?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
push 553365173a7b5952fd5f4a32453e8e96b9d8a845
[wine/hacks.git] / dlls / wined3d / arb_program_shader.c
blob5d3bd76df68463fd7bfa4b7519851fc724486475
1 /*
2 * Pixel and vertex shaders implementation using ARB_vertex_program
3 * and ARB_fragment_program GL extensions.
4 *
5 * Copyright 2002-2003 Jason Edmeades
6 * Copyright 2002-2003 Raphael Junqueira
7 * Copyright 2004 Christian Costa
8 * Copyright 2005 Oliver Stieber
9 * Copyright 2006 Ivan Gyurdiev
10 * Copyright 2006 Jason Green
11 * Copyright 2006 Henri Verbeet
12 * Copyright 2007-2008 Stefan Dösinger for CodeWeavers
13 *
14 * This library is free software; you can redistribute it and/or
15 * modify it under the terms of the GNU Lesser General Public
16 * License as published by the Free Software Foundation; either
17 * version 2.1 of the License, or (at your option) any later version.
18 *
19 * This library is distributed in the hope that it will be useful,
20 * but WITHOUT ANY WARRANTY; without even the implied warranty of
21 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
22 * Lesser General Public License for more details.
23 *
24 * You should have received a copy of the GNU Lesser General Public
25 * License along with this library; if not, write to the Free Software
26 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
27 */
28
29 #include "config.h"
30
31 #include <math.h>
32 #include <stdio.h>
33
34 #include "wined3d_private.h"
35
36 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
37 WINE_DECLARE_DEBUG_CHANNEL(d3d_constants);
38 WINE_DECLARE_DEBUG_CHANNEL(d3d_caps);
39 WINE_DECLARE_DEBUG_CHANNEL(d3d);
40
41 #define GLINFO_LOCATION (*gl_info)
42
43 /* Internally used shader constants. Applications can use constants 0 to GL_LIMITS(vshader_constantsF) - 1,
44 * so upload them above that
45 */
46 #define ARB_SHADER_PRIVCONST_BASE GL_LIMITS(vshader_constantsF)
47 #define ARB_SHADER_PRIVCONST_POS ARB_SHADER_PRIVCONST_BASE + 0
48
49 /* ARB_program_shader private data */
50 struct shader_arb_priv {
51 GLuint current_vprogram_id;
52 GLuint current_fprogram_id;
53 GLuint depth_blt_vprogram_id;
54 GLuint depth_blt_fprogram_id[tex_type_count];
55 BOOL use_arbfp_fixed_func;
56 struct hash_table_t *fragment_shaders;
57 };
58
59 /********************************************************
60 * ARB_[vertex/fragment]_program helper functions follow
61 ********************************************************/
62
63 /**
64 * Loads floating point constants into the currently set ARB_vertex/fragment_program.
65 * When constant_list == NULL, it will load all the constants.
66 *
67 * @target_type should be either GL_VERTEX_PROGRAM_ARB (for vertex shaders)
68 * or GL_FRAGMENT_PROGRAM_ARB (for pixel shaders)
69 */
70 static unsigned int shader_arb_load_constantsF(IWineD3DBaseShaderImpl* This, const WineD3D_GL_Info *gl_info,
71 GLuint target_type, unsigned int max_constants, const float *constants, char *dirty_consts)
72 {
73 DWORD shader_version = This->baseShader.reg_maps.shader_version;
74 local_constant* lconst;
75 DWORD i, j;
76 unsigned int ret;
77
78 if (TRACE_ON(d3d_shader)) {
79 for(i = 0; i < max_constants; i++) {
80 if(!dirty_consts[i]) continue;
81 TRACE_(d3d_constants)("Loading constants %i: %f, %f, %f, %f\n", i,
82 constants[i * 4 + 0], constants[i * 4 + 1],
83 constants[i * 4 + 2], constants[i * 4 + 3]);
84 }
85 }
86 /* In 1.X pixel shaders constants are implicitly clamped in the range [-1;1] */
87 if (target_type == GL_FRAGMENT_PROGRAM_ARB && WINED3DSHADER_VERSION_MAJOR(shader_version) == 1)
88 {
89 float lcl_const[4];
90 for(i = 0; i < max_constants; i++) {
91 if(!dirty_consts[i]) continue;
92 dirty_consts[i] = 0;
93
94 j = 4 * i;
95 if(constants[j + 0] > 1.0) lcl_const[0] = 1.0;
96 else if(constants[j + 0] < -1.0) lcl_const[0] = -1.0;
97 else lcl_const[0] = constants[j + 0];
98
99 if(constants[j + 1] > 1.0) lcl_const[1] = 1.0;
100 else if(constants[j + 1] < -1.0) lcl_const[1] = -1.0;
101 else lcl_const[1] = constants[j + 1];
102
103 if(constants[j + 2] > 1.0) lcl_const[2] = 1.0;
104 else if(constants[j + 2] < -1.0) lcl_const[2] = -1.0;
105 else lcl_const[2] = constants[j + 2];
106
107 if(constants[j + 3] > 1.0) lcl_const[3] = 1.0;
108 else if(constants[j + 3] < -1.0) lcl_const[3] = -1.0;
109 else lcl_const[3] = constants[j + 3];
110
111 GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, i, lcl_const));
112 }
113 } else {
114 if(GL_SUPPORT(EXT_GPU_PROGRAM_PARAMETERS)) {
115 /* TODO: Benchmark if we're better of with finding the dirty constants ourselves,
116 * or just reloading *all* constants at once
117 *
118 GL_EXTCALL(glProgramEnvParameters4fvEXT(target_type, 0, max_constants, constants));
119 */
120 for(i = 0; i < max_constants; i++) {
121 if(!dirty_consts[i]) continue;
122
123 /* Find the next block of dirty constants */
124 dirty_consts[i] = 0;
125 j = i;
126 for(i++; (i < max_constants) && dirty_consts[i]; i++) {
127 dirty_consts[i] = 0;
128 }
129
130 GL_EXTCALL(glProgramEnvParameters4fvEXT(target_type, j, i - j, constants + (j * 4)));
131 }
132 } else {
133 for(i = 0; i < max_constants; i++) {
134 if(dirty_consts[i]) {
135 dirty_consts[i] = 0;
136 GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, i, constants + (i * 4)));
137 }
138 }
139 }
140 }
141 checkGLcall("glProgramEnvParameter4fvARB()");
142
143 /* Load immediate constants */
144 if(This->baseShader.load_local_constsF) {
145 if (TRACE_ON(d3d_shader)) {
146 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
147 GLfloat* values = (GLfloat*)lconst->value;
148 TRACE_(d3d_constants)("Loading local constants %i: %f, %f, %f, %f\n", lconst->idx,
149 values[0], values[1], values[2], values[3]);
150 }
151 }
152 /* Immediate constants are clamped for 1.X shaders at loading times */
153 ret = 0;
154 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
155 dirty_consts[lconst->idx] = 1; /* Dirtify so the non-immediate constant overwrites it next time */
156 ret = max(ret, lconst->idx + 1);
157 GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, lconst->idx, (GLfloat*)lconst->value));
158 }
159 checkGLcall("glProgramEnvParameter4fvARB()");
160 return ret; /* The loaded immediate constants need reloading for the next shader */
161 } else {
162 return 0; /* No constants are dirty now */
163 }
164 }
165
166 /**
167 * Loads the app-supplied constants into the currently set ARB_[vertex/fragment]_programs.
168 *
169 * We only support float constants in ARB at the moment, so don't
170 * worry about the Integers or Booleans
171 */
172 static void shader_arb_load_constants(
173 IWineD3DDevice* device,
174 char usePixelShader,
175 char useVertexShader) {
176
177 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) device;
178 IWineD3DStateBlockImpl* stateBlock = deviceImpl->stateBlock;
179 const WineD3D_GL_Info *gl_info = &deviceImpl->adapter->gl_info;
180 unsigned char i;
181
182 if (useVertexShader) {
183 IWineD3DBaseShaderImpl* vshader = (IWineD3DBaseShaderImpl*) stateBlock->vertexShader;
184
185 /* Load DirectX 9 float constants for vertex shader */
186 deviceImpl->highest_dirty_vs_const = shader_arb_load_constantsF(
187 vshader, gl_info, GL_VERTEX_PROGRAM_ARB,
188 deviceImpl->highest_dirty_vs_const,
189 stateBlock->vertexShaderConstantF,
190 deviceImpl->activeContext->vshader_const_dirty);
191
192 /* Upload the position fixup */
193 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_VERTEX_PROGRAM_ARB, ARB_SHADER_PRIVCONST_POS, deviceImpl->posFixup));
194 }
195
196 if (usePixelShader) {
197
198 IWineD3DBaseShaderImpl* pshader = (IWineD3DBaseShaderImpl*) stateBlock->pixelShader;
199 IWineD3DPixelShaderImpl *psi = (IWineD3DPixelShaderImpl *) pshader;
200
201 /* Load DirectX 9 float constants for pixel shader */
202 deviceImpl->highest_dirty_ps_const = shader_arb_load_constantsF(
203 pshader, gl_info, GL_FRAGMENT_PROGRAM_ARB,
204 deviceImpl->highest_dirty_ps_const,
205 stateBlock->pixelShaderConstantF,
206 deviceImpl->activeContext->pshader_const_dirty);
207
208 for(i = 0; i < psi->numbumpenvmatconsts; i++) {
209 /* The state manager takes care that this function is always called if the bump env matrix changes
210 */
211 const float *data = (const float *)&stateBlock->textureState[(int) psi->bumpenvmatconst[i].texunit][WINED3DTSS_BUMPENVMAT00];
212 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, psi->bumpenvmatconst[i].const_num, data));
213 deviceImpl->activeContext->pshader_const_dirty[psi->bumpenvmatconst[i].const_num] = 1;
214
215 if (psi->luminanceconst[i].const_num != WINED3D_CONST_NUM_UNUSED)
216 {
217 /* WINED3DTSS_BUMPENVLSCALE and WINED3DTSS_BUMPENVLOFFSET are next to each other.
218 * point gl to the scale, and load 4 floats. x = scale, y = offset, z and w are junk, we
219 * don't care about them. The pointers are valid for sure because the stateblock is bigger.
220 * (they're WINED3DTSS_TEXTURETRANSFORMFLAGS and WINED3DTSS_ADDRESSW, so most likely 0 or NaN
221 */
222 const float *scale = (const float *)&stateBlock->textureState[(int) psi->luminanceconst[i].texunit][WINED3DTSS_BUMPENVLSCALE];
223 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, psi->luminanceconst[i].const_num, scale));
224 deviceImpl->activeContext->pshader_const_dirty[psi->luminanceconst[i].const_num] = 1;
225 }
226 }
227 }
228 }
229
230 static void shader_arb_update_float_vertex_constants(IWineD3DDevice *iface, UINT start, UINT count)
231 {
232 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
233
234 /* We don't want shader constant dirtification to be an O(contexts), so just dirtify the active
235 * context. On a context switch the old context will be fully dirtified */
236 memset(This->activeContext->vshader_const_dirty + start, 1,
237 sizeof(*This->activeContext->vshader_const_dirty) * count);
238 This->highest_dirty_vs_const = max(This->highest_dirty_vs_const, start + count + 1);
239 }
240
241 static void shader_arb_update_float_pixel_constants(IWineD3DDevice *iface, UINT start, UINT count)
242 {
243 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
244
245 /* We don't want shader constant dirtification to be an O(contexts), so just dirtify the active
246 * context. On a context switch the old context will be fully dirtified */
247 memset(This->activeContext->pshader_const_dirty + start, 1,
248 sizeof(*This->activeContext->pshader_const_dirty) * count);
249 This->highest_dirty_ps_const = max(This->highest_dirty_ps_const, start + count + 1);
250 }
251
252 /* Generate the variable & register declarations for the ARB_vertex_program output target */
253 static void shader_generate_arb_declarations(IWineD3DBaseShader *iface, const shader_reg_maps *reg_maps,
254 SHADER_BUFFER *buffer, const WineD3D_GL_Info *gl_info)
255 {
256 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
257 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device;
258 DWORD i, cur;
259 char pshader = shader_is_pshader_version(reg_maps->shader_version);
260 unsigned max_constantsF = min(This->baseShader.limits.constant_float,
261 (pshader ? GL_LIMITS(pshader_constantsF) : GL_LIMITS(vshader_constantsF)));
262 UINT extra_constants_needed = 0;
263 const local_constant *lconst;
264
265 /* Temporary Output register */
266 shader_addline(buffer, "TEMP TMP_OUT;\n");
267
268 for(i = 0; i < This->baseShader.limits.temporary; i++) {
269 if (reg_maps->temporary[i])
270 shader_addline(buffer, "TEMP R%u;\n", i);
271 }
272
273 for (i = 0; i < This->baseShader.limits.address; i++) {
274 if (reg_maps->address[i])
275 shader_addline(buffer, "ADDRESS A%d;\n", i);
276 }
277
278 for(i = 0; i < This->baseShader.limits.texcoord; i++) {
279 if (reg_maps->texcoord[i])
280 shader_addline(buffer,"TEMP T%u;\n", i);
281 }
282
283 /* Texture coordinate registers must be pre-loaded */
284 for (i = 0; i < This->baseShader.limits.texcoord; i++) {
285 if (reg_maps->texcoord[i])
286 shader_addline(buffer, "MOV T%u, fragment.texcoord[%u];\n", i, i);
287 }
288
289 for(i = 0; i < (sizeof(reg_maps->bumpmat) / sizeof(reg_maps->bumpmat[0])); i++) {
290 IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *) This;
291 if(!reg_maps->bumpmat[i]) continue;
292
293 cur = ps->numbumpenvmatconsts;
294 ps->bumpenvmatconst[cur].const_num = -1;
295 ps->bumpenvmatconst[cur].texunit = i;
296 ps->luminanceconst[cur].const_num = -1;
297 ps->luminanceconst[cur].texunit = i;
298
299 /* If the shader does not use all available constants, use the next free constant to load the bump mapping environment matrix from
300 * the stateblock into the shader. If no constant is available don't load, texbem will then just sample the texture without applying
301 * bump mapping.
302 */
303 if(max_constantsF + extra_constants_needed < GL_LIMITS(pshader_constantsF)) {
304 ps->bumpenvmatconst[cur].const_num = max_constantsF + extra_constants_needed;
305 shader_addline(buffer, "PARAM bumpenvmat%d = program.env[%d];\n",
306 i, ps->bumpenvmatconst[cur].const_num);
307 extra_constants_needed++;
308
309 if(reg_maps->luminanceparams && max_constantsF + extra_constants_needed < GL_LIMITS(pshader_constantsF)) {
310 ((IWineD3DPixelShaderImpl *)This)->luminanceconst[cur].const_num = max_constantsF + extra_constants_needed;
311 shader_addline(buffer, "PARAM luminance%d = program.env[%d];\n",
312 i, ps->luminanceconst[cur].const_num);
313 extra_constants_needed++;
314 } else if(reg_maps->luminanceparams) {
315 FIXME("No free constant to load the luminance parameters\n");
316 }
317 } else {
318 FIXME("No free constant found to load environment bump mapping matrix into the shader. texbem instruction will not apply bump mapping\n");
319 }
320
321 ps->numbumpenvmatconsts = cur + 1;
322 }
323
324 if(device->stateBlock->renderState[WINED3DRS_SRGBWRITEENABLE] && pshader) {
325 shader_addline(buffer, "PARAM srgb_mul_low = {%f, %f, %f, 1.0};\n",
326 srgb_mul_low, srgb_mul_low, srgb_mul_low);
327 shader_addline(buffer, "PARAM srgb_comparison = {%f, %f, %f, %f};\n",
328 srgb_cmp, srgb_cmp, srgb_cmp, srgb_cmp);
329 shader_addline(buffer, "PARAM srgb_pow = {%f, %f, %f, 1.0};\n",
330 srgb_pow, srgb_pow, srgb_pow);
331 shader_addline(buffer, "PARAM srgb_mul_hi = {%f, %f, %f, 1.0};\n",
332 srgb_mul_high, srgb_mul_high, srgb_mul_high);
333 shader_addline(buffer, "PARAM srgb_sub_hi = {%f, %f, %f, 0.0};\n",
334 srgb_sub_high, srgb_sub_high, srgb_sub_high);
335 }
336
337 /* Load local constants using the program-local space,
338 * this avoids reloading them each time the shader is used
339 */
340 if(!This->baseShader.load_local_constsF) {
341 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
342 shader_addline(buffer, "PARAM C%u = program.local[%u];\n", lconst->idx,
343 lconst->idx);
344 }
345 }
346
347 /* we use the array-based constants array if the local constants are marked for loading,
348 * because then we use indirect addressing, or when the local constant list is empty,
349 * because then we don't know if we're using indirect addressing or not. If we're hardcoding
350 * local constants do not declare the loaded constants as an array because ARB compilers usually
351 * do not optimize unused constants away
352 */
353 if(This->baseShader.load_local_constsF || list_empty(&This->baseShader.constantsF)) {
354 /* Need to PARAM the environment parameters (constants) so we can use relative addressing */
355 shader_addline(buffer, "PARAM C[%d] = { program.env[0..%d] };\n",
356 max_constantsF, max_constantsF - 1);
357 } else {
358 for(i = 0; i < max_constantsF; i++) {
359 if(!shader_constant_is_local(This, i)) {
360 shader_addline(buffer, "PARAM C%d = program.env[%d];\n",i, i);
361 }
362 }
363 }
364 }
365
366 static const char * const shift_tab[] = {
367 "dummy", /* 0 (none) */
368 "coefmul.x", /* 1 (x2) */
369 "coefmul.y", /* 2 (x4) */
370 "coefmul.z", /* 3 (x8) */
371 "coefmul.w", /* 4 (x16) */
372 "dummy", /* 5 (x32) */
373 "dummy", /* 6 (x64) */
374 "dummy", /* 7 (x128) */
375 "dummy", /* 8 (d256) */
376 "dummy", /* 9 (d128) */
377 "dummy", /* 10 (d64) */
378 "dummy", /* 11 (d32) */
379 "coefdiv.w", /* 12 (d16) */
380 "coefdiv.z", /* 13 (d8) */
381 "coefdiv.y", /* 14 (d4) */
382 "coefdiv.x" /* 15 (d2) */
383 };
384
385 static void shader_arb_get_write_mask(const struct wined3d_shader_instruction *ins,
386 const DWORD param, char *write_mask)
387 {
388 char *ptr = write_mask;
389 char vshader = shader_is_vshader_version(ins->reg_maps->shader_version);
390
391 if(vshader && shader_get_regtype(param) == WINED3DSPR_ADDR) {
392 *ptr++ = '.';
393 *ptr++ = 'x';
394 } else if ((param & WINED3DSP_WRITEMASK_ALL) != WINED3DSP_WRITEMASK_ALL) {
395 *ptr++ = '.';
396 if (param & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
397 if (param & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
398 if (param & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
399 if (param & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
400 }
401
402 *ptr = '\0';
403 }
404
405 static void shader_arb_get_swizzle(const DWORD param, BOOL fixup, char *swizzle_str) {
406 /* For registers of type WINED3DDECLTYPE_D3DCOLOR, data is stored as "bgra",
407 * but addressed as "rgba". To fix this we need to swap the register's x
408 * and z components. */
409 const char *swizzle_chars = fixup ? "zyxw" : "xyzw";
410 char *ptr = swizzle_str;
411
412 /* swizzle bits fields: wwzzyyxx */
413 DWORD swizzle = (param & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
414 DWORD swizzle_x = swizzle & 0x03;
415 DWORD swizzle_y = (swizzle >> 2) & 0x03;
416 DWORD swizzle_z = (swizzle >> 4) & 0x03;
417 DWORD swizzle_w = (swizzle >> 6) & 0x03;
418
419 /* If the swizzle is the default swizzle (ie, "xyzw"), we don't need to
420 * generate a swizzle string. Unless we need to our own swizzling. */
421 if ((WINED3DSP_NOSWIZZLE >> WINED3DSP_SWIZZLE_SHIFT) != swizzle || fixup) {
422 *ptr++ = '.';
423 if (swizzle_x == swizzle_y && swizzle_x == swizzle_z && swizzle_x == swizzle_w) {
424 *ptr++ = swizzle_chars[swizzle_x];
425 } else {
426 *ptr++ = swizzle_chars[swizzle_x];
427 *ptr++ = swizzle_chars[swizzle_y];
428 *ptr++ = swizzle_chars[swizzle_z];
429 *ptr++ = swizzle_chars[swizzle_w];
430 }
431 }
432
433 *ptr = '\0';
434 }
435
436 static void pshader_get_register_name(IWineD3DBaseShader* iface,
437 const DWORD param, char* regstr) {
438
439 DWORD reg = param & WINED3DSP_REGNUM_MASK;
440 DWORD regtype = shader_get_regtype(param);
441 IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *) iface;
442
443 switch (regtype) {
444 case WINED3DSPR_TEMP:
445 sprintf(regstr, "R%u", reg);
446 break;
447 case WINED3DSPR_INPUT:
448 if (reg==0) {
449 strcpy(regstr, "fragment.color.primary");
450 } else {
451 strcpy(regstr, "fragment.color.secondary");
452 }
453 break;
454 case WINED3DSPR_CONST:
455 if(This->baseShader.load_local_constsF || list_empty(&This->baseShader.constantsF)) {
456 sprintf(regstr, "C[%u]", reg);
457 } else {
458 sprintf(regstr, "C%u", reg);
459 }
460 break;
461 case WINED3DSPR_TEXTURE: /* case WINED3DSPR_ADDR: */
462 sprintf(regstr,"T%u", reg);
463 break;
464 case WINED3DSPR_COLOROUT:
465 if (reg == 0)
466 sprintf(regstr, "TMP_COLOR");
467 else {
468 /* TODO: See GL_ARB_draw_buffers */
469 FIXME("Unsupported write to render target %u\n", reg);
470 sprintf(regstr, "unsupported_register");
471 }
472 break;
473 case WINED3DSPR_DEPTHOUT:
474 sprintf(regstr, "result.depth");
475 break;
476 case WINED3DSPR_ATTROUT:
477 sprintf(regstr, "oD[%u]", reg);
478 break;
479 case WINED3DSPR_TEXCRDOUT:
480 sprintf(regstr, "oT[%u]", reg);
481 break;
482 default:
483 FIXME("Unhandled register name Type(%d)\n", regtype);
484 sprintf(regstr, "unrecognized_register");
485 break;
486 }
487 }
488
489 /* TODO: merge with pixel shader */
490 static void vshader_program_add_param(const struct wined3d_shader_instruction *ins,
491 const DWORD param, BOOL is_input, char *hwLine)
492 {
493 IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *)ins->shader;
494
495 /* oPos, oFog and oPts in D3D */
496 static const char * const hwrastout_reg_names[] = { "TMP_OUT", "result.fogcoord", "result.pointsize" };
497
498 DWORD reg = param & WINED3DSP_REGNUM_MASK;
499 DWORD regtype = shader_get_regtype(param);
500 char tmpReg[255];
501 BOOL is_color = FALSE;
502
503 if ((param & WINED3DSP_SRCMOD_MASK) == WINED3DSPSM_NEG) {
504 strcat(hwLine, " -");
505 } else {
506 strcat(hwLine, " ");
507 }
508
509 switch (regtype) {
510 case WINED3DSPR_TEMP:
511 sprintf(tmpReg, "R%u", reg);
512 strcat(hwLine, tmpReg);
513 break;
514 case WINED3DSPR_INPUT:
515
516 if (This->cur_args->swizzle_map & (1 << reg)) is_color = TRUE;
517
518 sprintf(tmpReg, "vertex.attrib[%u]", reg);
519 strcat(hwLine, tmpReg);
520 break;
521 case WINED3DSPR_CONST:
522 if(param & WINED3DSHADER_ADDRMODE_RELATIVE) {
523 if(reg >= This->rel_offset) {
524 sprintf(tmpReg, "C[A0.x + %u]", reg - This->rel_offset);
525 } else {
526 sprintf(tmpReg, "C[A0.x - %u]", -reg + This->rel_offset);
527 }
528 } else {
529 if(This->baseShader.load_local_constsF || list_empty(&This->baseShader.constantsF)) {
530 sprintf(tmpReg, "C[%u]", reg);
531 } else {
532 sprintf(tmpReg, "C%u", reg);
533 }
534 }
535 strcat(hwLine, tmpReg);
536 break;
537 case WINED3DSPR_ADDR: /*case D3DSPR_TEXTURE:*/
538 sprintf(tmpReg, "A%u", reg);
539 strcat(hwLine, tmpReg);
540 break;
541 case WINED3DSPR_RASTOUT:
542 sprintf(tmpReg, "%s", hwrastout_reg_names[reg]);
543 strcat(hwLine, tmpReg);
544 break;
545 case WINED3DSPR_ATTROUT:
546 if (reg==0) {
547 strcat(hwLine, "result.color.primary");
548 } else {
549 strcat(hwLine, "result.color.secondary");
550 }
551 break;
552 case WINED3DSPR_TEXCRDOUT:
553 sprintf(tmpReg, "result.texcoord[%u]", reg);
554 strcat(hwLine, tmpReg);
555 break;
556 default:
557 FIXME("Unknown reg type %d %d\n", regtype, reg);
558 strcat(hwLine, "unrecognized_register");
559 break;
560 }
561
562 if (!is_input) {
563 char write_mask[6];
564 shader_arb_get_write_mask(ins, param, write_mask);
565 strcat(hwLine, write_mask);
566 } else {
567 char swizzle[6];
568 shader_arb_get_swizzle(param, is_color, swizzle);
569 strcat(hwLine, swizzle);
570 }
571 }
572
573 static const char *shader_arb_get_fixup_swizzle(enum fixup_channel_source channel_source)
574 {
575 switch(channel_source)
576 {
577 case CHANNEL_SOURCE_ZERO: return "0";
578 case CHANNEL_SOURCE_ONE: return "1";
579 case CHANNEL_SOURCE_X: return "x";
580 case CHANNEL_SOURCE_Y: return "y";
581 case CHANNEL_SOURCE_Z: return "z";
582 case CHANNEL_SOURCE_W: return "w";
583 default:
584 FIXME("Unhandled channel source %#x\n", channel_source);
585 return "undefined";
586 }
587 }
588
589 static void gen_color_correction(SHADER_BUFFER *buffer, const char *reg, DWORD dst_mask,
590 const char *one, const char *two, struct color_fixup_desc fixup)
591 {
592 DWORD mask;
593
594 if (is_yuv_fixup(fixup))
595 {
596 enum yuv_fixup yuv_fixup = get_yuv_fixup(fixup);
597 FIXME("YUV fixup (%#x) not supported\n", yuv_fixup);
598 return;
599 }
600
601 mask = 0;
602 if (fixup.x_source != CHANNEL_SOURCE_X) mask |= WINED3DSP_WRITEMASK_0;
603 if (fixup.y_source != CHANNEL_SOURCE_Y) mask |= WINED3DSP_WRITEMASK_1;
604 if (fixup.z_source != CHANNEL_SOURCE_Z) mask |= WINED3DSP_WRITEMASK_2;
605 if (fixup.w_source != CHANNEL_SOURCE_W) mask |= WINED3DSP_WRITEMASK_3;
606 mask &= dst_mask;
607
608 if (mask)
609 {
610 shader_addline(buffer, "SWZ %s, %s, %s, %s, %s, %s;\n", reg, reg,
611 shader_arb_get_fixup_swizzle(fixup.x_source), shader_arb_get_fixup_swizzle(fixup.y_source),
612 shader_arb_get_fixup_swizzle(fixup.z_source), shader_arb_get_fixup_swizzle(fixup.w_source));
613 }
614
615 mask = 0;
616 if (fixup.x_sign_fixup) mask |= WINED3DSP_WRITEMASK_0;
617 if (fixup.y_sign_fixup) mask |= WINED3DSP_WRITEMASK_1;
618 if (fixup.z_sign_fixup) mask |= WINED3DSP_WRITEMASK_2;
619 if (fixup.w_sign_fixup) mask |= WINED3DSP_WRITEMASK_3;
620 mask &= dst_mask;
621
622 if (mask)
623 {
624 char reg_mask[6];
625 char *ptr = reg_mask;
626
627 if (mask != WINED3DSP_WRITEMASK_ALL)
628 {
629 *ptr++ = '.';
630 if (mask & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
631 if (mask & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
632 if (mask & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
633 if (mask & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
634 }
635 *ptr = '\0';
636
637 shader_addline(buffer, "MAD %s%s, %s, %s, -%s;\n", reg, reg_mask, reg, two, one);
638 }
639 }
640
641 static void shader_hw_sample(const struct wined3d_shader_instruction *ins, DWORD sampler_idx,
642 const char *dst_str, const char *coord_reg, BOOL projected, BOOL bias)
643 {
644 SHADER_BUFFER *buffer = ins->buffer;
645 DWORD sampler_type = ins->reg_maps->samplers[sampler_idx] & WINED3DSP_TEXTURETYPE_MASK;
646 const char *tex_type;
647 IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)ins->shader;
648 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device;
649
650 switch(sampler_type) {
651 case WINED3DSTT_1D:
652 tex_type = "1D";
653 break;
654
655 case WINED3DSTT_2D:
656 if(device->stateBlock->textures[sampler_idx] &&
657 IWineD3DBaseTexture_GetTextureDimensions(device->stateBlock->textures[sampler_idx]) == GL_TEXTURE_RECTANGLE_ARB) {
658 tex_type = "RECT";
659 } else {
660 tex_type = "2D";
661 }
662 break;
663
664 case WINED3DSTT_VOLUME:
665 tex_type = "3D";
666 break;
667
668 case WINED3DSTT_CUBE:
669 tex_type = "CUBE";
670 break;
671
672 default:
673 ERR("Unexpected texture type %d\n", sampler_type);
674 tex_type = "";
675 }
676
677 if (bias) {
678 /* Shouldn't be possible, but let's check for it */
679 if(projected) FIXME("Biased and Projected texture sampling\n");
680 /* TXB takes the 4th component of the source vector automatically, as d3d. Nothing more to do */
681 shader_addline(buffer, "TXB %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
682 } else if (projected) {
683 shader_addline(buffer, "TXP %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
684 } else {
685 shader_addline(buffer, "TEX %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
686 }
687
688 if (shader_is_pshader_version(ins->reg_maps->shader_version))
689 {
690 IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)ins->shader;
691 gen_color_correction(buffer, dst_str, ins->dst[0].write_mask,
692 "one", "coefmul.x", ps->cur_args->color_fixup[sampler_idx]);
693 }
694 }
695
696 static void pshader_gen_input_modifier_line (
697 IWineD3DBaseShader *iface,
698 SHADER_BUFFER* buffer,
699 const DWORD instr,
700 int tmpreg,
701 char *outregstr) {
702
703 /* Generate a line that does the input modifier computation and return the input register to use */
704 char regstr[256];
705 char swzstr[20];
706 int insert_line;
707
708 /* Assume a new line will be added */
709 insert_line = 1;
710
711 /* Get register name */
712 pshader_get_register_name(iface, instr, regstr);
713 shader_arb_get_swizzle(instr, FALSE, swzstr);
714
715 switch (instr & WINED3DSP_SRCMOD_MASK) {
716 case WINED3DSPSM_NONE:
717 sprintf(outregstr, "%s%s", regstr, swzstr);
718 insert_line = 0;
719 break;
720 case WINED3DSPSM_NEG:
721 sprintf(outregstr, "-%s%s", regstr, swzstr);
722 insert_line = 0;
723 break;
724 case WINED3DSPSM_BIAS:
725 shader_addline(buffer, "ADD T%c, %s, -coefdiv.x;\n", 'A' + tmpreg, regstr);
726 break;
727 case WINED3DSPSM_BIASNEG:
728 shader_addline(buffer, "ADD T%c, -%s, coefdiv.x;\n", 'A' + tmpreg, regstr);
729 break;
730 case WINED3DSPSM_SIGN:
731 shader_addline(buffer, "MAD T%c, %s, coefmul.x, -one.x;\n", 'A' + tmpreg, regstr);
732 break;
733 case WINED3DSPSM_SIGNNEG:
734 shader_addline(buffer, "MAD T%c, %s, -coefmul.x, one.x;\n", 'A' + tmpreg, regstr);
735 break;
736 case WINED3DSPSM_COMP:
737 shader_addline(buffer, "SUB T%c, one.x, %s;\n", 'A' + tmpreg, regstr);
738 break;
739 case WINED3DSPSM_X2:
740 shader_addline(buffer, "ADD T%c, %s, %s;\n", 'A' + tmpreg, regstr, regstr);
741 break;
742 case WINED3DSPSM_X2NEG:
743 shader_addline(buffer, "ADD T%c, -%s, -%s;\n", 'A' + tmpreg, regstr, regstr);
744 break;
745 case WINED3DSPSM_DZ:
746 shader_addline(buffer, "RCP T%c, %s.z;\n", 'A' + tmpreg, regstr);
747 shader_addline(buffer, "MUL T%c, %s, T%c;\n", 'A' + tmpreg, regstr, 'A' + tmpreg);
748 break;
749 case WINED3DSPSM_DW:
750 shader_addline(buffer, "RCP T%c, %s.w;\n", 'A' + tmpreg, regstr);
751 shader_addline(buffer, "MUL T%c, %s, T%c;\n", 'A' + tmpreg, regstr, 'A' + tmpreg);
752 break;
753 default:
754 sprintf(outregstr, "%s%s", regstr, swzstr);
755 insert_line = 0;
756 }
757
758 /* Return modified or original register, with swizzle */
759 if (insert_line)
760 sprintf(outregstr, "T%c%s", 'A' + tmpreg, swzstr);
761 }
762
763 static inline void pshader_gen_output_modifier_line(SHADER_BUFFER *buffer, int saturate, const char *write_mask,
764 int shift, const char *regstr)
765 {
766 /* Generate a line that does the output modifier computation */
767 shader_addline(buffer, "MUL%s %s%s, %s, %s;\n", saturate ? "_SAT" : "",
768 regstr, write_mask, regstr, shift_tab[shift]);
769 }
770
771 static void pshader_hw_bem(const struct wined3d_shader_instruction *ins)
772 {
773 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->shader;
774 SHADER_BUFFER *buffer = ins->buffer;
775 char dst_name[50];
776 char src_name[2][50];
777 char dst_wmask[20];
778 DWORD sampler_code = ins->dst[0].register_idx;
779 BOOL has_bumpmat = FALSE;
780 int i;
781
782 for(i = 0; i < This->numbumpenvmatconsts; i++) {
783 if (This->bumpenvmatconst[i].const_num != WINED3D_CONST_NUM_UNUSED
784 && This->bumpenvmatconst[i].texunit == sampler_code)
785 {
786 has_bumpmat = TRUE;
787 break;
788 }
789 }
790
791 pshader_get_register_name(ins->shader, ins->dst[0].token, dst_name);
792 shader_arb_get_write_mask(ins, ins->dst[0].token, dst_wmask);
793 strcat(dst_name, dst_wmask);
794
795 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src_name[0]);
796 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[1], 1, src_name[1]);
797
798 if(has_bumpmat) {
799 /* Sampling the perturbation map in Tsrc was done already, including the signedness correction if needed */
800 shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, x, z, 0, 0;\n", sampler_code);
801 shader_addline(buffer, "DP3 TMP.r, TMP2, %s;\n", src_name[1]);
802 shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, y, w, 0, 0;\n", sampler_code);
803 shader_addline(buffer, "DP3 TMP.g, TMP2, %s;\n", src_name[1]);
804
805 shader_addline(buffer, "ADD %s, %s, TMP;\n", dst_name, src_name[0]);
806 } else {
807 shader_addline(buffer, "MOV %s, %s;\n", dst_name, src_name[0]);
808 }
809 }
810
811 static void pshader_hw_cnd(const struct wined3d_shader_instruction *ins)
812 {
813 SHADER_BUFFER *buffer = ins->buffer;
814 char dst_wmask[20];
815 char dst_name[50];
816 char src_name[3][50];
817 BOOL sat = ins->dst[0].modifiers & WINED3DSPDM_SATURATE;
818 DWORD shift = (ins->dst[0].token & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
819
820 /* FIXME: support output modifiers */
821
822 /* Handle output register */
823 pshader_get_register_name(ins->shader, ins->dst[0].token, dst_name);
824 shader_arb_get_write_mask(ins, ins->dst[0].token, dst_wmask);
825
826 /* Generate input register names (with modifiers) */
827 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src_name[0]);
828 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[1], 1, src_name[1]);
829 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[2], 2, src_name[2]);
830
831 /* The coissue flag changes the semantic of the cnd instruction in <= 1.3 shaders */
832 if (ins->reg_maps->shader_version <= WINED3DPS_VERSION(1, 3) && ins->coissue)
833 {
834 shader_addline(buffer, "MOV%s %s%s, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask, src_name[1]);
835 } else {
836 shader_addline(buffer, "ADD TMP, -%s, coefdiv.x;\n", src_name[0]);
837 shader_addline(buffer, "CMP%s %s%s, TMP, %s, %s;\n",
838 sat ? "_SAT" : "", dst_name, dst_wmask, src_name[1], src_name[2]);
839 }
840 if (shift != 0)
841 pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
842 }
843
844 static void pshader_hw_cmp(const struct wined3d_shader_instruction *ins)
845 {
846 SHADER_BUFFER *buffer = ins->buffer;
847 char dst_wmask[20];
848 char dst_name[50];
849 char src_name[3][50];
850 DWORD shift = (ins->dst[0].token & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
851 BOOL sat = ins->dst[0].modifiers & WINED3DSPDM_SATURATE;
852
853 /* FIXME: support output modifiers */
854
855 /* Handle output register */
856 pshader_get_register_name(ins->shader, ins->dst[0].token, dst_name);
857 shader_arb_get_write_mask(ins, ins->dst[0].token, dst_wmask);
858
859 /* Generate input register names (with modifiers) */
860 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src_name[0]);
861 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[1], 1, src_name[1]);
862 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[2], 2, src_name[2]);
863
864 shader_addline(buffer, "CMP%s %s%s, %s, %s, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask,
865 src_name[0], src_name[2], src_name[1]);
866
867 if (shift != 0)
868 pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
869 }
870
871 /** Process the WINED3DSIO_DP2ADD instruction in ARB.
872 * dst = dot2(src0, src1) + src2 */
873 static void pshader_hw_dp2add(const struct wined3d_shader_instruction *ins)
874 {
875 SHADER_BUFFER *buffer = ins->buffer;
876 char dst_wmask[20];
877 char dst_name[50];
878 char src_name[3][50];
879 DWORD shift = (ins->dst[0].token & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
880 BOOL sat = ins->dst[0].modifiers & WINED3DSPDM_SATURATE;
881
882 pshader_get_register_name(ins->shader, ins->dst[0].token, dst_name);
883 shader_arb_get_write_mask(ins, ins->dst[0].token, dst_wmask);
884
885 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src_name[0]);
886 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[1], 1, src_name[1]);
887 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[2], 2, src_name[2]);
888
889 /* Emulate a DP2 with a DP3 and 0.0 */
890 shader_addline(buffer, "MOV TMP, %s;\n", src_name[0]);
891 shader_addline(buffer, "MOV TMP.z, 0.0;\n");
892 shader_addline(buffer, "DP3 TMP2, TMP, %s;\n", src_name[1]);
893 shader_addline(buffer, "ADD%s %s%s, TMP2, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask, src_name[2]);
894
895 if (shift != 0)
896 pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
897 }
898
899 /* Map the opcode 1-to-1 to the GL code */
900 static void shader_hw_map2gl(const struct wined3d_shader_instruction *ins)
901 {
902 SHADER_BUFFER *buffer = ins->buffer;
903 const DWORD *src = ins->src;
904 const char *instruction;
905 char arguments[256];
906 unsigned int i;
907
908 switch (ins->handler_idx)
909 {
910 case WINED3DSIH_ABS: instruction = "ABS"; break;
911 case WINED3DSIH_ADD: instruction = "ADD"; break;
912 case WINED3DSIH_CRS: instruction = "XPD"; break;
913 case WINED3DSIH_DP3: instruction = "DP3"; break;
914 case WINED3DSIH_DP4: instruction = "DP4"; break;
915 case WINED3DSIH_DST: instruction = "DST"; break;
916 case WINED3DSIH_EXP: instruction = "EX2"; break;
917 case WINED3DSIH_EXPP: instruction = "EXP"; break;
918 case WINED3DSIH_FRC: instruction = "FRC"; break;
919 case WINED3DSIH_LIT: instruction = "LIT"; break;
920 case WINED3DSIH_LOG: instruction = "LG2"; break;
921 case WINED3DSIH_LOGP: instruction = "LOG"; break;
922 case WINED3DSIH_LRP: instruction = "LRP"; break;
923 case WINED3DSIH_MAD: instruction = "MAD"; break;
924 case WINED3DSIH_MAX: instruction = "MAX"; break;
925 case WINED3DSIH_MIN: instruction = "MIN"; break;
926 case WINED3DSIH_MOV: instruction = "MOV"; break;
927 case WINED3DSIH_MUL: instruction = "MUL"; break;
928 case WINED3DSIH_NOP: instruction = "NOP"; break;
929 case WINED3DSIH_POW: instruction = "POW"; break;
930 case WINED3DSIH_SGE: instruction = "SGE"; break;
931 case WINED3DSIH_SLT: instruction = "SLT"; break;
932 case WINED3DSIH_SUB: instruction = "SUB"; break;
933 default: instruction = "";
934 FIXME("Unhandled opcode %#x\n", ins->handler_idx);
935 break;
936 }
937
938 if (shader_is_pshader_version(ins->reg_maps->shader_version))
939 {
940 /* Output token related */
941 const struct wined3d_shader_dst_param *dst;
942 char output_rname[256];
943 char output_wmask[20];
944 char operands[4][100];
945 BOOL saturate = FALSE;
946 BOOL centroid = FALSE;
947 BOOL partialprecision = FALSE;
948 const char *modifier;
949 DWORD shift;
950
951 if (!(ins->dst_count + ins->src_count))
952 {
953 ERR("Opcode \"%#x\" has no parameters\n", ins->handler_idx);
954 return;
955 }
956 dst = &ins->dst[0];
957
958 /* Process modifiers */
959 if (dst->modifiers)
960 {
961 DWORD mask = dst->modifiers;
962
963 saturate = mask & WINED3DSPDM_SATURATE;
964 centroid = mask & WINED3DSPDM_MSAMPCENTROID;
965 partialprecision = mask & WINED3DSPDM_PARTIALPRECISION;
966 mask &= ~(WINED3DSPDM_MSAMPCENTROID | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_SATURATE);
967 if (mask)
968 FIXME("Unrecognized modifier(%#x)\n", mask >> WINED3DSP_DSTMOD_SHIFT);
969
970 if (centroid)
971 FIXME("Unhandled modifier(%#x)\n", mask >> WINED3DSP_DSTMOD_SHIFT);
972 }
973 shift = (dst->token & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
974 modifier = (saturate && !shift) ? "_SAT" : "";
975
976 /* Generate input register names (with modifiers) */
977 for (i = 0; i < ins->src_count; ++i)
978 {
979 pshader_gen_input_modifier_line(ins->shader, buffer, src[i], i, operands[i + 1]);
980 }
981
982 /* Handle output register */
983 pshader_get_register_name(ins->shader, dst->token, output_rname);
984 strcpy(operands[0], output_rname);
985 shader_arb_get_write_mask(ins, dst->token, output_wmask);
986 strcat(operands[0], output_wmask);
987
988 arguments[0] = '\0';
989 strcat(arguments, operands[0]);
990 for (i = 0; i < ins->src_count; ++i)
991 {
992 strcat(arguments, ", ");
993 strcat(arguments, operands[i + 1]);
994 }
995 shader_addline(buffer, "%s%s %s;\n", instruction, modifier, arguments);
996
997 /* A shift requires another line. */
998 if (shift) pshader_gen_output_modifier_line(buffer, saturate, output_wmask, shift, output_rname);
999 } else {
1000 /* Note that vshader_program_add_param() adds spaces. */
1001
1002 arguments[0] = '\0';
1003 if (ins->dst_count)
1004 {
1005 vshader_program_add_param(ins, ins->dst[0].token, FALSE, arguments);
1006 for (i = 0; i < ins->src_count; ++i)
1007 {
1008 strcat(arguments, ",");
1009 vshader_program_add_param(ins, src[i], TRUE, arguments);
1010 }
1011 }
1012 shader_addline(buffer, "%s%s;\n", instruction, arguments);
1013 }
1014 }
1015
1016 static void shader_hw_mov(const struct wined3d_shader_instruction *ins)
1017 {
1018 IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)ins->shader;
1019
1020 if ((WINED3DSHADER_VERSION_MAJOR(ins->reg_maps->shader_version) == 1
1021 && !shader_is_pshader_version(ins->reg_maps->shader_version)
1022 && ins->dst[0].register_type == WINED3DSPR_ADDR)
1023 || ins->handler_idx == WINED3DSIH_MOVA)
1024 {
1025 SHADER_BUFFER *buffer = ins->buffer;
1026 char src0_param[256];
1027
1028 if (ins->handler_idx == WINED3DSIH_MOVA)
1029 FIXME("mova should round\n");
1030
1031 src0_param[0] = '\0';
1032 if (((IWineD3DVertexShaderImpl *)shader)->rel_offset)
1033 {
1034 vshader_program_add_param(ins, ins->src[0], TRUE, src0_param);
1035 shader_addline(buffer, "ADD TMP.x, %s, helper_const.z;\n", src0_param);
1036 shader_addline(buffer, "ARL A0.x, TMP.x;\n");
1037 }
1038 else
1039 {
1040 /* Apple's ARB_vertex_program implementation does not accept an ARL source argument
1041 * with more than one component. Thus replicate the first source argument over all
1042 * 4 components. For example, .xyzw -> .x (or better: .xxxx), .zwxy -> .z, etc)
1043 */
1044 DWORD parm = ins->src[0] & ~(WINED3DVS_SWIZZLE_MASK);
1045 if((ins->src[0] & WINED3DVS_X_W) == WINED3DVS_X_W)
1046 parm |= WINED3DVS_X_W | WINED3DVS_Y_W | WINED3DVS_Z_W | WINED3DVS_W_W;
1047 else if((ins->src[0] & WINED3DVS_X_Z) == WINED3DVS_X_Z)
1048 parm |= WINED3DVS_X_Z | WINED3DVS_Y_Z | WINED3DVS_Z_Z | WINED3DVS_W_Z;
1049 else if((ins->src[0] & WINED3DVS_X_Y) == WINED3DVS_X_Y)
1050 parm |= WINED3DVS_X_Y | WINED3DVS_Y_Y | WINED3DVS_Z_Y | WINED3DVS_W_Y;
1051 else if((ins->src[0] & WINED3DVS_X_X) == WINED3DVS_X_X)
1052 parm |= WINED3DVS_X_X | WINED3DVS_Y_X | WINED3DVS_Z_X | WINED3DVS_W_X;
1053 vshader_program_add_param(ins, parm, TRUE, src0_param);
1054 shader_addline(buffer, "ARL A0.x, %s;\n", src0_param);
1055 }
1056 }
1057 else
1058 {
1059 shader_hw_map2gl(ins);
1060 }
1061 }
1062
1063 static void pshader_hw_texkill(const struct wined3d_shader_instruction *ins)
1064 {
1065 DWORD shader_version = ins->reg_maps->shader_version;
1066 SHADER_BUFFER *buffer = ins->buffer;
1067 char reg_dest[40];
1068
1069 /* No swizzles are allowed in d3d's texkill. PS 1.x ignores the 4th component as documented,
1070 * but >= 2.0 honors it(undocumented, but tested by the d3d9 testsuit)
1071 */
1072 pshader_get_register_name(ins->shader, ins->dst[0].token, reg_dest);
1073
1074 if (shader_version >= WINED3DPS_VERSION(2,0))
1075 {
1076 /* The arb backend doesn't claim ps 2.0 support, but try to eat what the app feeds to us */
1077 shader_addline(buffer, "KIL %s;\n", reg_dest);
1078 } else {
1079 /* ARB fp doesn't like swizzles on the parameter of the KIL instruction. To mask the 4th component,
1080 * copy the register into our general purpose TMP variable, overwrite .w and pass TMP to KIL
1081 */
1082 shader_addline(buffer, "MOV TMP, %s;\n", reg_dest);
1083 shader_addline(buffer, "MOV TMP.w, one.w;\n");
1084 shader_addline(buffer, "KIL TMP;\n");
1085 }
1086 }
1087
1088 static void pshader_hw_tex(const struct wined3d_shader_instruction *ins)
1089 {
1090 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->shader;
1091 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1092
1093 DWORD dst = ins->dst[0].token;
1094 const DWORD *src = ins->src;
1095 SHADER_BUFFER* buffer = ins->buffer;
1096 DWORD shader_version = ins->reg_maps->shader_version;
1097 BOOL projected = FALSE, bias = FALSE;
1098
1099 char reg_dest[40];
1100 char reg_coord[40];
1101 DWORD reg_dest_code;
1102 DWORD reg_sampler_code;
1103
1104 /* All versions have a destination register */
1105 reg_dest_code = ins->dst[0].register_idx;
1106 pshader_get_register_name(ins->shader, dst, reg_dest);
1107
1108 /* 1.0-1.3: Use destination register as coordinate source.
1109 1.4+: Use provided coordinate source register. */
1110 if (shader_version < WINED3DPS_VERSION(1,4))
1111 strcpy(reg_coord, reg_dest);
1112 else
1113 pshader_gen_input_modifier_line(ins->shader, buffer, src[0], 0, reg_coord);
1114
1115 /* 1.0-1.4: Use destination register number as texture code.
1116 2.0+: Use provided sampler number as texure code. */
1117 if (shader_version < WINED3DPS_VERSION(2,0))
1118 reg_sampler_code = reg_dest_code;
1119 else
1120 reg_sampler_code = src[1] & WINED3DSP_REGNUM_MASK;
1121
1122 /* projection flag:
1123 * 1.1, 1.2, 1.3: Use WINED3DTSS_TEXTURETRANSFORMFLAGS
1124 * 1.4: Use WINED3DSPSM_DZ or WINED3DSPSM_DW on src[0]
1125 * 2.0+: Use WINED3DSI_TEXLD_PROJECT on the opcode
1126 */
1127 if (shader_version < WINED3DPS_VERSION(1,4))
1128 {
1129 DWORD flags = 0;
1130 if(reg_sampler_code < MAX_TEXTURES) {
1131 flags = deviceImpl->stateBlock->textureState[reg_sampler_code][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1132 }
1133 if (flags & WINED3DTTFF_PROJECTED) {
1134 projected = TRUE;
1135 }
1136 }
1137 else if (shader_version < WINED3DPS_VERSION(2,0))
1138 {
1139 DWORD src_mod = ins->src[0] & WINED3DSP_SRCMOD_MASK;
1140 if (src_mod == WINED3DSPSM_DZ) {
1141 projected = TRUE;
1142 } else if(src_mod == WINED3DSPSM_DW) {
1143 projected = TRUE;
1144 }
1145 } else {
1146 if (ins->flags & WINED3DSI_TEXLD_PROJECT) projected = TRUE;
1147 if (ins->flags & WINED3DSI_TEXLD_BIAS) bias = TRUE;
1148 }
1149 shader_hw_sample(ins, reg_sampler_code, reg_dest, reg_coord, projected, bias);
1150 }
1151
1152 static void pshader_hw_texcoord(const struct wined3d_shader_instruction *ins)
1153 {
1154 DWORD dst = ins->dst[0].token;
1155 SHADER_BUFFER *buffer = ins->buffer;
1156
1157 char tmp[20];
1158 shader_arb_get_write_mask(ins, dst, tmp);
1159 if (ins->reg_maps->shader_version != WINED3DPS_VERSION(1,4))
1160 {
1161 DWORD reg = ins->dst[0].register_idx;
1162 shader_addline(buffer, "MOV_SAT T%u%s, fragment.texcoord[%u];\n", reg, tmp, reg);
1163 } else {
1164 DWORD reg1 = ins->dst[0].register_idx;
1165 char reg_src[40];
1166
1167 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, reg_src);
1168 shader_addline(buffer, "MOV R%u%s, %s;\n", reg1, tmp, reg_src);
1169 }
1170 }
1171
1172 static void pshader_hw_texreg2ar(const struct wined3d_shader_instruction *ins)
1173 {
1174 SHADER_BUFFER *buffer = ins->buffer;
1175 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->shader;
1176 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1177 DWORD flags;
1178
1179 DWORD reg1 = ins->dst[0].register_idx;
1180 char dst_str[8];
1181 char src_str[50];
1182
1183 sprintf(dst_str, "T%u", reg1);
1184 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src_str);
1185 shader_addline(buffer, "MOV TMP.x, %s.w;\n", src_str);
1186 shader_addline(buffer, "MOV TMP.y, %s.x;\n", src_str);
1187 flags = reg1 < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg1][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1188 shader_hw_sample(ins, reg1, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1189 }
1190
1191 static void pshader_hw_texreg2gb(const struct wined3d_shader_instruction *ins)
1192 {
1193 SHADER_BUFFER *buffer = ins->buffer;
1194
1195 DWORD reg1 = ins->dst[0].register_idx;
1196 char dst_str[8];
1197 char src_str[50];
1198
1199 sprintf(dst_str, "T%u", reg1);
1200 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src_str);
1201 shader_addline(buffer, "MOV TMP.x, %s.y;\n", src_str);
1202 shader_addline(buffer, "MOV TMP.y, %s.z;\n", src_str);
1203 shader_hw_sample(ins, reg1, dst_str, "TMP", FALSE, FALSE);
1204 }
1205
1206 static void pshader_hw_texreg2rgb(const struct wined3d_shader_instruction *ins)
1207 {
1208 SHADER_BUFFER *buffer = ins->buffer;
1209 DWORD reg1 = ins->dst[0].register_idx;
1210 char dst_str[8];
1211 char src_str[50];
1212
1213 sprintf(dst_str, "T%u", reg1);
1214 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src_str);
1215 shader_hw_sample(ins, reg1, dst_str, src_str, FALSE, FALSE);
1216 }
1217
1218 static void pshader_hw_texbem(const struct wined3d_shader_instruction *ins)
1219 {
1220 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->shader;
1221 BOOL has_bumpmat = FALSE;
1222 BOOL has_luminance = FALSE;
1223 int i;
1224
1225 DWORD dst = ins->dst[0].token;
1226 DWORD src = ins->src[0] & WINED3DSP_REGNUM_MASK;
1227 SHADER_BUFFER* buffer = ins->buffer;
1228
1229 char reg_coord[40];
1230 DWORD reg_dest_code;
1231
1232 /* All versions have a destination register */
1233 reg_dest_code = ins->dst[0].register_idx;
1234 /* Can directly use the name because texbem is only valid for <= 1.3 shaders */
1235 pshader_get_register_name(ins->shader, dst, reg_coord);
1236
1237 for(i = 0; i < This->numbumpenvmatconsts; i++) {
1238 if (This->bumpenvmatconst[i].const_num != WINED3D_CONST_NUM_UNUSED
1239 && reg_dest_code == This->bumpenvmatconst[i].texunit)
1240 {
1241 has_bumpmat = TRUE;
1242 break;
1243 }
1244 }
1245 for(i = 0; i < This->numbumpenvmatconsts; i++) {
1246 if (This->luminanceconst[i].const_num != WINED3D_CONST_NUM_UNUSED
1247 && reg_dest_code == This->luminanceconst[i].texunit)
1248 {
1249 has_luminance = TRUE;
1250 break;
1251 }
1252 }
1253
1254 if(has_bumpmat) {
1255 /* Sampling the perturbation map in Tsrc was done already, including the signedness correction if needed */
1256
1257 shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, x, z, 0, 0;\n", reg_dest_code);
1258 shader_addline(buffer, "DP3 TMP.x, TMP2, T%u;\n", src);
1259 shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, y, w, 0, 0;\n", reg_dest_code);
1260 shader_addline(buffer, "DP3 TMP.y, TMP2, T%u;\n", src);
1261
1262 /* with projective textures, texbem only divides the static texture coord, not the displacement,
1263 * so we can't let the GL handle this.
1264 */
1265 if (((IWineD3DDeviceImpl*) This->baseShader.device)->stateBlock->textureState[reg_dest_code][WINED3DTSS_TEXTURETRANSFORMFLAGS]
1266 & WINED3DTTFF_PROJECTED) {
1267 shader_addline(buffer, "RCP TMP2.w, %s.w;\n", reg_coord);
1268 shader_addline(buffer, "MUL TMP2.xy, %s, TMP2.w;\n", reg_coord);
1269 shader_addline(buffer, "ADD TMP.xy, TMP, TMP2;\n");
1270 } else {
1271 shader_addline(buffer, "ADD TMP.xy, TMP, %s;\n", reg_coord);
1272 }
1273
1274 shader_hw_sample(ins, reg_dest_code, reg_coord, "TMP", FALSE, FALSE);
1275
1276 if (ins->handler_idx == WINED3DSIH_TEXBEML && has_luminance)
1277 {
1278 shader_addline(buffer, "MAD TMP, T%u.z, luminance%d.x, luminance%d.y;\n",
1279 src, reg_dest_code, reg_dest_code);
1280 shader_addline(buffer, "MUL %s, %s, TMP;\n", reg_coord, reg_coord);
1281 }
1282
1283 } else {
1284 DWORD tf;
1285 if(reg_dest_code < MAX_TEXTURES) {
1286 tf = ((IWineD3DDeviceImpl*) This->baseShader.device)->stateBlock->textureState[reg_dest_code][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1287 } else {
1288 tf = 0;
1289 }
1290 /* Without a bump matrix loaded, just sample with the unmodified coordinates */
1291 shader_hw_sample(ins, reg_dest_code, reg_coord, reg_coord, tf & WINED3DTTFF_PROJECTED, FALSE);
1292 }
1293 }
1294
1295 static void pshader_hw_texm3x2pad(const struct wined3d_shader_instruction *ins)
1296 {
1297 DWORD reg = ins->dst[0].register_idx;
1298 SHADER_BUFFER *buffer = ins->buffer;
1299 char src0_name[50];
1300
1301 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src0_name);
1302 shader_addline(buffer, "DP3 TMP.x, T%u, %s;\n", reg, src0_name);
1303 }
1304
1305 static void pshader_hw_texm3x2tex(const struct wined3d_shader_instruction *ins)
1306 {
1307 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->shader;
1308 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1309 DWORD flags;
1310 DWORD reg = ins->dst[0].register_idx;
1311 SHADER_BUFFER *buffer = ins->buffer;
1312 char dst_str[8];
1313 char src0_name[50];
1314
1315 sprintf(dst_str, "T%u", reg);
1316 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src0_name);
1317 shader_addline(buffer, "DP3 TMP.y, T%u, %s;\n", reg, src0_name);
1318 flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1319 shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1320 }
1321
1322 static void pshader_hw_texm3x3pad(const struct wined3d_shader_instruction *ins)
1323 {
1324 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->shader;
1325 DWORD reg = ins->dst[0].register_idx;
1326 SHADER_BUFFER *buffer = ins->buffer;
1327 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1328 char src0_name[50];
1329
1330 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src0_name);
1331 shader_addline(buffer, "DP3 TMP.%c, T%u, %s;\n", 'x' + current_state->current_row, reg, src0_name);
1332 current_state->texcoord_w[current_state->current_row++] = reg;
1333 }
1334
1335 static void pshader_hw_texm3x3tex(const struct wined3d_shader_instruction *ins)
1336 {
1337 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->shader;
1338 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1339 DWORD flags;
1340 DWORD reg = ins->dst[0].register_idx;
1341 SHADER_BUFFER *buffer = ins->buffer;
1342 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1343 char dst_str[8];
1344 char src0_name[50];
1345
1346 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src0_name);
1347 shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
1348
1349 /* Sample the texture using the calculated coordinates */
1350 sprintf(dst_str, "T%u", reg);
1351 flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1352 shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1353 current_state->current_row = 0;
1354 }
1355
1356 static void pshader_hw_texm3x3vspec(const struct wined3d_shader_instruction *ins)
1357 {
1358 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->shader;
1359 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1360 DWORD flags;
1361 DWORD reg = ins->dst[0].register_idx;
1362 SHADER_BUFFER *buffer = ins->buffer;
1363 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1364 char dst_str[8];
1365 char src0_name[50];
1366
1367 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src0_name);
1368 shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
1369
1370 /* Construct the eye-ray vector from w coordinates */
1371 shader_addline(buffer, "MOV TMP2.x, fragment.texcoord[%u].w;\n", current_state->texcoord_w[0]);
1372 shader_addline(buffer, "MOV TMP2.y, fragment.texcoord[%u].w;\n", current_state->texcoord_w[1]);
1373 shader_addline(buffer, "MOV TMP2.z, fragment.texcoord[%u].w;\n", reg);
1374
1375 /* Calculate reflection vector
1376 */
1377 shader_addline(buffer, "DP3 TMP.w, TMP, TMP2;\n");
1378 /* The .w is ignored when sampling, so I can use TMP2.w to calculate dot(N, N) */
1379 shader_addline(buffer, "DP3 TMP2.w, TMP, TMP;\n");
1380 shader_addline(buffer, "RCP TMP2.w, TMP2.w;\n");
1381 shader_addline(buffer, "MUL TMP.w, TMP.w, TMP2.w;\n");
1382 shader_addline(buffer, "MUL TMP, TMP.w, TMP;\n");
1383 shader_addline(buffer, "MAD TMP, coefmul.x, TMP, -TMP2;\n");
1384
1385 /* Sample the texture using the calculated coordinates */
1386 sprintf(dst_str, "T%u", reg);
1387 flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1388 shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1389 current_state->current_row = 0;
1390 }
1391
1392 static void pshader_hw_texm3x3spec(const struct wined3d_shader_instruction *ins)
1393 {
1394 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->shader;
1395 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1396 DWORD flags;
1397 DWORD reg = ins->dst[0].register_idx;
1398 DWORD reg3 = ins->src[1] & WINED3DSP_REGNUM_MASK;
1399 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1400 SHADER_BUFFER *buffer = ins->buffer;
1401 char dst_str[8];
1402 char src0_name[50];
1403
1404 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src0_name);
1405 shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
1406
1407 /* Calculate reflection vector.
1408 *
1409 * dot(N, E)
1410 * TMP.xyz = 2 * --------- * N - E
1411 * dot(N, N)
1412 *
1413 * Which normalizes the normal vector
1414 */
1415 shader_addline(buffer, "DP3 TMP.w, TMP, C[%u];\n", reg3);
1416 shader_addline(buffer, "DP3 TMP2.w, TMP, TMP;\n");
1417 shader_addline(buffer, "RCP TMP2.w, TMP2.w;\n");
1418 shader_addline(buffer, "MUL TMP.w, TMP.w, TMP2.w;\n");
1419 shader_addline(buffer, "MUL TMP, TMP.w, TMP;\n");
1420 shader_addline(buffer, "MAD TMP, coefmul.x, TMP, -C[%u];\n", reg3);
1421
1422 /* Sample the texture using the calculated coordinates */
1423 sprintf(dst_str, "T%u", reg);
1424 flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1425 shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1426 current_state->current_row = 0;
1427 }
1428
1429 static void pshader_hw_texdepth(const struct wined3d_shader_instruction *ins)
1430 {
1431 SHADER_BUFFER *buffer = ins->buffer;
1432 char dst_name[50];
1433
1434 /* texdepth has an implicit destination, the fragment depth value. It's only parameter,
1435 * which is essentially an input, is the destination register because it is the first
1436 * parameter. According to the msdn, this must be register r5, but let's keep it more flexible
1437 * here
1438 */
1439 pshader_get_register_name(ins->shader, ins->dst[0].token, dst_name);
1440
1441 /* According to the msdn, the source register(must be r5) is unusable after
1442 * the texdepth instruction, so we're free to modify it
1443 */
1444 shader_addline(buffer, "MIN %s.y, %s.y, one.y;\n", dst_name, dst_name);
1445
1446 /* How to deal with the special case dst_name.g == 0? if r != 0, then
1447 * the r * (1 / 0) will give infinity, which is clamped to 1.0, the correct
1448 * result. But if r = 0.0, then 0 * inf = 0, which is incorrect.
1449 */
1450 shader_addline(buffer, "RCP %s.y, %s.y;\n", dst_name, dst_name);
1451 shader_addline(buffer, "MUL TMP.x, %s.x, %s.y;\n", dst_name, dst_name);
1452 shader_addline(buffer, "MIN TMP.x, TMP.x, one.x;\n");
1453 shader_addline(buffer, "MAX result.depth, TMP.x, 0.0;\n");
1454 }
1455
1456 /** Process the WINED3DSIO_TEXDP3TEX instruction in ARB:
1457 * Take a 3-component dot product of the TexCoord[dstreg] and src,
1458 * then perform a 1D texture lookup from stage dstregnum, place into dst. */
1459 static void pshader_hw_texdp3tex(const struct wined3d_shader_instruction *ins)
1460 {
1461 SHADER_BUFFER *buffer = ins->buffer;
1462 DWORD sampler_idx = ins->dst[0].register_idx;
1463 char src0[50];
1464 char dst_str[8];
1465
1466 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src0);
1467 shader_addline(buffer, "MOV TMP, 0.0;\n");
1468 shader_addline(buffer, "DP3 TMP.x, T%u, %s;\n", sampler_idx, src0);
1469
1470 sprintf(dst_str, "T%u", sampler_idx);
1471 shader_hw_sample(ins, sampler_idx, dst_str, "TMP", FALSE /* Only one coord, can't be projected */, FALSE);
1472 }
1473
1474 /** Process the WINED3DSIO_TEXDP3 instruction in ARB:
1475 * Take a 3-component dot product of the TexCoord[dstreg] and src. */
1476 static void pshader_hw_texdp3(const struct wined3d_shader_instruction *ins)
1477 {
1478 char src0[50];
1479 char dst_str[50];
1480 char dst_mask[6];
1481 DWORD dstreg = ins->dst[0].register_idx;
1482 SHADER_BUFFER *buffer = ins->buffer;
1483
1484 /* Handle output register */
1485 pshader_get_register_name(ins->shader, ins->dst[0].token, dst_str);
1486 shader_arb_get_write_mask(ins, ins->dst[0].token, dst_mask);
1487
1488 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src0);
1489 shader_addline(buffer, "DP3 %s%s, T%u, %s;\n", dst_str, dst_mask, dstreg, src0);
1490
1491 /* TODO: Handle output modifiers */
1492 }
1493
1494 /** Process the WINED3DSIO_TEXM3X3 instruction in ARB
1495 * Perform the 3rd row of a 3x3 matrix multiply */
1496 static void pshader_hw_texm3x3(const struct wined3d_shader_instruction *ins)
1497 {
1498 SHADER_BUFFER *buffer = ins->buffer;
1499 char dst_str[50];
1500 char dst_mask[6];
1501 char src0[50];
1502 DWORD dst_reg = ins->dst[0].register_idx;
1503
1504 pshader_get_register_name(ins->shader, ins->dst[0].token, dst_str);
1505 shader_arb_get_write_mask(ins, ins->dst[0].token, dst_mask);
1506
1507 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src0);
1508 shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", dst_reg, src0);
1509 shader_addline(buffer, "MOV %s%s, TMP;\n", dst_str, dst_mask);
1510
1511 /* TODO: Handle output modifiers */
1512 }
1513
1514 /** Process the WINED3DSIO_TEXM3X2DEPTH instruction in ARB:
1515 * Last row of a 3x2 matrix multiply, use the result to calculate the depth:
1516 * Calculate tmp0.y = TexCoord[dstreg] . src.xyz; (tmp0.x has already been calculated)
1517 * depth = (tmp0.y == 0.0) ? 1.0 : tmp0.x / tmp0.y
1518 */
1519 static void pshader_hw_texm3x2depth(const struct wined3d_shader_instruction *ins)
1520 {
1521 SHADER_BUFFER *buffer = ins->buffer;
1522 DWORD dst_reg = ins->dst[0].register_idx;
1523 char src0[50];
1524
1525 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src0);
1526 shader_addline(buffer, "DP3 TMP.y, T%u, %s;\n", dst_reg, src0);
1527
1528 /* How to deal with the special case dst_name.g == 0? if r != 0, then
1529 * the r * (1 / 0) will give infinity, which is clamped to 1.0, the correct
1530 * result. But if r = 0.0, then 0 * inf = 0, which is incorrect.
1531 */
1532 shader_addline(buffer, "RCP TMP.y, TMP.y;\n");
1533 shader_addline(buffer, "MUL TMP.x, TMP.x, TMP.y;\n");
1534 shader_addline(buffer, "MIN TMP.x, TMP.x, one.x;\n");
1535 shader_addline(buffer, "MAX result.depth, TMP.x, 0.0;\n");
1536 }
1537
1538 /** Handles transforming all WINED3DSIO_M?x? opcodes for
1539 Vertex/Pixel shaders to ARB_vertex_program codes */
1540 static void shader_hw_mnxn(const struct wined3d_shader_instruction *ins)
1541 {
1542 int i;
1543 int nComponents = 0;
1544 struct wined3d_shader_dst_param tmp_dst = {0};
1545 struct wined3d_shader_instruction tmp_ins;
1546
1547 memset(&tmp_ins, 0, sizeof(tmp_ins));
1548
1549 /* Set constants for the temporary argument */
1550 tmp_ins.shader = ins->shader;
1551 tmp_ins.buffer = ins->buffer;
1552 tmp_ins.src[0] = ins->src[0];
1553 tmp_ins.src_addr[0] = ins->src_addr[0];
1554 tmp_ins.src_addr[1] = ins->src_addr[1];
1555 tmp_ins.reg_maps = ins->reg_maps;
1556 tmp_ins.dst_count = 1;
1557 tmp_ins.dst = &tmp_dst;
1558 tmp_ins.src_count = 2;
1559
1560 switch(ins->handler_idx)
1561 {
1562 case WINED3DSIH_M4x4:
1563 nComponents = 4;
1564 tmp_ins.handler_idx = WINED3DSIH_DP4;
1565 break;
1566 case WINED3DSIH_M4x3:
1567 nComponents = 3;
1568 tmp_ins.handler_idx = WINED3DSIH_DP4;
1569 break;
1570 case WINED3DSIH_M3x4:
1571 nComponents = 4;
1572 tmp_ins.handler_idx = WINED3DSIH_DP3;
1573 break;
1574 case WINED3DSIH_M3x3:
1575 nComponents = 3;
1576 tmp_ins.handler_idx = WINED3DSIH_DP3;
1577 break;
1578 case WINED3DSIH_M3x2:
1579 nComponents = 2;
1580 tmp_ins.handler_idx = WINED3DSIH_DP3;
1581 break;
1582 default:
1583 FIXME("Unhandled opcode %#x\n", ins->handler_idx);
1584 break;
1585 }
1586
1587 tmp_dst = ins->dst[0];
1588 for (i = 0; i < nComponents; i++) {
1589 tmp_dst.write_mask = WINED3DSP_WRITEMASK_0 << i;
1590 tmp_dst.token = (tmp_dst.token & ~WINED3DSP_WRITEMASK_ALL) | tmp_dst.write_mask;
1591 tmp_ins.src[1] = ins->src[1]+i;
1592 shader_hw_map2gl(&tmp_ins);
1593 }
1594 }
1595
1596 static void vshader_hw_rsq_rcp(const struct wined3d_shader_instruction *ins)
1597 {
1598 SHADER_BUFFER *buffer = ins->buffer;
1599 DWORD dst = ins->dst[0].token;
1600 DWORD src = ins->src[0];
1601 DWORD swizzle = (src & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
1602 const char *instruction;
1603
1604 char tmpLine[256];
1605
1606 switch(ins->handler_idx)
1607 {
1608 case WINED3DSIH_RSQ: instruction = "RSQ"; break;
1609 case WINED3DSIH_RCP: instruction = "RCP"; break;
1610 default: instruction = "";
1611 FIXME("Unhandled opcode %#x\n", ins->handler_idx);
1612 break;
1613 }
1614
1615 strcpy(tmpLine, instruction);
1616 vshader_program_add_param(ins, dst, FALSE, tmpLine); /* Destination */
1617 strcat(tmpLine, ",");
1618 vshader_program_add_param(ins, src, TRUE, tmpLine);
1619 if ((WINED3DSP_NOSWIZZLE >> WINED3DSP_SWIZZLE_SHIFT) == swizzle) {
1620 /* Dx sdk says .x is used if no swizzle is given, but our test shows that
1621 * .w is used
1622 */
1623 strcat(tmpLine, ".w");
1624 }
1625
1626 shader_addline(buffer, "%s;\n", tmpLine);
1627 }
1628
1629 static void shader_hw_nrm(const struct wined3d_shader_instruction *ins)
1630 {
1631 SHADER_BUFFER *buffer = ins->buffer;
1632 char dst_name[50];
1633 char src_name[50];
1634 char dst_wmask[20];
1635 DWORD shift = (ins->dst[0].token & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
1636 BOOL sat = ins->dst[0].modifiers & WINED3DSPDM_SATURATE;
1637
1638 pshader_get_register_name(ins->shader, ins->dst[0].token, dst_name);
1639 shader_arb_get_write_mask(ins, ins->dst[0].token, dst_wmask);
1640
1641 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src_name);
1642 shader_addline(buffer, "DP3 TMP, %s, %s;\n", src_name, src_name);
1643 shader_addline(buffer, "RSQ TMP, TMP.x;\n");
1644 /* dst.w = src[0].w * 1 / (src.x^2 + src.y^2 + src.z^2)^(1/2) according to msdn*/
1645 shader_addline(buffer, "MUL%s %s%s, %s, TMP;\n", sat ? "_SAT" : "", dst_name, dst_wmask,
1646 src_name);
1647
1648 if (shift != 0)
1649 pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
1650 }
1651
1652 static void shader_hw_sincos(const struct wined3d_shader_instruction *ins)
1653 {
1654 /* This instruction exists in ARB, but the d3d instruction takes two extra parameters which
1655 * must contain fixed constants. So we need a separate function to filter those constants and
1656 * can't use map2gl
1657 */
1658 SHADER_BUFFER *buffer = ins->buffer;
1659 char dst_name[50];
1660 char src_name[50];
1661 char dst_wmask[20];
1662 DWORD shift = (ins->dst[0].token & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
1663 BOOL sat = ins->dst[0].modifiers & WINED3DSPDM_SATURATE;
1664
1665 pshader_get_register_name(ins->shader, ins->dst[0].token, dst_name);
1666 shader_arb_get_write_mask(ins, ins->dst[0].token, dst_wmask);
1667
1668 pshader_gen_input_modifier_line(ins->shader, buffer, ins->src[0], 0, src_name);
1669 shader_addline(buffer, "SCS%s %s%s, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask,
1670 src_name);
1671
1672 if (shift != 0)
1673 pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
1674
1675 }
1676
1677 static GLuint create_arb_blt_vertex_program(const WineD3D_GL_Info *gl_info)
1678 {
1679 GLuint program_id = 0;
1680 const char *blt_vprogram =
1681 "!!ARBvp1.0\n"
1682 "PARAM c[1] = { { 1, 0.5 } };\n"
1683 "MOV result.position, vertex.position;\n"
1684 "MOV result.color, c[0].x;\n"
1685 "MOV result.texcoord[0], vertex.texcoord[0];\n"
1686 "END\n";
1687
1688 GL_EXTCALL(glGenProgramsARB(1, &program_id));
1689 GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, program_id));
1690 GL_EXTCALL(glProgramStringARB(GL_VERTEX_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(blt_vprogram), blt_vprogram));
1691
1692 if (glGetError() == GL_INVALID_OPERATION) {
1693 GLint pos;
1694 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
1695 FIXME("Vertex program error at position %d: %s\n", pos,
1696 debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
1697 }
1698
1699 return program_id;
1700 }
1701
1702 static GLuint create_arb_blt_fragment_program(const WineD3D_GL_Info *gl_info, enum tex_types tex_type)
1703 {
1704 GLuint program_id = 0;
1705 static const char * const blt_fprograms[tex_type_count] =
1706 {
1707 /* tex_1d */
1708 NULL,
1709 /* tex_2d */
1710 "!!ARBfp1.0\n"
1711 "TEMP R0;\n"
1712 "TEX R0.x, fragment.texcoord[0], texture[0], 2D;\n"
1713 "MOV result.depth.z, R0.x;\n"
1714 "END\n",
1715 /* tex_3d */
1716 NULL,
1717 /* tex_cube */
1718 "!!ARBfp1.0\n"
1719 "TEMP R0;\n"
1720 "TEX R0.x, fragment.texcoord[0], texture[0], CUBE;\n"
1721 "MOV result.depth.z, R0.x;\n"
1722 "END\n",
1723 /* tex_rect */
1724 "!!ARBfp1.0\n"
1725 "TEMP R0;\n"
1726 "TEX R0.x, fragment.texcoord[0], texture[0], RECT;\n"
1727 "MOV result.depth.z, R0.x;\n"
1728 "END\n",
1729 };
1730
1731 if (!blt_fprograms[tex_type])
1732 {
1733 FIXME("tex_type %#x not supported\n", tex_type);
1734 tex_type = tex_2d;
1735 }
1736
1737 GL_EXTCALL(glGenProgramsARB(1, &program_id));
1738 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, program_id));
1739 GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(blt_fprograms[tex_type]), blt_fprograms[tex_type]));
1740
1741 if (glGetError() == GL_INVALID_OPERATION) {
1742 GLint pos;
1743 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
1744 FIXME("Fragment program error at position %d: %s\n", pos,
1745 debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
1746 }
1747
1748 return program_id;
1749 }
1750
1751 static void shader_arb_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {
1752 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1753 struct shader_arb_priv *priv = This->shader_priv;
1754 const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1755
1756 if (useVS) {
1757 struct vs_compile_args compile_args;
1758
1759 TRACE("Using vertex shader\n");
1760 find_vs_compile_args((IWineD3DVertexShaderImpl *) This->stateBlock->vertexShader, This->stateBlock, &compile_args);
1761 priv->current_vprogram_id = find_gl_vshader((IWineD3DVertexShaderImpl *) This->stateBlock->vertexShader, &compile_args);
1762
1763 /* Bind the vertex program */
1764 GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id));
1765 checkGLcall("glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id);");
1766
1767 /* Enable OpenGL vertex programs */
1768 glEnable(GL_VERTEX_PROGRAM_ARB);
1769 checkGLcall("glEnable(GL_VERTEX_PROGRAM_ARB);");
1770 TRACE("(%p) : Bound vertex program %u and enabled GL_VERTEX_PROGRAM_ARB\n", This, priv->current_vprogram_id);
1771 } else if(GL_SUPPORT(ARB_VERTEX_PROGRAM)) {
1772 priv->current_vprogram_id = 0;
1773 glDisable(GL_VERTEX_PROGRAM_ARB);
1774 checkGLcall("glDisable(GL_VERTEX_PROGRAM_ARB)");
1775 }
1776
1777 if (usePS) {
1778 struct ps_compile_args compile_args;
1779 TRACE("Using pixel shader\n");
1780 find_ps_compile_args((IWineD3DPixelShaderImpl *) This->stateBlock->pixelShader, This->stateBlock, &compile_args);
1781 priv->current_fprogram_id = find_gl_pshader((IWineD3DPixelShaderImpl *) This->stateBlock->pixelShader,
1782 &compile_args);
1783
1784 /* Bind the fragment program */
1785 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id));
1786 checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id);");
1787
1788 if(!priv->use_arbfp_fixed_func) {
1789 /* Enable OpenGL fragment programs */
1790 glEnable(GL_FRAGMENT_PROGRAM_ARB);
1791 checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB);");
1792 }
1793 TRACE("(%p) : Bound fragment program %u and enabled GL_FRAGMENT_PROGRAM_ARB\n", This, priv->current_fprogram_id);
1794 } else if(GL_SUPPORT(ARB_FRAGMENT_PROGRAM) && !priv->use_arbfp_fixed_func) {
1795 /* Disable only if we're not using arbfp fixed function fragment processing. If this is used,
1796 * keep GL_FRAGMENT_PROGRAM_ARB enabled, and the fixed function pipeline will bind the fixed function
1797 * replacement shader
1798 */
1799 glDisable(GL_FRAGMENT_PROGRAM_ARB);
1800 checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
1801 priv->current_fprogram_id = 0;
1802 }
1803 }
1804
1805 static void shader_arb_select_depth_blt(IWineD3DDevice *iface, enum tex_types tex_type) {
1806 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1807 struct shader_arb_priv *priv = This->shader_priv;
1808 GLuint *blt_fprogram = &priv->depth_blt_fprogram_id[tex_type];
1809 const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1810
1811 if (!priv->depth_blt_vprogram_id) priv->depth_blt_vprogram_id = create_arb_blt_vertex_program(gl_info);
1812 GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->depth_blt_vprogram_id));
1813 glEnable(GL_VERTEX_PROGRAM_ARB);
1814
1815 if (!*blt_fprogram) *blt_fprogram = create_arb_blt_fragment_program(gl_info, tex_type);
1816 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, *blt_fprogram));
1817 glEnable(GL_FRAGMENT_PROGRAM_ARB);
1818 }
1819
1820 static void shader_arb_deselect_depth_blt(IWineD3DDevice *iface) {
1821 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1822 struct shader_arb_priv *priv = This->shader_priv;
1823 const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1824
1825 if (priv->current_vprogram_id) {
1826 GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id));
1827 checkGLcall("glBindProgramARB(GL_VERTEX_PROGRAM_ARB, vertexShader->prgId);");
1828
1829 glEnable(GL_VERTEX_PROGRAM_ARB);
1830 checkGLcall("glEnable(GL_VERTEX_PROGRAM_ARB);");
1831
1832 TRACE("(%p) : Bound vertex program %u and enabled GL_VERTEX_PROGRAM_ARB\n", This, priv->current_vprogram_id);
1833 } else {
1834 glDisable(GL_VERTEX_PROGRAM_ARB);
1835 checkGLcall("glDisable(GL_VERTEX_PROGRAM_ARB)");
1836 }
1837
1838 if (priv->current_fprogram_id) {
1839 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id));
1840 checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, pixelShader->prgId);");
1841
1842 glEnable(GL_FRAGMENT_PROGRAM_ARB);
1843 checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB);");
1844
1845 TRACE("(%p) : Bound fragment program %u and enabled GL_FRAGMENT_PROGRAM_ARB\n", This, priv->current_fprogram_id);
1846 } else {
1847 glDisable(GL_FRAGMENT_PROGRAM_ARB);
1848 checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
1849 }
1850 }
1851
1852 static void shader_arb_destroy(IWineD3DBaseShader *iface) {
1853 IWineD3DBaseShaderImpl *baseShader = (IWineD3DBaseShaderImpl *) iface;
1854 const WineD3D_GL_Info *gl_info = &((IWineD3DDeviceImpl *)baseShader->baseShader.device)->adapter->gl_info;
1855
1856 if (shader_is_pshader_version(baseShader->baseShader.reg_maps.shader_version))
1857 {
1858 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *) iface;
1859 UINT i;
1860
1861 ENTER_GL();
1862 for(i = 0; i < This->num_gl_shaders; i++) {
1863 GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId));
1864 checkGLcall("GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId))");
1865 }
1866 LEAVE_GL();
1867 HeapFree(GetProcessHeap(), 0, This->gl_shaders);
1868 This->gl_shaders = NULL;
1869 This->num_gl_shaders = 0;
1870 This->shader_array_size = 0;
1871 } else {
1872 IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *) iface;
1873 UINT i;
1874
1875 ENTER_GL();
1876 for(i = 0; i < This->num_gl_shaders; i++) {
1877 GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId));
1878 checkGLcall("GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId))");
1879 }
1880 LEAVE_GL();
1881 HeapFree(GetProcessHeap(), 0, This->gl_shaders);
1882 This->gl_shaders = NULL;
1883 This->num_gl_shaders = 0;
1884 This->shader_array_size = 0;
1885 }
1886 }
1887
1888 static HRESULT shader_arb_alloc(IWineD3DDevice *iface) {
1889 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1890 This->shader_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct shader_arb_priv));
1891 return WINED3D_OK;
1892 }
1893
1894 static void shader_arb_free(IWineD3DDevice *iface) {
1895 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1896 const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1897 struct shader_arb_priv *priv = This->shader_priv;
1898 int i;
1899
1900 if(priv->depth_blt_vprogram_id) {
1901 GL_EXTCALL(glDeleteProgramsARB(1, &priv->depth_blt_vprogram_id));
1902 }
1903 for (i = 0; i < tex_type_count; ++i) {
1904 if (priv->depth_blt_fprogram_id[i]) {
1905 GL_EXTCALL(glDeleteProgramsARB(1, &priv->depth_blt_fprogram_id[i]));
1906 }
1907 }
1908
1909 HeapFree(GetProcessHeap(), 0, This->shader_priv);
1910 }
1911
1912 static BOOL shader_arb_dirty_const(IWineD3DDevice *iface) {
1913 return TRUE;
1914 }
1915
1916 static void arbfp_add_sRGB_correction(SHADER_BUFFER *buffer, const char *fragcolor, const char *tmp1,
1917 const char *tmp2, const char *tmp3, const char *tmp4) {
1918 /* Perform sRGB write correction. See GLX_EXT_framebuffer_sRGB */
1919
1920 /* Calculate the > 0.0031308 case */
1921 shader_addline(buffer, "POW %s.x, %s.x, srgb_pow.x;\n", tmp1, fragcolor);
1922 shader_addline(buffer, "POW %s.y, %s.y, srgb_pow.y;\n", tmp1, fragcolor);
1923 shader_addline(buffer, "POW %s.z, %s.z, srgb_pow.z;\n", tmp1, fragcolor);
1924 shader_addline(buffer, "MUL %s, %s, srgb_mul_hi;\n", tmp1, tmp1);
1925 shader_addline(buffer, "SUB %s, %s, srgb_sub_hi;\n", tmp1, tmp1);
1926 /* Calculate the < case */
1927 shader_addline(buffer, "MUL %s, srgb_mul_low, %s;\n", tmp2, fragcolor);
1928 /* Get 1.0 / 0.0 masks for > 0.0031308 and < 0.0031308 */
1929 shader_addline(buffer, "SLT %s, srgb_comparison, %s;\n", tmp3, fragcolor);
1930 shader_addline(buffer, "SGE %s, srgb_comparison, %s;\n", tmp4, fragcolor);
1931 /* Store the components > 0.0031308 in the destination */
1932 shader_addline(buffer, "MUL %s, %s, %s;\n", fragcolor, tmp1, tmp3);
1933 /* Add the components that are < 0.0031308 */
1934 shader_addline(buffer, "MAD result.color.xyz, %s, %s, %s;\n", tmp2, tmp4, fragcolor);
1935 /* [0.0;1.0] clamping. Not needed, this is done implicitly */
1936 }
1937
1938 static GLuint shader_arb_generate_pshader(IWineD3DPixelShader *iface, SHADER_BUFFER *buffer, const struct ps_compile_args *args) {
1939 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
1940 const shader_reg_maps* reg_maps = &This->baseShader.reg_maps;
1941 CONST DWORD *function = This->baseShader.function;
1942 DWORD shader_version = reg_maps->shader_version;
1943 const WineD3D_GL_Info *gl_info = &((IWineD3DDeviceImpl *)This->baseShader.device)->adapter->gl_info;
1944 const local_constant *lconst;
1945 GLuint retval;
1946 const char *fragcolor;
1947
1948 /* Create the hw ARB shader */
1949 shader_addline(buffer, "!!ARBfp1.0\n");
1950
1951 if (shader_version < WINED3DPS_VERSION(3,0)) {
1952 switch(args->fog) {
1953 case FOG_OFF:
1954 break;
1955 case FOG_LINEAR:
1956 shader_addline(buffer, "OPTION ARB_fog_linear;\n");
1957 break;
1958 case FOG_EXP:
1959 shader_addline(buffer, "OPTION ARB_fog_exp;\n");
1960 break;
1961 case FOG_EXP2:
1962 shader_addline(buffer, "OPTION ARB_fog_exp2;\n");
1963 break;
1964 }
1965 }
1966
1967 shader_addline(buffer, "TEMP TMP;\n"); /* Used in matrix ops */
1968 shader_addline(buffer, "TEMP TMP2;\n"); /* Used in matrix ops */
1969 shader_addline(buffer, "TEMP TA;\n"); /* Used for modifiers */
1970 shader_addline(buffer, "TEMP TB;\n"); /* Used for modifiers */
1971 shader_addline(buffer, "TEMP TC;\n"); /* Used for modifiers */
1972 shader_addline(buffer, "PARAM coefdiv = { 0.5, 0.25, 0.125, 0.0625 };\n");
1973 shader_addline(buffer, "PARAM coefmul = { 2, 4, 8, 16 };\n");
1974 shader_addline(buffer, "PARAM one = { 1.0, 1.0, 1.0, 1.0 };\n");
1975
1976 if (shader_version < WINED3DPS_VERSION(2,0)) {
1977 fragcolor = "R0";
1978 } else {
1979 shader_addline(buffer, "TEMP TMP_COLOR;\n");
1980 fragcolor = "TMP_COLOR";
1981 }
1982
1983 /* Base Declarations */
1984 shader_generate_arb_declarations( (IWineD3DBaseShader*) This, reg_maps, buffer, &GLINFO_LOCATION);
1985
1986 /* Base Shader Body */
1987 shader_generate_main( (IWineD3DBaseShader*) This, buffer, reg_maps, function);
1988
1989 if(args->srgb_correction) {
1990 arbfp_add_sRGB_correction(buffer, fragcolor, "TMP", "TMP2", "TA", "TB");
1991 }
1992 shader_addline(buffer, "MOV result.color, %s;\n", fragcolor);
1993 shader_addline(buffer, "END\n");
1994
1995 /* TODO: change to resource.glObjectHandle or something like that */
1996 GL_EXTCALL(glGenProgramsARB(1, &retval));
1997
1998 TRACE("Creating a hw pixel shader, prg=%d\n", retval);
1999 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, retval));
2000
2001 TRACE("Created hw pixel shader, prg=%d\n", retval);
2002 /* Create the program and check for errors */
2003 GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB,
2004 buffer->bsize, buffer->buffer));
2005
2006 if (glGetError() == GL_INVALID_OPERATION) {
2007 GLint errPos;
2008 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &errPos);
2009 FIXME("HW PixelShader Error at position %d: %s\n",
2010 errPos, debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
2011 retval = 0;
2012 }
2013
2014 /* Load immediate constants */
2015 if(!This->baseShader.load_local_constsF) {
2016 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
2017 const float *value = (const float *)lconst->value;
2018 GL_EXTCALL(glProgramLocalParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, lconst->idx, value));
2019 checkGLcall("glProgramLocalParameter4fvARB");
2020 }
2021 }
2022
2023 return retval;
2024 }
2025
2026 static GLuint shader_arb_generate_vshader(IWineD3DVertexShader *iface, SHADER_BUFFER *buffer, const struct vs_compile_args *args) {
2027 IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *)iface;
2028 const shader_reg_maps *reg_maps = &This->baseShader.reg_maps;
2029 CONST DWORD *function = This->baseShader.function;
2030 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)This->baseShader.device;
2031 const WineD3D_GL_Info *gl_info = &device->adapter->gl_info;
2032 const local_constant *lconst;
2033 GLuint ret;
2034
2035 /* Create the hw ARB shader */
2036 shader_addline(buffer, "!!ARBvp1.0\n");
2037 shader_addline(buffer, "PARAM helper_const = { 2.0, -1.0, %d.0, 0.0 };\n", This->rel_offset);
2038
2039 /* Mesa supports only 95 constants */
2040 if (GL_VEND(MESA) || GL_VEND(WINE))
2041 This->baseShader.limits.constant_float =
2042 min(95, This->baseShader.limits.constant_float);
2043
2044 shader_addline(buffer, "TEMP TMP;\n");
2045
2046 /* Base Declarations */
2047 shader_generate_arb_declarations( (IWineD3DBaseShader*) This, reg_maps, buffer, &GLINFO_LOCATION);
2048
2049 /* We need a constant to fixup the final position */
2050 shader_addline(buffer, "PARAM posFixup = program.env[%d];\n", ARB_SHADER_PRIVCONST_POS);
2051
2052 /* Initialize output parameters. GL_ARB_vertex_program does not require special initialization values
2053 * for output parameters. D3D in theory does not do that either, but some applications depend on a
2054 * proper initialization of the secondary color, and programs using the fixed function pipeline without
2055 * a replacement shader depend on the texcoord.w being set properly.
2056 *
2057 * GL_NV_vertex_program defines that all output values are initialized to {0.0, 0.0, 0.0, 1.0}. This
2058 * assertion is in effect even when using GL_ARB_vertex_program without any NV specific additions. So
2059 * skip this if NV_vertex_program is supported. Otherwise, initialize the secondary color. For the tex-
2060 * coords, we have a flag in the opengl caps. Many cards do not require the texcoord being set, and
2061 * this can eat a number of instructions, so skip it unless this cap is set as well
2062 */
2063 if(!GL_SUPPORT(NV_VERTEX_PROGRAM)) {
2064 shader_addline(buffer, "MOV result.color.secondary, -helper_const.wwwy;\n");
2065
2066 if((GLINFO_LOCATION).set_texcoord_w && !device->frag_pipe->ffp_proj_control) {
2067 int i;
2068 for(i = 0; i < min(8, MAX_REG_TEXCRD); i++) {
2069 if(This->baseShader.reg_maps.texcoord_mask[i] != 0 &&
2070 This->baseShader.reg_maps.texcoord_mask[i] != WINED3DSP_WRITEMASK_ALL) {
2071 shader_addline(buffer, "MOV result.texcoord[%u].w, -helper_const.y;\n", i);
2072 }
2073 }
2074 }
2075 }
2076
2077 /* Base Shader Body */
2078 shader_generate_main( (IWineD3DBaseShader*) This, buffer, reg_maps, function);
2079
2080 /* The D3DRS_FOGTABLEMODE render state defines if the shader-generated fog coord is used
2081 * or if the fragment depth is used. If the fragment depth is used(FOGTABLEMODE != NONE),
2082 * the fog frag coord is thrown away. If the fog frag coord is used, but not written by
2083 * the shader, it is set to 0.0(fully fogged, since start = 1.0, end = 0.0)
2084 */
2085 if(args->fog_src == VS_FOG_Z) {
2086 shader_addline(buffer, "MOV result.fogcoord, TMP_OUT.z;\n");
2087 } else if (!reg_maps->fog) {
2088 shader_addline(buffer, "MOV result.fogcoord, 0.0;\n");
2089 }
2090
2091 /* Write the final position.
2092 *
2093 * OpenGL coordinates specify the center of the pixel while d3d coords specify
2094 * the corner. The offsets are stored in z and w in posFixup. posFixup.y contains
2095 * 1.0 or -1.0 to turn the rendering upside down for offscreen rendering. PosFixup.x
2096 * contains 1.0 to allow a mad, but arb vs swizzles are too restricted for that.
2097 */
2098 shader_addline(buffer, "MUL TMP, posFixup, TMP_OUT.w;\n");
2099 shader_addline(buffer, "ADD TMP_OUT.x, TMP_OUT.x, TMP.z;\n");
2100 shader_addline(buffer, "MAD TMP_OUT.y, TMP_OUT.y, posFixup.y, TMP.w;\n");
2101
2102 /* Z coord [0;1]->[-1;1] mapping, see comment in transform_projection in state.c
2103 * and the glsl equivalent
2104 */
2105 shader_addline(buffer, "MAD TMP_OUT.z, TMP_OUT.z, helper_const.x, -TMP_OUT.w;\n");
2106
2107 shader_addline(buffer, "MOV result.position, TMP_OUT;\n");
2108
2109 shader_addline(buffer, "END\n");
2110
2111 /* TODO: change to resource.glObjectHandle or something like that */
2112 GL_EXTCALL(glGenProgramsARB(1, &ret));
2113
2114 TRACE("Creating a hw vertex shader, prg=%d\n", ret);
2115 GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, ret));
2116
2117 TRACE("Created hw vertex shader, prg=%d\n", ret);
2118 /* Create the program and check for errors */
2119 GL_EXTCALL(glProgramStringARB(GL_VERTEX_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB,
2120 buffer->bsize, buffer->buffer));
2121
2122 if (glGetError() == GL_INVALID_OPERATION) {
2123 GLint errPos;
2124 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &errPos);
2125 FIXME("HW VertexShader Error at position %d: %s\n",
2126 errPos, debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
2127 ret = -1;
2128 } else {
2129 /* Load immediate constants */
2130 if(!This->baseShader.load_local_constsF) {
2131 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
2132 const float *value = (const float *)lconst->value;
2133 GL_EXTCALL(glProgramLocalParameter4fvARB(GL_VERTEX_PROGRAM_ARB, lconst->idx, value));
2134 }
2135 }
2136 }
2137 return ret;
2138 }
2139
2140 static void shader_arb_get_caps(WINED3DDEVTYPE devtype, const WineD3D_GL_Info *gl_info, struct shader_caps *pCaps)
2141 {
2142 /* We don't have an ARB fixed function pipeline yet, so let the none backend set its caps,
2143 * then overwrite the shader specific ones
2144 */
2145 none_shader_backend.shader_get_caps(devtype, gl_info, pCaps);
2146
2147 if(GL_SUPPORT(ARB_VERTEX_PROGRAM)) {
2148 pCaps->VertexShaderVersion = WINED3DVS_VERSION(1,1);
2149 TRACE_(d3d_caps)("Hardware vertex shader version 1.1 enabled (ARB_PROGRAM)\n");
2150 pCaps->MaxVertexShaderConst = GL_LIMITS(vshader_constantsF);
2151 }
2152
2153 if(GL_SUPPORT(ARB_FRAGMENT_PROGRAM)) {
2154 pCaps->PixelShaderVersion = WINED3DPS_VERSION(1,4);
2155 pCaps->PixelShader1xMaxValue = 8.0;
2156 TRACE_(d3d_caps)("Hardware pixel shader version 1.4 enabled (ARB_PROGRAM)\n");
2157 }
2158 }
2159
2160 static BOOL shader_arb_color_fixup_supported(struct color_fixup_desc fixup)
2161 {
2162 if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
2163 {
2164 TRACE("Checking support for color_fixup:\n");
2165 dump_color_fixup_desc(fixup);
2166 }
2167
2168 /* We support everything except YUV conversions. */
2169 if (!is_yuv_fixup(fixup))
2170 {
2171 TRACE("[OK]\n");
2172 return TRUE;
2173 }
2174
2175 TRACE("[FAILED]\n");
2176 return FALSE;
2177 }
2178
2179 static const SHADER_HANDLER shader_arb_instruction_handler_table[WINED3DSIH_TABLE_SIZE] =
2180 {
2181 /* WINED3DSIH_ABS */ shader_hw_map2gl,
2182 /* WINED3DSIH_ADD */ shader_hw_map2gl,
2183 /* WINED3DSIH_BEM */ pshader_hw_bem,
2184 /* WINED3DSIH_BREAK */ NULL,
2185 /* WINED3DSIH_BREAKC */ NULL,
2186 /* WINED3DSIH_BREAKP */ NULL,
2187 /* WINED3DSIH_CALL */ NULL,
2188 /* WINED3DSIH_CALLNZ */ NULL,
2189 /* WINED3DSIH_CMP */ pshader_hw_cmp,
2190 /* WINED3DSIH_CND */ pshader_hw_cnd,
2191 /* WINED3DSIH_CRS */ shader_hw_map2gl,
2192 /* WINED3DSIH_DCL */ NULL,
2193 /* WINED3DSIH_DEF */ NULL,
2194 /* WINED3DSIH_DEFB */ NULL,
2195 /* WINED3DSIH_DEFI */ NULL,
2196 /* WINED3DSIH_DP2ADD */ pshader_hw_dp2add,
2197 /* WINED3DSIH_DP3 */ shader_hw_map2gl,
2198 /* WINED3DSIH_DP4 */ shader_hw_map2gl,
2199 /* WINED3DSIH_DST */ shader_hw_map2gl,
2200 /* WINED3DSIH_DSX */ NULL,
2201 /* WINED3DSIH_DSY */ NULL,
2202 /* WINED3DSIH_ELSE */ NULL,
2203 /* WINED3DSIH_ENDIF */ NULL,
2204 /* WINED3DSIH_ENDLOOP */ NULL,
2205 /* WINED3DSIH_ENDREP */ NULL,
2206 /* WINED3DSIH_EXP */ shader_hw_map2gl,
2207 /* WINED3DSIH_EXPP */ shader_hw_map2gl,
2208 /* WINED3DSIH_FRC */ shader_hw_map2gl,
2209 /* WINED3DSIH_IF */ NULL,
2210 /* WINED3DSIH_IFC */ NULL,
2211 /* WINED3DSIH_LABEL */ NULL,
2212 /* WINED3DSIH_LIT */ shader_hw_map2gl,
2213 /* WINED3DSIH_LOG */ shader_hw_map2gl,
2214 /* WINED3DSIH_LOGP */ shader_hw_map2gl,
2215 /* WINED3DSIH_LOOP */ NULL,
2216 /* WINED3DSIH_LRP */ shader_hw_map2gl,
2217 /* WINED3DSIH_M3x2 */ shader_hw_mnxn,
2218 /* WINED3DSIH_M3x3 */ shader_hw_mnxn,
2219 /* WINED3DSIH_M3x4 */ shader_hw_mnxn,
2220 /* WINED3DSIH_M4x3 */ shader_hw_mnxn,
2221 /* WINED3DSIH_M4x4 */ shader_hw_mnxn,
2222 /* WINED3DSIH_MAD */ shader_hw_map2gl,
2223 /* WINED3DSIH_MAX */ shader_hw_map2gl,
2224 /* WINED3DSIH_MIN */ shader_hw_map2gl,
2225 /* WINED3DSIH_MOV */ shader_hw_mov,
2226 /* WINED3DSIH_MOVA */ shader_hw_mov,
2227 /* WINED3DSIH_MUL */ shader_hw_map2gl,
2228 /* WINED3DSIH_NOP */ shader_hw_map2gl,
2229 /* WINED3DSIH_NRM */ shader_hw_nrm,
2230 /* WINED3DSIH_PHASE */ NULL,
2231 /* WINED3DSIH_POW */ shader_hw_map2gl,
2232 /* WINED3DSIH_RCP */ vshader_hw_rsq_rcp,
2233 /* WINED3DSIH_REP */ NULL,
2234 /* WINED3DSIH_RET */ NULL,
2235 /* WINED3DSIH_RSQ */ vshader_hw_rsq_rcp,
2236 /* WINED3DSIH_SETP */ NULL,
2237 /* WINED3DSIH_SGE */ shader_hw_map2gl,
2238 /* WINED3DSIH_SGN */ NULL,
2239 /* WINED3DSIH_SINCOS */ shader_hw_sincos,
2240 /* WINED3DSIH_SLT */ shader_hw_map2gl,
2241 /* WINED3DSIH_SUB */ shader_hw_map2gl,
2242 /* WINED3DSIH_TEX */ pshader_hw_tex,
2243 /* WINED3DSIH_TEXBEM */ pshader_hw_texbem,
2244 /* WINED3DSIH_TEXBEML */ pshader_hw_texbem,
2245 /* WINED3DSIH_TEXCOORD */ pshader_hw_texcoord,
2246 /* WINED3DSIH_TEXDEPTH */ pshader_hw_texdepth,
2247 /* WINED3DSIH_TEXDP3 */ pshader_hw_texdp3,
2248 /* WINED3DSIH_TEXDP3TEX */ pshader_hw_texdp3tex,
2249 /* WINED3DSIH_TEXKILL */ pshader_hw_texkill,
2250 /* WINED3DSIH_TEXLDD */ NULL,
2251 /* WINED3DSIH_TEXLDL */ NULL,
2252 /* WINED3DSIH_TEXM3x2DEPTH */ pshader_hw_texm3x2depth,
2253 /* WINED3DSIH_TEXM3x2PAD */ pshader_hw_texm3x2pad,
2254 /* WINED3DSIH_TEXM3x2TEX */ pshader_hw_texm3x2tex,
2255 /* WINED3DSIH_TEXM3x3 */ pshader_hw_texm3x3,
2256 /* WINED3DSIH_TEXM3x3DIFF */ NULL,
2257 /* WINED3DSIH_TEXM3x3PAD */ pshader_hw_texm3x3pad,
2258 /* WINED3DSIH_TEXM3x3SPEC */ pshader_hw_texm3x3spec,
2259 /* WINED3DSIH_TEXM3x3TEX */ pshader_hw_texm3x3tex,
2260 /* WINED3DSIH_TEXM3x3VSPEC */ pshader_hw_texm3x3vspec,
2261 /* WINED3DSIH_TEXREG2AR */ pshader_hw_texreg2ar,
2262 /* WINED3DSIH_TEXREG2GB */ pshader_hw_texreg2gb,
2263 /* WINED3DSIH_TEXREG2RGB */ pshader_hw_texreg2rgb,
2264 };
2265
2266 const shader_backend_t arb_program_shader_backend = {
2267 shader_arb_instruction_handler_table,
2268 shader_arb_select,
2269 shader_arb_select_depth_blt,
2270 shader_arb_deselect_depth_blt,
2271 shader_arb_update_float_vertex_constants,
2272 shader_arb_update_float_pixel_constants,
2273 shader_arb_load_constants,
2274 shader_arb_destroy,
2275 shader_arb_alloc,
2276 shader_arb_free,
2277 shader_arb_dirty_const,
2278 shader_arb_generate_pshader,
2279 shader_arb_generate_vshader,
2280 shader_arb_get_caps,
2281 shader_arb_color_fixup_supported,
2282 };
2283
2284 /* ARB_fragment_program fixed function pipeline replacement definitions */
2285 #define ARB_FFP_CONST_TFACTOR 0
2286 #define ARB_FFP_CONST_SPECULAR_ENABLE ((ARB_FFP_CONST_TFACTOR) + 1)
2287 #define ARB_FFP_CONST_CONSTANT(i) ((ARB_FFP_CONST_SPECULAR_ENABLE) + 1 + i)
2288 #define ARB_FFP_CONST_BUMPMAT(i) ((ARB_FFP_CONST_CONSTANT(7)) + 1 + i)
2289 #define ARB_FFP_CONST_LUMINANCE(i) ((ARB_FFP_CONST_BUMPMAT(7)) + 1 + i)
2290
2291 struct arbfp_ffp_desc
2292 {
2293 struct ffp_frag_desc parent;
2294 GLuint shader;
2295 unsigned int num_textures_used;
2296 };
2297
2298 static void arbfp_enable(IWineD3DDevice *iface, BOOL enable) {
2299 if(enable) {
2300 glEnable(GL_FRAGMENT_PROGRAM_ARB);
2301 checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB)");
2302 } else {
2303 glDisable(GL_FRAGMENT_PROGRAM_ARB);
2304 checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
2305 }
2306 }
2307
2308 static HRESULT arbfp_alloc(IWineD3DDevice *iface) {
2309 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
2310 struct shader_arb_priv *priv;
2311 /* Share private data between the shader backend and the pipeline replacement, if both
2312 * are the arb implementation. This is needed to figure out whether ARBfp should be disabled
2313 * if no pixel shader is bound or not
2314 */
2315 if(This->shader_backend == &arb_program_shader_backend) {
2316 This->fragment_priv = This->shader_priv;
2317 } else {
2318 This->fragment_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct shader_arb_priv));
2319 if(!This->fragment_priv) return E_OUTOFMEMORY;
2320 }
2321 priv = This->fragment_priv;
2322 priv->fragment_shaders = hash_table_create(ffp_frag_program_key_hash, ffp_frag_program_key_compare);
2323 priv->use_arbfp_fixed_func = TRUE;
2324 return WINED3D_OK;
2325 }
2326
2327 static void arbfp_free_ffpshader(void *value, void *gli) {
2328 const WineD3D_GL_Info *gl_info = gli;
2329 struct arbfp_ffp_desc *entry_arb = value;
2330
2331 ENTER_GL();
2332 GL_EXTCALL(glDeleteProgramsARB(1, &entry_arb->shader));
2333 checkGLcall("glDeleteProgramsARB(1, &entry_arb->shader)");
2334 HeapFree(GetProcessHeap(), 0, entry_arb);
2335 LEAVE_GL();
2336 }
2337
2338 static void arbfp_free(IWineD3DDevice *iface) {
2339 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
2340 struct shader_arb_priv *priv = This->fragment_priv;
2341
2342 hash_table_destroy(priv->fragment_shaders, arbfp_free_ffpshader, &This->adapter->gl_info);
2343 priv->use_arbfp_fixed_func = FALSE;
2344
2345 if(This->shader_backend != &arb_program_shader_backend) {
2346 HeapFree(GetProcessHeap(), 0, This->fragment_priv);
2347 }
2348 }
2349
2350 static void arbfp_get_caps(WINED3DDEVTYPE devtype, const WineD3D_GL_Info *gl_info, struct fragment_caps *caps)
2351 {
2352 caps->TextureOpCaps = WINED3DTEXOPCAPS_DISABLE |
2353 WINED3DTEXOPCAPS_SELECTARG1 |
2354 WINED3DTEXOPCAPS_SELECTARG2 |
2355 WINED3DTEXOPCAPS_MODULATE4X |
2356 WINED3DTEXOPCAPS_MODULATE2X |
2357 WINED3DTEXOPCAPS_MODULATE |
2358 WINED3DTEXOPCAPS_ADDSIGNED2X |
2359 WINED3DTEXOPCAPS_ADDSIGNED |
2360 WINED3DTEXOPCAPS_ADD |
2361 WINED3DTEXOPCAPS_SUBTRACT |
2362 WINED3DTEXOPCAPS_ADDSMOOTH |
2363 WINED3DTEXOPCAPS_BLENDCURRENTALPHA |
2364 WINED3DTEXOPCAPS_BLENDFACTORALPHA |
2365 WINED3DTEXOPCAPS_BLENDTEXTUREALPHA |
2366 WINED3DTEXOPCAPS_BLENDDIFFUSEALPHA |
2367 WINED3DTEXOPCAPS_BLENDTEXTUREALPHAPM |
2368 WINED3DTEXOPCAPS_MODULATEALPHA_ADDCOLOR |
2369 WINED3DTEXOPCAPS_MODULATECOLOR_ADDALPHA |
2370 WINED3DTEXOPCAPS_MODULATEINVCOLOR_ADDALPHA |
2371 WINED3DTEXOPCAPS_MODULATEINVALPHA_ADDCOLOR |
2372 WINED3DTEXOPCAPS_DOTPRODUCT3 |
2373 WINED3DTEXOPCAPS_MULTIPLYADD |
2374 WINED3DTEXOPCAPS_LERP |
2375 WINED3DTEXOPCAPS_BUMPENVMAP |
2376 WINED3DTEXOPCAPS_BUMPENVMAPLUMINANCE;
2377
2378 /* TODO: Implement WINED3DTEXOPCAPS_PREMODULATE */
2379
2380 caps->MaxTextureBlendStages = 8;
2381 caps->MaxSimultaneousTextures = min(GL_LIMITS(fragment_samplers), 8);
2382
2383 caps->PrimitiveMiscCaps |= WINED3DPMISCCAPS_TSSARGTEMP;
2384 }
2385 #undef GLINFO_LOCATION
2386
2387 #define GLINFO_LOCATION stateblock->wineD3DDevice->adapter->gl_info
2388 static void state_texfactor_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2389 float col[4];
2390 IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2391
2392 /* Don't load the parameter if we're using an arbfp pixel shader, otherwise we'll overwrite
2393 * application provided constants
2394 */
2395 if(device->shader_backend == &arb_program_shader_backend) {
2396 if (use_ps(stateblock)) return;
2397
2398 device = stateblock->wineD3DDevice;
2399 device->activeContext->pshader_const_dirty[ARB_FFP_CONST_TFACTOR] = 1;
2400 device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_TFACTOR + 1);
2401 }
2402
2403 D3DCOLORTOGLFLOAT4(stateblock->renderState[WINED3DRS_TEXTUREFACTOR], col);
2404 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_TFACTOR, col));
2405 checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_TFACTOR, col)");
2406
2407 }
2408
2409 static void state_arb_specularenable(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2410 float col[4];
2411 IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2412
2413 /* Don't load the parameter if we're using an arbfp pixel shader, otherwise we'll overwrite
2414 * application provided constants
2415 */
2416 if(device->shader_backend == &arb_program_shader_backend) {
2417 if (use_ps(stateblock)) return;
2418
2419 device = stateblock->wineD3DDevice;
2420 device->activeContext->pshader_const_dirty[ARB_FFP_CONST_SPECULAR_ENABLE] = 1;
2421 device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_SPECULAR_ENABLE + 1);
2422 }
2423
2424 if(stateblock->renderState[WINED3DRS_SPECULARENABLE]) {
2425 /* The specular color has no alpha */
2426 col[0] = 1.0; col[1] = 1.0;
2427 col[2] = 1.0; col[3] = 0.0;
2428 } else {
2429 col[0] = 0.0; col[1] = 0.0;
2430 col[2] = 0.0; col[3] = 0.0;
2431 }
2432 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_SPECULAR_ENABLE, col));
2433 checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_SPECULAR_ENABLE, col)");
2434 }
2435
2436 static void set_bumpmat_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2437 DWORD stage = (state - STATE_TEXTURESTAGE(0, 0)) / (WINED3D_HIGHEST_TEXTURE_STATE + 1);
2438 IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2439 float mat[2][2];
2440
2441 if (use_ps(stateblock))
2442 {
2443 if(stage != 0 &&
2444 ((IWineD3DPixelShaderImpl *) stateblock->pixelShader)->baseShader.reg_maps.bumpmat[stage]) {
2445 /* The pixel shader has to know the bump env matrix. Do a constants update if it isn't scheduled
2446 * anyway
2447 */
2448 if(!isStateDirty(context, STATE_PIXELSHADERCONSTANT)) {
2449 device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
2450 }
2451 }
2452
2453 if(device->shader_backend == &arb_program_shader_backend) {
2454 /* Exit now, don't set the bumpmat below, otherwise we may overwrite pixel shader constants */
2455 return;
2456 }
2457 } else if(device->shader_backend == &arb_program_shader_backend) {
2458 device->activeContext->pshader_const_dirty[ARB_FFP_CONST_BUMPMAT(stage)] = 1;
2459 device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_BUMPMAT(stage) + 1);
2460 }
2461
2462 mat[0][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT00]);
2463 mat[0][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT01]);
2464 mat[1][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT10]);
2465 mat[1][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT11]);
2466
2467 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_BUMPMAT(stage), &mat[0][0]));
2468 checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_BUMPMAT(stage), &mat[0][0])");
2469 }
2470
2471 static void tex_bumpenvlum_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2472 DWORD stage = (state - STATE_TEXTURESTAGE(0, 0)) / (WINED3D_HIGHEST_TEXTURE_STATE + 1);
2473 IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2474 float param[4];
2475
2476 if (use_ps(stateblock))
2477 {
2478 if(stage != 0 &&
2479 ((IWineD3DPixelShaderImpl *) stateblock->pixelShader)->baseShader.reg_maps.luminanceparams[stage]) {
2480 /* The pixel shader has to know the luminance offset. Do a constants update if it
2481 * isn't scheduled anyway
2482 */
2483 if(!isStateDirty(context, STATE_PIXELSHADERCONSTANT)) {
2484 device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
2485 }
2486 }
2487
2488 if(device->shader_backend == &arb_program_shader_backend) {
2489 /* Exit now, don't set the bumpmat below, otherwise we may overwrite pixel shader constants */
2490 return;
2491 }
2492 } else if(device->shader_backend == &arb_program_shader_backend) {
2493 device->activeContext->pshader_const_dirty[ARB_FFP_CONST_LUMINANCE(stage)] = 1;
2494 device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_LUMINANCE(stage) + 1);
2495 }
2496
2497 param[0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVLSCALE]);
2498 param[1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVLOFFSET]);
2499 param[2] = 0.0;
2500 param[3] = 0.0;
2501
2502 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_LUMINANCE(stage), param));
2503 checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_LUMINANCE(stage), param)");
2504 }
2505
2506 static const char *get_argreg(SHADER_BUFFER *buffer, DWORD argnum, unsigned int stage, DWORD arg) {
2507 const char *ret;
2508
2509 if(arg == ARG_UNUSED) return "unused"; /* This is the marker for unused registers */
2510
2511 switch(arg & WINED3DTA_SELECTMASK) {
2512 case WINED3DTA_DIFFUSE:
2513 ret = "fragment.color.primary"; break;
2514
2515 case WINED3DTA_CURRENT:
2516 if(stage == 0) ret = "fragment.color.primary";
2517 else ret = "ret";
2518 break;
2519
2520 case WINED3DTA_TEXTURE:
2521 switch(stage) {
2522 case 0: ret = "tex0"; break;
2523 case 1: ret = "tex1"; break;
2524 case 2: ret = "tex2"; break;
2525 case 3: ret = "tex3"; break;
2526 case 4: ret = "tex4"; break;
2527 case 5: ret = "tex5"; break;
2528 case 6: ret = "tex6"; break;
2529 case 7: ret = "tex7"; break;
2530 default: ret = "unknown texture";
2531 }
2532 break;
2533
2534 case WINED3DTA_TFACTOR:
2535 ret = "tfactor"; break;
2536
2537 case WINED3DTA_SPECULAR:
2538 ret = "fragment.color.secondary"; break;
2539
2540 case WINED3DTA_TEMP:
2541 ret = "tempreg"; break;
2542
2543 case WINED3DTA_CONSTANT:
2544 FIXME("Implement perstage constants\n");
2545 switch(stage) {
2546 case 0: ret = "const0"; break;
2547 case 1: ret = "const1"; break;
2548 case 2: ret = "const2"; break;
2549 case 3: ret = "const3"; break;
2550 case 4: ret = "const4"; break;
2551 case 5: ret = "const5"; break;
2552 case 6: ret = "const6"; break;
2553 case 7: ret = "const7"; break;
2554 default: ret = "unknown constant";
2555 }
2556 break;
2557
2558 default:
2559 return "unknown";
2560 }
2561
2562 if(arg & WINED3DTA_COMPLEMENT) {
2563 shader_addline(buffer, "SUB arg%u, const.x, %s;\n", argnum, ret);
2564 if(argnum == 0) ret = "arg0";
2565 if(argnum == 1) ret = "arg1";
2566 if(argnum == 2) ret = "arg2";
2567 }
2568 if(arg & WINED3DTA_ALPHAREPLICATE) {
2569 shader_addline(buffer, "MOV arg%u, %s.w;\n", argnum, ret);
2570 if(argnum == 0) ret = "arg0";
2571 if(argnum == 1) ret = "arg1";
2572 if(argnum == 2) ret = "arg2";
2573 }
2574 return ret;
2575 }
2576
2577 static void gen_ffp_instr(SHADER_BUFFER *buffer, unsigned int stage, BOOL color, BOOL alpha,
2578 DWORD dst, DWORD op, DWORD dw_arg0, DWORD dw_arg1, DWORD dw_arg2) {
2579 const char *dstmask, *dstreg, *arg0, *arg1, *arg2;
2580 unsigned int mul = 1;
2581 BOOL mul_final_dest = FALSE;
2582
2583 if(color && alpha) dstmask = "";
2584 else if(color) dstmask = ".xyz";
2585 else dstmask = ".w";
2586
2587 if(dst == tempreg) dstreg = "tempreg";
2588 else dstreg = "ret";
2589
2590 arg0 = get_argreg(buffer, 0, stage, dw_arg0);
2591 arg1 = get_argreg(buffer, 1, stage, dw_arg1);
2592 arg2 = get_argreg(buffer, 2, stage, dw_arg2);
2593
2594 switch(op) {
2595 case WINED3DTOP_DISABLE:
2596 if(stage == 0) shader_addline(buffer, "MOV %s%s, fragment.color.primary;\n", dstreg, dstmask);
2597 break;
2598
2599 case WINED3DTOP_SELECTARG2:
2600 arg1 = arg2;
2601 case WINED3DTOP_SELECTARG1:
2602 shader_addline(buffer, "MOV %s%s, %s;\n", dstreg, dstmask, arg1);
2603 break;
2604
2605 case WINED3DTOP_MODULATE4X:
2606 mul = 2;
2607 case WINED3DTOP_MODULATE2X:
2608 mul *= 2;
2609 if(strcmp(dstreg, "result.color") == 0) {
2610 dstreg = "ret";
2611 mul_final_dest = TRUE;
2612 }
2613 case WINED3DTOP_MODULATE:
2614 shader_addline(buffer, "MUL %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2615 break;
2616
2617 case WINED3DTOP_ADDSIGNED2X:
2618 mul = 2;
2619 if(strcmp(dstreg, "result.color") == 0) {
2620 dstreg = "ret";
2621 mul_final_dest = TRUE;
2622 }
2623 case WINED3DTOP_ADDSIGNED:
2624 shader_addline(buffer, "SUB arg2, %s, const.w;\n", arg2);
2625 arg2 = "arg2";
2626 case WINED3DTOP_ADD:
2627 shader_addline(buffer, "ADD_SAT %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2628 break;
2629
2630 case WINED3DTOP_SUBTRACT:
2631 shader_addline(buffer, "SUB_SAT %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2632 break;
2633
2634 case WINED3DTOP_ADDSMOOTH:
2635 shader_addline(buffer, "SUB arg1, const.x, %s;\n", arg1);
2636 shader_addline(buffer, "MAD_SAT %s%s, arg1, %s, %s;\n", dstreg, dstmask, arg2, arg1);
2637 break;
2638
2639 case WINED3DTOP_BLENDCURRENTALPHA:
2640 arg0 = get_argreg(buffer, 0, stage, WINED3DTA_CURRENT);
2641 shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2642 break;
2643 case WINED3DTOP_BLENDFACTORALPHA:
2644 arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TFACTOR);
2645 shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2646 break;
2647 case WINED3DTOP_BLENDTEXTUREALPHA:
2648 arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TEXTURE);
2649 shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2650 break;
2651 case WINED3DTOP_BLENDDIFFUSEALPHA:
2652 arg0 = get_argreg(buffer, 0, stage, WINED3DTA_DIFFUSE);
2653 shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2654 break;
2655
2656 case WINED3DTOP_BLENDTEXTUREALPHAPM:
2657 arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TEXTURE);
2658 shader_addline(buffer, "SUB arg0.w, const.x, %s.w;\n", arg0);
2659 shader_addline(buffer, "MAD_SAT %s%s, %s, arg0.w, %s;\n", dstreg, dstmask, arg2, arg1);
2660 break;
2661
2662 /* D3DTOP_PREMODULATE ???? */
2663
2664 case WINED3DTOP_MODULATEINVALPHA_ADDCOLOR:
2665 shader_addline(buffer, "SUB arg0.w, const.x, %s;\n", arg1);
2666 shader_addline(buffer, "MAD_SAT %s%s, arg0.w, %s, %s;\n", dstreg, dstmask, arg2, arg1);
2667 break;
2668 case WINED3DTOP_MODULATEALPHA_ADDCOLOR:
2669 shader_addline(buffer, "MAD_SAT %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg1, arg2, arg1);
2670 break;
2671 case WINED3DTOP_MODULATEINVCOLOR_ADDALPHA:
2672 shader_addline(buffer, "SUB arg0, const.x, %s;\n", arg1);
2673 shader_addline(buffer, "MAD_SAT %s%s, arg0, %s, %s.w;\n", dstreg, dstmask, arg2, arg1);
2674 break;
2675 case WINED3DTOP_MODULATECOLOR_ADDALPHA:
2676 shader_addline(buffer, "MAD_SAT %s%s, %s, %s, %s.w;\n", dstreg, dstmask, arg1, arg2, arg1);
2677 break;
2678
2679 case WINED3DTOP_DOTPRODUCT3:
2680 mul = 4;
2681 if(strcmp(dstreg, "result.color") == 0) {
2682 dstreg = "ret";
2683 mul_final_dest = TRUE;
2684 }
2685 shader_addline(buffer, "SUB arg1, %s, const.w;\n", arg1);
2686 shader_addline(buffer, "SUB arg2, %s, const.w;\n", arg2);
2687 shader_addline(buffer, "DP3_SAT %s%s, arg1, arg2;\n", dstreg, dstmask);
2688 break;
2689
2690 case WINED3DTOP_MULTIPLYADD:
2691 shader_addline(buffer, "MAD_SAT %s%s, %s, %s, %s;\n", dstreg, dstmask, arg1, arg2, arg0);
2692 break;
2693
2694 case WINED3DTOP_LERP:
2695 /* The msdn is not quite right here */
2696 shader_addline(buffer, "LRP %s%s, %s, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2697 break;
2698
2699 case WINED3DTOP_BUMPENVMAP:
2700 case WINED3DTOP_BUMPENVMAPLUMINANCE:
2701 /* Those are handled in the first pass of the shader(generation pass 1 and 2) already */
2702 break;
2703
2704 default:
2705 FIXME("Unhandled texture op %08x\n", op);
2706 }
2707
2708 if(mul == 2) {
2709 shader_addline(buffer, "MUL_SAT %s%s, %s, const.y;\n", mul_final_dest ? "result.color" : dstreg, dstmask, dstreg);
2710 } else if(mul == 4) {
2711 shader_addline(buffer, "MUL_SAT %s%s, %s, const.z;\n", mul_final_dest ? "result.color" : dstreg, dstmask, dstreg);
2712 }
2713 }
2714
2715 /* The stateblock is passed for GLINFO_LOCATION */
2716 static GLuint gen_arbfp_ffp_shader(const struct ffp_frag_settings *settings, IWineD3DStateBlockImpl *stateblock)
2717 {
2718 unsigned int stage;
2719 SHADER_BUFFER buffer;
2720 BOOL tex_read[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2721 BOOL bump_used[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2722 BOOL luminance_used[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2723 const char *textype;
2724 const char *instr, *sat;
2725 char colorcor_dst[8];
2726 GLuint ret;
2727 DWORD arg0, arg1, arg2;
2728 BOOL tempreg_used = FALSE, tfactor_used = FALSE;
2729 BOOL op_equal;
2730 const char *final_combiner_src = "ret";
2731
2732 /* Find out which textures are read */
2733 for(stage = 0; stage < MAX_TEXTURES; stage++) {
2734 if(settings->op[stage].cop == WINED3DTOP_DISABLE) break;
2735 arg0 = settings->op[stage].carg0 & WINED3DTA_SELECTMASK;
2736 arg1 = settings->op[stage].carg1 & WINED3DTA_SELECTMASK;
2737 arg2 = settings->op[stage].carg2 & WINED3DTA_SELECTMASK;
2738 if(arg0 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2739 if(arg1 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2740 if(arg2 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2741
2742 if(settings->op[stage].cop == WINED3DTOP_BLENDTEXTUREALPHA) tex_read[stage] = TRUE;
2743 if(settings->op[stage].cop == WINED3DTOP_BLENDTEXTUREALPHAPM) tex_read[stage] = TRUE;
2744 if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAP) {
2745 bump_used[stage] = TRUE;
2746 tex_read[stage] = TRUE;
2747 }
2748 if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2749 bump_used[stage] = TRUE;
2750 tex_read[stage] = TRUE;
2751 luminance_used[stage] = TRUE;
2752 } else if(settings->op[stage].cop == WINED3DTOP_BLENDFACTORALPHA) {
2753 tfactor_used = TRUE;
2754 }
2755
2756 if(arg0 == WINED3DTA_TFACTOR || arg1 == WINED3DTA_TFACTOR || arg2 == WINED3DTA_TFACTOR) {
2757 tfactor_used = TRUE;
2758 }
2759
2760 if(settings->op[stage].dst == tempreg) tempreg_used = TRUE;
2761 if(arg0 == WINED3DTA_TEMP || arg1 == WINED3DTA_TEMP || arg2 == WINED3DTA_TEMP) {
2762 tempreg_used = TRUE;
2763 }
2764
2765 if(settings->op[stage].aop == WINED3DTOP_DISABLE) continue;
2766 arg0 = settings->op[stage].aarg0 & WINED3DTA_SELECTMASK;
2767 arg1 = settings->op[stage].aarg1 & WINED3DTA_SELECTMASK;
2768 arg2 = settings->op[stage].aarg2 & WINED3DTA_SELECTMASK;
2769 if(arg0 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2770 if(arg1 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2771 if(arg2 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2772
2773 if(arg0 == WINED3DTA_TEMP || arg1 == WINED3DTA_TEMP || arg2 == WINED3DTA_TEMP) {
2774 tempreg_used = TRUE;
2775 }
2776 if(arg0 == WINED3DTA_TFACTOR || arg1 == WINED3DTA_TFACTOR || arg2 == WINED3DTA_TFACTOR) {
2777 tfactor_used = TRUE;
2778 }
2779 }
2780
2781 /* Shader header */
2782 shader_buffer_init(&buffer);
2783
2784 shader_addline(&buffer, "!!ARBfp1.0\n");
2785
2786 switch(settings->fog) {
2787 case FOG_OFF: break;
2788 case FOG_LINEAR: shader_addline(&buffer, "OPTION ARB_fog_linear;\n"); break;
2789 case FOG_EXP: shader_addline(&buffer, "OPTION ARB_fog_exp;\n"); break;
2790 case FOG_EXP2: shader_addline(&buffer, "OPTION ARB_fog_exp2;\n"); break;
2791 default: FIXME("Unexpected fog setting %d\n", settings->fog);
2792 }
2793
2794 shader_addline(&buffer, "PARAM const = {1, 2, 4, 0.5};\n");
2795 shader_addline(&buffer, "TEMP TMP;\n");
2796 shader_addline(&buffer, "TEMP ret;\n");
2797 if(tempreg_used || settings->sRGB_write) shader_addline(&buffer, "TEMP tempreg;\n");
2798 shader_addline(&buffer, "TEMP arg0;\n");
2799 shader_addline(&buffer, "TEMP arg1;\n");
2800 shader_addline(&buffer, "TEMP arg2;\n");
2801 for(stage = 0; stage < MAX_TEXTURES; stage++) {
2802 if(!tex_read[stage]) continue;
2803 shader_addline(&buffer, "TEMP tex%u;\n", stage);
2804 if(!bump_used[stage]) continue;
2805 shader_addline(&buffer, "PARAM bumpmat%u = program.env[%u];\n", stage, ARB_FFP_CONST_BUMPMAT(stage));
2806 if(!luminance_used[stage]) continue;
2807 shader_addline(&buffer, "PARAM luminance%u = program.env[%u];\n", stage, ARB_FFP_CONST_LUMINANCE(stage));
2808 }
2809 if(tfactor_used) {
2810 shader_addline(&buffer, "PARAM tfactor = program.env[%u];\n", ARB_FFP_CONST_TFACTOR);
2811 }
2812 shader_addline(&buffer, "PARAM specular_enable = program.env[%u];\n", ARB_FFP_CONST_SPECULAR_ENABLE);
2813
2814 if(settings->sRGB_write) {
2815 shader_addline(&buffer, "PARAM srgb_mul_low = {%f, %f, %f, 1.0};\n",
2816 srgb_mul_low, srgb_mul_low, srgb_mul_low);
2817 shader_addline(&buffer, "PARAM srgb_comparison = {%f, %f, %f, %f};\n",
2818 srgb_cmp, srgb_cmp, srgb_cmp, srgb_cmp);
2819 shader_addline(&buffer, "PARAM srgb_pow = {%f, %f, %f, 1.0};\n",
2820 srgb_pow, srgb_pow, srgb_pow);
2821 shader_addline(&buffer, "PARAM srgb_mul_hi = {%f, %f, %f, 1.0};\n",
2822 srgb_mul_high, srgb_mul_high, srgb_mul_high);
2823 shader_addline(&buffer, "PARAM srgb_sub_hi = {%f, %f, %f, 0.0};\n",
2824 srgb_sub_high, srgb_sub_high, srgb_sub_high);
2825 }
2826
2827 /* Generate texture sampling instructions) */
2828 for(stage = 0; stage < MAX_TEXTURES && settings->op[stage].cop != WINED3DTOP_DISABLE; stage++) {
2829 if(!tex_read[stage]) continue;
2830
2831 switch(settings->op[stage].tex_type) {
2832 case tex_1d: textype = "1D"; break;
2833 case tex_2d: textype = "2D"; break;
2834 case tex_3d: textype = "3D"; break;
2835 case tex_cube: textype = "CUBE"; break;
2836 case tex_rect: textype = "RECT"; break;
2837 default: textype = "unexpected_textype"; break;
2838 }
2839
2840 if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAP ||
2841 settings->op[stage].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2842 sat = "";
2843 } else {
2844 sat = "_SAT";
2845 }
2846
2847 if(settings->op[stage].projected == proj_none) {
2848 instr = "TEX";
2849 } else if(settings->op[stage].projected == proj_count4 ||
2850 settings->op[stage].projected == proj_count3) {
2851 instr = "TXP";
2852 } else {
2853 FIXME("Unexpected projection mode %d\n", settings->op[stage].projected);
2854 instr = "TXP";
2855 }
2856
2857 if(stage > 0 &&
2858 (settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAP ||
2859 settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAPLUMINANCE)) {
2860 shader_addline(&buffer, "SWZ arg1, bumpmat%u, x, z, 0, 0;\n", stage - 1);
2861 shader_addline(&buffer, "DP3 ret.x, arg1, tex%u;\n", stage - 1);
2862 shader_addline(&buffer, "SWZ arg1, bumpmat%u, y, w, 0, 0;\n", stage - 1);
2863 shader_addline(&buffer, "DP3 ret.y, arg1, tex%u;\n", stage - 1);
2864
2865 /* with projective textures, texbem only divides the static texture coord, not the displacement,
2866 * so multiply the displacement with the dividing parameter before passing it to TXP
2867 */
2868 if (settings->op[stage].projected != proj_none) {
2869 if(settings->op[stage].projected == proj_count4) {
2870 shader_addline(&buffer, "MOV ret.w, fragment.texcoord[%u].w;\n", stage);
2871 shader_addline(&buffer, "MUL ret.xyz, ret, fragment.texcoord[%u].w, fragment.texcoord[%u];\n", stage, stage);
2872 } else {
2873 shader_addline(&buffer, "MOV ret.w, fragment.texcoord[%u].z;\n", stage);
2874 shader_addline(&buffer, "MAD ret.xyz, ret, fragment.texcoord[%u].z, fragment.texcoord[%u];\n", stage, stage);
2875 }
2876 } else {
2877 shader_addline(&buffer, "ADD ret, ret, fragment.texcoord[%u];\n", stage);
2878 }
2879
2880 shader_addline(&buffer, "%s%s tex%u, ret, texture[%u], %s;\n",
2881 instr, sat, stage, stage, textype);
2882 if(settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2883 shader_addline(&buffer, "MAD_SAT ret.x, tex%u.z, luminance%u.x, luminance%u.y;\n",
2884 stage - 1, stage - 1, stage - 1);
2885 shader_addline(&buffer, "MUL tex%u, tex%u, ret.x;\n", stage, stage);
2886 }
2887 } else if(settings->op[stage].projected == proj_count3) {
2888 shader_addline(&buffer, "MOV ret, fragment.texcoord[%u];\n", stage);
2889 shader_addline(&buffer, "MOV ret.w, ret.z;\n");
2890 shader_addline(&buffer, "%s%s tex%u, ret, texture[%u], %s;\n",
2891 instr, sat, stage, stage, textype);
2892 } else {
2893 shader_addline(&buffer, "%s%s tex%u, fragment.texcoord[%u], texture[%u], %s;\n",
2894 instr, sat, stage, stage, stage, textype);
2895 }
2896
2897 sprintf(colorcor_dst, "tex%u", stage);
2898 gen_color_correction(&buffer, colorcor_dst, WINED3DSP_WRITEMASK_ALL, "const.x", "const.y",
2899 settings->op[stage].color_fixup);
2900 }
2901
2902 /* Generate the main shader */
2903 for(stage = 0; stage < MAX_TEXTURES; stage++) {
2904 if(settings->op[stage].cop == WINED3DTOP_DISABLE) {
2905 if(stage == 0) {
2906 final_combiner_src = "fragment.color.primary";
2907 }
2908 break;
2909 }
2910
2911 if(settings->op[stage].cop == WINED3DTOP_SELECTARG1 &&
2912 settings->op[stage].aop == WINED3DTOP_SELECTARG1) {
2913 op_equal = settings->op[stage].carg1 == settings->op[stage].aarg1;
2914 } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG1 &&
2915 settings->op[stage].aop == WINED3DTOP_SELECTARG2) {
2916 op_equal = settings->op[stage].carg1 == settings->op[stage].aarg2;
2917 } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG2 &&
2918 settings->op[stage].aop == WINED3DTOP_SELECTARG1) {
2919 op_equal = settings->op[stage].carg2 == settings->op[stage].aarg1;
2920 } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG2 &&
2921 settings->op[stage].aop == WINED3DTOP_SELECTARG2) {
2922 op_equal = settings->op[stage].carg2 == settings->op[stage].aarg2;
2923 } else {
2924 op_equal = settings->op[stage].aop == settings->op[stage].cop &&
2925 settings->op[stage].carg0 == settings->op[stage].aarg0 &&
2926 settings->op[stage].carg1 == settings->op[stage].aarg1 &&
2927 settings->op[stage].carg2 == settings->op[stage].aarg2;
2928 }
2929
2930 if(settings->op[stage].aop == WINED3DTOP_DISABLE) {
2931 gen_ffp_instr(&buffer, stage, TRUE, FALSE, settings->op[stage].dst,
2932 settings->op[stage].cop, settings->op[stage].carg0,
2933 settings->op[stage].carg1, settings->op[stage].carg2);
2934 if(stage == 0) {
2935 shader_addline(&buffer, "MOV ret.w, fragment.color.primary.w;\n");
2936 }
2937 } else if(op_equal) {
2938 gen_ffp_instr(&buffer, stage, TRUE, TRUE, settings->op[stage].dst,
2939 settings->op[stage].cop, settings->op[stage].carg0,
2940 settings->op[stage].carg1, settings->op[stage].carg2);
2941 } else {
2942 gen_ffp_instr(&buffer, stage, TRUE, FALSE, settings->op[stage].dst,
2943 settings->op[stage].cop, settings->op[stage].carg0,
2944 settings->op[stage].carg1, settings->op[stage].carg2);
2945 gen_ffp_instr(&buffer, stage, FALSE, TRUE, settings->op[stage].dst,
2946 settings->op[stage].aop, settings->op[stage].aarg0,
2947 settings->op[stage].aarg1, settings->op[stage].aarg2);
2948 }
2949 }
2950
2951 if(settings->sRGB_write) {
2952 shader_addline(&buffer, "MAD ret, fragment.color.secondary, specular_enable, %s;\n", final_combiner_src);
2953 arbfp_add_sRGB_correction(&buffer, "ret", "arg0", "arg1", "arg2", "tempreg");
2954 shader_addline(&buffer, "MOV result.color.w, ret.w;\n");
2955 } else {
2956 shader_addline(&buffer, "MAD result.color, fragment.color.secondary, specular_enable, %s;\n", final_combiner_src);
2957 }
2958
2959 /* Footer */
2960 shader_addline(&buffer, "END\n");
2961
2962 /* Generate the shader */
2963 GL_EXTCALL(glGenProgramsARB(1, &ret));
2964 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, ret));
2965 GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(buffer.buffer), buffer.buffer));
2966
2967 if (glGetError() == GL_INVALID_OPERATION) {
2968 GLint pos;
2969 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
2970 FIXME("Fragment program error at position %d: %s\n", pos,
2971 debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
2972 }
2973 shader_buffer_free(&buffer);
2974 return ret;
2975 }
2976
2977 static void fragment_prog_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2978 IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2979 struct shader_arb_priv *priv = device->fragment_priv;
2980 BOOL use_pshader = use_ps(stateblock);
2981 BOOL use_vshader = use_vs(stateblock);
2982 struct ffp_frag_settings settings;
2983 const struct arbfp_ffp_desc *desc;
2984 unsigned int i;
2985
2986 TRACE("state %#x, stateblock %p, context %p\n", state, stateblock, context);
2987
2988 if(isStateDirty(context, STATE_RENDER(WINED3DRS_FOGENABLE))) {
2989 if(!use_pshader && device->shader_backend == &arb_program_shader_backend && context->last_was_pshader) {
2990 /* Reload fixed function constants since they collide with the pixel shader constants */
2991 for(i = 0; i < MAX_TEXTURES; i++) {
2992 set_bumpmat_arbfp(STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT00), stateblock, context);
2993 }
2994 state_texfactor_arbfp(STATE_RENDER(WINED3DRS_TEXTUREFACTOR), stateblock, context);
2995 state_arb_specularenable(STATE_RENDER(WINED3DRS_SPECULARENABLE), stateblock, context);
2996 } else if(use_pshader && !isStateDirty(context, device->StateTable[STATE_VSHADER].representative)) {
2997 device->shader_backend->shader_select((IWineD3DDevice *)stateblock->wineD3DDevice, use_pshader, use_vshader);
2998 }
2999 return;
3000 }
3001
3002 if(!use_pshader) {
3003 /* Find or create a shader implementing the fixed function pipeline settings, then activate it */
3004 gen_ffp_frag_op(stateblock, &settings, FALSE);
3005 desc = (const struct arbfp_ffp_desc *)find_ffp_frag_shader(priv->fragment_shaders, &settings);
3006 if(!desc) {
3007 struct arbfp_ffp_desc *new_desc = HeapAlloc(GetProcessHeap(), 0, sizeof(*new_desc));
3008 if (!new_desc)
3009 {
3010 ERR("Out of memory\n");
3011 return;
3012 }
3013 new_desc->num_textures_used = 0;
3014 for(i = 0; i < GL_LIMITS(texture_stages); i++) {
3015 if(settings.op[i].cop == WINED3DTOP_DISABLE) break;
3016 new_desc->num_textures_used = i;
3017 }
3018
3019 memcpy(&new_desc->parent.settings, &settings, sizeof(settings));
3020 new_desc->shader = gen_arbfp_ffp_shader(&settings, stateblock);
3021 add_ffp_frag_shader(priv->fragment_shaders, &new_desc->parent);
3022 TRACE("Allocated fixed function replacement shader descriptor %p\n", new_desc);
3023 desc = new_desc;
3024 }
3025
3026 /* Now activate the replacement program. GL_FRAGMENT_PROGRAM_ARB is already active(however, note the
3027 * comment above the shader_select call below). If e.g. GLSL is active, the shader_select call will
3028 * deactivate it.
3029 */
3030 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, desc->shader));
3031 checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, desc->shader)");
3032 priv->current_fprogram_id = desc->shader;
3033
3034 if(device->shader_backend == &arb_program_shader_backend && context->last_was_pshader) {
3035 /* Reload fixed function constants since they collide with the pixel shader constants */
3036 for(i = 0; i < MAX_TEXTURES; i++) {
3037 set_bumpmat_arbfp(STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT00), stateblock, context);
3038 }
3039 state_texfactor_arbfp(STATE_RENDER(WINED3DRS_TEXTUREFACTOR), stateblock, context);
3040 state_arb_specularenable(STATE_RENDER(WINED3DRS_SPECULARENABLE), stateblock, context);
3041 }
3042 context->last_was_pshader = FALSE;
3043 } else {
3044 context->last_was_pshader = TRUE;
3045 }
3046
3047 /* Finally, select the shader. If a pixel shader is used, it will be set and enabled by the shader backend.
3048 * If this shader backend is arbfp(most likely), then it will simply overwrite the last fixed function replace-
3049 * ment shader. If the shader backend is not ARB, it currently is important that the opengl implementation
3050 * type overwrites GL_ARB_fragment_program. This is currently the case with GLSL. If we really want to use
3051 * atifs or nvrc pixel shaders with arb fragment programs we'd have to disable GL_FRAGMENT_PROGRAM_ARB here
3052 *
3053 * Don't call shader_select if the vertex shader is dirty, because it will be called later on by the vertex
3054 * shader handler
3055 */
3056 if(!isStateDirty(context, device->StateTable[STATE_VSHADER].representative)) {
3057 device->shader_backend->shader_select((IWineD3DDevice *)stateblock->wineD3DDevice, use_pshader, use_vshader);
3058
3059 if (!isStateDirty(context, STATE_VERTEXSHADERCONSTANT) && (use_vshader || use_pshader)) {
3060 device->StateTable[STATE_VERTEXSHADERCONSTANT].apply(STATE_VERTEXSHADERCONSTANT, stateblock, context);
3061 }
3062 }
3063 if(use_pshader) {
3064 device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
3065 }
3066 }
3067
3068 /* We can't link the fog states to the fragment state directly since the vertex pipeline links them
3069 * to FOGENABLE. A different linking in different pipeline parts can't be expressed in the combined
3070 * state table, so we need to handle that with a forwarding function. The other invisible side effect
3071 * is that changing the fog start and fog end(which links to FOGENABLE in vertex) results in the
3072 * fragment_prog_arbfp function being called because FOGENABLE is dirty, which calls this function here
3073 */
3074 static void state_arbfp_fog(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
3075 enum fogsource new_source;
3076
3077 TRACE("state %#x, stateblock %p, context %p\n", state, stateblock, context);
3078
3079 if(!isStateDirty(context, STATE_PIXELSHADER)) {
3080 fragment_prog_arbfp(state, stateblock, context);
3081 }
3082
3083 if(!stateblock->renderState[WINED3DRS_FOGENABLE]) return;
3084
3085 if(stateblock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE) {
3086 if(use_vs(stateblock)) {
3087 new_source = FOGSOURCE_VS;
3088 } else {
3089 if(stateblock->renderState[WINED3DRS_FOGVERTEXMODE] == WINED3DFOG_NONE || context->last_was_rhw) {
3090 new_source = FOGSOURCE_COORD;
3091 } else {
3092 new_source = FOGSOURCE_FFP;
3093 }
3094 }
3095 } else {
3096 new_source = FOGSOURCE_FFP;
3097 }
3098 if(new_source != context->fog_source) {
3099 context->fog_source = new_source;
3100 state_fogstartend(STATE_RENDER(WINED3DRS_FOGSTART), stateblock, context);
3101 }
3102 }
3103
3104 static void textransform(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
3105 if(!isStateDirty(context, STATE_PIXELSHADER)) {
3106 fragment_prog_arbfp(state, stateblock, context);
3107 }
3108 }
3109
3110 #undef GLINFO_LOCATION
3111
3112 static const struct StateEntryTemplate arbfp_fragmentstate_template[] = {
3113 {STATE_RENDER(WINED3DRS_TEXTUREFACTOR), { STATE_RENDER(WINED3DRS_TEXTUREFACTOR), state_texfactor_arbfp }, 0 },
3114 {STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3115 {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3116 {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3117 {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3118 {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3119 {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3120 {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3121 {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3122 {STATE_TEXTURESTAGE(0, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3123 {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3124 {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3125 {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3126 {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3127 {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3128 {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3129 {STATE_TEXTURESTAGE(1, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3130 {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3131 {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3132 {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3133 {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3134 {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3135 {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3136 {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3137 {STATE_TEXTURESTAGE(1, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3138 {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3139 {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3140 {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3141 {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3142 {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3143 {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3144 {STATE_TEXTURESTAGE(2, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3145 {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3146 {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3147 {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3148 {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3149 {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3150 {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3151 {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3152 {STATE_TEXTURESTAGE(2, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3153 {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3154 {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3155 {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3156 {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3157 {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3158 {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3159 {STATE_TEXTURESTAGE(3, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3160 {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3161 {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3162 {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3163 {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3164 {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3165 {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3166 {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3167 {STATE_TEXTURESTAGE(3, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3168 {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3169 {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3170 {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3171 {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3172 {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3173 {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3174 {STATE_TEXTURESTAGE(4, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3175 {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3176 {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3177 {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3178 {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3179 {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3180 {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3181 {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3182 {STATE_TEXTURESTAGE(4, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3183 {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3184 {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3185 {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3186 {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3187 {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3188 {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3189 {STATE_TEXTURESTAGE(5, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3190 {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3191 {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3192 {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3193 {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3194 {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3195 {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3196 {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3197 {STATE_TEXTURESTAGE(5, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3198 {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3199 {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3200 {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3201 {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3202 {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3203 {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3204 {STATE_TEXTURESTAGE(6, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3205 {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3206 {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3207 {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3208 {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3209 {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3210 {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3211 {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3212 {STATE_TEXTURESTAGE(6, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3213 {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3214 {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3215 {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3216 {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3217 {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3218 {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3219 {STATE_TEXTURESTAGE(7, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3220 {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3221 {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3222 {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3223 {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3224 {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3225 {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3226 {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3227 {STATE_TEXTURESTAGE(7, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3228 {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3229 {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3230 {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3231 {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3232 {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3233 {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3234 {STATE_SAMPLER(0), { STATE_SAMPLER(0), sampler_texdim }, 0 },
3235 {STATE_SAMPLER(1), { STATE_SAMPLER(1), sampler_texdim }, 0 },
3236 {STATE_SAMPLER(2), { STATE_SAMPLER(2), sampler_texdim }, 0 },
3237 {STATE_SAMPLER(3), { STATE_SAMPLER(3), sampler_texdim }, 0 },
3238 {STATE_SAMPLER(4), { STATE_SAMPLER(4), sampler_texdim }, 0 },
3239 {STATE_SAMPLER(5), { STATE_SAMPLER(5), sampler_texdim }, 0 },
3240 {STATE_SAMPLER(6), { STATE_SAMPLER(6), sampler_texdim }, 0 },
3241 {STATE_SAMPLER(7), { STATE_SAMPLER(7), sampler_texdim }, 0 },
3242 {STATE_PIXELSHADER, { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3243 {STATE_RENDER(WINED3DRS_FOGENABLE), { STATE_RENDER(WINED3DRS_FOGENABLE), state_arbfp_fog }, 0 },
3244 {STATE_RENDER(WINED3DRS_FOGTABLEMODE), { STATE_RENDER(WINED3DRS_FOGENABLE), state_arbfp_fog }, 0 },
3245 {STATE_RENDER(WINED3DRS_FOGVERTEXMODE), { STATE_RENDER(WINED3DRS_FOGENABLE), state_arbfp_fog }, 0 },
3246 {STATE_RENDER(WINED3DRS_FOGSTART), { STATE_RENDER(WINED3DRS_FOGSTART), state_fogstartend }, 0 },
3247 {STATE_RENDER(WINED3DRS_FOGEND), { STATE_RENDER(WINED3DRS_FOGSTART), state_fogstartend }, 0 },
3248 {STATE_RENDER(WINED3DRS_SRGBWRITEENABLE), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3249 {STATE_RENDER(WINED3DRS_FOGCOLOR), { STATE_RENDER(WINED3DRS_FOGCOLOR), state_fogcolor }, 0 },
3250 {STATE_RENDER(WINED3DRS_FOGDENSITY), { STATE_RENDER(WINED3DRS_FOGDENSITY), state_fogdensity }, 0 },
3251 {STATE_TEXTURESTAGE(0,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(0, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3252 {STATE_TEXTURESTAGE(1,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(1, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3253 {STATE_TEXTURESTAGE(2,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(2, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3254 {STATE_TEXTURESTAGE(3,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(3, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3255 {STATE_TEXTURESTAGE(4,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(4, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3256 {STATE_TEXTURESTAGE(5,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(5, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3257 {STATE_TEXTURESTAGE(6,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(6, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3258 {STATE_TEXTURESTAGE(7,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(7, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3259 {STATE_RENDER(WINED3DRS_SPECULARENABLE), { STATE_RENDER(WINED3DRS_SPECULARENABLE), state_arb_specularenable}, 0 },
3260 {0 /* Terminate */, { 0, 0 }, 0 },
3261 };
3262
3263 const struct fragment_pipeline arbfp_fragment_pipeline = {
3264 arbfp_enable,
3265 arbfp_get_caps,
3266 arbfp_alloc,
3267 arbfp_free,
3268 shader_arb_color_fixup_supported,
3269 arbfp_fragmentstate_template,
3270 TRUE /* We can disable projected textures */
3271 };
3272
3273 #define GLINFO_LOCATION device->adapter->gl_info
3274
3275 struct arbfp_blit_priv {
3276 GLenum yuy2_rect_shader, yuy2_2d_shader;
3277 GLenum uyvy_rect_shader, uyvy_2d_shader;
3278 GLenum yv12_rect_shader, yv12_2d_shader;
3279 };
3280
3281 static HRESULT arbfp_blit_alloc(IWineD3DDevice *iface) {
3282 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3283 device->blit_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct arbfp_blit_priv));
3284 if(!device->blit_priv) {
3285 ERR("Out of memory\n");
3286 return E_OUTOFMEMORY;
3287 }
3288 return WINED3D_OK;
3289 }
3290 static void arbfp_blit_free(IWineD3DDevice *iface) {
3291 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3292 struct arbfp_blit_priv *priv = device->blit_priv;
3293
3294 ENTER_GL();
3295 GL_EXTCALL(glDeleteProgramsARB(1, &priv->yuy2_rect_shader));
3296 GL_EXTCALL(glDeleteProgramsARB(1, &priv->yuy2_2d_shader));
3297 GL_EXTCALL(glDeleteProgramsARB(1, &priv->uyvy_rect_shader));
3298 GL_EXTCALL(glDeleteProgramsARB(1, &priv->uyvy_2d_shader));
3299 GL_EXTCALL(glDeleteProgramsARB(1, &priv->yv12_rect_shader));
3300 GL_EXTCALL(glDeleteProgramsARB(1, &priv->yv12_2d_shader));
3301 checkGLcall("Delete yuv programs\n");
3302 LEAVE_GL();
3303 }
3304
3305 static BOOL gen_planar_yuv_read(SHADER_BUFFER *buffer, enum yuv_fixup yuv_fixup, GLenum textype, char *luminance)
3306 {
3307 char chroma;
3308 const char *tex, *texinstr;
3309
3310 if (yuv_fixup == YUV_FIXUP_UYVY) {
3311 chroma = 'x';
3312 *luminance = 'w';
3313 } else {
3314 chroma = 'w';
3315 *luminance = 'x';
3316 }
3317 switch(textype) {
3318 case GL_TEXTURE_2D: tex = "2D"; texinstr = "TXP"; break;
3319 case GL_TEXTURE_RECTANGLE_ARB: tex = "RECT"; texinstr = "TEX"; break;
3320 default:
3321 /* This is more tricky than just replacing the texture type - we have to navigate
3322 * properly in the texture to find the correct chroma values
3323 */
3324 FIXME("Implement yuv correction for non-2d, non-rect textures\n");
3325 return FALSE;
3326 }
3327
3328 /* First we have to read the chroma values. This means we need at least two pixels(no filtering),
3329 * or 4 pixels(with filtering). To get the unmodified chromas, we have to rid ourselves of the
3330 * filtering when we sample the texture.
3331 *
3332 * These are the rules for reading the chroma:
3333 *
3334 * Even pixel: Cr
3335 * Even pixel: U
3336 * Odd pixel: V
3337 *
3338 * So we have to get the sampling x position in non-normalized coordinates in integers
3339 */
3340 if(textype != GL_TEXTURE_RECTANGLE_ARB) {
3341 shader_addline(buffer, "MUL texcrd.xy, fragment.texcoord[0], size.x;\n");
3342 shader_addline(buffer, "MOV texcrd.w, size.x;\n");
3343 } else {
3344 shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3345 }
3346 /* We must not allow filtering between pixel x and x+1, this would mix U and V
3347 * Vertical filtering is ok. However, bear in mind that the pixel center is at
3348 * 0.5, so add 0.5.
3349 */
3350 shader_addline(buffer, "FLR texcrd.x, texcrd.x;\n");
3351 shader_addline(buffer, "ADD texcrd.x, texcrd.x, coef.y;\n");
3352
3353 /* Divide the x coordinate by 0.5 and get the fraction. This gives 0.25 and 0.75 for the
3354 * even and odd pixels respectively
3355 */
3356 shader_addline(buffer, "MUL texcrd2, texcrd, coef.y;\n");
3357 shader_addline(buffer, "FRC texcrd2, texcrd2;\n");
3358
3359 /* Sample Pixel 1 */
3360 shader_addline(buffer, "%s luminance, texcrd, texture[0], %s;\n", texinstr, tex);
3361
3362 /* Put the value into either of the chroma values */
3363 shader_addline(buffer, "SGE temp.x, texcrd2.x, coef.y;\n");
3364 shader_addline(buffer, "MUL chroma.x, luminance.%c, temp.x;\n", chroma);
3365 shader_addline(buffer, "SLT temp.x, texcrd2.x, coef.y;\n");
3366 shader_addline(buffer, "MUL chroma.y, luminance.%c, temp.x;\n", chroma);
3367
3368 /* Sample pixel 2. If we read an even pixel(SLT above returned 1), sample
3369 * the pixel right to the current one. Otherwise, sample the left pixel.
3370 * Bias and scale the SLT result to -1;1 and add it to the texcrd.x.
3371 */
3372 shader_addline(buffer, "MAD temp.x, temp.x, coef.z, -coef.x;\n");
3373 shader_addline(buffer, "ADD texcrd.x, texcrd, temp.x;\n");
3374 shader_addline(buffer, "%s luminance, texcrd, texture[0], %s;\n", texinstr, tex);
3375
3376 /* Put the value into the other chroma */
3377 shader_addline(buffer, "SGE temp.x, texcrd2.x, coef.y;\n");
3378 shader_addline(buffer, "MAD chroma.y, luminance.%c, temp.x, chroma.y;\n", chroma);
3379 shader_addline(buffer, "SLT temp.x, texcrd2.x, coef.y;\n");
3380 shader_addline(buffer, "MAD chroma.x, luminance.%c, temp.x, chroma.x;\n", chroma);
3381
3382 /* TODO: If filtering is enabled, sample a 2nd pair of pixels left or right of
3383 * the current one and lerp the two U and V values
3384 */
3385
3386 /* This gives the correctly filtered luminance value */
3387 shader_addline(buffer, "TEX luminance, fragment.texcoord[0], texture[0], %s;\n", tex);
3388
3389 return TRUE;
3390 }
3391
3392 static BOOL gen_yv12_read(SHADER_BUFFER *buffer, GLenum textype, char *luminance)
3393 {
3394 const char *tex;
3395
3396 switch(textype) {
3397 case GL_TEXTURE_2D: tex = "2D"; break;
3398 case GL_TEXTURE_RECTANGLE_ARB: tex = "RECT"; break;
3399 default:
3400 FIXME("Implement yv12 correction for non-2d, non-rect textures\n");
3401 return FALSE;
3402 }
3403
3404 /* YV12 surfaces contain a WxH sized luminance plane, followed by a (W/2)x(H/2)
3405 * V and a (W/2)x(H/2) U plane, each with 8 bit per pixel. So the effective
3406 * bitdepth is 12 bits per pixel. Since the U and V planes have only half the
3407 * pitch of the luminance plane, the packing into the gl texture is a bit
3408 * unfortunate. If the whole texture is interpreted as luminance data it looks
3409 * approximately like this:
3410 *
3411 * +----------------------------------+----
3412 * | |
3413 * | |
3414 * | |
3415 * | |
3416 * | | 2
3417 * | LUMINANCE | -
3418 * | | 3
3419 * | |
3420 * | |
3421 * | |
3422 * | |
3423 * +----------------+-----------------+----
3424 * | | |
3425 * | U even rows | U odd rows |
3426 * | | | 1
3427 * +----------------+------------------ -
3428 * | | | 3
3429 * | V even rows | V odd rows |
3430 * | | |
3431 * +----------------+-----------------+----
3432 * | | |
3433 * | 0.5 | 0.5 |
3434 *
3435 * So it appears as if there are 4 chroma images, but in fact the odd rows
3436 * in the chroma images are in the same row as the even ones. So its is
3437 * kinda tricky to read
3438 *
3439 * When reading from rectangle textures, keep in mind that the input y coordinates
3440 * go from 0 to d3d_height, whereas the opengl texture height is 1.5 * d3d_height
3441 */
3442 shader_addline(buffer, "PARAM yv12_coef = {%f, %f, %f, %f};\n",
3443 2.0 / 3.0, 1.0 / 6.0, (2.0 / 3.0) + (1.0 / 6.0), 1.0 / 3.0);
3444
3445 shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3446 /* the chroma planes have only half the width */
3447 shader_addline(buffer, "MUL texcrd.x, texcrd.x, coef.y;\n");
3448
3449 /* The first value is between 2/3 and 5/6th of the texture's height, so scale+bias
3450 * the coordinate. Also read the right side of the image when reading odd lines
3451 *
3452 * Don't forget to clamp the y values in into the range, otherwise we'll get filtering
3453 * bleeding
3454 */
3455 if(textype == GL_TEXTURE_2D) {
3456
3457 shader_addline(buffer, "RCP chroma.w, size.y;\n");
3458
3459 shader_addline(buffer, "MUL texcrd2.y, texcrd.y, size.y;\n");
3460
3461 shader_addline(buffer, "FLR texcrd2.y, texcrd2.y;\n");
3462 shader_addline(buffer, "MAD texcrd.y, texcrd.y, yv12_coef.y, yv12_coef.x;\n");
3463
3464 /* Read odd lines from the right side(add size * 0.5 to the x coordinate */
3465 shader_addline(buffer, "ADD texcrd2.x, texcrd2.y, yv12_coef.y;\n"); /* To avoid 0.5 == 0.5 comparisons */
3466 shader_addline(buffer, "FRC texcrd2.x, texcrd2.x;\n");
3467 shader_addline(buffer, "SGE texcrd2.x, texcrd2.x, coef.y;\n");
3468 shader_addline(buffer, "MAD texcrd.x, texcrd2.x, coef.y, texcrd.x;\n");
3469
3470 /* clamp, keep the half pixel origin in mind */
3471 shader_addline(buffer, "MAD temp.y, coef.y, chroma.w, yv12_coef.x;\n");
3472 shader_addline(buffer, "MAX texcrd.y, temp.y, texcrd.y;\n");
3473 shader_addline(buffer, "MAD temp.y, -coef.y, chroma.w, yv12_coef.z;\n");
3474 shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3475 } else {
3476 /* Read from [size - size+size/4] */
3477 shader_addline(buffer, "FLR texcrd.y, texcrd.y;\n");
3478 shader_addline(buffer, "MAD texcrd.y, texcrd.y, coef.w, size.y;\n");
3479
3480 /* Read odd lines from the right side(add size * 0.5 to the x coordinate */
3481 shader_addline(buffer, "ADD texcrd2.x, texcrd.y, yv12_coef.y;\n"); /* To avoid 0.5 == 0.5 comparisons */
3482 shader_addline(buffer, "FRC texcrd2.x, texcrd2.x;\n");
3483 shader_addline(buffer, "SGE texcrd2.x, texcrd2.x, coef.y;\n");
3484 shader_addline(buffer, "MUL texcrd2.x, texcrd2.x, size.x;\n");
3485 shader_addline(buffer, "MAD texcrd.x, texcrd2.x, coef.y, texcrd.x;\n");
3486
3487 /* Make sure to read exactly from the pixel center */
3488 shader_addline(buffer, "FLR texcrd.y, texcrd.y;\n");
3489 shader_addline(buffer, "ADD texcrd.y, texcrd.y, coef.y;\n");
3490
3491 /* Clamp */
3492 shader_addline(buffer, "MAD temp.y, size.y, coef.w, size.y;\n");
3493 shader_addline(buffer, "ADD temp.y, temp.y, -coef.y;\n");
3494 shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3495 shader_addline(buffer, "ADD temp.y, size.y, -coef.y;\n");
3496 shader_addline(buffer, "MAX texcrd.y, temp.y, texcrd.y;\n");
3497 }
3498 /* Read the texture, put the result into the output register */
3499 shader_addline(buffer, "TEX temp, texcrd, texture[0], %s;\n", tex);
3500 shader_addline(buffer, "MOV chroma.x, temp.w;\n");
3501
3502 /* The other chroma value is 1/6th of the texture lower, from 5/6th to 6/6th
3503 * No need to clamp because we're just reusing the already clamped value from above
3504 */
3505 if(textype == GL_TEXTURE_2D) {
3506 shader_addline(buffer, "ADD texcrd.y, texcrd.y, yv12_coef.y;\n");
3507 } else {
3508 shader_addline(buffer, "MAD texcrd.y, size.y, coef.w, texcrd.y;\n");
3509 }
3510 shader_addline(buffer, "TEX temp, texcrd, texture[0], %s;\n", tex);
3511 shader_addline(buffer, "MOV chroma.y, temp.w;\n");
3512
3513 /* Sample the luminance value. It is in the top 2/3rd of the texture, so scale the y coordinate.
3514 * Clamp the y coordinate to prevent the chroma values from bleeding into the sampled luminance
3515 * values due to filtering
3516 */
3517 shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3518 if(textype == GL_TEXTURE_2D) {
3519 /* Multiply the y coordinate by 2/3 and clamp it */
3520 shader_addline(buffer, "MUL texcrd.y, texcrd.y, yv12_coef.x;\n");
3521 shader_addline(buffer, "MAD temp.y, -coef.y, chroma.w, yv12_coef.x;\n");
3522 shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3523 shader_addline(buffer, "TEX luminance, texcrd, texture[0], %s;\n", tex);
3524 } else {
3525 /* Reading from texture_rectangles is pretty straightforward, just use the unmodified
3526 * texture coordinate. It is still a good idea to clamp it though, since the opengl texture
3527 * is bigger
3528 */
3529 shader_addline(buffer, "ADD temp.x, size.y, -coef.y;\n");
3530 shader_addline(buffer, "MIN texcrd.y, texcrd.y, size.x;\n");
3531 shader_addline(buffer, "TEX luminance, texcrd, texture[0], %s;\n", tex);
3532 }
3533 *luminance = 'a';
3534
3535 return TRUE;
3536 }
3537
3538 static GLuint gen_yuv_shader(IWineD3DDeviceImpl *device, enum yuv_fixup yuv_fixup, GLenum textype)
3539 {
3540 GLenum shader;
3541 SHADER_BUFFER buffer;
3542 char luminance_component;
3543 struct arbfp_blit_priv *priv = device->blit_priv;
3544
3545 /* Shader header */
3546 shader_buffer_init(&buffer);
3547
3548 ENTER_GL();
3549 GL_EXTCALL(glGenProgramsARB(1, &shader));
3550 checkGLcall("GL_EXTCALL(glGenProgramsARB(1, &shader))");
3551 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader));
3552 checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader)");
3553 LEAVE_GL();
3554 if(!shader) {
3555 shader_buffer_free(&buffer);
3556 return 0;
3557 }
3558
3559 /* The YUY2 and UYVY formats contain two pixels packed into a 32 bit macropixel,
3560 * giving effectively 16 bit per pixel. The color consists of a luminance(Y) and
3561 * two chroma(U and V) values. Each macropixel has two luminance values, one for
3562 * each single pixel it contains, and one U and one V value shared between both
3563 * pixels.
3564 *
3565 * The data is loaded into an A8L8 texture. With YUY2, the luminance component
3566 * contains the luminance and alpha the chroma. With UYVY it is vice versa. Thus
3567 * take the format into account when generating the read swizzles
3568 *
3569 * Reading the Y value is straightforward - just sample the texture. The hardware
3570 * takes care of filtering in the horizontal and vertical direction.
3571 *
3572 * Reading the U and V values is harder. We have to avoid filtering horizontally,
3573 * because that would mix the U and V values of one pixel or two adjacent pixels.
3574 * Thus floor the texture coordinate and add 0.5 to get an unfiltered read,
3575 * regardless of the filtering setting. Vertical filtering works automatically
3576 * though - the U and V values of two rows are mixed nicely.
3577 *
3578 * Appart of avoiding filtering issues, the code has to know which value it just
3579 * read, and where it can find the other one. To determine this, it checks if
3580 * it sampled an even or odd pixel, and shifts the 2nd read accordingly.
3581 *
3582 * Handling horizontal filtering of U and V values requires reading a 2nd pair
3583 * of pixels, extracting U and V and mixing them. This is not implemented yet.
3584 *
3585 * An alternative implementation idea is to load the texture as A8R8G8B8 texture,
3586 * with width / 2. This way one read gives all 3 values, finding U and V is easy
3587 * in an unfiltered situation. Finding the luminance on the other hand requires
3588 * finding out if it is an odd or even pixel. The real drawback of this approach
3589 * is filtering. This would have to be emulated completely in the shader, reading
3590 * up two 2 packed pixels in up to 2 rows and interpolating both horizontally and
3591 * vertically. Beyond that it would require adjustments to the texture handling
3592 * code to deal with the width scaling
3593 */
3594 shader_addline(&buffer, "!!ARBfp1.0\n");
3595 shader_addline(&buffer, "TEMP luminance;\n");
3596 shader_addline(&buffer, "TEMP temp;\n");
3597 shader_addline(&buffer, "TEMP chroma;\n");
3598 shader_addline(&buffer, "TEMP texcrd;\n");
3599 shader_addline(&buffer, "TEMP texcrd2;\n");
3600 shader_addline(&buffer, "PARAM coef = {1.0, 0.5, 2.0, 0.25};\n");
3601 shader_addline(&buffer, "PARAM yuv_coef = {1.403, 0.344, 0.714, 1.770};\n");
3602 shader_addline(&buffer, "PARAM size = program.local[0];\n");
3603
3604 switch (yuv_fixup)
3605 {
3606 case YUV_FIXUP_UYVY:
3607 case YUV_FIXUP_YUY2:
3608 if (!gen_planar_yuv_read(&buffer, yuv_fixup, textype, &luminance_component))
3609 {
3610 shader_buffer_free(&buffer);
3611 return 0;
3612 }
3613 break;
3614
3615 case YUV_FIXUP_YV12:
3616 if (!gen_yv12_read(&buffer, textype, &luminance_component))
3617 {
3618 shader_buffer_free(&buffer);
3619 return 0;
3620 }
3621 break;
3622
3623 default:
3624 FIXME("Unsupported YUV fixup %#x\n", yuv_fixup);
3625 shader_buffer_free(&buffer);
3626 return 0;
3627 }
3628
3629 /* Calculate the final result. Formula is taken from
3630 * http://www.fourcc.org/fccyvrgb.php. Note that the chroma
3631 * ranges from -0.5 to 0.5
3632 */
3633 shader_addline(&buffer, "SUB chroma.xy, chroma, coef.y;\n");
3634
3635 shader_addline(&buffer, "MAD result.color.x, chroma.x, yuv_coef.x, luminance.%c;\n", luminance_component);
3636 shader_addline(&buffer, "MAD temp.x, -chroma.y, yuv_coef.y, luminance.%c;\n", luminance_component);
3637 shader_addline(&buffer, "MAD result.color.y, -chroma.x, yuv_coef.z, temp.x;\n");
3638 shader_addline(&buffer, "MAD result.color.z, chroma.y, yuv_coef.w, luminance.%c;\n", luminance_component);
3639 shader_addline(&buffer, "END\n");
3640
3641 ENTER_GL();
3642 GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(buffer.buffer), buffer.buffer));
3643
3644 if (glGetError() == GL_INVALID_OPERATION) {
3645 GLint pos;
3646 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
3647 FIXME("Fragment program error at position %d: %s\n", pos,
3648 debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
3649 }
3650 shader_buffer_free(&buffer);
3651 LEAVE_GL();
3652
3653 switch (yuv_fixup)
3654 {
3655 case YUV_FIXUP_YUY2:
3656 if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->yuy2_rect_shader = shader;
3657 else priv->yuy2_2d_shader = shader;
3658 break;
3659
3660 case YUV_FIXUP_UYVY:
3661 if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->uyvy_rect_shader = shader;
3662 else priv->uyvy_2d_shader = shader;
3663 break;
3664
3665 case YUV_FIXUP_YV12:
3666 if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->yv12_rect_shader = shader;
3667 else priv->yv12_2d_shader = shader;
3668 break;
3669 }
3670
3671 return shader;
3672 }
3673
3674 static HRESULT arbfp_blit_set(IWineD3DDevice *iface, const struct GlPixelFormatDesc *format_desc,
3675 GLenum textype, UINT width, UINT height)
3676 {
3677 GLenum shader;
3678 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3679 float size[4] = {width, height, 1, 1};
3680 struct arbfp_blit_priv *priv = device->blit_priv;
3681 enum yuv_fixup yuv_fixup;
3682
3683 if (!is_yuv_fixup(format_desc->color_fixup))
3684 {
3685 TRACE("Fixup:\n");
3686 dump_color_fixup_desc(format_desc->color_fixup);
3687 /* Don't bother setting up a shader for unconverted formats */
3688 ENTER_GL();
3689 glEnable(textype);
3690 checkGLcall("glEnable(textype)");
3691 LEAVE_GL();
3692 return WINED3D_OK;
3693 }
3694
3695 yuv_fixup = get_yuv_fixup(format_desc->color_fixup);
3696
3697 switch(yuv_fixup)
3698 {
3699 case YUV_FIXUP_YUY2:
3700 shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->yuy2_rect_shader : priv->yuy2_2d_shader;
3701 break;
3702
3703 case YUV_FIXUP_UYVY:
3704 shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->uyvy_rect_shader : priv->uyvy_2d_shader;
3705 break;
3706
3707 case YUV_FIXUP_YV12:
3708 shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->yv12_rect_shader : priv->yv12_2d_shader;
3709 break;
3710
3711 default:
3712 FIXME("Unsupported YUV fixup %#x, not setting a shader\n", yuv_fixup);
3713 ENTER_GL();
3714 glEnable(textype);
3715 checkGLcall("glEnable(textype)");
3716 LEAVE_GL();
3717 return E_NOTIMPL;
3718 }
3719
3720 if (!shader) shader = gen_yuv_shader(device, yuv_fixup, textype);
3721
3722 ENTER_GL();
3723 glEnable(GL_FRAGMENT_PROGRAM_ARB);
3724 checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB)");
3725 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader));
3726 checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader)");
3727 GL_EXTCALL(glProgramLocalParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, 0, size));
3728 checkGLcall("glProgramLocalParameter4fvARB");
3729 LEAVE_GL();
3730
3731 return WINED3D_OK;
3732 }
3733
3734 static void arbfp_blit_unset(IWineD3DDevice *iface) {
3735 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3736
3737 ENTER_GL();
3738 glDisable(GL_FRAGMENT_PROGRAM_ARB);
3739 checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
3740 glDisable(GL_TEXTURE_2D);
3741 checkGLcall("glDisable(GL_TEXTURE_2D)");
3742 if(GL_SUPPORT(ARB_TEXTURE_CUBE_MAP)) {
3743 glDisable(GL_TEXTURE_CUBE_MAP_ARB);
3744 checkGLcall("glDisable(GL_TEXTURE_CUBE_MAP_ARB)");
3745 }
3746 if(GL_SUPPORT(ARB_TEXTURE_RECTANGLE)) {
3747 glDisable(GL_TEXTURE_RECTANGLE_ARB);
3748 checkGLcall("glDisable(GL_TEXTURE_RECTANGLE_ARB)");
3749 }
3750 LEAVE_GL();
3751 }
3752
3753 static BOOL arbfp_blit_color_fixup_supported(struct color_fixup_desc fixup)
3754 {
3755 enum yuv_fixup yuv_fixup;
3756
3757 if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
3758 {
3759 TRACE("Checking support for fixup:\n");
3760 dump_color_fixup_desc(fixup);
3761 }
3762
3763 if (is_identity_fixup(fixup))
3764 {
3765 TRACE("[OK]\n");
3766 return TRUE;
3767 }
3768
3769 /* We only support YUV conversions. */
3770 if (!is_yuv_fixup(fixup))
3771 {
3772 TRACE("[FAILED]\n");
3773 return FALSE;
3774 }
3775
3776 yuv_fixup = get_yuv_fixup(fixup);
3777 switch(yuv_fixup)
3778 {
3779 case YUV_FIXUP_YUY2:
3780 case YUV_FIXUP_UYVY:
3781 case YUV_FIXUP_YV12:
3782 TRACE("[OK]\n");
3783 return TRUE;
3784
3785 default:
3786 FIXME("Unsupported YUV fixup %#x\n", yuv_fixup);
3787 TRACE("[FAILED]\n");
3788 return FALSE;
3789 }
3790 }
3791
3792 const struct blit_shader arbfp_blit = {
3793 arbfp_blit_alloc,
3794 arbfp_blit_free,
3795 arbfp_blit_set,
3796 arbfp_blit_unset,
3797 arbfp_blit_color_fixup_supported,
3798 };
3799
3800 #undef GLINFO_LOCATION
useful hacks and other patches not (yet) suitable for mainline