?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
push 8724397e7ede0f147b6e05994a72142d44d4fecc
[wine/hacks.git] / dlls / wined3d / arb_program_shader.c
blobb05b541d5a4c2087c10763b3e23eae4786b1f8f3
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 != -1) {
216 /* WINED3DTSS_BUMPENVLSCALE and WINED3DTSS_BUMPENVLOFFSET are next to each other.
217 * point gl to the scale, and load 4 floats. x = scale, y = offset, z and w are junk, we
218 * don't care about them. The pointers are valid for sure because the stateblock is bigger.
219 * (they're WINED3DTSS_TEXTURETRANSFORMFLAGS and WINED3DTSS_ADDRESSW, so most likely 0 or NaN
220 */
221 const float *scale = (const float *)&stateBlock->textureState[(int) psi->luminanceconst[i].texunit][WINED3DTSS_BUMPENVLSCALE];
222 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, psi->luminanceconst[i].const_num, scale));
223 deviceImpl->activeContext->pshader_const_dirty[psi->luminanceconst[i].const_num] = 1;
224 }
225 }
226 }
227 }
228
229 static void shader_arb_update_float_vertex_constants(IWineD3DDevice *iface, UINT start, UINT count)
230 {
231 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
232
233 /* We don't want shader constant dirtification to be an O(contexts), so just dirtify the active
234 * context. On a context switch the old context will be fully dirtified */
235 memset(This->activeContext->vshader_const_dirty + start, 1,
236 sizeof(*This->activeContext->vshader_const_dirty) * count);
237 This->highest_dirty_vs_const = max(This->highest_dirty_vs_const, start + count + 1);
238 }
239
240 static void shader_arb_update_float_pixel_constants(IWineD3DDevice *iface, UINT start, UINT count)
241 {
242 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
243
244 /* We don't want shader constant dirtification to be an O(contexts), so just dirtify the active
245 * context. On a context switch the old context will be fully dirtified */
246 memset(This->activeContext->pshader_const_dirty + start, 1,
247 sizeof(*This->activeContext->pshader_const_dirty) * count);
248 This->highest_dirty_ps_const = max(This->highest_dirty_ps_const, start + count + 1);
249 }
250
251 /* Generate the variable & register declarations for the ARB_vertex_program output target */
252 static void shader_generate_arb_declarations(IWineD3DBaseShader *iface, const shader_reg_maps *reg_maps,
253 SHADER_BUFFER *buffer, const WineD3D_GL_Info *gl_info)
254 {
255 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
256 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device;
257 DWORD i, cur;
258 char pshader = shader_is_pshader_version(reg_maps->shader_version);
259 unsigned max_constantsF = min(This->baseShader.limits.constant_float,
260 (pshader ? GL_LIMITS(pshader_constantsF) : GL_LIMITS(vshader_constantsF)));
261 UINT extra_constants_needed = 0;
262 const local_constant *lconst;
263
264 /* Temporary Output register */
265 shader_addline(buffer, "TEMP TMP_OUT;\n");
266
267 for(i = 0; i < This->baseShader.limits.temporary; i++) {
268 if (reg_maps->temporary[i])
269 shader_addline(buffer, "TEMP R%u;\n", i);
270 }
271
272 for (i = 0; i < This->baseShader.limits.address; i++) {
273 if (reg_maps->address[i])
274 shader_addline(buffer, "ADDRESS A%d;\n", i);
275 }
276
277 for(i = 0; i < This->baseShader.limits.texcoord; i++) {
278 if (reg_maps->texcoord[i])
279 shader_addline(buffer,"TEMP T%u;\n", i);
280 }
281
282 /* Texture coordinate registers must be pre-loaded */
283 for (i = 0; i < This->baseShader.limits.texcoord; i++) {
284 if (reg_maps->texcoord[i])
285 shader_addline(buffer, "MOV T%u, fragment.texcoord[%u];\n", i, i);
286 }
287
288 for(i = 0; i < (sizeof(reg_maps->bumpmat) / sizeof(reg_maps->bumpmat[0])); i++) {
289 IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *) This;
290 if(!reg_maps->bumpmat[i]) continue;
291
292 cur = ps->numbumpenvmatconsts;
293 ps->bumpenvmatconst[cur].const_num = -1;
294 ps->bumpenvmatconst[cur].texunit = i;
295 ps->luminanceconst[cur].const_num = -1;
296 ps->luminanceconst[cur].texunit = i;
297
298 /* If the shader does not use all available constants, use the next free constant to load the bump mapping environment matrix from
299 * the stateblock into the shader. If no constant is available don't load, texbem will then just sample the texture without applying
300 * bump mapping.
301 */
302 if(max_constantsF + extra_constants_needed < GL_LIMITS(pshader_constantsF)) {
303 ps->bumpenvmatconst[cur].const_num = max_constantsF + extra_constants_needed;
304 shader_addline(buffer, "PARAM bumpenvmat%d = program.env[%d];\n",
305 i, ps->bumpenvmatconst[cur].const_num);
306 extra_constants_needed++;
307
308 if(reg_maps->luminanceparams && max_constantsF + extra_constants_needed < GL_LIMITS(pshader_constantsF)) {
309 ((IWineD3DPixelShaderImpl *)This)->luminanceconst[cur].const_num = max_constantsF + extra_constants_needed;
310 shader_addline(buffer, "PARAM luminance%d = program.env[%d];\n",
311 i, ps->luminanceconst[cur].const_num);
312 extra_constants_needed++;
313 } else if(reg_maps->luminanceparams) {
314 FIXME("No free constant to load the luminance parameters\n");
315 }
316 } else {
317 FIXME("No free constant found to load environment bump mapping matrix into the shader. texbem instruction will not apply bump mapping\n");
318 }
319
320 ps->numbumpenvmatconsts = cur + 1;
321 }
322
323 if(device->stateBlock->renderState[WINED3DRS_SRGBWRITEENABLE] && pshader) {
324 shader_addline(buffer, "PARAM srgb_mul_low = {%f, %f, %f, 1.0};\n",
325 srgb_mul_low, srgb_mul_low, srgb_mul_low);
326 shader_addline(buffer, "PARAM srgb_comparison = {%f, %f, %f, %f};\n",
327 srgb_cmp, srgb_cmp, srgb_cmp, srgb_cmp);
328 shader_addline(buffer, "PARAM srgb_pow = {%f, %f, %f, 1.0};\n",
329 srgb_pow, srgb_pow, srgb_pow);
330 shader_addline(buffer, "PARAM srgb_mul_hi = {%f, %f, %f, 1.0};\n",
331 srgb_mul_high, srgb_mul_high, srgb_mul_high);
332 shader_addline(buffer, "PARAM srgb_sub_hi = {%f, %f, %f, 0.0};\n",
333 srgb_sub_high, srgb_sub_high, srgb_sub_high);
334 }
335
336 /* Load local constants using the program-local space,
337 * this avoids reloading them each time the shader is used
338 */
339 if(!This->baseShader.load_local_constsF) {
340 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
341 shader_addline(buffer, "PARAM C%u = program.local[%u];\n", lconst->idx,
342 lconst->idx);
343 }
344 }
345
346 /* we use the array-based constants array if the local constants are marked for loading,
347 * because then we use indirect addressing, or when the local constant list is empty,
348 * because then we don't know if we're using indirect addressing or not. If we're hardcoding
349 * local constants do not declare the loaded constants as an array because ARB compilers usually
350 * do not optimize unused constants away
351 */
352 if(This->baseShader.load_local_constsF || list_empty(&This->baseShader.constantsF)) {
353 /* Need to PARAM the environment parameters (constants) so we can use relative addressing */
354 shader_addline(buffer, "PARAM C[%d] = { program.env[0..%d] };\n",
355 max_constantsF, max_constantsF - 1);
356 } else {
357 for(i = 0; i < max_constantsF; i++) {
358 if(!shader_constant_is_local(This, i)) {
359 shader_addline(buffer, "PARAM C%d = program.env[%d];\n",i, i);
360 }
361 }
362 }
363 }
364
365 static const char * const shift_tab[] = {
366 "dummy", /* 0 (none) */
367 "coefmul.x", /* 1 (x2) */
368 "coefmul.y", /* 2 (x4) */
369 "coefmul.z", /* 3 (x8) */
370 "coefmul.w", /* 4 (x16) */
371 "dummy", /* 5 (x32) */
372 "dummy", /* 6 (x64) */
373 "dummy", /* 7 (x128) */
374 "dummy", /* 8 (d256) */
375 "dummy", /* 9 (d128) */
376 "dummy", /* 10 (d64) */
377 "dummy", /* 11 (d32) */
378 "coefdiv.w", /* 12 (d16) */
379 "coefdiv.z", /* 13 (d8) */
380 "coefdiv.y", /* 14 (d4) */
381 "coefdiv.x" /* 15 (d2) */
382 };
383
384 static void shader_arb_get_write_mask(const SHADER_OPCODE_ARG *arg, const DWORD param, char *write_mask)
385 {
386 char *ptr = write_mask;
387 char vshader = shader_is_vshader_version(arg->reg_maps->shader_version);
388
389 if(vshader && shader_get_regtype(param) == WINED3DSPR_ADDR) {
390 *ptr++ = '.';
391 *ptr++ = 'x';
392 } else if ((param & WINED3DSP_WRITEMASK_ALL) != WINED3DSP_WRITEMASK_ALL) {
393 *ptr++ = '.';
394 if (param & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
395 if (param & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
396 if (param & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
397 if (param & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
398 }
399
400 *ptr = '\0';
401 }
402
403 static void shader_arb_get_swizzle(const DWORD param, BOOL fixup, char *swizzle_str) {
404 /* For registers of type WINED3DDECLTYPE_D3DCOLOR, data is stored as "bgra",
405 * but addressed as "rgba". To fix this we need to swap the register's x
406 * and z components. */
407 const char *swizzle_chars = fixup ? "zyxw" : "xyzw";
408 char *ptr = swizzle_str;
409
410 /* swizzle bits fields: wwzzyyxx */
411 DWORD swizzle = (param & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
412 DWORD swizzle_x = swizzle & 0x03;
413 DWORD swizzle_y = (swizzle >> 2) & 0x03;
414 DWORD swizzle_z = (swizzle >> 4) & 0x03;
415 DWORD swizzle_w = (swizzle >> 6) & 0x03;
416
417 /* If the swizzle is the default swizzle (ie, "xyzw"), we don't need to
418 * generate a swizzle string. Unless we need to our own swizzling. */
419 if ((WINED3DSP_NOSWIZZLE >> WINED3DSP_SWIZZLE_SHIFT) != swizzle || fixup) {
420 *ptr++ = '.';
421 if (swizzle_x == swizzle_y && swizzle_x == swizzle_z && swizzle_x == swizzle_w) {
422 *ptr++ = swizzle_chars[swizzle_x];
423 } else {
424 *ptr++ = swizzle_chars[swizzle_x];
425 *ptr++ = swizzle_chars[swizzle_y];
426 *ptr++ = swizzle_chars[swizzle_z];
427 *ptr++ = swizzle_chars[swizzle_w];
428 }
429 }
430
431 *ptr = '\0';
432 }
433
434 static void pshader_get_register_name(IWineD3DBaseShader* iface,
435 const DWORD param, char* regstr) {
436
437 DWORD reg = param & WINED3DSP_REGNUM_MASK;
438 DWORD regtype = shader_get_regtype(param);
439 IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *) iface;
440
441 switch (regtype) {
442 case WINED3DSPR_TEMP:
443 sprintf(regstr, "R%u", reg);
444 break;
445 case WINED3DSPR_INPUT:
446 if (reg==0) {
447 strcpy(regstr, "fragment.color.primary");
448 } else {
449 strcpy(regstr, "fragment.color.secondary");
450 }
451 break;
452 case WINED3DSPR_CONST:
453 if(This->baseShader.load_local_constsF || list_empty(&This->baseShader.constantsF)) {
454 sprintf(regstr, "C[%u]", reg);
455 } else {
456 sprintf(regstr, "C%u", reg);
457 }
458 break;
459 case WINED3DSPR_TEXTURE: /* case WINED3DSPR_ADDR: */
460 sprintf(regstr,"T%u", reg);
461 break;
462 case WINED3DSPR_COLOROUT:
463 if (reg == 0)
464 sprintf(regstr, "TMP_COLOR");
465 else {
466 /* TODO: See GL_ARB_draw_buffers */
467 FIXME("Unsupported write to render target %u\n", reg);
468 sprintf(regstr, "unsupported_register");
469 }
470 break;
471 case WINED3DSPR_DEPTHOUT:
472 sprintf(regstr, "result.depth");
473 break;
474 case WINED3DSPR_ATTROUT:
475 sprintf(regstr, "oD[%u]", reg);
476 break;
477 case WINED3DSPR_TEXCRDOUT:
478 sprintf(regstr, "oT[%u]", reg);
479 break;
480 default:
481 FIXME("Unhandled register name Type(%d)\n", regtype);
482 sprintf(regstr, "unrecognized_register");
483 break;
484 }
485 }
486
487 /* TODO: merge with pixel shader */
488 static void vshader_program_add_param(const SHADER_OPCODE_ARG *arg, const DWORD param, BOOL is_input, char *hwLine)
489 {
490 IWineD3DVertexShaderImpl* This = (IWineD3DVertexShaderImpl*) arg->shader;
491
492 /* oPos, oFog and oPts in D3D */
493 static const char * const hwrastout_reg_names[] = { "TMP_OUT", "result.fogcoord", "result.pointsize" };
494
495 DWORD reg = param & WINED3DSP_REGNUM_MASK;
496 DWORD regtype = shader_get_regtype(param);
497 char tmpReg[255];
498 BOOL is_color = FALSE;
499
500 if ((param & WINED3DSP_SRCMOD_MASK) == WINED3DSPSM_NEG) {
501 strcat(hwLine, " -");
502 } else {
503 strcat(hwLine, " ");
504 }
505
506 switch (regtype) {
507 case WINED3DSPR_TEMP:
508 sprintf(tmpReg, "R%u", reg);
509 strcat(hwLine, tmpReg);
510 break;
511 case WINED3DSPR_INPUT:
512
513 if (This->cur_args->swizzle_map & (1 << reg)) is_color = TRUE;
514
515 sprintf(tmpReg, "vertex.attrib[%u]", reg);
516 strcat(hwLine, tmpReg);
517 break;
518 case WINED3DSPR_CONST:
519 if(param & WINED3DSHADER_ADDRMODE_RELATIVE) {
520 if(reg >= This->rel_offset) {
521 sprintf(tmpReg, "C[A0.x + %u]", reg - This->rel_offset);
522 } else {
523 sprintf(tmpReg, "C[A0.x - %u]", -reg + This->rel_offset);
524 }
525 } else {
526 if(This->baseShader.load_local_constsF || list_empty(&This->baseShader.constantsF)) {
527 sprintf(tmpReg, "C[%u]", reg);
528 } else {
529 sprintf(tmpReg, "C%u", reg);
530 }
531 }
532 strcat(hwLine, tmpReg);
533 break;
534 case WINED3DSPR_ADDR: /*case D3DSPR_TEXTURE:*/
535 sprintf(tmpReg, "A%u", reg);
536 strcat(hwLine, tmpReg);
537 break;
538 case WINED3DSPR_RASTOUT:
539 sprintf(tmpReg, "%s", hwrastout_reg_names[reg]);
540 strcat(hwLine, tmpReg);
541 break;
542 case WINED3DSPR_ATTROUT:
543 if (reg==0) {
544 strcat(hwLine, "result.color.primary");
545 } else {
546 strcat(hwLine, "result.color.secondary");
547 }
548 break;
549 case WINED3DSPR_TEXCRDOUT:
550 sprintf(tmpReg, "result.texcoord[%u]", reg);
551 strcat(hwLine, tmpReg);
552 break;
553 default:
554 FIXME("Unknown reg type %d %d\n", regtype, reg);
555 strcat(hwLine, "unrecognized_register");
556 break;
557 }
558
559 if (!is_input) {
560 char write_mask[6];
561 shader_arb_get_write_mask(arg, param, write_mask);
562 strcat(hwLine, write_mask);
563 } else {
564 char swizzle[6];
565 shader_arb_get_swizzle(param, is_color, swizzle);
566 strcat(hwLine, swizzle);
567 }
568 }
569
570 static const char *shader_arb_get_fixup_swizzle(enum fixup_channel_source channel_source)
571 {
572 switch(channel_source)
573 {
574 case CHANNEL_SOURCE_ZERO: return "0";
575 case CHANNEL_SOURCE_ONE: return "1";
576 case CHANNEL_SOURCE_X: return "x";
577 case CHANNEL_SOURCE_Y: return "y";
578 case CHANNEL_SOURCE_Z: return "z";
579 case CHANNEL_SOURCE_W: return "w";
580 default:
581 FIXME("Unhandled channel source %#x\n", channel_source);
582 return "undefined";
583 }
584 }
585
586 static void gen_color_correction(SHADER_BUFFER *buffer, const char *reg, DWORD dst_mask,
587 const char *one, const char *two, struct color_fixup_desc fixup)
588 {
589 DWORD mask;
590
591 if (is_yuv_fixup(fixup))
592 {
593 enum yuv_fixup yuv_fixup = get_yuv_fixup(fixup);
594 FIXME("YUV fixup (%#x) not supported\n", yuv_fixup);
595 return;
596 }
597
598 mask = 0;
599 if (fixup.x_source != CHANNEL_SOURCE_X) mask |= WINED3DSP_WRITEMASK_0;
600 if (fixup.y_source != CHANNEL_SOURCE_Y) mask |= WINED3DSP_WRITEMASK_1;
601 if (fixup.z_source != CHANNEL_SOURCE_Z) mask |= WINED3DSP_WRITEMASK_2;
602 if (fixup.w_source != CHANNEL_SOURCE_W) mask |= WINED3DSP_WRITEMASK_3;
603 mask &= dst_mask;
604
605 if (mask)
606 {
607 shader_addline(buffer, "SWZ %s, %s, %s, %s, %s, %s;\n", reg, reg,
608 shader_arb_get_fixup_swizzle(fixup.x_source), shader_arb_get_fixup_swizzle(fixup.y_source),
609 shader_arb_get_fixup_swizzle(fixup.z_source), shader_arb_get_fixup_swizzle(fixup.w_source));
610 }
611
612 mask = 0;
613 if (fixup.x_sign_fixup) mask |= WINED3DSP_WRITEMASK_0;
614 if (fixup.y_sign_fixup) mask |= WINED3DSP_WRITEMASK_1;
615 if (fixup.z_sign_fixup) mask |= WINED3DSP_WRITEMASK_2;
616 if (fixup.w_sign_fixup) mask |= WINED3DSP_WRITEMASK_3;
617 mask &= dst_mask;
618
619 if (mask)
620 {
621 char reg_mask[6];
622 char *ptr = reg_mask;
623
624 if (mask != WINED3DSP_WRITEMASK_ALL)
625 {
626 *ptr++ = '.';
627 if (mask & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
628 if (mask & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
629 if (mask & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
630 if (mask & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
631 }
632 *ptr = '\0';
633
634 shader_addline(buffer, "MAD %s%s, %s, %s, -%s;\n", reg, reg_mask, reg, two, one);
635 }
636 }
637
638 static void shader_hw_sample(const SHADER_OPCODE_ARG *arg, DWORD sampler_idx, const char *dst_str,
639 const char *coord_reg, BOOL projected, BOOL bias)
640 {
641 SHADER_BUFFER* buffer = arg->buffer;
642 DWORD sampler_type = arg->reg_maps->samplers[sampler_idx] & WINED3DSP_TEXTURETYPE_MASK;
643 const char *tex_type;
644 IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *) arg->shader;
645 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device;
646
647 switch(sampler_type) {
648 case WINED3DSTT_1D:
649 tex_type = "1D";
650 break;
651
652 case WINED3DSTT_2D:
653 if(device->stateBlock->textures[sampler_idx] &&
654 IWineD3DBaseTexture_GetTextureDimensions(device->stateBlock->textures[sampler_idx]) == GL_TEXTURE_RECTANGLE_ARB) {
655 tex_type = "RECT";
656 } else {
657 tex_type = "2D";
658 }
659 break;
660
661 case WINED3DSTT_VOLUME:
662 tex_type = "3D";
663 break;
664
665 case WINED3DSTT_CUBE:
666 tex_type = "CUBE";
667 break;
668
669 default:
670 ERR("Unexpected texture type %d\n", sampler_type);
671 tex_type = "";
672 }
673
674 if (bias) {
675 /* Shouldn't be possible, but let's check for it */
676 if(projected) FIXME("Biased and Projected texture sampling\n");
677 /* TXB takes the 4th component of the source vector automatically, as d3d. Nothing more to do */
678 shader_addline(buffer, "TXB %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
679 } else if (projected) {
680 shader_addline(buffer, "TXP %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
681 } else {
682 shader_addline(buffer, "TEX %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
683 }
684
685 if(shader_is_pshader_version(arg->reg_maps->shader_version)) {
686 IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *) arg->shader;
687 gen_color_correction(buffer, dst_str, arg->dst & WINED3DSP_WRITEMASK_ALL, "one", "coefmul.x",
688 ps->cur_args->color_fixup[sampler_idx]);
689 }
690 }
691
692 static void pshader_gen_input_modifier_line (
693 IWineD3DBaseShader *iface,
694 SHADER_BUFFER* buffer,
695 const DWORD instr,
696 int tmpreg,
697 char *outregstr) {
698
699 /* Generate a line that does the input modifier computation and return the input register to use */
700 char regstr[256];
701 char swzstr[20];
702 int insert_line;
703
704 /* Assume a new line will be added */
705 insert_line = 1;
706
707 /* Get register name */
708 pshader_get_register_name(iface, instr, regstr);
709 shader_arb_get_swizzle(instr, FALSE, swzstr);
710
711 switch (instr & WINED3DSP_SRCMOD_MASK) {
712 case WINED3DSPSM_NONE:
713 sprintf(outregstr, "%s%s", regstr, swzstr);
714 insert_line = 0;
715 break;
716 case WINED3DSPSM_NEG:
717 sprintf(outregstr, "-%s%s", regstr, swzstr);
718 insert_line = 0;
719 break;
720 case WINED3DSPSM_BIAS:
721 shader_addline(buffer, "ADD T%c, %s, -coefdiv.x;\n", 'A' + tmpreg, regstr);
722 break;
723 case WINED3DSPSM_BIASNEG:
724 shader_addline(buffer, "ADD T%c, -%s, coefdiv.x;\n", 'A' + tmpreg, regstr);
725 break;
726 case WINED3DSPSM_SIGN:
727 shader_addline(buffer, "MAD T%c, %s, coefmul.x, -one.x;\n", 'A' + tmpreg, regstr);
728 break;
729 case WINED3DSPSM_SIGNNEG:
730 shader_addline(buffer, "MAD T%c, %s, -coefmul.x, one.x;\n", 'A' + tmpreg, regstr);
731 break;
732 case WINED3DSPSM_COMP:
733 shader_addline(buffer, "SUB T%c, one.x, %s;\n", 'A' + tmpreg, regstr);
734 break;
735 case WINED3DSPSM_X2:
736 shader_addline(buffer, "ADD T%c, %s, %s;\n", 'A' + tmpreg, regstr, regstr);
737 break;
738 case WINED3DSPSM_X2NEG:
739 shader_addline(buffer, "ADD T%c, -%s, -%s;\n", 'A' + tmpreg, regstr, regstr);
740 break;
741 case WINED3DSPSM_DZ:
742 shader_addline(buffer, "RCP T%c, %s.z;\n", 'A' + tmpreg, regstr);
743 shader_addline(buffer, "MUL T%c, %s, T%c;\n", 'A' + tmpreg, regstr, 'A' + tmpreg);
744 break;
745 case WINED3DSPSM_DW:
746 shader_addline(buffer, "RCP T%c, %s.w;\n", 'A' + tmpreg, regstr);
747 shader_addline(buffer, "MUL T%c, %s, T%c;\n", 'A' + tmpreg, regstr, 'A' + tmpreg);
748 break;
749 default:
750 sprintf(outregstr, "%s%s", regstr, swzstr);
751 insert_line = 0;
752 }
753
754 /* Return modified or original register, with swizzle */
755 if (insert_line)
756 sprintf(outregstr, "T%c%s", 'A' + tmpreg, swzstr);
757 }
758
759 static inline void pshader_gen_output_modifier_line(SHADER_BUFFER *buffer, int saturate, const char *write_mask,
760 int shift, const char *regstr)
761 {
762 /* Generate a line that does the output modifier computation */
763 shader_addline(buffer, "MUL%s %s%s, %s, %s;\n", saturate ? "_SAT" : "",
764 regstr, write_mask, regstr, shift_tab[shift]);
765 }
766
767 static void pshader_hw_bem(const SHADER_OPCODE_ARG *arg)
768 {
769 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
770
771 SHADER_BUFFER* buffer = arg->buffer;
772 char dst_name[50];
773 char src_name[2][50];
774 char dst_wmask[20];
775 DWORD sampler_code = arg->dst & WINED3DSP_REGNUM_MASK;
776 BOOL has_bumpmat = FALSE;
777 int i;
778
779 for(i = 0; i < This->numbumpenvmatconsts; i++) {
780 if(This->bumpenvmatconst[i].const_num != -1 && This->bumpenvmatconst[i].texunit == sampler_code) {
781 has_bumpmat = TRUE;
782 break;
783 }
784 }
785
786 pshader_get_register_name(arg->shader, arg->dst, dst_name);
787 shader_arb_get_write_mask(arg, arg->dst, dst_wmask);
788 strcat(dst_name, dst_wmask);
789
790 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src_name[0]);
791 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[1], 1, src_name[1]);
792
793 if(has_bumpmat) {
794 /* Sampling the perturbation map in Tsrc was done already, including the signedness correction if needed */
795 shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, x, z, 0, 0;\n", sampler_code);
796 shader_addline(buffer, "DP3 TMP.r, TMP2, %s;\n", src_name[1]);
797 shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, y, w, 0, 0;\n", sampler_code);
798 shader_addline(buffer, "DP3 TMP.g, TMP2, %s;\n", src_name[1]);
799
800 shader_addline(buffer, "ADD %s, %s, TMP;\n", dst_name, src_name[0]);
801 } else {
802 shader_addline(buffer, "MOV %s, %s;\n", dst_name, src_name[0]);
803 }
804 }
805
806 static void pshader_hw_cnd(const SHADER_OPCODE_ARG *arg)
807 {
808 SHADER_BUFFER* buffer = arg->buffer;
809 char dst_wmask[20];
810 char dst_name[50];
811 char src_name[3][50];
812 BOOL sat = (arg->dst & WINED3DSP_DSTMOD_MASK) & WINED3DSPDM_SATURATE;
813 DWORD shift = (arg->dst & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
814
815 /* FIXME: support output modifiers */
816
817 /* Handle output register */
818 pshader_get_register_name(arg->shader, arg->dst, dst_name);
819 shader_arb_get_write_mask(arg, arg->dst, dst_wmask);
820
821 /* Generate input register names (with modifiers) */
822 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src_name[0]);
823 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[1], 1, src_name[1]);
824 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[2], 2, src_name[2]);
825
826 /* The coissue flag changes the semantic of the cnd instruction in <= 1.3 shaders */
827 if (arg->reg_maps->shader_version <= WINED3DPS_VERSION(1, 3) && arg->opcode_token & WINED3DSI_COISSUE)
828 {
829 shader_addline(buffer, "MOV%s %s%s, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask, src_name[1]);
830 } else {
831 shader_addline(buffer, "ADD TMP, -%s, coefdiv.x;\n", src_name[0]);
832 shader_addline(buffer, "CMP%s %s%s, TMP, %s, %s;\n",
833 sat ? "_SAT" : "", dst_name, dst_wmask, src_name[1], src_name[2]);
834 }
835 if (shift != 0)
836 pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
837 }
838
839 static void pshader_hw_cmp(const SHADER_OPCODE_ARG *arg)
840 {
841 SHADER_BUFFER* buffer = arg->buffer;
842 char dst_wmask[20];
843 char dst_name[50];
844 char src_name[3][50];
845 DWORD shift = (arg->dst & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
846 BOOL sat = (arg->dst & WINED3DSP_DSTMOD_MASK) & WINED3DSPDM_SATURATE;
847
848 /* FIXME: support output modifiers */
849
850 /* Handle output register */
851 pshader_get_register_name(arg->shader, arg->dst, dst_name);
852 shader_arb_get_write_mask(arg, arg->dst, dst_wmask);
853
854 /* Generate input register names (with modifiers) */
855 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src_name[0]);
856 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[1], 1, src_name[1]);
857 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[2], 2, src_name[2]);
858
859 shader_addline(buffer, "CMP%s %s%s, %s, %s, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask,
860 src_name[0], src_name[2], src_name[1]);
861
862 if (shift != 0)
863 pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
864 }
865
866 /** Process the WINED3DSIO_DP2ADD instruction in ARB.
867 * dst = dot2(src0, src1) + src2 */
868 static void pshader_hw_dp2add(const SHADER_OPCODE_ARG *arg)
869 {
870 SHADER_BUFFER* buffer = arg->buffer;
871 char dst_wmask[20];
872 char dst_name[50];
873 char src_name[3][50];
874 DWORD shift = (arg->dst & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
875 BOOL sat = (arg->dst & WINED3DSP_DSTMOD_MASK) & WINED3DSPDM_SATURATE;
876
877 pshader_get_register_name(arg->shader, arg->dst, dst_name);
878 shader_arb_get_write_mask(arg, arg->dst, dst_wmask);
879
880 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src_name[0]);
881 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[1], 1, src_name[1]);
882 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[2], 2, src_name[2]);
883
884 /* Emulate a DP2 with a DP3 and 0.0 */
885 shader_addline(buffer, "MOV TMP, %s;\n", src_name[0]);
886 shader_addline(buffer, "MOV TMP.z, 0.0;\n");
887 shader_addline(buffer, "DP3 TMP2, TMP, %s;\n", src_name[1]);
888 shader_addline(buffer, "ADD%s %s%s, TMP2, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask, src_name[2]);
889
890 if (shift != 0)
891 pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
892 }
893
894 /* Map the opcode 1-to-1 to the GL code */
895 static void shader_hw_map2gl(const SHADER_OPCODE_ARG *arg)
896 {
897 CONST SHADER_OPCODE* curOpcode = arg->opcode;
898 SHADER_BUFFER* buffer = arg->buffer;
899 DWORD dst = arg->dst;
900 const DWORD *src = arg->src;
901 const char *instruction;
902 char arguments[256];
903 unsigned int i;
904
905 switch (curOpcode->opcode)
906 {
907 case WINED3DSIO_ABS: instruction = "ABS"; break;
908 case WINED3DSIO_ADD: instruction = "ADD"; break;
909 case WINED3DSIO_CRS: instruction = "XPD"; break;
910 case WINED3DSIO_DP3: instruction = "DP3"; break;
911 case WINED3DSIO_DP4: instruction = "DP4"; break;
912 case WINED3DSIO_DST: instruction = "DST"; break;
913 case WINED3DSIO_EXP: instruction = "EX2"; break;
914 case WINED3DSIO_EXPP: instruction = "EXP"; break;
915 case WINED3DSIO_FRC: instruction = "FRC"; break;
916 case WINED3DSIO_LIT: instruction = "LIT"; break;
917 case WINED3DSIO_LOG: instruction = "LG2"; break;
918 case WINED3DSIO_LOGP: instruction = "LOG"; break;
919 case WINED3DSIO_LRP: instruction = "LRP"; break;
920 case WINED3DSIO_MAD: instruction = "MAD"; break;
921 case WINED3DSIO_MAX: instruction = "MAX"; break;
922 case WINED3DSIO_MIN: instruction = "MIN"; break;
923 case WINED3DSIO_MOV: instruction = "MOV"; break;
924 case WINED3DSIO_MUL: instruction = "MUL"; break;
925 case WINED3DSIO_NOP: instruction = "NOP"; break;
926 case WINED3DSIO_POW: instruction = "POW"; break;
927 case WINED3DSIO_SGE: instruction = "SGE"; break;
928 case WINED3DSIO_SLT: instruction = "SLT"; break;
929 case WINED3DSIO_SUB: instruction = "SUB"; break;
930 default: instruction = "";
931 FIXME("Unhandled opcode %s\n", curOpcode->name);
932 break;
933 }
934
935 if (shader_is_pshader_version(arg->reg_maps->shader_version))
936 {
937 /* Output token related */
938 char output_rname[256];
939 char output_wmask[20];
940 char operands[4][100];
941 BOOL saturate = FALSE;
942 BOOL centroid = FALSE;
943 BOOL partialprecision = FALSE;
944 const char *modifier;
945 DWORD shift;
946
947 if (!curOpcode->num_params)
948 {
949 ERR("Opcode \"%s\" has no parameters\n", curOpcode->name);
950 return;
951 }
952
953 /* Process modifiers */
954 if (dst & WINED3DSP_DSTMOD_MASK)
955 {
956 DWORD mask = dst & WINED3DSP_DSTMOD_MASK;
957
958 saturate = mask & WINED3DSPDM_SATURATE;
959 centroid = mask & WINED3DSPDM_MSAMPCENTROID;
960 partialprecision = mask & WINED3DSPDM_PARTIALPRECISION;
961 mask &= ~(WINED3DSPDM_MSAMPCENTROID | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_SATURATE);
962 if (mask)
963 FIXME("Unrecognized modifier(%#x)\n", mask >> WINED3DSP_DSTMOD_SHIFT);
964
965 if (centroid)
966 FIXME("Unhandled modifier(%#x)\n", mask >> WINED3DSP_DSTMOD_SHIFT);
967 }
968 shift = (dst & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
969 modifier = (saturate && !shift) ? "_SAT" : "";
970
971 /* Generate input register names (with modifiers) */
972 for (i = 1; i < curOpcode->num_params; ++i)
973 pshader_gen_input_modifier_line(arg->shader, buffer, src[i-1], i-1, operands[i]);
974
975 /* Handle output register */
976 pshader_get_register_name(arg->shader, dst, output_rname);
977 strcpy(operands[0], output_rname);
978 shader_arb_get_write_mask(arg, dst, output_wmask);
979 strcat(operands[0], output_wmask);
980
981 arguments[0] = '\0';
982 strcat(arguments, operands[0]);
983 for (i = 1; i < curOpcode->num_params; i++)
984 {
985 strcat(arguments, ", ");
986 strcat(arguments, operands[i]);
987 }
988 shader_addline(buffer, "%s%s %s;\n", instruction, modifier, arguments);
989
990 /* A shift requires another line. */
991 if (shift) pshader_gen_output_modifier_line(buffer, saturate, output_wmask, shift, output_rname);
992 } else {
993 /* Note that vshader_program_add_param() adds spaces. */
994
995 arguments[0] = '\0';
996 if (curOpcode->num_params > 0)
997 {
998 vshader_program_add_param(arg, dst, FALSE, arguments);
999 for (i = 1; i < curOpcode->num_params; ++i)
1000 {
1001 strcat(arguments, ",");
1002 vshader_program_add_param(arg, src[i-1], TRUE, arguments);
1003 }
1004 }
1005 shader_addline(buffer, "%s%s;\n", instruction, arguments);
1006 }
1007 }
1008
1009 static void shader_hw_mov(const SHADER_OPCODE_ARG *arg)
1010 {
1011 IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl*)arg->shader;
1012
1013 if ((WINED3DSHADER_VERSION_MAJOR(arg->reg_maps->shader_version) == 1
1014 && !shader_is_pshader_version(arg->reg_maps->shader_version)
1015 && shader_get_regtype(arg->dst) == WINED3DSPR_ADDR)
1016 || arg->opcode->opcode == WINED3DSIO_MOVA)
1017 {
1018 SHADER_BUFFER *buffer = arg->buffer;
1019 char src0_param[256];
1020
1021 if (arg->opcode->opcode == WINED3DSIO_MOVA)
1022 FIXME("mova should round\n");
1023
1024 src0_param[0] = '\0';
1025 if (((IWineD3DVertexShaderImpl *)shader)->rel_offset)
1026 {
1027 vshader_program_add_param(arg, arg->src[0], TRUE, src0_param);
1028 shader_addline(buffer, "ADD TMP.x, %s, helper_const.z;\n", src0_param);
1029 shader_addline(buffer, "ARL A0.x, TMP.x;\n");
1030 }
1031 else
1032 {
1033 /* Apple's ARB_vertex_program implementation does not accept an ARL source argument
1034 * with more than one component. Thus replicate the first source argument over all
1035 * 4 components. For example, .xyzw -> .x (or better: .xxxx), .zwxy -> .z, etc)
1036 */
1037 DWORD parm = arg->src[0] & ~(WINED3DVS_SWIZZLE_MASK);
1038 if((arg->src[0] & WINED3DVS_X_W) == WINED3DVS_X_W)
1039 parm |= WINED3DVS_X_W | WINED3DVS_Y_W | WINED3DVS_Z_W | WINED3DVS_W_W;
1040 else if((arg->src[0] & WINED3DVS_X_Z) == WINED3DVS_X_Z)
1041 parm |= WINED3DVS_X_Z | WINED3DVS_Y_Z | WINED3DVS_Z_Z | WINED3DVS_W_Z;
1042 else if((arg->src[0] & WINED3DVS_X_Y) == WINED3DVS_X_Y)
1043 parm |= WINED3DVS_X_Y | WINED3DVS_Y_Y | WINED3DVS_Z_Y | WINED3DVS_W_Y;
1044 else if((arg->src[0] & WINED3DVS_X_X) == WINED3DVS_X_X)
1045 parm |= WINED3DVS_X_X | WINED3DVS_Y_X | WINED3DVS_Z_X | WINED3DVS_W_X;
1046 vshader_program_add_param(arg, parm, TRUE, src0_param);
1047 shader_addline(buffer, "ARL A0.x, %s;\n", src0_param);
1048 }
1049 }
1050 else
1051 {
1052 shader_hw_map2gl(arg);
1053 }
1054 }
1055
1056 static void pshader_hw_texkill(const SHADER_OPCODE_ARG *arg)
1057 {
1058 DWORD shader_version = arg->reg_maps->shader_version;
1059 SHADER_BUFFER* buffer = arg->buffer;
1060 char reg_dest[40];
1061
1062 /* No swizzles are allowed in d3d's texkill. PS 1.x ignores the 4th component as documented,
1063 * but >= 2.0 honors it(undocumented, but tested by the d3d9 testsuit)
1064 */
1065 pshader_get_register_name(arg->shader, arg->dst, reg_dest);
1066
1067 if (shader_version >= WINED3DPS_VERSION(2,0))
1068 {
1069 /* The arb backend doesn't claim ps 2.0 support, but try to eat what the app feeds to us */
1070 shader_addline(buffer, "KIL %s;\n", reg_dest);
1071 } else {
1072 /* ARB fp doesn't like swizzles on the parameter of the KIL instruction. To mask the 4th component,
1073 * copy the register into our general purpose TMP variable, overwrite .w and pass TMP to KIL
1074 */
1075 shader_addline(buffer, "MOV TMP, %s;\n", reg_dest);
1076 shader_addline(buffer, "MOV TMP.w, one.w;\n");
1077 shader_addline(buffer, "KIL TMP;\n");
1078 }
1079 }
1080
1081 static void pshader_hw_tex(const SHADER_OPCODE_ARG *arg)
1082 {
1083 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1084 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1085
1086 DWORD dst = arg->dst;
1087 const DWORD *src = arg->src;
1088 SHADER_BUFFER* buffer = arg->buffer;
1089 DWORD shader_version = arg->reg_maps->shader_version;
1090 BOOL projected = FALSE, bias = FALSE;
1091
1092 char reg_dest[40];
1093 char reg_coord[40];
1094 DWORD reg_dest_code;
1095 DWORD reg_sampler_code;
1096
1097 /* All versions have a destination register */
1098 reg_dest_code = dst & WINED3DSP_REGNUM_MASK;
1099 pshader_get_register_name(arg->shader, dst, reg_dest);
1100
1101 /* 1.0-1.3: Use destination register as coordinate source.
1102 1.4+: Use provided coordinate source register. */
1103 if (shader_version < WINED3DPS_VERSION(1,4))
1104 strcpy(reg_coord, reg_dest);
1105 else
1106 pshader_gen_input_modifier_line(arg->shader, buffer, src[0], 0, reg_coord);
1107
1108 /* 1.0-1.4: Use destination register number as texture code.
1109 2.0+: Use provided sampler number as texure code. */
1110 if (shader_version < WINED3DPS_VERSION(2,0))
1111 reg_sampler_code = reg_dest_code;
1112 else
1113 reg_sampler_code = src[1] & WINED3DSP_REGNUM_MASK;
1114
1115 /* projection flag:
1116 * 1.1, 1.2, 1.3: Use WINED3DTSS_TEXTURETRANSFORMFLAGS
1117 * 1.4: Use WINED3DSPSM_DZ or WINED3DSPSM_DW on src[0]
1118 * 2.0+: Use WINED3DSI_TEXLD_PROJECT on the opcode
1119 */
1120 if (shader_version < WINED3DPS_VERSION(1,4))
1121 {
1122 DWORD flags = 0;
1123 if(reg_sampler_code < MAX_TEXTURES) {
1124 flags = deviceImpl->stateBlock->textureState[reg_sampler_code][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1125 }
1126 if (flags & WINED3DTTFF_PROJECTED) {
1127 projected = TRUE;
1128 }
1129 }
1130 else if (shader_version < WINED3DPS_VERSION(2,0))
1131 {
1132 DWORD src_mod = arg->src[0] & WINED3DSP_SRCMOD_MASK;
1133 if (src_mod == WINED3DSPSM_DZ) {
1134 projected = TRUE;
1135 } else if(src_mod == WINED3DSPSM_DW) {
1136 projected = TRUE;
1137 }
1138 } else {
1139 if(arg->opcode_token & WINED3DSI_TEXLD_PROJECT) {
1140 projected = TRUE;
1141 }
1142 if(arg->opcode_token & WINED3DSI_TEXLD_BIAS) {
1143 bias = TRUE;
1144 }
1145 }
1146 shader_hw_sample(arg, reg_sampler_code, reg_dest, reg_coord, projected, bias);
1147 }
1148
1149 static void pshader_hw_texcoord(const SHADER_OPCODE_ARG *arg)
1150 {
1151 DWORD dst = arg->dst;
1152 SHADER_BUFFER* buffer = arg->buffer;
1153
1154 char tmp[20];
1155 shader_arb_get_write_mask(arg, dst, tmp);
1156 if (arg->reg_maps->shader_version != WINED3DPS_VERSION(1,4))
1157 {
1158 DWORD reg = dst & WINED3DSP_REGNUM_MASK;
1159 shader_addline(buffer, "MOV_SAT T%u%s, fragment.texcoord[%u];\n", reg, tmp, reg);
1160 } else {
1161 DWORD reg1 = dst & WINED3DSP_REGNUM_MASK;
1162 char reg_src[40];
1163
1164 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, reg_src);
1165 shader_addline(buffer, "MOV R%u%s, %s;\n", reg1, tmp, reg_src);
1166 }
1167 }
1168
1169 static void pshader_hw_texreg2ar(const SHADER_OPCODE_ARG *arg)
1170 {
1171 SHADER_BUFFER* buffer = arg->buffer;
1172 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1173 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1174 DWORD flags;
1175
1176 DWORD reg1 = arg->dst & WINED3DSP_REGNUM_MASK;
1177 char dst_str[8];
1178 char src_str[50];
1179
1180 sprintf(dst_str, "T%u", reg1);
1181 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src_str);
1182 shader_addline(buffer, "MOV TMP.x, %s.w;\n", src_str);
1183 shader_addline(buffer, "MOV TMP.y, %s.x;\n", src_str);
1184 flags = reg1 < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg1][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1185 shader_hw_sample(arg, reg1, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1186 }
1187
1188 static void pshader_hw_texreg2gb(const SHADER_OPCODE_ARG *arg)
1189 {
1190 SHADER_BUFFER* buffer = arg->buffer;
1191
1192 DWORD reg1 = arg->dst & WINED3DSP_REGNUM_MASK;
1193 char dst_str[8];
1194 char src_str[50];
1195
1196 sprintf(dst_str, "T%u", reg1);
1197 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src_str);
1198 shader_addline(buffer, "MOV TMP.x, %s.y;\n", src_str);
1199 shader_addline(buffer, "MOV TMP.y, %s.z;\n", src_str);
1200 shader_hw_sample(arg, reg1, dst_str, "TMP", FALSE, FALSE);
1201 }
1202
1203 static void pshader_hw_texreg2rgb(const SHADER_OPCODE_ARG *arg)
1204 {
1205 SHADER_BUFFER* buffer = arg->buffer;
1206 DWORD reg1 = arg->dst & WINED3DSP_REGNUM_MASK;
1207 char dst_str[8];
1208 char src_str[50];
1209
1210 sprintf(dst_str, "T%u", reg1);
1211 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src_str);
1212 shader_hw_sample(arg, reg1, dst_str, src_str, FALSE, FALSE);
1213 }
1214
1215 static void pshader_hw_texbem(const SHADER_OPCODE_ARG *arg)
1216 {
1217 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1218 BOOL has_bumpmat = FALSE;
1219 BOOL has_luminance = FALSE;
1220 int i;
1221
1222 DWORD dst = arg->dst;
1223 DWORD src = arg->src[0] & WINED3DSP_REGNUM_MASK;
1224 SHADER_BUFFER* buffer = arg->buffer;
1225
1226 char reg_coord[40];
1227 DWORD reg_dest_code;
1228
1229 /* All versions have a destination register */
1230 reg_dest_code = dst & WINED3DSP_REGNUM_MASK;
1231 /* Can directly use the name because texbem is only valid for <= 1.3 shaders */
1232 pshader_get_register_name(arg->shader, dst, reg_coord);
1233
1234 for(i = 0; i < This->numbumpenvmatconsts; i++) {
1235 if(This->bumpenvmatconst[i].const_num != -1 && reg_dest_code == This->bumpenvmatconst[i].texunit) {
1236 has_bumpmat = TRUE;
1237 break;
1238 }
1239 }
1240 for(i = 0; i < This->numbumpenvmatconsts; i++) {
1241 if(This->luminanceconst[i].const_num != -1 && reg_dest_code == This->luminanceconst[i].texunit) {
1242 has_luminance = TRUE;
1243 break;
1244 }
1245 }
1246
1247 if(has_bumpmat) {
1248 /* Sampling the perturbation map in Tsrc was done already, including the signedness correction if needed */
1249
1250 shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, x, z, 0, 0;\n", reg_dest_code);
1251 shader_addline(buffer, "DP3 TMP.x, TMP2, T%u;\n", src);
1252 shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, y, w, 0, 0;\n", reg_dest_code);
1253 shader_addline(buffer, "DP3 TMP.y, TMP2, T%u;\n", src);
1254
1255 /* with projective textures, texbem only divides the static texture coord, not the displacement,
1256 * so we can't let the GL handle this.
1257 */
1258 if (((IWineD3DDeviceImpl*) This->baseShader.device)->stateBlock->textureState[reg_dest_code][WINED3DTSS_TEXTURETRANSFORMFLAGS]
1259 & WINED3DTTFF_PROJECTED) {
1260 shader_addline(buffer, "RCP TMP2.w, %s.w;\n", reg_coord);
1261 shader_addline(buffer, "MUL TMP2.xy, %s, TMP2.w;\n", reg_coord);
1262 shader_addline(buffer, "ADD TMP.xy, TMP, TMP2;\n");
1263 } else {
1264 shader_addline(buffer, "ADD TMP.xy, TMP, %s;\n", reg_coord);
1265 }
1266
1267 shader_hw_sample(arg, reg_dest_code, reg_coord, "TMP", FALSE, FALSE);
1268
1269 if(arg->opcode->opcode == WINED3DSIO_TEXBEML && has_luminance) {
1270 shader_addline(buffer, "MAD TMP, T%u.z, luminance%d.x, luminance%d.y;\n",
1271 src, reg_dest_code, reg_dest_code);
1272 shader_addline(buffer, "MUL %s, %s, TMP;\n", reg_coord, reg_coord);
1273 }
1274
1275 } else {
1276 DWORD tf;
1277 if(reg_dest_code < MAX_TEXTURES) {
1278 tf = ((IWineD3DDeviceImpl*) This->baseShader.device)->stateBlock->textureState[reg_dest_code][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1279 } else {
1280 tf = 0;
1281 }
1282 /* Without a bump matrix loaded, just sample with the unmodified coordinates */
1283 shader_hw_sample(arg, reg_dest_code, reg_coord, reg_coord, tf & WINED3DTTFF_PROJECTED, FALSE);
1284 }
1285 }
1286
1287 static void pshader_hw_texm3x2pad(const SHADER_OPCODE_ARG *arg)
1288 {
1289 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1290 SHADER_BUFFER* buffer = arg->buffer;
1291 char src0_name[50];
1292
1293 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src0_name);
1294 shader_addline(buffer, "DP3 TMP.x, T%u, %s;\n", reg, src0_name);
1295 }
1296
1297 static void pshader_hw_texm3x2tex(const SHADER_OPCODE_ARG *arg)
1298 {
1299 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1300 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1301 DWORD flags;
1302 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1303 SHADER_BUFFER* buffer = arg->buffer;
1304 char dst_str[8];
1305 char src0_name[50];
1306
1307 sprintf(dst_str, "T%u", reg);
1308 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src0_name);
1309 shader_addline(buffer, "DP3 TMP.y, T%u, %s;\n", reg, src0_name);
1310 flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1311 shader_hw_sample(arg, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1312 }
1313
1314 static void pshader_hw_texm3x3pad(const SHADER_OPCODE_ARG *arg)
1315 {
1316 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1317 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1318 SHADER_BUFFER* buffer = arg->buffer;
1319 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1320 char src0_name[50];
1321
1322 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src0_name);
1323 shader_addline(buffer, "DP3 TMP.%c, T%u, %s;\n", 'x' + current_state->current_row, reg, src0_name);
1324 current_state->texcoord_w[current_state->current_row++] = reg;
1325 }
1326
1327 static void pshader_hw_texm3x3tex(const SHADER_OPCODE_ARG *arg)
1328 {
1329 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1330 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1331 DWORD flags;
1332 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1333 SHADER_BUFFER* buffer = arg->buffer;
1334 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1335 char dst_str[8];
1336 char src0_name[50];
1337
1338 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src0_name);
1339 shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
1340
1341 /* Sample the texture using the calculated coordinates */
1342 sprintf(dst_str, "T%u", reg);
1343 flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1344 shader_hw_sample(arg, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1345 current_state->current_row = 0;
1346 }
1347
1348 static void pshader_hw_texm3x3vspec(const SHADER_OPCODE_ARG *arg)
1349 {
1350 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1351 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1352 DWORD flags;
1353 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1354 SHADER_BUFFER* buffer = arg->buffer;
1355 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1356 char dst_str[8];
1357 char src0_name[50];
1358
1359 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src0_name);
1360 shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
1361
1362 /* Construct the eye-ray vector from w coordinates */
1363 shader_addline(buffer, "MOV TMP2.x, fragment.texcoord[%u].w;\n", current_state->texcoord_w[0]);
1364 shader_addline(buffer, "MOV TMP2.y, fragment.texcoord[%u].w;\n", current_state->texcoord_w[1]);
1365 shader_addline(buffer, "MOV TMP2.z, fragment.texcoord[%u].w;\n", reg);
1366
1367 /* Calculate reflection vector
1368 */
1369 shader_addline(buffer, "DP3 TMP.w, TMP, TMP2;\n");
1370 /* The .w is ignored when sampling, so I can use TMP2.w to calculate dot(N, N) */
1371 shader_addline(buffer, "DP3 TMP2.w, TMP, TMP;\n");
1372 shader_addline(buffer, "RCP TMP2.w, TMP2.w;\n");
1373 shader_addline(buffer, "MUL TMP.w, TMP.w, TMP2.w;\n");
1374 shader_addline(buffer, "MUL TMP, TMP.w, TMP;\n");
1375 shader_addline(buffer, "MAD TMP, coefmul.x, TMP, -TMP2;\n");
1376
1377 /* Sample the texture using the calculated coordinates */
1378 sprintf(dst_str, "T%u", reg);
1379 flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1380 shader_hw_sample(arg, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1381 current_state->current_row = 0;
1382 }
1383
1384 static void pshader_hw_texm3x3spec(const SHADER_OPCODE_ARG *arg)
1385 {
1386 IWineD3DPixelShaderImpl* This = (IWineD3DPixelShaderImpl*) arg->shader;
1387 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1388 DWORD flags;
1389 DWORD reg = arg->dst & WINED3DSP_REGNUM_MASK;
1390 DWORD reg3 = arg->src[1] & WINED3DSP_REGNUM_MASK;
1391 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1392 SHADER_BUFFER* buffer = arg->buffer;
1393 char dst_str[8];
1394 char src0_name[50];
1395
1396 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src0_name);
1397 shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
1398
1399 /* Calculate reflection vector.
1400 *
1401 * dot(N, E)
1402 * TMP.xyz = 2 * --------- * N - E
1403 * dot(N, N)
1404 *
1405 * Which normalizes the normal vector
1406 */
1407 shader_addline(buffer, "DP3 TMP.w, TMP, C[%u];\n", reg3);
1408 shader_addline(buffer, "DP3 TMP2.w, TMP, TMP;\n");
1409 shader_addline(buffer, "RCP TMP2.w, TMP2.w;\n");
1410 shader_addline(buffer, "MUL TMP.w, TMP.w, TMP2.w;\n");
1411 shader_addline(buffer, "MUL TMP, TMP.w, TMP;\n");
1412 shader_addline(buffer, "MAD TMP, coefmul.x, TMP, -C[%u];\n", reg3);
1413
1414 /* Sample the texture using the calculated coordinates */
1415 sprintf(dst_str, "T%u", reg);
1416 flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1417 shader_hw_sample(arg, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1418 current_state->current_row = 0;
1419 }
1420
1421 static void pshader_hw_texdepth(const SHADER_OPCODE_ARG *arg)
1422 {
1423 SHADER_BUFFER* buffer = arg->buffer;
1424 char dst_name[50];
1425
1426 /* texdepth has an implicit destination, the fragment depth value. It's only parameter,
1427 * which is essentially an input, is the destination register because it is the first
1428 * parameter. According to the msdn, this must be register r5, but let's keep it more flexible
1429 * here
1430 */
1431 pshader_get_register_name(arg->shader, arg->dst, dst_name);
1432
1433 /* According to the msdn, the source register(must be r5) is unusable after
1434 * the texdepth instruction, so we're free to modify it
1435 */
1436 shader_addline(buffer, "MIN %s.y, %s.y, one.y;\n", dst_name, dst_name);
1437
1438 /* How to deal with the special case dst_name.g == 0? if r != 0, then
1439 * the r * (1 / 0) will give infinity, which is clamped to 1.0, the correct
1440 * result. But if r = 0.0, then 0 * inf = 0, which is incorrect.
1441 */
1442 shader_addline(buffer, "RCP %s.y, %s.y;\n", dst_name, dst_name);
1443 shader_addline(buffer, "MUL TMP.x, %s.x, %s.y;\n", dst_name, dst_name);
1444 shader_addline(buffer, "MIN TMP.x, TMP.x, one.x;\n");
1445 shader_addline(buffer, "MAX result.depth, TMP.x, 0.0;\n");
1446 }
1447
1448 /** Process the WINED3DSIO_TEXDP3TEX instruction in ARB:
1449 * Take a 3-component dot product of the TexCoord[dstreg] and src,
1450 * then perform a 1D texture lookup from stage dstregnum, place into dst. */
1451 static void pshader_hw_texdp3tex(const SHADER_OPCODE_ARG *arg)
1452 {
1453 SHADER_BUFFER* buffer = arg->buffer;
1454 DWORD sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
1455 char src0[50];
1456 char dst_str[8];
1457
1458 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src0);
1459 shader_addline(buffer, "MOV TMP, 0.0;\n");
1460 shader_addline(buffer, "DP3 TMP.x, T%u, %s;\n", sampler_idx, src0);
1461
1462 sprintf(dst_str, "T%u", sampler_idx);
1463 shader_hw_sample(arg, sampler_idx, dst_str, "TMP", FALSE /* Only one coord, can't be projected */, FALSE);
1464 }
1465
1466 /** Process the WINED3DSIO_TEXDP3 instruction in ARB:
1467 * Take a 3-component dot product of the TexCoord[dstreg] and src. */
1468 static void pshader_hw_texdp3(const SHADER_OPCODE_ARG *arg)
1469 {
1470 char src0[50];
1471 char dst_str[50];
1472 char dst_mask[6];
1473 DWORD dstreg = arg->dst & WINED3DSP_REGNUM_MASK;
1474 SHADER_BUFFER* buffer = arg->buffer;
1475
1476 /* Handle output register */
1477 pshader_get_register_name(arg->shader, arg->dst, dst_str);
1478 shader_arb_get_write_mask(arg, arg->dst, dst_mask);
1479
1480 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src0);
1481 shader_addline(buffer, "DP3 %s%s, T%u, %s;\n", dst_str, dst_mask, dstreg, src0);
1482
1483 /* TODO: Handle output modifiers */
1484 }
1485
1486 /** Process the WINED3DSIO_TEXM3X3 instruction in ARB
1487 * Perform the 3rd row of a 3x3 matrix multiply */
1488 static void pshader_hw_texm3x3(const SHADER_OPCODE_ARG *arg)
1489 {
1490 SHADER_BUFFER* buffer = arg->buffer;
1491 char dst_str[50];
1492 char dst_mask[6];
1493 char src0[50];
1494 DWORD dst_reg = arg->dst & WINED3DSP_REGNUM_MASK;
1495
1496 pshader_get_register_name(arg->shader, arg->dst, dst_str);
1497 shader_arb_get_write_mask(arg, arg->dst, dst_mask);
1498
1499 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src0);
1500 shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", dst_reg, src0);
1501 shader_addline(buffer, "MOV %s%s, TMP;\n", dst_str, dst_mask);
1502
1503 /* TODO: Handle output modifiers */
1504 }
1505
1506 /** Process the WINED3DSIO_TEXM3X2DEPTH instruction in ARB:
1507 * Last row of a 3x2 matrix multiply, use the result to calculate the depth:
1508 * Calculate tmp0.y = TexCoord[dstreg] . src.xyz; (tmp0.x has already been calculated)
1509 * depth = (tmp0.y == 0.0) ? 1.0 : tmp0.x / tmp0.y
1510 */
1511 static void pshader_hw_texm3x2depth(const SHADER_OPCODE_ARG *arg)
1512 {
1513 SHADER_BUFFER* buffer = arg->buffer;
1514 DWORD dst_reg = arg->dst & WINED3DSP_REGNUM_MASK;
1515 char src0[50];
1516
1517 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src0);
1518 shader_addline(buffer, "DP3 TMP.y, T%u, %s;\n", dst_reg, src0);
1519
1520 /* How to deal with the special case dst_name.g == 0? if r != 0, then
1521 * the r * (1 / 0) will give infinity, which is clamped to 1.0, the correct
1522 * result. But if r = 0.0, then 0 * inf = 0, which is incorrect.
1523 */
1524 shader_addline(buffer, "RCP TMP.y, TMP.y;\n");
1525 shader_addline(buffer, "MUL TMP.x, TMP.x, TMP.y;\n");
1526 shader_addline(buffer, "MIN TMP.x, TMP.x, one.x;\n");
1527 shader_addline(buffer, "MAX result.depth, TMP.x, 0.0;\n");
1528 }
1529
1530 /** Handles transforming all WINED3DSIO_M?x? opcodes for
1531 Vertex/Pixel shaders to ARB_vertex_program codes */
1532 static void shader_hw_mnxn(const SHADER_OPCODE_ARG *arg)
1533 {
1534 int i;
1535 int nComponents = 0;
1536 SHADER_OPCODE_ARG tmpArg;
1537 IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)arg->shader;
1538 const SHADER_OPCODE *opcode_table = shader->baseShader.shader_ins;
1539 DWORD shader_version = arg->reg_maps->shader_version;
1540
1541 memset(&tmpArg, 0, sizeof(SHADER_OPCODE_ARG));
1542
1543 /* Set constants for the temporary argument */
1544 tmpArg.shader = arg->shader;
1545 tmpArg.buffer = arg->buffer;
1546 tmpArg.src[0] = arg->src[0];
1547 tmpArg.src_addr[0] = arg->src_addr[0];
1548 tmpArg.src_addr[1] = arg->src_addr[1];
1549 tmpArg.reg_maps = arg->reg_maps;
1550
1551 switch(arg->opcode->opcode) {
1552 case WINED3DSIO_M4x4:
1553 nComponents = 4;
1554 tmpArg.opcode = shader_get_opcode(opcode_table, shader_version, WINED3DSIO_DP4);
1555 break;
1556 case WINED3DSIO_M4x3:
1557 nComponents = 3;
1558 tmpArg.opcode = shader_get_opcode(opcode_table, shader_version, WINED3DSIO_DP4);
1559 break;
1560 case WINED3DSIO_M3x4:
1561 nComponents = 4;
1562 tmpArg.opcode = shader_get_opcode(opcode_table, shader_version, WINED3DSIO_DP3);
1563 break;
1564 case WINED3DSIO_M3x3:
1565 nComponents = 3;
1566 tmpArg.opcode = shader_get_opcode(opcode_table, shader_version, WINED3DSIO_DP3);
1567 break;
1568 case WINED3DSIO_M3x2:
1569 nComponents = 2;
1570 tmpArg.opcode = shader_get_opcode(opcode_table, shader_version, WINED3DSIO_DP3);
1571 break;
1572 default:
1573 break;
1574 }
1575
1576 for (i = 0; i < nComponents; i++) {
1577 tmpArg.dst = ((arg->dst) & ~WINED3DSP_WRITEMASK_ALL)|(WINED3DSP_WRITEMASK_0<<i);
1578 tmpArg.src[1] = arg->src[1]+i;
1579 shader_hw_map2gl(&tmpArg);
1580 }
1581 }
1582
1583 static void vshader_hw_rsq_rcp(const SHADER_OPCODE_ARG *arg)
1584 {
1585 CONST SHADER_OPCODE* curOpcode = arg->opcode;
1586 SHADER_BUFFER* buffer = arg->buffer;
1587 DWORD dst = arg->dst;
1588 DWORD src = arg->src[0];
1589 DWORD swizzle = (src & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
1590 const char *instruction;
1591
1592 char tmpLine[256];
1593
1594 switch(curOpcode->opcode)
1595 {
1596 case WINED3DSIO_RSQ: instruction = "RSQ"; break;
1597 case WINED3DSIO_RCP: instruction = "RCP"; break;
1598 default: instruction = "";
1599 FIXME("Unhandled opcode %s\n", curOpcode->name);
1600 break;
1601 }
1602
1603 strcpy(tmpLine, instruction);
1604 vshader_program_add_param(arg, dst, FALSE, tmpLine); /* Destination */
1605 strcat(tmpLine, ",");
1606 vshader_program_add_param(arg, src, TRUE, tmpLine);
1607 if ((WINED3DSP_NOSWIZZLE >> WINED3DSP_SWIZZLE_SHIFT) == swizzle) {
1608 /* Dx sdk says .x is used if no swizzle is given, but our test shows that
1609 * .w is used
1610 */
1611 strcat(tmpLine, ".w");
1612 }
1613
1614 shader_addline(buffer, "%s;\n", tmpLine);
1615 }
1616
1617 static void shader_hw_nrm(const SHADER_OPCODE_ARG *arg)
1618 {
1619 SHADER_BUFFER* buffer = arg->buffer;
1620 char dst_name[50];
1621 char src_name[50];
1622 char dst_wmask[20];
1623 DWORD shift = (arg->dst & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
1624 BOOL sat = (arg->dst & WINED3DSP_DSTMOD_MASK) & WINED3DSPDM_SATURATE;
1625
1626 pshader_get_register_name(arg->shader, arg->dst, dst_name);
1627 shader_arb_get_write_mask(arg, arg->dst, dst_wmask);
1628
1629 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src_name);
1630 shader_addline(buffer, "DP3 TMP, %s, %s;\n", src_name, src_name);
1631 shader_addline(buffer, "RSQ TMP, TMP.x;\n");
1632 /* dst.w = src[0].w * 1 / (src.x^2 + src.y^2 + src.z^2)^(1/2) according to msdn*/
1633 shader_addline(buffer, "MUL%s %s%s, %s, TMP;\n", sat ? "_SAT" : "", dst_name, dst_wmask,
1634 src_name);
1635
1636 if (shift != 0)
1637 pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
1638 }
1639
1640 static void shader_hw_sincos(const SHADER_OPCODE_ARG *arg)
1641 {
1642 /* This instruction exists in ARB, but the d3d instruction takes two extra parameters which
1643 * must contain fixed constants. So we need a separate function to filter those constants and
1644 * can't use map2gl
1645 */
1646 SHADER_BUFFER* buffer = arg->buffer;
1647 char dst_name[50];
1648 char src_name[50];
1649 char dst_wmask[20];
1650 DWORD shift = (arg->dst & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
1651 BOOL sat = (arg->dst & WINED3DSP_DSTMOD_MASK) & WINED3DSPDM_SATURATE;
1652
1653 pshader_get_register_name(arg->shader, arg->dst, dst_name);
1654 shader_arb_get_write_mask(arg, arg->dst, dst_wmask);
1655
1656 pshader_gen_input_modifier_line(arg->shader, buffer, arg->src[0], 0, src_name);
1657 shader_addline(buffer, "SCS%s %s%s, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask,
1658 src_name);
1659
1660 if (shift != 0)
1661 pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
1662
1663 }
1664
1665 static GLuint create_arb_blt_vertex_program(const WineD3D_GL_Info *gl_info)
1666 {
1667 GLuint program_id = 0;
1668 const char *blt_vprogram =
1669 "!!ARBvp1.0\n"
1670 "PARAM c[1] = { { 1, 0.5 } };\n"
1671 "MOV result.position, vertex.position;\n"
1672 "MOV result.color, c[0].x;\n"
1673 "MOV result.texcoord[0], vertex.texcoord[0];\n"
1674 "END\n";
1675
1676 GL_EXTCALL(glGenProgramsARB(1, &program_id));
1677 GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, program_id));
1678 GL_EXTCALL(glProgramStringARB(GL_VERTEX_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(blt_vprogram), blt_vprogram));
1679
1680 if (glGetError() == GL_INVALID_OPERATION) {
1681 GLint pos;
1682 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
1683 FIXME("Vertex program error at position %d: %s\n", pos,
1684 debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
1685 }
1686
1687 return program_id;
1688 }
1689
1690 static GLuint create_arb_blt_fragment_program(const WineD3D_GL_Info *gl_info, enum tex_types tex_type)
1691 {
1692 GLuint program_id = 0;
1693 static const char * const blt_fprograms[tex_type_count] =
1694 {
1695 /* tex_1d */
1696 NULL,
1697 /* tex_2d */
1698 "!!ARBfp1.0\n"
1699 "TEMP R0;\n"
1700 "TEX R0.x, fragment.texcoord[0], texture[0], 2D;\n"
1701 "MOV result.depth.z, R0.x;\n"
1702 "END\n",
1703 /* tex_3d */
1704 NULL,
1705 /* tex_cube */
1706 "!!ARBfp1.0\n"
1707 "TEMP R0;\n"
1708 "TEX R0.x, fragment.texcoord[0], texture[0], CUBE;\n"
1709 "MOV result.depth.z, R0.x;\n"
1710 "END\n",
1711 /* tex_rect */
1712 "!!ARBfp1.0\n"
1713 "TEMP R0;\n"
1714 "TEX R0.x, fragment.texcoord[0], texture[0], RECT;\n"
1715 "MOV result.depth.z, R0.x;\n"
1716 "END\n",
1717 };
1718
1719 if (!blt_fprograms[tex_type])
1720 {
1721 FIXME("tex_type %#x not supported\n", tex_type);
1722 tex_type = tex_2d;
1723 }
1724
1725 GL_EXTCALL(glGenProgramsARB(1, &program_id));
1726 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, program_id));
1727 GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(blt_fprograms[tex_type]), blt_fprograms[tex_type]));
1728
1729 if (glGetError() == GL_INVALID_OPERATION) {
1730 GLint pos;
1731 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
1732 FIXME("Fragment program error at position %d: %s\n", pos,
1733 debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
1734 }
1735
1736 return program_id;
1737 }
1738
1739 static void shader_arb_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {
1740 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1741 struct shader_arb_priv *priv = This->shader_priv;
1742 const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1743
1744 if (useVS) {
1745 struct vs_compile_args compile_args;
1746
1747 TRACE("Using vertex shader\n");
1748 find_vs_compile_args((IWineD3DVertexShaderImpl *) This->stateBlock->vertexShader, This->stateBlock, &compile_args);
1749 priv->current_vprogram_id = find_gl_vshader((IWineD3DVertexShaderImpl *) This->stateBlock->vertexShader, &compile_args);
1750
1751 /* Bind the vertex program */
1752 GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id));
1753 checkGLcall("glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id);");
1754
1755 /* Enable OpenGL vertex programs */
1756 glEnable(GL_VERTEX_PROGRAM_ARB);
1757 checkGLcall("glEnable(GL_VERTEX_PROGRAM_ARB);");
1758 TRACE("(%p) : Bound vertex program %u and enabled GL_VERTEX_PROGRAM_ARB\n", This, priv->current_vprogram_id);
1759 } else if(GL_SUPPORT(ARB_VERTEX_PROGRAM)) {
1760 priv->current_vprogram_id = 0;
1761 glDisable(GL_VERTEX_PROGRAM_ARB);
1762 checkGLcall("glDisable(GL_VERTEX_PROGRAM_ARB)");
1763 }
1764
1765 if (usePS) {
1766 struct ps_compile_args compile_args;
1767 TRACE("Using pixel shader\n");
1768 find_ps_compile_args((IWineD3DPixelShaderImpl *) This->stateBlock->pixelShader, This->stateBlock, &compile_args);
1769 priv->current_fprogram_id = find_gl_pshader((IWineD3DPixelShaderImpl *) This->stateBlock->pixelShader,
1770 &compile_args);
1771
1772 /* Bind the fragment program */
1773 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id));
1774 checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id);");
1775
1776 if(!priv->use_arbfp_fixed_func) {
1777 /* Enable OpenGL fragment programs */
1778 glEnable(GL_FRAGMENT_PROGRAM_ARB);
1779 checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB);");
1780 }
1781 TRACE("(%p) : Bound fragment program %u and enabled GL_FRAGMENT_PROGRAM_ARB\n", This, priv->current_fprogram_id);
1782 } else if(GL_SUPPORT(ARB_FRAGMENT_PROGRAM) && !priv->use_arbfp_fixed_func) {
1783 /* Disable only if we're not using arbfp fixed function fragment processing. If this is used,
1784 * keep GL_FRAGMENT_PROGRAM_ARB enabled, and the fixed function pipeline will bind the fixed function
1785 * replacement shader
1786 */
1787 glDisable(GL_FRAGMENT_PROGRAM_ARB);
1788 checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
1789 priv->current_fprogram_id = 0;
1790 }
1791 }
1792
1793 static void shader_arb_select_depth_blt(IWineD3DDevice *iface, enum tex_types tex_type) {
1794 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1795 struct shader_arb_priv *priv = This->shader_priv;
1796 GLuint *blt_fprogram = &priv->depth_blt_fprogram_id[tex_type];
1797 const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1798
1799 if (!priv->depth_blt_vprogram_id) priv->depth_blt_vprogram_id = create_arb_blt_vertex_program(gl_info);
1800 GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->depth_blt_vprogram_id));
1801 glEnable(GL_VERTEX_PROGRAM_ARB);
1802
1803 if (!*blt_fprogram) *blt_fprogram = create_arb_blt_fragment_program(gl_info, tex_type);
1804 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, *blt_fprogram));
1805 glEnable(GL_FRAGMENT_PROGRAM_ARB);
1806 }
1807
1808 static void shader_arb_deselect_depth_blt(IWineD3DDevice *iface) {
1809 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1810 struct shader_arb_priv *priv = This->shader_priv;
1811 const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1812
1813 if (priv->current_vprogram_id) {
1814 GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id));
1815 checkGLcall("glBindProgramARB(GL_VERTEX_PROGRAM_ARB, vertexShader->prgId);");
1816
1817 glEnable(GL_VERTEX_PROGRAM_ARB);
1818 checkGLcall("glEnable(GL_VERTEX_PROGRAM_ARB);");
1819
1820 TRACE("(%p) : Bound vertex program %u and enabled GL_VERTEX_PROGRAM_ARB\n", This, priv->current_vprogram_id);
1821 } else {
1822 glDisable(GL_VERTEX_PROGRAM_ARB);
1823 checkGLcall("glDisable(GL_VERTEX_PROGRAM_ARB)");
1824 }
1825
1826 if (priv->current_fprogram_id) {
1827 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id));
1828 checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, pixelShader->prgId);");
1829
1830 glEnable(GL_FRAGMENT_PROGRAM_ARB);
1831 checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB);");
1832
1833 TRACE("(%p) : Bound fragment program %u and enabled GL_FRAGMENT_PROGRAM_ARB\n", This, priv->current_fprogram_id);
1834 } else {
1835 glDisable(GL_FRAGMENT_PROGRAM_ARB);
1836 checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
1837 }
1838 }
1839
1840 static void shader_arb_destroy(IWineD3DBaseShader *iface) {
1841 IWineD3DBaseShaderImpl *baseShader = (IWineD3DBaseShaderImpl *) iface;
1842 const WineD3D_GL_Info *gl_info = &((IWineD3DDeviceImpl *)baseShader->baseShader.device)->adapter->gl_info;
1843
1844 if (shader_is_pshader_version(baseShader->baseShader.reg_maps.shader_version))
1845 {
1846 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *) iface;
1847 UINT i;
1848
1849 ENTER_GL();
1850 for(i = 0; i < This->num_gl_shaders; i++) {
1851 GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId));
1852 checkGLcall("GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId))");
1853 }
1854 LEAVE_GL();
1855 HeapFree(GetProcessHeap(), 0, This->gl_shaders);
1856 This->gl_shaders = NULL;
1857 This->num_gl_shaders = 0;
1858 This->shader_array_size = 0;
1859 } else {
1860 IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *) iface;
1861 UINT i;
1862
1863 ENTER_GL();
1864 for(i = 0; i < This->num_gl_shaders; i++) {
1865 GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId));
1866 checkGLcall("GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId))");
1867 }
1868 LEAVE_GL();
1869 HeapFree(GetProcessHeap(), 0, This->gl_shaders);
1870 This->gl_shaders = NULL;
1871 This->num_gl_shaders = 0;
1872 This->shader_array_size = 0;
1873 }
1874 }
1875
1876 static HRESULT shader_arb_alloc(IWineD3DDevice *iface) {
1877 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1878 This->shader_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct shader_arb_priv));
1879 return WINED3D_OK;
1880 }
1881
1882 static void shader_arb_free(IWineD3DDevice *iface) {
1883 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1884 const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1885 struct shader_arb_priv *priv = This->shader_priv;
1886 int i;
1887
1888 if(priv->depth_blt_vprogram_id) {
1889 GL_EXTCALL(glDeleteProgramsARB(1, &priv->depth_blt_vprogram_id));
1890 }
1891 for (i = 0; i < tex_type_count; ++i) {
1892 if (priv->depth_blt_fprogram_id[i]) {
1893 GL_EXTCALL(glDeleteProgramsARB(1, &priv->depth_blt_fprogram_id[i]));
1894 }
1895 }
1896
1897 HeapFree(GetProcessHeap(), 0, This->shader_priv);
1898 }
1899
1900 static BOOL shader_arb_dirty_const(IWineD3DDevice *iface) {
1901 return TRUE;
1902 }
1903
1904 static void arbfp_add_sRGB_correction(SHADER_BUFFER *buffer, const char *fragcolor, const char *tmp1,
1905 const char *tmp2, const char *tmp3, const char *tmp4) {
1906 /* Perform sRGB write correction. See GLX_EXT_framebuffer_sRGB */
1907
1908 /* Calculate the > 0.0031308 case */
1909 shader_addline(buffer, "POW %s.x, %s.x, srgb_pow.x;\n", tmp1, fragcolor);
1910 shader_addline(buffer, "POW %s.y, %s.y, srgb_pow.y;\n", tmp1, fragcolor);
1911 shader_addline(buffer, "POW %s.z, %s.z, srgb_pow.z;\n", tmp1, fragcolor);
1912 shader_addline(buffer, "MUL %s, %s, srgb_mul_hi;\n", tmp1, tmp1);
1913 shader_addline(buffer, "SUB %s, %s, srgb_sub_hi;\n", tmp1, tmp1);
1914 /* Calculate the < case */
1915 shader_addline(buffer, "MUL %s, srgb_mul_low, %s;\n", tmp2, fragcolor);
1916 /* Get 1.0 / 0.0 masks for > 0.0031308 and < 0.0031308 */
1917 shader_addline(buffer, "SLT %s, srgb_comparison, %s;\n", tmp3, fragcolor);
1918 shader_addline(buffer, "SGE %s, srgb_comparison, %s;\n", tmp4, fragcolor);
1919 /* Store the components > 0.0031308 in the destination */
1920 shader_addline(buffer, "MUL %s, %s, %s;\n", fragcolor, tmp1, tmp3);
1921 /* Add the components that are < 0.0031308 */
1922 shader_addline(buffer, "MAD result.color.xyz, %s, %s, %s;\n", tmp2, tmp4, fragcolor);
1923 /* [0.0;1.0] clamping. Not needed, this is done implicitly */
1924 }
1925
1926 static GLuint shader_arb_generate_pshader(IWineD3DPixelShader *iface, SHADER_BUFFER *buffer, const struct ps_compile_args *args) {
1927 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
1928 const shader_reg_maps* reg_maps = &This->baseShader.reg_maps;
1929 CONST DWORD *function = This->baseShader.function;
1930 DWORD shader_version = reg_maps->shader_version;
1931 const WineD3D_GL_Info *gl_info = &((IWineD3DDeviceImpl *)This->baseShader.device)->adapter->gl_info;
1932 const local_constant *lconst;
1933 GLuint retval;
1934 const char *fragcolor;
1935
1936 /* Create the hw ARB shader */
1937 shader_addline(buffer, "!!ARBfp1.0\n");
1938
1939 if (shader_version < WINED3DPS_VERSION(3,0)) {
1940 switch(args->fog) {
1941 case FOG_OFF:
1942 break;
1943 case FOG_LINEAR:
1944 shader_addline(buffer, "OPTION ARB_fog_linear;\n");
1945 break;
1946 case FOG_EXP:
1947 shader_addline(buffer, "OPTION ARB_fog_exp;\n");
1948 break;
1949 case FOG_EXP2:
1950 shader_addline(buffer, "OPTION ARB_fog_exp2;\n");
1951 break;
1952 }
1953 }
1954
1955 shader_addline(buffer, "TEMP TMP;\n"); /* Used in matrix ops */
1956 shader_addline(buffer, "TEMP TMP2;\n"); /* Used in matrix ops */
1957 shader_addline(buffer, "TEMP TA;\n"); /* Used for modifiers */
1958 shader_addline(buffer, "TEMP TB;\n"); /* Used for modifiers */
1959 shader_addline(buffer, "TEMP TC;\n"); /* Used for modifiers */
1960 shader_addline(buffer, "PARAM coefdiv = { 0.5, 0.25, 0.125, 0.0625 };\n");
1961 shader_addline(buffer, "PARAM coefmul = { 2, 4, 8, 16 };\n");
1962 shader_addline(buffer, "PARAM one = { 1.0, 1.0, 1.0, 1.0 };\n");
1963
1964 if (shader_version < WINED3DPS_VERSION(2,0)) {
1965 fragcolor = "R0";
1966 } else {
1967 shader_addline(buffer, "TEMP TMP_COLOR;\n");
1968 fragcolor = "TMP_COLOR";
1969 }
1970
1971 /* Base Declarations */
1972 shader_generate_arb_declarations( (IWineD3DBaseShader*) This, reg_maps, buffer, &GLINFO_LOCATION);
1973
1974 /* Base Shader Body */
1975 shader_generate_main( (IWineD3DBaseShader*) This, buffer, reg_maps, function);
1976
1977 if(args->srgb_correction) {
1978 arbfp_add_sRGB_correction(buffer, fragcolor, "TMP", "TMP2", "TA", "TB");
1979 }
1980 shader_addline(buffer, "MOV result.color, %s;\n", fragcolor);
1981 shader_addline(buffer, "END\n");
1982
1983 /* TODO: change to resource.glObjectHandle or something like that */
1984 GL_EXTCALL(glGenProgramsARB(1, &retval));
1985
1986 TRACE("Creating a hw pixel shader, prg=%d\n", retval);
1987 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, retval));
1988
1989 TRACE("Created hw pixel shader, prg=%d\n", retval);
1990 /* Create the program and check for errors */
1991 GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB,
1992 buffer->bsize, buffer->buffer));
1993
1994 if (glGetError() == GL_INVALID_OPERATION) {
1995 GLint errPos;
1996 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &errPos);
1997 FIXME("HW PixelShader Error at position %d: %s\n",
1998 errPos, debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
1999 retval = 0;
2000 }
2001
2002 /* Load immediate constants */
2003 if(!This->baseShader.load_local_constsF) {
2004 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
2005 const float *value = (const float *)lconst->value;
2006 GL_EXTCALL(glProgramLocalParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, lconst->idx, value));
2007 checkGLcall("glProgramLocalParameter4fvARB");
2008 }
2009 }
2010
2011 return retval;
2012 }
2013
2014 static GLuint shader_arb_generate_vshader(IWineD3DVertexShader *iface, SHADER_BUFFER *buffer, const struct vs_compile_args *args) {
2015 IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *)iface;
2016 const shader_reg_maps *reg_maps = &This->baseShader.reg_maps;
2017 CONST DWORD *function = This->baseShader.function;
2018 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)This->baseShader.device;
2019 const WineD3D_GL_Info *gl_info = &device->adapter->gl_info;
2020 const local_constant *lconst;
2021 GLuint ret;
2022
2023 /* Create the hw ARB shader */
2024 shader_addline(buffer, "!!ARBvp1.0\n");
2025 shader_addline(buffer, "PARAM helper_const = { 2.0, -1.0, %d.0, 0.0 };\n", This->rel_offset);
2026
2027 /* Mesa supports only 95 constants */
2028 if (GL_VEND(MESA) || GL_VEND(WINE))
2029 This->baseShader.limits.constant_float =
2030 min(95, This->baseShader.limits.constant_float);
2031
2032 shader_addline(buffer, "TEMP TMP;\n");
2033
2034 /* Base Declarations */
2035 shader_generate_arb_declarations( (IWineD3DBaseShader*) This, reg_maps, buffer, &GLINFO_LOCATION);
2036
2037 /* We need a constant to fixup the final position */
2038 shader_addline(buffer, "PARAM posFixup = program.env[%d];\n", ARB_SHADER_PRIVCONST_POS);
2039
2040 /* Initialize output parameters. GL_ARB_vertex_program does not require special initialization values
2041 * for output parameters. D3D in theory does not do that either, but some applications depend on a
2042 * proper initialization of the secondary color, and programs using the fixed function pipeline without
2043 * a replacement shader depend on the texcoord.w being set properly.
2044 *
2045 * GL_NV_vertex_program defines that all output values are initialized to {0.0, 0.0, 0.0, 1.0}. This
2046 * assertion is in effect even when using GL_ARB_vertex_program without any NV specific additions. So
2047 * skip this if NV_vertex_program is supported. Otherwise, initialize the secondary color. For the tex-
2048 * coords, we have a flag in the opengl caps. Many cards do not require the texcoord being set, and
2049 * this can eat a number of instructions, so skip it unless this cap is set as well
2050 */
2051 if(!GL_SUPPORT(NV_VERTEX_PROGRAM)) {
2052 shader_addline(buffer, "MOV result.color.secondary, -helper_const.wwwy;\n");
2053
2054 if((GLINFO_LOCATION).set_texcoord_w && !device->frag_pipe->ffp_proj_control) {
2055 int i;
2056 for(i = 0; i < min(8, MAX_REG_TEXCRD); i++) {
2057 if(This->baseShader.reg_maps.texcoord_mask[i] != 0 &&
2058 This->baseShader.reg_maps.texcoord_mask[i] != WINED3DSP_WRITEMASK_ALL) {
2059 shader_addline(buffer, "MOV result.texcoord[%u].w, -helper_const.y;\n", i);
2060 }
2061 }
2062 }
2063 }
2064
2065 /* Base Shader Body */
2066 shader_generate_main( (IWineD3DBaseShader*) This, buffer, reg_maps, function);
2067
2068 /* The D3DRS_FOGTABLEMODE render state defines if the shader-generated fog coord is used
2069 * or if the fragment depth is used. If the fragment depth is used(FOGTABLEMODE != NONE),
2070 * the fog frag coord is thrown away. If the fog frag coord is used, but not written by
2071 * the shader, it is set to 0.0(fully fogged, since start = 1.0, end = 0.0)
2072 */
2073 if(args->fog_src == VS_FOG_Z) {
2074 shader_addline(buffer, "MOV result.fogcoord, TMP_OUT.z;\n");
2075 } else if (!reg_maps->fog) {
2076 shader_addline(buffer, "MOV result.fogcoord, 0.0;\n");
2077 }
2078
2079 /* Write the final position.
2080 *
2081 * OpenGL coordinates specify the center of the pixel while d3d coords specify
2082 * the corner. The offsets are stored in z and w in posFixup. posFixup.y contains
2083 * 1.0 or -1.0 to turn the rendering upside down for offscreen rendering. PosFixup.x
2084 * contains 1.0 to allow a mad, but arb vs swizzles are too restricted for that.
2085 */
2086 shader_addline(buffer, "MUL TMP, posFixup, TMP_OUT.w;\n");
2087 shader_addline(buffer, "ADD TMP_OUT.x, TMP_OUT.x, TMP.z;\n");
2088 shader_addline(buffer, "MAD TMP_OUT.y, TMP_OUT.y, posFixup.y, TMP.w;\n");
2089
2090 /* Z coord [0;1]->[-1;1] mapping, see comment in transform_projection in state.c
2091 * and the glsl equivalent
2092 */
2093 shader_addline(buffer, "MAD TMP_OUT.z, TMP_OUT.z, helper_const.x, -TMP_OUT.w;\n");
2094
2095 shader_addline(buffer, "MOV result.position, TMP_OUT;\n");
2096
2097 shader_addline(buffer, "END\n");
2098
2099 /* TODO: change to resource.glObjectHandle or something like that */
2100 GL_EXTCALL(glGenProgramsARB(1, &ret));
2101
2102 TRACE("Creating a hw vertex shader, prg=%d\n", ret);
2103 GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, ret));
2104
2105 TRACE("Created hw vertex shader, prg=%d\n", ret);
2106 /* Create the program and check for errors */
2107 GL_EXTCALL(glProgramStringARB(GL_VERTEX_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB,
2108 buffer->bsize, buffer->buffer));
2109
2110 if (glGetError() == GL_INVALID_OPERATION) {
2111 GLint errPos;
2112 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &errPos);
2113 FIXME("HW VertexShader Error at position %d: %s\n",
2114 errPos, debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
2115 ret = -1;
2116 } else {
2117 /* Load immediate constants */
2118 if(!This->baseShader.load_local_constsF) {
2119 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
2120 const float *value = (const float *)lconst->value;
2121 GL_EXTCALL(glProgramLocalParameter4fvARB(GL_VERTEX_PROGRAM_ARB, lconst->idx, value));
2122 }
2123 }
2124 }
2125 return ret;
2126 }
2127
2128 static void shader_arb_get_caps(WINED3DDEVTYPE devtype, const WineD3D_GL_Info *gl_info, struct shader_caps *pCaps)
2129 {
2130 /* We don't have an ARB fixed function pipeline yet, so let the none backend set its caps,
2131 * then overwrite the shader specific ones
2132 */
2133 none_shader_backend.shader_get_caps(devtype, gl_info, pCaps);
2134
2135 if(GL_SUPPORT(ARB_VERTEX_PROGRAM)) {
2136 pCaps->VertexShaderVersion = WINED3DVS_VERSION(1,1);
2137 TRACE_(d3d_caps)("Hardware vertex shader version 1.1 enabled (ARB_PROGRAM)\n");
2138 pCaps->MaxVertexShaderConst = GL_LIMITS(vshader_constantsF);
2139 }
2140
2141 if(GL_SUPPORT(ARB_FRAGMENT_PROGRAM)) {
2142 pCaps->PixelShaderVersion = WINED3DPS_VERSION(1,4);
2143 pCaps->PixelShader1xMaxValue = 8.0;
2144 TRACE_(d3d_caps)("Hardware pixel shader version 1.4 enabled (ARB_PROGRAM)\n");
2145 }
2146 }
2147
2148 static BOOL shader_arb_color_fixup_supported(struct color_fixup_desc fixup)
2149 {
2150 if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
2151 {
2152 TRACE("Checking support for color_fixup:\n");
2153 dump_color_fixup_desc(fixup);
2154 }
2155
2156 /* We support everything except YUV conversions. */
2157 if (!is_yuv_fixup(fixup))
2158 {
2159 TRACE("[OK]\n");
2160 return TRUE;
2161 }
2162
2163 TRACE("[FAILED]\n");
2164 return FALSE;
2165 }
2166
2167 static const SHADER_HANDLER shader_arb_instruction_handler_table[WINED3DSIH_TABLE_SIZE] =
2168 {
2169 /* WINED3DSIH_ABS */ shader_hw_map2gl,
2170 /* WINED3DSIH_ADD */ shader_hw_map2gl,
2171 /* WINED3DSIH_BEM */ pshader_hw_bem,
2172 /* WINED3DSIH_BREAK */ NULL,
2173 /* WINED3DSIH_BREAKC */ NULL,
2174 /* WINED3DSIH_BREAKP */ NULL,
2175 /* WINED3DSIH_CALL */ NULL,
2176 /* WINED3DSIH_CALLNZ */ NULL,
2177 /* WINED3DSIH_CMP */ pshader_hw_cmp,
2178 /* WINED3DSIH_CND */ pshader_hw_cnd,
2179 /* WINED3DSIH_CRS */ shader_hw_map2gl,
2180 /* WINED3DSIH_DCL */ NULL,
2181 /* WINED3DSIH_DEF */ NULL,
2182 /* WINED3DSIH_DEFB */ NULL,
2183 /* WINED3DSIH_DEFI */ NULL,
2184 /* WINED3DSIH_DP2ADD */ pshader_hw_dp2add,
2185 /* WINED3DSIH_DP3 */ shader_hw_map2gl,
2186 /* WINED3DSIH_DP4 */ shader_hw_map2gl,
2187 /* WINED3DSIH_DST */ shader_hw_map2gl,
2188 /* WINED3DSIH_DSX */ NULL,
2189 /* WINED3DSIH_DSY */ NULL,
2190 /* WINED3DSIH_ELSE */ NULL,
2191 /* WINED3DSIH_ENDIF */ NULL,
2192 /* WINED3DSIH_ENDLOOP */ NULL,
2193 /* WINED3DSIH_ENDREP */ NULL,
2194 /* WINED3DSIH_EXP */ shader_hw_map2gl,
2195 /* WINED3DSIH_EXPP */ shader_hw_map2gl,
2196 /* WINED3DSIH_FRC */ shader_hw_map2gl,
2197 /* WINED3DSIH_IF */ NULL,
2198 /* WINED3DSIH_IFC */ NULL,
2199 /* WINED3DSIH_LABEL */ NULL,
2200 /* WINED3DSIH_LIT */ shader_hw_map2gl,
2201 /* WINED3DSIH_LOG */ shader_hw_map2gl,
2202 /* WINED3DSIH_LOGP */ shader_hw_map2gl,
2203 /* WINED3DSIH_LOOP */ NULL,
2204 /* WINED3DSIH_LRP */ shader_hw_map2gl,
2205 /* WINED3DSIH_M3x2 */ shader_hw_mnxn,
2206 /* WINED3DSIH_M3x3 */ shader_hw_mnxn,
2207 /* WINED3DSIH_M3x4 */ shader_hw_mnxn,
2208 /* WINED3DSIH_M4x3 */ shader_hw_mnxn,
2209 /* WINED3DSIH_M4x4 */ shader_hw_mnxn,
2210 /* WINED3DSIH_MAD */ shader_hw_map2gl,
2211 /* WINED3DSIH_MAX */ shader_hw_map2gl,
2212 /* WINED3DSIH_MIN */ shader_hw_map2gl,
2213 /* WINED3DSIH_MOV */ shader_hw_mov,
2214 /* WINED3DSIH_MOVA */ shader_hw_mov,
2215 /* WINED3DSIH_MUL */ shader_hw_map2gl,
2216 /* WINED3DSIH_NOP */ shader_hw_map2gl,
2217 /* WINED3DSIH_NRM */ shader_hw_nrm,
2218 /* WINED3DSIH_PHASE */ NULL,
2219 /* WINED3DSIH_POW */ shader_hw_map2gl,
2220 /* WINED3DSIH_RCP */ vshader_hw_rsq_rcp,
2221 /* WINED3DSIH_REP */ NULL,
2222 /* WINED3DSIH_RET */ NULL,
2223 /* WINED3DSIH_RSQ */ vshader_hw_rsq_rcp,
2224 /* WINED3DSIH_SETP */ NULL,
2225 /* WINED3DSIH_SGE */ shader_hw_map2gl,
2226 /* WINED3DSIH_SGN */ NULL,
2227 /* WINED3DSIH_SINCOS */ shader_hw_sincos,
2228 /* WINED3DSIH_SLT */ shader_hw_map2gl,
2229 /* WINED3DSIH_SUB */ shader_hw_map2gl,
2230 /* WINED3DSIH_TEX */ pshader_hw_tex,
2231 /* WINED3DSIH_TEXBEM */ pshader_hw_texbem,
2232 /* WINED3DSIH_TEXBEML */ pshader_hw_texbem,
2233 /* WINED3DSIH_TEXCOORD */ pshader_hw_texcoord,
2234 /* WINED3DSIH_TEXDEPTH */ pshader_hw_texdepth,
2235 /* WINED3DSIH_TEXDP3 */ pshader_hw_texdp3,
2236 /* WINED3DSIH_TEXDP3TEX */ pshader_hw_texdp3tex,
2237 /* WINED3DSIH_TEXKILL */ pshader_hw_texkill,
2238 /* WINED3DSIH_TEXLDD */ NULL,
2239 /* WINED3DSIH_TEXLDL */ NULL,
2240 /* WINED3DSIH_TEXM3x2DEPTH */ pshader_hw_texm3x2depth,
2241 /* WINED3DSIH_TEXM3x2PAD */ pshader_hw_texm3x2pad,
2242 /* WINED3DSIH_TEXM3x2TEX */ pshader_hw_texm3x2tex,
2243 /* WINED3DSIH_TEXM3x3 */ pshader_hw_texm3x3,
2244 /* WINED3DSIH_TEXM3x3DIFF */ NULL,
2245 /* WINED3DSIH_TEXM3x3PAD */ pshader_hw_texm3x3pad,
2246 /* WINED3DSIH_TEXM3x3SPEC */ pshader_hw_texm3x3spec,
2247 /* WINED3DSIH_TEXM3x3TEX */ pshader_hw_texm3x3tex,
2248 /* WINED3DSIH_TEXM3x3VSPEC */ pshader_hw_texm3x3vspec,
2249 /* WINED3DSIH_TEXREG2AR */ pshader_hw_texreg2ar,
2250 /* WINED3DSIH_TEXREG2GB */ pshader_hw_texreg2gb,
2251 /* WINED3DSIH_TEXREG2RGB */ pshader_hw_texreg2rgb,
2252 };
2253
2254 const shader_backend_t arb_program_shader_backend = {
2255 shader_arb_instruction_handler_table,
2256 shader_arb_select,
2257 shader_arb_select_depth_blt,
2258 shader_arb_deselect_depth_blt,
2259 shader_arb_update_float_vertex_constants,
2260 shader_arb_update_float_pixel_constants,
2261 shader_arb_load_constants,
2262 shader_arb_destroy,
2263 shader_arb_alloc,
2264 shader_arb_free,
2265 shader_arb_dirty_const,
2266 shader_arb_generate_pshader,
2267 shader_arb_generate_vshader,
2268 shader_arb_get_caps,
2269 shader_arb_color_fixup_supported,
2270 };
2271
2272 /* ARB_fragment_program fixed function pipeline replacement definitions */
2273 #define ARB_FFP_CONST_TFACTOR 0
2274 #define ARB_FFP_CONST_SPECULAR_ENABLE ((ARB_FFP_CONST_TFACTOR) + 1)
2275 #define ARB_FFP_CONST_CONSTANT(i) ((ARB_FFP_CONST_SPECULAR_ENABLE) + 1 + i)
2276 #define ARB_FFP_CONST_BUMPMAT(i) ((ARB_FFP_CONST_CONSTANT(7)) + 1 + i)
2277 #define ARB_FFP_CONST_LUMINANCE(i) ((ARB_FFP_CONST_BUMPMAT(7)) + 1 + i)
2278
2279 struct arbfp_ffp_desc
2280 {
2281 struct ffp_frag_desc parent;
2282 GLuint shader;
2283 unsigned int num_textures_used;
2284 };
2285
2286 static void arbfp_enable(IWineD3DDevice *iface, BOOL enable) {
2287 if(enable) {
2288 glEnable(GL_FRAGMENT_PROGRAM_ARB);
2289 checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB)");
2290 } else {
2291 glDisable(GL_FRAGMENT_PROGRAM_ARB);
2292 checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
2293 }
2294 }
2295
2296 static HRESULT arbfp_alloc(IWineD3DDevice *iface) {
2297 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
2298 struct shader_arb_priv *priv;
2299 /* Share private data between the shader backend and the pipeline replacement, if both
2300 * are the arb implementation. This is needed to figure out whether ARBfp should be disabled
2301 * if no pixel shader is bound or not
2302 */
2303 if(This->shader_backend == &arb_program_shader_backend) {
2304 This->fragment_priv = This->shader_priv;
2305 } else {
2306 This->fragment_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct shader_arb_priv));
2307 if(!This->fragment_priv) return E_OUTOFMEMORY;
2308 }
2309 priv = This->fragment_priv;
2310 priv->fragment_shaders = hash_table_create(ffp_frag_program_key_hash, ffp_frag_program_key_compare);
2311 priv->use_arbfp_fixed_func = TRUE;
2312 return WINED3D_OK;
2313 }
2314
2315 static void arbfp_free_ffpshader(void *value, void *gli) {
2316 const WineD3D_GL_Info *gl_info = gli;
2317 struct arbfp_ffp_desc *entry_arb = value;
2318
2319 ENTER_GL();
2320 GL_EXTCALL(glDeleteProgramsARB(1, &entry_arb->shader));
2321 checkGLcall("glDeleteProgramsARB(1, &entry_arb->shader)");
2322 HeapFree(GetProcessHeap(), 0, entry_arb);
2323 LEAVE_GL();
2324 }
2325
2326 static void arbfp_free(IWineD3DDevice *iface) {
2327 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
2328 struct shader_arb_priv *priv = This->fragment_priv;
2329
2330 hash_table_destroy(priv->fragment_shaders, arbfp_free_ffpshader, &This->adapter->gl_info);
2331 priv->use_arbfp_fixed_func = FALSE;
2332
2333 if(This->shader_backend != &arb_program_shader_backend) {
2334 HeapFree(GetProcessHeap(), 0, This->fragment_priv);
2335 }
2336 }
2337
2338 static void arbfp_get_caps(WINED3DDEVTYPE devtype, const WineD3D_GL_Info *gl_info, struct fragment_caps *caps)
2339 {
2340 caps->TextureOpCaps = WINED3DTEXOPCAPS_DISABLE |
2341 WINED3DTEXOPCAPS_SELECTARG1 |
2342 WINED3DTEXOPCAPS_SELECTARG2 |
2343 WINED3DTEXOPCAPS_MODULATE4X |
2344 WINED3DTEXOPCAPS_MODULATE2X |
2345 WINED3DTEXOPCAPS_MODULATE |
2346 WINED3DTEXOPCAPS_ADDSIGNED2X |
2347 WINED3DTEXOPCAPS_ADDSIGNED |
2348 WINED3DTEXOPCAPS_ADD |
2349 WINED3DTEXOPCAPS_SUBTRACT |
2350 WINED3DTEXOPCAPS_ADDSMOOTH |
2351 WINED3DTEXOPCAPS_BLENDCURRENTALPHA |
2352 WINED3DTEXOPCAPS_BLENDFACTORALPHA |
2353 WINED3DTEXOPCAPS_BLENDTEXTUREALPHA |
2354 WINED3DTEXOPCAPS_BLENDDIFFUSEALPHA |
2355 WINED3DTEXOPCAPS_BLENDTEXTUREALPHAPM |
2356 WINED3DTEXOPCAPS_MODULATEALPHA_ADDCOLOR |
2357 WINED3DTEXOPCAPS_MODULATECOLOR_ADDALPHA |
2358 WINED3DTEXOPCAPS_MODULATEINVCOLOR_ADDALPHA |
2359 WINED3DTEXOPCAPS_MODULATEINVALPHA_ADDCOLOR |
2360 WINED3DTEXOPCAPS_DOTPRODUCT3 |
2361 WINED3DTEXOPCAPS_MULTIPLYADD |
2362 WINED3DTEXOPCAPS_LERP |
2363 WINED3DTEXOPCAPS_BUMPENVMAP |
2364 WINED3DTEXOPCAPS_BUMPENVMAPLUMINANCE;
2365
2366 /* TODO: Implement WINED3DTEXOPCAPS_PREMODULATE */
2367
2368 caps->MaxTextureBlendStages = 8;
2369 caps->MaxSimultaneousTextures = min(GL_LIMITS(fragment_samplers), 8);
2370
2371 caps->PrimitiveMiscCaps |= WINED3DPMISCCAPS_TSSARGTEMP;
2372 }
2373 #undef GLINFO_LOCATION
2374
2375 #define GLINFO_LOCATION stateblock->wineD3DDevice->adapter->gl_info
2376 static void state_texfactor_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2377 float col[4];
2378 IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2379
2380 /* Don't load the parameter if we're using an arbfp pixel shader, otherwise we'll overwrite
2381 * application provided constants
2382 */
2383 if(device->shader_backend == &arb_program_shader_backend) {
2384 if (use_ps(stateblock)) return;
2385
2386 device = stateblock->wineD3DDevice;
2387 device->activeContext->pshader_const_dirty[ARB_FFP_CONST_TFACTOR] = 1;
2388 device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_TFACTOR + 1);
2389 }
2390
2391 D3DCOLORTOGLFLOAT4(stateblock->renderState[WINED3DRS_TEXTUREFACTOR], col);
2392 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_TFACTOR, col));
2393 checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_TFACTOR, col)");
2394
2395 }
2396
2397 static void state_arb_specularenable(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2398 float col[4];
2399 IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2400
2401 /* Don't load the parameter if we're using an arbfp pixel shader, otherwise we'll overwrite
2402 * application provided constants
2403 */
2404 if(device->shader_backend == &arb_program_shader_backend) {
2405 if (use_ps(stateblock)) return;
2406
2407 device = stateblock->wineD3DDevice;
2408 device->activeContext->pshader_const_dirty[ARB_FFP_CONST_SPECULAR_ENABLE] = 1;
2409 device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_SPECULAR_ENABLE + 1);
2410 }
2411
2412 if(stateblock->renderState[WINED3DRS_SPECULARENABLE]) {
2413 /* The specular color has no alpha */
2414 col[0] = 1.0; col[1] = 1.0;
2415 col[2] = 1.0; col[3] = 0.0;
2416 } else {
2417 col[0] = 0.0; col[1] = 0.0;
2418 col[2] = 0.0; col[3] = 0.0;
2419 }
2420 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_SPECULAR_ENABLE, col));
2421 checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_SPECULAR_ENABLE, col)");
2422 }
2423
2424 static void set_bumpmat_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2425 DWORD stage = (state - STATE_TEXTURESTAGE(0, 0)) / (WINED3D_HIGHEST_TEXTURE_STATE + 1);
2426 IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2427 float mat[2][2];
2428
2429 if (use_ps(stateblock))
2430 {
2431 if(stage != 0 &&
2432 ((IWineD3DPixelShaderImpl *) stateblock->pixelShader)->baseShader.reg_maps.bumpmat[stage]) {
2433 /* The pixel shader has to know the bump env matrix. Do a constants update if it isn't scheduled
2434 * anyway
2435 */
2436 if(!isStateDirty(context, STATE_PIXELSHADERCONSTANT)) {
2437 device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
2438 }
2439 }
2440
2441 if(device->shader_backend == &arb_program_shader_backend) {
2442 /* Exit now, don't set the bumpmat below, otherwise we may overwrite pixel shader constants */
2443 return;
2444 }
2445 } else if(device->shader_backend == &arb_program_shader_backend) {
2446 device->activeContext->pshader_const_dirty[ARB_FFP_CONST_BUMPMAT(stage)] = 1;
2447 device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_BUMPMAT(stage) + 1);
2448 }
2449
2450 mat[0][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT00]);
2451 mat[0][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT01]);
2452 mat[1][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT10]);
2453 mat[1][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT11]);
2454
2455 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_BUMPMAT(stage), &mat[0][0]));
2456 checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_BUMPMAT(stage), &mat[0][0])");
2457 }
2458
2459 static void tex_bumpenvlum_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2460 DWORD stage = (state - STATE_TEXTURESTAGE(0, 0)) / (WINED3D_HIGHEST_TEXTURE_STATE + 1);
2461 IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2462 float param[4];
2463
2464 if (use_ps(stateblock))
2465 {
2466 if(stage != 0 &&
2467 ((IWineD3DPixelShaderImpl *) stateblock->pixelShader)->baseShader.reg_maps.luminanceparams[stage]) {
2468 /* The pixel shader has to know the luminance offset. Do a constants update if it
2469 * isn't scheduled anyway
2470 */
2471 if(!isStateDirty(context, STATE_PIXELSHADERCONSTANT)) {
2472 device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
2473 }
2474 }
2475
2476 if(device->shader_backend == &arb_program_shader_backend) {
2477 /* Exit now, don't set the bumpmat below, otherwise we may overwrite pixel shader constants */
2478 return;
2479 }
2480 } else if(device->shader_backend == &arb_program_shader_backend) {
2481 device->activeContext->pshader_const_dirty[ARB_FFP_CONST_LUMINANCE(stage)] = 1;
2482 device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_LUMINANCE(stage) + 1);
2483 }
2484
2485 param[0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVLSCALE]);
2486 param[1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVLOFFSET]);
2487 param[2] = 0.0;
2488 param[3] = 0.0;
2489
2490 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_LUMINANCE(stage), param));
2491 checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_LUMINANCE(stage), param)");
2492 }
2493
2494 static const char *get_argreg(SHADER_BUFFER *buffer, DWORD argnum, unsigned int stage, DWORD arg) {
2495 const char *ret;
2496
2497 if(arg == ARG_UNUSED) return "unused"; /* This is the marker for unused registers */
2498
2499 switch(arg & WINED3DTA_SELECTMASK) {
2500 case WINED3DTA_DIFFUSE:
2501 ret = "fragment.color.primary"; break;
2502
2503 case WINED3DTA_CURRENT:
2504 if(stage == 0) ret = "fragment.color.primary";
2505 else ret = "ret";
2506 break;
2507
2508 case WINED3DTA_TEXTURE:
2509 switch(stage) {
2510 case 0: ret = "tex0"; break;
2511 case 1: ret = "tex1"; break;
2512 case 2: ret = "tex2"; break;
2513 case 3: ret = "tex3"; break;
2514 case 4: ret = "tex4"; break;
2515 case 5: ret = "tex5"; break;
2516 case 6: ret = "tex6"; break;
2517 case 7: ret = "tex7"; break;
2518 default: ret = "unknown texture";
2519 }
2520 break;
2521
2522 case WINED3DTA_TFACTOR:
2523 ret = "tfactor"; break;
2524
2525 case WINED3DTA_SPECULAR:
2526 ret = "fragment.color.secondary"; break;
2527
2528 case WINED3DTA_TEMP:
2529 ret = "tempreg"; break;
2530
2531 case WINED3DTA_CONSTANT:
2532 FIXME("Implement perstage constants\n");
2533 switch(stage) {
2534 case 0: ret = "const0"; break;
2535 case 1: ret = "const1"; break;
2536 case 2: ret = "const2"; break;
2537 case 3: ret = "const3"; break;
2538 case 4: ret = "const4"; break;
2539 case 5: ret = "const5"; break;
2540 case 6: ret = "const6"; break;
2541 case 7: ret = "const7"; break;
2542 default: ret = "unknown constant";
2543 }
2544 break;
2545
2546 default:
2547 return "unknown";
2548 }
2549
2550 if(arg & WINED3DTA_COMPLEMENT) {
2551 shader_addline(buffer, "SUB arg%u, const.x, %s;\n", argnum, ret);
2552 if(argnum == 0) ret = "arg0";
2553 if(argnum == 1) ret = "arg1";
2554 if(argnum == 2) ret = "arg2";
2555 }
2556 if(arg & WINED3DTA_ALPHAREPLICATE) {
2557 shader_addline(buffer, "MOV arg%u, %s.w;\n", argnum, ret);
2558 if(argnum == 0) ret = "arg0";
2559 if(argnum == 1) ret = "arg1";
2560 if(argnum == 2) ret = "arg2";
2561 }
2562 return ret;
2563 }
2564
2565 static void gen_ffp_instr(SHADER_BUFFER *buffer, unsigned int stage, BOOL color, BOOL alpha,
2566 DWORD dst, DWORD op, DWORD dw_arg0, DWORD dw_arg1, DWORD dw_arg2) {
2567 const char *dstmask, *dstreg, *arg0, *arg1, *arg2;
2568 unsigned int mul = 1;
2569 BOOL mul_final_dest = FALSE;
2570
2571 if(color && alpha) dstmask = "";
2572 else if(color) dstmask = ".xyz";
2573 else dstmask = ".w";
2574
2575 if(dst == tempreg) dstreg = "tempreg";
2576 else dstreg = "ret";
2577
2578 arg0 = get_argreg(buffer, 0, stage, dw_arg0);
2579 arg1 = get_argreg(buffer, 1, stage, dw_arg1);
2580 arg2 = get_argreg(buffer, 2, stage, dw_arg2);
2581
2582 switch(op) {
2583 case WINED3DTOP_DISABLE:
2584 if(stage == 0) shader_addline(buffer, "MOV %s%s, fragment.color.primary;\n", dstreg, dstmask);
2585 break;
2586
2587 case WINED3DTOP_SELECTARG2:
2588 arg1 = arg2;
2589 case WINED3DTOP_SELECTARG1:
2590 shader_addline(buffer, "MOV %s%s, %s;\n", dstreg, dstmask, arg1);
2591 break;
2592
2593 case WINED3DTOP_MODULATE4X:
2594 mul = 2;
2595 case WINED3DTOP_MODULATE2X:
2596 mul *= 2;
2597 if(strcmp(dstreg, "result.color") == 0) {
2598 dstreg = "ret";
2599 mul_final_dest = TRUE;
2600 }
2601 case WINED3DTOP_MODULATE:
2602 shader_addline(buffer, "MUL %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2603 break;
2604
2605 case WINED3DTOP_ADDSIGNED2X:
2606 mul = 2;
2607 if(strcmp(dstreg, "result.color") == 0) {
2608 dstreg = "ret";
2609 mul_final_dest = TRUE;
2610 }
2611 case WINED3DTOP_ADDSIGNED:
2612 shader_addline(buffer, "SUB arg2, %s, const.w;\n", arg2);
2613 arg2 = "arg2";
2614 case WINED3DTOP_ADD:
2615 shader_addline(buffer, "ADD_SAT %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2616 break;
2617
2618 case WINED3DTOP_SUBTRACT:
2619 shader_addline(buffer, "SUB_SAT %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2620 break;
2621
2622 case WINED3DTOP_ADDSMOOTH:
2623 shader_addline(buffer, "SUB arg1, const.x, %s;\n", arg1);
2624 shader_addline(buffer, "MAD_SAT %s%s, arg1, %s, %s;\n", dstreg, dstmask, arg2, arg1);
2625 break;
2626
2627 case WINED3DTOP_BLENDCURRENTALPHA:
2628 arg0 = get_argreg(buffer, 0, stage, WINED3DTA_CURRENT);
2629 shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2630 break;
2631 case WINED3DTOP_BLENDFACTORALPHA:
2632 arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TFACTOR);
2633 shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2634 break;
2635 case WINED3DTOP_BLENDTEXTUREALPHA:
2636 arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TEXTURE);
2637 shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2638 break;
2639 case WINED3DTOP_BLENDDIFFUSEALPHA:
2640 arg0 = get_argreg(buffer, 0, stage, WINED3DTA_DIFFUSE);
2641 shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2642 break;
2643
2644 case WINED3DTOP_BLENDTEXTUREALPHAPM:
2645 arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TEXTURE);
2646 shader_addline(buffer, "SUB arg0.w, const.x, %s.w;\n", arg0);
2647 shader_addline(buffer, "MAD_SAT %s%s, %s, arg0.w, %s;\n", dstreg, dstmask, arg2, arg1);
2648 break;
2649
2650 /* D3DTOP_PREMODULATE ???? */
2651
2652 case WINED3DTOP_MODULATEINVALPHA_ADDCOLOR:
2653 shader_addline(buffer, "SUB arg0.w, const.x, %s;\n", arg1);
2654 shader_addline(buffer, "MAD_SAT %s%s, arg0.w, %s, %s;\n", dstreg, dstmask, arg2, arg1);
2655 break;
2656 case WINED3DTOP_MODULATEALPHA_ADDCOLOR:
2657 shader_addline(buffer, "MAD_SAT %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg1, arg2, arg1);
2658 break;
2659 case WINED3DTOP_MODULATEINVCOLOR_ADDALPHA:
2660 shader_addline(buffer, "SUB arg0, const.x, %s;\n", arg1);
2661 shader_addline(buffer, "MAD_SAT %s%s, arg0, %s, %s.w;\n", dstreg, dstmask, arg2, arg1);
2662 break;
2663 case WINED3DTOP_MODULATECOLOR_ADDALPHA:
2664 shader_addline(buffer, "MAD_SAT %s%s, %s, %s, %s.w;\n", dstreg, dstmask, arg1, arg2, arg1);
2665 break;
2666
2667 case WINED3DTOP_DOTPRODUCT3:
2668 mul = 4;
2669 if(strcmp(dstreg, "result.color") == 0) {
2670 dstreg = "ret";
2671 mul_final_dest = TRUE;
2672 }
2673 shader_addline(buffer, "SUB arg1, %s, const.w;\n", arg1);
2674 shader_addline(buffer, "SUB arg2, %s, const.w;\n", arg2);
2675 shader_addline(buffer, "DP3_SAT %s%s, arg1, arg2;\n", dstreg, dstmask);
2676 break;
2677
2678 case WINED3DTOP_MULTIPLYADD:
2679 shader_addline(buffer, "MAD_SAT %s%s, %s, %s, %s;\n", dstreg, dstmask, arg1, arg2, arg0);
2680 break;
2681
2682 case WINED3DTOP_LERP:
2683 /* The msdn is not quite right here */
2684 shader_addline(buffer, "LRP %s%s, %s, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2685 break;
2686
2687 case WINED3DTOP_BUMPENVMAP:
2688 case WINED3DTOP_BUMPENVMAPLUMINANCE:
2689 /* Those are handled in the first pass of the shader(generation pass 1 and 2) already */
2690 break;
2691
2692 default:
2693 FIXME("Unhandled texture op %08x\n", op);
2694 }
2695
2696 if(mul == 2) {
2697 shader_addline(buffer, "MUL_SAT %s%s, %s, const.y;\n", mul_final_dest ? "result.color" : dstreg, dstmask, dstreg);
2698 } else if(mul == 4) {
2699 shader_addline(buffer, "MUL_SAT %s%s, %s, const.z;\n", mul_final_dest ? "result.color" : dstreg, dstmask, dstreg);
2700 }
2701 }
2702
2703 /* The stateblock is passed for GLINFO_LOCATION */
2704 static GLuint gen_arbfp_ffp_shader(const struct ffp_frag_settings *settings, IWineD3DStateBlockImpl *stateblock)
2705 {
2706 unsigned int stage;
2707 SHADER_BUFFER buffer;
2708 BOOL tex_read[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2709 BOOL bump_used[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2710 BOOL luminance_used[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2711 const char *textype;
2712 const char *instr, *sat;
2713 char colorcor_dst[8];
2714 GLuint ret;
2715 DWORD arg0, arg1, arg2;
2716 BOOL tempreg_used = FALSE, tfactor_used = FALSE;
2717 BOOL op_equal;
2718 const char *final_combiner_src = "ret";
2719
2720 /* Find out which textures are read */
2721 for(stage = 0; stage < MAX_TEXTURES; stage++) {
2722 if(settings->op[stage].cop == WINED3DTOP_DISABLE) break;
2723 arg0 = settings->op[stage].carg0 & WINED3DTA_SELECTMASK;
2724 arg1 = settings->op[stage].carg1 & WINED3DTA_SELECTMASK;
2725 arg2 = settings->op[stage].carg2 & WINED3DTA_SELECTMASK;
2726 if(arg0 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2727 if(arg1 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2728 if(arg2 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2729
2730 if(settings->op[stage].cop == WINED3DTOP_BLENDTEXTUREALPHA) tex_read[stage] = TRUE;
2731 if(settings->op[stage].cop == WINED3DTOP_BLENDTEXTUREALPHAPM) tex_read[stage] = TRUE;
2732 if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAP) {
2733 bump_used[stage] = TRUE;
2734 tex_read[stage] = TRUE;
2735 }
2736 if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2737 bump_used[stage] = TRUE;
2738 tex_read[stage] = TRUE;
2739 luminance_used[stage] = TRUE;
2740 } else if(settings->op[stage].cop == WINED3DTOP_BLENDFACTORALPHA) {
2741 tfactor_used = TRUE;
2742 }
2743
2744 if(arg0 == WINED3DTA_TFACTOR || arg1 == WINED3DTA_TFACTOR || arg2 == WINED3DTA_TFACTOR) {
2745 tfactor_used = TRUE;
2746 }
2747
2748 if(settings->op[stage].dst == tempreg) tempreg_used = TRUE;
2749 if(arg0 == WINED3DTA_TEMP || arg1 == WINED3DTA_TEMP || arg2 == WINED3DTA_TEMP) {
2750 tempreg_used = TRUE;
2751 }
2752
2753 if(settings->op[stage].aop == WINED3DTOP_DISABLE) continue;
2754 arg0 = settings->op[stage].aarg0 & WINED3DTA_SELECTMASK;
2755 arg1 = settings->op[stage].aarg1 & WINED3DTA_SELECTMASK;
2756 arg2 = settings->op[stage].aarg2 & WINED3DTA_SELECTMASK;
2757 if(arg0 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2758 if(arg1 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2759 if(arg2 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2760
2761 if(arg0 == WINED3DTA_TEMP || arg1 == WINED3DTA_TEMP || arg2 == WINED3DTA_TEMP) {
2762 tempreg_used = TRUE;
2763 }
2764 if(arg0 == WINED3DTA_TFACTOR || arg1 == WINED3DTA_TFACTOR || arg2 == WINED3DTA_TFACTOR) {
2765 tfactor_used = TRUE;
2766 }
2767 }
2768
2769 /* Shader header */
2770 shader_buffer_init(&buffer);
2771
2772 shader_addline(&buffer, "!!ARBfp1.0\n");
2773
2774 switch(settings->fog) {
2775 case FOG_OFF: break;
2776 case FOG_LINEAR: shader_addline(&buffer, "OPTION ARB_fog_linear;\n"); break;
2777 case FOG_EXP: shader_addline(&buffer, "OPTION ARB_fog_exp;\n"); break;
2778 case FOG_EXP2: shader_addline(&buffer, "OPTION ARB_fog_exp2;\n"); break;
2779 default: FIXME("Unexpected fog setting %d\n", settings->fog);
2780 }
2781
2782 shader_addline(&buffer, "PARAM const = {1, 2, 4, 0.5};\n");
2783 shader_addline(&buffer, "TEMP TMP;\n");
2784 shader_addline(&buffer, "TEMP ret;\n");
2785 if(tempreg_used || settings->sRGB_write) shader_addline(&buffer, "TEMP tempreg;\n");
2786 shader_addline(&buffer, "TEMP arg0;\n");
2787 shader_addline(&buffer, "TEMP arg1;\n");
2788 shader_addline(&buffer, "TEMP arg2;\n");
2789 for(stage = 0; stage < MAX_TEXTURES; stage++) {
2790 if(!tex_read[stage]) continue;
2791 shader_addline(&buffer, "TEMP tex%u;\n", stage);
2792 if(!bump_used[stage]) continue;
2793 shader_addline(&buffer, "PARAM bumpmat%u = program.env[%u];\n", stage, ARB_FFP_CONST_BUMPMAT(stage));
2794 if(!luminance_used[stage]) continue;
2795 shader_addline(&buffer, "PARAM luminance%u = program.env[%u];\n", stage, ARB_FFP_CONST_LUMINANCE(stage));
2796 }
2797 if(tfactor_used) {
2798 shader_addline(&buffer, "PARAM tfactor = program.env[%u];\n", ARB_FFP_CONST_TFACTOR);
2799 }
2800 shader_addline(&buffer, "PARAM specular_enable = program.env[%u];\n", ARB_FFP_CONST_SPECULAR_ENABLE);
2801
2802 if(settings->sRGB_write) {
2803 shader_addline(&buffer, "PARAM srgb_mul_low = {%f, %f, %f, 1.0};\n",
2804 srgb_mul_low, srgb_mul_low, srgb_mul_low);
2805 shader_addline(&buffer, "PARAM srgb_comparison = {%f, %f, %f, %f};\n",
2806 srgb_cmp, srgb_cmp, srgb_cmp, srgb_cmp);
2807 shader_addline(&buffer, "PARAM srgb_pow = {%f, %f, %f, 1.0};\n",
2808 srgb_pow, srgb_pow, srgb_pow);
2809 shader_addline(&buffer, "PARAM srgb_mul_hi = {%f, %f, %f, 1.0};\n",
2810 srgb_mul_high, srgb_mul_high, srgb_mul_high);
2811 shader_addline(&buffer, "PARAM srgb_sub_hi = {%f, %f, %f, 0.0};\n",
2812 srgb_sub_high, srgb_sub_high, srgb_sub_high);
2813 }
2814
2815 /* Generate texture sampling instructions) */
2816 for(stage = 0; stage < MAX_TEXTURES && settings->op[stage].cop != WINED3DTOP_DISABLE; stage++) {
2817 if(!tex_read[stage]) continue;
2818
2819 switch(settings->op[stage].tex_type) {
2820 case tex_1d: textype = "1D"; break;
2821 case tex_2d: textype = "2D"; break;
2822 case tex_3d: textype = "3D"; break;
2823 case tex_cube: textype = "CUBE"; break;
2824 case tex_rect: textype = "RECT"; break;
2825 default: textype = "unexpected_textype"; break;
2826 }
2827
2828 if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAP ||
2829 settings->op[stage].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2830 sat = "";
2831 } else {
2832 sat = "_SAT";
2833 }
2834
2835 if(settings->op[stage].projected == proj_none) {
2836 instr = "TEX";
2837 } else if(settings->op[stage].projected == proj_count4 ||
2838 settings->op[stage].projected == proj_count3) {
2839 instr = "TXP";
2840 } else {
2841 FIXME("Unexpected projection mode %d\n", settings->op[stage].projected);
2842 instr = "TXP";
2843 }
2844
2845 if(stage > 0 &&
2846 (settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAP ||
2847 settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAPLUMINANCE)) {
2848 shader_addline(&buffer, "SWZ arg1, bumpmat%u, x, z, 0, 0;\n", stage - 1);
2849 shader_addline(&buffer, "DP3 ret.x, arg1, tex%u;\n", stage - 1);
2850 shader_addline(&buffer, "SWZ arg1, bumpmat%u, y, w, 0, 0;\n", stage - 1);
2851 shader_addline(&buffer, "DP3 ret.y, arg1, tex%u;\n", stage - 1);
2852
2853 /* with projective textures, texbem only divides the static texture coord, not the displacement,
2854 * so multiply the displacement with the dividing parameter before passing it to TXP
2855 */
2856 if (settings->op[stage].projected != proj_none) {
2857 if(settings->op[stage].projected == proj_count4) {
2858 shader_addline(&buffer, "MOV ret.w, fragment.texcoord[%u].w;\n", stage);
2859 shader_addline(&buffer, "MUL ret.xyz, ret, fragment.texcoord[%u].w, fragment.texcoord[%u];\n", stage, stage);
2860 } else {
2861 shader_addline(&buffer, "MOV ret.w, fragment.texcoord[%u].z;\n", stage);
2862 shader_addline(&buffer, "MAD ret.xyz, ret, fragment.texcoord[%u].z, fragment.texcoord[%u];\n", stage, stage);
2863 }
2864 } else {
2865 shader_addline(&buffer, "ADD ret, ret, fragment.texcoord[%u];\n", stage);
2866 }
2867
2868 shader_addline(&buffer, "%s%s tex%u, ret, texture[%u], %s;\n",
2869 instr, sat, stage, stage, textype);
2870 if(settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2871 shader_addline(&buffer, "MAD_SAT ret.x, tex%u.z, luminance%u.x, luminance%u.y;\n",
2872 stage - 1, stage - 1, stage - 1);
2873 shader_addline(&buffer, "MUL tex%u, tex%u, ret.x;\n", stage, stage);
2874 }
2875 } else if(settings->op[stage].projected == proj_count3) {
2876 shader_addline(&buffer, "MOV ret, fragment.texcoord[%u];\n", stage);
2877 shader_addline(&buffer, "MOV ret.w, ret.z;\n");
2878 shader_addline(&buffer, "%s%s tex%u, ret, texture[%u], %s;\n",
2879 instr, sat, stage, stage, textype);
2880 } else {
2881 shader_addline(&buffer, "%s%s tex%u, fragment.texcoord[%u], texture[%u], %s;\n",
2882 instr, sat, stage, stage, stage, textype);
2883 }
2884
2885 sprintf(colorcor_dst, "tex%u", stage);
2886 gen_color_correction(&buffer, colorcor_dst, WINED3DSP_WRITEMASK_ALL, "const.x", "const.y",
2887 settings->op[stage].color_fixup);
2888 }
2889
2890 /* Generate the main shader */
2891 for(stage = 0; stage < MAX_TEXTURES; stage++) {
2892 if(settings->op[stage].cop == WINED3DTOP_DISABLE) {
2893 if(stage == 0) {
2894 final_combiner_src = "fragment.color.primary";
2895 }
2896 break;
2897 }
2898
2899 if(settings->op[stage].cop == WINED3DTOP_SELECTARG1 &&
2900 settings->op[stage].aop == WINED3DTOP_SELECTARG1) {
2901 op_equal = settings->op[stage].carg1 == settings->op[stage].aarg1;
2902 } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG1 &&
2903 settings->op[stage].aop == WINED3DTOP_SELECTARG2) {
2904 op_equal = settings->op[stage].carg1 == settings->op[stage].aarg2;
2905 } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG2 &&
2906 settings->op[stage].aop == WINED3DTOP_SELECTARG1) {
2907 op_equal = settings->op[stage].carg2 == settings->op[stage].aarg1;
2908 } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG2 &&
2909 settings->op[stage].aop == WINED3DTOP_SELECTARG2) {
2910 op_equal = settings->op[stage].carg2 == settings->op[stage].aarg2;
2911 } else {
2912 op_equal = settings->op[stage].aop == settings->op[stage].cop &&
2913 settings->op[stage].carg0 == settings->op[stage].aarg0 &&
2914 settings->op[stage].carg1 == settings->op[stage].aarg1 &&
2915 settings->op[stage].carg2 == settings->op[stage].aarg2;
2916 }
2917
2918 if(settings->op[stage].aop == WINED3DTOP_DISABLE) {
2919 gen_ffp_instr(&buffer, stage, TRUE, FALSE, settings->op[stage].dst,
2920 settings->op[stage].cop, settings->op[stage].carg0,
2921 settings->op[stage].carg1, settings->op[stage].carg2);
2922 if(stage == 0) {
2923 shader_addline(&buffer, "MOV ret.w, fragment.color.primary.w;\n");
2924 }
2925 } else if(op_equal) {
2926 gen_ffp_instr(&buffer, stage, TRUE, TRUE, settings->op[stage].dst,
2927 settings->op[stage].cop, settings->op[stage].carg0,
2928 settings->op[stage].carg1, settings->op[stage].carg2);
2929 } else {
2930 gen_ffp_instr(&buffer, stage, TRUE, FALSE, settings->op[stage].dst,
2931 settings->op[stage].cop, settings->op[stage].carg0,
2932 settings->op[stage].carg1, settings->op[stage].carg2);
2933 gen_ffp_instr(&buffer, stage, FALSE, TRUE, settings->op[stage].dst,
2934 settings->op[stage].aop, settings->op[stage].aarg0,
2935 settings->op[stage].aarg1, settings->op[stage].aarg2);
2936 }
2937 }
2938
2939 if(settings->sRGB_write) {
2940 shader_addline(&buffer, "MAD ret, fragment.color.secondary, specular_enable, %s;\n", final_combiner_src);
2941 arbfp_add_sRGB_correction(&buffer, "ret", "arg0", "arg1", "arg2", "tempreg");
2942 shader_addline(&buffer, "MOV result.color.w, ret.w;\n");
2943 } else {
2944 shader_addline(&buffer, "MAD result.color, fragment.color.secondary, specular_enable, %s;\n", final_combiner_src);
2945 }
2946
2947 /* Footer */
2948 shader_addline(&buffer, "END\n");
2949
2950 /* Generate the shader */
2951 GL_EXTCALL(glGenProgramsARB(1, &ret));
2952 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, ret));
2953 GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(buffer.buffer), buffer.buffer));
2954
2955 if (glGetError() == GL_INVALID_OPERATION) {
2956 GLint pos;
2957 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
2958 FIXME("Fragment program error at position %d: %s\n", pos,
2959 debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
2960 }
2961 shader_buffer_free(&buffer);
2962 return ret;
2963 }
2964
2965 static void fragment_prog_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2966 IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2967 struct shader_arb_priv *priv = device->fragment_priv;
2968 BOOL use_pshader = use_ps(stateblock);
2969 BOOL use_vshader = use_vs(stateblock);
2970 struct ffp_frag_settings settings;
2971 const struct arbfp_ffp_desc *desc;
2972 unsigned int i;
2973
2974 TRACE("state %#x, stateblock %p, context %p\n", state, stateblock, context);
2975
2976 if(isStateDirty(context, STATE_RENDER(WINED3DRS_FOGENABLE))) {
2977 if(!use_pshader && device->shader_backend == &arb_program_shader_backend && context->last_was_pshader) {
2978 /* Reload fixed function constants since they collide with the pixel shader constants */
2979 for(i = 0; i < MAX_TEXTURES; i++) {
2980 set_bumpmat_arbfp(STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT00), stateblock, context);
2981 }
2982 state_texfactor_arbfp(STATE_RENDER(WINED3DRS_TEXTUREFACTOR), stateblock, context);
2983 state_arb_specularenable(STATE_RENDER(WINED3DRS_SPECULARENABLE), stateblock, context);
2984 } else if(use_pshader && !isStateDirty(context, device->StateTable[STATE_VSHADER].representative)) {
2985 device->shader_backend->shader_select((IWineD3DDevice *)stateblock->wineD3DDevice, use_pshader, use_vshader);
2986 }
2987 return;
2988 }
2989
2990 if(!use_pshader) {
2991 /* Find or create a shader implementing the fixed function pipeline settings, then activate it */
2992 gen_ffp_frag_op(stateblock, &settings, FALSE);
2993 desc = (const struct arbfp_ffp_desc *)find_ffp_frag_shader(priv->fragment_shaders, &settings);
2994 if(!desc) {
2995 struct arbfp_ffp_desc *new_desc = HeapAlloc(GetProcessHeap(), 0, sizeof(*new_desc));
2996 if (!new_desc)
2997 {
2998 ERR("Out of memory\n");
2999 return;
3000 }
3001 new_desc->num_textures_used = 0;
3002 for(i = 0; i < GL_LIMITS(texture_stages); i++) {
3003 if(settings.op[i].cop == WINED3DTOP_DISABLE) break;
3004 new_desc->num_textures_used = i;
3005 }
3006
3007 memcpy(&new_desc->parent.settings, &settings, sizeof(settings));
3008 new_desc->shader = gen_arbfp_ffp_shader(&settings, stateblock);
3009 add_ffp_frag_shader(priv->fragment_shaders, &new_desc->parent);
3010 TRACE("Allocated fixed function replacement shader descriptor %p\n", new_desc);
3011 desc = new_desc;
3012 }
3013
3014 /* Now activate the replacement program. GL_FRAGMENT_PROGRAM_ARB is already active(however, note the
3015 * comment above the shader_select call below). If e.g. GLSL is active, the shader_select call will
3016 * deactivate it.
3017 */
3018 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, desc->shader));
3019 checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, desc->shader)");
3020 priv->current_fprogram_id = desc->shader;
3021
3022 if(device->shader_backend == &arb_program_shader_backend && context->last_was_pshader) {
3023 /* Reload fixed function constants since they collide with the pixel shader constants */
3024 for(i = 0; i < MAX_TEXTURES; i++) {
3025 set_bumpmat_arbfp(STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT00), stateblock, context);
3026 }
3027 state_texfactor_arbfp(STATE_RENDER(WINED3DRS_TEXTUREFACTOR), stateblock, context);
3028 state_arb_specularenable(STATE_RENDER(WINED3DRS_SPECULARENABLE), stateblock, context);
3029 }
3030 context->last_was_pshader = FALSE;
3031 } else {
3032 context->last_was_pshader = TRUE;
3033 }
3034
3035 /* Finally, select the shader. If a pixel shader is used, it will be set and enabled by the shader backend.
3036 * If this shader backend is arbfp(most likely), then it will simply overwrite the last fixed function replace-
3037 * ment shader. If the shader backend is not ARB, it currently is important that the opengl implementation
3038 * type overwrites GL_ARB_fragment_program. This is currently the case with GLSL. If we really want to use
3039 * atifs or nvrc pixel shaders with arb fragment programs we'd have to disable GL_FRAGMENT_PROGRAM_ARB here
3040 *
3041 * Don't call shader_select if the vertex shader is dirty, because some shader backends(GLSL) need both shaders
3042 * to be compiled before activating them(needs some cleanups in the shader backend interface)
3043 */
3044 if(!isStateDirty(context, device->StateTable[STATE_VSHADER].representative)) {
3045 device->shader_backend->shader_select((IWineD3DDevice *)stateblock->wineD3DDevice, use_pshader, use_vshader);
3046
3047 if (!isStateDirty(context, STATE_VERTEXSHADERCONSTANT) && (use_vshader || use_pshader)) {
3048 device->StateTable[STATE_VERTEXSHADERCONSTANT].apply(STATE_VERTEXSHADERCONSTANT, stateblock, context);
3049 }
3050 }
3051 if(use_pshader) {
3052 device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
3053 }
3054 }
3055
3056 /* We can't link the fog states to the fragment state directly since the vertex pipeline links them
3057 * to FOGENABLE. A different linking in different pipeline parts can't be expressed in the combined
3058 * state table, so we need to handle that with a forwarding function. The other invisible side effect
3059 * is that changing the fog start and fog end(which links to FOGENABLE in vertex) results in the
3060 * fragment_prog_arbfp function being called because FOGENABLE is dirty, which calls this function here
3061 */
3062 static void state_arbfp_fog(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
3063 enum fogsource new_source;
3064
3065 TRACE("state %#x, stateblock %p, context %p\n", state, stateblock, context);
3066
3067 if(!isStateDirty(context, STATE_PIXELSHADER)) {
3068 fragment_prog_arbfp(state, stateblock, context);
3069 }
3070
3071 if(!stateblock->renderState[WINED3DRS_FOGENABLE]) return;
3072
3073 if(stateblock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE) {
3074 if(use_vs(stateblock)) {
3075 new_source = FOGSOURCE_VS;
3076 } else {
3077 if(stateblock->renderState[WINED3DRS_FOGVERTEXMODE] == WINED3DFOG_NONE || context->last_was_rhw) {
3078 new_source = FOGSOURCE_COORD;
3079 } else {
3080 new_source = FOGSOURCE_FFP;
3081 }
3082 }
3083 } else {
3084 new_source = FOGSOURCE_FFP;
3085 }
3086 if(new_source != context->fog_source) {
3087 context->fog_source = new_source;
3088 state_fogstartend(STATE_RENDER(WINED3DRS_FOGSTART), stateblock, context);
3089 }
3090 }
3091
3092 static void textransform(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
3093 if(!isStateDirty(context, STATE_PIXELSHADER)) {
3094 fragment_prog_arbfp(state, stateblock, context);
3095 }
3096 }
3097
3098 #undef GLINFO_LOCATION
3099
3100 static const struct StateEntryTemplate arbfp_fragmentstate_template[] = {
3101 {STATE_RENDER(WINED3DRS_TEXTUREFACTOR), { STATE_RENDER(WINED3DRS_TEXTUREFACTOR), state_texfactor_arbfp }, 0 },
3102 {STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3103 {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3104 {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3105 {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3106 {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3107 {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3108 {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3109 {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3110 {STATE_TEXTURESTAGE(0, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3111 {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3112 {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3113 {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3114 {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3115 {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3116 {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3117 {STATE_TEXTURESTAGE(1, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3118 {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3119 {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3120 {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3121 {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3122 {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3123 {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3124 {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3125 {STATE_TEXTURESTAGE(1, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3126 {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3127 {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3128 {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3129 {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3130 {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3131 {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3132 {STATE_TEXTURESTAGE(2, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3133 {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3134 {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3135 {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3136 {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3137 {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3138 {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3139 {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3140 {STATE_TEXTURESTAGE(2, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3141 {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3142 {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3143 {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3144 {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3145 {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3146 {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3147 {STATE_TEXTURESTAGE(3, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3148 {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3149 {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3150 {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3151 {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3152 {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3153 {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3154 {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3155 {STATE_TEXTURESTAGE(3, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3156 {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3157 {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3158 {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3159 {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3160 {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3161 {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3162 {STATE_TEXTURESTAGE(4, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3163 {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3164 {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3165 {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3166 {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3167 {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3168 {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3169 {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3170 {STATE_TEXTURESTAGE(4, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3171 {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3172 {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3173 {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3174 {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3175 {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3176 {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3177 {STATE_TEXTURESTAGE(5, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3178 {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3179 {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3180 {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3181 {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3182 {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3183 {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3184 {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3185 {STATE_TEXTURESTAGE(5, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3186 {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3187 {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3188 {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3189 {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3190 {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3191 {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3192 {STATE_TEXTURESTAGE(6, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3193 {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3194 {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3195 {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3196 {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3197 {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3198 {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3199 {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3200 {STATE_TEXTURESTAGE(6, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3201 {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3202 {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3203 {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3204 {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3205 {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3206 {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3207 {STATE_TEXTURESTAGE(7, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3208 {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3209 {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3210 {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3211 {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3212 {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3213 {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3214 {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3215 {STATE_TEXTURESTAGE(7, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3216 {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3217 {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3218 {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3219 {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3220 {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3221 {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3222 { STATE_SAMPLER(0), { STATE_SAMPLER(0), sampler_texdim }, 0 },
3223 { STATE_SAMPLER(1), { STATE_SAMPLER(1), sampler_texdim }, 0 },
3224 { STATE_SAMPLER(2), { STATE_SAMPLER(2), sampler_texdim }, 0 },
3225 { STATE_SAMPLER(3), { STATE_SAMPLER(3), sampler_texdim }, 0 },
3226 { STATE_SAMPLER(4), { STATE_SAMPLER(4), sampler_texdim }, 0 },
3227 { STATE_SAMPLER(5), { STATE_SAMPLER(5), sampler_texdim }, 0 },
3228 { STATE_SAMPLER(6), { STATE_SAMPLER(6), sampler_texdim }, 0 },
3229 { STATE_SAMPLER(7), { STATE_SAMPLER(7), sampler_texdim }, 0 },
3230 { STATE_PIXELSHADER, { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3231 { STATE_RENDER(WINED3DRS_FOGENABLE), { STATE_RENDER(WINED3DRS_FOGENABLE), state_arbfp_fog }, 0 },
3232 { STATE_RENDER(WINED3DRS_FOGTABLEMODE), { STATE_RENDER(WINED3DRS_FOGENABLE), state_arbfp_fog }, 0 },
3233 { STATE_RENDER(WINED3DRS_FOGVERTEXMODE), { STATE_RENDER(WINED3DRS_FOGENABLE), state_arbfp_fog }, 0 },
3234 { STATE_RENDER(WINED3DRS_FOGSTART), { STATE_RENDER(WINED3DRS_FOGSTART), state_fogstartend }, 0 },
3235 { STATE_RENDER(WINED3DRS_FOGEND), { STATE_RENDER(WINED3DRS_FOGSTART), state_fogstartend }, 0 },
3236 { STATE_RENDER(WINED3DRS_SRGBWRITEENABLE), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3237 { STATE_RENDER(WINED3DRS_FOGCOLOR), { STATE_RENDER(WINED3DRS_FOGCOLOR), state_fogcolor }, 0 },
3238 { STATE_RENDER(WINED3DRS_FOGDENSITY), { STATE_RENDER(WINED3DRS_FOGDENSITY), state_fogdensity }, 0 },
3239 {STATE_TEXTURESTAGE(0,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(0, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3240 {STATE_TEXTURESTAGE(1,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(1, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3241 {STATE_TEXTURESTAGE(2,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(2, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3242 {STATE_TEXTURESTAGE(3,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(3, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3243 {STATE_TEXTURESTAGE(4,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(4, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3244 {STATE_TEXTURESTAGE(5,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(5, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3245 {STATE_TEXTURESTAGE(6,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(6, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3246 {STATE_TEXTURESTAGE(7,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(7, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3247 { STATE_RENDER(WINED3DRS_SPECULARENABLE), { STATE_RENDER(WINED3DRS_SPECULARENABLE), state_arb_specularenable}, 0 },
3248 {0 /* Terminate */, { 0, 0 }, 0 },
3249 };
3250
3251 const struct fragment_pipeline arbfp_fragment_pipeline = {
3252 arbfp_enable,
3253 arbfp_get_caps,
3254 arbfp_alloc,
3255 arbfp_free,
3256 shader_arb_color_fixup_supported,
3257 arbfp_fragmentstate_template,
3258 TRUE /* We can disable projected textures */
3259 };
3260
3261 #define GLINFO_LOCATION device->adapter->gl_info
3262
3263 struct arbfp_blit_priv {
3264 GLenum yuy2_rect_shader, yuy2_2d_shader;
3265 GLenum uyvy_rect_shader, uyvy_2d_shader;
3266 GLenum yv12_rect_shader, yv12_2d_shader;
3267 };
3268
3269 static HRESULT arbfp_blit_alloc(IWineD3DDevice *iface) {
3270 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3271 device->blit_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct arbfp_blit_priv));
3272 if(!device->blit_priv) {
3273 ERR("Out of memory\n");
3274 return E_OUTOFMEMORY;
3275 }
3276 return WINED3D_OK;
3277 }
3278 static void arbfp_blit_free(IWineD3DDevice *iface) {
3279 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3280 struct arbfp_blit_priv *priv = device->blit_priv;
3281
3282 ENTER_GL();
3283 GL_EXTCALL(glDeleteProgramsARB(1, &priv->yuy2_rect_shader));
3284 GL_EXTCALL(glDeleteProgramsARB(1, &priv->yuy2_2d_shader));
3285 GL_EXTCALL(glDeleteProgramsARB(1, &priv->uyvy_rect_shader));
3286 GL_EXTCALL(glDeleteProgramsARB(1, &priv->uyvy_2d_shader));
3287 GL_EXTCALL(glDeleteProgramsARB(1, &priv->yv12_rect_shader));
3288 GL_EXTCALL(glDeleteProgramsARB(1, &priv->yv12_2d_shader));
3289 checkGLcall("Delete yuv programs\n");
3290 LEAVE_GL();
3291 }
3292
3293 static BOOL gen_planar_yuv_read(SHADER_BUFFER *buffer, enum yuv_fixup yuv_fixup, GLenum textype, char *luminance)
3294 {
3295 char chroma;
3296 const char *tex, *texinstr;
3297
3298 if (yuv_fixup == YUV_FIXUP_UYVY) {
3299 chroma = 'x';
3300 *luminance = 'w';
3301 } else {
3302 chroma = 'w';
3303 *luminance = 'x';
3304 }
3305 switch(textype) {
3306 case GL_TEXTURE_2D: tex = "2D"; texinstr = "TXP"; break;
3307 case GL_TEXTURE_RECTANGLE_ARB: tex = "RECT"; texinstr = "TEX"; break;
3308 default:
3309 /* This is more tricky than just replacing the texture type - we have to navigate
3310 * properly in the texture to find the correct chroma values
3311 */
3312 FIXME("Implement yuv correction for non-2d, non-rect textures\n");
3313 return FALSE;
3314 }
3315
3316 /* First we have to read the chroma values. This means we need at least two pixels(no filtering),
3317 * or 4 pixels(with filtering). To get the unmodified chromas, we have to rid ourselves of the
3318 * filtering when we sample the texture.
3319 *
3320 * These are the rules for reading the chroma:
3321 *
3322 * Even pixel: Cr
3323 * Even pixel: U
3324 * Odd pixel: V
3325 *
3326 * So we have to get the sampling x position in non-normalized coordinates in integers
3327 */
3328 if(textype != GL_TEXTURE_RECTANGLE_ARB) {
3329 shader_addline(buffer, "MUL texcrd.xy, fragment.texcoord[0], size.x;\n");
3330 shader_addline(buffer, "MOV texcrd.w, size.x;\n");
3331 } else {
3332 shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3333 }
3334 /* We must not allow filtering between pixel x and x+1, this would mix U and V
3335 * Vertical filtering is ok. However, bear in mind that the pixel center is at
3336 * 0.5, so add 0.5.
3337 */
3338 shader_addline(buffer, "FLR texcrd.x, texcrd.x;\n");
3339 shader_addline(buffer, "ADD texcrd.x, texcrd.x, coef.y;\n");
3340
3341 /* Divide the x coordinate by 0.5 and get the fraction. This gives 0.25 and 0.75 for the
3342 * even and odd pixels respectively
3343 */
3344 shader_addline(buffer, "MUL texcrd2, texcrd, coef.y;\n");
3345 shader_addline(buffer, "FRC texcrd2, texcrd2;\n");
3346
3347 /* Sample Pixel 1 */
3348 shader_addline(buffer, "%s luminance, texcrd, texture[0], %s;\n", texinstr, tex);
3349
3350 /* Put the value into either of the chroma values */
3351 shader_addline(buffer, "SGE temp.x, texcrd2.x, coef.y;\n");
3352 shader_addline(buffer, "MUL chroma.x, luminance.%c, temp.x;\n", chroma);
3353 shader_addline(buffer, "SLT temp.x, texcrd2.x, coef.y;\n");
3354 shader_addline(buffer, "MUL chroma.y, luminance.%c, temp.x;\n", chroma);
3355
3356 /* Sample pixel 2. If we read an even pixel(SLT above returned 1), sample
3357 * the pixel right to the current one. Otherwise, sample the left pixel.
3358 * Bias and scale the SLT result to -1;1 and add it to the texcrd.x.
3359 */
3360 shader_addline(buffer, "MAD temp.x, temp.x, coef.z, -coef.x;\n");
3361 shader_addline(buffer, "ADD texcrd.x, texcrd, temp.x;\n");
3362 shader_addline(buffer, "%s luminance, texcrd, texture[0], %s;\n", texinstr, tex);
3363
3364 /* Put the value into the other chroma */
3365 shader_addline(buffer, "SGE temp.x, texcrd2.x, coef.y;\n");
3366 shader_addline(buffer, "MAD chroma.y, luminance.%c, temp.x, chroma.y;\n", chroma);
3367 shader_addline(buffer, "SLT temp.x, texcrd2.x, coef.y;\n");
3368 shader_addline(buffer, "MAD chroma.x, luminance.%c, temp.x, chroma.x;\n", chroma);
3369
3370 /* TODO: If filtering is enabled, sample a 2nd pair of pixels left or right of
3371 * the current one and lerp the two U and V values
3372 */
3373
3374 /* This gives the correctly filtered luminance value */
3375 shader_addline(buffer, "TEX luminance, fragment.texcoord[0], texture[0], %s;\n", tex);
3376
3377 return TRUE;
3378 }
3379
3380 static BOOL gen_yv12_read(SHADER_BUFFER *buffer, GLenum textype, char *luminance)
3381 {
3382 const char *tex;
3383
3384 switch(textype) {
3385 case GL_TEXTURE_2D: tex = "2D"; break;
3386 case GL_TEXTURE_RECTANGLE_ARB: tex = "RECT"; break;
3387 default:
3388 FIXME("Implement yv12 correction for non-2d, non-rect textures\n");
3389 return FALSE;
3390 }
3391
3392 /* YV12 surfaces contain a WxH sized luminance plane, followed by a (W/2)x(H/2)
3393 * V and a (W/2)x(H/2) U plane, each with 8 bit per pixel. So the effective
3394 * bitdepth is 12 bits per pixel. Since the U and V planes have only half the
3395 * pitch of the luminance plane, the packing into the gl texture is a bit
3396 * unfortunate. If the whole texture is interpreted as luminance data it looks
3397 * approximately like this:
3398 *
3399 * +----------------------------------+----
3400 * | |
3401 * | |
3402 * | |
3403 * | |
3404 * | | 2
3405 * | LUMINANCE | -
3406 * | | 3
3407 * | |
3408 * | |
3409 * | |
3410 * | |
3411 * +----------------+-----------------+----
3412 * | | |
3413 * | U even rows | U odd rows |
3414 * | | | 1
3415 * +----------------+------------------ -
3416 * | | | 3
3417 * | V even rows | V odd rows |
3418 * | | |
3419 * +----------------+-----------------+----
3420 * | | |
3421 * | 0.5 | 0.5 |
3422 *
3423 * So it appears as if there are 4 chroma images, but in fact the odd rows
3424 * in the chroma images are in the same row as the even ones. So its is
3425 * kinda tricky to read
3426 *
3427 * When reading from rectangle textures, keep in mind that the input y coordinates
3428 * go from 0 to d3d_height, whereas the opengl texture height is 1.5 * d3d_height
3429 */
3430 shader_addline(buffer, "PARAM yv12_coef = {%f, %f, %f, %f};\n",
3431 2.0 / 3.0, 1.0 / 6.0, (2.0 / 3.0) + (1.0 / 6.0), 1.0 / 3.0);
3432
3433 shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3434 /* the chroma planes have only half the width */
3435 shader_addline(buffer, "MUL texcrd.x, texcrd.x, coef.y;\n");
3436
3437 /* The first value is between 2/3 and 5/6th of the texture's height, so scale+bias
3438 * the coordinate. Also read the right side of the image when reading odd lines
3439 *
3440 * Don't forget to clamp the y values in into the range, otherwise we'll get filtering
3441 * bleeding
3442 */
3443 if(textype == GL_TEXTURE_2D) {
3444
3445 shader_addline(buffer, "RCP chroma.w, size.y;\n");
3446
3447 shader_addline(buffer, "MUL texcrd2.y, texcrd.y, size.y;\n");
3448
3449 shader_addline(buffer, "FLR texcrd2.y, texcrd2.y;\n");
3450 shader_addline(buffer, "MAD texcrd.y, texcrd.y, yv12_coef.y, yv12_coef.x;\n");
3451
3452 /* Read odd lines from the right side(add size * 0.5 to the x coordinate */
3453 shader_addline(buffer, "ADD texcrd2.x, texcrd2.y, yv12_coef.y;\n"); /* To avoid 0.5 == 0.5 comparisons */
3454 shader_addline(buffer, "FRC texcrd2.x, texcrd2.x;\n");
3455 shader_addline(buffer, "SGE texcrd2.x, texcrd2.x, coef.y;\n");
3456 shader_addline(buffer, "MAD texcrd.x, texcrd2.x, coef.y, texcrd.x;\n");
3457
3458 /* clamp, keep the half pixel origin in mind */
3459 shader_addline(buffer, "MAD temp.y, coef.y, chroma.w, yv12_coef.x;\n");
3460 shader_addline(buffer, "MAX texcrd.y, temp.y, texcrd.y;\n");
3461 shader_addline(buffer, "MAD temp.y, -coef.y, chroma.w, yv12_coef.z;\n");
3462 shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3463 } else {
3464 /* Read from [size - size+size/4] */
3465 shader_addline(buffer, "FLR texcrd.y, texcrd.y;\n");
3466 shader_addline(buffer, "MAD texcrd.y, texcrd.y, coef.w, size.y;\n");
3467
3468 /* Read odd lines from the right side(add size * 0.5 to the x coordinate */
3469 shader_addline(buffer, "ADD texcrd2.x, texcrd.y, yv12_coef.y;\n"); /* To avoid 0.5 == 0.5 comparisons */
3470 shader_addline(buffer, "FRC texcrd2.x, texcrd2.x;\n");
3471 shader_addline(buffer, "SGE texcrd2.x, texcrd2.x, coef.y;\n");
3472 shader_addline(buffer, "MUL texcrd2.x, texcrd2.x, size.x;\n");
3473 shader_addline(buffer, "MAD texcrd.x, texcrd2.x, coef.y, texcrd.x;\n");
3474
3475 /* Make sure to read exactly from the pixel center */
3476 shader_addline(buffer, "FLR texcrd.y, texcrd.y;\n");
3477 shader_addline(buffer, "ADD texcrd.y, texcrd.y, coef.y;\n");
3478
3479 /* Clamp */
3480 shader_addline(buffer, "MAD temp.y, size.y, coef.w, size.y;\n");
3481 shader_addline(buffer, "ADD temp.y, temp.y, -coef.y;\n");
3482 shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3483 shader_addline(buffer, "ADD temp.y, size.y, -coef.y;\n");
3484 shader_addline(buffer, "MAX texcrd.y, temp.y, texcrd.y;\n");
3485 }
3486 /* Read the texture, put the result into the output register */
3487 shader_addline(buffer, "TEX temp, texcrd, texture[0], %s;\n", tex);
3488 shader_addline(buffer, "MOV chroma.x, temp.w;\n");
3489
3490 /* The other chroma value is 1/6th of the texture lower, from 5/6th to 6/6th
3491 * No need to clamp because we're just reusing the already clamped value from above
3492 */
3493 if(textype == GL_TEXTURE_2D) {
3494 shader_addline(buffer, "ADD texcrd.y, texcrd.y, yv12_coef.y;\n");
3495 } else {
3496 shader_addline(buffer, "MAD texcrd.y, size.y, coef.w, texcrd.y;\n");
3497 }
3498 shader_addline(buffer, "TEX temp, texcrd, texture[0], %s;\n", tex);
3499 shader_addline(buffer, "MOV chroma.y, temp.w;\n");
3500
3501 /* Sample the luminance value. It is in the top 2/3rd of the texture, so scale the y coordinate.
3502 * Clamp the y coordinate to prevent the chroma values from bleeding into the sampled luminance
3503 * values due to filtering
3504 */
3505 shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3506 if(textype == GL_TEXTURE_2D) {
3507 /* Multiply the y coordinate by 2/3 and clamp it */
3508 shader_addline(buffer, "MUL texcrd.y, texcrd.y, yv12_coef.x;\n");
3509 shader_addline(buffer, "MAD temp.y, -coef.y, chroma.w, yv12_coef.x;\n");
3510 shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3511 shader_addline(buffer, "TEX luminance, texcrd, texture[0], %s;\n", tex);
3512 } else {
3513 /* Reading from texture_rectangles is pretty straightforward, just use the unmodified
3514 * texture coordinate. It is still a good idea to clamp it though, since the opengl texture
3515 * is bigger
3516 */
3517 shader_addline(buffer, "ADD temp.x, size.y, -coef.y;\n");
3518 shader_addline(buffer, "MIN texcrd.y, texcrd.y, size.x;\n");
3519 shader_addline(buffer, "TEX luminance, texcrd, texture[0], %s;\n", tex);
3520 }
3521 *luminance = 'a';
3522
3523 return TRUE;
3524 }
3525
3526 static GLuint gen_yuv_shader(IWineD3DDeviceImpl *device, enum yuv_fixup yuv_fixup, GLenum textype)
3527 {
3528 GLenum shader;
3529 SHADER_BUFFER buffer;
3530 char luminance_component;
3531 struct arbfp_blit_priv *priv = device->blit_priv;
3532
3533 /* Shader header */
3534 shader_buffer_init(&buffer);
3535
3536 ENTER_GL();
3537 GL_EXTCALL(glGenProgramsARB(1, &shader));
3538 checkGLcall("GL_EXTCALL(glGenProgramsARB(1, &shader))");
3539 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader));
3540 checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader)");
3541 LEAVE_GL();
3542 if(!shader) {
3543 shader_buffer_free(&buffer);
3544 return 0;
3545 }
3546
3547 /* The YUY2 and UYVY formats contain two pixels packed into a 32 bit macropixel,
3548 * giving effectively 16 bit per pixel. The color consists of a luminance(Y) and
3549 * two chroma(U and V) values. Each macropixel has two luminance values, one for
3550 * each single pixel it contains, and one U and one V value shared between both
3551 * pixels.
3552 *
3553 * The data is loaded into an A8L8 texture. With YUY2, the luminance component
3554 * contains the luminance and alpha the chroma. With UYVY it is vice versa. Thus
3555 * take the format into account when generating the read swizzles
3556 *
3557 * Reading the Y value is straightforward - just sample the texture. The hardware
3558 * takes care of filtering in the horizontal and vertical direction.
3559 *
3560 * Reading the U and V values is harder. We have to avoid filtering horizontally,
3561 * because that would mix the U and V values of one pixel or two adjacent pixels.
3562 * Thus floor the texture coordinate and add 0.5 to get an unfiltered read,
3563 * regardless of the filtering setting. Vertical filtering works automatically
3564 * though - the U and V values of two rows are mixed nicely.
3565 *
3566 * Appart of avoiding filtering issues, the code has to know which value it just
3567 * read, and where it can find the other one. To determine this, it checks if
3568 * it sampled an even or odd pixel, and shifts the 2nd read accordingly.
3569 *
3570 * Handling horizontal filtering of U and V values requires reading a 2nd pair
3571 * of pixels, extracting U and V and mixing them. This is not implemented yet.
3572 *
3573 * An alternative implementation idea is to load the texture as A8R8G8B8 texture,
3574 * with width / 2. This way one read gives all 3 values, finding U and V is easy
3575 * in an unfiltered situation. Finding the luminance on the other hand requires
3576 * finding out if it is an odd or even pixel. The real drawback of this approach
3577 * is filtering. This would have to be emulated completely in the shader, reading
3578 * up two 2 packed pixels in up to 2 rows and interpolating both horizontally and
3579 * vertically. Beyond that it would require adjustments to the texture handling
3580 * code to deal with the width scaling
3581 */
3582 shader_addline(&buffer, "!!ARBfp1.0\n");
3583 shader_addline(&buffer, "TEMP luminance;\n");
3584 shader_addline(&buffer, "TEMP temp;\n");
3585 shader_addline(&buffer, "TEMP chroma;\n");
3586 shader_addline(&buffer, "TEMP texcrd;\n");
3587 shader_addline(&buffer, "TEMP texcrd2;\n");
3588 shader_addline(&buffer, "PARAM coef = {1.0, 0.5, 2.0, 0.25};\n");
3589 shader_addline(&buffer, "PARAM yuv_coef = {1.403, 0.344, 0.714, 1.770};\n");
3590 shader_addline(&buffer, "PARAM size = program.local[0];\n");
3591
3592 switch (yuv_fixup)
3593 {
3594 case YUV_FIXUP_UYVY:
3595 case YUV_FIXUP_YUY2:
3596 if (!gen_planar_yuv_read(&buffer, yuv_fixup, textype, &luminance_component))
3597 {
3598 shader_buffer_free(&buffer);
3599 return 0;
3600 }
3601 break;
3602
3603 case YUV_FIXUP_YV12:
3604 if (!gen_yv12_read(&buffer, textype, &luminance_component))
3605 {
3606 shader_buffer_free(&buffer);
3607 return 0;
3608 }
3609 break;
3610
3611 default:
3612 FIXME("Unsupported YUV fixup %#x\n", yuv_fixup);
3613 shader_buffer_free(&buffer);
3614 return 0;
3615 }
3616
3617 /* Calculate the final result. Formula is taken from
3618 * http://www.fourcc.org/fccyvrgb.php. Note that the chroma
3619 * ranges from -0.5 to 0.5
3620 */
3621 shader_addline(&buffer, "SUB chroma.xy, chroma, coef.y;\n");
3622
3623 shader_addline(&buffer, "MAD result.color.x, chroma.x, yuv_coef.x, luminance.%c;\n", luminance_component);
3624 shader_addline(&buffer, "MAD temp.x, -chroma.y, yuv_coef.y, luminance.%c;\n", luminance_component);
3625 shader_addline(&buffer, "MAD result.color.y, -chroma.x, yuv_coef.z, temp.x;\n");
3626 shader_addline(&buffer, "MAD result.color.z, chroma.y, yuv_coef.w, luminance.%c;\n", luminance_component);
3627 shader_addline(&buffer, "END\n");
3628
3629 ENTER_GL();
3630 GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(buffer.buffer), buffer.buffer));
3631
3632 if (glGetError() == GL_INVALID_OPERATION) {
3633 GLint pos;
3634 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
3635 FIXME("Fragment program error at position %d: %s\n", pos,
3636 debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
3637 }
3638 shader_buffer_free(&buffer);
3639 LEAVE_GL();
3640
3641 switch (yuv_fixup)
3642 {
3643 case YUV_FIXUP_YUY2:
3644 if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->yuy2_rect_shader = shader;
3645 else priv->yuy2_2d_shader = shader;
3646 break;
3647
3648 case YUV_FIXUP_UYVY:
3649 if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->uyvy_rect_shader = shader;
3650 else priv->uyvy_2d_shader = shader;
3651 break;
3652
3653 case YUV_FIXUP_YV12:
3654 if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->yv12_rect_shader = shader;
3655 else priv->yv12_2d_shader = shader;
3656 break;
3657 }
3658
3659 return shader;
3660 }
3661
3662 static HRESULT arbfp_blit_set(IWineD3DDevice *iface, WINED3DFORMAT fmt, GLenum textype, UINT width, UINT height) {
3663 GLenum shader;
3664 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3665 float size[4] = {width, height, 1, 1};
3666 struct arbfp_blit_priv *priv = device->blit_priv;
3667 const struct GlPixelFormatDesc *glDesc;
3668 enum yuv_fixup yuv_fixup;
3669
3670 getFormatDescEntry(fmt, &GLINFO_LOCATION, &glDesc);
3671
3672 if (!is_yuv_fixup(glDesc->color_fixup))
3673 {
3674 TRACE("Fixup:\n");
3675 dump_color_fixup_desc(glDesc->color_fixup);
3676 /* Don't bother setting up a shader for unconverted formats */
3677 ENTER_GL();
3678 glEnable(textype);
3679 checkGLcall("glEnable(textype)");
3680 LEAVE_GL();
3681 return WINED3D_OK;
3682 }
3683
3684 yuv_fixup = get_yuv_fixup(glDesc->color_fixup);
3685
3686 switch(yuv_fixup)
3687 {
3688 case YUV_FIXUP_YUY2:
3689 shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->yuy2_rect_shader : priv->yuy2_2d_shader;
3690 break;
3691
3692 case YUV_FIXUP_UYVY:
3693 shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->uyvy_rect_shader : priv->uyvy_2d_shader;
3694 break;
3695
3696 case YUV_FIXUP_YV12:
3697 shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->yv12_rect_shader : priv->yv12_2d_shader;
3698 break;
3699
3700 default:
3701 FIXME("Unsupported YUV fixup %#x, not setting a shader\n", yuv_fixup);
3702 ENTER_GL();
3703 glEnable(textype);
3704 checkGLcall("glEnable(textype)");
3705 LEAVE_GL();
3706 return E_NOTIMPL;
3707 }
3708
3709 if (!shader) shader = gen_yuv_shader(device, yuv_fixup, textype);
3710
3711 ENTER_GL();
3712 glEnable(GL_FRAGMENT_PROGRAM_ARB);
3713 checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB)");
3714 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader));
3715 checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader)");
3716 GL_EXTCALL(glProgramLocalParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, 0, size));
3717 checkGLcall("glProgramLocalParameter4fvARB");
3718 LEAVE_GL();
3719
3720 return WINED3D_OK;
3721 }
3722
3723 static void arbfp_blit_unset(IWineD3DDevice *iface) {
3724 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3725
3726 ENTER_GL();
3727 glDisable(GL_FRAGMENT_PROGRAM_ARB);
3728 checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
3729 glDisable(GL_TEXTURE_2D);
3730 checkGLcall("glDisable(GL_TEXTURE_2D)");
3731 if(GL_SUPPORT(ARB_TEXTURE_CUBE_MAP)) {
3732 glDisable(GL_TEXTURE_CUBE_MAP_ARB);
3733 checkGLcall("glDisable(GL_TEXTURE_CUBE_MAP_ARB)");
3734 }
3735 if(GL_SUPPORT(ARB_TEXTURE_RECTANGLE)) {
3736 glDisable(GL_TEXTURE_RECTANGLE_ARB);
3737 checkGLcall("glDisable(GL_TEXTURE_RECTANGLE_ARB)");
3738 }
3739 LEAVE_GL();
3740 }
3741
3742 static BOOL arbfp_blit_color_fixup_supported(struct color_fixup_desc fixup)
3743 {
3744 enum yuv_fixup yuv_fixup;
3745
3746 if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
3747 {
3748 TRACE("Checking support for fixup:\n");
3749 dump_color_fixup_desc(fixup);
3750 }
3751
3752 if (is_identity_fixup(fixup))
3753 {
3754 TRACE("[OK]\n");
3755 return TRUE;
3756 }
3757
3758 /* We only support YUV conversions. */
3759 if (!is_yuv_fixup(fixup))
3760 {
3761 TRACE("[FAILED]\n");
3762 return FALSE;
3763 }
3764
3765 yuv_fixup = get_yuv_fixup(fixup);
3766 switch(yuv_fixup)
3767 {
3768 case YUV_FIXUP_YUY2:
3769 case YUV_FIXUP_UYVY:
3770 case YUV_FIXUP_YV12:
3771 TRACE("[OK]\n");
3772 return TRUE;
3773
3774 default:
3775 FIXME("Unsupported YUV fixup %#x\n", yuv_fixup);
3776 TRACE("[FAILED]\n");
3777 return FALSE;
3778 }
3779 }
3780
3781 const struct blit_shader arbfp_blit = {
3782 arbfp_blit_alloc,
3783 arbfp_blit_free,
3784 arbfp_blit_set,
3785 arbfp_blit_unset,
3786 arbfp_blit_color_fixup_supported,
3787 };
3788
3789 #undef GLINFO_LOCATION
useful hacks and other patches not (yet) suitable for mainline