Merge remote-tracking branch 'public/master' into vulkan
[mesa.git] / src / gallium / auxiliary / gallivm / lp_bld_tgsi.h
1 /**************************************************************************
2 *
3 * Copyright 2011-2012 Advanced Micro Devices, Inc.
4 * Copyright 2009 VMware, Inc.
5 * 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 /**
30 * @file
31 * TGSI to LLVM IR translation.
32 *
33 * @author Jose Fonseca <jfonseca@vmware.com>
34 * @author Tom Stellard <thomas.stellard@amd.com>
35 */
36
37 #ifndef LP_BLD_TGSI_H
38 #define LP_BLD_TGSI_H
39
40 #include "gallivm/lp_bld.h"
41 #include "gallivm/lp_bld_tgsi_action.h"
42 #include "gallivm/lp_bld_limits.h"
43 #include "gallivm/lp_bld_sample.h"
44 #include "lp_bld_type.h"
45 #include "pipe/p_compiler.h"
46 #include "pipe/p_state.h"
47 #include "tgsi/tgsi_exec.h"
48 #include "tgsi/tgsi_scan.h"
49 #include "tgsi/tgsi_info.h"
50
51 #ifdef __cplusplus
52 extern "C" {
53 #endif
54
55 #define LP_CHAN_ALL ~0
56
57 #define LP_MAX_INSTRUCTIONS 256
58
59 struct tgsi_full_declaration;
60 struct tgsi_full_immediate;
61 struct tgsi_full_instruction;
62 struct tgsi_full_src_register;
63 struct tgsi_opcode_info;
64 struct tgsi_token;
65 struct tgsi_shader_info;
66 struct lp_build_mask_context;
67 struct gallivm_state;
68 struct lp_derivatives;
69 struct lp_build_tgsi_gs_iface;
70
71
72 enum lp_build_tex_modifier {
73 LP_BLD_TEX_MODIFIER_NONE = 0,
74 LP_BLD_TEX_MODIFIER_PROJECTED,
75 LP_BLD_TEX_MODIFIER_LOD_BIAS,
76 LP_BLD_TEX_MODIFIER_EXPLICIT_LOD,
77 LP_BLD_TEX_MODIFIER_EXPLICIT_DERIV,
78 LP_BLD_TEX_MODIFIER_LOD_ZERO
79 };
80
81
82 /**
83 * Describe a channel of a register.
84 *
85 * The value can be a:
86 * - immediate value (i.e. derived from a IMM register)
87 * - CONST[n].x/y/z/w
88 * - IN[n].x/y/z/w
89 * - undetermined (when .file == TGSI_FILE_NULL)
90 *
91 * This is one of the analysis results, and is used to described
92 * the output color in terms of inputs.
93 */
94 struct lp_tgsi_channel_info
95 {
96 unsigned file:4; /* TGSI_FILE_* */
97 unsigned swizzle:3; /* PIPE_SWIZZLE_x */
98 union {
99 uint32_t index;
100 float value; /* for TGSI_FILE_IMMEDIATE */
101 } u;
102 };
103
104
105 /**
106 * Describe a texture sampler interpolator.
107 *
108 * The interpolation is described in terms of regular inputs.
109 */
110 struct lp_tgsi_texture_info
111 {
112 struct lp_tgsi_channel_info coord[4];
113 unsigned target:8; /* TGSI_TEXTURE_* */
114 unsigned sampler_unit:8; /* Sampler unit */
115 unsigned texture_unit:8; /* Texture unit */
116 unsigned modifier:8; /* LP_BLD_TEX_MODIFIER_* */
117 };
118
119
120 struct lp_tgsi_info
121 {
122 struct tgsi_shader_info base;
123
124 /*
125 * Whether any of the texture opcodes access a register file other than
126 * TGSI_FILE_INPUT.
127 *
128 * We could also handle TGSI_FILE_CONST/IMMEDIATE here, but there is little
129 * benefit.
130 */
131 unsigned indirect_textures:1;
132
133 /*
134 * Whether any of the texture (sample) ocpodes use different sampler
135 * and sampler view unit.
136 */
137 unsigned sampler_texture_units_different:1;
138
139 /*
140 * Whether any immediate values are outside the range of 0 and 1
141 */
142 unsigned unclamped_immediates:1;
143
144 /*
145 * Texture opcode description. Aimed at detecting and described direct
146 * texture opcodes.
147 */
148 unsigned num_texs;
149 struct lp_tgsi_texture_info tex[PIPE_MAX_SAMPLERS];
150
151 /*
152 * Output description. Aimed at detecting and describing simple blit
153 * shaders.
154 */
155 struct lp_tgsi_channel_info output[PIPE_MAX_SHADER_OUTPUTS][4];
156
157 /*
158 * Shortcut pointers into the above (for fragment shaders).
159 */
160 const struct lp_tgsi_channel_info *cbuf[PIPE_MAX_COLOR_BUFS];
161 };
162
163 /**
164 * Reference to system values.
165 */
166 struct lp_bld_tgsi_system_values {
167 LLVMValueRef instance_id;
168 LLVMValueRef vertex_id;
169 LLVMValueRef vertex_id_nobase;
170 LLVMValueRef prim_id;
171 LLVMValueRef basevertex;
172 LLVMValueRef invocation_id;
173 };
174
175
176 /**
177 * Sampler code generation interface.
178 *
179 * Although texture sampling is a requirement for TGSI translation, it is
180 * a very different problem with several different approaches to it. This
181 * structure establishes an interface for texture sampling code generation, so
182 * that we can easily use different texture sampling strategies.
183 */
184 struct lp_build_sampler_soa
185 {
186 void
187 (*destroy)( struct lp_build_sampler_soa *sampler );
188
189 void
190 (*emit_tex_sample)(const struct lp_build_sampler_soa *sampler,
191 struct gallivm_state *gallivm,
192 const struct lp_sampler_params *params);
193
194 void
195 (*emit_size_query)( const struct lp_build_sampler_soa *sampler,
196 struct gallivm_state *gallivm,
197 struct lp_type type,
198 unsigned unit,
199 unsigned target,
200 LLVMValueRef context_ptr,
201 boolean need_nr_mips,
202 enum lp_sampler_lod_property,
203 LLVMValueRef explicit_lod, /* optional */
204 LLVMValueRef *sizes_out);
205 };
206
207
208 struct lp_build_sampler_aos
209 {
210 LLVMValueRef
211 (*emit_fetch_texel)( struct lp_build_sampler_aos *sampler,
212 struct lp_build_context *bld,
213 unsigned target, /* TGSI_TEXTURE_* */
214 unsigned unit,
215 LLVMValueRef coords,
216 const struct lp_derivatives derivs,
217 enum lp_build_tex_modifier modifier);
218 };
219
220
221 void
222 lp_build_tgsi_info(const struct tgsi_token *tokens,
223 struct lp_tgsi_info *info);
224
225
226 void
227 lp_build_tgsi_soa(struct gallivm_state *gallivm,
228 const struct tgsi_token *tokens,
229 struct lp_type type,
230 struct lp_build_mask_context *mask,
231 LLVMValueRef consts_ptr,
232 LLVMValueRef const_sizes_ptr,
233 const struct lp_bld_tgsi_system_values *system_values,
234 const LLVMValueRef (*inputs)[4],
235 LLVMValueRef (*outputs)[4],
236 LLVMValueRef context_ptr,
237 LLVMValueRef thread_data_ptr,
238 struct lp_build_sampler_soa *sampler,
239 const struct tgsi_shader_info *info,
240 const struct lp_build_tgsi_gs_iface *gs_iface);
241
242
243 void
244 lp_build_tgsi_aos(struct gallivm_state *gallivm,
245 const struct tgsi_token *tokens,
246 struct lp_type type,
247 const unsigned char swizzles[4],
248 LLVMValueRef consts_ptr,
249 const LLVMValueRef *inputs,
250 LLVMValueRef *outputs,
251 struct lp_build_sampler_aos *sampler,
252 const struct tgsi_shader_info *info);
253
254
255 enum lp_exec_mask_break_type {
256 LP_EXEC_MASK_BREAK_TYPE_LOOP,
257 LP_EXEC_MASK_BREAK_TYPE_SWITCH
258 };
259
260
261 struct lp_exec_mask {
262 struct lp_build_context *bld;
263
264 boolean has_mask;
265 boolean ret_in_main;
266
267 LLVMTypeRef int_vec_type;
268
269 LLVMValueRef exec_mask;
270
271 LLVMValueRef ret_mask;
272 LLVMValueRef cond_mask;
273 LLVMValueRef switch_mask; /* current switch exec mask */
274 LLVMValueRef cont_mask;
275 LLVMValueRef break_mask;
276
277 struct function_ctx {
278 int pc;
279 LLVMValueRef ret_mask;
280
281 LLVMValueRef cond_stack[LP_MAX_TGSI_NESTING];
282 int cond_stack_size;
283
284 /* keep track if break belongs to switch or loop */
285 enum lp_exec_mask_break_type break_type_stack[LP_MAX_TGSI_NESTING];
286 enum lp_exec_mask_break_type break_type;
287
288 struct {
289 LLVMValueRef switch_val;
290 LLVMValueRef switch_mask;
291 LLVMValueRef switch_mask_default;
292 boolean switch_in_default;
293 unsigned switch_pc;
294 } switch_stack[LP_MAX_TGSI_NESTING];
295 int switch_stack_size;
296 LLVMValueRef switch_val;
297 LLVMValueRef switch_mask_default; /* reverse of switch mask used for default */
298 boolean switch_in_default; /* if switch exec is currently in default */
299 unsigned switch_pc; /* when used points to default or endswitch-1 */
300
301 LLVMValueRef loop_limiter;
302 LLVMBasicBlockRef loop_block;
303 LLVMValueRef break_var;
304 struct {
305 LLVMBasicBlockRef loop_block;
306 LLVMValueRef cont_mask;
307 LLVMValueRef break_mask;
308 LLVMValueRef break_var;
309 } loop_stack[LP_MAX_TGSI_NESTING];
310 int loop_stack_size;
311
312 } *function_stack;
313 int function_stack_size;
314 };
315
316 struct lp_build_tgsi_inst_list
317 {
318 struct tgsi_full_instruction *instructions;
319 uint max_instructions;
320 uint num_instructions;
321 };
322
323 unsigned lp_bld_tgsi_list_init(struct lp_build_tgsi_context * bld_base);
324
325
326 unsigned lp_bld_tgsi_add_instruction(
327 struct lp_build_tgsi_context * bld_base,
328 const struct tgsi_full_instruction *inst_to_add);
329
330
331 struct lp_build_tgsi_context;
332
333
334 typedef LLVMValueRef (*lp_build_emit_fetch_fn)(struct lp_build_tgsi_context *,
335 const struct tgsi_full_src_register *,
336 enum tgsi_opcode_type,
337 unsigned);
338
339 struct lp_build_tgsi_context
340 {
341 struct lp_build_context base;
342
343 struct lp_build_context uint_bld;
344 struct lp_build_context int_bld;
345
346 struct lp_build_context dbl_bld;
347 /** This array stores functions that are used to transform TGSI opcodes to
348 * LLVM instructions.
349 */
350 struct lp_build_tgsi_action op_actions[TGSI_OPCODE_LAST];
351
352 /* TGSI_OPCODE_RSQ is defined as 1 / sqrt( abs(src0.x) ), rsq_action
353 * should compute 1 / sqrt (src0.x) */
354 struct lp_build_tgsi_action rsq_action;
355
356 struct lp_build_tgsi_action sqrt_action;
357
358 struct lp_build_tgsi_action drsq_action;
359
360 struct lp_build_tgsi_action dsqrt_action;
361 const struct tgsi_shader_info *info;
362
363 lp_build_emit_fetch_fn emit_fetch_funcs[TGSI_FILE_COUNT];
364
365 LLVMValueRef (*emit_swizzle)(struct lp_build_tgsi_context *,
366 LLVMValueRef, unsigned, unsigned, unsigned, unsigned);
367
368
369 void (*emit_debug)(struct lp_build_tgsi_context *,
370 const struct tgsi_full_instruction *,
371 const struct tgsi_opcode_info *);
372
373 void (*emit_store)(struct lp_build_tgsi_context *,
374 const struct tgsi_full_instruction *,
375 const struct tgsi_opcode_info *,
376 LLVMValueRef dst[4]);
377
378 void (*emit_declaration)(struct lp_build_tgsi_context *,
379 const struct tgsi_full_declaration *decl);
380
381 void (*emit_immediate)(struct lp_build_tgsi_context *,
382 const struct tgsi_full_immediate *imm);
383
384
385 /* Allow the user to store data in this structure rather than passing it
386 * to every function. */
387 void * userdata;
388
389 boolean soa;
390
391 int pc;
392
393 struct tgsi_full_instruction *instructions;
394 uint max_instructions;
395 uint num_instructions;
396
397 /** This function allows the user to insert some instructions at the
398 * beginning of the program. It is optional and does not need to be
399 * implemented.
400 */
401 void (*emit_prologue)(struct lp_build_tgsi_context*);
402
403 /** This function allows the user to insert some instructions at the end of
404 * the program. This callback is intended to be used for emitting
405 * instructions to handle the export for the output registers, but it can
406 * be used for any purpose. Implementing this function is optiona, but
407 * recommended.
408 */
409 void (*emit_epilogue)(struct lp_build_tgsi_context*);
410 };
411
412 struct lp_build_tgsi_gs_iface
413 {
414 LLVMValueRef (*fetch_input)(const struct lp_build_tgsi_gs_iface *gs_iface,
415 struct lp_build_tgsi_context * bld_base,
416 boolean is_vindex_indirect,
417 LLVMValueRef vertex_index,
418 boolean is_aindex_indirect,
419 LLVMValueRef attrib_index,
420 LLVMValueRef swizzle_index);
421 void (*emit_vertex)(const struct lp_build_tgsi_gs_iface *gs_iface,
422 struct lp_build_tgsi_context * bld_base,
423 LLVMValueRef (*outputs)[4],
424 LLVMValueRef emitted_vertices_vec);
425 void (*end_primitive)(const struct lp_build_tgsi_gs_iface *gs_iface,
426 struct lp_build_tgsi_context * bld_base,
427 LLVMValueRef verts_per_prim_vec,
428 LLVMValueRef emitted_prims_vec);
429 void (*gs_epilogue)(const struct lp_build_tgsi_gs_iface *gs_iface,
430 struct lp_build_tgsi_context * bld_base,
431 LLVMValueRef total_emitted_vertices_vec,
432 LLVMValueRef emitted_prims_vec);
433 };
434
435 struct lp_build_tgsi_soa_context
436 {
437 struct lp_build_tgsi_context bld_base;
438
439 /* Builder for scalar elements of shader's data type (float) */
440 struct lp_build_context elem_bld;
441
442 const struct lp_build_tgsi_gs_iface *gs_iface;
443 LLVMValueRef emitted_prims_vec_ptr;
444 LLVMValueRef total_emitted_vertices_vec_ptr;
445 LLVMValueRef emitted_vertices_vec_ptr;
446 LLVMValueRef max_output_vertices_vec;
447
448 LLVMValueRef consts_ptr;
449 LLVMValueRef const_sizes_ptr;
450 LLVMValueRef consts[LP_MAX_TGSI_CONST_BUFFERS];
451 LLVMValueRef consts_sizes[LP_MAX_TGSI_CONST_BUFFERS];
452 const LLVMValueRef (*inputs)[TGSI_NUM_CHANNELS];
453 LLVMValueRef (*outputs)[TGSI_NUM_CHANNELS];
454 LLVMValueRef context_ptr;
455 LLVMValueRef thread_data_ptr;
456
457 const struct lp_build_sampler_soa *sampler;
458
459 struct tgsi_declaration_sampler_view sv[PIPE_MAX_SHADER_SAMPLER_VIEWS];
460
461 LLVMValueRef immediates[LP_MAX_INLINED_IMMEDIATES][TGSI_NUM_CHANNELS];
462 LLVMValueRef temps[LP_MAX_INLINED_TEMPS][TGSI_NUM_CHANNELS];
463 LLVMValueRef addr[LP_MAX_TGSI_ADDRS][TGSI_NUM_CHANNELS];
464 LLVMValueRef preds[LP_MAX_TGSI_PREDS][TGSI_NUM_CHANNELS];
465
466 /* We allocate/use this array of temps if (1 << TGSI_FILE_TEMPORARY) is
467 * set in the indirect_files field.
468 * The temps[] array above is unused then.
469 */
470 LLVMValueRef temps_array;
471
472 /* We allocate/use this array of output if (1 << TGSI_FILE_OUTPUT) is
473 * set in the indirect_files field.
474 * The outputs[] array above is unused then.
475 */
476 LLVMValueRef outputs_array;
477
478 /* We allocate/use this array of inputs if (1 << TGSI_FILE_INPUT) is
479 * set in the indirect_files field.
480 * The inputs[] array above is unused then.
481 */
482 LLVMValueRef inputs_array;
483
484 /* We allocate/use this array of temps if (1 << TGSI_FILE_IMMEDIATE) is
485 * set in the indirect_files field.
486 */
487 LLVMValueRef imms_array;
488
489
490 struct lp_bld_tgsi_system_values system_values;
491
492 /** bitmask indicating which register files are accessed indirectly */
493 unsigned indirect_files;
494
495 struct lp_build_mask_context *mask;
496 struct lp_exec_mask exec_mask;
497
498 uint num_immediates;
499 boolean use_immediates_array;
500 };
501
502 void
503 lp_emit_declaration_soa(
504 struct lp_build_tgsi_context *bld,
505 const struct tgsi_full_declaration *decl);
506
507 void lp_emit_immediate_soa(
508 struct lp_build_tgsi_context *bld_base,
509 const struct tgsi_full_immediate *imm);
510
511 boolean
512 lp_emit_instruction_soa(
513 struct lp_build_tgsi_soa_context *bld,
514 const struct tgsi_full_instruction *inst,
515 const struct tgsi_opcode_info *info);
516
517
518 LLVMValueRef
519 lp_get_temp_ptr_soa(
520 struct lp_build_tgsi_soa_context *bld,
521 unsigned index,
522 unsigned chan);
523
524 LLVMValueRef
525 lp_get_output_ptr(
526 struct lp_build_tgsi_soa_context *bld,
527 unsigned index,
528 unsigned chan);
529
530 struct lp_build_tgsi_aos_context
531 {
532 struct lp_build_tgsi_context bld_base;
533
534 /* Builder for integer masks and indices */
535 struct lp_build_context int_bld;
536
537 /*
538 * AoS swizzle used:
539 * - swizzles[0] = red index
540 * - swizzles[1] = green index
541 * - swizzles[2] = blue index
542 * - swizzles[3] = alpha index
543 */
544 unsigned char swizzles[4];
545 unsigned char inv_swizzles[4];
546
547 LLVMValueRef consts_ptr;
548 const LLVMValueRef *inputs;
549 LLVMValueRef *outputs;
550
551 struct lp_build_sampler_aos *sampler;
552
553 struct tgsi_declaration_sampler_view sv[PIPE_MAX_SHADER_SAMPLER_VIEWS];
554
555 LLVMValueRef immediates[LP_MAX_INLINED_IMMEDIATES];
556 LLVMValueRef temps[LP_MAX_INLINED_TEMPS];
557 LLVMValueRef addr[LP_MAX_TGSI_ADDRS];
558 LLVMValueRef preds[LP_MAX_TGSI_PREDS];
559
560 /* We allocate/use this array of temps if (1 << TGSI_FILE_TEMPORARY) is
561 * set in the indirect_files field.
562 * The temps[] array above is unused then.
563 */
564 LLVMValueRef temps_array;
565
566 /** bitmask indicating which register files are accessed indirectly */
567 unsigned indirect_files;
568
569 };
570
571 static inline struct lp_build_tgsi_soa_context *
572 lp_soa_context(struct lp_build_tgsi_context *bld_base)
573 {
574 return (struct lp_build_tgsi_soa_context *)bld_base;
575 }
576
577 static inline struct lp_build_tgsi_aos_context *
578 lp_aos_context(struct lp_build_tgsi_context *bld_base)
579 {
580 return (struct lp_build_tgsi_aos_context *)bld_base;
581 }
582
583 void
584 lp_emit_declaration_aos(
585 struct lp_build_tgsi_aos_context *bld,
586 const struct tgsi_full_declaration *decl);
587
588
589 boolean
590 lp_emit_instruction_aos(
591 struct lp_build_tgsi_aos_context *bld,
592 const struct tgsi_full_instruction *inst,
593 const struct tgsi_opcode_info *info,
594 int *pc);
595
596 void
597 lp_emit_store_aos(
598 struct lp_build_tgsi_aos_context *bld,
599 const struct tgsi_full_instruction *inst,
600 unsigned index,
601 LLVMValueRef value);
602
603 void lp_build_fetch_args(
604 struct lp_build_tgsi_context * bld_base,
605 struct lp_build_emit_data * emit_data);
606
607 LLVMValueRef
608 lp_build_tgsi_inst_llvm_aos(
609 struct lp_build_tgsi_context * bld_base,
610 const struct tgsi_full_instruction *inst);
611
612 void
613 lp_build_tgsi_intrinsic(
614 const struct lp_build_tgsi_action * action,
615 struct lp_build_tgsi_context * bld_base,
616 struct lp_build_emit_data * emit_data);
617
618 LLVMValueRef
619 lp_build_emit_llvm(
620 struct lp_build_tgsi_context *bld_base,
621 unsigned tgsi_opcode,
622 struct lp_build_emit_data * emit_data);
623
624 LLVMValueRef
625 lp_build_emit_llvm_unary(
626 struct lp_build_tgsi_context *bld_base,
627 unsigned tgsi_opcode,
628 LLVMValueRef arg0);
629
630 LLVMValueRef
631 lp_build_emit_llvm_binary(
632 struct lp_build_tgsi_context *bld_base,
633 unsigned tgsi_opcode,
634 LLVMValueRef arg0,
635 LLVMValueRef arg1);
636
637 LLVMValueRef
638 lp_build_emit_llvm_ternary(
639 struct lp_build_tgsi_context *bld_base,
640 unsigned tgsi_opcode,
641 LLVMValueRef arg0,
642 LLVMValueRef arg1,
643 LLVMValueRef arg2);
644
645 boolean
646 lp_build_tgsi_inst_llvm(
647 struct lp_build_tgsi_context * bld_base,
648 const struct tgsi_full_instruction *inst);
649
650 LLVMValueRef
651 lp_build_emit_fetch(
652 struct lp_build_tgsi_context *bld_base,
653 const struct tgsi_full_instruction *inst,
654 unsigned src_op,
655 const unsigned chan_index);
656
657
658 LLVMValueRef
659 lp_build_emit_fetch_texoffset(
660 struct lp_build_tgsi_context *bld_base,
661 const struct tgsi_full_instruction *inst,
662 unsigned tex_off_op,
663 const unsigned chan_index);
664
665 boolean
666 lp_build_tgsi_llvm(
667 struct lp_build_tgsi_context * bld_base,
668 const struct tgsi_token *tokens);
669
670 #ifdef __cplusplus
671 }
672 #endif
673
674 #endif /* LP_BLD_TGSI_H */