Merge remote-tracking branch 'public/master' into vulkan
[mesa.git] / src / gallium / auxiliary / tgsi / tgsi_exec.h
1 /**************************************************************************
2 *
3 * Copyright 2007-2008 VMware, Inc.
4 * All Rights Reserved.
5 * Copyright 2009-2010 VMware, Inc. All rights Reserved.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the
9 * "Software"), to deal in the Software without restriction, including
10 * without limitation the rights to use, copy, modify, merge, publish,
11 * distribute, sub license, and/or sell copies of the Software, and to
12 * permit persons to whom the Software is furnished to do so, subject to
13 * the following conditions:
14 *
15 * The above copyright notice and this permission notice (including the
16 * next paragraph) shall be included in all copies or substantial portions
17 * of the Software.
18 *
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
21 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
22 * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
23 * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
24 * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
25 * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 *
27 **************************************************************************/
28
29 #ifndef TGSI_EXEC_H
30 #define TGSI_EXEC_H
31
32 #include "pipe/p_compiler.h"
33 #include "pipe/p_state.h"
34 #include "pipe/p_shader_tokens.h"
35
36 #if defined __cplusplus
37 extern "C" {
38 #endif
39
40 #define TGSI_CHAN_X 0
41 #define TGSI_CHAN_Y 1
42 #define TGSI_CHAN_Z 2
43 #define TGSI_CHAN_W 3
44
45 #define TGSI_NUM_CHANNELS 4 /* R,G,B,A */
46 #define TGSI_QUAD_SIZE 4 /* 4 pixel/quad */
47
48 #define TGSI_FOR_EACH_CHANNEL( CHAN )\
49 for (CHAN = 0; CHAN < TGSI_NUM_CHANNELS; CHAN++)
50
51 #define TGSI_IS_DST0_CHANNEL_ENABLED( INST, CHAN )\
52 ((INST)->Dst[0].Register.WriteMask & (1 << (CHAN)))
53
54 #define TGSI_IF_IS_DST0_CHANNEL_ENABLED( INST, CHAN )\
55 if (TGSI_IS_DST0_CHANNEL_ENABLED( INST, CHAN ))
56
57 #define TGSI_FOR_EACH_DST0_ENABLED_CHANNEL( INST, CHAN )\
58 TGSI_FOR_EACH_CHANNEL( CHAN )\
59 TGSI_IF_IS_DST0_CHANNEL_ENABLED( INST, CHAN )
60
61
62 /**
63 * Registers may be treated as float, signed int or unsigned int.
64 */
65 union tgsi_exec_channel
66 {
67 float f[TGSI_QUAD_SIZE];
68 int i[TGSI_QUAD_SIZE];
69 unsigned u[TGSI_QUAD_SIZE];
70 };
71
72 /**
73 * A vector[RGBA] of channels[4 pixels]
74 */
75 struct tgsi_exec_vector
76 {
77 union tgsi_exec_channel xyzw[TGSI_NUM_CHANNELS];
78 };
79
80 /**
81 * For fragment programs, information for computing fragment input
82 * values from plane equation of the triangle/line.
83 */
84 struct tgsi_interp_coef
85 {
86 float a0[TGSI_NUM_CHANNELS]; /* in an xyzw layout */
87 float dadx[TGSI_NUM_CHANNELS];
88 float dady[TGSI_NUM_CHANNELS];
89 };
90
91 enum tgsi_sampler_control
92 {
93 TGSI_SAMPLER_LOD_NONE,
94 TGSI_SAMPLER_LOD_BIAS,
95 TGSI_SAMPLER_LOD_EXPLICIT,
96 TGSI_SAMPLER_LOD_ZERO,
97 TGSI_SAMPLER_DERIVS_EXPLICIT,
98 TGSI_SAMPLER_GATHER,
99 };
100
101 struct tgsi_image_params {
102 unsigned unit;
103 unsigned tgsi_tex_instr;
104 enum pipe_format format;
105 unsigned execmask;
106 };
107
108 struct tgsi_image {
109 /* image interfaces */
110 void (*load)(const struct tgsi_image *image,
111 const struct tgsi_image_params *params,
112 const int s[TGSI_QUAD_SIZE],
113 const int t[TGSI_QUAD_SIZE],
114 const int r[TGSI_QUAD_SIZE],
115 const int sample[TGSI_QUAD_SIZE],
116 float rgba[TGSI_NUM_CHANNELS][TGSI_QUAD_SIZE]);
117
118 void (*store)(const struct tgsi_image *image,
119 const struct tgsi_image_params *params,
120 const int s[TGSI_QUAD_SIZE],
121 const int t[TGSI_QUAD_SIZE],
122 const int r[TGSI_QUAD_SIZE],
123 const int sample[TGSI_QUAD_SIZE],
124 float rgba[TGSI_NUM_CHANNELS][TGSI_QUAD_SIZE]);
125
126 void (*op)(const struct tgsi_image *image,
127 const struct tgsi_image_params *params,
128 unsigned opcode,
129 const int s[TGSI_QUAD_SIZE],
130 const int t[TGSI_QUAD_SIZE],
131 const int r[TGSI_QUAD_SIZE],
132 const int sample[TGSI_QUAD_SIZE],
133 float rgba[TGSI_NUM_CHANNELS][TGSI_QUAD_SIZE],
134 float rgba2[TGSI_NUM_CHANNELS][TGSI_QUAD_SIZE]);
135
136 void (*get_dims)(const struct tgsi_image *image,
137 const struct tgsi_image_params *params,
138 int dims[4]);
139 };
140
141 /**
142 * Information for sampling textures, which must be implemented
143 * by code outside the TGSI executor.
144 */
145 struct tgsi_sampler
146 {
147 /** Get samples for four fragments in a quad */
148 /* this interface contains 5 sets of channels that vary
149 * depending on the sampler.
150 * s - the first texture coordinate for sampling.
151 * t - the second texture coordinate for sampling - unused for 1D,
152 layer for 1D arrays.
153 * r - the third coordinate for sampling for 3D, cube, cube arrays,
154 * layer for 2D arrays. Compare value for 1D/2D shadows.
155 * c0 - Compare value for shadow cube and shadow 2d arrays,
156 * layer for cube arrays.
157 * derivs - explicit derivatives.
158 * offset - texel offsets
159 * lod - lod value, except for shadow cube arrays (compare value there).
160 */
161 void (*get_samples)(struct tgsi_sampler *sampler,
162 const unsigned sview_index,
163 const unsigned sampler_index,
164 const float s[TGSI_QUAD_SIZE],
165 const float t[TGSI_QUAD_SIZE],
166 const float r[TGSI_QUAD_SIZE],
167 const float c0[TGSI_QUAD_SIZE],
168 const float c1[TGSI_QUAD_SIZE],
169 float derivs[3][2][TGSI_QUAD_SIZE],
170 const int8_t offset[3],
171 enum tgsi_sampler_control control,
172 float rgba[TGSI_NUM_CHANNELS][TGSI_QUAD_SIZE]);
173 void (*get_dims)(struct tgsi_sampler *sampler,
174 const unsigned sview_index,
175 int level, int dims[4]);
176 void (*get_texel)(struct tgsi_sampler *sampler,
177 const unsigned sview_index,
178 const int i[TGSI_QUAD_SIZE],
179 const int j[TGSI_QUAD_SIZE], const int k[TGSI_QUAD_SIZE],
180 const int lod[TGSI_QUAD_SIZE], const int8_t offset[3],
181 float rgba[TGSI_NUM_CHANNELS][TGSI_QUAD_SIZE]);
182 void (*query_lod)(const struct tgsi_sampler *tgsi_sampler,
183 const unsigned sview_index,
184 const unsigned sampler_index,
185 const float s[TGSI_QUAD_SIZE],
186 const float t[TGSI_QUAD_SIZE],
187 const float p[TGSI_QUAD_SIZE],
188 const float c0[TGSI_QUAD_SIZE],
189 const enum tgsi_sampler_control control,
190 float mipmap[TGSI_QUAD_SIZE],
191 float lod[TGSI_QUAD_SIZE]);
192 };
193
194 #define TGSI_EXEC_NUM_TEMPS 4096
195 #define TGSI_EXEC_NUM_IMMEDIATES 256
196
197 /*
198 * Locations of various utility registers (_I = Index, _C = Channel)
199 */
200 #define TGSI_EXEC_TEMP_00000000_I (TGSI_EXEC_NUM_TEMPS + 0)
201 #define TGSI_EXEC_TEMP_00000000_C 0
202
203 #define TGSI_EXEC_TEMP_7FFFFFFF_I (TGSI_EXEC_NUM_TEMPS + 0)
204 #define TGSI_EXEC_TEMP_7FFFFFFF_C 1
205
206 #define TGSI_EXEC_TEMP_80000000_I (TGSI_EXEC_NUM_TEMPS + 0)
207 #define TGSI_EXEC_TEMP_80000000_C 2
208
209 #define TGSI_EXEC_TEMP_FFFFFFFF_I (TGSI_EXEC_NUM_TEMPS + 0)
210 #define TGSI_EXEC_TEMP_FFFFFFFF_C 3
211
212 #define TGSI_EXEC_TEMP_ONE_I (TGSI_EXEC_NUM_TEMPS + 1)
213 #define TGSI_EXEC_TEMP_ONE_C 0
214
215 #define TGSI_EXEC_TEMP_TWO_I (TGSI_EXEC_NUM_TEMPS + 1)
216 #define TGSI_EXEC_TEMP_TWO_C 1
217
218 #define TGSI_EXEC_TEMP_128_I (TGSI_EXEC_NUM_TEMPS + 1)
219 #define TGSI_EXEC_TEMP_128_C 2
220
221 #define TGSI_EXEC_TEMP_MINUS_128_I (TGSI_EXEC_NUM_TEMPS + 1)
222 #define TGSI_EXEC_TEMP_MINUS_128_C 3
223
224 #define TGSI_EXEC_TEMP_KILMASK_I (TGSI_EXEC_NUM_TEMPS + 2)
225 #define TGSI_EXEC_TEMP_KILMASK_C 0
226
227 #define TGSI_EXEC_TEMP_OUTPUT_I (TGSI_EXEC_NUM_TEMPS + 2)
228 #define TGSI_EXEC_TEMP_OUTPUT_C 1
229
230 #define TGSI_EXEC_TEMP_PRIMITIVE_I (TGSI_EXEC_NUM_TEMPS + 2)
231 #define TGSI_EXEC_TEMP_PRIMITIVE_C 2
232
233 #define TGSI_EXEC_TEMP_THREE_I (TGSI_EXEC_NUM_TEMPS + 2)
234 #define TGSI_EXEC_TEMP_THREE_C 3
235
236 #define TGSI_EXEC_TEMP_HALF_I (TGSI_EXEC_NUM_TEMPS + 3)
237 #define TGSI_EXEC_TEMP_HALF_C 0
238
239 /* 4 register buffer for various purposes */
240 #define TGSI_EXEC_TEMP_R0 (TGSI_EXEC_NUM_TEMPS + 4)
241 #define TGSI_EXEC_NUM_TEMP_R 4
242
243 #define TGSI_EXEC_TEMP_ADDR (TGSI_EXEC_NUM_TEMPS + 8)
244 #define TGSI_EXEC_NUM_ADDRS 3
245
246 /* predicate register */
247 #define TGSI_EXEC_TEMP_P0 (TGSI_EXEC_NUM_TEMPS + 11)
248 #define TGSI_EXEC_NUM_PREDS 1
249
250 #define TGSI_EXEC_NUM_TEMP_EXTRAS 12
251
252
253
254 #define TGSI_EXEC_MAX_NESTING 32
255 #define TGSI_EXEC_MAX_COND_NESTING TGSI_EXEC_MAX_NESTING
256 #define TGSI_EXEC_MAX_LOOP_NESTING TGSI_EXEC_MAX_NESTING
257 #define TGSI_EXEC_MAX_SWITCH_NESTING TGSI_EXEC_MAX_NESTING
258 #define TGSI_EXEC_MAX_CALL_NESTING TGSI_EXEC_MAX_NESTING
259
260 /* The maximum number of input attributes per vertex. For 2D
261 * input register files, this is the stride between two 1D
262 * arrays.
263 */
264 #define TGSI_EXEC_MAX_INPUT_ATTRIBS 32
265
266 /* The maximum number of bytes per constant buffer.
267 */
268 #define TGSI_EXEC_MAX_CONST_BUFFER_SIZE (4096 * sizeof(float[4]))
269
270 /* The maximum number of vertices per primitive */
271 #define TGSI_MAX_PRIM_VERTICES 6
272
273 /* The maximum number of primitives to be generated */
274 #define TGSI_MAX_PRIMITIVES 64
275
276 /* The maximum total number of vertices */
277 #define TGSI_MAX_TOTAL_VERTICES (TGSI_MAX_PRIM_VERTICES * TGSI_MAX_PRIMITIVES * PIPE_MAX_ATTRIBS)
278
279 #define TGSI_MAX_MISC_INPUTS 8
280
281 /** function call/activation record */
282 struct tgsi_call_record
283 {
284 uint CondStackTop;
285 uint LoopStackTop;
286 uint ContStackTop;
287 int SwitchStackTop;
288 int BreakStackTop;
289 uint ReturnAddr;
290 };
291
292
293 /* Switch-case block state. */
294 struct tgsi_switch_record {
295 uint mask; /**< execution mask */
296 union tgsi_exec_channel selector; /**< a value case statements are compared to */
297 uint defaultMask; /**< non-execute mask for default case */
298 };
299
300
301 enum tgsi_break_type {
302 TGSI_EXEC_BREAK_INSIDE_LOOP,
303 TGSI_EXEC_BREAK_INSIDE_SWITCH
304 };
305
306
307 #define TGSI_EXEC_MAX_BREAK_STACK (TGSI_EXEC_MAX_LOOP_NESTING + TGSI_EXEC_MAX_SWITCH_NESTING)
308
309
310 /**
311 * Run-time virtual machine state for executing TGSI shader.
312 */
313 struct tgsi_exec_machine
314 {
315 /* Total = program temporaries + internal temporaries
316 */
317 struct tgsi_exec_vector Temps[TGSI_EXEC_NUM_TEMPS +
318 TGSI_EXEC_NUM_TEMP_EXTRAS];
319
320 float Imms[TGSI_EXEC_NUM_IMMEDIATES][4];
321
322 float ImmArray[TGSI_EXEC_NUM_IMMEDIATES][4];
323
324 struct tgsi_exec_vector *Inputs;
325 struct tgsi_exec_vector *Outputs;
326
327 /* System values */
328 unsigned SysSemanticToIndex[TGSI_SEMANTIC_COUNT];
329 union tgsi_exec_channel SystemValue[TGSI_MAX_MISC_INPUTS];
330
331 struct tgsi_exec_vector *Addrs;
332 struct tgsi_exec_vector *Predicates;
333
334 struct tgsi_sampler *Sampler;
335
336 struct tgsi_image *Image;
337 unsigned ImmLimit;
338
339 const void *Consts[PIPE_MAX_CONSTANT_BUFFERS];
340 unsigned ConstsSize[PIPE_MAX_CONSTANT_BUFFERS];
341
342 const struct tgsi_token *Tokens; /**< Declarations, instructions */
343 unsigned Processor; /**< TGSI_PROCESSOR_x */
344
345 /* GEOMETRY processor only. */
346 unsigned *Primitives;
347 unsigned NumOutputs;
348 unsigned MaxGeometryShaderOutputs;
349 unsigned MaxOutputVertices;
350
351 /* FRAGMENT processor only. */
352 const struct tgsi_interp_coef *InterpCoefs;
353 struct tgsi_exec_vector QuadPos;
354 float Face; /**< +1 if front facing, -1 if back facing */
355 bool flatshade_color;
356
357 /* See GLSL 4.50 specification for definition of helper invocations */
358 uint NonHelperMask; /**< non-helpers */
359 /* Conditional execution masks */
360 uint CondMask; /**< For IF/ELSE/ENDIF */
361 uint LoopMask; /**< For BGNLOOP/ENDLOOP */
362 uint ContMask; /**< For loop CONT statements */
363 uint FuncMask; /**< For function calls */
364 uint ExecMask; /**< = CondMask & LoopMask */
365
366 /* Current switch-case state. */
367 struct tgsi_switch_record Switch;
368
369 /* Current break type. */
370 enum tgsi_break_type BreakType;
371
372 /** Condition mask stack (for nested conditionals) */
373 uint CondStack[TGSI_EXEC_MAX_COND_NESTING];
374 int CondStackTop;
375
376 /** Loop mask stack (for nested loops) */
377 uint LoopStack[TGSI_EXEC_MAX_LOOP_NESTING];
378 int LoopStackTop;
379
380 /** Loop label stack */
381 uint LoopLabelStack[TGSI_EXEC_MAX_LOOP_NESTING];
382 int LoopLabelStackTop;
383
384 /** Loop continue mask stack (see comments in tgsi_exec.c) */
385 uint ContStack[TGSI_EXEC_MAX_LOOP_NESTING];
386 int ContStackTop;
387
388 /** Switch case stack */
389 struct tgsi_switch_record SwitchStack[TGSI_EXEC_MAX_SWITCH_NESTING];
390 int SwitchStackTop;
391
392 enum tgsi_break_type BreakStack[TGSI_EXEC_MAX_BREAK_STACK];
393 int BreakStackTop;
394
395 /** Function execution mask stack (for executing subroutine code) */
396 uint FuncStack[TGSI_EXEC_MAX_CALL_NESTING];
397 int FuncStackTop;
398
399 /** Function call stack for saving/restoring the program counter */
400 struct tgsi_call_record CallStack[TGSI_EXEC_MAX_CALL_NESTING];
401 int CallStackTop;
402
403 struct tgsi_full_instruction *Instructions;
404 uint NumInstructions;
405
406 struct tgsi_full_declaration *Declarations;
407 uint NumDeclarations;
408
409 struct tgsi_declaration_sampler_view
410 SamplerViews[PIPE_MAX_SHADER_SAMPLER_VIEWS];
411
412 boolean UsedGeometryShader;
413 };
414
415 struct tgsi_exec_machine *
416 tgsi_exec_machine_create( void );
417
418 void
419 tgsi_exec_machine_destroy(struct tgsi_exec_machine *mach);
420
421
422 void
423 tgsi_exec_machine_bind_shader(
424 struct tgsi_exec_machine *mach,
425 const struct tgsi_token *tokens,
426 struct tgsi_sampler *sampler,
427 struct tgsi_image *image);
428
429 uint
430 tgsi_exec_machine_run(
431 struct tgsi_exec_machine *mach );
432
433
434 void
435 tgsi_exec_machine_free_data(struct tgsi_exec_machine *mach);
436
437
438 boolean
439 tgsi_check_soa_dependencies(const struct tgsi_full_instruction *inst);
440
441
442 extern void
443 tgsi_exec_set_constant_buffers(struct tgsi_exec_machine *mach,
444 unsigned num_bufs,
445 const void **bufs,
446 const unsigned *buf_sizes);
447
448
449 static inline int
450 tgsi_exec_get_shader_param(enum pipe_shader_cap param)
451 {
452 switch(param) {
453 case PIPE_SHADER_CAP_MAX_INSTRUCTIONS:
454 case PIPE_SHADER_CAP_MAX_ALU_INSTRUCTIONS:
455 case PIPE_SHADER_CAP_MAX_TEX_INSTRUCTIONS:
456 case PIPE_SHADER_CAP_MAX_TEX_INDIRECTIONS:
457 return INT_MAX;
458 case PIPE_SHADER_CAP_MAX_CONTROL_FLOW_DEPTH:
459 return TGSI_EXEC_MAX_NESTING;
460 case PIPE_SHADER_CAP_MAX_INPUTS:
461 return TGSI_EXEC_MAX_INPUT_ATTRIBS;
462 case PIPE_SHADER_CAP_MAX_OUTPUTS:
463 return 32;
464 case PIPE_SHADER_CAP_MAX_CONST_BUFFER_SIZE:
465 return TGSI_EXEC_MAX_CONST_BUFFER_SIZE;
466 case PIPE_SHADER_CAP_MAX_CONST_BUFFERS:
467 return PIPE_MAX_CONSTANT_BUFFERS;
468 case PIPE_SHADER_CAP_MAX_TEMPS:
469 return TGSI_EXEC_NUM_TEMPS;
470 case PIPE_SHADER_CAP_MAX_PREDS:
471 return TGSI_EXEC_NUM_PREDS;
472 case PIPE_SHADER_CAP_TGSI_CONT_SUPPORTED:
473 return 1;
474 case PIPE_SHADER_CAP_INDIRECT_INPUT_ADDR:
475 case PIPE_SHADER_CAP_INDIRECT_OUTPUT_ADDR:
476 case PIPE_SHADER_CAP_INDIRECT_TEMP_ADDR:
477 case PIPE_SHADER_CAP_INDIRECT_CONST_ADDR:
478 return 1;
479 case PIPE_SHADER_CAP_SUBROUTINES:
480 return 1;
481 case PIPE_SHADER_CAP_INTEGERS:
482 return 1;
483 case PIPE_SHADER_CAP_MAX_TEXTURE_SAMPLERS:
484 return PIPE_MAX_SAMPLERS;
485 case PIPE_SHADER_CAP_MAX_SAMPLER_VIEWS:
486 return PIPE_MAX_SHADER_SAMPLER_VIEWS;
487 case PIPE_SHADER_CAP_PREFERRED_IR:
488 return PIPE_SHADER_IR_TGSI;
489 case PIPE_SHADER_CAP_SUPPORTED_IRS:
490 return 1 << PIPE_SHADER_IR_TGSI;
491 case PIPE_SHADER_CAP_TGSI_SQRT_SUPPORTED:
492 return 1;
493 case PIPE_SHADER_CAP_DOUBLES:
494 case PIPE_SHADER_CAP_TGSI_DFRACEXP_DLDEXP_SUPPORTED:
495 case PIPE_SHADER_CAP_TGSI_ANY_INOUT_DECL_RANGE:
496 return 1;
497 case PIPE_SHADER_CAP_TGSI_DROUND_SUPPORTED:
498 case PIPE_SHADER_CAP_TGSI_FMA_SUPPORTED:
499 case PIPE_SHADER_CAP_MAX_SHADER_BUFFERS:
500 return 0;
501 case PIPE_SHADER_CAP_MAX_SHADER_IMAGES:
502 return PIPE_MAX_SHADER_IMAGES;
503
504 case PIPE_SHADER_CAP_MAX_UNROLL_ITERATIONS_HINT:
505 return 32;
506 }
507 /* if we get here, we missed a shader cap above (and should have seen
508 * a compiler warning.)
509 */
510 return 0;
511 }
512
513 #if defined __cplusplus
514 } /* extern "C" */
515 #endif
516
517 #endif /* TGSI_EXEC_H */