?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
push d0760662e4c9af92a52cfc83b053a1228de45087
[wine/hacks.git] / dlls / wined3d / arb_program_shader.c
bloba7a79c190f4d3514916dfd7e150de3abea790930
1 /*
2 * Pixel and vertex shaders implementation using ARB_vertex_program
3 * and ARB_fragment_program GL extensions.
4 *
5 * Copyright 2002-2003 Jason Edmeades
6 * Copyright 2002-2003 Raphael Junqueira
7 * Copyright 2004 Christian Costa
8 * Copyright 2005 Oliver Stieber
9 * Copyright 2006 Ivan Gyurdiev
10 * Copyright 2006 Jason Green
11 * Copyright 2006 Henri Verbeet
12 * Copyright 2007-2008 Stefan Dösinger for CodeWeavers
13 *
14 * This library is free software; you can redistribute it and/or
15 * modify it under the terms of the GNU Lesser General Public
16 * License as published by the Free Software Foundation; either
17 * version 2.1 of the License, or (at your option) any later version.
18 *
19 * This library is distributed in the hope that it will be useful,
20 * but WITHOUT ANY WARRANTY; without even the implied warranty of
21 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
22 * Lesser General Public License for more details.
23 *
24 * You should have received a copy of the GNU Lesser General Public
25 * License along with this library; if not, write to the Free Software
26 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
27 */
28
29 #include "config.h"
30
31 #include <math.h>
32 #include <stdio.h>
33
34 #include "wined3d_private.h"
35
36 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
37 WINE_DECLARE_DEBUG_CHANNEL(d3d_constants);
38 WINE_DECLARE_DEBUG_CHANNEL(d3d_caps);
39 WINE_DECLARE_DEBUG_CHANNEL(d3d);
40
41 #define GLINFO_LOCATION (*gl_info)
42
43 /* We have to subtract any other PARAMs that we might use in our shader programs.
44 * ATI seems to count 2 implicit PARAMs when we use fog and NVIDIA counts 1,
45 * and we reference one row of the PROJECTION matrix which counts as 1 PARAM. */
46 #define ARB_SHADER_RESERVED_VS_CONSTS 3
47
48 /* The arb shader only loads the bump mapping environment matrix into the shader if it finds
49 * a free constant to do that, so only reduce the number of available constants by 2 for the fog states.
50 */
51 #define ARB_SHADER_RESERVED_PS_CONSTS 2
52
53 /* Internally used shader constants. Applications can use constants 0 to GL_LIMITS(vshader_constantsF) - 1,
54 * so upload them above that
55 */
56 #define ARB_SHADER_PRIVCONST_BASE (GL_LIMITS(vshader_constantsF) - ARB_SHADER_RESERVED_VS_CONSTS)
57 #define ARB_SHADER_PRIVCONST_POS ARB_SHADER_PRIVCONST_BASE + 0
58
59 /* ARB_program_shader private data */
60 struct shader_arb_priv {
61 GLuint current_vprogram_id;
62 GLuint current_fprogram_id;
63 GLuint depth_blt_vprogram_id;
64 GLuint depth_blt_fprogram_id[tex_type_count];
65 BOOL use_arbfp_fixed_func;
66 struct hash_table_t *fragment_shaders;
67 };
68
69 /********************************************************
70 * ARB_[vertex/fragment]_program helper functions follow
71 ********************************************************/
72
73 /**
74 * Loads floating point constants into the currently set ARB_vertex/fragment_program.
75 * When constant_list == NULL, it will load all the constants.
76 *
77 * @target_type should be either GL_VERTEX_PROGRAM_ARB (for vertex shaders)
78 * or GL_FRAGMENT_PROGRAM_ARB (for pixel shaders)
79 */
80 static unsigned int shader_arb_load_constantsF(IWineD3DBaseShaderImpl* This, const WineD3D_GL_Info *gl_info,
81 GLuint target_type, unsigned int max_constants, const float *constants, char *dirty_consts)
82 {
83 DWORD shader_version = This->baseShader.reg_maps.shader_version;
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 && WINED3DSHADER_VERSION_MAJOR(shader_version) == 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 /* Generate the variable & register declarations for the ARB_vertex_program output target */
273 static void shader_generate_arb_declarations(IWineD3DBaseShader *iface, const shader_reg_maps *reg_maps,
274 SHADER_BUFFER *buffer, const WineD3D_GL_Info *gl_info)
275 {
276 IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
277 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device;
278 DWORD i, cur;
279 char pshader = shader_is_pshader_version(reg_maps->shader_version);
280 unsigned max_constantsF = min(This->baseShader.limits.constant_float,
281 (pshader ? GL_LIMITS(pshader_constantsF) - ARB_SHADER_RESERVED_PS_CONSTS :
282 GL_LIMITS(vshader_constantsF) - ARB_SHADER_RESERVED_VS_CONSTS));
283 UINT extra_constants_needed = 0;
284 const local_constant *lconst;
285
286 /* Temporary Output register */
287 shader_addline(buffer, "TEMP TMP_OUT;\n");
288
289 for(i = 0; i < This->baseShader.limits.temporary; i++) {
290 if (reg_maps->temporary[i])
291 shader_addline(buffer, "TEMP R%u;\n", i);
292 }
293
294 for (i = 0; i < This->baseShader.limits.address; i++) {
295 if (reg_maps->address[i])
296 shader_addline(buffer, "ADDRESS A%d;\n", i);
297 }
298
299 for(i = 0; i < This->baseShader.limits.texcoord; i++) {
300 if (reg_maps->texcoord[i])
301 shader_addline(buffer,"TEMP T%u;\n", i);
302 }
303
304 /* Texture coordinate registers must be pre-loaded */
305 for (i = 0; i < This->baseShader.limits.texcoord; i++) {
306 if (reg_maps->texcoord[i])
307 shader_addline(buffer, "MOV T%u, fragment.texcoord[%u];\n", i, i);
308 }
309
310 for(i = 0; i < (sizeof(reg_maps->bumpmat) / sizeof(reg_maps->bumpmat[0])); i++) {
311 IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *) This;
312 if(!reg_maps->bumpmat[i]) continue;
313
314 cur = ps->numbumpenvmatconsts;
315 ps->bumpenvmatconst[cur].const_num = -1;
316 ps->bumpenvmatconst[cur].texunit = i;
317 ps->luminanceconst[cur].const_num = -1;
318 ps->luminanceconst[cur].texunit = i;
319
320 /* If the shader does not use all available constants, use the next free constant to load the bump mapping environment matrix from
321 * the stateblock into the shader. If no constant is available don't load, texbem will then just sample the texture without applying
322 * bump mapping.
323 */
324 if(max_constantsF + extra_constants_needed < GL_LIMITS(pshader_constantsF) - ARB_SHADER_RESERVED_PS_CONSTS) {
325 ps->bumpenvmatconst[cur].const_num = max_constantsF + extra_constants_needed;
326 shader_addline(buffer, "PARAM bumpenvmat%d = program.env[%d];\n",
327 i, ps->bumpenvmatconst[cur].const_num);
328 extra_constants_needed++;
329
330 if(reg_maps->luminanceparams && max_constantsF + extra_constants_needed < GL_LIMITS(pshader_constantsF) - ARB_SHADER_RESERVED_PS_CONSTS) {
331 ((IWineD3DPixelShaderImpl *)This)->luminanceconst[cur].const_num = max_constantsF + extra_constants_needed;
332 shader_addline(buffer, "PARAM luminance%d = program.env[%d];\n",
333 i, ps->luminanceconst[cur].const_num);
334 extra_constants_needed++;
335 } else if(reg_maps->luminanceparams) {
336 FIXME("No free constant to load the luminance parameters\n");
337 }
338 } else {
339 FIXME("No free constant found to load environment bump mapping matrix into the shader. texbem instruction will not apply bump mapping\n");
340 }
341
342 ps->numbumpenvmatconsts = cur + 1;
343 }
344
345 if(device->stateBlock->renderState[WINED3DRS_SRGBWRITEENABLE] && pshader) {
346 shader_addline(buffer, "PARAM srgb_mul_low = {%f, %f, %f, 1.0};\n",
347 srgb_mul_low, srgb_mul_low, srgb_mul_low);
348 shader_addline(buffer, "PARAM srgb_comparison = {%f, %f, %f, %f};\n",
349 srgb_cmp, srgb_cmp, srgb_cmp, srgb_cmp);
350 shader_addline(buffer, "PARAM srgb_pow = {%f, %f, %f, 1.0};\n",
351 srgb_pow, srgb_pow, srgb_pow);
352 shader_addline(buffer, "PARAM srgb_mul_hi = {%f, %f, %f, 1.0};\n",
353 srgb_mul_high, srgb_mul_high, srgb_mul_high);
354 shader_addline(buffer, "PARAM srgb_sub_hi = {%f, %f, %f, 0.0};\n",
355 srgb_sub_high, srgb_sub_high, srgb_sub_high);
356 }
357
358 /* Load local constants using the program-local space,
359 * this avoids reloading them each time the shader is used
360 */
361 if(!This->baseShader.load_local_constsF) {
362 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
363 shader_addline(buffer, "PARAM C%u = program.local[%u];\n", lconst->idx,
364 lconst->idx);
365 }
366 }
367
368 /* we use the array-based constants array if the local constants are marked for loading,
369 * because then we use indirect addressing, or when the local constant list is empty,
370 * because then we don't know if we're using indirect addressing or not. If we're hardcoding
371 * local constants do not declare the loaded constants as an array because ARB compilers usually
372 * do not optimize unused constants away
373 */
374 if(This->baseShader.load_local_constsF || list_empty(&This->baseShader.constantsF)) {
375 /* Need to PARAM the environment parameters (constants) so we can use relative addressing */
376 shader_addline(buffer, "PARAM C[%d] = { program.env[0..%d] };\n",
377 max_constantsF, max_constantsF - 1);
378 } else {
379 for(i = 0; i < max_constantsF; i++) {
380 if(!shader_constant_is_local(This, i)) {
381 shader_addline(buffer, "PARAM C%d = program.env[%d];\n",i, i);
382 }
383 }
384 }
385 }
386
387 static const char * const shift_tab[] = {
388 "dummy", /* 0 (none) */
389 "coefmul.x", /* 1 (x2) */
390 "coefmul.y", /* 2 (x4) */
391 "coefmul.z", /* 3 (x8) */
392 "coefmul.w", /* 4 (x16) */
393 "dummy", /* 5 (x32) */
394 "dummy", /* 6 (x64) */
395 "dummy", /* 7 (x128) */
396 "dummy", /* 8 (d256) */
397 "dummy", /* 9 (d128) */
398 "dummy", /* 10 (d64) */
399 "dummy", /* 11 (d32) */
400 "coefdiv.w", /* 12 (d16) */
401 "coefdiv.z", /* 13 (d8) */
402 "coefdiv.y", /* 14 (d4) */
403 "coefdiv.x" /* 15 (d2) */
404 };
405
406 static void shader_arb_get_write_mask(const struct wined3d_shader_instruction *ins,
407 const struct wined3d_shader_dst_param *dst, char *write_mask)
408 {
409 char *ptr = write_mask;
410 char vshader = shader_is_vshader_version(ins->ctx->reg_maps->shader_version);
411
412 if (vshader && dst->register_type == WINED3DSPR_ADDR)
413 {
414 *ptr++ = '.';
415 *ptr++ = 'x';
416 }
417 else if (dst->write_mask != WINED3DSP_WRITEMASK_ALL)
418 {
419 *ptr++ = '.';
420 if (dst->write_mask & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
421 if (dst->write_mask & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
422 if (dst->write_mask & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
423 if (dst->write_mask & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
424 }
425
426 *ptr = '\0';
427 }
428
429 static void shader_arb_get_swizzle(const DWORD param, BOOL fixup, char *swizzle_str) {
430 /* For registers of type WINED3DDECLTYPE_D3DCOLOR, data is stored as "bgra",
431 * but addressed as "rgba". To fix this we need to swap the register's x
432 * and z components. */
433 const char *swizzle_chars = fixup ? "zyxw" : "xyzw";
434 char *ptr = swizzle_str;
435
436 /* swizzle bits fields: wwzzyyxx */
437 DWORD swizzle = (param & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
438 DWORD swizzle_x = swizzle & 0x03;
439 DWORD swizzle_y = (swizzle >> 2) & 0x03;
440 DWORD swizzle_z = (swizzle >> 4) & 0x03;
441 DWORD swizzle_w = (swizzle >> 6) & 0x03;
442
443 /* If the swizzle is the default swizzle (ie, "xyzw"), we don't need to
444 * generate a swizzle string. Unless we need to our own swizzling. */
445 if ((WINED3DSP_NOSWIZZLE >> WINED3DSP_SWIZZLE_SHIFT) != swizzle || fixup) {
446 *ptr++ = '.';
447 if (swizzle_x == swizzle_y && swizzle_x == swizzle_z && swizzle_x == swizzle_w) {
448 *ptr++ = swizzle_chars[swizzle_x];
449 } else {
450 *ptr++ = swizzle_chars[swizzle_x];
451 *ptr++ = swizzle_chars[swizzle_y];
452 *ptr++ = swizzle_chars[swizzle_z];
453 *ptr++ = swizzle_chars[swizzle_w];
454 }
455 }
456
457 *ptr = '\0';
458 }
459
460 static void shader_arb_get_register_name(IWineD3DBaseShader *iface, WINED3DSHADER_PARAM_REGISTER_TYPE register_type,
461 UINT register_idx, BOOL rel_addr, char *register_name, BOOL *is_color)
462 {
463 /* oPos, oFog and oPts in D3D */
464 static const char * const rastout_reg_names[] = {"TMP_OUT", "result.fogcoord", "result.pointsize"};
465 IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)iface;
466 DWORD shader_version = This->baseShader.reg_maps.shader_version;
467 BOOL pshader = shader_is_pshader_version(shader_version);
468
469 *is_color = FALSE;
470
471 switch (register_type)
472 {
473 case WINED3DSPR_TEMP:
474 sprintf(register_name, "R%u", register_idx);
475 break;
476
477 case WINED3DSPR_INPUT:
478 if (pshader)
479 {
480 if (register_idx == 0) strcpy(register_name, "fragment.color.primary");
481 else strcpy(register_name, "fragment.color.secondary");
482 }
483 else
484 {
485 if (((IWineD3DVertexShaderImpl *)This)->cur_args->swizzle_map & (1 << register_idx)) *is_color = TRUE;
486 sprintf(register_name, "vertex.attrib[%u]", register_idx);
487 }
488 break;
489
490 case WINED3DSPR_CONST:
491 if (!pshader && rel_addr)
492 {
493 UINT rel_offset = ((IWineD3DVertexShaderImpl *)This)->rel_offset;
494 if (register_idx >= rel_offset)
495 sprintf(register_name, "C[A0.x + %u]", register_idx - rel_offset);
496 else
497 sprintf(register_name, "C[A0.x - %u]", -register_idx + rel_offset);
498 }
499 else
500 {
501 if (This->baseShader.load_local_constsF || list_empty(&This->baseShader.constantsF))
502 sprintf(register_name, "C[%u]", register_idx);
503 else
504 sprintf(register_name, "C%u", register_idx);
505 }
506 break;
507
508 case WINED3DSPR_TEXTURE: /* case WINED3DSPR_ADDR: */
509 if (pshader) sprintf(register_name, "T%u", register_idx);
510 else sprintf(register_name, "A%u", register_idx);
511 break;
512
513 case WINED3DSPR_COLOROUT:
514 if (register_idx == 0)
515 {
516 strcpy(register_name, "TMP_COLOR");
517 }
518 else
519 {
520 /* TODO: See GL_ARB_draw_buffers */
521 FIXME("Unsupported write to render target %u\n", register_idx);
522 sprintf(register_name, "unsupported_register");
523 }
524 break;
525
526 case WINED3DSPR_RASTOUT:
527 sprintf(register_name, "%s", rastout_reg_names[register_idx]);
528 break;
529
530 case WINED3DSPR_DEPTHOUT:
531 strcpy(register_name, "result.depth");
532 break;
533
534 case WINED3DSPR_ATTROUT:
535 if (pshader) sprintf(register_name, "oD[%u]", register_idx);
536 else if (register_idx == 0) strcpy(register_name, "result.color.primary");
537 else strcpy(register_name, "result.color.secondary");
538 break;
539
540 case WINED3DSPR_TEXCRDOUT:
541 if (pshader) sprintf(register_name, "oT[%u]", register_idx);
542 else sprintf(register_name, "result.texcoord[%u]", register_idx);
543 break;
544
545 default:
546 FIXME("Unhandled register type %#x[%u]\n", register_type, register_idx);
547 sprintf(register_name, "unrecognized_register[%u]", register_idx);
548 break;
549 }
550 }
551
552 static void shader_arb_add_src_param(const struct wined3d_shader_instruction *ins,
553 DWORD param, char *str)
554 {
555 char register_name[255];
556 char swizzle[6];
557 BOOL is_color;
558
559 if ((param & WINED3DSP_SRCMOD_MASK) == WINED3DSPSM_NEG) strcat(str, " -");
560 else strcat(str, " ");
561
562 shader_arb_get_register_name(ins->ctx->shader, shader_get_regtype(param), param & WINED3DSP_REGNUM_MASK,
563 param & WINED3DSHADER_ADDRMODE_RELATIVE, register_name, &is_color);
564 strcat(str, register_name);
565
566 shader_arb_get_swizzle(param, is_color, swizzle);
567 strcat(str, swizzle);
568 }
569
570 static void shader_arb_add_dst_param(const struct wined3d_shader_instruction *ins,
571 const struct wined3d_shader_dst_param *wined3d_dst, char *str)
572 {
573 char register_name[255];
574 char write_mask[6];
575 BOOL is_color;
576
577 strcat(str, " ");
578
579 shader_arb_get_register_name(ins->ctx->shader, wined3d_dst->register_type,
580 wined3d_dst->register_idx, wined3d_dst->has_rel_addr, register_name, &is_color);
581 strcat(str, register_name);
582
583 shader_arb_get_write_mask(ins, wined3d_dst, write_mask);
584 strcat(str, write_mask);
585 }
586
587 static const char *shader_arb_get_fixup_swizzle(enum fixup_channel_source channel_source)
588 {
589 switch(channel_source)
590 {
591 case CHANNEL_SOURCE_ZERO: return "0";
592 case CHANNEL_SOURCE_ONE: return "1";
593 case CHANNEL_SOURCE_X: return "x";
594 case CHANNEL_SOURCE_Y: return "y";
595 case CHANNEL_SOURCE_Z: return "z";
596 case CHANNEL_SOURCE_W: return "w";
597 default:
598 FIXME("Unhandled channel source %#x\n", channel_source);
599 return "undefined";
600 }
601 }
602
603 static void gen_color_correction(SHADER_BUFFER *buffer, const char *reg, DWORD dst_mask,
604 const char *one, const char *two, struct color_fixup_desc fixup)
605 {
606 DWORD mask;
607
608 if (is_yuv_fixup(fixup))
609 {
610 enum yuv_fixup yuv_fixup = get_yuv_fixup(fixup);
611 FIXME("YUV fixup (%#x) not supported\n", yuv_fixup);
612 return;
613 }
614
615 mask = 0;
616 if (fixup.x_source != CHANNEL_SOURCE_X) mask |= WINED3DSP_WRITEMASK_0;
617 if (fixup.y_source != CHANNEL_SOURCE_Y) mask |= WINED3DSP_WRITEMASK_1;
618 if (fixup.z_source != CHANNEL_SOURCE_Z) mask |= WINED3DSP_WRITEMASK_2;
619 if (fixup.w_source != CHANNEL_SOURCE_W) mask |= WINED3DSP_WRITEMASK_3;
620 mask &= dst_mask;
621
622 if (mask)
623 {
624 shader_addline(buffer, "SWZ %s, %s, %s, %s, %s, %s;\n", reg, reg,
625 shader_arb_get_fixup_swizzle(fixup.x_source), shader_arb_get_fixup_swizzle(fixup.y_source),
626 shader_arb_get_fixup_swizzle(fixup.z_source), shader_arb_get_fixup_swizzle(fixup.w_source));
627 }
628
629 mask = 0;
630 if (fixup.x_sign_fixup) mask |= WINED3DSP_WRITEMASK_0;
631 if (fixup.y_sign_fixup) mask |= WINED3DSP_WRITEMASK_1;
632 if (fixup.z_sign_fixup) mask |= WINED3DSP_WRITEMASK_2;
633 if (fixup.w_sign_fixup) mask |= WINED3DSP_WRITEMASK_3;
634 mask &= dst_mask;
635
636 if (mask)
637 {
638 char reg_mask[6];
639 char *ptr = reg_mask;
640
641 if (mask != WINED3DSP_WRITEMASK_ALL)
642 {
643 *ptr++ = '.';
644 if (mask & WINED3DSP_WRITEMASK_0) *ptr++ = 'x';
645 if (mask & WINED3DSP_WRITEMASK_1) *ptr++ = 'y';
646 if (mask & WINED3DSP_WRITEMASK_2) *ptr++ = 'z';
647 if (mask & WINED3DSP_WRITEMASK_3) *ptr++ = 'w';
648 }
649 *ptr = '\0';
650
651 shader_addline(buffer, "MAD %s%s, %s, %s, -%s;\n", reg, reg_mask, reg, two, one);
652 }
653 }
654
655 static void shader_hw_sample(const struct wined3d_shader_instruction *ins, DWORD sampler_idx,
656 const char *dst_str, const char *coord_reg, BOOL projected, BOOL bias)
657 {
658 SHADER_BUFFER *buffer = ins->ctx->buffer;
659 DWORD sampler_type = ins->ctx->reg_maps->samplers[sampler_idx] & WINED3DSP_TEXTURETYPE_MASK;
660 const char *tex_type;
661 IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)ins->ctx->shader;
662 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device;
663
664 switch(sampler_type) {
665 case WINED3DSTT_1D:
666 tex_type = "1D";
667 break;
668
669 case WINED3DSTT_2D:
670 if(device->stateBlock->textures[sampler_idx] &&
671 IWineD3DBaseTexture_GetTextureDimensions(device->stateBlock->textures[sampler_idx]) == GL_TEXTURE_RECTANGLE_ARB) {
672 tex_type = "RECT";
673 } else {
674 tex_type = "2D";
675 }
676 if (shader_is_pshader_version(ins->ctx->reg_maps->shader_version))
677 {
678 const IWineD3DPixelShaderImpl* const ps = (const IWineD3DPixelShaderImpl*)This;
679 if(ps->cur_args->np2_fixup & (1 << sampler_idx)) {
680 FIXME("NP2 texcoord fixup is currently not implemented in ARB mode (use GLSL instead).\n");
681 }
682 }
683 break;
684
685 case WINED3DSTT_VOLUME:
686 tex_type = "3D";
687 break;
688
689 case WINED3DSTT_CUBE:
690 tex_type = "CUBE";
691 break;
692
693 default:
694 ERR("Unexpected texture type %d\n", sampler_type);
695 tex_type = "";
696 }
697
698 if (bias) {
699 /* Shouldn't be possible, but let's check for it */
700 if(projected) FIXME("Biased and Projected texture sampling\n");
701 /* TXB takes the 4th component of the source vector automatically, as d3d. Nothing more to do */
702 shader_addline(buffer, "TXB %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
703 } else if (projected) {
704 shader_addline(buffer, "TXP %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
705 } else {
706 shader_addline(buffer, "TEX %s, %s, texture[%u], %s;\n", dst_str, coord_reg, sampler_idx, tex_type);
707 }
708
709 if (shader_is_pshader_version(ins->ctx->reg_maps->shader_version))
710 {
711 IWineD3DPixelShaderImpl *ps = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
712 gen_color_correction(buffer, dst_str, ins->dst[0].write_mask,
713 "one", "coefmul.x", ps->cur_args->color_fixup[sampler_idx]);
714 }
715 }
716
717 static void pshader_gen_input_modifier_line (
718 IWineD3DBaseShader *iface,
719 SHADER_BUFFER* buffer,
720 const DWORD instr,
721 int tmpreg,
722 char *outregstr) {
723
724 /* Generate a line that does the input modifier computation and return the input register to use */
725 BOOL is_color = FALSE;
726 char regstr[256];
727 char swzstr[20];
728 int insert_line;
729
730 /* Assume a new line will be added */
731 insert_line = 1;
732
733 /* Get register name */
734 shader_arb_get_register_name(iface, shader_get_regtype(instr), instr & WINED3DSP_REGNUM_MASK,
735 instr & WINED3DSHADER_ADDRMODE_RELATIVE, regstr, &is_color);
736 shader_arb_get_swizzle(instr, is_color, swzstr);
737
738 switch (instr & WINED3DSP_SRCMOD_MASK) {
739 case WINED3DSPSM_NONE:
740 sprintf(outregstr, "%s%s", regstr, swzstr);
741 insert_line = 0;
742 break;
743 case WINED3DSPSM_NEG:
744 sprintf(outregstr, "-%s%s", regstr, swzstr);
745 insert_line = 0;
746 break;
747 case WINED3DSPSM_BIAS:
748 shader_addline(buffer, "ADD T%c, %s, -coefdiv.x;\n", 'A' + tmpreg, regstr);
749 break;
750 case WINED3DSPSM_BIASNEG:
751 shader_addline(buffer, "ADD T%c, -%s, coefdiv.x;\n", 'A' + tmpreg, regstr);
752 break;
753 case WINED3DSPSM_SIGN:
754 shader_addline(buffer, "MAD T%c, %s, coefmul.x, -one.x;\n", 'A' + tmpreg, regstr);
755 break;
756 case WINED3DSPSM_SIGNNEG:
757 shader_addline(buffer, "MAD T%c, %s, -coefmul.x, one.x;\n", 'A' + tmpreg, regstr);
758 break;
759 case WINED3DSPSM_COMP:
760 shader_addline(buffer, "SUB T%c, one.x, %s;\n", 'A' + tmpreg, regstr);
761 break;
762 case WINED3DSPSM_X2:
763 shader_addline(buffer, "ADD T%c, %s, %s;\n", 'A' + tmpreg, regstr, regstr);
764 break;
765 case WINED3DSPSM_X2NEG:
766 shader_addline(buffer, "ADD T%c, -%s, -%s;\n", 'A' + tmpreg, regstr, regstr);
767 break;
768 case WINED3DSPSM_DZ:
769 shader_addline(buffer, "RCP T%c, %s.z;\n", 'A' + tmpreg, regstr);
770 shader_addline(buffer, "MUL T%c, %s, T%c;\n", 'A' + tmpreg, regstr, 'A' + tmpreg);
771 break;
772 case WINED3DSPSM_DW:
773 shader_addline(buffer, "RCP T%c, %s.w;\n", 'A' + tmpreg, regstr);
774 shader_addline(buffer, "MUL T%c, %s, T%c;\n", 'A' + tmpreg, regstr, 'A' + tmpreg);
775 break;
776 default:
777 sprintf(outregstr, "%s%s", regstr, swzstr);
778 insert_line = 0;
779 }
780
781 /* Return modified or original register, with swizzle */
782 if (insert_line)
783 sprintf(outregstr, "T%c%s", 'A' + tmpreg, swzstr);
784 }
785
786 static inline void pshader_gen_output_modifier_line(SHADER_BUFFER *buffer, int saturate, const char *write_mask,
787 int shift, const char *regstr)
788 {
789 /* Generate a line that does the output modifier computation */
790 shader_addline(buffer, "MUL%s %s%s, %s, %s;\n", saturate ? "_SAT" : "",
791 regstr, write_mask, regstr, shift_tab[shift]);
792 }
793
794 static void pshader_hw_bem(const struct wined3d_shader_instruction *ins)
795 {
796 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
797 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
798 SHADER_BUFFER *buffer = ins->ctx->buffer;
799 char dst_name[50];
800 char src_name[2][50];
801 char dst_wmask[20];
802 DWORD sampler_code = dst->register_idx;
803 BOOL has_bumpmat = FALSE;
804 BOOL is_color;
805 int i;
806
807 for(i = 0; i < This->numbumpenvmatconsts; i++) {
808 if (This->bumpenvmatconst[i].const_num != WINED3D_CONST_NUM_UNUSED
809 && This->bumpenvmatconst[i].texunit == sampler_code)
810 {
811 has_bumpmat = TRUE;
812 break;
813 }
814 }
815
816 shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
817 dst->register_idx, dst->has_rel_addr, dst_name, &is_color);
818 shader_arb_get_write_mask(ins, dst, dst_wmask);
819 strcat(dst_name, dst_wmask);
820
821 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[0].token, 0, src_name[0]);
822 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[1].token, 1, src_name[1]);
823
824 if(has_bumpmat) {
825 /* Sampling the perturbation map in Tsrc was done already, including the signedness correction if needed */
826 shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, x, z, 0, 0;\n", sampler_code);
827 shader_addline(buffer, "DP3 TMP.r, TMP2, %s;\n", src_name[1]);
828 shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, y, w, 0, 0;\n", sampler_code);
829 shader_addline(buffer, "DP3 TMP.g, TMP2, %s;\n", src_name[1]);
830
831 shader_addline(buffer, "ADD %s, %s, TMP;\n", dst_name, src_name[0]);
832 } else {
833 shader_addline(buffer, "MOV %s, %s;\n", dst_name, src_name[0]);
834 }
835 }
836
837 static void pshader_hw_cnd(const struct wined3d_shader_instruction *ins)
838 {
839 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
840 SHADER_BUFFER *buffer = ins->ctx->buffer;
841 char dst_wmask[20];
842 char dst_name[50];
843 char src_name[3][50];
844 BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
845 DWORD shift = dst->shift;
846 BOOL is_color;
847
848 /* FIXME: support output modifiers */
849
850 /* Handle output register */
851 shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
852 dst->register_idx, dst->has_rel_addr, dst_name, &is_color);
853 shader_arb_get_write_mask(ins, dst, dst_wmask);
854
855 /* Generate input register names (with modifiers) */
856 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[0].token, 0, src_name[0]);
857 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[1].token, 1, src_name[1]);
858 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[2].token, 2, src_name[2]);
859
860 /* The coissue flag changes the semantic of the cnd instruction in <= 1.3 shaders */
861 if (ins->ctx->reg_maps->shader_version <= WINED3DPS_VERSION(1, 3) && ins->coissue)
862 {
863 shader_addline(buffer, "MOV%s %s%s, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask, src_name[1]);
864 } else {
865 shader_addline(buffer, "ADD TMP, -%s, coefdiv.x;\n", src_name[0]);
866 shader_addline(buffer, "CMP%s %s%s, TMP, %s, %s;\n",
867 sat ? "_SAT" : "", dst_name, dst_wmask, src_name[1], src_name[2]);
868 }
869 if (shift != 0)
870 pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
871 }
872
873 static void pshader_hw_cmp(const struct wined3d_shader_instruction *ins)
874 {
875 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
876 SHADER_BUFFER *buffer = ins->ctx->buffer;
877 char dst_wmask[20];
878 char dst_name[50];
879 char src_name[3][50];
880 DWORD shift = dst->shift;
881 BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
882 BOOL is_color;
883
884 /* FIXME: support output modifiers */
885
886 /* Handle output register */
887 shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
888 dst->register_idx, dst->has_rel_addr, dst_name, &is_color);
889 shader_arb_get_write_mask(ins, dst, dst_wmask);
890
891 /* Generate input register names (with modifiers) */
892 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[0].token, 0, src_name[0]);
893 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[1].token, 1, src_name[1]);
894 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[2].token, 2, src_name[2]);
895
896 shader_addline(buffer, "CMP%s %s%s, %s, %s, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask,
897 src_name[0], src_name[2], src_name[1]);
898
899 if (shift != 0)
900 pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
901 }
902
903 /** Process the WINED3DSIO_DP2ADD instruction in ARB.
904 * dst = dot2(src0, src1) + src2 */
905 static void pshader_hw_dp2add(const struct wined3d_shader_instruction *ins)
906 {
907 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
908 SHADER_BUFFER *buffer = ins->ctx->buffer;
909 char dst_wmask[20];
910 char dst_name[50];
911 char src_name[3][50];
912 DWORD shift = dst->shift;
913 BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
914 BOOL is_color;
915
916 shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
917 dst->register_idx, dst->has_rel_addr, dst_name, &is_color);
918 shader_arb_get_write_mask(ins, dst, dst_wmask);
919
920 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[0].token, 0, src_name[0]);
921 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[1].token, 1, src_name[1]);
922 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[2].token, 2, src_name[2]);
923
924 /* Emulate a DP2 with a DP3 and 0.0 */
925 shader_addline(buffer, "MOV TMP, %s;\n", src_name[0]);
926 shader_addline(buffer, "MOV TMP.z, 0.0;\n");
927 shader_addline(buffer, "DP3 TMP2, TMP, %s;\n", src_name[1]);
928 shader_addline(buffer, "ADD%s %s%s, TMP2, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask, src_name[2]);
929
930 if (shift != 0)
931 pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
932 }
933
934 /* Map the opcode 1-to-1 to the GL code */
935 static void shader_hw_map2gl(const struct wined3d_shader_instruction *ins)
936 {
937 SHADER_BUFFER *buffer = ins->ctx->buffer;
938 const char *instruction;
939 char arguments[256];
940 unsigned int i;
941
942 switch (ins->handler_idx)
943 {
944 case WINED3DSIH_ABS: instruction = "ABS"; break;
945 case WINED3DSIH_ADD: instruction = "ADD"; break;
946 case WINED3DSIH_CRS: instruction = "XPD"; break;
947 case WINED3DSIH_DP3: instruction = "DP3"; break;
948 case WINED3DSIH_DP4: instruction = "DP4"; break;
949 case WINED3DSIH_DST: instruction = "DST"; break;
950 case WINED3DSIH_EXP: instruction = "EX2"; break;
951 case WINED3DSIH_EXPP: instruction = "EXP"; break;
952 case WINED3DSIH_FRC: instruction = "FRC"; break;
953 case WINED3DSIH_LIT: instruction = "LIT"; break;
954 case WINED3DSIH_LOG: instruction = "LG2"; break;
955 case WINED3DSIH_LOGP: instruction = "LOG"; break;
956 case WINED3DSIH_LRP: instruction = "LRP"; break;
957 case WINED3DSIH_MAD: instruction = "MAD"; break;
958 case WINED3DSIH_MAX: instruction = "MAX"; break;
959 case WINED3DSIH_MIN: instruction = "MIN"; break;
960 case WINED3DSIH_MOV: instruction = "MOV"; break;
961 case WINED3DSIH_MUL: instruction = "MUL"; break;
962 case WINED3DSIH_NOP: instruction = "NOP"; break;
963 case WINED3DSIH_POW: instruction = "POW"; break;
964 case WINED3DSIH_SGE: instruction = "SGE"; break;
965 case WINED3DSIH_SLT: instruction = "SLT"; break;
966 case WINED3DSIH_SUB: instruction = "SUB"; break;
967 default: instruction = "";
968 FIXME("Unhandled opcode %#x\n", ins->handler_idx);
969 break;
970 }
971
972 if (shader_is_pshader_version(ins->ctx->reg_maps->shader_version))
973 {
974 /* Output token related */
975 const struct wined3d_shader_dst_param *dst;
976 char output_rname[256];
977 char output_wmask[20];
978 char operands[4][100];
979 BOOL saturate = FALSE;
980 BOOL centroid = FALSE;
981 BOOL partialprecision = FALSE;
982 const char *modifier;
983 BOOL is_color;
984 DWORD shift;
985
986 if (!(ins->dst_count + ins->src_count))
987 {
988 ERR("Opcode \"%#x\" has no parameters\n", ins->handler_idx);
989 return;
990 }
991 dst = &ins->dst[0];
992
993 /* Process modifiers */
994 if (dst->modifiers)
995 {
996 DWORD mask = dst->modifiers;
997
998 saturate = mask & WINED3DSPDM_SATURATE;
999 centroid = mask & WINED3DSPDM_MSAMPCENTROID;
1000 partialprecision = mask & WINED3DSPDM_PARTIALPRECISION;
1001 mask &= ~(WINED3DSPDM_MSAMPCENTROID | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_SATURATE);
1002 if (mask)
1003 FIXME("Unrecognized modifier(%#x)\n", mask >> WINED3DSP_DSTMOD_SHIFT);
1004
1005 if (centroid)
1006 FIXME("Unhandled modifier(%#x)\n", mask >> WINED3DSP_DSTMOD_SHIFT);
1007 }
1008 shift = dst->shift;
1009 modifier = (saturate && !shift) ? "_SAT" : "";
1010
1011 /* Generate input register names (with modifiers) */
1012 for (i = 0; i < ins->src_count; ++i)
1013 {
1014 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[i].token, i, operands[i + 1]);
1015 }
1016
1017 /* Handle output register */
1018 shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
1019 dst->register_idx, dst->has_rel_addr, output_rname, &is_color);
1020 strcpy(operands[0], output_rname);
1021 shader_arb_get_write_mask(ins, dst, output_wmask);
1022 strcat(operands[0], output_wmask);
1023
1024 arguments[0] = '\0';
1025 strcat(arguments, operands[0]);
1026 for (i = 0; i < ins->src_count; ++i)
1027 {
1028 strcat(arguments, ", ");
1029 strcat(arguments, operands[i + 1]);
1030 }
1031 shader_addline(buffer, "%s%s %s;\n", instruction, modifier, arguments);
1032
1033 /* A shift requires another line. */
1034 if (shift) pshader_gen_output_modifier_line(buffer, saturate, output_wmask, shift, output_rname);
1035 } else {
1036 /* Note that shader_arb_add_*_param() adds spaces. */
1037
1038 arguments[0] = '\0';
1039 if (ins->dst_count)
1040 {
1041 shader_arb_add_dst_param(ins, &ins->dst[0], arguments);
1042 for (i = 0; i < ins->src_count; ++i)
1043 {
1044 strcat(arguments, ",");
1045 shader_arb_add_src_param(ins, ins->src[i].token, arguments);
1046 }
1047 }
1048 shader_addline(buffer, "%s%s;\n", instruction, arguments);
1049 }
1050 }
1051
1052 static void shader_hw_mov(const struct wined3d_shader_instruction *ins)
1053 {
1054 IWineD3DBaseShaderImpl *shader = (IWineD3DBaseShaderImpl *)ins->ctx->shader;
1055
1056 if ((WINED3DSHADER_VERSION_MAJOR(ins->ctx->reg_maps->shader_version) == 1
1057 && !shader_is_pshader_version(ins->ctx->reg_maps->shader_version)
1058 && ins->dst[0].register_type == WINED3DSPR_ADDR)
1059 || ins->handler_idx == WINED3DSIH_MOVA)
1060 {
1061 SHADER_BUFFER *buffer = ins->ctx->buffer;
1062 char src0_param[256];
1063
1064 if (ins->handler_idx == WINED3DSIH_MOVA)
1065 FIXME("mova should round\n");
1066
1067 src0_param[0] = '\0';
1068 if (((IWineD3DVertexShaderImpl *)shader)->rel_offset)
1069 {
1070 shader_arb_add_src_param(ins, ins->src[0].token, src0_param);
1071 shader_addline(buffer, "ADD TMP.x, %s, helper_const.z;\n", src0_param);
1072 shader_addline(buffer, "ARL A0.x, TMP.x;\n");
1073 }
1074 else
1075 {
1076 /* Apple's ARB_vertex_program implementation does not accept an ARL source argument
1077 * with more than one component. Thus replicate the first source argument over all
1078 * 4 components. For example, .xyzw -> .x (or better: .xxxx), .zwxy -> .z, etc)
1079 */
1080 DWORD parm = ins->src[0].token & ~(WINED3DVS_SWIZZLE_MASK);
1081 if ((ins->src[0].token & WINED3DVS_X_W) == WINED3DVS_X_W)
1082 parm |= WINED3DVS_X_W | WINED3DVS_Y_W | WINED3DVS_Z_W | WINED3DVS_W_W;
1083 else if ((ins->src[0].token & WINED3DVS_X_Z) == WINED3DVS_X_Z)
1084 parm |= WINED3DVS_X_Z | WINED3DVS_Y_Z | WINED3DVS_Z_Z | WINED3DVS_W_Z;
1085 else if ((ins->src[0].token & WINED3DVS_X_Y) == WINED3DVS_X_Y)
1086 parm |= WINED3DVS_X_Y | WINED3DVS_Y_Y | WINED3DVS_Z_Y | WINED3DVS_W_Y;
1087 else if ((ins->src[0].token & WINED3DVS_X_X) == WINED3DVS_X_X)
1088 parm |= WINED3DVS_X_X | WINED3DVS_Y_X | WINED3DVS_Z_X | WINED3DVS_W_X;
1089 shader_arb_add_src_param(ins, parm, src0_param);
1090 shader_addline(buffer, "ARL A0.x, %s;\n", src0_param);
1091 }
1092 }
1093 else
1094 {
1095 shader_hw_map2gl(ins);
1096 }
1097 }
1098
1099 static void pshader_hw_texkill(const struct wined3d_shader_instruction *ins)
1100 {
1101 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1102 DWORD shader_version = ins->ctx->reg_maps->shader_version;
1103 SHADER_BUFFER *buffer = ins->ctx->buffer;
1104 char reg_dest[40];
1105 BOOL is_color;
1106
1107 /* No swizzles are allowed in d3d's texkill. PS 1.x ignores the 4th component as documented,
1108 * but >= 2.0 honors it(undocumented, but tested by the d3d9 testsuit)
1109 */
1110 shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
1111 dst->register_idx, dst->has_rel_addr, reg_dest, &is_color);
1112
1113 if (shader_version >= WINED3DPS_VERSION(2,0))
1114 {
1115 /* The arb backend doesn't claim ps 2.0 support, but try to eat what the app feeds to us */
1116 shader_addline(buffer, "KIL %s;\n", reg_dest);
1117 } else {
1118 /* ARB fp doesn't like swizzles on the parameter of the KIL instruction. To mask the 4th component,
1119 * copy the register into our general purpose TMP variable, overwrite .w and pass TMP to KIL
1120 */
1121 shader_addline(buffer, "MOV TMP, %s;\n", reg_dest);
1122 shader_addline(buffer, "MOV TMP.w, one.w;\n");
1123 shader_addline(buffer, "KIL TMP;\n");
1124 }
1125 }
1126
1127 static void pshader_hw_tex(const struct wined3d_shader_instruction *ins)
1128 {
1129 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1130 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1131 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1132 BOOL is_color;
1133 SHADER_BUFFER *buffer = ins->ctx->buffer;
1134 DWORD shader_version = ins->ctx->reg_maps->shader_version;
1135 BOOL projected = FALSE, bias = FALSE;
1136
1137 char reg_dest[40];
1138 char reg_coord[40];
1139 DWORD reg_sampler_code;
1140
1141 /* All versions have a destination register */
1142 shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
1143 dst->register_idx, dst->has_rel_addr, reg_dest, &is_color);
1144
1145 /* 1.0-1.3: Use destination register as coordinate source.
1146 1.4+: Use provided coordinate source register. */
1147 if (shader_version < WINED3DPS_VERSION(1,4))
1148 strcpy(reg_coord, reg_dest);
1149 else
1150 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[0].token, 0, reg_coord);
1151
1152 /* 1.0-1.4: Use destination register number as texture code.
1153 2.0+: Use provided sampler number as texure code. */
1154 if (shader_version < WINED3DPS_VERSION(2,0))
1155 reg_sampler_code = dst->register_idx;
1156 else
1157 reg_sampler_code = ins->src[1].register_idx;
1158
1159 /* projection flag:
1160 * 1.1, 1.2, 1.3: Use WINED3DTSS_TEXTURETRANSFORMFLAGS
1161 * 1.4: Use WINED3DSPSM_DZ or WINED3DSPSM_DW on src[0]
1162 * 2.0+: Use WINED3DSI_TEXLD_PROJECT on the opcode
1163 */
1164 if (shader_version < WINED3DPS_VERSION(1,4))
1165 {
1166 DWORD flags = 0;
1167 if(reg_sampler_code < MAX_TEXTURES) {
1168 flags = deviceImpl->stateBlock->textureState[reg_sampler_code][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1169 }
1170 if (flags & WINED3DTTFF_PROJECTED) {
1171 projected = TRUE;
1172 }
1173 }
1174 else if (shader_version < WINED3DPS_VERSION(2,0))
1175 {
1176 DWORD src_mod = ins->src[0].modifiers;
1177 if (src_mod == WINED3DSPSM_DZ) {
1178 projected = TRUE;
1179 } else if(src_mod == WINED3DSPSM_DW) {
1180 projected = TRUE;
1181 }
1182 } else {
1183 if (ins->flags & WINED3DSI_TEXLD_PROJECT) projected = TRUE;
1184 if (ins->flags & WINED3DSI_TEXLD_BIAS) bias = TRUE;
1185 }
1186 shader_hw_sample(ins, reg_sampler_code, reg_dest, reg_coord, projected, bias);
1187 }
1188
1189 static void pshader_hw_texcoord(const struct wined3d_shader_instruction *ins)
1190 {
1191 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1192 SHADER_BUFFER *buffer = ins->ctx->buffer;
1193
1194 char tmp[20];
1195 shader_arb_get_write_mask(ins, dst, tmp);
1196 if (ins->ctx->reg_maps->shader_version != WINED3DPS_VERSION(1,4))
1197 {
1198 DWORD reg = dst->register_idx;
1199 shader_addline(buffer, "MOV_SAT T%u%s, fragment.texcoord[%u];\n", reg, tmp, reg);
1200 } else {
1201 char reg_src[40];
1202
1203 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[0].token, 0, reg_src);
1204 shader_addline(buffer, "MOV R%u%s, %s;\n", dst->register_idx, tmp, reg_src);
1205 }
1206 }
1207
1208 static void pshader_hw_texreg2ar(const struct wined3d_shader_instruction *ins)
1209 {
1210 SHADER_BUFFER *buffer = ins->ctx->buffer;
1211 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1212 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1213 DWORD flags;
1214
1215 DWORD reg1 = ins->dst[0].register_idx;
1216 char dst_str[8];
1217 char src_str[50];
1218
1219 sprintf(dst_str, "T%u", reg1);
1220 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[0].token, 0, src_str);
1221 shader_addline(buffer, "MOV TMP.x, %s.w;\n", src_str);
1222 shader_addline(buffer, "MOV TMP.y, %s.x;\n", src_str);
1223 flags = reg1 < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg1][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1224 shader_hw_sample(ins, reg1, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1225 }
1226
1227 static void pshader_hw_texreg2gb(const struct wined3d_shader_instruction *ins)
1228 {
1229 SHADER_BUFFER *buffer = ins->ctx->buffer;
1230
1231 DWORD reg1 = ins->dst[0].register_idx;
1232 char dst_str[8];
1233 char src_str[50];
1234
1235 sprintf(dst_str, "T%u", reg1);
1236 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[0].token, 0, src_str);
1237 shader_addline(buffer, "MOV TMP.x, %s.y;\n", src_str);
1238 shader_addline(buffer, "MOV TMP.y, %s.z;\n", src_str);
1239 shader_hw_sample(ins, reg1, dst_str, "TMP", FALSE, FALSE);
1240 }
1241
1242 static void pshader_hw_texreg2rgb(const struct wined3d_shader_instruction *ins)
1243 {
1244 SHADER_BUFFER *buffer = ins->ctx->buffer;
1245 DWORD reg1 = ins->dst[0].register_idx;
1246 char dst_str[8];
1247 char src_str[50];
1248
1249 sprintf(dst_str, "T%u", reg1);
1250 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[0].token, 0, src_str);
1251 shader_hw_sample(ins, reg1, dst_str, src_str, FALSE, FALSE);
1252 }
1253
1254 static void pshader_hw_texbem(const struct wined3d_shader_instruction *ins)
1255 {
1256 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1257 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1258 BOOL has_bumpmat = FALSE;
1259 BOOL has_luminance = FALSE;
1260 BOOL is_color;
1261 int i;
1262
1263 SHADER_BUFFER *buffer = ins->ctx->buffer;
1264
1265 char reg_coord[40];
1266 DWORD reg_dest_code;
1267
1268 /* All versions have a destination register */
1269 reg_dest_code = dst->register_idx;
1270 /* Can directly use the name because texbem is only valid for <= 1.3 shaders */
1271 shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
1272 dst->register_idx, dst->has_rel_addr, reg_coord, &is_color);
1273
1274 for(i = 0; i < This->numbumpenvmatconsts; i++) {
1275 if (This->bumpenvmatconst[i].const_num != WINED3D_CONST_NUM_UNUSED
1276 && reg_dest_code == This->bumpenvmatconst[i].texunit)
1277 {
1278 has_bumpmat = TRUE;
1279 break;
1280 }
1281 }
1282 for(i = 0; i < This->numbumpenvmatconsts; i++) {
1283 if (This->luminanceconst[i].const_num != WINED3D_CONST_NUM_UNUSED
1284 && reg_dest_code == This->luminanceconst[i].texunit)
1285 {
1286 has_luminance = TRUE;
1287 break;
1288 }
1289 }
1290
1291 if(has_bumpmat) {
1292 DWORD src = ins->src[0].register_idx;
1293
1294 /* Sampling the perturbation map in Tsrc was done already, including the signedness correction if needed */
1295
1296 shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, x, z, 0, 0;\n", reg_dest_code);
1297 shader_addline(buffer, "DP3 TMP.x, TMP2, T%u;\n", src);
1298 shader_addline(buffer, "SWZ TMP2, bumpenvmat%d, y, w, 0, 0;\n", reg_dest_code);
1299 shader_addline(buffer, "DP3 TMP.y, TMP2, T%u;\n", src);
1300
1301 /* with projective textures, texbem only divides the static texture coord, not the displacement,
1302 * so we can't let the GL handle this.
1303 */
1304 if (((IWineD3DDeviceImpl*) This->baseShader.device)->stateBlock->textureState[reg_dest_code][WINED3DTSS_TEXTURETRANSFORMFLAGS]
1305 & WINED3DTTFF_PROJECTED) {
1306 shader_addline(buffer, "RCP TMP2.w, %s.w;\n", reg_coord);
1307 shader_addline(buffer, "MUL TMP2.xy, %s, TMP2.w;\n", reg_coord);
1308 shader_addline(buffer, "ADD TMP.xy, TMP, TMP2;\n");
1309 } else {
1310 shader_addline(buffer, "ADD TMP.xy, TMP, %s;\n", reg_coord);
1311 }
1312
1313 shader_hw_sample(ins, reg_dest_code, reg_coord, "TMP", FALSE, FALSE);
1314
1315 if (ins->handler_idx == WINED3DSIH_TEXBEML && has_luminance)
1316 {
1317 shader_addline(buffer, "MAD TMP, T%u.z, luminance%d.x, luminance%d.y;\n",
1318 src, reg_dest_code, reg_dest_code);
1319 shader_addline(buffer, "MUL %s, %s, TMP;\n", reg_coord, reg_coord);
1320 }
1321
1322 } else {
1323 DWORD tf;
1324 if(reg_dest_code < MAX_TEXTURES) {
1325 tf = ((IWineD3DDeviceImpl*) This->baseShader.device)->stateBlock->textureState[reg_dest_code][WINED3DTSS_TEXTURETRANSFORMFLAGS];
1326 } else {
1327 tf = 0;
1328 }
1329 /* Without a bump matrix loaded, just sample with the unmodified coordinates */
1330 shader_hw_sample(ins, reg_dest_code, reg_coord, reg_coord, tf & WINED3DTTFF_PROJECTED, FALSE);
1331 }
1332 }
1333
1334 static void pshader_hw_texm3x2pad(const struct wined3d_shader_instruction *ins)
1335 {
1336 DWORD reg = ins->dst[0].register_idx;
1337 SHADER_BUFFER *buffer = ins->ctx->buffer;
1338 char src0_name[50];
1339
1340 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[0].token, 0, src0_name);
1341 shader_addline(buffer, "DP3 TMP.x, T%u, %s;\n", reg, src0_name);
1342 }
1343
1344 static void pshader_hw_texm3x2tex(const struct wined3d_shader_instruction *ins)
1345 {
1346 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1347 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1348 DWORD flags;
1349 DWORD reg = ins->dst[0].register_idx;
1350 SHADER_BUFFER *buffer = ins->ctx->buffer;
1351 char dst_str[8];
1352 char src0_name[50];
1353
1354 sprintf(dst_str, "T%u", reg);
1355 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[0].token, 0, src0_name);
1356 shader_addline(buffer, "DP3 TMP.y, T%u, %s;\n", reg, src0_name);
1357 flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1358 shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1359 }
1360
1361 static void pshader_hw_texm3x3pad(const struct wined3d_shader_instruction *ins)
1362 {
1363 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1364 DWORD reg = ins->dst[0].register_idx;
1365 SHADER_BUFFER *buffer = ins->ctx->buffer;
1366 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1367 char src0_name[50];
1368
1369 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[0].token, 0, src0_name);
1370 shader_addline(buffer, "DP3 TMP.%c, T%u, %s;\n", 'x' + current_state->current_row, reg, src0_name);
1371 current_state->texcoord_w[current_state->current_row++] = reg;
1372 }
1373
1374 static void pshader_hw_texm3x3tex(const struct wined3d_shader_instruction *ins)
1375 {
1376 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1377 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1378 DWORD flags;
1379 DWORD reg = ins->dst[0].register_idx;
1380 SHADER_BUFFER *buffer = ins->ctx->buffer;
1381 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1382 char dst_str[8];
1383 char src0_name[50];
1384
1385 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[0].token, 0, src0_name);
1386 shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
1387
1388 /* Sample the texture using the calculated coordinates */
1389 sprintf(dst_str, "T%u", reg);
1390 flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1391 shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1392 current_state->current_row = 0;
1393 }
1394
1395 static void pshader_hw_texm3x3vspec(const struct wined3d_shader_instruction *ins)
1396 {
1397 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1398 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1399 DWORD flags;
1400 DWORD reg = ins->dst[0].register_idx;
1401 SHADER_BUFFER *buffer = ins->ctx->buffer;
1402 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1403 char dst_str[8];
1404 char src0_name[50];
1405
1406 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[0].token, 0, src0_name);
1407 shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
1408
1409 /* Construct the eye-ray vector from w coordinates */
1410 shader_addline(buffer, "MOV TMP2.x, fragment.texcoord[%u].w;\n", current_state->texcoord_w[0]);
1411 shader_addline(buffer, "MOV TMP2.y, fragment.texcoord[%u].w;\n", current_state->texcoord_w[1]);
1412 shader_addline(buffer, "MOV TMP2.z, fragment.texcoord[%u].w;\n", reg);
1413
1414 /* Calculate reflection vector
1415 */
1416 shader_addline(buffer, "DP3 TMP.w, TMP, TMP2;\n");
1417 /* The .w is ignored when sampling, so I can use TMP2.w to calculate dot(N, N) */
1418 shader_addline(buffer, "DP3 TMP2.w, TMP, TMP;\n");
1419 shader_addline(buffer, "RCP TMP2.w, TMP2.w;\n");
1420 shader_addline(buffer, "MUL TMP.w, TMP.w, TMP2.w;\n");
1421 shader_addline(buffer, "MUL TMP, TMP.w, TMP;\n");
1422 shader_addline(buffer, "MAD TMP, coefmul.x, TMP, -TMP2;\n");
1423
1424 /* Sample the texture using the calculated coordinates */
1425 sprintf(dst_str, "T%u", reg);
1426 flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1427 shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1428 current_state->current_row = 0;
1429 }
1430
1431 static void pshader_hw_texm3x3spec(const struct wined3d_shader_instruction *ins)
1432 {
1433 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)ins->ctx->shader;
1434 IWineD3DDeviceImpl* deviceImpl = (IWineD3DDeviceImpl*) This->baseShader.device;
1435 DWORD flags;
1436 DWORD reg = ins->dst[0].register_idx;
1437 DWORD reg3 = ins->src[1].register_idx;
1438 SHADER_PARSE_STATE* current_state = &This->baseShader.parse_state;
1439 SHADER_BUFFER *buffer = ins->ctx->buffer;
1440 char dst_str[8];
1441 char src0_name[50];
1442
1443 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[0].token, 0, src0_name);
1444 shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", reg, src0_name);
1445
1446 /* Calculate reflection vector.
1447 *
1448 * dot(N, E)
1449 * TMP.xyz = 2 * --------- * N - E
1450 * dot(N, N)
1451 *
1452 * Which normalizes the normal vector
1453 */
1454 shader_addline(buffer, "DP3 TMP.w, TMP, C[%u];\n", reg3);
1455 shader_addline(buffer, "DP3 TMP2.w, TMP, TMP;\n");
1456 shader_addline(buffer, "RCP TMP2.w, TMP2.w;\n");
1457 shader_addline(buffer, "MUL TMP.w, TMP.w, TMP2.w;\n");
1458 shader_addline(buffer, "MUL TMP, TMP.w, TMP;\n");
1459 shader_addline(buffer, "MAD TMP, coefmul.x, TMP, -C[%u];\n", reg3);
1460
1461 /* Sample the texture using the calculated coordinates */
1462 sprintf(dst_str, "T%u", reg);
1463 flags = reg < MAX_TEXTURES ? deviceImpl->stateBlock->textureState[reg][WINED3DTSS_TEXTURETRANSFORMFLAGS] : 0;
1464 shader_hw_sample(ins, reg, dst_str, "TMP", flags & WINED3DTTFF_PROJECTED, FALSE);
1465 current_state->current_row = 0;
1466 }
1467
1468 static void pshader_hw_texdepth(const struct wined3d_shader_instruction *ins)
1469 {
1470 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1471 SHADER_BUFFER *buffer = ins->ctx->buffer;
1472 char dst_name[50];
1473 BOOL is_color;
1474
1475 /* texdepth has an implicit destination, the fragment depth value. It's only parameter,
1476 * which is essentially an input, is the destination register because it is the first
1477 * parameter. According to the msdn, this must be register r5, but let's keep it more flexible
1478 * here
1479 */
1480 shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
1481 dst->register_idx, dst->has_rel_addr, dst_name, &is_color);
1482
1483 /* According to the msdn, the source register(must be r5) is unusable after
1484 * the texdepth instruction, so we're free to modify it
1485 */
1486 shader_addline(buffer, "MIN %s.y, %s.y, one.y;\n", dst_name, dst_name);
1487
1488 /* How to deal with the special case dst_name.g == 0? if r != 0, then
1489 * the r * (1 / 0) will give infinity, which is clamped to 1.0, the correct
1490 * result. But if r = 0.0, then 0 * inf = 0, which is incorrect.
1491 */
1492 shader_addline(buffer, "RCP %s.y, %s.y;\n", dst_name, dst_name);
1493 shader_addline(buffer, "MUL TMP.x, %s.x, %s.y;\n", dst_name, dst_name);
1494 shader_addline(buffer, "MIN TMP.x, TMP.x, one.x;\n");
1495 shader_addline(buffer, "MAX result.depth, TMP.x, 0.0;\n");
1496 }
1497
1498 /** Process the WINED3DSIO_TEXDP3TEX instruction in ARB:
1499 * Take a 3-component dot product of the TexCoord[dstreg] and src,
1500 * then perform a 1D texture lookup from stage dstregnum, place into dst. */
1501 static void pshader_hw_texdp3tex(const struct wined3d_shader_instruction *ins)
1502 {
1503 SHADER_BUFFER *buffer = ins->ctx->buffer;
1504 DWORD sampler_idx = ins->dst[0].register_idx;
1505 char src0[50];
1506 char dst_str[8];
1507
1508 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[0].token, 0, src0);
1509 shader_addline(buffer, "MOV TMP, 0.0;\n");
1510 shader_addline(buffer, "DP3 TMP.x, T%u, %s;\n", sampler_idx, src0);
1511
1512 sprintf(dst_str, "T%u", sampler_idx);
1513 shader_hw_sample(ins, sampler_idx, dst_str, "TMP", FALSE /* Only one coord, can't be projected */, FALSE);
1514 }
1515
1516 /** Process the WINED3DSIO_TEXDP3 instruction in ARB:
1517 * Take a 3-component dot product of the TexCoord[dstreg] and src. */
1518 static void pshader_hw_texdp3(const struct wined3d_shader_instruction *ins)
1519 {
1520 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1521 char src0[50];
1522 char dst_str[50];
1523 char dst_mask[6];
1524 SHADER_BUFFER *buffer = ins->ctx->buffer;
1525 BOOL is_color;
1526
1527 /* Handle output register */
1528 shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
1529 dst->register_idx, dst->has_rel_addr, dst_str, &is_color);
1530 shader_arb_get_write_mask(ins, dst, dst_mask);
1531
1532 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[0].token, 0, src0);
1533 shader_addline(buffer, "DP3 %s%s, T%u, %s;\n", dst_str, dst_mask, dst->register_idx, src0);
1534
1535 /* TODO: Handle output modifiers */
1536 }
1537
1538 /** Process the WINED3DSIO_TEXM3X3 instruction in ARB
1539 * Perform the 3rd row of a 3x3 matrix multiply */
1540 static void pshader_hw_texm3x3(const struct wined3d_shader_instruction *ins)
1541 {
1542 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1543 SHADER_BUFFER *buffer = ins->ctx->buffer;
1544 char dst_str[50];
1545 char dst_mask[6];
1546 char src0[50];
1547 BOOL is_color;
1548
1549 shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
1550 dst->register_idx, dst->has_rel_addr, dst_str, &is_color);
1551 shader_arb_get_write_mask(ins, dst, dst_mask);
1552
1553 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[0].token, 0, src0);
1554 shader_addline(buffer, "DP3 TMP.z, T%u, %s;\n", dst->register_idx, src0);
1555 shader_addline(buffer, "MOV %s%s, TMP;\n", dst_str, dst_mask);
1556
1557 /* TODO: Handle output modifiers */
1558 }
1559
1560 /** Process the WINED3DSIO_TEXM3X2DEPTH instruction in ARB:
1561 * Last row of a 3x2 matrix multiply, use the result to calculate the depth:
1562 * Calculate tmp0.y = TexCoord[dstreg] . src.xyz; (tmp0.x has already been calculated)
1563 * depth = (tmp0.y == 0.0) ? 1.0 : tmp0.x / tmp0.y
1564 */
1565 static void pshader_hw_texm3x2depth(const struct wined3d_shader_instruction *ins)
1566 {
1567 SHADER_BUFFER *buffer = ins->ctx->buffer;
1568 DWORD dst_reg = ins->dst[0].register_idx;
1569 char src0[50];
1570
1571 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[0].token, 0, src0);
1572 shader_addline(buffer, "DP3 TMP.y, T%u, %s;\n", dst_reg, src0);
1573
1574 /* How to deal with the special case dst_name.g == 0? if r != 0, then
1575 * the r * (1 / 0) will give infinity, which is clamped to 1.0, the correct
1576 * result. But if r = 0.0, then 0 * inf = 0, which is incorrect.
1577 */
1578 shader_addline(buffer, "RCP TMP.y, TMP.y;\n");
1579 shader_addline(buffer, "MUL TMP.x, TMP.x, TMP.y;\n");
1580 shader_addline(buffer, "MIN TMP.x, TMP.x, one.x;\n");
1581 shader_addline(buffer, "MAX result.depth, TMP.x, 0.0;\n");
1582 }
1583
1584 /** Handles transforming all WINED3DSIO_M?x? opcodes for
1585 Vertex/Pixel shaders to ARB_vertex_program codes */
1586 static void shader_hw_mnxn(const struct wined3d_shader_instruction *ins)
1587 {
1588 int i;
1589 int nComponents = 0;
1590 struct wined3d_shader_dst_param tmp_dst = {0};
1591 struct wined3d_shader_src_param tmp_src[2] = {{0}};
1592 struct wined3d_shader_instruction tmp_ins;
1593
1594 memset(&tmp_ins, 0, sizeof(tmp_ins));
1595
1596 /* Set constants for the temporary argument */
1597 tmp_ins.ctx = ins->ctx;
1598 tmp_ins.dst_count = 1;
1599 tmp_ins.dst = &tmp_dst;
1600 tmp_ins.src_count = 2;
1601 tmp_ins.src = tmp_src;
1602
1603 switch(ins->handler_idx)
1604 {
1605 case WINED3DSIH_M4x4:
1606 nComponents = 4;
1607 tmp_ins.handler_idx = WINED3DSIH_DP4;
1608 break;
1609 case WINED3DSIH_M4x3:
1610 nComponents = 3;
1611 tmp_ins.handler_idx = WINED3DSIH_DP4;
1612 break;
1613 case WINED3DSIH_M3x4:
1614 nComponents = 4;
1615 tmp_ins.handler_idx = WINED3DSIH_DP3;
1616 break;
1617 case WINED3DSIH_M3x3:
1618 nComponents = 3;
1619 tmp_ins.handler_idx = WINED3DSIH_DP3;
1620 break;
1621 case WINED3DSIH_M3x2:
1622 nComponents = 2;
1623 tmp_ins.handler_idx = WINED3DSIH_DP3;
1624 break;
1625 default:
1626 FIXME("Unhandled opcode %#x\n", ins->handler_idx);
1627 break;
1628 }
1629
1630 tmp_dst = ins->dst[0];
1631 tmp_src[0] = ins->src[0];
1632 tmp_src[1] = ins->src[1];
1633 for (i = 0; i < nComponents; i++) {
1634 tmp_dst.write_mask = WINED3DSP_WRITEMASK_0 << i;
1635 shader_hw_map2gl(&tmp_ins);
1636 ++tmp_src[1].register_idx;
1637 ++tmp_src[1].token;
1638 }
1639 }
1640
1641 static void vshader_hw_rsq_rcp(const struct wined3d_shader_instruction *ins)
1642 {
1643 SHADER_BUFFER *buffer = ins->ctx->buffer;
1644 DWORD swizzle = (ins->src[0].token & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
1645 const char *instruction;
1646
1647 char tmpLine[256];
1648
1649 switch(ins->handler_idx)
1650 {
1651 case WINED3DSIH_RSQ: instruction = "RSQ"; break;
1652 case WINED3DSIH_RCP: instruction = "RCP"; break;
1653 default: instruction = "";
1654 FIXME("Unhandled opcode %#x\n", ins->handler_idx);
1655 break;
1656 }
1657
1658 strcpy(tmpLine, instruction);
1659 shader_arb_add_dst_param(ins, &ins->dst[0], tmpLine); /* Destination */
1660 strcat(tmpLine, ",");
1661 shader_arb_add_src_param(ins, ins->src[0].token, tmpLine);
1662 if ((WINED3DSP_NOSWIZZLE >> WINED3DSP_SWIZZLE_SHIFT) == swizzle) {
1663 /* Dx sdk says .x is used if no swizzle is given, but our test shows that
1664 * .w is used
1665 */
1666 strcat(tmpLine, ".w");
1667 }
1668
1669 shader_addline(buffer, "%s;\n", tmpLine);
1670 }
1671
1672 static void shader_hw_nrm(const struct wined3d_shader_instruction *ins)
1673 {
1674 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1675 SHADER_BUFFER *buffer = ins->ctx->buffer;
1676 char dst_name[50];
1677 char src_name[50];
1678 char dst_wmask[20];
1679 DWORD shift = dst->shift;
1680 BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
1681 BOOL is_color;
1682
1683 shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
1684 dst->register_idx, dst->has_rel_addr, dst_name, &is_color);
1685 shader_arb_get_write_mask(ins, dst, dst_wmask);
1686
1687 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[0].token, 0, src_name);
1688 shader_addline(buffer, "DP3 TMP, %s, %s;\n", src_name, src_name);
1689 shader_addline(buffer, "RSQ TMP, TMP.x;\n");
1690 /* dst.w = src[0].w * 1 / (src.x^2 + src.y^2 + src.z^2)^(1/2) according to msdn*/
1691 shader_addline(buffer, "MUL%s %s%s, %s, TMP;\n", sat ? "_SAT" : "", dst_name, dst_wmask,
1692 src_name);
1693
1694 if (shift != 0)
1695 pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
1696 }
1697
1698 static void shader_hw_sincos(const struct wined3d_shader_instruction *ins)
1699 {
1700 /* This instruction exists in ARB, but the d3d instruction takes two extra parameters which
1701 * must contain fixed constants. So we need a separate function to filter those constants and
1702 * can't use map2gl
1703 */
1704 const struct wined3d_shader_dst_param *dst = &ins->dst[0];
1705 SHADER_BUFFER *buffer = ins->ctx->buffer;
1706 char dst_name[50];
1707 char src_name[50];
1708 char dst_wmask[20];
1709 DWORD shift = dst->shift;
1710 BOOL sat = dst->modifiers & WINED3DSPDM_SATURATE;
1711 BOOL is_color;
1712
1713 shader_arb_get_register_name(ins->ctx->shader, dst->register_type,
1714 dst->register_idx, dst->has_rel_addr, dst_name, &is_color);
1715 shader_arb_get_write_mask(ins, dst, dst_wmask);
1716
1717 pshader_gen_input_modifier_line(ins->ctx->shader, buffer, ins->src[0].token, 0, src_name);
1718 shader_addline(buffer, "SCS%s %s%s, %s;\n", sat ? "_SAT" : "", dst_name, dst_wmask,
1719 src_name);
1720
1721 if (shift != 0)
1722 pshader_gen_output_modifier_line(buffer, FALSE, dst_wmask, shift, dst_name);
1723
1724 }
1725
1726 static GLuint create_arb_blt_vertex_program(const WineD3D_GL_Info *gl_info)
1727 {
1728 GLuint program_id = 0;
1729 const char *blt_vprogram =
1730 "!!ARBvp1.0\n"
1731 "PARAM c[1] = { { 1, 0.5 } };\n"
1732 "MOV result.position, vertex.position;\n"
1733 "MOV result.color, c[0].x;\n"
1734 "MOV result.texcoord[0], vertex.texcoord[0];\n"
1735 "END\n";
1736
1737 GL_EXTCALL(glGenProgramsARB(1, &program_id));
1738 GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, program_id));
1739 GL_EXTCALL(glProgramStringARB(GL_VERTEX_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(blt_vprogram), blt_vprogram));
1740
1741 if (glGetError() == GL_INVALID_OPERATION) {
1742 GLint pos;
1743 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
1744 FIXME("Vertex program error at position %d: %s\n", pos,
1745 debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
1746 }
1747
1748 return program_id;
1749 }
1750
1751 static GLuint create_arb_blt_fragment_program(const WineD3D_GL_Info *gl_info, enum tex_types tex_type)
1752 {
1753 GLuint program_id = 0;
1754 static const char * const blt_fprograms[tex_type_count] =
1755 {
1756 /* tex_1d */
1757 NULL,
1758 /* tex_2d */
1759 "!!ARBfp1.0\n"
1760 "TEMP R0;\n"
1761 "TEX R0.x, fragment.texcoord[0], texture[0], 2D;\n"
1762 "MOV result.depth.z, R0.x;\n"
1763 "END\n",
1764 /* tex_3d */
1765 NULL,
1766 /* tex_cube */
1767 "!!ARBfp1.0\n"
1768 "TEMP R0;\n"
1769 "TEX R0.x, fragment.texcoord[0], texture[0], CUBE;\n"
1770 "MOV result.depth.z, R0.x;\n"
1771 "END\n",
1772 /* tex_rect */
1773 "!!ARBfp1.0\n"
1774 "TEMP R0;\n"
1775 "TEX R0.x, fragment.texcoord[0], texture[0], RECT;\n"
1776 "MOV result.depth.z, R0.x;\n"
1777 "END\n",
1778 };
1779
1780 if (!blt_fprograms[tex_type])
1781 {
1782 FIXME("tex_type %#x not supported\n", tex_type);
1783 tex_type = tex_2d;
1784 }
1785
1786 GL_EXTCALL(glGenProgramsARB(1, &program_id));
1787 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, program_id));
1788 GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(blt_fprograms[tex_type]), blt_fprograms[tex_type]));
1789
1790 if (glGetError() == GL_INVALID_OPERATION) {
1791 GLint pos;
1792 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
1793 FIXME("Fragment program error at position %d: %s\n", pos,
1794 debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
1795 }
1796
1797 return program_id;
1798 }
1799
1800 static void shader_arb_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {
1801 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1802 struct shader_arb_priv *priv = This->shader_priv;
1803 const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1804
1805 if (useVS) {
1806 struct vs_compile_args compile_args;
1807
1808 TRACE("Using vertex shader\n");
1809 find_vs_compile_args((IWineD3DVertexShaderImpl *) This->stateBlock->vertexShader, This->stateBlock, &compile_args);
1810 priv->current_vprogram_id = find_gl_vshader((IWineD3DVertexShaderImpl *) This->stateBlock->vertexShader, &compile_args);
1811
1812 /* Bind the vertex program */
1813 GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id));
1814 checkGLcall("glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id);");
1815
1816 /* Enable OpenGL vertex programs */
1817 glEnable(GL_VERTEX_PROGRAM_ARB);
1818 checkGLcall("glEnable(GL_VERTEX_PROGRAM_ARB);");
1819 TRACE("(%p) : Bound vertex program %u and enabled GL_VERTEX_PROGRAM_ARB\n", This, priv->current_vprogram_id);
1820 } else if(GL_SUPPORT(ARB_VERTEX_PROGRAM)) {
1821 priv->current_vprogram_id = 0;
1822 glDisable(GL_VERTEX_PROGRAM_ARB);
1823 checkGLcall("glDisable(GL_VERTEX_PROGRAM_ARB)");
1824 }
1825
1826 if (usePS) {
1827 struct ps_compile_args compile_args;
1828 TRACE("Using pixel shader\n");
1829 find_ps_compile_args((IWineD3DPixelShaderImpl *) This->stateBlock->pixelShader, This->stateBlock, &compile_args);
1830 priv->current_fprogram_id = find_gl_pshader((IWineD3DPixelShaderImpl *) This->stateBlock->pixelShader,
1831 &compile_args);
1832
1833 /* Bind the fragment program */
1834 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id));
1835 checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id);");
1836
1837 if(!priv->use_arbfp_fixed_func) {
1838 /* Enable OpenGL fragment programs */
1839 glEnable(GL_FRAGMENT_PROGRAM_ARB);
1840 checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB);");
1841 }
1842 TRACE("(%p) : Bound fragment program %u and enabled GL_FRAGMENT_PROGRAM_ARB\n", This, priv->current_fprogram_id);
1843 } else if(GL_SUPPORT(ARB_FRAGMENT_PROGRAM) && !priv->use_arbfp_fixed_func) {
1844 /* Disable only if we're not using arbfp fixed function fragment processing. If this is used,
1845 * keep GL_FRAGMENT_PROGRAM_ARB enabled, and the fixed function pipeline will bind the fixed function
1846 * replacement shader
1847 */
1848 glDisable(GL_FRAGMENT_PROGRAM_ARB);
1849 checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
1850 priv->current_fprogram_id = 0;
1851 }
1852 }
1853
1854 static void shader_arb_select_depth_blt(IWineD3DDevice *iface, enum tex_types tex_type) {
1855 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1856 struct shader_arb_priv *priv = This->shader_priv;
1857 GLuint *blt_fprogram = &priv->depth_blt_fprogram_id[tex_type];
1858 const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1859
1860 if (!priv->depth_blt_vprogram_id) priv->depth_blt_vprogram_id = create_arb_blt_vertex_program(gl_info);
1861 GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->depth_blt_vprogram_id));
1862 glEnable(GL_VERTEX_PROGRAM_ARB);
1863
1864 if (!*blt_fprogram) *blt_fprogram = create_arb_blt_fragment_program(gl_info, tex_type);
1865 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, *blt_fprogram));
1866 glEnable(GL_FRAGMENT_PROGRAM_ARB);
1867 }
1868
1869 static void shader_arb_deselect_depth_blt(IWineD3DDevice *iface) {
1870 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1871 struct shader_arb_priv *priv = This->shader_priv;
1872 const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1873
1874 if (priv->current_vprogram_id) {
1875 GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, priv->current_vprogram_id));
1876 checkGLcall("glBindProgramARB(GL_VERTEX_PROGRAM_ARB, vertexShader->prgId);");
1877
1878 glEnable(GL_VERTEX_PROGRAM_ARB);
1879 checkGLcall("glEnable(GL_VERTEX_PROGRAM_ARB);");
1880
1881 TRACE("(%p) : Bound vertex program %u and enabled GL_VERTEX_PROGRAM_ARB\n", This, priv->current_vprogram_id);
1882 } else {
1883 glDisable(GL_VERTEX_PROGRAM_ARB);
1884 checkGLcall("glDisable(GL_VERTEX_PROGRAM_ARB)");
1885 }
1886
1887 if (priv->current_fprogram_id) {
1888 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, priv->current_fprogram_id));
1889 checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, pixelShader->prgId);");
1890
1891 glEnable(GL_FRAGMENT_PROGRAM_ARB);
1892 checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB);");
1893
1894 TRACE("(%p) : Bound fragment program %u and enabled GL_FRAGMENT_PROGRAM_ARB\n", This, priv->current_fprogram_id);
1895 } else {
1896 glDisable(GL_FRAGMENT_PROGRAM_ARB);
1897 checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
1898 }
1899 }
1900
1901 static void shader_arb_destroy(IWineD3DBaseShader *iface) {
1902 IWineD3DBaseShaderImpl *baseShader = (IWineD3DBaseShaderImpl *) iface;
1903 const WineD3D_GL_Info *gl_info = &((IWineD3DDeviceImpl *)baseShader->baseShader.device)->adapter->gl_info;
1904
1905 if (shader_is_pshader_version(baseShader->baseShader.reg_maps.shader_version))
1906 {
1907 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *) iface;
1908 UINT i;
1909
1910 ENTER_GL();
1911 for(i = 0; i < This->num_gl_shaders; i++) {
1912 GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId));
1913 checkGLcall("GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId))");
1914 }
1915 LEAVE_GL();
1916 HeapFree(GetProcessHeap(), 0, This->gl_shaders);
1917 This->gl_shaders = NULL;
1918 This->num_gl_shaders = 0;
1919 This->shader_array_size = 0;
1920 } else {
1921 IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *) iface;
1922 UINT i;
1923
1924 ENTER_GL();
1925 for(i = 0; i < This->num_gl_shaders; i++) {
1926 GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId));
1927 checkGLcall("GL_EXTCALL(glDeleteProgramsARB(1, &This->gl_shaders[i].prgId))");
1928 }
1929 LEAVE_GL();
1930 HeapFree(GetProcessHeap(), 0, This->gl_shaders);
1931 This->gl_shaders = NULL;
1932 This->num_gl_shaders = 0;
1933 This->shader_array_size = 0;
1934 }
1935 }
1936
1937 static HRESULT shader_arb_alloc(IWineD3DDevice *iface) {
1938 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1939 This->shader_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct shader_arb_priv));
1940 return WINED3D_OK;
1941 }
1942
1943 static void shader_arb_free(IWineD3DDevice *iface) {
1944 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *)iface;
1945 const WineD3D_GL_Info *gl_info = &This->adapter->gl_info;
1946 struct shader_arb_priv *priv = This->shader_priv;
1947 int i;
1948
1949 if(priv->depth_blt_vprogram_id) {
1950 GL_EXTCALL(glDeleteProgramsARB(1, &priv->depth_blt_vprogram_id));
1951 }
1952 for (i = 0; i < tex_type_count; ++i) {
1953 if (priv->depth_blt_fprogram_id[i]) {
1954 GL_EXTCALL(glDeleteProgramsARB(1, &priv->depth_blt_fprogram_id[i]));
1955 }
1956 }
1957
1958 HeapFree(GetProcessHeap(), 0, This->shader_priv);
1959 }
1960
1961 static BOOL shader_arb_dirty_const(IWineD3DDevice *iface) {
1962 return TRUE;
1963 }
1964
1965 static void arbfp_add_sRGB_correction(SHADER_BUFFER *buffer, const char *fragcolor, const char *tmp1,
1966 const char *tmp2, const char *tmp3, const char *tmp4) {
1967 /* Perform sRGB write correction. See GLX_EXT_framebuffer_sRGB */
1968
1969 /* Calculate the > 0.0031308 case */
1970 shader_addline(buffer, "POW %s.x, %s.x, srgb_pow.x;\n", tmp1, fragcolor);
1971 shader_addline(buffer, "POW %s.y, %s.y, srgb_pow.y;\n", tmp1, fragcolor);
1972 shader_addline(buffer, "POW %s.z, %s.z, srgb_pow.z;\n", tmp1, fragcolor);
1973 shader_addline(buffer, "MUL %s, %s, srgb_mul_hi;\n", tmp1, tmp1);
1974 shader_addline(buffer, "SUB %s, %s, srgb_sub_hi;\n", tmp1, tmp1);
1975 /* Calculate the < case */
1976 shader_addline(buffer, "MUL %s, srgb_mul_low, %s;\n", tmp2, fragcolor);
1977 /* Get 1.0 / 0.0 masks for > 0.0031308 and < 0.0031308 */
1978 shader_addline(buffer, "SLT %s, srgb_comparison, %s;\n", tmp3, fragcolor);
1979 shader_addline(buffer, "SGE %s, srgb_comparison, %s;\n", tmp4, fragcolor);
1980 /* Store the components > 0.0031308 in the destination */
1981 shader_addline(buffer, "MUL %s, %s, %s;\n", fragcolor, tmp1, tmp3);
1982 /* Add the components that are < 0.0031308 */
1983 shader_addline(buffer, "MAD result.color.xyz, %s, %s, %s;\n", tmp2, tmp4, fragcolor);
1984 /* [0.0;1.0] clamping. Not needed, this is done implicitly */
1985 }
1986
1987 static GLuint shader_arb_generate_pshader(IWineD3DPixelShader *iface, SHADER_BUFFER *buffer, const struct ps_compile_args *args) {
1988 IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
1989 const shader_reg_maps* reg_maps = &This->baseShader.reg_maps;
1990 CONST DWORD *function = This->baseShader.function;
1991 DWORD shader_version = reg_maps->shader_version;
1992 const WineD3D_GL_Info *gl_info = &((IWineD3DDeviceImpl *)This->baseShader.device)->adapter->gl_info;
1993 const local_constant *lconst;
1994 GLuint retval;
1995 const char *fragcolor;
1996
1997 /* Create the hw ARB shader */
1998 shader_addline(buffer, "!!ARBfp1.0\n");
1999
2000 if (shader_version < WINED3DPS_VERSION(3,0)) {
2001 switch(args->fog) {
2002 case FOG_OFF:
2003 break;
2004 case FOG_LINEAR:
2005 shader_addline(buffer, "OPTION ARB_fog_linear;\n");
2006 break;
2007 case FOG_EXP:
2008 shader_addline(buffer, "OPTION ARB_fog_exp;\n");
2009 break;
2010 case FOG_EXP2:
2011 shader_addline(buffer, "OPTION ARB_fog_exp2;\n");
2012 break;
2013 }
2014 }
2015
2016 shader_addline(buffer, "TEMP TMP;\n"); /* Used in matrix ops */
2017 shader_addline(buffer, "TEMP TMP2;\n"); /* Used in matrix ops */
2018 shader_addline(buffer, "TEMP TA;\n"); /* Used for modifiers */
2019 shader_addline(buffer, "TEMP TB;\n"); /* Used for modifiers */
2020 shader_addline(buffer, "TEMP TC;\n"); /* Used for modifiers */
2021 shader_addline(buffer, "PARAM coefdiv = { 0.5, 0.25, 0.125, 0.0625 };\n");
2022 shader_addline(buffer, "PARAM coefmul = { 2, 4, 8, 16 };\n");
2023 shader_addline(buffer, "PARAM one = { 1.0, 1.0, 1.0, 1.0 };\n");
2024
2025 if (shader_version < WINED3DPS_VERSION(2,0)) {
2026 fragcolor = "R0";
2027 } else {
2028 shader_addline(buffer, "TEMP TMP_COLOR;\n");
2029 fragcolor = "TMP_COLOR";
2030 }
2031
2032 /* Base Declarations */
2033 shader_generate_arb_declarations( (IWineD3DBaseShader*) This, reg_maps, buffer, &GLINFO_LOCATION);
2034
2035 /* Base Shader Body */
2036 shader_generate_main( (IWineD3DBaseShader*) This, buffer, reg_maps, function);
2037
2038 if(args->srgb_correction) {
2039 arbfp_add_sRGB_correction(buffer, fragcolor, "TMP", "TMP2", "TA", "TB");
2040 }
2041 shader_addline(buffer, "MOV result.color, %s;\n", fragcolor);
2042 shader_addline(buffer, "END\n");
2043
2044 /* TODO: change to resource.glObjectHandle or something like that */
2045 GL_EXTCALL(glGenProgramsARB(1, &retval));
2046
2047 TRACE("Creating a hw pixel shader, prg=%d\n", retval);
2048 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, retval));
2049
2050 TRACE("Created hw pixel shader, prg=%d\n", retval);
2051 /* Create the program and check for errors */
2052 GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB,
2053 buffer->bsize, buffer->buffer));
2054
2055 if (glGetError() == GL_INVALID_OPERATION) {
2056 GLint errPos;
2057 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &errPos);
2058 FIXME("HW PixelShader Error at position %d: %s\n",
2059 errPos, debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
2060 retval = 0;
2061 }
2062
2063 /* Load immediate constants */
2064 if(!This->baseShader.load_local_constsF) {
2065 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
2066 const float *value = (const float *)lconst->value;
2067 GL_EXTCALL(glProgramLocalParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, lconst->idx, value));
2068 checkGLcall("glProgramLocalParameter4fvARB");
2069 }
2070 }
2071
2072 return retval;
2073 }
2074
2075 static GLuint shader_arb_generate_vshader(IWineD3DVertexShader *iface, SHADER_BUFFER *buffer, const struct vs_compile_args *args) {
2076 IWineD3DVertexShaderImpl *This = (IWineD3DVertexShaderImpl *)iface;
2077 const shader_reg_maps *reg_maps = &This->baseShader.reg_maps;
2078 CONST DWORD *function = This->baseShader.function;
2079 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *)This->baseShader.device;
2080 const WineD3D_GL_Info *gl_info = &device->adapter->gl_info;
2081 const local_constant *lconst;
2082 GLuint ret;
2083
2084 /* Create the hw ARB shader */
2085 shader_addline(buffer, "!!ARBvp1.0\n");
2086 shader_addline(buffer, "PARAM helper_const = { 2.0, -1.0, %d.0, 0.0 };\n", This->rel_offset);
2087
2088 /* Mesa supports only 95 constants */
2089 if (GL_VEND(MESA) || GL_VEND(WINE))
2090 This->baseShader.limits.constant_float =
2091 min(95, This->baseShader.limits.constant_float);
2092
2093 shader_addline(buffer, "TEMP TMP;\n");
2094
2095 /* Base Declarations */
2096 shader_generate_arb_declarations( (IWineD3DBaseShader*) This, reg_maps, buffer, &GLINFO_LOCATION);
2097
2098 /* We need a constant to fixup the final position */
2099 shader_addline(buffer, "PARAM posFixup = program.env[%d];\n", ARB_SHADER_PRIVCONST_POS);
2100
2101 /* Initialize output parameters. GL_ARB_vertex_program does not require special initialization values
2102 * for output parameters. D3D in theory does not do that either, but some applications depend on a
2103 * proper initialization of the secondary color, and programs using the fixed function pipeline without
2104 * a replacement shader depend on the texcoord.w being set properly.
2105 *
2106 * GL_NV_vertex_program defines that all output values are initialized to {0.0, 0.0, 0.0, 1.0}. This
2107 * assertion is in effect even when using GL_ARB_vertex_program without any NV specific additions. So
2108 * skip this if NV_vertex_program is supported. Otherwise, initialize the secondary color. For the tex-
2109 * coords, we have a flag in the opengl caps. Many cards do not require the texcoord being set, and
2110 * this can eat a number of instructions, so skip it unless this cap is set as well
2111 */
2112 if(!GL_SUPPORT(NV_VERTEX_PROGRAM)) {
2113 shader_addline(buffer, "MOV result.color.secondary, -helper_const.wwwy;\n");
2114
2115 if((GLINFO_LOCATION).set_texcoord_w && !device->frag_pipe->ffp_proj_control) {
2116 int i;
2117 for(i = 0; i < min(8, MAX_REG_TEXCRD); i++) {
2118 if(This->baseShader.reg_maps.texcoord_mask[i] != 0 &&
2119 This->baseShader.reg_maps.texcoord_mask[i] != WINED3DSP_WRITEMASK_ALL) {
2120 shader_addline(buffer, "MOV result.texcoord[%u].w, -helper_const.y;\n", i);
2121 }
2122 }
2123 }
2124 }
2125
2126 /* Base Shader Body */
2127 shader_generate_main( (IWineD3DBaseShader*) This, buffer, reg_maps, function);
2128
2129 /* The D3DRS_FOGTABLEMODE render state defines if the shader-generated fog coord is used
2130 * or if the fragment depth is used. If the fragment depth is used(FOGTABLEMODE != NONE),
2131 * the fog frag coord is thrown away. If the fog frag coord is used, but not written by
2132 * the shader, it is set to 0.0(fully fogged, since start = 1.0, end = 0.0)
2133 */
2134 if(args->fog_src == VS_FOG_Z) {
2135 shader_addline(buffer, "MOV result.fogcoord, TMP_OUT.z;\n");
2136 } else if (!reg_maps->fog) {
2137 shader_addline(buffer, "MOV result.fogcoord, 0.0;\n");
2138 }
2139
2140 /* Write the final position.
2141 *
2142 * OpenGL coordinates specify the center of the pixel while d3d coords specify
2143 * the corner. The offsets are stored in z and w in posFixup. posFixup.y contains
2144 * 1.0 or -1.0 to turn the rendering upside down for offscreen rendering. PosFixup.x
2145 * contains 1.0 to allow a mad, but arb vs swizzles are too restricted for that.
2146 */
2147 shader_addline(buffer, "MUL TMP, posFixup, TMP_OUT.w;\n");
2148 shader_addline(buffer, "ADD TMP_OUT.x, TMP_OUT.x, TMP.z;\n");
2149 shader_addline(buffer, "MAD TMP_OUT.y, TMP_OUT.y, posFixup.y, TMP.w;\n");
2150
2151 /* Z coord [0;1]->[-1;1] mapping, see comment in transform_projection in state.c
2152 * and the glsl equivalent
2153 */
2154 shader_addline(buffer, "MAD TMP_OUT.z, TMP_OUT.z, helper_const.x, -TMP_OUT.w;\n");
2155
2156 shader_addline(buffer, "MOV result.position, TMP_OUT;\n");
2157
2158 shader_addline(buffer, "END\n");
2159
2160 /* TODO: change to resource.glObjectHandle or something like that */
2161 GL_EXTCALL(glGenProgramsARB(1, &ret));
2162
2163 TRACE("Creating a hw vertex shader, prg=%d\n", ret);
2164 GL_EXTCALL(glBindProgramARB(GL_VERTEX_PROGRAM_ARB, ret));
2165
2166 TRACE("Created hw vertex shader, prg=%d\n", ret);
2167 /* Create the program and check for errors */
2168 GL_EXTCALL(glProgramStringARB(GL_VERTEX_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB,
2169 buffer->bsize, buffer->buffer));
2170
2171 if (glGetError() == GL_INVALID_OPERATION) {
2172 GLint errPos;
2173 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &errPos);
2174 FIXME("HW VertexShader Error at position %d: %s\n",
2175 errPos, debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
2176 ret = -1;
2177 } else {
2178 /* Load immediate constants */
2179 if(!This->baseShader.load_local_constsF) {
2180 LIST_FOR_EACH_ENTRY(lconst, &This->baseShader.constantsF, local_constant, entry) {
2181 const float *value = (const float *)lconst->value;
2182 GL_EXTCALL(glProgramLocalParameter4fvARB(GL_VERTEX_PROGRAM_ARB, lconst->idx, value));
2183 }
2184 }
2185 }
2186 return ret;
2187 }
2188
2189 static void shader_arb_get_caps(WINED3DDEVTYPE devtype, const WineD3D_GL_Info *gl_info, struct shader_caps *pCaps)
2190 {
2191 /* We don't have an ARB fixed function pipeline yet, so let the none backend set its caps,
2192 * then overwrite the shader specific ones
2193 */
2194 none_shader_backend.shader_get_caps(devtype, gl_info, pCaps);
2195
2196 if(GL_SUPPORT(ARB_VERTEX_PROGRAM)) {
2197 pCaps->VertexShaderVersion = WINED3DVS_VERSION(1,1);
2198 TRACE_(d3d_caps)("Hardware vertex shader version 1.1 enabled (ARB_PROGRAM)\n");
2199 pCaps->MaxVertexShaderConst = GL_LIMITS(vshader_constantsF) - ARB_SHADER_RESERVED_VS_CONSTS;
2200 }
2201
2202 if(GL_SUPPORT(ARB_FRAGMENT_PROGRAM)) {
2203 pCaps->PixelShaderVersion = WINED3DPS_VERSION(1,4);
2204 pCaps->PixelShader1xMaxValue = 8.0;
2205 TRACE_(d3d_caps)("Hardware pixel shader version 1.4 enabled (ARB_PROGRAM)\n");
2206 pCaps->MaxPixelShaderConst = GL_LIMITS(pshader_constantsF) - ARB_SHADER_RESERVED_PS_CONSTS;
2207 }
2208 }
2209
2210 static BOOL shader_arb_color_fixup_supported(struct color_fixup_desc fixup)
2211 {
2212 if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
2213 {
2214 TRACE("Checking support for color_fixup:\n");
2215 dump_color_fixup_desc(fixup);
2216 }
2217
2218 /* We support everything except YUV conversions. */
2219 if (!is_yuv_fixup(fixup))
2220 {
2221 TRACE("[OK]\n");
2222 return TRUE;
2223 }
2224
2225 TRACE("[FAILED]\n");
2226 return FALSE;
2227 }
2228
2229 static const SHADER_HANDLER shader_arb_instruction_handler_table[WINED3DSIH_TABLE_SIZE] =
2230 {
2231 /* WINED3DSIH_ABS */ shader_hw_map2gl,
2232 /* WINED3DSIH_ADD */ shader_hw_map2gl,
2233 /* WINED3DSIH_BEM */ pshader_hw_bem,
2234 /* WINED3DSIH_BREAK */ NULL,
2235 /* WINED3DSIH_BREAKC */ NULL,
2236 /* WINED3DSIH_BREAKP */ NULL,
2237 /* WINED3DSIH_CALL */ NULL,
2238 /* WINED3DSIH_CALLNZ */ NULL,
2239 /* WINED3DSIH_CMP */ pshader_hw_cmp,
2240 /* WINED3DSIH_CND */ pshader_hw_cnd,
2241 /* WINED3DSIH_CRS */ shader_hw_map2gl,
2242 /* WINED3DSIH_DCL */ NULL,
2243 /* WINED3DSIH_DEF */ NULL,
2244 /* WINED3DSIH_DEFB */ NULL,
2245 /* WINED3DSIH_DEFI */ NULL,
2246 /* WINED3DSIH_DP2ADD */ pshader_hw_dp2add,
2247 /* WINED3DSIH_DP3 */ shader_hw_map2gl,
2248 /* WINED3DSIH_DP4 */ shader_hw_map2gl,
2249 /* WINED3DSIH_DST */ shader_hw_map2gl,
2250 /* WINED3DSIH_DSX */ NULL,
2251 /* WINED3DSIH_DSY */ NULL,
2252 /* WINED3DSIH_ELSE */ NULL,
2253 /* WINED3DSIH_ENDIF */ NULL,
2254 /* WINED3DSIH_ENDLOOP */ NULL,
2255 /* WINED3DSIH_ENDREP */ NULL,
2256 /* WINED3DSIH_EXP */ shader_hw_map2gl,
2257 /* WINED3DSIH_EXPP */ shader_hw_map2gl,
2258 /* WINED3DSIH_FRC */ shader_hw_map2gl,
2259 /* WINED3DSIH_IF */ NULL,
2260 /* WINED3DSIH_IFC */ NULL,
2261 /* WINED3DSIH_LABEL */ NULL,
2262 /* WINED3DSIH_LIT */ shader_hw_map2gl,
2263 /* WINED3DSIH_LOG */ shader_hw_map2gl,
2264 /* WINED3DSIH_LOGP */ shader_hw_map2gl,
2265 /* WINED3DSIH_LOOP */ NULL,
2266 /* WINED3DSIH_LRP */ shader_hw_map2gl,
2267 /* WINED3DSIH_M3x2 */ shader_hw_mnxn,
2268 /* WINED3DSIH_M3x3 */ shader_hw_mnxn,
2269 /* WINED3DSIH_M3x4 */ shader_hw_mnxn,
2270 /* WINED3DSIH_M4x3 */ shader_hw_mnxn,
2271 /* WINED3DSIH_M4x4 */ shader_hw_mnxn,
2272 /* WINED3DSIH_MAD */ shader_hw_map2gl,
2273 /* WINED3DSIH_MAX */ shader_hw_map2gl,
2274 /* WINED3DSIH_MIN */ shader_hw_map2gl,
2275 /* WINED3DSIH_MOV */ shader_hw_mov,
2276 /* WINED3DSIH_MOVA */ shader_hw_mov,
2277 /* WINED3DSIH_MUL */ shader_hw_map2gl,
2278 /* WINED3DSIH_NOP */ shader_hw_map2gl,
2279 /* WINED3DSIH_NRM */ shader_hw_nrm,
2280 /* WINED3DSIH_PHASE */ NULL,
2281 /* WINED3DSIH_POW */ shader_hw_map2gl,
2282 /* WINED3DSIH_RCP */ vshader_hw_rsq_rcp,
2283 /* WINED3DSIH_REP */ NULL,
2284 /* WINED3DSIH_RET */ NULL,
2285 /* WINED3DSIH_RSQ */ vshader_hw_rsq_rcp,
2286 /* WINED3DSIH_SETP */ NULL,
2287 /* WINED3DSIH_SGE */ shader_hw_map2gl,
2288 /* WINED3DSIH_SGN */ NULL,
2289 /* WINED3DSIH_SINCOS */ shader_hw_sincos,
2290 /* WINED3DSIH_SLT */ shader_hw_map2gl,
2291 /* WINED3DSIH_SUB */ shader_hw_map2gl,
2292 /* WINED3DSIH_TEX */ pshader_hw_tex,
2293 /* WINED3DSIH_TEXBEM */ pshader_hw_texbem,
2294 /* WINED3DSIH_TEXBEML */ pshader_hw_texbem,
2295 /* WINED3DSIH_TEXCOORD */ pshader_hw_texcoord,
2296 /* WINED3DSIH_TEXDEPTH */ pshader_hw_texdepth,
2297 /* WINED3DSIH_TEXDP3 */ pshader_hw_texdp3,
2298 /* WINED3DSIH_TEXDP3TEX */ pshader_hw_texdp3tex,
2299 /* WINED3DSIH_TEXKILL */ pshader_hw_texkill,
2300 /* WINED3DSIH_TEXLDD */ NULL,
2301 /* WINED3DSIH_TEXLDL */ NULL,
2302 /* WINED3DSIH_TEXM3x2DEPTH */ pshader_hw_texm3x2depth,
2303 /* WINED3DSIH_TEXM3x2PAD */ pshader_hw_texm3x2pad,
2304 /* WINED3DSIH_TEXM3x2TEX */ pshader_hw_texm3x2tex,
2305 /* WINED3DSIH_TEXM3x3 */ pshader_hw_texm3x3,
2306 /* WINED3DSIH_TEXM3x3DIFF */ NULL,
2307 /* WINED3DSIH_TEXM3x3PAD */ pshader_hw_texm3x3pad,
2308 /* WINED3DSIH_TEXM3x3SPEC */ pshader_hw_texm3x3spec,
2309 /* WINED3DSIH_TEXM3x3TEX */ pshader_hw_texm3x3tex,
2310 /* WINED3DSIH_TEXM3x3VSPEC */ pshader_hw_texm3x3vspec,
2311 /* WINED3DSIH_TEXREG2AR */ pshader_hw_texreg2ar,
2312 /* WINED3DSIH_TEXREG2GB */ pshader_hw_texreg2gb,
2313 /* WINED3DSIH_TEXREG2RGB */ pshader_hw_texreg2rgb,
2314 };
2315
2316 const shader_backend_t arb_program_shader_backend = {
2317 shader_arb_instruction_handler_table,
2318 shader_arb_select,
2319 shader_arb_select_depth_blt,
2320 shader_arb_deselect_depth_blt,
2321 shader_arb_update_float_vertex_constants,
2322 shader_arb_update_float_pixel_constants,
2323 shader_arb_load_constants,
2324 shader_arb_load_np2fixup_constants,
2325 shader_arb_destroy,
2326 shader_arb_alloc,
2327 shader_arb_free,
2328 shader_arb_dirty_const,
2329 shader_arb_generate_pshader,
2330 shader_arb_generate_vshader,
2331 shader_arb_get_caps,
2332 shader_arb_color_fixup_supported,
2333 };
2334
2335 /* ARB_fragment_program fixed function pipeline replacement definitions */
2336 #define ARB_FFP_CONST_TFACTOR 0
2337 #define ARB_FFP_CONST_SPECULAR_ENABLE ((ARB_FFP_CONST_TFACTOR) + 1)
2338 #define ARB_FFP_CONST_CONSTANT(i) ((ARB_FFP_CONST_SPECULAR_ENABLE) + 1 + i)
2339 #define ARB_FFP_CONST_BUMPMAT(i) ((ARB_FFP_CONST_CONSTANT(7)) + 1 + i)
2340 #define ARB_FFP_CONST_LUMINANCE(i) ((ARB_FFP_CONST_BUMPMAT(7)) + 1 + i)
2341
2342 struct arbfp_ffp_desc
2343 {
2344 struct ffp_frag_desc parent;
2345 GLuint shader;
2346 unsigned int num_textures_used;
2347 };
2348
2349 static void arbfp_enable(IWineD3DDevice *iface, BOOL enable) {
2350 if(enable) {
2351 glEnable(GL_FRAGMENT_PROGRAM_ARB);
2352 checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB)");
2353 } else {
2354 glDisable(GL_FRAGMENT_PROGRAM_ARB);
2355 checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
2356 }
2357 }
2358
2359 static HRESULT arbfp_alloc(IWineD3DDevice *iface) {
2360 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
2361 struct shader_arb_priv *priv;
2362 /* Share private data between the shader backend and the pipeline replacement, if both
2363 * are the arb implementation. This is needed to figure out whether ARBfp should be disabled
2364 * if no pixel shader is bound or not
2365 */
2366 if(This->shader_backend == &arb_program_shader_backend) {
2367 This->fragment_priv = This->shader_priv;
2368 } else {
2369 This->fragment_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct shader_arb_priv));
2370 if(!This->fragment_priv) return E_OUTOFMEMORY;
2371 }
2372 priv = This->fragment_priv;
2373 priv->fragment_shaders = hash_table_create(ffp_frag_program_key_hash, ffp_frag_program_key_compare);
2374 priv->use_arbfp_fixed_func = TRUE;
2375 return WINED3D_OK;
2376 }
2377
2378 static void arbfp_free_ffpshader(void *value, void *gli) {
2379 const WineD3D_GL_Info *gl_info = gli;
2380 struct arbfp_ffp_desc *entry_arb = value;
2381
2382 ENTER_GL();
2383 GL_EXTCALL(glDeleteProgramsARB(1, &entry_arb->shader));
2384 checkGLcall("glDeleteProgramsARB(1, &entry_arb->shader)");
2385 HeapFree(GetProcessHeap(), 0, entry_arb);
2386 LEAVE_GL();
2387 }
2388
2389 static void arbfp_free(IWineD3DDevice *iface) {
2390 IWineD3DDeviceImpl *This = (IWineD3DDeviceImpl *) iface;
2391 struct shader_arb_priv *priv = This->fragment_priv;
2392
2393 hash_table_destroy(priv->fragment_shaders, arbfp_free_ffpshader, &This->adapter->gl_info);
2394 priv->use_arbfp_fixed_func = FALSE;
2395
2396 if(This->shader_backend != &arb_program_shader_backend) {
2397 HeapFree(GetProcessHeap(), 0, This->fragment_priv);
2398 }
2399 }
2400
2401 static void arbfp_get_caps(WINED3DDEVTYPE devtype, const WineD3D_GL_Info *gl_info, struct fragment_caps *caps)
2402 {
2403 caps->TextureOpCaps = WINED3DTEXOPCAPS_DISABLE |
2404 WINED3DTEXOPCAPS_SELECTARG1 |
2405 WINED3DTEXOPCAPS_SELECTARG2 |
2406 WINED3DTEXOPCAPS_MODULATE4X |
2407 WINED3DTEXOPCAPS_MODULATE2X |
2408 WINED3DTEXOPCAPS_MODULATE |
2409 WINED3DTEXOPCAPS_ADDSIGNED2X |
2410 WINED3DTEXOPCAPS_ADDSIGNED |
2411 WINED3DTEXOPCAPS_ADD |
2412 WINED3DTEXOPCAPS_SUBTRACT |
2413 WINED3DTEXOPCAPS_ADDSMOOTH |
2414 WINED3DTEXOPCAPS_BLENDCURRENTALPHA |
2415 WINED3DTEXOPCAPS_BLENDFACTORALPHA |
2416 WINED3DTEXOPCAPS_BLENDTEXTUREALPHA |
2417 WINED3DTEXOPCAPS_BLENDDIFFUSEALPHA |
2418 WINED3DTEXOPCAPS_BLENDTEXTUREALPHAPM |
2419 WINED3DTEXOPCAPS_MODULATEALPHA_ADDCOLOR |
2420 WINED3DTEXOPCAPS_MODULATECOLOR_ADDALPHA |
2421 WINED3DTEXOPCAPS_MODULATEINVCOLOR_ADDALPHA |
2422 WINED3DTEXOPCAPS_MODULATEINVALPHA_ADDCOLOR |
2423 WINED3DTEXOPCAPS_DOTPRODUCT3 |
2424 WINED3DTEXOPCAPS_MULTIPLYADD |
2425 WINED3DTEXOPCAPS_LERP |
2426 WINED3DTEXOPCAPS_BUMPENVMAP |
2427 WINED3DTEXOPCAPS_BUMPENVMAPLUMINANCE;
2428
2429 /* TODO: Implement WINED3DTEXOPCAPS_PREMODULATE */
2430
2431 caps->MaxTextureBlendStages = 8;
2432 caps->MaxSimultaneousTextures = min(GL_LIMITS(fragment_samplers), 8);
2433
2434 caps->PrimitiveMiscCaps |= WINED3DPMISCCAPS_TSSARGTEMP;
2435 }
2436 #undef GLINFO_LOCATION
2437
2438 #define GLINFO_LOCATION stateblock->wineD3DDevice->adapter->gl_info
2439 static void state_texfactor_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2440 float col[4];
2441 IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2442
2443 /* Don't load the parameter if we're using an arbfp pixel shader, otherwise we'll overwrite
2444 * application provided constants
2445 */
2446 if(device->shader_backend == &arb_program_shader_backend) {
2447 if (use_ps(stateblock)) return;
2448
2449 device = stateblock->wineD3DDevice;
2450 device->activeContext->pshader_const_dirty[ARB_FFP_CONST_TFACTOR] = 1;
2451 device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_TFACTOR + 1);
2452 }
2453
2454 D3DCOLORTOGLFLOAT4(stateblock->renderState[WINED3DRS_TEXTUREFACTOR], col);
2455 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_TFACTOR, col));
2456 checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_TFACTOR, col)");
2457
2458 }
2459
2460 static void state_arb_specularenable(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2461 float col[4];
2462 IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2463
2464 /* Don't load the parameter if we're using an arbfp pixel shader, otherwise we'll overwrite
2465 * application provided constants
2466 */
2467 if(device->shader_backend == &arb_program_shader_backend) {
2468 if (use_ps(stateblock)) return;
2469
2470 device = stateblock->wineD3DDevice;
2471 device->activeContext->pshader_const_dirty[ARB_FFP_CONST_SPECULAR_ENABLE] = 1;
2472 device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_SPECULAR_ENABLE + 1);
2473 }
2474
2475 if(stateblock->renderState[WINED3DRS_SPECULARENABLE]) {
2476 /* The specular color has no alpha */
2477 col[0] = 1.0; col[1] = 1.0;
2478 col[2] = 1.0; col[3] = 0.0;
2479 } else {
2480 col[0] = 0.0; col[1] = 0.0;
2481 col[2] = 0.0; col[3] = 0.0;
2482 }
2483 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_SPECULAR_ENABLE, col));
2484 checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_SPECULAR_ENABLE, col)");
2485 }
2486
2487 static void set_bumpmat_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2488 DWORD stage = (state - STATE_TEXTURESTAGE(0, 0)) / (WINED3D_HIGHEST_TEXTURE_STATE + 1);
2489 IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2490 float mat[2][2];
2491
2492 if (use_ps(stateblock))
2493 {
2494 if(stage != 0 &&
2495 ((IWineD3DPixelShaderImpl *) stateblock->pixelShader)->baseShader.reg_maps.bumpmat[stage]) {
2496 /* The pixel shader has to know the bump env matrix. Do a constants update if it isn't scheduled
2497 * anyway
2498 */
2499 if(!isStateDirty(context, STATE_PIXELSHADERCONSTANT)) {
2500 device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
2501 }
2502 }
2503
2504 if(device->shader_backend == &arb_program_shader_backend) {
2505 /* Exit now, don't set the bumpmat below, otherwise we may overwrite pixel shader constants */
2506 return;
2507 }
2508 } else if(device->shader_backend == &arb_program_shader_backend) {
2509 device->activeContext->pshader_const_dirty[ARB_FFP_CONST_BUMPMAT(stage)] = 1;
2510 device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_BUMPMAT(stage) + 1);
2511 }
2512
2513 mat[0][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT00]);
2514 mat[0][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT01]);
2515 mat[1][0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT10]);
2516 mat[1][1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVMAT11]);
2517
2518 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_BUMPMAT(stage), &mat[0][0]));
2519 checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_BUMPMAT(stage), &mat[0][0])");
2520 }
2521
2522 static void tex_bumpenvlum_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
2523 DWORD stage = (state - STATE_TEXTURESTAGE(0, 0)) / (WINED3D_HIGHEST_TEXTURE_STATE + 1);
2524 IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
2525 float param[4];
2526
2527 if (use_ps(stateblock))
2528 {
2529 if(stage != 0 &&
2530 ((IWineD3DPixelShaderImpl *) stateblock->pixelShader)->baseShader.reg_maps.luminanceparams[stage]) {
2531 /* The pixel shader has to know the luminance offset. Do a constants update if it
2532 * isn't scheduled anyway
2533 */
2534 if(!isStateDirty(context, STATE_PIXELSHADERCONSTANT)) {
2535 device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
2536 }
2537 }
2538
2539 if(device->shader_backend == &arb_program_shader_backend) {
2540 /* Exit now, don't set the bumpmat below, otherwise we may overwrite pixel shader constants */
2541 return;
2542 }
2543 } else if(device->shader_backend == &arb_program_shader_backend) {
2544 device->activeContext->pshader_const_dirty[ARB_FFP_CONST_LUMINANCE(stage)] = 1;
2545 device->highest_dirty_ps_const = max(device->highest_dirty_ps_const, ARB_FFP_CONST_LUMINANCE(stage) + 1);
2546 }
2547
2548 param[0] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVLSCALE]);
2549 param[1] = *((float *) &stateblock->textureState[stage][WINED3DTSS_BUMPENVLOFFSET]);
2550 param[2] = 0.0;
2551 param[3] = 0.0;
2552
2553 GL_EXTCALL(glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_LUMINANCE(stage), param));
2554 checkGLcall("glProgramEnvParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, ARB_FFP_CONST_LUMINANCE(stage), param)");
2555 }
2556
2557 static const char *get_argreg(SHADER_BUFFER *buffer, DWORD argnum, unsigned int stage, DWORD arg) {
2558 const char *ret;
2559
2560 if(arg == ARG_UNUSED) return "unused"; /* This is the marker for unused registers */
2561
2562 switch(arg & WINED3DTA_SELECTMASK) {
2563 case WINED3DTA_DIFFUSE:
2564 ret = "fragment.color.primary"; break;
2565
2566 case WINED3DTA_CURRENT:
2567 if(stage == 0) ret = "fragment.color.primary";
2568 else ret = "ret";
2569 break;
2570
2571 case WINED3DTA_TEXTURE:
2572 switch(stage) {
2573 case 0: ret = "tex0"; break;
2574 case 1: ret = "tex1"; break;
2575 case 2: ret = "tex2"; break;
2576 case 3: ret = "tex3"; break;
2577 case 4: ret = "tex4"; break;
2578 case 5: ret = "tex5"; break;
2579 case 6: ret = "tex6"; break;
2580 case 7: ret = "tex7"; break;
2581 default: ret = "unknown texture";
2582 }
2583 break;
2584
2585 case WINED3DTA_TFACTOR:
2586 ret = "tfactor"; break;
2587
2588 case WINED3DTA_SPECULAR:
2589 ret = "fragment.color.secondary"; break;
2590
2591 case WINED3DTA_TEMP:
2592 ret = "tempreg"; break;
2593
2594 case WINED3DTA_CONSTANT:
2595 FIXME("Implement perstage constants\n");
2596 switch(stage) {
2597 case 0: ret = "const0"; break;
2598 case 1: ret = "const1"; break;
2599 case 2: ret = "const2"; break;
2600 case 3: ret = "const3"; break;
2601 case 4: ret = "const4"; break;
2602 case 5: ret = "const5"; break;
2603 case 6: ret = "const6"; break;
2604 case 7: ret = "const7"; break;
2605 default: ret = "unknown constant";
2606 }
2607 break;
2608
2609 default:
2610 return "unknown";
2611 }
2612
2613 if(arg & WINED3DTA_COMPLEMENT) {
2614 shader_addline(buffer, "SUB arg%u, const.x, %s;\n", argnum, ret);
2615 if(argnum == 0) ret = "arg0";
2616 if(argnum == 1) ret = "arg1";
2617 if(argnum == 2) ret = "arg2";
2618 }
2619 if(arg & WINED3DTA_ALPHAREPLICATE) {
2620 shader_addline(buffer, "MOV arg%u, %s.w;\n", argnum, ret);
2621 if(argnum == 0) ret = "arg0";
2622 if(argnum == 1) ret = "arg1";
2623 if(argnum == 2) ret = "arg2";
2624 }
2625 return ret;
2626 }
2627
2628 static void gen_ffp_instr(SHADER_BUFFER *buffer, unsigned int stage, BOOL color, BOOL alpha,
2629 DWORD dst, DWORD op, DWORD dw_arg0, DWORD dw_arg1, DWORD dw_arg2) {
2630 const char *dstmask, *dstreg, *arg0, *arg1, *arg2;
2631 unsigned int mul = 1;
2632 BOOL mul_final_dest = FALSE;
2633
2634 if(color && alpha) dstmask = "";
2635 else if(color) dstmask = ".xyz";
2636 else dstmask = ".w";
2637
2638 if(dst == tempreg) dstreg = "tempreg";
2639 else dstreg = "ret";
2640
2641 arg0 = get_argreg(buffer, 0, stage, dw_arg0);
2642 arg1 = get_argreg(buffer, 1, stage, dw_arg1);
2643 arg2 = get_argreg(buffer, 2, stage, dw_arg2);
2644
2645 switch(op) {
2646 case WINED3DTOP_DISABLE:
2647 if(stage == 0) shader_addline(buffer, "MOV %s%s, fragment.color.primary;\n", dstreg, dstmask);
2648 break;
2649
2650 case WINED3DTOP_SELECTARG2:
2651 arg1 = arg2;
2652 case WINED3DTOP_SELECTARG1:
2653 shader_addline(buffer, "MOV %s%s, %s;\n", dstreg, dstmask, arg1);
2654 break;
2655
2656 case WINED3DTOP_MODULATE4X:
2657 mul = 2;
2658 case WINED3DTOP_MODULATE2X:
2659 mul *= 2;
2660 if(strcmp(dstreg, "result.color") == 0) {
2661 dstreg = "ret";
2662 mul_final_dest = TRUE;
2663 }
2664 case WINED3DTOP_MODULATE:
2665 shader_addline(buffer, "MUL %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2666 break;
2667
2668 case WINED3DTOP_ADDSIGNED2X:
2669 mul = 2;
2670 if(strcmp(dstreg, "result.color") == 0) {
2671 dstreg = "ret";
2672 mul_final_dest = TRUE;
2673 }
2674 case WINED3DTOP_ADDSIGNED:
2675 shader_addline(buffer, "SUB arg2, %s, const.w;\n", arg2);
2676 arg2 = "arg2";
2677 case WINED3DTOP_ADD:
2678 shader_addline(buffer, "ADD_SAT %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2679 break;
2680
2681 case WINED3DTOP_SUBTRACT:
2682 shader_addline(buffer, "SUB_SAT %s%s, %s, %s;\n", dstreg, dstmask, arg1, arg2);
2683 break;
2684
2685 case WINED3DTOP_ADDSMOOTH:
2686 shader_addline(buffer, "SUB arg1, const.x, %s;\n", arg1);
2687 shader_addline(buffer, "MAD_SAT %s%s, arg1, %s, %s;\n", dstreg, dstmask, arg2, arg1);
2688 break;
2689
2690 case WINED3DTOP_BLENDCURRENTALPHA:
2691 arg0 = get_argreg(buffer, 0, stage, WINED3DTA_CURRENT);
2692 shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2693 break;
2694 case WINED3DTOP_BLENDFACTORALPHA:
2695 arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TFACTOR);
2696 shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2697 break;
2698 case WINED3DTOP_BLENDTEXTUREALPHA:
2699 arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TEXTURE);
2700 shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2701 break;
2702 case WINED3DTOP_BLENDDIFFUSEALPHA:
2703 arg0 = get_argreg(buffer, 0, stage, WINED3DTA_DIFFUSE);
2704 shader_addline(buffer, "LRP %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2705 break;
2706
2707 case WINED3DTOP_BLENDTEXTUREALPHAPM:
2708 arg0 = get_argreg(buffer, 0, stage, WINED3DTA_TEXTURE);
2709 shader_addline(buffer, "SUB arg0.w, const.x, %s.w;\n", arg0);
2710 shader_addline(buffer, "MAD_SAT %s%s, %s, arg0.w, %s;\n", dstreg, dstmask, arg2, arg1);
2711 break;
2712
2713 /* D3DTOP_PREMODULATE ???? */
2714
2715 case WINED3DTOP_MODULATEINVALPHA_ADDCOLOR:
2716 shader_addline(buffer, "SUB arg0.w, const.x, %s;\n", arg1);
2717 shader_addline(buffer, "MAD_SAT %s%s, arg0.w, %s, %s;\n", dstreg, dstmask, arg2, arg1);
2718 break;
2719 case WINED3DTOP_MODULATEALPHA_ADDCOLOR:
2720 shader_addline(buffer, "MAD_SAT %s%s, %s.w, %s, %s;\n", dstreg, dstmask, arg1, arg2, arg1);
2721 break;
2722 case WINED3DTOP_MODULATEINVCOLOR_ADDALPHA:
2723 shader_addline(buffer, "SUB arg0, const.x, %s;\n", arg1);
2724 shader_addline(buffer, "MAD_SAT %s%s, arg0, %s, %s.w;\n", dstreg, dstmask, arg2, arg1);
2725 break;
2726 case WINED3DTOP_MODULATECOLOR_ADDALPHA:
2727 shader_addline(buffer, "MAD_SAT %s%s, %s, %s, %s.w;\n", dstreg, dstmask, arg1, arg2, arg1);
2728 break;
2729
2730 case WINED3DTOP_DOTPRODUCT3:
2731 mul = 4;
2732 if(strcmp(dstreg, "result.color") == 0) {
2733 dstreg = "ret";
2734 mul_final_dest = TRUE;
2735 }
2736 shader_addline(buffer, "SUB arg1, %s, const.w;\n", arg1);
2737 shader_addline(buffer, "SUB arg2, %s, const.w;\n", arg2);
2738 shader_addline(buffer, "DP3_SAT %s%s, arg1, arg2;\n", dstreg, dstmask);
2739 break;
2740
2741 case WINED3DTOP_MULTIPLYADD:
2742 shader_addline(buffer, "MAD_SAT %s%s, %s, %s, %s;\n", dstreg, dstmask, arg1, arg2, arg0);
2743 break;
2744
2745 case WINED3DTOP_LERP:
2746 /* The msdn is not quite right here */
2747 shader_addline(buffer, "LRP %s%s, %s, %s, %s;\n", dstreg, dstmask, arg0, arg1, arg2);
2748 break;
2749
2750 case WINED3DTOP_BUMPENVMAP:
2751 case WINED3DTOP_BUMPENVMAPLUMINANCE:
2752 /* Those are handled in the first pass of the shader(generation pass 1 and 2) already */
2753 break;
2754
2755 default:
2756 FIXME("Unhandled texture op %08x\n", op);
2757 }
2758
2759 if(mul == 2) {
2760 shader_addline(buffer, "MUL_SAT %s%s, %s, const.y;\n", mul_final_dest ? "result.color" : dstreg, dstmask, dstreg);
2761 } else if(mul == 4) {
2762 shader_addline(buffer, "MUL_SAT %s%s, %s, const.z;\n", mul_final_dest ? "result.color" : dstreg, dstmask, dstreg);
2763 }
2764 }
2765
2766 /* The stateblock is passed for GLINFO_LOCATION */
2767 static GLuint gen_arbfp_ffp_shader(const struct ffp_frag_settings *settings, IWineD3DStateBlockImpl *stateblock)
2768 {
2769 unsigned int stage;
2770 SHADER_BUFFER buffer;
2771 BOOL tex_read[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2772 BOOL bump_used[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2773 BOOL luminance_used[MAX_TEXTURES] = {FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE, FALSE};
2774 const char *textype;
2775 const char *instr, *sat;
2776 char colorcor_dst[8];
2777 GLuint ret;
2778 DWORD arg0, arg1, arg2;
2779 BOOL tempreg_used = FALSE, tfactor_used = FALSE;
2780 BOOL op_equal;
2781 const char *final_combiner_src = "ret";
2782
2783 /* Find out which textures are read */
2784 for(stage = 0; stage < MAX_TEXTURES; stage++) {
2785 if(settings->op[stage].cop == WINED3DTOP_DISABLE) break;
2786 arg0 = settings->op[stage].carg0 & WINED3DTA_SELECTMASK;
2787 arg1 = settings->op[stage].carg1 & WINED3DTA_SELECTMASK;
2788 arg2 = settings->op[stage].carg2 & WINED3DTA_SELECTMASK;
2789 if(arg0 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2790 if(arg1 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2791 if(arg2 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2792
2793 if(settings->op[stage].cop == WINED3DTOP_BLENDTEXTUREALPHA) tex_read[stage] = TRUE;
2794 if(settings->op[stage].cop == WINED3DTOP_BLENDTEXTUREALPHAPM) tex_read[stage] = TRUE;
2795 if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAP) {
2796 bump_used[stage] = TRUE;
2797 tex_read[stage] = TRUE;
2798 }
2799 if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2800 bump_used[stage] = TRUE;
2801 tex_read[stage] = TRUE;
2802 luminance_used[stage] = TRUE;
2803 } else if(settings->op[stage].cop == WINED3DTOP_BLENDFACTORALPHA) {
2804 tfactor_used = TRUE;
2805 }
2806
2807 if(arg0 == WINED3DTA_TFACTOR || arg1 == WINED3DTA_TFACTOR || arg2 == WINED3DTA_TFACTOR) {
2808 tfactor_used = TRUE;
2809 }
2810
2811 if(settings->op[stage].dst == tempreg) tempreg_used = TRUE;
2812 if(arg0 == WINED3DTA_TEMP || arg1 == WINED3DTA_TEMP || arg2 == WINED3DTA_TEMP) {
2813 tempreg_used = TRUE;
2814 }
2815
2816 if(settings->op[stage].aop == WINED3DTOP_DISABLE) continue;
2817 arg0 = settings->op[stage].aarg0 & WINED3DTA_SELECTMASK;
2818 arg1 = settings->op[stage].aarg1 & WINED3DTA_SELECTMASK;
2819 arg2 = settings->op[stage].aarg2 & WINED3DTA_SELECTMASK;
2820 if(arg0 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2821 if(arg1 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2822 if(arg2 == WINED3DTA_TEXTURE) tex_read[stage] = TRUE;
2823
2824 if(arg0 == WINED3DTA_TEMP || arg1 == WINED3DTA_TEMP || arg2 == WINED3DTA_TEMP) {
2825 tempreg_used = TRUE;
2826 }
2827 if(arg0 == WINED3DTA_TFACTOR || arg1 == WINED3DTA_TFACTOR || arg2 == WINED3DTA_TFACTOR) {
2828 tfactor_used = TRUE;
2829 }
2830 }
2831
2832 /* Shader header */
2833 shader_buffer_init(&buffer);
2834
2835 shader_addline(&buffer, "!!ARBfp1.0\n");
2836
2837 switch(settings->fog) {
2838 case FOG_OFF: break;
2839 case FOG_LINEAR: shader_addline(&buffer, "OPTION ARB_fog_linear;\n"); break;
2840 case FOG_EXP: shader_addline(&buffer, "OPTION ARB_fog_exp;\n"); break;
2841 case FOG_EXP2: shader_addline(&buffer, "OPTION ARB_fog_exp2;\n"); break;
2842 default: FIXME("Unexpected fog setting %d\n", settings->fog);
2843 }
2844
2845 shader_addline(&buffer, "PARAM const = {1, 2, 4, 0.5};\n");
2846 shader_addline(&buffer, "TEMP TMP;\n");
2847 shader_addline(&buffer, "TEMP ret;\n");
2848 if(tempreg_used || settings->sRGB_write) shader_addline(&buffer, "TEMP tempreg;\n");
2849 shader_addline(&buffer, "TEMP arg0;\n");
2850 shader_addline(&buffer, "TEMP arg1;\n");
2851 shader_addline(&buffer, "TEMP arg2;\n");
2852 for(stage = 0; stage < MAX_TEXTURES; stage++) {
2853 if(!tex_read[stage]) continue;
2854 shader_addline(&buffer, "TEMP tex%u;\n", stage);
2855 if(!bump_used[stage]) continue;
2856 shader_addline(&buffer, "PARAM bumpmat%u = program.env[%u];\n", stage, ARB_FFP_CONST_BUMPMAT(stage));
2857 if(!luminance_used[stage]) continue;
2858 shader_addline(&buffer, "PARAM luminance%u = program.env[%u];\n", stage, ARB_FFP_CONST_LUMINANCE(stage));
2859 }
2860 if(tfactor_used) {
2861 shader_addline(&buffer, "PARAM tfactor = program.env[%u];\n", ARB_FFP_CONST_TFACTOR);
2862 }
2863 shader_addline(&buffer, "PARAM specular_enable = program.env[%u];\n", ARB_FFP_CONST_SPECULAR_ENABLE);
2864
2865 if(settings->sRGB_write) {
2866 shader_addline(&buffer, "PARAM srgb_mul_low = {%f, %f, %f, 1.0};\n",
2867 srgb_mul_low, srgb_mul_low, srgb_mul_low);
2868 shader_addline(&buffer, "PARAM srgb_comparison = {%f, %f, %f, %f};\n",
2869 srgb_cmp, srgb_cmp, srgb_cmp, srgb_cmp);
2870 shader_addline(&buffer, "PARAM srgb_pow = {%f, %f, %f, 1.0};\n",
2871 srgb_pow, srgb_pow, srgb_pow);
2872 shader_addline(&buffer, "PARAM srgb_mul_hi = {%f, %f, %f, 1.0};\n",
2873 srgb_mul_high, srgb_mul_high, srgb_mul_high);
2874 shader_addline(&buffer, "PARAM srgb_sub_hi = {%f, %f, %f, 0.0};\n",
2875 srgb_sub_high, srgb_sub_high, srgb_sub_high);
2876 }
2877
2878 /* Generate texture sampling instructions) */
2879 for(stage = 0; stage < MAX_TEXTURES && settings->op[stage].cop != WINED3DTOP_DISABLE; stage++) {
2880 if(!tex_read[stage]) continue;
2881
2882 switch(settings->op[stage].tex_type) {
2883 case tex_1d: textype = "1D"; break;
2884 case tex_2d: textype = "2D"; break;
2885 case tex_3d: textype = "3D"; break;
2886 case tex_cube: textype = "CUBE"; break;
2887 case tex_rect: textype = "RECT"; break;
2888 default: textype = "unexpected_textype"; break;
2889 }
2890
2891 if(settings->op[stage].cop == WINED3DTOP_BUMPENVMAP ||
2892 settings->op[stage].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2893 sat = "";
2894 } else {
2895 sat = "_SAT";
2896 }
2897
2898 if(settings->op[stage].projected == proj_none) {
2899 instr = "TEX";
2900 } else if(settings->op[stage].projected == proj_count4 ||
2901 settings->op[stage].projected == proj_count3) {
2902 instr = "TXP";
2903 } else {
2904 FIXME("Unexpected projection mode %d\n", settings->op[stage].projected);
2905 instr = "TXP";
2906 }
2907
2908 if(stage > 0 &&
2909 (settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAP ||
2910 settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAPLUMINANCE)) {
2911 shader_addline(&buffer, "SWZ arg1, bumpmat%u, x, z, 0, 0;\n", stage - 1);
2912 shader_addline(&buffer, "DP3 ret.x, arg1, tex%u;\n", stage - 1);
2913 shader_addline(&buffer, "SWZ arg1, bumpmat%u, y, w, 0, 0;\n", stage - 1);
2914 shader_addline(&buffer, "DP3 ret.y, arg1, tex%u;\n", stage - 1);
2915
2916 /* with projective textures, texbem only divides the static texture coord, not the displacement,
2917 * so multiply the displacement with the dividing parameter before passing it to TXP
2918 */
2919 if (settings->op[stage].projected != proj_none) {
2920 if(settings->op[stage].projected == proj_count4) {
2921 shader_addline(&buffer, "MOV ret.w, fragment.texcoord[%u].w;\n", stage);
2922 shader_addline(&buffer, "MUL ret.xyz, ret, fragment.texcoord[%u].w, fragment.texcoord[%u];\n", stage, stage);
2923 } else {
2924 shader_addline(&buffer, "MOV ret.w, fragment.texcoord[%u].z;\n", stage);
2925 shader_addline(&buffer, "MAD ret.xyz, ret, fragment.texcoord[%u].z, fragment.texcoord[%u];\n", stage, stage);
2926 }
2927 } else {
2928 shader_addline(&buffer, "ADD ret, ret, fragment.texcoord[%u];\n", stage);
2929 }
2930
2931 shader_addline(&buffer, "%s%s tex%u, ret, texture[%u], %s;\n",
2932 instr, sat, stage, stage, textype);
2933 if(settings->op[stage - 1].cop == WINED3DTOP_BUMPENVMAPLUMINANCE) {
2934 shader_addline(&buffer, "MAD_SAT ret.x, tex%u.z, luminance%u.x, luminance%u.y;\n",
2935 stage - 1, stage - 1, stage - 1);
2936 shader_addline(&buffer, "MUL tex%u, tex%u, ret.x;\n", stage, stage);
2937 }
2938 } else if(settings->op[stage].projected == proj_count3) {
2939 shader_addline(&buffer, "MOV ret, fragment.texcoord[%u];\n", stage);
2940 shader_addline(&buffer, "MOV ret.w, ret.z;\n");
2941 shader_addline(&buffer, "%s%s tex%u, ret, texture[%u], %s;\n",
2942 instr, sat, stage, stage, textype);
2943 } else {
2944 shader_addline(&buffer, "%s%s tex%u, fragment.texcoord[%u], texture[%u], %s;\n",
2945 instr, sat, stage, stage, stage, textype);
2946 }
2947
2948 sprintf(colorcor_dst, "tex%u", stage);
2949 gen_color_correction(&buffer, colorcor_dst, WINED3DSP_WRITEMASK_ALL, "const.x", "const.y",
2950 settings->op[stage].color_fixup);
2951 }
2952
2953 /* Generate the main shader */
2954 for(stage = 0; stage < MAX_TEXTURES; stage++) {
2955 if(settings->op[stage].cop == WINED3DTOP_DISABLE) {
2956 if(stage == 0) {
2957 final_combiner_src = "fragment.color.primary";
2958 }
2959 break;
2960 }
2961
2962 if(settings->op[stage].cop == WINED3DTOP_SELECTARG1 &&
2963 settings->op[stage].aop == WINED3DTOP_SELECTARG1) {
2964 op_equal = settings->op[stage].carg1 == settings->op[stage].aarg1;
2965 } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG1 &&
2966 settings->op[stage].aop == WINED3DTOP_SELECTARG2) {
2967 op_equal = settings->op[stage].carg1 == settings->op[stage].aarg2;
2968 } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG2 &&
2969 settings->op[stage].aop == WINED3DTOP_SELECTARG1) {
2970 op_equal = settings->op[stage].carg2 == settings->op[stage].aarg1;
2971 } else if(settings->op[stage].cop == WINED3DTOP_SELECTARG2 &&
2972 settings->op[stage].aop == WINED3DTOP_SELECTARG2) {
2973 op_equal = settings->op[stage].carg2 == settings->op[stage].aarg2;
2974 } else {
2975 op_equal = settings->op[stage].aop == settings->op[stage].cop &&
2976 settings->op[stage].carg0 == settings->op[stage].aarg0 &&
2977 settings->op[stage].carg1 == settings->op[stage].aarg1 &&
2978 settings->op[stage].carg2 == settings->op[stage].aarg2;
2979 }
2980
2981 if(settings->op[stage].aop == WINED3DTOP_DISABLE) {
2982 gen_ffp_instr(&buffer, stage, TRUE, FALSE, settings->op[stage].dst,
2983 settings->op[stage].cop, settings->op[stage].carg0,
2984 settings->op[stage].carg1, settings->op[stage].carg2);
2985 if(stage == 0) {
2986 shader_addline(&buffer, "MOV ret.w, fragment.color.primary.w;\n");
2987 }
2988 } else if(op_equal) {
2989 gen_ffp_instr(&buffer, stage, TRUE, TRUE, settings->op[stage].dst,
2990 settings->op[stage].cop, settings->op[stage].carg0,
2991 settings->op[stage].carg1, settings->op[stage].carg2);
2992 } else {
2993 gen_ffp_instr(&buffer, stage, TRUE, FALSE, settings->op[stage].dst,
2994 settings->op[stage].cop, settings->op[stage].carg0,
2995 settings->op[stage].carg1, settings->op[stage].carg2);
2996 gen_ffp_instr(&buffer, stage, FALSE, TRUE, settings->op[stage].dst,
2997 settings->op[stage].aop, settings->op[stage].aarg0,
2998 settings->op[stage].aarg1, settings->op[stage].aarg2);
2999 }
3000 }
3001
3002 if(settings->sRGB_write) {
3003 shader_addline(&buffer, "MAD ret, fragment.color.secondary, specular_enable, %s;\n", final_combiner_src);
3004 arbfp_add_sRGB_correction(&buffer, "ret", "arg0", "arg1", "arg2", "tempreg");
3005 shader_addline(&buffer, "MOV result.color.w, ret.w;\n");
3006 } else {
3007 shader_addline(&buffer, "MAD result.color, fragment.color.secondary, specular_enable, %s;\n", final_combiner_src);
3008 }
3009
3010 /* Footer */
3011 shader_addline(&buffer, "END\n");
3012
3013 /* Generate the shader */
3014 GL_EXTCALL(glGenProgramsARB(1, &ret));
3015 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, ret));
3016 GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(buffer.buffer), buffer.buffer));
3017
3018 if (glGetError() == GL_INVALID_OPERATION) {
3019 GLint pos;
3020 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
3021 FIXME("Fragment program error at position %d: %s\n", pos,
3022 debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
3023 }
3024 shader_buffer_free(&buffer);
3025 return ret;
3026 }
3027
3028 static void fragment_prog_arbfp(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
3029 IWineD3DDeviceImpl *device = stateblock->wineD3DDevice;
3030 struct shader_arb_priv *priv = device->fragment_priv;
3031 BOOL use_pshader = use_ps(stateblock);
3032 BOOL use_vshader = use_vs(stateblock);
3033 struct ffp_frag_settings settings;
3034 const struct arbfp_ffp_desc *desc;
3035 unsigned int i;
3036
3037 TRACE("state %#x, stateblock %p, context %p\n", state, stateblock, context);
3038
3039 if(isStateDirty(context, STATE_RENDER(WINED3DRS_FOGENABLE))) {
3040 if(!use_pshader && device->shader_backend == &arb_program_shader_backend && context->last_was_pshader) {
3041 /* Reload fixed function constants since they collide with the pixel shader constants */
3042 for(i = 0; i < MAX_TEXTURES; i++) {
3043 set_bumpmat_arbfp(STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT00), stateblock, context);
3044 }
3045 state_texfactor_arbfp(STATE_RENDER(WINED3DRS_TEXTUREFACTOR), stateblock, context);
3046 state_arb_specularenable(STATE_RENDER(WINED3DRS_SPECULARENABLE), stateblock, context);
3047 } else if(use_pshader && !isStateDirty(context, device->StateTable[STATE_VSHADER].representative)) {
3048 device->shader_backend->shader_select((IWineD3DDevice *)stateblock->wineD3DDevice, use_pshader, use_vshader);
3049 }
3050 return;
3051 }
3052
3053 if(!use_pshader) {
3054 /* Find or create a shader implementing the fixed function pipeline settings, then activate it */
3055 gen_ffp_frag_op(stateblock, &settings, FALSE);
3056 desc = (const struct arbfp_ffp_desc *)find_ffp_frag_shader(priv->fragment_shaders, &settings);
3057 if(!desc) {
3058 struct arbfp_ffp_desc *new_desc = HeapAlloc(GetProcessHeap(), 0, sizeof(*new_desc));
3059 if (!new_desc)
3060 {
3061 ERR("Out of memory\n");
3062 return;
3063 }
3064 new_desc->num_textures_used = 0;
3065 for(i = 0; i < GL_LIMITS(texture_stages); i++) {
3066 if(settings.op[i].cop == WINED3DTOP_DISABLE) break;
3067 new_desc->num_textures_used = i;
3068 }
3069
3070 memcpy(&new_desc->parent.settings, &settings, sizeof(settings));
3071 new_desc->shader = gen_arbfp_ffp_shader(&settings, stateblock);
3072 add_ffp_frag_shader(priv->fragment_shaders, &new_desc->parent);
3073 TRACE("Allocated fixed function replacement shader descriptor %p\n", new_desc);
3074 desc = new_desc;
3075 }
3076
3077 /* Now activate the replacement program. GL_FRAGMENT_PROGRAM_ARB is already active(however, note the
3078 * comment above the shader_select call below). If e.g. GLSL is active, the shader_select call will
3079 * deactivate it.
3080 */
3081 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, desc->shader));
3082 checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, desc->shader)");
3083 priv->current_fprogram_id = desc->shader;
3084
3085 if(device->shader_backend == &arb_program_shader_backend && context->last_was_pshader) {
3086 /* Reload fixed function constants since they collide with the pixel shader constants */
3087 for(i = 0; i < MAX_TEXTURES; i++) {
3088 set_bumpmat_arbfp(STATE_TEXTURESTAGE(i, WINED3DTSS_BUMPENVMAT00), stateblock, context);
3089 }
3090 state_texfactor_arbfp(STATE_RENDER(WINED3DRS_TEXTUREFACTOR), stateblock, context);
3091 state_arb_specularenable(STATE_RENDER(WINED3DRS_SPECULARENABLE), stateblock, context);
3092 }
3093 context->last_was_pshader = FALSE;
3094 } else {
3095 context->last_was_pshader = TRUE;
3096 }
3097
3098 /* Finally, select the shader. If a pixel shader is used, it will be set and enabled by the shader backend.
3099 * If this shader backend is arbfp(most likely), then it will simply overwrite the last fixed function replace-
3100 * ment shader. If the shader backend is not ARB, it currently is important that the opengl implementation
3101 * type overwrites GL_ARB_fragment_program. This is currently the case with GLSL. If we really want to use
3102 * atifs or nvrc pixel shaders with arb fragment programs we'd have to disable GL_FRAGMENT_PROGRAM_ARB here
3103 *
3104 * Don't call shader_select if the vertex shader is dirty, because it will be called later on by the vertex
3105 * shader handler
3106 */
3107 if(!isStateDirty(context, device->StateTable[STATE_VSHADER].representative)) {
3108 device->shader_backend->shader_select((IWineD3DDevice *)stateblock->wineD3DDevice, use_pshader, use_vshader);
3109
3110 if (!isStateDirty(context, STATE_VERTEXSHADERCONSTANT) && (use_vshader || use_pshader)) {
3111 device->StateTable[STATE_VERTEXSHADERCONSTANT].apply(STATE_VERTEXSHADERCONSTANT, stateblock, context);
3112 }
3113 }
3114 if(use_pshader) {
3115 device->StateTable[STATE_PIXELSHADERCONSTANT].apply(STATE_PIXELSHADERCONSTANT, stateblock, context);
3116 }
3117 }
3118
3119 /* We can't link the fog states to the fragment state directly since the vertex pipeline links them
3120 * to FOGENABLE. A different linking in different pipeline parts can't be expressed in the combined
3121 * state table, so we need to handle that with a forwarding function. The other invisible side effect
3122 * is that changing the fog start and fog end(which links to FOGENABLE in vertex) results in the
3123 * fragment_prog_arbfp function being called because FOGENABLE is dirty, which calls this function here
3124 */
3125 static void state_arbfp_fog(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
3126 enum fogsource new_source;
3127
3128 TRACE("state %#x, stateblock %p, context %p\n", state, stateblock, context);
3129
3130 if(!isStateDirty(context, STATE_PIXELSHADER)) {
3131 fragment_prog_arbfp(state, stateblock, context);
3132 }
3133
3134 if(!stateblock->renderState[WINED3DRS_FOGENABLE]) return;
3135
3136 if(stateblock->renderState[WINED3DRS_FOGTABLEMODE] == WINED3DFOG_NONE) {
3137 if(use_vs(stateblock)) {
3138 new_source = FOGSOURCE_VS;
3139 } else {
3140 if(stateblock->renderState[WINED3DRS_FOGVERTEXMODE] == WINED3DFOG_NONE || context->last_was_rhw) {
3141 new_source = FOGSOURCE_COORD;
3142 } else {
3143 new_source = FOGSOURCE_FFP;
3144 }
3145 }
3146 } else {
3147 new_source = FOGSOURCE_FFP;
3148 }
3149 if(new_source != context->fog_source) {
3150 context->fog_source = new_source;
3151 state_fogstartend(STATE_RENDER(WINED3DRS_FOGSTART), stateblock, context);
3152 }
3153 }
3154
3155 static void textransform(DWORD state, IWineD3DStateBlockImpl *stateblock, WineD3DContext *context) {
3156 if(!isStateDirty(context, STATE_PIXELSHADER)) {
3157 fragment_prog_arbfp(state, stateblock, context);
3158 }
3159 }
3160
3161 #undef GLINFO_LOCATION
3162
3163 static const struct StateEntryTemplate arbfp_fragmentstate_template[] = {
3164 {STATE_RENDER(WINED3DRS_TEXTUREFACTOR), { STATE_RENDER(WINED3DRS_TEXTUREFACTOR), state_texfactor_arbfp }, 0 },
3165 {STATE_TEXTURESTAGE(0, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3166 {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3167 {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3168 {STATE_TEXTURESTAGE(0, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3169 {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3170 {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3171 {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3172 {STATE_TEXTURESTAGE(0, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3173 {STATE_TEXTURESTAGE(0, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3174 {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3175 {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3176 {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3177 {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3178 {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3179 {STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(0, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3180 {STATE_TEXTURESTAGE(1, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3181 {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3182 {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3183 {STATE_TEXTURESTAGE(1, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3184 {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3185 {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3186 {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3187 {STATE_TEXTURESTAGE(1, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3188 {STATE_TEXTURESTAGE(1, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3189 {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3190 {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3191 {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3192 {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3193 {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3194 {STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(1, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3195 {STATE_TEXTURESTAGE(2, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3196 {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3197 {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3198 {STATE_TEXTURESTAGE(2, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3199 {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3200 {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3201 {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3202 {STATE_TEXTURESTAGE(2, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3203 {STATE_TEXTURESTAGE(2, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3204 {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3205 {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3206 {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3207 {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3208 {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3209 {STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(2, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3210 {STATE_TEXTURESTAGE(3, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3211 {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3212 {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3213 {STATE_TEXTURESTAGE(3, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3214 {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3215 {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3216 {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3217 {STATE_TEXTURESTAGE(3, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3218 {STATE_TEXTURESTAGE(3, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3219 {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3220 {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3221 {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3222 {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3223 {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3224 {STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(3, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3225 {STATE_TEXTURESTAGE(4, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3226 {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3227 {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3228 {STATE_TEXTURESTAGE(4, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3229 {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3230 {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3231 {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3232 {STATE_TEXTURESTAGE(4, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3233 {STATE_TEXTURESTAGE(4, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3234 {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3235 {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3236 {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3237 {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3238 {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3239 {STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(4, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3240 {STATE_TEXTURESTAGE(5, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3241 {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3242 {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3243 {STATE_TEXTURESTAGE(5, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3244 {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3245 {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3246 {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3247 {STATE_TEXTURESTAGE(5, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3248 {STATE_TEXTURESTAGE(5, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3249 {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3250 {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3251 {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3252 {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3253 {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3254 {STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(5, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3255 {STATE_TEXTURESTAGE(6, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3256 {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3257 {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3258 {STATE_TEXTURESTAGE(6, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3259 {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3260 {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3261 {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3262 {STATE_TEXTURESTAGE(6, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3263 {STATE_TEXTURESTAGE(6, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3264 {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3265 {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3266 {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3267 {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3268 {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3269 {STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(6, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3270 {STATE_TEXTURESTAGE(7, WINED3DTSS_COLOROP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3271 {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3272 {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3273 {STATE_TEXTURESTAGE(7, WINED3DTSS_COLORARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3274 {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAOP), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3275 {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG1), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3276 {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG2), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3277 {STATE_TEXTURESTAGE(7, WINED3DTSS_ALPHAARG0), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3278 {STATE_TEXTURESTAGE(7, WINED3DTSS_RESULTARG), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3279 {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00), { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3280 {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT01), { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3281 {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT10), { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3282 {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT11), { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVMAT00), set_bumpmat_arbfp }, 0 },
3283 {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE), { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3284 {STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLOFFSET), { STATE_TEXTURESTAGE(7, WINED3DTSS_BUMPENVLSCALE), tex_bumpenvlum_arbfp }, 0 },
3285 {STATE_SAMPLER(0), { STATE_SAMPLER(0), sampler_texdim }, 0 },
3286 {STATE_SAMPLER(1), { STATE_SAMPLER(1), sampler_texdim }, 0 },
3287 {STATE_SAMPLER(2), { STATE_SAMPLER(2), sampler_texdim }, 0 },
3288 {STATE_SAMPLER(3), { STATE_SAMPLER(3), sampler_texdim }, 0 },
3289 {STATE_SAMPLER(4), { STATE_SAMPLER(4), sampler_texdim }, 0 },
3290 {STATE_SAMPLER(5), { STATE_SAMPLER(5), sampler_texdim }, 0 },
3291 {STATE_SAMPLER(6), { STATE_SAMPLER(6), sampler_texdim }, 0 },
3292 {STATE_SAMPLER(7), { STATE_SAMPLER(7), sampler_texdim }, 0 },
3293 {STATE_PIXELSHADER, { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3294 {STATE_RENDER(WINED3DRS_FOGENABLE), { STATE_RENDER(WINED3DRS_FOGENABLE), state_arbfp_fog }, 0 },
3295 {STATE_RENDER(WINED3DRS_FOGTABLEMODE), { STATE_RENDER(WINED3DRS_FOGENABLE), state_arbfp_fog }, 0 },
3296 {STATE_RENDER(WINED3DRS_FOGVERTEXMODE), { STATE_RENDER(WINED3DRS_FOGENABLE), state_arbfp_fog }, 0 },
3297 {STATE_RENDER(WINED3DRS_FOGSTART), { STATE_RENDER(WINED3DRS_FOGSTART), state_fogstartend }, 0 },
3298 {STATE_RENDER(WINED3DRS_FOGEND), { STATE_RENDER(WINED3DRS_FOGSTART), state_fogstartend }, 0 },
3299 {STATE_RENDER(WINED3DRS_SRGBWRITEENABLE), { STATE_PIXELSHADER, fragment_prog_arbfp }, 0 },
3300 {STATE_RENDER(WINED3DRS_FOGCOLOR), { STATE_RENDER(WINED3DRS_FOGCOLOR), state_fogcolor }, 0 },
3301 {STATE_RENDER(WINED3DRS_FOGDENSITY), { STATE_RENDER(WINED3DRS_FOGDENSITY), state_fogdensity }, 0 },
3302 {STATE_TEXTURESTAGE(0,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(0, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3303 {STATE_TEXTURESTAGE(1,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(1, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3304 {STATE_TEXTURESTAGE(2,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(2, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3305 {STATE_TEXTURESTAGE(3,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(3, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3306 {STATE_TEXTURESTAGE(4,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(4, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3307 {STATE_TEXTURESTAGE(5,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(5, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3308 {STATE_TEXTURESTAGE(6,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(6, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3309 {STATE_TEXTURESTAGE(7,WINED3DTSS_TEXTURETRANSFORMFLAGS),{STATE_TEXTURESTAGE(7, WINED3DTSS_TEXTURETRANSFORMFLAGS), textransform }, 0 },
3310 {STATE_RENDER(WINED3DRS_SPECULARENABLE), { STATE_RENDER(WINED3DRS_SPECULARENABLE), state_arb_specularenable}, 0 },
3311 {0 /* Terminate */, { 0, 0 }, 0 },
3312 };
3313
3314 const struct fragment_pipeline arbfp_fragment_pipeline = {
3315 arbfp_enable,
3316 arbfp_get_caps,
3317 arbfp_alloc,
3318 arbfp_free,
3319 shader_arb_color_fixup_supported,
3320 arbfp_fragmentstate_template,
3321 TRUE /* We can disable projected textures */
3322 };
3323
3324 #define GLINFO_LOCATION device->adapter->gl_info
3325
3326 struct arbfp_blit_priv {
3327 GLenum yuy2_rect_shader, yuy2_2d_shader;
3328 GLenum uyvy_rect_shader, uyvy_2d_shader;
3329 GLenum yv12_rect_shader, yv12_2d_shader;
3330 };
3331
3332 static HRESULT arbfp_blit_alloc(IWineD3DDevice *iface) {
3333 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3334 device->blit_priv = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(struct arbfp_blit_priv));
3335 if(!device->blit_priv) {
3336 ERR("Out of memory\n");
3337 return E_OUTOFMEMORY;
3338 }
3339 return WINED3D_OK;
3340 }
3341 static void arbfp_blit_free(IWineD3DDevice *iface) {
3342 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3343 struct arbfp_blit_priv *priv = device->blit_priv;
3344
3345 ENTER_GL();
3346 GL_EXTCALL(glDeleteProgramsARB(1, &priv->yuy2_rect_shader));
3347 GL_EXTCALL(glDeleteProgramsARB(1, &priv->yuy2_2d_shader));
3348 GL_EXTCALL(glDeleteProgramsARB(1, &priv->uyvy_rect_shader));
3349 GL_EXTCALL(glDeleteProgramsARB(1, &priv->uyvy_2d_shader));
3350 GL_EXTCALL(glDeleteProgramsARB(1, &priv->yv12_rect_shader));
3351 GL_EXTCALL(glDeleteProgramsARB(1, &priv->yv12_2d_shader));
3352 checkGLcall("Delete yuv programs\n");
3353 LEAVE_GL();
3354 }
3355
3356 static BOOL gen_planar_yuv_read(SHADER_BUFFER *buffer, enum yuv_fixup yuv_fixup, GLenum textype, char *luminance)
3357 {
3358 char chroma;
3359 const char *tex, *texinstr;
3360
3361 if (yuv_fixup == YUV_FIXUP_UYVY) {
3362 chroma = 'x';
3363 *luminance = 'w';
3364 } else {
3365 chroma = 'w';
3366 *luminance = 'x';
3367 }
3368 switch(textype) {
3369 case GL_TEXTURE_2D: tex = "2D"; texinstr = "TXP"; break;
3370 case GL_TEXTURE_RECTANGLE_ARB: tex = "RECT"; texinstr = "TEX"; break;
3371 default:
3372 /* This is more tricky than just replacing the texture type - we have to navigate
3373 * properly in the texture to find the correct chroma values
3374 */
3375 FIXME("Implement yuv correction for non-2d, non-rect textures\n");
3376 return FALSE;
3377 }
3378
3379 /* First we have to read the chroma values. This means we need at least two pixels(no filtering),
3380 * or 4 pixels(with filtering). To get the unmodified chromas, we have to rid ourselves of the
3381 * filtering when we sample the texture.
3382 *
3383 * These are the rules for reading the chroma:
3384 *
3385 * Even pixel: Cr
3386 * Even pixel: U
3387 * Odd pixel: V
3388 *
3389 * So we have to get the sampling x position in non-normalized coordinates in integers
3390 */
3391 if(textype != GL_TEXTURE_RECTANGLE_ARB) {
3392 shader_addline(buffer, "MUL texcrd.xy, fragment.texcoord[0], size.x;\n");
3393 shader_addline(buffer, "MOV texcrd.w, size.x;\n");
3394 } else {
3395 shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3396 }
3397 /* We must not allow filtering between pixel x and x+1, this would mix U and V
3398 * Vertical filtering is ok. However, bear in mind that the pixel center is at
3399 * 0.5, so add 0.5.
3400 */
3401 shader_addline(buffer, "FLR texcrd.x, texcrd.x;\n");
3402 shader_addline(buffer, "ADD texcrd.x, texcrd.x, coef.y;\n");
3403
3404 /* Divide the x coordinate by 0.5 and get the fraction. This gives 0.25 and 0.75 for the
3405 * even and odd pixels respectively
3406 */
3407 shader_addline(buffer, "MUL texcrd2, texcrd, coef.y;\n");
3408 shader_addline(buffer, "FRC texcrd2, texcrd2;\n");
3409
3410 /* Sample Pixel 1 */
3411 shader_addline(buffer, "%s luminance, texcrd, texture[0], %s;\n", texinstr, tex);
3412
3413 /* Put the value into either of the chroma values */
3414 shader_addline(buffer, "SGE temp.x, texcrd2.x, coef.y;\n");
3415 shader_addline(buffer, "MUL chroma.x, luminance.%c, temp.x;\n", chroma);
3416 shader_addline(buffer, "SLT temp.x, texcrd2.x, coef.y;\n");
3417 shader_addline(buffer, "MUL chroma.y, luminance.%c, temp.x;\n", chroma);
3418
3419 /* Sample pixel 2. If we read an even pixel(SLT above returned 1), sample
3420 * the pixel right to the current one. Otherwise, sample the left pixel.
3421 * Bias and scale the SLT result to -1;1 and add it to the texcrd.x.
3422 */
3423 shader_addline(buffer, "MAD temp.x, temp.x, coef.z, -coef.x;\n");
3424 shader_addline(buffer, "ADD texcrd.x, texcrd, temp.x;\n");
3425 shader_addline(buffer, "%s luminance, texcrd, texture[0], %s;\n", texinstr, tex);
3426
3427 /* Put the value into the other chroma */
3428 shader_addline(buffer, "SGE temp.x, texcrd2.x, coef.y;\n");
3429 shader_addline(buffer, "MAD chroma.y, luminance.%c, temp.x, chroma.y;\n", chroma);
3430 shader_addline(buffer, "SLT temp.x, texcrd2.x, coef.y;\n");
3431 shader_addline(buffer, "MAD chroma.x, luminance.%c, temp.x, chroma.x;\n", chroma);
3432
3433 /* TODO: If filtering is enabled, sample a 2nd pair of pixels left or right of
3434 * the current one and lerp the two U and V values
3435 */
3436
3437 /* This gives the correctly filtered luminance value */
3438 shader_addline(buffer, "TEX luminance, fragment.texcoord[0], texture[0], %s;\n", tex);
3439
3440 return TRUE;
3441 }
3442
3443 static BOOL gen_yv12_read(SHADER_BUFFER *buffer, GLenum textype, char *luminance)
3444 {
3445 const char *tex;
3446
3447 switch(textype) {
3448 case GL_TEXTURE_2D: tex = "2D"; break;
3449 case GL_TEXTURE_RECTANGLE_ARB: tex = "RECT"; break;
3450 default:
3451 FIXME("Implement yv12 correction for non-2d, non-rect textures\n");
3452 return FALSE;
3453 }
3454
3455 /* YV12 surfaces contain a WxH sized luminance plane, followed by a (W/2)x(H/2)
3456 * V and a (W/2)x(H/2) U plane, each with 8 bit per pixel. So the effective
3457 * bitdepth is 12 bits per pixel. Since the U and V planes have only half the
3458 * pitch of the luminance plane, the packing into the gl texture is a bit
3459 * unfortunate. If the whole texture is interpreted as luminance data it looks
3460 * approximately like this:
3461 *
3462 * +----------------------------------+----
3463 * | |
3464 * | |
3465 * | |
3466 * | |
3467 * | | 2
3468 * | LUMINANCE | -
3469 * | | 3
3470 * | |
3471 * | |
3472 * | |
3473 * | |
3474 * +----------------+-----------------+----
3475 * | | |
3476 * | U even rows | U odd rows |
3477 * | | | 1
3478 * +----------------+------------------ -
3479 * | | | 3
3480 * | V even rows | V odd rows |
3481 * | | |
3482 * +----------------+-----------------+----
3483 * | | |
3484 * | 0.5 | 0.5 |
3485 *
3486 * So it appears as if there are 4 chroma images, but in fact the odd rows
3487 * in the chroma images are in the same row as the even ones. So its is
3488 * kinda tricky to read
3489 *
3490 * When reading from rectangle textures, keep in mind that the input y coordinates
3491 * go from 0 to d3d_height, whereas the opengl texture height is 1.5 * d3d_height
3492 */
3493 shader_addline(buffer, "PARAM yv12_coef = {%f, %f, %f, %f};\n",
3494 2.0 / 3.0, 1.0 / 6.0, (2.0 / 3.0) + (1.0 / 6.0), 1.0 / 3.0);
3495
3496 shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3497 /* the chroma planes have only half the width */
3498 shader_addline(buffer, "MUL texcrd.x, texcrd.x, coef.y;\n");
3499
3500 /* The first value is between 2/3 and 5/6th of the texture's height, so scale+bias
3501 * the coordinate. Also read the right side of the image when reading odd lines
3502 *
3503 * Don't forget to clamp the y values in into the range, otherwise we'll get filtering
3504 * bleeding
3505 */
3506 if(textype == GL_TEXTURE_2D) {
3507
3508 shader_addline(buffer, "RCP chroma.w, size.y;\n");
3509
3510 shader_addline(buffer, "MUL texcrd2.y, texcrd.y, size.y;\n");
3511
3512 shader_addline(buffer, "FLR texcrd2.y, texcrd2.y;\n");
3513 shader_addline(buffer, "MAD texcrd.y, texcrd.y, yv12_coef.y, yv12_coef.x;\n");
3514
3515 /* Read odd lines from the right side(add size * 0.5 to the x coordinate */
3516 shader_addline(buffer, "ADD texcrd2.x, texcrd2.y, yv12_coef.y;\n"); /* To avoid 0.5 == 0.5 comparisons */
3517 shader_addline(buffer, "FRC texcrd2.x, texcrd2.x;\n");
3518 shader_addline(buffer, "SGE texcrd2.x, texcrd2.x, coef.y;\n");
3519 shader_addline(buffer, "MAD texcrd.x, texcrd2.x, coef.y, texcrd.x;\n");
3520
3521 /* clamp, keep the half pixel origin in mind */
3522 shader_addline(buffer, "MAD temp.y, coef.y, chroma.w, yv12_coef.x;\n");
3523 shader_addline(buffer, "MAX texcrd.y, temp.y, texcrd.y;\n");
3524 shader_addline(buffer, "MAD temp.y, -coef.y, chroma.w, yv12_coef.z;\n");
3525 shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3526 } else {
3527 /* Read from [size - size+size/4] */
3528 shader_addline(buffer, "FLR texcrd.y, texcrd.y;\n");
3529 shader_addline(buffer, "MAD texcrd.y, texcrd.y, coef.w, size.y;\n");
3530
3531 /* Read odd lines from the right side(add size * 0.5 to the x coordinate */
3532 shader_addline(buffer, "ADD texcrd2.x, texcrd.y, yv12_coef.y;\n"); /* To avoid 0.5 == 0.5 comparisons */
3533 shader_addline(buffer, "FRC texcrd2.x, texcrd2.x;\n");
3534 shader_addline(buffer, "SGE texcrd2.x, texcrd2.x, coef.y;\n");
3535 shader_addline(buffer, "MUL texcrd2.x, texcrd2.x, size.x;\n");
3536 shader_addline(buffer, "MAD texcrd.x, texcrd2.x, coef.y, texcrd.x;\n");
3537
3538 /* Make sure to read exactly from the pixel center */
3539 shader_addline(buffer, "FLR texcrd.y, texcrd.y;\n");
3540 shader_addline(buffer, "ADD texcrd.y, texcrd.y, coef.y;\n");
3541
3542 /* Clamp */
3543 shader_addline(buffer, "MAD temp.y, size.y, coef.w, size.y;\n");
3544 shader_addline(buffer, "ADD temp.y, temp.y, -coef.y;\n");
3545 shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3546 shader_addline(buffer, "ADD temp.y, size.y, -coef.y;\n");
3547 shader_addline(buffer, "MAX texcrd.y, temp.y, texcrd.y;\n");
3548 }
3549 /* Read the texture, put the result into the output register */
3550 shader_addline(buffer, "TEX temp, texcrd, texture[0], %s;\n", tex);
3551 shader_addline(buffer, "MOV chroma.x, temp.w;\n");
3552
3553 /* The other chroma value is 1/6th of the texture lower, from 5/6th to 6/6th
3554 * No need to clamp because we're just reusing the already clamped value from above
3555 */
3556 if(textype == GL_TEXTURE_2D) {
3557 shader_addline(buffer, "ADD texcrd.y, texcrd.y, yv12_coef.y;\n");
3558 } else {
3559 shader_addline(buffer, "MAD texcrd.y, size.y, coef.w, texcrd.y;\n");
3560 }
3561 shader_addline(buffer, "TEX temp, texcrd, texture[0], %s;\n", tex);
3562 shader_addline(buffer, "MOV chroma.y, temp.w;\n");
3563
3564 /* Sample the luminance value. It is in the top 2/3rd of the texture, so scale the y coordinate.
3565 * Clamp the y coordinate to prevent the chroma values from bleeding into the sampled luminance
3566 * values due to filtering
3567 */
3568 shader_addline(buffer, "MOV texcrd, fragment.texcoord[0];\n");
3569 if(textype == GL_TEXTURE_2D) {
3570 /* Multiply the y coordinate by 2/3 and clamp it */
3571 shader_addline(buffer, "MUL texcrd.y, texcrd.y, yv12_coef.x;\n");
3572 shader_addline(buffer, "MAD temp.y, -coef.y, chroma.w, yv12_coef.x;\n");
3573 shader_addline(buffer, "MIN texcrd.y, temp.y, texcrd.y;\n");
3574 shader_addline(buffer, "TEX luminance, texcrd, texture[0], %s;\n", tex);
3575 } else {
3576 /* Reading from texture_rectangles is pretty straightforward, just use the unmodified
3577 * texture coordinate. It is still a good idea to clamp it though, since the opengl texture
3578 * is bigger
3579 */
3580 shader_addline(buffer, "ADD temp.x, size.y, -coef.y;\n");
3581 shader_addline(buffer, "MIN texcrd.y, texcrd.y, size.x;\n");
3582 shader_addline(buffer, "TEX luminance, texcrd, texture[0], %s;\n", tex);
3583 }
3584 *luminance = 'a';
3585
3586 return TRUE;
3587 }
3588
3589 static GLuint gen_yuv_shader(IWineD3DDeviceImpl *device, enum yuv_fixup yuv_fixup, GLenum textype)
3590 {
3591 GLenum shader;
3592 SHADER_BUFFER buffer;
3593 char luminance_component;
3594 struct arbfp_blit_priv *priv = device->blit_priv;
3595
3596 /* Shader header */
3597 shader_buffer_init(&buffer);
3598
3599 ENTER_GL();
3600 GL_EXTCALL(glGenProgramsARB(1, &shader));
3601 checkGLcall("GL_EXTCALL(glGenProgramsARB(1, &shader))");
3602 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader));
3603 checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader)");
3604 LEAVE_GL();
3605 if(!shader) {
3606 shader_buffer_free(&buffer);
3607 return 0;
3608 }
3609
3610 /* The YUY2 and UYVY formats contain two pixels packed into a 32 bit macropixel,
3611 * giving effectively 16 bit per pixel. The color consists of a luminance(Y) and
3612 * two chroma(U and V) values. Each macropixel has two luminance values, one for
3613 * each single pixel it contains, and one U and one V value shared between both
3614 * pixels.
3615 *
3616 * The data is loaded into an A8L8 texture. With YUY2, the luminance component
3617 * contains the luminance and alpha the chroma. With UYVY it is vice versa. Thus
3618 * take the format into account when generating the read swizzles
3619 *
3620 * Reading the Y value is straightforward - just sample the texture. The hardware
3621 * takes care of filtering in the horizontal and vertical direction.
3622 *
3623 * Reading the U and V values is harder. We have to avoid filtering horizontally,
3624 * because that would mix the U and V values of one pixel or two adjacent pixels.
3625 * Thus floor the texture coordinate and add 0.5 to get an unfiltered read,
3626 * regardless of the filtering setting. Vertical filtering works automatically
3627 * though - the U and V values of two rows are mixed nicely.
3628 *
3629 * Appart of avoiding filtering issues, the code has to know which value it just
3630 * read, and where it can find the other one. To determine this, it checks if
3631 * it sampled an even or odd pixel, and shifts the 2nd read accordingly.
3632 *
3633 * Handling horizontal filtering of U and V values requires reading a 2nd pair
3634 * of pixels, extracting U and V and mixing them. This is not implemented yet.
3635 *
3636 * An alternative implementation idea is to load the texture as A8R8G8B8 texture,
3637 * with width / 2. This way one read gives all 3 values, finding U and V is easy
3638 * in an unfiltered situation. Finding the luminance on the other hand requires
3639 * finding out if it is an odd or even pixel. The real drawback of this approach
3640 * is filtering. This would have to be emulated completely in the shader, reading
3641 * up two 2 packed pixels in up to 2 rows and interpolating both horizontally and
3642 * vertically. Beyond that it would require adjustments to the texture handling
3643 * code to deal with the width scaling
3644 */
3645 shader_addline(&buffer, "!!ARBfp1.0\n");
3646 shader_addline(&buffer, "TEMP luminance;\n");
3647 shader_addline(&buffer, "TEMP temp;\n");
3648 shader_addline(&buffer, "TEMP chroma;\n");
3649 shader_addline(&buffer, "TEMP texcrd;\n");
3650 shader_addline(&buffer, "TEMP texcrd2;\n");
3651 shader_addline(&buffer, "PARAM coef = {1.0, 0.5, 2.0, 0.25};\n");
3652 shader_addline(&buffer, "PARAM yuv_coef = {1.403, 0.344, 0.714, 1.770};\n");
3653 shader_addline(&buffer, "PARAM size = program.local[0];\n");
3654
3655 switch (yuv_fixup)
3656 {
3657 case YUV_FIXUP_UYVY:
3658 case YUV_FIXUP_YUY2:
3659 if (!gen_planar_yuv_read(&buffer, yuv_fixup, textype, &luminance_component))
3660 {
3661 shader_buffer_free(&buffer);
3662 return 0;
3663 }
3664 break;
3665
3666 case YUV_FIXUP_YV12:
3667 if (!gen_yv12_read(&buffer, textype, &luminance_component))
3668 {
3669 shader_buffer_free(&buffer);
3670 return 0;
3671 }
3672 break;
3673
3674 default:
3675 FIXME("Unsupported YUV fixup %#x\n", yuv_fixup);
3676 shader_buffer_free(&buffer);
3677 return 0;
3678 }
3679
3680 /* Calculate the final result. Formula is taken from
3681 * http://www.fourcc.org/fccyvrgb.php. Note that the chroma
3682 * ranges from -0.5 to 0.5
3683 */
3684 shader_addline(&buffer, "SUB chroma.xy, chroma, coef.y;\n");
3685
3686 shader_addline(&buffer, "MAD result.color.x, chroma.x, yuv_coef.x, luminance.%c;\n", luminance_component);
3687 shader_addline(&buffer, "MAD temp.x, -chroma.y, yuv_coef.y, luminance.%c;\n", luminance_component);
3688 shader_addline(&buffer, "MAD result.color.y, -chroma.x, yuv_coef.z, temp.x;\n");
3689 shader_addline(&buffer, "MAD result.color.z, chroma.y, yuv_coef.w, luminance.%c;\n", luminance_component);
3690 shader_addline(&buffer, "END\n");
3691
3692 ENTER_GL();
3693 GL_EXTCALL(glProgramStringARB(GL_FRAGMENT_PROGRAM_ARB, GL_PROGRAM_FORMAT_ASCII_ARB, strlen(buffer.buffer), buffer.buffer));
3694
3695 if (glGetError() == GL_INVALID_OPERATION) {
3696 GLint pos;
3697 glGetIntegerv(GL_PROGRAM_ERROR_POSITION_ARB, &pos);
3698 FIXME("Fragment program error at position %d: %s\n", pos,
3699 debugstr_a((const char *)glGetString(GL_PROGRAM_ERROR_STRING_ARB)));
3700 }
3701 shader_buffer_free(&buffer);
3702 LEAVE_GL();
3703
3704 switch (yuv_fixup)
3705 {
3706 case YUV_FIXUP_YUY2:
3707 if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->yuy2_rect_shader = shader;
3708 else priv->yuy2_2d_shader = shader;
3709 break;
3710
3711 case YUV_FIXUP_UYVY:
3712 if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->uyvy_rect_shader = shader;
3713 else priv->uyvy_2d_shader = shader;
3714 break;
3715
3716 case YUV_FIXUP_YV12:
3717 if (textype == GL_TEXTURE_RECTANGLE_ARB) priv->yv12_rect_shader = shader;
3718 else priv->yv12_2d_shader = shader;
3719 break;
3720 }
3721
3722 return shader;
3723 }
3724
3725 static HRESULT arbfp_blit_set(IWineD3DDevice *iface, const struct GlPixelFormatDesc *format_desc,
3726 GLenum textype, UINT width, UINT height)
3727 {
3728 GLenum shader;
3729 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3730 float size[4] = {width, height, 1, 1};
3731 struct arbfp_blit_priv *priv = device->blit_priv;
3732 enum yuv_fixup yuv_fixup;
3733
3734 if (!is_yuv_fixup(format_desc->color_fixup))
3735 {
3736 TRACE("Fixup:\n");
3737 dump_color_fixup_desc(format_desc->color_fixup);
3738 /* Don't bother setting up a shader for unconverted formats */
3739 ENTER_GL();
3740 glEnable(textype);
3741 checkGLcall("glEnable(textype)");
3742 LEAVE_GL();
3743 return WINED3D_OK;
3744 }
3745
3746 yuv_fixup = get_yuv_fixup(format_desc->color_fixup);
3747
3748 switch(yuv_fixup)
3749 {
3750 case YUV_FIXUP_YUY2:
3751 shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->yuy2_rect_shader : priv->yuy2_2d_shader;
3752 break;
3753
3754 case YUV_FIXUP_UYVY:
3755 shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->uyvy_rect_shader : priv->uyvy_2d_shader;
3756 break;
3757
3758 case YUV_FIXUP_YV12:
3759 shader = textype == GL_TEXTURE_RECTANGLE_ARB ? priv->yv12_rect_shader : priv->yv12_2d_shader;
3760 break;
3761
3762 default:
3763 FIXME("Unsupported YUV fixup %#x, not setting a shader\n", yuv_fixup);
3764 ENTER_GL();
3765 glEnable(textype);
3766 checkGLcall("glEnable(textype)");
3767 LEAVE_GL();
3768 return E_NOTIMPL;
3769 }
3770
3771 if (!shader) shader = gen_yuv_shader(device, yuv_fixup, textype);
3772
3773 ENTER_GL();
3774 glEnable(GL_FRAGMENT_PROGRAM_ARB);
3775 checkGLcall("glEnable(GL_FRAGMENT_PROGRAM_ARB)");
3776 GL_EXTCALL(glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader));
3777 checkGLcall("glBindProgramARB(GL_FRAGMENT_PROGRAM_ARB, shader)");
3778 GL_EXTCALL(glProgramLocalParameter4fvARB(GL_FRAGMENT_PROGRAM_ARB, 0, size));
3779 checkGLcall("glProgramLocalParameter4fvARB");
3780 LEAVE_GL();
3781
3782 return WINED3D_OK;
3783 }
3784
3785 static void arbfp_blit_unset(IWineD3DDevice *iface) {
3786 IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) iface;
3787
3788 ENTER_GL();
3789 glDisable(GL_FRAGMENT_PROGRAM_ARB);
3790 checkGLcall("glDisable(GL_FRAGMENT_PROGRAM_ARB)");
3791 glDisable(GL_TEXTURE_2D);
3792 checkGLcall("glDisable(GL_TEXTURE_2D)");
3793 if(GL_SUPPORT(ARB_TEXTURE_CUBE_MAP)) {
3794 glDisable(GL_TEXTURE_CUBE_MAP_ARB);
3795 checkGLcall("glDisable(GL_TEXTURE_CUBE_MAP_ARB)");
3796 }
3797 if(GL_SUPPORT(ARB_TEXTURE_RECTANGLE)) {
3798 glDisable(GL_TEXTURE_RECTANGLE_ARB);
3799 checkGLcall("glDisable(GL_TEXTURE_RECTANGLE_ARB)");
3800 }
3801 LEAVE_GL();
3802 }
3803
3804 static BOOL arbfp_blit_color_fixup_supported(struct color_fixup_desc fixup)
3805 {
3806 enum yuv_fixup yuv_fixup;
3807
3808 if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
3809 {
3810 TRACE("Checking support for fixup:\n");
3811 dump_color_fixup_desc(fixup);
3812 }
3813
3814 if (is_identity_fixup(fixup))
3815 {
3816 TRACE("[OK]\n");
3817 return TRUE;
3818 }
3819
3820 /* We only support YUV conversions. */
3821 if (!is_yuv_fixup(fixup))
3822 {
3823 TRACE("[FAILED]\n");
3824 return FALSE;
3825 }
3826
3827 yuv_fixup = get_yuv_fixup(fixup);
3828 switch(yuv_fixup)
3829 {
3830 case YUV_FIXUP_YUY2:
3831 case YUV_FIXUP_UYVY:
3832 case YUV_FIXUP_YV12:
3833 TRACE("[OK]\n");
3834 return TRUE;
3835
3836 default:
3837 FIXME("Unsupported YUV fixup %#x\n", yuv_fixup);
3838 TRACE("[FAILED]\n");
3839 return FALSE;
3840 }
3841 }
3842
3843 const struct blit_shader arbfp_blit = {
3844 arbfp_blit_alloc,
3845 arbfp_blit_free,
3846 arbfp_blit_set,
3847 arbfp_blit_unset,
3848 arbfp_blit_color_fixup_supported,
3849 };
3850
3851 #undef GLINFO_LOCATION
useful hacks and other patches not (yet) suitable for mainline