DSPSM_DW:
555 case WINED3DSPSM_NONE:
556 sprintf(out_str, "%s%s", in_reg, in_regswizzle);
557 break;
558 case WINED3DSPSM_NEG:
559 sprintf(out_str, "-%s%s", in_reg, in_regswizzle);
560 break;
561 case WINED3DSPSM_NOT:
562 sprintf(out_str, "!%s%s", in_reg, in_regswizzle);
563 break;
564 case WINED3DSPSM_BIAS:
565 sprintf(out_str, "(%s%s - vec4(0.5)%s)", in_reg, in_regswizzle, in_regswizzle);
566 break;
567 case WINED3DSPSM_BIASNEG:
568 sprintf(out_str, "-(%s%s - vec4(0.5)%s)", in_reg, in_regswizzle, in_regswizzle);
569 break;
570 case WINED3DSPSM_SIGN:
571 sprintf(out_str, "(2.0 * (%s%s - 0.5))", in_reg, in_regswizzle);
572 break;
573 case WINED3DSPSM_SIGNNEG:
574 sprintf(out_str, "-(2.0 * (%s%s - 0.5))", in_reg, in_regswizzle);
575 break;
576 case WINED3DSPSM_COMP:
577 sprintf(out_str, "(1.0 - %s%s)", in_reg, in_regswizzle);
578 break;
579 case WINED3DSPSM_X2:
580 sprintf(out_str, "(2.0 * %s%s)", in_reg, in_regswizzle);
581 break;
582 case WINED3DSPSM_X2NEG:
583 sprintf(out_str, "-(2.0 * %s%s)", in_reg, in_regswizzle);
584 break;
585 case WINED3DSPSM_ABS:
586 sprintf(out_str, "abs(%s%s)", in_reg, in_regswizzle);
587 break;
588 case WINED3DSPSM_ABSNEG:
589 sprintf(out_str, "-abs(%s%s)", in_reg, in_regswizzle);
590 break;
591 default:
592 FIXME("Unhandled modifier %u\n", (instr & WINED3DSP_SRCMOD_MASK));
593 sprintf(out_str, "%s%s", in_reg, in_regswizzle);
594 }
595 }
596
597 /** Writes the GLSL variable name that corresponds to the register that the
598 * DX opcode parameter is trying to access */
599 static void shader_glsl_get_register_name(
600 const DWORD param,
601 const DWORD addr_token,
602 char* regstr,
603 BOOL* is_color,
604 SHADER_OPCODE_ARG* arg) {
605
606 /* oPos, oFog and oPts in D3D */
607 static const char * const hwrastout_reg_names[] = { "gl_Position", "gl_FogFragCoord", "gl_PointSize" };
608
609 DWORD reg = param & WINED3DSP_REGNUM_MASK;
610 DWORD regtype = shader_get_regtype(param);
611 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) arg->shader;
612 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
613 WineD3D_GL_Info* gl_info = &deviceImpl->adapter->gl_info;
614
615 char pshader = shader_is_pshader_version(This->baseShader.hex_version);
616 char tmpStr[50];
617
618 *is_color = FALSE;
619
620 switch (regtype) {
621 case WINED3DSPR_TEMP:
622 sprintf(tmpStr, "R%u", reg);
623 break;
624 case WINED3DSPR_INPUT:
625 if (pshader) {
626 /* Pixel shaders >= 3.0 */
627 if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3)
628 sprintf(tmpStr, "IN%u", reg);
629 else {
630 if (reg==0)
631 strcpy(tmpStr, "gl_Color");
632 else
633 strcpy(tmpStr, "gl_SecondaryColor");
634 }
635 } else {
636 if (vshader_input_is_color((IWineD3DVertexShader*) This, reg))
637 *is_color = TRUE;
638 sprintf(tmpStr, "attrib%u", reg);
639 }
640 break;
641 case WINED3DSPR_CONST:
642 {
643 const char* prefix = pshader? "PC":"VC";
644
645 /* Relative addressing */
646 if (param & WINED3DSHADER_ADDRMODE_RELATIVE) {
647
648 /* Relative addressing on shaders 2.0+ have a relative address token,
649 * prior to that, it was hard-coded as "A0.x" because there's only 1 register */
650 if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 2) {
651 glsl_src_param_t rel_param;
652 shader_glsl_add_src_param(arg, addr_token, 0, WINED3DSP_WRITEMASK_0, &rel_param);
653 sprintf(tmpStr, "%s[%s + %u]", prefix, rel_param.param_str, reg);
654 } else
655 sprintf(tmpStr, "%s[A0.x + %u]", prefix, reg);
656
657 } else
658 sprintf(tmpStr, "%s[%u]", prefix, reg);
659
660 break;
661 }
662 case WINED3DSPR_CONSTINT:
663 if (pshader)
664 sprintf(tmpStr, "PI[%u]", reg);
665 else
666 sprintf(tmpStr, "VI[%u]", reg);
667 break;
668 case WINED3DSPR_CONSTBOOL:
669 if (pshader)
670 sprintf(tmpStr, "PB[%u]", reg);
671 else
672 sprintf(tmpStr, "VB[%u]", reg);
673 break;
674 case WINED3DSPR_TEXTURE: /* case WINED3DSPR_ADDR: */
675 if (pshader) {
676 sprintf(tmpStr, "T%u", reg);
677 } else {
678 sprintf(tmpStr, "A%u", reg);
679 }
680 break;
681 case WINED3DSPR_LOOP:
682 sprintf(tmpStr, "aL");
683 break;
684 case WINED3DSPR_SAMPLER:
685 if (pshader)
686 sprintf(tmpStr, "Psampler%u", reg);
687 else
688 sprintf(tmpStr, "Vsampler%u", reg);
689 break;
690 case WINED3DSPR_COLOROUT:
691 if (reg >= GL_LIMITS(buffers)) {
692 WARN("Write to render target %u, only %d supported\n", reg, 4);
693 }
694 if (GL_SUPPORT(ARB_DRAW_BUFFERS)) {
695 sprintf(tmpStr, "gl_FragData[%u]", reg);
696 } else { /* On older cards with GLSL support like the GeforceFX there's only one buffer. */
697 sprintf(tmpStr, "gl_FragColor");
698 }
699 break;
700 case WINED3DSPR_RASTOUT:
701 sprintf(tmpStr, "%s", hwrastout_reg_names[reg]);
702 break;
703 case WINED3DSPR_DEPTHOUT:
704 sprintf(tmpStr, "gl_FragDepth");
705 break;
706 case WINED3DSPR_ATTROUT:
707 if (reg == 0) {
708 sprintf(tmpStr, "gl_FrontColor");
709 } else {
710 sprintf(tmpStr, "gl_FrontSecondaryColor");
711 }
712 break;
713 case WINED3DSPR_TEXCRDOUT:
714 /* Vertex shaders >= 3.0: WINED3DSPR_OUTPUT */
715 if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 3)
716 sprintf(tmpStr, "OUT%u", reg);
717 else
718 sprintf(tmpStr, "gl_TexCoord[%u]", reg);
719 break;
720 case WINED3DSPR_MISCTYPE:
721 if (reg == 0) {
722 /* vPos */
723 sprintf(tmpStr, "gl_FragCoord");
724 } else {
725 /* gl_FrontFacing could be used for vFace, but note that
726 * gl_FrontFacing is a bool, while vFace is a float for
727 * which the sign determines front/back */
728 FIXME("Unhandled misctype register %d\n", reg);
729 sprintf(tmpStr, "unrecognized_register");
730 }
731 break;
732 default:
733 FIXME("Unhandled register name Type(%d)\n", regtype);
734 sprintf(tmpStr, "unrecognized_register");
735 break;
736 }
737
738 strcat(regstr, tmpStr);
739 }
740
741 /* Get the GLSL write mask for the destination register */
742 static DWORD shader_glsl_get_write_mask(const DWORD param, char *write_mask) {
743 char *ptr = write_mask;
744 DWORD mask = param & WINED3DSP_WRITEMASK_ALL;
745
746 if (shader_is_scalar(param)) {
747 mask = WINED3DSP_WRITEMASK_0;
748 } else {
749 *ptr++ = '.';
750 if (param & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
751 if (param & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
752 if (param & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
753 if (param & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
754 }
755
756 *ptr = '\0';
757
758 return mask;
759 }
760
761 static size_t shader_glsl_get_write_mask_size(DWORD write_mask) {
762 size_t size = 0;
763
764 if (write_mask & WINED3DSP_WRITEMASK_0) ++size;
765 if (write_mask & WINED3DSP_WRITEMASK_1) ++size;
766 if (write_mask & WINED3DSP_WRITEMASK_2) ++size;
767 if (write_mask & WINED3DSP_WRITEMASK_3) ++size;
768
769 return size;
770 }
771
772 static void shader_glsl_get_swizzle(const DWORD param, BOOL fixup, DWORD mask, char *swizzle_str) {
773 /* For registers of type WINED3DDECLTYPE_D3DCOLOR, data is stored as "bgra",
774 * but addressed as "rgba". To fix this we need to swap the register's x
775 * and z components. */
776 DWORD swizzle = (param & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
777 const char *swizzle_chars = fixup ? "zyxw" : "xyzw";
778 char *ptr = swizzle_str;
779
780 if (!shader_is_scalar(param)) {
781 *ptr++ = '.';
782 /* swizzle bits fields: wwzzyyxx */
783 if (mask & WINED3DSP_WRITEMASK_0) *ptr++ = swizzle_chars[swizzle & 0x03];
784 if (mask & WINED3DSP_WRITEMASK_1) *ptr++ = swizzle_chars[(swizzle >> 2) & 0x03];
785 if (mask & WINED3DSP_WRITEMASK_2) *ptr++ = swizzle_chars[(swizzle >> 4) & 0x03];
786 if (mask & WINED3DSP_WRITEMASK_3) *ptr++ = swizzle_chars[(swizzle >> 6) & 0x03];
787 }
788
789 *ptr = '\0';
790 }
791
792 /* From a given parameter token, generate the corresponding GLSL string.
793 * Also, return the actual register name and swizzle in case the
794 * caller needs this information as well. */
795 static void shader_glsl_add_src_param(SHADER_OPCODE_ARG* arg, const DWORD param,
796 const DWORD addr_token, DWORD mask, glsl_src_param_t *src_param) {
797 BOOL is_color = FALSE;
798 char swizzle_str[6];
799
800 src_param->reg_name[0] = '\0';
801 src_param->param_str[0] = '\0';
802 swizzle_str[0] = '\0';
803
804 shader_glsl_get_register_name(param, addr_token, src_param->reg_name, &is_color, arg);
805
806 shader_glsl_get_swizzle(param, is_color, mask, swizzle_str);
807 shader_glsl_gen_modifier(param, src_param->reg_name, swizzle_str, src_param->param_str);
808 }
809
810 /* From a given parameter token, generate the corresponding GLSL string.
811 * Also, return the actual register name and swizzle in case the
812 * caller needs this information as well. */
813 static DWORD shader_glsl_add_dst_param(SHADER_OPCODE_ARG* arg, const DWORD param,
814 const DWORD addr_token, glsl_dst_param_t *dst_param) {
815 BOOL is_color = FALSE;
816
817 dst_param->mask_str[0] = '\0';
818 dst_param->reg_name[0] = '\0';
819
820 shader_glsl_get_register_name(param, addr_token, dst_param->reg_name, &is_color, arg);
821 return shader_glsl_get_write_mask(param, dst_param->mask_str);
822 }
823
824 /* Append the destination part of the instruction to the buffer, return the effective write mask */
825 static DWORD shader_glsl_append_dst_ext(SHADER_BUFFER *buffer, SHADER_OPCODE_ARG *arg, const DWORD param) {
826 glsl_dst_param_t dst_param;
827 DWORD mask;
828 int shift;
829
830 mask = shader_glsl_add_dst_param(arg, param, arg->dst_addr, &dst_param);
831
832 if(mask) {
833 shift = (param & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
834 shader_addline(buffer, "%s%s = %s(", dst_param.reg_name, dst_param.mask_str, shift_glsl_tab[shift]);
835 }
836
837 return mask;
838 }
839
840 /* Append the destination part of the instruction to the buffer, return the effective write mask */
841 static DWORD shader_glsl_append_dst(SHADER_BUFFER *buffer, SHADER_OPCODE_ARG *arg) {
842 return shader_glsl_append_dst_ext(buffer, arg, arg->dst);
843 }
844
845 /** Process GLSL instruction modifiers */
846 void shader_glsl_add_instruction_modifiers(SHADER_OPCODE_ARG* arg) {
847
848 DWORD mask = arg->dst & WINED3DSP_DSTMOD_MASK;
849
850 if (arg->opcode->dst_token && mask != 0) {
851 glsl_dst_param_t dst_param;
852
853 shader_glsl_add_dst_param(arg, arg->dst, 0, &dst_param);
854
855 if (mask & WINED3DSPDM_SATURATE) {
856 /* _SAT means to clamp the value of the register to between 0 and 1 */
857 shader_addline(arg->buffer, "%s%s = clamp(%s%s, 0.0, 1.0);\n", dst_param.reg_name,
858 dst_param.mask_str, dst_param.reg_name, dst_param.mask_str);
859 }
860 if (mask & WINED3DSPDM_MSAMPCENTROID) {
861 FIXME("_centroid modifier not handled\n");
862 }
863 if (mask & WINED3DSPDM_PARTIALPRECISION) {
864 /* MSDN says this modifier can be safely ignored, so that's what we'll do. */
865 }
866 }
867 }
868
869 static inline const char* shader_get_comp_op(
870 const DWORD opcode) {
871
872 DWORD op = (opcode & INST_CONTROLS_MASK) >> INST_CONTROLS_SHIFT;
873 switch (op) {
874 case COMPARISON_GT: return ">";
875 case COMPARISON_EQ: return "==";
876 case COMPARISON_GE: return ">=";
877 case COMPARISON_LT: return "<";
878 case COMPARISON_NE: return "!=";
879 case COMPARISON_LE: return "<=";
880 default:
881 FIXME("Unrecognized comparison value: %u\n", op);
882 return "(\?\?)";
883 }
884 }
885
886 static void shader_glsl_get_sample_function(DWORD sampler_type, BOOL projected, glsl_sample_function_t *sample_function) {
887 /* Note that there's no such thing as a projected cube texture. */
888 switch(sampler_type) {
889 case WINED3DSTT_1D:
890 sample_function->name = projected ? "texture1DProj" : "texture1D";
891 sample_function->coord_mask = WINED3DSP_WRITEMASK_0;
892 break;
893 case WINED3DSTT_2D:
894 sample_function->name = projected ? "texture2DProj" : "texture2D";
895 sample_function->coord_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1;
896 break;
897 case WINED3DSTT_CUBE:
898 sample_function->name = "textureCube";
899 sample_function->coord_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
900 break;
901 case WINED3DSTT_VOLUME:
902 sample_function->name = projected ? "texture3DProj" : "texture3D";
903 sample_function->coord_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
904 break;
905 default:
906 sample_function->name = "";
907 FIXME("Unrecognized sampler type: %#x;\n", sampler_type);
908 break;
909 }
910 }
911
912
913 /*****************************************************************************
914 *
915 * Begin processing individual instruction opcodes
916 *
917 ****************************************************************************/
918
919 /* Generate GLSL arithmetic functions (dst = src1 + src2) */
920 void shader_glsl_arith(SHADER_OPCODE_ARG* arg) {
921 CONST SHADER_OPCODE* curOpcode = arg->opcode;
922 SHADER_BUFFER* buffer = arg->buffer;
923 glsl_src_param_t src0_param;
924 glsl_src_param_t src1_param;
925 DWORD write_mask;
926 char op;
927
928 /* Determine the GLSL operator to use based on the opcode */
929 switch (curOpcode->opcode) {
930 case WINED3DSIO_MUL: op = '*'; break;
931 case WINED3DSIO_ADD: op = '+'; break;
932 case WINED3DSIO_SUB: op = '-'; break;
933 default:
934 op = ' ';
935 FIXME("Opcode %s not yet handled in GLSL\n", curOpcode->name);
936 break;
937 }
938
939 write_mask = shader_glsl_append_dst(buffer, arg);
940 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, &src0_param);
941 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, &src1_param);
942 shader_addline(buffer, "%s %c %s);\n", src0_param.param_str, op, src1_param.param_str);
943 }
944
945 /* Process the WINED3DSIO_MOV opcode using GLSL (dst = src) */
946 void shader_glsl_mov(SHADER_OPCODE_ARG* arg) {
947 IWineD3DBaseShaderImpl* shader = (IWineD3DBaseShaderImpl*) arg->shader;
948 SHADER_BUFFER* buffer = arg->buffer;
949 glsl_src_param_t src0_param;
950 DWORD write_mask;
951
952 write_mask = shader_glsl_append_dst(buffer, arg);
953 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, &src0_param);
954
955 /* In vs_1_1 WINED3DSIO_MOV can write to the address register. In later
956 * shader versions WINED3DSIO_MOVA is used for this. */
957 if ((WINED3DSHADER_VERSION_MAJOR(shader->baseShader.hex_version) == 1 &&
958 !shader_is_pshader_version(shader->baseShader.hex_version) &&
959 shader_get_regtype(arg->dst) == WINED3DSPR_ADDR) ||
960 arg->opcode->opcode == WINED3DSIO_MOVA) {
961 /* We need to *round* to the nearest int here. */
962 size_t mask_size = shader_glsl_get_write_mask_size(write_mask);
963 if (mask_size > 1) {
964 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);
965 } else {
966 shader_addline(buffer, "int(floor(abs(%s) + 0.5) * sign(%s)));\n", src0_param.param_str, src0_param.param_str);
967 }
968 } else {
969 shader_addline(buffer, "%s);\n", src0_param.param_str);
970 }
971 }
972
973 /* Process the dot product operators DP3 and DP4 in GLSL (dst = dot(src0, src1)) */
974 void shader_glsl_dot(SHADER_OPCODE_ARG* arg) {
975 CONST SHADER_OPCODE* curOpcode = arg->opcode;
976 SHADER_BUFFER* buffer = arg->buffer;
977 glsl_src_param_t src0_param;
978 glsl_src_param_t src1_param;
979 DWORD dst_write_mask, src_write_mask;
980 size_t dst_size = 0;
981
982 dst_write_mask = shader_glsl_append_dst(buffer, arg);
983 dst_size = shader_glsl_get_write_mask_size(dst_write_mask);
984
985 /* dp3 works on vec3, dp4 on vec4 */
986 if (curOpcode->opcode == WINED3DSIO_DP4) {
987 src_write_mask = WINED3DSP_WRITEMASK_ALL;
988 } else {
989 src_write_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
990 }
991
992 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_write_mask, &src0_param);
993 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], src_write_mask, &src1_param);
994
995 if (dst_size > 1) {
996 shader_addline(buffer, "vec%d(dot(%s, %s)));\n", dst_size, src0_param.param_str, src1_param.param_str);
997 } else {
998 shader_addline(buffer, "dot(%s, %s));\n", src0_param.param_str, src1_param.param_str);
999 }
1000 }
1001
1002 /* Note that this instruction has some restrictions. The destination write mask
1003 * can't contain the w component, and the source swizzles have to be .xyzw */
1004 void shader_glsl_cross(SHADER_OPCODE_ARG *arg) {
1005 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1006 glsl_src_param_t src0_param;
1007 glsl_src_param_t src1_param;
1008 char dst_mask[6];
1009
1010 shader_glsl_get_write_mask(arg->dst, dst_mask);
1011 shader_glsl_append_dst(arg->buffer, arg);
1012 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1013 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], src_mask, &src1_param);
1014 shader_addline(arg->buffer, "cross(%s, %s).%s);\n", src0_param.param_str, src1_param.param_str, dst_mask);
1015 }
1016
1017 /* Process the WINED3DSIO_POW instruction in GLSL (dst = |src0|^src1)
1018 * Src0 and src1 are scalars. Note that D3D uses the absolute of src0, while
1019 * GLSL uses the value as-is. */
1020 void shader_glsl_pow(SHADER_OPCODE_ARG *arg) {
1021 SHADER_BUFFER *buffer = arg->buffer;
1022 glsl_src_param_t src0_param;
1023 glsl_src_param_t src1_param;
1024 DWORD dst_write_mask;
1025 size_t dst_size;
1026
1027 dst_write_mask = shader_glsl_append_dst(buffer, arg);
1028 dst_size = shader_glsl_get_write_mask_size(dst_write_mask);
1029
1030 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src0_param);
1031 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_0, &src1_param);
1032
1033 if (dst_size > 1) {
1034 shader_addline(buffer, "vec%d(pow(abs(%s), %s)));\n", dst_size, src0_param.param_str, src1_param.param_str);
1035 } else {
1036 shader_addline(buffer, "pow(abs(%s), %s));\n", src0_param.param_str, src1_param.param_str);
1037 }
1038 }
1039
1040 /* Map the opcode 1-to-1 to the GL code (arg->dst = instruction(src0, src1, ...) */
1041 void shader_glsl_map2gl(SHADER_OPCODE_ARG* arg) {
1042 CONST SHADER_OPCODE* curOpcode = arg->opcode;
1043 SHADER_BUFFER* buffer = arg->buffer;
1044 glsl_src_param_t src_param;
1045 const char *instruction;
1046 char arguments[256];
1047 DWORD write_mask;
1048 unsigned i;
1049
1050 /* Determine the GLSL function to use based on the opcode */
1051 /* TODO: Possibly make this a table for faster lookups */
1052 switch (curOpcode->opcode) {
1053 case WINED3DSIO_MIN: instruction = "min"; break;
1054 case WINED3DSIO_MAX: instruction = "max"; break;
1055 case WINED3DSIO_RSQ: instruction = "inversesqrt"; break;
1056 case WINED3DSIO_ABS: instruction = "abs"; break;
1057 case WINED3DSIO_FRC: instruction = "fract"; break;
1058 case WINED3DSIO_NRM: instruction = "normalize"; break;
1059 case WINED3DSIO_LOGP:
1060 case WINED3DSIO_LOG: instruction = "log2"; break;
1061 case WINED3DSIO_EXP: instruction = "exp2"; break;
1062 case WINED3DSIO_SGN: instruction = "sign"; break;
1063 default: instruction = "";
1064 FIXME("Opcode %s not yet handled in GLSL\n", curOpcode->name);
1065 break;
1066 }
1067
1068 write_mask = shader_glsl_append_dst(buffer, arg);
1069
1070 arguments[0] = '\0';
1071 if (curOpcode->num_params > 0) {
1072 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, &src_param);
1073 strcat(arguments, src_param.param_str);
1074 for (i = 2; i < curOpcode->num_params; ++i) {
1075 strcat(arguments, ", ");
1076 shader_glsl_add_src_param(arg, arg->src[i-1], arg->src_addr[i-1], write_mask, &src_param);
1077 strcat(arguments, src_param.param_str);
1078 }
1079 }
1080
1081 shader_addline(buffer, "%s(%s));\n", instruction, arguments);
1082 }
1083
1084 /** Process the WINED3DSIO_EXPP instruction in GLSL:
1085 * For shader model 1.x, do the following (and honor the writemask, so use a temporary variable):
1086 * dst.x = 2^(floor(src))
1087 * dst.y = src - floor(src)
1088 * dst.z = 2^src (partial precision is allowed, but optional)
1089 * dst.w = 1.0;
1090 * For 2.0 shaders, just do this (honoring writemask and swizzle):
1091 * dst = 2^src; (partial precision is allowed, but optional)
1092 */
1093 void shader_glsl_expp(SHADER_OPCODE_ARG* arg) {
1094 IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)arg->shader;
1095 glsl_src_param_t src_param;
1096
1097 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src_param);
1098
1099 if (shader->baseShader.hex_version < WINED3DPS_VERSION(2,0)) {
1100 char dst_mask[6];
1101
1102 shader_addline(arg->buffer, "tmp0.x = exp2(floor(%s));\n", src_param.param_str);
1103 shader_addline(arg->buffer, "tmp0.y = %s - floor(%s);\n", src_param.param_str, src_param.param_str);
1104 shader_addline(arg->buffer, "tmp0.z = exp2(%s);\n", src_param.param_str);
1105 shader_addline(arg->buffer, "tmp0.w = 1.0;\n");
1106
1107 shader_glsl_append_dst(arg->buffer, arg);
1108 shader_glsl_get_write_mask(arg->dst, dst_mask);
1109 shader_addline(arg->buffer, "tmp0%s);\n", dst_mask);
1110 } else {
1111 DWORD write_mask;
1112 size_t mask_size;
1113
1114 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1115 mask_size = shader_glsl_get_write_mask_size(write_mask);
1116
1117 if (mask_size > 1) {
1118 shader_addline(arg->buffer, "vec%d(exp2(%s)));\n", mask_size, src_param.param_str);
1119 } else {
1120 shader_addline(arg->buffer, "exp2(%s));\n", src_param.param_str);
1121 }
1122 }
1123 }
1124
1125 /** Process the RCP (reciprocal or inverse) opcode in GLSL (dst = 1 / src) */
1126 void shader_glsl_rcp(SHADER_OPCODE_ARG* arg) {
1127 glsl_src_param_t src_param;
1128 DWORD write_mask;
1129 size_t mask_size;
1130
1131 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1132 mask_size = shader_glsl_get_write_mask_size(write_mask);
1133 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src_param);
1134
1135 if (mask_size > 1) {
1136 shader_addline(arg->buffer, "vec%d(1.0 / %s));\n", mask_size, src_param.param_str);
1137 } else {
1138 shader_addline(arg->buffer, "1.0 / %s);\n", src_param.param_str);
1139 }
1140 }
1141
1142 /** Process signed comparison opcodes in GLSL. */
1143 void shader_glsl_compare(SHADER_OPCODE_ARG* arg) {
1144 glsl_src_param_t src0_param;
1145 glsl_src_param_t src1_param;
1146 DWORD write_mask;
1147 size_t mask_size;
1148
1149 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1150 mask_size = shader_glsl_get_write_mask_size(write_mask);
1151 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, &src0_param);
1152 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, &src1_param);
1153
1154 if (mask_size > 1) {
1155 const char *compare;
1156
1157 switch(arg->opcode->opcode) {
1158 case WINED3DSIO_SLT: compare = "lessThan"; break;
1159 case WINED3DSIO_SGE: compare = "greaterThanEqual"; break;
1160 default: compare = "";
1161 FIXME("Can't handle opcode %s\n", arg->opcode->name);
1162 }
1163
1164 shader_addline(arg->buffer, "vec%d(%s(%s, %s)));\n", mask_size, compare,
1165 src0_param.param_str, src1_param.param_str);
1166 } else {
1167 const char *compare;
1168
1169 switch(arg->opcode->opcode) {
1170 case WINED3DSIO_SLT: compare = "<"; break;
1171 case WINED3DSIO_SGE: compare = ">="; break;
1172 default: compare = "";
1173 FIXME("Can't handle opcode %s\n", arg->opcode->name);
1174 }
1175
1176 shader_addline(arg->buffer, "(%s %s %s) ? 1.0 : 0.0);\n",
1177 src0_param.param_str, compare, src1_param.param_str);
1178 }
1179 }
1180
1181 /** Process CMP instruction in GLSL (dst = src0 >= 0.0 ? src1 : src2), per channel */
1182 void shader_glsl_cmp(SHADER_OPCODE_ARG* arg) {
1183 glsl_src_param_t src0_param;
1184 glsl_src_param_t src1_param;
1185 glsl_src_param_t src2_param;
1186 DWORD write_mask, cmp_channel = 0;
1187 unsigned int i, j;
1188
1189 /* Cycle through all source0 channels */
1190 for (i=0; i<4; i++) {
1191 write_mask = 0;
1192 /* Find the destination channels which use the current source0 channel */
1193 for (j=0; j<4; j++) {
1194 if ( ((arg->src[0] >> (WINED3DSP_SWIZZLE_SHIFT + 2*j)) & 0x3) == i ) {
1195 write_mask |= WINED3DSP_WRITEMASK_0 << j;
1196 cmp_channel = WINED3DSP_WRITEMASK_0 << j;
1197 }
1198 }
1199 write_mask = shader_glsl_append_dst_ext(arg->buffer, arg, arg->dst & (~WINED3DSP_SWIZZLE_MASK | write_mask));
1200 if (!write_mask) continue;
1201
1202 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], cmp_channel, &src0_param);
1203 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, &src1_param);
1204 shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], write_mask, &src2_param);
1205
1206 shader_addline(arg->buffer, "%s >= 0.0 ? %s : %s);\n",
1207 src0_param.param_str, src1_param.param_str, src2_param.param_str);
1208 }
1209 }
1210
1211 /** Process the CND opcode in GLSL (dst = (src0 > 0.5) ? src1 : src2) */
1212 /* For ps 1.1-1.3, only a single component of src0 is used. For ps 1.4
1213 * the compare is done per component of src0. */
1214 void shader_glsl_cnd(SHADER_OPCODE_ARG* arg) {
1215 IWineD3DBaseShaderImpl* shader = (IWineD3DBaseShaderImpl*) arg->shader;
1216 glsl_src_param_t src0_param;
1217 glsl_src_param_t src1_param;
1218 glsl_src_param_t src2_param;
1219 DWORD write_mask, cmp_channel = 0;
1220 unsigned int i, j;
1221
1222 if (shader->baseShader.hex_version < WINED3DPS_VERSION(1, 4)) {
1223 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1224 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src0_param);
1225 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, &src1_param);
1226 shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], write_mask, &src2_param);
1227 shader_addline(arg->buffer, "%s > 0.5 ? %s : %s);\n",
1228 src0_param.param_str, src1_param.param_str, src2_param.param_str);
1229 return;
1230 }
1231 /* Cycle through all source0 channels */
1232 for (i=0; i<4; i++) {
1233 write_mask = 0;
1234 /* Find the destination channels which use the current source0 channel */
1235 for (j=0; j<4; j++) {
1236 if ( ((arg->src[0] >> (WINED3DSP_SWIZZLE_SHIFT + 2*j)) & 0x3) == i ) {
1237 write_mask |= WINED3DSP_WRITEMASK_0 << j;
1238 cmp_channel = WINED3DSP_WRITEMASK_0 << j;
1239 }
1240 }
1241 write_mask = shader_glsl_append_dst_ext(arg->buffer, arg, arg->dst & (~WINED3DSP_SWIZZLE_MASK | write_mask));
1242 if (!write_mask) continue;
1243
1244 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], cmp_channel, &src0_param);
1245 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, &src1_param);
1246 shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], write_mask, &src2_param);
1247
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 }
1251 }
1252
1253 /** GLSL code generation for WINED3DSIO_MAD: Multiply the first 2 opcodes, then add the last */
1254 void shader_glsl_mad(SHADER_OPCODE_ARG* arg) {
1255 glsl_src_param_t src0_param;
1256 glsl_src_param_t src1_param;
1257 glsl_src_param_t src2_param;
1258 DWORD write_mask;
1259
1260 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1261 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, &src0_param);
1262 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, &src1_param);
1263 shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], write_mask, &src2_param);
1264 shader_addline(arg->buffer, "(%s * %s) + %s);\n",
1265 src0_param.param_str, src1_param.param_str, src2_param.param_str);
1266 }
1267
1268 /** Handles transforming all WINED3DSIO_M?x? opcodes for
1269 Vertex shaders to GLSL codes */
1270 void shader_glsl_mnxn(SHADER_OPCODE_ARG* arg) {
1271 int i;
1272 int nComponents = 0;
1273 SHADER_OPCODE_ARG tmpArg;
1274
1275 memset(&tmpArg, 0, sizeof(SHADER_OPCODE_ARG));
1276
1277 /* Set constants for the temporary argument */
1278 tmpArg.shader = arg->shader;
1279 tmpArg.buffer = arg->buffer;
1280 tmpArg.src[0] = arg->src[0];
1281 tmpArg.src_addr[0] = arg->src_addr[0];
1282 tmpArg.src_addr[1] = arg->src_addr[1];
1283 tmpArg.reg_maps = arg->reg_maps;
1284
1285 switch(arg->opcode->opcode) {
1286 case WINED3DSIO_M4x4:
1287 nComponents = 4;
1288 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP4);
1289 break;
1290 case WINED3DSIO_M4x3:
1291 nComponents = 3;
1292 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP4);
1293 break;
1294 case WINED3DSIO_M3x4:
1295 nComponents = 4;
1296 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP3);
1297 break;
1298 case WINED3DSIO_M3x3:
1299 nComponents = 3;
1300 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP3);
1301 break;
1302 case WINED3DSIO_M3x2:
1303 nComponents = 2;
1304 tmpArg.opcode = shader_get_opcode(arg->shader, WINED3DSIO_DP3);
1305 break;
1306 default:
1307 break;
1308 }
1309
1310 for (i = 0; i < nComponents; i++) {
1311 tmpArg.dst = ((arg->dst) & ~WINED3DSP_WRITEMASK_ALL)|(WINED3DSP_WRITEMASK_0<<i);
1312 tmpArg.src[1] = arg->src[1]+i;
1313 shader_glsl_dot(&tmpArg);
1314 }
1315 }
1316
1317 /**
1318 The LRP instruction performs a component-wise linear interpolation
1319 between the second and third operands using the first operand as the
1320 blend factor. Equation: (dst = src2 + src0 * (src1 - src2))
1321 This is equivalent to mix(src2, src1, src0);
1322 */
1323 void shader_glsl_lrp(SHADER_OPCODE_ARG* arg) {
1324 glsl_src_param_t src0_param;
1325 glsl_src_param_t src1_param;
1326 glsl_src_param_t src2_param;
1327 DWORD write_mask;
1328
1329 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1330
1331 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], write_mask, &src0_param);
1332 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], write_mask, &src1_param);
1333 shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], write_mask, &src2_param);
1334
1335 shader_addline(arg->buffer, "mix(%s, %s, %s));\n",
1336 src2_param.param_str, src1_param.param_str, src0_param.param_str);
1337 }
1338
1339 /** Process the WINED3DSIO_LIT instruction in GLSL:
1340 * dst.x = dst.w = 1.0
1341 * dst.y = (src0.x > 0) ? src0.x
1342 * dst.z = (src0.x > 0) ? ((src0.y > 0) ? pow(src0.y, src.w) : 0) : 0
1343 * where src.w is clamped at +- 128
1344 */
1345 void shader_glsl_lit(SHADER_OPCODE_ARG* arg) {
1346 glsl_src_param_t src0_param;
1347 glsl_src_param_t src1_param;
1348 glsl_src_param_t src3_param;
1349 char dst_mask[6];
1350
1351 shader_glsl_append_dst(arg->buffer, arg);
1352 shader_glsl_get_write_mask(arg->dst, dst_mask);
1353
1354 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src0_param);
1355 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_1, &src1_param);
1356 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_3, &src3_param);
1357
1358 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",
1359 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);
1360 }
1361
1362 /** Process the WINED3DSIO_DST instruction in GLSL:
1363 * dst.x = 1.0
1364 * dst.y = src0.x * src0.y
1365 * dst.z = src0.z
1366 * dst.w = src1.w
1367 */
1368 void shader_glsl_dst(SHADER_OPCODE_ARG* arg) {
1369 glsl_src_param_t src0y_param;
1370 glsl_src_param_t src0z_param;
1371 glsl_src_param_t src1y_param;
1372 glsl_src_param_t src1w_param;
1373 char dst_mask[6];
1374
1375 shader_glsl_append_dst(arg->buffer, arg);
1376 shader_glsl_get_write_mask(arg->dst, dst_mask);
1377
1378 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_1, &src0y_param);
1379 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_2, &src0z_param);
1380 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_1, &src1y_param);
1381 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_3, &src1w_param);
1382
1383 shader_addline(arg->buffer, "vec4(1.0, %s * %s, %s, %s))%s;\n",
1384 src0y_param.param_str, src1y_param.param_str, src0z_param.param_str, src1w_param.param_str, dst_mask);
1385 }
1386
1387 /** Process the WINED3DSIO_SINCOS instruction in GLSL:
1388 * VS 2.0 requires that specific cosine and sine constants be passed to this instruction so the hardware
1389 * can handle it. But, these functions are built-in for GLSL, so we can just ignore the last 2 params.
1390 *
1391 * dst.x = cos(src0.?)
1392 * dst.y = sin(src0.?)
1393 * dst.z = dst.z
1394 * dst.w = dst.w
1395 */
1396 void shader_glsl_sincos(SHADER_OPCODE_ARG* arg) {
1397 glsl_src_param_t src0_param;
1398 DWORD write_mask;
1399
1400 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1401 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src0_param);
1402
1403 switch (write_mask) {
1404 case WINED3DSP_WRITEMASK_0:
1405 shader_addline(arg->buffer, "cos(%s));\n", src0_param.param_str);
1406 break;
1407
1408 case WINED3DSP_WRITEMASK_1:
1409 shader_addline(arg->buffer, "sin(%s));\n", src0_param.param_str);
1410 break;
1411
1412 case (WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1):
1413 shader_addline(arg->buffer, "vec2(cos(%s), sin(%s)));\n", src0_param.param_str, src0_param.param_str);
1414 break;
1415
1416 default:
1417 ERR("Write mask should be .x, .y or .xy\n");
1418 break;
1419 }
1420 }
1421
1422 /** Process the WINED3DSIO_LOOP instruction in GLSL:
1423 * Start a for() loop where src1.y is the initial value of aL,
1424 * increment aL by src1.z for a total of src1.x iterations.
1425 * Need to use a temporary variable for this operation.
1426 */
1427 /* FIXME: I don't think nested loops will work correctly this way. */
1428 void shader_glsl_loop(SHADER_OPCODE_ARG* arg) {
1429 glsl_src_param_t src1_param;
1430
1431 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_ALL, &src1_param);
1432
1433 shader_addline(arg->buffer, "for (tmpInt = 0, aL = %s.y; tmpInt < %s.x; tmpInt++, aL += %s.z) {\n",
1434 src1_param.reg_name, src1_param.reg_name, src1_param.reg_name);
1435 }
1436
1437 void shader_glsl_end(SHADER_OPCODE_ARG* arg) {
1438 shader_addline(arg->buffer, "}\n");
1439 }
1440
1441 void shader_glsl_rep(SHADER_OPCODE_ARG* arg) {
1442 glsl_src_param_t src0_param;
1443
1444 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src0_param);
1445 shader_addline(arg->buffer, "for (tmpInt = 0; tmpInt < %s; tmpInt++) {\n", src0_param.param_str);
1446 }
1447
1448 void shader_glsl_if(SHADER_OPCODE_ARG* arg) {
1449 glsl_src_param_t src0_param;
1450
1451 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src0_param);
1452 shader_addline(arg->buffer, "if (%s) {\n", src0_param.param_str);
1453 }
1454
1455 void shader_glsl_ifc(SHADER_OPCODE_ARG* arg) {
1456 glsl_src_param_t src0_param;
1457 glsl_src_param_t src1_param;
1458
1459 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src0_param);
1460 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_0, &src1_param);
1461
1462 shader_addline(arg->buffer, "if (%s %s %s) {\n",
1463 src0_param.param_str, shader_get_comp_op(arg->opcode_token), src1_param.param_str);
1464 }
1465
1466 void shader_glsl_else(SHADER_OPCODE_ARG* arg) {
1467 shader_addline(arg->buffer, "} else {\n");
1468 }
1469
1470 void shader_glsl_break(SHADER_OPCODE_ARG* arg) {
1471 shader_addline(arg->buffer, "break;\n");
1472 }
1473
1474 /* FIXME: According to MSDN the compare is done per component. */
1475 void shader_glsl_breakc(SHADER_OPCODE_ARG* arg) {
1476 glsl_src_param_t src0_param;
1477 glsl_src_param_t src1_param;
1478
1479 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0, &src0_param);
1480 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_0, &src1_param);
1481
1482 shader_addline(arg->buffer, "if (%s %s %s) break;\n",
1483 src0_param.param_str, shader_get_comp_op(arg->opcode_token), src1_param.param_str);
1484 }
1485
1486 void shader_glsl_label(SHADER_OPCODE_ARG* arg) {
1487
1488 DWORD snum = (arg->src[0]) & WINED3DSP_REGNUM_MASK;
1489 shader_addline(arg->buffer, "}\n");
1490 shader_addline(arg->buffer, "void subroutine%lu () {\n", snum);
1491 }
1492
1493 void shader_glsl_call(SHADER_OPCODE_ARG* arg) {
1494 DWORD snum = (arg->src[0]) & WINED3DSP_REGNUM_MASK;
1495 shader_addline(arg->buffer, "subroutine%lu();\n", snum);
1496 }
1497
1498 void shader_glsl_callnz(SHADER_OPCODE_ARG* arg) {
1499 glsl_src_param_t src1_param;
1500
1501 DWORD snum = (arg->src[0]) & WINED3DSP_REGNUM_MASK;
1502 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_0, &src1_param);
1503 shader_addline(arg->buffer, "if (%s) subroutine%lu();\n", src1_param.param_str, snum);
1504 }
1505
1506 /*********************************************
1507 * Pixel Shader Specific Code begins here
1508 ********************************************/
1509 void pshader_glsl_tex(SHADER_OPCODE_ARG* arg) {
1510 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1511 DWORD hex_version = This->baseShader.hex_version;
1512 char dst_swizzle[6];
1513 glsl_sample_function_t sample_function;
1514 DWORD sampler_type;
1515 DWORD sampler_idx;
1516 BOOL projected;
1517 DWORD mask = 0;
1518
1519 /* All versions have a destination register */
1520 shader_glsl_append_dst(arg->buffer, arg);
1521
1522 /* 1.0-1.4: Use destination register as sampler source.
1523 * 2.0+: Use provided sampler source. */
1524 if (hex_version < WINED3DPS_VERSION(1,4)) {
1525 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1526 DWORD flags;
1527
1528 sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1529 flags = deviceImpl->stateBlock->textureState[sampler_idx][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1530
1531 if (flags & WINED3DTTFF_PROJECTED) {
1532 projected = TRUE;
1533 switch (flags & ~WINED3DTTFF_PROJECTED) {
1534 case WINED3DTTFF_COUNT1: FIXME("WINED3DTTFF_PROJECTED with WINED3DTTFF_COUNT1?\n"); break;
1535 case WINED3DTTFF_COUNT2: mask = WINED3DSP_WRITEMASK_1; break;
1536 case WINED3DTTFF_COUNT3: mask = WINED3DSP_WRITEMASK_2; break;
1537 case WINED3DTTFF_COUNT4:
1538 case WINED3DTTFF_DISABLE: mask = WINED3DSP_WRITEMASK_3; break;
1539 }
1540 } else {
1541 projected = FALSE;
1542 }
1543 } else if (hex_version < WINED3DPS_VERSION(2,0)) {
1544 DWORD src_mod = arg->src[0] & WINED3DSP_SRCMOD_MASK;
1545 sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1546
1547 if (src_mod == WINED3DSPSM_DZ) {
1548 projected = TRUE;
1549 mask = WINED3DSP_WRITEMASK_2;
1550 } else if (src_mod == WINED3DSPSM_DW) {
1551 projected = TRUE;
1552 mask = WINED3DSP_WRITEMASK_3;
1553 } else {
1554 projected = FALSE;
1555 }
1556 } else {
1557 sampler_idx = arg->src[1] & WINED3DSP_REGNUM_MASK;
1558 if(arg->opcode_token & WINED3DSI_TEXLD_PROJECT) {
1559 /* ps 2.0 texldp instruction always divides by the fourth component. */
1560 projected = TRUE;
1561 mask = WINED3DSP_WRITEMASK_3;
1562 } else {
1563 projected = FALSE;
1564 }
1565 }
1566
1567 sampler_type = arg->reg_maps->samplers[sampler_idx] & WINED3DSP_TEXTURETYPE_MASK;
1568 shader_glsl_get_sample_function(sampler_type, projected, &sample_function);
1569 mask |= sample_function.coord_mask;
1570
1571 if (hex_version < WINED3DPS_VERSION(2,0)) {
1572 shader_glsl_get_write_mask(arg->dst, dst_swizzle);
1573 } else {
1574 shader_glsl_get_swizzle(arg->src[1], FALSE, arg->dst, dst_swizzle);
1575 }
1576
1577 /* 1.0-1.3: Use destination register as coordinate source.
1578 1.4+: Use provided coordinate source register. */
1579 if (hex_version < WINED3DPS_VERSION(1,4)) {
1580 char coord_mask[6];
1581 shader_glsl_get_write_mask(mask, coord_mask);
1582 shader_addline(arg->buffer, "%s(Psampler%u, T%u%s)%s);\n",
1583 sample_function.name, sampler_idx, sampler_idx, coord_mask, dst_swizzle);
1584 } else {
1585 glsl_src_param_t coord_param;
1586 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], mask, &coord_param);
1587 shader_addline(arg->buffer, "%s(Psampler%u, %s)%s);\n",
1588 sample_function.name, sampler_idx, coord_param.param_str, dst_swizzle);
1589 }
1590 }
1591
1592 void shader_glsl_texldl(SHADER_OPCODE_ARG* arg) {
1593 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*)arg->shader;
1594 glsl_sample_function_t sample_function;
1595 glsl_src_param_t coord_param, lod_param;
1596 char dst_swizzle[6];
1597 DWORD sampler_type;
1598 DWORD sampler_idx;
1599
1600 shader_glsl_append_dst(arg->buffer, arg);
1601 shader_glsl_get_swizzle(arg->src[1], FALSE, arg->dst, dst_swizzle);
1602
1603 sampler_idx = arg->src[1] & WINED3DSP_REGNUM_MASK;
1604 sampler_type = arg->reg_maps->samplers[sampler_idx] & WINED3DSP_TEXTURETYPE_MASK;
1605 shader_glsl_get_sample_function(sampler_type, FALSE, &sample_function);
1606 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], sample_function.coord_mask, &coord_param);
1607
1608 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_3, &lod_param);
1609
1610 if (shader_is_pshader_version(This->baseShader.hex_version)) {
1611 /* The GLSL spec claims the Lod sampling functions are only supported in vertex shaders.
1612 * However, they seem to work just fine in fragment shaders as well. */
1613 WARN("Using %sLod in fragment shader.\n", sample_function.name);
1614 shader_addline(arg->buffer, "%sLod(Psampler%u, %s, %s)%s);\n",
1615 sample_function.name, sampler_idx, coord_param.param_str, lod_param.param_str, dst_swizzle);
1616 } else {
1617 shader_addline(arg->buffer, "%sLod(Vsampler%u, %s, %s)%s);\n",
1618 sample_function.name, sampler_idx, coord_param.param_str, lod_param.param_str, dst_swizzle);
1619 }
1620 }
1621
1622 void pshader_glsl_texcoord(SHADER_OPCODE_ARG* arg) {
1623
1624 /* FIXME: Make this work for more than just 2D textures */
1625
1626 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1627 SHADER_BUFFER* buffer = arg->buffer;
1628 DWORD hex_version = This->baseShader.hex_version;
1629 DWORD write_mask;
1630 char dst_mask[6];
1631
1632 write_mask = shader_glsl_append_dst(arg->buffer, arg);
1633 shader_glsl_get_write_mask(write_mask, dst_mask);
1634
1635 if (hex_version != WINED3DPS_VERSION(1,4)) {
1636 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1637 shader_addline(buffer, "clamp(gl_TexCoord[%u], 0.0, 1.0)%s);\n", reg, dst_mask);
1638 } else {
1639 DWORD reg = arg->src[0] & WINED3DSP_REGNUM_MASK;
1640 DWORD src_mod = arg->src[0] & WINED3DSP_SRCMOD_MASK;
1641 char dst_swizzle[6];
1642
1643 shader_glsl_get_swizzle(arg->src[0], FALSE, write_mask, dst_swizzle);
1644
1645 if (src_mod == WINED3DSPSM_DZ) {
1646 glsl_src_param_t div_param;
1647 size_t mask_size = shader_glsl_get_write_mask_size(write_mask);
1648 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_2, &div_param);
1649
1650 if (mask_size > 1) {
1651 shader_addline(buffer, "gl_TexCoord[%u]%s / vec%d(%s));\n", reg, dst_swizzle, mask_size, div_param.param_str);
1652 } else {
1653 shader_addline(buffer, "gl_TexCoord[%u]%s / %s);\n", reg, dst_swizzle, div_param.param_str);
1654 }
1655 } else if (src_mod == WINED3DSPSM_DW) {
1656 glsl_src_param_t div_param;
1657 size_t mask_size = shader_glsl_get_write_mask_size(write_mask);
1658 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_3, &div_param);
1659
1660 if (mask_size > 1) {
1661 shader_addline(buffer, "gl_TexCoord[%u]%s / vec%d(%s));\n", reg, dst_swizzle, mask_size, div_param.param_str);
1662 } else {
1663 shader_addline(buffer, "gl_TexCoord[%u]%s / %s);\n", reg, dst_swizzle, div_param.param_str);
1664 }
1665 } else {
1666 shader_addline(buffer, "gl_TexCoord[%u]%s);\n", reg, dst_swizzle);
1667 }
1668 }
1669 }
1670
1671 /** Process the WINED3DSIO_TEXDP3TEX instruction in GLSL:
1672 * Take a 3-component dot product of the TexCoord[dstreg] and src,
1673 * then perform a 1D texture lookup from stage dstregnum, place into dst. */
1674 void pshader_glsl_texdp3tex(SHADER_OPCODE_ARG* arg) {
1675 glsl_src_param_t src0_param;
1676 char dst_mask[6];
1677 DWORD sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1678 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1679
1680 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1681
1682 shader_glsl_append_dst(arg->buffer, arg);
1683 shader_glsl_get_write_mask(arg->dst, dst_mask);
1684 shader_addline(arg->buffer, "texture2D(Psampler%u, vec2(dot(gl_TexCoord[%u].xyz, %s), 0.5))%s);\n",
1685 sampler_idx, sampler_idx, src0_param.param_str, dst_mask);
1686 }
1687
1688 /** Process the WINED3DSIO_TEXDP3 instruction in GLSL:
1689 * Take a 3-component dot product of the TexCoord[dstreg] and src. */
1690 void pshader_glsl_texdp3(SHADER_OPCODE_ARG* arg) {
1691 glsl_src_param_t src0_param;
1692 DWORD dstreg = arg->dst & WINED3DSP_REGNUM_MASK;
1693 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1694 DWORD dst_mask;
1695 size_t mask_size;
1696
1697 dst_mask = shader_glsl_append_dst(arg->buffer, arg);
1698 mask_size = shader_glsl_get_write_mask_size(dst_mask);
1699 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1700
1701 if (mask_size > 1) {
1702 shader_addline(arg->buffer, "vec%d(dot(T%u.xyz, %s)));\n", mask_size, dstreg, src0_param.param_str);
1703 } else {
1704 shader_addline(arg->buffer, "dot(T%u.xyz, %s));\n", dstreg, src0_param.param_str);
1705 }
1706 }
1707
1708 /** Process the WINED3DSIO_TEXDEPTH instruction in GLSL:
1709 * Calculate the depth as dst.x / dst.y */
1710 void pshader_glsl_texdepth(SHADER_OPCODE_ARG* arg) {
1711 glsl_dst_param_t dst_param;
1712
1713 shader_glsl_add_dst_param(arg, arg->dst, 0, &dst_param);
1714
1715 shader_addline(arg->buffer, "gl_FragDepth = %s.x / %s.y;\n", dst_param.reg_name, dst_param.reg_name);
1716 }
1717
1718 /** Process the WINED3DSIO_TEXM3X2DEPTH instruction in GLSL:
1719 * Last row of a 3x2 matrix multiply, use the result to calculate the depth:
1720 * Calculate tmp0.y = TexCoord[dstreg] . src.xyz; (tmp0.x has already been calculated)
1721 * depth = (tmp0.y == 0.0) ? 1.0 : tmp0.x / tmp0.y
1722 */
1723 void pshader_glsl_texm3x2depth(SHADER_OPCODE_ARG* arg) {
1724 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1725 DWORD dstreg = arg->dst & WINED3DSP_REGNUM_MASK;
1726 glsl_src_param_t src0_param;
1727
1728 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1729
1730 shader_addline(arg->buffer, "tmp0.y = dot(T%u.xyz, %s);\n", dstreg, src0_param.param_str);
1731 shader_addline(arg->buffer, "gl_FragDepth = (tmp0.y == 0.0) ? 1.0 : clamp(tmp0.x / tmp0.y, 0.0, 1.0);\n");
1732 }
1733
1734 /** Process the WINED3DSIO_TEXM3X2PAD instruction in GLSL
1735 * Calculate the 1st of a 2-row matrix multiplication. */
1736 void pshader_glsl_texm3x2pad(SHADER_OPCODE_ARG* arg) {
1737 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1738 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1739 SHADER_BUFFER* buffer = arg->buffer;
1740 glsl_src_param_t src0_param;
1741
1742 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1743 shader_addline(buffer, "tmp0.x = dot(T%u.xyz, %s);\n", reg, src0_param.param_str);
1744 }
1745
1746 /** Process the WINED3DSIO_TEXM3X3PAD instruction in GLSL
1747 * Calculate the 1st or 2nd row of a 3-row matrix multiplication. */
1748 void pshader_glsl_texm3x3pad(SHADER_OPCODE_ARG* arg) {
1749
1750 IWineD3DPixelShaderImpl* shader = (IWineD3DPixelShaderImpl*) arg->shader;
1751 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1752 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1753 SHADER_BUFFER* buffer = arg->buffer;
1754 SHADER_PARSE_STATE* current_state = &shader->baseShader.parse_state;
1755 glsl_src_param_t src0_param;
1756
1757 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1758 shader_addline(buffer, "tmp0.%c = dot(T%u.xyz, %s);\n", 'x' + current_state->current_row, reg, src0_param.param_str);
1759 current_state->texcoord_w[current_state->current_row++] = reg;
1760 }
1761
1762 void pshader_glsl_texm3x2tex(SHADER_OPCODE_ARG* arg) {
1763 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1764 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1765 SHADER_BUFFER* buffer = arg->buffer;
1766 glsl_src_param_t src0_param;
1767 char dst_mask[6];
1768
1769 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1770 shader_addline(buffer, "tmp0.y = dot(T%u.xyz, %s);\n", reg, src0_param.param_str);
1771
1772 shader_glsl_append_dst(buffer, arg);
1773 shader_glsl_get_write_mask(arg->dst, dst_mask);
1774
1775 /* Sample the texture using the calculated coordinates */
1776 shader_addline(buffer, "texture2D(Psampler%u, tmp0.xy)%s);\n", reg, dst_mask);
1777 }
1778
1779 /** Process the WINED3DSIO_TEXM3X3TEX instruction in GLSL
1780 * Perform the 3rd row of a 3x3 matrix multiply, then sample the texture using the calculated coordinates */
1781 void pshader_glsl_texm3x3tex(SHADER_OPCODE_ARG* arg) {
1782 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1783 glsl_src_param_t src0_param;
1784 char dst_mask[6];
1785 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1786 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1787 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1788 DWORD sampler_type = arg->reg_maps->samplers[reg] & WINED3DSP_TEXTURETYPE_MASK;
1789 glsl_sample_function_t sample_function;
1790
1791 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1792 shader_addline(arg->buffer, "tmp0.z = dot(T%u.xyz, %s);\n", reg, src0_param.param_str);
1793
1794 shader_glsl_append_dst(arg->buffer, arg);
1795 shader_glsl_get_write_mask(arg->dst, dst_mask);
1796 shader_glsl_get_sample_function(sampler_type, FALSE, &sample_function);
1797
1798 /* Sample the texture using the calculated coordinates */
1799 shader_addline(arg->buffer, "%s(Psampler%u, tmp0.xyz)%s);\n", sample_function.name, reg, dst_mask);
1800
1801 current_state->current_row = 0;
1802 }
1803
1804 /** Process the WINED3DSIO_TEXM3X3 instruction in GLSL
1805 * Perform the 3rd row of a 3x3 matrix multiply */
1806 void pshader_glsl_texm3x3(SHADER_OPCODE_ARG* arg) {
1807 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1808 glsl_src_param_t src0_param;
1809 char dst_mask[6];
1810 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1811 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1812 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1813
1814 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1815
1816 shader_glsl_append_dst(arg->buffer, arg);
1817 shader_glsl_get_write_mask(arg->dst, dst_mask);
1818 shader_addline(arg->buffer, "vec4(tmp.xy, dot(T%u.xyz, %s), 1.0)%s);\n", reg, src0_param.param_str, dst_mask);
1819
1820 current_state->current_row = 0;
1821 }
1822
1823 /** Process the WINED3DSIO_TEXM3X3SPEC instruction in GLSL
1824 * Peform the final texture lookup based on the previous 2 3x3 matrix multiplies */
1825 void pshader_glsl_texm3x3spec(SHADER_OPCODE_ARG* arg) {
1826
1827 IWineD3DPixelShaderImpl* shader = (IWineD3DPixelShaderImpl*) arg->shader;
1828 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1829 glsl_src_param_t src0_param;
1830 glsl_src_param_t src1_param;
1831 char dst_mask[6];
1832 SHADER_BUFFER* buffer = arg->buffer;
1833 SHADER_PARSE_STATE* current_state = &shader->baseShader.parse_state;
1834 DWORD stype = arg->reg_maps->samplers[reg] & WINED3DSP_TEXTURETYPE_MASK;
1835 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1836 glsl_sample_function_t sample_function;
1837
1838 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1839 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], src_mask, &src1_param);
1840
1841 /* Perform the last matrix multiply operation */
1842 shader_addline(buffer, "tmp0.z = dot(T%u.xyz, %s);\n", reg, src0_param.param_str);
1843
1844 /* Calculate reflection vector, 2*(tmp0.src1)*tmp0-src1
1845 * This is equivalent to reflect(-src1, tmp0); */
1846 shader_addline(buffer, "tmp0.xyz = reflect(-(%s), tmp0.xyz);\n", src1_param.param_str);
1847
1848 shader_glsl_append_dst(buffer, arg);
1849 shader_glsl_get_write_mask(arg->dst, dst_mask);
1850 shader_glsl_get_sample_function(stype, FALSE, &sample_function);
1851
1852 /* Sample the texture */
1853 shader_addline(buffer, "%s(Psampler%u, tmp0.xyz)%s);\n", sample_function.name, reg, dst_mask);
1854
1855 current_state->current_row = 0;
1856 }
1857
1858 /** Process the WINED3DSIO_TEXM3X3VSPEC instruction in GLSL
1859 * Peform the final texture lookup based on the previous 2 3x3 matrix multiplies */
1860 void pshader_glsl_texm3x3vspec(SHADER_OPCODE_ARG* arg) {
1861
1862 IWineD3DPixelShaderImpl* shader = (IWineD3DPixelShaderImpl*) arg->shader;
1863 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1864 SHADER_BUFFER* buffer = arg->buffer;
1865 SHADER_PARSE_STATE* current_state = &shader->baseShader.parse_state;
1866 glsl_src_param_t src0_param;
1867 char dst_mask[6];
1868 DWORD src_mask = WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1 | WINED3DSP_WRITEMASK_2;
1869 DWORD sampler_type = arg->reg_maps->samplers[reg] & WINED3DSP_TEXTURETYPE_MASK;
1870 glsl_sample_function_t sample_function;
1871
1872 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], src_mask, &src0_param);
1873
1874 /* Perform the last matrix multiply operation */
1875 shader_addline(buffer, "tmp0.z = dot(vec3(T%u), vec3(%s));\n", reg, src0_param.param_str);
1876
1877 /* Construct the eye-ray vector from w coordinates */
1878 shader_addline(buffer, "tmp1.xyz = normalize(vec3(gl_TexCoord[%u].w, gl_TexCoord[%u].w, gl_TexCoord[%u].w));\n",
1879 current_state->texcoord_w[0], current_state->texcoord_w[1], reg);
1880
1881 /* Calculate reflection vector (Assume normal is normalized): RF = 2*(tmp0.tmp1)*tmp0-tmp1
1882 * This is equivalent to reflect(-tmp1, tmp0); */
1883 shader_addline(buffer, "tmp0.xyz = reflect(-tmp1.xyz, tmp0.xyz);\n");
1884
1885 shader_glsl_append_dst(buffer, arg);
1886 shader_glsl_get_write_mask(arg->dst, dst_mask);
1887 shader_glsl_get_sample_function(sampler_type, FALSE, &sample_function);
1888
1889 /* Sample the texture using the calculated coordinates */
1890 shader_addline(buffer, "%s(Psampler%u, tmp0.xyz)%s);\n", sample_function.name, reg, dst_mask);
1891
1892 current_state->current_row = 0;
1893 }
1894
1895 /** Process the WINED3DSIO_TEXBEM instruction in GLSL.
1896 * Apply a fake bump map transform.
1897 * texbem is pshader <= 1.3 only, this saves a few version checks
1898 */
1899 void pshader_glsl_texbem(SHADER_OPCODE_ARG* arg) {
1900 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1901 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1902 char dst_swizzle[6];
1903 glsl_sample_function_t sample_function;
1904 glsl_src_param_t coord_param;
1905 DWORD sampler_type;
1906 DWORD sampler_idx;
1907 DWORD mask;
1908 DWORD flags;
1909 char coord_mask[6];
1910
1911 sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1912 flags = deviceImpl->stateBlock->textureState[sampler_idx][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1913
1914 sampler_type = arg->reg_maps->samplers[sampler_idx] & WINED3DSP_TEXTURETYPE_MASK;
1915 shader_glsl_get_sample_function(sampler_type, FALSE, &sample_function);
1916 mask = sample_function.coord_mask;
1917
1918 shader_glsl_get_write_mask(arg->dst, dst_swizzle);
1919
1920 shader_glsl_get_write_mask(mask, coord_mask);
1921
1922 /* with projective textures, texbem only divides the static texture coord, not the displacement,
1923 * so we can't let the GL handle this.
1924 */
1925 if (flags & WINED3DTTFF_PROJECTED) {
1926 DWORD div_mask=0;
1927 char coord_div_mask[3];
1928 switch (flags & ~WINED3DTTFF_PROJECTED) {
1929 case WINED3DTTFF_COUNT1: FIXME("WINED3DTTFF_PROJECTED with WINED3DTTFF_COUNT1?\n"); break;
1930 case WINED3DTTFF_COUNT2: div_mask = WINED3DSP_WRITEMASK_1; break;
1931 case WINED3DTTFF_COUNT3: div_mask = WINED3DSP_WRITEMASK_2; break;
1932 case WINED3DTTFF_COUNT4:
1933 case WINED3DTTFF_DISABLE: div_mask = WINED3DSP_WRITEMASK_3; break;
1934 }
1935 shader_glsl_get_write_mask(div_mask, coord_div_mask);
1936 shader_addline(arg->buffer, "T%u%s /= T%u%s;\n", sampler_idx, coord_mask, sampler_idx, coord_div_mask);
1937 }
1938
1939 shader_glsl_append_dst(arg->buffer, arg);
1940 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0|WINED3DSP_WRITEMASK_1, &coord_param);
1941 shader_addline(arg->buffer, "%s(Psampler%u, T%u%s + vec4(bumpenvmat * %s, 0.0, 0.0)%s )%s);\n",
1942 sample_function.name, sampler_idx, sampler_idx, coord_mask, coord_param.param_str, coord_mask, dst_swizzle);
1943 }
1944
1945 void pshader_glsl_bem(SHADER_OPCODE_ARG* arg) {
1946 glsl_src_param_t src0_param, src1_param;
1947
1948 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0|WINED3DSP_WRITEMASK_1, &src0_param);
1949 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_0|WINED3DSP_WRITEMASK_1, &src1_param);
1950
1951 shader_glsl_append_dst(arg->buffer, arg);
1952 shader_addline(arg->buffer, "%s + bumpenvmat * %s);\n",
1953 src0_param.param_str, src1_param.param_str);
1954 }
1955
1956 /** Process the WINED3DSIO_TEXREG2AR instruction in GLSL
1957 * Sample 2D texture at dst using the alpha & red (wx) components of src as texture coordinates */
1958 void pshader_glsl_texreg2ar(SHADER_OPCODE_ARG* arg) {
1959 glsl_src_param_t src0_param;
1960 DWORD sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1961 char dst_mask[6];
1962
1963 shader_glsl_append_dst(arg->buffer, arg);
1964 shader_glsl_get_write_mask(arg->dst, dst_mask);
1965 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_ALL, &src0_param);
1966
1967 shader_addline(arg->buffer, "texture2D(Psampler%u, %s.wx)%s);\n", sampler_idx, src0_param.reg_name, dst_mask);
1968 }
1969
1970 /** Process the WINED3DSIO_TEXREG2GB instruction in GLSL
1971 * Sample 2D texture at dst using the green & blue (yz) components of src as texture coordinates */
1972 void pshader_glsl_texreg2gb(SHADER_OPCODE_ARG* arg) {
1973 glsl_src_param_t src0_param;
1974 DWORD sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1975 char dst_mask[6];
1976
1977 shader_glsl_append_dst(arg->buffer, arg);
1978 shader_glsl_get_write_mask(arg->dst, dst_mask);
1979 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_ALL, &src0_param);
1980
1981 shader_addline(arg->buffer, "texture2D(Psampler%u, %s.yz)%s);\n", sampler_idx, src0_param.reg_name, dst_mask);
1982 }
1983
1984 /** Process the WINED3DSIO_TEXREG2RGB instruction in GLSL
1985 * Sample texture at dst using the rgb (xyz) components of src as texture coordinates */
1986 void pshader_glsl_texreg2rgb(SHADER_OPCODE_ARG* arg) {
1987 glsl_src_param_t src0_param;
1988 char dst_mask[6];
1989 DWORD sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1990 DWORD sampler_type = arg->reg_maps->samplers[sampler_idx] & WINED3DSP_TEXTURETYPE_MASK;
1991 glsl_sample_function_t sample_function;
1992
1993 shader_glsl_append_dst(arg->buffer, arg);
1994 shader_glsl_get_write_mask(arg->dst, dst_mask);
1995 shader_glsl_get_sample_function(sampler_type, FALSE, &sample_function);
1996 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], sample_function.coord_mask, &src0_param);
1997
1998 shader_addline(arg->buffer, "%s(Psampler%u, %s)%s);\n", sample_function.name, sampler_idx, src0_param.param_str, dst_mask);
1999 }
2000
2001 /** Process the WINED3DSIO_TEXKILL instruction in GLSL.
2002 * If any of the first 3 components are < 0, discard this pixel */
2003 void pshader_glsl_texkill(SHADER_OPCODE_ARG* arg) {
2004 glsl_dst_param_t dst_param;
2005
2006 shader_glsl_add_dst_param(arg, arg->dst, 0, &dst_param);
2007 shader_addline(arg->buffer, "if (any(lessThan(%s.xyz, vec3(0.0)))) discard;\n", dst_param.reg_name);
2008 }
2009
2010 /** Process the WINED3DSIO_DP2ADD instruction in GLSL.
2011 * dst = dot2(src0, src1) + src2 */
2012 void pshader_glsl_dp2add(SHADER_OPCODE_ARG* arg) {
2013 glsl_src_param_t src0_param;
2014 glsl_src_param_t src1_param;
2015 glsl_src_param_t src2_param;
2016 DWORD write_mask;
2017 size_t mask_size;
2018
2019 write_mask = shader_glsl_append_dst(arg->buffer, arg);
2020 mask_size = shader_glsl_get_write_mask_size(write_mask);
2021
2022 shader_glsl_add_src_param(arg, arg->src[0], arg->src_addr[0], WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1, &src0_param);
2023 shader_glsl_add_src_param(arg, arg->src[1], arg->src_addr[1], WINED3DSP_WRITEMASK_0 | WINED3DSP_WRITEMASK_1, &src1_param);
2024 shader_glsl_add_src_param(arg, arg->src[2], arg->src_addr[2], WINED3DSP_WRITEMASK_0, &src2_param);
2025
2026 shader_addline(arg->buffer, "dot(%s, %s) + %s);\n", src0_param.param_str, src1_param.param_str, src2_param.param_str);
2027 }
2028
2029 void pshader_glsl_input_pack(
2030 SHADER_BUFFER* buffer,
2031 semantic* semantics_in) {
2032
2033 unsigned int i;
2034
2035 for (i = 0; i < MAX_REG_INPUT; i++) {
2036
2037 DWORD usage_token = semantics_in[i].usage;
2038 DWORD register_token = semantics_in[i].reg;
2039 DWORD usage, usage_idx;
2040 char reg_mask[6];
2041
2042 /* Uninitialized */
2043 if (!usage_token) continue;
2044 usage = (usage_token & WINED3DSP_DCL_USAGE_MASK) >> WINED3DSP_DCL_USAGE_SHIFT;
2045 usage_idx = (usage_token & WINED3DSP_DCL_USAGEINDEX_MASK) >> WINED3DSP_DCL_USAGEINDEX_SHIFT;
2046 shader_glsl_get_write_mask(register_token, reg_mask);
2047
2048 switch(usage) {
2049
2050 case WINED3DDECLUSAGE_COLOR:
2051 if (usage_idx == 0)
2052 shader_addline(buffer, "IN%u%s = vec4(gl_Color)%s;\n",
2053 i, reg_mask, reg_mask);
2054 else if (usage_idx == 1)
2055 shader_addline(buffer, "IN%u%s = vec4(gl_SecondaryColor)%s;\n",
2056 i, reg_mask, reg_mask);
2057 else
2058 shader_addline(buffer, "IN%u%s = vec4(unsupported_color_input)%s;\n",
2059 i, reg_mask, reg_mask);
2060 break;
2061
2062 case WINED3DDECLUSAGE_TEXCOORD:
2063 shader_addline(buffer, "IN%u%s = vec4(gl_TexCoord[%u])%s;\n",
2064 i, reg_mask, usage_idx, reg_mask );
2065 break;
2066
2067 case WINED3DDECLUSAGE_FOG:
2068 shader_addline(buffer, "IN%u%s = vec4(gl_FogFragCoord)%s;\n",
2069 i, reg_mask, reg_mask);
2070 break;
2071
2072 default:
2073 shader_addline(buffer, "IN%u%s = vec4(unsupported_input)%s;\n",
2074 i, reg_mask, reg_mask);
2075 }
2076 }
2077 }
2078
2079 /*********************************************
2080 * Vertex Shader Specific Code begins here
2081 ********************************************/
2082
2083 void vshader_glsl_output_unpack(
2084 SHADER_BUFFER* buffer,
2085 semantic* semantics_out) {
2086
2087 unsigned int i;
2088
2089 for (i = 0; i < MAX_REG_OUTPUT; i++) {
2090
2091 DWORD usage_token = semantics_out[i].usage;
2092 DWORD register_token = semantics_out[i].reg;
2093 DWORD usage, usage_idx;
2094 char reg_mask[6];
2095
2096 /* Uninitialized */
2097 if (!usage_token) continue;
2098
2099 usage = (usage_token & WINED3DSP_DCL_USAGE_MASK) >> WINED3DSP_DCL_USAGE_SHIFT;
2100 usage_idx = (usage_token & WINED3DSP_DCL_USAGEINDEX_MASK) >> WINED3DSP_DCL_USAGEINDEX_SHIFT;
2101 shader_glsl_get_write_mask(register_token, reg_mask);
2102
2103 switch(usage) {
2104
2105 case WINED3DDECLUSAGE_COLOR:
2106 if (usage_idx == 0)
2107 shader_addline(buffer, "gl_FrontColor%s = OUT%u%s;\n", reg_mask, i, reg_mask);
2108 else if (usage_idx == 1)
2109 shader_addline(buffer, "gl_FrontSecondaryColor%s = OUT%u%s;\n", reg_mask, i, reg_mask);
2110 else
2111 shader_addline(buffer, "unsupported_color_output%s = OUT%u%s;\n", reg_mask, i, reg_mask);
2112 break;
2113
2114 case WINED3DDECLUSAGE_POSITION:
2115 shader_addline(buffer, "gl_Position%s = OUT%u%s;\n", reg_mask, i, reg_mask);
2116 break;
2117
2118 case WINED3DDECLUSAGE_TEXCOORD:
2119 shader_addline(buffer, "gl_TexCoord[%u]%s = OUT%u%s;\n",
2120 usage_idx, reg_mask, i, reg_mask);
2121 break;
2122
2123 case WINED3DDECLUSAGE_PSIZE:
2124 shader_addline(buffer, "gl_PointSize = OUT%u.x;\n", i);
2125 break;
2126
2127 case WINED3DDECLUSAGE_FOG:
2128 shader_addline(buffer, "gl_FogFragCoord = OUT%u%s;\n", i, reg_mask);
2129 break;
2130
2131 default:
2132 shader_addline(buffer, "unsupported_output%s = OUT%u%s;\n", reg_mask, i, reg_mask);
2133 }
2134 }
2135 }
2136
2137 static void add_glsl_program_entry(IWineD3DDeviceImpl *device, struct glsl_shader_prog_link *entry) {
2138 glsl_program_key_t *key;
2139
2140 key = HeapAlloc(GetProcessHeap(), 0, sizeof(glsl_program_key_t));
2141 key->vshader = entry->vshader;
2142 key->pshader = entry->pshader;
2143
2144 hash_table_put(device->glsl_program_lookup, key, entry);
2145 }
2146
2147 static struct glsl_shader_prog_link *get_glsl_program_entry(IWineD3DDeviceImpl *device,
2148 GLhandleARB vshader, GLhandleARB pshader) {
2149 glsl_program_key_t key;
2150
2151 key.vshader = vshader;
2152 key.pshader = pshader;
2153
2154 return (struct glsl_shader_prog_link *)hash_table_get(device->glsl_program_lookup, &key);
2155 }
2156
2157 void delete_glsl_program_entry(IWineD3DDevice *iface, struct glsl_shader_prog_link *entry) {
2158 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
2159 WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
2160 glsl_program_key_t *key;
2161
2162 key = HeapAlloc(GetProcessHeap(), 0, sizeof(glsl_program_key_t));
2163 key->vshader = entry->vshader;
2164 key->pshader = entry->pshader;
2165 hash_table_remove(This->glsl_program_lookup, key);
2166
2167 GL_EXTCALL(glDeleteObjectARB(entry->programId));
2168 if (entry->vshader) list_remove(&entry->vshader_entry);
2169 if (entry->pshader) list_remove(&entry->pshader_entry);
2170 HeapFree(GetProcessHeap(), 0, entry->vuniformF_locations);
2171 HeapFree(GetProcessHeap(), 0, entry->puniformF_locations);
2172 HeapFree(GetProcessHeap(), 0, entry);
2173 }
2174
2175 /** Sets the GLSL program ID for the given pixel and vertex shader combination.
2176 * It sets the programId on the current StateBlock (because it should be called
2177 * inside of the DrawPrimitive() part of the render loop).
2178 *
2179 * If a program for the given combination does not exist, create one, and store
2180 * the program in the hash table. If it creates a program, it will link the
2181 * given objects, too.
2182 */
2183 static void set_glsl_shader_program(IWineD3DDevice *iface, BOOL use_ps, BOOL use_vs) {
2184 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
2185 WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
2186 IWineD3DPixelShader *pshader = This->stateBlock->pixelShader;
2187 IWineD3DVertexShader *vshader = This->stateBlock->vertexShader;
2188 struct glsl_shader_prog_link *entry = NULL;
2189 GLhandleARB programId = 0;
2190 int i;
2191 char glsl_name[8];
2192
2193 GLhandleARB vshader_id = use_vs ? ((IWineD3DBaseShaderImpl*)vshader)->baseShader.prgId : 0;
2194 GLhandleARB pshader_id = use_ps ? ((IWineD3DBaseShaderImpl*)pshader)->baseShader.prgId : 0;
2195 entry = get_glsl_program_entry(This, vshader_id, pshader_id);
2196 if (entry) {
2197 This->stateBlock->glsl_program = entry;
2198 return;
2199 }
2200
2201 /* If we get to this point, then no matching program exists, so we create one */
2202 programId = GL_EXTCALL(glCreateProgramObjectARB());
2203 TRACE("Created new GLSL shader program %u\n", programId);
2204
2205 /* Create the entry */
2206 entry = HeapAlloc(GetProcessHeap(), 0, sizeof(struct glsl_shader_prog_link));
2207 entry->programId = programId;
2208 entry->vshader = vshader_id;
2209 entry->pshader = pshader_id;
2210 /* Add the hash table entry */
2211 add_glsl_program_entry(This, entry);
2212
2213 /* Set the current program */
2214 This->stateBlock->glsl_program = entry;
2215
2216 /* Attach GLSL vshader */
2217 if (vshader_id) {
2218 int max_attribs = 16; /* TODO: Will this always be the case? It is at the moment... */
2219 char tmp_name[10];
2220
2221 TRACE("Attaching GLSL shader object %u to program %u\n", vshader_id, programId);
2222 GL_EXTCALL(glAttachObjectARB(programId, vshader_id));
2223 checkGLcall("glAttachObjectARB");
2224
2225 /* Bind vertex attributes to a corresponding index number to match
2226 * the same index numbers as ARB_vertex_programs (makes loading
2227 * vertex attributes simpler). With this method, we can use the
2228 * exact same code to load the attributes later for both ARB and
2229 * GLSL shaders.
2230 *
2231 * We have to do this here because we need to know the Program ID
2232 * in order to make the bindings work, and it has to be done prior
2233 * to linking the GLSL program. */
2234 for (i = 0; i < max_attribs; ++i) {
2235 snprintf(tmp_name, sizeof(tmp_name), "attrib%i", i);
2236 GL_EXTCALL(glBindAttribLocationARB(programId, i, tmp_name));
2237 }
2238 checkGLcall("glBindAttribLocationARB");
2239
2240 list_add_head(&((IWineD3DBaseShaderImpl *)vshader)->baseShader.linked_programs, &entry->vshader_entry);
2241 }
2242
2243 /* Attach GLSL pshader */
2244 if (pshader_id) {
2245 TRACE("Attaching GLSL shader object %u to program %u\n", pshader_id, programId);
2246 GL_EXTCALL(glAttachObjectARB(programId, pshader_id));
2247 checkGLcall("glAttachObjectARB");
2248
2249 list_add_head(&((IWineD3DBaseShaderImpl *)pshader)->baseShader.linked_programs, &entry->pshader_entry);
2250 }
2251
2252 /* Link the program */
2253 TRACE("Linking GLSL shader program %u\n", programId);
2254 GL_EXTCALL(glLinkProgramARB(programId));
2255 print_glsl_info_log(&GLINFO_LOCATION, programId);
2256
2257 entry->vuniformF_locations = HeapAlloc(GetProcessHeap(), 0, sizeof(GLhandleARB) * GL_LIMITS(vshader_constantsF));
2258 for (i = 0; i < GL_LIMITS(vshader_constantsF); ++i) {
2259 snprintf(glsl_name, sizeof(glsl_name), "VC[%i]", i);
2260 entry->vuniformF_locations[i] = GL_EXTCALL(glGetUniformLocationARB(programId, glsl_name));
2261 }
2262 entry->puniformF_locations = HeapAlloc(GetProcessHeap(), 0, sizeof(GLhandleARB) * GL_LIMITS(pshader_constantsF));
2263 for (i = 0; i < GL_LIMITS(pshader_constantsF); ++i) {
2264 snprintf(glsl_name, sizeof(glsl_name), "PC[%i]", i);
2265 entry->puniformF_locations[i] = GL_EXTCALL(glGetUniformLocationARB(programId, glsl_name));
2266 }
2267 }
2268
2269 static GLhandleARB create_glsl_blt_shader(WineD3D_GL_Info *gl_info) {
2270 GLhandleARB program_id;
2271 GLhandleARB vshader_id, pshader_id;
2272 const char *blt_vshader[] = {
2273 "void main(void)\n"
2274 "{\n"
2275 " gl_Position = gl_Vertex;\n"
2276 " gl_FrontColor = vec4(1.0);\n"
2277 " gl_TexCoord[0].x = (gl_Vertex.x * 0.5) + 0.5;\n"
2278 " gl_TexCoord[0].y = (-gl_Vertex.y * 0.5) + 0.5;\n"
2279 "}\n"
2280 };
2281
2282 const char *blt_pshader[] = {
2283 "uniform sampler2D sampler;\n"
2284 "void main(void)\n"
2285 "{\n"
2286 " gl_FragDepth = texture2D(sampler, gl_TexCoord[0].xy).x;\n"
2287 "}\n"
2288 };
2289
2290 vshader_id = GL_EXTCALL(glCreateShaderObjectARB(GL_VERTEX_SHADER_ARB));
2291 GL_EXTCALL(glShaderSourceARB(vshader_id, 1, blt_vshader, NULL));
2292 GL_EXTCALL(glCompileShaderARB(vshader_id));
2293
2294 pshader_id = GL_EXTCALL(glCreateShaderObjectARB(GL_FRAGMENT_SHADER_ARB));
2295 GL_EXTCALL(glShaderSourceARB(pshader_id, 1, blt_pshader, NULL));
2296 GL_EXTCALL(glCompileShaderARB(pshader_id));
2297
2298 program_id = GL_EXTCALL(glCreateProgramObjectARB());
2299 GL_EXTCALL(glAttachObjectARB(program_id, vshader_id));
2300 GL_EXTCALL(glAttachObjectARB(program_id, pshader_id));
2301 GL_EXTCALL(glLinkProgramARB(program_id));
2302
2303 print_glsl_info_log(&GLINFO_LOCATION, program_id);
2304
2305 return program_id;
2306 }
2307
2308 static void shader_glsl_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {
2309 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
2310 WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
2311 GLhandleARB program_id = 0;
2312
2313 if (useVS || usePS) set_glsl_shader_program(iface, usePS, useVS);
2314 else This->stateBlock->glsl_program = NULL;
2315
2316 program_id = This->stateBlock->glsl_program ? This->stateBlock->glsl_program->programId : 0;
2317 if (program_id) TRACE("Using GLSL program %u\n", program_id);
2318 GL_EXTCALL(glUseProgramObjectARB(program_id));
2319 checkGLcall("glUseProgramObjectARB");
2320 }
2321
2322 static void shader_glsl_select_depth_blt(IWineD3DDevice *iface) {
2323 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
2324 WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
2325 static GLhandleARB program_id = 0;
2326 static GLhandleARB loc = -1;
2327
2328 if (!program_id) {
2329 program_id = create_glsl_blt_shader(gl_info);
2330 loc = GL_EXTCALL(glGetUniformLocationARB(program_id, "sampler"));
2331 }
2332
2333 GL_EXTCALL(glUseProgramObjectARB(program_id));
2334 GL_EXTCALL(glUniform1iARB(loc, 0));
2335 }
2336
2337 static void shader_glsl_cleanup(IWineD3DDevice *iface) {
2338 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
2339 WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
2340 GL_EXTCALL(glUseProgramObjectARB(0));
2341 }
2342
2343 const shader_backend_t glsl_shader_backend = {
2344 &shader_glsl_select,
2345 &shader_glsl_select_depth_blt,
2346 &shader_glsl_load_constants,
2347 &shader_glsl_cleanup
2348 };
useful hacks and other patches not (yet) suitable for mainline