1 | /*
|
---|
2 | * shaders implementation
|
---|
3 | *
|
---|
4 | * Copyright 2002-2003 Jason Edmeades
|
---|
5 | * Copyright 2002-2003 Raphael Junqueira
|
---|
6 | * Copyright 2004 Christian Costa
|
---|
7 | * Copyright 2005 Oliver Stieber
|
---|
8 | * Copyright 2006 Ivan Gyurdiev
|
---|
9 | * Copyright 2007-2008 Stefan Dösinger for CodeWeavers
|
---|
10 | *
|
---|
11 | * This library is free software; you can redistribute it and/or
|
---|
12 | * modify it under the terms of the GNU Lesser General Public
|
---|
13 | * License as published by the Free Software Foundation; either
|
---|
14 | * version 2.1 of the License, or (at your option) any later version.
|
---|
15 | *
|
---|
16 | * This library is distributed in the hope that it will be useful,
|
---|
17 | * but WITHOUT ANY WARRANTY; without even the implied warranty of
|
---|
18 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
---|
19 | * Lesser General Public License for more details.
|
---|
20 | *
|
---|
21 | * You should have received a copy of the GNU Lesser General Public
|
---|
22 | * License along with this library; if not, write to the Free Software
|
---|
23 | * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
|
---|
24 | */
|
---|
25 |
|
---|
26 | /*
|
---|
27 | * Sun LGPL Disclaimer: For the avoidance of doubt, except that if any license choice
|
---|
28 | * other than GPL or LGPL is available it will apply instead, Sun elects to use only
|
---|
29 | * the Lesser General Public License version 2.1 (LGPLv2) at this time for any software where
|
---|
30 | * a choice of LGPL license versions is made available with the language indicating
|
---|
31 | * that LGPLv2 or any later version may be used, or where a choice of which version
|
---|
32 | * of the LGPL is applied is otherwise unspecified.
|
---|
33 | */
|
---|
34 |
|
---|
35 | #include "config.h"
|
---|
36 | #include <string.h>
|
---|
37 | #include <stdio.h>
|
---|
38 | #include "wined3d_private.h"
|
---|
39 |
|
---|
40 | WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
|
---|
41 | WINE_DECLARE_DEBUG_CHANNEL(d3d);
|
---|
42 |
|
---|
43 | #define GLNAME_REQUIRE_GLSL ((const char *)1)
|
---|
44 |
|
---|
45 | static void shader_dump_param(const DWORD param, const DWORD addr_token, int input, DWORD shader_version);
|
---|
46 |
|
---|
47 | static inline BOOL shader_is_version_token(DWORD token) {
|
---|
48 | return shader_is_pshader_version(token) ||
|
---|
49 | shader_is_vshader_version(token);
|
---|
50 | }
|
---|
51 |
|
---|
52 | int shader_addline(
|
---|
53 | SHADER_BUFFER* buffer,
|
---|
54 | const char *format, ...) {
|
---|
55 |
|
---|
56 | char* base = buffer->buffer + buffer->bsize;
|
---|
57 | int rc;
|
---|
58 |
|
---|
59 | va_list args;
|
---|
60 | va_start(args, format);
|
---|
61 | rc = _vsnprintf(base, SHADER_PGMSIZE - 1 - buffer->bsize, format, args);
|
---|
62 | va_end(args);
|
---|
63 |
|
---|
64 | if (rc < 0 || /* C89 */
|
---|
65 | rc > SHADER_PGMSIZE - 1 - buffer->bsize) { /* C99 */
|
---|
66 |
|
---|
67 | ERR("The buffer allocated for the shader program string "
|
---|
68 | "is too small at %d bytes.\n", SHADER_PGMSIZE);
|
---|
69 | buffer->bsize = SHADER_PGMSIZE - 1;
|
---|
70 | return -1;
|
---|
71 | }
|
---|
72 |
|
---|
73 | if (buffer->newline) {
|
---|
74 | TRACE("GL HW (%u, %u) : %s", buffer->lineNo + 1, buffer->bsize, base);
|
---|
75 | buffer->newline = FALSE;
|
---|
76 | } else {
|
---|
77 | TRACE("%s", base);
|
---|
78 | }
|
---|
79 |
|
---|
80 | buffer->bsize += rc;
|
---|
81 | if (buffer->buffer[buffer->bsize-1] == '\n') {
|
---|
82 | buffer->lineNo++;
|
---|
83 | buffer->newline = TRUE;
|
---|
84 | }
|
---|
85 | return 0;
|
---|
86 | }
|
---|
87 |
|
---|
88 | const SHADER_OPCODE *shader_get_opcode(const SHADER_OPCODE *opcode_table, DWORD shader_version, DWORD code)
|
---|
89 | {
|
---|
90 | DWORD i = 0;
|
---|
91 |
|
---|
92 | /** TODO: use dichotomic search */
|
---|
93 | while (opcode_table[i].name)
|
---|
94 | {
|
---|
95 | if ((code & WINED3DSI_OPCODE_MASK) == opcode_table[i].opcode
|
---|
96 | && shader_version >= opcode_table[i].min_version
|
---|
97 | && (!opcode_table[i].max_version || shader_version <= opcode_table[i].max_version))
|
---|
98 | {
|
---|
99 | return &opcode_table[i];
|
---|
100 | }
|
---|
101 | ++i;
|
---|
102 | }
|
---|
103 |
|
---|
104 | FIXME("Unsupported opcode %#x(%d) masked %#x, shader version %#x\n",
|
---|
105 | code, code, code & WINED3DSI_OPCODE_MASK, shader_version);
|
---|
106 |
|
---|
107 | return NULL;
|
---|
108 | }
|
---|
109 |
|
---|
110 | /* Read a parameter opcode from the input stream,
|
---|
111 | * and possibly a relative addressing token.
|
---|
112 | * Return the number of tokens read */
|
---|
113 | static int shader_get_param(IWineD3DBaseShader *iface, const DWORD *pToken, DWORD *param, DWORD *addr_token)
|
---|
114 | {
|
---|
115 | /* PS >= 3.0 have relative addressing (with token)
|
---|
116 | * VS >= 2.0 have relative addressing (with token)
|
---|
117 | * VS >= 1.0 < 2.0 have relative addressing (without token)
|
---|
118 | * The version check below should work in general */
|
---|
119 |
|
---|
120 | IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
|
---|
121 | char rel_token = WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 2 &&
|
---|
122 | ((*pToken & WINED3DSHADER_ADDRESSMODE_MASK) == WINED3DSHADER_ADDRMODE_RELATIVE);
|
---|
123 |
|
---|
124 | *param = *pToken;
|
---|
125 | *addr_token = rel_token? *(pToken + 1): 0;
|
---|
126 | return rel_token? 2:1;
|
---|
127 | }
|
---|
128 |
|
---|
129 | /* Return the number of parameters to skip for an opcode */
|
---|
130 | static inline int shader_skip_opcode(
|
---|
131 | IWineD3DBaseShaderImpl* This,
|
---|
132 | const SHADER_OPCODE* curOpcode,
|
---|
133 | DWORD opcode_token) {
|
---|
134 |
|
---|
135 | /* Shaders >= 2.0 may contain address tokens, but fortunately they
|
---|
136 | * have a useful length mask - use it here. Shaders 1.0 contain no such tokens */
|
---|
137 |
|
---|
138 | return (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) >= 2)?
|
---|
139 | ((opcode_token & WINED3DSI_INSTLENGTH_MASK) >> WINED3DSI_INSTLENGTH_SHIFT):
|
---|
140 | curOpcode->num_params;
|
---|
141 | }
|
---|
142 |
|
---|
143 | /* Read the parameters of an unrecognized opcode from the input stream
|
---|
144 | * Return the number of tokens read.
|
---|
145 | *
|
---|
146 | * Note: This function assumes source or destination token format.
|
---|
147 | * It will not work with specially-formatted tokens like DEF or DCL,
|
---|
148 | * but hopefully those would be recognized */
|
---|
149 | static int shader_skip_unrecognized(IWineD3DBaseShader *iface, const DWORD *pToken)
|
---|
150 | {
|
---|
151 | DWORD shader_version = ((IWineD3DBaseShaderImpl *)iface)->baseShader.hex_version;
|
---|
152 | int tokens_read = 0;
|
---|
153 | int i = 0;
|
---|
154 |
|
---|
155 | /* TODO: Think of a good name for 0x80000000 and replace it with a constant */
|
---|
156 | while (*pToken & 0x80000000) {
|
---|
157 |
|
---|
158 | DWORD param, addr_token;
|
---|
159 | tokens_read += shader_get_param(iface, pToken, ¶m, &addr_token);
|
---|
160 | pToken += tokens_read;
|
---|
161 |
|
---|
162 | FIXME("Unrecognized opcode param: token=0x%08x "
|
---|
163 | "addr_token=0x%08x name=", param, addr_token);
|
---|
164 | shader_dump_param(param, addr_token, i, shader_version);
|
---|
165 | FIXME("\n");
|
---|
166 | ++i;
|
---|
167 | }
|
---|
168 | return tokens_read;
|
---|
169 | }
|
---|
170 |
|
---|
171 | /* Convert floating point offset relative
|
---|
172 | * to a register file to an absolute offset for float constants */
|
---|
173 | static unsigned int shader_get_float_offset(const DWORD reg)
|
---|
174 | {
|
---|
175 | unsigned int regnum = reg & WINED3DSP_REGNUM_MASK;
|
---|
176 | int regtype = shader_get_regtype(reg);
|
---|
177 |
|
---|
178 | switch (regtype) {
|
---|
179 | case WINED3DSPR_CONST: return regnum;
|
---|
180 | case WINED3DSPR_CONST2: return 2048 + regnum;
|
---|
181 | case WINED3DSPR_CONST3: return 4096 + regnum;
|
---|
182 | case WINED3DSPR_CONST4: return 6144 + regnum;
|
---|
183 | default:
|
---|
184 | FIXME("Unsupported register type: %d\n", regtype);
|
---|
185 | return regnum;
|
---|
186 | }
|
---|
187 | }
|
---|
188 |
|
---|
189 | static void shader_delete_constant_list(struct list* clist) {
|
---|
190 |
|
---|
191 | struct list *ptr;
|
---|
192 | struct local_constant* constant;
|
---|
193 |
|
---|
194 | ptr = list_head(clist);
|
---|
195 | while (ptr) {
|
---|
196 | constant = LIST_ENTRY(ptr, struct local_constant, entry);
|
---|
197 | ptr = list_next(clist, ptr);
|
---|
198 | HeapFree(GetProcessHeap(), 0, constant);
|
---|
199 | }
|
---|
200 | list_init(clist);
|
---|
201 | }
|
---|
202 |
|
---|
203 | /* Note that this does not count the loop register
|
---|
204 | * as an address register. */
|
---|
205 |
|
---|
206 | HRESULT shader_get_registers_used(
|
---|
207 | IWineD3DBaseShader *iface,
|
---|
208 | shader_reg_maps* reg_maps,
|
---|
209 | semantic* semantics_in,
|
---|
210 | semantic* semantics_out,
|
---|
211 | CONST DWORD* pToken,
|
---|
212 | IWineD3DStateBlockImpl *stateBlock) {
|
---|
213 |
|
---|
214 | IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
|
---|
215 | const SHADER_OPCODE *shader_ins = This->baseShader.shader_ins;
|
---|
216 | DWORD shader_version = This->baseShader.hex_version;
|
---|
217 | unsigned int cur_loop_depth = 0, max_loop_depth = 0;
|
---|
218 |
|
---|
219 | /* There are some minor differences between pixel and vertex shaders */
|
---|
220 | char pshader = shader_is_pshader_version(This->baseShader.hex_version);
|
---|
221 |
|
---|
222 | memset(reg_maps->bumpmat, 0, sizeof(reg_maps->bumpmat));
|
---|
223 | memset(reg_maps->luminanceparams, 0, sizeof(reg_maps->luminanceparams));
|
---|
224 |
|
---|
225 | if (pToken == NULL)
|
---|
226 | return WINED3D_OK;
|
---|
227 |
|
---|
228 | /* get_registers_used is called on every compile on some 1.x shaders, which can result
|
---|
229 | * in stacking up a collection of local constants. Delete the old constants if existing
|
---|
230 | */
|
---|
231 | shader_delete_constant_list(&This->baseShader.constantsF);
|
---|
232 | shader_delete_constant_list(&This->baseShader.constantsB);
|
---|
233 | shader_delete_constant_list(&This->baseShader.constantsI);
|
---|
234 |
|
---|
235 | while (WINED3DVS_END() != *pToken) {
|
---|
236 | CONST SHADER_OPCODE* curOpcode;
|
---|
237 | DWORD opcode_token;
|
---|
238 |
|
---|
239 | /* Skip version */
|
---|
240 | if (shader_is_version_token(*pToken)) {
|
---|
241 | ++pToken;
|
---|
242 | continue;
|
---|
243 |
|
---|
244 | /* Skip comments */
|
---|
245 | } else if (shader_is_comment(*pToken)) {
|
---|
246 | DWORD comment_len = (*pToken & WINED3DSI_COMMENTSIZE_MASK) >> WINED3DSI_COMMENTSIZE_SHIFT;
|
---|
247 | ++pToken;
|
---|
248 | pToken += comment_len;
|
---|
249 | continue;
|
---|
250 | }
|
---|
251 |
|
---|
252 | /* Fetch opcode */
|
---|
253 | opcode_token = *pToken++;
|
---|
254 | curOpcode = shader_get_opcode(shader_ins, shader_version, opcode_token);
|
---|
255 |
|
---|
256 | /* Unhandled opcode, and its parameters */
|
---|
257 | if (NULL == curOpcode) {
|
---|
258 | while (*pToken & 0x80000000)
|
---|
259 | ++pToken;
|
---|
260 |
|
---|
261 | /* Handle declarations */
|
---|
262 | } else if (WINED3DSIO_DCL == curOpcode->opcode) {
|
---|
263 |
|
---|
264 | DWORD usage = *pToken++;
|
---|
265 | DWORD param = *pToken++;
|
---|
266 | DWORD regtype = shader_get_regtype(param);
|
---|
267 | unsigned int regnum = param & WINED3DSP_REGNUM_MASK;
|
---|
268 |
|
---|
269 | /* Vshader: mark attributes used
|
---|
270 | Pshader: mark 3.0 input registers used, save token */
|
---|
271 | if (WINED3DSPR_INPUT == regtype) {
|
---|
272 |
|
---|
273 | if (!pshader)
|
---|
274 | reg_maps->attributes[regnum] = 1;
|
---|
275 | else
|
---|
276 | reg_maps->packed_input[regnum] = 1;
|
---|
277 |
|
---|
278 | semantics_in[regnum].usage = usage;
|
---|
279 | semantics_in[regnum].reg = param;
|
---|
280 |
|
---|
281 | /* Vshader: mark 3.0 output registers used, save token */
|
---|
282 | } else if (WINED3DSPR_OUTPUT == regtype) {
|
---|
283 | reg_maps->packed_output[regnum] = 1;
|
---|
284 | semantics_out[regnum].usage = usage;
|
---|
285 | semantics_out[regnum].reg = param;
|
---|
286 | if (usage & (WINED3DDECLUSAGE_FOG << WINED3DSP_DCL_USAGE_SHIFT))
|
---|
287 | reg_maps->fog = 1;
|
---|
288 |
|
---|
289 | /* Save sampler usage token */
|
---|
290 | } else if (WINED3DSPR_SAMPLER == regtype)
|
---|
291 | reg_maps->samplers[regnum] = usage;
|
---|
292 |
|
---|
293 | } else if (WINED3DSIO_DEF == curOpcode->opcode) {
|
---|
294 |
|
---|
295 | local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
|
---|
296 | if (!lconst) return E_OUTOFMEMORY;
|
---|
297 | lconst->idx = *pToken & WINED3DSP_REGNUM_MASK;
|
---|
298 | memcpy(lconst->value, pToken + 1, 4 * sizeof(DWORD));
|
---|
299 |
|
---|
300 | /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
|
---|
301 | if(WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) == 1 && pshader) {
|
---|
302 | float *value = (float *) lconst->value;
|
---|
303 | if(value[0] < -1.0) value[0] = -1.0;
|
---|
304 | else if(value[0] > 1.0) value[0] = 1.0;
|
---|
305 | if(value[1] < -1.0) value[1] = -1.0;
|
---|
306 | else if(value[1] > 1.0) value[1] = 1.0;
|
---|
307 | if(value[2] < -1.0) value[2] = -1.0;
|
---|
308 | else if(value[2] > 1.0) value[2] = 1.0;
|
---|
309 | if(value[3] < -1.0) value[3] = -1.0;
|
---|
310 | else if(value[3] > 1.0) value[3] = 1.0;
|
---|
311 | }
|
---|
312 |
|
---|
313 | list_add_head(&This->baseShader.constantsF, &lconst->entry);
|
---|
314 | pToken += curOpcode->num_params;
|
---|
315 |
|
---|
316 | } else if (WINED3DSIO_DEFI == curOpcode->opcode) {
|
---|
317 |
|
---|
318 | local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
|
---|
319 | if (!lconst) return E_OUTOFMEMORY;
|
---|
320 | lconst->idx = *pToken & WINED3DSP_REGNUM_MASK;
|
---|
321 | memcpy(lconst->value, pToken + 1, 4 * sizeof(DWORD));
|
---|
322 | list_add_head(&This->baseShader.constantsI, &lconst->entry);
|
---|
323 | pToken += curOpcode->num_params;
|
---|
324 |
|
---|
325 | } else if (WINED3DSIO_DEFB == curOpcode->opcode) {
|
---|
326 |
|
---|
327 | local_constant* lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(local_constant));
|
---|
328 | if (!lconst) return E_OUTOFMEMORY;
|
---|
329 | lconst->idx = *pToken & WINED3DSP_REGNUM_MASK;
|
---|
330 | memcpy(lconst->value, pToken + 1, 1 * sizeof(DWORD));
|
---|
331 | list_add_head(&This->baseShader.constantsB, &lconst->entry);
|
---|
332 | pToken += curOpcode->num_params;
|
---|
333 |
|
---|
334 | /* If there's a loop in the shader */
|
---|
335 | } else if (WINED3DSIO_LOOP == curOpcode->opcode ||
|
---|
336 | WINED3DSIO_REP == curOpcode->opcode) {
|
---|
337 | cur_loop_depth++;
|
---|
338 | if(cur_loop_depth > max_loop_depth)
|
---|
339 | max_loop_depth = cur_loop_depth;
|
---|
340 | pToken += curOpcode->num_params;
|
---|
341 |
|
---|
342 | } else if (WINED3DSIO_ENDLOOP == curOpcode->opcode ||
|
---|
343 | WINED3DSIO_ENDREP == curOpcode->opcode) {
|
---|
344 | cur_loop_depth--;
|
---|
345 |
|
---|
346 | /* For subroutine prototypes */
|
---|
347 | } else if (WINED3DSIO_LABEL == curOpcode->opcode) {
|
---|
348 |
|
---|
349 | DWORD snum = *pToken & WINED3DSP_REGNUM_MASK;
|
---|
350 | reg_maps->labels[snum] = 1;
|
---|
351 | pToken += curOpcode->num_params;
|
---|
352 |
|
---|
353 | /* Set texture, address, temporary registers */
|
---|
354 | } else {
|
---|
355 | int i, limit;
|
---|
356 |
|
---|
357 | /* Declare 1.X samplers implicitly, based on the destination reg. number */
|
---|
358 | if (WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) == 1 &&
|
---|
359 | pshader /* Filter different instructions with the same enum values in VS */ &&
|
---|
360 | (WINED3DSIO_TEX == curOpcode->opcode ||
|
---|
361 | WINED3DSIO_TEXBEM == curOpcode->opcode ||
|
---|
362 | WINED3DSIO_TEXBEML == curOpcode->opcode ||
|
---|
363 | WINED3DSIO_TEXDP3TEX == curOpcode->opcode ||
|
---|
364 | WINED3DSIO_TEXM3x2TEX == curOpcode->opcode ||
|
---|
365 | WINED3DSIO_TEXM3x3SPEC == curOpcode->opcode ||
|
---|
366 | WINED3DSIO_TEXM3x3TEX == curOpcode->opcode ||
|
---|
367 | WINED3DSIO_TEXM3x3VSPEC == curOpcode->opcode ||
|
---|
368 | WINED3DSIO_TEXREG2AR == curOpcode->opcode ||
|
---|
369 | WINED3DSIO_TEXREG2GB == curOpcode->opcode ||
|
---|
370 | WINED3DSIO_TEXREG2RGB == curOpcode->opcode)) {
|
---|
371 |
|
---|
372 | /* Fake sampler usage, only set reserved bit and ttype */
|
---|
373 | DWORD sampler_code = *pToken & WINED3DSP_REGNUM_MASK;
|
---|
374 |
|
---|
375 | if(!stateBlock->textures[sampler_code]) {
|
---|
376 | ERR("No texture bound to sampler %d\n", sampler_code);
|
---|
377 | reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
|
---|
378 | } else {
|
---|
379 | int texType = IWineD3DBaseTexture_GetTextureDimensions(stateBlock->textures[sampler_code]);
|
---|
380 | switch(texType) {
|
---|
381 | /* We have to select between texture rectangles and 2D textures later because 2.0 and
|
---|
382 | * 3.0 shaders only have WINED3DSTT_2D as well
|
---|
383 | */
|
---|
384 | case GL_TEXTURE_RECTANGLE_ARB:
|
---|
385 | case GL_TEXTURE_2D:
|
---|
386 | reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
|
---|
387 | break;
|
---|
388 |
|
---|
389 | case GL_TEXTURE_3D:
|
---|
390 | reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_VOLUME;
|
---|
391 | break;
|
---|
392 |
|
---|
393 | case GL_TEXTURE_CUBE_MAP_ARB:
|
---|
394 | reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_CUBE;
|
---|
395 | break;
|
---|
396 |
|
---|
397 | default:
|
---|
398 | ERR("Unexpected gl texture type found: %d\n", texType);
|
---|
399 | reg_maps->samplers[sampler_code] = (0x1 << 31) | WINED3DSTT_2D;
|
---|
400 | }
|
---|
401 | }
|
---|
402 |
|
---|
403 | /* texbem is only valid with < 1.4 pixel shaders */
|
---|
404 | if(WINED3DSIO_TEXBEM == curOpcode->opcode ||
|
---|
405 | WINED3DSIO_TEXBEML == curOpcode->opcode) {
|
---|
406 | reg_maps->bumpmat[sampler_code] = TRUE;
|
---|
407 | if(WINED3DSIO_TEXBEML == curOpcode->opcode) {
|
---|
408 | reg_maps->luminanceparams[sampler_code] = TRUE;
|
---|
409 | }
|
---|
410 | }
|
---|
411 | }
|
---|
412 | if(WINED3DSIO_NRM == curOpcode->opcode) {
|
---|
413 | reg_maps->usesnrm = 1;
|
---|
414 | } else if(WINED3DSIO_BEM == curOpcode->opcode && pshader) {
|
---|
415 | DWORD regnum = *pToken & WINED3DSP_REGNUM_MASK;
|
---|
416 | reg_maps->bumpmat[regnum] = TRUE;
|
---|
417 | } else if(WINED3DSIO_DSY == curOpcode->opcode) {
|
---|
418 | reg_maps->usesdsy = 1;
|
---|
419 | }
|
---|
420 |
|
---|
421 | /* This will loop over all the registers and try to
|
---|
422 | * make a bitmask of the ones we're interested in.
|
---|
423 | *
|
---|
424 | * Relative addressing tokens are ignored, but that's
|
---|
425 | * okay, since we'll catch any address registers when
|
---|
426 | * they are initialized (required by spec) */
|
---|
427 |
|
---|
428 | limit = (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED)?
|
---|
429 | curOpcode->num_params + 1: curOpcode->num_params;
|
---|
430 |
|
---|
431 | for (i = 0; i < limit; ++i) {
|
---|
432 |
|
---|
433 | DWORD param, addr_token, reg, regtype;
|
---|
434 | pToken += shader_get_param(iface, pToken, ¶m, &addr_token);
|
---|
435 |
|
---|
436 | regtype = shader_get_regtype(param);
|
---|
437 | reg = param & WINED3DSP_REGNUM_MASK;
|
---|
438 |
|
---|
439 | if (WINED3DSPR_TEXTURE == regtype) { /* vs: WINED3DSPR_ADDR */
|
---|
440 |
|
---|
441 | if (pshader)
|
---|
442 | reg_maps->texcoord[reg] = 1;
|
---|
443 | else
|
---|
444 | reg_maps->address[reg] = 1;
|
---|
445 | }
|
---|
446 |
|
---|
447 | else if (WINED3DSPR_TEMP == regtype)
|
---|
448 | reg_maps->temporary[reg] = 1;
|
---|
449 |
|
---|
450 | else if (WINED3DSPR_INPUT == regtype) {
|
---|
451 | if( !pshader)
|
---|
452 | reg_maps->attributes[reg] = 1;
|
---|
453 | else {
|
---|
454 | if(param & WINED3DSHADER_ADDRMODE_RELATIVE) {
|
---|
455 | /* If relative addressing is used, we must assume that all registers
|
---|
456 | * are used. Even if it is a construct like v3[aL], we can't assume
|
---|
457 | * that v0, v1 and v2 aren't read because aL can be negative
|
---|
458 | */
|
---|
459 | unsigned int i;
|
---|
460 | for(i = 0; i < MAX_REG_INPUT; i++) {
|
---|
461 | ((IWineD3DPixelShaderImpl *) This)->input_reg_used[i] = TRUE;
|
---|
462 | }
|
---|
463 | } else {
|
---|
464 | ((IWineD3DPixelShaderImpl *) This)->input_reg_used[reg] = TRUE;
|
---|
465 | }
|
---|
466 | }
|
---|
467 | }
|
---|
468 |
|
---|
469 | else if (WINED3DSPR_RASTOUT == regtype && reg == 1)
|
---|
470 | reg_maps->fog = 1;
|
---|
471 |
|
---|
472 | else if (WINED3DSPR_MISCTYPE == regtype && reg == 0 && pshader)
|
---|
473 | reg_maps->vpos = 1;
|
---|
474 |
|
---|
475 | else if(WINED3DSPR_CONST == regtype) {
|
---|
476 | if(param & WINED3DSHADER_ADDRMODE_RELATIVE) {
|
---|
477 | if(!pshader) {
|
---|
478 | if(reg <= ((IWineD3DVertexShaderImpl *) This)->min_rel_offset) {
|
---|
479 | ((IWineD3DVertexShaderImpl *) This)->min_rel_offset = reg;
|
---|
480 | } else if(reg >= ((IWineD3DVertexShaderImpl *) This)->max_rel_offset) {
|
---|
481 | ((IWineD3DVertexShaderImpl *) This)->max_rel_offset = reg;
|
---|
482 | }
|
---|
483 | }
|
---|
484 | reg_maps->usesrelconstF = TRUE;
|
---|
485 | }
|
---|
486 | }
|
---|
487 |
|
---|
488 | /* WINED3DSPR_TEXCRDOUT is the same as WINED3DSPR_OUTPUT. _OUTPUT can be > MAX_REG_TEXCRD and is used
|
---|
489 | * in >= 3.0 shaders. Filter 3.0 shaders to prevent overflows, and also filter pixel shaders because TECRDOUT
|
---|
490 | * isn't used in them, but future register types might cause issues
|
---|
491 | */
|
---|
492 | else if(WINED3DSPR_TEXCRDOUT == regtype && i == 0 /* Only look at writes */ &&
|
---|
493 | !pshader && WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version) < 3) {
|
---|
494 | reg_maps->texcoord_mask[reg] |= shader_get_writemask(param);
|
---|
495 | }
|
---|
496 | }
|
---|
497 | }
|
---|
498 | }
|
---|
499 | reg_maps->loop_depth = max_loop_depth;
|
---|
500 |
|
---|
501 | return WINED3D_OK;
|
---|
502 | }
|
---|
503 |
|
---|
504 | static void shader_dump_decl_usage(DWORD decl, DWORD param, DWORD shader_version)
|
---|
505 | {
|
---|
506 | DWORD regtype = shader_get_regtype(param);
|
---|
507 |
|
---|
508 | TRACE("dcl");
|
---|
509 |
|
---|
510 | if (regtype == WINED3DSPR_SAMPLER) {
|
---|
511 | DWORD ttype = decl & WINED3DSP_TEXTURETYPE_MASK;
|
---|
512 |
|
---|
513 | switch (ttype) {
|
---|
514 | case WINED3DSTT_2D: TRACE("_2d"); break;
|
---|
515 | case WINED3DSTT_CUBE: TRACE("_cube"); break;
|
---|
516 | case WINED3DSTT_VOLUME: TRACE("_volume"); break;
|
---|
517 | default: TRACE("_unknown_ttype(0x%08x)", ttype);
|
---|
518 | }
|
---|
519 |
|
---|
520 | } else {
|
---|
521 |
|
---|
522 | DWORD usage = decl & WINED3DSP_DCL_USAGE_MASK;
|
---|
523 | DWORD idx = (decl & WINED3DSP_DCL_USAGEINDEX_MASK) >> WINED3DSP_DCL_USAGEINDEX_SHIFT;
|
---|
524 |
|
---|
525 | /* Pixel shaders 3.0 don't have usage semantics */
|
---|
526 | if (shader_is_pshader_version(shader_version) && shader_version < WINED3DPS_VERSION(3,0))
|
---|
527 | return;
|
---|
528 | else
|
---|
529 | TRACE("_");
|
---|
530 |
|
---|
531 | switch(usage) {
|
---|
532 | case WINED3DDECLUSAGE_POSITION:
|
---|
533 | TRACE("position%d", idx);
|
---|
534 | break;
|
---|
535 | case WINED3DDECLUSAGE_BLENDINDICES:
|
---|
536 | TRACE("blend");
|
---|
537 | break;
|
---|
538 | case WINED3DDECLUSAGE_BLENDWEIGHT:
|
---|
539 | TRACE("weight");
|
---|
540 | break;
|
---|
541 | case WINED3DDECLUSAGE_NORMAL:
|
---|
542 | TRACE("normal%d", idx);
|
---|
543 | break;
|
---|
544 | case WINED3DDECLUSAGE_PSIZE:
|
---|
545 | TRACE("psize");
|
---|
546 | break;
|
---|
547 | case WINED3DDECLUSAGE_COLOR:
|
---|
548 | if(idx == 0) {
|
---|
549 | TRACE("color");
|
---|
550 | } else {
|
---|
551 | TRACE("specular%d", (idx - 1));
|
---|
552 | }
|
---|
553 | break;
|
---|
554 | case WINED3DDECLUSAGE_TEXCOORD:
|
---|
555 | TRACE("texture%d", idx);
|
---|
556 | break;
|
---|
557 | case WINED3DDECLUSAGE_TANGENT:
|
---|
558 | TRACE("tangent");
|
---|
559 | break;
|
---|
560 | case WINED3DDECLUSAGE_BINORMAL:
|
---|
561 | TRACE("binormal");
|
---|
562 | break;
|
---|
563 | case WINED3DDECLUSAGE_TESSFACTOR:
|
---|
564 | TRACE("tessfactor");
|
---|
565 | break;
|
---|
566 | case WINED3DDECLUSAGE_POSITIONT:
|
---|
567 | TRACE("positionT%d", idx);
|
---|
568 | break;
|
---|
569 | case WINED3DDECLUSAGE_FOG:
|
---|
570 | TRACE("fog");
|
---|
571 | break;
|
---|
572 | case WINED3DDECLUSAGE_DEPTH:
|
---|
573 | TRACE("depth");
|
---|
574 | break;
|
---|
575 | case WINED3DDECLUSAGE_SAMPLE:
|
---|
576 | TRACE("sample");
|
---|
577 | break;
|
---|
578 | default:
|
---|
579 | FIXME("unknown_semantics(0x%08x)", usage);
|
---|
580 | }
|
---|
581 | }
|
---|
582 | }
|
---|
583 |
|
---|
584 | static void shader_dump_arr_entry(const DWORD param, const DWORD addr_token,
|
---|
585 | unsigned int reg, int input, DWORD shader_version)
|
---|
586 | {
|
---|
587 | char relative =
|
---|
588 | ((param & WINED3DSHADER_ADDRESSMODE_MASK) == WINED3DSHADER_ADDRMODE_RELATIVE);
|
---|
589 |
|
---|
590 | if (relative) {
|
---|
591 | TRACE("[");
|
---|
592 | if (addr_token)
|
---|
593 | shader_dump_param(addr_token, 0, input, shader_version);
|
---|
594 | else
|
---|
595 | TRACE("a0.x");
|
---|
596 | TRACE(" + ");
|
---|
597 | }
|
---|
598 | TRACE("%u", reg);
|
---|
599 | if (relative)
|
---|
600 | TRACE("]");
|
---|
601 | }
|
---|
602 |
|
---|
603 | static void shader_dump_param(const DWORD param, const DWORD addr_token, int input, DWORD shader_version)
|
---|
604 | {
|
---|
605 | static const char * const rastout_reg_names[] = { "oPos", "oFog", "oPts" };
|
---|
606 | static const char * const misctype_reg_names[] = { "vPos", "vFace"};
|
---|
607 | char swizzle_reg_chars[4];
|
---|
608 |
|
---|
609 | DWORD reg = param & WINED3DSP_REGNUM_MASK;
|
---|
610 | DWORD regtype = shader_get_regtype(param);
|
---|
611 | DWORD modifier = param & WINED3DSP_SRCMOD_MASK;
|
---|
612 |
|
---|
613 | /* There are some minor differences between pixel and vertex shaders */
|
---|
614 | char pshader = shader_is_pshader_version(shader_version);
|
---|
615 |
|
---|
616 | /* For one, we'd prefer color components to be shown for pshaders.
|
---|
617 | * FIXME: use the swizzle function for this */
|
---|
618 |
|
---|
619 | swizzle_reg_chars[0] = pshader? 'r': 'x';
|
---|
620 | swizzle_reg_chars[1] = pshader? 'g': 'y';
|
---|
621 | swizzle_reg_chars[2] = pshader? 'b': 'z';
|
---|
622 | swizzle_reg_chars[3] = pshader? 'a': 'w';
|
---|
623 |
|
---|
624 | if (input) {
|
---|
625 | if ( (modifier == WINED3DSPSM_NEG) ||
|
---|
626 | (modifier == WINED3DSPSM_BIASNEG) ||
|
---|
627 | (modifier == WINED3DSPSM_SIGNNEG) ||
|
---|
628 | (modifier == WINED3DSPSM_X2NEG) ||
|
---|
629 | (modifier == WINED3DSPSM_ABSNEG) )
|
---|
630 | TRACE("-");
|
---|
631 | else if (modifier == WINED3DSPSM_COMP)
|
---|
632 | TRACE("1-");
|
---|
633 | else if (modifier == WINED3DSPSM_NOT)
|
---|
634 | TRACE("!");
|
---|
635 |
|
---|
636 | if (modifier == WINED3DSPSM_ABS || modifier == WINED3DSPSM_ABSNEG)
|
---|
637 | TRACE("abs(");
|
---|
638 | }
|
---|
639 |
|
---|
640 | switch (regtype) {
|
---|
641 | case WINED3DSPR_TEMP:
|
---|
642 | TRACE("r%u", reg);
|
---|
643 | break;
|
---|
644 | case WINED3DSPR_INPUT:
|
---|
645 | TRACE("v");
|
---|
646 | shader_dump_arr_entry(param, addr_token, reg, input, shader_version);
|
---|
647 | break;
|
---|
648 | case WINED3DSPR_CONST:
|
---|
649 | case WINED3DSPR_CONST2:
|
---|
650 | case WINED3DSPR_CONST3:
|
---|
651 | case WINED3DSPR_CONST4:
|
---|
652 | TRACE("c");
|
---|
653 | shader_dump_arr_entry(param, addr_token, shader_get_float_offset(param), input, shader_version);
|
---|
654 | break;
|
---|
655 | case WINED3DSPR_TEXTURE: /* vs: case D3DSPR_ADDR */
|
---|
656 | TRACE("%c%u", (pshader? 't':'a'), reg);
|
---|
657 | break;
|
---|
658 | case WINED3DSPR_RASTOUT:
|
---|
659 | TRACE("%s", rastout_reg_names[reg]);
|
---|
660 | break;
|
---|
661 | case WINED3DSPR_COLOROUT:
|
---|
662 | TRACE("oC%u", reg);
|
---|
663 | break;
|
---|
664 | case WINED3DSPR_DEPTHOUT:
|
---|
665 | TRACE("oDepth");
|
---|
666 | break;
|
---|
667 | case WINED3DSPR_ATTROUT:
|
---|
668 | TRACE("oD%u", reg);
|
---|
669 | break;
|
---|
670 | case WINED3DSPR_TEXCRDOUT:
|
---|
671 |
|
---|
672 | /* Vertex shaders >= 3.0 use general purpose output registers
|
---|
673 | * (WINED3DSPR_OUTPUT), which can include an address token */
|
---|
674 |
|
---|
675 | if (WINED3DSHADER_VERSION_MAJOR(shader_version) >= 3) {
|
---|
676 | TRACE("o");
|
---|
677 | shader_dump_arr_entry(param, addr_token, reg, input, shader_version);
|
---|
678 | }
|
---|
679 | else
|
---|
680 | TRACE("oT%u", reg);
|
---|
681 | break;
|
---|
682 | case WINED3DSPR_CONSTINT:
|
---|
683 | TRACE("i");
|
---|
684 | shader_dump_arr_entry(param, addr_token, reg, input, shader_version);
|
---|
685 | break;
|
---|
686 | case WINED3DSPR_CONSTBOOL:
|
---|
687 | TRACE("b");
|
---|
688 | shader_dump_arr_entry(param, addr_token, reg, input, shader_version);
|
---|
689 | break;
|
---|
690 | case WINED3DSPR_LABEL:
|
---|
691 | TRACE("l%u", reg);
|
---|
692 | break;
|
---|
693 | case WINED3DSPR_LOOP:
|
---|
694 | TRACE("aL");
|
---|
695 | break;
|
---|
696 | case WINED3DSPR_SAMPLER:
|
---|
697 | TRACE("s%u", reg);
|
---|
698 | break;
|
---|
699 | case WINED3DSPR_MISCTYPE:
|
---|
700 | if (reg > 1) {
|
---|
701 | FIXME("Unhandled misctype register %d\n", reg);
|
---|
702 | } else {
|
---|
703 | TRACE("%s", misctype_reg_names[reg]);
|
---|
704 | }
|
---|
705 | break;
|
---|
706 | case WINED3DSPR_PREDICATE:
|
---|
707 | TRACE("p%u", reg);
|
---|
708 | break;
|
---|
709 | default:
|
---|
710 | TRACE("unhandled_rtype(%#x)", regtype);
|
---|
711 | break;
|
---|
712 | }
|
---|
713 |
|
---|
714 | if (!input) {
|
---|
715 | /* operand output (for modifiers and shift, see dump_ins_modifiers) */
|
---|
716 |
|
---|
717 | if ((param & WINED3DSP_WRITEMASK_ALL) != WINED3DSP_WRITEMASK_ALL) {
|
---|
718 | TRACE(".");
|
---|
719 | if (param & WINED3DSP_WRITEMASK_0) TRACE("%c", swizzle_reg_chars[0]);
|
---|
720 | if (param & WINED3DSP_WRITEMASK_1) TRACE("%c", swizzle_reg_chars[1]);
|
---|
721 | if (param & WINED3DSP_WRITEMASK_2) TRACE("%c", swizzle_reg_chars[2]);
|
---|
722 | if (param & WINED3DSP_WRITEMASK_3) TRACE("%c", swizzle_reg_chars[3]);
|
---|
723 | }
|
---|
724 |
|
---|
725 | } else {
|
---|
726 | /** operand input */
|
---|
727 | DWORD swizzle = (param & WINED3DSP_SWIZZLE_MASK) >> WINED3DSP_SWIZZLE_SHIFT;
|
---|
728 | DWORD swizzle_r = swizzle & 0x03;
|
---|
729 | DWORD swizzle_g = (swizzle >> 2) & 0x03;
|
---|
730 | DWORD swizzle_b = (swizzle >> 4) & 0x03;
|
---|
731 | DWORD swizzle_a = (swizzle >> 6) & 0x03;
|
---|
732 |
|
---|
733 | if (0 != modifier) {
|
---|
734 | switch (modifier) {
|
---|
735 | case WINED3DSPSM_NONE: break;
|
---|
736 | case WINED3DSPSM_NEG: break;
|
---|
737 | case WINED3DSPSM_NOT: break;
|
---|
738 | case WINED3DSPSM_BIAS: TRACE("_bias"); break;
|
---|
739 | case WINED3DSPSM_BIASNEG: TRACE("_bias"); break;
|
---|
740 | case WINED3DSPSM_SIGN: TRACE("_bx2"); break;
|
---|
741 | case WINED3DSPSM_SIGNNEG: TRACE("_bx2"); break;
|
---|
742 | case WINED3DSPSM_COMP: break;
|
---|
743 | case WINED3DSPSM_X2: TRACE("_x2"); break;
|
---|
744 | case WINED3DSPSM_X2NEG: TRACE("_x2"); break;
|
---|
745 | case WINED3DSPSM_DZ: TRACE("_dz"); break;
|
---|
746 | case WINED3DSPSM_DW: TRACE("_dw"); break;
|
---|
747 | case WINED3DSPSM_ABSNEG: TRACE(")"); break;
|
---|
748 | case WINED3DSPSM_ABS: TRACE(")"); break;
|
---|
749 | default:
|
---|
750 | TRACE("_unknown_modifier(%#x)", modifier >> WINED3DSP_SRCMOD_SHIFT);
|
---|
751 | }
|
---|
752 | }
|
---|
753 |
|
---|
754 | /**
|
---|
755 | * swizzle bits fields:
|
---|
756 | * RRGGBBAA
|
---|
757 | */
|
---|
758 | if ((WINED3DVS_NOSWIZZLE >> WINED3DVS_SWIZZLE_SHIFT) != swizzle) {
|
---|
759 | if (swizzle_r == swizzle_g &&
|
---|
760 | swizzle_r == swizzle_b &&
|
---|
761 | swizzle_r == swizzle_a) {
|
---|
762 | TRACE(".%c", swizzle_reg_chars[swizzle_r]);
|
---|
763 | } else {
|
---|
764 | TRACE(".%c%c%c%c",
|
---|
765 | swizzle_reg_chars[swizzle_r],
|
---|
766 | swizzle_reg_chars[swizzle_g],
|
---|
767 | swizzle_reg_chars[swizzle_b],
|
---|
768 | swizzle_reg_chars[swizzle_a]);
|
---|
769 | }
|
---|
770 | }
|
---|
771 | }
|
---|
772 | }
|
---|
773 |
|
---|
774 | static void shader_color_correction(IWineD3DBaseShaderImpl *shader,
|
---|
775 | IWineD3DDeviceImpl *device, const struct SHADER_OPCODE_ARG *arg)
|
---|
776 | {
|
---|
777 | DWORD hex_version = shader->baseShader.hex_version;
|
---|
778 | IWineD3DBaseTextureImpl *texture;
|
---|
779 | struct color_fixup_desc fixup;
|
---|
780 | BOOL recorded = FALSE;
|
---|
781 | DWORD sampler_idx;
|
---|
782 | UINT i;
|
---|
783 |
|
---|
784 | switch(arg->opcode->opcode)
|
---|
785 | {
|
---|
786 | case WINED3DSIO_TEX:
|
---|
787 | if (hex_version < WINED3DPS_VERSION(2,0)) sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
|
---|
788 | else sampler_idx = arg->src[1] & WINED3DSP_REGNUM_MASK;
|
---|
789 | break;
|
---|
790 |
|
---|
791 | case WINED3DSIO_TEXLDL:
|
---|
792 | FIXME("Add color fixup for vertex texture WINED3DSIO_TEXLDL\n");
|
---|
793 | return;
|
---|
794 |
|
---|
795 | case WINED3DSIO_TEXDP3TEX:
|
---|
796 | case WINED3DSIO_TEXM3x3TEX:
|
---|
797 | case WINED3DSIO_TEXM3x3SPEC:
|
---|
798 | case WINED3DSIO_TEXM3x3VSPEC:
|
---|
799 | case WINED3DSIO_TEXBEM:
|
---|
800 | case WINED3DSIO_TEXREG2AR:
|
---|
801 | case WINED3DSIO_TEXREG2GB:
|
---|
802 | case WINED3DSIO_TEXREG2RGB:
|
---|
803 | sampler_idx = arg->dst & WINED3DSP_REGNUM_MASK;
|
---|
804 | break;
|
---|
805 |
|
---|
806 | default:
|
---|
807 | /* Not a texture sampling instruction, nothing to do */
|
---|
808 | return;
|
---|
809 | };
|
---|
810 |
|
---|
811 | texture = (IWineD3DBaseTextureImpl *)device->stateBlock->textures[sampler_idx];
|
---|
812 | if (texture) fixup = texture->baseTexture.shader_color_fixup;
|
---|
813 | else fixup = COLOR_FIXUP_IDENTITY;
|
---|
814 |
|
---|
815 | /* before doing anything, record the sampler with the format in the format conversion list,
|
---|
816 | * but check if it's not there already */
|
---|
817 | for (i = 0; i < shader->baseShader.num_sampled_samplers; ++i)
|
---|
818 | {
|
---|
819 | if (shader->baseShader.sampled_samplers[i] == sampler_idx)
|
---|
820 | {
|
---|
821 | recorded = TRUE;
|
---|
822 | break;
|
---|
823 | }
|
---|
824 | }
|
---|
825 |
|
---|
826 | if (!recorded)
|
---|
827 | {
|
---|
828 | shader->baseShader.sampled_samplers[shader->baseShader.num_sampled_samplers] = sampler_idx;
|
---|
829 | ++shader->baseShader.num_sampled_samplers;
|
---|
830 | }
|
---|
831 |
|
---|
832 | device->shader_backend->shader_color_correction(arg, fixup);
|
---|
833 | }
|
---|
834 |
|
---|
835 | /** Shared code in order to generate the bulk of the shader string.
|
---|
836 | Use the shader_header_fct & shader_footer_fct to add strings
|
---|
837 | that are specific to pixel or vertex functions
|
---|
838 | NOTE: A description of how to parse tokens can be found on msdn */
|
---|
839 | void shader_generate_main(IWineD3DBaseShader *iface, SHADER_BUFFER* buffer,
|
---|
840 | const shader_reg_maps* reg_maps, CONST DWORD* pFunction)
|
---|
841 | {
|
---|
842 | IWineD3DBaseShaderImpl* This = (IWineD3DBaseShaderImpl*) iface;
|
---|
843 | IWineD3DDeviceImpl *device = (IWineD3DDeviceImpl *) This->baseShader.device; /* To access shader backend callbacks */
|
---|
844 | const SHADER_OPCODE *opcode_table = This->baseShader.shader_ins;
|
---|
845 | const SHADER_HANDLER *handler_table = device->shader_backend->shader_instruction_handler_table;
|
---|
846 | DWORD shader_version = This->baseShader.hex_version;
|
---|
847 | const DWORD *pToken = pFunction;
|
---|
848 | const SHADER_OPCODE *curOpcode = NULL;
|
---|
849 | SHADER_HANDLER hw_fct = NULL;
|
---|
850 | DWORD i;
|
---|
851 | SHADER_OPCODE_ARG hw_arg;
|
---|
852 |
|
---|
853 | /* Initialize current parsing state */
|
---|
854 | hw_arg.shader = iface;
|
---|
855 | hw_arg.buffer = buffer;
|
---|
856 | hw_arg.reg_maps = reg_maps;
|
---|
857 | This->baseShader.parse_state.current_row = 0;
|
---|
858 |
|
---|
859 | /* Second pass, process opcodes */
|
---|
860 | if (NULL != pToken) {
|
---|
861 | while (WINED3DPS_END() != *pToken) {
|
---|
862 |
|
---|
863 | /* Skip version token */
|
---|
864 | if (shader_is_version_token(*pToken)) {
|
---|
865 | ++pToken;
|
---|
866 | continue;
|
---|
867 | }
|
---|
868 |
|
---|
869 | /* Skip comment tokens */
|
---|
870 | if (shader_is_comment(*pToken)) {
|
---|
871 | DWORD comment_len = (*pToken & WINED3DSI_COMMENTSIZE_MASK) >> WINED3DSI_COMMENTSIZE_SHIFT;
|
---|
872 | ++pToken;
|
---|
873 | TRACE("#%s\n", (const char*)pToken);
|
---|
874 | pToken += comment_len;
|
---|
875 | continue;
|
---|
876 | }
|
---|
877 |
|
---|
878 | /* Read opcode */
|
---|
879 | hw_arg.opcode_token = *pToken++;
|
---|
880 | curOpcode = shader_get_opcode(opcode_table, shader_version, hw_arg.opcode_token);
|
---|
881 |
|
---|
882 | /* Select handler */
|
---|
883 | if (curOpcode == NULL)
|
---|
884 | hw_fct = NULL;
|
---|
885 | else
|
---|
886 | hw_fct = handler_table[curOpcode->handler_idx];
|
---|
887 |
|
---|
888 | /* Unknown opcode and its parameters */
|
---|
889 | if (NULL == curOpcode) {
|
---|
890 | FIXME("Unrecognized opcode: token=0x%08x\n", hw_arg.opcode_token);
|
---|
891 | pToken += shader_skip_unrecognized(iface, pToken);
|
---|
892 |
|
---|
893 | /* Nothing to do */
|
---|
894 | } else if (WINED3DSIO_DCL == curOpcode->opcode ||
|
---|
895 | WINED3DSIO_NOP == curOpcode->opcode ||
|
---|
896 | WINED3DSIO_DEF == curOpcode->opcode ||
|
---|
897 | WINED3DSIO_DEFI == curOpcode->opcode ||
|
---|
898 | WINED3DSIO_DEFB == curOpcode->opcode ||
|
---|
899 | WINED3DSIO_PHASE == curOpcode->opcode ||
|
---|
900 | WINED3DSIO_RET == curOpcode->opcode) {
|
---|
901 |
|
---|
902 | pToken += shader_skip_opcode(This, curOpcode, hw_arg.opcode_token);
|
---|
903 |
|
---|
904 | /* If a generator function is set for current shader target, use it */
|
---|
905 | } else if (hw_fct != NULL) {
|
---|
906 |
|
---|
907 | hw_arg.opcode = curOpcode;
|
---|
908 |
|
---|
909 | /* Destination token */
|
---|
910 | if (curOpcode->dst_token) {
|
---|
911 |
|
---|
912 | DWORD param, addr_token = 0;
|
---|
913 | pToken += shader_get_param(iface, pToken, ¶m, &addr_token);
|
---|
914 | hw_arg.dst = param;
|
---|
915 | hw_arg.dst_addr = addr_token;
|
---|
916 | }
|
---|
917 |
|
---|
918 | /* Predication token */
|
---|
919 | if (hw_arg.opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED)
|
---|
920 | hw_arg.predicate = *pToken++;
|
---|
921 |
|
---|
922 | /* Other source tokens */
|
---|
923 | for (i = 0; i < (curOpcode->num_params - curOpcode->dst_token); i++) {
|
---|
924 |
|
---|
925 | DWORD param, addr_token = 0;
|
---|
926 | pToken += shader_get_param(iface, pToken, ¶m, &addr_token);
|
---|
927 | hw_arg.src[i] = param;
|
---|
928 | hw_arg.src_addr[i] = addr_token;
|
---|
929 | }
|
---|
930 |
|
---|
931 | /* Call appropriate function for output target */
|
---|
932 | hw_fct(&hw_arg);
|
---|
933 |
|
---|
934 | /* Add color correction if needed */
|
---|
935 | shader_color_correction(This, device, &hw_arg);
|
---|
936 |
|
---|
937 | /* Process instruction modifiers for GLSL apps ( _sat, etc. ) */
|
---|
938 | /* FIXME: This should be internal to the shader backend.
|
---|
939 | * Also, right now this is the only reason "shader_mode" exists. */
|
---|
940 | if (This->baseShader.shader_mode == SHADER_GLSL)
|
---|
941 | shader_glsl_add_instruction_modifiers(&hw_arg);
|
---|
942 |
|
---|
943 | /* Unhandled opcode */
|
---|
944 | } else {
|
---|
945 |
|
---|
946 | FIXME("Can't handle opcode %s in hwShader\n", curOpcode->name);
|
---|
947 | pToken += shader_skip_opcode(This, curOpcode, hw_arg.opcode_token);
|
---|
948 | }
|
---|
949 | }
|
---|
950 | /* TODO: What about result.depth? */
|
---|
951 |
|
---|
952 | }
|
---|
953 | }
|
---|
954 |
|
---|
955 | static void shader_dump_ins_modifiers(const DWORD output)
|
---|
956 | {
|
---|
957 | DWORD shift = (output & WINED3DSP_DSTSHIFT_MASK) >> WINED3DSP_DSTSHIFT_SHIFT;
|
---|
958 | DWORD mmask = output & WINED3DSP_DSTMOD_MASK;
|
---|
959 |
|
---|
960 | switch (shift) {
|
---|
961 | case 0: break;
|
---|
962 | case 13: TRACE("_d8"); break;
|
---|
963 | case 14: TRACE("_d4"); break;
|
---|
964 | case 15: TRACE("_d2"); break;
|
---|
965 | case 1: TRACE("_x2"); break;
|
---|
966 | case 2: TRACE("_x4"); break;
|
---|
967 | case 3: TRACE("_x8"); break;
|
---|
968 | default: TRACE("_unhandled_shift(%d)", shift); break;
|
---|
969 | }
|
---|
970 |
|
---|
971 | if (mmask & WINED3DSPDM_SATURATE) TRACE("_sat");
|
---|
972 | if (mmask & WINED3DSPDM_PARTIALPRECISION) TRACE("_pp");
|
---|
973 | if (mmask & WINED3DSPDM_MSAMPCENTROID) TRACE("_centroid");
|
---|
974 |
|
---|
975 | mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
|
---|
976 | if (mmask)
|
---|
977 | FIXME("_unrecognized_modifier(%#x)", mmask >> WINED3DSP_DSTMOD_SHIFT);
|
---|
978 | }
|
---|
979 |
|
---|
980 | /* First pass: trace shader, initialize length and version */
|
---|
981 | void shader_trace_init(
|
---|
982 | IWineD3DBaseShader *iface,
|
---|
983 | const DWORD* pFunction) {
|
---|
984 |
|
---|
985 | IWineD3DBaseShaderImpl *This =(IWineD3DBaseShaderImpl *)iface;
|
---|
986 |
|
---|
987 | const DWORD* pToken = pFunction;
|
---|
988 | const SHADER_OPCODE* curOpcode = NULL;
|
---|
989 | DWORD opcode_token;
|
---|
990 | unsigned int len = 0;
|
---|
991 | DWORD i;
|
---|
992 |
|
---|
993 | TRACE("(%p) : Parsing program\n", This);
|
---|
994 |
|
---|
995 | if (NULL != pToken) {
|
---|
996 | while (WINED3DVS_END() != *pToken) {
|
---|
997 | if (shader_is_version_token(*pToken)) { /** version */
|
---|
998 | This->baseShader.hex_version = *pToken;
|
---|
999 | TRACE("%s_%u_%u\n", shader_is_pshader_version(This->baseShader.hex_version)? "ps": "vs",
|
---|
1000 | WINED3DSHADER_VERSION_MAJOR(This->baseShader.hex_version),
|
---|
1001 | WINED3DSHADER_VERSION_MINOR(This->baseShader.hex_version));
|
---|
1002 | ++pToken;
|
---|
1003 | ++len;
|
---|
1004 | continue;
|
---|
1005 | }
|
---|
1006 | if (shader_is_comment(*pToken)) { /** comment */
|
---|
1007 | DWORD comment_len = (*pToken & WINED3DSI_COMMENTSIZE_MASK) >> WINED3DSI_COMMENTSIZE_SHIFT;
|
---|
1008 | ++pToken;
|
---|
1009 | TRACE("//%s\n", (const char*)pToken);
|
---|
1010 | pToken += comment_len;
|
---|
1011 | len += comment_len + 1;
|
---|
1012 | continue;
|
---|
1013 | }
|
---|
1014 | opcode_token = *pToken++;
|
---|
1015 | curOpcode = shader_get_opcode(This->baseShader.shader_ins, This->baseShader.hex_version, opcode_token);
|
---|
1016 | len++;
|
---|
1017 |
|
---|
1018 | if (NULL == curOpcode) {
|
---|
1019 | int tokens_read;
|
---|
1020 | FIXME("Unrecognized opcode: token=0x%08x\n", opcode_token);
|
---|
1021 | tokens_read = shader_skip_unrecognized(iface, pToken);
|
---|
1022 | pToken += tokens_read;
|
---|
1023 | len += tokens_read;
|
---|
1024 |
|
---|
1025 | } else {
|
---|
1026 | if (curOpcode->opcode == WINED3DSIO_DCL) {
|
---|
1027 |
|
---|
1028 | DWORD usage = *pToken;
|
---|
1029 | DWORD param = *(pToken + 1);
|
---|
1030 |
|
---|
1031 | shader_dump_decl_usage(usage, param, This->baseShader.hex_version);
|
---|
1032 | shader_dump_ins_modifiers(param);
|
---|
1033 | TRACE(" ");
|
---|
1034 | shader_dump_param(param, 0, 0, This->baseShader.hex_version);
|
---|
1035 | pToken += 2;
|
---|
1036 | len += 2;
|
---|
1037 |
|
---|
1038 | } else if (curOpcode->opcode == WINED3DSIO_DEF) {
|
---|
1039 |
|
---|
1040 | unsigned int offset = shader_get_float_offset(*pToken);
|
---|
1041 |
|
---|
1042 | TRACE("def c%u = %f, %f, %f, %f", offset,
|
---|
1043 | *(const float *)(pToken + 1),
|
---|
1044 | *(const float *)(pToken + 2),
|
---|
1045 | *(const float *)(pToken + 3),
|
---|
1046 | *(const float *)(pToken + 4));
|
---|
1047 |
|
---|
1048 | pToken += 5;
|
---|
1049 | len += 5;
|
---|
1050 | } else if (curOpcode->opcode == WINED3DSIO_DEFI) {
|
---|
1051 |
|
---|
1052 | TRACE("defi i%u = %d, %d, %d, %d", *pToken & WINED3DSP_REGNUM_MASK,
|
---|
1053 | *(pToken + 1),
|
---|
1054 | *(pToken + 2),
|
---|
1055 | *(pToken + 3),
|
---|
1056 | *(pToken + 4));
|
---|
1057 |
|
---|
1058 | pToken += 5;
|
---|
1059 | len += 5;
|
---|
1060 |
|
---|
1061 | } else if (curOpcode->opcode == WINED3DSIO_DEFB) {
|
---|
1062 |
|
---|
1063 | TRACE("defb b%u = %s", *pToken & WINED3DSP_REGNUM_MASK,
|
---|
1064 | *(pToken + 1)? "true": "false");
|
---|
1065 |
|
---|
1066 | pToken += 2;
|
---|
1067 | len += 2;
|
---|
1068 |
|
---|
1069 | } else {
|
---|
1070 |
|
---|
1071 | DWORD param, addr_token;
|
---|
1072 | int tokens_read;
|
---|
1073 |
|
---|
1074 | /* Print out predication source token first - it follows
|
---|
1075 | * the destination token. */
|
---|
1076 | if (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED) {
|
---|
1077 | TRACE("(");
|
---|
1078 | shader_dump_param(*(pToken + 2), 0, 1, This->baseShader.hex_version);
|
---|
1079 | TRACE(") ");
|
---|
1080 | }
|
---|
1081 | if (opcode_token & WINED3DSI_COISSUE) {
|
---|
1082 | /* PixWin marks instructions with the coissue flag with a '+' */
|
---|
1083 | TRACE("+");
|
---|
1084 | }
|
---|
1085 |
|
---|
1086 | TRACE("%s", curOpcode->name);
|
---|
1087 |
|
---|
1088 | if (curOpcode->opcode == WINED3DSIO_IFC ||
|
---|
1089 | curOpcode->opcode == WINED3DSIO_BREAKC) {
|
---|
1090 |
|
---|
1091 | DWORD op = (opcode_token & INST_CONTROLS_MASK) >> INST_CONTROLS_SHIFT;
|
---|
1092 | switch (op) {
|
---|
1093 | case COMPARISON_GT: TRACE("_gt"); break;
|
---|
1094 | case COMPARISON_EQ: TRACE("_eq"); break;
|
---|
1095 | case COMPARISON_GE: TRACE("_ge"); break;
|
---|
1096 | case COMPARISON_LT: TRACE("_lt"); break;
|
---|
1097 | case COMPARISON_NE: TRACE("_ne"); break;
|
---|
1098 | case COMPARISON_LE: TRACE("_le"); break;
|
---|
1099 | default:
|
---|
1100 | TRACE("_(%u)", op);
|
---|
1101 | }
|
---|
1102 | } else if (curOpcode->opcode == WINED3DSIO_TEX &&
|
---|
1103 | This->baseShader.hex_version >= WINED3DPS_VERSION(2,0)) {
|
---|
1104 | if(opcode_token & WINED3DSI_TEXLD_PROJECT) TRACE("p");
|
---|
1105 | }
|
---|
1106 |
|
---|
1107 | /* Destination token */
|
---|
1108 | if (curOpcode->dst_token) {
|
---|
1109 |
|
---|
1110 | /* Destination token */
|
---|
1111 | tokens_read = shader_get_param(iface, pToken, ¶m, &addr_token);
|
---|
1112 | pToken += tokens_read;
|
---|
1113 | len += tokens_read;
|
---|
1114 |
|
---|
1115 | shader_dump_ins_modifiers(param);
|
---|
1116 | TRACE(" ");
|
---|
1117 | shader_dump_param(param, addr_token, 0, This->baseShader.hex_version);
|
---|
1118 | }
|
---|
1119 |
|
---|
1120 | /* Predication token - already printed out, just skip it */
|
---|
1121 | if (opcode_token & WINED3DSHADER_INSTRUCTION_PREDICATED) {
|
---|
1122 | pToken++;
|
---|
1123 | len++;
|
---|
1124 | }
|
---|
1125 |
|
---|
1126 | /* Other source tokens */
|
---|
1127 | for (i = curOpcode->dst_token; i < curOpcode->num_params; ++i) {
|
---|
1128 |
|
---|
1129 | tokens_read = shader_get_param(iface, pToken, ¶m, &addr_token);
|
---|
1130 | pToken += tokens_read;
|
---|
1131 | len += tokens_read;
|
---|
1132 |
|
---|
1133 | TRACE((i == 0)? " " : ", ");
|
---|
1134 | shader_dump_param(param, addr_token, 1, This->baseShader.hex_version);
|
---|
1135 | }
|
---|
1136 | }
|
---|
1137 | TRACE("\n");
|
---|
1138 | }
|
---|
1139 | }
|
---|
1140 | This->baseShader.functionLength = (len + 1) * sizeof(DWORD);
|
---|
1141 | } else {
|
---|
1142 | This->baseShader.functionLength = 1; /* no Function defined use fixed function vertex processing */
|
---|
1143 | }
|
---|
1144 | }
|
---|
1145 |
|
---|
1146 | static const SHADER_HANDLER shader_none_instruction_handler_table[WINED3DSIH_TABLE_SIZE] = {0};
|
---|
1147 | static void shader_none_select(IWineD3DDevice *iface, BOOL usePS, BOOL useVS) {}
|
---|
1148 | static void shader_none_select_depth_blt(IWineD3DDevice *iface, enum tex_types tex_type) {}
|
---|
1149 | static void shader_none_deselect_depth_blt(IWineD3DDevice *iface) {}
|
---|
1150 | static void shader_none_load_constants(IWineD3DDevice *iface, char usePS, char useVS) {}
|
---|
1151 | static void shader_none_cleanup(IWineD3DDevice *iface) {}
|
---|
1152 | static void shader_none_color_correction(const struct SHADER_OPCODE_ARG *arg, struct color_fixup_desc fixup) {}
|
---|
1153 | static void shader_none_destroy(IWineD3DBaseShader *iface) {}
|
---|
1154 | static HRESULT shader_none_alloc(IWineD3DDevice *iface) {return WINED3D_OK;}
|
---|
1155 | static void shader_none_free(IWineD3DDevice *iface) {}
|
---|
1156 | static BOOL shader_none_dirty_const(IWineD3DDevice *iface) {return FALSE;}
|
---|
1157 | static GLuint shader_none_generate_pshader(IWineD3DPixelShader *iface, SHADER_BUFFER *buffer) {
|
---|
1158 | FIXME("NONE shader backend asked to generate a pixel shader\n");
|
---|
1159 | return 0;
|
---|
1160 | }
|
---|
1161 | static void shader_none_generate_vshader(IWineD3DVertexShader *iface, SHADER_BUFFER *buffer) {
|
---|
1162 | FIXME("NONE shader backend asked to generate a vertex shader\n");
|
---|
1163 | }
|
---|
1164 |
|
---|
1165 | #define GLINFO_LOCATION (*gl_info)
|
---|
1166 | static void shader_none_get_caps(WINED3DDEVTYPE devtype, const WineD3D_GL_Info *gl_info, struct shader_caps *pCaps)
|
---|
1167 | {
|
---|
1168 | /* Set the shader caps to 0 for the none shader backend */
|
---|
1169 | pCaps->VertexShaderVersion = 0;
|
---|
1170 | pCaps->PixelShaderVersion = 0;
|
---|
1171 | pCaps->PixelShader1xMaxValue = 0.0;
|
---|
1172 | }
|
---|
1173 | #undef GLINFO_LOCATION
|
---|
1174 | static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
|
---|
1175 | {
|
---|
1176 | if (TRACE_ON(d3d_shader) && TRACE_ON(d3d))
|
---|
1177 | {
|
---|
1178 | TRACE("Checking support for fixup:\n");
|
---|
1179 | dump_color_fixup_desc(fixup);
|
---|
1180 | }
|
---|
1181 |
|
---|
1182 | /* Faked to make some apps happy. */
|
---|
1183 | if (!is_yuv_fixup(fixup))
|
---|
1184 | {
|
---|
1185 | TRACE("[OK]\n");
|
---|
1186 | return TRUE;
|
---|
1187 | }
|
---|
1188 |
|
---|
1189 | TRACE("[FAILED]\n");
|
---|
1190 | return FALSE;
|
---|
1191 | }
|
---|
1192 |
|
---|
1193 | const shader_backend_t none_shader_backend = {
|
---|
1194 | shader_none_instruction_handler_table,
|
---|
1195 | shader_none_select,
|
---|
1196 | shader_none_select_depth_blt,
|
---|
1197 | shader_none_deselect_depth_blt,
|
---|
1198 | shader_none_load_constants,
|
---|
1199 | shader_none_cleanup,
|
---|
1200 | shader_none_color_correction,
|
---|
1201 | shader_none_destroy,
|
---|
1202 | shader_none_alloc,
|
---|
1203 | shader_none_free,
|
---|
1204 | shader_none_dirty_const,
|
---|
1205 | shader_none_generate_pshader,
|
---|
1206 | shader_none_generate_vshader,
|
---|
1207 | shader_none_get_caps,
|
---|
1208 | shader_none_color_fixup_supported,
|
---|
1209 | };
|
---|
1210 |
|
---|
1211 | /* *******************************************
|
---|
1212 | IWineD3DPixelShader IUnknown parts follow
|
---|
1213 | ******************************************* */
|
---|
1214 | HRESULT WINAPI IWineD3DBaseShaderImpl_QueryInterface(IWineD3DBaseShader *iface, REFIID riid, LPVOID *ppobj)
|
---|
1215 | {
|
---|
1216 | IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)iface;
|
---|
1217 | TRACE("(%p)->(%s,%p)\n",This,debugstr_guid(riid),ppobj);
|
---|
1218 | if (IsEqualGUID(riid, &IID_IUnknown)
|
---|
1219 | || IsEqualGUID(riid, &IID_IWineD3DBase)
|
---|
1220 | || IsEqualGUID(riid, &IID_IWineD3DBaseShader)
|
---|
1221 | || IsEqualGUID(riid, &IID_IWineD3DPixelShader)) {
|
---|
1222 | IUnknown_AddRef(iface);
|
---|
1223 | *ppobj = This;
|
---|
1224 | return S_OK;
|
---|
1225 | }
|
---|
1226 | *ppobj = NULL;
|
---|
1227 | return E_NOINTERFACE;
|
---|
1228 | }
|
---|
1229 |
|
---|
1230 | ULONG WINAPI IWineD3DBaseShaderImpl_AddRef(IWineD3DBaseShader *iface) {
|
---|
1231 | IWineD3DPixelShaderImpl *This = (IWineD3DPixelShaderImpl *)iface;
|
---|
1232 | TRACE("(%p) : AddRef increasing from %d\n", This, This->baseShader.ref);
|
---|
1233 | return InterlockedIncrement(&This->baseShader.ref);
|
---|
1234 | }
|
---|
1235 |
|
---|
1236 | ULONG WINAPI IWineD3DBaseShaderImpl_Release(IWineD3DBaseShader *iface) {
|
---|
1237 | IWineD3DBaseShaderImpl *This = (IWineD3DBaseShaderImpl *)iface;
|
---|
1238 | IWineD3DDeviceImpl *deviceImpl = (IWineD3DDeviceImpl *) This->baseShader.device;
|
---|
1239 | ULONG ref;
|
---|
1240 | TRACE("(%p) : Releasing from %d\n", This, This->baseShader.ref);
|
---|
1241 | ref = InterlockedDecrement(&This->baseShader.ref);
|
---|
1242 | if (ref == 0) {
|
---|
1243 | deviceImpl->shader_backend->shader_destroy(iface);
|
---|
1244 | HeapFree(GetProcessHeap(), 0, This->baseShader.function);
|
---|
1245 | shader_delete_constant_list(&This->baseShader.constantsF);
|
---|
1246 | shader_delete_constant_list(&This->baseShader.constantsB);
|
---|
1247 | shader_delete_constant_list(&This->baseShader.constantsI);
|
---|
1248 | list_remove(&This->baseShader.shader_list_entry);
|
---|
1249 | HeapFree(GetProcessHeap(), 0, This);
|
---|
1250 | }
|
---|
1251 | return ref;
|
---|
1252 | }
|
---|