?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
windows.applicationmodel/tests: Use PathRemoveFileSpecW() instead of PathCchRemoveFil...
[wine.git] / dlls / wined3d / shader.c
blobb6bfcd1184f7169a2c726f3df1eeed06ec9a8957
1 /*
2 * Copyright 2002-2003 Jason Edmeades
3 * Copyright 2002-2003 Raphael Junqueira
4 * Copyright 2004 Christian Costa
5 * Copyright 2005 Oliver Stieber
6 * Copyright 2006 Ivan Gyurdiev
7 * Copyright 2007-2008, 2013 Stefan Dösinger for CodeWeavers
8 * Copyright 2009-2011 Henri Verbeet for CodeWeavers
9 *
10 * This library is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU Lesser General Public
12 * License as published by the Free Software Foundation; either
13 * version 2.1 of the License, or (at your option) any later version.
14 *
15 * This library is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * Lesser General Public License for more details.
19 *
20 * You should have received a copy of the GNU Lesser General Public
21 * License along with this library; if not, write to the Free Software
22 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA
23 */
24
25 #include <stdio.h>
26 #include <string.h>
27
28 #include "wined3d_private.h"
29 #include "wined3d_gl.h"
30
31 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
32
33 const struct wined3d_vec4 wined3d_srgb_const[] =
34 {
35 /* pow, mul_high, sub_high, mul_low */
36 {0.41666f, 1.055f, 0.055f, 12.92f},
37 /* cmp */
38 {0.0031308f, 0.0f, 0.0f, 0.0f},
39 };
40
41 static const char * const shader_opcode_names[] =
42 {
43 /* WINED3DSIH_ABS */ "abs",
44 /* WINED3DSIH_ADD */ "add",
45 /* WINED3DSIH_AND */ "and",
46 /* WINED3DSIH_ATOMIC_AND */ "atomic_and",
47 /* WINED3DSIH_ATOMIC_CMP_STORE */ "atomic_cmp_store",
48 /* WINED3DSIH_ATOMIC_IADD */ "atomic_iadd",
49 /* WINED3DSIH_ATOMIC_IMAX */ "atomic_imax",
50 /* WINED3DSIH_ATOMIC_IMIN */ "atomic_imin",
51 /* WINED3DSIH_ATOMIC_OR */ "atomic_or",
52 /* WINED3DSIH_ATOMIC_UMAX */ "atomic_umax",
53 /* WINED3DSIH_ATOMIC_UMIN */ "atomic_umin",
54 /* WINED3DSIH_ATOMIC_XOR */ "atomic_xor",
55 /* WINED3DSIH_BEM */ "bem",
56 /* WINED3DSIH_BFI */ "bfi",
57 /* WINED3DSIH_BFREV */ "bfrev",
58 /* WINED3DSIH_BREAK */ "break",
59 /* WINED3DSIH_BREAKC */ "breakc",
60 /* WINED3DSIH_BREAKP */ "breakp",
61 /* WINED3DSIH_BUFINFO */ "bufinfo",
62 /* WINED3DSIH_CALL */ "call",
63 /* WINED3DSIH_CALLNZ */ "callnz",
64 /* WINED3DSIH_CASE */ "case",
65 /* WINED3DSIH_CMP */ "cmp",
66 /* WINED3DSIH_CND */ "cnd",
67 /* WINED3DSIH_CONTINUE */ "continue",
68 /* WINED3DSIH_CONTINUEP */ "continuec",
69 /* WINED3DSIH_COUNTBITS */ "countbits",
70 /* WINED3DSIH_CRS */ "crs",
71 /* WINED3DSIH_CUT */ "cut",
72 /* WINED3DSIH_CUT_STREAM */ "cut_stream",
73 /* WINED3DSIH_DCL */ "dcl",
74 /* WINED3DSIH_DCL_CONSTANT_BUFFER */ "dcl_constantBuffer",
75 /* WINED3DSIH_DCL_FUNCTION_BODY */ "dcl_function_body",
76 /* WINED3DSIH_DCL_FUNCTION_TABLE */ "dcl_function_table",
77 /* WINED3DSIH_DCL_GLOBAL_FLAGS */ "dcl_globalFlags",
78 /* WINED3DSIH_DCL_GS_INSTANCES */ "dcl_gs_instances",
79 /* WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT */ "dcl_hs_fork_phase_instance_count",
80 /* WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT */ "dcl_hs_join_phase_instance_count",
81 /* WINED3DSIH_DCL_HS_MAX_TESSFACTOR */ "dcl_hs_max_tessfactor",
82 /* WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER */ "dcl_immediateConstantBuffer",
83 /* WINED3DSIH_DCL_INDEX_RANGE */ "dcl_index_range",
84 /* WINED3DSIH_DCL_INDEXABLE_TEMP */ "dcl_indexableTemp",
85 /* WINED3DSIH_DCL_INPUT */ "dcl_input",
86 /* WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT */ "dcl_input_control_point_count",
87 /* WINED3DSIH_DCL_INPUT_PRIMITIVE */ "dcl_inputPrimitive",
88 /* WINED3DSIH_DCL_INPUT_PS */ "dcl_input_ps",
89 /* WINED3DSIH_DCL_INPUT_PS_SGV */ "dcl_input_ps_sgv",
90 /* WINED3DSIH_DCL_INPUT_PS_SIV */ "dcl_input_ps_siv",
91 /* WINED3DSIH_DCL_INPUT_SGV */ "dcl_input_sgv",
92 /* WINED3DSIH_DCL_INPUT_SIV */ "dcl_input_siv",
93 /* WINED3DSIH_DCL_INTERFACE */ "dcl_interface",
94 /* WINED3DSIH_DCL_OUTPUT */ "dcl_output",
95 /* WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT */ "dcl_output_control_point_count",
96 /* WINED3DSIH_DCL_OUTPUT_SIV */ "dcl_output_siv",
97 /* WINED3DSIH_DCL_OUTPUT_TOPOLOGY */ "dcl_outputTopology",
98 /* WINED3DSIH_DCL_RESOURCE_RAW */ "dcl_resource_raw",
99 /* WINED3DSIH_DCL_RESOURCE_STRUCTURED */ "dcl_resource_structured",
100 /* WINED3DSIH_DCL_SAMPLER */ "dcl_sampler",
101 /* WINED3DSIH_DCL_STREAM */ "dcl_stream",
102 /* WINED3DSIH_DCL_TEMPS */ "dcl_temps",
103 /* WINED3DSIH_DCL_TESSELLATOR_DOMAIN */ "dcl_tessellator_domain",
104 /* WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE */ "dcl_tessellator_output_primitive",
105 /* WINED3DSIH_DCL_TESSELLATOR_PARTITIONING */ "dcl_tessellator_partitioning",
106 /* WINED3DSIH_DCL_TGSM_RAW */ "dcl_tgsm_raw",
107 /* WINED3DSIH_DCL_TGSM_STRUCTURED */ "dcl_tgsm_structured",
108 /* WINED3DSIH_DCL_THREAD_GROUP */ "dcl_thread_group",
109 /* WINED3DSIH_DCL_UAV_RAW */ "dcl_uav_raw",
110 /* WINED3DSIH_DCL_UAV_STRUCTURED */ "dcl_uav_structured",
111 /* WINED3DSIH_DCL_UAV_TYPED */ "dcl_uav_typed",
112 /* WINED3DSIH_DCL_VERTICES_OUT */ "dcl_maxOutputVertexCount",
113 /* WINED3DSIH_DEF */ "def",
114 /* WINED3DSIH_DEFAULT */ "default",
115 /* WINED3DSIH_DEFB */ "defb",
116 /* WINED3DSIH_DEFI */ "defi",
117 /* WINED3DSIH_DIV */ "div",
118 /* WINED3DSIH_DP2 */ "dp2",
119 /* WINED3DSIH_DP2ADD */ "dp2add",
120 /* WINED3DSIH_DP3 */ "dp3",
121 /* WINED3DSIH_DP4 */ "dp4",
122 /* WINED3DSIH_DST */ "dst",
123 /* WINED3DSIH_DSX */ "dsx",
124 /* WINED3DSIH_DSX_COARSE */ "deriv_rtx_coarse",
125 /* WINED3DSIH_DSX_FINE */ "deriv_rtx_fine",
126 /* WINED3DSIH_DSY */ "dsy",
127 /* WINED3DSIH_DSY_COARSE */ "deriv_rty_coarse",
128 /* WINED3DSIH_DSY_FINE */ "deriv_rty_fine",
129 /* WINED3DSIH_ELSE */ "else",
130 /* WINED3DSIH_EMIT */ "emit",
131 /* WINED3DSIH_EMIT_STREAM */ "emit_stream",
132 /* WINED3DSIH_ENDIF */ "endif",
133 /* WINED3DSIH_ENDLOOP */ "endloop",
134 /* WINED3DSIH_ENDREP */ "endrep",
135 /* WINED3DSIH_ENDSWITCH */ "endswitch",
136 /* WINED3DSIH_EQ */ "eq",
137 /* WINED3DSIH_EVAL_CENTROID */ "eval_centroid",
138 /* WINED3DSIH_EVAL_SAMPLE_INDEX */ "eval_sample_index",
139 /* WINED3DSIH_EXP */ "exp",
140 /* WINED3DSIH_EXPP */ "expp",
141 /* WINED3DSIH_F16TOF32 */ "f16tof32",
142 /* WINED3DSIH_F32TOF16 */ "f32tof16",
143 /* WINED3DSIH_FCALL */ "fcall",
144 /* WINED3DSIH_FIRSTBIT_HI */ "firstbit_hi",
145 /* WINED3DSIH_FIRSTBIT_LO */ "firstbit_lo",
146 /* WINED3DSIH_FIRSTBIT_SHI */ "firstbit_shi",
147 /* WINED3DSIH_FRC */ "frc",
148 /* WINED3DSIH_FTOI */ "ftoi",
149 /* WINED3DSIH_FTOU */ "ftou",
150 /* WINED3DSIH_GATHER4 */ "gather4",
151 /* WINED3DSIH_GATHER4_C */ "gather4_c",
152 /* WINED3DSIH_GATHER4_PO */ "gather4_po",
153 /* WINED3DSIH_GATHER4_PO_C */ "gather4_po_c",
154 /* WINED3DSIH_GE */ "ge",
155 /* WINED3DSIH_HS_CONTROL_POINT_PHASE */ "hs_control_point_phase",
156 /* WINED3DSIH_HS_DECLS */ "hs_decls",
157 /* WINED3DSIH_HS_FORK_PHASE */ "hs_fork_phase",
158 /* WINED3DSIH_HS_JOIN_PHASE */ "hs_join_phase",
159 /* WINED3DSIH_IADD */ "iadd",
160 /* WINED3DSIH_IBFE */ "ibfe",
161 /* WINED3DSIH_IEQ */ "ieq",
162 /* WINED3DSIH_IF */ "if",
163 /* WINED3DSIH_IFC */ "ifc",
164 /* WINED3DSIH_IGE */ "ige",
165 /* WINED3DSIH_ILT */ "ilt",
166 /* WINED3DSIH_IMAD */ "imad",
167 /* WINED3DSIH_IMAX */ "imax",
168 /* WINED3DSIH_IMIN */ "imin",
169 /* WINED3DSIH_IMM_ATOMIC_ALLOC */ "imm_atomic_alloc",
170 /* WINED3DSIH_IMM_ATOMIC_AND */ "imm_atomic_and",
171 /* WINED3DSIH_IMM_ATOMIC_CMP_EXCH */ "imm_atomic_cmp_exch",
172 /* WINED3DSIH_IMM_ATOMIC_CONSUME */ "imm_atomic_consume",
173 /* WINED3DSIH_IMM_ATOMIC_EXCH */ "imm_atomic_exch",
174 /* WINED3DSIH_IMM_ATOMIC_IADD */ "imm_atomic_iadd",
175 /* WINED3DSIH_IMM_ATOMIC_IMAX */ "imm_atomic_imax",
176 /* WINED3DSIH_IMM_ATOMIC_IMIN */ "imm_atomic_imin",
177 /* WINED3DSIH_IMM_ATOMIC_OR */ "imm_atomic_or",
178 /* WINED3DSIH_IMM_ATOMIC_UMAX */ "imm_atomic_umax",
179 /* WINED3DSIH_IMM_ATOMIC_UMIN */ "imm_atomic_umin",
180 /* WINED3DSIH_IMM_ATOMIC_XOR */ "imm_atomic_xor",
181 /* WINED3DSIH_IMUL */ "imul",
182 /* WINED3DSIH_INE */ "ine",
183 /* WINED3DSIH_INEG */ "ineg",
184 /* WINED3DSIH_ISHL */ "ishl",
185 /* WINED3DSIH_ISHR */ "ishr",
186 /* WINED3DSIH_ITOF */ "itof",
187 /* WINED3DSIH_LABEL */ "label",
188 /* WINED3DSIH_LD */ "ld",
189 /* WINED3DSIH_LD2DMS */ "ld2dms",
190 /* WINED3DSIH_LD_RAW */ "ld_raw",
191 /* WINED3DSIH_LD_STRUCTURED */ "ld_structured",
192 /* WINED3DSIH_LD_UAV_TYPED */ "ld_uav_typed",
193 /* WINED3DSIH_LIT */ "lit",
194 /* WINED3DSIH_LOD */ "lod",
195 /* WINED3DSIH_LOG */ "log",
196 /* WINED3DSIH_LOGP */ "logp",
197 /* WINED3DSIH_LOOP */ "loop",
198 /* WINED3DSIH_LRP */ "lrp",
199 /* WINED3DSIH_LT */ "lt",
200 /* WINED3DSIH_M3x2 */ "m3x2",
201 /* WINED3DSIH_M3x3 */ "m3x3",
202 /* WINED3DSIH_M3x4 */ "m3x4",
203 /* WINED3DSIH_M4x3 */ "m4x3",
204 /* WINED3DSIH_M4x4 */ "m4x4",
205 /* WINED3DSIH_MAD */ "mad",
206 /* WINED3DSIH_MAX */ "max",
207 /* WINED3DSIH_MIN */ "min",
208 /* WINED3DSIH_MOV */ "mov",
209 /* WINED3DSIH_MOVA */ "mova",
210 /* WINED3DSIH_MOVC */ "movc",
211 /* WINED3DSIH_MUL */ "mul",
212 /* WINED3DSIH_NE */ "ne",
213 /* WINED3DSIH_NOP */ "nop",
214 /* WINED3DSIH_NOT */ "not",
215 /* WINED3DSIH_NRM */ "nrm",
216 /* WINED3DSIH_OR */ "or",
217 /* WINED3DSIH_PHASE */ "phase",
218 /* WINED3DSIH_POW */ "pow",
219 /* WINED3DSIH_RCP */ "rcp",
220 /* WINED3DSIH_REP */ "rep",
221 /* WINED3DSIH_RESINFO */ "resinfo",
222 /* WINED3DSIH_RET */ "ret",
223 /* WINED3DSIH_RETP */ "retp",
224 /* WINED3DSIH_ROUND_NE */ "round_ne",
225 /* WINED3DSIH_ROUND_NI */ "round_ni",
226 /* WINED3DSIH_ROUND_PI */ "round_pi",
227 /* WINED3DSIH_ROUND_Z */ "round_z",
228 /* WINED3DSIH_RSQ */ "rsq",
229 /* WINED3DSIH_SAMPLE */ "sample",
230 /* WINED3DSIH_SAMPLE_B */ "sample_b",
231 /* WINED3DSIH_SAMPLE_C */ "sample_c",
232 /* WINED3DSIH_SAMPLE_C_LZ */ "sample_c_lz",
233 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
234 /* WINED3DSIH_SAMPLE_INFO */ "sample_info",
235 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
236 /* WINED3DSIH_SAMPLE_POS */ "sample_pos",
237 /* WINED3DSIH_SETP */ "setp",
238 /* WINED3DSIH_SGE */ "sge",
239 /* WINED3DSIH_SGN */ "sgn",
240 /* WINED3DSIH_SINCOS */ "sincos",
241 /* WINED3DSIH_SLT */ "slt",
242 /* WINED3DSIH_SQRT */ "sqrt",
243 /* WINED3DSIH_STORE_RAW */ "store_raw",
244 /* WINED3DSIH_STORE_STRUCTURED */ "store_structured",
245 /* WINED3DSIH_STORE_UAV_TYPED */ "store_uav_typed",
246 /* WINED3DSIH_SUB */ "sub",
247 /* WINED3DSIH_SWAPC */ "swapc",
248 /* WINED3DSIH_SWITCH */ "switch",
249 /* WINED3DSIH_SYNC */ "sync",
250 /* WINED3DSIH_TEX */ "texld",
251 /* WINED3DSIH_TEXBEM */ "texbem",
252 /* WINED3DSIH_TEXBEML */ "texbeml",
253 /* WINED3DSIH_TEXCOORD */ "texcrd",
254 /* WINED3DSIH_TEXDEPTH */ "texdepth",
255 /* WINED3DSIH_TEXDP3 */ "texdp3",
256 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
257 /* WINED3DSIH_TEXKILL */ "texkill",
258 /* WINED3DSIH_TEXLDD */ "texldd",
259 /* WINED3DSIH_TEXLDL */ "texldl",
260 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
261 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
262 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
263 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
264 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
265 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
266 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
267 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
268 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
269 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
270 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
271 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
272 /* WINED3DSIH_UBFE */ "ubfe",
273 /* WINED3DSIH_UDIV */ "udiv",
274 /* WINED3DSIH_UGE */ "uge",
275 /* WINED3DSIH_ULT */ "ult",
276 /* WINED3DSIH_UMAX */ "umax",
277 /* WINED3DSIH_UMIN */ "umin",
278 /* WINED3DSIH_UMUL */ "umul",
279 /* WINED3DSIH_USHR */ "ushr",
280 /* WINED3DSIH_UTOF */ "utof",
281 /* WINED3DSIH_XOR */ "xor",
282 };
283
284 static const char * const semantic_names[] =
285 {
286 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
287 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
288 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
289 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
290 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
291 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
292 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
293 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
294 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
295 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
296 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
297 /* WINED3D_DECL_USAGE_FOG */ "FOG",
298 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
299 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
300 };
301
302 const char *debug_d3dshaderinstructionhandler(enum WINED3D_SHADER_INSTRUCTION_HANDLER handler_idx)
303 {
304 if (handler_idx >= ARRAY_SIZE(shader_opcode_names))
305 return wine_dbg_sprintf("UNRECOGNIZED(%#x)", handler_idx);
306
307 return shader_opcode_names[handler_idx];
308 }
309
310 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
311 {
312 if (usage >= ARRAY_SIZE(semantic_names))
313 {
314 FIXME("Unrecognized usage %#x.\n", usage);
315 return "UNRECOGNIZED";
316 }
317
318 return semantic_names[usage];
319 }
320
321 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
322 {
323 unsigned int i;
324
325 for (i = 0; i < ARRAY_SIZE(semantic_names); ++i)
326 {
327 if (!strcmp(name, semantic_names[i]))
328 return i;
329 }
330
331 return ~0U;
332 }
333
334 static enum wined3d_sysval_semantic shader_sysval_semantic_from_usage(enum wined3d_decl_usage usage)
335 {
336 switch (usage)
337 {
338 case WINED3D_DECL_USAGE_POSITION:
339 return WINED3D_SV_POSITION;
340 default:
341 return 0;
342 }
343 }
344
345 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
346 {
347 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
348 }
349
350 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
351 const struct wined3d_shader_semantic *s)
352 {
353 e->semantic_name = shader_semantic_name_from_usage(s->usage);
354 e->semantic_idx = s->usage_idx;
355 e->stream_idx = 0;
356 e->sysval_semantic = shader_sysval_semantic_from_usage(s->usage);
357 e->component_type = WINED3D_TYPE_FLOAT;
358 e->register_idx = s->reg.reg.idx[0].offset;
359 e->mask = s->reg.write_mask;
360 }
361
362 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
363 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
364 {
365 e->semantic_name = shader_semantic_name_from_usage(usage);
366 e->semantic_idx = usage_idx;
367 e->stream_idx = 0;
368 e->sysval_semantic = shader_sysval_semantic_from_usage(usage);
369 e->component_type = WINED3D_TYPE_FLOAT;
370 e->register_idx = reg_idx;
371 e->mask = write_mask;
372 }
373
374 static const struct wined3d_shader_frontend *shader_select_frontend(enum vkd3d_shader_source_type source_type)
375 {
376 switch (source_type)
377 {
378 case VKD3D_SHADER_SOURCE_D3D_BYTECODE:
379 return &sm1_shader_frontend;
380
381 case VKD3D_SHADER_SOURCE_DXBC_TPF:
382 return &sm4_shader_frontend;
383
384 default:
385 WARN("Invalid source type %#x specified.\n", source_type);
386 return NULL;
387 }
388 }
389
390 void string_buffer_clear(struct wined3d_string_buffer *buffer)
391 {
392 buffer->buffer[0] = '\0';
393 buffer->content_size = 0;
394 }
395
396 BOOL string_buffer_init(struct wined3d_string_buffer *buffer)
397 {
398 buffer->buffer_size = 32;
399 if (!(buffer->buffer = heap_alloc(buffer->buffer_size)))
400 {
401 ERR("Failed to allocate shader buffer memory.\n");
402 return FALSE;
403 }
404
405 string_buffer_clear(buffer);
406 return TRUE;
407 }
408
409 void string_buffer_free(struct wined3d_string_buffer *buffer)
410 {
411 heap_free(buffer->buffer);
412 }
413
414 BOOL string_buffer_resize(struct wined3d_string_buffer *buffer, int rc)
415 {
416 char *new_buffer;
417 unsigned int new_buffer_size = buffer->buffer_size * 2;
418
419 while (rc > 0 && (unsigned int)rc >= new_buffer_size - buffer->content_size)
420 new_buffer_size *= 2;
421 if (!(new_buffer = heap_realloc(buffer->buffer, new_buffer_size)))
422 {
423 ERR("Failed to grow buffer.\n");
424 buffer->buffer[buffer->content_size] = '\0';
425 return FALSE;
426 }
427 buffer->buffer = new_buffer;
428 buffer->buffer_size = new_buffer_size;
429 return TRUE;
430 }
431
432 int shader_vaddline(struct wined3d_string_buffer *buffer, const char *format, va_list args)
433 {
434 unsigned int rem;
435 int rc;
436
437 rem = buffer->buffer_size - buffer->content_size;
438 rc = vsnprintf(&buffer->buffer[buffer->content_size], rem, format, args);
439 if (rc < 0 /* C89 */ || (unsigned int)rc >= rem /* C99 */)
440 return rc;
441
442 buffer->content_size += rc;
443 return 0;
444 }
445
446 int shader_addline(struct wined3d_string_buffer *buffer, const char *format, ...)
447 {
448 va_list args;
449 int ret;
450
451 for (;;)
452 {
453 va_start(args, format);
454 ret = shader_vaddline(buffer, format, args);
455 va_end(args);
456 if (!ret)
457 return ret;
458 if (!string_buffer_resize(buffer, ret))
459 return -1;
460 }
461 }
462
463 struct wined3d_string_buffer *string_buffer_get(struct wined3d_string_buffer_list *list)
464 {
465 struct wined3d_string_buffer *buffer;
466
467 if (list_empty(&list->list))
468 {
469 buffer = heap_alloc(sizeof(*buffer));
470 if (!buffer || !string_buffer_init(buffer))
471 {
472 ERR("Couldn't allocate buffer for temporary string.\n");
473 heap_free(buffer);
474 return NULL;
475 }
476 }
477 else
478 {
479 buffer = LIST_ENTRY(list_head(&list->list), struct wined3d_string_buffer, entry);
480 list_remove(&buffer->entry);
481 }
482 string_buffer_clear(buffer);
483 return buffer;
484 }
485
486 static int string_buffer_vsprintf(struct wined3d_string_buffer *buffer, const char *format, va_list args)
487 {
488 if (!buffer)
489 return 0;
490 string_buffer_clear(buffer);
491 return shader_vaddline(buffer, format, args);
492 }
493
494 void string_buffer_sprintf(struct wined3d_string_buffer *buffer, const char *format, ...)
495 {
496 va_list args;
497 int ret;
498
499 for (;;)
500 {
501 va_start(args, format);
502 ret = string_buffer_vsprintf(buffer, format, args);
503 va_end(args);
504 if (!ret)
505 return;
506 if (!string_buffer_resize(buffer, ret))
507 return;
508 }
509 }
510
511 void string_buffer_release(struct wined3d_string_buffer_list *list, struct wined3d_string_buffer *buffer)
512 {
513 if (!buffer)
514 return;
515 list_add_head(&list->list, &buffer->entry);
516 }
517
518 void string_buffer_list_init(struct wined3d_string_buffer_list *list)
519 {
520 list_init(&list->list);
521 }
522
523 void string_buffer_list_cleanup(struct wined3d_string_buffer_list *list)
524 {
525 struct wined3d_string_buffer *buffer, *buffer_next;
526
527 LIST_FOR_EACH_ENTRY_SAFE(buffer, buffer_next, &list->list, struct wined3d_string_buffer, entry)
528 {
529 string_buffer_free(buffer);
530 heap_free(buffer);
531 }
532 list_init(&list->list);
533 }
534
535 static void shader_delete_constant_list(struct list *clist)
536 {
537 struct wined3d_shader_lconst *constant, *constant_next;
538
539 LIST_FOR_EACH_ENTRY_SAFE(constant, constant_next, clist, struct wined3d_shader_lconst, entry)
540 heap_free(constant);
541 list_init(clist);
542 }
543
544 static void shader_set_limits(struct wined3d_shader *shader)
545 {
546 static const struct limits_entry
547 {
548 unsigned int min_version;
549 unsigned int max_version;
550 struct wined3d_shader_limits limits;
551 }
552 vs_limits[] =
553 {
554 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
555 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 1), { 0, 0, 256, 0, 12, 0}},
556 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 1), { 0, 16, 256, 16, 12, 0}},
557 /* DX10 cards on Windows advertise a D3D9 constant limit of 256
558 * even though they are capable of supporting much more (GL
559 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
560 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
561 * shaders to 256. */
562 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), { 4, 16, 256, 16, 12, 0}},
563 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 16, 0}},
564 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 0}},
565 {0}
566 },
567 hs_limits[] =
568 {
569 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
570 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
571 },
572 ds_limits[] =
573 {
574 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
575 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
576 },
577 gs_limits[] =
578 {
579 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
580 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 32, 16}},
581 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
582 {0}
583 },
584 ps_limits[] =
585 {
586 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
587 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 3), { 4, 0, 8, 0, 0, 0}},
588 {WINED3D_SHADER_VERSION(1, 4), WINED3D_SHADER_VERSION(1, 4), { 6, 0, 8, 0, 0, 0}},
589 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 0), {16, 0, 32, 0, 0, 0}},
590 {WINED3D_SHADER_VERSION(2, 1), WINED3D_SHADER_VERSION(2, 1), {16, 16, 32, 16, 0, 0}},
591 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), {16, 16, 224, 16, 0, 10}},
592 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 32}},
593 {0}
594 },
595 cs_limits[] =
596 {
597 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
598 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 0}},
599 };
600 const struct limits_entry *limits_array;
601 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
602 shader->reg_maps.shader_version.minor);
603 int i = 0;
604
605 switch (shader->reg_maps.shader_version.type)
606 {
607 default:
608 FIXME("Unexpected shader type %u found.\n", shader->reg_maps.shader_version.type);
609 /* Fall-through. */
610 case WINED3D_SHADER_TYPE_VERTEX:
611 limits_array = vs_limits;
612 break;
613 case WINED3D_SHADER_TYPE_HULL:
614 limits_array = hs_limits;
615 break;
616 case WINED3D_SHADER_TYPE_DOMAIN:
617 limits_array = ds_limits;
618 break;
619 case WINED3D_SHADER_TYPE_GEOMETRY:
620 limits_array = gs_limits;
621 break;
622 case WINED3D_SHADER_TYPE_PIXEL:
623 limits_array = ps_limits;
624 break;
625 case WINED3D_SHADER_TYPE_COMPUTE:
626 limits_array = cs_limits;
627 break;
628 }
629
630 while (limits_array[i].min_version && limits_array[i].min_version <= shader_version)
631 {
632 if (shader_version <= limits_array[i].max_version)
633 {
634 shader->limits = &limits_array[i].limits;
635 break;
636 }
637 ++i;
638 }
639 if (!shader->limits)
640 {
641 FIXME("Unexpected shader version \"%u.%u\".\n",
642 shader->reg_maps.shader_version.major,
643 shader->reg_maps.shader_version.minor);
644 shader->limits = &limits_array[max(0, i - 1)].limits;
645 }
646 }
647
648 static BOOL shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
649 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type, unsigned int constf_size)
650 {
651 switch (reg->type)
652 {
653 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
654 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
655 reg_maps->texcoord |= 1u << reg->idx[0].offset;
656 else
657 reg_maps->address |= 1u << reg->idx[0].offset;
658 break;
659
660 case WINED3DSPR_TEMP:
661 reg_maps->temporary |= 1u << reg->idx[0].offset;
662 break;
663
664 case WINED3DSPR_INPUT:
665 if (reg->idx[0].rel_addr)
666 reg_maps->input_rel_addressing = 1;
667 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
668 {
669 /* If relative addressing is used, we must assume that all
670 * registers are used. Even if it is a construct like v3[aL],
671 * we can't assume that v0, v1 and v2 aren't read because aL
672 * can be negative. */
673 if (reg->idx[0].rel_addr)
674 shader->u.ps.input_reg_used = ~0u;
675 else
676 shader->u.ps.input_reg_used |= 1u << reg->idx[0].offset;
677 }
678 else
679 {
680 reg_maps->input_registers |= 1u << reg->idx[0].offset;
681 }
682 break;
683
684 case WINED3DSPR_RASTOUT:
685 if (reg->idx[0].offset == 1)
686 reg_maps->fog = 1;
687 if (reg->idx[0].offset == 2)
688 reg_maps->point_size = 1;
689 break;
690
691 case WINED3DSPR_MISCTYPE:
692 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
693 {
694 if (!reg->idx[0].offset)
695 reg_maps->vpos = 1;
696 else if (reg->idx[0].offset == 1)
697 reg_maps->usesfacing = 1;
698 }
699 break;
700
701 case WINED3DSPR_CONST:
702 if (reg->idx[0].rel_addr)
703 {
704 if (reg->idx[0].offset < reg_maps->min_rel_offset)
705 reg_maps->min_rel_offset = reg->idx[0].offset;
706 if (reg->idx[0].offset > reg_maps->max_rel_offset)
707 reg_maps->max_rel_offset = reg->idx[0].offset;
708 reg_maps->usesrelconstF = TRUE;
709 }
710 else
711 {
712 if (reg->idx[0].offset >= min(shader->limits->constant_float, constf_size))
713 {
714 WARN("Shader using float constant %u which is not supported.\n", reg->idx[0].offset);
715 return FALSE;
716 }
717 else
718 {
719 wined3d_insert_bits(reg_maps->constf, reg->idx[0].offset, 1, 0x1);
720 }
721 }
722 break;
723
724 case WINED3DSPR_CONSTINT:
725 if (reg->idx[0].offset >= shader->limits->constant_int)
726 {
727 WARN("Shader using integer constant %u which is not supported.\n", reg->idx[0].offset);
728 return FALSE;
729 }
730 else
731 {
732 reg_maps->integer_constants |= (1u << reg->idx[0].offset);
733 }
734 break;
735
736 case WINED3DSPR_CONSTBOOL:
737 if (reg->idx[0].offset >= shader->limits->constant_bool)
738 {
739 WARN("Shader using bool constant %u which is not supported.\n", reg->idx[0].offset);
740 return FALSE;
741 }
742 else
743 {
744 reg_maps->boolean_constants |= (1u << reg->idx[0].offset);
745 }
746 break;
747
748 case WINED3DSPR_COLOROUT:
749 reg_maps->rt_mask |= (1u << reg->idx[0].offset);
750 break;
751
752 case WINED3DSPR_OUTCONTROLPOINT:
753 reg_maps->vocp = 1;
754 break;
755
756 case WINED3DSPR_SAMPLEMASK:
757 reg_maps->sample_mask = 1;
758 break;
759
760 default:
761 TRACE("Not recording register of type %#x and [%#x][%#x].\n",
762 reg->type, reg->idx[0].offset, reg->idx[1].offset);
763 break;
764 }
765 return TRUE;
766 }
767
768 static void shader_record_sample(struct wined3d_shader_reg_maps *reg_maps,
769 unsigned int resource_idx, unsigned int sampler_idx, unsigned int bind_idx)
770 {
771 struct wined3d_shader_sampler_map_entry *entries, *entry;
772 struct wined3d_shader_sampler_map *map;
773 unsigned int i;
774
775 map = &reg_maps->sampler_map;
776 entries = map->entries;
777 for (i = 0; i < map->count; ++i)
778 {
779 if (entries[i].resource_idx == resource_idx && entries[i].sampler_idx == sampler_idx)
780 return;
781 }
782
783 if (!map->size)
784 {
785 if (!(entries = heap_calloc(4, sizeof(*entries))))
786 {
787 ERR("Failed to allocate sampler map entries.\n");
788 return;
789 }
790 map->size = 4;
791 map->entries = entries;
792 }
793 else if (map->count == map->size)
794 {
795 size_t new_size = map->size * 2;
796
797 if (sizeof(*entries) * new_size <= sizeof(*entries) * map->size
798 || !(entries = heap_realloc(entries, sizeof(*entries) * new_size)))
799 {
800 ERR("Failed to resize sampler map entries.\n");
801 return;
802 }
803 map->size = new_size;
804 map->entries = entries;
805 }
806
807 entry = &entries[map->count++];
808 entry->resource_idx = resource_idx;
809 entry->sampler_idx = sampler_idx;
810 entry->bind_idx = bind_idx;
811 }
812
813 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
814 {
815 switch (instr)
816 {
817 case WINED3DSIH_M4x4:
818 case WINED3DSIH_M3x4:
819 return param == 1 ? 3 : 0;
820
821 case WINED3DSIH_M4x3:
822 case WINED3DSIH_M3x3:
823 return param == 1 ? 2 : 0;
824
825 case WINED3DSIH_M3x2:
826 return param == 1 ? 1 : 0;
827
828 default:
829 return 0;
830 }
831 }
832
833 static HRESULT shader_reg_maps_add_tgsm(struct wined3d_shader_reg_maps *reg_maps,
834 unsigned int register_idx, unsigned int size, unsigned int stride)
835 {
836 struct wined3d_shader_tgsm *tgsm;
837
838 if (register_idx >= MAX_TGSM_REGISTERS)
839 {
840 ERR("Invalid TGSM register index %u.\n", register_idx);
841 return S_OK;
842 }
843 if (reg_maps->shader_version.type != WINED3D_SHADER_TYPE_COMPUTE)
844 {
845 FIXME("TGSM declarations are allowed only in compute shaders.\n");
846 return S_OK;
847 }
848
849 if (!wined3d_array_reserve((void **)&reg_maps->tgsm, &reg_maps->tgsm_capacity,
850 register_idx + 1, sizeof(*reg_maps->tgsm)))
851 return E_OUTOFMEMORY;
852
853 reg_maps->tgsm_count = max(register_idx + 1, reg_maps->tgsm_count);
854 tgsm = &reg_maps->tgsm[register_idx];
855 tgsm->size = size;
856 tgsm->stride = stride;
857 return S_OK;
858 }
859
860 static HRESULT shader_record_shader_phase(struct wined3d_shader *shader,
861 struct wined3d_shader_phase **current_phase, const struct wined3d_shader_instruction *ins,
862 const DWORD *current_instruction_ptr, const DWORD *previous_instruction_ptr)
863 {
864 struct wined3d_shader_phase *phase;
865
866 if ((phase = *current_phase))
867 {
868 phase->end = previous_instruction_ptr;
869 *current_phase = NULL;
870 }
871
872 if (shader->reg_maps.shader_version.type != WINED3D_SHADER_TYPE_HULL)
873 {
874 ERR("Unexpected shader type %s.\n", debug_shader_type(shader->reg_maps.shader_version.type));
875 return E_FAIL;
876 }
877
878 switch (ins->handler_idx)
879 {
880 case WINED3DSIH_HS_CONTROL_POINT_PHASE:
881 if (shader->u.hs.phases.control_point)
882 {
883 FIXME("Multiple control point phases.\n");
884 heap_free(shader->u.hs.phases.control_point);
885 }
886 if (!(shader->u.hs.phases.control_point = heap_alloc_zero(sizeof(*shader->u.hs.phases.control_point))))
887 return E_OUTOFMEMORY;
888 phase = shader->u.hs.phases.control_point;
889 break;
890 case WINED3DSIH_HS_FORK_PHASE:
891 if (!wined3d_array_reserve((void **)&shader->u.hs.phases.fork,
892 &shader->u.hs.phases.fork_size, shader->u.hs.phases.fork_count + 1,
893 sizeof(*shader->u.hs.phases.fork)))
894 return E_OUTOFMEMORY;
895 phase = &shader->u.hs.phases.fork[shader->u.hs.phases.fork_count++];
896 break;
897 case WINED3DSIH_HS_JOIN_PHASE:
898 if (!wined3d_array_reserve((void **)&shader->u.hs.phases.join,
899 &shader->u.hs.phases.join_size, shader->u.hs.phases.join_count + 1,
900 sizeof(*shader->u.hs.phases.join)))
901 return E_OUTOFMEMORY;
902 phase = &shader->u.hs.phases.join[shader->u.hs.phases.join_count++];
903 break;
904 default:
905 ERR("Unexpected opcode %s.\n", debug_d3dshaderinstructionhandler(ins->handler_idx));
906 return E_FAIL;
907 }
908
909 phase->start = current_instruction_ptr;
910 *current_phase = phase;
911
912 return WINED3D_OK;
913 }
914
915 static HRESULT shader_calculate_clip_or_cull_distance_mask(
916 const struct wined3d_shader_signature_element *e, unsigned int *mask)
917 {
918 /* Clip and cull distances are packed in 4 component registers. 0 and 1 are
919 * the only allowed semantic indices.
920 */
921 if (e->semantic_idx >= WINED3D_MAX_CLIP_DISTANCES / 4)
922 {
923 *mask = 0;
924 WARN("Invalid clip/cull distance index %u.\n", e->semantic_idx);
925 return WINED3DERR_INVALIDCALL;
926 }
927
928 *mask = (e->mask & WINED3DSP_WRITEMASK_ALL) << (4 * e->semantic_idx);
929 return WINED3D_OK;
930 }
931
932 static void wined3d_insert_interpolation_mode(uint32_t *packed_interpolation_mode,
933 unsigned int register_idx, enum wined3d_shader_interpolation_mode mode)
934 {
935 if (mode > WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE)
936 FIXME("Unexpected interpolation mode %#x.\n", mode);
937
938 wined3d_insert_bits(packed_interpolation_mode,
939 register_idx * WINED3D_PACKED_INTERPOLATION_BIT_COUNT, WINED3D_PACKED_INTERPOLATION_BIT_COUNT, mode);
940 }
941
942 static HRESULT shader_scan_output_signature(struct wined3d_shader *shader)
943 {
944 const struct wined3d_shader_signature *output_signature = &shader->output_signature;
945 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
946 unsigned int i;
947 HRESULT hr;
948
949 for (i = 0; i < output_signature->element_count; ++i)
950 {
951 const struct wined3d_shader_signature_element *e = &output_signature->elements[i];
952 unsigned int mask;
953
954 reg_maps->output_registers |= 1u << e->register_idx;
955 if (e->sysval_semantic == WINED3D_SV_CLIP_DISTANCE)
956 {
957 if (FAILED(hr = shader_calculate_clip_or_cull_distance_mask(e, &mask)))
958 return hr;
959 reg_maps->clip_distance_mask |= mask;
960 }
961 else if (e->sysval_semantic == WINED3D_SV_CULL_DISTANCE)
962 {
963 if (FAILED(hr = shader_calculate_clip_or_cull_distance_mask(e, &mask)))
964 return hr;
965 reg_maps->cull_distance_mask |= mask;
966 }
967 else if (e->sysval_semantic == WINED3D_SV_VIEWPORT_ARRAY_INDEX)
968 {
969 reg_maps->viewport_array = 1;
970 }
971 }
972
973 return WINED3D_OK;
974 }
975
976 /* Note that this does not count the loop register as an address register. */
977 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, DWORD constf_size)
978 {
979 struct wined3d_shader_signature_element input_signature_elements[max(MAX_ATTRIBS, MAX_REG_INPUT)];
980 struct wined3d_shader_signature_element output_signature_elements[MAX_REG_OUTPUT];
981 struct wined3d_shader_signature *output_signature = &shader->output_signature;
982 struct wined3d_shader_signature *input_signature = &shader->input_signature;
983 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
984 const struct wined3d_shader_frontend *fe = shader->frontend;
985 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
986 struct wined3d_shader_version shader_version;
987 struct wined3d_shader_phase *phase = NULL;
988 const DWORD *ptr, *prev_ins, *current_ins;
989 void *fe_data = shader->frontend_data;
990 unsigned int i;
991 HRESULT hr;
992
993 memset(reg_maps, 0, sizeof(*reg_maps));
994 memset(input_signature_elements, 0, sizeof(input_signature_elements));
995 memset(output_signature_elements, 0, sizeof(output_signature_elements));
996 reg_maps->min_rel_offset = ~0U;
997 list_init(&reg_maps->indexable_temps);
998
999 fe->shader_read_header(fe_data, &ptr, &shader_version);
1000 prev_ins = current_ins = ptr;
1001 reg_maps->shader_version = shader_version;
1002
1003 shader_set_limits(shader);
1004
1005 if (!(reg_maps->constf = heap_calloc(((min(shader->limits->constant_float, constf_size) + 31) / 32),
1006 sizeof(*reg_maps->constf))))
1007 {
1008 ERR("Failed to allocate constant map memory.\n");
1009 return E_OUTOFMEMORY;
1010 }
1011
1012 while (!fe->shader_is_end(fe_data, &ptr))
1013 {
1014 struct wined3d_shader_instruction ins;
1015
1016 current_ins = ptr;
1017 /* Fetch opcode. */
1018 fe->shader_read_instruction(fe_data, &ptr, &ins);
1019
1020 /* Unhandled opcode, and its parameters. */
1021 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1022 {
1023 WARN("Encountered unrecognised or invalid instruction.\n");
1024 return WINED3DERR_INVALIDCALL;
1025 }
1026
1027 /* Handle declarations. */
1028 if (ins.handler_idx == WINED3DSIH_DCL
1029 || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
1030 {
1031 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
1032 unsigned int reg_idx = semantic->reg.reg.idx[0].offset;
1033
1034 switch (semantic->reg.reg.type)
1035 {
1036 /* Mark input registers used. */
1037 case WINED3DSPR_INPUT:
1038 if (reg_idx >= MAX_REG_INPUT)
1039 {
1040 ERR("Invalid input register index %u.\n", reg_idx);
1041 break;
1042 }
1043 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL && shader_version.major == 3
1044 && semantic->usage == WINED3D_DECL_USAGE_POSITION && !semantic->usage_idx)
1045 return WINED3DERR_INVALIDCALL;
1046 reg_maps->input_registers |= 1u << reg_idx;
1047 shader_signature_from_semantic(&input_signature_elements[reg_idx], semantic);
1048 break;
1049
1050 /* Vertex shader: mark 3.0 output registers used, save token. */
1051 case WINED3DSPR_OUTPUT:
1052 if (reg_idx >= MAX_REG_OUTPUT)
1053 {
1054 ERR("Invalid output register index %u.\n", reg_idx);
1055 break;
1056 }
1057 reg_maps->output_registers |= 1u << reg_idx;
1058 shader_signature_from_semantic(&output_signature_elements[reg_idx], semantic);
1059 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
1060 reg_maps->fog = 1;
1061 if (semantic->usage == WINED3D_DECL_USAGE_PSIZE)
1062 reg_maps->point_size = 1;
1063 break;
1064
1065 case WINED3DSPR_SAMPLER:
1066 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1067 case WINED3DSPR_RESOURCE:
1068 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1069 {
1070 ERR("Invalid resource index %u.\n", reg_idx);
1071 break;
1072 }
1073 reg_maps->resource_info[reg_idx].type = semantic->resource_type;
1074 if (semantic->resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2DMS && semantic->sample_count == 1)
1075 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1076 if (semantic->resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY && semantic->sample_count == 1)
1077 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY;
1078 reg_maps->resource_info[reg_idx].data_type = semantic->resource_data_type;
1079 wined3d_bitmap_set(reg_maps->resource_map, reg_idx);
1080 break;
1081
1082 case WINED3DSPR_UAV:
1083 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1084 {
1085 ERR("Invalid UAV resource index %u.\n", reg_idx);
1086 break;
1087 }
1088 reg_maps->uav_resource_info[reg_idx].type = semantic->resource_type;
1089 reg_maps->uav_resource_info[reg_idx].data_type = semantic->resource_data_type;
1090 if (ins.flags)
1091 FIXME("Ignoring typed UAV flags %#x.\n", ins.flags);
1092 break;
1093
1094 default:
1095 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
1096 break;
1097 }
1098 }
1099 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
1100 {
1101 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
1102 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
1103 {
1104 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
1105 }
1106 else
1107 {
1108 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
1109 wined3d_bitmap_set(&reg_maps->cb_map, reg->idx[0].offset);
1110 }
1111 }
1112 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
1113 {
1114 if (ins.flags & WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL)
1115 {
1116 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1117 shader->u.ps.force_early_depth_stencil = TRUE;
1118 else
1119 FIXME("Invalid instruction %#x for shader type %#x.\n",
1120 ins.handler_idx, shader_version.type);
1121 }
1122 else
1123 {
1124 WARN("Ignoring global flags %#x.\n", ins.flags);
1125 }
1126 }
1127 else if (ins.handler_idx == WINED3DSIH_DCL_GS_INSTANCES)
1128 {
1129 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1130 shader->u.gs.instance_count = ins.declaration.count;
1131 else
1132 FIXME("Invalid instruction %#x for shader type %#x.\n",
1133 ins.handler_idx, shader_version.type);
1134 }
1135 else if (ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
1136 || ins.handler_idx == WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT)
1137 {
1138 if (phase)
1139 phase->instance_count = ins.declaration.count;
1140 else
1141 FIXME("Instruction %s outside of shader phase.\n",
1142 debug_d3dshaderinstructionhandler(ins.handler_idx));
1143 }
1144 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
1145 {
1146 if (reg_maps->icb)
1147 FIXME("Multiple immediate constant buffers.\n");
1148 reg_maps->icb = ins.declaration.icb;
1149 }
1150 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
1151 {
1152 if (phase)
1153 {
1154 FIXME("Indexable temporary registers not supported.\n");
1155 }
1156 else
1157 {
1158 struct wined3d_shader_indexable_temp *reg;
1159
1160 if (!(reg = heap_alloc(sizeof(*reg))))
1161 return E_OUTOFMEMORY;
1162
1163 *reg = ins.declaration.indexable_temp;
1164 list_add_tail(&reg_maps->indexable_temps, &reg->entry);
1165 }
1166 }
1167 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
1168 {
1169 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1170 shader->u.gs.input_type = ins.declaration.primitive_type.type;
1171 else
1172 FIXME("Invalid instruction %#x for shader type %#x.\n",
1173 ins.handler_idx, shader_version.type);
1174 }
1175 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
1176 {
1177 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1178 if (reg_idx >= MAX_REG_INPUT)
1179 {
1180 ERR("Invalid register index %u.\n", reg_idx);
1181 break;
1182 }
1183 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1184 wined3d_insert_interpolation_mode(shader->u.ps.interpolation_mode, reg_idx, ins.flags);
1185 else
1186 FIXME("Invalid instruction %#x for shader type %#x.\n",
1187 ins.handler_idx, shader_version.type);
1188 }
1189 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
1190 {
1191 if (ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUT
1192 || ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUTGE
1193 || ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUTLE)
1194 {
1195 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1196 shader->u.ps.depth_output = ins.declaration.dst.reg.type;
1197 else
1198 FIXME("Invalid instruction %#x for shader type %#x.\n",
1199 ins.handler_idx, shader_version.type);
1200 }
1201 }
1202 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT)
1203 {
1204 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1205 shader->u.hs.output_vertex_count = ins.declaration.count;
1206 else
1207 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1208 }
1209 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
1210 {
1211 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1212 shader->u.gs.output_type = ins.declaration.primitive_type.type;
1213 else
1214 FIXME("Invalid instruction %#x for shader type %#x.\n",
1215 ins.handler_idx, shader_version.type);
1216 }
1217 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_RAW)
1218 {
1219 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1220 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1221 {
1222 ERR("Invalid resource index %u.\n", reg_idx);
1223 break;
1224 }
1225 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1226 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1227 reg_maps->resource_info[reg_idx].flags = WINED3D_VIEW_BUFFER_RAW;
1228 wined3d_bitmap_set(reg_maps->resource_map, reg_idx);
1229 }
1230 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
1231 {
1232 unsigned int reg_idx = ins.declaration.structured_resource.reg.reg.idx[0].offset;
1233 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1234 {
1235 ERR("Invalid resource index %u.\n", reg_idx);
1236 break;
1237 }
1238 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1239 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1240 reg_maps->resource_info[reg_idx].flags = 0;
1241 reg_maps->resource_info[reg_idx].stride = ins.declaration.structured_resource.byte_stride / 4;
1242 wined3d_bitmap_set(reg_maps->resource_map, reg_idx);
1243 }
1244 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
1245 {
1246 if (ins.flags & WINED3DSI_SAMPLER_COMPARISON_MODE)
1247 reg_maps->sampler_comparison_mode |= (1u << ins.declaration.dst.reg.idx[0].offset);
1248 }
1249 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS)
1250 {
1251 if (phase)
1252 phase->temporary_count = ins.declaration.count;
1253 else
1254 reg_maps->temporary_count = ins.declaration.count;
1255 }
1256 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
1257 {
1258 if (shader_version.type == WINED3D_SHADER_TYPE_DOMAIN)
1259 shader->u.ds.tessellator_domain = ins.declaration.tessellator_domain;
1260 else if (shader_version.type != WINED3D_SHADER_TYPE_HULL)
1261 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1262 }
1263 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
1264 {
1265 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1266 shader->u.hs.tessellator_output_primitive = ins.declaration.tessellator_output_primitive;
1267 else
1268 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1269 }
1270 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
1271 {
1272 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1273 shader->u.hs.tessellator_partitioning = ins.declaration.tessellator_partitioning;
1274 else
1275 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1276 }
1277 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
1278 {
1279 if (FAILED(hr = shader_reg_maps_add_tgsm(reg_maps, ins.declaration.tgsm_raw.reg.reg.idx[0].offset,
1280 ins.declaration.tgsm_raw.byte_count / 4, 0)))
1281 return hr;
1282 }
1283 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
1284 {
1285 unsigned int stride = ins.declaration.tgsm_structured.byte_stride / 4;
1286 unsigned int size = stride * ins.declaration.tgsm_structured.structure_count;
1287 if (FAILED(hr = shader_reg_maps_add_tgsm(reg_maps,
1288 ins.declaration.tgsm_structured.reg.reg.idx[0].offset, size, stride)))
1289 return hr;
1290 }
1291 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
1292 {
1293 if (shader_version.type == WINED3D_SHADER_TYPE_COMPUTE)
1294 {
1295 shader->u.cs.thread_group_size = ins.declaration.thread_group_size;
1296 }
1297 else
1298 {
1299 FIXME("Invalid instruction %#x for shader type %#x.\n",
1300 ins.handler_idx, shader_version.type);
1301 }
1302 }
1303 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
1304 {
1305 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1306 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1307 {
1308 ERR("Invalid UAV resource index %u.\n", reg_idx);
1309 break;
1310 }
1311 if (ins.flags)
1312 FIXME("Ignoring raw UAV flags %#x.\n", ins.flags);
1313 reg_maps->uav_resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1314 reg_maps->uav_resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1315 reg_maps->uav_resource_info[reg_idx].flags = WINED3D_VIEW_BUFFER_RAW;
1316 }
1317 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
1318 {
1319 unsigned int reg_idx = ins.declaration.structured_resource.reg.reg.idx[0].offset;
1320 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1321 {
1322 ERR("Invalid UAV resource index %u.\n", reg_idx);
1323 break;
1324 }
1325 if (ins.flags)
1326 FIXME("Ignoring structured UAV flags %#x.\n", ins.flags);
1327 reg_maps->uav_resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1328 reg_maps->uav_resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1329 reg_maps->uav_resource_info[reg_idx].flags = 0;
1330 reg_maps->uav_resource_info[reg_idx].stride = ins.declaration.structured_resource.byte_stride / 4;
1331 }
1332 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1333 {
1334 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1335 shader->u.gs.vertices_out = ins.declaration.count;
1336 else
1337 FIXME("Invalid instruction %#x for shader type %#x.\n",
1338 ins.handler_idx, shader_version.type);
1339 }
1340 else if (ins.handler_idx == WINED3DSIH_DEF)
1341 {
1342 struct wined3d_shader_lconst *lconst;
1343 float *value;
1344
1345 if (!(lconst = heap_alloc(sizeof(*lconst))))
1346 return E_OUTOFMEMORY;
1347
1348 lconst->idx = ins.dst[0].reg.idx[0].offset;
1349 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1350 value = (float *)lconst->value;
1351
1352 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
1353 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1354 {
1355 if (value[0] < -1.0f) value[0] = -1.0f;
1356 else if (value[0] > 1.0f) value[0] = 1.0f;
1357 if (value[1] < -1.0f) value[1] = -1.0f;
1358 else if (value[1] > 1.0f) value[1] = 1.0f;
1359 if (value[2] < -1.0f) value[2] = -1.0f;
1360 else if (value[2] > 1.0f) value[2] = 1.0f;
1361 if (value[3] < -1.0f) value[3] = -1.0f;
1362 else if (value[3] > 1.0f) value[3] = 1.0f;
1363 }
1364
1365 list_add_head(&shader->constantsF, &lconst->entry);
1366
1367 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
1368 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
1369 {
1370 shader->lconst_inf_or_nan = TRUE;
1371 }
1372 }
1373 else if (ins.handler_idx == WINED3DSIH_DEFI)
1374 {
1375 struct wined3d_shader_lconst *lconst;
1376
1377 if (!(lconst = heap_alloc(sizeof(*lconst))))
1378 return E_OUTOFMEMORY;
1379
1380 lconst->idx = ins.dst[0].reg.idx[0].offset;
1381 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1382
1383 list_add_head(&shader->constantsI, &lconst->entry);
1384 reg_maps->local_int_consts |= (1u << lconst->idx);
1385 }
1386 else if (ins.handler_idx == WINED3DSIH_DEFB)
1387 {
1388 struct wined3d_shader_lconst *lconst;
1389
1390 if (!(lconst = heap_alloc(sizeof(*lconst))))
1391 return E_OUTOFMEMORY;
1392
1393 lconst->idx = ins.dst[0].reg.idx[0].offset;
1394 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, sizeof(DWORD));
1395
1396 list_add_head(&shader->constantsB, &lconst->entry);
1397 reg_maps->local_bool_consts |= (1u << lconst->idx);
1398 }
1399 /* Handle shader phases. */
1400 else if (ins.handler_idx == WINED3DSIH_HS_CONTROL_POINT_PHASE
1401 || ins.handler_idx == WINED3DSIH_HS_FORK_PHASE
1402 || ins.handler_idx == WINED3DSIH_HS_JOIN_PHASE)
1403 {
1404 if (FAILED(hr = shader_record_shader_phase(shader, &phase, &ins, current_ins, prev_ins)))
1405 return hr;
1406 }
1407 /* For subroutine prototypes. */
1408 else if (ins.handler_idx == WINED3DSIH_LABEL)
1409 {
1410 reg_maps->labels |= 1u << ins.src[0].reg.idx[0].offset;
1411 }
1412 /* Set texture, address, temporary registers. */
1413 else
1414 {
1415 BOOL color0_mov = FALSE;
1416 unsigned int i;
1417
1418 /* This will loop over all the registers and try to
1419 * make a bitmask of the ones we're interested in.
1420 *
1421 * Relative addressing tokens are ignored, but that's
1422 * okay, since we'll catch any address registers when
1423 * they are initialized (required by spec). */
1424 for (i = 0; i < ins.dst_count; ++i)
1425 {
1426 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
1427 shader_version.type, constf_size))
1428 return WINED3DERR_INVALIDCALL;
1429
1430 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1431 {
1432 UINT idx = ins.dst[i].reg.idx[0].offset;
1433
1434 switch (ins.dst[i].reg.type)
1435 {
1436 case WINED3DSPR_RASTOUT:
1437 if (shader_version.major >= 3)
1438 break;
1439 switch (idx)
1440 {
1441 case 0: /* oPos */
1442 reg_maps->output_registers |= 1u << 10;
1443 shader_signature_from_usage(&output_signature_elements[10],
1444 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
1445 break;
1446
1447 case 1: /* oFog */
1448 reg_maps->output_registers |= 1u << 11;
1449 shader_signature_from_usage(&output_signature_elements[11],
1450 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
1451 break;
1452
1453 case 2: /* oPts */
1454 reg_maps->output_registers |= 1u << 11;
1455 shader_signature_from_usage(&output_signature_elements[11],
1456 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
1457 break;
1458 }
1459 break;
1460
1461 case WINED3DSPR_ATTROUT:
1462 if (shader_version.major >= 3)
1463 break;
1464 if (idx < 2)
1465 {
1466 idx += 8;
1467 if (reg_maps->output_registers & (1u << idx))
1468 {
1469 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1470 }
1471 else
1472 {
1473 reg_maps->output_registers |= 1u << idx;
1474 shader_signature_from_usage(&output_signature_elements[idx],
1475 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
1476 }
1477 }
1478 break;
1479
1480 case WINED3DSPR_TEXCRDOUT: /* WINED3DSPR_OUTPUT */
1481 if (shader_version.major >= 3)
1482 {
1483 if (idx >= ARRAY_SIZE(reg_maps->u.output_registers_mask))
1484 {
1485 WARN("Invalid output register index %u.\n", idx);
1486 break;
1487 }
1488 reg_maps->u.output_registers_mask[idx] |= ins.dst[i].write_mask;
1489 break;
1490 }
1491 if (idx >= ARRAY_SIZE(reg_maps->u.texcoord_mask))
1492 {
1493 WARN("Invalid texcoord index %u.\n", idx);
1494 break;
1495 }
1496 reg_maps->u.texcoord_mask[idx] |= ins.dst[i].write_mask;
1497 if (reg_maps->output_registers & (1u << idx))
1498 {
1499 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1500 }
1501 else
1502 {
1503 reg_maps->output_registers |= 1u << idx;
1504 shader_signature_from_usage(&output_signature_elements[idx],
1505 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
1506 }
1507 break;
1508
1509 default:
1510 break;
1511 }
1512 }
1513
1514 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1515 {
1516 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
1517 {
1518 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
1519 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
1520 * the mov and perform the sRGB write correction from the source register.
1521 *
1522 * However, if the mov is only partial, we can't do this, and if the write
1523 * comes from an instruction other than MOV it is hard to do as well. If
1524 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
1525 shader->u.ps.color0_mov = FALSE;
1526 if (ins.handler_idx == WINED3DSIH_MOV
1527 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
1528 {
1529 /* Used later when the source register is read. */
1530 color0_mov = TRUE;
1531 }
1532 }
1533 /* Also drop the MOV marker if the source register is overwritten prior to the shader
1534 * end
1535 */
1536 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
1537 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
1538 {
1539 shader->u.ps.color0_mov = FALSE;
1540 }
1541 }
1542
1543 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
1544 if (shader_version.major == 1
1545 && (ins.handler_idx == WINED3DSIH_TEX
1546 || ins.handler_idx == WINED3DSIH_TEXBEM
1547 || ins.handler_idx == WINED3DSIH_TEXBEML
1548 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
1549 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
1550 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
1551 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
1552 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
1553 || ins.handler_idx == WINED3DSIH_TEXREG2AR
1554 || ins.handler_idx == WINED3DSIH_TEXREG2GB
1555 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
1556 {
1557 unsigned int reg_idx = ins.dst[i].reg.idx[0].offset;
1558
1559 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1560 {
1561 WARN("Invalid 1.x sampler index %u.\n", reg_idx);
1562 continue;
1563 }
1564
1565 TRACE("Setting fake 2D resource for 1.x pixelshader.\n");
1566 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1567 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_FLOAT;
1568 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1569 wined3d_bitmap_set(reg_maps->resource_map, reg_idx);
1570
1571 /* texbem is only valid with < 1.4 pixel shaders */
1572 if (ins.handler_idx == WINED3DSIH_TEXBEM
1573 || ins.handler_idx == WINED3DSIH_TEXBEML)
1574 {
1575 reg_maps->bumpmat |= 1u << reg_idx;
1576 if (ins.handler_idx == WINED3DSIH_TEXBEML)
1577 {
1578 reg_maps->luminanceparams |= 1u << reg_idx;
1579 }
1580 }
1581 }
1582 else if (ins.handler_idx == WINED3DSIH_BEM)
1583 {
1584 reg_maps->bumpmat |= 1u << ins.dst[i].reg.idx[0].offset;
1585 }
1586 }
1587
1588 if (ins.handler_idx == WINED3DSIH_IMM_ATOMIC_ALLOC || ins.handler_idx == WINED3DSIH_IMM_ATOMIC_CONSUME)
1589 {
1590 unsigned int reg_idx = ins.src[0].reg.idx[0].offset;
1591 if (reg_idx >= MAX_UNORDERED_ACCESS_VIEWS)
1592 {
1593 ERR("Invalid UAV index %u.\n", reg_idx);
1594 break;
1595 }
1596 reg_maps->uav_counter_mask |= (1u << reg_idx);
1597 }
1598 else if ((WINED3DSIH_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_ATOMIC_XOR)
1599 || (WINED3DSIH_IMM_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_IMM_ATOMIC_XOR)
1600 || (ins.handler_idx == WINED3DSIH_BUFINFO && ins.src[0].reg.type == WINED3DSPR_UAV)
1601 || ins.handler_idx == WINED3DSIH_LD_UAV_TYPED
1602 || (ins.handler_idx == WINED3DSIH_LD_RAW && ins.src[1].reg.type == WINED3DSPR_UAV)
1603 || (ins.handler_idx == WINED3DSIH_LD_STRUCTURED && ins.src[2].reg.type == WINED3DSPR_UAV))
1604 {
1605 const struct wined3d_shader_register *reg;
1606
1607 if (ins.handler_idx == WINED3DSIH_LD_UAV_TYPED || ins.handler_idx == WINED3DSIH_LD_RAW)
1608 reg = &ins.src[1].reg;
1609 else if (ins.handler_idx == WINED3DSIH_LD_STRUCTURED)
1610 reg = &ins.src[2].reg;
1611 else if (WINED3DSIH_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_ATOMIC_XOR)
1612 reg = &ins.dst[0].reg;
1613 else if (ins.handler_idx == WINED3DSIH_BUFINFO)
1614 reg = &ins.src[0].reg;
1615 else
1616 reg = &ins.dst[1].reg;
1617
1618 if (reg->type == WINED3DSPR_UAV)
1619 {
1620 if (reg->idx[0].offset >= MAX_UNORDERED_ACCESS_VIEWS)
1621 {
1622 ERR("Invalid UAV index %u.\n", reg->idx[0].offset);
1623 break;
1624 }
1625 reg_maps->uav_read_mask |= (1u << reg->idx[0].offset);
1626 }
1627 }
1628 else if (ins.handler_idx == WINED3DSIH_NRM)
1629 {
1630 reg_maps->usesnrm = 1;
1631 }
1632 else if (ins.handler_idx == WINED3DSIH_DSY
1633 || ins.handler_idx == WINED3DSIH_DSY_COARSE
1634 || ins.handler_idx == WINED3DSIH_DSY_FINE)
1635 {
1636 reg_maps->usesdsy = 1;
1637 }
1638 else if (ins.handler_idx == WINED3DSIH_DSX
1639 || ins.handler_idx == WINED3DSIH_DSX_COARSE
1640 || ins.handler_idx == WINED3DSIH_DSX_FINE)
1641 {
1642 reg_maps->usesdsx = 1;
1643 }
1644 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
1645 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
1646 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
1647 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
1648 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
1649 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
1650 else if (ins.handler_idx == WINED3DSIH_LOOP
1651 || ins.handler_idx == WINED3DSIH_REP)
1652 {
1653 ++cur_loop_depth;
1654 if (cur_loop_depth > max_loop_depth)
1655 max_loop_depth = cur_loop_depth;
1656 }
1657 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
1658 || ins.handler_idx == WINED3DSIH_ENDREP)
1659 {
1660 --cur_loop_depth;
1661 }
1662 else if (ins.handler_idx == WINED3DSIH_GATHER4
1663 || ins.handler_idx == WINED3DSIH_GATHER4_C
1664 || ins.handler_idx == WINED3DSIH_SAMPLE
1665 || ins.handler_idx == WINED3DSIH_SAMPLE_B
1666 || ins.handler_idx == WINED3DSIH_SAMPLE_C
1667 || ins.handler_idx == WINED3DSIH_SAMPLE_C_LZ
1668 || ins.handler_idx == WINED3DSIH_SAMPLE_GRAD
1669 || ins.handler_idx == WINED3DSIH_SAMPLE_LOD)
1670 {
1671 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1672 ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
1673 }
1674 else if (ins.handler_idx == WINED3DSIH_GATHER4_PO
1675 || ins.handler_idx == WINED3DSIH_GATHER4_PO_C)
1676 {
1677 shader_record_sample(reg_maps, ins.src[2].reg.idx[0].offset,
1678 ins.src[3].reg.idx[0].offset, reg_maps->sampler_map.count);
1679 }
1680 else if ((ins.handler_idx == WINED3DSIH_BUFINFO && ins.src[0].reg.type == WINED3DSPR_RESOURCE)
1681 || (ins.handler_idx == WINED3DSIH_SAMPLE_INFO && ins.src[0].reg.type == WINED3DSPR_RESOURCE))
1682 {
1683 shader_record_sample(reg_maps, ins.src[0].reg.idx[0].offset,
1684 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1685 }
1686 else if (ins.handler_idx == WINED3DSIH_LD
1687 || ins.handler_idx == WINED3DSIH_LD2DMS
1688 || (ins.handler_idx == WINED3DSIH_LD_RAW && ins.src[1].reg.type == WINED3DSPR_RESOURCE)
1689 || (ins.handler_idx == WINED3DSIH_RESINFO && ins.src[1].reg.type == WINED3DSPR_RESOURCE))
1690 {
1691 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1692 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1693 }
1694 else if (ins.handler_idx == WINED3DSIH_LD_STRUCTURED
1695 && ins.src[2].reg.type == WINED3DSPR_RESOURCE)
1696 {
1697 shader_record_sample(reg_maps, ins.src[2].reg.idx[0].offset,
1698 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1699 }
1700
1701 if (ins.predicate)
1702 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
1703 shader_version.type, constf_size))
1704 return WINED3DERR_INVALIDCALL;
1705
1706 for (i = 0; i < ins.src_count; ++i)
1707 {
1708 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
1709 struct wined3d_shader_register reg = ins.src[i].reg;
1710
1711 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
1712 shader_version.type, constf_size))
1713 return WINED3DERR_INVALIDCALL;
1714 while (count)
1715 {
1716 ++reg.idx[0].offset;
1717 if (!shader_record_register_usage(shader, reg_maps, &reg,
1718 shader_version.type, constf_size))
1719 return WINED3DERR_INVALIDCALL;
1720 --count;
1721 }
1722
1723 if (color0_mov)
1724 {
1725 if (ins.src[i].reg.type == WINED3DSPR_TEMP
1726 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
1727 {
1728 shader->u.ps.color0_mov = TRUE;
1729 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
1730 }
1731 }
1732 }
1733 }
1734
1735 prev_ins = current_ins;
1736 }
1737 reg_maps->loop_depth = max_loop_depth;
1738
1739 if (phase)
1740 {
1741 phase->end = prev_ins;
1742 phase = NULL;
1743 }
1744
1745 /* PS before 2.0 don't have explicit color outputs. Instead the value of
1746 * R0 is written to the render target. */
1747 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1748 reg_maps->rt_mask |= (1u << 0);
1749
1750 if (input_signature->elements)
1751 {
1752 for (i = 0; i < input_signature->element_count; ++i)
1753 {
1754 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1755 {
1756 if (input_signature->elements[i].register_idx >= ARRAY_SIZE(shader->u.vs.attributes))
1757 {
1758 WARN("Invalid input signature register index %u.\n", input_signature->elements[i].register_idx);
1759 return WINED3DERR_INVALIDCALL;
1760 }
1761 }
1762 else if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1763 {
1764 if (input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1765 reg_maps->vpos = 1;
1766 else if (input_signature->elements[i].sysval_semantic == WINED3D_SV_IS_FRONT_FACE)
1767 reg_maps->usesfacing = 1;
1768 }
1769 reg_maps->input_registers |= 1u << input_signature->elements[i].register_idx;
1770 }
1771 }
1772 else if (!input_signature->elements && reg_maps->input_registers)
1773 {
1774 unsigned int count = wined3d_popcount(reg_maps->input_registers);
1775 struct wined3d_shader_signature_element *e;
1776 unsigned int i;
1777
1778 if (!(input_signature->elements = heap_calloc(count, sizeof(*input_signature->elements))))
1779 return E_OUTOFMEMORY;
1780 input_signature->element_count = count;
1781
1782 e = input_signature->elements;
1783 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1784 {
1785 if (!(reg_maps->input_registers & (1u << i)))
1786 continue;
1787 input_signature_elements[i].register_idx = i;
1788 *e++ = input_signature_elements[i];
1789 }
1790 }
1791
1792 if (output_signature->elements)
1793 {
1794 if (FAILED(hr = shader_scan_output_signature(shader)))
1795 return hr;
1796 }
1797 else if (reg_maps->output_registers)
1798 {
1799 unsigned int count = wined3d_popcount(reg_maps->output_registers);
1800 struct wined3d_shader_signature_element *e;
1801
1802 if (!(output_signature->elements = heap_calloc(count, sizeof(*output_signature->elements))))
1803 return E_OUTOFMEMORY;
1804 output_signature->element_count = count;
1805
1806 e = output_signature->elements;
1807 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1808 {
1809 if (!(reg_maps->output_registers & (1u << i)))
1810 continue;
1811 *e++ = output_signature_elements[i];
1812 }
1813 }
1814
1815 return WINED3D_OK;
1816 }
1817
1818 static void shader_cleanup_reg_maps(struct wined3d_shader_reg_maps *reg_maps)
1819 {
1820 struct wined3d_shader_indexable_temp *reg, *reg_next;
1821
1822 heap_free(reg_maps->constf);
1823 heap_free(reg_maps->sampler_map.entries);
1824
1825 LIST_FOR_EACH_ENTRY_SAFE(reg, reg_next, &reg_maps->indexable_temps, struct wined3d_shader_indexable_temp, entry)
1826 heap_free(reg);
1827 list_init(&reg_maps->indexable_temps);
1828
1829 heap_free(reg_maps->tgsm);
1830 }
1831
1832 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1833 {
1834 DWORD map = 1u << max;
1835 map |= map - 1;
1836 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1837
1838 return wined3d_log2i(map);
1839 }
1840
1841 /* Shared code in order to generate the bulk of the shader string. */
1842 HRESULT shader_generate_code(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
1843 const struct wined3d_shader_reg_maps *reg_maps, void *backend_ctx,
1844 const DWORD *start, const DWORD *end)
1845 {
1846 struct wined3d_device *device = shader->device;
1847 const struct wined3d_shader_frontend *fe = shader->frontend;
1848 void *fe_data = shader->frontend_data;
1849 struct wined3d_shader_version shader_version;
1850 struct wined3d_shader_parser_state state;
1851 struct wined3d_shader_instruction ins;
1852 struct wined3d_shader_tex_mx tex_mx;
1853 struct wined3d_shader_context ctx;
1854 const DWORD *ptr;
1855
1856 /* Initialize current parsing state. */
1857 tex_mx.current_row = 0;
1858 state.current_loop_depth = 0;
1859 state.current_loop_reg = 0;
1860 state.in_subroutine = FALSE;
1861
1862 ctx.shader = shader;
1863 ctx.reg_maps = reg_maps;
1864 ctx.buffer = buffer;
1865 ctx.tex_mx = &tex_mx;
1866 ctx.state = &state;
1867 ctx.backend_data = backend_ctx;
1868 ins.ctx = &ctx;
1869
1870 fe->shader_read_header(fe_data, &ptr, &shader_version);
1871 if (start)
1872 ptr = start;
1873
1874 while (!fe->shader_is_end(fe_data, &ptr) && ptr != end)
1875 {
1876 /* Read opcode. */
1877 fe->shader_read_instruction(fe_data, &ptr, &ins);
1878
1879 /* Unknown opcode and its parameters. */
1880 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1881 {
1882 WARN("Encountered unrecognised or invalid instruction.\n");
1883 return WINED3DERR_INVALIDCALL;
1884 }
1885
1886 if (ins.predicate)
1887 FIXME("Predicates not implemented.\n");
1888
1889 /* Call appropriate function for output target */
1890 device->shader_backend->shader_handle_instruction(&ins);
1891 }
1892
1893 return WINED3D_OK;
1894 }
1895
1896 static void shader_cleanup(struct wined3d_shader *shader)
1897 {
1898 if (shader->reg_maps.shader_version.type == WINED3D_SHADER_TYPE_HULL)
1899 {
1900 heap_free(shader->u.hs.phases.control_point);
1901 heap_free(shader->u.hs.phases.fork);
1902 heap_free(shader->u.hs.phases.join);
1903 }
1904
1905 heap_free(shader->patch_constant_signature.elements);
1906 heap_free(shader->output_signature.elements);
1907 heap_free(shader->input_signature.elements);
1908 shader->device->shader_backend->shader_destroy(shader);
1909 shader_cleanup_reg_maps(&shader->reg_maps);
1910 heap_free(shader->byte_code);
1911 shader_delete_constant_list(&shader->constantsF);
1912 shader_delete_constant_list(&shader->constantsB);
1913 shader_delete_constant_list(&shader->constantsI);
1914 list_remove(&shader->shader_list_entry);
1915
1916 if (shader->frontend && shader->frontend_data)
1917 shader->frontend->shader_free(shader->frontend_data);
1918 }
1919
1920 struct shader_none_priv
1921 {
1922 const struct wined3d_vertex_pipe_ops *vertex_pipe;
1923 const struct wined3d_fragment_pipe_ops *fragment_pipe;
1924 BOOL ffp_proj_control;
1925 };
1926
1927 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1928 static void shader_none_precompile(void *shader_priv, struct wined3d_shader *shader) {}
1929 static void shader_none_select_compute(void *shader_priv, struct wined3d_context *context,
1930 const struct wined3d_state *state) {}
1931 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
1932 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
1933 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
1934 const struct wined3d_state *state) {}
1935 static void shader_none_destroy(struct wined3d_shader *shader) {}
1936 static void shader_none_free_context_data(struct wined3d_context *context) {}
1937 static void shader_none_init_context_state(struct wined3d_context *context) {}
1938
1939 /* Context activation is done by the caller. */
1940 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
1941 const struct wined3d_state *state)
1942 {
1943 struct shader_none_priv *priv = shader_priv;
1944
1945 priv->vertex_pipe->vp_enable(context, !use_vs(state));
1946 priv->fragment_pipe->fp_enable(context, !use_ps(state));
1947 }
1948
1949 /* Context activation is done by the caller. */
1950 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
1951 {
1952 struct shader_none_priv *priv = shader_priv;
1953
1954 priv->vertex_pipe->vp_enable(context, FALSE);
1955 priv->fragment_pipe->fp_enable(context, FALSE);
1956
1957 context->shader_update_mask = (1u << WINED3D_SHADER_TYPE_PIXEL)
1958 | (1u << WINED3D_SHADER_TYPE_VERTEX)
1959 | (1u << WINED3D_SHADER_TYPE_GEOMETRY)
1960 | (1u << WINED3D_SHADER_TYPE_HULL)
1961 | (1u << WINED3D_SHADER_TYPE_DOMAIN)
1962 | (1u << WINED3D_SHADER_TYPE_COMPUTE);
1963 }
1964
1965 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
1966 const struct wined3d_fragment_pipe_ops *fragment_pipe)
1967 {
1968 struct fragment_caps fragment_caps;
1969 void *vertex_priv, *fragment_priv;
1970 struct shader_none_priv *priv;
1971
1972 if (!(priv = heap_alloc(sizeof(*priv))))
1973 return E_OUTOFMEMORY;
1974
1975 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
1976 {
1977 ERR("Failed to initialize vertex pipe.\n");
1978 heap_free(priv);
1979 return E_FAIL;
1980 }
1981
1982 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
1983 {
1984 ERR("Failed to initialize fragment pipe.\n");
1985 vertex_pipe->vp_free(device, NULL);
1986 heap_free(priv);
1987 return E_FAIL;
1988 }
1989
1990 priv->vertex_pipe = vertex_pipe;
1991 priv->fragment_pipe = fragment_pipe;
1992 fragment_pipe->get_caps(device->adapter, &fragment_caps);
1993 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
1994
1995 device->vertex_priv = vertex_priv;
1996 device->fragment_priv = fragment_priv;
1997 device->shader_priv = priv;
1998
1999 return WINED3D_OK;
2000 }
2001
2002 static void shader_none_free(struct wined3d_device *device, struct wined3d_context *context)
2003 {
2004 struct shader_none_priv *priv = device->shader_priv;
2005
2006 priv->fragment_pipe->free_private(device, context);
2007 priv->vertex_pipe->vp_free(device, context);
2008 heap_free(priv);
2009 }
2010
2011 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
2012 {
2013 return TRUE;
2014 }
2015
2016 static void shader_none_get_caps(const struct wined3d_adapter *adapter, struct shader_caps *caps)
2017 {
2018 /* Set the shader caps to 0 for the none shader backend */
2019 memset(caps, 0, sizeof(*caps));
2020 }
2021
2022 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
2023 {
2024 /* We "support" every possible fixup, since we don't support any shader
2025 * model, and will never have to actually sample a texture. */
2026 return TRUE;
2027 }
2028
2029 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
2030 {
2031 struct shader_none_priv *priv = shader_priv;
2032
2033 return priv->ffp_proj_control;
2034 }
2035
2036 static uint64_t shader_none_shader_compile(struct wined3d_context *context, const struct wined3d_shader_desc *shader_desc,
2037 enum wined3d_shader_type shader_type)
2038 {
2039 return 0;
2040 }
2041
2042 const struct wined3d_shader_backend_ops none_shader_backend =
2043 {
2044 shader_none_handle_instruction,
2045 shader_none_precompile,
2046 shader_none_select,
2047 shader_none_select_compute,
2048 shader_none_disable,
2049 shader_none_update_float_vertex_constants,
2050 shader_none_update_float_pixel_constants,
2051 shader_none_load_constants,
2052 shader_none_destroy,
2053 shader_none_alloc,
2054 shader_none_free,
2055 shader_none_allocate_context_data,
2056 shader_none_free_context_data,
2057 shader_none_init_context_state,
2058 shader_none_get_caps,
2059 shader_none_color_fixup_supported,
2060 shader_none_has_ffp_proj_control,
2061 shader_none_shader_compile,
2062 };
2063
2064 static unsigned int shader_max_version_from_feature_level(enum wined3d_feature_level level)
2065 {
2066 switch (level)
2067 {
2068 case WINED3D_FEATURE_LEVEL_11_1:
2069 case WINED3D_FEATURE_LEVEL_11:
2070 return 5;
2071 case WINED3D_FEATURE_LEVEL_10_1:
2072 case WINED3D_FEATURE_LEVEL_10:
2073 return 4;
2074 case WINED3D_FEATURE_LEVEL_9_3:
2075 return 3;
2076 case WINED3D_FEATURE_LEVEL_9_2:
2077 case WINED3D_FEATURE_LEVEL_9_1:
2078 return 2;
2079 default:
2080 return 1;
2081 }
2082 }
2083
2084 static HRESULT shader_set_function(struct wined3d_shader *shader, struct wined3d_device *device,
2085 enum wined3d_shader_type type, unsigned int float_const_count)
2086 {
2087 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
2088 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2089 const struct wined3d_shader_version *version = &reg_maps->shader_version;
2090 const struct wined3d_shader_frontend *fe;
2091 unsigned int backend_version;
2092 HRESULT hr;
2093
2094 TRACE("shader %p, device %p, type %s, float_const_count %u.\n",
2095 shader, device, debug_shader_type(type), float_const_count);
2096
2097 fe = shader->frontend;
2098 if (!(shader->frontend_data = fe->shader_init(shader->function,
2099 shader->functionLength, &shader->output_signature)))
2100 {
2101 FIXME("Failed to initialize frontend.\n");
2102 return WINED3DERR_INVALIDCALL;
2103 }
2104
2105 if (FAILED(hr = shader_get_registers_used(shader, float_const_count)))
2106 return hr;
2107
2108 if (version->type != type)
2109 {
2110 WARN("Wrong shader type %s.\n", debug_shader_type(reg_maps->shader_version.type));
2111 return WINED3DERR_INVALIDCALL;
2112 }
2113 if (version->major > shader_max_version_from_feature_level(device->cs->c.state->feature_level))
2114 {
2115 WARN("Shader version %u not supported by this device.\n", version->major);
2116 return WINED3DERR_INVALIDCALL;
2117 }
2118 switch (type)
2119 {
2120 case WINED3D_SHADER_TYPE_VERTEX:
2121 backend_version = d3d_info->limits.vs_version;
2122 break;
2123 case WINED3D_SHADER_TYPE_HULL:
2124 backend_version = d3d_info->limits.hs_version;
2125 break;
2126 case WINED3D_SHADER_TYPE_DOMAIN:
2127 backend_version = d3d_info->limits.ds_version;
2128 break;
2129 case WINED3D_SHADER_TYPE_GEOMETRY:
2130 backend_version = d3d_info->limits.gs_version;
2131 break;
2132 case WINED3D_SHADER_TYPE_PIXEL:
2133 backend_version = d3d_info->limits.ps_version;
2134 break;
2135 case WINED3D_SHADER_TYPE_COMPUTE:
2136 backend_version = d3d_info->limits.cs_version;
2137 break;
2138 default:
2139 FIXME("No backend version-checking for this shader type.\n");
2140 backend_version = 0;
2141 }
2142 if (version->major > backend_version)
2143 {
2144 WARN("Shader version %u.%u not supported by the current shader backend.\n",
2145 version->major, version->minor);
2146 return WINED3DERR_INVALIDCALL;
2147 }
2148
2149 shader->load_local_constsF = shader->lconst_inf_or_nan;
2150
2151 return WINED3D_OK;
2152 }
2153
2154 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
2155 {
2156 unsigned int refcount = InterlockedIncrement(&shader->ref);
2157
2158 TRACE("%p increasing refcount to %u.\n", shader, refcount);
2159
2160 return refcount;
2161 }
2162
2163 static void wined3d_shader_init_object(void *object)
2164 {
2165 struct wined3d_shader *shader = object;
2166 struct wined3d_device *device = shader->device;
2167
2168 TRACE("shader %p.\n", shader);
2169
2170 list_add_head(&device->shaders, &shader->shader_list_entry);
2171
2172 device->shader_backend->shader_precompile(device->shader_priv, shader);
2173 }
2174
2175 static void wined3d_shader_destroy_object(void *object)
2176 {
2177 TRACE("object %p.\n", object);
2178
2179 shader_cleanup(object);
2180 heap_free(object);
2181 }
2182
2183 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
2184 {
2185 unsigned int refcount = InterlockedDecrement(&shader->ref);
2186
2187 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
2188
2189 if (!refcount)
2190 {
2191 wined3d_mutex_lock();
2192 shader->parent_ops->wined3d_object_destroyed(shader->parent);
2193 wined3d_cs_destroy_object(shader->device->cs, wined3d_shader_destroy_object, shader);
2194 wined3d_mutex_unlock();
2195 }
2196
2197 return refcount;
2198 }
2199
2200 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
2201 {
2202 TRACE("shader %p.\n", shader);
2203
2204 return shader->parent;
2205 }
2206
2207 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
2208 void *byte_code, UINT *byte_code_size)
2209 {
2210 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
2211
2212 if (!byte_code)
2213 {
2214 *byte_code_size = shader->byte_code_size;
2215 return WINED3D_OK;
2216 }
2217
2218 if (*byte_code_size < shader->byte_code_size)
2219 {
2220 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
2221 * than the required size we should write the required size and
2222 * return D3DERR_MOREDATA. That's not actually true. */
2223 return WINED3DERR_INVALIDCALL;
2224 }
2225
2226 memcpy(byte_code, shader->byte_code, shader->byte_code_size);
2227
2228 return WINED3D_OK;
2229 }
2230
2231 /* Set local constants for d3d8 shaders. */
2232 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
2233 UINT start_idx, const float *src_data, UINT count)
2234 {
2235 UINT end_idx = start_idx + count;
2236 UINT i;
2237
2238 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
2239
2240 if (end_idx > shader->limits->constant_float)
2241 {
2242 WARN("end_idx %u > float constants limit %u.\n",
2243 end_idx, shader->limits->constant_float);
2244 end_idx = shader->limits->constant_float;
2245 }
2246
2247 for (i = start_idx; i < end_idx; ++i)
2248 {
2249 struct wined3d_shader_lconst *lconst;
2250 float *value;
2251
2252 if (!(lconst = heap_alloc(sizeof(*lconst))))
2253 return E_OUTOFMEMORY;
2254
2255 lconst->idx = i;
2256 value = (float *)lconst->value;
2257 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
2258 list_add_head(&shader->constantsF, &lconst->entry);
2259
2260 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
2261 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
2262 {
2263 shader->lconst_inf_or_nan = TRUE;
2264 }
2265 }
2266
2267 return WINED3D_OK;
2268 }
2269
2270 static void init_interpolation_compile_args(uint32_t *interpolation_args,
2271 const struct wined3d_shader *pixel_shader, const struct wined3d_d3d_info *d3d_info)
2272 {
2273 if (!d3d_info->shader_output_interpolation || !pixel_shader
2274 || pixel_shader->reg_maps.shader_version.major < 4)
2275 {
2276 memset(interpolation_args, 0, sizeof(pixel_shader->u.ps.interpolation_mode));
2277 return;
2278 }
2279
2280 memcpy(interpolation_args, pixel_shader->u.ps.interpolation_mode,
2281 sizeof(pixel_shader->u.ps.interpolation_mode));
2282 }
2283
2284 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2285 struct vs_compile_args *args, const struct wined3d_context *context)
2286 {
2287 const struct wined3d_shader *geometry_shader = state->shader[WINED3D_SHADER_TYPE_GEOMETRY];
2288 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
2289 const struct wined3d_shader *hull_shader = state->shader[WINED3D_SHADER_TYPE_HULL];
2290 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
2291 WORD swizzle_map = context->stream_info.swizzle_map;
2292
2293 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
2294 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
2295 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
2296 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
2297 args->point_size = state->primitive_type == WINED3D_PT_POINTLIST;
2298 args->per_vertex_point_size = shader->reg_maps.point_size;
2299 args->next_shader_type = hull_shader ? WINED3D_SHADER_TYPE_HULL
2300 : geometry_shader ? WINED3D_SHADER_TYPE_GEOMETRY : WINED3D_SHADER_TYPE_PIXEL;
2301 if (shader->reg_maps.shader_version.major >= 4)
2302 args->next_shader_input_count = hull_shader ? hull_shader->limits->packed_input
2303 : geometry_shader ? geometry_shader->limits->packed_input
2304 : pixel_shader ? pixel_shader->limits->packed_input : 0;
2305 else
2306 args->next_shader_input_count = 0;
2307 args->swizzle_map = swizzle_map;
2308 if (d3d_info->emulated_flatshading)
2309 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
2310 else
2311 args->flatshading = 0;
2312
2313 init_interpolation_compile_args(args->interpolation_mode,
2314 args->next_shader_type == WINED3D_SHADER_TYPE_PIXEL ? pixel_shader : NULL, d3d_info);
2315 }
2316
2317 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
2318 {
2319 if (usage_idx1 != usage_idx2)
2320 return FALSE;
2321 if (usage1 == usage2)
2322 return TRUE;
2323 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
2324 return TRUE;
2325 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
2326 return TRUE;
2327
2328 return FALSE;
2329 }
2330
2331 bool vshader_get_input(const struct wined3d_shader *shader,
2332 uint8_t usage_req, uint8_t usage_idx_req, unsigned int *regnum)
2333 {
2334 uint32_t map = shader->reg_maps.input_registers & 0xffff;
2335 unsigned int i;
2336
2337 while (map)
2338 {
2339 i = wined3d_bit_scan(&map);
2340 if (match_usage(shader->u.vs.attributes[i].usage,
2341 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
2342 {
2343 *regnum = i;
2344 return true;
2345 }
2346 }
2347
2348 return false;
2349 }
2350
2351 static void shader_trace(const void *code, size_t size, enum vkd3d_shader_source_type source_type)
2352 {
2353 struct vkd3d_shader_compile_info info;
2354 struct vkd3d_shader_code d3d_asm;
2355 const char *ptr, *end, *line;
2356 char *messages;
2357 int ret;
2358
2359 static const struct vkd3d_shader_compile_option compile_options[] =
2360 {
2361 {VKD3D_SHADER_COMPILE_OPTION_API_VERSION, VKD3D_SHADER_API_VERSION_1_6},
2362 };
2363
2364 info.type = VKD3D_SHADER_STRUCTURE_TYPE_COMPILE_INFO;
2365 info.next = NULL;
2366 info.source.code = code;
2367 info.source.size = size;
2368 info.source_type = source_type;
2369 info.target_type = VKD3D_SHADER_TARGET_D3D_ASM;
2370 info.options = compile_options;
2371 info.option_count = ARRAY_SIZE(compile_options);
2372 info.log_level = VKD3D_SHADER_LOG_WARNING;
2373 info.source_name = NULL;
2374
2375 ret = vkd3d_shader_compile(&info, &d3d_asm, &messages);
2376 if (messages && *messages && FIXME_ON(d3d_shader))
2377 {
2378 FIXME("Shader log:\n");
2379 ptr = messages;
2380 end = ptr + strlen(ptr);
2381 while ((line = wined3d_get_line(&ptr, end)))
2382 {
2383 FIXME(" %.*s", (int)(ptr - line), line);
2384 }
2385 FIXME("\n");
2386 }
2387 vkd3d_shader_free_messages(messages);
2388
2389 if (ret < 0)
2390 {
2391 ERR("Failed to disassemble, ret %d.\n", ret);
2392 return;
2393 }
2394
2395 ptr = d3d_asm.code;
2396 end = ptr + d3d_asm.size;
2397 while ((line = wined3d_get_line(&ptr, end)))
2398 {
2399 TRACE(" %.*s", (int)(ptr - line), line);
2400 }
2401 TRACE("\n");
2402
2403 vkd3d_shader_free_shader_code(&d3d_asm);
2404 }
2405
2406 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2407 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2408 {
2409 enum vkd3d_shader_source_type source_type;
2410 HRESULT hr;
2411
2412 TRACE("byte_code %p, byte_code_size %#lx.\n", desc->byte_code, (long)desc->byte_code_size);
2413
2414 if (!desc->byte_code)
2415 return WINED3DERR_INVALIDCALL;
2416
2417 shader->ref = 1;
2418 shader->device = device;
2419 shader->parent = parent;
2420 shader->parent_ops = parent_ops;
2421
2422 list_init(&shader->linked_programs);
2423 list_init(&shader->constantsF);
2424 list_init(&shader->constantsB);
2425 list_init(&shader->constantsI);
2426 shader->lconst_inf_or_nan = FALSE;
2427 list_init(&shader->reg_maps.indexable_temps);
2428 list_init(&shader->shader_list_entry);
2429
2430 if (desc->byte_code_size == ~(size_t)0)
2431 {
2432 struct wined3d_shader_version shader_version;
2433 const struct wined3d_shader_frontend *fe;
2434 struct wined3d_shader_instruction ins;
2435 const DWORD *ptr;
2436 void *fe_data;
2437
2438 source_type = VKD3D_SHADER_SOURCE_D3D_BYTECODE;
2439 if (!(shader->frontend = shader_select_frontend(source_type)))
2440 {
2441 FIXME("Unable to find frontend for shader.\n");
2442 hr = WINED3DERR_INVALIDCALL;
2443 goto fail;
2444 }
2445
2446 fe = shader->frontend;
2447 if (!(fe_data = fe->shader_init(desc->byte_code, desc->byte_code_size, &shader->output_signature)))
2448 {
2449 WARN("Failed to initialise frontend data.\n");
2450 hr = WINED3DERR_INVALIDCALL;
2451 goto fail;
2452 }
2453
2454 fe->shader_read_header(fe_data, &ptr, &shader_version);
2455 while (!fe->shader_is_end(fe_data, &ptr))
2456 fe->shader_read_instruction(fe_data, &ptr, &ins);
2457
2458 fe->shader_free(fe_data);
2459
2460 shader->byte_code_size = (ptr - desc->byte_code) * sizeof(*ptr);
2461
2462 if (!(shader->byte_code = heap_alloc(shader->byte_code_size)))
2463 {
2464 hr = E_OUTOFMEMORY;
2465 goto fail;
2466 }
2467 memcpy(shader->byte_code, desc->byte_code, shader->byte_code_size);
2468
2469 shader->function = shader->byte_code;
2470 shader->functionLength = shader->byte_code_size;
2471 }
2472 else
2473 {
2474 unsigned int max_version;
2475
2476 if (!(shader->byte_code = heap_alloc(desc->byte_code_size)))
2477 {
2478 hr = E_OUTOFMEMORY;
2479 goto fail;
2480 }
2481 memcpy(shader->byte_code, desc->byte_code, desc->byte_code_size);
2482 shader->byte_code_size = desc->byte_code_size;
2483
2484 max_version = shader_max_version_from_feature_level(device->cs->c.state->feature_level);
2485 if (FAILED(hr = wined3d_shader_extract_from_dxbc(shader, max_version, &source_type)))
2486 goto fail;
2487
2488 if (!(shader->frontend = shader_select_frontend(source_type)))
2489 {
2490 FIXME("Unable to find frontend for shader.\n");
2491 hr = WINED3DERR_INVALIDCALL;
2492 goto fail;
2493 }
2494 }
2495
2496 if (TRACE_ON(d3d_shader))
2497 {
2498 if (source_type == VKD3D_SHADER_SOURCE_D3D_BYTECODE)
2499 shader_trace(shader->function, shader->functionLength, source_type);
2500 else
2501 shader_trace(shader->byte_code, shader->byte_code_size, source_type);
2502 }
2503
2504
2505 return WINED3D_OK;
2506
2507 fail:
2508 shader_cleanup(shader);
2509 return hr;
2510 }
2511
2512 static HRESULT vertex_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2513 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2514 {
2515 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2516 unsigned int i;
2517 HRESULT hr;
2518
2519 if (FAILED(hr = shader_init(shader, device, desc, parent, parent_ops)))
2520 return hr;
2521
2522 if (FAILED(hr = shader_set_function(shader, device,
2523 WINED3D_SHADER_TYPE_VERTEX, device->adapter->d3d_info.limits.vs_uniform_count)))
2524 {
2525 shader_cleanup(shader);
2526 return hr;
2527 }
2528
2529 for (i = 0; i < shader->input_signature.element_count; ++i)
2530 {
2531 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
2532
2533 if (!(reg_maps->input_registers & (1u << input->register_idx)) || !input->semantic_name)
2534 continue;
2535
2536 shader->u.vs.attributes[input->register_idx].usage =
2537 shader_usage_from_semantic_name(input->semantic_name);
2538 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
2539 }
2540
2541 if (reg_maps->usesrelconstF && !list_empty(&shader->constantsF))
2542 shader->load_local_constsF = TRUE;
2543
2544 return WINED3D_OK;
2545 }
2546
2547 static struct wined3d_shader_signature_element *shader_find_signature_element(const struct wined3d_shader_signature *s,
2548 unsigned int stream_idx, const char *semantic_name, unsigned int semantic_idx)
2549 {
2550 struct wined3d_shader_signature_element *e = s->elements;
2551 unsigned int i;
2552
2553 for (i = 0; i < s->element_count; ++i)
2554 {
2555 if (e[i].stream_idx == stream_idx
2556 && !stricmp(e[i].semantic_name, semantic_name)
2557 && e[i].semantic_idx == semantic_idx)
2558 return &e[i];
2559 }
2560
2561 return NULL;
2562 }
2563
2564 BOOL shader_get_stream_output_register_info(const struct wined3d_shader *shader,
2565 const struct wined3d_stream_output_element *so_element, unsigned int *register_idx, unsigned int *component_idx)
2566 {
2567 const struct wined3d_shader_signature_element *output;
2568 unsigned int idx;
2569
2570 if (!(output = shader_find_signature_element(&shader->output_signature,
2571 so_element->stream_idx, so_element->semantic_name, so_element->semantic_idx)))
2572 return FALSE;
2573
2574 for (idx = 0; idx < 4; ++idx)
2575 {
2576 if (output->mask & (1u << idx))
2577 break;
2578 }
2579 idx += so_element->component_idx;
2580
2581 *register_idx = output->register_idx;
2582 *component_idx = idx;
2583 return TRUE;
2584 }
2585
2586 static HRESULT geometry_shader_init_so_desc(struct wined3d_geometry_shader *gs, struct wined3d_device *device,
2587 const struct wined3d_stream_output_desc *so_desc)
2588 {
2589 struct wined3d_so_desc_entry *s;
2590 struct wine_rb_entry *entry;
2591 unsigned int i;
2592 size_t size;
2593 char *name;
2594
2595 if ((entry = wine_rb_get(&device->so_descs, so_desc)))
2596 {
2597 gs->so_desc = &WINE_RB_ENTRY_VALUE(entry, struct wined3d_so_desc_entry, entry)->desc;
2598 return WINED3D_OK;
2599 }
2600
2601 size = FIELD_OFFSET(struct wined3d_so_desc_entry, elements[so_desc->element_count]);
2602 for (i = 0; i < so_desc->element_count; ++i)
2603 {
2604 const char *n = so_desc->elements[i].semantic_name;
2605
2606 if (n)
2607 size += strlen(n) + 1;
2608 }
2609 if (!(s = heap_alloc(size)))
2610 return E_OUTOFMEMORY;
2611
2612 s->desc = *so_desc;
2613
2614 memcpy(s->elements, so_desc->elements, so_desc->element_count * sizeof(*s->elements));
2615 s->desc.elements = s->elements;
2616
2617 name = (char *)&s->elements[s->desc.element_count];
2618 for (i = 0; i < so_desc->element_count; ++i)
2619 {
2620 struct wined3d_stream_output_element *e = &s->elements[i];
2621
2622 if (!e->semantic_name)
2623 continue;
2624
2625 size = strlen(e->semantic_name) + 1;
2626 memcpy(name, e->semantic_name, size);
2627 e->semantic_name = name;
2628 name += size;
2629 }
2630
2631 if (wine_rb_put(&device->so_descs, &s->desc, &s->entry) == -1)
2632 {
2633 heap_free(s);
2634 return E_FAIL;
2635 }
2636 gs->so_desc = &s->desc;
2637
2638 return WINED3D_OK;
2639 }
2640
2641 static HRESULT geometry_shader_init_stream_output(struct wined3d_shader *shader,
2642 const struct wined3d_stream_output_desc *so_desc)
2643 {
2644 const struct wined3d_shader_frontend *fe = shader->frontend;
2645 const struct wined3d_shader_signature_element *output;
2646 unsigned int i, component_idx, register_idx, mask;
2647 struct wined3d_shader_version shader_version;
2648 const DWORD *ptr;
2649 void *fe_data;
2650 HRESULT hr;
2651
2652 if (!so_desc)
2653 return WINED3D_OK;
2654
2655 if (!(fe_data = fe->shader_init(shader->function, shader->functionLength, &shader->output_signature)))
2656 {
2657 WARN("Failed to initialise frontend data.\n");
2658 return WINED3DERR_INVALIDCALL;
2659 }
2660 fe->shader_read_header(fe_data, &ptr, &shader_version);
2661 fe->shader_free(fe_data);
2662
2663 switch (shader_version.type)
2664 {
2665 case WINED3D_SHADER_TYPE_VERTEX:
2666 case WINED3D_SHADER_TYPE_DOMAIN:
2667 shader->function = NULL;
2668 shader->functionLength = 0;
2669 break;
2670 case WINED3D_SHADER_TYPE_GEOMETRY:
2671 break;
2672 default:
2673 WARN("Wrong shader type %s.\n", debug_shader_type(shader_version.type));
2674 return E_INVALIDARG;
2675 }
2676
2677 if (!shader->function)
2678 {
2679 shader->reg_maps.shader_version = shader_version;
2680 shader->reg_maps.shader_version.type = WINED3D_SHADER_TYPE_GEOMETRY;
2681 shader_set_limits(shader);
2682 if (FAILED(hr = shader_scan_output_signature(shader)))
2683 return hr;
2684 }
2685
2686 for (i = 0; i < so_desc->element_count; ++i)
2687 {
2688 const struct wined3d_stream_output_element *e = &so_desc->elements[i];
2689
2690 if (!e->semantic_name)
2691 continue;
2692 if (!(output = shader_find_signature_element(&shader->output_signature,
2693 e->stream_idx, e->semantic_name, e->semantic_idx))
2694 || !shader_get_stream_output_register_info(shader, e, &register_idx, &component_idx))
2695 {
2696 WARN("Failed to find output signature element for stream output entry.\n");
2697 return E_INVALIDARG;
2698 }
2699
2700 mask = wined3d_mask_from_size(e->component_count) << component_idx;
2701 if ((output->mask & 0xff & mask) != mask)
2702 {
2703 WARN("Invalid component range %u-%u (mask %#x), output mask %#x.\n",
2704 component_idx, e->component_count, mask, output->mask & 0xff);
2705 return E_INVALIDARG;
2706 }
2707 }
2708
2709 if (FAILED(hr = geometry_shader_init_so_desc(&shader->u.gs, shader->device, so_desc)))
2710 {
2711 WARN("Failed to initialise stream output description, hr %#lx.\n", hr);
2712 return hr;
2713 }
2714
2715 return WINED3D_OK;
2716 }
2717
2718 static HRESULT geometry_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2719 const struct wined3d_shader_desc *desc, const struct wined3d_stream_output_desc *so_desc,
2720 void *parent, const struct wined3d_parent_ops *parent_ops)
2721 {
2722 HRESULT hr;
2723
2724 if (FAILED(hr = shader_init(shader, device, desc, parent, parent_ops)))
2725 return hr;
2726
2727 if (FAILED(hr = geometry_shader_init_stream_output(shader, so_desc)))
2728 goto fail;
2729
2730 if (shader->function
2731 && FAILED(hr = shader_set_function(shader, device, WINED3D_SHADER_TYPE_GEOMETRY, 0)))
2732 goto fail;
2733
2734 return WINED3D_OK;
2735
2736 fail:
2737 shader_cleanup(shader);
2738 return hr;
2739 }
2740
2741 void find_ds_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2742 struct ds_compile_args *args, const struct wined3d_context *context)
2743 {
2744 const struct wined3d_shader *geometry_shader = state->shader[WINED3D_SHADER_TYPE_GEOMETRY];
2745 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
2746 const struct wined3d_shader *hull_shader = state->shader[WINED3D_SHADER_TYPE_HULL];
2747
2748 args->tessellator_output_primitive = hull_shader->u.hs.tessellator_output_primitive;
2749 args->tessellator_partitioning = hull_shader->u.hs.tessellator_partitioning;
2750
2751 args->output_count = geometry_shader ? geometry_shader->limits->packed_input
2752 : pixel_shader ? pixel_shader->limits->packed_input : shader->limits->packed_output;
2753 args->next_shader_type = geometry_shader ? WINED3D_SHADER_TYPE_GEOMETRY : WINED3D_SHADER_TYPE_PIXEL;
2754
2755 args->render_offscreen = context->render_offscreen;
2756
2757 init_interpolation_compile_args(args->interpolation_mode,
2758 args->next_shader_type == WINED3D_SHADER_TYPE_PIXEL ? pixel_shader : NULL, context->d3d_info);
2759
2760 args->padding = 0;
2761 }
2762
2763 void find_gs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2764 struct gs_compile_args *args, const struct wined3d_context *context)
2765 {
2766 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
2767
2768 args->output_count = pixel_shader ? pixel_shader->limits->packed_input : shader->limits->packed_output;
2769
2770 if (!(args->primitive_type = shader->u.gs.input_type))
2771 args->primitive_type = state->primitive_type;
2772
2773 init_interpolation_compile_args(args->interpolation_mode, pixel_shader, context->d3d_info);
2774 }
2775
2776 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2777 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_context *context)
2778 {
2779 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
2780 struct wined3d_texture *texture;
2781 unsigned int i;
2782
2783 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
2784 if (!d3d_info->srgb_write_control && needs_srgb_write(d3d_info, state, &state->fb))
2785 {
2786 static unsigned int warned = 0;
2787
2788 args->srgb_correction = 1;
2789 if (state->blend_state && state->blend_state->desc.rt[0].enable && !warned++)
2790 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
2791 "support, expect rendering artifacts.\n");
2792 }
2793
2794 if (shader->reg_maps.shader_version.major == 1
2795 && shader->reg_maps.shader_version.minor <= 3)
2796 {
2797 for (i = 0; i < shader->limits->sampler; ++i)
2798 {
2799 uint32_t flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
2800
2801 if (flags & WINED3D_TTFF_PROJECTED)
2802 {
2803 uint32_t tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
2804
2805 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
2806 {
2807 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
2808 unsigned int j;
2809 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
2810 uint32_t max_valid = WINED3D_TTFF_COUNT4;
2811
2812 for (j = 0; j < state->vertex_declaration->element_count; ++j)
2813 {
2814 struct wined3d_vertex_declaration_element *element =
2815 &state->vertex_declaration->elements[j];
2816
2817 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
2818 && element->usage_idx == index)
2819 {
2820 max_valid = element->format->component_count;
2821 break;
2822 }
2823 }
2824 if (!tex_transform || tex_transform > max_valid)
2825 {
2826 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
2827 tex_transform, max_valid);
2828 tex_transform = max_valid;
2829 }
2830 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
2831 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
2832 && tex_transform > WINED3D_TTFF_COUNT2)
2833 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
2834 && tex_transform > WINED3D_TTFF_COUNT3))
2835 tex_transform |= WINED3D_PSARGS_PROJECTED;
2836 else
2837 {
2838 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
2839 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
2840 i, tex_transform, resource_type);
2841 }
2842 }
2843 else
2844 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
2845
2846 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
2847 }
2848 }
2849 }
2850 if (shader->reg_maps.shader_version.major == 1
2851 && shader->reg_maps.shader_version.minor <= 4)
2852 {
2853 for (i = 0; i < shader->limits->sampler; ++i)
2854 {
2855 if (!shader->reg_maps.resource_info[i].type)
2856 continue;
2857
2858 /* Treat unbound textures as 2D. The dummy texture will provide
2859 * the proper sample value. The tex_types bitmap defaults to
2860 * 2D because of the memset. */
2861 if (!(texture = state->textures[i]))
2862 continue;
2863
2864 switch (wined3d_texture_gl(texture)->target)
2865 {
2866 /* RECT textures are distinguished from 2D textures via np2_fixup */
2867 default:
2868 break;
2869
2870 case GL_TEXTURE_3D:
2871 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
2872 break;
2873
2874 case GL_TEXTURE_CUBE_MAP_ARB:
2875 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
2876 break;
2877 }
2878 }
2879 }
2880 else if (shader->reg_maps.shader_version.major <= 3)
2881 {
2882 for (i = 0; i < shader->limits->sampler; ++i)
2883 {
2884 enum wined3d_shader_resource_type resource_type;
2885 enum wined3d_shader_tex_types tex_type;
2886
2887 if (!(resource_type = shader->reg_maps.resource_info[i].type))
2888 continue;
2889
2890 switch (resource_type)
2891 {
2892 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
2893 tex_type = WINED3D_SHADER_TEX_3D;
2894 break;
2895 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
2896 tex_type = WINED3D_SHADER_TEX_CUBE;
2897 break;
2898 default:
2899 tex_type = WINED3D_SHADER_TEX_2D;
2900 break;
2901 }
2902
2903 if ((texture = state->textures[i]))
2904 {
2905 /* Star Wars: The Old Republic uses mismatched samplers for rendering water. */
2906 if (texture->resource.type == WINED3D_RTYPE_TEXTURE_2D
2907 && resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
2908 && !(texture->resource.usage & WINED3DUSAGE_LEGACY_CUBEMAP))
2909 tex_type = WINED3D_SHADER_TEX_2D;
2910 else if (texture->resource.type == WINED3D_RTYPE_TEXTURE_3D
2911 && resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D)
2912 tex_type = WINED3D_SHADER_TEX_3D;
2913 }
2914 args->tex_types |= tex_type << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
2915 }
2916 }
2917
2918 if (shader->reg_maps.shader_version.major >= 4)
2919 {
2920 /* In SM4+ we use dcl_sampler in order to determine if we should use shadow sampler. */
2921 args->shadow = 0;
2922 for (i = 0 ; i < WINED3D_MAX_FRAGMENT_SAMPLERS; ++i)
2923 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2924 args->np2_fixup = 0;
2925 }
2926 else
2927 {
2928 for (i = 0; i < WINED3D_MAX_FRAGMENT_SAMPLERS; ++i)
2929 {
2930 if (!shader->reg_maps.resource_info[i].type)
2931 continue;
2932
2933 texture = state->textures[i];
2934 if (!texture)
2935 {
2936 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2937 continue;
2938 }
2939 if (can_use_texture_swizzle(d3d_info, texture->resource.format))
2940 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2941 else
2942 args->color_fixup[i] = texture->resource.format->color_fixup;
2943
2944 if (texture->resource.format_caps & WINED3D_FORMAT_CAP_SHADOW)
2945 args->shadow |= 1u << i;
2946
2947 /* Flag samplers that need NP2 texcoord fixup. */
2948 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
2949 args->np2_fixup |= (1u << i);
2950 }
2951 }
2952
2953 if (shader->reg_maps.shader_version.major >= 3)
2954 {
2955 if (position_transformed)
2956 args->vp_mode = WINED3D_VP_MODE_NONE;
2957 else if (use_vs(state))
2958 args->vp_mode = WINED3D_VP_MODE_SHADER;
2959 else
2960 args->vp_mode = WINED3D_VP_MODE_FF;
2961 args->fog = WINED3D_FFP_PS_FOG_OFF;
2962 }
2963 else
2964 {
2965 args->vp_mode = WINED3D_VP_MODE_SHADER;
2966 if (state->render_states[WINED3D_RS_FOGENABLE])
2967 {
2968 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
2969 {
2970 case WINED3D_FOG_NONE:
2971 if (position_transformed || use_vs(state))
2972 {
2973 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
2974 break;
2975 }
2976
2977 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
2978 {
2979 case WINED3D_FOG_NONE: /* Fall through. */
2980 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
2981 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
2982 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
2983 }
2984 break;
2985
2986 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
2987 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
2988 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
2989 }
2990 }
2991 else
2992 {
2993 args->fog = WINED3D_FFP_PS_FOG_OFF;
2994 }
2995 }
2996
2997 if (!d3d_info->full_ffp_varyings)
2998 {
2999 const struct wined3d_shader *vs = state->shader[WINED3D_SHADER_TYPE_VERTEX];
3000
3001 args->texcoords_initialized = 0;
3002 for (i = 0; i < WINED3D_MAX_TEXTURES; ++i)
3003 {
3004 if (vs)
3005 {
3006 if (state->shader[WINED3D_SHADER_TYPE_VERTEX]->reg_maps.output_registers & (1u << i))
3007 args->texcoords_initialized |= 1u << i;
3008 }
3009 else
3010 {
3011 const struct wined3d_stream_info *si = &context->stream_info;
3012 unsigned int coord_idx = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3013
3014 if ((state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX] >> WINED3D_FFP_TCI_SHIFT)
3015 & WINED3D_FFP_TCI_MASK
3016 || (coord_idx < WINED3D_MAX_TEXTURES && (si->use_map & (1u << (WINED3D_FFP_TEXCOORD0 + coord_idx)))))
3017 args->texcoords_initialized |= 1u << i;
3018 }
3019 }
3020 }
3021 else
3022 {
3023 args->texcoords_initialized = wined3d_mask_from_size(WINED3D_MAX_TEXTURES);
3024 }
3025
3026 args->pointsprite = state->render_states[WINED3D_RS_POINTSPRITEENABLE]
3027 && state->primitive_type == WINED3D_PT_POINTLIST;
3028
3029 if (d3d_info->ffp_alpha_test)
3030 args->alpha_test_func = WINED3D_CMP_ALWAYS - 1;
3031 else
3032 args->alpha_test_func = (state->render_states[WINED3D_RS_ALPHATESTENABLE]
3033 ? wined3d_sanitize_cmp_func(state->render_states[WINED3D_RS_ALPHAFUNC])
3034 : WINED3D_CMP_ALWAYS) - 1;
3035
3036 if (d3d_info->emulated_flatshading)
3037 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3038
3039 args->y_correction = (shader->reg_maps.vpos && d3d_info->frag_coord_correction)
3040 || (shader->reg_maps.usesdsy && wined3d_settings.offscreen_rendering_mode != ORM_FBO)
3041 ? !context->render_offscreen : 0;
3042
3043 for (i = 0; i < ARRAY_SIZE(state->fb.render_targets); ++i)
3044 {
3045 struct wined3d_rendertarget_view *rtv = state->fb.render_targets[i];
3046 if (rtv && rtv->format->id == WINED3DFMT_A8_UNORM && !is_identity_fixup(rtv->format->color_fixup))
3047 args->rt_alpha_swizzle |= 1u << i;
3048 }
3049
3050 args->dual_source_blend = state->blend_state && state->blend_state->dual_source;
3051 }
3052
3053 static HRESULT pixel_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3054 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3055 {
3056 const struct wined3d_d3d_info *d3d_info = &device->adapter->d3d_info;
3057 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
3058 HRESULT hr;
3059
3060 if (FAILED(hr = shader_init(shader, device, desc, parent, parent_ops)))
3061 return hr;
3062
3063 if (FAILED(hr = shader_set_function(shader, device,
3064 WINED3D_SHADER_TYPE_PIXEL, d3d_info->limits.ps_uniform_count)))
3065 {
3066 shader_cleanup(shader);
3067 return hr;
3068 }
3069
3070 for (i = 0; i < MAX_REG_INPUT; ++i)
3071 {
3072 if (shader->u.ps.input_reg_used & (1u << i))
3073 {
3074 ++num_regs_used;
3075 highest_reg_used = i;
3076 }
3077 }
3078
3079 /* Don't do any register mapping magic if it is not needed, or if we can't
3080 * achieve anything anyway */
3081 if (highest_reg_used < (d3d_info->limits.varying_count / 4)
3082 || num_regs_used > (d3d_info->limits.varying_count / 4)
3083 || shader->reg_maps.shader_version.major >= 4)
3084 {
3085 if (num_regs_used > (d3d_info->limits.varying_count / 4))
3086 {
3087 /* This happens with relative addressing. The input mapper function
3088 * warns about this if the higher registers are declared too, so
3089 * don't write a FIXME here */
3090 WARN("More varying registers used than supported\n");
3091 }
3092
3093 for (i = 0; i < MAX_REG_INPUT; ++i)
3094 {
3095 shader->u.ps.input_reg_map[i] = i;
3096 }
3097
3098 shader->u.ps.declared_in_count = highest_reg_used + 1;
3099 }
3100 else
3101 {
3102 shader->u.ps.declared_in_count = 0;
3103 for (i = 0; i < MAX_REG_INPUT; ++i)
3104 {
3105 if (shader->u.ps.input_reg_used & (1u << i))
3106 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
3107 else shader->u.ps.input_reg_map[i] = ~0U;
3108 }
3109 }
3110
3111 return WINED3D_OK;
3112 }
3113
3114 enum wined3d_shader_resource_type pixelshader_get_resource_type(const struct wined3d_shader_reg_maps *reg_maps,
3115 unsigned int resource_idx, DWORD tex_types)
3116 {
3117 static enum wined3d_shader_resource_type shader_resource_type_from_shader_tex_types[] =
3118 {
3119 WINED3D_SHADER_RESOURCE_TEXTURE_2D, /* WINED3D_SHADER_TEX_2D */
3120 WINED3D_SHADER_RESOURCE_TEXTURE_3D, /* WINED3D_SHADER_TEX_3D */
3121 WINED3D_SHADER_RESOURCE_TEXTURE_CUBE, /* WINED3D_SHADER_TEX_CUBE */
3122 };
3123
3124 unsigned int idx;
3125
3126 if (reg_maps->shader_version.major > 3)
3127 return reg_maps->resource_info[resource_idx].type;
3128
3129 if (!reg_maps->resource_info[resource_idx].type)
3130 return 0;
3131
3132 idx = (tex_types >> resource_idx * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK;
3133 assert(idx < ARRAY_SIZE(shader_resource_type_from_shader_tex_types));
3134 return shader_resource_type_from_shader_tex_types[idx];
3135 }
3136
3137 HRESULT CDECL wined3d_shader_create_cs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3138 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3139 {
3140 struct wined3d_shader *object;
3141 HRESULT hr;
3142
3143 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3144 device, desc, parent, parent_ops, shader);
3145
3146 if (!(object = heap_alloc_zero(sizeof(*object))))
3147 return E_OUTOFMEMORY;
3148
3149 if (FAILED(hr = shader_init(object, device, desc, parent, parent_ops)))
3150 {
3151 WARN("Failed to initialize compute shader, hr %#lx.\n", hr);
3152 heap_free(object);
3153 return hr;
3154 }
3155
3156 if (FAILED(hr = shader_set_function(object, device, WINED3D_SHADER_TYPE_COMPUTE, 0)))
3157 {
3158 shader_cleanup(object);
3159 heap_free(object);
3160 return hr;
3161 }
3162
3163 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
3164
3165 TRACE("Created compute shader %p.\n", object);
3166 *shader = object;
3167
3168 return WINED3D_OK;
3169 }
3170
3171 HRESULT CDECL wined3d_shader_create_ds(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3172 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3173 {
3174 struct wined3d_shader *object;
3175 HRESULT hr;
3176
3177 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3178 device, desc, parent, parent_ops, shader);
3179
3180 if (!(object = heap_alloc_zero(sizeof(*object))))
3181 return E_OUTOFMEMORY;
3182
3183 if (FAILED(hr = shader_init(object, device, desc, parent, parent_ops)))
3184 {
3185 WARN("Failed to initialize domain shader, hr %#lx.\n", hr);
3186 heap_free(object);
3187 return hr;
3188 }
3189
3190 if (FAILED(hr = shader_set_function(object, device, WINED3D_SHADER_TYPE_DOMAIN, 0)))
3191 {
3192 shader_cleanup(object);
3193 heap_free(object);
3194 return hr;
3195 }
3196
3197 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
3198
3199 TRACE("Created domain shader %p.\n", object);
3200 *shader = object;
3201
3202 return WINED3D_OK;
3203 }
3204
3205 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3206 const struct wined3d_stream_output_desc *so_desc, void *parent,
3207 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3208 {
3209 struct wined3d_shader *object;
3210 HRESULT hr;
3211
3212 TRACE("device %p, desc %p, so_desc %p, parent %p, parent_ops %p, shader %p.\n",
3213 device, desc, so_desc, parent, parent_ops, shader);
3214
3215 if (!(object = heap_alloc_zero(sizeof(*object))))
3216 return E_OUTOFMEMORY;
3217
3218 if (FAILED(hr = geometry_shader_init(object, device, desc, so_desc, parent, parent_ops)))
3219 {
3220 WARN("Failed to initialize geometry shader, hr %#lx.\n", hr);
3221 heap_free(object);
3222 return hr;
3223 }
3224
3225 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
3226
3227 TRACE("Created geometry shader %p.\n", object);
3228 *shader = object;
3229
3230 return WINED3D_OK;
3231 }
3232
3233 HRESULT CDECL wined3d_shader_create_hs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3234 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3235 {
3236 struct wined3d_shader *object;
3237 HRESULT hr;
3238
3239 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3240 device, desc, parent, parent_ops, shader);
3241
3242 if (!(object = heap_alloc_zero(sizeof(*object))))
3243 return E_OUTOFMEMORY;
3244
3245 if (FAILED(hr = shader_init(object, device, desc, parent, parent_ops)))
3246 {
3247 WARN("Failed to initialize hull shader, hr %#lx.\n", hr);
3248 heap_free(object);
3249 return hr;
3250 }
3251
3252 if (FAILED(hr = shader_set_function(object, device, WINED3D_SHADER_TYPE_HULL, 0)))
3253 {
3254 shader_cleanup(object);
3255 heap_free(object);
3256 return hr;
3257 }
3258
3259 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
3260
3261 TRACE("Created hull shader %p.\n", object);
3262 *shader = object;
3263
3264 return WINED3D_OK;
3265 }
3266
3267 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3268 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3269 {
3270 struct wined3d_shader *object;
3271 HRESULT hr;
3272
3273 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3274 device, desc, parent, parent_ops, shader);
3275
3276 if (!(object = heap_alloc_zero(sizeof(*object))))
3277 return E_OUTOFMEMORY;
3278
3279 if (FAILED(hr = pixel_shader_init(object, device, desc, parent, parent_ops)))
3280 {
3281 WARN("Failed to initialize pixel shader, hr %#lx.\n", hr);
3282 heap_free(object);
3283 return hr;
3284 }
3285
3286 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
3287
3288 TRACE("Created pixel shader %p.\n", object);
3289 *shader = object;
3290
3291 return WINED3D_OK;
3292 }
3293
3294 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3295 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3296 {
3297 struct wined3d_shader *object;
3298 HRESULT hr;
3299
3300 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3301 device, desc, parent, parent_ops, shader);
3302
3303 if (!(object = heap_alloc_zero(sizeof(*object))))
3304 return E_OUTOFMEMORY;
3305
3306 if (FAILED(hr = vertex_shader_init(object, device, desc, parent, parent_ops)))
3307 {
3308 WARN("Failed to initialize vertex shader, hr %#lx.\n", hr);
3309 heap_free(object);
3310 return hr;
3311 }
3312
3313 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
3314
3315 TRACE("Created vertex shader %p.\n", object);
3316 *shader = object;
3317
3318 return WINED3D_OK;
3319 }
Windows API implementation