v3d: Drop shadow comparison state from shader variant key.
[mesa.git] / src / broadcom / compiler / v3d_compiler.h
1 /*
2 * Copyright © 2016 Broadcom
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24 #ifndef V3D_COMPILER_H
25 #define V3D_COMPILER_H
26
27 #include <assert.h>
28 #include <stdio.h>
29 #include <stdlib.h>
30 #include <stdbool.h>
31 #include <stdint.h>
32 #include <string.h>
33
34 #include "util/macros.h"
35 #include "common/v3d_debug.h"
36 #include "common/v3d_device_info.h"
37 #include "compiler/nir/nir.h"
38 #include "util/list.h"
39 #include "util/u_math.h"
40
41 #include "qpu/qpu_instr.h"
42 #include "pipe/p_state.h"
43
44 #define V3D_MAX_TEXTURE_SAMPLERS 32
45 #define V3D_MAX_SAMPLES 4
46 #define V3D_MAX_FS_INPUTS 64
47 #define V3D_MAX_VS_INPUTS 64
48
49 struct nir_builder;
50
51 struct v3d_fs_inputs {
52 /**
53 * Array of the meanings of the VPM inputs this shader needs.
54 *
55 * It doesn't include those that aren't part of the VPM, like
56 * point/line coordinates.
57 */
58 struct v3d_varying_slot *input_slots;
59 uint32_t num_inputs;
60 };
61
62 enum qfile {
63 /** An unused source or destination register. */
64 QFILE_NULL,
65
66 /** A physical register, such as the W coordinate payload. */
67 QFILE_REG,
68 /** One of the regsiters for fixed function interactions. */
69 QFILE_MAGIC,
70
71 /**
72 * A virtual register, that will be allocated to actual accumulator
73 * or physical registers later.
74 */
75 QFILE_TEMP,
76 QFILE_UNIF,
77 QFILE_TLB,
78 QFILE_TLBU,
79
80 /**
81 * VPM reads use this with an index value to say what part of the VPM
82 * is being read.
83 */
84 QFILE_VPM,
85
86 /**
87 * Stores an immediate value in the index field that will be used
88 * directly by qpu_load_imm().
89 */
90 QFILE_LOAD_IMM,
91
92 /**
93 * Stores an immediate value in the index field that can be turned
94 * into a small immediate field by qpu_encode_small_immediate().
95 */
96 QFILE_SMALL_IMM,
97 };
98
99 /**
100 * A reference to a QPU register or a virtual temp register.
101 */
102 struct qreg {
103 enum qfile file;
104 uint32_t index;
105 };
106
107 static inline struct qreg vir_reg(enum qfile file, uint32_t index)
108 {
109 return (struct qreg){file, index};
110 }
111
112 /**
113 * A reference to an actual register at the QPU level, for register
114 * allocation.
115 */
116 struct qpu_reg {
117 bool magic;
118 bool smimm;
119 int index;
120 };
121
122 struct qinst {
123 /** Entry in qblock->instructions */
124 struct list_head link;
125
126 /**
127 * The instruction being wrapped. Its condition codes, pack flags,
128 * signals, etc. will all be used, with just the register references
129 * being replaced by the contents of qinst->dst and qinst->src[].
130 */
131 struct v3d_qpu_instr qpu;
132
133 /* Pre-register-allocation references to src/dst registers */
134 struct qreg dst;
135 struct qreg src[3];
136 bool cond_is_exec_mask;
137 bool has_implicit_uniform;
138 bool is_last_thrsw;
139
140 /* After vir_to_qpu.c: If instr reads a uniform, which uniform from
141 * the uncompiled stream it is.
142 */
143 int uniform;
144 };
145
146 enum quniform_contents {
147 /**
148 * Indicates that a constant 32-bit value is copied from the program's
149 * uniform contents.
150 */
151 QUNIFORM_CONSTANT,
152 /**
153 * Indicates that the program's uniform contents are used as an index
154 * into the GL uniform storage.
155 */
156 QUNIFORM_UNIFORM,
157
158 /** @{
159 * Scaling factors from clip coordinates to relative to the viewport
160 * center.
161 *
162 * This is used by the coordinate and vertex shaders to produce the
163 * 32-bit entry consisting of 2 16-bit fields with 12.4 signed fixed
164 * point offsets from the viewport ccenter.
165 */
166 QUNIFORM_VIEWPORT_X_SCALE,
167 QUNIFORM_VIEWPORT_Y_SCALE,
168 /** @} */
169
170 QUNIFORM_VIEWPORT_Z_OFFSET,
171 QUNIFORM_VIEWPORT_Z_SCALE,
172
173 QUNIFORM_USER_CLIP_PLANE,
174
175 /**
176 * A reference to a V3D 3.x texture config parameter 0 uniform.
177 *
178 * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
179 * defines texture type, miplevels, and such. It will be found as a
180 * parameter to the first QOP_TEX_[STRB] instruction in a sequence.
181 */
182 QUNIFORM_TEXTURE_CONFIG_P0_0,
183 QUNIFORM_TEXTURE_CONFIG_P0_1,
184 QUNIFORM_TEXTURE_CONFIG_P0_2,
185 QUNIFORM_TEXTURE_CONFIG_P0_3,
186 QUNIFORM_TEXTURE_CONFIG_P0_4,
187 QUNIFORM_TEXTURE_CONFIG_P0_5,
188 QUNIFORM_TEXTURE_CONFIG_P0_6,
189 QUNIFORM_TEXTURE_CONFIG_P0_7,
190 QUNIFORM_TEXTURE_CONFIG_P0_8,
191 QUNIFORM_TEXTURE_CONFIG_P0_9,
192 QUNIFORM_TEXTURE_CONFIG_P0_10,
193 QUNIFORM_TEXTURE_CONFIG_P0_11,
194 QUNIFORM_TEXTURE_CONFIG_P0_12,
195 QUNIFORM_TEXTURE_CONFIG_P0_13,
196 QUNIFORM_TEXTURE_CONFIG_P0_14,
197 QUNIFORM_TEXTURE_CONFIG_P0_15,
198 QUNIFORM_TEXTURE_CONFIG_P0_16,
199 QUNIFORM_TEXTURE_CONFIG_P0_17,
200 QUNIFORM_TEXTURE_CONFIG_P0_18,
201 QUNIFORM_TEXTURE_CONFIG_P0_19,
202 QUNIFORM_TEXTURE_CONFIG_P0_20,
203 QUNIFORM_TEXTURE_CONFIG_P0_21,
204 QUNIFORM_TEXTURE_CONFIG_P0_22,
205 QUNIFORM_TEXTURE_CONFIG_P0_23,
206 QUNIFORM_TEXTURE_CONFIG_P0_24,
207 QUNIFORM_TEXTURE_CONFIG_P0_25,
208 QUNIFORM_TEXTURE_CONFIG_P0_26,
209 QUNIFORM_TEXTURE_CONFIG_P0_27,
210 QUNIFORM_TEXTURE_CONFIG_P0_28,
211 QUNIFORM_TEXTURE_CONFIG_P0_29,
212 QUNIFORM_TEXTURE_CONFIG_P0_30,
213 QUNIFORM_TEXTURE_CONFIG_P0_31,
214 QUNIFORM_TEXTURE_CONFIG_P0_32,
215
216 /**
217 * A reference to a V3D 3.x texture config parameter 1 uniform.
218 *
219 * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
220 * has the pointer to the indirect texture state. Our data[] field
221 * will have a packed p1 value, but the address field will be just
222 * which texture unit's texture should be referenced.
223 */
224 QUNIFORM_TEXTURE_CONFIG_P1,
225
226 /* A V3D 4.x texture config parameter. The high 8 bits will be
227 * which texture or sampler is being sampled, and the driver must
228 * replace the address field with the appropriate address.
229 */
230 QUNIFORM_TMU_CONFIG_P0,
231 QUNIFORM_TMU_CONFIG_P1,
232
233 QUNIFORM_TEXTURE_FIRST_LEVEL,
234
235 QUNIFORM_TEXTURE_WIDTH,
236 QUNIFORM_TEXTURE_HEIGHT,
237 QUNIFORM_TEXTURE_DEPTH,
238 QUNIFORM_TEXTURE_ARRAY_SIZE,
239 QUNIFORM_TEXTURE_LEVELS,
240
241 QUNIFORM_UBO_ADDR,
242
243 QUNIFORM_TEXRECT_SCALE_X,
244 QUNIFORM_TEXRECT_SCALE_Y,
245
246 QUNIFORM_ALPHA_REF,
247
248 /**
249 * Returns the the offset of the scratch buffer for register spilling.
250 */
251 QUNIFORM_SPILL_OFFSET,
252 QUNIFORM_SPILL_SIZE_PER_THREAD,
253 };
254
255 static inline uint32_t v3d_tmu_config_data_create(uint32_t unit, uint32_t value)
256 {
257 return unit << 24 | value;
258 }
259
260 static inline uint32_t v3d_tmu_config_data_get_unit(uint32_t data)
261 {
262 return data >> 24;
263 }
264
265 static inline uint32_t v3d_tmu_config_data_get_value(uint32_t data)
266 {
267 return data & 0xffffff;
268 }
269
270 struct v3d_varying_slot {
271 uint8_t slot_and_component;
272 };
273
274 static inline struct v3d_varying_slot
275 v3d_slot_from_slot_and_component(uint8_t slot, uint8_t component)
276 {
277 assert(slot < 255 / 4);
278 return (struct v3d_varying_slot){ (slot << 2) + component };
279 }
280
281 static inline uint8_t v3d_slot_get_slot(struct v3d_varying_slot slot)
282 {
283 return slot.slot_and_component >> 2;
284 }
285
286 static inline uint8_t v3d_slot_get_component(struct v3d_varying_slot slot)
287 {
288 return slot.slot_and_component & 3;
289 }
290
291 struct v3d_ubo_range {
292 /**
293 * offset in bytes from the start of the ubo where this range is
294 * uploaded.
295 *
296 * Only set once used is set.
297 */
298 uint32_t dst_offset;
299
300 /**
301 * offset in bytes from the start of the gallium uniforms where the
302 * data comes from.
303 */
304 uint32_t src_offset;
305
306 /** size in bytes of this ubo range */
307 uint32_t size;
308 };
309
310 struct v3d_key {
311 void *shader_state;
312 struct {
313 uint8_t swizzle[4];
314 uint8_t return_size;
315 uint8_t return_channels;
316 bool clamp_s:1;
317 bool clamp_t:1;
318 bool clamp_r:1;
319 } tex[V3D_MAX_TEXTURE_SAMPLERS];
320 uint8_t ucp_enables;
321 };
322
323 struct v3d_fs_key {
324 struct v3d_key base;
325 bool depth_enabled;
326 bool is_points;
327 bool is_lines;
328 bool alpha_test;
329 bool point_coord_upper_left;
330 bool light_twoside;
331 bool msaa;
332 bool sample_coverage;
333 bool sample_alpha_to_coverage;
334 bool sample_alpha_to_one;
335 bool clamp_color;
336 bool shade_model_flat;
337 uint8_t nr_cbufs;
338 uint8_t swap_color_rb;
339 /* Mask of which render targets need to be written as 32-bit floats */
340 uint8_t f32_color_rb;
341 /* Masks of which render targets need to be written as ints/uints.
342 * Used by gallium to work around lost information in TGSI.
343 */
344 uint8_t int_color_rb;
345 uint8_t uint_color_rb;
346 uint8_t alpha_test_func;
347 uint8_t logicop_func;
348 uint32_t point_sprite_mask;
349
350 struct pipe_rt_blend_state blend;
351 };
352
353 struct v3d_vs_key {
354 struct v3d_key base;
355
356 struct v3d_varying_slot fs_inputs[V3D_MAX_FS_INPUTS];
357 uint8_t num_fs_inputs;
358
359 bool is_coord;
360 bool per_vertex_point_size;
361 bool clamp_color;
362 };
363
364 /** A basic block of VIR intructions. */
365 struct qblock {
366 struct list_head link;
367
368 struct list_head instructions;
369
370 struct set *predecessors;
371 struct qblock *successors[2];
372
373 int index;
374
375 /* Instruction IPs for the first and last instruction of the block.
376 * Set by qpu_schedule.c.
377 */
378 uint32_t start_qpu_ip;
379 uint32_t end_qpu_ip;
380
381 /* Instruction IP for the branch instruction of the block. Set by
382 * qpu_schedule.c.
383 */
384 uint32_t branch_qpu_ip;
385
386 /** Offset within the uniform stream at the start of the block. */
387 uint32_t start_uniform;
388 /** Offset within the uniform stream of the branch instruction */
389 uint32_t branch_uniform;
390
391 /** @{ used by v3d_vir_live_variables.c */
392 BITSET_WORD *def;
393 BITSET_WORD *use;
394 BITSET_WORD *live_in;
395 BITSET_WORD *live_out;
396 int start_ip, end_ip;
397 /** @} */
398 };
399
400 /** Which util/list.h add mode we should use when inserting an instruction. */
401 enum vir_cursor_mode {
402 vir_cursor_add,
403 vir_cursor_addtail,
404 };
405
406 /**
407 * Tracking structure for where new instructions should be inserted. Create
408 * with one of the vir_after_inst()-style helper functions.
409 *
410 * This does not protect against removal of the block or instruction, so we
411 * have an assert in instruction removal to try to catch it.
412 */
413 struct vir_cursor {
414 enum vir_cursor_mode mode;
415 struct list_head *link;
416 };
417
418 static inline struct vir_cursor
419 vir_before_inst(struct qinst *inst)
420 {
421 return (struct vir_cursor){ vir_cursor_addtail, &inst->link };
422 }
423
424 static inline struct vir_cursor
425 vir_after_inst(struct qinst *inst)
426 {
427 return (struct vir_cursor){ vir_cursor_add, &inst->link };
428 }
429
430 static inline struct vir_cursor
431 vir_before_block(struct qblock *block)
432 {
433 return (struct vir_cursor){ vir_cursor_add, &block->instructions };
434 }
435
436 static inline struct vir_cursor
437 vir_after_block(struct qblock *block)
438 {
439 return (struct vir_cursor){ vir_cursor_addtail, &block->instructions };
440 }
441
442 /**
443 * Compiler state saved across compiler invocations, for any expensive global
444 * setup.
445 */
446 struct v3d_compiler {
447 const struct v3d_device_info *devinfo;
448 struct ra_regs *regs;
449 unsigned int reg_class_phys[3];
450 unsigned int reg_class_phys_or_acc[3];
451 };
452
453 struct v3d_compile {
454 const struct v3d_device_info *devinfo;
455 nir_shader *s;
456 nir_function_impl *impl;
457 struct exec_list *cf_node_list;
458 const struct v3d_compiler *compiler;
459
460 /**
461 * Mapping from nir_register * or nir_ssa_def * to array of struct
462 * qreg for the values.
463 */
464 struct hash_table *def_ht;
465
466 /* For each temp, the instruction generating its value. */
467 struct qinst **defs;
468 uint32_t defs_array_size;
469
470 /**
471 * Inputs to the shader, arranged by TGSI declaration order.
472 *
473 * Not all fragment shader QFILE_VARY reads are present in this array.
474 */
475 struct qreg *inputs;
476 struct qreg *outputs;
477 bool msaa_per_sample_output;
478 struct qreg color_reads[V3D_MAX_SAMPLES];
479 struct qreg sample_colors[V3D_MAX_SAMPLES];
480 uint32_t inputs_array_size;
481 uint32_t outputs_array_size;
482 uint32_t uniforms_array_size;
483
484 /* Booleans for whether the corresponding QFILE_VARY[i] is
485 * flat-shaded. This includes gl_FragColor flat-shading, which is
486 * customized based on the shademodel_flat shader key.
487 */
488 uint32_t flat_shade_flags[BITSET_WORDS(V3D_MAX_FS_INPUTS)];
489
490 uint32_t noperspective_flags[BITSET_WORDS(V3D_MAX_FS_INPUTS)];
491
492 uint32_t centroid_flags[BITSET_WORDS(V3D_MAX_FS_INPUTS)];
493
494 bool uses_center_w;
495
496 struct v3d_ubo_range *ubo_ranges;
497 bool *ubo_range_used;
498 uint32_t ubo_ranges_array_size;
499 /** Number of uniform areas tracked in ubo_ranges. */
500 uint32_t num_ubo_ranges;
501 uint32_t next_ubo_dst_offset;
502
503 /* State for whether we're executing on each channel currently. 0 if
504 * yes, otherwise a block number + 1 that the channel jumped to.
505 */
506 struct qreg execute;
507
508 struct qreg line_x, point_x, point_y;
509
510 /**
511 * Instance ID, which comes in before the vertex attribute payload if
512 * the shader record requests it.
513 */
514 struct qreg iid;
515
516 /**
517 * Vertex ID, which comes in before the vertex attribute payload
518 * (after Instance ID) if the shader record requests it.
519 */
520 struct qreg vid;
521
522 /* Fragment shader payload regs. */
523 struct qreg payload_w, payload_w_centroid, payload_z;
524
525 uint8_t vattr_sizes[V3D_MAX_VS_INPUTS];
526 uint32_t num_vpm_writes;
527
528 /* Size in bytes of registers that have been spilled. This is how much
529 * space needs to be available in the spill BO per thread per QPU.
530 */
531 uint32_t spill_size;
532 /* Shader-db stats for register spilling. */
533 uint32_t spills, fills;
534 /**
535 * Register spilling's per-thread base address, shared between each
536 * spill/fill's addressing calculations.
537 */
538 struct qreg spill_base;
539 /* Bit vector of which temps may be spilled */
540 BITSET_WORD *spillable;
541
542 /**
543 * Array of the VARYING_SLOT_* of all FS QFILE_VARY reads.
544 *
545 * This includes those that aren't part of the VPM varyings, like
546 * point/line coordinates.
547 */
548 struct v3d_varying_slot input_slots[V3D_MAX_FS_INPUTS];
549
550 /**
551 * An entry per outputs[] in the VS indicating what the VARYING_SLOT_*
552 * of the output is. Used to emit from the VS in the order that the
553 * FS needs.
554 */
555 struct v3d_varying_slot *output_slots;
556
557 struct pipe_shader_state *shader_state;
558 struct v3d_key *key;
559 struct v3d_fs_key *fs_key;
560 struct v3d_vs_key *vs_key;
561
562 /* Live ranges of temps. */
563 int *temp_start, *temp_end;
564 bool live_intervals_valid;
565
566 uint32_t *uniform_data;
567 enum quniform_contents *uniform_contents;
568 uint32_t uniform_array_size;
569 uint32_t num_uniforms;
570 uint32_t num_outputs;
571 uint32_t output_position_index;
572 nir_variable *output_color_var[4];
573 uint32_t output_point_size_index;
574 uint32_t output_sample_mask_index;
575
576 struct qreg undef;
577 uint32_t num_temps;
578
579 struct vir_cursor cursor;
580 struct list_head blocks;
581 int next_block_index;
582 struct qblock *cur_block;
583 struct qblock *loop_cont_block;
584 struct qblock *loop_break_block;
585
586 uint64_t *qpu_insts;
587 uint32_t qpu_inst_count;
588 uint32_t qpu_inst_size;
589
590 /* For the FS, the number of varying inputs not counting the
591 * point/line varyings payload
592 */
593 uint32_t num_inputs;
594
595 /**
596 * Number of inputs from num_inputs remaining to be queued to the read
597 * FIFO in the VS/CS.
598 */
599 uint32_t num_inputs_remaining;
600
601 /* Number of inputs currently in the read FIFO for the VS/CS */
602 uint32_t num_inputs_in_fifo;
603
604 /** Next offset in the VPM to read from in the VS/CS */
605 uint32_t vpm_read_offset;
606
607 uint32_t program_id;
608 uint32_t variant_id;
609
610 /* Set to compile program in in 1x, 2x, or 4x threaded mode, where
611 * SIG_THREAD_SWITCH is used to hide texturing latency at the cost of
612 * limiting ourselves to the part of the physical reg space.
613 *
614 * On V3D 3.x, 2x or 4x divide the physical reg space by 2x or 4x. On
615 * V3D 4.x, all shaders are 2x threaded, and 4x only divides the
616 * physical reg space in half.
617 */
618 uint8_t threads;
619 struct qinst *last_thrsw;
620 bool last_thrsw_at_top_level;
621
622 bool failed;
623 };
624
625 struct v3d_uniform_list {
626 enum quniform_contents *contents;
627 uint32_t *data;
628 uint32_t count;
629 };
630
631 struct v3d_prog_data {
632 struct v3d_uniform_list uniforms;
633
634 struct v3d_ubo_range *ubo_ranges;
635 uint32_t num_ubo_ranges;
636 uint32_t ubo_size;
637 uint32_t spill_size;
638
639 uint8_t num_inputs;
640 uint8_t threads;
641
642 /* For threads > 1, whether the program should be dispatched in the
643 * after-final-THRSW state.
644 */
645 bool single_seg;
646 };
647
648 struct v3d_vs_prog_data {
649 struct v3d_prog_data base;
650
651 bool uses_iid, uses_vid;
652
653 /* Number of components read from each vertex attribute. */
654 uint8_t vattr_sizes[32];
655
656 /* Total number of components read, for the shader state record. */
657 uint32_t vpm_input_size;
658
659 /* Total number of components written, for the shader state record. */
660 uint32_t vpm_output_size;
661
662 /* Set if there should be separate VPM segments for input and output.
663 * If unset, vpm_input_size will be 0.
664 */
665 bool separate_segments;
666
667 /* Value to be programmed in VCM_CACHE_SIZE. */
668 uint8_t vcm_cache_size;
669 };
670
671 struct v3d_fs_prog_data {
672 struct v3d_prog_data base;
673
674 struct v3d_varying_slot input_slots[V3D_MAX_FS_INPUTS];
675
676 /* Array of flat shade flags.
677 *
678 * Each entry is only 24 bits (high 8 bits 0), to match the hardware
679 * packet layout.
680 */
681 uint32_t flat_shade_flags[((V3D_MAX_FS_INPUTS - 1) / 24) + 1];
682
683 uint32_t noperspective_flags[((V3D_MAX_FS_INPUTS - 1) / 24) + 1];
684
685 uint32_t centroid_flags[((V3D_MAX_FS_INPUTS - 1) / 24) + 1];
686
687 bool writes_z;
688 bool discard;
689 bool uses_center_w;
690 };
691
692 /* Special nir_load_input intrinsic index for loading the current TLB
693 * destination color.
694 */
695 #define V3D_NIR_TLB_COLOR_READ_INPUT 2000000000
696
697 #define V3D_NIR_MS_MASK_OUTPUT 2000000000
698
699 extern const nir_shader_compiler_options v3d_nir_options;
700
701 const struct v3d_compiler *v3d_compiler_init(const struct v3d_device_info *devinfo);
702 void v3d_compiler_free(const struct v3d_compiler *compiler);
703 void v3d_optimize_nir(struct nir_shader *s);
704
705 uint64_t *v3d_compile_vs(const struct v3d_compiler *compiler,
706 struct v3d_vs_key *key,
707 struct v3d_vs_prog_data *prog_data,
708 nir_shader *s,
709 int program_id, int variant_id,
710 uint32_t *final_assembly_size);
711
712 uint64_t *v3d_compile_fs(const struct v3d_compiler *compiler,
713 struct v3d_fs_key *key,
714 struct v3d_fs_prog_data *prog_data,
715 nir_shader *s,
716 int program_id, int variant_id,
717 uint32_t *final_assembly_size);
718
719 void v3d_nir_to_vir(struct v3d_compile *c);
720
721 void vir_compile_destroy(struct v3d_compile *c);
722 const char *vir_get_stage_name(struct v3d_compile *c);
723 struct qblock *vir_new_block(struct v3d_compile *c);
724 void vir_set_emit_block(struct v3d_compile *c, struct qblock *block);
725 void vir_link_blocks(struct qblock *predecessor, struct qblock *successor);
726 struct qblock *vir_entry_block(struct v3d_compile *c);
727 struct qblock *vir_exit_block(struct v3d_compile *c);
728 struct qinst *vir_add_inst(enum v3d_qpu_add_op op, struct qreg dst,
729 struct qreg src0, struct qreg src1);
730 struct qinst *vir_mul_inst(enum v3d_qpu_mul_op op, struct qreg dst,
731 struct qreg src0, struct qreg src1);
732 struct qinst *vir_branch_inst(enum v3d_qpu_branch_cond cond, struct qreg src0);
733 void vir_remove_instruction(struct v3d_compile *c, struct qinst *qinst);
734 struct qreg vir_uniform(struct v3d_compile *c,
735 enum quniform_contents contents,
736 uint32_t data);
737 void vir_schedule_instructions(struct v3d_compile *c);
738 struct v3d_qpu_instr v3d_qpu_nop(void);
739
740 struct qreg vir_emit_def(struct v3d_compile *c, struct qinst *inst);
741 struct qinst *vir_emit_nondef(struct v3d_compile *c, struct qinst *inst);
742 void vir_set_cond(struct qinst *inst, enum v3d_qpu_cond cond);
743 void vir_set_pf(struct qinst *inst, enum v3d_qpu_pf pf);
744 void vir_set_unpack(struct qinst *inst, int src,
745 enum v3d_qpu_input_unpack unpack);
746
747 struct qreg vir_get_temp(struct v3d_compile *c);
748 void vir_emit_last_thrsw(struct v3d_compile *c);
749 void vir_calculate_live_intervals(struct v3d_compile *c);
750 bool vir_has_implicit_uniform(struct qinst *inst);
751 int vir_get_implicit_uniform_src(struct qinst *inst);
752 int vir_get_non_sideband_nsrc(struct qinst *inst);
753 int vir_get_nsrc(struct qinst *inst);
754 bool vir_has_side_effects(struct v3d_compile *c, struct qinst *inst);
755 bool vir_get_add_op(struct qinst *inst, enum v3d_qpu_add_op *op);
756 bool vir_get_mul_op(struct qinst *inst, enum v3d_qpu_mul_op *op);
757 bool vir_is_raw_mov(struct qinst *inst);
758 bool vir_is_tex(struct qinst *inst);
759 bool vir_is_add(struct qinst *inst);
760 bool vir_is_mul(struct qinst *inst);
761 bool vir_is_float_input(struct qinst *inst);
762 bool vir_depends_on_flags(struct qinst *inst);
763 bool vir_writes_r3(const struct v3d_device_info *devinfo, struct qinst *inst);
764 bool vir_writes_r4(const struct v3d_device_info *devinfo, struct qinst *inst);
765 struct qreg vir_follow_movs(struct v3d_compile *c, struct qreg reg);
766 uint8_t vir_channels_written(struct qinst *inst);
767 struct qreg ntq_get_src(struct v3d_compile *c, nir_src src, int i);
768 void ntq_store_dest(struct v3d_compile *c, nir_dest *dest, int chan,
769 struct qreg result);
770 void vir_emit_thrsw(struct v3d_compile *c);
771
772 void vir_dump(struct v3d_compile *c);
773 void vir_dump_inst(struct v3d_compile *c, struct qinst *inst);
774 void vir_dump_uniform(enum quniform_contents contents, uint32_t data);
775
776 void vir_validate(struct v3d_compile *c);
777
778 void vir_optimize(struct v3d_compile *c);
779 bool vir_opt_algebraic(struct v3d_compile *c);
780 bool vir_opt_constant_folding(struct v3d_compile *c);
781 bool vir_opt_copy_propagate(struct v3d_compile *c);
782 bool vir_opt_dead_code(struct v3d_compile *c);
783 bool vir_opt_peephole_sf(struct v3d_compile *c);
784 bool vir_opt_small_immediates(struct v3d_compile *c);
785 bool vir_opt_vpm(struct v3d_compile *c);
786 void v3d_nir_lower_blend(nir_shader *s, struct v3d_compile *c);
787 void v3d_nir_lower_io(nir_shader *s, struct v3d_compile *c);
788 void v3d_nir_lower_txf_ms(nir_shader *s, struct v3d_compile *c);
789 void vir_lower_uniforms(struct v3d_compile *c);
790
791 void v3d33_vir_vpm_read_setup(struct v3d_compile *c, int num_components);
792 void v3d33_vir_vpm_write_setup(struct v3d_compile *c);
793 void v3d33_vir_emit_tex(struct v3d_compile *c, nir_tex_instr *instr);
794 void v3d40_vir_emit_tex(struct v3d_compile *c, nir_tex_instr *instr);
795
796 void v3d_vir_to_qpu(struct v3d_compile *c, struct qpu_reg *temp_registers);
797 uint32_t v3d_qpu_schedule_instructions(struct v3d_compile *c);
798 void qpu_validate(struct v3d_compile *c);
799 struct qpu_reg *v3d_register_allocate(struct v3d_compile *c, bool *spilled);
800 bool vir_init_reg_sets(struct v3d_compiler *compiler);
801
802 void vir_PF(struct v3d_compile *c, struct qreg src, enum v3d_qpu_pf pf);
803
804 static inline bool
805 quniform_contents_is_texture_p0(enum quniform_contents contents)
806 {
807 return (contents >= QUNIFORM_TEXTURE_CONFIG_P0_0 &&
808 contents < (QUNIFORM_TEXTURE_CONFIG_P0_0 +
809 V3D_MAX_TEXTURE_SAMPLERS));
810 }
811
812 static inline struct qreg
813 vir_uniform_ui(struct v3d_compile *c, uint32_t ui)
814 {
815 return vir_uniform(c, QUNIFORM_CONSTANT, ui);
816 }
817
818 static inline struct qreg
819 vir_uniform_f(struct v3d_compile *c, float f)
820 {
821 return vir_uniform(c, QUNIFORM_CONSTANT, fui(f));
822 }
823
824 #define VIR_ALU0(name, vir_inst, op) \
825 static inline struct qreg \
826 vir_##name(struct v3d_compile *c) \
827 { \
828 return vir_emit_def(c, vir_inst(op, c->undef, \
829 c->undef, c->undef)); \
830 } \
831 static inline struct qinst * \
832 vir_##name##_dest(struct v3d_compile *c, struct qreg dest) \
833 { \
834 return vir_emit_nondef(c, vir_inst(op, dest, \
835 c->undef, c->undef)); \
836 }
837
838 #define VIR_ALU1(name, vir_inst, op) \
839 static inline struct qreg \
840 vir_##name(struct v3d_compile *c, struct qreg a) \
841 { \
842 return vir_emit_def(c, vir_inst(op, c->undef, \
843 a, c->undef)); \
844 } \
845 static inline struct qinst * \
846 vir_##name##_dest(struct v3d_compile *c, struct qreg dest, \
847 struct qreg a) \
848 { \
849 return vir_emit_nondef(c, vir_inst(op, dest, a, \
850 c->undef)); \
851 }
852
853 #define VIR_ALU2(name, vir_inst, op) \
854 static inline struct qreg \
855 vir_##name(struct v3d_compile *c, struct qreg a, struct qreg b) \
856 { \
857 return vir_emit_def(c, vir_inst(op, c->undef, a, b)); \
858 } \
859 static inline struct qinst * \
860 vir_##name##_dest(struct v3d_compile *c, struct qreg dest, \
861 struct qreg a, struct qreg b) \
862 { \
863 return vir_emit_nondef(c, vir_inst(op, dest, a, b)); \
864 }
865
866 #define VIR_NODST_0(name, vir_inst, op) \
867 static inline struct qinst * \
868 vir_##name(struct v3d_compile *c) \
869 { \
870 return vir_emit_nondef(c, vir_inst(op, c->undef, \
871 c->undef, c->undef)); \
872 }
873
874 #define VIR_NODST_1(name, vir_inst, op) \
875 static inline struct qinst * \
876 vir_##name(struct v3d_compile *c, struct qreg a) \
877 { \
878 return vir_emit_nondef(c, vir_inst(op, c->undef, \
879 a, c->undef)); \
880 }
881
882 #define VIR_NODST_2(name, vir_inst, op) \
883 static inline struct qinst * \
884 vir_##name(struct v3d_compile *c, struct qreg a, struct qreg b) \
885 { \
886 return vir_emit_nondef(c, vir_inst(op, c->undef, \
887 a, b)); \
888 }
889
890 #define VIR_SFU(name) \
891 static inline struct qreg \
892 vir_##name(struct v3d_compile *c, struct qreg a) \
893 { \
894 if (c->devinfo->ver >= 41) { \
895 return vir_emit_def(c, vir_add_inst(V3D_QPU_A_##name, \
896 c->undef, \
897 a, c->undef)); \
898 } else { \
899 vir_FMOV_dest(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_##name), a); \
900 return vir_FMOV(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_R4)); \
901 } \
902 } \
903 static inline struct qinst * \
904 vir_##name##_dest(struct v3d_compile *c, struct qreg dest, \
905 struct qreg a) \
906 { \
907 if (c->devinfo->ver >= 41) { \
908 return vir_emit_nondef(c, vir_add_inst(V3D_QPU_A_##name, \
909 dest, \
910 a, c->undef)); \
911 } else { \
912 vir_FMOV_dest(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_##name), a); \
913 return vir_FMOV_dest(c, dest, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_R4)); \
914 } \
915 }
916
917 #define VIR_A_ALU2(name) VIR_ALU2(name, vir_add_inst, V3D_QPU_A_##name)
918 #define VIR_M_ALU2(name) VIR_ALU2(name, vir_mul_inst, V3D_QPU_M_##name)
919 #define VIR_A_ALU1(name) VIR_ALU1(name, vir_add_inst, V3D_QPU_A_##name)
920 #define VIR_M_ALU1(name) VIR_ALU1(name, vir_mul_inst, V3D_QPU_M_##name)
921 #define VIR_A_ALU0(name) VIR_ALU0(name, vir_add_inst, V3D_QPU_A_##name)
922 #define VIR_M_ALU0(name) VIR_ALU0(name, vir_mul_inst, V3D_QPU_M_##name)
923 #define VIR_A_NODST_2(name) VIR_NODST_2(name, vir_add_inst, V3D_QPU_A_##name)
924 #define VIR_M_NODST_2(name) VIR_NODST_2(name, vir_mul_inst, V3D_QPU_M_##name)
925 #define VIR_A_NODST_1(name) VIR_NODST_1(name, vir_add_inst, V3D_QPU_A_##name)
926 #define VIR_M_NODST_1(name) VIR_NODST_1(name, vir_mul_inst, V3D_QPU_M_##name)
927 #define VIR_A_NODST_0(name) VIR_NODST_0(name, vir_add_inst, V3D_QPU_A_##name)
928
929 VIR_A_ALU2(FADD)
930 VIR_A_ALU2(VFPACK)
931 VIR_A_ALU2(FSUB)
932 VIR_A_ALU2(FMIN)
933 VIR_A_ALU2(FMAX)
934
935 VIR_A_ALU2(ADD)
936 VIR_A_ALU2(SUB)
937 VIR_A_ALU2(SHL)
938 VIR_A_ALU2(SHR)
939 VIR_A_ALU2(ASR)
940 VIR_A_ALU2(ROR)
941 VIR_A_ALU2(MIN)
942 VIR_A_ALU2(MAX)
943 VIR_A_ALU2(UMIN)
944 VIR_A_ALU2(UMAX)
945 VIR_A_ALU2(AND)
946 VIR_A_ALU2(OR)
947 VIR_A_ALU2(XOR)
948 VIR_A_ALU2(VADD)
949 VIR_A_ALU2(VSUB)
950 VIR_A_NODST_2(STVPMV)
951 VIR_A_ALU1(NOT)
952 VIR_A_ALU1(NEG)
953 VIR_A_ALU1(FLAPUSH)
954 VIR_A_ALU1(FLBPUSH)
955 VIR_A_ALU1(FLPOP)
956 VIR_A_ALU1(SETMSF)
957 VIR_A_ALU1(SETREVF)
958 VIR_A_ALU0(TIDX)
959 VIR_A_ALU0(EIDX)
960 VIR_A_ALU1(LDVPMV_IN)
961 VIR_A_ALU1(LDVPMV_OUT)
962 VIR_A_ALU0(TMUWT)
963
964 VIR_A_ALU0(FXCD)
965 VIR_A_ALU0(XCD)
966 VIR_A_ALU0(FYCD)
967 VIR_A_ALU0(YCD)
968 VIR_A_ALU0(MSF)
969 VIR_A_ALU0(REVF)
970 VIR_A_NODST_1(VPMSETUP)
971 VIR_A_NODST_0(VPMWT)
972 VIR_A_ALU2(FCMP)
973 VIR_A_ALU2(VFMAX)
974
975 VIR_A_ALU1(FROUND)
976 VIR_A_ALU1(FTOIN)
977 VIR_A_ALU1(FTRUNC)
978 VIR_A_ALU1(FTOIZ)
979 VIR_A_ALU1(FFLOOR)
980 VIR_A_ALU1(FTOUZ)
981 VIR_A_ALU1(FCEIL)
982 VIR_A_ALU1(FTOC)
983
984 VIR_A_ALU1(FDX)
985 VIR_A_ALU1(FDY)
986
987 VIR_A_ALU1(ITOF)
988 VIR_A_ALU1(CLZ)
989 VIR_A_ALU1(UTOF)
990
991 VIR_M_ALU2(UMUL24)
992 VIR_M_ALU2(FMUL)
993 VIR_M_ALU2(SMUL24)
994 VIR_M_NODST_2(MULTOP)
995
996 VIR_M_ALU1(MOV)
997 VIR_M_ALU1(FMOV)
998
999 VIR_SFU(RECIP)
1000 VIR_SFU(RSQRT)
1001 VIR_SFU(EXP)
1002 VIR_SFU(LOG)
1003 VIR_SFU(SIN)
1004 VIR_SFU(RSQRT2)
1005
1006 static inline struct qinst *
1007 vir_MOV_cond(struct v3d_compile *c, enum v3d_qpu_cond cond,
1008 struct qreg dest, struct qreg src)
1009 {
1010 struct qinst *mov = vir_MOV_dest(c, dest, src);
1011 vir_set_cond(mov, cond);
1012 return mov;
1013 }
1014
1015 static inline struct qreg
1016 vir_SEL(struct v3d_compile *c, enum v3d_qpu_cond cond,
1017 struct qreg src0, struct qreg src1)
1018 {
1019 struct qreg t = vir_get_temp(c);
1020 vir_MOV_dest(c, t, src1);
1021 vir_MOV_cond(c, cond, t, src0);
1022 return t;
1023 }
1024
1025 static inline struct qinst *
1026 vir_NOP(struct v3d_compile *c)
1027 {
1028 return vir_emit_nondef(c, vir_add_inst(V3D_QPU_A_NOP,
1029 c->undef, c->undef, c->undef));
1030 }
1031
1032 static inline struct qreg
1033 vir_LDTMU(struct v3d_compile *c)
1034 {
1035 if (c->devinfo->ver >= 41) {
1036 struct qinst *ldtmu = vir_add_inst(V3D_QPU_A_NOP, c->undef,
1037 c->undef, c->undef);
1038 ldtmu->qpu.sig.ldtmu = true;
1039
1040 return vir_emit_def(c, ldtmu);
1041 } else {
1042 vir_NOP(c)->qpu.sig.ldtmu = true;
1043 return vir_MOV(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_R4));
1044 }
1045 }
1046
1047 static inline struct qreg
1048 vir_UMUL(struct v3d_compile *c, struct qreg src0, struct qreg src1)
1049 {
1050 vir_MULTOP(c, src0, src1);
1051 return vir_UMUL24(c, src0, src1);
1052 }
1053
1054 /*
1055 static inline struct qreg
1056 vir_LOAD_IMM(struct v3d_compile *c, uint32_t val)
1057 {
1058 return vir_emit_def(c, vir_inst(QOP_LOAD_IMM, c->undef,
1059 vir_reg(QFILE_LOAD_IMM, val), c->undef));
1060 }
1061
1062 static inline struct qreg
1063 vir_LOAD_IMM_U2(struct v3d_compile *c, uint32_t val)
1064 {
1065 return vir_emit_def(c, vir_inst(QOP_LOAD_IMM_U2, c->undef,
1066 vir_reg(QFILE_LOAD_IMM, val),
1067 c->undef));
1068 }
1069 static inline struct qreg
1070 vir_LOAD_IMM_I2(struct v3d_compile *c, uint32_t val)
1071 {
1072 return vir_emit_def(c, vir_inst(QOP_LOAD_IMM_I2, c->undef,
1073 vir_reg(QFILE_LOAD_IMM, val),
1074 c->undef));
1075 }
1076 */
1077
1078 static inline struct qinst *
1079 vir_BRANCH(struct v3d_compile *c, enum v3d_qpu_branch_cond cond)
1080 {
1081 /* The actual uniform_data value will be set at scheduling time */
1082 return vir_emit_nondef(c, vir_branch_inst(cond, vir_uniform_ui(c, 0)));
1083 }
1084
1085 #define vir_for_each_block(block, c) \
1086 list_for_each_entry(struct qblock, block, &c->blocks, link)
1087
1088 #define vir_for_each_block_rev(block, c) \
1089 list_for_each_entry_rev(struct qblock, block, &c->blocks, link)
1090
1091 /* Loop over the non-NULL members of the successors array. */
1092 #define vir_for_each_successor(succ, block) \
1093 for (struct qblock *succ = block->successors[0]; \
1094 succ != NULL; \
1095 succ = (succ == block->successors[1] ? NULL : \
1096 block->successors[1]))
1097
1098 #define vir_for_each_inst(inst, block) \
1099 list_for_each_entry(struct qinst, inst, &block->instructions, link)
1100
1101 #define vir_for_each_inst_rev(inst, block) \
1102 list_for_each_entry_rev(struct qinst, inst, &block->instructions, link)
1103
1104 #define vir_for_each_inst_safe(inst, block) \
1105 list_for_each_entry_safe(struct qinst, inst, &block->instructions, link)
1106
1107 #define vir_for_each_inst_inorder(inst, c) \
1108 vir_for_each_block(_block, c) \
1109 vir_for_each_inst(inst, _block)
1110
1111 #endif /* V3D_COMPILER_H */