v3d: Implement a small immediates optimization, based on VC4's.
[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 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_TEXTURE_BORDER_COLOR,
247
248 QUNIFORM_ALPHA_REF,
249 QUNIFORM_SAMPLE_MASK,
250
251 /**
252 * Returns the the offset of the scratch buffer for register spilling.
253 */
254 QUNIFORM_SPILL_OFFSET,
255 QUNIFORM_SPILL_SIZE_PER_THREAD,
256 };
257
258 struct v3d_varying_slot {
259 uint8_t slot_and_component;
260 };
261
262 static inline struct v3d_varying_slot
263 v3d_slot_from_slot_and_component(uint8_t slot, uint8_t component)
264 {
265 assert(slot < 255 / 4);
266 return (struct v3d_varying_slot){ (slot << 2) + component };
267 }
268
269 static inline uint8_t v3d_slot_get_slot(struct v3d_varying_slot slot)
270 {
271 return slot.slot_and_component >> 2;
272 }
273
274 static inline uint8_t v3d_slot_get_component(struct v3d_varying_slot slot)
275 {
276 return slot.slot_and_component & 3;
277 }
278
279 struct v3d_ubo_range {
280 /**
281 * offset in bytes from the start of the ubo where this range is
282 * uploaded.
283 *
284 * Only set once used is set.
285 */
286 uint32_t dst_offset;
287
288 /**
289 * offset in bytes from the start of the gallium uniforms where the
290 * data comes from.
291 */
292 uint32_t src_offset;
293
294 /** size in bytes of this ubo range */
295 uint32_t size;
296 };
297
298 struct v3d_key {
299 void *shader_state;
300 struct {
301 uint8_t swizzle[4];
302 uint8_t return_size;
303 uint8_t return_channels;
304 unsigned compare_mode:1;
305 unsigned compare_func:3;
306 bool clamp_s:1;
307 bool clamp_t:1;
308 bool clamp_r:1;
309 } tex[V3D_MAX_TEXTURE_SAMPLERS];
310 uint8_t ucp_enables;
311 };
312
313 struct v3d_fs_key {
314 struct v3d_key base;
315 bool depth_enabled;
316 bool is_points;
317 bool is_lines;
318 bool alpha_test;
319 bool point_coord_upper_left;
320 bool light_twoside;
321 bool msaa;
322 bool sample_coverage;
323 bool sample_alpha_to_coverage;
324 bool sample_alpha_to_one;
325 bool clamp_color;
326 bool shade_model_flat;
327 uint8_t nr_cbufs;
328 uint8_t swap_color_rb;
329 /* Mask of which render targets need to be written as 32-bit floats */
330 uint8_t f32_color_rb;
331 /* Masks of which render targets need to be written as ints/uints.
332 * Used by gallium to work around lost information in TGSI.
333 */
334 uint8_t int_color_rb;
335 uint8_t uint_color_rb;
336 uint8_t alpha_test_func;
337 uint8_t logicop_func;
338 uint32_t point_sprite_mask;
339
340 struct pipe_rt_blend_state blend;
341 };
342
343 struct v3d_vs_key {
344 struct v3d_key base;
345
346 struct v3d_varying_slot fs_inputs[V3D_MAX_FS_INPUTS];
347 uint8_t num_fs_inputs;
348
349 bool is_coord;
350 bool per_vertex_point_size;
351 bool clamp_color;
352 };
353
354 /** A basic block of VIR intructions. */
355 struct qblock {
356 struct list_head link;
357
358 struct list_head instructions;
359
360 struct set *predecessors;
361 struct qblock *successors[2];
362
363 int index;
364
365 /* Instruction IPs for the first and last instruction of the block.
366 * Set by qpu_schedule.c.
367 */
368 uint32_t start_qpu_ip;
369 uint32_t end_qpu_ip;
370
371 /* Instruction IP for the branch instruction of the block. Set by
372 * qpu_schedule.c.
373 */
374 uint32_t branch_qpu_ip;
375
376 /** Offset within the uniform stream at the start of the block. */
377 uint32_t start_uniform;
378 /** Offset within the uniform stream of the branch instruction */
379 uint32_t branch_uniform;
380
381 /** @{ used by v3d_vir_live_variables.c */
382 BITSET_WORD *def;
383 BITSET_WORD *use;
384 BITSET_WORD *live_in;
385 BITSET_WORD *live_out;
386 int start_ip, end_ip;
387 /** @} */
388 };
389
390 /** Which util/list.h add mode we should use when inserting an instruction. */
391 enum vir_cursor_mode {
392 vir_cursor_add,
393 vir_cursor_addtail,
394 };
395
396 /**
397 * Tracking structure for where new instructions should be inserted. Create
398 * with one of the vir_after_inst()-style helper functions.
399 *
400 * This does not protect against removal of the block or instruction, so we
401 * have an assert in instruction removal to try to catch it.
402 */
403 struct vir_cursor {
404 enum vir_cursor_mode mode;
405 struct list_head *link;
406 };
407
408 static inline struct vir_cursor
409 vir_before_inst(struct qinst *inst)
410 {
411 return (struct vir_cursor){ vir_cursor_addtail, &inst->link };
412 }
413
414 static inline struct vir_cursor
415 vir_after_inst(struct qinst *inst)
416 {
417 return (struct vir_cursor){ vir_cursor_add, &inst->link };
418 }
419
420 static inline struct vir_cursor
421 vir_before_block(struct qblock *block)
422 {
423 return (struct vir_cursor){ vir_cursor_add, &block->instructions };
424 }
425
426 static inline struct vir_cursor
427 vir_after_block(struct qblock *block)
428 {
429 return (struct vir_cursor){ vir_cursor_addtail, &block->instructions };
430 }
431
432 /**
433 * Compiler state saved across compiler invocations, for any expensive global
434 * setup.
435 */
436 struct v3d_compiler {
437 const struct v3d_device_info *devinfo;
438 struct ra_regs *regs;
439 unsigned int reg_class_phys[3];
440 unsigned int reg_class_phys_or_acc[3];
441 };
442
443 struct v3d_compile {
444 const struct v3d_device_info *devinfo;
445 nir_shader *s;
446 nir_function_impl *impl;
447 struct exec_list *cf_node_list;
448 const struct v3d_compiler *compiler;
449
450 /**
451 * Mapping from nir_register * or nir_ssa_def * to array of struct
452 * qreg for the values.
453 */
454 struct hash_table *def_ht;
455
456 /* For each temp, the instruction generating its value. */
457 struct qinst **defs;
458 uint32_t defs_array_size;
459
460 /**
461 * Inputs to the shader, arranged by TGSI declaration order.
462 *
463 * Not all fragment shader QFILE_VARY reads are present in this array.
464 */
465 struct qreg *inputs;
466 struct qreg *outputs;
467 bool msaa_per_sample_output;
468 struct qreg color_reads[V3D_MAX_SAMPLES];
469 struct qreg sample_colors[V3D_MAX_SAMPLES];
470 uint32_t inputs_array_size;
471 uint32_t outputs_array_size;
472 uint32_t uniforms_array_size;
473
474 /* Booleans for whether the corresponding QFILE_VARY[i] is
475 * flat-shaded. This includes gl_FragColor flat-shading, which is
476 * customized based on the shademodel_flat shader key.
477 */
478 uint32_t flat_shade_flags[BITSET_WORDS(V3D_MAX_FS_INPUTS)];
479
480 uint32_t noperspective_flags[BITSET_WORDS(V3D_MAX_FS_INPUTS)];
481
482 uint32_t centroid_flags[BITSET_WORDS(V3D_MAX_FS_INPUTS)];
483
484 bool uses_center_w;
485
486 struct v3d_ubo_range *ubo_ranges;
487 bool *ubo_range_used;
488 uint32_t ubo_ranges_array_size;
489 /** Number of uniform areas tracked in ubo_ranges. */
490 uint32_t num_ubo_ranges;
491 uint32_t next_ubo_dst_offset;
492
493 /* State for whether we're executing on each channel currently. 0 if
494 * yes, otherwise a block number + 1 that the channel jumped to.
495 */
496 struct qreg execute;
497
498 struct qreg line_x, point_x, point_y;
499
500 /**
501 * Instance ID, which comes in before the vertex attribute payload if
502 * the shader record requests it.
503 */
504 struct qreg iid;
505
506 /**
507 * Vertex ID, which comes in before the vertex attribute payload
508 * (after Instance ID) if the shader record requests it.
509 */
510 struct qreg vid;
511
512 /* Fragment shader payload regs. */
513 struct qreg payload_w, payload_w_centroid, payload_z;
514
515 uint8_t vattr_sizes[V3D_MAX_VS_INPUTS];
516 uint32_t num_vpm_writes;
517
518 /* Size in bytes of registers that have been spilled. This is how much
519 * space needs to be available in the spill BO per thread per QPU.
520 */
521 uint32_t spill_size;
522 /* Shader-db stats for register spilling. */
523 uint32_t spills, fills;
524 /**
525 * Register spilling's per-thread base address, shared between each
526 * spill/fill's addressing calculations.
527 */
528 struct qreg spill_base;
529 /* Bit vector of which temps may be spilled */
530 BITSET_WORD *spillable;
531
532 /**
533 * Array of the VARYING_SLOT_* of all FS QFILE_VARY reads.
534 *
535 * This includes those that aren't part of the VPM varyings, like
536 * point/line coordinates.
537 */
538 struct v3d_varying_slot input_slots[V3D_MAX_FS_INPUTS];
539
540 /**
541 * An entry per outputs[] in the VS indicating what the VARYING_SLOT_*
542 * of the output is. Used to emit from the VS in the order that the
543 * FS needs.
544 */
545 struct v3d_varying_slot *output_slots;
546
547 struct pipe_shader_state *shader_state;
548 struct v3d_key *key;
549 struct v3d_fs_key *fs_key;
550 struct v3d_vs_key *vs_key;
551
552 /* Live ranges of temps. */
553 int *temp_start, *temp_end;
554 bool live_intervals_valid;
555
556 uint32_t *uniform_data;
557 enum quniform_contents *uniform_contents;
558 uint32_t uniform_array_size;
559 uint32_t num_uniforms;
560 uint32_t num_outputs;
561 uint32_t output_position_index;
562 nir_variable *output_color_var[4];
563 uint32_t output_point_size_index;
564 uint32_t output_sample_mask_index;
565
566 struct qreg undef;
567 uint32_t num_temps;
568
569 struct vir_cursor cursor;
570 struct list_head blocks;
571 int next_block_index;
572 struct qblock *cur_block;
573 struct qblock *loop_cont_block;
574 struct qblock *loop_break_block;
575
576 uint64_t *qpu_insts;
577 uint32_t qpu_inst_count;
578 uint32_t qpu_inst_size;
579
580 /* For the FS, the number of varying inputs not counting the
581 * point/line varyings payload
582 */
583 uint32_t num_inputs;
584
585 /**
586 * Number of inputs from num_inputs remaining to be queued to the read
587 * FIFO in the VS/CS.
588 */
589 uint32_t num_inputs_remaining;
590
591 /* Number of inputs currently in the read FIFO for the VS/CS */
592 uint32_t num_inputs_in_fifo;
593
594 /** Next offset in the VPM to read from in the VS/CS */
595 uint32_t vpm_read_offset;
596
597 uint32_t program_id;
598 uint32_t variant_id;
599
600 /* Set to compile program in in 1x, 2x, or 4x threaded mode, where
601 * SIG_THREAD_SWITCH is used to hide texturing latency at the cost of
602 * limiting ourselves to the part of the physical reg space.
603 *
604 * On V3D 3.x, 2x or 4x divide the physical reg space by 2x or 4x. On
605 * V3D 4.x, all shaders are 2x threaded, and 4x only divides the
606 * physical reg space in half.
607 */
608 uint8_t threads;
609 struct qinst *last_thrsw;
610 bool last_thrsw_at_top_level;
611
612 bool failed;
613 };
614
615 struct v3d_uniform_list {
616 enum quniform_contents *contents;
617 uint32_t *data;
618 uint32_t count;
619 };
620
621 struct v3d_prog_data {
622 struct v3d_uniform_list uniforms;
623
624 struct v3d_ubo_range *ubo_ranges;
625 uint32_t num_ubo_ranges;
626 uint32_t ubo_size;
627 uint32_t spill_size;
628
629 uint8_t num_inputs;
630 uint8_t threads;
631
632 /* For threads > 1, whether the program should be dispatched in the
633 * after-final-THRSW state.
634 */
635 bool single_seg;
636 };
637
638 struct v3d_vs_prog_data {
639 struct v3d_prog_data base;
640
641 bool uses_iid, uses_vid;
642
643 /* Number of components read from each vertex attribute. */
644 uint8_t vattr_sizes[32];
645
646 /* Total number of components read, for the shader state record. */
647 uint32_t vpm_input_size;
648
649 /* Total number of components written, for the shader state record. */
650 uint32_t vpm_output_size;
651 };
652
653 struct v3d_fs_prog_data {
654 struct v3d_prog_data base;
655
656 struct v3d_varying_slot input_slots[V3D_MAX_FS_INPUTS];
657
658 /* Array of flat shade flags.
659 *
660 * Each entry is only 24 bits (high 8 bits 0), to match the hardware
661 * packet layout.
662 */
663 uint32_t flat_shade_flags[((V3D_MAX_FS_INPUTS - 1) / 24) + 1];
664
665 uint32_t noperspective_flags[((V3D_MAX_FS_INPUTS - 1) / 24) + 1];
666
667 uint32_t centroid_flags[((V3D_MAX_FS_INPUTS - 1) / 24) + 1];
668
669 bool writes_z;
670 bool discard;
671 bool uses_center_w;
672 };
673
674 /* Special nir_load_input intrinsic index for loading the current TLB
675 * destination color.
676 */
677 #define V3D_NIR_TLB_COLOR_READ_INPUT 2000000000
678
679 #define V3D_NIR_MS_MASK_OUTPUT 2000000000
680
681 extern const nir_shader_compiler_options v3d_nir_options;
682
683 const struct v3d_compiler *v3d_compiler_init(const struct v3d_device_info *devinfo);
684 void v3d_compiler_free(const struct v3d_compiler *compiler);
685 void v3d_optimize_nir(struct nir_shader *s);
686
687 uint64_t *v3d_compile_vs(const struct v3d_compiler *compiler,
688 struct v3d_vs_key *key,
689 struct v3d_vs_prog_data *prog_data,
690 nir_shader *s,
691 int program_id, int variant_id,
692 uint32_t *final_assembly_size);
693
694 uint64_t *v3d_compile_fs(const struct v3d_compiler *compiler,
695 struct v3d_fs_key *key,
696 struct v3d_fs_prog_data *prog_data,
697 nir_shader *s,
698 int program_id, int variant_id,
699 uint32_t *final_assembly_size);
700
701 void v3d_nir_to_vir(struct v3d_compile *c);
702
703 void vir_compile_destroy(struct v3d_compile *c);
704 const char *vir_get_stage_name(struct v3d_compile *c);
705 struct qblock *vir_new_block(struct v3d_compile *c);
706 void vir_set_emit_block(struct v3d_compile *c, struct qblock *block);
707 void vir_link_blocks(struct qblock *predecessor, struct qblock *successor);
708 struct qblock *vir_entry_block(struct v3d_compile *c);
709 struct qblock *vir_exit_block(struct v3d_compile *c);
710 struct qinst *vir_add_inst(enum v3d_qpu_add_op op, struct qreg dst,
711 struct qreg src0, struct qreg src1);
712 struct qinst *vir_mul_inst(enum v3d_qpu_mul_op op, struct qreg dst,
713 struct qreg src0, struct qreg src1);
714 struct qinst *vir_branch_inst(enum v3d_qpu_branch_cond cond, struct qreg src0);
715 void vir_remove_instruction(struct v3d_compile *c, struct qinst *qinst);
716 struct qreg vir_uniform(struct v3d_compile *c,
717 enum quniform_contents contents,
718 uint32_t data);
719 void vir_schedule_instructions(struct v3d_compile *c);
720 struct v3d_qpu_instr v3d_qpu_nop(void);
721
722 struct qreg vir_emit_def(struct v3d_compile *c, struct qinst *inst);
723 struct qinst *vir_emit_nondef(struct v3d_compile *c, struct qinst *inst);
724 void vir_set_cond(struct qinst *inst, enum v3d_qpu_cond cond);
725 void vir_set_pf(struct qinst *inst, enum v3d_qpu_pf pf);
726 void vir_set_unpack(struct qinst *inst, int src,
727 enum v3d_qpu_input_unpack unpack);
728
729 struct qreg vir_get_temp(struct v3d_compile *c);
730 void vir_emit_last_thrsw(struct v3d_compile *c);
731 void vir_calculate_live_intervals(struct v3d_compile *c);
732 bool vir_has_implicit_uniform(struct qinst *inst);
733 int vir_get_implicit_uniform_src(struct qinst *inst);
734 int vir_get_non_sideband_nsrc(struct qinst *inst);
735 int vir_get_nsrc(struct qinst *inst);
736 bool vir_has_side_effects(struct v3d_compile *c, struct qinst *inst);
737 bool vir_get_add_op(struct qinst *inst, enum v3d_qpu_add_op *op);
738 bool vir_get_mul_op(struct qinst *inst, enum v3d_qpu_mul_op *op);
739 bool vir_is_raw_mov(struct qinst *inst);
740 bool vir_is_tex(struct qinst *inst);
741 bool vir_is_add(struct qinst *inst);
742 bool vir_is_mul(struct qinst *inst);
743 bool vir_is_float_input(struct qinst *inst);
744 bool vir_depends_on_flags(struct qinst *inst);
745 bool vir_writes_r3(const struct v3d_device_info *devinfo, struct qinst *inst);
746 bool vir_writes_r4(const struct v3d_device_info *devinfo, struct qinst *inst);
747 struct qreg vir_follow_movs(struct v3d_compile *c, struct qreg reg);
748 uint8_t vir_channels_written(struct qinst *inst);
749 struct qreg ntq_get_src(struct v3d_compile *c, nir_src src, int i);
750 void ntq_store_dest(struct v3d_compile *c, nir_dest *dest, int chan,
751 struct qreg result);
752 void vir_emit_thrsw(struct v3d_compile *c);
753
754 void vir_dump(struct v3d_compile *c);
755 void vir_dump_inst(struct v3d_compile *c, struct qinst *inst);
756
757 void vir_validate(struct v3d_compile *c);
758
759 void vir_optimize(struct v3d_compile *c);
760 bool vir_opt_algebraic(struct v3d_compile *c);
761 bool vir_opt_constant_folding(struct v3d_compile *c);
762 bool vir_opt_copy_propagate(struct v3d_compile *c);
763 bool vir_opt_dead_code(struct v3d_compile *c);
764 bool vir_opt_peephole_sf(struct v3d_compile *c);
765 bool vir_opt_small_immediates(struct v3d_compile *c);
766 bool vir_opt_vpm(struct v3d_compile *c);
767 void v3d_nir_lower_blend(nir_shader *s, struct v3d_compile *c);
768 void v3d_nir_lower_io(nir_shader *s, struct v3d_compile *c);
769 void v3d_nir_lower_txf_ms(nir_shader *s, struct v3d_compile *c);
770 void vir_lower_uniforms(struct v3d_compile *c);
771
772 void v3d33_vir_vpm_read_setup(struct v3d_compile *c, int num_components);
773 void v3d33_vir_vpm_write_setup(struct v3d_compile *c);
774 void v3d33_vir_emit_tex(struct v3d_compile *c, nir_tex_instr *instr);
775 void v3d40_vir_emit_tex(struct v3d_compile *c, nir_tex_instr *instr);
776
777 void v3d_vir_to_qpu(struct v3d_compile *c, struct qpu_reg *temp_registers);
778 uint32_t v3d_qpu_schedule_instructions(struct v3d_compile *c);
779 void qpu_validate(struct v3d_compile *c);
780 struct qpu_reg *v3d_register_allocate(struct v3d_compile *c, bool *spilled);
781 bool vir_init_reg_sets(struct v3d_compiler *compiler);
782
783 void vir_PF(struct v3d_compile *c, struct qreg src, enum v3d_qpu_pf pf);
784
785 static inline bool
786 quniform_contents_is_texture_p0(enum quniform_contents contents)
787 {
788 return (contents >= QUNIFORM_TEXTURE_CONFIG_P0_0 &&
789 contents < (QUNIFORM_TEXTURE_CONFIG_P0_0 +
790 V3D_MAX_TEXTURE_SAMPLERS));
791 }
792
793 static inline struct qreg
794 vir_uniform_ui(struct v3d_compile *c, uint32_t ui)
795 {
796 return vir_uniform(c, QUNIFORM_CONSTANT, ui);
797 }
798
799 static inline struct qreg
800 vir_uniform_f(struct v3d_compile *c, float f)
801 {
802 return vir_uniform(c, QUNIFORM_CONSTANT, fui(f));
803 }
804
805 #define VIR_ALU0(name, vir_inst, op) \
806 static inline struct qreg \
807 vir_##name(struct v3d_compile *c) \
808 { \
809 return vir_emit_def(c, vir_inst(op, c->undef, \
810 c->undef, c->undef)); \
811 } \
812 static inline struct qinst * \
813 vir_##name##_dest(struct v3d_compile *c, struct qreg dest) \
814 { \
815 return vir_emit_nondef(c, vir_inst(op, dest, \
816 c->undef, c->undef)); \
817 }
818
819 #define VIR_ALU1(name, vir_inst, op) \
820 static inline struct qreg \
821 vir_##name(struct v3d_compile *c, struct qreg a) \
822 { \
823 return vir_emit_def(c, vir_inst(op, c->undef, \
824 a, c->undef)); \
825 } \
826 static inline struct qinst * \
827 vir_##name##_dest(struct v3d_compile *c, struct qreg dest, \
828 struct qreg a) \
829 { \
830 return vir_emit_nondef(c, vir_inst(op, dest, a, \
831 c->undef)); \
832 }
833
834 #define VIR_ALU2(name, vir_inst, op) \
835 static inline struct qreg \
836 vir_##name(struct v3d_compile *c, struct qreg a, struct qreg b) \
837 { \
838 return vir_emit_def(c, vir_inst(op, c->undef, a, b)); \
839 } \
840 static inline struct qinst * \
841 vir_##name##_dest(struct v3d_compile *c, struct qreg dest, \
842 struct qreg a, struct qreg b) \
843 { \
844 return vir_emit_nondef(c, vir_inst(op, dest, a, b)); \
845 }
846
847 #define VIR_NODST_0(name, vir_inst, op) \
848 static inline struct qinst * \
849 vir_##name(struct v3d_compile *c) \
850 { \
851 return vir_emit_nondef(c, vir_inst(op, c->undef, \
852 c->undef, c->undef)); \
853 }
854
855 #define VIR_NODST_1(name, vir_inst, op) \
856 static inline struct qinst * \
857 vir_##name(struct v3d_compile *c, struct qreg a) \
858 { \
859 return vir_emit_nondef(c, vir_inst(op, c->undef, \
860 a, c->undef)); \
861 }
862
863 #define VIR_NODST_2(name, vir_inst, op) \
864 static inline struct qinst * \
865 vir_##name(struct v3d_compile *c, struct qreg a, struct qreg b) \
866 { \
867 return vir_emit_nondef(c, vir_inst(op, c->undef, \
868 a, b)); \
869 }
870
871 #define VIR_SFU(name) \
872 static inline struct qreg \
873 vir_##name(struct v3d_compile *c, struct qreg a) \
874 { \
875 if (c->devinfo->ver >= 41) { \
876 return vir_emit_def(c, vir_add_inst(V3D_QPU_A_##name, \
877 c->undef, \
878 a, c->undef)); \
879 } else { \
880 vir_FMOV_dest(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_##name), a); \
881 return vir_FMOV(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_R4)); \
882 } \
883 } \
884 static inline struct qinst * \
885 vir_##name##_dest(struct v3d_compile *c, struct qreg dest, \
886 struct qreg a) \
887 { \
888 if (c->devinfo->ver >= 41) { \
889 return vir_emit_nondef(c, vir_add_inst(V3D_QPU_A_##name, \
890 dest, \
891 a, c->undef)); \
892 } else { \
893 vir_FMOV_dest(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_##name), a); \
894 return vir_FMOV_dest(c, dest, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_R4)); \
895 } \
896 }
897
898 #define VIR_A_ALU2(name) VIR_ALU2(name, vir_add_inst, V3D_QPU_A_##name)
899 #define VIR_M_ALU2(name) VIR_ALU2(name, vir_mul_inst, V3D_QPU_M_##name)
900 #define VIR_A_ALU1(name) VIR_ALU1(name, vir_add_inst, V3D_QPU_A_##name)
901 #define VIR_M_ALU1(name) VIR_ALU1(name, vir_mul_inst, V3D_QPU_M_##name)
902 #define VIR_A_ALU0(name) VIR_ALU0(name, vir_add_inst, V3D_QPU_A_##name)
903 #define VIR_M_ALU0(name) VIR_ALU0(name, vir_mul_inst, V3D_QPU_M_##name)
904 #define VIR_A_NODST_2(name) VIR_NODST_2(name, vir_add_inst, V3D_QPU_A_##name)
905 #define VIR_M_NODST_2(name) VIR_NODST_2(name, vir_mul_inst, V3D_QPU_M_##name)
906 #define VIR_A_NODST_1(name) VIR_NODST_1(name, vir_add_inst, V3D_QPU_A_##name)
907 #define VIR_M_NODST_1(name) VIR_NODST_1(name, vir_mul_inst, V3D_QPU_M_##name)
908 #define VIR_A_NODST_0(name) VIR_NODST_0(name, vir_add_inst, V3D_QPU_A_##name)
909
910 VIR_A_ALU2(FADD)
911 VIR_A_ALU2(VFPACK)
912 VIR_A_ALU2(FSUB)
913 VIR_A_ALU2(FMIN)
914 VIR_A_ALU2(FMAX)
915
916 VIR_A_ALU2(ADD)
917 VIR_A_ALU2(SUB)
918 VIR_A_ALU2(SHL)
919 VIR_A_ALU2(SHR)
920 VIR_A_ALU2(ASR)
921 VIR_A_ALU2(ROR)
922 VIR_A_ALU2(MIN)
923 VIR_A_ALU2(MAX)
924 VIR_A_ALU2(UMIN)
925 VIR_A_ALU2(UMAX)
926 VIR_A_ALU2(AND)
927 VIR_A_ALU2(OR)
928 VIR_A_ALU2(XOR)
929 VIR_A_ALU2(VADD)
930 VIR_A_ALU2(VSUB)
931 VIR_A_ALU2(STVPMV)
932 VIR_A_ALU1(NOT)
933 VIR_A_ALU1(NEG)
934 VIR_A_ALU1(FLAPUSH)
935 VIR_A_ALU1(FLBPUSH)
936 VIR_A_ALU1(FLPOP)
937 VIR_A_ALU1(SETMSF)
938 VIR_A_ALU1(SETREVF)
939 VIR_A_ALU0(TIDX)
940 VIR_A_ALU0(EIDX)
941 VIR_A_ALU1(LDVPMV_IN)
942 VIR_A_ALU1(LDVPMV_OUT)
943
944 VIR_A_ALU0(FXCD)
945 VIR_A_ALU0(XCD)
946 VIR_A_ALU0(FYCD)
947 VIR_A_ALU0(YCD)
948 VIR_A_ALU0(MSF)
949 VIR_A_ALU0(REVF)
950 VIR_A_NODST_1(VPMSETUP)
951 VIR_A_NODST_0(VPMWT)
952 VIR_A_ALU2(FCMP)
953 VIR_A_ALU2(VFMAX)
954
955 VIR_A_ALU1(FROUND)
956 VIR_A_ALU1(FTOIN)
957 VIR_A_ALU1(FTRUNC)
958 VIR_A_ALU1(FTOIZ)
959 VIR_A_ALU1(FFLOOR)
960 VIR_A_ALU1(FTOUZ)
961 VIR_A_ALU1(FCEIL)
962 VIR_A_ALU1(FTOC)
963
964 VIR_A_ALU1(FDX)
965 VIR_A_ALU1(FDY)
966
967 VIR_A_ALU1(ITOF)
968 VIR_A_ALU1(CLZ)
969 VIR_A_ALU1(UTOF)
970
971 VIR_M_ALU2(UMUL24)
972 VIR_M_ALU2(FMUL)
973 VIR_M_ALU2(SMUL24)
974 VIR_M_NODST_2(MULTOP)
975
976 VIR_M_ALU1(MOV)
977 VIR_M_ALU1(FMOV)
978
979 VIR_SFU(RECIP)
980 VIR_SFU(RSQRT)
981 VIR_SFU(EXP)
982 VIR_SFU(LOG)
983 VIR_SFU(SIN)
984 VIR_SFU(RSQRT2)
985
986 static inline struct qinst *
987 vir_MOV_cond(struct v3d_compile *c, enum v3d_qpu_cond cond,
988 struct qreg dest, struct qreg src)
989 {
990 struct qinst *mov = vir_MOV_dest(c, dest, src);
991 vir_set_cond(mov, cond);
992 return mov;
993 }
994
995 static inline struct qreg
996 vir_SEL(struct v3d_compile *c, enum v3d_qpu_cond cond,
997 struct qreg src0, struct qreg src1)
998 {
999 struct qreg t = vir_get_temp(c);
1000 vir_MOV_dest(c, t, src1);
1001 vir_MOV_cond(c, cond, t, src0);
1002 return t;
1003 }
1004
1005 static inline struct qinst *
1006 vir_NOP(struct v3d_compile *c)
1007 {
1008 return vir_emit_nondef(c, vir_add_inst(V3D_QPU_A_NOP,
1009 c->undef, c->undef, c->undef));
1010 }
1011
1012 static inline struct qreg
1013 vir_LDTMU(struct v3d_compile *c)
1014 {
1015 if (c->devinfo->ver >= 41) {
1016 struct qinst *ldtmu = vir_add_inst(V3D_QPU_A_NOP, c->undef,
1017 c->undef, c->undef);
1018 ldtmu->qpu.sig.ldtmu = true;
1019
1020 return vir_emit_def(c, ldtmu);
1021 } else {
1022 vir_NOP(c)->qpu.sig.ldtmu = true;
1023 return vir_MOV(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_R4));
1024 }
1025 }
1026
1027 static inline struct qreg
1028 vir_UMUL(struct v3d_compile *c, struct qreg src0, struct qreg src1)
1029 {
1030 vir_MULTOP(c, src0, src1);
1031 return vir_UMUL24(c, src0, src1);
1032 }
1033
1034 /*
1035 static inline struct qreg
1036 vir_LOAD_IMM(struct v3d_compile *c, uint32_t val)
1037 {
1038 return vir_emit_def(c, vir_inst(QOP_LOAD_IMM, c->undef,
1039 vir_reg(QFILE_LOAD_IMM, val), c->undef));
1040 }
1041
1042 static inline struct qreg
1043 vir_LOAD_IMM_U2(struct v3d_compile *c, uint32_t val)
1044 {
1045 return vir_emit_def(c, vir_inst(QOP_LOAD_IMM_U2, c->undef,
1046 vir_reg(QFILE_LOAD_IMM, val),
1047 c->undef));
1048 }
1049 static inline struct qreg
1050 vir_LOAD_IMM_I2(struct v3d_compile *c, uint32_t val)
1051 {
1052 return vir_emit_def(c, vir_inst(QOP_LOAD_IMM_I2, c->undef,
1053 vir_reg(QFILE_LOAD_IMM, val),
1054 c->undef));
1055 }
1056 */
1057
1058 static inline struct qinst *
1059 vir_BRANCH(struct v3d_compile *c, enum v3d_qpu_cond cond)
1060 {
1061 /* The actual uniform_data value will be set at scheduling time */
1062 return vir_emit_nondef(c, vir_branch_inst(cond, vir_uniform_ui(c, 0)));
1063 }
1064
1065 #define vir_for_each_block(block, c) \
1066 list_for_each_entry(struct qblock, block, &c->blocks, link)
1067
1068 #define vir_for_each_block_rev(block, c) \
1069 list_for_each_entry_rev(struct qblock, block, &c->blocks, link)
1070
1071 /* Loop over the non-NULL members of the successors array. */
1072 #define vir_for_each_successor(succ, block) \
1073 for (struct qblock *succ = block->successors[0]; \
1074 succ != NULL; \
1075 succ = (succ == block->successors[1] ? NULL : \
1076 block->successors[1]))
1077
1078 #define vir_for_each_inst(inst, block) \
1079 list_for_each_entry(struct qinst, inst, &block->instructions, link)
1080
1081 #define vir_for_each_inst_rev(inst, block) \
1082 list_for_each_entry_rev(struct qinst, inst, &block->instructions, link)
1083
1084 #define vir_for_each_inst_safe(inst, block) \
1085 list_for_each_entry_safe(struct qinst, inst, &block->instructions, link)
1086
1087 #define vir_for_each_inst_inorder(inst, c) \
1088 vir_for_each_block(_block, c) \
1089 vir_for_each_inst(inst, _block)
1090
1091 #endif /* V3D_COMPILER_H */