?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
wined3d: Introduce the "glsl-vkd3d" shader backend.
[wine.git] / dlls / wined3d / shader.c
blobe6a3b4a33bbe1fa073edb4c4cf09027beb05a4b6
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
30 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
31
32 const struct wined3d_vec4 wined3d_srgb_const[] =
33 {
34 /* pow, mul_high, sub_high, mul_low */
35 {0.41666f, 1.055f, 0.055f, 12.92f},
36 /* cmp */
37 {0.0031308f, 0.0f, 0.0f, 0.0f},
38 };
39
40 static const char * const shader_opcode_names[] =
41 {
42 /* WINED3DSIH_ABS */ "abs",
43 /* WINED3DSIH_ADD */ "add",
44 /* WINED3DSIH_AND */ "and",
45 /* WINED3DSIH_ATOMIC_AND */ "atomic_and",
46 /* WINED3DSIH_ATOMIC_CMP_STORE */ "atomic_cmp_store",
47 /* WINED3DSIH_ATOMIC_IADD */ "atomic_iadd",
48 /* WINED3DSIH_ATOMIC_IMAX */ "atomic_imax",
49 /* WINED3DSIH_ATOMIC_IMIN */ "atomic_imin",
50 /* WINED3DSIH_ATOMIC_OR */ "atomic_or",
51 /* WINED3DSIH_ATOMIC_UMAX */ "atomic_umax",
52 /* WINED3DSIH_ATOMIC_UMIN */ "atomic_umin",
53 /* WINED3DSIH_ATOMIC_XOR */ "atomic_xor",
54 /* WINED3DSIH_BEM */ "bem",
55 /* WINED3DSIH_BFI */ "bfi",
56 /* WINED3DSIH_BFREV */ "bfrev",
57 /* WINED3DSIH_BREAK */ "break",
58 /* WINED3DSIH_BREAKC */ "breakc",
59 /* WINED3DSIH_BREAKP */ "breakp",
60 /* WINED3DSIH_BUFINFO */ "bufinfo",
61 /* WINED3DSIH_CALL */ "call",
62 /* WINED3DSIH_CALLNZ */ "callnz",
63 /* WINED3DSIH_CASE */ "case",
64 /* WINED3DSIH_CMP */ "cmp",
65 /* WINED3DSIH_CND */ "cnd",
66 /* WINED3DSIH_CONTINUE */ "continue",
67 /* WINED3DSIH_CONTINUEP */ "continuec",
68 /* WINED3DSIH_COUNTBITS */ "countbits",
69 /* WINED3DSIH_CRS */ "crs",
70 /* WINED3DSIH_CUT */ "cut",
71 /* WINED3DSIH_CUT_STREAM */ "cut_stream",
72 /* WINED3DSIH_DCL */ "dcl",
73 /* WINED3DSIH_DCL_CONSTANT_BUFFER */ "dcl_constantBuffer",
74 /* WINED3DSIH_DCL_FUNCTION_BODY */ "dcl_function_body",
75 /* WINED3DSIH_DCL_FUNCTION_TABLE */ "dcl_function_table",
76 /* WINED3DSIH_DCL_GLOBAL_FLAGS */ "dcl_globalFlags",
77 /* WINED3DSIH_DCL_GS_INSTANCES */ "dcl_gs_instances",
78 /* WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT */ "dcl_hs_fork_phase_instance_count",
79 /* WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT */ "dcl_hs_join_phase_instance_count",
80 /* WINED3DSIH_DCL_HS_MAX_TESSFACTOR */ "dcl_hs_max_tessfactor",
81 /* WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER */ "dcl_immediateConstantBuffer",
82 /* WINED3DSIH_DCL_INDEX_RANGE */ "dcl_index_range",
83 /* WINED3DSIH_DCL_INDEXABLE_TEMP */ "dcl_indexableTemp",
84 /* WINED3DSIH_DCL_INPUT */ "dcl_input",
85 /* WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT */ "dcl_input_control_point_count",
86 /* WINED3DSIH_DCL_INPUT_PRIMITIVE */ "dcl_inputPrimitive",
87 /* WINED3DSIH_DCL_INPUT_PS */ "dcl_input_ps",
88 /* WINED3DSIH_DCL_INPUT_PS_SGV */ "dcl_input_ps_sgv",
89 /* WINED3DSIH_DCL_INPUT_PS_SIV */ "dcl_input_ps_siv",
90 /* WINED3DSIH_DCL_INPUT_SGV */ "dcl_input_sgv",
91 /* WINED3DSIH_DCL_INPUT_SIV */ "dcl_input_siv",
92 /* WINED3DSIH_DCL_INTERFACE */ "dcl_interface",
93 /* WINED3DSIH_DCL_OUTPUT */ "dcl_output",
94 /* WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT */ "dcl_output_control_point_count",
95 /* WINED3DSIH_DCL_OUTPUT_SIV */ "dcl_output_siv",
96 /* WINED3DSIH_DCL_OUTPUT_TOPOLOGY */ "dcl_outputTopology",
97 /* WINED3DSIH_DCL_RESOURCE_RAW */ "dcl_resource_raw",
98 /* WINED3DSIH_DCL_RESOURCE_STRUCTURED */ "dcl_resource_structured",
99 /* WINED3DSIH_DCL_SAMPLER */ "dcl_sampler",
100 /* WINED3DSIH_DCL_STREAM */ "dcl_stream",
101 /* WINED3DSIH_DCL_TEMPS */ "dcl_temps",
102 /* WINED3DSIH_DCL_TESSELLATOR_DOMAIN */ "dcl_tessellator_domain",
103 /* WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE */ "dcl_tessellator_output_primitive",
104 /* WINED3DSIH_DCL_TESSELLATOR_PARTITIONING */ "dcl_tessellator_partitioning",
105 /* WINED3DSIH_DCL_TGSM_RAW */ "dcl_tgsm_raw",
106 /* WINED3DSIH_DCL_TGSM_STRUCTURED */ "dcl_tgsm_structured",
107 /* WINED3DSIH_DCL_THREAD_GROUP */ "dcl_thread_group",
108 /* WINED3DSIH_DCL_UAV_RAW */ "dcl_uav_raw",
109 /* WINED3DSIH_DCL_UAV_STRUCTURED */ "dcl_uav_structured",
110 /* WINED3DSIH_DCL_UAV_TYPED */ "dcl_uav_typed",
111 /* WINED3DSIH_DCL_VERTICES_OUT */ "dcl_maxOutputVertexCount",
112 /* WINED3DSIH_DEF */ "def",
113 /* WINED3DSIH_DEFAULT */ "default",
114 /* WINED3DSIH_DEFB */ "defb",
115 /* WINED3DSIH_DEFI */ "defi",
116 /* WINED3DSIH_DIV */ "div",
117 /* WINED3DSIH_DP2 */ "dp2",
118 /* WINED3DSIH_DP2ADD */ "dp2add",
119 /* WINED3DSIH_DP3 */ "dp3",
120 /* WINED3DSIH_DP4 */ "dp4",
121 /* WINED3DSIH_DST */ "dst",
122 /* WINED3DSIH_DSX */ "dsx",
123 /* WINED3DSIH_DSX_COARSE */ "deriv_rtx_coarse",
124 /* WINED3DSIH_DSX_FINE */ "deriv_rtx_fine",
125 /* WINED3DSIH_DSY */ "dsy",
126 /* WINED3DSIH_DSY_COARSE */ "deriv_rty_coarse",
127 /* WINED3DSIH_DSY_FINE */ "deriv_rty_fine",
128 /* WINED3DSIH_ELSE */ "else",
129 /* WINED3DSIH_EMIT */ "emit",
130 /* WINED3DSIH_EMIT_STREAM */ "emit_stream",
131 /* WINED3DSIH_ENDIF */ "endif",
132 /* WINED3DSIH_ENDLOOP */ "endloop",
133 /* WINED3DSIH_ENDREP */ "endrep",
134 /* WINED3DSIH_ENDSWITCH */ "endswitch",
135 /* WINED3DSIH_EQ */ "eq",
136 /* WINED3DSIH_EVAL_CENTROID */ "eval_centroid",
137 /* WINED3DSIH_EVAL_SAMPLE_INDEX */ "eval_sample_index",
138 /* WINED3DSIH_EXP */ "exp",
139 /* WINED3DSIH_EXPP */ "expp",
140 /* WINED3DSIH_F16TOF32 */ "f16tof32",
141 /* WINED3DSIH_F32TOF16 */ "f32tof16",
142 /* WINED3DSIH_FCALL */ "fcall",
143 /* WINED3DSIH_FIRSTBIT_HI */ "firstbit_hi",
144 /* WINED3DSIH_FIRSTBIT_LO */ "firstbit_lo",
145 /* WINED3DSIH_FIRSTBIT_SHI */ "firstbit_shi",
146 /* WINED3DSIH_FRC */ "frc",
147 /* WINED3DSIH_FTOI */ "ftoi",
148 /* WINED3DSIH_FTOU */ "ftou",
149 /* WINED3DSIH_GATHER4 */ "gather4",
150 /* WINED3DSIH_GATHER4_C */ "gather4_c",
151 /* WINED3DSIH_GATHER4_PO */ "gather4_po",
152 /* WINED3DSIH_GATHER4_PO_C */ "gather4_po_c",
153 /* WINED3DSIH_GE */ "ge",
154 /* WINED3DSIH_HS_CONTROL_POINT_PHASE */ "hs_control_point_phase",
155 /* WINED3DSIH_HS_DECLS */ "hs_decls",
156 /* WINED3DSIH_HS_FORK_PHASE */ "hs_fork_phase",
157 /* WINED3DSIH_HS_JOIN_PHASE */ "hs_join_phase",
158 /* WINED3DSIH_IADD */ "iadd",
159 /* WINED3DSIH_IBFE */ "ibfe",
160 /* WINED3DSIH_IEQ */ "ieq",
161 /* WINED3DSIH_IF */ "if",
162 /* WINED3DSIH_IFC */ "ifc",
163 /* WINED3DSIH_IGE */ "ige",
164 /* WINED3DSIH_ILT */ "ilt",
165 /* WINED3DSIH_IMAD */ "imad",
166 /* WINED3DSIH_IMAX */ "imax",
167 /* WINED3DSIH_IMIN */ "imin",
168 /* WINED3DSIH_IMM_ATOMIC_ALLOC */ "imm_atomic_alloc",
169 /* WINED3DSIH_IMM_ATOMIC_AND */ "imm_atomic_and",
170 /* WINED3DSIH_IMM_ATOMIC_CMP_EXCH */ "imm_atomic_cmp_exch",
171 /* WINED3DSIH_IMM_ATOMIC_CONSUME */ "imm_atomic_consume",
172 /* WINED3DSIH_IMM_ATOMIC_EXCH */ "imm_atomic_exch",
173 /* WINED3DSIH_IMM_ATOMIC_IADD */ "imm_atomic_iadd",
174 /* WINED3DSIH_IMM_ATOMIC_IMAX */ "imm_atomic_imax",
175 /* WINED3DSIH_IMM_ATOMIC_IMIN */ "imm_atomic_imin",
176 /* WINED3DSIH_IMM_ATOMIC_OR */ "imm_atomic_or",
177 /* WINED3DSIH_IMM_ATOMIC_UMAX */ "imm_atomic_umax",
178 /* WINED3DSIH_IMM_ATOMIC_UMIN */ "imm_atomic_umin",
179 /* WINED3DSIH_IMM_ATOMIC_XOR */ "imm_atomic_xor",
180 /* WINED3DSIH_IMUL */ "imul",
181 /* WINED3DSIH_INE */ "ine",
182 /* WINED3DSIH_INEG */ "ineg",
183 /* WINED3DSIH_ISHL */ "ishl",
184 /* WINED3DSIH_ISHR */ "ishr",
185 /* WINED3DSIH_ITOF */ "itof",
186 /* WINED3DSIH_LABEL */ "label",
187 /* WINED3DSIH_LD */ "ld",
188 /* WINED3DSIH_LD2DMS */ "ld2dms",
189 /* WINED3DSIH_LD_RAW */ "ld_raw",
190 /* WINED3DSIH_LD_STRUCTURED */ "ld_structured",
191 /* WINED3DSIH_LD_UAV_TYPED */ "ld_uav_typed",
192 /* WINED3DSIH_LIT */ "lit",
193 /* WINED3DSIH_LOD */ "lod",
194 /* WINED3DSIH_LOG */ "log",
195 /* WINED3DSIH_LOGP */ "logp",
196 /* WINED3DSIH_LOOP */ "loop",
197 /* WINED3DSIH_LRP */ "lrp",
198 /* WINED3DSIH_LT */ "lt",
199 /* WINED3DSIH_M3x2 */ "m3x2",
200 /* WINED3DSIH_M3x3 */ "m3x3",
201 /* WINED3DSIH_M3x4 */ "m3x4",
202 /* WINED3DSIH_M4x3 */ "m4x3",
203 /* WINED3DSIH_M4x4 */ "m4x4",
204 /* WINED3DSIH_MAD */ "mad",
205 /* WINED3DSIH_MAX */ "max",
206 /* WINED3DSIH_MIN */ "min",
207 /* WINED3DSIH_MOV */ "mov",
208 /* WINED3DSIH_MOVA */ "mova",
209 /* WINED3DSIH_MOVC */ "movc",
210 /* WINED3DSIH_MUL */ "mul",
211 /* WINED3DSIH_NE */ "ne",
212 /* WINED3DSIH_NOP */ "nop",
213 /* WINED3DSIH_NOT */ "not",
214 /* WINED3DSIH_NRM */ "nrm",
215 /* WINED3DSIH_OR */ "or",
216 /* WINED3DSIH_PHASE */ "phase",
217 /* WINED3DSIH_POW */ "pow",
218 /* WINED3DSIH_RCP */ "rcp",
219 /* WINED3DSIH_REP */ "rep",
220 /* WINED3DSIH_RESINFO */ "resinfo",
221 /* WINED3DSIH_RET */ "ret",
222 /* WINED3DSIH_RETP */ "retp",
223 /* WINED3DSIH_ROUND_NE */ "round_ne",
224 /* WINED3DSIH_ROUND_NI */ "round_ni",
225 /* WINED3DSIH_ROUND_PI */ "round_pi",
226 /* WINED3DSIH_ROUND_Z */ "round_z",
227 /* WINED3DSIH_RSQ */ "rsq",
228 /* WINED3DSIH_SAMPLE */ "sample",
229 /* WINED3DSIH_SAMPLE_B */ "sample_b",
230 /* WINED3DSIH_SAMPLE_C */ "sample_c",
231 /* WINED3DSIH_SAMPLE_C_LZ */ "sample_c_lz",
232 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
233 /* WINED3DSIH_SAMPLE_INFO */ "sample_info",
234 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
235 /* WINED3DSIH_SAMPLE_POS */ "sample_pos",
236 /* WINED3DSIH_SETP */ "setp",
237 /* WINED3DSIH_SGE */ "sge",
238 /* WINED3DSIH_SGN */ "sgn",
239 /* WINED3DSIH_SINCOS */ "sincos",
240 /* WINED3DSIH_SLT */ "slt",
241 /* WINED3DSIH_SQRT */ "sqrt",
242 /* WINED3DSIH_STORE_RAW */ "store_raw",
243 /* WINED3DSIH_STORE_STRUCTURED */ "store_structured",
244 /* WINED3DSIH_STORE_UAV_TYPED */ "store_uav_typed",
245 /* WINED3DSIH_SUB */ "sub",
246 /* WINED3DSIH_SWAPC */ "swapc",
247 /* WINED3DSIH_SWITCH */ "switch",
248 /* WINED3DSIH_SYNC */ "sync",
249 /* WINED3DSIH_TEX */ "texld",
250 /* WINED3DSIH_TEXBEM */ "texbem",
251 /* WINED3DSIH_TEXBEML */ "texbeml",
252 /* WINED3DSIH_TEXCOORD */ "texcrd",
253 /* WINED3DSIH_TEXDEPTH */ "texdepth",
254 /* WINED3DSIH_TEXDP3 */ "texdp3",
255 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
256 /* WINED3DSIH_TEXKILL */ "texkill",
257 /* WINED3DSIH_TEXLDD */ "texldd",
258 /* WINED3DSIH_TEXLDL */ "texldl",
259 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
260 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
261 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
262 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
263 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
264 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
265 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
266 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
267 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
268 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
269 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
270 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
271 /* WINED3DSIH_UBFE */ "ubfe",
272 /* WINED3DSIH_UDIV */ "udiv",
273 /* WINED3DSIH_UGE */ "uge",
274 /* WINED3DSIH_ULT */ "ult",
275 /* WINED3DSIH_UMAX */ "umax",
276 /* WINED3DSIH_UMIN */ "umin",
277 /* WINED3DSIH_UMUL */ "umul",
278 /* WINED3DSIH_USHR */ "ushr",
279 /* WINED3DSIH_UTOF */ "utof",
280 /* WINED3DSIH_XOR */ "xor",
281 };
282
283 static const char * const semantic_names[] =
284 {
285 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
286 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
287 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
288 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
289 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
290 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
291 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
292 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
293 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
294 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
295 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
296 /* WINED3D_DECL_USAGE_FOG */ "FOG",
297 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
298 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
299 };
300
301 const char *debug_d3dshaderinstructionhandler(enum WINED3D_SHADER_INSTRUCTION_HANDLER handler_idx)
302 {
303 if (handler_idx >= ARRAY_SIZE(shader_opcode_names))
304 return wine_dbg_sprintf("UNRECOGNIZED(%#x)", handler_idx);
305
306 return shader_opcode_names[handler_idx];
307 }
308
309 enum vkd3d_shader_visibility vkd3d_shader_visibility_from_wined3d(enum wined3d_shader_type shader_type)
310 {
311 switch (shader_type)
312 {
313 case WINED3D_SHADER_TYPE_VERTEX:
314 return VKD3D_SHADER_VISIBILITY_VERTEX;
315 case WINED3D_SHADER_TYPE_HULL:
316 return VKD3D_SHADER_VISIBILITY_HULL;
317 case WINED3D_SHADER_TYPE_DOMAIN:
318 return VKD3D_SHADER_VISIBILITY_DOMAIN;
319 case WINED3D_SHADER_TYPE_GEOMETRY:
320 return VKD3D_SHADER_VISIBILITY_GEOMETRY;
321 case WINED3D_SHADER_TYPE_PIXEL:
322 return VKD3D_SHADER_VISIBILITY_PIXEL;
323 case WINED3D_SHADER_TYPE_COMPUTE:
324 return VKD3D_SHADER_VISIBILITY_COMPUTE;
325 default:
326 ERR("Invalid shader type %s.\n", debug_shader_type(shader_type));
327 return VKD3D_SHADER_VISIBILITY_ALL;
328 }
329 }
330
331 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
332 {
333 if (usage >= ARRAY_SIZE(semantic_names))
334 {
335 FIXME("Unrecognized usage %#x.\n", usage);
336 return "UNRECOGNIZED";
337 }
338
339 return semantic_names[usage];
340 }
341
342 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
343 {
344 unsigned int i;
345
346 for (i = 0; i < ARRAY_SIZE(semantic_names); ++i)
347 {
348 if (!strcmp(name, semantic_names[i]))
349 return i;
350 }
351
352 return ~0U;
353 }
354
355 static enum wined3d_sysval_semantic shader_sysval_semantic_from_usage(enum wined3d_decl_usage usage)
356 {
357 switch (usage)
358 {
359 case WINED3D_DECL_USAGE_POSITION:
360 return WINED3D_SV_POSITION;
361 default:
362 return 0;
363 }
364 }
365
366 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
367 {
368 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
369 }
370
371 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
372 const struct wined3d_shader_semantic *s)
373 {
374 e->semantic_name = shader_semantic_name_from_usage(s->usage);
375 e->semantic_idx = s->usage_idx;
376 e->stream_idx = 0;
377 e->sysval_semantic = shader_sysval_semantic_from_usage(s->usage);
378 e->component_type = WINED3D_TYPE_FLOAT;
379 e->register_idx = s->reg.reg.idx[0].offset;
380 e->mask = s->reg.write_mask;
381 }
382
383 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
384 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
385 {
386 e->semantic_name = shader_semantic_name_from_usage(usage);
387 e->semantic_idx = usage_idx;
388 e->stream_idx = 0;
389 e->sysval_semantic = shader_sysval_semantic_from_usage(usage);
390 e->component_type = WINED3D_TYPE_FLOAT;
391 e->register_idx = reg_idx;
392 e->mask = write_mask;
393 }
394
395 static const struct wined3d_shader_frontend *shader_select_frontend(enum vkd3d_shader_source_type source_type)
396 {
397 switch (source_type)
398 {
399 case VKD3D_SHADER_SOURCE_D3D_BYTECODE:
400 return &sm1_shader_frontend;
401
402 case VKD3D_SHADER_SOURCE_DXBC_TPF:
403 return &sm4_shader_frontend;
404
405 default:
406 WARN("Invalid source type %#x specified.\n", source_type);
407 return NULL;
408 }
409 }
410
411 void string_buffer_clear(struct wined3d_string_buffer *buffer)
412 {
413 buffer->buffer[0] = '\0';
414 buffer->content_size = 0;
415 }
416
417 BOOL string_buffer_init(struct wined3d_string_buffer *buffer)
418 {
419 buffer->buffer_size = 32;
420 if (!(buffer->buffer = malloc(buffer->buffer_size)))
421 {
422 ERR("Failed to allocate shader buffer memory.\n");
423 return FALSE;
424 }
425
426 string_buffer_clear(buffer);
427 return TRUE;
428 }
429
430 void string_buffer_free(struct wined3d_string_buffer *buffer)
431 {
432 free(buffer->buffer);
433 }
434
435 BOOL string_buffer_resize(struct wined3d_string_buffer *buffer, int rc)
436 {
437 char *new_buffer;
438 unsigned int new_buffer_size = buffer->buffer_size * 2;
439
440 while (rc > 0 && (unsigned int)rc >= new_buffer_size - buffer->content_size)
441 new_buffer_size *= 2;
442 if (!(new_buffer = realloc(buffer->buffer, new_buffer_size)))
443 {
444 ERR("Failed to grow buffer.\n");
445 buffer->buffer[buffer->content_size] = '\0';
446 return FALSE;
447 }
448 buffer->buffer = new_buffer;
449 buffer->buffer_size = new_buffer_size;
450 return TRUE;
451 }
452
453 int shader_vaddline(struct wined3d_string_buffer *buffer, const char *format, va_list args)
454 {
455 unsigned int rem;
456 int rc;
457
458 rem = buffer->buffer_size - buffer->content_size;
459 rc = vsnprintf(&buffer->buffer[buffer->content_size], rem, format, args);
460 if (rc < 0 /* C89 */ || (unsigned int)rc >= rem /* C99 */)
461 return rc;
462
463 buffer->content_size += rc;
464 return 0;
465 }
466
467 int shader_addline(struct wined3d_string_buffer *buffer, const char *format, ...)
468 {
469 va_list args;
470 int ret;
471
472 for (;;)
473 {
474 va_start(args, format);
475 ret = shader_vaddline(buffer, format, args);
476 va_end(args);
477 if (!ret)
478 return ret;
479 if (!string_buffer_resize(buffer, ret))
480 return -1;
481 }
482 }
483
484 struct wined3d_string_buffer *string_buffer_get(struct wined3d_string_buffer_list *list)
485 {
486 struct wined3d_string_buffer *buffer;
487
488 if (list_empty(&list->list))
489 {
490 buffer = malloc(sizeof(*buffer));
491 if (!buffer || !string_buffer_init(buffer))
492 {
493 ERR("Couldn't allocate buffer for temporary string.\n");
494 free(buffer);
495 return NULL;
496 }
497 }
498 else
499 {
500 buffer = LIST_ENTRY(list_head(&list->list), struct wined3d_string_buffer, entry);
501 list_remove(&buffer->entry);
502 }
503 string_buffer_clear(buffer);
504 return buffer;
505 }
506
507 static int string_buffer_vsprintf(struct wined3d_string_buffer *buffer, const char *format, va_list args)
508 {
509 if (!buffer)
510 return 0;
511 string_buffer_clear(buffer);
512 return shader_vaddline(buffer, format, args);
513 }
514
515 void string_buffer_sprintf(struct wined3d_string_buffer *buffer, const char *format, ...)
516 {
517 va_list args;
518 int ret;
519
520 for (;;)
521 {
522 va_start(args, format);
523 ret = string_buffer_vsprintf(buffer, format, args);
524 va_end(args);
525 if (!ret)
526 return;
527 if (!string_buffer_resize(buffer, ret))
528 return;
529 }
530 }
531
532 void string_buffer_release(struct wined3d_string_buffer_list *list, struct wined3d_string_buffer *buffer)
533 {
534 if (!buffer)
535 return;
536 list_add_head(&list->list, &buffer->entry);
537 }
538
539 void string_buffer_list_init(struct wined3d_string_buffer_list *list)
540 {
541 list_init(&list->list);
542 }
543
544 void string_buffer_list_cleanup(struct wined3d_string_buffer_list *list)
545 {
546 struct wined3d_string_buffer *buffer, *buffer_next;
547
548 LIST_FOR_EACH_ENTRY_SAFE(buffer, buffer_next, &list->list, struct wined3d_string_buffer, entry)
549 {
550 string_buffer_free(buffer);
551 free(buffer);
552 }
553 list_init(&list->list);
554 }
555
556 static void shader_delete_constant_list(struct list *clist)
557 {
558 struct wined3d_shader_lconst *constant, *constant_next;
559
560 LIST_FOR_EACH_ENTRY_SAFE(constant, constant_next, clist, struct wined3d_shader_lconst, entry)
561 free(constant);
562 list_init(clist);
563 }
564
565 static void shader_set_limits(struct wined3d_shader *shader)
566 {
567 static const struct limits_entry
568 {
569 unsigned int min_version;
570 unsigned int max_version;
571 struct wined3d_shader_limits limits;
572 }
573 vs_limits[] =
574 {
575 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
576 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 1), { 0, 0, 256, 0, 12, 0}},
577 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 1), { 0, 16, 256, 16, 12, 0}},
578 /* DX10 cards on Windows advertise a D3D9 constant limit of 256
579 * even though they are capable of supporting much more (GL
580 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
581 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
582 * shaders to 256. */
583 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), { 4, 16, 256, 16, 12, 0}},
584 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 16, 0}},
585 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 0}},
586 {0}
587 },
588 hs_limits[] =
589 {
590 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
591 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
592 },
593 ds_limits[] =
594 {
595 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
596 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
597 },
598 gs_limits[] =
599 {
600 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
601 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 32, 16}},
602 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
603 {0}
604 },
605 ps_limits[] =
606 {
607 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
608 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 3), { 4, 0, 8, 0, 0, 0}},
609 {WINED3D_SHADER_VERSION(1, 4), WINED3D_SHADER_VERSION(1, 4), { 6, 0, 8, 0, 0, 0}},
610 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 0), {16, 0, 32, 0, 0, 0}},
611 {WINED3D_SHADER_VERSION(2, 1), WINED3D_SHADER_VERSION(2, 1), {16, 16, 32, 16, 0, 0}},
612 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), {16, 16, 224, 16, 0, 10}},
613 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 32}},
614 {0}
615 },
616 cs_limits[] =
617 {
618 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
619 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 0}},
620 };
621 const struct limits_entry *limits_array;
622 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
623 shader->reg_maps.shader_version.minor);
624 int i = 0;
625
626 switch (shader->reg_maps.shader_version.type)
627 {
628 default:
629 FIXME("Unexpected shader type %u found.\n", shader->reg_maps.shader_version.type);
630 /* Fall-through. */
631 case WINED3D_SHADER_TYPE_VERTEX:
632 limits_array = vs_limits;
633 break;
634 case WINED3D_SHADER_TYPE_HULL:
635 limits_array = hs_limits;
636 break;
637 case WINED3D_SHADER_TYPE_DOMAIN:
638 limits_array = ds_limits;
639 break;
640 case WINED3D_SHADER_TYPE_GEOMETRY:
641 limits_array = gs_limits;
642 break;
643 case WINED3D_SHADER_TYPE_PIXEL:
644 limits_array = ps_limits;
645 break;
646 case WINED3D_SHADER_TYPE_COMPUTE:
647 limits_array = cs_limits;
648 break;
649 }
650
651 while (limits_array[i].min_version && limits_array[i].min_version <= shader_version)
652 {
653 if (shader_version <= limits_array[i].max_version)
654 {
655 shader->limits = &limits_array[i].limits;
656 break;
657 }
658 ++i;
659 }
660 if (!shader->limits)
661 {
662 FIXME("Unexpected shader version \"%u.%u\" (shader type %u).\n",
663 shader->reg_maps.shader_version.major,
664 shader->reg_maps.shader_version.minor,
665 shader->reg_maps.shader_version.type);
666 shader->limits = &limits_array[max(0, i - 1)].limits;
667 }
668 }
669
670 static BOOL shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
671 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type, unsigned int constf_size)
672 {
673 switch (reg->type)
674 {
675 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
676 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
677 reg_maps->texcoord |= 1u << reg->idx[0].offset;
678 else
679 reg_maps->address |= 1u << reg->idx[0].offset;
680 break;
681
682 case WINED3DSPR_TEMP:
683 reg_maps->temporary |= 1u << reg->idx[0].offset;
684 break;
685
686 case WINED3DSPR_INPUT:
687 if (reg->idx[0].rel_addr)
688 reg_maps->input_rel_addressing = 1;
689 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
690 {
691 /* If relative addressing is used, we must assume that all
692 * registers are used. Even if it is a construct like v3[aL],
693 * we can't assume that v0, v1 and v2 aren't read because aL
694 * can be negative. */
695 if (reg->idx[0].rel_addr)
696 shader->u.ps.input_reg_used = ~0u;
697 else
698 shader->u.ps.input_reg_used |= 1u << reg->idx[0].offset;
699 }
700 else
701 {
702 reg_maps->input_registers |= 1u << reg->idx[0].offset;
703 }
704 break;
705
706 case WINED3DSPR_RASTOUT:
707 if (reg->idx[0].offset == 1)
708 reg_maps->fog = 1;
709 if (reg->idx[0].offset == 2)
710 reg_maps->point_size = 1;
711 break;
712
713 case WINED3DSPR_MISCTYPE:
714 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
715 {
716 if (!reg->idx[0].offset)
717 reg_maps->vpos = 1;
718 else if (reg->idx[0].offset == 1)
719 reg_maps->usesfacing = 1;
720 }
721 break;
722
723 case WINED3DSPR_CONST:
724 if (reg->idx[0].rel_addr)
725 {
726 if (reg->idx[0].offset < reg_maps->min_rel_offset)
727 reg_maps->min_rel_offset = reg->idx[0].offset;
728 if (reg->idx[0].offset > reg_maps->max_rel_offset)
729 reg_maps->max_rel_offset = reg->idx[0].offset;
730 reg_maps->usesrelconstF = TRUE;
731 }
732 else
733 {
734 if (reg->idx[0].offset >= min(shader->limits->constant_float, constf_size))
735 {
736 WARN("Shader using float constant %u which is not supported.\n", reg->idx[0].offset);
737 return FALSE;
738 }
739 else
740 {
741 wined3d_insert_bits(reg_maps->constf, reg->idx[0].offset, 1, 0x1);
742 }
743 }
744 break;
745
746 case WINED3DSPR_CONSTINT:
747 if (reg->idx[0].offset >= shader->limits->constant_int)
748 {
749 WARN("Shader using integer constant %u which is not supported.\n", reg->idx[0].offset);
750 return FALSE;
751 }
752 else
753 {
754 reg_maps->integer_constants |= (1u << reg->idx[0].offset);
755 }
756 break;
757
758 case WINED3DSPR_CONSTBOOL:
759 if (reg->idx[0].offset >= shader->limits->constant_bool)
760 {
761 WARN("Shader using bool constant %u which is not supported.\n", reg->idx[0].offset);
762 return FALSE;
763 }
764 else
765 {
766 reg_maps->boolean_constants |= (1u << reg->idx[0].offset);
767 }
768 break;
769
770 case WINED3DSPR_COLOROUT:
771 reg_maps->rt_mask |= (1u << reg->idx[0].offset);
772 break;
773
774 case WINED3DSPR_OUTCONTROLPOINT:
775 reg_maps->vocp = 1;
776 break;
777
778 case WINED3DSPR_SAMPLEMASK:
779 reg_maps->sample_mask = 1;
780 break;
781
782 case WINED3DSPR_STENCILREF:
783 reg_maps->stencil_ref = 1;
784 break;
785
786 default:
787 TRACE("Not recording register of type %#x and [%#x][%#x].\n",
788 reg->type, reg->idx[0].offset, reg->idx[1].offset);
789 break;
790 }
791 return TRUE;
792 }
793
794 static void shader_record_sample(struct wined3d_shader_reg_maps *reg_maps,
795 unsigned int resource_idx, unsigned int sampler_idx, unsigned int bind_idx)
796 {
797 struct wined3d_shader_sampler_map_entry *entries, *entry;
798 struct wined3d_shader_sampler_map *map;
799 unsigned int i;
800
801 map = &reg_maps->sampler_map;
802 entries = map->entries;
803 for (i = 0; i < map->count; ++i)
804 {
805 if (entries[i].resource_idx == resource_idx && entries[i].sampler_idx == sampler_idx)
806 return;
807 }
808
809 if (!map->size)
810 {
811 if (!(entries = calloc(4, sizeof(*entries))))
812 {
813 ERR("Failed to allocate sampler map entries.\n");
814 return;
815 }
816 map->size = 4;
817 map->entries = entries;
818 }
819 else if (map->count == map->size)
820 {
821 size_t new_size = map->size * 2;
822
823 if (sizeof(*entries) * new_size <= sizeof(*entries) * map->size
824 || !(entries = realloc(entries, sizeof(*entries) * new_size)))
825 {
826 ERR("Failed to resize sampler map entries.\n");
827 return;
828 }
829 map->size = new_size;
830 map->entries = entries;
831 }
832
833 entry = &entries[map->count++];
834 entry->resource_idx = resource_idx;
835 entry->sampler_idx = sampler_idx;
836 entry->bind_idx = bind_idx;
837 }
838
839 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
840 {
841 switch (instr)
842 {
843 case WINED3DSIH_M4x4:
844 case WINED3DSIH_M3x4:
845 return param == 1 ? 3 : 0;
846
847 case WINED3DSIH_M4x3:
848 case WINED3DSIH_M3x3:
849 return param == 1 ? 2 : 0;
850
851 case WINED3DSIH_M3x2:
852 return param == 1 ? 1 : 0;
853
854 default:
855 return 0;
856 }
857 }
858
859 static HRESULT shader_reg_maps_add_tgsm(struct wined3d_shader_reg_maps *reg_maps,
860 unsigned int register_idx, unsigned int size, unsigned int stride)
861 {
862 struct wined3d_shader_tgsm *tgsm;
863
864 if (register_idx >= MAX_TGSM_REGISTERS)
865 {
866 ERR("Invalid TGSM register index %u.\n", register_idx);
867 return S_OK;
868 }
869 if (reg_maps->shader_version.type != WINED3D_SHADER_TYPE_COMPUTE)
870 {
871 FIXME("TGSM declarations are allowed only in compute shaders.\n");
872 return S_OK;
873 }
874
875 if (!wined3d_array_reserve((void **)&reg_maps->tgsm, &reg_maps->tgsm_capacity,
876 register_idx + 1, sizeof(*reg_maps->tgsm)))
877 return E_OUTOFMEMORY;
878
879 reg_maps->tgsm_count = max(register_idx + 1, reg_maps->tgsm_count);
880 tgsm = &reg_maps->tgsm[register_idx];
881 tgsm->size = size;
882 tgsm->stride = stride;
883 return S_OK;
884 }
885
886 static HRESULT shader_record_shader_phase(struct wined3d_shader *shader,
887 struct wined3d_shader_phase **current_phase, const struct wined3d_shader_instruction *ins,
888 const DWORD *current_instruction_ptr, const DWORD *previous_instruction_ptr)
889 {
890 struct wined3d_shader_phase *phase;
891
892 if ((phase = *current_phase))
893 {
894 phase->end = previous_instruction_ptr;
895 *current_phase = NULL;
896 }
897
898 if (shader->reg_maps.shader_version.type != WINED3D_SHADER_TYPE_HULL)
899 {
900 ERR("Unexpected shader type %s.\n", debug_shader_type(shader->reg_maps.shader_version.type));
901 return E_FAIL;
902 }
903
904 switch (ins->handler_idx)
905 {
906 case WINED3DSIH_HS_CONTROL_POINT_PHASE:
907 if (shader->u.hs.phases.control_point)
908 {
909 FIXME("Multiple control point phases.\n");
910 free(shader->u.hs.phases.control_point);
911 }
912 if (!(shader->u.hs.phases.control_point = calloc(1, sizeof(*shader->u.hs.phases.control_point))))
913 return E_OUTOFMEMORY;
914 phase = shader->u.hs.phases.control_point;
915 break;
916 case WINED3DSIH_HS_FORK_PHASE:
917 if (!wined3d_array_reserve((void **)&shader->u.hs.phases.fork,
918 &shader->u.hs.phases.fork_size, shader->u.hs.phases.fork_count + 1,
919 sizeof(*shader->u.hs.phases.fork)))
920 return E_OUTOFMEMORY;
921 phase = &shader->u.hs.phases.fork[shader->u.hs.phases.fork_count++];
922 break;
923 case WINED3DSIH_HS_JOIN_PHASE:
924 if (!wined3d_array_reserve((void **)&shader->u.hs.phases.join,
925 &shader->u.hs.phases.join_size, shader->u.hs.phases.join_count + 1,
926 sizeof(*shader->u.hs.phases.join)))
927 return E_OUTOFMEMORY;
928 phase = &shader->u.hs.phases.join[shader->u.hs.phases.join_count++];
929 break;
930 default:
931 ERR("Unexpected opcode %s.\n", debug_d3dshaderinstructionhandler(ins->handler_idx));
932 return E_FAIL;
933 }
934
935 phase->start = current_instruction_ptr;
936 *current_phase = phase;
937
938 return WINED3D_OK;
939 }
940
941 static HRESULT shader_calculate_clip_or_cull_distance_mask(
942 const struct wined3d_shader_signature_element *e, unsigned int *mask)
943 {
944 /* Clip and cull distances are packed in 4 component registers. 0 and 1 are
945 * the only allowed semantic indices.
946 */
947 if (e->semantic_idx >= WINED3D_MAX_CLIP_DISTANCES / 4)
948 {
949 *mask = 0;
950 WARN("Invalid clip/cull distance index %u.\n", e->semantic_idx);
951 return WINED3DERR_INVALIDCALL;
952 }
953
954 *mask = (e->mask & WINED3DSP_WRITEMASK_ALL) << (4 * e->semantic_idx);
955 return WINED3D_OK;
956 }
957
958 static void wined3d_insert_interpolation_mode(uint32_t *packed_interpolation_mode,
959 unsigned int register_idx, enum wined3d_shader_interpolation_mode mode)
960 {
961 if (mode > WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE)
962 FIXME("Unexpected interpolation mode %#x.\n", mode);
963
964 wined3d_insert_bits(packed_interpolation_mode,
965 register_idx * WINED3D_PACKED_INTERPOLATION_BIT_COUNT, WINED3D_PACKED_INTERPOLATION_BIT_COUNT, mode);
966 }
967
968 static HRESULT shader_scan_output_signature(struct wined3d_shader *shader)
969 {
970 const struct wined3d_shader_signature *output_signature = &shader->output_signature;
971 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
972 unsigned int i;
973 HRESULT hr;
974
975 for (i = 0; i < output_signature->element_count; ++i)
976 {
977 const struct wined3d_shader_signature_element *e = &output_signature->elements[i];
978 unsigned int mask;
979
980 reg_maps->output_registers |= 1u << e->register_idx;
981 if (e->sysval_semantic == WINED3D_SV_CLIP_DISTANCE)
982 {
983 if (FAILED(hr = shader_calculate_clip_or_cull_distance_mask(e, &mask)))
984 return hr;
985 reg_maps->clip_distance_mask |= mask;
986 }
987 else if (e->sysval_semantic == WINED3D_SV_CULL_DISTANCE)
988 {
989 if (FAILED(hr = shader_calculate_clip_or_cull_distance_mask(e, &mask)))
990 return hr;
991 reg_maps->cull_distance_mask |= mask;
992 }
993 else if (e->sysval_semantic == WINED3D_SV_VIEWPORT_ARRAY_INDEX)
994 {
995 reg_maps->viewport_array = 1;
996 }
997 }
998
999 return WINED3D_OK;
1000 }
1001
1002 /* Note that this does not count the loop register as an address register. */
1003 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, DWORD constf_size)
1004 {
1005 struct wined3d_shader_signature_element input_signature_elements[max(MAX_ATTRIBS, MAX_REG_INPUT)];
1006 struct wined3d_shader_signature_element output_signature_elements[MAX_REG_OUTPUT];
1007 struct wined3d_shader_signature *output_signature = &shader->output_signature;
1008 struct wined3d_shader_signature *input_signature = &shader->input_signature;
1009 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
1010 const struct wined3d_shader_frontend *fe = shader->frontend;
1011 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
1012 struct wined3d_shader_version shader_version;
1013 struct wined3d_shader_phase *phase = NULL;
1014 const DWORD *ptr, *prev_ins, *current_ins;
1015 void *fe_data = shader->frontend_data;
1016 unsigned int i;
1017 HRESULT hr;
1018
1019 memset(reg_maps, 0, sizeof(*reg_maps));
1020 memset(input_signature_elements, 0, sizeof(input_signature_elements));
1021 memset(output_signature_elements, 0, sizeof(output_signature_elements));
1022 reg_maps->min_rel_offset = ~0U;
1023 list_init(&reg_maps->indexable_temps);
1024
1025 fe->shader_read_header(fe_data, &ptr, &shader_version);
1026 prev_ins = current_ins = ptr;
1027 reg_maps->shader_version = shader_version;
1028
1029 shader_set_limits(shader);
1030
1031 if (!(reg_maps->constf = calloc(((min(shader->limits->constant_float, constf_size) + 31) / 32),
1032 sizeof(*reg_maps->constf))))
1033 {
1034 ERR("Failed to allocate constant map memory.\n");
1035 return E_OUTOFMEMORY;
1036 }
1037
1038 while (!fe->shader_is_end(fe_data, &ptr))
1039 {
1040 struct wined3d_shader_instruction ins;
1041
1042 current_ins = ptr;
1043 /* Fetch opcode. */
1044 fe->shader_read_instruction(fe_data, &ptr, &ins);
1045
1046 /* Unhandled opcode, and its parameters. */
1047 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1048 {
1049 WARN("Encountered unrecognised or invalid instruction.\n");
1050 return WINED3DERR_INVALIDCALL;
1051 }
1052
1053 /* Handle declarations. */
1054 if (ins.handler_idx == WINED3DSIH_DCL
1055 || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
1056 {
1057 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
1058 unsigned int reg_idx = semantic->reg.reg.idx[0].offset;
1059
1060 switch (semantic->reg.reg.type)
1061 {
1062 /* Mark input registers used. */
1063 case WINED3DSPR_INPUT:
1064 if (reg_idx >= MAX_REG_INPUT)
1065 {
1066 ERR("Invalid input register index %u.\n", reg_idx);
1067 break;
1068 }
1069 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL && shader_version.major == 3
1070 && semantic->usage == WINED3D_DECL_USAGE_POSITION && !semantic->usage_idx)
1071 return WINED3DERR_INVALIDCALL;
1072 reg_maps->input_registers |= 1u << reg_idx;
1073 shader_signature_from_semantic(&input_signature_elements[reg_idx], semantic);
1074 break;
1075
1076 /* Vertex shader: mark 3.0 output registers used, save token. */
1077 case WINED3DSPR_OUTPUT:
1078 if (reg_idx >= MAX_REG_OUTPUT)
1079 {
1080 ERR("Invalid output register index %u.\n", reg_idx);
1081 break;
1082 }
1083 reg_maps->output_registers |= 1u << reg_idx;
1084 shader_signature_from_semantic(&output_signature_elements[reg_idx], semantic);
1085 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
1086 reg_maps->fog = 1;
1087 if (semantic->usage == WINED3D_DECL_USAGE_PSIZE)
1088 reg_maps->point_size = 1;
1089 break;
1090
1091 case WINED3DSPR_SAMPLER:
1092 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1093 case WINED3DSPR_RESOURCE:
1094 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1095 {
1096 ERR("Invalid resource index %u.\n", reg_idx);
1097 break;
1098 }
1099 reg_maps->resource_info[reg_idx].type = semantic->resource_type;
1100 if (semantic->resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2DMS && semantic->sample_count == 1)
1101 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1102 if (semantic->resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY && semantic->sample_count == 1)
1103 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY;
1104 reg_maps->resource_info[reg_idx].data_type = semantic->resource_data_type;
1105 wined3d_bitmap_set(reg_maps->resource_map, reg_idx);
1106 break;
1107
1108 case WINED3DSPR_UAV:
1109 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1110 {
1111 ERR("Invalid UAV resource index %u.\n", reg_idx);
1112 break;
1113 }
1114 reg_maps->uav_resource_info[reg_idx].type = semantic->resource_type;
1115 reg_maps->uav_resource_info[reg_idx].data_type = semantic->resource_data_type;
1116 if (ins.flags)
1117 FIXME("Ignoring typed UAV flags %#x.\n", ins.flags);
1118 break;
1119
1120 default:
1121 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
1122 break;
1123 }
1124 }
1125 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
1126 {
1127 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
1128 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
1129 {
1130 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
1131 }
1132 else
1133 {
1134 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
1135 wined3d_bitmap_set(&reg_maps->cb_map, reg->idx[0].offset);
1136 }
1137 }
1138 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
1139 {
1140 if (ins.flags & WINED3DSGF_FORCE_EARLY_DEPTH_STENCIL)
1141 {
1142 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1143 shader->u.ps.force_early_depth_stencil = TRUE;
1144 else
1145 FIXME("Invalid instruction %#x for shader type %#x.\n",
1146 ins.handler_idx, shader_version.type);
1147 }
1148 else
1149 {
1150 WARN("Ignoring global flags %#x.\n", ins.flags);
1151 }
1152 }
1153 else if (ins.handler_idx == WINED3DSIH_DCL_GS_INSTANCES)
1154 {
1155 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1156 shader->u.gs.instance_count = ins.declaration.count;
1157 else
1158 FIXME("Invalid instruction %#x for shader type %#x.\n",
1159 ins.handler_idx, shader_version.type);
1160 }
1161 else if (ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
1162 || ins.handler_idx == WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT)
1163 {
1164 if (phase)
1165 phase->instance_count = ins.declaration.count;
1166 else
1167 FIXME("Instruction %s outside of shader phase.\n",
1168 debug_d3dshaderinstructionhandler(ins.handler_idx));
1169 }
1170 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
1171 {
1172 if (reg_maps->icb)
1173 FIXME("Multiple immediate constant buffers.\n");
1174 reg_maps->icb = ins.declaration.icb;
1175 }
1176 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
1177 {
1178 if (phase)
1179 {
1180 FIXME("Indexable temporary registers not supported.\n");
1181 }
1182 else
1183 {
1184 struct wined3d_shader_indexable_temp *reg;
1185
1186 if (!(reg = malloc(sizeof(*reg))))
1187 return E_OUTOFMEMORY;
1188
1189 *reg = ins.declaration.indexable_temp;
1190 list_add_tail(&reg_maps->indexable_temps, &reg->entry);
1191 }
1192 }
1193 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
1194 {
1195 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1196 shader->u.gs.input_type = ins.declaration.primitive_type.type;
1197 else
1198 FIXME("Invalid instruction %#x for shader type %#x.\n",
1199 ins.handler_idx, shader_version.type);
1200 }
1201 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
1202 {
1203 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1204 if (reg_idx >= MAX_REG_INPUT)
1205 {
1206 ERR("Invalid register index %u.\n", reg_idx);
1207 break;
1208 }
1209 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1210 wined3d_insert_interpolation_mode(shader->u.ps.interpolation_mode, reg_idx, ins.flags);
1211 else
1212 FIXME("Invalid instruction %#x for shader type %#x.\n",
1213 ins.handler_idx, shader_version.type);
1214 }
1215 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
1216 {
1217 if (ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUT
1218 || ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUTGE
1219 || ins.declaration.dst.reg.type == WINED3DSPR_DEPTHOUTLE)
1220 {
1221 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1222 shader->u.ps.depth_output = ins.declaration.dst.reg.type;
1223 else
1224 FIXME("Invalid instruction %#x for shader type %#x.\n",
1225 ins.handler_idx, shader_version.type);
1226 }
1227 }
1228 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT)
1229 {
1230 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1231 shader->u.hs.output_vertex_count = ins.declaration.count;
1232 else
1233 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1234 }
1235 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
1236 {
1237 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1238 shader->u.gs.output_type = ins.declaration.primitive_type.type;
1239 else
1240 FIXME("Invalid instruction %#x for shader type %#x.\n",
1241 ins.handler_idx, shader_version.type);
1242 }
1243 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_RAW)
1244 {
1245 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1246 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1247 {
1248 ERR("Invalid resource index %u.\n", reg_idx);
1249 break;
1250 }
1251 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1252 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1253 reg_maps->resource_info[reg_idx].flags = WINED3D_VIEW_BUFFER_RAW;
1254 wined3d_bitmap_set(reg_maps->resource_map, reg_idx);
1255 }
1256 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
1257 {
1258 unsigned int reg_idx = ins.declaration.structured_resource.reg.reg.idx[0].offset;
1259 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1260 {
1261 ERR("Invalid resource index %u.\n", reg_idx);
1262 break;
1263 }
1264 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1265 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1266 reg_maps->resource_info[reg_idx].flags = 0;
1267 reg_maps->resource_info[reg_idx].stride = ins.declaration.structured_resource.byte_stride / 4;
1268 wined3d_bitmap_set(reg_maps->resource_map, reg_idx);
1269 }
1270 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
1271 {
1272 if (ins.flags & WINED3DSI_SAMPLER_COMPARISON_MODE)
1273 reg_maps->sampler_comparison_mode |= (1u << ins.declaration.dst.reg.idx[0].offset);
1274 }
1275 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS)
1276 {
1277 if (phase)
1278 phase->temporary_count = ins.declaration.count;
1279 else
1280 reg_maps->temporary_count = ins.declaration.count;
1281 }
1282 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
1283 {
1284 if (shader_version.type == WINED3D_SHADER_TYPE_DOMAIN)
1285 shader->u.ds.tessellator_domain = ins.declaration.tessellator_domain;
1286 else if (shader_version.type != WINED3D_SHADER_TYPE_HULL)
1287 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1288 }
1289 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
1290 {
1291 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1292 shader->u.hs.tessellator_output_primitive = ins.declaration.tessellator_output_primitive;
1293 else
1294 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1295 }
1296 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
1297 {
1298 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1299 shader->u.hs.tessellator_partitioning = ins.declaration.tessellator_partitioning;
1300 else
1301 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1302 }
1303 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
1304 {
1305 if (FAILED(hr = shader_reg_maps_add_tgsm(reg_maps, ins.declaration.tgsm_raw.reg.reg.idx[0].offset,
1306 ins.declaration.tgsm_raw.byte_count / 4, 0)))
1307 return hr;
1308 }
1309 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
1310 {
1311 unsigned int stride = ins.declaration.tgsm_structured.byte_stride / 4;
1312 unsigned int size = stride * ins.declaration.tgsm_structured.structure_count;
1313 if (FAILED(hr = shader_reg_maps_add_tgsm(reg_maps,
1314 ins.declaration.tgsm_structured.reg.reg.idx[0].offset, size, stride)))
1315 return hr;
1316 }
1317 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
1318 {
1319 if (shader_version.type == WINED3D_SHADER_TYPE_COMPUTE)
1320 {
1321 shader->u.cs.thread_group_size = ins.declaration.thread_group_size;
1322 }
1323 else
1324 {
1325 FIXME("Invalid instruction %#x for shader type %#x.\n",
1326 ins.handler_idx, shader_version.type);
1327 }
1328 }
1329 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
1330 {
1331 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1332 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1333 {
1334 ERR("Invalid UAV resource index %u.\n", reg_idx);
1335 break;
1336 }
1337 if (ins.flags)
1338 FIXME("Ignoring raw UAV flags %#x.\n", ins.flags);
1339 reg_maps->uav_resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1340 reg_maps->uav_resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1341 reg_maps->uav_resource_info[reg_idx].flags = WINED3D_VIEW_BUFFER_RAW;
1342 }
1343 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
1344 {
1345 unsigned int reg_idx = ins.declaration.structured_resource.reg.reg.idx[0].offset;
1346 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1347 {
1348 ERR("Invalid UAV resource index %u.\n", reg_idx);
1349 break;
1350 }
1351 if (ins.flags)
1352 FIXME("Ignoring structured UAV flags %#x.\n", ins.flags);
1353 reg_maps->uav_resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1354 reg_maps->uav_resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1355 reg_maps->uav_resource_info[reg_idx].flags = 0;
1356 reg_maps->uav_resource_info[reg_idx].stride = ins.declaration.structured_resource.byte_stride / 4;
1357 }
1358 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1359 {
1360 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1361 shader->u.gs.vertices_out = ins.declaration.count;
1362 else
1363 FIXME("Invalid instruction %#x for shader type %#x.\n",
1364 ins.handler_idx, shader_version.type);
1365 }
1366 else if (ins.handler_idx == WINED3DSIH_DEF)
1367 {
1368 struct wined3d_shader_lconst *lconst;
1369 float *value;
1370
1371 if (!(lconst = malloc(sizeof(*lconst))))
1372 return E_OUTOFMEMORY;
1373
1374 lconst->idx = ins.dst[0].reg.idx[0].offset;
1375 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1376 value = (float *)lconst->value;
1377
1378 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
1379 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1380 {
1381 if (value[0] < -1.0f) value[0] = -1.0f;
1382 else if (value[0] > 1.0f) value[0] = 1.0f;
1383 if (value[1] < -1.0f) value[1] = -1.0f;
1384 else if (value[1] > 1.0f) value[1] = 1.0f;
1385 if (value[2] < -1.0f) value[2] = -1.0f;
1386 else if (value[2] > 1.0f) value[2] = 1.0f;
1387 if (value[3] < -1.0f) value[3] = -1.0f;
1388 else if (value[3] > 1.0f) value[3] = 1.0f;
1389 }
1390
1391 list_add_head(&shader->constantsF, &lconst->entry);
1392
1393 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
1394 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
1395 {
1396 shader->lconst_inf_or_nan = TRUE;
1397 }
1398 }
1399 else if (ins.handler_idx == WINED3DSIH_DEFI)
1400 {
1401 struct wined3d_shader_lconst *lconst;
1402
1403 if (!(lconst = malloc(sizeof(*lconst))))
1404 return E_OUTOFMEMORY;
1405
1406 lconst->idx = ins.dst[0].reg.idx[0].offset;
1407 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1408
1409 list_add_head(&shader->constantsI, &lconst->entry);
1410 reg_maps->local_int_consts |= (1u << lconst->idx);
1411 }
1412 else if (ins.handler_idx == WINED3DSIH_DEFB)
1413 {
1414 struct wined3d_shader_lconst *lconst;
1415
1416 if (!(lconst = malloc(sizeof(*lconst))))
1417 return E_OUTOFMEMORY;
1418
1419 lconst->idx = ins.dst[0].reg.idx[0].offset;
1420 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, sizeof(DWORD));
1421
1422 list_add_head(&shader->constantsB, &lconst->entry);
1423 reg_maps->local_bool_consts |= (1u << lconst->idx);
1424 }
1425 /* Handle shader phases. */
1426 else if (ins.handler_idx == WINED3DSIH_HS_CONTROL_POINT_PHASE
1427 || ins.handler_idx == WINED3DSIH_HS_FORK_PHASE
1428 || ins.handler_idx == WINED3DSIH_HS_JOIN_PHASE)
1429 {
1430 if (FAILED(hr = shader_record_shader_phase(shader, &phase, &ins, current_ins, prev_ins)))
1431 return hr;
1432 }
1433 /* For subroutine prototypes. */
1434 else if (ins.handler_idx == WINED3DSIH_LABEL)
1435 {
1436 reg_maps->labels |= 1u << ins.src[0].reg.idx[0].offset;
1437 }
1438 /* Set texture, address, temporary registers. */
1439 else
1440 {
1441 BOOL color0_mov = FALSE;
1442 unsigned int i;
1443
1444 /* This will loop over all the registers and try to
1445 * make a bitmask of the ones we're interested in.
1446 *
1447 * Relative addressing tokens are ignored, but that's
1448 * okay, since we'll catch any address registers when
1449 * they are initialized (required by spec). */
1450 for (i = 0; i < ins.dst_count; ++i)
1451 {
1452 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
1453 shader_version.type, constf_size))
1454 return WINED3DERR_INVALIDCALL;
1455
1456 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1457 {
1458 UINT idx = ins.dst[i].reg.idx[0].offset;
1459
1460 switch (ins.dst[i].reg.type)
1461 {
1462 case WINED3DSPR_RASTOUT:
1463 if (shader_version.major >= 3)
1464 break;
1465 switch (idx)
1466 {
1467 case 0: /* oPos */
1468 reg_maps->output_registers |= 1u << 10;
1469 shader_signature_from_usage(&output_signature_elements[10],
1470 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
1471 break;
1472
1473 case 1: /* oFog */
1474 reg_maps->output_registers |= 1u << 11;
1475 shader_signature_from_usage(&output_signature_elements[11],
1476 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
1477 break;
1478
1479 case 2: /* oPts */
1480 reg_maps->output_registers |= 1u << 11;
1481 shader_signature_from_usage(&output_signature_elements[11],
1482 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
1483 break;
1484 }
1485 break;
1486
1487 case WINED3DSPR_ATTROUT:
1488 if (shader_version.major >= 3)
1489 break;
1490 if (idx < 2)
1491 {
1492 idx += 8;
1493 if (reg_maps->output_registers & (1u << idx))
1494 {
1495 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1496 }
1497 else
1498 {
1499 reg_maps->output_registers |= 1u << idx;
1500 shader_signature_from_usage(&output_signature_elements[idx],
1501 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
1502 }
1503 }
1504 break;
1505
1506 case WINED3DSPR_TEXCRDOUT: /* WINED3DSPR_OUTPUT */
1507 if (shader_version.major >= 3)
1508 {
1509 if (idx >= ARRAY_SIZE(reg_maps->u.output_registers_mask))
1510 {
1511 WARN("Invalid output register index %u.\n", idx);
1512 break;
1513 }
1514 reg_maps->u.output_registers_mask[idx] |= ins.dst[i].write_mask;
1515 break;
1516 }
1517 if (idx >= ARRAY_SIZE(reg_maps->u.texcoord_mask))
1518 {
1519 WARN("Invalid texcoord index %u.\n", idx);
1520 break;
1521 }
1522 reg_maps->u.texcoord_mask[idx] |= ins.dst[i].write_mask;
1523 if (reg_maps->output_registers & (1u << idx))
1524 {
1525 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1526 }
1527 else
1528 {
1529 reg_maps->output_registers |= 1u << idx;
1530 shader_signature_from_usage(&output_signature_elements[idx],
1531 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
1532 }
1533 break;
1534
1535 default:
1536 break;
1537 }
1538 }
1539
1540 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1541 {
1542 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
1543 {
1544 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
1545 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
1546 * the mov and perform the sRGB write correction from the source register.
1547 *
1548 * However, if the mov is only partial, we can't do this, and if the write
1549 * comes from an instruction other than MOV it is hard to do as well. If
1550 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
1551 shader->u.ps.color0_mov = FALSE;
1552 if (ins.handler_idx == WINED3DSIH_MOV
1553 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
1554 {
1555 /* Used later when the source register is read. */
1556 color0_mov = TRUE;
1557 }
1558 }
1559 /* Also drop the MOV marker if the source register is overwritten prior to the shader
1560 * end
1561 */
1562 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
1563 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
1564 {
1565 shader->u.ps.color0_mov = FALSE;
1566 }
1567 }
1568
1569 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
1570 if (shader_version.major == 1
1571 && (ins.handler_idx == WINED3DSIH_TEX
1572 || ins.handler_idx == WINED3DSIH_TEXBEM
1573 || ins.handler_idx == WINED3DSIH_TEXBEML
1574 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
1575 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
1576 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
1577 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
1578 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
1579 || ins.handler_idx == WINED3DSIH_TEXREG2AR
1580 || ins.handler_idx == WINED3DSIH_TEXREG2GB
1581 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
1582 {
1583 unsigned int reg_idx = ins.dst[i].reg.idx[0].offset;
1584
1585 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1586 {
1587 WARN("Invalid 1.x sampler index %u.\n", reg_idx);
1588 continue;
1589 }
1590
1591 TRACE("Setting fake 2D resource for 1.x pixelshader.\n");
1592 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1593 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_FLOAT;
1594 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1595 wined3d_bitmap_set(reg_maps->resource_map, reg_idx);
1596
1597 /* texbem is only valid with < 1.4 pixel shaders */
1598 if (ins.handler_idx == WINED3DSIH_TEXBEM
1599 || ins.handler_idx == WINED3DSIH_TEXBEML)
1600 {
1601 reg_maps->bumpmat |= 1u << reg_idx;
1602 if (ins.handler_idx == WINED3DSIH_TEXBEML)
1603 {
1604 reg_maps->luminanceparams |= 1u << reg_idx;
1605 }
1606 }
1607 }
1608 else if (ins.handler_idx == WINED3DSIH_BEM)
1609 {
1610 reg_maps->bumpmat |= 1u << ins.dst[i].reg.idx[0].offset;
1611 }
1612 }
1613
1614 if (ins.handler_idx == WINED3DSIH_IMM_ATOMIC_ALLOC || ins.handler_idx == WINED3DSIH_IMM_ATOMIC_CONSUME)
1615 {
1616 unsigned int reg_idx = ins.src[0].reg.idx[0].offset;
1617 if (reg_idx >= MAX_UNORDERED_ACCESS_VIEWS)
1618 {
1619 ERR("Invalid UAV index %u.\n", reg_idx);
1620 break;
1621 }
1622 reg_maps->uav_counter_mask |= (1u << reg_idx);
1623 }
1624 else if ((WINED3DSIH_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_ATOMIC_XOR)
1625 || (WINED3DSIH_IMM_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_IMM_ATOMIC_XOR)
1626 || (ins.handler_idx == WINED3DSIH_BUFINFO && ins.src[0].reg.type == WINED3DSPR_UAV)
1627 || ins.handler_idx == WINED3DSIH_LD_UAV_TYPED
1628 || (ins.handler_idx == WINED3DSIH_LD_RAW && ins.src[1].reg.type == WINED3DSPR_UAV)
1629 || (ins.handler_idx == WINED3DSIH_LD_STRUCTURED && ins.src[2].reg.type == WINED3DSPR_UAV))
1630 {
1631 const struct wined3d_shader_register *reg;
1632
1633 if (ins.handler_idx == WINED3DSIH_LD_UAV_TYPED || ins.handler_idx == WINED3DSIH_LD_RAW)
1634 reg = &ins.src[1].reg;
1635 else if (ins.handler_idx == WINED3DSIH_LD_STRUCTURED)
1636 reg = &ins.src[2].reg;
1637 else if (WINED3DSIH_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_ATOMIC_XOR)
1638 reg = &ins.dst[0].reg;
1639 else if (ins.handler_idx == WINED3DSIH_BUFINFO)
1640 reg = &ins.src[0].reg;
1641 else
1642 reg = &ins.dst[1].reg;
1643
1644 if (reg->type == WINED3DSPR_UAV)
1645 {
1646 if (reg->idx[0].offset >= MAX_UNORDERED_ACCESS_VIEWS)
1647 {
1648 ERR("Invalid UAV index %u.\n", reg->idx[0].offset);
1649 break;
1650 }
1651 reg_maps->uav_read_mask |= (1u << reg->idx[0].offset);
1652 }
1653 }
1654 else if (ins.handler_idx == WINED3DSIH_NRM)
1655 {
1656 reg_maps->usesnrm = 1;
1657 }
1658 else if (ins.handler_idx == WINED3DSIH_DSY
1659 || ins.handler_idx == WINED3DSIH_DSY_COARSE
1660 || ins.handler_idx == WINED3DSIH_DSY_FINE)
1661 {
1662 reg_maps->usesdsy = 1;
1663 }
1664 else if (ins.handler_idx == WINED3DSIH_DSX
1665 || ins.handler_idx == WINED3DSIH_DSX_COARSE
1666 || ins.handler_idx == WINED3DSIH_DSX_FINE)
1667 {
1668 reg_maps->usesdsx = 1;
1669 }
1670 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
1671 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
1672 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
1673 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
1674 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
1675 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
1676 else if (ins.handler_idx == WINED3DSIH_LOOP
1677 || ins.handler_idx == WINED3DSIH_REP)
1678 {
1679 ++cur_loop_depth;
1680 if (cur_loop_depth > max_loop_depth)
1681 max_loop_depth = cur_loop_depth;
1682 }
1683 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
1684 || ins.handler_idx == WINED3DSIH_ENDREP)
1685 {
1686 --cur_loop_depth;
1687 }
1688 else if (ins.handler_idx == WINED3DSIH_GATHER4
1689 || ins.handler_idx == WINED3DSIH_GATHER4_C
1690 || ins.handler_idx == WINED3DSIH_SAMPLE
1691 || ins.handler_idx == WINED3DSIH_SAMPLE_B
1692 || ins.handler_idx == WINED3DSIH_SAMPLE_C
1693 || ins.handler_idx == WINED3DSIH_SAMPLE_C_LZ
1694 || ins.handler_idx == WINED3DSIH_SAMPLE_GRAD
1695 || ins.handler_idx == WINED3DSIH_SAMPLE_LOD)
1696 {
1697 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1698 ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
1699 }
1700 else if (ins.handler_idx == WINED3DSIH_GATHER4_PO
1701 || ins.handler_idx == WINED3DSIH_GATHER4_PO_C)
1702 {
1703 shader_record_sample(reg_maps, ins.src[2].reg.idx[0].offset,
1704 ins.src[3].reg.idx[0].offset, reg_maps->sampler_map.count);
1705 }
1706 else if ((ins.handler_idx == WINED3DSIH_BUFINFO && ins.src[0].reg.type == WINED3DSPR_RESOURCE)
1707 || (ins.handler_idx == WINED3DSIH_SAMPLE_INFO && ins.src[0].reg.type == WINED3DSPR_RESOURCE))
1708 {
1709 shader_record_sample(reg_maps, ins.src[0].reg.idx[0].offset,
1710 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1711 }
1712 else if (ins.handler_idx == WINED3DSIH_LD
1713 || ins.handler_idx == WINED3DSIH_LD2DMS
1714 || (ins.handler_idx == WINED3DSIH_LD_RAW && ins.src[1].reg.type == WINED3DSPR_RESOURCE)
1715 || (ins.handler_idx == WINED3DSIH_RESINFO && ins.src[1].reg.type == WINED3DSPR_RESOURCE))
1716 {
1717 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1718 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1719 }
1720 else if (ins.handler_idx == WINED3DSIH_LD_STRUCTURED
1721 && ins.src[2].reg.type == WINED3DSPR_RESOURCE)
1722 {
1723 shader_record_sample(reg_maps, ins.src[2].reg.idx[0].offset,
1724 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1725 }
1726
1727 if (ins.predicate)
1728 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
1729 shader_version.type, constf_size))
1730 return WINED3DERR_INVALIDCALL;
1731
1732 for (i = 0; i < ins.src_count; ++i)
1733 {
1734 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
1735 struct wined3d_shader_register reg = ins.src[i].reg;
1736
1737 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
1738 shader_version.type, constf_size))
1739 return WINED3DERR_INVALIDCALL;
1740 while (count)
1741 {
1742 ++reg.idx[0].offset;
1743 if (!shader_record_register_usage(shader, reg_maps, &reg,
1744 shader_version.type, constf_size))
1745 return WINED3DERR_INVALIDCALL;
1746 --count;
1747 }
1748
1749 if (color0_mov)
1750 {
1751 if (ins.src[i].reg.type == WINED3DSPR_TEMP
1752 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
1753 {
1754 shader->u.ps.color0_mov = TRUE;
1755 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
1756 }
1757 }
1758 }
1759 }
1760
1761 prev_ins = current_ins;
1762 }
1763 reg_maps->loop_depth = max_loop_depth;
1764
1765 if (phase)
1766 {
1767 phase->end = prev_ins;
1768 phase = NULL;
1769 }
1770
1771 /* PS before 2.0 don't have explicit color outputs. Instead the value of
1772 * R0 is written to the render target. */
1773 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1774 reg_maps->rt_mask |= (1u << 0);
1775
1776 if (input_signature->elements)
1777 {
1778 for (i = 0; i < input_signature->element_count; ++i)
1779 {
1780 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1781 {
1782 if (input_signature->elements[i].register_idx >= ARRAY_SIZE(shader->u.vs.attributes))
1783 {
1784 WARN("Invalid input signature register index %u.\n", input_signature->elements[i].register_idx);
1785 return WINED3DERR_INVALIDCALL;
1786 }
1787 }
1788 else if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1789 {
1790 if (input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1791 reg_maps->vpos = 1;
1792 else if (input_signature->elements[i].sysval_semantic == WINED3D_SV_IS_FRONT_FACE)
1793 reg_maps->usesfacing = 1;
1794 }
1795 reg_maps->input_registers |= 1u << input_signature->elements[i].register_idx;
1796 }
1797 }
1798 else if (!input_signature->elements && reg_maps->input_registers)
1799 {
1800 unsigned int count = wined3d_popcount(reg_maps->input_registers);
1801 struct wined3d_shader_signature_element *e;
1802 unsigned int i;
1803
1804 if (!(input_signature->elements = calloc(count, sizeof(*input_signature->elements))))
1805 return E_OUTOFMEMORY;
1806 input_signature->element_count = count;
1807
1808 e = input_signature->elements;
1809 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1810 {
1811 if (!(reg_maps->input_registers & (1u << i)))
1812 continue;
1813 input_signature_elements[i].register_idx = i;
1814 *e++ = input_signature_elements[i];
1815 }
1816 }
1817
1818 if (output_signature->elements)
1819 {
1820 if (FAILED(hr = shader_scan_output_signature(shader)))
1821 return hr;
1822 }
1823 else if (reg_maps->output_registers)
1824 {
1825 unsigned int count = wined3d_popcount(reg_maps->output_registers);
1826 struct wined3d_shader_signature_element *e;
1827
1828 if (!(output_signature->elements = calloc(count, sizeof(*output_signature->elements))))
1829 return E_OUTOFMEMORY;
1830 output_signature->element_count = count;
1831
1832 e = output_signature->elements;
1833 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1834 {
1835 if (!(reg_maps->output_registers & (1u << i)))
1836 continue;
1837 *e++ = output_signature_elements[i];
1838 }
1839 }
1840
1841 return WINED3D_OK;
1842 }
1843
1844 static void shader_cleanup_reg_maps(struct wined3d_shader_reg_maps *reg_maps)
1845 {
1846 struct wined3d_shader_indexable_temp *reg, *reg_next;
1847
1848 free(reg_maps->constf);
1849 free(reg_maps->sampler_map.entries);
1850
1851 LIST_FOR_EACH_ENTRY_SAFE(reg, reg_next, &reg_maps->indexable_temps, struct wined3d_shader_indexable_temp, entry)
1852 free(reg);
1853 list_init(&reg_maps->indexable_temps);
1854
1855 free(reg_maps->tgsm);
1856 }
1857
1858 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1859 {
1860 DWORD map = 1u << max;
1861 map |= map - 1;
1862 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1863
1864 return wined3d_log2i(map);
1865 }
1866
1867 /* Shared code in order to generate the bulk of the shader string. */
1868 HRESULT shader_generate_code(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
1869 const struct wined3d_shader_reg_maps *reg_maps, void *backend_ctx,
1870 const DWORD *start, const DWORD *end)
1871 {
1872 struct wined3d_device *device = shader->device;
1873 const struct wined3d_shader_frontend *fe = shader->frontend;
1874 void *fe_data = shader->frontend_data;
1875 struct wined3d_shader_version shader_version;
1876 struct wined3d_shader_parser_state state;
1877 struct wined3d_shader_instruction ins;
1878 struct wined3d_shader_tex_mx tex_mx;
1879 struct wined3d_shader_context ctx;
1880 const DWORD *ptr;
1881
1882 /* Initialize current parsing state. */
1883 tex_mx.current_row = 0;
1884 state.current_loop_depth = 0;
1885 state.current_loop_reg = 0;
1886 state.in_subroutine = FALSE;
1887
1888 ctx.shader = shader;
1889 ctx.reg_maps = reg_maps;
1890 ctx.buffer = buffer;
1891 ctx.tex_mx = &tex_mx;
1892 ctx.state = &state;
1893 ctx.backend_data = backend_ctx;
1894 ins.ctx = &ctx;
1895
1896 fe->shader_read_header(fe_data, &ptr, &shader_version);
1897 if (start)
1898 ptr = start;
1899
1900 while (!fe->shader_is_end(fe_data, &ptr) && ptr != end)
1901 {
1902 /* Read opcode. */
1903 fe->shader_read_instruction(fe_data, &ptr, &ins);
1904
1905 /* Unknown opcode and its parameters. */
1906 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1907 {
1908 WARN("Encountered unrecognised or invalid instruction.\n");
1909 return WINED3DERR_INVALIDCALL;
1910 }
1911
1912 if (ins.predicate)
1913 FIXME("Predicates not implemented.\n");
1914
1915 /* Call appropriate function for output target */
1916 device->shader_backend->shader_handle_instruction(&ins);
1917 }
1918
1919 return WINED3D_OK;
1920 }
1921
1922 static void shader_cleanup(struct wined3d_shader *shader)
1923 {
1924 if (shader->reg_maps.shader_version.type == WINED3D_SHADER_TYPE_HULL)
1925 {
1926 free(shader->u.hs.phases.control_point);
1927 free(shader->u.hs.phases.fork);
1928 free(shader->u.hs.phases.join);
1929 }
1930
1931 free(shader->patch_constant_signature.elements);
1932 free(shader->output_signature.elements);
1933 free(shader->input_signature.elements);
1934 shader->device->shader_backend->shader_destroy(shader);
1935 shader_cleanup_reg_maps(&shader->reg_maps);
1936 free(shader->byte_code);
1937 shader_delete_constant_list(&shader->constantsF);
1938 shader_delete_constant_list(&shader->constantsB);
1939 shader_delete_constant_list(&shader->constantsI);
1940 list_remove(&shader->shader_list_entry);
1941
1942 if (shader->frontend && shader->frontend_data)
1943 shader->frontend->shader_free(shader->frontend_data);
1944 }
1945
1946 struct shader_none_priv
1947 {
1948 const struct wined3d_vertex_pipe_ops *vertex_pipe;
1949 const struct wined3d_fragment_pipe_ops *fragment_pipe;
1950 };
1951
1952 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
1953 static void shader_none_precompile(void *shader_priv, struct wined3d_shader *shader) {}
1954 static void shader_none_apply_compute_state(void *shader_priv, struct wined3d_context *context,
1955 const struct wined3d_state *state) {}
1956 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
1957 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
1958 static void shader_none_destroy(struct wined3d_shader *shader) {}
1959 static void shader_none_free_context_data(struct wined3d_context *context) {}
1960 static void shader_none_init_context_state(struct wined3d_context *context) {}
1961
1962 /* Context activation is done by the caller. */
1963 static void shader_none_apply_draw_state(void *shader_priv, struct wined3d_context *context,
1964 const struct wined3d_state *state)
1965 {
1966 struct shader_none_priv *priv = shader_priv;
1967
1968 priv->vertex_pipe->vp_apply_draw_state(context, state);
1969 priv->fragment_pipe->fp_apply_draw_state(context, state);
1970 }
1971
1972 /* Context activation is done by the caller. */
1973 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
1974 {
1975 struct shader_none_priv *priv = shader_priv;
1976
1977 priv->vertex_pipe->vp_disable(context);
1978 priv->fragment_pipe->fp_disable(context);
1979
1980 context->shader_update_mask = (1u << WINED3D_SHADER_TYPE_PIXEL)
1981 | (1u << WINED3D_SHADER_TYPE_VERTEX)
1982 | (1u << WINED3D_SHADER_TYPE_GEOMETRY)
1983 | (1u << WINED3D_SHADER_TYPE_HULL)
1984 | (1u << WINED3D_SHADER_TYPE_DOMAIN)
1985 | (1u << WINED3D_SHADER_TYPE_COMPUTE);
1986 }
1987
1988 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
1989 const struct wined3d_fragment_pipe_ops *fragment_pipe)
1990 {
1991 void *vertex_priv, *fragment_priv;
1992 struct shader_none_priv *priv;
1993
1994 if (!(priv = malloc(sizeof(*priv))))
1995 return E_OUTOFMEMORY;
1996
1997 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
1998 {
1999 ERR("Failed to initialize vertex pipe.\n");
2000 free(priv);
2001 return E_FAIL;
2002 }
2003
2004 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
2005 {
2006 ERR("Failed to initialize fragment pipe.\n");
2007 vertex_pipe->vp_free(device, NULL);
2008 free(priv);
2009 return E_FAIL;
2010 }
2011
2012 priv->vertex_pipe = vertex_pipe;
2013 priv->fragment_pipe = fragment_pipe;
2014
2015 device->vertex_priv = vertex_priv;
2016 device->fragment_priv = fragment_priv;
2017 device->shader_priv = priv;
2018
2019 return WINED3D_OK;
2020 }
2021
2022 static void shader_none_free(struct wined3d_device *device, struct wined3d_context *context)
2023 {
2024 struct shader_none_priv *priv = device->shader_priv;
2025
2026 priv->fragment_pipe->free_private(device, context);
2027 priv->vertex_pipe->vp_free(device, context);
2028 free(priv);
2029 }
2030
2031 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
2032 {
2033 return TRUE;
2034 }
2035
2036 static void shader_none_get_caps(const struct wined3d_adapter *adapter, struct shader_caps *caps)
2037 {
2038 /* Set the shader caps to 0 for the none shader backend */
2039 memset(caps, 0, sizeof(*caps));
2040 }
2041
2042 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
2043 {
2044 /* We "support" every possible fixup, since we don't support any shader
2045 * model, and will never have to actually sample a texture. */
2046 return TRUE;
2047 }
2048
2049 static uint64_t shader_none_shader_compile(struct wined3d_context *context, const struct wined3d_shader_desc *shader_desc,
2050 enum wined3d_shader_type shader_type)
2051 {
2052 return 0;
2053 }
2054
2055 const struct wined3d_shader_backend_ops none_shader_backend =
2056 {
2057 shader_none_handle_instruction,
2058 shader_none_precompile,
2059 shader_none_apply_draw_state,
2060 shader_none_apply_compute_state,
2061 shader_none_disable,
2062 shader_none_update_float_vertex_constants,
2063 shader_none_update_float_pixel_constants,
2064 shader_none_destroy,
2065 shader_none_alloc,
2066 shader_none_free,
2067 shader_none_allocate_context_data,
2068 shader_none_free_context_data,
2069 shader_none_init_context_state,
2070 shader_none_get_caps,
2071 shader_none_color_fixup_supported,
2072 shader_none_shader_compile,
2073 };
2074
2075 static unsigned int shader_max_version_from_feature_level(enum wined3d_feature_level level)
2076 {
2077 switch (level)
2078 {
2079 case WINED3D_FEATURE_LEVEL_11_1:
2080 case WINED3D_FEATURE_LEVEL_11:
2081 return 5;
2082 case WINED3D_FEATURE_LEVEL_10_1:
2083 case WINED3D_FEATURE_LEVEL_10:
2084 return 4;
2085 case WINED3D_FEATURE_LEVEL_9_3:
2086 return 3;
2087 case WINED3D_FEATURE_LEVEL_9_2:
2088 case WINED3D_FEATURE_LEVEL_9_1:
2089 return 2;
2090 default:
2091 return 1;
2092 }
2093 }
2094
2095 static HRESULT shader_set_function(struct wined3d_shader *shader, struct wined3d_device *device,
2096 enum wined3d_shader_type type, unsigned int float_const_count)
2097 {
2098 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
2099 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2100 const struct wined3d_shader_version *version = &reg_maps->shader_version;
2101 const struct wined3d_shader_frontend *fe;
2102 unsigned int backend_version;
2103 HRESULT hr;
2104
2105 TRACE("shader %p, device %p, type %s, float_const_count %u.\n",
2106 shader, device, debug_shader_type(type), float_const_count);
2107
2108 fe = shader->frontend;
2109 if (!(shader->frontend_data = fe->shader_init(shader->function,
2110 shader->functionLength, &shader->output_signature)))
2111 {
2112 FIXME("Failed to initialize frontend.\n");
2113 return WINED3DERR_INVALIDCALL;
2114 }
2115
2116 if (FAILED(hr = shader_get_registers_used(shader, float_const_count)))
2117 return hr;
2118
2119 if (version->type != type)
2120 {
2121 WARN("Wrong shader type %s.\n", debug_shader_type(reg_maps->shader_version.type));
2122 return WINED3DERR_INVALIDCALL;
2123 }
2124 if (version->major > shader_max_version_from_feature_level(device->cs->c.state->feature_level))
2125 {
2126 WARN("Shader version %u not supported by this device.\n", version->major);
2127 return WINED3DERR_INVALIDCALL;
2128 }
2129 switch (type)
2130 {
2131 case WINED3D_SHADER_TYPE_VERTEX:
2132 backend_version = d3d_info->limits.vs_version;
2133 break;
2134 case WINED3D_SHADER_TYPE_HULL:
2135 backend_version = d3d_info->limits.hs_version;
2136 break;
2137 case WINED3D_SHADER_TYPE_DOMAIN:
2138 backend_version = d3d_info->limits.ds_version;
2139 break;
2140 case WINED3D_SHADER_TYPE_GEOMETRY:
2141 backend_version = d3d_info->limits.gs_version;
2142 break;
2143 case WINED3D_SHADER_TYPE_PIXEL:
2144 backend_version = d3d_info->limits.ps_version;
2145 break;
2146 case WINED3D_SHADER_TYPE_COMPUTE:
2147 backend_version = d3d_info->limits.cs_version;
2148 break;
2149 default:
2150 FIXME("No backend version-checking for this shader type.\n");
2151 backend_version = 0;
2152 }
2153 if (version->major > backend_version)
2154 {
2155 WARN("Shader version %u.%u not supported by the current shader backend.\n",
2156 version->major, version->minor);
2157 return WINED3DERR_INVALIDCALL;
2158 }
2159
2160 shader->load_local_constsF = shader->lconst_inf_or_nan;
2161
2162 return WINED3D_OK;
2163 }
2164
2165 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
2166 {
2167 unsigned int refcount = InterlockedIncrement(&shader->ref);
2168
2169 TRACE("%p increasing refcount to %u.\n", shader, refcount);
2170
2171 return refcount;
2172 }
2173
2174 static void wined3d_shader_init_object(void *object)
2175 {
2176 struct wined3d_shader *shader = object;
2177 struct wined3d_device *device = shader->device;
2178
2179 TRACE("shader %p.\n", shader);
2180
2181 list_add_head(&device->shaders, &shader->shader_list_entry);
2182
2183 device->shader_backend->shader_precompile(device->shader_priv, shader);
2184 }
2185
2186 static void wined3d_shader_destroy_object(void *object)
2187 {
2188 TRACE("object %p.\n", object);
2189
2190 shader_cleanup(object);
2191 free(object);
2192 }
2193
2194 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
2195 {
2196 unsigned int refcount = InterlockedDecrement(&shader->ref);
2197
2198 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
2199
2200 if (!refcount)
2201 {
2202 wined3d_mutex_lock();
2203 shader->parent_ops->wined3d_object_destroyed(shader->parent);
2204 wined3d_cs_destroy_object(shader->device->cs, wined3d_shader_destroy_object, shader);
2205 wined3d_mutex_unlock();
2206 }
2207
2208 return refcount;
2209 }
2210
2211 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
2212 {
2213 TRACE("shader %p.\n", shader);
2214
2215 return shader->parent;
2216 }
2217
2218 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
2219 void *byte_code, UINT *byte_code_size)
2220 {
2221 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
2222
2223 if (!byte_code)
2224 {
2225 *byte_code_size = shader->byte_code_size;
2226 return WINED3D_OK;
2227 }
2228
2229 if (*byte_code_size < shader->byte_code_size)
2230 {
2231 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
2232 * than the required size we should write the required size and
2233 * return D3DERR_MOREDATA. That's not actually true. */
2234 return WINED3DERR_INVALIDCALL;
2235 }
2236
2237 memcpy(byte_code, shader->byte_code, shader->byte_code_size);
2238
2239 return WINED3D_OK;
2240 }
2241
2242 /* Set local constants for d3d8 shaders. */
2243 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
2244 UINT start_idx, const float *src_data, UINT count)
2245 {
2246 UINT end_idx = start_idx + count;
2247 UINT i;
2248
2249 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
2250
2251 if (end_idx > shader->limits->constant_float)
2252 {
2253 WARN("end_idx %u > float constants limit %u.\n",
2254 end_idx, shader->limits->constant_float);
2255 end_idx = shader->limits->constant_float;
2256 }
2257
2258 for (i = start_idx; i < end_idx; ++i)
2259 {
2260 struct wined3d_shader_lconst *lconst;
2261 float *value;
2262
2263 if (!(lconst = malloc(sizeof(*lconst))))
2264 return E_OUTOFMEMORY;
2265
2266 lconst->idx = i;
2267 value = (float *)lconst->value;
2268 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
2269 list_add_head(&shader->constantsF, &lconst->entry);
2270
2271 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
2272 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
2273 {
2274 shader->lconst_inf_or_nan = TRUE;
2275 }
2276 }
2277
2278 return WINED3D_OK;
2279 }
2280
2281 static void init_interpolation_compile_args(uint32_t *interpolation_args,
2282 const struct wined3d_shader *pixel_shader, const struct wined3d_d3d_info *d3d_info)
2283 {
2284 if (!d3d_info->shader_output_interpolation || !pixel_shader
2285 || pixel_shader->reg_maps.shader_version.major < 4)
2286 {
2287 memset(interpolation_args, 0, sizeof(pixel_shader->u.ps.interpolation_mode));
2288 return;
2289 }
2290
2291 memcpy(interpolation_args, pixel_shader->u.ps.interpolation_mode,
2292 sizeof(pixel_shader->u.ps.interpolation_mode));
2293 }
2294
2295 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2296 struct vs_compile_args *args, const struct wined3d_context *context)
2297 {
2298 const struct wined3d_shader *geometry_shader = state->shader[WINED3D_SHADER_TYPE_GEOMETRY];
2299 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
2300 const struct wined3d_shader *hull_shader = state->shader[WINED3D_SHADER_TYPE_HULL];
2301 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
2302 WORD swizzle_map = context->stream_info.swizzle_map;
2303
2304 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
2305 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
2306 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
2307 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
2308 args->point_size = state->primitive_type == WINED3D_PT_POINTLIST;
2309 args->per_vertex_point_size = shader->reg_maps.point_size;
2310 args->next_shader_type = hull_shader ? WINED3D_SHADER_TYPE_HULL
2311 : geometry_shader ? WINED3D_SHADER_TYPE_GEOMETRY : WINED3D_SHADER_TYPE_PIXEL;
2312 if (shader->reg_maps.shader_version.major >= 4)
2313 args->next_shader_input_count = hull_shader ? hull_shader->limits->packed_input
2314 : geometry_shader ? geometry_shader->limits->packed_input
2315 : pixel_shader ? pixel_shader->limits->packed_input : 0;
2316 else
2317 args->next_shader_input_count = 0;
2318 args->swizzle_map = swizzle_map;
2319 if (d3d_info->emulated_flatshading)
2320 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
2321 else
2322 args->flatshading = 0;
2323
2324 init_interpolation_compile_args(args->interpolation_mode,
2325 args->next_shader_type == WINED3D_SHADER_TYPE_PIXEL ? pixel_shader : NULL, d3d_info);
2326 }
2327
2328 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
2329 {
2330 if (usage_idx1 != usage_idx2)
2331 return FALSE;
2332 if (usage1 == usage2)
2333 return TRUE;
2334 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
2335 return TRUE;
2336 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
2337 return TRUE;
2338
2339 return FALSE;
2340 }
2341
2342 bool vshader_get_input(const struct wined3d_shader *shader,
2343 uint8_t usage_req, uint8_t usage_idx_req, unsigned int *regnum)
2344 {
2345 uint32_t map = shader->reg_maps.input_registers & 0xffff;
2346 unsigned int i;
2347
2348 while (map)
2349 {
2350 i = wined3d_bit_scan(&map);
2351 if (match_usage(shader->u.vs.attributes[i].usage,
2352 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
2353 {
2354 *regnum = i;
2355 return true;
2356 }
2357 }
2358
2359 return false;
2360 }
2361
2362 static void shader_trace(const void *code, size_t size, enum vkd3d_shader_source_type source_type)
2363 {
2364 struct vkd3d_shader_compile_info info;
2365 struct vkd3d_shader_code d3d_asm;
2366 const char *ptr, *end, *line;
2367 char *messages;
2368 int ret;
2369
2370 static const struct vkd3d_shader_compile_option compile_options[] =
2371 {
2372 {VKD3D_SHADER_COMPILE_OPTION_API_VERSION, VKD3D_SHADER_API_VERSION_1_6},
2373 };
2374
2375 info.type = VKD3D_SHADER_STRUCTURE_TYPE_COMPILE_INFO;
2376 info.next = NULL;
2377 info.source.code = code;
2378 info.source.size = size;
2379 info.source_type = source_type;
2380 info.target_type = VKD3D_SHADER_TARGET_D3D_ASM;
2381 info.options = compile_options;
2382 info.option_count = ARRAY_SIZE(compile_options);
2383 info.log_level = VKD3D_SHADER_LOG_WARNING;
2384 info.source_name = NULL;
2385
2386 ret = vkd3d_shader_compile(&info, &d3d_asm, &messages);
2387 if (messages && *messages && FIXME_ON(d3d_shader))
2388 {
2389 FIXME("Shader log:\n");
2390 ptr = messages;
2391 end = ptr + strlen(ptr);
2392 while ((line = wined3d_get_line(&ptr, end)))
2393 {
2394 FIXME(" %.*s", (int)(ptr - line), line);
2395 }
2396 FIXME("\n");
2397 }
2398 vkd3d_shader_free_messages(messages);
2399
2400 if (ret < 0)
2401 {
2402 ERR("Failed to disassemble, ret %d.\n", ret);
2403 return;
2404 }
2405
2406 ptr = d3d_asm.code;
2407 end = ptr + d3d_asm.size;
2408 while ((line = wined3d_get_line(&ptr, end)))
2409 {
2410 TRACE(" %.*s", (int)(ptr - line), line);
2411 }
2412 TRACE("\n");
2413
2414 vkd3d_shader_free_shader_code(&d3d_asm);
2415 }
2416
2417 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2418 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2419 {
2420 HRESULT hr;
2421
2422 TRACE("byte_code %p, byte_code_size %#lx.\n", desc->byte_code, (long)desc->byte_code_size);
2423
2424 if (!desc->byte_code)
2425 return WINED3DERR_INVALIDCALL;
2426
2427 shader->ref = 1;
2428 shader->device = device;
2429 shader->parent = parent;
2430 shader->parent_ops = parent_ops;
2431
2432 list_init(&shader->linked_programs);
2433 list_init(&shader->constantsF);
2434 list_init(&shader->constantsB);
2435 list_init(&shader->constantsI);
2436 shader->lconst_inf_or_nan = FALSE;
2437 list_init(&shader->reg_maps.indexable_temps);
2438 list_init(&shader->shader_list_entry);
2439
2440 if (desc->byte_code_size == ~(size_t)0)
2441 {
2442 struct wined3d_shader_version shader_version;
2443 const struct wined3d_shader_frontend *fe;
2444 struct wined3d_shader_instruction ins;
2445 const DWORD *ptr;
2446 void *fe_data;
2447
2448 shader->source_type = VKD3D_SHADER_SOURCE_D3D_BYTECODE;
2449 if (!(shader->frontend = shader_select_frontend(shader->source_type)))
2450 {
2451 FIXME("Unable to find frontend for shader.\n");
2452 hr = WINED3DERR_INVALIDCALL;
2453 goto fail;
2454 }
2455
2456 fe = shader->frontend;
2457 if (!(fe_data = fe->shader_init(desc->byte_code, desc->byte_code_size, &shader->output_signature)))
2458 {
2459 WARN("Failed to initialise frontend data.\n");
2460 hr = WINED3DERR_INVALIDCALL;
2461 goto fail;
2462 }
2463
2464 fe->shader_read_header(fe_data, &ptr, &shader_version);
2465 while (!fe->shader_is_end(fe_data, &ptr))
2466 fe->shader_read_instruction(fe_data, &ptr, &ins);
2467
2468 fe->shader_free(fe_data);
2469
2470 shader->byte_code_size = (ptr - desc->byte_code) * sizeof(*ptr);
2471
2472 if (!(shader->byte_code = malloc(shader->byte_code_size)))
2473 {
2474 hr = E_OUTOFMEMORY;
2475 goto fail;
2476 }
2477 memcpy(shader->byte_code, desc->byte_code, shader->byte_code_size);
2478
2479 shader->function = shader->byte_code;
2480 shader->functionLength = shader->byte_code_size;
2481 }
2482 else
2483 {
2484 unsigned int max_version;
2485
2486 if (!(shader->byte_code = malloc(desc->byte_code_size)))
2487 {
2488 hr = E_OUTOFMEMORY;
2489 goto fail;
2490 }
2491 memcpy(shader->byte_code, desc->byte_code, desc->byte_code_size);
2492 shader->byte_code_size = desc->byte_code_size;
2493
2494 max_version = shader_max_version_from_feature_level(device->cs->c.state->feature_level);
2495 if (FAILED(hr = wined3d_shader_extract_from_dxbc(shader, max_version, &shader->source_type)))
2496 goto fail;
2497
2498 if (!(shader->frontend = shader_select_frontend(shader->source_type)))
2499 {
2500 FIXME("Unable to find frontend for shader.\n");
2501 hr = WINED3DERR_INVALIDCALL;
2502 goto fail;
2503 }
2504 }
2505
2506 if (TRACE_ON(d3d_shader))
2507 {
2508 if (shader->source_type == VKD3D_SHADER_SOURCE_D3D_BYTECODE)
2509 shader_trace(shader->function, shader->functionLength, shader->source_type);
2510 else
2511 shader_trace(shader->byte_code, shader->byte_code_size, shader->source_type);
2512 }
2513
2514
2515 return WINED3D_OK;
2516
2517 fail:
2518 shader_cleanup(shader);
2519 return hr;
2520 }
2521
2522 static HRESULT vertex_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2523 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
2524 {
2525 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
2526 unsigned int i;
2527 HRESULT hr;
2528
2529 if (FAILED(hr = shader_init(shader, device, desc, parent, parent_ops)))
2530 return hr;
2531
2532 if (FAILED(hr = shader_set_function(shader, device,
2533 WINED3D_SHADER_TYPE_VERTEX, device->adapter->d3d_info.limits.vs_uniform_count)))
2534 {
2535 shader_cleanup(shader);
2536 return hr;
2537 }
2538
2539 for (i = 0; i < shader->input_signature.element_count; ++i)
2540 {
2541 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
2542
2543 if (!(reg_maps->input_registers & (1u << input->register_idx)) || !input->semantic_name)
2544 continue;
2545
2546 shader->u.vs.attributes[input->register_idx].usage =
2547 shader_usage_from_semantic_name(input->semantic_name);
2548 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
2549 }
2550
2551 if (reg_maps->usesrelconstF && !list_empty(&shader->constantsF))
2552 shader->load_local_constsF = TRUE;
2553
2554 return WINED3D_OK;
2555 }
2556
2557 static struct wined3d_shader_signature_element *shader_find_signature_element(const struct wined3d_shader_signature *s,
2558 unsigned int stream_idx, const char *semantic_name, unsigned int semantic_idx)
2559 {
2560 struct wined3d_shader_signature_element *e = s->elements;
2561 unsigned int i;
2562
2563 for (i = 0; i < s->element_count; ++i)
2564 {
2565 if (e[i].stream_idx == stream_idx
2566 && !stricmp(e[i].semantic_name, semantic_name)
2567 && e[i].semantic_idx == semantic_idx)
2568 return &e[i];
2569 }
2570
2571 return NULL;
2572 }
2573
2574 BOOL shader_get_stream_output_register_info(const struct wined3d_shader *shader,
2575 const struct wined3d_stream_output_element *so_element, unsigned int *register_idx, unsigned int *component_idx)
2576 {
2577 const struct wined3d_shader_signature_element *output;
2578 unsigned int idx;
2579
2580 if (!(output = shader_find_signature_element(&shader->output_signature,
2581 so_element->stream_idx, so_element->semantic_name, so_element->semantic_idx)))
2582 return FALSE;
2583
2584 for (idx = 0; idx < 4; ++idx)
2585 {
2586 if (output->mask & (1u << idx))
2587 break;
2588 }
2589 idx += so_element->component_idx;
2590
2591 *register_idx = output->register_idx;
2592 *component_idx = idx;
2593 return TRUE;
2594 }
2595
2596 static HRESULT geometry_shader_init_so_desc(struct wined3d_geometry_shader *gs, struct wined3d_device *device,
2597 const struct wined3d_stream_output_desc *so_desc)
2598 {
2599 struct wined3d_so_desc_entry *s;
2600 struct wine_rb_entry *entry;
2601 unsigned int i;
2602 size_t size;
2603 char *name;
2604
2605 if ((entry = wine_rb_get(&device->so_descs, so_desc)))
2606 {
2607 gs->so_desc = &WINE_RB_ENTRY_VALUE(entry, struct wined3d_so_desc_entry, entry)->desc;
2608 return WINED3D_OK;
2609 }
2610
2611 size = FIELD_OFFSET(struct wined3d_so_desc_entry, elements[so_desc->element_count]);
2612 for (i = 0; i < so_desc->element_count; ++i)
2613 {
2614 const char *n = so_desc->elements[i].semantic_name;
2615
2616 if (n)
2617 size += strlen(n) + 1;
2618 }
2619 if (!(s = malloc(size)))
2620 return E_OUTOFMEMORY;
2621
2622 s->desc = *so_desc;
2623
2624 memcpy(s->elements, so_desc->elements, so_desc->element_count * sizeof(*s->elements));
2625 s->desc.elements = s->elements;
2626
2627 name = (char *)&s->elements[s->desc.element_count];
2628 for (i = 0; i < so_desc->element_count; ++i)
2629 {
2630 struct wined3d_stream_output_element *e = &s->elements[i];
2631
2632 if (!e->semantic_name)
2633 continue;
2634
2635 size = strlen(e->semantic_name) + 1;
2636 memcpy(name, e->semantic_name, size);
2637 e->semantic_name = name;
2638 name += size;
2639 }
2640
2641 if (wine_rb_put(&device->so_descs, &s->desc, &s->entry) == -1)
2642 {
2643 free(s);
2644 return E_FAIL;
2645 }
2646 gs->so_desc = &s->desc;
2647
2648 return WINED3D_OK;
2649 }
2650
2651 static HRESULT geometry_shader_init_stream_output(struct wined3d_shader *shader,
2652 const struct wined3d_stream_output_desc *so_desc)
2653 {
2654 const struct wined3d_shader_frontend *fe = shader->frontend;
2655 const struct wined3d_shader_signature_element *output;
2656 unsigned int i, component_idx, register_idx, mask;
2657 struct wined3d_shader_version shader_version;
2658 const DWORD *ptr;
2659 void *fe_data;
2660 HRESULT hr;
2661
2662 if (!so_desc)
2663 return WINED3D_OK;
2664
2665 if (!(fe_data = fe->shader_init(shader->function, shader->functionLength, &shader->output_signature)))
2666 {
2667 WARN("Failed to initialise frontend data.\n");
2668 return WINED3DERR_INVALIDCALL;
2669 }
2670 fe->shader_read_header(fe_data, &ptr, &shader_version);
2671 fe->shader_free(fe_data);
2672
2673 switch (shader_version.type)
2674 {
2675 case WINED3D_SHADER_TYPE_VERTEX:
2676 case WINED3D_SHADER_TYPE_DOMAIN:
2677 shader->function = NULL;
2678 shader->functionLength = 0;
2679 break;
2680 case WINED3D_SHADER_TYPE_GEOMETRY:
2681 break;
2682 default:
2683 WARN("Wrong shader type %s.\n", debug_shader_type(shader_version.type));
2684 return E_INVALIDARG;
2685 }
2686
2687 if (!shader->function)
2688 {
2689 shader->reg_maps.shader_version = shader_version;
2690 shader->reg_maps.shader_version.type = WINED3D_SHADER_TYPE_GEOMETRY;
2691 shader_set_limits(shader);
2692 if (FAILED(hr = shader_scan_output_signature(shader)))
2693 return hr;
2694 }
2695
2696 for (i = 0; i < so_desc->element_count; ++i)
2697 {
2698 const struct wined3d_stream_output_element *e = &so_desc->elements[i];
2699
2700 if (!e->semantic_name)
2701 continue;
2702 if (!(output = shader_find_signature_element(&shader->output_signature,
2703 e->stream_idx, e->semantic_name, e->semantic_idx))
2704 || !shader_get_stream_output_register_info(shader, e, &register_idx, &component_idx))
2705 {
2706 WARN("Failed to find output signature element for stream output entry.\n");
2707 return E_INVALIDARG;
2708 }
2709
2710 mask = wined3d_mask_from_size(e->component_count) << component_idx;
2711 if ((output->mask & 0xff & mask) != mask)
2712 {
2713 WARN("Invalid component range %u-%u (mask %#x), output mask %#x.\n",
2714 component_idx, e->component_count, mask, output->mask & 0xff);
2715 return E_INVALIDARG;
2716 }
2717 }
2718
2719 if (FAILED(hr = geometry_shader_init_so_desc(&shader->u.gs, shader->device, so_desc)))
2720 {
2721 WARN("Failed to initialise stream output description, hr %#lx.\n", hr);
2722 return hr;
2723 }
2724
2725 return WINED3D_OK;
2726 }
2727
2728 static HRESULT geometry_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
2729 const struct wined3d_shader_desc *desc, const struct wined3d_stream_output_desc *so_desc,
2730 void *parent, const struct wined3d_parent_ops *parent_ops)
2731 {
2732 HRESULT hr;
2733
2734 if (FAILED(hr = shader_init(shader, device, desc, parent, parent_ops)))
2735 return hr;
2736
2737 if (FAILED(hr = geometry_shader_init_stream_output(shader, so_desc)))
2738 goto fail;
2739
2740 if (shader->function
2741 && FAILED(hr = shader_set_function(shader, device, WINED3D_SHADER_TYPE_GEOMETRY, 0)))
2742 goto fail;
2743
2744 return WINED3D_OK;
2745
2746 fail:
2747 shader_cleanup(shader);
2748 return hr;
2749 }
2750
2751 void find_ds_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2752 struct ds_compile_args *args, const struct wined3d_context *context)
2753 {
2754 const struct wined3d_shader *geometry_shader = state->shader[WINED3D_SHADER_TYPE_GEOMETRY];
2755 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
2756 const struct wined3d_shader *hull_shader = state->shader[WINED3D_SHADER_TYPE_HULL];
2757
2758 args->tessellator_output_primitive = hull_shader->u.hs.tessellator_output_primitive;
2759 args->tessellator_partitioning = hull_shader->u.hs.tessellator_partitioning;
2760
2761 args->output_count = geometry_shader ? geometry_shader->limits->packed_input
2762 : pixel_shader ? pixel_shader->limits->packed_input : shader->limits->packed_output;
2763 args->next_shader_type = geometry_shader ? WINED3D_SHADER_TYPE_GEOMETRY : WINED3D_SHADER_TYPE_PIXEL;
2764
2765 args->render_offscreen = context->render_offscreen;
2766
2767 init_interpolation_compile_args(args->interpolation_mode,
2768 args->next_shader_type == WINED3D_SHADER_TYPE_PIXEL ? pixel_shader : NULL, context->d3d_info);
2769
2770 args->padding = 0;
2771 }
2772
2773 void find_gs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2774 struct gs_compile_args *args, const struct wined3d_context *context)
2775 {
2776 const struct wined3d_shader *pixel_shader = state->shader[WINED3D_SHADER_TYPE_PIXEL];
2777
2778 args->output_count = pixel_shader ? pixel_shader->limits->packed_input : shader->limits->packed_output;
2779
2780 if (!(args->primitive_type = shader->u.gs.input_type))
2781 args->primitive_type = state->primitive_type;
2782
2783 init_interpolation_compile_args(args->interpolation_mode, pixel_shader, context->d3d_info);
2784 }
2785
2786 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
2787 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_context *context)
2788 {
2789 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
2790 struct wined3d_shader_resource_view *view;
2791 struct wined3d_texture *texture;
2792 unsigned int i;
2793
2794 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
2795 if (!d3d_info->srgb_write_control && needs_srgb_write(d3d_info, state, &state->fb))
2796 {
2797 static unsigned int warned = 0;
2798
2799 args->srgb_correction = 1;
2800 if (state->blend_state && state->blend_state->desc.rt[0].enable && !warned++)
2801 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
2802 "support, expect rendering artifacts.\n");
2803 }
2804
2805 if (shader->reg_maps.shader_version.major == 1
2806 && shader->reg_maps.shader_version.minor <= 3)
2807 {
2808 for (i = 0; i < shader->limits->sampler; ++i)
2809 {
2810 uint32_t flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
2811
2812 if (flags & WINED3D_TTFF_PROJECTED)
2813 {
2814 uint32_t tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
2815
2816 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
2817 {
2818 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
2819 unsigned int j;
2820 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
2821 uint32_t max_valid = WINED3D_TTFF_COUNT4;
2822
2823 for (j = 0; j < state->vertex_declaration->element_count; ++j)
2824 {
2825 struct wined3d_vertex_declaration_element *element =
2826 &state->vertex_declaration->elements[j];
2827
2828 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
2829 && element->usage_idx == index)
2830 {
2831 max_valid = element->format->component_count;
2832 break;
2833 }
2834 }
2835 if (!tex_transform || tex_transform > max_valid)
2836 {
2837 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
2838 tex_transform, max_valid);
2839 tex_transform = max_valid;
2840 }
2841 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
2842 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
2843 && tex_transform > WINED3D_TTFF_COUNT2)
2844 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
2845 && tex_transform > WINED3D_TTFF_COUNT3))
2846 tex_transform |= WINED3D_PSARGS_PROJECTED;
2847 else
2848 {
2849 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
2850 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
2851 i, tex_transform, resource_type);
2852 }
2853 }
2854 else
2855 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
2856
2857 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
2858 }
2859 }
2860 }
2861 if (shader->reg_maps.shader_version.major == 1
2862 && shader->reg_maps.shader_version.minor <= 4)
2863 {
2864 for (i = 0; i < shader->limits->sampler; ++i)
2865 {
2866 enum wined3d_shader_tex_types type = WINED3D_SHADER_TEX_2D;
2867
2868 if (!shader->reg_maps.resource_info[i].type)
2869 continue;
2870
2871 /* Treat unbound textures as 2D. The dummy texture will provide
2872 * the proper sample value. The tex_types bitmap defaults to
2873 * 2D because of the memset. */
2874 if (!(view = state->shader_resource_view[WINED3D_SHADER_TYPE_PIXEL][i]))
2875 continue;
2876 texture = texture_from_resource(view->resource);
2877
2878 switch (texture->resource.type)
2879 {
2880 case WINED3D_RTYPE_NONE:
2881 case WINED3D_RTYPE_BUFFER:
2882 case WINED3D_RTYPE_TEXTURE_1D:
2883 assert(0);
2884
2885 case WINED3D_RTYPE_TEXTURE_2D:
2886 if (texture->resource.usage & WINED3DUSAGE_LEGACY_CUBEMAP)
2887 type = WINED3D_SHADER_TEX_CUBE;
2888 break;
2889
2890 case WINED3D_RTYPE_TEXTURE_3D:
2891 type = WINED3D_SHADER_TEX_3D;
2892 break;
2893 }
2894
2895 args->tex_types |= (type << (i * WINED3D_PSARGS_TEXTYPE_SHIFT));
2896 }
2897 }
2898 else if (shader->reg_maps.shader_version.major <= 3)
2899 {
2900 for (i = 0; i < shader->limits->sampler; ++i)
2901 {
2902 enum wined3d_shader_resource_type resource_type;
2903 enum wined3d_shader_tex_types tex_type;
2904
2905 if (!(resource_type = shader->reg_maps.resource_info[i].type))
2906 continue;
2907
2908 switch (resource_type)
2909 {
2910 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
2911 tex_type = WINED3D_SHADER_TEX_3D;
2912 break;
2913 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
2914 tex_type = WINED3D_SHADER_TEX_CUBE;
2915 break;
2916 default:
2917 tex_type = WINED3D_SHADER_TEX_2D;
2918 break;
2919 }
2920
2921 if ((view = state->shader_resource_view[WINED3D_SHADER_TYPE_PIXEL][i]))
2922 {
2923 texture = texture_from_resource(view->resource);
2924 /* Star Wars: The Old Republic uses mismatched samplers for rendering water. */
2925 if (texture->resource.type == WINED3D_RTYPE_TEXTURE_2D
2926 && resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
2927 && !(texture->resource.usage & WINED3DUSAGE_LEGACY_CUBEMAP))
2928 tex_type = WINED3D_SHADER_TEX_2D;
2929 else if (texture->resource.type == WINED3D_RTYPE_TEXTURE_3D
2930 && resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D)
2931 tex_type = WINED3D_SHADER_TEX_3D;
2932 }
2933 args->tex_types |= tex_type << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
2934 }
2935 }
2936
2937 if (shader->reg_maps.shader_version.major >= 4)
2938 {
2939 /* In SM4+ we use dcl_sampler in order to determine if we should use shadow sampler. */
2940 args->shadow = 0;
2941 for (i = 0 ; i < WINED3D_MAX_FRAGMENT_SAMPLERS; ++i)
2942 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2943 args->np2_fixup = 0;
2944 }
2945 else
2946 {
2947 for (i = 0; i < WINED3D_MAX_FRAGMENT_SAMPLERS; ++i)
2948 {
2949 if (!shader->reg_maps.resource_info[i].type)
2950 continue;
2951
2952 if (!(view = state->shader_resource_view[WINED3D_SHADER_TYPE_PIXEL][i]))
2953 {
2954 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2955 continue;
2956 }
2957 texture = texture_from_resource(view->resource);
2958
2959 if (can_use_texture_swizzle(d3d_info, texture->resource.format))
2960 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
2961 else
2962 args->color_fixup[i] = texture->resource.format->color_fixup;
2963
2964 if (texture->resource.format_caps & WINED3D_FORMAT_CAP_SHADOW)
2965 args->shadow |= 1u << i;
2966
2967 /* Flag samplers that need NP2 texcoord fixup. */
2968 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
2969 args->np2_fixup |= (1u << i);
2970 }
2971 }
2972
2973 if (shader->reg_maps.shader_version.major >= 3)
2974 {
2975 if (position_transformed)
2976 args->vp_mode = WINED3D_VP_MODE_NONE;
2977 else if (use_vs(state))
2978 args->vp_mode = WINED3D_VP_MODE_SHADER;
2979 else
2980 args->vp_mode = WINED3D_VP_MODE_FF;
2981 args->fog = WINED3D_FFP_PS_FOG_OFF;
2982 }
2983 else
2984 {
2985 args->vp_mode = WINED3D_VP_MODE_SHADER;
2986 if (state->render_states[WINED3D_RS_FOGENABLE])
2987 {
2988 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
2989 {
2990 case WINED3D_FOG_NONE:
2991 if (position_transformed || use_vs(state))
2992 {
2993 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
2994 break;
2995 }
2996
2997 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
2998 {
2999 case WINED3D_FOG_NONE: /* Fall through. */
3000 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3001 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3002 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3003 }
3004 break;
3005
3006 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3007 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3008 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3009 }
3010 }
3011 else
3012 {
3013 args->fog = WINED3D_FFP_PS_FOG_OFF;
3014 }
3015 }
3016
3017 if (!d3d_info->full_ffp_varyings)
3018 {
3019 const struct wined3d_shader *vs = state->shader[WINED3D_SHADER_TYPE_VERTEX];
3020
3021 args->texcoords_initialized = 0;
3022 for (i = 0; i < WINED3D_MAX_FFP_TEXTURES; ++i)
3023 {
3024 if (vs)
3025 {
3026 if (state->shader[WINED3D_SHADER_TYPE_VERTEX]->reg_maps.output_registers & (1u << i))
3027 args->texcoords_initialized |= 1u << i;
3028 }
3029 else
3030 {
3031 const struct wined3d_stream_info *si = &context->stream_info;
3032 unsigned int coord_idx = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3033
3034 if ((state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX] >> WINED3D_FFP_TCI_SHIFT)
3035 & WINED3D_FFP_TCI_MASK
3036 || (coord_idx < WINED3D_MAX_FFP_TEXTURES && (si->use_map & (1u << (WINED3D_FFP_TEXCOORD0 + coord_idx)))))
3037 args->texcoords_initialized |= 1u << i;
3038 }
3039 }
3040 }
3041 else
3042 {
3043 args->texcoords_initialized = wined3d_mask_from_size(WINED3D_MAX_FFP_TEXTURES);
3044 }
3045
3046 args->pointsprite = state->render_states[WINED3D_RS_POINTSPRITEENABLE]
3047 && state->primitive_type == WINED3D_PT_POINTLIST;
3048
3049 if (d3d_info->ffp_alpha_test)
3050 args->alpha_test_func = WINED3D_CMP_ALWAYS - 1;
3051 else
3052 args->alpha_test_func = (state->render_states[WINED3D_RS_ALPHATESTENABLE]
3053 ? wined3d_sanitize_cmp_func(state->render_states[WINED3D_RS_ALPHAFUNC])
3054 : WINED3D_CMP_ALWAYS) - 1;
3055
3056 if (d3d_info->emulated_flatshading)
3057 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3058
3059 args->y_correction = (shader->reg_maps.vpos && d3d_info->frag_coord_correction)
3060 || (shader->reg_maps.usesdsy && wined3d_settings.offscreen_rendering_mode != ORM_FBO)
3061 ? !context->render_offscreen : 0;
3062
3063 for (i = 0; i < ARRAY_SIZE(state->fb.render_targets); ++i)
3064 {
3065 struct wined3d_rendertarget_view *rtv = state->fb.render_targets[i];
3066 if (rtv && rtv->format->id == WINED3DFMT_A8_UNORM && !is_identity_fixup(rtv->format->color_fixup))
3067 args->rt_alpha_swizzle |= 1u << i;
3068 }
3069
3070 args->dual_source_blend = state->blend_state && state->blend_state->dual_source;
3071 }
3072
3073 static HRESULT pixel_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3074 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3075 {
3076 const struct wined3d_d3d_info *d3d_info = &device->adapter->d3d_info;
3077 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
3078 HRESULT hr;
3079
3080 if (FAILED(hr = shader_init(shader, device, desc, parent, parent_ops)))
3081 return hr;
3082
3083 if (FAILED(hr = shader_set_function(shader, device,
3084 WINED3D_SHADER_TYPE_PIXEL, d3d_info->limits.ps_uniform_count)))
3085 {
3086 shader_cleanup(shader);
3087 return hr;
3088 }
3089
3090 for (i = 0; i < MAX_REG_INPUT; ++i)
3091 {
3092 if (shader->u.ps.input_reg_used & (1u << i))
3093 {
3094 ++num_regs_used;
3095 highest_reg_used = i;
3096 }
3097 }
3098
3099 /* Don't do any register mapping magic if it is not needed, or if we can't
3100 * achieve anything anyway */
3101 if (highest_reg_used < (d3d_info->limits.varying_count / 4)
3102 || num_regs_used > (d3d_info->limits.varying_count / 4)
3103 || shader->reg_maps.shader_version.major >= 4)
3104 {
3105 if (num_regs_used > (d3d_info->limits.varying_count / 4))
3106 {
3107 /* This happens with relative addressing. The input mapper function
3108 * warns about this if the higher registers are declared too, so
3109 * don't write a FIXME here */
3110 WARN("More varying registers used than supported\n");
3111 }
3112
3113 for (i = 0; i < MAX_REG_INPUT; ++i)
3114 {
3115 shader->u.ps.input_reg_map[i] = i;
3116 }
3117
3118 shader->u.ps.declared_in_count = highest_reg_used + 1;
3119 }
3120 else
3121 {
3122 shader->u.ps.declared_in_count = 0;
3123 for (i = 0; i < MAX_REG_INPUT; ++i)
3124 {
3125 if (shader->u.ps.input_reg_used & (1u << i))
3126 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
3127 else shader->u.ps.input_reg_map[i] = ~0U;
3128 }
3129 }
3130
3131 return WINED3D_OK;
3132 }
3133
3134 enum wined3d_shader_resource_type pixelshader_get_resource_type(const struct wined3d_shader_reg_maps *reg_maps,
3135 unsigned int resource_idx, DWORD tex_types)
3136 {
3137 static enum wined3d_shader_resource_type shader_resource_type_from_shader_tex_types[] =
3138 {
3139 WINED3D_SHADER_RESOURCE_TEXTURE_2D, /* WINED3D_SHADER_TEX_2D */
3140 WINED3D_SHADER_RESOURCE_TEXTURE_3D, /* WINED3D_SHADER_TEX_3D */
3141 WINED3D_SHADER_RESOURCE_TEXTURE_CUBE, /* WINED3D_SHADER_TEX_CUBE */
3142 };
3143
3144 unsigned int idx;
3145
3146 if (reg_maps->shader_version.major > 3)
3147 return reg_maps->resource_info[resource_idx].type;
3148
3149 if (!reg_maps->resource_info[resource_idx].type)
3150 return 0;
3151
3152 idx = (tex_types >> resource_idx * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK;
3153 assert(idx < ARRAY_SIZE(shader_resource_type_from_shader_tex_types));
3154 return shader_resource_type_from_shader_tex_types[idx];
3155 }
3156
3157 HRESULT CDECL wined3d_shader_create_cs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3158 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3159 {
3160 struct wined3d_shader *object;
3161 HRESULT hr;
3162
3163 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3164 device, desc, parent, parent_ops, shader);
3165
3166 if (!(object = calloc(1, sizeof(*object))))
3167 return E_OUTOFMEMORY;
3168
3169 if (FAILED(hr = shader_init(object, device, desc, parent, parent_ops)))
3170 {
3171 WARN("Failed to initialize compute shader, hr %#lx.\n", hr);
3172 free(object);
3173 return hr;
3174 }
3175
3176 if (FAILED(hr = shader_set_function(object, device, WINED3D_SHADER_TYPE_COMPUTE, 0)))
3177 {
3178 shader_cleanup(object);
3179 free(object);
3180 return hr;
3181 }
3182
3183 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
3184
3185 TRACE("Created compute shader %p.\n", object);
3186 *shader = object;
3187
3188 return WINED3D_OK;
3189 }
3190
3191 HRESULT CDECL wined3d_shader_create_ds(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3192 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3193 {
3194 struct wined3d_shader *object;
3195 HRESULT hr;
3196
3197 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3198 device, desc, parent, parent_ops, shader);
3199
3200 if (!(object = calloc(1, sizeof(*object))))
3201 return E_OUTOFMEMORY;
3202
3203 if (FAILED(hr = shader_init(object, device, desc, parent, parent_ops)))
3204 {
3205 WARN("Failed to initialize domain shader, hr %#lx.\n", hr);
3206 free(object);
3207 return hr;
3208 }
3209
3210 if (FAILED(hr = shader_set_function(object, device, WINED3D_SHADER_TYPE_DOMAIN, 0)))
3211 {
3212 shader_cleanup(object);
3213 free(object);
3214 return hr;
3215 }
3216
3217 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
3218
3219 TRACE("Created domain shader %p.\n", object);
3220 *shader = object;
3221
3222 return WINED3D_OK;
3223 }
3224
3225 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3226 const struct wined3d_stream_output_desc *so_desc, void *parent,
3227 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3228 {
3229 struct wined3d_shader *object;
3230 HRESULT hr;
3231
3232 TRACE("device %p, desc %p, so_desc %p, parent %p, parent_ops %p, shader %p.\n",
3233 device, desc, so_desc, parent, parent_ops, shader);
3234
3235 if (!(object = calloc(1, sizeof(*object))))
3236 return E_OUTOFMEMORY;
3237
3238 if (FAILED(hr = geometry_shader_init(object, device, desc, so_desc, parent, parent_ops)))
3239 {
3240 WARN("Failed to initialize geometry shader, hr %#lx.\n", hr);
3241 free(object);
3242 return hr;
3243 }
3244
3245 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
3246
3247 TRACE("Created geometry shader %p.\n", object);
3248 *shader = object;
3249
3250 return WINED3D_OK;
3251 }
3252
3253 HRESULT CDECL wined3d_shader_create_hs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3254 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3255 {
3256 struct wined3d_shader *object;
3257 HRESULT hr;
3258
3259 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3260 device, desc, parent, parent_ops, shader);
3261
3262 if (!(object = calloc(1, sizeof(*object))))
3263 return E_OUTOFMEMORY;
3264
3265 if (FAILED(hr = shader_init(object, device, desc, parent, parent_ops)))
3266 {
3267 WARN("Failed to initialize hull shader, hr %#lx.\n", hr);
3268 free(object);
3269 return hr;
3270 }
3271
3272 if (FAILED(hr = shader_set_function(object, device, WINED3D_SHADER_TYPE_HULL, 0)))
3273 {
3274 shader_cleanup(object);
3275 free(object);
3276 return hr;
3277 }
3278
3279 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
3280
3281 TRACE("Created hull shader %p.\n", object);
3282 *shader = object;
3283
3284 return WINED3D_OK;
3285 }
3286
3287 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3288 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3289 {
3290 struct wined3d_shader *object;
3291 HRESULT hr;
3292
3293 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3294 device, desc, parent, parent_ops, shader);
3295
3296 if (!(object = calloc(1, sizeof(*object))))
3297 return E_OUTOFMEMORY;
3298
3299 if (FAILED(hr = pixel_shader_init(object, device, desc, parent, parent_ops)))
3300 {
3301 WARN("Failed to initialize pixel shader, hr %#lx.\n", hr);
3302 free(object);
3303 return hr;
3304 }
3305
3306 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
3307
3308 TRACE("Created pixel shader %p.\n", object);
3309 *shader = object;
3310
3311 return WINED3D_OK;
3312 }
3313
3314 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3315 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3316 {
3317 struct wined3d_shader *object;
3318 HRESULT hr;
3319
3320 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3321 device, desc, parent, parent_ops, shader);
3322
3323 if (!(object = calloc(1, sizeof(*object))))
3324 return E_OUTOFMEMORY;
3325
3326 if (FAILED(hr = vertex_shader_init(object, device, desc, parent, parent_ops)))
3327 {
3328 WARN("Failed to initialize vertex shader, hr %#lx.\n", hr);
3329 free(object);
3330 return hr;
3331 }
3332
3333 wined3d_cs_init_object(device->cs, wined3d_shader_init_object, object);
3334
3335 TRACE("Created vertex shader %p.\n", object);
3336 *shader = object;
3337
3338 return WINED3D_OK;
3339 }
Windows API implementation