?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
push 5f793418e14616d83a4fb368b755a8821b49820b
[wine/hacks.git] / dlls / wined3d / arb_program_shader.c
blob2f8a989fead382c51491efa500c430ca13419bbd
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 * Copyright 2009 Henri Verbeet for CodeWeavers
14 *
15 * This library is free software; you can redistribute it and/or
16 * modify it under the terms of the GNU Lesser General Public
17 * License as published by the Free Software Foundation; either
18 * version 2.1 of the License, or (at your option) any later version.
19 *
20 * This library is distributed in the hope that it will be useful,
21 * but WITHOUT ANY WARRANTY; without even the implied warranty of
22 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
23 * Lesser General Public License for more details.
24 *
25 * You should have received a copy of the GNU Lesser General Public
26 * License along with this library; if not, write to the Free Software
27 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
28 */
29
30 #include "config.h"
31
32 #include <math.h>
33 #include <stdio.h>
34
35 #include "wined3d_private.h"
36
37 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
38 WINE_DECLARE_DEBUG_CHANNEL(d3d_constants);
39 WINE_DECLARE_DEBUG_CHANNEL(d3d_caps);
40 WINE_DECLARE_DEBUG_CHANNEL(d3d);
41
42 #define GLINFO_LOCATION (*gl_info)
43
44 /* We have to subtract any other PARAMs that we might use in our shader programs.
45 * ATI seems to count 2 implicit PARAMs when we use fog and NVIDIA counts 1,
46 * and we reference one row of the PROJECTION matrix which counts as 1 PARAM. */
47 #define ARB_SHADER_RESERVED_VS_CONSTS 3
48
49 /* The arb shader only loads the bump mapping environment matrix into the shader if it finds
50 * a free constant to do that, so only reduce the number of available constants by 2 for the fog states.
51 */
52 #define ARB_SHADER_RESERVED_PS_CONSTS 2
53
54 /* Internally used shader constants. Applications can use constants 0 to GL_LIMITS(vshader_constantsF) - 1,
55 * so upload them above that
56 */
57 #define ARB_SHADER_PRIVCONST_BASE (GL_LIMITS(vshader_constantsF) - ARB_SHADER_RESERVED_VS_CONSTS)
58 #define ARB_SHADER_PRIVCONST_POS ARB_SHADER_PRIVCONST_BASE + 0
59
60 /* ARB_program_shader private data */
61 struct shader_arb_priv {
62 GLuint current_vprogram_id;
63 GLuint current_fprogram_id;
64 GLuint depth_blt_vprogram_id;
65 GLuint depth_blt_fprogram_id[tex_type_count];
66 BOOL use_arbfp_fixed_func;
67 struct hash_table_t *fragment_shaders;
68 };
69
70 /********************************************************
71 * ARB_[vertex/fragment]_program helper functions follow
72 ********************************************************/
73
74 /**
75 * Loads floating point constants into the currently set ARB_vertex/fragment_program.
76 * When constant_list == NULL, it will load all the constants.
77 *
78 * @target_type should be either GL_VERTEX_PROGRAM_ARB (for vertex shaders)
79 * or GL_FRAGMENT_PROGRAM_ARB (for pixel shaders)
80 */
81 static unsigned int shader_arb_load_constantsF(IWineD3DBaseShaderImpl* This, const WineD3D_GL_Info *gl_info,
82 GLuint target_type, unsigned int max_constants, const float *constants, char *dirty_consts)
83 {
84 local_constant* lconst;
85 DWORD i, j;
86 unsigned int ret;
87
88 if (TRACE_ON(d3d_shader)) {
89 for(i = 0; i < max_constants; i++) {
90 if(!dirty_consts[i]) continue;
91 TRACE_(d3d_constants)("Loading constants %i: %f, %f, %f, %f\n", i,
92 constants[i * 4 + 0], constants[i * 4 + 1],
93 constants[i * 4 + 2], constants[i * 4 + 3]);
94 }
95 }
96 /* In 1.X pixel shaders constants are implicitly clamped in the range [-1;1] */
97 if (target_type == GL_FRAGMENT_PROGRAM_ARB && This->baseShader.reg_maps.shader_version.major == 1)
98 {
99 float lcl_const[4];
100 for(i = 0; i < max_constants; i++) {
101 if(!dirty_consts[i]) continue;
102 dirty_consts[i] = 0;
103
104 j = 4 * i;
105 if(constants[j + 0] > 1.0) lcl_const[0] = 1.0;
106 else if(constants[j + 0] < -1.0) lcl_const[0] = -1.0;
107 else lcl_const[0] = constants[j + 0];
108
109 if(constants[j + 1] > 1.0) lcl_const[1] = 1.0;
110 else if(constants[j + 1] < -1.0) lcl_const[1] = -1.0;
111 else lcl_const[1] = constants[j + 1];
112
113 if(constants[j + 2] > 1.0) lcl_const[2] = 1.0;
114 else if(constants[j + 2] < -1.0) lcl_const[2] = -1.0;
115 else lcl_const[2] = constants[j + 2];
116
117 if(constants[j + 3] > 1.0) lcl_const[3] = 1.0;
118 else if(constants[j + 3] < -1.0) lcl_const[3] = -1.0;
119 else lcl_const[3] = constants[j + 3];
120
121 GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, i, lcl_const));
122 }
123 } else {
124 if(GL_SUPPORT(EXT_GPU_PROGRAM_PARAMETERS)) {
125 /* TODO: Benchmark if we're better of with finding the dirty constants ourselves,
126 * or just reloading *all* constants at once
127 *
128 GL_EXTCALL(glProgramEnvParameters4fvEXT(target_type, 0, max_constants, constants));
129 */
130 for(i = 0; i < max_constants; i++) {
131 if(!dirty_consts[i]) continue;
132
133 /* Find the next block of dirty constants */
134 dirty_consts[i] = 0;
135 j = i;
136 for(i++; (i < max_constants) && dirty_consts[i]; i++) {
137 dirty_consts[i] = 0;
138 }
139
140 GL_EXTCALL(glProgramEnvParameters4fvEXT(target_type, j, i - j, constants + (j * 4)));
141 }
142 } else {
143 for(i = 0; i < max_constants; i++) {
144 if(dirty_consts[i]) {
145 dirty_consts[i] = 0;
146 GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, i, constants + (i * 4)));
147 }
148 }
149 }
150 }
151 checkGLcall("glProgramEnvParameter4fvARB()");
152
153 /* Load immediate constants */
154 if(This->baseShader.load_local_constsF) {
155 if (TRACE_ON(d3d_shader)) {
156 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
157 GLfloat* values = (GLfloat*)lconst->value;
158 TRACE_(d3d_constants)("Loading local constants %i: %f, %f, %f, %f\n", lconst->idx,
159 values[0], values[1], values[2], values[3]);
160 }
161 }
162 /* Immediate constants are clamped for 1.X shaders at loading times */
163 ret = 0;
164 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
165 dirty_consts[lconst->idx] = 1; /* Dirtify so the non-immediate constant overwrites it next time */
166 ret = max(ret, lconst->idx + 1);
167 GL_EXTCALL(glProgramEnvParameter4fvARB(target_type, lconst->idx, (GLfloat*)lconst->value));
168 }
169 checkGLcall("glProgramEnvParameter4fvARB()");
170 return ret; /* The loaded immediate constants need reloading for the next shader */
171 } else {
172 return 0; /* No constants are dirty now */
173 }
174 }
175
176 /**
177 * Loads the texture dimensions for NP2 fixup into the currently set ARB_[vertex/fragment]_programs.
178 */
179 static void shader_arb_load_np2fixup_constants(
180 IWineD3DDevice* device,
181 char usePixelShader,
182 char useVertexShader) {
183 /* not implemented */
184 }
185
186 /**
187 * Loads the app-supplied constants into the currently set ARB_[vertex/fragment]_programs.
188 *
189 * We only support float constants in ARB at the moment, so don't
190 * worry about the Integers or Booleans
191 */
192 static void shader_arb_load_constants(
193 IWineD3DDevice* device,
194 char usePixelShader,
195 char useVertexShader) {
196
197 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) device;
198 IWineD3DStateBlockImpl* stateBlock = deviceImpl->stateBlock;
199 const WineD3D_GL_Info *gl_info = &deviceImpl->adapter->gl_info;
200 unsigned char i;
201
202 if (useVertexShader) {
203 IWineD3DBaseShaderImpl* vshader = (IWineD3DBaseShaderImpl*) stateBlock->vertexShader;
204
205 /* Load DirectX 9 float constants for vertex shader */
206 deviceImpl->highest_dirty_vs_const = shader_arb_load_constantsF(
207 vshader, gl_info, GL_VERTEX_PROGRAM_ARB,
208 deviceImpl->highest_dirty_vs_const,
209 stateBlock->vertexShaderConstantF,
210 deviceImpl->activeContext->vshader_const_dirty);
211
212 /* Upload the position fixup */
213 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_VERTEX_PROGRAM_ARB, ARB_SHADER_PRIVCONST_POS, deviceImpl->posFixup));
214 }
215
216 if (usePixelShader) {
217
218 IWineD3DBaseShaderImpl* pshader = (IWineD3DBaseShaderImpl*) stateBlock->pixelShader;
219 IWineD3DPixelShaderImpl *psi = (IWineD3DPixelShaderImpl *) pshader;
220
221 /* Load DirectX 9 float constants for pixel shader */
222 deviceImpl->highest_dirty_ps_const = shader_arb_load_constantsF(
223 pshader, gl_info, GL_FRAGMENT_PROGRAM_ARB,
224 deviceImpl->highest_dirty_ps_const,
225 stateBlock->pixelShaderConstantF,
226 deviceImpl->activeContext->pshader_const_dirty);
227
228 for(i = 0; i < psi->numbumpenvmatconsts; i++) {
229 /* The state manager takes care that this function is always called if the bump env matrix changes
230 */
231 const float *data = (const float *)&stateBlock->textureState[(int) psi->bumpenvmatconst[i].texunit][WINED3DTSS_BUMPENVMAT00];
232 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, psi->bumpenvmatconst[i].const_num, data));
233 deviceImpl->activeContext->pshader_const_dirty[psi->bumpenvmatconst[i].const_num] = 1;
234
235 if (psi->luminanceconst[i].const_num != WINED3D_CONST_NUM_UNUSED)
236 {
237 /* WINED3DTSS_BUMPENVLSCALE and WINED3DTSS_BUMPENVLOFFSET are next to each other.
238 * point gl to the scale, and load 4 floats. x = scale, y = offset, z and w are junk, we
239 * don't care about them. The pointers are valid for sure because the stateblock is bigger.
240 * (they're WINED3DTSS_TEXTURETRANSFORMFLAGS and WINED3DTSS_ADDRESSW, so most likely 0 or NaN
241 */
242 const float *scale = (const float *)&stateBlock->textureState[(int) psi->luminanceconst[i].texunit][WINED3DTSS_BUMPENVLSCALE];
243 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, psi->luminanceconst[i].const_num, scale));
244 deviceImpl->activeContext->pshader_const_dirty[psi->luminanceconst[i].const_num] = 1;
245 }
246 }
247 }
248 }
249
250 static void shader_arb_update_float_vertex_constants(IWineD3DDevice *iface, UINT start, UINT count)
251 {
252 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
253
254 /* We don't want shader constant dirtification to be an O(contexts), so just dirtify the active
255 * context. On a context switch the old context will be fully dirtified */
256 memset(This->activeContext->vshader_const_dirty + start, 1,
257 sizeof(*This->activeContext->vshader_const_dirty) * count);
258 This->highest_dirty_vs_const = max(This->highest_dirty_vs_const, start + count + 1);
259 }
260
261 static void shader_arb_update_float_pixel_constants(IWineD3DDevice *iface, UINT start, UINT count)
262 {
263 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
264
265 /* We don't want shader constant dirtification to be an O(contexts), so just dirtify the active
266 * context. On a context switch the old context will be fully dirtified */
267 memset(This->activeContext->pshader_const_dirty + start, 1,
268 sizeof(*This->activeContext->pshader_const_dirty) * count);
269 This->highest_dirty_ps_const = max(This->highest_dirty_ps_const, start + count + 1);
270 }
271
272 static DWORD *local_const_mapping(IWineD3DBaseShaderImpl *This)
273 {
274 DWORD *ret;
275 DWORD idx = 0;
276 const local_constant *lconst;
277
278 if(This->baseShader.load_local_constsF || list_empty(&This->baseShader.constantsF)) return NULL;
279
280 ret = HeapAlloc(GetProcessHeap(), 0, sizeof(DWORD) * This->baseShader.limits.temporary);
281 if(!ret) {
282 ERR("Out of memory\n");
283 return NULL;
284 }
285
286 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
287 ret[lconst->idx] = idx++;
288 }
289 return ret;
290 }
291
292 /* Generate the variable & register declarations for the ARB_vertex_program output target */
293 static void shader_generate_arb_declarations(IWineD3DBaseShader *iface, const shader_reg_maps *reg_maps,
294 SHADER_BUFFER *buffer, const WineD3D_GL_Info *gl_info, DWORD *lconst_map)
295 {
296 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
297 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device;
298 DWORD i, cur;
299 char pshader = shader_is_pshader_version(reg_maps->shader_version.type);
300 unsigned max_constantsF = min(This->baseShader.limits.constant_float,
301 (pshader ? GL_LIMITS(pshader_constantsF) - ARB_SHADER_RESERVED_PS_CONSTS :
302 GL_LIMITS(vshader_constantsF) - ARB_SHADER_RESERVED_VS_CONSTS));
303 UINT extra_constants_needed = 0;
304 const local_constant *lconst;
305
306 /* Temporary Output register */
307 shader_addline(buffer, "TEMP TMP_OUT;\n");
308
309 for(i = 0; i < This->baseShader.limits.temporary; i++) {
310 if (reg_maps->temporary[i])
311 shader_addline(buffer, "TEMP R%u;\n", i);
312 }
313
314 for (i = 0; i < This->baseShader.limits.address; i++) {
315 if (reg_maps->address[i])
316 shader_addline(buffer, "ADDRESS A%d;\n", i);
317 }
318
319 for(i = 0; i < This->baseShader.limits.texcoord; i++) {
320 if (reg_maps->texcoord[i])
321 shader_addline(buffer,"TEMP T%u;\n", i);
322 }
323
324 /* Texture coordinate registers must be pre-loaded */
325 for (i = 0; i < This->baseShader.limits.texcoord; i++) {
326 if (reg_maps->texcoord[i])
327 shader_addline(buffer, "MOV T%u, fragment.texcoord[%u];\n", i, i);
328 }
329
330 for(i = 0; i < (sizeof(reg_maps->bumpmat) / sizeof(reg_maps->bumpmat[0])); i++) {
331 IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *) This;
332 if(!reg_maps->bumpmat[i]) continue;
333
334 cur = ps->numbumpenvmatconsts;
335 ps->bumpenvmatconst[cur].const_num = -1;
336 ps->bumpenvmatconst[cur].texunit = i;
337 ps->luminanceconst[cur].const_num = -1;
338 ps->luminanceconst[cur].texunit = i;
339
340 /* If the shader does not use all available constants, use the next free constant to load the bump mapping environment matrix from
341 * the stateblock into the shader. If no constant is available don't load, texbem will then just sample the texture without applying
342 * bump mapping.
343 */
344 if(max_constantsF + extra_constants_needed < GL_LIMITS(pshader_constantsF) - ARB_SHADER_RESERVED_PS_CONSTS) {
345 ps->bumpenvmatconst[cur].const_num = max_constantsF + extra_constants_needed;
346 shader_addline(buffer, "PARAM bumpenvmat%d = program.env[%d];\n",
347 i, ps->bumpenvmatconst[cur].const_num);
348 extra_constants_needed++;
349
350 if(reg_maps->luminanceparams && max_constantsF + extra_constants_needed < GL_LIMITS(pshader_constantsF) - ARB_SHADER_RESERVED_PS_CONSTS) {
351 ((IWineD3DPixelShaderImpl *)This)->luminanceconst[cur].const_num = max_constantsF + extra_constants_needed;
352 shader_addline(buffer, "PARAM luminance%d = program.env[%d];\n",
353 i, ps->luminanceconst[cur].const_num);
354 extra_constants_needed++;
355 } else if(reg_maps->luminanceparams) {
356 FIXME("No free constant to load the luminance parameters\n");
357 }
358 } else {
359 FIXME("No free constant found to load environment bump mapping matrix into the shader. texbem instruction will not apply bump mapping\n");
360 }
361
362 ps->numbumpenvmatconsts = cur + 1;
363 }
364
365 if(device->stateBlock->renderState[WINED3DRS_SRGBWRITEENABLE] && pshader) {
366 shader_addline(buffer, "PARAM srgb_consts1 = {%f, %f, %f, %f};\n",
367 srgb_mul_low, srgb_cmp, srgb_pow, srgb_mul_high);
368 shader_addline(buffer, "PARAM srgb_consts2 = {%f, %f, %f, %f};\n",
369 srgb_sub_high, 0.0, 0.0, 0.0);
370 }
371
372 /* Load local constants using the program-local space,
373 * this avoids reloading them each time the shader is used
374 */
375 if(lconst_map) {
376 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
377 shader_addline(buffer, "PARAM C%u = program.local[%u];\n", lconst->idx,
378 lconst_map[lconst->idx]);
379 }
380 }
381
382 /* we use the array-based constants array if the local constants are marked for loading,
383 * because then we use indirect addressing, or when the local constant list is empty,
384 * because then we don't know if we're using indirect addressing or not. If we're hardcoding
385 * local constants do not declare the loaded constants as an array because ARB compilers usually
386 * do not optimize unused constants away
387 */
388 if(This->baseShader.reg_maps.usesrelconstF) {
389 /* Need to PARAM the environment parameters (constants) so we can use relative addressing */
390 shader_addline(buffer, "PARAM C[%d] = { program.env[0..%d] };\n",
391 max_constantsF, max_constantsF - 1);
392 } else {
393 for(i = 0; i < max_constantsF; i++) {
394 DWORD idx, mask;
395 idx = i >> 5;
396 mask = 1 << (i & 0x1f);
397 if(!shader_constant_is_local(This, i) && (This->baseShader.reg_maps.constf[idx] & mask)) {
398 shader_addline(buffer, "PARAM C%d = program.env[%d];\n",i, i);
399 }
400 }
401 }
402 }
403
404 static const char * const shift_tab[] = {
405 "dummy", /* 0 (none) */
406 "coefmul.x", /* 1 (x2) */
407 "coefmul.y", /* 2 (x4) */
408 "coefmul.z", /* 3 (x8) */
409 "coefmul.w", /* 4 (x16) */
410 "dummy", /* 5 (x32) */
411 "dummy", /* 6 (x64) */
412 "dummy", /* 7 (x128) */
413 "dummy", /* 8 (d256) */
414 "dummy", /* 9 (d128) */
415 "dummy", /* 10 (d64) */
416 "dummy", /* 11 (d32) */
417 "coefdiv.w", /* 12 (d16) */
418 "coefdiv.z", /* 13 (d8) */
419 "coefdiv.y", /* 14 (d4) */
420 "coefdiv.x" /* 15 (d2) */
421 };
422
423 static void shader_arb_get_write_mask(const struct wined3d_shader_instruction *ins,
424 const struct wined3d_shader_dst_param *dst, char *write_mask)
425 {
426 char *ptr = write_mask;
427 char vshader = shader_is_vshader_version(ins->ctx->reg_maps->shader_version.type);
428
429 if (vshader && dst->reg.type == WINED3DSPR_ADDR)
430 {
431 *ptr++ = '.';
432 *ptr++ = 'x';
433 }
434 else if (dst->write_mask != WINED3DSP_WRITEMASK_ALL)
435 {
436 *ptr++ = '.';
437 if (dst->write_mask & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
438 if (dst->write_mask & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
439 if (dst->write_mask & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
440 if (dst->write_mask & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
441 }
442
443 *ptr = '\0';
444 }
445
446 static void shader_arb_get_swizzle(const struct wined3d_shader_src_param *param, BOOL fixup, char *swizzle_str)
447 {
448 /* For registers of type WINED3DDECLTYPE_D3DCOLOR, data is stored as "bgra",
449 * but addressed as "rgba". To fix this we need to swap the register's x
450 * and z components. */
451 const char *swizzle_chars = fixup ? "zyxw" : "xyzw";
452 char *ptr = swizzle_str;
453
454 /* swizzle bits fields: wwzzyyxx */
455 DWORD swizzle = param->swizzle;
456 DWORD swizzle_x = swizzle & 0x03;
457 DWORD swizzle_y = (swizzle >> 2) & 0x03;
458 DWORD swizzle_z = (swizzle >> 4) & 0x03;
459 DWORD swizzle_w = (swizzle >> 6) & 0x03;
460
461 /* If the swizzle is the default swizzle (ie, "xyzw"), we don't need to
462 * generate a swizzle string. Unless we need to our own swizzling. */
463 if (swizzle != WINED3DSP_NOSWIZZLE || fixup)
464 {
465 *ptr++ = '.';
466 if (swizzle_x == swizzle_y && swizzle_x == swizzle_z && swizzle_x == swizzle_w) {
467 *ptr++ = swizzle_chars[swizzle_x];
468 } else {
469 *ptr++ = swizzle_chars[swizzle_x];
470 *ptr++ = swizzle_chars[swizzle_y];
471 *ptr++ = swizzle_chars[swizzle_z];
472 *ptr++ = swizzle_chars[swizzle_w];
473 }
474 }
475
476 *ptr = '\0';
477 }
478
479 static void shader_arb_get_register_name(IWineD3DBaseShader *iface, WINED3DSHADER_PARAM_REGISTER_TYPE register_type,
480 UINT register_idx, BOOL rel_addr, char *register_name, BOOL *is_color)
481 {
482 /* oPos, oFog and oPts in D3D */
483 static const char * const rastout_reg_names[] = {"TMP_OUT", "result.fogcoord", "result.pointsize"};
484 IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)iface;
485 BOOL pshader = shader_is_pshader_version(This->baseShader.reg_maps.shader_version.type);
486
487 *is_color = FALSE;
488
489 switch (register_type)
490 {
491 case WINED3DSPR_TEMP:
492 sprintf(register_name, "R%u", register_idx);
493 break;
494
495 case WINED3DSPR_INPUT:
496 if (pshader)
497 {
498 if (register_idx == 0) strcpy(register_name, "fragment.color.primary");
499 else strcpy(register_name, "fragment.color.secondary");
500 }
501 else
502 {
503 if (((IWineD3DVertexShaderImpl *)This)->cur_args->swizzle_map & (1 << register_idx)) *is_color = TRUE;
504 sprintf(register_name, "vertex.attrib[%u]", register_idx);
505 }
506 break;
507
508 case WINED3DSPR_CONST:
509 if (!pshader && rel_addr)
510 {
511 UINT rel_offset = ((IWineD3DVertexShaderImpl *)This)->rel_offset;
512 if (register_idx >= rel_offset)
513 sprintf(register_name, "C[A0.x + %u]", register_idx - rel_offset);
514 else
515 sprintf(register_name, "C[A0.x - %u]", -register_idx + rel_offset);
516 }
517 else
518 {
519 if (This->baseShader.reg_maps.usesrelconstF)
520 sprintf(register_name, "C[%u]", register_idx);
521 else
522 sprintf(register_name, "C%u", register_idx);
523 }
524 break;
525
526 case WINED3DSPR_TEXTURE: /* case WINED3DSPR_ADDR: */
527 if (pshader) sprintf(register_name, "T%u", register_idx);
528 else sprintf(register_name, "A%u", register_idx);
529 break;
530
531 case WINED3DSPR_COLOROUT:
532 if (register_idx == 0)
533 {
534 strcpy(register_name, "TMP_COLOR");
535 }
536 else
537 {
538 /* TODO: See GL_ARB_draw_buffers */
539 FIXME("Unsupported write to render target %u\n", register_idx);
540 sprintf(register_name, "unsupported_register");
541 }
542 break;
543
544 case WINED3DSPR_RASTOUT:
545 sprintf(register_name, "%s", rastout_reg_names[register_idx]);
546 break;
547
548 case WINED3DSPR_DEPTHOUT:
549 strcpy(register_name, "result.depth");
550 break;
551
552 case WINED3DSPR_ATTROUT:
553 if (pshader) sprintf(register_name, "oD[%u]", register_idx);
554 else if (register_idx == 0) strcpy(register_name, "result.color.primary");
555 else strcpy(register_name, "result.color.secondary");
556 break;
557
558 case WINED3DSPR_TEXCRDOUT:
559 if (pshader) sprintf(register_name, "oT[%u]", register_idx);
560 else sprintf(register_name, "result.texcoord[%u]", register_idx);
561 break;
562
563 default:
564 FIXME("Unhandled register type %#x[%u]\n", register_type, register_idx);
565 sprintf(register_name, "unrecognized_register[%u]", register_idx);
566 break;
567 }
568 }
569
570 static void shader_arb_add_src_param(const struct wined3d_shader_instruction *ins,
571 const struct wined3d_shader_src_param *wined3d_src, char *str)
572 {
573 char register_name[255];
574 char swizzle[6];
575 BOOL is_color;
576
577 if (wined3d_src->modifiers == WINED3DSPSM_NEG) strcat(str, " -");
578 else strcat(str, " ");
579
580 shader_arb_get_register_name(ins->ctx->shader, wined3d_src->reg.type,
581 wined3d_src->reg.idx, !!wined3d_src->reg.rel_addr, register_name, &is_color);
582 strcat(str, register_name);
583
584 shader_arb_get_swizzle(wined3d_src, is_color, swizzle);
585 strcat(str, swizzle);
586 }
587
588 static void shader_arb_add_dst_param(const struct wined3d_shader_instruction *ins,
589 const struct wined3d_shader_dst_param *wined3d_dst, char *str)
590 {
591 char register_name[255];
592 char write_mask[6];
593 BOOL is_color;
594
595 strcat(str, " ");
596
597 shader_arb_get_register_name(ins->ctx->shader, wined3d_dst->reg.type,
598 wined3d_dst->reg.idx, !!wined3d_dst->reg.rel_addr, register_name, &is_color);
599 strcat(str, register_name);
600
601 shader_arb_get_write_mask(ins, wined3d_dst, write_mask);
602 strcat(str, write_mask);
603 }
604
605 static const char *shader_arb_get_fixup_swizzle(enum fixup_channel_source channel_source)
606 {
607 switch(channel_source)
608 {
609 case CHANNEL_SOURCE_ZERO: return "0";
610 case CHANNEL_SOURCE_ONE: return "1";
611 case CHANNEL_SOURCE_X: return "x";
612 case CHANNEL_SOURCE_Y: return "y";
613 case CHANNEL_SOURCE_Z: return "z";
614 case CHANNEL_SOURCE_W: return "w";
615 default:
616 FIXME("Unhandled channel source %#x\n", channel_source);
617 return "undefined";
618 }
619 }
620
621 static void gen_color_correction(SHADER_BUFFER *buffer, const char *reg, DWORD dst_mask,
622 const char *one, const char *two, struct color_fixup_desc fixup)
623 {
624 DWORD mask;
625
626 if (is_yuv_fixup(fixup))
627 {
628 enum yuv_fixup yuv_fixup = get_yuv_fixup(fixup);
629 FIXME("YUV fixup (%#x) not supported\n", yuv_fixup);
630 return;
631 }
632
633 mask = 0;
634 if (fixup.x_source != CHANNEL_SOURCE_X) mask |= WINED3DSP_WRITEMASK_0;
635 if (fixup.y_source != CHANNEL_SOURCE_Y) mask |= WINED3DSP_WRITEMASK_1;
636 if (fixup.z_source != CHANNEL_SOURCE_Z) mask |= WINED3DSP_WRITEMASK_2;
637 if (fixup.w_source != CHANNEL_SOURCE_W) mask |= WINED3DSP_WRITEMASK_3;
638 mask &= dst_mask;
639
640 if (mask)
641 {
642 shader_addline(buffer, "SWZ %s, %s, %s, %s, %s, %s;\n", reg, reg,
643 shader_arb_get_fixup_swizzle(fixup.x_source), shader_arb_get_fixup_swizzle(fixup.y_source),
644 shader_arb_get_fixup_swizzle(fixup.z_source), shader_arb_get_fixup_swizzle(fixup.w_source));
645 }
646
647 mask = 0;
648 if (fixup.x_sign_fixup) mask |= WINED3DSP_WRITEMASK_0;
649 if (fixup.y_sign_fixup) mask |= WINED3DSP_WRITEMASK_1;
650 if (fixup.z_sign_fixup) mask |= WINED3DSP_WRITEMASK_2;
651 if (fixup.w_sign_fixup) mask |= WINED3DSP_WRITEMASK_3;
652 mask &= dst_mask;
653
654 if (mask)
655 {
656 char reg_mask[6];
657 char *ptr = reg_mask;
658
659 if (mask != WINED3DSP_WRITEMASK_ALL)
660 {
661 *ptr++ = '.';
662 if (mask & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
663 if (mask & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
664 if (mask & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
665 if (mask & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
666 }
667 *ptr = '\0';
668
669 shader_addline(buffer, "MAD %s%s, %s, %s, -%s;\n", reg, reg_mask, reg, two, one);
670 }
671 }
672
673 static void shader_hw_sample(const struct wined3d_shader_instruction *ins, DWORD sampler_idx,
674 const char *dst_str, const char *coord_reg, BOOL projected, BOOL bias)
675 {
676 SHADER_BUFFER *buffer = ins->ctx->buffer;
677 DWORD sampler_type = ins->ctx->reg_maps->sampler_type[sampler_idx];
678 const char *tex_type;
679 IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)ins->ctx->shader;
680 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device;
681
682 switch(sampler_type) {
683 case WINED3DSTT_1D:
684 tex_type = "1D";
685 break;
686
687 case WINED3DSTT_2D:
688 if(device->stateBlock->textures[sampler_idx] &&
689 IWineD3DBaseTexture_GetTextureDimensions(device->stateBlock->textures[sampler_idx]) == GL_TEXTURE_RECTANGLE_ARB) {
690 tex_type = "RECT";
691 } else {
692 tex_type = "2D";
693 }
694 if (shader_is_pshader_version(ins->ctx->reg_maps->shader_version.type))
695 {
696 const IWineD3DPixelShaderImpl* const ps = (const IWineD3DPixelShaderImpl*)This;
697 if(ps->cur_args->np2_fixup & (1 << sampler_idx)) {
698 FIXME("NP2 texcoord fixup is currently not implemented in ARB mode (use GLSL instead).\n");
699 }
700 }
701 break;
702
703 case WINED3DSTT_VOLUME:
704 tex_type = "3D";
705 break;
706
707 case WINED3DSTT_CUBE:
708 tex_type = "CUBE";
709 break;
710
711 default:
712 ERR("Unexpected texture type %d\n", sampler_type);
713 tex_type = "";
714 }
715
716 if (bias) {
717 /* Shouldn't be possible, but let's check for it */
718 if(projected) FIXME("Biased and Projected texture sampling\n");
719 /* TXB takes the 4th component of the source vector automatically, as d3d. Nothing more to do */
720 shader_addline(buffer, "TXB %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
721 } else if (projected) {
722 shader_addline(buffer, "TXP %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
723 } else {
724 shader_addline(buffer, "TEX %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
725 }
726
727 if (shader_is_pshader_version(ins->ctx->reg_maps->shader_version.type))
728 {
729 IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
730 gen_color_correction(buffer, dst_str, ins->dst[0].write_mask,
731 "one", "coefmul.x", ps->cur_args->color_fixup[sampler_idx]);
732 }
733 }
734
735 static void pshader_gen_input_modifier_line(IWineD3DBaseShader *iface, SHADER_BUFFER *buffer,
736 const struct wined3d_shader_src_param *src, unsigned int tmpreg, char *outregstr)
737 {
738 /* Generate a line that does the input modifier computation and return the input register to use */
739 BOOL is_color = FALSE;
740 char regstr[256];
741 char swzstr[20];
742 int insert_line;
743
744 /* Assume a new line will be added */
745 insert_line = 1;
746
747 /* Get register name */
748 shader_arb_get_register_name(iface, src->reg.type,
749 src->reg.idx, !!src->reg.rel_addr, regstr, &is_color);
750 shader_arb_get_swizzle(src, is_color, swzstr);
751
752 switch (src->modifiers)
753 {
754 case WINED3DSPSM_NONE:
755 sprintf(outregstr, "%s%s", regstr, swzstr);
756 insert_line = 0;
757 break;
758 case WINED3DSPSM_NEG:
759 sprintf(outregstr, "-%s%s", regstr, swzstr);
760 insert_line = 0;
761 break;
762 case WINED3DSPSM_BIAS:
763 shader_addline(buffer, "ADD T%c, %s, -coefdiv.x;\n", 'A' + tmpreg, regstr);
764 break;
765 case WINED3DSPSM_BIASNEG:
766 shader_addline(buffer, "ADD T%c, -%s, coefdiv.x;\n", 'A' + tmpreg, regstr);
767 break;
768 case WINED3DSPSM_SIGN:
769 shader_addline(buffer, "MAD T%c, %s, coefmul.x, -one.x;\n", 'A' + tmpreg, regstr);
770 break;
771 case WINED3DSPSM_SIGNNEG:
772 shader_addline(buffer, "MAD T%c, %s, -coefmul.x, one.x;\n", 'A' + tmpreg, regstr);
773 break;
774 case WINED3DSPSM_COMP:
775 shader_addline(buffer, "SUB T%c, one.x, %s;\n", 'A' + tmpreg, regstr);
776 break;
777 case WINED3DSPSM_X2:
778 shader_addline(buffer, "ADD T%c, %s, %s;\n", 'A' + tmpreg, regstr, regstr);
779 break;
780 case WINED3DSPSM_X2NEG:
781 shader_addline(buffer, "ADD T%c, -%s, -%s;\n", 'A' + tmpreg, regstr, regstr);
782 break;
783 case WINED3DSPSM_DZ:
784 shader_addline(buffer, "RCP T%c, %s.z;\n", 'A' + tmpreg, regstr);
785 shader_addline(buffer, "MUL T%c, %s, T%c;\n", 'A' + tmpreg, regstr, 'A' + tmpreg);
786 break;
787 case WINED3DSPSM_DW:
788 shader_addline(buffer, "RCP T%c, %s.w;\n", 'A' + tmpreg, regstr);
789 shader_addline(buffer, "MUL T%c, %s, T%c;\n", 'A' + tmpreg, regstr, 'A' + tmpreg);
790 break;
791 default:
792 sprintf(outregstr, "%s%s", regstr, swzstr);
793 insert_line = 0;
794 }
795
796 /* Return modified or original register, with swizzle */
797 if (insert_line)
798 sprintf(outregstr, "T%c%s", 'A' + tmpreg, swzstr);
799 }
800
801 static void pshader_hw_bem(const struct wined3d_shader_instruction *ins)
802 {
803 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
804 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
805 SHADER_BUFFER *buffer = ins->ctx->buffer;
806 char dst_name[50];
807 char src_name[2][50];
808 char dst_wmask[20];
809 DWORD sampler_code = dst->reg.idx;
810 BOOL has_bumpmat = FALSE;
811 BOOL is_color;
812 int i;
813
814 for(i = 0; i < This->numbumpenvmatconsts; i++) {
815 if (This->bumpenvmatconst[i].const_num != WINED3D_CONST_NUM_UNUSED
816 && This->bumpenvmatconst[i].texunit == sampler_code)
817 {
818 has_bumpmat = TRUE;
819 break;
820 }
821 }
822
823 shader_arb_get_register_name(ins->ctx->shader, dst->reg.type,
824 dst->reg.idx, !!dst->reg.rel_addr, dst_name, &is_color);
825 shader_arb_get_write_mask(ins, dst, dst_wmask);
826 strcat(dst_name, dst_wmask);
827
828 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src_name[0]);
829 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[1], 1, src_name[1]);
830
831 if(has_bumpmat) {
832 /* Sampling the perturbation map in Tsrc was done already, including the signedness correction if needed */
833 shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, x, z, 0, 0;\n", sampler_code);
834 shader_addline(buffer, "DP3 TMP.r, TMP2, %s;\n", src_name[1]);
835 shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, y, w, 0, 0;\n", sampler_code);
836 shader_addline(buffer, "DP3 TMP.g, TMP2, %s;\n", src_name[1]);
837
838 shader_addline(buffer, "ADD %s, %s, TMP;\n", dst_name, src_name[0]);
839 } else {
840 shader_addline(buffer, "MOV %s, %s;\n", dst_name, src_name[0]);
841 }
842 }
843
844 static void pshader_hw_cnd(const struct wined3d_shader_instruction *ins)
845 {
846 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
847 SHADER_BUFFER *buffer = ins->ctx->buffer;
848 char dst_wmask[20];
849 char dst_name[50];
850 char src_name[3][50];
851 BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
852 BOOL is_color;
853 DWORD shader_version = WINED3D_SHADER_VERSION(ins->ctx->reg_maps->shader_version.major,
854 ins->ctx->reg_maps->shader_version.minor);
855
856 /* FIXME: support output modifiers */
857
858 /* Handle output register */
859 shader_arb_get_register_name(ins->ctx->shader, dst->reg.type,
860 dst->reg.idx, !!dst->reg.rel_addr, dst_name, &is_color);
861 shader_arb_get_write_mask(ins, dst, dst_wmask);
862
863 /* Generate input register names (with modifiers) */
864 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src_name[0]);
865 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[1], 1, src_name[1]);
866 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[2], 2, src_name[2]);
867
868 /* The coissue flag changes the semantic of the cnd instruction in <= 1.3 shaders */
869 if (shader_version <= WINED3D_SHADER_VERSION(1, 3) && ins->coissue)
870 {
871 shader_addline(buffer, "MOV%s %s%s, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask, src_name[1]);
872 } else {
873 shader_addline(buffer, "ADD TMP, -%s, coefdiv.x;\n", src_name[0]);
874 shader_addline(buffer, "CMP%s %s%s, TMP, %s, %s;\n",
875 sat ? "_SAT" : "", dst_name, dst_wmask, src_name[1], src_name[2]);
876 }
877 }
878
879 static void pshader_hw_cmp(const struct wined3d_shader_instruction *ins)
880 {
881 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
882 SHADER_BUFFER *buffer = ins->ctx->buffer;
883 char dst_wmask[20];
884 char dst_name[50];
885 char src_name[3][50];
886 BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
887 BOOL is_color;
888
889 /* FIXME: support output modifiers */
890
891 /* Handle output register */
892 shader_arb_get_register_name(ins->ctx->shader, dst->reg.type,
893 dst->reg.idx, !!dst->reg.rel_addr, dst_name, &is_color);
894 shader_arb_get_write_mask(ins, dst, dst_wmask);
895
896 /* Generate input register names (with modifiers) */
897 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src_name[0]);
898 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[1], 1, src_name[1]);
899 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[2], 2, src_name[2]);
900
901 shader_addline(buffer, "CMP%s %s%s, %s, %s, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask,
902 src_name[0], src_name[2], src_name[1]);
903 }
904
905 /** Process the WINED3DSIO_DP2ADD instruction in ARB.
906 * dst = dot2(src0, src1) + src2 */
907 static void pshader_hw_dp2add(const struct wined3d_shader_instruction *ins)
908 {
909 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
910 SHADER_BUFFER *buffer = ins->ctx->buffer;
911 char dst_wmask[20];
912 char dst_name[50];
913 char src_name[3][50];
914 BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
915 BOOL is_color;
916
917 shader_arb_get_register_name(ins->ctx->shader, dst->reg.type,
918 dst->reg.idx, !!dst->reg.rel_addr, dst_name, &is_color);
919 shader_arb_get_write_mask(ins, dst, dst_wmask);
920
921 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src_name[0]);
922 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[1], 1, src_name[1]);
923 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[2], 2, src_name[2]);
924
925 /* Emulate a DP2 with a DP3 and 0.0 */
926 shader_addline(buffer, "MOV TMP, %s;\n", src_name[0]);
927 shader_addline(buffer, "MOV TMP.z, 0.0;\n");
928 shader_addline(buffer, "DP3 TMP2, TMP, %s;\n", src_name[1]);
929 shader_addline(buffer, "ADD%s %s%s, TMP2, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask, src_name[2]);
930 }
931
932 /* Map the opcode 1-to-1 to the GL code */
933 static void shader_hw_map2gl(const struct wined3d_shader_instruction *ins)
934 {
935 SHADER_BUFFER *buffer = ins->ctx->buffer;
936 const char *instruction;
937 char arguments[256];
938 unsigned int i;
939
940 switch (ins->handler_idx)
941 {
942 case WINED3DSIH_ABS: instruction = "ABS"; break;
943 case WINED3DSIH_ADD: instruction = "ADD"; break;
944 case WINED3DSIH_CRS: instruction = "XPD"; break;
945 case WINED3DSIH_DP3: instruction = "DP3"; break;
946 case WINED3DSIH_DP4: instruction = "DP4"; break;
947 case WINED3DSIH_DST: instruction = "DST"; break;
948 case WINED3DSIH_EXP: instruction = "EX2"; break;
949 case WINED3DSIH_EXPP: instruction = "EXP"; break;
950 case WINED3DSIH_FRC: instruction = "FRC"; break;
951 case WINED3DSIH_LIT: instruction = "LIT"; break;
952 case WINED3DSIH_LOG: instruction = "LG2"; break;
953 case WINED3DSIH_LOGP: instruction = "LOG"; break;
954 case WINED3DSIH_LRP: instruction = "LRP"; break;
955 case WINED3DSIH_MAD: instruction = "MAD"; break;
956 case WINED3DSIH_MAX: instruction = "MAX"; break;
957 case WINED3DSIH_MIN: instruction = "MIN"; break;
958 case WINED3DSIH_MOV: instruction = "MOV"; break;
959 case WINED3DSIH_MUL: instruction = "MUL"; break;
960 case WINED3DSIH_NOP: instruction = "NOP"; break;
961 case WINED3DSIH_POW: instruction = "POW"; break;
962 case WINED3DSIH_SGE: instruction = "SGE"; break;
963 case WINED3DSIH_SLT: instruction = "SLT"; break;
964 case WINED3DSIH_SUB: instruction = "SUB"; break;
965 default: instruction = "";
966 FIXME("Unhandled opcode %#x\n", ins->handler_idx);
967 break;
968 }
969
970 if (shader_is_pshader_version(ins->ctx->reg_maps->shader_version.type))
971 {
972 /* Output token related */
973 const struct wined3d_shader_dst_param *dst;
974 char output_rname[256];
975 char output_wmask[20];
976 char operands[4][100];
977 BOOL saturate = FALSE;
978 BOOL centroid = FALSE;
979 BOOL partialprecision = FALSE;
980 const char *modifier;
981 BOOL is_color;
982 DWORD shift;
983
984 if (!(ins->dst_count + ins->src_count))
985 {
986 ERR("Opcode \"%#x\" has no parameters\n", ins->handler_idx);
987 return;
988 }
989 dst = &ins->dst[0];
990
991 /* Process modifiers */
992 if (dst->modifiers)
993 {
994 DWORD mask = dst->modifiers;
995
996 saturate = mask & WINED3DSPDM_SATURATE;
997 centroid = mask & WINED3DSPDM_MSAMPCENTROID;
998 partialprecision = mask & WINED3DSPDM_PARTIALPRECISION;
999 mask &= ~(WINED3DSPDM_MSAMPCENTROID | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_SATURATE);
1000 if (mask)
1001 FIXME("Unrecognized modifier(%#x)\n", mask);
1002
1003 if (centroid)
1004 FIXME("Unhandled modifier(%#x)\n", mask);
1005 }
1006 shift = dst->shift;
1007 modifier = (saturate && !shift) ? "_SAT" : "";
1008
1009 /* Generate input register names (with modifiers) */
1010 for (i = 0; i < ins->src_count; ++i)
1011 {
1012 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[i], i, operands[i + 1]);
1013 }
1014
1015 /* Handle output register */
1016 shader_arb_get_register_name(ins->ctx->shader, dst->reg.type,
1017 dst->reg.idx, !!dst->reg.rel_addr, output_rname, &is_color);
1018 strcpy(operands[0], output_rname);
1019 shader_arb_get_write_mask(ins, dst, output_wmask);
1020 strcat(operands[0], output_wmask);
1021
1022 arguments[0] = '\0';
1023 strcat(arguments, operands[0]);
1024 for (i = 0; i < ins->src_count; ++i)
1025 {
1026 strcat(arguments, ", ");
1027 strcat(arguments, operands[i + 1]);
1028 }
1029 shader_addline(buffer, "%s%s %s;\n", instruction, modifier, arguments);
1030 } else {
1031 /* Note that shader_arb_add_*_param() adds spaces. */
1032
1033 arguments[0] = '\0';
1034 if (ins->dst_count)
1035 {
1036 shader_arb_add_dst_param(ins, &ins->dst[0], arguments);
1037 for (i = 0; i < ins->src_count; ++i)
1038 {
1039 strcat(arguments, ",");
1040 shader_arb_add_src_param(ins, &ins->src[i], arguments);
1041 }
1042 }
1043 shader_addline(buffer, "%s%s;\n", instruction, arguments);
1044 }
1045 }
1046
1047 static void shader_hw_mov(const struct wined3d_shader_instruction *ins)
1048 {
1049 IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)ins->ctx->shader;
1050
1051 if ((ins->ctx->reg_maps->shader_version.major == 1
1052 && !shader_is_pshader_version(ins->ctx->reg_maps->shader_version.type)
1053 && ins->dst[0].reg.type == WINED3DSPR_ADDR)
1054 || ins->handler_idx == WINED3DSIH_MOVA)
1055 {
1056 SHADER_BUFFER *buffer = ins->ctx->buffer;
1057 char src0_param[256];
1058
1059 if (ins->handler_idx == WINED3DSIH_MOVA)
1060 FIXME("mova should round\n");
1061
1062 src0_param[0] = '\0';
1063 if (((IWineD3DVertexShaderImpl *)shader)->rel_offset)
1064 {
1065 shader_arb_add_src_param(ins, &ins->src[0], src0_param);
1066 shader_addline(buffer, "ADD TMP.x, %s, helper_const.z;\n", src0_param);
1067 shader_addline(buffer, "ARL A0.x, TMP.x;\n");
1068 }
1069 else
1070 {
1071 /* Apple's ARB_vertex_program implementation does not accept an ARL source argument
1072 * with more than one component. Thus replicate the first source argument over all
1073 * 4 components. For example, .xyzw -> .x (or better: .xxxx), .zwxy -> .z, etc) */
1074 struct wined3d_shader_src_param tmp_src = ins->src[0];
1075 tmp_src.swizzle = (tmp_src.swizzle & 0x3) * 0x55;
1076 shader_arb_add_src_param(ins, &tmp_src, src0_param);
1077 shader_addline(buffer, "ARL A0.x, %s;\n", src0_param);
1078 }
1079 }
1080 else
1081 {
1082 shader_hw_map2gl(ins);
1083 }
1084 }
1085
1086 static void pshader_hw_texkill(const struct wined3d_shader_instruction *ins)
1087 {
1088 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1089 SHADER_BUFFER *buffer = ins->ctx->buffer;
1090 char reg_dest[40];
1091 BOOL is_color;
1092
1093 /* No swizzles are allowed in d3d's texkill. PS 1.x ignores the 4th component as documented,
1094 * but >= 2.0 honors it(undocumented, but tested by the d3d9 testsuit)
1095 */
1096 shader_arb_get_register_name(ins->ctx->shader, dst->reg.type,
1097 dst->reg.idx, !!dst->reg.rel_addr, reg_dest, &is_color);
1098
1099 if (ins->ctx->reg_maps->shader_version.major >= 2)
1100 {
1101 /* The arb backend doesn't claim ps 2.0 support, but try to eat what the app feeds to us */
1102 shader_addline(buffer, "KIL %s;\n", reg_dest);
1103 } else {
1104 /* ARB fp doesn't like swizzles on the parameter of the KIL instruction. To mask the 4th component,
1105 * copy the register into our general purpose TMP variable, overwrite .w and pass TMP to KIL
1106 */
1107 shader_addline(buffer, "MOV TMP, %s;\n", reg_dest);
1108 shader_addline(buffer, "MOV TMP.w, one.w;\n");
1109 shader_addline(buffer, "KIL TMP;\n");
1110 }
1111 }
1112
1113 static void pshader_hw_tex(const struct wined3d_shader_instruction *ins)
1114 {
1115 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1116 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1117 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1118 BOOL is_color;
1119 SHADER_BUFFER *buffer = ins->ctx->buffer;
1120 DWORD shader_version = WINED3D_SHADER_VERSION(ins->ctx->reg_maps->shader_version.major,
1121 ins->ctx->reg_maps->shader_version.minor);
1122 BOOL projected = FALSE, bias = FALSE;
1123
1124 char reg_dest[40];
1125 char reg_coord[40];
1126 DWORD reg_sampler_code;
1127
1128 /* All versions have a destination register */
1129 shader_arb_get_register_name(ins->ctx->shader, dst->reg.type,
1130 dst->reg.idx, !!dst->reg.rel_addr, reg_dest, &is_color);
1131
1132 /* 1.0-1.3: Use destination register as coordinate source.
1133 1.4+: Use provided coordinate source register. */
1134 if (shader_version < WINED3D_SHADER_VERSION(1,4))
1135 strcpy(reg_coord, reg_dest);
1136 else
1137 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, reg_coord);
1138
1139 /* 1.0-1.4: Use destination register number as texture code.
1140 2.0+: Use provided sampler number as texure code. */
1141 if (shader_version < WINED3D_SHADER_VERSION(2,0))
1142 reg_sampler_code = dst->reg.idx;
1143 else
1144 reg_sampler_code = ins->src[1].reg.idx;
1145
1146 /* projection flag:
1147 * 1.1, 1.2, 1.3: Use WINED3DTSS_TEXTURETRANSFORMFLAGS
1148 * 1.4: Use WINED3DSPSM_DZ or WINED3DSPSM_DW on src[0]
1149 * 2.0+: Use WINED3DSI_TEXLD_PROJECT on the opcode
1150 */
1151 if (shader_version < WINED3D_SHADER_VERSION(1,4))
1152 {
1153 DWORD flags = 0;
1154 if(reg_sampler_code < MAX_TEXTURES) {
1155 flags = deviceImpl->stateBlock->textureState[reg_sampler_code][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1156 }
1157 if (flags & WINED3DTTFF_PROJECTED) {
1158 projected = TRUE;
1159 }
1160 }
1161 else if (shader_version < WINED3D_SHADER_VERSION(2,0))
1162 {
1163 DWORD src_mod = ins->src[0].modifiers;
1164 if (src_mod == WINED3DSPSM_DZ) {
1165 projected = TRUE;
1166 } else if(src_mod == WINED3DSPSM_DW) {
1167 projected = TRUE;
1168 }
1169 } else {
1170 if (ins->flags & WINED3DSI_TEXLD_PROJECT) projected = TRUE;
1171 if (ins->flags & WINED3DSI_TEXLD_BIAS) bias = TRUE;
1172 }
1173 shader_hw_sample(ins, reg_sampler_code, reg_dest, reg_coord, projected, bias);
1174 }
1175
1176 static void pshader_hw_texcoord(const struct wined3d_shader_instruction *ins)
1177 {
1178 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1179 SHADER_BUFFER *buffer = ins->ctx->buffer;
1180 DWORD shader_version = WINED3D_SHADER_VERSION(ins->ctx->reg_maps->shader_version.major,
1181 ins->ctx->reg_maps->shader_version.minor);
1182
1183 char tmp[20];
1184 shader_arb_get_write_mask(ins, dst, tmp);
1185 if (shader_version != WINED3D_SHADER_VERSION(1,4))
1186 {
1187 DWORD reg = dst->reg.idx;
1188 shader_addline(buffer, "MOV_SAT T%u%s, fragment.texcoord[%u];\n", reg, tmp, reg);
1189 } else {
1190 char reg_src[40];
1191
1192 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, reg_src);
1193 shader_addline(buffer, "MOV R%u%s, %s;\n", dst->reg.idx, tmp, reg_src);
1194 }
1195 }
1196
1197 static void pshader_hw_texreg2ar(const struct wined3d_shader_instruction *ins)
1198 {
1199 SHADER_BUFFER *buffer = ins->ctx->buffer;
1200 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1201 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1202 DWORD flags;
1203
1204 DWORD reg1 = ins->dst[0].reg.idx;
1205 char dst_str[8];
1206 char src_str[50];
1207
1208 sprintf(dst_str, "T%u", reg1);
1209 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src_str);
1210 shader_addline(buffer, "MOV TMP.x, %s.w;\n", src_str);
1211 shader_addline(buffer, "MOV TMP.y, %s.x;\n", src_str);
1212 flags = reg1 < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg1][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1213 shader_hw_sample(ins, reg1, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1214 }
1215
1216 static void pshader_hw_texreg2gb(const struct wined3d_shader_instruction *ins)
1217 {
1218 SHADER_BUFFER *buffer = ins->ctx->buffer;
1219
1220 DWORD reg1 = ins->dst[0].reg.idx;
1221 char dst_str[8];
1222 char src_str[50];
1223
1224 sprintf(dst_str, "T%u", reg1);
1225 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src_str);
1226 shader_addline(buffer, "MOV TMP.x, %s.y;\n", src_str);
1227 shader_addline(buffer, "MOV TMP.y, %s.z;\n", src_str);
1228 shader_hw_sample(ins, reg1, dst_str, "TMP", FALSE, FALSE);
1229 }
1230
1231 static void pshader_hw_texreg2rgb(const struct wined3d_shader_instruction *ins)
1232 {
1233 SHADER_BUFFER *buffer = ins->ctx->buffer;
1234 DWORD reg1 = ins->dst[0].reg.idx;
1235 char dst_str[8];
1236 char src_str[50];
1237
1238 sprintf(dst_str, "T%u", reg1);
1239 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src_str);
1240 shader_hw_sample(ins, reg1, dst_str, src_str, FALSE, FALSE);
1241 }
1242
1243 static void pshader_hw_texbem(const struct wined3d_shader_instruction *ins)
1244 {
1245 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1246 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1247 BOOL has_bumpmat = FALSE;
1248 BOOL has_luminance = FALSE;
1249 BOOL is_color;
1250 int i;
1251
1252 SHADER_BUFFER *buffer = ins->ctx->buffer;
1253
1254 char reg_coord[40];
1255 DWORD reg_dest_code;
1256
1257 /* All versions have a destination register */
1258 reg_dest_code = dst->reg.idx;
1259 /* Can directly use the name because texbem is only valid for <= 1.3 shaders */
1260 shader_arb_get_register_name(ins->ctx->shader, dst->reg.type,
1261 dst->reg.idx, !!dst->reg.rel_addr, reg_coord, &is_color);
1262
1263 for(i = 0; i < This->numbumpenvmatconsts; i++) {
1264 if (This->bumpenvmatconst[i].const_num != WINED3D_CONST_NUM_UNUSED
1265 && reg_dest_code == This->bumpenvmatconst[i].texunit)
1266 {
1267 has_bumpmat = TRUE;
1268 break;
1269 }
1270 }
1271 for(i = 0; i < This->numbumpenvmatconsts; i++) {
1272 if (This->luminanceconst[i].const_num != WINED3D_CONST_NUM_UNUSED
1273 && reg_dest_code == This->luminanceconst[i].texunit)
1274 {
1275 has_luminance = TRUE;
1276 break;
1277 }
1278 }
1279
1280 if(has_bumpmat) {
1281 DWORD src = ins->src[0].reg.idx;
1282
1283 /* Sampling the perturbation map in Tsrc was done already, including the signedness correction if needed */
1284
1285 shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, x, z, 0, 0;\n", reg_dest_code);
1286 shader_addline(buffer, "DP3 TMP.x, TMP2, T%u;\n", src);
1287 shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, y, w, 0, 0;\n", reg_dest_code);
1288 shader_addline(buffer, "DP3 TMP.y, TMP2, T%u;\n", src);
1289
1290 /* with projective textures, texbem only divides the static texture coord, not the displacement,
1291 * so we can't let the GL handle this.
1292 */
1293 if (((IWineD3DDeviceImpl*) This->baseShader.device)->stateBlock->textureState[reg_dest_code][WINED3DTSS_TEXTURETRANSFORMFLAGS]
1294 & WINED3DTTFF_PROJECTED) {
1295 shader_addline(buffer, "RCP TMP2.w, %s.w;\n", reg_coord);
1296 shader_addline(buffer, "MUL TMP2.xy, %s, TMP2.w;\n", reg_coord);
1297 shader_addline(buffer, "ADD TMP.xy, TMP, TMP2;\n");
1298 } else {
1299 shader_addline(buffer, "ADD TMP.xy, TMP, %s;\n", reg_coord);
1300 }
1301
1302 shader_hw_sample(ins, reg_dest_code, reg_coord, "TMP", FALSE, FALSE);
1303
1304 if (ins->handler_idx == WINED3DSIH_TEXBEML && has_luminance)
1305 {
1306 shader_addline(buffer, "MAD TMP, T%u.z, luminance%d.x, luminance%d.y;\n",
1307 src, reg_dest_code, reg_dest_code);
1308 shader_addline(buffer, "MUL %s, %s, TMP;\n", reg_coord, reg_coord);
1309 }
1310
1311 } else {
1312 DWORD tf;
1313 if(reg_dest_code < MAX_TEXTURES) {
1314 tf = ((IWineD3DDeviceImpl*) This->baseShader.device)->stateBlock->textureState[reg_dest_code][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1315 } else {
1316 tf = 0;
1317 }
1318 /* Without a bump matrix loaded, just sample with the unmodified coordinates */
1319 shader_hw_sample(ins, reg_dest_code, reg_coord, reg_coord, tf & WINED3DTTFF_PROJECTED, FALSE);
1320 }
1321 }
1322
1323 static void pshader_hw_texm3x2pad(const struct wined3d_shader_instruction *ins)
1324 {
1325 DWORD reg = ins->dst[0].reg.idx;
1326 SHADER_BUFFER *buffer = ins->ctx->buffer;
1327 char src0_name[50];
1328
1329 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src0_name);
1330 shader_addline(buffer, "DP3 TMP.x, T%u, %s;\n", reg, src0_name);
1331 }
1332
1333 static void pshader_hw_texm3x2tex(const struct wined3d_shader_instruction *ins)
1334 {
1335 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1336 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1337 DWORD flags;
1338 DWORD reg = ins->dst[0].reg.idx;
1339 SHADER_BUFFER *buffer = ins->ctx->buffer;
1340 char dst_str[8];
1341 char src0_name[50];
1342
1343 sprintf(dst_str, "T%u", reg);
1344 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src0_name);
1345 shader_addline(buffer, "DP3 TMP.y, T%u, %s;\n", reg, src0_name);
1346 flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1347 shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1348 }
1349
1350 static void pshader_hw_texm3x3pad(const struct wined3d_shader_instruction *ins)
1351 {
1352 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1353 DWORD reg = ins->dst[0].reg.idx;
1354 SHADER_BUFFER *buffer = ins->ctx->buffer;
1355 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1356 char src0_name[50];
1357
1358 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src0_name);
1359 shader_addline(buffer, "DP3 TMP.%c, T%u, %s;\n", 'x' + current_state->current_row, reg, src0_name);
1360 current_state->texcoord_w[current_state->current_row++] = reg;
1361 }
1362
1363 static void pshader_hw_texm3x3tex(const struct wined3d_shader_instruction *ins)
1364 {
1365 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1366 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1367 DWORD flags;
1368 DWORD reg = ins->dst[0].reg.idx;
1369 SHADER_BUFFER *buffer = ins->ctx->buffer;
1370 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1371 char dst_str[8];
1372 char src0_name[50];
1373
1374 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src0_name);
1375 shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
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(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1381 current_state->current_row = 0;
1382 }
1383
1384 static void pshader_hw_texm3x3vspec(const struct wined3d_shader_instruction *ins)
1385 {
1386 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1387 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1388 DWORD flags;
1389 DWORD reg = ins->dst[0].reg.idx;
1390 SHADER_BUFFER *buffer = ins->ctx->buffer;
1391 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1392 char dst_str[8];
1393 char src0_name[50];
1394
1395 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src0_name);
1396 shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
1397
1398 /* Construct the eye-ray vector from w coordinates */
1399 shader_addline(buffer, "MOV TMP2.x, fragment.texcoord[%u].w;\n", current_state->texcoord_w[0]);
1400 shader_addline(buffer, "MOV TMP2.y, fragment.texcoord[%u].w;\n", current_state->texcoord_w[1]);
1401 shader_addline(buffer, "MOV TMP2.z, fragment.texcoord[%u].w;\n", reg);
1402
1403 /* Calculate reflection vector
1404 */
1405 shader_addline(buffer, "DP3 TMP.w, TMP, TMP2;\n");
1406 /* The .w is ignored when sampling, so I can use TMP2.w to calculate dot(N, N) */
1407 shader_addline(buffer, "DP3 TMP2.w, TMP, TMP;\n");
1408 shader_addline(buffer, "RCP TMP2.w, TMP2.w;\n");
1409 shader_addline(buffer, "MUL TMP.w, TMP.w, TMP2.w;\n");
1410 shader_addline(buffer, "MUL TMP, TMP.w, TMP;\n");
1411 shader_addline(buffer, "MAD TMP, coefmul.x, TMP, -TMP2;\n");
1412
1413 /* Sample the texture using the calculated coordinates */
1414 sprintf(dst_str, "T%u", reg);
1415 flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1416 shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1417 current_state->current_row = 0;
1418 }
1419
1420 static void pshader_hw_texm3x3spec(const struct wined3d_shader_instruction *ins)
1421 {
1422 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1423 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1424 DWORD flags;
1425 DWORD reg = ins->dst[0].reg.idx;
1426 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1427 SHADER_BUFFER *buffer = ins->ctx->buffer;
1428 char dst_str[8];
1429 char src0_name[50];
1430 char src1_name[50];
1431
1432 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src0_name);
1433 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[1], 1, src1_name);
1434 shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
1435
1436 /* Calculate reflection vector.
1437 *
1438 * dot(N, E)
1439 * TMP.xyz = 2 * --------- * N - E
1440 * dot(N, N)
1441 *
1442 * Which normalizes the normal vector
1443 */
1444 shader_addline(buffer, "DP3 TMP.w, TMP, %s;\n", src1_name);
1445 shader_addline(buffer, "DP3 TMP2.w, TMP, TMP;\n");
1446 shader_addline(buffer, "RCP TMP2.w, TMP2.w;\n");
1447 shader_addline(buffer, "MUL TMP.w, TMP.w, TMP2.w;\n");
1448 shader_addline(buffer, "MUL TMP, TMP.w, TMP;\n");
1449 shader_addline(buffer, "MAD TMP, coefmul.x, TMP, -%s;\n", src1_name);
1450
1451 /* Sample the texture using the calculated coordinates */
1452 sprintf(dst_str, "T%u", reg);
1453 flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1454 shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1455 current_state->current_row = 0;
1456 }
1457
1458 static void pshader_hw_texdepth(const struct wined3d_shader_instruction *ins)
1459 {
1460 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1461 SHADER_BUFFER *buffer = ins->ctx->buffer;
1462 char dst_name[50];
1463 BOOL is_color;
1464
1465 /* texdepth has an implicit destination, the fragment depth value. It's only parameter,
1466 * which is essentially an input, is the destination register because it is the first
1467 * parameter. According to the msdn, this must be register r5, but let's keep it more flexible
1468 * here
1469 */
1470 shader_arb_get_register_name(ins->ctx->shader, dst->reg.type,
1471 dst->reg.idx, !!dst->reg.rel_addr, dst_name, &is_color);
1472
1473 /* According to the msdn, the source register(must be r5) is unusable after
1474 * the texdepth instruction, so we're free to modify it
1475 */
1476 shader_addline(buffer, "MIN %s.y, %s.y, one.y;\n", dst_name, dst_name);
1477
1478 /* How to deal with the special case dst_name.g == 0? if r != 0, then
1479 * the r * (1 / 0) will give infinity, which is clamped to 1.0, the correct
1480 * result. But if r = 0.0, then 0 * inf = 0, which is incorrect.
1481 */
1482 shader_addline(buffer, "RCP %s.y, %s.y;\n", dst_name, dst_name);
1483 shader_addline(buffer, "MUL TMP.x, %s.x, %s.y;\n", dst_name, dst_name);
1484 shader_addline(buffer, "MIN TMP.x, TMP.x, one.x;\n");
1485 shader_addline(buffer, "MAX result.depth, TMP.x, 0.0;\n");
1486 }
1487
1488 /** Process the WINED3DSIO_TEXDP3TEX instruction in ARB:
1489 * Take a 3-component dot product of the TexCoord[dstreg] and src,
1490 * then perform a 1D texture lookup from stage dstregnum, place into dst. */
1491 static void pshader_hw_texdp3tex(const struct wined3d_shader_instruction *ins)
1492 {
1493 SHADER_BUFFER *buffer = ins->ctx->buffer;
1494 DWORD sampler_idx = ins->dst[0].reg.idx;
1495 char src0[50];
1496 char dst_str[8];
1497
1498 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src0);
1499 shader_addline(buffer, "MOV TMP, 0.0;\n");
1500 shader_addline(buffer, "DP3 TMP.x, T%u, %s;\n", sampler_idx, src0);
1501
1502 sprintf(dst_str, "T%u", sampler_idx);
1503 shader_hw_sample(ins, sampler_idx, dst_str, "TMP", FALSE /* Only one coord, can't be projected */, FALSE);
1504 }
1505
1506 /** Process the WINED3DSIO_TEXDP3 instruction in ARB:
1507 * Take a 3-component dot product of the TexCoord[dstreg] and src. */
1508 static void pshader_hw_texdp3(const struct wined3d_shader_instruction *ins)
1509 {
1510 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1511 char src0[50];
1512 char dst_str[50];
1513 char dst_mask[6];
1514 SHADER_BUFFER *buffer = ins->ctx->buffer;
1515 BOOL is_color;
1516
1517 /* Handle output register */
1518 shader_arb_get_register_name(ins->ctx->shader, dst->reg.type,
1519 dst->reg.idx, !!dst->reg.rel_addr, dst_str, &is_color);
1520 shader_arb_get_write_mask(ins, dst, dst_mask);
1521
1522 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src0);
1523 shader_addline(buffer, "DP3 %s%s, T%u, %s;\n", dst_str, dst_mask, dst->reg.idx, src0);
1524
1525 /* TODO: Handle output modifiers */
1526 }
1527
1528 /** Process the WINED3DSIO_TEXM3X3 instruction in ARB
1529 * Perform the 3rd row of a 3x3 matrix multiply */
1530 static void pshader_hw_texm3x3(const struct wined3d_shader_instruction *ins)
1531 {
1532 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1533 SHADER_BUFFER *buffer = ins->ctx->buffer;
1534 char dst_str[50];
1535 char dst_mask[6];
1536 char src0[50];
1537 BOOL is_color;
1538
1539 shader_arb_get_register_name(ins->ctx->shader, dst->reg.type,
1540 dst->reg.idx, !!dst->reg.rel_addr, dst_str, &is_color);
1541 shader_arb_get_write_mask(ins, dst, dst_mask);
1542
1543 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src0);
1544 shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", dst->reg.idx, src0);
1545 shader_addline(buffer, "MOV %s%s, TMP;\n", dst_str, dst_mask);
1546
1547 /* TODO: Handle output modifiers */
1548 }
1549
1550 /** Process the WINED3DSIO_TEXM3X2DEPTH instruction in ARB:
1551 * Last row of a 3x2 matrix multiply, use the result to calculate the depth:
1552 * Calculate tmp0.y = TexCoord[dstreg] . src.xyz; (tmp0.x has already been calculated)
1553 * depth = (tmp0.y == 0.0) ? 1.0 : tmp0.x / tmp0.y
1554 */
1555 static void pshader_hw_texm3x2depth(const struct wined3d_shader_instruction *ins)
1556 {
1557 SHADER_BUFFER *buffer = ins->ctx->buffer;
1558 DWORD dst_reg = ins->dst[0].reg.idx;
1559 char src0[50];
1560
1561 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src0);
1562 shader_addline(buffer, "DP3 TMP.y, T%u, %s;\n", dst_reg, src0);
1563
1564 /* How to deal with the special case dst_name.g == 0? if r != 0, then
1565 * the r * (1 / 0) will give infinity, which is clamped to 1.0, the correct
1566 * result. But if r = 0.0, then 0 * inf = 0, which is incorrect.
1567 */
1568 shader_addline(buffer, "RCP TMP.y, TMP.y;\n");
1569 shader_addline(buffer, "MUL TMP.x, TMP.x, TMP.y;\n");
1570 shader_addline(buffer, "MIN TMP.x, TMP.x, one.x;\n");
1571 shader_addline(buffer, "MAX result.depth, TMP.x, 0.0;\n");
1572 }
1573
1574 /** Handles transforming all WINED3DSIO_M?x? opcodes for
1575 Vertex/Pixel shaders to ARB_vertex_program codes */
1576 static void shader_hw_mnxn(const struct wined3d_shader_instruction *ins)
1577 {
1578 int i;
1579 int nComponents = 0;
1580 struct wined3d_shader_dst_param tmp_dst = {{0}};
1581 struct wined3d_shader_src_param tmp_src[2] = {{{0}}};
1582 struct wined3d_shader_instruction tmp_ins;
1583
1584 memset(&tmp_ins, 0, sizeof(tmp_ins));
1585
1586 /* Set constants for the temporary argument */
1587 tmp_ins.ctx = ins->ctx;
1588 tmp_ins.dst_count = 1;
1589 tmp_ins.dst = &tmp_dst;
1590 tmp_ins.src_count = 2;
1591 tmp_ins.src = tmp_src;
1592
1593 switch(ins->handler_idx)
1594 {
1595 case WINED3DSIH_M4x4:
1596 nComponents = 4;
1597 tmp_ins.handler_idx = WINED3DSIH_DP4;
1598 break;
1599 case WINED3DSIH_M4x3:
1600 nComponents = 3;
1601 tmp_ins.handler_idx = WINED3DSIH_DP4;
1602 break;
1603 case WINED3DSIH_M3x4:
1604 nComponents = 4;
1605 tmp_ins.handler_idx = WINED3DSIH_DP3;
1606 break;
1607 case WINED3DSIH_M3x3:
1608 nComponents = 3;
1609 tmp_ins.handler_idx = WINED3DSIH_DP3;
1610 break;
1611 case WINED3DSIH_M3x2:
1612 nComponents = 2;
1613 tmp_ins.handler_idx = WINED3DSIH_DP3;
1614 break;
1615 default:
1616 FIXME("Unhandled opcode %#x\n", ins->handler_idx);
1617 break;
1618 }
1619
1620 tmp_dst = ins->dst[0];
1621 tmp_src[0] = ins->src[0];
1622 tmp_src[1] = ins->src[1];
1623 for (i = 0; i < nComponents; i++) {
1624 tmp_dst.write_mask = WINED3DSP_WRITEMASK_0 << i;
1625 shader_hw_map2gl(&tmp_ins);
1626 ++tmp_src[1].reg.idx;
1627 }
1628 }
1629
1630 static void vshader_hw_rsq_rcp(const struct wined3d_shader_instruction *ins)
1631 {
1632 SHADER_BUFFER *buffer = ins->ctx->buffer;
1633 const char *instruction;
1634
1635 char tmpLine[256];
1636
1637 switch(ins->handler_idx)
1638 {
1639 case WINED3DSIH_RSQ: instruction = "RSQ"; break;
1640 case WINED3DSIH_RCP: instruction = "RCP"; break;
1641 default: instruction = "";
1642 FIXME("Unhandled opcode %#x\n", ins->handler_idx);
1643 break;
1644 }
1645
1646 strcpy(tmpLine, instruction);
1647 shader_arb_add_dst_param(ins, &ins->dst[0], tmpLine); /* Destination */
1648 strcat(tmpLine, ",");
1649 shader_arb_add_src_param(ins, &ins->src[0], tmpLine);
1650 if (ins->src[0].swizzle == WINED3DSP_NOSWIZZLE)
1651 {
1652 /* Dx sdk says .x is used if no swizzle is given, but our test shows that
1653 * .w is used
1654 */
1655 strcat(tmpLine, ".w");
1656 }
1657
1658 shader_addline(buffer, "%s;\n", tmpLine);
1659 }
1660
1661 static void shader_hw_nrm(const struct wined3d_shader_instruction *ins)
1662 {
1663 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1664 SHADER_BUFFER *buffer = ins->ctx->buffer;
1665 char dst_name[50];
1666 char src_name[50];
1667 char dst_wmask[20];
1668 BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
1669 BOOL is_color;
1670
1671 shader_arb_get_register_name(ins->ctx->shader, dst->reg.type,
1672 dst->reg.idx, !!dst->reg.rel_addr, dst_name, &is_color);
1673 shader_arb_get_write_mask(ins, dst, dst_wmask);
1674
1675 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src_name);
1676 shader_addline(buffer, "DP3 TMP, %s, %s;\n", src_name, src_name);
1677 shader_addline(buffer, "RSQ TMP, TMP.x;\n");
1678 /* dst.w = src[0].w * 1 / (src.x^2 + src.y^2 + src.z^2)^(1/2) according to msdn*/
1679 shader_addline(buffer, "MUL%s %s%s, %s, TMP;\n", sat ? "_SAT" : "", dst_name, dst_wmask,
1680 src_name);
1681 }
1682
1683 static void shader_hw_sincos(const struct wined3d_shader_instruction *ins)
1684 {
1685 /* This instruction exists in ARB, but the d3d instruction takes two extra parameters which
1686 * must contain fixed constants. So we need a separate function to filter those constants and
1687 * can't use map2gl
1688 */
1689 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1690 SHADER_BUFFER *buffer = ins->ctx->buffer;
1691 char dst_name[50];
1692 char src_name[50];
1693 char dst_wmask[20];
1694 BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
1695 BOOL is_color;
1696
1697 shader_arb_get_register_name(ins->ctx->shader, dst->reg.type,
1698 dst->reg.idx, !!dst->reg.rel_addr, dst_name, &is_color);
1699 shader_arb_get_write_mask(ins, dst, dst_wmask);
1700
1701 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, &ins->src[0], 0, src_name);
1702 shader_addline(buffer, "SCS%s %s%s, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask,
1703 src_name);
1704 }
1705
1706 static GLuint create_arb_blt_vertex_program(const WineD3D_GL_Info *gl_info)
1707 {
1708 GLuint program_id = 0;
1709 const char *blt_vprogram =
1710 "!!ARBvp1.0\n"
1711 "PARAM c[1] = { { 1, 0.5 } };\n"
1712 "MOV result.position, vertex.position;\n"
1713 "MOV result.color, c[0].x;\n"
1714 "MOV result.texcoord[0], vertex.texcoord[0];\n"
1715 "END\n";
1716
1717 GL_EXTCALL(glGenProgramsARB(1, &program_id));
1718 GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, program_id));
1719 GL_EXTCALL(glProgramStringARB(GL_VERTEX_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(blt_vprogram), blt_vprogram));
1720
1721 if (glGetError() == GL_INVALID_OPERATION) {
1722 GLint pos;
1723 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
1724 FIXME("Vertex program error at position %d: %s\n", pos,
1725 debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
1726 }
1727
1728 return program_id;
1729 }
1730
1731 static GLuint create_arb_blt_fragment_program(const WineD3D_GL_Info *gl_info, enum tex_types tex_type)
1732 {
1733 GLuint program_id = 0;
1734 static const char * const blt_fprograms[tex_type_count] =
1735 {
1736 /* tex_1d */
1737 NULL,
1738 /* tex_2d */
1739 "!!ARBfp1.0\n"
1740 "TEMP R0;\n"
1741 "TEX R0.x, fragment.texcoord[0], texture[0], 2D;\n"
1742 "MOV result.depth.z, R0.x;\n"
1743 "END\n",
1744 /* tex_3d */
1745 NULL,
1746 /* tex_cube */
1747 "!!ARBfp1.0\n"
1748 "TEMP R0;\n"
1749 "TEX R0.x, fragment.texcoord[0], texture[0], CUBE;\n"
1750 "MOV result.depth.z, R0.x;\n"
1751 "END\n",
1752 /* tex_rect */
1753 "!!ARBfp1.0\n"
1754 "TEMP R0;\n"
1755 "TEX R0.x, fragment.texcoord[0], texture[0], RECT;\n"
1756 "MOV result.depth.z, R0.x;\n"
1757 "END\n",
1758 };
1759
1760 if (!blt_fprograms[tex_type])
1761 {
1762 FIXME("tex_type %#x not supported\n", tex_type);
1763 tex_type = tex_2d;
1764 }
1765
1766 GL_EXTCALL(glGenProgramsARB(1, &program_id));
1767 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, program_id));
1768 GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(blt_fprograms[tex_type]), blt_fprograms[tex_type]));
1769
1770 if (glGetError() == GL_INVALID_OPERATION) {
1771 GLint pos;
1772 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
1773 FIXME("Fragment program error at position %d: %s\n", pos,
1774 debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
1775 }
1776
1777 return program_id;
1778 }
1779
1780 static void shader_arb_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {
1781 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1782 struct shader_arb_priv *priv = This->shader_priv;
1783 const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1784
1785 if (useVS) {
1786 struct vs_compile_args compile_args;
1787
1788 TRACE("Using vertex shader\n");
1789 find_vs_compile_args((IWineD3DVertexShaderImpl *) This->stateBlock->vertexShader, This->stateBlock, &compile_args);
1790 priv->current_vprogram_id = find_gl_vshader((IWineD3DVertexShaderImpl *) This->stateBlock->vertexShader, &compile_args);
1791
1792 /* Bind the vertex program */
1793 GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id));
1794 checkGLcall("glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id);");
1795
1796 /* Enable OpenGL vertex programs */
1797 glEnable(GL_VERTEX_PROGRAM_ARB);
1798 checkGLcall("glEnable(GL_VERTEX_PROGRAM_ARB);");
1799 TRACE("(%p) : Bound vertex program %u and enabled GL_VERTEX_PROGRAM_ARB\n", This, priv->current_vprogram_id);
1800 } else if(GL_SUPPORT(ARB_VERTEX_PROGRAM)) {
1801 priv->current_vprogram_id = 0;
1802 glDisable(GL_VERTEX_PROGRAM_ARB);
1803 checkGLcall("glDisable(GL_VERTEX_PROGRAM_ARB)");
1804 }
1805
1806 if (usePS) {
1807 struct ps_compile_args compile_args;
1808 TRACE("Using pixel shader\n");
1809 find_ps_compile_args((IWineD3DPixelShaderImpl *) This->stateBlock->pixelShader, This->stateBlock, &compile_args);
1810 priv->current_fprogram_id = find_gl_pshader((IWineD3DPixelShaderImpl *) This->stateBlock->pixelShader,
1811 &compile_args);
1812
1813 /* Bind the fragment program */
1814 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id));
1815 checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id);");
1816
1817 if(!priv->use_arbfp_fixed_func) {
1818 /* Enable OpenGL fragment programs */
1819 glEnable(GL_FRAGMENT_PROGRAM_ARB);
1820 checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB);");
1821 }
1822 TRACE("(%p) : Bound fragment program %u and enabled GL_FRAGMENT_PROGRAM_ARB\n", This, priv->current_fprogram_id);
1823 } else if(GL_SUPPORT(ARB_FRAGMENT_PROGRAM) && !priv->use_arbfp_fixed_func) {
1824 /* Disable only if we're not using arbfp fixed function fragment processing. If this is used,
1825 * keep GL_FRAGMENT_PROGRAM_ARB enabled, and the fixed function pipeline will bind the fixed function
1826 * replacement shader
1827 */
1828 glDisable(GL_FRAGMENT_PROGRAM_ARB);
1829 checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
1830 priv->current_fprogram_id = 0;
1831 }
1832 }
1833
1834 static void shader_arb_select_depth_blt(IWineD3DDevice *iface, enum tex_types tex_type) {
1835 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1836 struct shader_arb_priv *priv = This->shader_priv;
1837 GLuint *blt_fprogram = &priv->depth_blt_fprogram_id[tex_type];
1838 const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1839
1840 if (!priv->depth_blt_vprogram_id) priv->depth_blt_vprogram_id = create_arb_blt_vertex_program(gl_info);
1841 GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->depth_blt_vprogram_id));
1842 glEnable(GL_VERTEX_PROGRAM_ARB);
1843
1844 if (!*blt_fprogram) *blt_fprogram = create_arb_blt_fragment_program(gl_info, tex_type);
1845 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, *blt_fprogram));
1846 glEnable(GL_FRAGMENT_PROGRAM_ARB);
1847 }
1848
1849 static void shader_arb_deselect_depth_blt(IWineD3DDevice *iface) {
1850 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1851 struct shader_arb_priv *priv = This->shader_priv;
1852 const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1853
1854 if (priv->current_vprogram_id) {
1855 GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id));
1856 checkGLcall("glBindProgramARB(GL_VERTEX_PROGRAM_ARB, vertexShader->prgId);");
1857
1858 glEnable(GL_VERTEX_PROGRAM_ARB);
1859 checkGLcall("glEnable(GL_VERTEX_PROGRAM_ARB);");
1860
1861 TRACE("(%p) : Bound vertex program %u and enabled GL_VERTEX_PROGRAM_ARB\n", This, priv->current_vprogram_id);
1862 } else {
1863 glDisable(GL_VERTEX_PROGRAM_ARB);
1864 checkGLcall("glDisable(GL_VERTEX_PROGRAM_ARB)");
1865 }
1866
1867 if (priv->current_fprogram_id) {
1868 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id));
1869 checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, pixelShader->prgId);");
1870
1871 glEnable(GL_FRAGMENT_PROGRAM_ARB);
1872 checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB);");
1873
1874 TRACE("(%p) : Bound fragment program %u and enabled GL_FRAGMENT_PROGRAM_ARB\n", This, priv->current_fprogram_id);
1875 } else {
1876 glDisable(GL_FRAGMENT_PROGRAM_ARB);
1877 checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
1878 }
1879 }
1880
1881 static void shader_arb_destroy(IWineD3DBaseShader *iface) {
1882 IWineD3DBaseShaderImpl *baseShader = (IWineD3DBaseShaderImpl *) iface;
1883 const WineD3D_GL_Info *gl_info = &((IWineD3DDeviceImpl *)baseShader->baseShader.device)->adapter->gl_info;
1884
1885 if (shader_is_pshader_version(baseShader->baseShader.reg_maps.shader_version.type))
1886 {
1887 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *) iface;
1888 UINT i;
1889
1890 ENTER_GL();
1891 for(i = 0; i < This->num_gl_shaders; i++) {
1892 GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId));
1893 checkGLcall("GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId))");
1894 }
1895 LEAVE_GL();
1896 HeapFree(GetProcessHeap(), 0, This->gl_shaders);
1897 This->gl_shaders = NULL;
1898 This->num_gl_shaders = 0;
1899 This->shader_array_size = 0;
1900 } else {
1901 IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *) iface;
1902 UINT i;
1903
1904 ENTER_GL();
1905 for(i = 0; i < This->num_gl_shaders; i++) {
1906 GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId));
1907 checkGLcall("GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId))");
1908 }
1909 LEAVE_GL();
1910 HeapFree(GetProcessHeap(), 0, This->gl_shaders);
1911 This->gl_shaders = NULL;
1912 This->num_gl_shaders = 0;
1913 This->shader_array_size = 0;
1914 }
1915 }
1916
1917 static HRESULT shader_arb_alloc(IWineD3DDevice *iface) {
1918 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1919 This->shader_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct shader_arb_priv));
1920 return WINED3D_OK;
1921 }
1922
1923 static void shader_arb_free(IWineD3DDevice *iface) {
1924 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1925 const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1926 struct shader_arb_priv *priv = This->shader_priv;
1927 int i;
1928
1929 if(priv->depth_blt_vprogram_id) {
1930 GL_EXTCALL(glDeleteProgramsARB(1, &priv->depth_blt_vprogram_id));
1931 }
1932 for (i = 0; i < tex_type_count; ++i) {
1933 if (priv->depth_blt_fprogram_id[i]) {
1934 GL_EXTCALL(glDeleteProgramsARB(1, &priv->depth_blt_fprogram_id[i]));
1935 }
1936 }
1937
1938 HeapFree(GetProcessHeap(), 0, This->shader_priv);
1939 }
1940
1941 static BOOL shader_arb_dirty_const(IWineD3DDevice *iface) {
1942 return TRUE;
1943 }
1944
1945 static void arbfp_add_sRGB_correction(SHADER_BUFFER *buffer, const char *fragcolor, const char *tmp1,
1946 const char *tmp2, const char *tmp3, const char *tmp4) {
1947 /* Perform sRGB write correction. See GLX_EXT_framebuffer_sRGB */
1948
1949 /* Calculate the > 0.0031308 case */
1950 shader_addline(buffer, "POW %s.x, %s.x, srgb_consts1.z;\n", tmp1, fragcolor);
1951 shader_addline(buffer, "POW %s.y, %s.y, srgb_consts1.z;\n", tmp1, fragcolor);
1952 shader_addline(buffer, "POW %s.z, %s.z, srgb_consts1.z;\n", tmp1, fragcolor);
1953 shader_addline(buffer, "MUL %s, %s, srgb_consts1.w;\n", tmp1, tmp1);
1954 shader_addline(buffer, "SUB %s, %s, srgb_consts2.x;\n", tmp1, tmp1);
1955 /* Calculate the < case */
1956 shader_addline(buffer, "MUL %s, srgb_consts1.x, %s;\n", tmp2, fragcolor);
1957 /* Get 1.0 / 0.0 masks for > 0.0031308 and < 0.0031308 */
1958 shader_addline(buffer, "SLT %s, srgb_consts1.y, %s;\n", tmp3, fragcolor);
1959 shader_addline(buffer, "SGE %s, srgb_consts1.y, %s;\n", tmp4, fragcolor);
1960 /* Store the components > 0.0031308 in the destination */
1961 shader_addline(buffer, "MUL %s, %s, %s;\n", fragcolor, tmp1, tmp3);
1962 /* Add the components that are < 0.0031308 */
1963 shader_addline(buffer, "MAD result.color.xyz, %s, %s, %s;\n", tmp2, tmp4, fragcolor);
1964 /* [0.0;1.0] clamping. Not needed, this is done implicitly */
1965 }
1966
1967 static GLuint shader_arb_generate_pshader(IWineD3DPixelShader *iface,
1968 SHADER_BUFFER *buffer, const struct ps_compile_args *args)
1969 {
1970 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
1971 const shader_reg_maps* reg_maps = &This->baseShader.reg_maps;
1972 CONST DWORD *function = This->baseShader.function;
1973 const WineD3D_GL_Info *gl_info = &((IWineD3DDeviceImpl *)This->baseShader.device)->adapter->gl_info;
1974 const local_constant *lconst;
1975 GLuint retval;
1976 const char *fragcolor;
1977 DWORD *lconst_map = local_const_mapping((IWineD3DBaseShaderImpl *) This);
1978
1979 /* Create the hw ARB shader */
1980 shader_addline(buffer, "!!ARBfp1.0\n");
1981
1982 if (reg_maps->shader_version.major < 3)
1983 {
1984 switch(args->fog) {
1985 case FOG_OFF:
1986 break;
1987 case FOG_LINEAR:
1988 shader_addline(buffer, "OPTION ARB_fog_linear;\n");
1989 break;
1990 case FOG_EXP:
1991 shader_addline(buffer, "OPTION ARB_fog_exp;\n");
1992 break;
1993 case FOG_EXP2:
1994 shader_addline(buffer, "OPTION ARB_fog_exp2;\n");
1995 break;
1996 }
1997 }
1998
1999 shader_addline(buffer, "TEMP TMP;\n"); /* Used in matrix ops */
2000 shader_addline(buffer, "TEMP TMP2;\n"); /* Used in matrix ops */
2001 shader_addline(buffer, "TEMP TA;\n"); /* Used for modifiers */
2002 shader_addline(buffer, "TEMP TB;\n"); /* Used for modifiers */
2003 shader_addline(buffer, "TEMP TC;\n"); /* Used for modifiers */
2004 shader_addline(buffer, "PARAM coefdiv = { 0.5, 0.25, 0.125, 0.0625 };\n");
2005 shader_addline(buffer, "PARAM coefmul = { 2, 4, 8, 16 };\n");
2006 shader_addline(buffer, "PARAM one = { 1.0, 1.0, 1.0, 1.0 };\n");
2007
2008 if (reg_maps->shader_version.major < 2)
2009 {
2010 fragcolor = "R0";
2011 } else {
2012 shader_addline(buffer, "TEMP TMP_COLOR;\n");
2013 fragcolor = "TMP_COLOR";
2014 }
2015
2016 /* Base Declarations */
2017 shader_generate_arb_declarations( (IWineD3DBaseShader*) This, reg_maps, buffer, &GLINFO_LOCATION, lconst_map);
2018
2019 /* Base Shader Body */
2020 shader_generate_main((IWineD3DBaseShader *)This, buffer, reg_maps, function);
2021
2022 if(args->srgb_correction) {
2023 arbfp_add_sRGB_correction(buffer, fragcolor, "TMP", "TMP2", "TA", "TB");
2024 }
2025 shader_addline(buffer, "MOV result.color, %s;\n", fragcolor);
2026 shader_addline(buffer, "END\n");
2027
2028 /* TODO: change to resource.glObjectHandle or something like that */
2029 GL_EXTCALL(glGenProgramsARB(1, &retval));
2030
2031 TRACE("Creating a hw pixel shader, prg=%d\n", retval);
2032 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, retval));
2033
2034 TRACE("Created hw pixel shader, prg=%d\n", retval);
2035 /* Create the program and check for errors */
2036 GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB,
2037 buffer->bsize, buffer->buffer));
2038
2039 if (glGetError() == GL_INVALID_OPERATION) {
2040 GLint errPos;
2041 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &errPos);
2042 FIXME("HW PixelShader Error at position %d: %s\n",
2043 errPos, debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
2044 retval = 0;
2045 }
2046
2047 /* Load immediate constants */
2048 if(lconst_map) {
2049 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
2050 const float *value = (const float *)lconst->value;
2051 GL_EXTCALL(glProgramLocalParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, lconst_map[lconst->idx], value));
2052 checkGLcall("glProgramLocalParameter4fvARB");
2053 }
2054 HeapFree(GetProcessHeap(), 0, lconst_map);
2055 }
2056
2057 return retval;
2058 }
2059
2060 static GLuint shader_arb_generate_vshader(IWineD3DVertexShader *iface,
2061 SHADER_BUFFER *buffer, const struct vs_compile_args *args)
2062 {
2063 IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *)iface;
2064 const shader_reg_maps *reg_maps = &This->baseShader.reg_maps;
2065 CONST DWORD *function = This->baseShader.function;
2066 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)This->baseShader.device;
2067 const WineD3D_GL_Info *gl_info = &device->adapter->gl_info;
2068 const local_constant *lconst;
2069 GLuint ret;
2070 DWORD *lconst_map = local_const_mapping((IWineD3DBaseShaderImpl *) This);
2071
2072 /* Create the hw ARB shader */
2073 shader_addline(buffer, "!!ARBvp1.0\n");
2074 shader_addline(buffer, "PARAM helper_const = { 2.0, -1.0, %d.0, 0.0 };\n", This->rel_offset);
2075
2076 /* Mesa supports only 95 constants */
2077 if (GL_VEND(MESA) || GL_VEND(WINE))
2078 This->baseShader.limits.constant_float =
2079 min(95, This->baseShader.limits.constant_float);
2080
2081 shader_addline(buffer, "TEMP TMP;\n");
2082
2083 /* Base Declarations */
2084 shader_generate_arb_declarations( (IWineD3DBaseShader*) This, reg_maps, buffer, &GLINFO_LOCATION, lconst_map);
2085
2086 /* We need a constant to fixup the final position */
2087 shader_addline(buffer, "PARAM posFixup = program.env[%d];\n", ARB_SHADER_PRIVCONST_POS);
2088
2089 /* Initialize output parameters. GL_ARB_vertex_program does not require special initialization values
2090 * for output parameters. D3D in theory does not do that either, but some applications depend on a
2091 * proper initialization of the secondary color, and programs using the fixed function pipeline without
2092 * a replacement shader depend on the texcoord.w being set properly.
2093 *
2094 * GL_NV_vertex_program defines that all output values are initialized to {0.0, 0.0, 0.0, 1.0}. This
2095 * assertion is in effect even when using GL_ARB_vertex_program without any NV specific additions. So
2096 * skip this if NV_vertex_program is supported. Otherwise, initialize the secondary color. For the tex-
2097 * coords, we have a flag in the opengl caps. Many cards do not require the texcoord being set, and
2098 * this can eat a number of instructions, so skip it unless this cap is set as well
2099 */
2100 if(!GL_SUPPORT(NV_VERTEX_PROGRAM)) {
2101 shader_addline(buffer, "MOV result.color.secondary, -helper_const.wwwy;\n");
2102
2103 if((GLINFO_LOCATION).set_texcoord_w && !device->frag_pipe->ffp_proj_control) {
2104 int i;
2105 for(i = 0; i < min(8, MAX_REG_TEXCRD); i++) {
2106 if(This->baseShader.reg_maps.texcoord_mask[i] != 0 &&
2107 This->baseShader.reg_maps.texcoord_mask[i] != WINED3DSP_WRITEMASK_ALL) {
2108 shader_addline(buffer, "MOV result.texcoord[%u].w, -helper_const.y;\n", i);
2109 }
2110 }
2111 }
2112 }
2113
2114 /* Base Shader Body */
2115 shader_generate_main((IWineD3DBaseShader *)This, buffer, reg_maps, function);
2116
2117 /* The D3DRS_FOGTABLEMODE render state defines if the shader-generated fog coord is used
2118 * or if the fragment depth is used. If the fragment depth is used(FOGTABLEMODE != NONE),
2119 * the fog frag coord is thrown away. If the fog frag coord is used, but not written by
2120 * the shader, it is set to 0.0(fully fogged, since start = 1.0, end = 0.0)
2121 */
2122 if(args->fog_src == VS_FOG_Z) {
2123 shader_addline(buffer, "MOV result.fogcoord, TMP_OUT.z;\n");
2124 } else if (!reg_maps->fog) {
2125 shader_addline(buffer, "MOV result.fogcoord, helper_const.w;\n");
2126 }
2127
2128 /* Write the final position.
2129 *
2130 * OpenGL coordinates specify the center of the pixel while d3d coords specify
2131 * the corner. The offsets are stored in z and w in posFixup. posFixup.y contains
2132 * 1.0 or -1.0 to turn the rendering upside down for offscreen rendering. PosFixup.x
2133 * contains 1.0 to allow a mad, but arb vs swizzles are too restricted for that.
2134 */
2135 shader_addline(buffer, "MUL TMP, posFixup, TMP_OUT.w;\n");
2136 shader_addline(buffer, "ADD TMP_OUT.x, TMP_OUT.x, TMP.z;\n");
2137 shader_addline(buffer, "MAD TMP_OUT.y, TMP_OUT.y, posFixup.y, TMP.w;\n");
2138
2139 /* Z coord [0;1]->[-1;1] mapping, see comment in transform_projection in state.c
2140 * and the glsl equivalent
2141 */
2142 shader_addline(buffer, "MAD TMP_OUT.z, TMP_OUT.z, helper_const.x, -TMP_OUT.w;\n");
2143
2144 shader_addline(buffer, "MOV result.position, TMP_OUT;\n");
2145
2146 shader_addline(buffer, "END\n");
2147
2148 /* TODO: change to resource.glObjectHandle or something like that */
2149 GL_EXTCALL(glGenProgramsARB(1, &ret));
2150
2151 TRACE("Creating a hw vertex shader, prg=%d\n", ret);
2152 GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, ret));
2153
2154 TRACE("Created hw vertex shader, prg=%d\n", ret);
2155 /* Create the program and check for errors */
2156 GL_EXTCALL(glProgramStringARB(GL_VERTEX_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB,
2157 buffer->bsize, buffer->buffer));
2158
2159 if (glGetError() == GL_INVALID_OPERATION) {
2160 GLint errPos;
2161 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &errPos);
2162 FIXME("HW VertexShader Error at position %d: %s\n",
2163 errPos, debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
2164 ret = -1;
2165 } else {
2166 /* Load immediate constants */
2167 if(lconst_map) {
2168 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
2169 const float *value = (const float *)lconst->value;
2170 GL_EXTCALL(glProgramLocalParameter4fvARB(GL_VERTEX_PROGRAM_ARB, lconst_map[lconst->idx], value));
2171 }
2172 }
2173 }
2174 HeapFree(GetProcessHeap(), 0, lconst_map);
2175
2176 return ret;
2177 }
2178
2179 static void shader_arb_get_caps(WINED3DDEVTYPE devtype, const WineD3D_GL_Info *gl_info, struct shader_caps *pCaps)
2180 {
2181 /* We don't have an ARB fixed function pipeline yet, so let the none backend set its caps,
2182 * then overwrite the shader specific ones
2183 */
2184 none_shader_backend.shader_get_caps(devtype, gl_info, pCaps);
2185
2186 if(GL_SUPPORT(ARB_VERTEX_PROGRAM)) {
2187 pCaps->VertexShaderVersion = WINED3DVS_VERSION(1,1);
2188 TRACE_(d3d_caps)("Hardware vertex shader version 1.1 enabled (ARB_PROGRAM)\n");
2189 pCaps->MaxVertexShaderConst = GL_LIMITS(vshader_constantsF) - ARB_SHADER_RESERVED_VS_CONSTS;
2190 }
2191
2192 if(GL_SUPPORT(ARB_FRAGMENT_PROGRAM)) {
2193 pCaps->PixelShaderVersion = WINED3DPS_VERSION(1,4);
2194 pCaps->PixelShader1xMaxValue = 8.0;
2195 TRACE_(d3d_caps)("Hardware pixel shader version 1.4 enabled (ARB_PROGRAM)\n");
2196 pCaps->MaxPixelShaderConst = GL_LIMITS(pshader_constantsF) - ARB_SHADER_RESERVED_PS_CONSTS;
2197 }
2198 }
2199
2200 static BOOL shader_arb_color_fixup_supported(struct color_fixup_desc fixup)
2201 {
2202 if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
2203 {
2204 TRACE("Checking support for color_fixup:\n");
2205 dump_color_fixup_desc(fixup);
2206 }
2207
2208 /* We support everything except YUV conversions. */
2209 if (!is_yuv_fixup(fixup))
2210 {
2211 TRACE("[OK]\n");
2212 return TRUE;
2213 }
2214
2215 TRACE("[FAILED]\n");
2216 return FALSE;
2217 }
2218
2219 static void shader_arb_add_instruction_modifiers(const struct wined3d_shader_instruction *ins) {
2220 BOOL saturate;
2221 DWORD shift;
2222 char write_mask[20], regstr[50];
2223 SHADER_BUFFER *buffer = ins->ctx->buffer;
2224 BOOL is_color = FALSE;
2225 const struct wined3d_shader_dst_param *dst;
2226
2227 if (!ins->dst_count) return;
2228
2229 dst = &ins->dst[0];
2230 shift = dst->shift;
2231 if(shift == 0) return; /* Saturate alone is handled by the instructions */
2232 saturate = dst->modifiers & WINED3DSPDM_SATURATE;
2233
2234 shader_arb_get_write_mask(ins, dst, write_mask);
2235 shader_arb_get_register_name(ins->ctx->shader, dst->reg.type,
2236 dst->reg.idx, !!dst->reg.rel_addr, regstr, &is_color);
2237
2238 /* Generate a line that does the output modifier computation */
2239 shader_addline(buffer, "MUL%s %s%s, %s, %s;\n", saturate ? "_SAT" : "",
2240 regstr, write_mask, regstr, shift_tab[shift]);
2241 }
2242
2243 static const SHADER_HANDLER shader_arb_instruction_handler_table[WINED3DSIH_TABLE_SIZE] =
2244 {
2245 /* WINED3DSIH_ABS */ shader_hw_map2gl,
2246 /* WINED3DSIH_ADD */ shader_hw_map2gl,
2247 /* WINED3DSIH_BEM */ pshader_hw_bem,
2248 /* WINED3DSIH_BREAK */ NULL,
2249 /* WINED3DSIH_BREAKC */ NULL,
2250 /* WINED3DSIH_BREAKP */ NULL,
2251 /* WINED3DSIH_CALL */ NULL,
2252 /* WINED3DSIH_CALLNZ */ NULL,
2253 /* WINED3DSIH_CMP */ pshader_hw_cmp,
2254 /* WINED3DSIH_CND */ pshader_hw_cnd,
2255 /* WINED3DSIH_CRS */ shader_hw_map2gl,
2256 /* WINED3DSIH_DCL */ NULL,
2257 /* WINED3DSIH_DEF */ NULL,
2258 /* WINED3DSIH_DEFB */ NULL,
2259 /* WINED3DSIH_DEFI */ NULL,
2260 /* WINED3DSIH_DP2ADD */ pshader_hw_dp2add,
2261 /* WINED3DSIH_DP3 */ shader_hw_map2gl,
2262 /* WINED3DSIH_DP4 */ shader_hw_map2gl,
2263 /* WINED3DSIH_DST */ shader_hw_map2gl,
2264 /* WINED3DSIH_DSX */ NULL,
2265 /* WINED3DSIH_DSY */ NULL,
2266 /* WINED3DSIH_ELSE */ NULL,
2267 /* WINED3DSIH_ENDIF */ NULL,
2268 /* WINED3DSIH_ENDLOOP */ NULL,
2269 /* WINED3DSIH_ENDREP */ NULL,
2270 /* WINED3DSIH_EXP */ shader_hw_map2gl,
2271 /* WINED3DSIH_EXPP */ shader_hw_map2gl,
2272 /* WINED3DSIH_FRC */ shader_hw_map2gl,
2273 /* WINED3DSIH_IF */ NULL,
2274 /* WINED3DSIH_IFC */ NULL,
2275 /* WINED3DSIH_LABEL */ NULL,
2276 /* WINED3DSIH_LIT */ shader_hw_map2gl,
2277 /* WINED3DSIH_LOG */ shader_hw_map2gl,
2278 /* WINED3DSIH_LOGP */ shader_hw_map2gl,
2279 /* WINED3DSIH_LOOP */ NULL,
2280 /* WINED3DSIH_LRP */ shader_hw_map2gl,
2281 /* WINED3DSIH_M3x2 */ shader_hw_mnxn,
2282 /* WINED3DSIH_M3x3 */ shader_hw_mnxn,
2283 /* WINED3DSIH_M3x4 */ shader_hw_mnxn,
2284 /* WINED3DSIH_M4x3 */ shader_hw_mnxn,
2285 /* WINED3DSIH_M4x4 */ shader_hw_mnxn,
2286 /* WINED3DSIH_MAD */ shader_hw_map2gl,
2287 /* WINED3DSIH_MAX */ shader_hw_map2gl,
2288 /* WINED3DSIH_MIN */ shader_hw_map2gl,
2289 /* WINED3DSIH_MOV */ shader_hw_mov,
2290 /* WINED3DSIH_MOVA */ shader_hw_mov,
2291 /* WINED3DSIH_MUL */ shader_hw_map2gl,
2292 /* WINED3DSIH_NOP */ shader_hw_map2gl,
2293 /* WINED3DSIH_NRM */ shader_hw_nrm,
2294 /* WINED3DSIH_PHASE */ NULL,
2295 /* WINED3DSIH_POW */ shader_hw_map2gl,
2296 /* WINED3DSIH_RCP */ vshader_hw_rsq_rcp,
2297 /* WINED3DSIH_REP */ NULL,
2298 /* WINED3DSIH_RET */ NULL,
2299 /* WINED3DSIH_RSQ */ vshader_hw_rsq_rcp,
2300 /* WINED3DSIH_SETP */ NULL,
2301 /* WINED3DSIH_SGE */ shader_hw_map2gl,
2302 /* WINED3DSIH_SGN */ NULL,
2303 /* WINED3DSIH_SINCOS */ shader_hw_sincos,
2304 /* WINED3DSIH_SLT */ shader_hw_map2gl,
2305 /* WINED3DSIH_SUB */ shader_hw_map2gl,
2306 /* WINED3DSIH_TEX */ pshader_hw_tex,
2307 /* WINED3DSIH_TEXBEM */ pshader_hw_texbem,
2308 /* WINED3DSIH_TEXBEML */ pshader_hw_texbem,
2309 /* WINED3DSIH_TEXCOORD */ pshader_hw_texcoord,
2310 /* WINED3DSIH_TEXDEPTH */ pshader_hw_texdepth,
2311 /* WINED3DSIH_TEXDP3 */ pshader_hw_texdp3,
2312 /* WINED3DSIH_TEXDP3TEX */ pshader_hw_texdp3tex,
2313 /* WINED3DSIH_TEXKILL */ pshader_hw_texkill,
2314 /* WINED3DSIH_TEXLDD */ NULL,
2315 /* WINED3DSIH_TEXLDL */ NULL,
2316 /* WINED3DSIH_TEXM3x2DEPTH */ pshader_hw_texm3x2depth,
2317 /* WINED3DSIH_TEXM3x2PAD */ pshader_hw_texm3x2pad,
2318 /* WINED3DSIH_TEXM3x2TEX */ pshader_hw_texm3x2tex,
2319 /* WINED3DSIH_TEXM3x3 */ pshader_hw_texm3x3,
2320 /* WINED3DSIH_TEXM3x3DIFF */ NULL,
2321 /* WINED3DSIH_TEXM3x3PAD */ pshader_hw_texm3x3pad,
2322 /* WINED3DSIH_TEXM3x3SPEC */ pshader_hw_texm3x3spec,
2323 /* WINED3DSIH_TEXM3x3TEX */ pshader_hw_texm3x3tex,
2324 /* WINED3DSIH_TEXM3x3VSPEC */ pshader_hw_texm3x3vspec,
2325 /* WINED3DSIH_TEXREG2AR */ pshader_hw_texreg2ar,
2326 /* WINED3DSIH_TEXREG2GB */ pshader_hw_texreg2gb,
2327 /* WINED3DSIH_TEXREG2RGB */ pshader_hw_texreg2rgb,
2328 };
2329
2330 const shader_backend_t arb_program_shader_backend = {
2331 shader_arb_instruction_handler_table,
2332 shader_arb_select,
2333 shader_arb_select_depth_blt,
2334 shader_arb_deselect_depth_blt,
2335 shader_arb_update_float_vertex_constants,
2336 shader_arb_update_float_pixel_constants,
2337 shader_arb_load_constants,
2338 shader_arb_load_np2fixup_constants,
2339 shader_arb_destroy,
2340 shader_arb_alloc,
2341 shader_arb_free,
2342 shader_arb_dirty_const,
2343 shader_arb_generate_pshader,
2344 shader_arb_generate_vshader,
2345 shader_arb_get_caps,
2346 shader_arb_color_fixup_supported,
2347 shader_arb_add_instruction_modifiers,
2348 };
2349
2350 /* ARB_fragment_program fixed function pipeline replacement definitions */
2351 #define ARB_FFP_CONST_TFACTOR 0
2352 #define ARB_FFP_CONST_SPECULAR_ENABLE ((ARB_FFP_CONST_TFACTOR) + 1)
2353 #define ARB_FFP_CONST_CONSTANT(i) ((ARB_FFP_CONST_SPECULAR_ENABLE) + 1 + i)
2354 #define ARB_FFP_CONST_BUMPMAT(i) ((ARB_FFP_CONST_CONSTANT(7)) + 1 + i)
2355 #define ARB_FFP_CONST_LUMINANCE(i) ((ARB_FFP_CONST_BUMPMAT(7)) + 1 + i)
2356
2357 struct arbfp_ffp_desc
2358 {
2359 struct ffp_frag_desc parent;
2360 GLuint shader;
2361 unsigned int num_textures_used;
2362 };
2363
2364 static void arbfp_enable(IWineD3DDevice *iface, BOOL enable) {
2365 if(enable) {
2366 glEnable(GL_FRAGMENT_PROGRAM_ARB);
2367 checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB)");
2368 } else {
2369 glDisable(GL_FRAGMENT_PROGRAM_ARB);
2370 checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
2371 }
2372 }
2373
2374 static HRESULT arbfp_alloc(IWineD3DDevice *iface) {
2375 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
2376 struct shader_arb_priv *priv;
2377 /* Share private data between the shader backend and the pipeline replacement, if both
2378 * are the arb implementation. This is needed to figure out whether ARBfp should be disabled
2379 * if no pixel shader is bound or not
2380 */
2381 if(This->shader_backend == &arb_program_shader_backend) {
2382 This->fragment_priv = This->shader_priv;
2383 } else {
2384 This->fragment_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct shader_arb_priv));
2385 if(!This->fragment_priv) return E_OUTOFMEMORY;
2386 }
2387 priv = This->fragment_priv;
2388 priv->fragment_shaders = hash_table_create(ffp_frag_program_key_hash, ffp_frag_program_key_compare);
2389 priv->use_arbfp_fixed_func = TRUE;
2390 return WINED3D_OK;
2391 }
2392
2393 static void arbfp_free_ffpshader(void *value, void *gli) {
2394 const WineD3D_GL_Info *gl_info = gli;
2395 struct arbfp_ffp_desc *entry_arb = value;
2396
2397 ENTER_GL();
2398 GL_EXTCALL(glDeleteProgramsARB(1, &entry_arb->shader));
2399 checkGLcall("glDeleteProgramsARB(1, &entry_arb->shader)");
2400 HeapFree(GetProcessHeap(), 0, entry_arb);
2401 LEAVE_GL();
2402 }
2403
2404 static void arbfp_free(IWineD3DDevice *iface) {
2405 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
2406 struct shader_arb_priv *priv = This->fragment_priv;
2407
2408 hash_table_destroy(priv->fragment_shaders, arbfp_free_ffpshader, &This->adapter->gl_info);
2409 priv->use_arbfp_fixed_func = FALSE;
2410
2411 if(This->shader_backend != &arb_program_shader_backend) {
2412 HeapFree(GetProcessHeap(), 0, This->fragment_priv);
2413 }
2414 }
2415
2416 static void arbfp_get_caps(WINED3DDEVTYPE devtype, const WineD3D_GL_Info *gl_info, struct fragment_caps *caps)
2417 {
2418 caps->TextureOpCaps = WINED3DTEXOPCAPS_DISABLE |
2419 WINED3DTEXOPCAPS_SELECTARG1 |
2420 WINED3DTEXOPCAPS_SELECTARG2 |
2421 WINED3DTEXOPCAPS_MODULATE4X |
2422 WINED3DTEXOPCAPS_MODULATE2X |
2423 WINED3DTEXOPCAPS_MODULATE |
2424 WINED3DTEXOPCAPS_ADDSIGNED2X |
2425 WINED3DTEXOPCAPS_ADDSIGNED |
2426 WINED3DTEXOPCAPS_ADD |
2427 WINED3DTEXOPCAPS_SUBTRACT |
2428 WINED3DTEXOPCAPS_ADDSMOOTH |
2429 WINED3DTEXOPCAPS_BLENDCURRENTALPHA |
2430 WINED3DTEXOPCAPS_BLENDFACTORALPHA |
2431 WINED3DTEXOPCAPS_BLENDTEXTUREALPHA |
2432 WINED3DTEXOPCAPS_BLENDDIFFUSEALPHA |
2433 WINED3DTEXOPCAPS_BLENDTEXTUREALPHAPM |
2434 WINED3DTEXOPCAPS_MODULATEALPHA_ADDCOLOR |
2435 WINED3DTEXOPCAPS_MODULATECOLOR_ADDALPHA |
2436 WINED3DTEXOPCAPS_MODULATEINVCOLOR_ADDALPHA |
2437 WINED3DTEXOPCAPS_MODULATEINVALPHA_ADDCOLOR |
2438 WINED3DTEXOPCAPS_DOTPRODUCT3 |
2439 WINED3DTEXOPCAPS_MULTIPLYADD |
2440 WINED3DTEXOPCAPS_LERP |
2441 WINED3DTEXOPCAPS_BUMPENVMAP |
2442 WINED3DTEXOPCAPS_BUMPENVMAPLUMINANCE;
2443
2444 /* TODO: Implement WINED3DTEXOPCAPS_PREMODULATE */
2445
2446 caps->MaxTextureBlendStages = 8;
2447 caps->MaxSimultaneousTextures = min(GL_LIMITS(fragment_samplers), 8);
2448
2449 caps->PrimitiveMiscCaps |= WINED3DPMISCCAPS_TSSARGTEMP;
2450 }
2451 #undef GLINFO_LOCATION
2452
2453 #define GLINFO_LOCATION stateblock->wineD3DDevice->adapter->gl_info
2454 static void state_texfactor_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2455 float col[4];
2456 IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2457
2458 /* Don't load the parameter if we're using an arbfp pixel shader, otherwise we'll overwrite
2459 * application provided constants
2460 */
2461 if(device->shader_backend == &arb_program_shader_backend) {
2462 if (use_ps(stateblock)) return;
2463
2464 device = stateblock->wineD3DDevice;
2465 device->activeContext->pshader_const_dirty[ARB_FFP_CONST_TFACTOR] = 1;
2466 device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_TFACTOR + 1);
2467 }
2468
2469 D3DCOLORTOGLFLOAT4(stateblock->renderState[WINED3DRS_TEXTUREFACTOR], col);
2470 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_TFACTOR, col));
2471 checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_TFACTOR, col)");
2472
2473 }
2474
2475 static void state_arb_specularenable(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2476 float col[4];
2477 IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2478
2479 /* Don't load the parameter if we're using an arbfp pixel shader, otherwise we'll overwrite
2480 * application provided constants
2481 */
2482 if(device->shader_backend == &arb_program_shader_backend) {
2483 if (use_ps(stateblock)) return;
2484
2485 device = stateblock->wineD3DDevice;
2486 device->activeContext->pshader_const_dirty[ARB_FFP_CONST_SPECULAR_ENABLE] = 1;
2487 device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_SPECULAR_ENABLE + 1);
2488 }
2489
2490 if(stateblock->renderState[WINED3DRS_SPECULARENABLE]) {
2491 /* The specular color has no alpha */
2492 col[0] = 1.0; col[1] = 1.0;
2493 col[2] = 1.0; col[3] = 0.0;
2494 } else {
2495 col[0] = 0.0; col[1] = 0.0;
2496 col[2] = 0.0; col[3] = 0.0;
2497 }
2498 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_SPECULAR_ENABLE, col));
2499 checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_SPECULAR_ENABLE, col)");
2500 }
2501
2502 static void set_bumpmat_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2503 DWORD stage = (state - STATE_TEXTURESTAGE(0, 0)) / (WINED3D_HIGHEST_TEXTURE_STATE + 1);
2504 IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2505 float mat[2][2];
2506
2507 if (use_ps(stateblock))
2508 {
2509 if(stage != 0 &&
2510 ((IWineD3DPixelShaderImpl *) stateblock->pixelShader)->baseShader.reg_maps.bumpmat[stage]) {
2511 /* The pixel shader has to know the bump env matrix. Do a constants update if it isn't scheduled
2512 * anyway
2513 */
2514 if(!isStateDirty(context, STATE_PIXELSHADERCONSTANT)) {
2515 device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
2516 }
2517 }
2518
2519 if(device->shader_backend == &arb_program_shader_backend) {
2520 /* Exit now, don't set the bumpmat below, otherwise we may overwrite pixel shader constants */
2521 return;
2522 }
2523 } else if(device->shader_backend == &arb_program_shader_backend) {
2524 device->activeContext->pshader_const_dirty[ARB_FFP_CONST_BUMPMAT(stage)] = 1;
2525 device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_BUMPMAT(stage) + 1);
2526 }
2527
2528 mat[0][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT00]);
2529 mat[0][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT01]);
2530 mat[1][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT10]);
2531 mat[1][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT11]);
2532
2533 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_BUMPMAT(stage), &mat[0][0]));
2534 checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_BUMPMAT(stage), &mat[0][0])");
2535 }
2536
2537 static void tex_bumpenvlum_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2538 DWORD stage = (state - STATE_TEXTURESTAGE(0, 0)) / (WINED3D_HIGHEST_TEXTURE_STATE + 1);
2539 IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2540 float param[4];
2541
2542 if (use_ps(stateblock))
2543 {
2544 if(stage != 0 &&
2545 ((IWineD3DPixelShaderImpl *) stateblock->pixelShader)->baseShader.reg_maps.luminanceparams[stage]) {
2546 /* The pixel shader has to know the luminance offset. Do a constants update if it
2547 * isn't scheduled anyway
2548 */
2549 if(!isStateDirty(context, STATE_PIXELSHADERCONSTANT)) {
2550 device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
2551 }
2552 }
2553
2554 if(device->shader_backend == &arb_program_shader_backend) {
2555 /* Exit now, don't set the bumpmat below, otherwise we may overwrite pixel shader constants */
2556 return;
2557 }
2558 } else if(device->shader_backend == &arb_program_shader_backend) {
2559 device->activeContext->pshader_const_dirty[ARB_FFP_CONST_LUMINANCE(stage)] = 1;
2560 device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_LUMINANCE(stage) + 1);
2561 }
2562
2563 param[0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVLSCALE]);
2564 param[1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVLOFFSET]);
2565 param[2] = 0.0;
2566 param[3] = 0.0;
2567
2568 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_LUMINANCE(stage), param));
2569 checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_LUMINANCE(stage), param)");
2570 }
2571
2572 static const char *get_argreg(SHADER_BUFFER *buffer, DWORD argnum, unsigned int stage, DWORD arg) {
2573 const char *ret;
2574
2575 if(arg == ARG_UNUSED) return "unused"; /* This is the marker for unused registers */
2576
2577 switch(arg & WINED3DTA_SELECTMASK) {
2578 case WINED3DTA_DIFFUSE:
2579 ret = "fragment.color.primary"; break;
2580
2581 case WINED3DTA_CURRENT:
2582 if(stage == 0) ret = "fragment.color.primary";
2583 else ret = "ret";
2584 break;
2585
2586 case WINED3DTA_TEXTURE:
2587 switch(stage) {
2588 case 0: ret = "tex0"; break;
2589 case 1: ret = "tex1"; break;
2590 case 2: ret = "tex2"; break;
2591 case 3: ret = "tex3"; break;
2592 case 4: ret = "tex4"; break;
2593 case 5: ret = "tex5"; break;
2594 case 6: ret = "tex6"; break;
2595 case 7: ret = "tex7"; break;
2596 default: ret = "unknown texture";
2597 }
2598 break;
2599
2600 case WINED3DTA_TFACTOR:
2601 ret = "tfactor"; break;
2602
2603 case WINED3DTA_SPECULAR:
2604 ret = "fragment.color.secondary"; break;
2605
2606 case WINED3DTA_TEMP:
2607 ret = "tempreg"; break;
2608
2609 case WINED3DTA_CONSTANT:
2610 FIXME("Implement perstage constants\n");
2611 switch(stage) {
2612 case 0: ret = "const0"; break;
2613 case 1: ret = "const1"; break;
2614 case 2: ret = "const2"; break;
2615 case 3: ret = "const3"; break;
2616 case 4: ret = "const4"; break;
2617 case 5: ret = "const5"; break;
2618 case 6: ret = "const6"; break;
2619 case 7: ret = "const7"; break;
2620 default: ret = "unknown constant";
2621 }
2622 break;
2623
2624 default:
2625 return "unknown";
2626 }
2627
2628 if(arg & WINED3DTA_COMPLEMENT) {
2629 shader_addline(buffer, "SUB arg%u, const.x, %s;\n", argnum, ret);
2630 if(argnum == 0) ret = "arg0";
2631 if(argnum == 1) ret = "arg1";
2632 if(argnum == 2) ret = "arg2";
2633 }
2634 if(arg & WINED3DTA_ALPHAREPLICATE) {
2635 shader_addline(buffer, "MOV arg%u, %s.w;\n", argnum, ret);
2636 if(argnum == 0) ret = "arg0";
2637 if(argnum == 1) ret = "arg1";
2638 if(argnum == 2) ret = "arg2";
2639 }
2640 return ret;
2641 }
2642
2643 static void gen_ffp_instr(SHADER_BUFFER *buffer, unsigned int stage, BOOL color, BOOL alpha,
2644 DWORD dst, DWORD op, DWORD dw_arg0, DWORD dw_arg1, DWORD dw_arg2) {
2645 const char *dstmask, *dstreg, *arg0, *arg1, *arg2;
2646 unsigned int mul = 1;
2647 BOOL mul_final_dest = FALSE;
2648
2649 if(color && alpha) dstmask = "";
2650 else if(color) dstmask = ".xyz";
2651 else dstmask = ".w";
2652
2653 if(dst == tempreg) dstreg = "tempreg";
2654 else dstreg = "ret";
2655
2656 arg0 = get_argreg(buffer, 0, stage, dw_arg0);
2657 arg1 = get_argreg(buffer, 1, stage, dw_arg1);
2658 arg2 = get_argreg(buffer, 2, stage, dw_arg2);
2659
2660 switch(op) {
2661 case WINED3DTOP_DISABLE:
2662 if(stage == 0) shader_addline(buffer, "MOV %s%s, fragment.color.primary;\n", dstreg, dstmask);
2663 break;
2664
2665 case WINED3DTOP_SELECTARG2:
2666 arg1 = arg2;
2667 case WINED3DTOP_SELECTARG1:
2668 shader_addline(buffer, "MOV %s%s, %s;\n", dstreg, dstmask, arg1);
2669 break;
2670
2671 case WINED3DTOP_MODULATE4X:
2672 mul = 2;
2673 case WINED3DTOP_MODULATE2X:
2674 mul *= 2;
2675 if(strcmp(dstreg, "result.color") == 0) {
2676 dstreg = "ret";
2677 mul_final_dest = TRUE;
2678 }
2679 case WINED3DTOP_MODULATE:
2680 shader_addline(buffer, "MUL %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2681 break;
2682
2683 case WINED3DTOP_ADDSIGNED2X:
2684 mul = 2;
2685 if(strcmp(dstreg, "result.color") == 0) {
2686 dstreg = "ret";
2687 mul_final_dest = TRUE;
2688 }
2689 case WINED3DTOP_ADDSIGNED:
2690 shader_addline(buffer, "SUB arg2, %s, const.w;\n", arg2);
2691 arg2 = "arg2";
2692 case WINED3DTOP_ADD:
2693 shader_addline(buffer, "ADD_SAT %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2694 break;
2695
2696 case WINED3DTOP_SUBTRACT:
2697 shader_addline(buffer, "SUB_SAT %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2698 break;
2699
2700 case WINED3DTOP_ADDSMOOTH:
2701 shader_addline(buffer, "SUB arg1, const.x, %s;\n", arg1);
2702 shader_addline(buffer, "MAD_SAT %s%s, arg1, %s, %s;\n", dstreg, dstmask, arg2, arg1);
2703 break;
2704
2705 case WINED3DTOP_BLENDCURRENTALPHA:
2706 arg0 = get_argreg(buffer, 0, stage, WINED3DTA_CURRENT);
2707 shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2708 break;
2709 case WINED3DTOP_BLENDFACTORALPHA:
2710 arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TFACTOR);
2711 shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2712 break;
2713 case WINED3DTOP_BLENDTEXTUREALPHA:
2714 arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TEXTURE);
2715 shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2716 break;
2717 case WINED3DTOP_BLENDDIFFUSEALPHA:
2718 arg0 = get_argreg(buffer, 0, stage, WINED3DTA_DIFFUSE);
2719 shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2720 break;
2721
2722 case WINED3DTOP_BLENDTEXTUREALPHAPM:
2723 arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TEXTURE);
2724 shader_addline(buffer, "SUB arg0.w, const.x, %s.w;\n", arg0);
2725 shader_addline(buffer, "MAD_SAT %s%s, %s, arg0.w, %s;\n", dstreg, dstmask, arg2, arg1);
2726 break;
2727
2728 /* D3DTOP_PREMODULATE ???? */
2729
2730 case WINED3DTOP_MODULATEINVALPHA_ADDCOLOR:
2731 shader_addline(buffer, "SUB arg0.w, const.x, %s;\n", arg1);
2732 shader_addline(buffer, "MAD_SAT %s%s, arg0.w, %s, %s;\n", dstreg, dstmask, arg2, arg1);
2733 break;
2734 case WINED3DTOP_MODULATEALPHA_ADDCOLOR:
2735 shader_addline(buffer, "MAD_SAT %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg1, arg2, arg1);
2736 break;
2737 case WINED3DTOP_MODULATEINVCOLOR_ADDALPHA:
2738 shader_addline(buffer, "SUB arg0, const.x, %s;\n", arg1);
2739 shader_addline(buffer, "MAD_SAT %s%s, arg0, %s, %s.w;\n", dstreg, dstmask, arg2, arg1);
2740 break;
2741 case WINED3DTOP_MODULATECOLOR_ADDALPHA:
2742 shader_addline(buffer, "MAD_SAT %s%s, %s, %s, %s.w;\n", dstreg, dstmask, arg1, arg2, arg1);
2743 break;
2744
2745 case WINED3DTOP_DOTPRODUCT3:
2746 mul = 4;
2747 if(strcmp(dstreg, "result.color") == 0) {
2748 dstreg = "ret";
2749 mul_final_dest = TRUE;
2750 }
2751 shader_addline(buffer, "SUB arg1, %s, const.w;\n", arg1);
2752 shader_addline(buffer, "SUB arg2, %s, const.w;\n", arg2);
2753 shader_addline(buffer, "DP3_SAT %s%s, arg1, arg2;\n", dstreg, dstmask);
2754 break;
2755
2756 case WINED3DTOP_MULTIPLYADD:
2757 shader_addline(buffer, "MAD_SAT %s%s, %s, %s, %s;\n", dstreg, dstmask, arg1, arg2, arg0);
2758 break;
2759
2760 case WINED3DTOP_LERP:
2761 /* The msdn is not quite right here */
2762 shader_addline(buffer, "LRP %s%s, %s, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2763 break;
2764
2765 case WINED3DTOP_BUMPENVMAP:
2766 case WINED3DTOP_BUMPENVMAPLUMINANCE:
2767 /* Those are handled in the first pass of the shader(generation pass 1 and 2) already */
2768 break;
2769
2770 default:
2771 FIXME("Unhandled texture op %08x\n", op);
2772 }
2773
2774 if(mul == 2) {
2775 shader_addline(buffer, "MUL_SAT %s%s, %s, const.y;\n", mul_final_dest ? "result.color" : dstreg, dstmask, dstreg);
2776 } else if(mul == 4) {
2777 shader_addline(buffer, "MUL_SAT %s%s, %s, const.z;\n", mul_final_dest ? "result.color" : dstreg, dstmask, dstreg);
2778 }
2779 }
2780
2781 /* The stateblock is passed for GLINFO_LOCATION */
2782 static GLuint gen_arbfp_ffp_shader(const struct ffp_frag_settings *settings, IWineD3DStateBlockImpl *stateblock)
2783 {
2784 unsigned int stage;
2785 SHADER_BUFFER buffer;
2786 BOOL tex_read[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2787 BOOL bump_used[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2788 BOOL luminance_used[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2789 const char *textype;
2790 const char *instr, *sat;
2791 char colorcor_dst[8];
2792 GLuint ret;
2793 DWORD arg0, arg1, arg2;
2794 BOOL tempreg_used = FALSE, tfactor_used = FALSE;
2795 BOOL op_equal;
2796 const char *final_combiner_src = "ret";
2797
2798 /* Find out which textures are read */
2799 for(stage = 0; stage < MAX_TEXTURES; stage++) {
2800 if(settings->op[stage].cop == WINED3DTOP_DISABLE) break;
2801 arg0 = settings->op[stage].carg0 & WINED3DTA_SELECTMASK;
2802 arg1 = settings->op[stage].carg1 & WINED3DTA_SELECTMASK;
2803 arg2 = settings->op[stage].carg2 & WINED3DTA_SELECTMASK;
2804 if(arg0 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2805 if(arg1 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2806 if(arg2 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2807
2808 if(settings->op[stage].cop == WINED3DTOP_BLENDTEXTUREALPHA) tex_read[stage] = TRUE;
2809 if(settings->op[stage].cop == WINED3DTOP_BLENDTEXTUREALPHAPM) tex_read[stage] = TRUE;
2810 if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAP) {
2811 bump_used[stage] = TRUE;
2812 tex_read[stage] = TRUE;
2813 }
2814 if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2815 bump_used[stage] = TRUE;
2816 tex_read[stage] = TRUE;
2817 luminance_used[stage] = TRUE;
2818 } else if(settings->op[stage].cop == WINED3DTOP_BLENDFACTORALPHA) {
2819 tfactor_used = TRUE;
2820 }
2821
2822 if(arg0 == WINED3DTA_TFACTOR || arg1 == WINED3DTA_TFACTOR || arg2 == WINED3DTA_TFACTOR) {
2823 tfactor_used = TRUE;
2824 }
2825
2826 if(settings->op[stage].dst == tempreg) tempreg_used = TRUE;
2827 if(arg0 == WINED3DTA_TEMP || arg1 == WINED3DTA_TEMP || arg2 == WINED3DTA_TEMP) {
2828 tempreg_used = TRUE;
2829 }
2830
2831 if(settings->op[stage].aop == WINED3DTOP_DISABLE) continue;
2832 arg0 = settings->op[stage].aarg0 & WINED3DTA_SELECTMASK;
2833 arg1 = settings->op[stage].aarg1 & WINED3DTA_SELECTMASK;
2834 arg2 = settings->op[stage].aarg2 & WINED3DTA_SELECTMASK;
2835 if(arg0 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2836 if(arg1 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2837 if(arg2 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2838
2839 if(arg0 == WINED3DTA_TEMP || arg1 == WINED3DTA_TEMP || arg2 == WINED3DTA_TEMP) {
2840 tempreg_used = TRUE;
2841 }
2842 if(arg0 == WINED3DTA_TFACTOR || arg1 == WINED3DTA_TFACTOR || arg2 == WINED3DTA_TFACTOR) {
2843 tfactor_used = TRUE;
2844 }
2845 }
2846
2847 /* Shader header */
2848 shader_buffer_init(&buffer);
2849
2850 shader_addline(&buffer, "!!ARBfp1.0\n");
2851
2852 switch(settings->fog) {
2853 case FOG_OFF: break;
2854 case FOG_LINEAR: shader_addline(&buffer, "OPTION ARB_fog_linear;\n"); break;
2855 case FOG_EXP: shader_addline(&buffer, "OPTION ARB_fog_exp;\n"); break;
2856 case FOG_EXP2: shader_addline(&buffer, "OPTION ARB_fog_exp2;\n"); break;
2857 default: FIXME("Unexpected fog setting %d\n", settings->fog);
2858 }
2859
2860 shader_addline(&buffer, "PARAM const = {1, 2, 4, 0.5};\n");
2861 shader_addline(&buffer, "TEMP TMP;\n");
2862 shader_addline(&buffer, "TEMP ret;\n");
2863 if(tempreg_used || settings->sRGB_write) shader_addline(&buffer, "TEMP tempreg;\n");
2864 shader_addline(&buffer, "TEMP arg0;\n");
2865 shader_addline(&buffer, "TEMP arg1;\n");
2866 shader_addline(&buffer, "TEMP arg2;\n");
2867 for(stage = 0; stage < MAX_TEXTURES; stage++) {
2868 if(!tex_read[stage]) continue;
2869 shader_addline(&buffer, "TEMP tex%u;\n", stage);
2870 if(!bump_used[stage]) continue;
2871 shader_addline(&buffer, "PARAM bumpmat%u = program.env[%u];\n", stage, ARB_FFP_CONST_BUMPMAT(stage));
2872 if(!luminance_used[stage]) continue;
2873 shader_addline(&buffer, "PARAM luminance%u = program.env[%u];\n", stage, ARB_FFP_CONST_LUMINANCE(stage));
2874 }
2875 if(tfactor_used) {
2876 shader_addline(&buffer, "PARAM tfactor = program.env[%u];\n", ARB_FFP_CONST_TFACTOR);
2877 }
2878 shader_addline(&buffer, "PARAM specular_enable = program.env[%u];\n", ARB_FFP_CONST_SPECULAR_ENABLE);
2879
2880 if(settings->sRGB_write) {
2881 shader_addline(&buffer, "PARAM srgb_consts1 = {%f, %f, %f, %f};\n",
2882 srgb_mul_low, srgb_cmp, srgb_pow, srgb_mul_high);
2883 shader_addline(&buffer, "PARAM srgb_consts2 = {%f, %f, %f, %f};\n",
2884 srgb_sub_high, 0.0, 0.0, 0.0);
2885 }
2886
2887 /* Generate texture sampling instructions) */
2888 for(stage = 0; stage < MAX_TEXTURES && settings->op[stage].cop != WINED3DTOP_DISABLE; stage++) {
2889 if(!tex_read[stage]) continue;
2890
2891 switch(settings->op[stage].tex_type) {
2892 case tex_1d: textype = "1D"; break;
2893 case tex_2d: textype = "2D"; break;
2894 case tex_3d: textype = "3D"; break;
2895 case tex_cube: textype = "CUBE"; break;
2896 case tex_rect: textype = "RECT"; break;
2897 default: textype = "unexpected_textype"; break;
2898 }
2899
2900 if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAP ||
2901 settings->op[stage].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2902 sat = "";
2903 } else {
2904 sat = "_SAT";
2905 }
2906
2907 if(settings->op[stage].projected == proj_none) {
2908 instr = "TEX";
2909 } else if(settings->op[stage].projected == proj_count4 ||
2910 settings->op[stage].projected == proj_count3) {
2911 instr = "TXP";
2912 } else {
2913 FIXME("Unexpected projection mode %d\n", settings->op[stage].projected);
2914 instr = "TXP";
2915 }
2916
2917 if(stage > 0 &&
2918 (settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAP ||
2919 settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAPLUMINANCE)) {
2920 shader_addline(&buffer, "SWZ arg1, bumpmat%u, x, z, 0, 0;\n", stage - 1);
2921 shader_addline(&buffer, "DP3 ret.x, arg1, tex%u;\n", stage - 1);
2922 shader_addline(&buffer, "SWZ arg1, bumpmat%u, y, w, 0, 0;\n", stage - 1);
2923 shader_addline(&buffer, "DP3 ret.y, arg1, tex%u;\n", stage - 1);
2924
2925 /* with projective textures, texbem only divides the static texture coord, not the displacement,
2926 * so multiply the displacement with the dividing parameter before passing it to TXP
2927 */
2928 if (settings->op[stage].projected != proj_none) {
2929 if(settings->op[stage].projected == proj_count4) {
2930 shader_addline(&buffer, "MOV ret.w, fragment.texcoord[%u].w;\n", stage);
2931 shader_addline(&buffer, "MUL ret.xyz, ret, fragment.texcoord[%u].w, fragment.texcoord[%u];\n", stage, stage);
2932 } else {
2933 shader_addline(&buffer, "MOV ret.w, fragment.texcoord[%u].z;\n", stage);
2934 shader_addline(&buffer, "MAD ret.xyz, ret, fragment.texcoord[%u].z, fragment.texcoord[%u];\n", stage, stage);
2935 }
2936 } else {
2937 shader_addline(&buffer, "ADD ret, ret, fragment.texcoord[%u];\n", stage);
2938 }
2939
2940 shader_addline(&buffer, "%s%s tex%u, ret, texture[%u], %s;\n",
2941 instr, sat, stage, stage, textype);
2942 if(settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2943 shader_addline(&buffer, "MAD_SAT ret.x, tex%u.z, luminance%u.x, luminance%u.y;\n",
2944 stage - 1, stage - 1, stage - 1);
2945 shader_addline(&buffer, "MUL tex%u, tex%u, ret.x;\n", stage, stage);
2946 }
2947 } else if(settings->op[stage].projected == proj_count3) {
2948 shader_addline(&buffer, "MOV ret, fragment.texcoord[%u];\n", stage);
2949 shader_addline(&buffer, "MOV ret.w, ret.z;\n");
2950 shader_addline(&buffer, "%s%s tex%u, ret, texture[%u], %s;\n",
2951 instr, sat, stage, stage, textype);
2952 } else {
2953 shader_addline(&buffer, "%s%s tex%u, fragment.texcoord[%u], texture[%u], %s;\n",
2954 instr, sat, stage, stage, stage, textype);
2955 }
2956
2957 sprintf(colorcor_dst, "tex%u", stage);
2958 gen_color_correction(&buffer, colorcor_dst, WINED3DSP_WRITEMASK_ALL, "const.x", "const.y",
2959 settings->op[stage].color_fixup);
2960 }
2961
2962 /* Generate the main shader */
2963 for(stage = 0; stage < MAX_TEXTURES; stage++) {
2964 if(settings->op[stage].cop == WINED3DTOP_DISABLE) {
2965 if(stage == 0) {
2966 final_combiner_src = "fragment.color.primary";
2967 }
2968 break;
2969 }
2970
2971 if(settings->op[stage].cop == WINED3DTOP_SELECTARG1 &&
2972 settings->op[stage].aop == WINED3DTOP_SELECTARG1) {
2973 op_equal = settings->op[stage].carg1 == settings->op[stage].aarg1;
2974 } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG1 &&
2975 settings->op[stage].aop == WINED3DTOP_SELECTARG2) {
2976 op_equal = settings->op[stage].carg1 == settings->op[stage].aarg2;
2977 } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG2 &&
2978 settings->op[stage].aop == WINED3DTOP_SELECTARG1) {
2979 op_equal = settings->op[stage].carg2 == settings->op[stage].aarg1;
2980 } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG2 &&
2981 settings->op[stage].aop == WINED3DTOP_SELECTARG2) {
2982 op_equal = settings->op[stage].carg2 == settings->op[stage].aarg2;
2983 } else {
2984 op_equal = settings->op[stage].aop == settings->op[stage].cop &&
2985 settings->op[stage].carg0 == settings->op[stage].aarg0 &&
2986 settings->op[stage].carg1 == settings->op[stage].aarg1 &&
2987 settings->op[stage].carg2 == settings->op[stage].aarg2;
2988 }
2989
2990 if(settings->op[stage].aop == WINED3DTOP_DISABLE) {
2991 gen_ffp_instr(&buffer, stage, TRUE, FALSE, settings->op[stage].dst,
2992 settings->op[stage].cop, settings->op[stage].carg0,
2993 settings->op[stage].carg1, settings->op[stage].carg2);
2994 if(stage == 0) {
2995 shader_addline(&buffer, "MOV ret.w, fragment.color.primary.w;\n");
2996 }
2997 } else if(op_equal) {
2998 gen_ffp_instr(&buffer, stage, TRUE, TRUE, settings->op[stage].dst,
2999 settings->op[stage].cop, settings->op[stage].carg0,
3000 settings->op[stage].carg1, settings->op[stage].carg2);
3001 } else {
3002 gen_ffp_instr(&buffer, stage, TRUE, FALSE, settings->op[stage].dst,
3003 settings->op[stage].cop, settings->op[stage].carg0,
3004 settings->op[stage].carg1, settings->op[stage].carg2);
3005 gen_ffp_instr(&buffer, stage, FALSE, TRUE, settings->op[stage].dst,
3006 settings->op[stage].aop, settings->op[stage].aarg0,
3007 settings->op[stage].aarg1, settings->op[stage].aarg2);
3008 }
3009 }
3010
3011 if(settings->sRGB_write) {
3012 shader_addline(&buffer, "MAD ret, fragment.color.secondary, specular_enable, %s;\n", final_combiner_src);
3013 arbfp_add_sRGB_correction(&buffer, "ret", "arg0", "arg1", "arg2", "tempreg");
3014 shader_addline(&buffer, "MOV result.color.w, ret.w;\n");
3015 } else {
3016 shader_addline(&buffer, "MAD result.color, fragment.color.secondary, specular_enable, %s;\n", final_combiner_src);
3017 }
3018
3019 /* Footer */
3020 shader_addline(&buffer, "END\n");
3021
3022 /* Generate the shader */
3023 GL_EXTCALL(glGenProgramsARB(1, &ret));
3024 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, ret));
3025 GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(buffer.buffer), buffer.buffer));
3026
3027 if (glGetError() == GL_INVALID_OPERATION) {
3028 GLint pos;
3029 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
3030 FIXME("Fragment program error at position %d: %s\n", pos,
3031 debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
3032 }
3033 shader_buffer_free(&buffer);
3034 return ret;
3035 }
3036
3037 static void fragment_prog_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
3038 IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
3039 struct shader_arb_priv *priv = device->fragment_priv;
3040 BOOL use_pshader = use_ps(stateblock);
3041 BOOL use_vshader = use_vs(stateblock);
3042 struct ffp_frag_settings settings;
3043 const struct arbfp_ffp_desc *desc;
3044 unsigned int i;
3045
3046 TRACE("state %#x, stateblock %p, context %p\n", state, stateblock, context);
3047
3048 if(isStateDirty(context, STATE_RENDER(WINED3DRS_FOGENABLE))) {
3049 if(!use_pshader && device->shader_backend == &arb_program_shader_backend && context->last_was_pshader) {
3050 /* Reload fixed function constants since they collide with the pixel shader constants */
3051 for(i = 0; i < MAX_TEXTURES; i++) {
3052 set_bumpmat_arbfp(STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT00), stateblock, context);
3053 }
3054 state_texfactor_arbfp(STATE_RENDER(WINED3DRS_TEXTUREFACTOR), stateblock, context);
3055 state_arb_specularenable(STATE_RENDER(WINED3DRS_SPECULARENABLE), stateblock, context);
3056 } else if(use_pshader && !isStateDirty(context, device->StateTable[STATE_VSHADER].representative)) {
3057 device->shader_backend->shader_select((IWineD3DDevice *)stateblock->wineD3DDevice, use_pshader, use_vshader);
3058 }
3059 return;
3060 }
3061
3062 if(!use_pshader) {
3063 /* Find or create a shader implementing the fixed function pipeline settings, then activate it */
3064 gen_ffp_frag_op(stateblock, &settings, FALSE);
3065 desc = (const struct arbfp_ffp_desc *)find_ffp_frag_shader(priv->fragment_shaders, &settings);
3066 if(!desc) {
3067 struct arbfp_ffp_desc *new_desc = HeapAlloc(GetProcessHeap(), 0, sizeof(*new_desc));
3068 if (!new_desc)
3069 {
3070 ERR("Out of memory\n");
3071 return;
3072 }
3073 new_desc->num_textures_used = 0;
3074 for(i = 0; i < GL_LIMITS(texture_stages); i++) {
3075 if(settings.op[i].cop == WINED3DTOP_DISABLE) break;
3076 new_desc->num_textures_used = i;
3077 }
3078
3079 memcpy(&new_desc->parent.settings, &settings, sizeof(settings));
3080 new_desc->shader = gen_arbfp_ffp_shader(&settings, stateblock);
3081 add_ffp_frag_shader(priv->fragment_shaders, &new_desc->parent);
3082 TRACE("Allocated fixed function replacement shader descriptor %p\n", new_desc);
3083 desc = new_desc;
3084 }
3085
3086 /* Now activate the replacement program. GL_FRAGMENT_PROGRAM_ARB is already active(however, note the
3087 * comment above the shader_select call below). If e.g. GLSL is active, the shader_select call will
3088 * deactivate it.
3089 */
3090 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, desc->shader));
3091 checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, desc->shader)");
3092 priv->current_fprogram_id = desc->shader;
3093
3094 if(device->shader_backend == &arb_program_shader_backend && context->last_was_pshader) {
3095 /* Reload fixed function constants since they collide with the pixel shader constants */
3096 for(i = 0; i < MAX_TEXTURES; i++) {
3097 set_bumpmat_arbfp(STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT00), stateblock, context);
3098 }
3099 state_texfactor_arbfp(STATE_RENDER(WINED3DRS_TEXTUREFACTOR), stateblock, context);
3100 state_arb_specularenable(STATE_RENDER(WINED3DRS_SPECULARENABLE), stateblock, context);
3101 }
3102 context->last_was_pshader = FALSE;
3103 } else {
3104 context->last_was_pshader = TRUE;
3105 }
3106
3107 /* Finally, select the shader. If a pixel shader is used, it will be set and enabled by the shader backend.
3108 * If this shader backend is arbfp(most likely), then it will simply overwrite the last fixed function replace-
3109 * ment shader. If the shader backend is not ARB, it currently is important that the opengl implementation
3110 * type overwrites GL_ARB_fragment_program. This is currently the case with GLSL. If we really want to use
3111 * atifs or nvrc pixel shaders with arb fragment programs we'd have to disable GL_FRAGMENT_PROGRAM_ARB here
3112 *
3113 * Don't call shader_select if the vertex shader is dirty, because it will be called later on by the vertex
3114 * shader handler
3115 */
3116 if(!isStateDirty(context, device->StateTable[STATE_VSHADER].representative)) {
3117 device->shader_backend->shader_select((IWineD3DDevice *)stateblock->wineD3DDevice, use_pshader, use_vshader);
3118
3119 if (!isStateDirty(context, STATE_VERTEXSHADERCONSTANT) && (use_vshader || use_pshader)) {
3120 device->StateTable[STATE_VERTEXSHADERCONSTANT].apply(STATE_VERTEXSHADERCONSTANT, stateblock, context);
3121 }
3122 }
3123 if(use_pshader) {
3124 device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
3125 }
3126 }
3127
3128 /* We can't link the fog states to the fragment state directly since the vertex pipeline links them
3129 * to FOGENABLE. A different linking in different pipeline parts can't be expressed in the combined
3130 * state table, so we need to handle that with a forwarding function. The other invisible side effect
3131 * is that changing the fog start and fog end(which links to FOGENABLE in vertex) results in the
3132 * fragment_prog_arbfp function being called because FOGENABLE is dirty, which calls this function here
3133 */
3134 static void state_arbfp_fog(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
3135 enum fogsource new_source;
3136
3137 TRACE("state %#x, stateblock %p, context %p\n", state, stateblock, context);
3138
3139 if(!isStateDirty(context, STATE_PIXELSHADER)) {
3140 fragment_prog_arbfp(state, stateblock, context);
3141 }
3142
3143 if(!stateblock->renderState[WINED3DRS_FOGENABLE]) return;
3144
3145 if(stateblock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE) {
3146 if(use_vs(stateblock)) {
3147 new_source = FOGSOURCE_VS;
3148 } else {
3149 if(stateblock->renderState[WINED3DRS_FOGVERTEXMODE] == WINED3DFOG_NONE || context->last_was_rhw) {
3150 new_source = FOGSOURCE_COORD;
3151 } else {
3152 new_source = FOGSOURCE_FFP;
3153 }
3154 }
3155 } else {
3156 new_source = FOGSOURCE_FFP;
3157 }
3158 if(new_source != context->fog_source) {
3159 context->fog_source = new_source;
3160 state_fogstartend(STATE_RENDER(WINED3DRS_FOGSTART), stateblock, context);
3161 }
3162 }
3163
3164 static void textransform(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
3165 if(!isStateDirty(context, STATE_PIXELSHADER)) {
3166 fragment_prog_arbfp(state, stateblock, context);
3167 }
3168 }
3169
3170 #undef GLINFO_LOCATION
3171
3172 static const struct StateEntryTemplate arbfp_fragmentstate_template[] = {
3173 {STATE_RENDER(WINED3DRS_TEXTUREFACTOR), { STATE_RENDER(WINED3DRS_TEXTUREFACTOR), state_texfactor_arbfp }, 0 },
3174 {STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3175 {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3176 {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3177 {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3178 {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3179 {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3180 {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3181 {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3182 {STATE_TEXTURESTAGE(0, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3183 {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3184 {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3185 {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3186 {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3187 {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3188 {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3189 {STATE_TEXTURESTAGE(1, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3190 {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3191 {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3192 {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3193 {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3194 {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3195 {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3196 {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3197 {STATE_TEXTURESTAGE(1, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3198 {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3199 {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3200 {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3201 {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3202 {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3203 {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3204 {STATE_TEXTURESTAGE(2, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3205 {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3206 {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3207 {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3208 {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3209 {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3210 {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3211 {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3212 {STATE_TEXTURESTAGE(2, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3213 {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3214 {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3215 {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3216 {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3217 {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3218 {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3219 {STATE_TEXTURESTAGE(3, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3220 {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3221 {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3222 {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3223 {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3224 {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3225 {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3226 {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3227 {STATE_TEXTURESTAGE(3, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3228 {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3229 {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3230 {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3231 {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3232 {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3233 {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3234 {STATE_TEXTURESTAGE(4, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3235 {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3236 {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3237 {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3238 {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3239 {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3240 {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3241 {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3242 {STATE_TEXTURESTAGE(4, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3243 {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3244 {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3245 {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3246 {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3247 {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3248 {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3249 {STATE_TEXTURESTAGE(5, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3250 {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3251 {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3252 {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3253 {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3254 {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3255 {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3256 {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3257 {STATE_TEXTURESTAGE(5, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3258 {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3259 {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3260 {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3261 {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3262 {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3263 {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3264 {STATE_TEXTURESTAGE(6, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3265 {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3266 {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3267 {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3268 {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3269 {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3270 {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3271 {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3272 {STATE_TEXTURESTAGE(6, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3273 {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3274 {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3275 {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3276 {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3277 {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3278 {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3279 {STATE_TEXTURESTAGE(7, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3280 {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3281 {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3282 {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3283 {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3284 {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3285 {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3286 {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3287 {STATE_TEXTURESTAGE(7, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3288 {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3289 {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3290 {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3291 {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3292 {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3293 {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3294 {STATE_SAMPLER(0), { STATE_SAMPLER(0), sampler_texdim }, 0 },
3295 {STATE_SAMPLER(1), { STATE_SAMPLER(1), sampler_texdim }, 0 },
3296 {STATE_SAMPLER(2), { STATE_SAMPLER(2), sampler_texdim }, 0 },
3297 {STATE_SAMPLER(3), { STATE_SAMPLER(3), sampler_texdim }, 0 },
3298 {STATE_SAMPLER(4), { STATE_SAMPLER(4), sampler_texdim }, 0 },
3299 {STATE_SAMPLER(5), { STATE_SAMPLER(5), sampler_texdim }, 0 },
3300 {STATE_SAMPLER(6), { STATE_SAMPLER(6), sampler_texdim }, 0 },
3301 {STATE_SAMPLER(7), { STATE_SAMPLER(7), sampler_texdim }, 0 },
3302 {STATE_PIXELSHADER, { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3303 {STATE_RENDER(WINED3DRS_FOGENABLE), { STATE_RENDER(WINED3DRS_FOGENABLE), state_arbfp_fog }, 0 },
3304 {STATE_RENDER(WINED3DRS_FOGTABLEMODE), { STATE_RENDER(WINED3DRS_FOGENABLE), state_arbfp_fog }, 0 },
3305 {STATE_RENDER(WINED3DRS_FOGVERTEXMODE), { STATE_RENDER(WINED3DRS_FOGENABLE), state_arbfp_fog }, 0 },
3306 {STATE_RENDER(WINED3DRS_FOGSTART), { STATE_RENDER(WINED3DRS_FOGSTART), state_fogstartend }, 0 },
3307 {STATE_RENDER(WINED3DRS_FOGEND), { STATE_RENDER(WINED3DRS_FOGSTART), state_fogstartend }, 0 },
3308 {STATE_RENDER(WINED3DRS_SRGBWRITEENABLE), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3309 {STATE_RENDER(WINED3DRS_FOGCOLOR), { STATE_RENDER(WINED3DRS_FOGCOLOR), state_fogcolor }, 0 },
3310 {STATE_RENDER(WINED3DRS_FOGDENSITY), { STATE_RENDER(WINED3DRS_FOGDENSITY), state_fogdensity }, 0 },
3311 {STATE_TEXTURESTAGE(0,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(0, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3312 {STATE_TEXTURESTAGE(1,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(1, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3313 {STATE_TEXTURESTAGE(2,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(2, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3314 {STATE_TEXTURESTAGE(3,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(3, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3315 {STATE_TEXTURESTAGE(4,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(4, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3316 {STATE_TEXTURESTAGE(5,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(5, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3317 {STATE_TEXTURESTAGE(6,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(6, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3318 {STATE_TEXTURESTAGE(7,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(7, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3319 {STATE_RENDER(WINED3DRS_SPECULARENABLE), { STATE_RENDER(WINED3DRS_SPECULARENABLE), state_arb_specularenable}, 0 },
3320 {0 /* Terminate */, { 0, 0 }, 0 },
3321 };
3322
3323 const struct fragment_pipeline arbfp_fragment_pipeline = {
3324 arbfp_enable,
3325 arbfp_get_caps,
3326 arbfp_alloc,
3327 arbfp_free,
3328 shader_arb_color_fixup_supported,
3329 arbfp_fragmentstate_template,
3330 TRUE /* We can disable projected textures */
3331 };
3332
3333 #define GLINFO_LOCATION device->adapter->gl_info
3334
3335 struct arbfp_blit_priv {
3336 GLenum yuy2_rect_shader, yuy2_2d_shader;
3337 GLenum uyvy_rect_shader, uyvy_2d_shader;
3338 GLenum yv12_rect_shader, yv12_2d_shader;
3339 };
3340
3341 static HRESULT arbfp_blit_alloc(IWineD3DDevice *iface) {
3342 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3343 device->blit_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct arbfp_blit_priv));
3344 if(!device->blit_priv) {
3345 ERR("Out of memory\n");
3346 return E_OUTOFMEMORY;
3347 }
3348 return WINED3D_OK;
3349 }
3350 static void arbfp_blit_free(IWineD3DDevice *iface) {
3351 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3352 struct arbfp_blit_priv *priv = device->blit_priv;
3353
3354 ENTER_GL();
3355 GL_EXTCALL(glDeleteProgramsARB(1, &priv->yuy2_rect_shader));
3356 GL_EXTCALL(glDeleteProgramsARB(1, &priv->yuy2_2d_shader));
3357 GL_EXTCALL(glDeleteProgramsARB(1, &priv->uyvy_rect_shader));
3358 GL_EXTCALL(glDeleteProgramsARB(1, &priv->uyvy_2d_shader));
3359 GL_EXTCALL(glDeleteProgramsARB(1, &priv->yv12_rect_shader));
3360 GL_EXTCALL(glDeleteProgramsARB(1, &priv->yv12_2d_shader));
3361 checkGLcall("Delete yuv programs\n");
3362 LEAVE_GL();
3363 }
3364
3365 static BOOL gen_planar_yuv_read(SHADER_BUFFER *buffer, enum yuv_fixup yuv_fixup, GLenum textype, char *luminance)
3366 {
3367 char chroma;
3368 const char *tex, *texinstr;
3369
3370 if (yuv_fixup == YUV_FIXUP_UYVY) {
3371 chroma = 'x';
3372 *luminance = 'w';
3373 } else {
3374 chroma = 'w';
3375 *luminance = 'x';
3376 }
3377 switch(textype) {
3378 case GL_TEXTURE_2D: tex = "2D"; texinstr = "TXP"; break;
3379 case GL_TEXTURE_RECTANGLE_ARB: tex = "RECT"; texinstr = "TEX"; break;
3380 default:
3381 /* This is more tricky than just replacing the texture type - we have to navigate
3382 * properly in the texture to find the correct chroma values
3383 */
3384 FIXME("Implement yuv correction for non-2d, non-rect textures\n");
3385 return FALSE;
3386 }
3387
3388 /* First we have to read the chroma values. This means we need at least two pixels(no filtering),
3389 * or 4 pixels(with filtering). To get the unmodified chromas, we have to rid ourselves of the
3390 * filtering when we sample the texture.
3391 *
3392 * These are the rules for reading the chroma:
3393 *
3394 * Even pixel: Cr
3395 * Even pixel: U
3396 * Odd pixel: V
3397 *
3398 * So we have to get the sampling x position in non-normalized coordinates in integers
3399 */
3400 if(textype != GL_TEXTURE_RECTANGLE_ARB) {
3401 shader_addline(buffer, "MUL texcrd.xy, fragment.texcoord[0], size.x;\n");
3402 shader_addline(buffer, "MOV texcrd.w, size.x;\n");
3403 } else {
3404 shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3405 }
3406 /* We must not allow filtering between pixel x and x+1, this would mix U and V
3407 * Vertical filtering is ok. However, bear in mind that the pixel center is at
3408 * 0.5, so add 0.5.
3409 */
3410 shader_addline(buffer, "FLR texcrd.x, texcrd.x;\n");
3411 shader_addline(buffer, "ADD texcrd.x, texcrd.x, coef.y;\n");
3412
3413 /* Divide the x coordinate by 0.5 and get the fraction. This gives 0.25 and 0.75 for the
3414 * even and odd pixels respectively
3415 */
3416 shader_addline(buffer, "MUL texcrd2, texcrd, coef.y;\n");
3417 shader_addline(buffer, "FRC texcrd2, texcrd2;\n");
3418
3419 /* Sample Pixel 1 */
3420 shader_addline(buffer, "%s luminance, texcrd, texture[0], %s;\n", texinstr, tex);
3421
3422 /* Put the value into either of the chroma values */
3423 shader_addline(buffer, "SGE temp.x, texcrd2.x, coef.y;\n");
3424 shader_addline(buffer, "MUL chroma.x, luminance.%c, temp.x;\n", chroma);
3425 shader_addline(buffer, "SLT temp.x, texcrd2.x, coef.y;\n");
3426 shader_addline(buffer, "MUL chroma.y, luminance.%c, temp.x;\n", chroma);
3427
3428 /* Sample pixel 2. If we read an even pixel(SLT above returned 1), sample
3429 * the pixel right to the current one. Otherwise, sample the left pixel.
3430 * Bias and scale the SLT result to -1;1 and add it to the texcrd.x.
3431 */
3432 shader_addline(buffer, "MAD temp.x, temp.x, coef.z, -coef.x;\n");
3433 shader_addline(buffer, "ADD texcrd.x, texcrd, temp.x;\n");
3434 shader_addline(buffer, "%s luminance, texcrd, texture[0], %s;\n", texinstr, tex);
3435
3436 /* Put the value into the other chroma */
3437 shader_addline(buffer, "SGE temp.x, texcrd2.x, coef.y;\n");
3438 shader_addline(buffer, "MAD chroma.y, luminance.%c, temp.x, chroma.y;\n", chroma);
3439 shader_addline(buffer, "SLT temp.x, texcrd2.x, coef.y;\n");
3440 shader_addline(buffer, "MAD chroma.x, luminance.%c, temp.x, chroma.x;\n", chroma);
3441
3442 /* TODO: If filtering is enabled, sample a 2nd pair of pixels left or right of
3443 * the current one and lerp the two U and V values
3444 */
3445
3446 /* This gives the correctly filtered luminance value */
3447 shader_addline(buffer, "TEX luminance, fragment.texcoord[0], texture[0], %s;\n", tex);
3448
3449 return TRUE;
3450 }
3451
3452 static BOOL gen_yv12_read(SHADER_BUFFER *buffer, GLenum textype, char *luminance)
3453 {
3454 const char *tex;
3455
3456 switch(textype) {
3457 case GL_TEXTURE_2D: tex = "2D"; break;
3458 case GL_TEXTURE_RECTANGLE_ARB: tex = "RECT"; break;
3459 default:
3460 FIXME("Implement yv12 correction for non-2d, non-rect textures\n");
3461 return FALSE;
3462 }
3463
3464 /* YV12 surfaces contain a WxH sized luminance plane, followed by a (W/2)x(H/2)
3465 * V and a (W/2)x(H/2) U plane, each with 8 bit per pixel. So the effective
3466 * bitdepth is 12 bits per pixel. Since the U and V planes have only half the
3467 * pitch of the luminance plane, the packing into the gl texture is a bit
3468 * unfortunate. If the whole texture is interpreted as luminance data it looks
3469 * approximately like this:
3470 *
3471 * +----------------------------------+----
3472 * | |
3473 * | |
3474 * | |
3475 * | |
3476 * | | 2
3477 * | LUMINANCE | -
3478 * | | 3
3479 * | |
3480 * | |
3481 * | |
3482 * | |
3483 * +----------------+-----------------+----
3484 * | | |
3485 * | U even rows | U odd rows |
3486 * | | | 1
3487 * +----------------+------------------ -
3488 * | | | 3
3489 * | V even rows | V odd rows |
3490 * | | |
3491 * +----------------+-----------------+----
3492 * | | |
3493 * | 0.5 | 0.5 |
3494 *
3495 * So it appears as if there are 4 chroma images, but in fact the odd rows
3496 * in the chroma images are in the same row as the even ones. So its is
3497 * kinda tricky to read
3498 *
3499 * When reading from rectangle textures, keep in mind that the input y coordinates
3500 * go from 0 to d3d_height, whereas the opengl texture height is 1.5 * d3d_height
3501 */
3502 shader_addline(buffer, "PARAM yv12_coef = {%f, %f, %f, %f};\n",
3503 2.0 / 3.0, 1.0 / 6.0, (2.0 / 3.0) + (1.0 / 6.0), 1.0 / 3.0);
3504
3505 shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3506 /* the chroma planes have only half the width */
3507 shader_addline(buffer, "MUL texcrd.x, texcrd.x, coef.y;\n");
3508
3509 /* The first value is between 2/3 and 5/6th of the texture's height, so scale+bias
3510 * the coordinate. Also read the right side of the image when reading odd lines
3511 *
3512 * Don't forget to clamp the y values in into the range, otherwise we'll get filtering
3513 * bleeding
3514 */
3515 if(textype == GL_TEXTURE_2D) {
3516
3517 shader_addline(buffer, "RCP chroma.w, size.y;\n");
3518
3519 shader_addline(buffer, "MUL texcrd2.y, texcrd.y, size.y;\n");
3520
3521 shader_addline(buffer, "FLR texcrd2.y, texcrd2.y;\n");
3522 shader_addline(buffer, "MAD texcrd.y, texcrd.y, yv12_coef.y, yv12_coef.x;\n");
3523
3524 /* Read odd lines from the right side(add size * 0.5 to the x coordinate */
3525 shader_addline(buffer, "ADD texcrd2.x, texcrd2.y, yv12_coef.y;\n"); /* To avoid 0.5 == 0.5 comparisons */
3526 shader_addline(buffer, "FRC texcrd2.x, texcrd2.x;\n");
3527 shader_addline(buffer, "SGE texcrd2.x, texcrd2.x, coef.y;\n");
3528 shader_addline(buffer, "MAD texcrd.x, texcrd2.x, coef.y, texcrd.x;\n");
3529
3530 /* clamp, keep the half pixel origin in mind */
3531 shader_addline(buffer, "MAD temp.y, coef.y, chroma.w, yv12_coef.x;\n");
3532 shader_addline(buffer, "MAX texcrd.y, temp.y, texcrd.y;\n");
3533 shader_addline(buffer, "MAD temp.y, -coef.y, chroma.w, yv12_coef.z;\n");
3534 shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3535 } else {
3536 /* Read from [size - size+size/4] */
3537 shader_addline(buffer, "FLR texcrd.y, texcrd.y;\n");
3538 shader_addline(buffer, "MAD texcrd.y, texcrd.y, coef.w, size.y;\n");
3539
3540 /* Read odd lines from the right side(add size * 0.5 to the x coordinate */
3541 shader_addline(buffer, "ADD texcrd2.x, texcrd.y, yv12_coef.y;\n"); /* To avoid 0.5 == 0.5 comparisons */
3542 shader_addline(buffer, "FRC texcrd2.x, texcrd2.x;\n");
3543 shader_addline(buffer, "SGE texcrd2.x, texcrd2.x, coef.y;\n");
3544 shader_addline(buffer, "MUL texcrd2.x, texcrd2.x, size.x;\n");
3545 shader_addline(buffer, "MAD texcrd.x, texcrd2.x, coef.y, texcrd.x;\n");
3546
3547 /* Make sure to read exactly from the pixel center */
3548 shader_addline(buffer, "FLR texcrd.y, texcrd.y;\n");
3549 shader_addline(buffer, "ADD texcrd.y, texcrd.y, coef.y;\n");
3550
3551 /* Clamp */
3552 shader_addline(buffer, "MAD temp.y, size.y, coef.w, size.y;\n");
3553 shader_addline(buffer, "ADD temp.y, temp.y, -coef.y;\n");
3554 shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3555 shader_addline(buffer, "ADD temp.y, size.y, -coef.y;\n");
3556 shader_addline(buffer, "MAX texcrd.y, temp.y, texcrd.y;\n");
3557 }
3558 /* Read the texture, put the result into the output register */
3559 shader_addline(buffer, "TEX temp, texcrd, texture[0], %s;\n", tex);
3560 shader_addline(buffer, "MOV chroma.x, temp.w;\n");
3561
3562 /* The other chroma value is 1/6th of the texture lower, from 5/6th to 6/6th
3563 * No need to clamp because we're just reusing the already clamped value from above
3564 */
3565 if(textype == GL_TEXTURE_2D) {
3566 shader_addline(buffer, "ADD texcrd.y, texcrd.y, yv12_coef.y;\n");
3567 } else {
3568 shader_addline(buffer, "MAD texcrd.y, size.y, coef.w, texcrd.y;\n");
3569 }
3570 shader_addline(buffer, "TEX temp, texcrd, texture[0], %s;\n", tex);
3571 shader_addline(buffer, "MOV chroma.y, temp.w;\n");
3572
3573 /* Sample the luminance value. It is in the top 2/3rd of the texture, so scale the y coordinate.
3574 * Clamp the y coordinate to prevent the chroma values from bleeding into the sampled luminance
3575 * values due to filtering
3576 */
3577 shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3578 if(textype == GL_TEXTURE_2D) {
3579 /* Multiply the y coordinate by 2/3 and clamp it */
3580 shader_addline(buffer, "MUL texcrd.y, texcrd.y, yv12_coef.x;\n");
3581 shader_addline(buffer, "MAD temp.y, -coef.y, chroma.w, yv12_coef.x;\n");
3582 shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3583 shader_addline(buffer, "TEX luminance, texcrd, texture[0], %s;\n", tex);
3584 } else {
3585 /* Reading from texture_rectangles is pretty straightforward, just use the unmodified
3586 * texture coordinate. It is still a good idea to clamp it though, since the opengl texture
3587 * is bigger
3588 */
3589 shader_addline(buffer, "ADD temp.x, size.y, -coef.y;\n");
3590 shader_addline(buffer, "MIN texcrd.y, texcrd.y, size.x;\n");
3591 shader_addline(buffer, "TEX luminance, texcrd, texture[0], %s;\n", tex);
3592 }
3593 *luminance = 'a';
3594
3595 return TRUE;
3596 }
3597
3598 static GLuint gen_yuv_shader(IWineD3DDeviceImpl *device, enum yuv_fixup yuv_fixup, GLenum textype)
3599 {
3600 GLenum shader;
3601 SHADER_BUFFER buffer;
3602 char luminance_component;
3603 struct arbfp_blit_priv *priv = device->blit_priv;
3604
3605 /* Shader header */
3606 shader_buffer_init(&buffer);
3607
3608 ENTER_GL();
3609 GL_EXTCALL(glGenProgramsARB(1, &shader));
3610 checkGLcall("GL_EXTCALL(glGenProgramsARB(1, &shader))");
3611 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader));
3612 checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader)");
3613 LEAVE_GL();
3614 if(!shader) {
3615 shader_buffer_free(&buffer);
3616 return 0;
3617 }
3618
3619 /* The YUY2 and UYVY formats contain two pixels packed into a 32 bit macropixel,
3620 * giving effectively 16 bit per pixel. The color consists of a luminance(Y) and
3621 * two chroma(U and V) values. Each macropixel has two luminance values, one for
3622 * each single pixel it contains, and one U and one V value shared between both
3623 * pixels.
3624 *
3625 * The data is loaded into an A8L8 texture. With YUY2, the luminance component
3626 * contains the luminance and alpha the chroma. With UYVY it is vice versa. Thus
3627 * take the format into account when generating the read swizzles
3628 *
3629 * Reading the Y value is straightforward - just sample the texture. The hardware
3630 * takes care of filtering in the horizontal and vertical direction.
3631 *
3632 * Reading the U and V values is harder. We have to avoid filtering horizontally,
3633 * because that would mix the U and V values of one pixel or two adjacent pixels.
3634 * Thus floor the texture coordinate and add 0.5 to get an unfiltered read,
3635 * regardless of the filtering setting. Vertical filtering works automatically
3636 * though - the U and V values of two rows are mixed nicely.
3637 *
3638 * Appart of avoiding filtering issues, the code has to know which value it just
3639 * read, and where it can find the other one. To determine this, it checks if
3640 * it sampled an even or odd pixel, and shifts the 2nd read accordingly.
3641 *
3642 * Handling horizontal filtering of U and V values requires reading a 2nd pair
3643 * of pixels, extracting U and V and mixing them. This is not implemented yet.
3644 *
3645 * An alternative implementation idea is to load the texture as A8R8G8B8 texture,
3646 * with width / 2. This way one read gives all 3 values, finding U and V is easy
3647 * in an unfiltered situation. Finding the luminance on the other hand requires
3648 * finding out if it is an odd or even pixel. The real drawback of this approach
3649 * is filtering. This would have to be emulated completely in the shader, reading
3650 * up two 2 packed pixels in up to 2 rows and interpolating both horizontally and
3651 * vertically. Beyond that it would require adjustments to the texture handling
3652 * code to deal with the width scaling
3653 */
3654 shader_addline(&buffer, "!!ARBfp1.0\n");
3655 shader_addline(&buffer, "TEMP luminance;\n");
3656 shader_addline(&buffer, "TEMP temp;\n");
3657 shader_addline(&buffer, "TEMP chroma;\n");
3658 shader_addline(&buffer, "TEMP texcrd;\n");
3659 shader_addline(&buffer, "TEMP texcrd2;\n");
3660 shader_addline(&buffer, "PARAM coef = {1.0, 0.5, 2.0, 0.25};\n");
3661 shader_addline(&buffer, "PARAM yuv_coef = {1.403, 0.344, 0.714, 1.770};\n");
3662 shader_addline(&buffer, "PARAM size = program.local[0];\n");
3663
3664 switch (yuv_fixup)
3665 {
3666 case YUV_FIXUP_UYVY:
3667 case YUV_FIXUP_YUY2:
3668 if (!gen_planar_yuv_read(&buffer, yuv_fixup, textype, &luminance_component))
3669 {
3670 shader_buffer_free(&buffer);
3671 return 0;
3672 }
3673 break;
3674
3675 case YUV_FIXUP_YV12:
3676 if (!gen_yv12_read(&buffer, textype, &luminance_component))
3677 {
3678 shader_buffer_free(&buffer);
3679 return 0;
3680 }
3681 break;
3682
3683 default:
3684 FIXME("Unsupported YUV fixup %#x\n", yuv_fixup);
3685 shader_buffer_free(&buffer);
3686 return 0;
3687 }
3688
3689 /* Calculate the final result. Formula is taken from
3690 * http://www.fourcc.org/fccyvrgb.php. Note that the chroma
3691 * ranges from -0.5 to 0.5
3692 */
3693 shader_addline(&buffer, "SUB chroma.xy, chroma, coef.y;\n");
3694
3695 shader_addline(&buffer, "MAD result.color.x, chroma.x, yuv_coef.x, luminance.%c;\n", luminance_component);
3696 shader_addline(&buffer, "MAD temp.x, -chroma.y, yuv_coef.y, luminance.%c;\n", luminance_component);
3697 shader_addline(&buffer, "MAD result.color.y, -chroma.x, yuv_coef.z, temp.x;\n");
3698 shader_addline(&buffer, "MAD result.color.z, chroma.y, yuv_coef.w, luminance.%c;\n", luminance_component);
3699 shader_addline(&buffer, "END\n");
3700
3701 ENTER_GL();
3702 GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(buffer.buffer), buffer.buffer));
3703
3704 if (glGetError() == GL_INVALID_OPERATION) {
3705 GLint pos;
3706 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
3707 FIXME("Fragment program error at position %d: %s\n", pos,
3708 debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
3709 }
3710 shader_buffer_free(&buffer);
3711 LEAVE_GL();
3712
3713 switch (yuv_fixup)
3714 {
3715 case YUV_FIXUP_YUY2:
3716 if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->yuy2_rect_shader = shader;
3717 else priv->yuy2_2d_shader = shader;
3718 break;
3719
3720 case YUV_FIXUP_UYVY:
3721 if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->uyvy_rect_shader = shader;
3722 else priv->uyvy_2d_shader = shader;
3723 break;
3724
3725 case YUV_FIXUP_YV12:
3726 if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->yv12_rect_shader = shader;
3727 else priv->yv12_2d_shader = shader;
3728 break;
3729 }
3730
3731 return shader;
3732 }
3733
3734 static HRESULT arbfp_blit_set(IWineD3DDevice *iface, const struct GlPixelFormatDesc *format_desc,
3735 GLenum textype, UINT width, UINT height)
3736 {
3737 GLenum shader;
3738 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3739 float size[4] = {width, height, 1, 1};
3740 struct arbfp_blit_priv *priv = device->blit_priv;
3741 enum yuv_fixup yuv_fixup;
3742
3743 if (!is_yuv_fixup(format_desc->color_fixup))
3744 {
3745 TRACE("Fixup:\n");
3746 dump_color_fixup_desc(format_desc->color_fixup);
3747 /* Don't bother setting up a shader for unconverted formats */
3748 ENTER_GL();
3749 glEnable(textype);
3750 checkGLcall("glEnable(textype)");
3751 LEAVE_GL();
3752 return WINED3D_OK;
3753 }
3754
3755 yuv_fixup = get_yuv_fixup(format_desc->color_fixup);
3756
3757 switch(yuv_fixup)
3758 {
3759 case YUV_FIXUP_YUY2:
3760 shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->yuy2_rect_shader : priv->yuy2_2d_shader;
3761 break;
3762
3763 case YUV_FIXUP_UYVY:
3764 shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->uyvy_rect_shader : priv->uyvy_2d_shader;
3765 break;
3766
3767 case YUV_FIXUP_YV12:
3768 shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->yv12_rect_shader : priv->yv12_2d_shader;
3769 break;
3770
3771 default:
3772 FIXME("Unsupported YUV fixup %#x, not setting a shader\n", yuv_fixup);
3773 ENTER_GL();
3774 glEnable(textype);
3775 checkGLcall("glEnable(textype)");
3776 LEAVE_GL();
3777 return E_NOTIMPL;
3778 }
3779
3780 if (!shader) shader = gen_yuv_shader(device, yuv_fixup, textype);
3781
3782 ENTER_GL();
3783 glEnable(GL_FRAGMENT_PROGRAM_ARB);
3784 checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB)");
3785 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader));
3786 checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader)");
3787 GL_EXTCALL(glProgramLocalParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, 0, size));
3788 checkGLcall("glProgramLocalParameter4fvARB");
3789 LEAVE_GL();
3790
3791 return WINED3D_OK;
3792 }
3793
3794 static void arbfp_blit_unset(IWineD3DDevice *iface) {
3795 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3796
3797 ENTER_GL();
3798 glDisable(GL_FRAGMENT_PROGRAM_ARB);
3799 checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
3800 glDisable(GL_TEXTURE_2D);
3801 checkGLcall("glDisable(GL_TEXTURE_2D)");
3802 if(GL_SUPPORT(ARB_TEXTURE_CUBE_MAP)) {
3803 glDisable(GL_TEXTURE_CUBE_MAP_ARB);
3804 checkGLcall("glDisable(GL_TEXTURE_CUBE_MAP_ARB)");
3805 }
3806 if(GL_SUPPORT(ARB_TEXTURE_RECTANGLE)) {
3807 glDisable(GL_TEXTURE_RECTANGLE_ARB);
3808 checkGLcall("glDisable(GL_TEXTURE_RECTANGLE_ARB)");
3809 }
3810 LEAVE_GL();
3811 }
3812
3813 static BOOL arbfp_blit_color_fixup_supported(struct color_fixup_desc fixup)
3814 {
3815 enum yuv_fixup yuv_fixup;
3816
3817 if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
3818 {
3819 TRACE("Checking support for fixup:\n");
3820 dump_color_fixup_desc(fixup);
3821 }
3822
3823 if (is_identity_fixup(fixup))
3824 {
3825 TRACE("[OK]\n");
3826 return TRUE;
3827 }
3828
3829 /* We only support YUV conversions. */
3830 if (!is_yuv_fixup(fixup))
3831 {
3832 TRACE("[FAILED]\n");
3833 return FALSE;
3834 }
3835
3836 yuv_fixup = get_yuv_fixup(fixup);
3837 switch(yuv_fixup)
3838 {
3839 case YUV_FIXUP_YUY2:
3840 case YUV_FIXUP_UYVY:
3841 case YUV_FIXUP_YV12:
3842 TRACE("[OK]\n");
3843 return TRUE;
3844
3845 default:
3846 FIXME("Unsupported YUV fixup %#x\n", yuv_fixup);
3847 TRACE("[FAILED]\n");
3848 return FALSE;
3849 }
3850 }
3851
3852 const struct blit_shader arbfp_blit = {
3853 arbfp_blit_alloc,
3854 arbfp_blit_free,
3855 arbfp_blit_set,
3856 arbfp_blit_unset,
3857 arbfp_blit_color_fixup_supported,
3858 };
3859
3860 #undef GLINFO_LOCATION
useful hacks and other patches not (yet) suitable for mainline