?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
wineandroid: Create the top-level view on desktop window creation.
[wine.git] / dlls / wined3d / shader.c
blob8827641e36824f1827e107b011a6357ff254ac8e
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 "config.h"
26 #include "wine/port.h"
27
28 #include <math.h>
29 #include <stdio.h>
30 #include <string.h>
31
32 #include "wined3d_private.h"
33
34 WINE_DEFAULT_DEBUG_CHANNEL(d3d_shader);
35
36 /* pow, mul_high, sub_high, mul_low */
37 const float wined3d_srgb_const0[] = {0.41666f, 1.055f, 0.055f, 12.92f};
38 /* cmp */
39 const float wined3d_srgb_const1[] = {0.0031308f, 0.0f, 0.0f, 0.0f};
40
41 static const char * const shader_opcode_names[] =
42 {
43 /* WINED3DSIH_ABS */ "abs",
44 /* WINED3DSIH_ADD */ "add",
45 /* WINED3DSIH_AND */ "and",
46 /* WINED3DSIH_ATOMIC_AND */ "atomic_and",
47 /* WINED3DSIH_ATOMIC_CMP_STORE */ "atomic_cmp_store",
48 /* WINED3DSIH_ATOMIC_IADD */ "atomic_iadd",
49 /* WINED3DSIH_ATOMIC_IMAX */ "atomic_imax",
50 /* WINED3DSIH_ATOMIC_IMIN */ "atomic_imin",
51 /* WINED3DSIH_ATOMIC_OR */ "atomic_or",
52 /* WINED3DSIH_ATOMIC_UMAX */ "atomic_umax",
53 /* WINED3DSIH_ATOMIC_UMIN */ "atomic_umin",
54 /* WINED3DSIH_ATOMIC_XOR */ "atomic_xor",
55 /* WINED3DSIH_BEM */ "bem",
56 /* WINED3DSIH_BFI */ "bfi",
57 /* WINED3DSIH_BFREV */ "bfrev",
58 /* WINED3DSIH_BREAK */ "break",
59 /* WINED3DSIH_BREAKC */ "breakc",
60 /* WINED3DSIH_BREAKP */ "breakp",
61 /* WINED3DSIH_BUFINFO */ "bufinfo",
62 /* WINED3DSIH_CALL */ "call",
63 /* WINED3DSIH_CALLNZ */ "callnz",
64 /* WINED3DSIH_CASE */ "case",
65 /* WINED3DSIH_CMP */ "cmp",
66 /* WINED3DSIH_CND */ "cnd",
67 /* WINED3DSIH_CONTINUE */ "continue",
68 /* WINED3DSIH_CONTINUEP */ "continuec",
69 /* WINED3DSIH_COUNTBITS */ "countbits",
70 /* WINED3DSIH_CRS */ "crs",
71 /* WINED3DSIH_CUT */ "cut",
72 /* WINED3DSIH_CUT_STREAM */ "cut_stream",
73 /* WINED3DSIH_DCL */ "dcl",
74 /* WINED3DSIH_DCL_CONSTANT_BUFFER */ "dcl_constantBuffer",
75 /* WINED3DSIH_DCL_FUNCTION_BODY */ "dcl_function_body",
76 /* WINED3DSIH_DCL_FUNCTION_TABLE */ "dcl_function_table",
77 /* WINED3DSIH_DCL_GLOBAL_FLAGS */ "dcl_globalFlags",
78 /* WINED3DSIH_DCL_GS_INSTANCES */ "dcl_gs_instances",
79 /* WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT */ "dcl_hs_fork_phase_instance_count",
80 /* WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT */ "dcl_hs_join_phase_instance_count",
81 /* WINED3DSIH_DCL_HS_MAX_TESSFACTOR */ "dcl_hs_max_tessfactor",
82 /* WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER */ "dcl_immediateConstantBuffer",
83 /* WINED3DSIH_DCL_INDEX_RANGE */ "dcl_index_range",
84 /* WINED3DSIH_DCL_INDEXABLE_TEMP */ "dcl_indexableTemp",
85 /* WINED3DSIH_DCL_INPUT */ "dcl_input",
86 /* WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT */ "dcl_input_control_point_count",
87 /* WINED3DSIH_DCL_INPUT_PRIMITIVE */ "dcl_inputPrimitive",
88 /* WINED3DSIH_DCL_INPUT_PS */ "dcl_input_ps",
89 /* WINED3DSIH_DCL_INPUT_PS_SGV */ "dcl_input_ps_sgv",
90 /* WINED3DSIH_DCL_INPUT_PS_SIV */ "dcl_input_ps_siv",
91 /* WINED3DSIH_DCL_INPUT_SGV */ "dcl_input_sgv",
92 /* WINED3DSIH_DCL_INPUT_SIV */ "dcl_input_siv",
93 /* WINED3DSIH_DCL_INTERFACE */ "dcl_interface",
94 /* WINED3DSIH_DCL_OUTPUT */ "dcl_output",
95 /* WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT */ "dcl_output_control_point_count",
96 /* WINED3DSIH_DCL_OUTPUT_SIV */ "dcl_output_siv",
97 /* WINED3DSIH_DCL_OUTPUT_TOPOLOGY */ "dcl_outputTopology",
98 /* WINED3DSIH_DCL_RESOURCE_RAW */ "dcl_resource_raw",
99 /* WINED3DSIH_DCL_RESOURCE_STRUCTURED */ "dcl_resource_structured",
100 /* WINED3DSIH_DCL_SAMPLER */ "dcl_sampler",
101 /* WINED3DSIH_DCL_STREAM */ "dcl_stream",
102 /* WINED3DSIH_DCL_TEMPS */ "dcl_temps",
103 /* WINED3DSIH_DCL_TESSELLATOR_DOMAIN */ "dcl_tessellator_domain",
104 /* WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE */ "dcl_tessellator_output_primitive",
105 /* WINED3DSIH_DCL_TESSELLATOR_PARTITIONING */ "dcl_tessellator_partitioning",
106 /* WINED3DSIH_DCL_TGSM_RAW */ "dcl_tgsm_raw",
107 /* WINED3DSIH_DCL_TGSM_STRUCTURED */ "dcl_tgsm_structured",
108 /* WINED3DSIH_DCL_THREAD_GROUP */ "dcl_thread_group",
109 /* WINED3DSIH_DCL_UAV_RAW */ "dcl_uav_raw",
110 /* WINED3DSIH_DCL_UAV_STRUCTURED */ "dcl_uav_structured",
111 /* WINED3DSIH_DCL_UAV_TYPED */ "dcl_uav_typed",
112 /* WINED3DSIH_DCL_VERTICES_OUT */ "dcl_maxOutputVertexCount",
113 /* WINED3DSIH_DEF */ "def",
114 /* WINED3DSIH_DEFAULT */ "default",
115 /* WINED3DSIH_DEFB */ "defb",
116 /* WINED3DSIH_DEFI */ "defi",
117 /* WINED3DSIH_DIV */ "div",
118 /* WINED3DSIH_DP2 */ "dp2",
119 /* WINED3DSIH_DP2ADD */ "dp2add",
120 /* WINED3DSIH_DP3 */ "dp3",
121 /* WINED3DSIH_DP4 */ "dp4",
122 /* WINED3DSIH_DST */ "dst",
123 /* WINED3DSIH_DSX */ "dsx",
124 /* WINED3DSIH_DSX_COARSE */ "deriv_rtx_coarse",
125 /* WINED3DSIH_DSX_FINE */ "deriv_rtx_fine",
126 /* WINED3DSIH_DSY */ "dsy",
127 /* WINED3DSIH_DSY_COARSE */ "deriv_rty_coarse",
128 /* WINED3DSIH_DSY_FINE */ "deriv_rty_fine",
129 /* WINED3DSIH_ELSE */ "else",
130 /* WINED3DSIH_EMIT */ "emit",
131 /* WINED3DSIH_EMIT_STREAM */ "emit_stream",
132 /* WINED3DSIH_ENDIF */ "endif",
133 /* WINED3DSIH_ENDLOOP */ "endloop",
134 /* WINED3DSIH_ENDREP */ "endrep",
135 /* WINED3DSIH_ENDSWITCH */ "endswitch",
136 /* WINED3DSIH_EQ */ "eq",
137 /* WINED3DSIH_EXP */ "exp",
138 /* WINED3DSIH_EXPP */ "expp",
139 /* WINED3DSIH_F16TOF32 */ "f16tof32",
140 /* WINED3DSIH_F32TOF16 */ "f32tof16",
141 /* WINED3DSIH_FCALL */ "fcall",
142 /* WINED3DSIH_FIRSTBIT_HI */ "firstbit_hi",
143 /* WINED3DSIH_FIRSTBIT_LO */ "firstbit_lo",
144 /* WINED3DSIH_FIRSTBIT_SHI */ "firstbit_shi",
145 /* WINED3DSIH_FRC */ "frc",
146 /* WINED3DSIH_FTOI */ "ftoi",
147 /* WINED3DSIH_FTOU */ "ftou",
148 /* WINED3DSIH_GATHER4 */ "gather4",
149 /* WINED3DSIH_GATHER4_C */ "gather4_c",
150 /* WINED3DSIH_GATHER4_PO */ "gather4_po",
151 /* WINED3DSIH_GATHER4_PO_C */ "gather4_po_c",
152 /* WINED3DSIH_GE */ "ge",
153 /* WINED3DSIH_HS_CONTROL_POINT_PHASE */ "hs_control_point_phase",
154 /* WINED3DSIH_HS_DECLS */ "hs_decls",
155 /* WINED3DSIH_HS_FORK_PHASE */ "hs_fork_phase",
156 /* WINED3DSIH_HS_JOIN_PHASE */ "hs_join_phase",
157 /* WINED3DSIH_IADD */ "iadd",
158 /* WINED3DSIH_IBFE */ "ibfe",
159 /* WINED3DSIH_IEQ */ "ieq",
160 /* WINED3DSIH_IF */ "if",
161 /* WINED3DSIH_IFC */ "ifc",
162 /* WINED3DSIH_IGE */ "ige",
163 /* WINED3DSIH_ILT */ "ilt",
164 /* WINED3DSIH_IMAD */ "imad",
165 /* WINED3DSIH_IMAX */ "imax",
166 /* WINED3DSIH_IMIN */ "imin",
167 /* WINED3DSIH_IMM_ATOMIC_ALLOC */ "imm_atomic_alloc",
168 /* WINED3DSIH_IMM_ATOMIC_AND */ "imm_atomic_and",
169 /* WINED3DSIH_IMM_ATOMIC_CMP_EXCH */ "imm_atomic_cmp_exch",
170 /* WINED3DSIH_IMM_ATOMIC_CONSUME */ "imm_atomic_consume",
171 /* WINED3DSIH_IMM_ATOMIC_EXCH */ "imm_atomic_exch",
172 /* WINED3DSIH_IMM_ATOMIC_IADD */ "imm_atomic_iadd",
173 /* WINED3DSIH_IMM_ATOMIC_IMAX */ "imm_atomic_imax",
174 /* WINED3DSIH_IMM_ATOMIC_IMIN */ "imm_atomic_imin",
175 /* WINED3DSIH_IMM_ATOMIC_OR */ "imm_atomic_or",
176 /* WINED3DSIH_IMM_ATOMIC_UMAX */ "imm_atomic_umax",
177 /* WINED3DSIH_IMM_ATOMIC_UMIN */ "imm_atomic_umin",
178 /* WINED3DSIH_IMM_ATOMIC_XOR */ "imm_atomic_xor",
179 /* WINED3DSIH_IMUL */ "imul",
180 /* WINED3DSIH_INE */ "ine",
181 /* WINED3DSIH_INEG */ "ineg",
182 /* WINED3DSIH_ISHL */ "ishl",
183 /* WINED3DSIH_ISHR */ "ishr",
184 /* WINED3DSIH_ITOF */ "itof",
185 /* WINED3DSIH_LABEL */ "label",
186 /* WINED3DSIH_LD */ "ld",
187 /* WINED3DSIH_LD2DMS */ "ld2dms",
188 /* WINED3DSIH_LD_RAW */ "ld_raw",
189 /* WINED3DSIH_LD_STRUCTURED */ "ld_structured",
190 /* WINED3DSIH_LD_UAV_TYPED */ "ld_uav_typed",
191 /* WINED3DSIH_LIT */ "lit",
192 /* WINED3DSIH_LOD */ "lod",
193 /* WINED3DSIH_LOG */ "log",
194 /* WINED3DSIH_LOGP */ "logp",
195 /* WINED3DSIH_LOOP */ "loop",
196 /* WINED3DSIH_LRP */ "lrp",
197 /* WINED3DSIH_LT */ "lt",
198 /* WINED3DSIH_M3x2 */ "m3x2",
199 /* WINED3DSIH_M3x3 */ "m3x3",
200 /* WINED3DSIH_M3x4 */ "m3x4",
201 /* WINED3DSIH_M4x3 */ "m4x3",
202 /* WINED3DSIH_M4x4 */ "m4x4",
203 /* WINED3DSIH_MAD */ "mad",
204 /* WINED3DSIH_MAX */ "max",
205 /* WINED3DSIH_MIN */ "min",
206 /* WINED3DSIH_MOV */ "mov",
207 /* WINED3DSIH_MOVA */ "mova",
208 /* WINED3DSIH_MOVC */ "movc",
209 /* WINED3DSIH_MUL */ "mul",
210 /* WINED3DSIH_NE */ "ne",
211 /* WINED3DSIH_NOP */ "nop",
212 /* WINED3DSIH_NOT */ "not",
213 /* WINED3DSIH_NRM */ "nrm",
214 /* WINED3DSIH_OR */ "or",
215 /* WINED3DSIH_PHASE */ "phase",
216 /* WINED3DSIH_POW */ "pow",
217 /* WINED3DSIH_RCP */ "rcp",
218 /* WINED3DSIH_REP */ "rep",
219 /* WINED3DSIH_RESINFO */ "resinfo",
220 /* WINED3DSIH_RET */ "ret",
221 /* WINED3DSIH_RETP */ "retp",
222 /* WINED3DSIH_ROUND_NE */ "round_ne",
223 /* WINED3DSIH_ROUND_NI */ "round_ni",
224 /* WINED3DSIH_ROUND_PI */ "round_pi",
225 /* WINED3DSIH_ROUND_Z */ "round_z",
226 /* WINED3DSIH_RSQ */ "rsq",
227 /* WINED3DSIH_SAMPLE */ "sample",
228 /* WINED3DSIH_SAMPLE_B */ "sample_b",
229 /* WINED3DSIH_SAMPLE_C */ "sample_c",
230 /* WINED3DSIH_SAMPLE_C_LZ */ "sample_c_lz",
231 /* WINED3DSIH_SAMPLE_GRAD */ "sample_d",
232 /* WINED3DSIH_SAMPLE_INFO */ "sample_info",
233 /* WINED3DSIH_SAMPLE_LOD */ "sample_l",
234 /* WINED3DSIH_SAMPLE_POS */ "sample_pos",
235 /* WINED3DSIH_SETP */ "setp",
236 /* WINED3DSIH_SGE */ "sge",
237 /* WINED3DSIH_SGN */ "sgn",
238 /* WINED3DSIH_SINCOS */ "sincos",
239 /* WINED3DSIH_SLT */ "slt",
240 /* WINED3DSIH_SQRT */ "sqrt",
241 /* WINED3DSIH_STORE_RAW */ "store_raw",
242 /* WINED3DSIH_STORE_STRUCTURED */ "store_structured",
243 /* WINED3DSIH_STORE_UAV_TYPED */ "store_uav_typed",
244 /* WINED3DSIH_SUB */ "sub",
245 /* WINED3DSIH_SWAPC */ "swapc",
246 /* WINED3DSIH_SWITCH */ "switch",
247 /* WINED3DSIH_SYNC */ "sync",
248 /* WINED3DSIH_TEX */ "texld",
249 /* WINED3DSIH_TEXBEM */ "texbem",
250 /* WINED3DSIH_TEXBEML */ "texbeml",
251 /* WINED3DSIH_TEXCOORD */ "texcrd",
252 /* WINED3DSIH_TEXDEPTH */ "texdepth",
253 /* WINED3DSIH_TEXDP3 */ "texdp3",
254 /* WINED3DSIH_TEXDP3TEX */ "texdp3tex",
255 /* WINED3DSIH_TEXKILL */ "texkill",
256 /* WINED3DSIH_TEXLDD */ "texldd",
257 /* WINED3DSIH_TEXLDL */ "texldl",
258 /* WINED3DSIH_TEXM3x2DEPTH */ "texm3x2depth",
259 /* WINED3DSIH_TEXM3x2PAD */ "texm3x2pad",
260 /* WINED3DSIH_TEXM3x2TEX */ "texm3x2tex",
261 /* WINED3DSIH_TEXM3x3 */ "texm3x3",
262 /* WINED3DSIH_TEXM3x3DIFF */ "texm3x3diff",
263 /* WINED3DSIH_TEXM3x3PAD */ "texm3x3pad",
264 /* WINED3DSIH_TEXM3x3SPEC */ "texm3x3spec",
265 /* WINED3DSIH_TEXM3x3TEX */ "texm3x3tex",
266 /* WINED3DSIH_TEXM3x3VSPEC */ "texm3x3vspec",
267 /* WINED3DSIH_TEXREG2AR */ "texreg2ar",
268 /* WINED3DSIH_TEXREG2GB */ "texreg2gb",
269 /* WINED3DSIH_TEXREG2RGB */ "texreg2rgb",
270 /* WINED3DSIH_UBFE */ "ubfe",
271 /* WINED3DSIH_UDIV */ "udiv",
272 /* WINED3DSIH_UGE */ "uge",
273 /* WINED3DSIH_ULT */ "ult",
274 /* WINED3DSIH_UMAX */ "umax",
275 /* WINED3DSIH_UMIN */ "umin",
276 /* WINED3DSIH_UMUL */ "umul",
277 /* WINED3DSIH_USHR */ "ushr",
278 /* WINED3DSIH_UTOF */ "utof",
279 /* WINED3DSIH_XOR */ "xor",
280 };
281
282 static const char * const semantic_names[] =
283 {
284 /* WINED3D_DECL_USAGE_POSITION */ "SV_POSITION",
285 /* WINED3D_DECL_USAGE_BLEND_WEIGHT */ "BLENDWEIGHT",
286 /* WINED3D_DECL_USAGE_BLEND_INDICES */ "BLENDINDICES",
287 /* WINED3D_DECL_USAGE_NORMAL */ "NORMAL",
288 /* WINED3D_DECL_USAGE_PSIZE */ "PSIZE",
289 /* WINED3D_DECL_USAGE_TEXCOORD */ "TEXCOORD",
290 /* WINED3D_DECL_USAGE_TANGENT */ "TANGENT",
291 /* WINED3D_DECL_USAGE_BINORMAL */ "BINORMAL",
292 /* WINED3D_DECL_USAGE_TESS_FACTOR */ "TESSFACTOR",
293 /* WINED3D_DECL_USAGE_POSITIONT */ "POSITIONT",
294 /* WINED3D_DECL_USAGE_COLOR */ "COLOR",
295 /* WINED3D_DECL_USAGE_FOG */ "FOG",
296 /* WINED3D_DECL_USAGE_DEPTH */ "DEPTH",
297 /* WINED3D_DECL_USAGE_SAMPLE */ "SAMPLE",
298 };
299
300 static const struct
301 {
302 enum wined3d_shader_input_sysval_semantic sysval_semantic;
303 const char *sysval_name;
304 }
305 shader_input_sysval_semantic_names[] =
306 {
307 {WINED3D_SIV_POSITION, "position"},
308 {WINED3D_SIV_CLIP_DISTANCE, "clip_distance"},
309 {WINED3D_SIV_CULL_DISTANCE, "cull_distance"},
310 {WINED3D_SIV_RENDER_TARGET_ARRAY_INDEX, "render_target_array_index"},
311 {WINED3D_SIV_VIEWPORT_ARRAY_INDEX, "viewport_array_index"},
312 {WINED3D_SIV_VERTEX_ID, "vertex_id"},
313 {WINED3D_SIV_INSTANCE_ID, "instance_id"},
314 {WINED3D_SIV_PRIMITIVE_ID, "primitive_id"},
315 {WINED3D_SIV_IS_FRONT_FACE, "is_front_face"},
316 {WINED3D_SIV_SAMPLE_INDEX, "sample_index"},
317 {WINED3D_SIV_QUAD_U0_TESS_FACTOR, "finalQuadUeq0EdgeTessFactor"},
318 {WINED3D_SIV_QUAD_V0_TESS_FACTOR, "finalQuadVeq0EdgeTessFactor"},
319 {WINED3D_SIV_QUAD_U1_TESS_FACTOR, "finalQuadUeq1EdgeTessFactor"},
320 {WINED3D_SIV_QUAD_V1_TESS_FACTOR, "finalQuadVeq1EdgeTessFactor"},
321 {WINED3D_SIV_QUAD_U_INNER_TESS_FACTOR, "finalQuadUInsideTessFactor"},
322 {WINED3D_SIV_QUAD_V_INNER_TESS_FACTOR, "finalQuadVInsideTessFactor"},
323 {WINED3D_SIV_TRIANGLE_U_TESS_FACTOR, "finalTriUeq0EdgeTessFactor"},
324 {WINED3D_SIV_TRIANGLE_V_TESS_FACTOR, "finalTriVeq0EdgeTessFactor"},
325 {WINED3D_SIV_TRIANGLE_W_TESS_FACTOR, "finalTriWeq0EdgeTessFactor"},
326 {WINED3D_SIV_TRIANGLE_INNER_TESS_FACTOR, "finalTriInsideTessFactor"},
327 {WINED3D_SIV_LINE_DETAIL_TESS_FACTOR, "finalLineDetailTessFactor"},
328 {WINED3D_SIV_LINE_DENSITY_TESS_FACTOR, "finalLineDensityTessFactor"},
329 };
330
331 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
332 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version);
333
334 const char *debug_d3dshaderinstructionhandler(enum WINED3D_SHADER_INSTRUCTION_HANDLER handler_idx)
335 {
336 if (handler_idx >= ARRAY_SIZE(shader_opcode_names))
337 return wine_dbg_sprintf("UNRECOGNIZED(%#x)", handler_idx);
338
339 return shader_opcode_names[handler_idx];
340 }
341
342 static const char *shader_semantic_name_from_usage(enum wined3d_decl_usage usage)
343 {
344 if (usage >= ARRAY_SIZE(semantic_names))
345 {
346 FIXME("Unrecognized usage %#x.\n", usage);
347 return "UNRECOGNIZED";
348 }
349
350 return semantic_names[usage];
351 }
352
353 static enum wined3d_decl_usage shader_usage_from_semantic_name(const char *name)
354 {
355 unsigned int i;
356
357 for (i = 0; i < ARRAY_SIZE(semantic_names); ++i)
358 {
359 if (!strcmp(name, semantic_names[i]))
360 return i;
361 }
362
363 return ~0U;
364 }
365
366 static enum wined3d_sysval_semantic shader_sysval_semantic_from_usage(enum wined3d_decl_usage usage)
367 {
368 switch (usage)
369 {
370 case WINED3D_DECL_USAGE_POSITION:
371 return WINED3D_SV_POSITION;
372 default:
373 return 0;
374 }
375 }
376
377 BOOL shader_match_semantic(const char *semantic_name, enum wined3d_decl_usage usage)
378 {
379 return !strcmp(semantic_name, shader_semantic_name_from_usage(usage));
380 }
381
382 static void shader_signature_from_semantic(struct wined3d_shader_signature_element *e,
383 const struct wined3d_shader_semantic *s)
384 {
385 e->semantic_name = shader_semantic_name_from_usage(s->usage);
386 e->semantic_idx = s->usage_idx;
387 e->stream_idx = 0;
388 e->sysval_semantic = shader_sysval_semantic_from_usage(s->usage);
389 e->component_type = WINED3D_TYPE_FLOAT;
390 e->register_idx = s->reg.reg.idx[0].offset;
391 e->mask = s->reg.write_mask;
392 }
393
394 static void shader_signature_from_usage(struct wined3d_shader_signature_element *e,
395 enum wined3d_decl_usage usage, UINT usage_idx, UINT reg_idx, DWORD write_mask)
396 {
397 e->semantic_name = shader_semantic_name_from_usage(usage);
398 e->semantic_idx = usage_idx;
399 e->stream_idx = 0;
400 e->sysval_semantic = shader_sysval_semantic_from_usage(usage);
401 e->component_type = WINED3D_TYPE_FLOAT;
402 e->register_idx = reg_idx;
403 e->mask = write_mask;
404 }
405
406 static const struct wined3d_shader_frontend *shader_select_frontend(enum wined3d_shader_byte_code_format format)
407 {
408 switch (format)
409 {
410 case WINED3D_SHADER_BYTE_CODE_FORMAT_SM1:
411 return &sm1_shader_frontend;
412
413 case WINED3D_SHADER_BYTE_CODE_FORMAT_SM4:
414 return &sm4_shader_frontend;
415
416 default:
417 WARN("Invalid byte code format %#x specified.\n", format);
418 return NULL;
419 }
420 }
421
422 void string_buffer_clear(struct wined3d_string_buffer *buffer)
423 {
424 buffer->buffer[0] = '\0';
425 buffer->content_size = 0;
426 }
427
428 BOOL string_buffer_init(struct wined3d_string_buffer *buffer)
429 {
430 buffer->buffer_size = 32;
431 if (!(buffer->buffer = HeapAlloc(GetProcessHeap(), 0, buffer->buffer_size)))
432 {
433 ERR("Failed to allocate shader buffer memory.\n");
434 return FALSE;
435 }
436
437 string_buffer_clear(buffer);
438 return TRUE;
439 }
440
441 void string_buffer_free(struct wined3d_string_buffer *buffer)
442 {
443 HeapFree(GetProcessHeap(), 0, buffer->buffer);
444 }
445
446 BOOL string_buffer_resize(struct wined3d_string_buffer *buffer, int rc)
447 {
448 char *new_buffer;
449 unsigned int new_buffer_size = buffer->buffer_size * 2;
450
451 while (rc > 0 && (unsigned int)rc >= new_buffer_size - buffer->content_size)
452 new_buffer_size *= 2;
453 if (!(new_buffer = HeapReAlloc(GetProcessHeap(), 0, buffer->buffer, new_buffer_size)))
454 {
455 ERR("Failed to grow buffer.\n");
456 buffer->buffer[buffer->content_size] = '\0';
457 return FALSE;
458 }
459 buffer->buffer = new_buffer;
460 buffer->buffer_size = new_buffer_size;
461 return TRUE;
462 }
463
464 int shader_vaddline(struct wined3d_string_buffer *buffer, const char *format, va_list args)
465 {
466 unsigned int rem;
467 int rc;
468
469 rem = buffer->buffer_size - buffer->content_size;
470 rc = vsnprintf(&buffer->buffer[buffer->content_size], rem, format, args);
471 if (rc < 0 /* C89 */ || (unsigned int)rc >= rem /* C99 */)
472 return rc;
473
474 buffer->content_size += rc;
475 return 0;
476 }
477
478 int shader_addline(struct wined3d_string_buffer *buffer, const char *format, ...)
479 {
480 va_list args;
481 int ret;
482
483 for (;;)
484 {
485 va_start(args, format);
486 ret = shader_vaddline(buffer, format, args);
487 va_end(args);
488 if (!ret)
489 return ret;
490 if (!string_buffer_resize(buffer, ret))
491 return -1;
492 }
493 }
494
495 struct wined3d_string_buffer *string_buffer_get(struct wined3d_string_buffer_list *list)
496 {
497 struct wined3d_string_buffer *buffer;
498
499 if (list_empty(&list->list))
500 {
501 buffer = HeapAlloc(GetProcessHeap(), 0, sizeof(*buffer));
502 if (!buffer || !string_buffer_init(buffer))
503 {
504 ERR("Couldn't allocate buffer for temporary string.\n");
505 HeapFree(GetProcessHeap(), 0, buffer);
506 return NULL;
507 }
508 }
509 else
510 {
511 buffer = LIST_ENTRY(list_head(&list->list), struct wined3d_string_buffer, entry);
512 list_remove(&buffer->entry);
513 }
514 string_buffer_clear(buffer);
515 return buffer;
516 }
517
518 static int string_buffer_vsprintf(struct wined3d_string_buffer *buffer, const char *format, va_list args)
519 {
520 if (!buffer)
521 return 0;
522 string_buffer_clear(buffer);
523 return shader_vaddline(buffer, format, args);
524 }
525
526 void string_buffer_sprintf(struct wined3d_string_buffer *buffer, const char *format, ...)
527 {
528 va_list args;
529 int ret;
530
531 for (;;)
532 {
533 va_start(args, format);
534 ret = string_buffer_vsprintf(buffer, format, args);
535 va_end(args);
536 if (!ret)
537 return;
538 if (!string_buffer_resize(buffer, ret))
539 return;
540 }
541 }
542
543 void string_buffer_release(struct wined3d_string_buffer_list *list, struct wined3d_string_buffer *buffer)
544 {
545 if (!buffer)
546 return;
547 list_add_head(&list->list, &buffer->entry);
548 }
549
550 void string_buffer_list_init(struct wined3d_string_buffer_list *list)
551 {
552 list_init(&list->list);
553 }
554
555 void string_buffer_list_cleanup(struct wined3d_string_buffer_list *list)
556 {
557 struct wined3d_string_buffer *buffer, *buffer_next;
558
559 LIST_FOR_EACH_ENTRY_SAFE(buffer, buffer_next, &list->list, struct wined3d_string_buffer, entry)
560 {
561 string_buffer_free(buffer);
562 HeapFree(GetProcessHeap(), 0, buffer);
563 }
564 list_init(&list->list);
565 }
566
567 /* Convert floating point offset relative to a register file to an absolute
568 * offset for float constants. */
569 static unsigned int shader_get_float_offset(enum wined3d_shader_register_type register_type, UINT register_idx)
570 {
571 switch (register_type)
572 {
573 case WINED3DSPR_CONST: return register_idx;
574 case WINED3DSPR_CONST2: return 2048 + register_idx;
575 case WINED3DSPR_CONST3: return 4096 + register_idx;
576 case WINED3DSPR_CONST4: return 6144 + register_idx;
577 default:
578 FIXME("Unsupported register type: %u.\n", register_type);
579 return register_idx;
580 }
581 }
582
583 static void shader_delete_constant_list(struct list *clist)
584 {
585 struct wined3d_shader_lconst *constant, *constant_next;
586
587 LIST_FOR_EACH_ENTRY_SAFE(constant, constant_next, clist, struct wined3d_shader_lconst, entry)
588 HeapFree(GetProcessHeap(), 0, constant);
589 list_init(clist);
590 }
591
592 static void shader_set_limits(struct wined3d_shader *shader)
593 {
594 static const struct limits_entry
595 {
596 unsigned int min_version;
597 unsigned int max_version;
598 struct wined3d_shader_limits limits;
599 }
600 vs_limits[] =
601 {
602 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
603 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 1), { 0, 0, 256, 0, 12, 0}},
604 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 1), { 0, 16, 256, 16, 12, 0}},
605 /* DX10 cards on Windows advertise a D3D9 constant limit of 256
606 * even though they are capable of supporting much more (GL
607 * drivers advertise 1024). d3d9.dll and d3d8.dll clamp the
608 * wined3d-advertised maximum. Clamp the constant limit for <= 3.0
609 * shaders to 256. */
610 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), { 4, 16, 256, 16, 12, 0}},
611 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 16, 0}},
612 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 0}},
613 {0}
614 },
615 hs_limits[] =
616 {
617 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
618 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
619 },
620 ds_limits[] =
621 {
622 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packet_input */
623 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
624 },
625 gs_limits[] =
626 {
627 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
628 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(4, 0), {16, 0, 0, 0, 32, 16}},
629 {WINED3D_SHADER_VERSION(4, 1), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 32, 32}},
630 {0}
631 },
632 ps_limits[] =
633 {
634 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
635 {WINED3D_SHADER_VERSION(1, 0), WINED3D_SHADER_VERSION(1, 3), { 4, 0, 8, 0, 0, 0}},
636 {WINED3D_SHADER_VERSION(1, 4), WINED3D_SHADER_VERSION(1, 4), { 6, 0, 8, 0, 0, 0}},
637 {WINED3D_SHADER_VERSION(2, 0), WINED3D_SHADER_VERSION(2, 0), {16, 0, 32, 0, 0, 0}},
638 {WINED3D_SHADER_VERSION(2, 1), WINED3D_SHADER_VERSION(2, 1), {16, 16, 32, 16, 0, 0}},
639 {WINED3D_SHADER_VERSION(3, 0), WINED3D_SHADER_VERSION(3, 0), {16, 16, 224, 16, 0, 10}},
640 {WINED3D_SHADER_VERSION(4, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 32}},
641 {0}
642 },
643 cs_limits[] =
644 {
645 /* min_version, max_version, sampler, constant_int, constant_float, constant_bool, packed_output, packed_input */
646 {WINED3D_SHADER_VERSION(5, 0), WINED3D_SHADER_VERSION(5, 0), {16, 0, 0, 0, 0, 0}},
647 };
648 const struct limits_entry *limits_array;
649 DWORD shader_version = WINED3D_SHADER_VERSION(shader->reg_maps.shader_version.major,
650 shader->reg_maps.shader_version.minor);
651 int i = 0;
652
653 switch (shader->reg_maps.shader_version.type)
654 {
655 default:
656 FIXME("Unexpected shader type %u found.\n", shader->reg_maps.shader_version.type);
657 /* Fall-through. */
658 case WINED3D_SHADER_TYPE_VERTEX:
659 limits_array = vs_limits;
660 break;
661 case WINED3D_SHADER_TYPE_HULL:
662 limits_array = hs_limits;
663 break;
664 case WINED3D_SHADER_TYPE_DOMAIN:
665 limits_array = ds_limits;
666 break;
667 case WINED3D_SHADER_TYPE_GEOMETRY:
668 limits_array = gs_limits;
669 break;
670 case WINED3D_SHADER_TYPE_PIXEL:
671 limits_array = ps_limits;
672 break;
673 case WINED3D_SHADER_TYPE_COMPUTE:
674 limits_array = cs_limits;
675 break;
676 }
677
678 while (limits_array[i].min_version && limits_array[i].min_version <= shader_version)
679 {
680 if (shader_version <= limits_array[i].max_version)
681 {
682 shader->limits = &limits_array[i].limits;
683 break;
684 }
685 ++i;
686 }
687 if (!shader->limits)
688 {
689 FIXME("Unexpected shader version \"%u.%u\".\n",
690 shader->reg_maps.shader_version.major,
691 shader->reg_maps.shader_version.minor);
692 shader->limits = &limits_array[max(0, i - 1)].limits;
693 }
694 }
695
696 static inline void set_bitmap_bit(DWORD *bitmap, DWORD bit)
697 {
698 DWORD idx, shift;
699 idx = bit >> 5;
700 shift = bit & 0x1f;
701 bitmap[idx] |= (1u << shift);
702 }
703
704 static BOOL shader_record_register_usage(struct wined3d_shader *shader, struct wined3d_shader_reg_maps *reg_maps,
705 const struct wined3d_shader_register *reg, enum wined3d_shader_type shader_type, unsigned int constf_size)
706 {
707 switch (reg->type)
708 {
709 case WINED3DSPR_TEXTURE: /* WINED3DSPR_ADDR */
710 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
711 reg_maps->texcoord |= 1u << reg->idx[0].offset;
712 else
713 reg_maps->address |= 1u << reg->idx[0].offset;
714 break;
715
716 case WINED3DSPR_TEMP:
717 reg_maps->temporary |= 1u << reg->idx[0].offset;
718 break;
719
720 case WINED3DSPR_INPUT:
721 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
722 {
723 /* If relative addressing is used, we must assume that all
724 * registers are used. Even if it is a construct like v3[aL],
725 * we can't assume that v0, v1 and v2 aren't read because aL
726 * can be negative. */
727 if (reg->idx[0].rel_addr)
728 shader->u.ps.input_reg_used = ~0u;
729 else
730 shader->u.ps.input_reg_used |= 1u << reg->idx[0].offset;
731 }
732 else
733 reg_maps->input_registers |= 1u << reg->idx[0].offset;
734 break;
735
736 case WINED3DSPR_RASTOUT:
737 if (reg->idx[0].offset == 1)
738 reg_maps->fog = 1;
739 if (reg->idx[0].offset == 2)
740 reg_maps->point_size = 1;
741 break;
742
743 case WINED3DSPR_MISCTYPE:
744 if (shader_type == WINED3D_SHADER_TYPE_PIXEL)
745 {
746 if (!reg->idx[0].offset)
747 reg_maps->vpos = 1;
748 else if (reg->idx[0].offset == 1)
749 reg_maps->usesfacing = 1;
750 }
751 break;
752
753 case WINED3DSPR_CONST:
754 if (reg->idx[0].rel_addr)
755 {
756 if (reg->idx[0].offset < reg_maps->min_rel_offset)
757 reg_maps->min_rel_offset = reg->idx[0].offset;
758 if (reg->idx[0].offset > reg_maps->max_rel_offset)
759 reg_maps->max_rel_offset = reg->idx[0].offset;
760 reg_maps->usesrelconstF = TRUE;
761 }
762 else
763 {
764 if (reg->idx[0].offset >= min(shader->limits->constant_float, constf_size))
765 {
766 WARN("Shader using float constant %u which is not supported.\n", reg->idx[0].offset);
767 return FALSE;
768 }
769 else
770 {
771 set_bitmap_bit(reg_maps->constf, reg->idx[0].offset);
772 }
773 }
774 break;
775
776 case WINED3DSPR_CONSTINT:
777 if (reg->idx[0].offset >= shader->limits->constant_int)
778 {
779 WARN("Shader using integer constant %u which is not supported.\n", reg->idx[0].offset);
780 return FALSE;
781 }
782 else
783 {
784 reg_maps->integer_constants |= (1u << reg->idx[0].offset);
785 }
786 break;
787
788 case WINED3DSPR_CONSTBOOL:
789 if (reg->idx[0].offset >= shader->limits->constant_bool)
790 {
791 WARN("Shader using bool constant %u which is not supported.\n", reg->idx[0].offset);
792 return FALSE;
793 }
794 else
795 {
796 reg_maps->boolean_constants |= (1u << reg->idx[0].offset);
797 }
798 break;
799
800 case WINED3DSPR_COLOROUT:
801 reg_maps->rt_mask |= (1u << reg->idx[0].offset);
802 break;
803
804 case WINED3DSPR_OUTCONTROLPOINT:
805 reg_maps->vocp = 1;
806 break;
807
808 default:
809 TRACE("Not recording register of type %#x and [%#x][%#x].\n",
810 reg->type, reg->idx[0].offset, reg->idx[1].offset);
811 break;
812 }
813 return TRUE;
814 }
815
816 static void shader_record_sample(struct wined3d_shader_reg_maps *reg_maps,
817 unsigned int resource_idx, unsigned int sampler_idx, unsigned int bind_idx)
818 {
819 struct wined3d_shader_sampler_map_entry *entries, *entry;
820 struct wined3d_shader_sampler_map *map;
821 unsigned int i;
822
823 map = &reg_maps->sampler_map;
824 entries = map->entries;
825 for (i = 0; i < map->count; ++i)
826 {
827 if (entries[i].resource_idx == resource_idx && entries[i].sampler_idx == sampler_idx)
828 return;
829 }
830
831 if (!map->size)
832 {
833 if (!(entries = wined3d_calloc(4, sizeof(*entries))))
834 {
835 ERR("Failed to allocate sampler map entries.\n");
836 return;
837 }
838 map->size = 4;
839 map->entries = entries;
840 }
841 else if (map->count == map->size)
842 {
843 size_t new_size = map->size * 2;
844
845 if (sizeof(*entries) * new_size <= sizeof(*entries) * map->size
846 || !(entries = HeapReAlloc(GetProcessHeap(), 0, entries, sizeof(*entries) * new_size)))
847 {
848 ERR("Failed to resize sampler map entries.\n");
849 return;
850 }
851 map->size = new_size;
852 map->entries = entries;
853 }
854
855 entry = &entries[map->count++];
856 entry->resource_idx = resource_idx;
857 entry->sampler_idx = sampler_idx;
858 entry->bind_idx = bind_idx;
859 }
860
861 static unsigned int get_instr_extra_regcount(enum WINED3D_SHADER_INSTRUCTION_HANDLER instr, unsigned int param)
862 {
863 switch (instr)
864 {
865 case WINED3DSIH_M4x4:
866 case WINED3DSIH_M3x4:
867 return param == 1 ? 3 : 0;
868
869 case WINED3DSIH_M4x3:
870 case WINED3DSIH_M3x3:
871 return param == 1 ? 2 : 0;
872
873 case WINED3DSIH_M3x2:
874 return param == 1 ? 1 : 0;
875
876 default:
877 return 0;
878 }
879 }
880
881 static HRESULT shader_reg_maps_add_tgsm(struct wined3d_shader_reg_maps *reg_maps,
882 unsigned int register_idx, unsigned int size, unsigned int stride)
883 {
884 struct wined3d_shader_tgsm *tgsm;
885
886 if (register_idx >= MAX_TGSM_REGISTERS)
887 {
888 ERR("Invalid TGSM register index %u.\n", register_idx);
889 return S_OK;
890 }
891 if (reg_maps->shader_version.type != WINED3D_SHADER_TYPE_COMPUTE)
892 {
893 FIXME("TGSM declarations are allowed only in compute shaders.\n");
894 return S_OK;
895 }
896
897 if (!wined3d_array_reserve((void **)&reg_maps->tgsm, &reg_maps->tgsm_capacity,
898 register_idx + 1, sizeof(*reg_maps->tgsm)))
899 return E_OUTOFMEMORY;
900
901 reg_maps->tgsm_count = max(register_idx + 1, reg_maps->tgsm_count);
902 tgsm = &reg_maps->tgsm[register_idx];
903 tgsm->size = size;
904 tgsm->stride = stride;
905 return S_OK;
906 }
907
908 static HRESULT shader_record_shader_phase(struct wined3d_shader *shader,
909 struct wined3d_shader_phase **current_phase, const struct wined3d_shader_instruction *ins,
910 const DWORD *current_instruction_ptr, const DWORD *previous_instruction_ptr)
911 {
912 struct wined3d_shader_phase *phase;
913
914 if ((phase = *current_phase))
915 {
916 phase->end = previous_instruction_ptr;
917 *current_phase = NULL;
918 }
919
920 if (shader->reg_maps.shader_version.type != WINED3D_SHADER_TYPE_HULL)
921 {
922 ERR("Unexpected shader type %#x.\n", shader->reg_maps.shader_version.type);
923 return E_FAIL;
924 }
925
926 switch (ins->handler_idx)
927 {
928 case WINED3DSIH_HS_CONTROL_POINT_PHASE:
929 if (shader->u.hs.phases.control_point)
930 {
931 FIXME("Multiple control point phases.\n");
932 HeapFree(GetProcessHeap(), 0, shader->u.hs.phases.control_point);
933 }
934 if (!(shader->u.hs.phases.control_point = HeapAlloc(GetProcessHeap(),
935 HEAP_ZERO_MEMORY, sizeof(*shader->u.hs.phases.control_point))))
936 return E_OUTOFMEMORY;
937 phase = shader->u.hs.phases.control_point;
938 break;
939 case WINED3DSIH_HS_FORK_PHASE:
940 if (!wined3d_array_reserve((void **)&shader->u.hs.phases.fork,
941 &shader->u.hs.phases.fork_size, shader->u.hs.phases.fork_count + 1,
942 sizeof(*shader->u.hs.phases.fork)))
943 return E_OUTOFMEMORY;
944 phase = &shader->u.hs.phases.fork[shader->u.hs.phases.fork_count++];
945 break;
946 case WINED3DSIH_HS_JOIN_PHASE:
947 if (!wined3d_array_reserve((void **)&shader->u.hs.phases.join,
948 &shader->u.hs.phases.join_size, shader->u.hs.phases.join_count + 1,
949 sizeof(*shader->u.hs.phases.join)))
950 return E_OUTOFMEMORY;
951 phase = &shader->u.hs.phases.join[shader->u.hs.phases.join_count++];
952 break;
953 default:
954 ERR("Unexpected opcode %s.\n", debug_d3dshaderinstructionhandler(ins->handler_idx));
955 return E_FAIL;
956 }
957
958 phase->start = current_instruction_ptr;
959 *current_phase = phase;
960
961 return WINED3D_OK;
962 }
963
964 /* Note that this does not count the loop register as an address register. */
965 static HRESULT shader_get_registers_used(struct wined3d_shader *shader, const struct wined3d_shader_frontend *fe,
966 struct wined3d_shader_reg_maps *reg_maps, struct wined3d_shader_signature *input_signature,
967 struct wined3d_shader_signature *output_signature, DWORD constf_size)
968 {
969 struct wined3d_shader_signature_element input_signature_elements[max(MAX_ATTRIBS, MAX_REG_INPUT)];
970 struct wined3d_shader_signature_element output_signature_elements[MAX_REG_OUTPUT];
971 unsigned int cur_loop_depth = 0, max_loop_depth = 0;
972 struct wined3d_shader_version shader_version;
973 struct wined3d_shader_phase *phase = NULL;
974 const DWORD *ptr, *prev_ins, *current_ins;
975 void *fe_data = shader->frontend_data;
976 unsigned int i;
977 HRESULT hr;
978
979 memset(reg_maps, 0, sizeof(*reg_maps));
980 memset(input_signature_elements, 0, sizeof(input_signature_elements));
981 memset(output_signature_elements, 0, sizeof(output_signature_elements));
982 reg_maps->min_rel_offset = ~0U;
983 list_init(&reg_maps->indexable_temps);
984
985 fe->shader_read_header(fe_data, &ptr, &shader_version);
986 prev_ins = current_ins = ptr;
987 reg_maps->shader_version = shader_version;
988
989 shader_set_limits(shader);
990
991 if (!(reg_maps->constf = wined3d_calloc(((min(shader->limits->constant_float, constf_size) + 31) / 32),
992 sizeof(*reg_maps->constf))))
993 {
994 ERR("Failed to allocate constant map memory.\n");
995 return E_OUTOFMEMORY;
996 }
997
998 while (!fe->shader_is_end(fe_data, &ptr))
999 {
1000 struct wined3d_shader_instruction ins;
1001
1002 current_ins = ptr;
1003 /* Fetch opcode. */
1004 fe->shader_read_instruction(fe_data, &ptr, &ins);
1005
1006 /* Unhandled opcode, and its parameters. */
1007 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
1008 {
1009 WARN("Encountered unrecognised or invalid instruction.\n");
1010 return WINED3DERR_INVALIDCALL;
1011 }
1012
1013 /* Handle declarations. */
1014 if (ins.handler_idx == WINED3DSIH_DCL
1015 || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
1016 {
1017 struct wined3d_shader_semantic *semantic = &ins.declaration.semantic;
1018 unsigned int reg_idx = semantic->reg.reg.idx[0].offset;
1019
1020 switch (semantic->reg.reg.type)
1021 {
1022 /* Mark input registers used. */
1023 case WINED3DSPR_INPUT:
1024 if (reg_idx >= MAX_REG_INPUT)
1025 {
1026 ERR("Invalid input register index %u.\n", reg_idx);
1027 break;
1028 }
1029 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL && shader_version.major == 3
1030 && semantic->usage == WINED3D_DECL_USAGE_POSITION && !semantic->usage_idx)
1031 return WINED3DERR_INVALIDCALL;
1032 reg_maps->input_registers |= 1u << reg_idx;
1033 shader_signature_from_semantic(&input_signature_elements[reg_idx], semantic);
1034 break;
1035
1036 /* Vertex shader: mark 3.0 output registers used, save token. */
1037 case WINED3DSPR_OUTPUT:
1038 if (reg_idx >= MAX_REG_OUTPUT)
1039 {
1040 ERR("Invalid output register index %u.\n", reg_idx);
1041 break;
1042 }
1043 reg_maps->output_registers |= 1u << reg_idx;
1044 shader_signature_from_semantic(&output_signature_elements[reg_idx], semantic);
1045 if (semantic->usage == WINED3D_DECL_USAGE_FOG)
1046 reg_maps->fog = 1;
1047 if (semantic->usage == WINED3D_DECL_USAGE_PSIZE)
1048 reg_maps->point_size = 1;
1049 break;
1050
1051 case WINED3DSPR_SAMPLER:
1052 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1053 case WINED3DSPR_RESOURCE:
1054 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1055 {
1056 ERR("Invalid resource index %u.\n", reg_idx);
1057 break;
1058 }
1059 reg_maps->resource_info[reg_idx].type = semantic->resource_type;
1060 reg_maps->resource_info[reg_idx].data_type = semantic->resource_data_type;
1061 break;
1062
1063 case WINED3DSPR_UAV:
1064 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1065 {
1066 ERR("Invalid UAV resource index %u.\n", reg_idx);
1067 break;
1068 }
1069 reg_maps->uav_resource_info[reg_idx].type = semantic->resource_type;
1070 reg_maps->uav_resource_info[reg_idx].data_type = semantic->resource_data_type;
1071 if (ins.flags)
1072 FIXME("Ignoring typed UAV flags %#x.\n", ins.flags);
1073 break;
1074
1075 default:
1076 TRACE("Not recording DCL register type %#x.\n", semantic->reg.reg.type);
1077 break;
1078 }
1079 }
1080 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
1081 {
1082 struct wined3d_shader_register *reg = &ins.declaration.src.reg;
1083 if (reg->idx[0].offset >= WINED3D_MAX_CBS)
1084 ERR("Invalid CB index %u.\n", reg->idx[0].offset);
1085 else
1086 reg_maps->cb_sizes[reg->idx[0].offset] = reg->idx[1].offset;
1087 }
1088 else if (ins.handler_idx == WINED3DSIH_DCL_GS_INSTANCES)
1089 {
1090 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1091 shader->u.gs.instance_count = ins.declaration.count;
1092 else
1093 FIXME("Invalid instruction %#x for shader type %#x.\n",
1094 ins.handler_idx, shader_version.type);
1095 }
1096 else if (ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
1097 || ins.handler_idx == WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT)
1098 {
1099 if (phase)
1100 phase->instance_count = ins.declaration.count;
1101 else
1102 FIXME("Instruction %s outside of shader phase.\n",
1103 debug_d3dshaderinstructionhandler(ins.handler_idx));
1104 }
1105 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
1106 {
1107 if (reg_maps->icb)
1108 FIXME("Multiple immediate constant buffers.\n");
1109 reg_maps->icb = ins.declaration.icb;
1110 }
1111 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
1112 {
1113 if (phase)
1114 {
1115 FIXME("Indexable temporary registers not supported.\n");
1116 }
1117 else
1118 {
1119 struct wined3d_shader_indexable_temp *reg;
1120
1121 if (!(reg = HeapAlloc(GetProcessHeap(), 0, sizeof(*reg))))
1122 return E_OUTOFMEMORY;
1123
1124 *reg = ins.declaration.indexable_temp;
1125 list_add_tail(&reg_maps->indexable_temps, &reg->entry);
1126 }
1127 }
1128 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE)
1129 {
1130 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1131 shader->u.gs.input_type = ins.declaration.primitive_type.type;
1132 else
1133 FIXME("Invalid instruction %#x for shader type %#x.\n",
1134 ins.handler_idx, shader_version.type);
1135 }
1136 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT)
1137 {
1138 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1139 shader->u.hs.output_vertex_count = ins.declaration.count;
1140 else
1141 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1142 }
1143 else if (ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
1144 {
1145 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1146 shader->u.gs.output_type = ins.declaration.primitive_type.type;
1147 else
1148 FIXME("Invalid instruction %#x for shader type %#x.\n",
1149 ins.handler_idx, shader_version.type);
1150 }
1151 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_RAW)
1152 {
1153 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1154 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1155 {
1156 ERR("Invalid resource index %u.\n", reg_idx);
1157 break;
1158 }
1159 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1160 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1161 reg_maps->resource_info[reg_idx].flags = WINED3D_VIEW_BUFFER_RAW;
1162 }
1163 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
1164 {
1165 unsigned int reg_idx = ins.declaration.structured_resource.reg.reg.idx[0].offset;
1166 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1167 {
1168 ERR("Invalid resource index %u.\n", reg_idx);
1169 break;
1170 }
1171 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1172 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1173 reg_maps->resource_info[reg_idx].flags = 0;
1174 reg_maps->resource_info[reg_idx].stride = ins.declaration.structured_resource.byte_stride / 4;
1175 }
1176 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
1177 {
1178 if (ins.flags & WINED3DSI_SAMPLER_COMPARISON_MODE)
1179 reg_maps->sampler_comparison_mode |= (1u << ins.declaration.dst.reg.idx[0].offset);
1180 }
1181 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS)
1182 {
1183 if (phase)
1184 phase->temporary_count = ins.declaration.count;
1185 else
1186 reg_maps->temporary_count = ins.declaration.count;
1187 }
1188 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
1189 {
1190 if (shader_version.type == WINED3D_SHADER_TYPE_DOMAIN)
1191 shader->u.ds.tessellator_domain = ins.declaration.tessellator_domain;
1192 else if (shader_version.type != WINED3D_SHADER_TYPE_HULL)
1193 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1194 }
1195 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
1196 {
1197 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1198 shader->u.hs.tessellator_output_primitive = ins.declaration.tessellator_output_primitive;
1199 else
1200 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1201 }
1202 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
1203 {
1204 if (shader_version.type == WINED3D_SHADER_TYPE_HULL)
1205 shader->u.hs.tessellator_partitioning = ins.declaration.tessellator_partitioning;
1206 else
1207 FIXME("Invalid instruction %#x for shader type %#x.\n", ins.handler_idx, shader_version.type);
1208 }
1209 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
1210 {
1211 if (FAILED(hr = shader_reg_maps_add_tgsm(reg_maps, ins.declaration.tgsm_raw.reg.reg.idx[0].offset,
1212 ins.declaration.tgsm_raw.byte_count / 4, 0)))
1213 return hr;
1214 }
1215 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
1216 {
1217 unsigned int stride = ins.declaration.tgsm_structured.byte_stride / 4;
1218 unsigned int size = stride * ins.declaration.tgsm_structured.structure_count;
1219 if (FAILED(hr = shader_reg_maps_add_tgsm(reg_maps,
1220 ins.declaration.tgsm_structured.reg.reg.idx[0].offset, size, stride)))
1221 return hr;
1222 }
1223 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
1224 {
1225 if (shader_version.type == WINED3D_SHADER_TYPE_COMPUTE)
1226 {
1227 shader->u.cs.thread_group_size = ins.declaration.thread_group_size;
1228 }
1229 else
1230 {
1231 FIXME("Invalid instruction %#x for shader type %#x.\n",
1232 ins.handler_idx, shader_version.type);
1233 }
1234 }
1235 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
1236 {
1237 unsigned int reg_idx = ins.declaration.dst.reg.idx[0].offset;
1238 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1239 {
1240 ERR("Invalid UAV resource index %u.\n", reg_idx);
1241 break;
1242 }
1243 if (ins.flags)
1244 FIXME("Ignoring raw UAV flags %#x.\n", ins.flags);
1245 reg_maps->uav_resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1246 reg_maps->uav_resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1247 reg_maps->uav_resource_info[reg_idx].flags = WINED3D_VIEW_BUFFER_RAW;
1248 }
1249 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
1250 {
1251 unsigned int reg_idx = ins.declaration.structured_resource.reg.reg.idx[0].offset;
1252 if (reg_idx >= ARRAY_SIZE(reg_maps->uav_resource_info))
1253 {
1254 ERR("Invalid UAV resource index %u.\n", reg_idx);
1255 break;
1256 }
1257 if (ins.flags)
1258 FIXME("Ignoring structured UAV flags %#x.\n", ins.flags);
1259 reg_maps->uav_resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_BUFFER;
1260 reg_maps->uav_resource_info[reg_idx].data_type = WINED3D_DATA_UINT;
1261 reg_maps->uav_resource_info[reg_idx].flags = 0;
1262 reg_maps->uav_resource_info[reg_idx].stride = ins.declaration.structured_resource.byte_stride / 4;
1263 }
1264 else if (ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
1265 {
1266 if (shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
1267 shader->u.gs.vertices_out = ins.declaration.count;
1268 else
1269 FIXME("Invalid instruction %#x for shader type %#x.\n",
1270 ins.handler_idx, shader_version.type);
1271 }
1272 else if (ins.handler_idx == WINED3DSIH_DEF)
1273 {
1274 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1275 float *value;
1276 if (!lconst) return E_OUTOFMEMORY;
1277
1278 lconst->idx = ins.dst[0].reg.idx[0].offset;
1279 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1280 value = (float *)lconst->value;
1281
1282 /* In pixel shader 1.X shaders, the constants are clamped between [-1;1] */
1283 if (shader_version.major == 1 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1284 {
1285 if (value[0] < -1.0f) value[0] = -1.0f;
1286 else if (value[0] > 1.0f) value[0] = 1.0f;
1287 if (value[1] < -1.0f) value[1] = -1.0f;
1288 else if (value[1] > 1.0f) value[1] = 1.0f;
1289 if (value[2] < -1.0f) value[2] = -1.0f;
1290 else if (value[2] > 1.0f) value[2] = 1.0f;
1291 if (value[3] < -1.0f) value[3] = -1.0f;
1292 else if (value[3] > 1.0f) value[3] = 1.0f;
1293 }
1294
1295 list_add_head(&shader->constantsF, &lconst->entry);
1296
1297 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
1298 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
1299 {
1300 shader->lconst_inf_or_nan = TRUE;
1301 }
1302 }
1303 else if (ins.handler_idx == WINED3DSIH_DEFI)
1304 {
1305 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1306 if (!lconst) return E_OUTOFMEMORY;
1307
1308 lconst->idx = ins.dst[0].reg.idx[0].offset;
1309 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, 4 * sizeof(DWORD));
1310
1311 list_add_head(&shader->constantsI, &lconst->entry);
1312 reg_maps->local_int_consts |= (1u << lconst->idx);
1313 }
1314 else if (ins.handler_idx == WINED3DSIH_DEFB)
1315 {
1316 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
1317 if (!lconst) return E_OUTOFMEMORY;
1318
1319 lconst->idx = ins.dst[0].reg.idx[0].offset;
1320 memcpy(lconst->value, ins.src[0].reg.u.immconst_data, sizeof(DWORD));
1321
1322 list_add_head(&shader->constantsB, &lconst->entry);
1323 reg_maps->local_bool_consts |= (1u << lconst->idx);
1324 }
1325 /* Handle shader phases. */
1326 else if (ins.handler_idx == WINED3DSIH_HS_CONTROL_POINT_PHASE
1327 || ins.handler_idx == WINED3DSIH_HS_FORK_PHASE
1328 || ins.handler_idx == WINED3DSIH_HS_JOIN_PHASE)
1329 {
1330 if (FAILED(hr = shader_record_shader_phase(shader, &phase, &ins, current_ins, prev_ins)))
1331 return hr;
1332 }
1333 /* For subroutine prototypes. */
1334 else if (ins.handler_idx == WINED3DSIH_LABEL)
1335 {
1336 reg_maps->labels |= 1u << ins.src[0].reg.idx[0].offset;
1337 }
1338 /* Set texture, address, temporary registers. */
1339 else
1340 {
1341 BOOL color0_mov = FALSE;
1342 unsigned int i;
1343
1344 /* This will loop over all the registers and try to
1345 * make a bitmask of the ones we're interested in.
1346 *
1347 * Relative addressing tokens are ignored, but that's
1348 * okay, since we'll catch any address registers when
1349 * they are initialized (required by spec). */
1350 for (i = 0; i < ins.dst_count; ++i)
1351 {
1352 if (!shader_record_register_usage(shader, reg_maps, &ins.dst[i].reg,
1353 shader_version.type, constf_size))
1354 return WINED3DERR_INVALIDCALL;
1355
1356 if (shader_version.type == WINED3D_SHADER_TYPE_VERTEX)
1357 {
1358 UINT idx = ins.dst[i].reg.idx[0].offset;
1359
1360 switch (ins.dst[i].reg.type)
1361 {
1362 case WINED3DSPR_RASTOUT:
1363 if (shader_version.major >= 3)
1364 break;
1365 switch (idx)
1366 {
1367 case 0: /* oPos */
1368 reg_maps->output_registers |= 1u << 10;
1369 shader_signature_from_usage(&output_signature_elements[10],
1370 WINED3D_DECL_USAGE_POSITION, 0, 10, WINED3DSP_WRITEMASK_ALL);
1371 break;
1372
1373 case 1: /* oFog */
1374 reg_maps->output_registers |= 1u << 11;
1375 shader_signature_from_usage(&output_signature_elements[11],
1376 WINED3D_DECL_USAGE_FOG, 0, 11, WINED3DSP_WRITEMASK_0);
1377 break;
1378
1379 case 2: /* oPts */
1380 reg_maps->output_registers |= 1u << 11;
1381 shader_signature_from_usage(&output_signature_elements[11],
1382 WINED3D_DECL_USAGE_PSIZE, 0, 11, WINED3DSP_WRITEMASK_1);
1383 break;
1384 }
1385 break;
1386
1387 case WINED3DSPR_ATTROUT:
1388 if (shader_version.major >= 3)
1389 break;
1390 if (idx < 2)
1391 {
1392 idx += 8;
1393 if (reg_maps->output_registers & (1u << idx))
1394 {
1395 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1396 }
1397 else
1398 {
1399 reg_maps->output_registers |= 1u << idx;
1400 shader_signature_from_usage(&output_signature_elements[idx],
1401 WINED3D_DECL_USAGE_COLOR, idx - 8, idx, ins.dst[i].write_mask);
1402 }
1403 }
1404 break;
1405
1406 case WINED3DSPR_TEXCRDOUT: /* WINED3DSPR_OUTPUT */
1407 if (shader_version.major >= 3)
1408 {
1409 if (idx >= ARRAY_SIZE(reg_maps->u.output_registers_mask))
1410 {
1411 WARN("Invalid output register index %u.\n", idx);
1412 break;
1413 }
1414 reg_maps->u.output_registers_mask[idx] |= ins.dst[i].write_mask;
1415 break;
1416 }
1417 if (idx >= ARRAY_SIZE(reg_maps->u.texcoord_mask))
1418 {
1419 WARN("Invalid texcoord index %u.\n", idx);
1420 break;
1421 }
1422 reg_maps->u.texcoord_mask[idx] |= ins.dst[i].write_mask;
1423 if (reg_maps->output_registers & (1u << idx))
1424 {
1425 output_signature_elements[idx].mask |= ins.dst[i].write_mask;
1426 }
1427 else
1428 {
1429 reg_maps->output_registers |= 1u << idx;
1430 shader_signature_from_usage(&output_signature_elements[idx],
1431 WINED3D_DECL_USAGE_TEXCOORD, idx, idx, ins.dst[i].write_mask);
1432 }
1433 break;
1434
1435 default:
1436 break;
1437 }
1438 }
1439
1440 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1441 {
1442 if (ins.dst[i].reg.type == WINED3DSPR_COLOROUT && !ins.dst[i].reg.idx[0].offset)
1443 {
1444 /* Many 2.0 and 3.0 pixel shaders end with a MOV from a temp register to
1445 * COLOROUT 0. If we know this in advance, the ARB shader backend can skip
1446 * the mov and perform the sRGB write correction from the source register.
1447 *
1448 * However, if the mov is only partial, we can't do this, and if the write
1449 * comes from an instruction other than MOV it is hard to do as well. If
1450 * COLOROUT 0 is overwritten partially later, the marker is dropped again. */
1451 shader->u.ps.color0_mov = FALSE;
1452 if (ins.handler_idx == WINED3DSIH_MOV
1453 && ins.dst[i].write_mask == WINED3DSP_WRITEMASK_ALL)
1454 {
1455 /* Used later when the source register is read. */
1456 color0_mov = TRUE;
1457 }
1458 }
1459 /* Also drop the MOV marker if the source register is overwritten prior to the shader
1460 * end
1461 */
1462 else if (ins.dst[i].reg.type == WINED3DSPR_TEMP
1463 && ins.dst[i].reg.idx[0].offset == shader->u.ps.color0_reg)
1464 {
1465 shader->u.ps.color0_mov = FALSE;
1466 }
1467 }
1468
1469 /* Declare 1.x samplers implicitly, based on the destination reg. number. */
1470 if (shader_version.major == 1
1471 && (ins.handler_idx == WINED3DSIH_TEX
1472 || ins.handler_idx == WINED3DSIH_TEXBEM
1473 || ins.handler_idx == WINED3DSIH_TEXBEML
1474 || ins.handler_idx == WINED3DSIH_TEXDP3TEX
1475 || ins.handler_idx == WINED3DSIH_TEXM3x2TEX
1476 || ins.handler_idx == WINED3DSIH_TEXM3x3SPEC
1477 || ins.handler_idx == WINED3DSIH_TEXM3x3TEX
1478 || ins.handler_idx == WINED3DSIH_TEXM3x3VSPEC
1479 || ins.handler_idx == WINED3DSIH_TEXREG2AR
1480 || ins.handler_idx == WINED3DSIH_TEXREG2GB
1481 || ins.handler_idx == WINED3DSIH_TEXREG2RGB))
1482 {
1483 unsigned int reg_idx = ins.dst[i].reg.idx[0].offset;
1484
1485 if (reg_idx >= ARRAY_SIZE(reg_maps->resource_info))
1486 {
1487 WARN("Invalid 1.x sampler index %u.\n", reg_idx);
1488 continue;
1489 }
1490
1491 TRACE("Setting fake 2D resource for 1.x pixelshader.\n");
1492 reg_maps->resource_info[reg_idx].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
1493 reg_maps->resource_info[reg_idx].data_type = WINED3D_DATA_FLOAT;
1494 shader_record_sample(reg_maps, reg_idx, reg_idx, reg_idx);
1495
1496 /* texbem is only valid with < 1.4 pixel shaders */
1497 if (ins.handler_idx == WINED3DSIH_TEXBEM
1498 || ins.handler_idx == WINED3DSIH_TEXBEML)
1499 {
1500 reg_maps->bumpmat |= 1u << reg_idx;
1501 if (ins.handler_idx == WINED3DSIH_TEXBEML)
1502 {
1503 reg_maps->luminanceparams |= 1u << reg_idx;
1504 }
1505 }
1506 }
1507 else if (ins.handler_idx == WINED3DSIH_BEM)
1508 {
1509 reg_maps->bumpmat |= 1u << ins.dst[i].reg.idx[0].offset;
1510 }
1511 }
1512
1513 if (ins.handler_idx == WINED3DSIH_IMM_ATOMIC_ALLOC || ins.handler_idx == WINED3DSIH_IMM_ATOMIC_CONSUME)
1514 {
1515 unsigned int reg_idx = ins.src[0].reg.idx[0].offset;
1516 if (reg_idx >= MAX_UNORDERED_ACCESS_VIEWS)
1517 {
1518 ERR("Invalid UAV index %u.\n", reg_idx);
1519 break;
1520 }
1521 reg_maps->uav_counter_mask |= (1u << reg_idx);
1522 }
1523 else if ((WINED3DSIH_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_ATOMIC_XOR)
1524 || (WINED3DSIH_IMM_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_IMM_ATOMIC_XOR)
1525 || (ins.handler_idx == WINED3DSIH_BUFINFO && ins.src[0].reg.type == WINED3DSPR_UAV)
1526 || ins.handler_idx == WINED3DSIH_LD_UAV_TYPED
1527 || (ins.handler_idx == WINED3DSIH_LD_RAW && ins.src[1].reg.type == WINED3DSPR_UAV)
1528 || (ins.handler_idx == WINED3DSIH_LD_STRUCTURED && ins.src[2].reg.type == WINED3DSPR_UAV))
1529 {
1530 unsigned int reg_idx;
1531 if (ins.handler_idx == WINED3DSIH_LD_UAV_TYPED || ins.handler_idx == WINED3DSIH_LD_RAW)
1532 reg_idx = ins.src[1].reg.idx[0].offset;
1533 else if (ins.handler_idx == WINED3DSIH_LD_STRUCTURED)
1534 reg_idx = ins.src[2].reg.idx[0].offset;
1535 else if (WINED3DSIH_ATOMIC_AND <= ins.handler_idx && ins.handler_idx <= WINED3DSIH_ATOMIC_XOR)
1536 reg_idx = ins.dst[0].reg.idx[0].offset;
1537 else if (ins.handler_idx == WINED3DSIH_BUFINFO)
1538 reg_idx = ins.src[0].reg.idx[0].offset;
1539 else
1540 reg_idx = ins.dst[1].reg.idx[0].offset;
1541 if (reg_idx >= MAX_UNORDERED_ACCESS_VIEWS)
1542 {
1543 ERR("Invalid UAV index %u.\n", reg_idx);
1544 break;
1545 }
1546 reg_maps->uav_read_mask |= (1u << reg_idx);
1547 }
1548 else if (ins.handler_idx == WINED3DSIH_NRM)
1549 {
1550 reg_maps->usesnrm = 1;
1551 }
1552 else if (ins.handler_idx == WINED3DSIH_DSY
1553 || ins.handler_idx == WINED3DSIH_DSY_COARSE
1554 || ins.handler_idx == WINED3DSIH_DSY_FINE)
1555 {
1556 reg_maps->usesdsy = 1;
1557 }
1558 else if (ins.handler_idx == WINED3DSIH_DSX
1559 || ins.handler_idx == WINED3DSIH_DSX_COARSE
1560 || ins.handler_idx == WINED3DSIH_DSX_FINE)
1561 {
1562 reg_maps->usesdsx = 1;
1563 }
1564 else if (ins.handler_idx == WINED3DSIH_TEXLDD) reg_maps->usestexldd = 1;
1565 else if (ins.handler_idx == WINED3DSIH_TEXLDL) reg_maps->usestexldl = 1;
1566 else if (ins.handler_idx == WINED3DSIH_MOVA) reg_maps->usesmova = 1;
1567 else if (ins.handler_idx == WINED3DSIH_IFC) reg_maps->usesifc = 1;
1568 else if (ins.handler_idx == WINED3DSIH_CALL) reg_maps->usescall = 1;
1569 else if (ins.handler_idx == WINED3DSIH_POW) reg_maps->usespow = 1;
1570 else if (ins.handler_idx == WINED3DSIH_LOOP
1571 || ins.handler_idx == WINED3DSIH_REP)
1572 {
1573 ++cur_loop_depth;
1574 if (cur_loop_depth > max_loop_depth)
1575 max_loop_depth = cur_loop_depth;
1576 }
1577 else if (ins.handler_idx == WINED3DSIH_ENDLOOP
1578 || ins.handler_idx == WINED3DSIH_ENDREP)
1579 {
1580 --cur_loop_depth;
1581 }
1582 else if (ins.handler_idx == WINED3DSIH_GATHER4
1583 || ins.handler_idx == WINED3DSIH_GATHER4_C
1584 || ins.handler_idx == WINED3DSIH_SAMPLE
1585 || ins.handler_idx == WINED3DSIH_SAMPLE_B
1586 || ins.handler_idx == WINED3DSIH_SAMPLE_C
1587 || ins.handler_idx == WINED3DSIH_SAMPLE_C_LZ
1588 || ins.handler_idx == WINED3DSIH_SAMPLE_GRAD
1589 || ins.handler_idx == WINED3DSIH_SAMPLE_LOD)
1590 {
1591 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1592 ins.src[2].reg.idx[0].offset, reg_maps->sampler_map.count);
1593 }
1594 else if (ins.handler_idx == WINED3DSIH_GATHER4_PO
1595 || ins.handler_idx == WINED3DSIH_GATHER4_PO_C)
1596 {
1597 shader_record_sample(reg_maps, ins.src[2].reg.idx[0].offset,
1598 ins.src[3].reg.idx[0].offset, reg_maps->sampler_map.count);
1599 }
1600 else if (ins.handler_idx == WINED3DSIH_BUFINFO && ins.src[0].reg.type == WINED3DSPR_RESOURCE)
1601 {
1602 shader_record_sample(reg_maps, ins.src[0].reg.idx[0].offset,
1603 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1604 }
1605 else if (ins.handler_idx == WINED3DSIH_LD
1606 || (ins.handler_idx == WINED3DSIH_LD_RAW && ins.src[1].reg.type == WINED3DSPR_RESOURCE)
1607 || (ins.handler_idx == WINED3DSIH_RESINFO && ins.src[1].reg.type == WINED3DSPR_RESOURCE))
1608 {
1609 shader_record_sample(reg_maps, ins.src[1].reg.idx[0].offset,
1610 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1611 }
1612 else if (ins.handler_idx == WINED3DSIH_LD_STRUCTURED
1613 && ins.src[2].reg.type == WINED3DSPR_RESOURCE)
1614 {
1615 shader_record_sample(reg_maps, ins.src[2].reg.idx[0].offset,
1616 WINED3D_SAMPLER_DEFAULT, reg_maps->sampler_map.count);
1617 }
1618
1619 if (ins.predicate)
1620 if (!shader_record_register_usage(shader, reg_maps, &ins.predicate->reg,
1621 shader_version.type, constf_size))
1622 return WINED3DERR_INVALIDCALL;
1623
1624 for (i = 0; i < ins.src_count; ++i)
1625 {
1626 unsigned int count = get_instr_extra_regcount(ins.handler_idx, i);
1627 struct wined3d_shader_register reg = ins.src[i].reg;
1628
1629 if (!shader_record_register_usage(shader, reg_maps, &ins.src[i].reg,
1630 shader_version.type, constf_size))
1631 return WINED3DERR_INVALIDCALL;
1632 while (count)
1633 {
1634 ++reg.idx[0].offset;
1635 if (!shader_record_register_usage(shader, reg_maps, &reg,
1636 shader_version.type, constf_size))
1637 return WINED3DERR_INVALIDCALL;
1638 --count;
1639 }
1640
1641 if (color0_mov)
1642 {
1643 if (ins.src[i].reg.type == WINED3DSPR_TEMP
1644 && ins.src[i].swizzle == WINED3DSP_NOSWIZZLE)
1645 {
1646 shader->u.ps.color0_mov = TRUE;
1647 shader->u.ps.color0_reg = ins.src[i].reg.idx[0].offset;
1648 }
1649 }
1650 }
1651 }
1652
1653 prev_ins = current_ins;
1654 }
1655 reg_maps->loop_depth = max_loop_depth;
1656
1657 if (phase)
1658 {
1659 phase->end = prev_ins;
1660 phase = NULL;
1661 }
1662
1663 /* PS before 2.0 don't have explicit color outputs. Instead the value of
1664 * R0 is written to the render target. */
1665 if (shader_version.major < 2 && shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1666 reg_maps->rt_mask |= (1u << 0);
1667
1668 if (input_signature->elements)
1669 {
1670 for (i = 0; i < input_signature->element_count; ++i)
1671 {
1672 reg_maps->input_registers |= 1u << input_signature->elements[i].register_idx;
1673 if (shader_version.type == WINED3D_SHADER_TYPE_PIXEL)
1674 {
1675 if (input_signature->elements[i].sysval_semantic == WINED3D_SV_POSITION)
1676 reg_maps->vpos = 1;
1677 else if (input_signature->elements[i].sysval_semantic == WINED3D_SV_IS_FRONT_FACE)
1678 reg_maps->usesfacing = 1;
1679 }
1680 }
1681 }
1682 else if (!input_signature->elements && reg_maps->input_registers)
1683 {
1684 unsigned int count = wined3d_popcount(reg_maps->input_registers);
1685 struct wined3d_shader_signature_element *e;
1686 unsigned int i;
1687
1688 if (!(input_signature->elements = wined3d_calloc(count, sizeof(*input_signature->elements))))
1689 return E_OUTOFMEMORY;
1690 input_signature->element_count = count;
1691
1692 e = input_signature->elements;
1693 for (i = 0; i < ARRAY_SIZE(input_signature_elements); ++i)
1694 {
1695 if (!(reg_maps->input_registers & (1u << i)))
1696 continue;
1697 input_signature_elements[i].register_idx = i;
1698 *e++ = input_signature_elements[i];
1699 }
1700 }
1701
1702 if (output_signature->elements)
1703 {
1704 for (i = 0; i < output_signature->element_count; ++i)
1705 {
1706 reg_maps->output_registers |= 1u << output_signature->elements[i].register_idx;
1707 }
1708 }
1709 else if (reg_maps->output_registers)
1710 {
1711 unsigned int count = wined3d_popcount(reg_maps->output_registers);
1712 struct wined3d_shader_signature_element *e;
1713
1714 if (!(output_signature->elements = wined3d_calloc(count, sizeof(*output_signature->elements))))
1715 return E_OUTOFMEMORY;
1716 output_signature->element_count = count;
1717
1718 e = output_signature->elements;
1719 for (i = 0; i < ARRAY_SIZE(output_signature_elements); ++i)
1720 {
1721 if (!(reg_maps->output_registers & (1u << i)))
1722 continue;
1723 *e++ = output_signature_elements[i];
1724 }
1725 }
1726
1727 return WINED3D_OK;
1728 }
1729
1730 static void shader_cleanup_reg_maps(struct wined3d_shader_reg_maps *reg_maps)
1731 {
1732 struct wined3d_shader_indexable_temp *reg, *reg_next;
1733
1734 HeapFree(GetProcessHeap(), 0, reg_maps->constf);
1735 HeapFree(GetProcessHeap(), 0, reg_maps->sampler_map.entries);
1736
1737 LIST_FOR_EACH_ENTRY_SAFE(reg, reg_next, &reg_maps->indexable_temps, struct wined3d_shader_indexable_temp, entry)
1738 HeapFree(GetProcessHeap(), 0, reg);
1739 list_init(&reg_maps->indexable_temps);
1740
1741 HeapFree(GetProcessHeap(), 0, reg_maps->tgsm);
1742 }
1743
1744 unsigned int shader_find_free_input_register(const struct wined3d_shader_reg_maps *reg_maps, unsigned int max)
1745 {
1746 DWORD map = 1u << max;
1747 map |= map - 1;
1748 map &= reg_maps->shader_version.major < 3 ? ~reg_maps->texcoord : ~reg_maps->input_registers;
1749
1750 return wined3d_log2i(map);
1751 }
1752
1753 static void shader_dump_global_flags(struct wined3d_string_buffer *buffer, DWORD global_flags)
1754 {
1755 if (global_flags & WINED3DSGF_REFACTORING_ALLOWED)
1756 {
1757 shader_addline(buffer, "refactoringAllowed");
1758 global_flags &= ~WINED3DSGF_REFACTORING_ALLOWED;
1759 if (global_flags)
1760 shader_addline(buffer, " | ");
1761 }
1762
1763 if (global_flags & WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS)
1764 {
1765 shader_addline(buffer, "enableRawAndStructuredBuffers");
1766 global_flags &= ~WINED3DSGF_ENABLE_RAW_AND_STRUCTURED_BUFFERS;
1767 }
1768
1769 if (global_flags)
1770 shader_addline(buffer, "unknown_flags(%#x)", global_flags);
1771 }
1772
1773 static void shader_dump_sync_flags(struct wined3d_string_buffer *buffer, DWORD sync_flags)
1774 {
1775 if (sync_flags & WINED3DSSF_GROUP_SHARED_MEMORY)
1776 {
1777 shader_addline(buffer, "_g");
1778 sync_flags &= ~WINED3DSSF_GROUP_SHARED_MEMORY;
1779 }
1780 if (sync_flags & WINED3DSSF_THREAD_GROUP)
1781 {
1782 shader_addline(buffer, "_t");
1783 sync_flags &= ~WINED3DSSF_THREAD_GROUP;
1784 }
1785
1786 if (sync_flags)
1787 shader_addline(buffer, "_unknown_flags(%#x)", sync_flags);
1788 }
1789
1790 static void shader_dump_uav_flags(struct wined3d_string_buffer *buffer, DWORD uav_flags)
1791 {
1792 if (uav_flags & WINED3DSUF_GLOBALLY_COHERENT)
1793 {
1794 shader_addline(buffer, "_glc");
1795 uav_flags &= ~WINED3DSUF_GLOBALLY_COHERENT;
1796 }
1797 if (uav_flags & WINED3DSUF_ORDER_PRESERVING_COUNTER)
1798 {
1799 shader_addline(buffer, "_opc");
1800 uav_flags &= ~WINED3DSUF_ORDER_PRESERVING_COUNTER;
1801 }
1802
1803 if (uav_flags)
1804 shader_addline(buffer, "_unknown_flags(%#x)", uav_flags);
1805 }
1806
1807 static void shader_dump_tessellator_domain(struct wined3d_string_buffer *buffer,
1808 enum wined3d_tessellator_domain domain)
1809 {
1810 switch (domain)
1811 {
1812 case WINED3D_TESSELLATOR_DOMAIN_LINE:
1813 shader_addline(buffer, "line");
1814 break;
1815 case WINED3D_TESSELLATOR_DOMAIN_TRIANGLE:
1816 shader_addline(buffer, "triangle");
1817 break;
1818 case WINED3D_TESSELLATOR_DOMAIN_QUAD:
1819 shader_addline(buffer, "quad");
1820 break;
1821 default:
1822 shader_addline(buffer, "unknown_tessellator_domain(%#x)", domain);
1823 break;
1824 }
1825 }
1826
1827 static void shader_dump_tessellator_output_primitive(struct wined3d_string_buffer *buffer,
1828 enum wined3d_tessellator_output_primitive output_primitive)
1829 {
1830 switch (output_primitive)
1831 {
1832 case WINED3D_TESSELLATOR_OUTPUT_POINT:
1833 shader_addline(buffer, "point");
1834 break;
1835 case WINED3D_TESSELLATOR_OUTPUT_LINE:
1836 shader_addline(buffer, "line");
1837 break;
1838 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CW:
1839 shader_addline(buffer, "triangle_cw");
1840 break;
1841 case WINED3D_TESSELLATOR_OUTPUT_TRIANGLE_CCW:
1842 shader_addline(buffer, "triangle_ccw");
1843 break;
1844 default:
1845 shader_addline(buffer, "unknown_tessellator_output_primitive(%#x)", output_primitive);
1846 break;
1847 }
1848 }
1849
1850 static void shader_dump_tessellator_partitioning(struct wined3d_string_buffer *buffer,
1851 enum wined3d_tessellator_partitioning partitioning)
1852 {
1853 switch (partitioning)
1854 {
1855 case WINED3D_TESSELLATOR_PARTITIONING_INTEGER:
1856 shader_addline(buffer, "integer");
1857 break;
1858 case WINED3D_TESSELLATOR_PARTITIONING_POW2:
1859 shader_addline(buffer, "pow2");
1860 break;
1861 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_ODD:
1862 shader_addline(buffer, "fractional_odd");
1863 break;
1864 case WINED3D_TESSELLATOR_PARTITIONING_FRACTIONAL_EVEN:
1865 shader_addline(buffer, "fractional_even");
1866 break;
1867 default:
1868 shader_addline(buffer, "unknown_tessellator_partitioning(%#x)", partitioning);
1869 break;
1870 }
1871 }
1872
1873 static void shader_dump_shader_input_sysval_semantic(struct wined3d_string_buffer *buffer,
1874 enum wined3d_shader_input_sysval_semantic semantic)
1875 {
1876 unsigned int i;
1877
1878 for (i = 0; i < ARRAY_SIZE(shader_input_sysval_semantic_names); ++i)
1879 {
1880 if (shader_input_sysval_semantic_names[i].sysval_semantic == semantic)
1881 {
1882 shader_addline(buffer, "%s", shader_input_sysval_semantic_names[i].sysval_name);
1883 return;
1884 }
1885 }
1886
1887 shader_addline(buffer, "unknown_shader_input_sysval_semantic(%#x)", semantic);
1888 }
1889
1890 static void shader_dump_decl_usage(struct wined3d_string_buffer *buffer,
1891 const struct wined3d_shader_semantic *semantic, unsigned int flags,
1892 const struct wined3d_shader_version *shader_version)
1893 {
1894 shader_addline(buffer, "dcl");
1895
1896 if (semantic->reg.reg.type == WINED3DSPR_SAMPLER)
1897 {
1898 switch (semantic->resource_type)
1899 {
1900 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1901 shader_addline(buffer, "_2d");
1902 break;
1903
1904 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1905 shader_addline(buffer, "_3d");
1906 break;
1907
1908 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1909 shader_addline(buffer, "_cube");
1910 break;
1911
1912 default:
1913 shader_addline(buffer, "_unknown_resource_type(%#x)", semantic->resource_type);
1914 break;
1915 }
1916 }
1917 else if (semantic->reg.reg.type == WINED3DSPR_RESOURCE || semantic->reg.reg.type == WINED3DSPR_UAV)
1918 {
1919 if (semantic->reg.reg.type == WINED3DSPR_RESOURCE)
1920 shader_addline(buffer, "_resource_");
1921 else
1922 shader_addline(buffer, "_uav_");
1923 switch (semantic->resource_type)
1924 {
1925 case WINED3D_SHADER_RESOURCE_BUFFER:
1926 shader_addline(buffer, "buffer");
1927 break;
1928
1929 case WINED3D_SHADER_RESOURCE_TEXTURE_1D:
1930 shader_addline(buffer, "texture1d");
1931 break;
1932
1933 case WINED3D_SHADER_RESOURCE_TEXTURE_2D:
1934 shader_addline(buffer, "texture2d");
1935 break;
1936
1937 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMS:
1938 shader_addline(buffer, "texture2dms");
1939 break;
1940
1941 case WINED3D_SHADER_RESOURCE_TEXTURE_3D:
1942 shader_addline(buffer, "texture3d");
1943 break;
1944
1945 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBE:
1946 shader_addline(buffer, "texturecube");
1947 break;
1948
1949 case WINED3D_SHADER_RESOURCE_TEXTURE_1DARRAY:
1950 shader_addline(buffer, "texture1darray");
1951 break;
1952
1953 case WINED3D_SHADER_RESOURCE_TEXTURE_2DARRAY:
1954 shader_addline(buffer, "texture2darray");
1955 break;
1956
1957 case WINED3D_SHADER_RESOURCE_TEXTURE_2DMSARRAY:
1958 shader_addline(buffer, "texture2dmsarray");
1959 break;
1960
1961 case WINED3D_SHADER_RESOURCE_TEXTURE_CUBEARRAY:
1962 shader_addline(buffer, "texturecubearray");
1963 break;
1964
1965 default:
1966 shader_addline(buffer, "unknown");
1967 break;
1968 }
1969 if (semantic->reg.reg.type == WINED3DSPR_UAV)
1970 shader_dump_uav_flags(buffer, flags);
1971 switch (semantic->resource_data_type)
1972 {
1973 case WINED3D_DATA_FLOAT:
1974 shader_addline(buffer, " (float)");
1975 break;
1976
1977 case WINED3D_DATA_INT:
1978 shader_addline(buffer, " (int)");
1979 break;
1980
1981 case WINED3D_DATA_UINT:
1982 shader_addline(buffer, " (uint)");
1983 break;
1984
1985 case WINED3D_DATA_UNORM:
1986 shader_addline(buffer, " (unorm)");
1987 break;
1988
1989 case WINED3D_DATA_SNORM:
1990 shader_addline(buffer, " (snorm)");
1991 break;
1992
1993 default:
1994 shader_addline(buffer, " (unknown)");
1995 break;
1996 }
1997 }
1998 else
1999 {
2000 /* Pixel shaders 3.0 don't have usage semantics. */
2001 if (shader_version->major < 3 && shader_version->type == WINED3D_SHADER_TYPE_PIXEL)
2002 return;
2003 else
2004 shader_addline(buffer, "_");
2005
2006 switch (semantic->usage)
2007 {
2008 case WINED3D_DECL_USAGE_POSITION:
2009 shader_addline(buffer, "position%u", semantic->usage_idx);
2010 break;
2011
2012 case WINED3D_DECL_USAGE_BLEND_INDICES:
2013 shader_addline(buffer, "blend");
2014 break;
2015
2016 case WINED3D_DECL_USAGE_BLEND_WEIGHT:
2017 shader_addline(buffer, "weight");
2018 break;
2019
2020 case WINED3D_DECL_USAGE_NORMAL:
2021 shader_addline(buffer, "normal%u", semantic->usage_idx);
2022 break;
2023
2024 case WINED3D_DECL_USAGE_PSIZE:
2025 shader_addline(buffer, "psize");
2026 break;
2027
2028 case WINED3D_DECL_USAGE_COLOR:
2029 if (!semantic->usage_idx)
2030 shader_addline(buffer, "color");
2031 else
2032 shader_addline(buffer, "specular%u", (semantic->usage_idx - 1));
2033 break;
2034
2035 case WINED3D_DECL_USAGE_TEXCOORD:
2036 shader_addline(buffer, "texture%u", semantic->usage_idx);
2037 break;
2038
2039 case WINED3D_DECL_USAGE_TANGENT:
2040 shader_addline(buffer, "tangent");
2041 break;
2042
2043 case WINED3D_DECL_USAGE_BINORMAL:
2044 shader_addline(buffer, "binormal");
2045 break;
2046
2047 case WINED3D_DECL_USAGE_TESS_FACTOR:
2048 shader_addline(buffer, "tessfactor");
2049 break;
2050
2051 case WINED3D_DECL_USAGE_POSITIONT:
2052 shader_addline(buffer, "positionT%u", semantic->usage_idx);
2053 break;
2054
2055 case WINED3D_DECL_USAGE_FOG:
2056 shader_addline(buffer, "fog");
2057 break;
2058
2059 case WINED3D_DECL_USAGE_DEPTH:
2060 shader_addline(buffer, "depth");
2061 break;
2062
2063 case WINED3D_DECL_USAGE_SAMPLE:
2064 shader_addline(buffer, "sample");
2065 break;
2066
2067 default:
2068 shader_addline(buffer, "<unknown_semantic(%#x)>", semantic->usage);
2069 FIXME("Unrecognised semantic usage %#x.\n", semantic->usage);
2070 }
2071 }
2072 }
2073
2074 static void shader_dump_register(struct wined3d_string_buffer *buffer,
2075 const struct wined3d_shader_register *reg, const struct wined3d_shader_version *shader_version)
2076 {
2077 static const char * const rastout_reg_names[] = {"oPos", "oFog", "oPts"};
2078 static const char * const misctype_reg_names[] = {"vPos", "vFace"};
2079 UINT offset = reg->idx[0].offset;
2080
2081 switch (reg->type)
2082 {
2083 case WINED3DSPR_TEMP:
2084 shader_addline(buffer, "r");
2085 break;
2086
2087 case WINED3DSPR_INPUT:
2088 shader_addline(buffer, "v");
2089 break;
2090
2091 case WINED3DSPR_CONST:
2092 case WINED3DSPR_CONST2:
2093 case WINED3DSPR_CONST3:
2094 case WINED3DSPR_CONST4:
2095 shader_addline(buffer, "c");
2096 offset = shader_get_float_offset(reg->type, offset);
2097 break;
2098
2099 case WINED3DSPR_TEXTURE: /* vs: case WINED3DSPR_ADDR */
2100 shader_addline(buffer, "%c", shader_version->type == WINED3D_SHADER_TYPE_PIXEL ? 't' : 'a');
2101 break;
2102
2103 case WINED3DSPR_RASTOUT:
2104 shader_addline(buffer, "%s", rastout_reg_names[offset]);
2105 break;
2106
2107 case WINED3DSPR_COLOROUT:
2108 shader_addline(buffer, "oC");
2109 break;
2110
2111 case WINED3DSPR_DEPTHOUT:
2112 shader_addline(buffer, "oDepth");
2113 break;
2114
2115 case WINED3DSPR_ATTROUT:
2116 shader_addline(buffer, "oD");
2117 break;
2118
2119 case WINED3DSPR_TEXCRDOUT:
2120 /* Vertex shaders >= 3.0 use general purpose output registers
2121 * (WINED3DSPR_OUTPUT), which can include an address token. */
2122 if (shader_version->major >= 3)
2123 shader_addline(buffer, "o");
2124 else
2125 shader_addline(buffer, "oT");
2126 break;
2127
2128 case WINED3DSPR_CONSTINT:
2129 shader_addline(buffer, "i");
2130 break;
2131
2132 case WINED3DSPR_CONSTBOOL:
2133 shader_addline(buffer, "b");
2134 break;
2135
2136 case WINED3DSPR_LABEL:
2137 shader_addline(buffer, "l");
2138 break;
2139
2140 case WINED3DSPR_LOOP:
2141 shader_addline(buffer, "aL");
2142 break;
2143
2144 case WINED3DSPR_SAMPLER:
2145 shader_addline(buffer, "s");
2146 break;
2147
2148 case WINED3DSPR_MISCTYPE:
2149 if (offset > 1)
2150 {
2151 FIXME("Unhandled misctype register %u.\n", offset);
2152 shader_addline(buffer, "<unhandled misctype %#x>", offset);
2153 }
2154 else
2155 {
2156 shader_addline(buffer, "%s", misctype_reg_names[offset]);
2157 }
2158 break;
2159
2160 case WINED3DSPR_PREDICATE:
2161 shader_addline(buffer, "p");
2162 break;
2163
2164 case WINED3DSPR_IMMCONST:
2165 shader_addline(buffer, "l");
2166 break;
2167
2168 case WINED3DSPR_CONSTBUFFER:
2169 shader_addline(buffer, "cb");
2170 break;
2171
2172 case WINED3DSPR_IMMCONSTBUFFER:
2173 shader_addline(buffer, "icb");
2174 break;
2175
2176 case WINED3DSPR_PRIMID:
2177 shader_addline(buffer, "primID");
2178 break;
2179
2180 case WINED3DSPR_NULL:
2181 shader_addline(buffer, "null");
2182 break;
2183
2184 case WINED3DSPR_RESOURCE:
2185 shader_addline(buffer, "t");
2186 break;
2187
2188 case WINED3DSPR_UAV:
2189 shader_addline(buffer, "u");
2190 break;
2191
2192 case WINED3DSPR_OUTPOINTID:
2193 shader_addline(buffer, "vOutputControlPointID");
2194 break;
2195
2196 case WINED3DSPR_FORKINSTID:
2197 shader_addline(buffer, "vForkInstanceId");
2198 break;
2199
2200 case WINED3DSPR_JOININSTID:
2201 shader_addline(buffer, "vJoinInstanceId");
2202 break;
2203
2204 case WINED3DSPR_INCONTROLPOINT:
2205 shader_addline(buffer, "vicp");
2206 break;
2207
2208 case WINED3DSPR_OUTCONTROLPOINT:
2209 shader_addline(buffer, "vocp");
2210 break;
2211
2212 case WINED3DSPR_PATCHCONST:
2213 shader_addline(buffer, "vpc");
2214 break;
2215
2216 case WINED3DSPR_TESSCOORD:
2217 shader_addline(buffer, "vDomainLocation");
2218 break;
2219
2220 case WINED3DSPR_GROUPSHAREDMEM:
2221 shader_addline(buffer, "g");
2222 break;
2223
2224 case WINED3DSPR_THREADID:
2225 shader_addline(buffer, "vThreadID");
2226 break;
2227
2228 case WINED3DSPR_THREADGROUPID:
2229 shader_addline(buffer, "vThreadGroupID");
2230 break;
2231
2232 case WINED3DSPR_LOCALTHREADID:
2233 shader_addline(buffer, "vThreadIDInGroup");
2234 break;
2235
2236 case WINED3DSPR_LOCALTHREADINDEX:
2237 shader_addline(buffer, "vThreadIDInGroupFlattened");
2238 break;
2239
2240 case WINED3DSPR_IDXTEMP:
2241 shader_addline(buffer, "x");
2242 break;
2243
2244 case WINED3DSPR_STREAM:
2245 shader_addline(buffer, "m");
2246 break;
2247
2248 case WINED3DSPR_FUNCTIONBODY:
2249 shader_addline(buffer, "fb");
2250 break;
2251
2252 case WINED3DSPR_FUNCTIONPOINTER:
2253 shader_addline(buffer, "fp");
2254 break;
2255
2256 case WINED3DSPR_COVERAGE:
2257 shader_addline(buffer, "vCoverage");
2258 break;
2259
2260 case WINED3DSPR_SAMPLEMASK:
2261 shader_addline(buffer, "oMask");
2262 break;
2263
2264 case WINED3DSPR_GSINSTID:
2265 shader_addline(buffer, "vGSInstanceID");
2266 break;
2267
2268 default:
2269 shader_addline(buffer, "<unhandled_rtype(%#x)>", reg->type);
2270 break;
2271 }
2272
2273 if (reg->type == WINED3DSPR_IMMCONST)
2274 {
2275 shader_addline(buffer, "(");
2276 switch (reg->immconst_type)
2277 {
2278 case WINED3D_IMMCONST_SCALAR:
2279 switch (reg->data_type)
2280 {
2281 case WINED3D_DATA_FLOAT:
2282 shader_addline(buffer, "%.8e", *(const float *)reg->u.immconst_data);
2283 break;
2284 case WINED3D_DATA_INT:
2285 shader_addline(buffer, "%d", reg->u.immconst_data[0]);
2286 break;
2287 case WINED3D_DATA_RESOURCE:
2288 case WINED3D_DATA_SAMPLER:
2289 case WINED3D_DATA_UINT:
2290 shader_addline(buffer, "%u", reg->u.immconst_data[0]);
2291 break;
2292 default:
2293 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
2294 break;
2295 }
2296 break;
2297
2298 case WINED3D_IMMCONST_VEC4:
2299 switch (reg->data_type)
2300 {
2301 case WINED3D_DATA_FLOAT:
2302 shader_addline(buffer, "%.8e, %.8e, %.8e, %.8e",
2303 *(const float *)&reg->u.immconst_data[0], *(const float *)&reg->u.immconst_data[1],
2304 *(const float *)&reg->u.immconst_data[2], *(const float *)&reg->u.immconst_data[3]);
2305 break;
2306 case WINED3D_DATA_INT:
2307 shader_addline(buffer, "%d, %d, %d, %d",
2308 reg->u.immconst_data[0], reg->u.immconst_data[1],
2309 reg->u.immconst_data[2], reg->u.immconst_data[3]);
2310 break;
2311 case WINED3D_DATA_RESOURCE:
2312 case WINED3D_DATA_SAMPLER:
2313 case WINED3D_DATA_UINT:
2314 shader_addline(buffer, "%u, %u, %u, %u",
2315 reg->u.immconst_data[0], reg->u.immconst_data[1],
2316 reg->u.immconst_data[2], reg->u.immconst_data[3]);
2317 break;
2318 default:
2319 shader_addline(buffer, "<unhandled data type %#x>", reg->data_type);
2320 break;
2321 }
2322 break;
2323
2324 default:
2325 shader_addline(buffer, "<unhandled immconst_type %#x>", reg->immconst_type);
2326 break;
2327 }
2328 shader_addline(buffer, ")");
2329 }
2330 else if (reg->type != WINED3DSPR_RASTOUT
2331 && reg->type != WINED3DSPR_MISCTYPE
2332 && reg->type != WINED3DSPR_NULL)
2333 {
2334 if (offset != ~0u)
2335 {
2336 shader_addline(buffer, "[");
2337 if (reg->idx[0].rel_addr)
2338 {
2339 shader_dump_src_param(buffer, reg->idx[0].rel_addr, shader_version);
2340 shader_addline(buffer, " + ");
2341 }
2342 shader_addline(buffer, "%u]", offset);
2343
2344 if (reg->idx[1].offset != ~0u)
2345 {
2346 shader_addline(buffer, "[");
2347 if (reg->idx[1].rel_addr)
2348 {
2349 shader_dump_src_param(buffer, reg->idx[1].rel_addr, shader_version);
2350 shader_addline(buffer, " + ");
2351 }
2352 shader_addline(buffer, "%u]", reg->idx[1].offset);
2353 }
2354 }
2355
2356 if (reg->type == WINED3DSPR_FUNCTIONPOINTER)
2357 shader_addline(buffer, "[%u]", reg->u.fp_body_idx);
2358 }
2359 }
2360
2361 static void shader_dump_dst_param(struct wined3d_string_buffer *buffer,
2362 const struct wined3d_shader_dst_param *param, const struct wined3d_shader_version *shader_version)
2363 {
2364 DWORD write_mask = param->write_mask;
2365
2366 shader_dump_register(buffer, &param->reg, shader_version);
2367
2368 if (write_mask && write_mask != WINED3DSP_WRITEMASK_ALL)
2369 {
2370 static const char write_mask_chars[] = "xyzw";
2371
2372 shader_addline(buffer, ".");
2373 if (write_mask & WINED3DSP_WRITEMASK_0)
2374 shader_addline(buffer, "%c", write_mask_chars[0]);
2375 if (write_mask & WINED3DSP_WRITEMASK_1)
2376 shader_addline(buffer, "%c", write_mask_chars[1]);
2377 if (write_mask & WINED3DSP_WRITEMASK_2)
2378 shader_addline(buffer, "%c", write_mask_chars[2]);
2379 if (write_mask & WINED3DSP_WRITEMASK_3)
2380 shader_addline(buffer, "%c", write_mask_chars[3]);
2381 }
2382 }
2383
2384 static void shader_dump_src_param(struct wined3d_string_buffer *buffer,
2385 const struct wined3d_shader_src_param *param, const struct wined3d_shader_version *shader_version)
2386 {
2387 enum wined3d_shader_src_modifier src_modifier = param->modifiers;
2388 DWORD swizzle = param->swizzle;
2389
2390 if (src_modifier == WINED3DSPSM_NEG
2391 || src_modifier == WINED3DSPSM_BIASNEG
2392 || src_modifier == WINED3DSPSM_SIGNNEG
2393 || src_modifier == WINED3DSPSM_X2NEG
2394 || src_modifier == WINED3DSPSM_ABSNEG)
2395 shader_addline(buffer, "-");
2396 else if (src_modifier == WINED3DSPSM_COMP)
2397 shader_addline(buffer, "1-");
2398 else if (src_modifier == WINED3DSPSM_NOT)
2399 shader_addline(buffer, "!");
2400
2401 if (src_modifier == WINED3DSPSM_ABS || src_modifier == WINED3DSPSM_ABSNEG)
2402 shader_addline(buffer, "abs(");
2403
2404 shader_dump_register(buffer, &param->reg, shader_version);
2405
2406 switch (src_modifier)
2407 {
2408 case WINED3DSPSM_NONE: break;
2409 case WINED3DSPSM_NEG: break;
2410 case WINED3DSPSM_NOT: break;
2411 case WINED3DSPSM_BIAS: shader_addline(buffer, "_bias"); break;
2412 case WINED3DSPSM_BIASNEG: shader_addline(buffer, "_bias"); break;
2413 case WINED3DSPSM_SIGN: shader_addline(buffer, "_bx2"); break;
2414 case WINED3DSPSM_SIGNNEG: shader_addline(buffer, "_bx2"); break;
2415 case WINED3DSPSM_COMP: break;
2416 case WINED3DSPSM_X2: shader_addline(buffer, "_x2"); break;
2417 case WINED3DSPSM_X2NEG: shader_addline(buffer, "_x2"); break;
2418 case WINED3DSPSM_DZ: shader_addline(buffer, "_dz"); break;
2419 case WINED3DSPSM_DW: shader_addline(buffer, "_dw"); break;
2420 case WINED3DSPSM_ABSNEG: shader_addline(buffer, ")"); break;
2421 case WINED3DSPSM_ABS: shader_addline(buffer, ")"); break;
2422 default: shader_addline(buffer, "_unknown_modifier(%#x)", src_modifier);
2423 }
2424
2425 if (swizzle != WINED3DSP_NOSWIZZLE)
2426 {
2427 static const char swizzle_chars[] = "xyzw";
2428 DWORD swizzle_x = swizzle & 0x03;
2429 DWORD swizzle_y = (swizzle >> 2) & 0x03;
2430 DWORD swizzle_z = (swizzle >> 4) & 0x03;
2431 DWORD swizzle_w = (swizzle >> 6) & 0x03;
2432
2433 if (swizzle_x == swizzle_y
2434 && swizzle_x == swizzle_z
2435 && swizzle_x == swizzle_w)
2436 {
2437 shader_addline(buffer, ".%c", swizzle_chars[swizzle_x]);
2438 }
2439 else
2440 {
2441 shader_addline(buffer, ".%c%c%c%c", swizzle_chars[swizzle_x], swizzle_chars[swizzle_y],
2442 swizzle_chars[swizzle_z], swizzle_chars[swizzle_w]);
2443 }
2444 }
2445 }
2446
2447 /* Shared code in order to generate the bulk of the shader string. */
2448 HRESULT shader_generate_code(const struct wined3d_shader *shader, struct wined3d_string_buffer *buffer,
2449 const struct wined3d_shader_reg_maps *reg_maps, void *backend_ctx,
2450 const DWORD *start, const DWORD *end)
2451 {
2452 struct wined3d_device *device = shader->device;
2453 const struct wined3d_shader_frontend *fe = shader->frontend;
2454 void *fe_data = shader->frontend_data;
2455 struct wined3d_shader_version shader_version;
2456 struct wined3d_shader_parser_state state;
2457 struct wined3d_shader_instruction ins;
2458 struct wined3d_shader_tex_mx tex_mx;
2459 struct wined3d_shader_context ctx;
2460 const DWORD *ptr;
2461
2462 /* Initialize current parsing state. */
2463 tex_mx.current_row = 0;
2464 state.current_loop_depth = 0;
2465 state.current_loop_reg = 0;
2466 state.in_subroutine = FALSE;
2467
2468 ctx.shader = shader;
2469 ctx.gl_info = &device->adapter->gl_info;
2470 ctx.reg_maps = reg_maps;
2471 ctx.buffer = buffer;
2472 ctx.tex_mx = &tex_mx;
2473 ctx.state = &state;
2474 ctx.backend_data = backend_ctx;
2475 ins.ctx = &ctx;
2476
2477 fe->shader_read_header(fe_data, &ptr, &shader_version);
2478 if (start)
2479 ptr = start;
2480
2481 while (!fe->shader_is_end(fe_data, &ptr) && ptr != end)
2482 {
2483 /* Read opcode. */
2484 fe->shader_read_instruction(fe_data, &ptr, &ins);
2485
2486 /* Unknown opcode and its parameters. */
2487 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2488 {
2489 WARN("Encountered unrecognised or invalid instruction.\n");
2490 return WINED3DERR_INVALIDCALL;
2491 }
2492
2493 if (ins.predicate)
2494 FIXME("Predicates not implemented.\n");
2495
2496 /* Call appropriate function for output target */
2497 device->shader_backend->shader_handle_instruction(&ins);
2498 }
2499
2500 return WINED3D_OK;
2501 }
2502
2503 static void shader_dump_ins_modifiers(struct wined3d_string_buffer *buffer,
2504 const struct wined3d_shader_dst_param *dst)
2505 {
2506 DWORD mmask = dst->modifiers;
2507
2508 switch (dst->shift)
2509 {
2510 case 0: break;
2511 case 13: shader_addline(buffer, "_d8"); break;
2512 case 14: shader_addline(buffer, "_d4"); break;
2513 case 15: shader_addline(buffer, "_d2"); break;
2514 case 1: shader_addline(buffer, "_x2"); break;
2515 case 2: shader_addline(buffer, "_x4"); break;
2516 case 3: shader_addline(buffer, "_x8"); break;
2517 default: shader_addline(buffer, "_unhandled_shift(%d)", dst->shift); break;
2518 }
2519
2520 if (mmask & WINED3DSPDM_SATURATE) shader_addline(buffer, "_sat");
2521 if (mmask & WINED3DSPDM_PARTIALPRECISION) shader_addline(buffer, "_pp");
2522 if (mmask & WINED3DSPDM_MSAMPCENTROID) shader_addline(buffer, "_centroid");
2523
2524 mmask &= ~(WINED3DSPDM_SATURATE | WINED3DSPDM_PARTIALPRECISION | WINED3DSPDM_MSAMPCENTROID);
2525 if (mmask) FIXME("Unrecognised modifier %#x.\n", mmask);
2526 }
2527
2528 static void shader_dump_primitive_type(struct wined3d_string_buffer *buffer,
2529 const struct wined3d_shader_primitive_type *primitive_type)
2530 {
2531 switch (primitive_type->type)
2532 {
2533 case WINED3D_PT_UNDEFINED:
2534 shader_addline(buffer, "undefined");
2535 break;
2536 case WINED3D_PT_POINTLIST:
2537 shader_addline(buffer, "pointlist");
2538 break;
2539 case WINED3D_PT_LINELIST:
2540 shader_addline(buffer, "linelist");
2541 break;
2542 case WINED3D_PT_LINESTRIP:
2543 shader_addline(buffer, "linestrip");
2544 break;
2545 case WINED3D_PT_TRIANGLELIST:
2546 shader_addline(buffer, "trianglelist");
2547 break;
2548 case WINED3D_PT_TRIANGLESTRIP:
2549 shader_addline(buffer, "trianglestrip");
2550 break;
2551 case WINED3D_PT_TRIANGLEFAN:
2552 shader_addline(buffer, "trianglefan");
2553 break;
2554 case WINED3D_PT_LINELIST_ADJ:
2555 shader_addline(buffer, "linelist_adj");
2556 break;
2557 case WINED3D_PT_LINESTRIP_ADJ:
2558 shader_addline(buffer, "linestrip_adj");
2559 break;
2560 case WINED3D_PT_TRIANGLELIST_ADJ:
2561 shader_addline(buffer, "trianglelist_adj");
2562 break;
2563 case WINED3D_PT_TRIANGLESTRIP_ADJ:
2564 shader_addline(buffer, "trianglestrip_adj");
2565 break;
2566 case WINED3D_PT_PATCH:
2567 shader_addline(buffer, "patch%u", primitive_type->patch_vertex_count);
2568 break;
2569 default:
2570 shader_addline(buffer, "<unrecognized_primitive_type %#x>", primitive_type->type);
2571 break;
2572 }
2573 }
2574
2575 static void shader_dump_interpolation_mode(struct wined3d_string_buffer *buffer,
2576 enum wined3d_shader_interpolation_mode interpolation_mode)
2577 {
2578 switch (interpolation_mode)
2579 {
2580 case WINED3DSIM_CONSTANT:
2581 shader_addline(buffer, "constant");
2582 break;
2583 case WINED3DSIM_LINEAR:
2584 shader_addline(buffer, "linear");
2585 break;
2586 case WINED3DSIM_LINEAR_CENTROID:
2587 shader_addline(buffer, "linear centroid");
2588 break;
2589 case WINED3DSIM_LINEAR_NOPERSPECTIVE:
2590 shader_addline(buffer, "linear noperspective");
2591 break;
2592 case WINED3DSIM_LINEAR_SAMPLE:
2593 shader_addline(buffer, "linear sample");
2594 break;
2595 case WINED3DSIM_LINEAR_NOPERSPECTIVE_CENTROID:
2596 shader_addline(buffer, "linear noperspective centroid");
2597 break;
2598 case WINED3DSIM_LINEAR_NOPERSPECTIVE_SAMPLE:
2599 shader_addline(buffer, "linear noperspective sample");
2600 break;
2601 default:
2602 shader_addline(buffer, "<unrecognized_interpolation_mode %#x>", interpolation_mode);
2603 break;
2604 }
2605 }
2606
2607 static void shader_trace_init(const struct wined3d_shader_frontend *fe, void *fe_data)
2608 {
2609 struct wined3d_shader_version shader_version;
2610 struct wined3d_string_buffer buffer;
2611 const char *type_prefix;
2612 const char *p, *q;
2613 const DWORD *ptr;
2614 DWORD i;
2615
2616 if (!string_buffer_init(&buffer))
2617 {
2618 ERR("Failed to initialize string buffer.\n");
2619 return;
2620 }
2621
2622 fe->shader_read_header(fe_data, &ptr, &shader_version);
2623
2624 TRACE("Parsing %p.\n", ptr);
2625
2626 switch (shader_version.type)
2627 {
2628 case WINED3D_SHADER_TYPE_VERTEX:
2629 type_prefix = "vs";
2630 break;
2631
2632 case WINED3D_SHADER_TYPE_HULL:
2633 type_prefix = "hs";
2634 break;
2635
2636 case WINED3D_SHADER_TYPE_DOMAIN:
2637 type_prefix = "ds";
2638 break;
2639
2640 case WINED3D_SHADER_TYPE_GEOMETRY:
2641 type_prefix = "gs";
2642 break;
2643
2644 case WINED3D_SHADER_TYPE_PIXEL:
2645 type_prefix = "ps";
2646 break;
2647
2648 case WINED3D_SHADER_TYPE_COMPUTE:
2649 type_prefix = "cs";
2650 break;
2651
2652 default:
2653 FIXME("Unhandled shader type %#x.\n", shader_version.type);
2654 type_prefix = "unknown";
2655 break;
2656 }
2657
2658 shader_addline(&buffer, "%s_%u_%u\n", type_prefix, shader_version.major, shader_version.minor);
2659
2660 while (!fe->shader_is_end(fe_data, &ptr))
2661 {
2662 struct wined3d_shader_instruction ins;
2663
2664 fe->shader_read_instruction(fe_data, &ptr, &ins);
2665 if (ins.handler_idx == WINED3DSIH_TABLE_SIZE)
2666 {
2667 WARN("Skipping unrecognized instruction.\n");
2668 shader_addline(&buffer, "<unrecognized instruction>\n");
2669 continue;
2670 }
2671
2672 if (ins.handler_idx == WINED3DSIH_DCL || ins.handler_idx == WINED3DSIH_DCL_UAV_TYPED)
2673 {
2674 shader_dump_decl_usage(&buffer, &ins.declaration.semantic, ins.flags, &shader_version);
2675 shader_dump_ins_modifiers(&buffer, &ins.declaration.semantic.reg);
2676 shader_addline(&buffer, " ");
2677 shader_dump_dst_param(&buffer, &ins.declaration.semantic.reg, &shader_version);
2678 }
2679 else if (ins.handler_idx == WINED3DSIH_DCL_CONSTANT_BUFFER)
2680 {
2681 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2682 shader_dump_src_param(&buffer, &ins.declaration.src, &shader_version);
2683 shader_addline(&buffer, ", %s",
2684 ins.flags & WINED3DSI_INDEXED_DYNAMIC ? "dynamicIndexed" : "immediateIndexed");
2685 }
2686 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_BODY)
2687 {
2688 shader_addline(&buffer, "%s fb%u",
2689 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2690 }
2691 else if (ins.handler_idx == WINED3DSIH_DCL_FUNCTION_TABLE)
2692 {
2693 shader_addline(&buffer, "%s ft%u = {...}",
2694 shader_opcode_names[ins.handler_idx], ins.declaration.index);
2695 }
2696 else if (ins.handler_idx == WINED3DSIH_DCL_GLOBAL_FLAGS)
2697 {
2698 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2699 shader_dump_global_flags(&buffer, ins.flags);
2700 }
2701 else if (ins.handler_idx == WINED3DSIH_DCL_HS_MAX_TESSFACTOR)
2702 {
2703 shader_addline(&buffer, "%s %.8e", shader_opcode_names[ins.handler_idx],
2704 ins.declaration.max_tessellation_factor);
2705 }
2706 else if (ins.handler_idx == WINED3DSIH_DCL_IMMEDIATE_CONSTANT_BUFFER)
2707 {
2708 shader_addline(&buffer, "%s {\n", shader_opcode_names[ins.handler_idx]);
2709 for (i = 0; i < ins.declaration.icb->vec4_count; ++i)
2710 {
2711 shader_addline(&buffer, " {0x%08x, 0x%08x, 0x%08x, 0x%08x},\n",
2712 ins.declaration.icb->data[4 * i + 0],
2713 ins.declaration.icb->data[4 * i + 1],
2714 ins.declaration.icb->data[4 * i + 2],
2715 ins.declaration.icb->data[4 * i + 3]);
2716 }
2717 shader_addline(&buffer, "}");
2718 }
2719 else if (ins.handler_idx == WINED3DSIH_DCL_INDEX_RANGE)
2720 {
2721 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2722 shader_dump_dst_param(&buffer, &ins.declaration.index_range.first_register, &shader_version);
2723 shader_addline(&buffer, " %u", ins.declaration.index_range.last_register);
2724 }
2725 else if (ins.handler_idx == WINED3DSIH_DCL_INDEXABLE_TEMP)
2726 {
2727 shader_addline(&buffer, "%s x[%u][%u], %u", shader_opcode_names[ins.handler_idx],
2728 ins.declaration.indexable_temp.register_idx,
2729 ins.declaration.indexable_temp.register_size,
2730 ins.declaration.indexable_temp.component_count);
2731 }
2732 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS)
2733 {
2734 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2735 shader_dump_interpolation_mode(&buffer, ins.flags);
2736 shader_addline(&buffer, " ");
2737 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2738 }
2739 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SGV
2740 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SGV
2741 || ins.handler_idx == WINED3DSIH_DCL_INPUT_SIV
2742 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_SIV)
2743 {
2744 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2745 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2746 shader_addline(&buffer, ", ");
2747 shader_dump_shader_input_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2748 }
2749 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PS_SIV)
2750 {
2751 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2752 shader_dump_interpolation_mode(&buffer, ins.flags);
2753 shader_addline(&buffer, " ");
2754 shader_dump_dst_param(&buffer, &ins.declaration.register_semantic.reg, &shader_version);
2755 shader_addline(&buffer, ", ");
2756 shader_dump_shader_input_sysval_semantic(&buffer, ins.declaration.register_semantic.sysval_semantic);
2757 }
2758 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT
2759 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT)
2760 {
2761 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2762 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2763 }
2764 else if (ins.handler_idx == WINED3DSIH_DCL_INPUT_PRIMITIVE
2765 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_TOPOLOGY)
2766 {
2767 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2768 shader_dump_primitive_type(&buffer, &ins.declaration.primitive_type);
2769 }
2770 else if (ins.handler_idx == WINED3DSIH_DCL_INTERFACE)
2771 {
2772 shader_addline(&buffer, "%s fp[%u][%u][%u] = {...}",
2773 shader_opcode_names[ins.handler_idx], ins.declaration.fp.index,
2774 ins.declaration.fp.array_size, ins.declaration.fp.body_count);
2775 }
2776 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_RAW)
2777 {
2778 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2779 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2780 }
2781 else if (ins.handler_idx == WINED3DSIH_DCL_RESOURCE_STRUCTURED)
2782 {
2783 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2784 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2785 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2786 }
2787 else if (ins.handler_idx == WINED3DSIH_DCL_SAMPLER)
2788 {
2789 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2790 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2791 if (ins.flags == WINED3DSI_SAMPLER_COMPARISON_MODE)
2792 shader_addline(&buffer, ", comparisonMode");
2793 }
2794 else if (ins.handler_idx == WINED3DSIH_DCL_TEMPS
2795 || ins.handler_idx == WINED3DSIH_DCL_GS_INSTANCES
2796 || ins.handler_idx == WINED3DSIH_DCL_HS_FORK_PHASE_INSTANCE_COUNT
2797 || ins.handler_idx == WINED3DSIH_DCL_HS_JOIN_PHASE_INSTANCE_COUNT
2798 || ins.handler_idx == WINED3DSIH_DCL_INPUT_CONTROL_POINT_COUNT
2799 || ins.handler_idx == WINED3DSIH_DCL_OUTPUT_CONTROL_POINT_COUNT
2800 || ins.handler_idx == WINED3DSIH_DCL_VERTICES_OUT)
2801 {
2802 shader_addline(&buffer, "%s %u", shader_opcode_names[ins.handler_idx], ins.declaration.count);
2803 }
2804 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_DOMAIN)
2805 {
2806 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2807 shader_dump_tessellator_domain(&buffer, ins.declaration.tessellator_domain);
2808 }
2809 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_OUTPUT_PRIMITIVE)
2810 {
2811 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2812 shader_dump_tessellator_output_primitive(&buffer, ins.declaration.tessellator_output_primitive);
2813 }
2814 else if (ins.handler_idx == WINED3DSIH_DCL_TESSELLATOR_PARTITIONING)
2815 {
2816 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2817 shader_dump_tessellator_partitioning(&buffer, ins.declaration.tessellator_partitioning);
2818 }
2819 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_RAW)
2820 {
2821 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2822 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_raw.reg, &shader_version);
2823 shader_addline(&buffer, ", %u", ins.declaration.tgsm_raw.byte_count);
2824 }
2825 else if (ins.handler_idx == WINED3DSIH_DCL_TGSM_STRUCTURED)
2826 {
2827 shader_addline(&buffer, "%s ", shader_opcode_names[ins.handler_idx]);
2828 shader_dump_dst_param(&buffer, &ins.declaration.tgsm_structured.reg, &shader_version);
2829 shader_addline(&buffer, ", %u, %u", ins.declaration.tgsm_structured.byte_stride,
2830 ins.declaration.tgsm_structured.structure_count);
2831 }
2832 else if (ins.handler_idx == WINED3DSIH_DCL_THREAD_GROUP)
2833 {
2834 shader_addline(&buffer, "%s %u, %u, %u", shader_opcode_names[ins.handler_idx],
2835 ins.declaration.thread_group_size.x,
2836 ins.declaration.thread_group_size.y,
2837 ins.declaration.thread_group_size.z);
2838 }
2839 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_RAW)
2840 {
2841 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2842 shader_dump_uav_flags(&buffer, ins.flags);
2843 shader_addline(&buffer, " ");
2844 shader_dump_dst_param(&buffer, &ins.declaration.dst, &shader_version);
2845 }
2846 else if (ins.handler_idx == WINED3DSIH_DCL_UAV_STRUCTURED)
2847 {
2848 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2849 shader_dump_uav_flags(&buffer, ins.flags);
2850 shader_addline(&buffer, " ");
2851 shader_dump_dst_param(&buffer, &ins.declaration.structured_resource.reg, &shader_version);
2852 shader_addline(&buffer, ", %u", ins.declaration.structured_resource.byte_stride);
2853 }
2854 else if (ins.handler_idx == WINED3DSIH_DEF)
2855 {
2856 shader_addline(&buffer, "def c%u = %.8e, %.8e, %.8e, %.8e", shader_get_float_offset(ins.dst[0].reg.type,
2857 ins.dst[0].reg.idx[0].offset),
2858 *(const float *)&ins.src[0].reg.u.immconst_data[0],
2859 *(const float *)&ins.src[0].reg.u.immconst_data[1],
2860 *(const float *)&ins.src[0].reg.u.immconst_data[2],
2861 *(const float *)&ins.src[0].reg.u.immconst_data[3]);
2862 }
2863 else if (ins.handler_idx == WINED3DSIH_DEFI)
2864 {
2865 shader_addline(&buffer, "defi i%u = %d, %d, %d, %d", ins.dst[0].reg.idx[0].offset,
2866 ins.src[0].reg.u.immconst_data[0],
2867 ins.src[0].reg.u.immconst_data[1],
2868 ins.src[0].reg.u.immconst_data[2],
2869 ins.src[0].reg.u.immconst_data[3]);
2870 }
2871 else if (ins.handler_idx == WINED3DSIH_DEFB)
2872 {
2873 shader_addline(&buffer, "defb b%u = %s",
2874 ins.dst[0].reg.idx[0].offset, ins.src[0].reg.u.immconst_data[0] ? "true" : "false");
2875 }
2876 else
2877 {
2878 if (ins.predicate)
2879 {
2880 shader_addline(&buffer, "(");
2881 shader_dump_src_param(&buffer, ins.predicate, &shader_version);
2882 shader_addline(&buffer, ") ");
2883 }
2884
2885 /* PixWin marks instructions with the coissue flag with a '+' */
2886 if (ins.coissue)
2887 shader_addline(&buffer, "+");
2888
2889 shader_addline(&buffer, "%s", shader_opcode_names[ins.handler_idx]);
2890
2891 if (ins.handler_idx == WINED3DSIH_BREAKP
2892 || ins.handler_idx == WINED3DSIH_CONTINUEP
2893 || ins.handler_idx == WINED3DSIH_IF
2894 || ins.handler_idx == WINED3DSIH_RETP)
2895 {
2896 switch (ins.flags)
2897 {
2898 case WINED3D_SHADER_CONDITIONAL_OP_NZ: shader_addline(&buffer, "_nz"); break;
2899 case WINED3D_SHADER_CONDITIONAL_OP_Z: shader_addline(&buffer, "_z"); break;
2900 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags); break;
2901 }
2902 }
2903 else if (ins.handler_idx == WINED3DSIH_IFC
2904 || ins.handler_idx == WINED3DSIH_BREAKC)
2905 {
2906 switch (ins.flags)
2907 {
2908 case WINED3D_SHADER_REL_OP_GT: shader_addline(&buffer, "_gt"); break;
2909 case WINED3D_SHADER_REL_OP_EQ: shader_addline(&buffer, "_eq"); break;
2910 case WINED3D_SHADER_REL_OP_GE: shader_addline(&buffer, "_ge"); break;
2911 case WINED3D_SHADER_REL_OP_LT: shader_addline(&buffer, "_lt"); break;
2912 case WINED3D_SHADER_REL_OP_NE: shader_addline(&buffer, "_ne"); break;
2913 case WINED3D_SHADER_REL_OP_LE: shader_addline(&buffer, "_le"); break;
2914 default: shader_addline(&buffer, "_(%u)", ins.flags);
2915 }
2916 }
2917 else if (ins.handler_idx == WINED3DSIH_TEX
2918 && shader_version.major >= 2
2919 && (ins.flags & WINED3DSI_TEXLD_PROJECT))
2920 {
2921 shader_addline(&buffer, "p");
2922 }
2923 else if (ins.handler_idx == WINED3DSIH_RESINFO && ins.flags)
2924 {
2925 switch (ins.flags)
2926 {
2927 case WINED3DSI_RESINFO_RCP_FLOAT: shader_addline(&buffer, "_rcpFloat"); break;
2928 case WINED3DSI_RESINFO_UINT: shader_addline(&buffer, "_uint"); break;
2929 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
2930 }
2931 }
2932 else if (ins.handler_idx == WINED3DSIH_SAMPLE_INFO && ins.flags)
2933 {
2934 switch (ins.flags)
2935 {
2936 case WINED3DSI_SAMPLE_INFO_UINT: shader_addline(&buffer, "_uint"); break;
2937 default: shader_addline(&buffer, "_unrecognized(%#x)", ins.flags);
2938 }
2939 }
2940 else if (ins.handler_idx == WINED3DSIH_SYNC)
2941 {
2942 shader_dump_sync_flags(&buffer, ins.flags);
2943 }
2944
2945 if (wined3d_shader_instruction_has_texel_offset(&ins))
2946 shader_addline(&buffer, "(%d,%d,%d)", ins.texel_offset.u, ins.texel_offset.v, ins.texel_offset.w);
2947
2948 for (i = 0; i < ins.dst_count; ++i)
2949 {
2950 shader_dump_ins_modifiers(&buffer, &ins.dst[i]);
2951 shader_addline(&buffer, !i ? " " : ", ");
2952 shader_dump_dst_param(&buffer, &ins.dst[i], &shader_version);
2953 }
2954
2955 /* Other source tokens */
2956 for (i = ins.dst_count; i < (ins.dst_count + ins.src_count); ++i)
2957 {
2958 shader_addline(&buffer, !i ? " " : ", ");
2959 shader_dump_src_param(&buffer, &ins.src[i - ins.dst_count], &shader_version);
2960 }
2961 }
2962 shader_addline(&buffer, "\n");
2963 }
2964
2965 for (p = buffer.buffer; *p; p = q)
2966 {
2967 if (!(q = strstr(p, "\n")))
2968 q = p + strlen(p);
2969 else
2970 ++q;
2971 TRACE(" %.*s", (int)(q - p), p);
2972 }
2973
2974 string_buffer_free(&buffer);
2975 }
2976
2977 static void shader_cleanup(struct wined3d_shader *shader)
2978 {
2979 if (shader->reg_maps.shader_version.type == WINED3D_SHADER_TYPE_HULL)
2980 {
2981 HeapFree(GetProcessHeap(), 0, shader->u.hs.phases.control_point);
2982 HeapFree(GetProcessHeap(), 0, shader->u.hs.phases.fork);
2983 HeapFree(GetProcessHeap(), 0, shader->u.hs.phases.join);
2984 }
2985 else if (shader->reg_maps.shader_version.type == WINED3D_SHADER_TYPE_GEOMETRY)
2986 {
2987 HeapFree(GetProcessHeap(), 0, shader->u.gs.so_desc.elements);
2988 }
2989
2990 HeapFree(GetProcessHeap(), 0, shader->patch_constant_signature.elements);
2991 HeapFree(GetProcessHeap(), 0, shader->output_signature.elements);
2992 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
2993 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
2994 shader->device->shader_backend->shader_destroy(shader);
2995 shader_cleanup_reg_maps(&shader->reg_maps);
2996 HeapFree(GetProcessHeap(), 0, shader->function);
2997 shader_delete_constant_list(&shader->constantsF);
2998 shader_delete_constant_list(&shader->constantsB);
2999 shader_delete_constant_list(&shader->constantsI);
3000 list_remove(&shader->shader_list_entry);
3001
3002 if (shader->frontend && shader->frontend_data)
3003 shader->frontend->shader_free(shader->frontend_data);
3004 }
3005
3006 struct shader_none_priv
3007 {
3008 const struct wined3d_vertex_pipe_ops *vertex_pipe;
3009 const struct fragment_pipeline *fragment_pipe;
3010 BOOL ffp_proj_control;
3011 };
3012
3013 static void shader_none_handle_instruction(const struct wined3d_shader_instruction *ins) {}
3014 static void shader_none_precompile(void *shader_priv, struct wined3d_shader *shader) {}
3015 static void shader_none_select_compute(void *shader_priv, struct wined3d_context *context,
3016 const struct wined3d_state *state) {}
3017 static void shader_none_update_float_vertex_constants(struct wined3d_device *device, UINT start, UINT count) {}
3018 static void shader_none_update_float_pixel_constants(struct wined3d_device *device, UINT start, UINT count) {}
3019 static void shader_none_load_constants(void *shader_priv, struct wined3d_context *context,
3020 const struct wined3d_state *state) {}
3021 static void shader_none_destroy(struct wined3d_shader *shader) {}
3022 static void shader_none_free_context_data(struct wined3d_context *context) {}
3023 static void shader_none_init_context_state(struct wined3d_context *context) {}
3024
3025 /* Context activation is done by the caller. */
3026 static void shader_none_select(void *shader_priv, struct wined3d_context *context,
3027 const struct wined3d_state *state)
3028 {
3029 const struct wined3d_gl_info *gl_info = context->gl_info;
3030 struct shader_none_priv *priv = shader_priv;
3031
3032 priv->vertex_pipe->vp_enable(gl_info, !use_vs(state));
3033 priv->fragment_pipe->enable_extension(gl_info, !use_ps(state));
3034 }
3035
3036 /* Context activation is done by the caller. */
3037 static void shader_none_disable(void *shader_priv, struct wined3d_context *context)
3038 {
3039 struct shader_none_priv *priv = shader_priv;
3040 const struct wined3d_gl_info *gl_info = context->gl_info;
3041
3042 priv->vertex_pipe->vp_enable(gl_info, FALSE);
3043 priv->fragment_pipe->enable_extension(gl_info, FALSE);
3044
3045 context->shader_update_mask = (1u << WINED3D_SHADER_TYPE_PIXEL)
3046 | (1u << WINED3D_SHADER_TYPE_VERTEX)
3047 | (1u << WINED3D_SHADER_TYPE_GEOMETRY)
3048 | (1u << WINED3D_SHADER_TYPE_HULL)
3049 | (1u << WINED3D_SHADER_TYPE_DOMAIN)
3050 | (1u << WINED3D_SHADER_TYPE_COMPUTE);
3051 }
3052
3053 static HRESULT shader_none_alloc(struct wined3d_device *device, const struct wined3d_vertex_pipe_ops *vertex_pipe,
3054 const struct fragment_pipeline *fragment_pipe)
3055 {
3056 struct fragment_caps fragment_caps;
3057 void *vertex_priv, *fragment_priv;
3058 struct shader_none_priv *priv;
3059
3060 if (!(priv = HeapAlloc(GetProcessHeap(), 0, sizeof(*priv))))
3061 return E_OUTOFMEMORY;
3062
3063 if (!(vertex_priv = vertex_pipe->vp_alloc(&none_shader_backend, priv)))
3064 {
3065 ERR("Failed to initialize vertex pipe.\n");
3066 HeapFree(GetProcessHeap(), 0, priv);
3067 return E_FAIL;
3068 }
3069
3070 if (!(fragment_priv = fragment_pipe->alloc_private(&none_shader_backend, priv)))
3071 {
3072 ERR("Failed to initialize fragment pipe.\n");
3073 vertex_pipe->vp_free(device);
3074 HeapFree(GetProcessHeap(), 0, priv);
3075 return E_FAIL;
3076 }
3077
3078 priv->vertex_pipe = vertex_pipe;
3079 priv->fragment_pipe = fragment_pipe;
3080 fragment_pipe->get_caps(&device->adapter->gl_info, &fragment_caps);
3081 priv->ffp_proj_control = fragment_caps.wined3d_caps & WINED3D_FRAGMENT_CAP_PROJ_CONTROL;
3082
3083 device->vertex_priv = vertex_priv;
3084 device->fragment_priv = fragment_priv;
3085 device->shader_priv = priv;
3086
3087 return WINED3D_OK;
3088 }
3089
3090 static void shader_none_free(struct wined3d_device *device)
3091 {
3092 struct shader_none_priv *priv = device->shader_priv;
3093
3094 priv->fragment_pipe->free_private(device);
3095 priv->vertex_pipe->vp_free(device);
3096 HeapFree(GetProcessHeap(), 0, priv);
3097 }
3098
3099 static BOOL shader_none_allocate_context_data(struct wined3d_context *context)
3100 {
3101 return TRUE;
3102 }
3103
3104 static void shader_none_get_caps(const struct wined3d_gl_info *gl_info, struct shader_caps *caps)
3105 {
3106 /* Set the shader caps to 0 for the none shader backend */
3107 caps->vs_version = 0;
3108 caps->hs_version = 0;
3109 caps->ds_version = 0;
3110 caps->gs_version = 0;
3111 caps->ps_version = 0;
3112 caps->cs_version = 0;
3113 caps->vs_uniform_count = 0;
3114 caps->ps_uniform_count = 0;
3115 caps->ps_1x_max_value = 0.0f;
3116 caps->varying_count = 0;
3117 caps->wined3d_caps = 0;
3118 }
3119
3120 static BOOL shader_none_color_fixup_supported(struct color_fixup_desc fixup)
3121 {
3122 /* We "support" every possible fixup, since we don't support any shader
3123 * model, and will never have to actually sample a texture. */
3124 return TRUE;
3125 }
3126
3127 static BOOL shader_none_has_ffp_proj_control(void *shader_priv)
3128 {
3129 struct shader_none_priv *priv = shader_priv;
3130
3131 return priv->ffp_proj_control;
3132 }
3133
3134 const struct wined3d_shader_backend_ops none_shader_backend =
3135 {
3136 shader_none_handle_instruction,
3137 shader_none_precompile,
3138 shader_none_select,
3139 shader_none_select_compute,
3140 shader_none_disable,
3141 shader_none_update_float_vertex_constants,
3142 shader_none_update_float_pixel_constants,
3143 shader_none_load_constants,
3144 shader_none_destroy,
3145 shader_none_alloc,
3146 shader_none_free,
3147 shader_none_allocate_context_data,
3148 shader_none_free_context_data,
3149 shader_none_init_context_state,
3150 shader_none_get_caps,
3151 shader_none_color_fixup_supported,
3152 shader_none_has_ffp_proj_control,
3153 };
3154
3155 static HRESULT shader_set_function(struct wined3d_shader *shader, DWORD float_const_count,
3156 enum wined3d_shader_type type, unsigned int max_version)
3157 {
3158 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3159 const struct wined3d_shader_frontend *fe;
3160 HRESULT hr;
3161 unsigned int backend_version;
3162 const struct wined3d_d3d_info *d3d_info = &shader->device->adapter->d3d_info;
3163
3164 TRACE("shader %p, float_const_count %u, type %#x, max_version %u.\n",
3165 shader, float_const_count, type, max_version);
3166
3167 fe = shader->frontend;
3168 if (!(shader->frontend_data = fe->shader_init(shader->function,
3169 shader->functionLength, &shader->output_signature)))
3170 {
3171 FIXME("Failed to initialize frontend.\n");
3172 return WINED3DERR_INVALIDCALL;
3173 }
3174
3175 /* First pass: trace shader. */
3176 if (TRACE_ON(d3d_shader))
3177 shader_trace_init(fe, shader->frontend_data);
3178
3179 /* Second pass: figure out which registers are used, what the semantics are, etc. */
3180 if (FAILED(hr = shader_get_registers_used(shader, fe, reg_maps, &shader->input_signature,
3181 &shader->output_signature, float_const_count)))
3182 return hr;
3183
3184 if (reg_maps->shader_version.type != type)
3185 {
3186 WARN("Wrong shader type %d.\n", reg_maps->shader_version.type);
3187 return WINED3DERR_INVALIDCALL;
3188 }
3189 if (reg_maps->shader_version.major > max_version)
3190 {
3191 WARN("Shader version %d not supported by this D3D API version.\n", reg_maps->shader_version.major);
3192 return WINED3DERR_INVALIDCALL;
3193 }
3194 switch (type)
3195 {
3196 case WINED3D_SHADER_TYPE_VERTEX:
3197 backend_version = d3d_info->limits.vs_version;
3198 break;
3199 case WINED3D_SHADER_TYPE_HULL:
3200 backend_version = d3d_info->limits.hs_version;
3201 break;
3202 case WINED3D_SHADER_TYPE_DOMAIN:
3203 backend_version = d3d_info->limits.ds_version;
3204 break;
3205 case WINED3D_SHADER_TYPE_GEOMETRY:
3206 backend_version = d3d_info->limits.gs_version;
3207 break;
3208 case WINED3D_SHADER_TYPE_PIXEL:
3209 backend_version = d3d_info->limits.ps_version;
3210 break;
3211 case WINED3D_SHADER_TYPE_COMPUTE:
3212 backend_version = d3d_info->limits.cs_version;
3213 break;
3214 default:
3215 FIXME("No backend version-checking for this shader type.\n");
3216 backend_version = 0;
3217 }
3218 if (reg_maps->shader_version.major > backend_version)
3219 {
3220 WARN("Shader version %d.%d not supported by your GPU with the current shader backend.\n",
3221 reg_maps->shader_version.major, reg_maps->shader_version.minor);
3222 return WINED3DERR_INVALIDCALL;
3223 }
3224
3225 return WINED3D_OK;
3226 }
3227
3228 ULONG CDECL wined3d_shader_incref(struct wined3d_shader *shader)
3229 {
3230 ULONG refcount = InterlockedIncrement(&shader->ref);
3231
3232 TRACE("%p increasing refcount to %u.\n", shader, refcount);
3233
3234 return refcount;
3235 }
3236
3237 static void wined3d_shader_init_object(void *object)
3238 {
3239 struct wined3d_shader *shader = object;
3240 struct wined3d_device *device = shader->device;
3241
3242 list_add_head(&device->shaders, &shader->shader_list_entry);
3243
3244 device->shader_backend->shader_precompile(device->shader_priv, shader);
3245 }
3246
3247 static void wined3d_shader_destroy_object(void *object)
3248 {
3249 shader_cleanup(object);
3250 HeapFree(GetProcessHeap(), 0, object);
3251 }
3252
3253 ULONG CDECL wined3d_shader_decref(struct wined3d_shader *shader)
3254 {
3255 ULONG refcount = InterlockedDecrement(&shader->ref);
3256
3257 TRACE("%p decreasing refcount to %u.\n", shader, refcount);
3258
3259 if (!refcount)
3260 {
3261 shader->parent_ops->wined3d_object_destroyed(shader->parent);
3262 wined3d_cs_destroy_object(shader->device->cs, wined3d_shader_destroy_object, shader);
3263 }
3264
3265 return refcount;
3266 }
3267
3268 void * CDECL wined3d_shader_get_parent(const struct wined3d_shader *shader)
3269 {
3270 TRACE("shader %p.\n", shader);
3271
3272 return shader->parent;
3273 }
3274
3275 HRESULT CDECL wined3d_shader_get_byte_code(const struct wined3d_shader *shader,
3276 void *byte_code, UINT *byte_code_size)
3277 {
3278 TRACE("shader %p, byte_code %p, byte_code_size %p.\n", shader, byte_code, byte_code_size);
3279
3280 if (!byte_code)
3281 {
3282 *byte_code_size = shader->functionLength;
3283 return WINED3D_OK;
3284 }
3285
3286 if (*byte_code_size < shader->functionLength)
3287 {
3288 /* MSDN claims (for d3d8 at least) that if *byte_code_size is smaller
3289 * than the required size we should write the required size and
3290 * return D3DERR_MOREDATA. That's not actually true. */
3291 return WINED3DERR_INVALIDCALL;
3292 }
3293
3294 memcpy(byte_code, shader->function, shader->functionLength);
3295
3296 return WINED3D_OK;
3297 }
3298
3299 /* Set local constants for d3d8 shaders. */
3300 HRESULT CDECL wined3d_shader_set_local_constants_float(struct wined3d_shader *shader,
3301 UINT start_idx, const float *src_data, UINT count)
3302 {
3303 UINT end_idx = start_idx + count;
3304 UINT i;
3305
3306 TRACE("shader %p, start_idx %u, src_data %p, count %u.\n", shader, start_idx, src_data, count);
3307
3308 if (end_idx > shader->limits->constant_float)
3309 {
3310 WARN("end_idx %u > float constants limit %u.\n",
3311 end_idx, shader->limits->constant_float);
3312 end_idx = shader->limits->constant_float;
3313 }
3314
3315 for (i = start_idx; i < end_idx; ++i)
3316 {
3317 struct wined3d_shader_lconst *lconst = HeapAlloc(GetProcessHeap(), 0, sizeof(*lconst));
3318 float *value;
3319 if (!lconst)
3320 return E_OUTOFMEMORY;
3321
3322 lconst->idx = i;
3323 value = (float *)lconst->value;
3324 memcpy(value, src_data + (i - start_idx) * 4 /* 4 components */, 4 * sizeof(float));
3325 list_add_head(&shader->constantsF, &lconst->entry);
3326
3327 if (isinf(value[0]) || isnan(value[0]) || isinf(value[1]) || isnan(value[1])
3328 || isinf(value[2]) || isnan(value[2]) || isinf(value[3]) || isnan(value[3]))
3329 {
3330 shader->lconst_inf_or_nan = TRUE;
3331 }
3332 }
3333
3334 return WINED3D_OK;
3335 }
3336
3337 void find_vs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3338 WORD swizzle_map, struct vs_compile_args *args, const struct wined3d_d3d_info *d3d_info)
3339 {
3340 args->fog_src = state->render_states[WINED3D_RS_FOGTABLEMODE]
3341 == WINED3D_FOG_NONE ? VS_FOG_COORD : VS_FOG_Z;
3342 args->clip_enabled = state->render_states[WINED3D_RS_CLIPPING]
3343 && state->render_states[WINED3D_RS_CLIPPLANEENABLE];
3344 args->point_size = state->gl_primitive_type == GL_POINTS;
3345 args->per_vertex_point_size = shader->reg_maps.point_size;
3346 args->next_shader_type = state->shader[WINED3D_SHADER_TYPE_HULL] ? WINED3D_SHADER_TYPE_HULL
3347 : state->shader[WINED3D_SHADER_TYPE_GEOMETRY] ? WINED3D_SHADER_TYPE_GEOMETRY
3348 : WINED3D_SHADER_TYPE_PIXEL;
3349 if (shader->reg_maps.shader_version.major >= 4)
3350 args->next_shader_input_count = state->shader[WINED3D_SHADER_TYPE_HULL]
3351 ? state->shader[WINED3D_SHADER_TYPE_HULL]->limits->packed_input
3352 : state->shader[WINED3D_SHADER_TYPE_GEOMETRY]
3353 ? state->shader[WINED3D_SHADER_TYPE_GEOMETRY]->limits->packed_input
3354 : state->shader[WINED3D_SHADER_TYPE_PIXEL]
3355 ? state->shader[WINED3D_SHADER_TYPE_PIXEL]->limits->packed_input : 0;
3356 else
3357 args->next_shader_input_count = 0;
3358 args->swizzle_map = swizzle_map;
3359 if (d3d_info->emulated_flatshading)
3360 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3361 else
3362 args->flatshading = 0;
3363 }
3364
3365 static BOOL match_usage(BYTE usage1, BYTE usage_idx1, BYTE usage2, BYTE usage_idx2)
3366 {
3367 if (usage_idx1 != usage_idx2)
3368 return FALSE;
3369 if (usage1 == usage2)
3370 return TRUE;
3371 if (usage1 == WINED3D_DECL_USAGE_POSITION && usage2 == WINED3D_DECL_USAGE_POSITIONT)
3372 return TRUE;
3373 if (usage2 == WINED3D_DECL_USAGE_POSITION && usage1 == WINED3D_DECL_USAGE_POSITIONT)
3374 return TRUE;
3375
3376 return FALSE;
3377 }
3378
3379 BOOL vshader_get_input(const struct wined3d_shader *shader,
3380 BYTE usage_req, BYTE usage_idx_req, unsigned int *regnum)
3381 {
3382 WORD map = shader->reg_maps.input_registers;
3383 unsigned int i;
3384
3385 for (i = 0; map; map >>= 1, ++i)
3386 {
3387 if (!(map & 1)) continue;
3388
3389 if (match_usage(shader->u.vs.attributes[i].usage,
3390 shader->u.vs.attributes[i].usage_idx, usage_req, usage_idx_req))
3391 {
3392 *regnum = i;
3393 return TRUE;
3394 }
3395 }
3396 return FALSE;
3397 }
3398
3399 static HRESULT shader_signature_calculate_strings_length(const struct wined3d_shader_signature *signature,
3400 SIZE_T *total)
3401 {
3402 struct wined3d_shader_signature_element *e;
3403 unsigned int i;
3404 SIZE_T len;
3405
3406 for (i = 0; i < signature->element_count; ++i)
3407 {
3408 e = &signature->elements[i];
3409 len = strlen(e->semantic_name);
3410 if (len >= ~(SIZE_T)0 - *total)
3411 return E_OUTOFMEMORY;
3412
3413 *total += len + 1;
3414 }
3415 return WINED3D_OK;
3416 }
3417
3418 static HRESULT shader_signature_copy(struct wined3d_shader_signature *dst,
3419 const struct wined3d_shader_signature *src, char **signature_strings)
3420 {
3421 struct wined3d_shader_signature_element *e;
3422 unsigned int i;
3423 SIZE_T len;
3424 char *ptr;
3425
3426 if (!src->element_count)
3427 return WINED3D_OK;
3428
3429 ptr = *signature_strings;
3430
3431 dst->element_count = src->element_count;
3432 if (!(dst->elements = wined3d_calloc(dst->element_count, sizeof(*dst->elements))))
3433 return E_OUTOFMEMORY;
3434
3435 for (i = 0; i < src->element_count; ++i)
3436 {
3437 e = &src->elements[i];
3438 dst->elements[i] = *e;
3439
3440 len = strlen(e->semantic_name);
3441 memcpy(ptr, e->semantic_name, len + 1);
3442 dst->elements[i].semantic_name = ptr;
3443 ptr += len + 1;
3444 }
3445
3446 *signature_strings = ptr;
3447
3448 return WINED3D_OK;
3449 }
3450
3451 static HRESULT shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3452 const struct wined3d_shader_desc *desc, DWORD float_const_count, enum wined3d_shader_type type,
3453 void *parent, const struct wined3d_parent_ops *parent_ops)
3454 {
3455 size_t byte_code_size;
3456 SIZE_T total;
3457 HRESULT hr;
3458 char *ptr;
3459
3460 TRACE("byte_code %p, byte_code_size %#lx, format %#x, max_version %#x.\n",
3461 desc->byte_code, (long)desc->byte_code_size, desc->format, desc->max_version);
3462
3463 if (!desc->byte_code)
3464 return WINED3DERR_INVALIDCALL;
3465
3466 if (!(shader->frontend = shader_select_frontend(desc->format)))
3467 {
3468 FIXME("Unable to find frontend for shader.\n");
3469 return WINED3DERR_INVALIDCALL;
3470 }
3471
3472 shader->ref = 1;
3473 shader->device = device;
3474 shader->parent = parent;
3475 shader->parent_ops = parent_ops;
3476
3477 total = 0;
3478 if (FAILED(hr = shader_signature_calculate_strings_length(&desc->input_signature, &total)))
3479 return hr;
3480 if (FAILED(hr = shader_signature_calculate_strings_length(&desc->output_signature, &total)))
3481 return hr;
3482 if (FAILED(hr = shader_signature_calculate_strings_length(&desc->patch_constant_signature, &total)))
3483 return hr;
3484 if (total && !(shader->signature_strings = HeapAlloc(GetProcessHeap(), 0, total)))
3485 return E_OUTOFMEMORY;
3486 ptr = shader->signature_strings;
3487
3488 if (FAILED(hr = shader_signature_copy(&shader->input_signature, &desc->input_signature, &ptr)))
3489 {
3490 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3491 return hr;
3492 }
3493 if (FAILED(hr = shader_signature_copy(&shader->output_signature, &desc->output_signature, &ptr)))
3494 {
3495 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
3496 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3497 return hr;
3498 }
3499 if (FAILED(hr = shader_signature_copy(&shader->patch_constant_signature, &desc->patch_constant_signature, &ptr)))
3500 {
3501 HeapFree(GetProcessHeap(), 0, shader->output_signature.elements);
3502 HeapFree(GetProcessHeap(), 0, shader->input_signature.elements);
3503 HeapFree(GetProcessHeap(), 0, shader->signature_strings);
3504 return hr;
3505 }
3506
3507 list_init(&shader->linked_programs);
3508 list_init(&shader->constantsF);
3509 list_init(&shader->constantsB);
3510 list_init(&shader->constantsI);
3511 shader->lconst_inf_or_nan = FALSE;
3512 list_init(&shader->reg_maps.indexable_temps);
3513 list_init(&shader->shader_list_entry);
3514
3515 byte_code_size = desc->byte_code_size;
3516 if (byte_code_size == ~(size_t)0)
3517 {
3518 const struct wined3d_shader_frontend *fe = shader->frontend;
3519 struct wined3d_shader_version shader_version;
3520 struct wined3d_shader_instruction ins;
3521 const DWORD *ptr;
3522 void *fe_data;
3523
3524 if (!(fe_data = fe->shader_init(desc->byte_code, byte_code_size, &shader->output_signature)))
3525 {
3526 WARN("Failed to initialise frontend data.\n");
3527 shader_cleanup(shader);
3528 return WINED3DERR_INVALIDCALL;
3529 }
3530
3531 fe->shader_read_header(fe_data, &ptr, &shader_version);
3532 while (!fe->shader_is_end(fe_data, &ptr))
3533 fe->shader_read_instruction(fe_data, &ptr, &ins);
3534
3535 fe->shader_free(fe_data);
3536
3537 byte_code_size = (ptr - desc->byte_code) * sizeof(*ptr);
3538 }
3539
3540 if (!(shader->function = HeapAlloc(GetProcessHeap(), 0, byte_code_size)))
3541 {
3542 shader_cleanup(shader);
3543 return E_OUTOFMEMORY;
3544 }
3545 memcpy(shader->function, desc->byte_code, byte_code_size);
3546 shader->functionLength = byte_code_size;
3547
3548 if (FAILED(hr = shader_set_function(shader, float_const_count, type, desc->max_version)))
3549 {
3550 WARN("Failed to set function, hr %#x.\n", hr);
3551 shader_cleanup(shader);
3552 return hr;
3553 }
3554
3555 shader->load_local_constsF = shader->lconst_inf_or_nan;
3556
3557 wined3d_cs_init_object(shader->device->cs, wined3d_shader_init_object, shader);
3558
3559 return hr;
3560 }
3561
3562 static HRESULT vertex_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3563 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3564 {
3565 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3566 unsigned int i;
3567 HRESULT hr;
3568
3569 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.vs_uniform_count,
3570 WINED3D_SHADER_TYPE_VERTEX, parent, parent_ops)))
3571 return hr;
3572
3573 for (i = 0; i < shader->input_signature.element_count; ++i)
3574 {
3575 const struct wined3d_shader_signature_element *input = &shader->input_signature.elements[i];
3576
3577 if (!(reg_maps->input_registers & (1u << input->register_idx)) || !input->semantic_name)
3578 continue;
3579
3580 shader->u.vs.attributes[input->register_idx].usage =
3581 shader_usage_from_semantic_name(input->semantic_name);
3582 shader->u.vs.attributes[input->register_idx].usage_idx = input->semantic_idx;
3583 }
3584
3585 if (reg_maps->usesrelconstF && !list_empty(&shader->constantsF))
3586 shader->load_local_constsF = TRUE;
3587
3588 return WINED3D_OK;
3589 }
3590
3591 static HRESULT domain_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3592 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3593 {
3594 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_DOMAIN, parent, parent_ops);
3595 }
3596
3597 static HRESULT hull_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3598 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3599 {
3600 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_HULL, parent, parent_ops);
3601 }
3602
3603 static HRESULT geometry_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3604 const struct wined3d_shader_desc *desc, const struct wined3d_stream_output_desc *so_desc,
3605 void *parent, const struct wined3d_parent_ops *parent_ops)
3606 {
3607 struct wined3d_stream_output_element *elements = NULL;
3608 HRESULT hr;
3609
3610 if (so_desc && !(elements = wined3d_calloc(so_desc->element_count, sizeof(*elements))))
3611 return E_OUTOFMEMORY;
3612
3613 if (FAILED(hr = shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_GEOMETRY, parent, parent_ops)))
3614 {
3615 HeapFree(GetProcessHeap(), 0, elements);
3616 return hr;
3617 }
3618
3619 if (so_desc)
3620 {
3621 shader->u.gs.so_desc = *so_desc;
3622 shader->u.gs.so_desc.elements = elements;
3623 memcpy(elements, so_desc->elements, so_desc->element_count * sizeof(*elements));
3624 }
3625
3626 return WINED3D_OK;
3627 }
3628
3629 void find_ds_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3630 struct ds_compile_args *args, const struct wined3d_context *context)
3631 {
3632 const struct wined3d_shader *hull_shader = state->shader[WINED3D_SHADER_TYPE_HULL];
3633
3634 args->tessellator_output_primitive = hull_shader->u.hs.tessellator_output_primitive;
3635 args->tessellator_partitioning = hull_shader->u.hs.tessellator_partitioning;
3636
3637 args->output_count = state->shader[WINED3D_SHADER_TYPE_GEOMETRY] ?
3638 state->shader[WINED3D_SHADER_TYPE_GEOMETRY]->limits->packed_input :
3639 state->shader[WINED3D_SHADER_TYPE_PIXEL] ? state->shader[WINED3D_SHADER_TYPE_PIXEL]->limits->packed_input
3640 : shader->limits->packed_output;
3641 args->next_shader_type = state->shader[WINED3D_SHADER_TYPE_GEOMETRY] ? WINED3D_SHADER_TYPE_GEOMETRY
3642 : WINED3D_SHADER_TYPE_PIXEL;
3643
3644 args->render_offscreen = context->render_offscreen;
3645
3646 args->padding = 0;
3647 }
3648
3649 void find_gs_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3650 struct gs_compile_args *args)
3651 {
3652 args->output_count = state->shader[WINED3D_SHADER_TYPE_PIXEL]
3653 ? state->shader[WINED3D_SHADER_TYPE_PIXEL]->limits->packed_input : shader->limits->packed_output;
3654 }
3655
3656 void find_ps_compile_args(const struct wined3d_state *state, const struct wined3d_shader *shader,
3657 BOOL position_transformed, struct ps_compile_args *args, const struct wined3d_context *context)
3658 {
3659 const struct wined3d_d3d_info *d3d_info = context->d3d_info;
3660 const struct wined3d_gl_info *gl_info = context->gl_info;
3661 const struct wined3d_texture *texture;
3662 unsigned int i;
3663
3664 memset(args, 0, sizeof(*args)); /* FIXME: Make sure all bits are set. */
3665 if (!gl_info->supported[ARB_FRAMEBUFFER_SRGB] && needs_srgb_write(context, state, state->fb))
3666 {
3667 static unsigned int warned = 0;
3668
3669 args->srgb_correction = 1;
3670 if (state->render_states[WINED3D_RS_ALPHABLENDENABLE] && !warned++)
3671 WARN("Blending into a sRGB render target with no GL_ARB_framebuffer_sRGB "
3672 "support, expect rendering artifacts.\n");
3673 }
3674
3675 if (shader->reg_maps.shader_version.major == 1
3676 && shader->reg_maps.shader_version.minor <= 3)
3677 {
3678 for (i = 0; i < shader->limits->sampler; ++i)
3679 {
3680 DWORD flags = state->texture_states[i][WINED3D_TSS_TEXTURE_TRANSFORM_FLAGS];
3681
3682 if (flags & WINED3D_TTFF_PROJECTED)
3683 {
3684 DWORD tex_transform = flags & ~WINED3D_TTFF_PROJECTED;
3685
3686 if (!state->shader[WINED3D_SHADER_TYPE_VERTEX])
3687 {
3688 enum wined3d_shader_resource_type resource_type = shader->reg_maps.resource_info[i].type;
3689 unsigned int j;
3690 unsigned int index = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3691 DWORD max_valid = WINED3D_TTFF_COUNT4;
3692
3693 for (j = 0; j < state->vertex_declaration->element_count; ++j)
3694 {
3695 struct wined3d_vertex_declaration_element *element =
3696 &state->vertex_declaration->elements[j];
3697
3698 if (element->usage == WINED3D_DECL_USAGE_TEXCOORD
3699 && element->usage_idx == index)
3700 {
3701 max_valid = element->format->component_count;
3702 break;
3703 }
3704 }
3705 if (!tex_transform || tex_transform > max_valid)
3706 {
3707 WARN("Fixing up projected texture transform flags from %#x to %#x.\n",
3708 tex_transform, max_valid);
3709 tex_transform = max_valid;
3710 }
3711 if ((resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_1D && tex_transform > WINED3D_TTFF_COUNT1)
3712 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_2D
3713 && tex_transform > WINED3D_TTFF_COUNT2)
3714 || (resource_type == WINED3D_SHADER_RESOURCE_TEXTURE_3D
3715 && tex_transform > WINED3D_TTFF_COUNT3))
3716 tex_transform |= WINED3D_PSARGS_PROJECTED;
3717 else
3718 {
3719 WARN("Application requested projected texture with unsuitable texture coordinates.\n");
3720 WARN("(texture unit %u, transform flags %#x, sampler type %u).\n",
3721 i, tex_transform, resource_type);
3722 }
3723 }
3724 else
3725 tex_transform = WINED3D_TTFF_COUNT4 | WINED3D_PSARGS_PROJECTED;
3726
3727 args->tex_transform |= tex_transform << i * WINED3D_PSARGS_TEXTRANSFORM_SHIFT;
3728 }
3729 }
3730 }
3731 if (shader->reg_maps.shader_version.major == 1
3732 && shader->reg_maps.shader_version.minor <= 4)
3733 {
3734 for (i = 0; i < shader->limits->sampler; ++i)
3735 {
3736 const struct wined3d_texture *texture = state->textures[i];
3737
3738 if (!shader->reg_maps.resource_info[i].type)
3739 continue;
3740
3741 /* Treat unbound textures as 2D. The dummy texture will provide
3742 * the proper sample value. The tex_types bitmap defaults to
3743 * 2D because of the memset. */
3744 if (!texture)
3745 continue;
3746
3747 switch (texture->target)
3748 {
3749 /* RECT textures are distinguished from 2D textures via np2_fixup */
3750 case GL_TEXTURE_RECTANGLE_ARB:
3751 case GL_TEXTURE_2D:
3752 break;
3753
3754 case GL_TEXTURE_3D:
3755 args->tex_types |= WINED3D_SHADER_TEX_3D << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3756 break;
3757
3758 case GL_TEXTURE_CUBE_MAP_ARB:
3759 args->tex_types |= WINED3D_SHADER_TEX_CUBE << i * WINED3D_PSARGS_TEXTYPE_SHIFT;
3760 break;
3761 }
3762 }
3763 }
3764
3765 if (shader->reg_maps.shader_version.major >= 4)
3766 {
3767 /* In SM4+ we use dcl_sampler in order to determine if we should use shadow sampler. */
3768 args->shadow = 0;
3769 for (i = 0 ; i < MAX_FRAGMENT_SAMPLERS; ++i)
3770 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3771 args->np2_fixup = 0;
3772 }
3773 else
3774 {
3775 for (i = 0; i < MAX_FRAGMENT_SAMPLERS; ++i)
3776 {
3777 if (!shader->reg_maps.resource_info[i].type)
3778 continue;
3779
3780 texture = state->textures[i];
3781 if (!texture)
3782 {
3783 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3784 continue;
3785 }
3786 if (can_use_texture_swizzle(gl_info, texture->resource.format))
3787 args->color_fixup[i] = COLOR_FIXUP_IDENTITY;
3788 else
3789 args->color_fixup[i] = texture->resource.format->color_fixup;
3790
3791 if (texture->resource.format_flags & WINED3DFMT_FLAG_SHADOW)
3792 args->shadow |= 1u << i;
3793
3794 /* Flag samplers that need NP2 texcoord fixup. */
3795 if (!(texture->flags & WINED3D_TEXTURE_POW2_MAT_IDENT))
3796 args->np2_fixup |= (1u << i);
3797 }
3798 }
3799
3800 if (shader->reg_maps.shader_version.major >= 3)
3801 {
3802 if (position_transformed)
3803 args->vp_mode = pretransformed;
3804 else if (use_vs(state))
3805 args->vp_mode = vertexshader;
3806 else
3807 args->vp_mode = fixedfunction;
3808 args->fog = WINED3D_FFP_PS_FOG_OFF;
3809 }
3810 else
3811 {
3812 args->vp_mode = vertexshader;
3813 if (state->render_states[WINED3D_RS_FOGENABLE])
3814 {
3815 switch (state->render_states[WINED3D_RS_FOGTABLEMODE])
3816 {
3817 case WINED3D_FOG_NONE:
3818 if (position_transformed || use_vs(state))
3819 {
3820 args->fog = WINED3D_FFP_PS_FOG_LINEAR;
3821 break;
3822 }
3823
3824 switch (state->render_states[WINED3D_RS_FOGVERTEXMODE])
3825 {
3826 case WINED3D_FOG_NONE: /* Fall through. */
3827 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3828 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3829 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3830 }
3831 break;
3832
3833 case WINED3D_FOG_LINEAR: args->fog = WINED3D_FFP_PS_FOG_LINEAR; break;
3834 case WINED3D_FOG_EXP: args->fog = WINED3D_FFP_PS_FOG_EXP; break;
3835 case WINED3D_FOG_EXP2: args->fog = WINED3D_FFP_PS_FOG_EXP2; break;
3836 }
3837 }
3838 else
3839 {
3840 args->fog = WINED3D_FFP_PS_FOG_OFF;
3841 }
3842 }
3843
3844 if (context->d3d_info->limits.varying_count < wined3d_max_compat_varyings(context->gl_info))
3845 {
3846 const struct wined3d_shader *vs = state->shader[WINED3D_SHADER_TYPE_VERTEX];
3847
3848 args->texcoords_initialized = 0;
3849 for (i = 0; i < MAX_TEXTURES; ++i)
3850 {
3851 if (vs)
3852 {
3853 if (state->shader[WINED3D_SHADER_TYPE_VERTEX]->reg_maps.output_registers & (1u << i))
3854 args->texcoords_initialized |= 1u << i;
3855 }
3856 else
3857 {
3858 const struct wined3d_stream_info *si = &context->stream_info;
3859 unsigned int coord_idx = state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX];
3860
3861 if ((state->texture_states[i][WINED3D_TSS_TEXCOORD_INDEX] >> WINED3D_FFP_TCI_SHIFT)
3862 & WINED3D_FFP_TCI_MASK
3863 || (coord_idx < MAX_TEXTURES && (si->use_map & (1u << (WINED3D_FFP_TEXCOORD0 + coord_idx)))))
3864 args->texcoords_initialized |= 1u << i;
3865 }
3866 }
3867 }
3868 else
3869 {
3870 args->texcoords_initialized = (1u << MAX_TEXTURES) - 1;
3871 }
3872
3873 args->pointsprite = state->render_states[WINED3D_RS_POINTSPRITEENABLE]
3874 && state->gl_primitive_type == GL_POINTS;
3875
3876 if (gl_info->supported[WINED3D_GL_LEGACY_CONTEXT])
3877 args->alpha_test_func = WINED3D_CMP_ALWAYS - 1;
3878 else
3879 args->alpha_test_func = (state->render_states[WINED3D_RS_ALPHATESTENABLE]
3880 ? wined3d_sanitize_cmp_func(state->render_states[WINED3D_RS_ALPHAFUNC])
3881 : WINED3D_CMP_ALWAYS) - 1;
3882
3883 if (d3d_info->emulated_flatshading)
3884 args->flatshading = state->render_states[WINED3D_RS_SHADEMODE] == WINED3D_SHADE_FLAT;
3885
3886 args->render_offscreen = shader->reg_maps.vpos && gl_info->supported[ARB_FRAGMENT_COORD_CONVENTIONS]
3887 ? context->render_offscreen : 0;
3888 }
3889
3890 static HRESULT pixel_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3891 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3892 {
3893 const struct wined3d_gl_info *gl_info = &device->adapter->gl_info;
3894 unsigned int i, highest_reg_used = 0, num_regs_used = 0;
3895 HRESULT hr;
3896
3897 if (FAILED(hr = shader_init(shader, device, desc, device->adapter->d3d_info.limits.ps_uniform_count,
3898 WINED3D_SHADER_TYPE_PIXEL, parent, parent_ops)))
3899 return hr;
3900
3901 for (i = 0; i < MAX_REG_INPUT; ++i)
3902 {
3903 if (shader->u.ps.input_reg_used & (1u << i))
3904 {
3905 ++num_regs_used;
3906 highest_reg_used = i;
3907 }
3908 }
3909
3910 /* Don't do any register mapping magic if it is not needed, or if we can't
3911 * achieve anything anyway */
3912 if (highest_reg_used < (gl_info->limits.glsl_varyings / 4)
3913 || num_regs_used > (gl_info->limits.glsl_varyings / 4)
3914 || shader->reg_maps.shader_version.major >= 4)
3915 {
3916 if (num_regs_used > (gl_info->limits.glsl_varyings / 4))
3917 {
3918 /* This happens with relative addressing. The input mapper function
3919 * warns about this if the higher registers are declared too, so
3920 * don't write a FIXME here */
3921 WARN("More varying registers used than supported\n");
3922 }
3923
3924 for (i = 0; i < MAX_REG_INPUT; ++i)
3925 {
3926 shader->u.ps.input_reg_map[i] = i;
3927 }
3928
3929 shader->u.ps.declared_in_count = highest_reg_used + 1;
3930 }
3931 else
3932 {
3933 shader->u.ps.declared_in_count = 0;
3934 for (i = 0; i < MAX_REG_INPUT; ++i)
3935 {
3936 if (shader->u.ps.input_reg_used & (1u << i))
3937 shader->u.ps.input_reg_map[i] = shader->u.ps.declared_in_count++;
3938 else shader->u.ps.input_reg_map[i] = ~0U;
3939 }
3940 }
3941
3942 return WINED3D_OK;
3943 }
3944
3945 void pixelshader_update_resource_types(struct wined3d_shader *shader, WORD tex_types)
3946 {
3947 struct wined3d_shader_reg_maps *reg_maps = &shader->reg_maps;
3948 struct wined3d_shader_resource_info *resource_info = reg_maps->resource_info;
3949 unsigned int i;
3950
3951 if (reg_maps->shader_version.major != 1) return;
3952
3953 for (i = 0; i < shader->limits->sampler; ++i)
3954 {
3955 /* We don't sample from this sampler. */
3956 if (!resource_info[i].type)
3957 continue;
3958
3959 switch ((tex_types >> i * WINED3D_PSARGS_TEXTYPE_SHIFT) & WINED3D_PSARGS_TEXTYPE_MASK)
3960 {
3961 case WINED3D_SHADER_TEX_2D:
3962 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_2D;
3963 break;
3964
3965 case WINED3D_SHADER_TEX_3D:
3966 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_3D;
3967 break;
3968
3969 case WINED3D_SHADER_TEX_CUBE:
3970 resource_info[i].type = WINED3D_SHADER_RESOURCE_TEXTURE_CUBE;
3971 break;
3972 }
3973 }
3974 }
3975
3976 static HRESULT compute_shader_init(struct wined3d_shader *shader, struct wined3d_device *device,
3977 const struct wined3d_shader_desc *desc, void *parent, const struct wined3d_parent_ops *parent_ops)
3978 {
3979 return shader_init(shader, device, desc, 0, WINED3D_SHADER_TYPE_COMPUTE, parent, parent_ops);
3980 }
3981
3982 HRESULT CDECL wined3d_shader_create_cs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
3983 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
3984 {
3985 struct wined3d_shader *object;
3986 HRESULT hr;
3987
3988 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
3989 device, desc, parent, parent_ops, shader);
3990
3991 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
3992 return E_OUTOFMEMORY;
3993
3994 if (FAILED(hr = compute_shader_init(object, device, desc, parent, parent_ops)))
3995 {
3996 WARN("Failed to initialize compute shader, hr %#x.\n", hr);
3997 HeapFree(GetProcessHeap(), 0, object);
3998 return hr;
3999 }
4000
4001 TRACE("Created compute shader %p.\n", object);
4002 *shader = object;
4003
4004 return WINED3D_OK;
4005 }
4006
4007 HRESULT CDECL wined3d_shader_create_ds(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4008 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4009 {
4010 struct wined3d_shader *object;
4011 HRESULT hr;
4012
4013 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4014 device, desc, parent, parent_ops, shader);
4015
4016 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
4017 return E_OUTOFMEMORY;
4018
4019 if (FAILED(hr = domain_shader_init(object, device, desc, parent, parent_ops)))
4020 {
4021 WARN("Failed to initialize domain shader, hr %#x.\n", hr);
4022 HeapFree(GetProcessHeap(), 0, object);
4023 return hr;
4024 }
4025
4026 TRACE("Created domain shader %p.\n", object);
4027 *shader = object;
4028
4029 return WINED3D_OK;
4030 }
4031
4032 HRESULT CDECL wined3d_shader_create_gs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4033 const struct wined3d_stream_output_desc *so_desc, void *parent,
4034 const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4035 {
4036 struct wined3d_shader *object;
4037 HRESULT hr;
4038
4039 TRACE("device %p, desc %p, so_desc %p, parent %p, parent_ops %p, shader %p.\n",
4040 device, desc, so_desc, parent, parent_ops, shader);
4041
4042 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
4043 return E_OUTOFMEMORY;
4044
4045 if (FAILED(hr = geometry_shader_init(object, device, desc, so_desc, parent, parent_ops)))
4046 {
4047 WARN("Failed to initialize geometry shader, hr %#x.\n", hr);
4048 HeapFree(GetProcessHeap(), 0, object);
4049 return hr;
4050 }
4051
4052 TRACE("Created geometry shader %p.\n", object);
4053 *shader = object;
4054
4055 return WINED3D_OK;
4056 }
4057
4058 HRESULT CDECL wined3d_shader_create_hs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4059 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4060 {
4061 struct wined3d_shader *object;
4062 HRESULT hr;
4063
4064 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4065 device, desc, parent, parent_ops, shader);
4066
4067 if (!(object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object))))
4068 return E_OUTOFMEMORY;
4069
4070 if (FAILED(hr = hull_shader_init(object, device, desc, parent, parent_ops)))
4071 {
4072 WARN("Failed to initialize hull shader, hr %#x.\n", hr);
4073 HeapFree(GetProcessHeap(), 0, object);
4074 return hr;
4075 }
4076
4077 TRACE("Created hull shader %p.\n", object);
4078 *shader = object;
4079
4080 return WINED3D_OK;
4081 }
4082
4083 HRESULT CDECL wined3d_shader_create_ps(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4084 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4085 {
4086 struct wined3d_shader *object;
4087 HRESULT hr;
4088
4089 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4090 device, desc, parent, parent_ops, shader);
4091
4092 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
4093 if (!object)
4094 return E_OUTOFMEMORY;
4095
4096 if (FAILED(hr = pixel_shader_init(object, device, desc, parent, parent_ops)))
4097 {
4098 WARN("Failed to initialize pixel shader, hr %#x.\n", hr);
4099 HeapFree(GetProcessHeap(), 0, object);
4100 return hr;
4101 }
4102
4103 TRACE("Created pixel shader %p.\n", object);
4104 *shader = object;
4105
4106 return WINED3D_OK;
4107 }
4108
4109 HRESULT CDECL wined3d_shader_create_vs(struct wined3d_device *device, const struct wined3d_shader_desc *desc,
4110 void *parent, const struct wined3d_parent_ops *parent_ops, struct wined3d_shader **shader)
4111 {
4112 struct wined3d_shader *object;
4113 HRESULT hr;
4114
4115 TRACE("device %p, desc %p, parent %p, parent_ops %p, shader %p.\n",
4116 device, desc, parent, parent_ops, shader);
4117
4118 object = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*object));
4119 if (!object)
4120 return E_OUTOFMEMORY;
4121
4122 if (FAILED(hr = vertex_shader_init(object, device, desc, parent, parent_ops)))
4123 {
4124 WARN("Failed to initialize vertex shader, hr %#x.\n", hr);
4125 HeapFree(GetProcessHeap(), 0, object);
4126 return hr;
4127 }
4128
4129 TRACE("Created vertex shader %p.\n", object);
4130 *shader = object;
4131
4132 return WINED3D_OK;
4133 }
Windows API implementation