v3d: Switch implicit uniforms over to being any qinst->uniform != ~0.
[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 "common/v3d_limits.h"
38 #include "compiler/nir/nir.h"
39 #include "util/list.h"
40 #include "util/u_math.h"
41
42 #include "qpu/qpu_instr.h"
43 #include "pipe/p_state.h"
44
45 struct nir_builder;
46
47 struct v3d_fs_inputs {
48 /**
49 * Array of the meanings of the VPM inputs this shader needs.
50 *
51 * It doesn't include those that aren't part of the VPM, like
52 * point/line coordinates.
53 */
54 struct v3d_varying_slot *input_slots;
55 uint32_t num_inputs;
56 };
57
58 enum qfile {
59 /** An unused source or destination register. */
60 QFILE_NULL,
61
62 /** A physical register, such as the W coordinate payload. */
63 QFILE_REG,
64 /** One of the regsiters for fixed function interactions. */
65 QFILE_MAGIC,
66
67 /**
68 * A virtual register, that will be allocated to actual accumulator
69 * or physical registers later.
70 */
71 QFILE_TEMP,
72 QFILE_UNIF,
73 QFILE_TLB,
74 QFILE_TLBU,
75
76 /**
77 * VPM reads use this with an index value to say what part of the VPM
78 * is being read.
79 */
80 QFILE_VPM,
81
82 /**
83 * Stores an immediate value in the index field that will be used
84 * directly by qpu_load_imm().
85 */
86 QFILE_LOAD_IMM,
87
88 /**
89 * Stores an immediate value in the index field that can be turned
90 * into a small immediate field by qpu_encode_small_immediate().
91 */
92 QFILE_SMALL_IMM,
93 };
94
95 /**
96 * A reference to a QPU register or a virtual temp register.
97 */
98 struct qreg {
99 enum qfile file;
100 uint32_t index;
101 };
102
103 static inline struct qreg vir_reg(enum qfile file, uint32_t index)
104 {
105 return (struct qreg){file, index};
106 }
107
108 static inline struct qreg vir_nop_reg(void)
109 {
110 return (struct qreg){QFILE_NULL, 0};
111 }
112
113 /**
114 * A reference to an actual register at the QPU level, for register
115 * allocation.
116 */
117 struct qpu_reg {
118 bool magic;
119 bool smimm;
120 int index;
121 };
122
123 struct qinst {
124 /** Entry in qblock->instructions */
125 struct list_head link;
126
127 /**
128 * The instruction being wrapped. Its condition codes, pack flags,
129 * signals, etc. will all be used, with just the register references
130 * being replaced by the contents of qinst->dst and qinst->src[].
131 */
132 struct v3d_qpu_instr qpu;
133
134 /* Pre-register-allocation references to src/dst registers */
135 struct qreg dst;
136 struct qreg src[3];
137 bool is_last_thrsw;
138
139 /* If the instruction reads a uniform (other than through src[i].file
140 * == QFILE_UNIF), that uniform's index in c->uniform_contents. ~0
141 * otherwise.
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_IMAGE_TMU_CONFIG_P0,
234
235 QUNIFORM_TEXTURE_FIRST_LEVEL,
236
237 QUNIFORM_TEXTURE_WIDTH,
238 QUNIFORM_TEXTURE_HEIGHT,
239 QUNIFORM_TEXTURE_DEPTH,
240 QUNIFORM_TEXTURE_ARRAY_SIZE,
241 QUNIFORM_TEXTURE_LEVELS,
242
243 QUNIFORM_UBO_ADDR,
244
245 QUNIFORM_TEXRECT_SCALE_X,
246 QUNIFORM_TEXRECT_SCALE_Y,
247
248 /* Returns the base offset of the SSBO given by the data value. */
249 QUNIFORM_SSBO_OFFSET,
250
251 /* Returns the size of the SSBO given by the data value. */
252 QUNIFORM_GET_BUFFER_SIZE,
253
254 /* Sizes (in pixels) of a shader image given by the data value. */
255 QUNIFORM_IMAGE_WIDTH,
256 QUNIFORM_IMAGE_HEIGHT,
257 QUNIFORM_IMAGE_DEPTH,
258 QUNIFORM_IMAGE_ARRAY_SIZE,
259
260 QUNIFORM_ALPHA_REF,
261
262 /* Number of workgroups passed to glDispatchCompute in the dimension
263 * selected by the data value.
264 */
265 QUNIFORM_NUM_WORK_GROUPS,
266
267 /**
268 * Returns the the offset of the scratch buffer for register spilling.
269 */
270 QUNIFORM_SPILL_OFFSET,
271 QUNIFORM_SPILL_SIZE_PER_THREAD,
272
273 /**
274 * Returns the offset of the shared memory for compute shaders.
275 *
276 * This will be accessed using TMU general memory operations, so the
277 * L2T cache will effectively be the shared memory area.
278 */
279 QUNIFORM_SHARED_OFFSET,
280 };
281
282 static inline uint32_t v3d_tmu_config_data_create(uint32_t unit, uint32_t value)
283 {
284 return unit << 24 | value;
285 }
286
287 static inline uint32_t v3d_tmu_config_data_get_unit(uint32_t data)
288 {
289 return data >> 24;
290 }
291
292 static inline uint32_t v3d_tmu_config_data_get_value(uint32_t data)
293 {
294 return data & 0xffffff;
295 }
296
297 struct v3d_varying_slot {
298 uint8_t slot_and_component;
299 };
300
301 static inline struct v3d_varying_slot
302 v3d_slot_from_slot_and_component(uint8_t slot, uint8_t component)
303 {
304 assert(slot < 255 / 4);
305 return (struct v3d_varying_slot){ (slot << 2) + component };
306 }
307
308 static inline uint8_t v3d_slot_get_slot(struct v3d_varying_slot slot)
309 {
310 return slot.slot_and_component >> 2;
311 }
312
313 static inline uint8_t v3d_slot_get_component(struct v3d_varying_slot slot)
314 {
315 return slot.slot_and_component & 3;
316 }
317
318 struct v3d_ubo_range {
319 /**
320 * offset in bytes from the start of the ubo where this range is
321 * uploaded.
322 *
323 * Only set once used is set.
324 */
325 uint32_t dst_offset;
326
327 /**
328 * offset in bytes from the start of the gallium uniforms where the
329 * data comes from.
330 */
331 uint32_t src_offset;
332
333 /** size in bytes of this ubo range */
334 uint32_t size;
335 };
336
337 struct v3d_key {
338 void *shader_state;
339 struct {
340 uint8_t swizzle[4];
341 uint8_t return_size;
342 uint8_t return_channels;
343 bool clamp_s:1;
344 bool clamp_t:1;
345 bool clamp_r:1;
346 } tex[V3D_MAX_TEXTURE_SAMPLERS];
347 uint8_t ucp_enables;
348 };
349
350 struct v3d_fs_key {
351 struct v3d_key base;
352 bool depth_enabled;
353 bool is_points;
354 bool is_lines;
355 bool alpha_test;
356 bool point_coord_upper_left;
357 bool light_twoside;
358 bool msaa;
359 bool sample_coverage;
360 bool sample_alpha_to_coverage;
361 bool sample_alpha_to_one;
362 bool clamp_color;
363 bool shade_model_flat;
364 /* Mask of which color render targets are present. */
365 uint8_t cbufs;
366 uint8_t swap_color_rb;
367 /* Mask of which render targets need to be written as 32-bit floats */
368 uint8_t f32_color_rb;
369 /* Masks of which render targets need to be written as ints/uints.
370 * Used by gallium to work around lost information in TGSI.
371 */
372 uint8_t int_color_rb;
373 uint8_t uint_color_rb;
374 uint8_t alpha_test_func;
375 uint8_t logicop_func;
376 uint32_t point_sprite_mask;
377
378 struct pipe_rt_blend_state blend;
379 };
380
381 struct v3d_vs_key {
382 struct v3d_key base;
383
384 struct v3d_varying_slot fs_inputs[V3D_MAX_FS_INPUTS];
385 uint8_t num_fs_inputs;
386
387 bool is_coord;
388 bool per_vertex_point_size;
389 bool clamp_color;
390 };
391
392 /** A basic block of VIR intructions. */
393 struct qblock {
394 struct list_head link;
395
396 struct list_head instructions;
397
398 struct set *predecessors;
399 struct qblock *successors[2];
400
401 int index;
402
403 /* Instruction IPs for the first and last instruction of the block.
404 * Set by qpu_schedule.c.
405 */
406 uint32_t start_qpu_ip;
407 uint32_t end_qpu_ip;
408
409 /* Instruction IP for the branch instruction of the block. Set by
410 * qpu_schedule.c.
411 */
412 uint32_t branch_qpu_ip;
413
414 /** Offset within the uniform stream at the start of the block. */
415 uint32_t start_uniform;
416 /** Offset within the uniform stream of the branch instruction */
417 uint32_t branch_uniform;
418
419 /** @{ used by v3d_vir_live_variables.c */
420 BITSET_WORD *def;
421 BITSET_WORD *defin;
422 BITSET_WORD *defout;
423 BITSET_WORD *use;
424 BITSET_WORD *live_in;
425 BITSET_WORD *live_out;
426 int start_ip, end_ip;
427 /** @} */
428 };
429
430 /** Which util/list.h add mode we should use when inserting an instruction. */
431 enum vir_cursor_mode {
432 vir_cursor_add,
433 vir_cursor_addtail,
434 };
435
436 /**
437 * Tracking structure for where new instructions should be inserted. Create
438 * with one of the vir_after_inst()-style helper functions.
439 *
440 * This does not protect against removal of the block or instruction, so we
441 * have an assert in instruction removal to try to catch it.
442 */
443 struct vir_cursor {
444 enum vir_cursor_mode mode;
445 struct list_head *link;
446 };
447
448 static inline struct vir_cursor
449 vir_before_inst(struct qinst *inst)
450 {
451 return (struct vir_cursor){ vir_cursor_addtail, &inst->link };
452 }
453
454 static inline struct vir_cursor
455 vir_after_inst(struct qinst *inst)
456 {
457 return (struct vir_cursor){ vir_cursor_add, &inst->link };
458 }
459
460 static inline struct vir_cursor
461 vir_before_block(struct qblock *block)
462 {
463 return (struct vir_cursor){ vir_cursor_add, &block->instructions };
464 }
465
466 static inline struct vir_cursor
467 vir_after_block(struct qblock *block)
468 {
469 return (struct vir_cursor){ vir_cursor_addtail, &block->instructions };
470 }
471
472 /**
473 * Compiler state saved across compiler invocations, for any expensive global
474 * setup.
475 */
476 struct v3d_compiler {
477 const struct v3d_device_info *devinfo;
478 struct ra_regs *regs;
479 unsigned int reg_class_phys[3];
480 unsigned int reg_class_phys_or_acc[3];
481 };
482
483 struct v3d_compile {
484 const struct v3d_device_info *devinfo;
485 nir_shader *s;
486 nir_function_impl *impl;
487 struct exec_list *cf_node_list;
488 const struct v3d_compiler *compiler;
489
490 void (*debug_output)(const char *msg,
491 void *debug_output_data);
492 void *debug_output_data;
493
494 /**
495 * Mapping from nir_register * or nir_ssa_def * to array of struct
496 * qreg for the values.
497 */
498 struct hash_table *def_ht;
499
500 /* For each temp, the instruction generating its value. */
501 struct qinst **defs;
502 uint32_t defs_array_size;
503
504 /**
505 * Inputs to the shader, arranged by TGSI declaration order.
506 *
507 * Not all fragment shader QFILE_VARY reads are present in this array.
508 */
509 struct qreg *inputs;
510 struct qreg *outputs;
511 bool msaa_per_sample_output;
512 struct qreg color_reads[V3D_MAX_SAMPLES];
513 struct qreg sample_colors[V3D_MAX_SAMPLES];
514 uint32_t inputs_array_size;
515 uint32_t outputs_array_size;
516 uint32_t uniforms_array_size;
517
518 /* Booleans for whether the corresponding QFILE_VARY[i] is
519 * flat-shaded. This includes gl_FragColor flat-shading, which is
520 * customized based on the shademodel_flat shader key.
521 */
522 uint32_t flat_shade_flags[BITSET_WORDS(V3D_MAX_FS_INPUTS)];
523
524 uint32_t noperspective_flags[BITSET_WORDS(V3D_MAX_FS_INPUTS)];
525
526 uint32_t centroid_flags[BITSET_WORDS(V3D_MAX_FS_INPUTS)];
527
528 bool uses_center_w;
529 bool writes_z;
530
531 struct v3d_ubo_range *ubo_ranges;
532 bool *ubo_range_used;
533 uint32_t ubo_ranges_array_size;
534 /** Number of uniform areas tracked in ubo_ranges. */
535 uint32_t num_ubo_ranges;
536 uint32_t next_ubo_dst_offset;
537
538 /* State for whether we're executing on each channel currently. 0 if
539 * yes, otherwise a block number + 1 that the channel jumped to.
540 */
541 struct qreg execute;
542 bool in_control_flow;
543
544 struct qreg line_x, point_x, point_y;
545
546 /**
547 * Instance ID, which comes in before the vertex attribute payload if
548 * the shader record requests it.
549 */
550 struct qreg iid;
551
552 /**
553 * Vertex ID, which comes in before the vertex attribute payload
554 * (after Instance ID) if the shader record requests it.
555 */
556 struct qreg vid;
557
558 /* Fragment shader payload regs. */
559 struct qreg payload_w, payload_w_centroid, payload_z;
560
561 struct qreg cs_payload[2];
562 struct qreg cs_shared_offset;
563 int local_invocation_index_bits;
564
565 uint8_t vattr_sizes[V3D_MAX_VS_INPUTS / 4];
566 uint32_t vpm_output_size;
567
568 /* Size in bytes of registers that have been spilled. This is how much
569 * space needs to be available in the spill BO per thread per QPU.
570 */
571 uint32_t spill_size;
572 /* Shader-db stats */
573 uint32_t spills, fills, loops;
574 /**
575 * Register spilling's per-thread base address, shared between each
576 * spill/fill's addressing calculations.
577 */
578 struct qreg spill_base;
579 /* Bit vector of which temps may be spilled */
580 BITSET_WORD *spillable;
581
582 /**
583 * Array of the VARYING_SLOT_* of all FS QFILE_VARY reads.
584 *
585 * This includes those that aren't part of the VPM varyings, like
586 * point/line coordinates.
587 */
588 struct v3d_varying_slot input_slots[V3D_MAX_FS_INPUTS];
589
590 /**
591 * An entry per outputs[] in the VS indicating what the VARYING_SLOT_*
592 * of the output is. Used to emit from the VS in the order that the
593 * FS needs.
594 */
595 struct v3d_varying_slot *output_slots;
596
597 struct pipe_shader_state *shader_state;
598 struct v3d_key *key;
599 struct v3d_fs_key *fs_key;
600 struct v3d_vs_key *vs_key;
601
602 /* Live ranges of temps. */
603 int *temp_start, *temp_end;
604 bool live_intervals_valid;
605
606 uint32_t *uniform_data;
607 enum quniform_contents *uniform_contents;
608 uint32_t uniform_array_size;
609 uint32_t num_uniforms;
610 uint32_t output_position_index;
611 nir_variable *output_color_var[4];
612 uint32_t output_sample_mask_index;
613
614 struct qreg undef;
615 uint32_t num_temps;
616
617 struct vir_cursor cursor;
618 struct list_head blocks;
619 int next_block_index;
620 struct qblock *cur_block;
621 struct qblock *loop_cont_block;
622 struct qblock *loop_break_block;
623
624 uint64_t *qpu_insts;
625 uint32_t qpu_inst_count;
626 uint32_t qpu_inst_size;
627
628 /* For the FS, the number of varying inputs not counting the
629 * point/line varyings payload
630 */
631 uint32_t num_inputs;
632
633 /**
634 * Number of inputs from num_inputs remaining to be queued to the read
635 * FIFO in the VS/CS.
636 */
637 uint32_t num_inputs_remaining;
638
639 /* Number of inputs currently in the read FIFO for the VS/CS */
640 uint32_t num_inputs_in_fifo;
641
642 /** Next offset in the VPM to read from in the VS/CS */
643 uint32_t vpm_read_offset;
644
645 uint32_t program_id;
646 uint32_t variant_id;
647
648 /* Set to compile program in in 1x, 2x, or 4x threaded mode, where
649 * SIG_THREAD_SWITCH is used to hide texturing latency at the cost of
650 * limiting ourselves to the part of the physical reg space.
651 *
652 * On V3D 3.x, 2x or 4x divide the physical reg space by 2x or 4x. On
653 * V3D 4.x, all shaders are 2x threaded, and 4x only divides the
654 * physical reg space in half.
655 */
656 uint8_t threads;
657 struct qinst *last_thrsw;
658 bool last_thrsw_at_top_level;
659
660 bool failed;
661 };
662
663 struct v3d_uniform_list {
664 enum quniform_contents *contents;
665 uint32_t *data;
666 uint32_t count;
667 };
668
669 struct v3d_prog_data {
670 struct v3d_uniform_list uniforms;
671
672 struct v3d_ubo_range *ubo_ranges;
673 uint32_t num_ubo_ranges;
674 uint32_t ubo_size;
675 uint32_t spill_size;
676
677 uint8_t threads;
678
679 /* For threads > 1, whether the program should be dispatched in the
680 * after-final-THRSW state.
681 */
682 bool single_seg;
683 };
684
685 struct v3d_vs_prog_data {
686 struct v3d_prog_data base;
687
688 bool uses_iid, uses_vid;
689
690 /* Number of components read from each vertex attribute. */
691 uint8_t vattr_sizes[V3D_MAX_VS_INPUTS / 4];
692
693 /* Total number of components read, for the shader state record. */
694 uint32_t vpm_input_size;
695
696 /* Total number of components written, for the shader state record. */
697 uint32_t vpm_output_size;
698
699 /* Set if there should be separate VPM segments for input and output.
700 * If unset, vpm_input_size will be 0.
701 */
702 bool separate_segments;
703
704 /* Value to be programmed in VCM_CACHE_SIZE. */
705 uint8_t vcm_cache_size;
706 };
707
708 struct v3d_fs_prog_data {
709 struct v3d_prog_data base;
710
711 struct v3d_varying_slot input_slots[V3D_MAX_FS_INPUTS];
712
713 /* Array of flat shade flags.
714 *
715 * Each entry is only 24 bits (high 8 bits 0), to match the hardware
716 * packet layout.
717 */
718 uint32_t flat_shade_flags[((V3D_MAX_FS_INPUTS - 1) / 24) + 1];
719
720 uint32_t noperspective_flags[((V3D_MAX_FS_INPUTS - 1) / 24) + 1];
721
722 uint32_t centroid_flags[((V3D_MAX_FS_INPUTS - 1) / 24) + 1];
723
724 uint8_t num_inputs;
725 bool writes_z;
726 bool disable_ez;
727 bool uses_center_w;
728 };
729
730 static inline bool
731 vir_has_uniform(struct qinst *inst)
732 {
733 return inst->uniform != ~0;
734 }
735
736 /* Special nir_load_input intrinsic index for loading the current TLB
737 * destination color.
738 */
739 #define V3D_NIR_TLB_COLOR_READ_INPUT 2000000000
740
741 #define V3D_NIR_MS_MASK_OUTPUT 2000000000
742
743 extern const nir_shader_compiler_options v3d_nir_options;
744
745 const struct v3d_compiler *v3d_compiler_init(const struct v3d_device_info *devinfo);
746 void v3d_compiler_free(const struct v3d_compiler *compiler);
747 void v3d_optimize_nir(struct nir_shader *s);
748
749 uint64_t *v3d_compile(const struct v3d_compiler *compiler,
750 struct v3d_key *key,
751 struct v3d_prog_data **prog_data,
752 nir_shader *s,
753 void (*debug_output)(const char *msg,
754 void *debug_output_data),
755 void *debug_output_data,
756 int program_id, int variant_id,
757 uint32_t *final_assembly_size);
758
759 void v3d_nir_to_vir(struct v3d_compile *c);
760
761 void vir_compile_destroy(struct v3d_compile *c);
762 const char *vir_get_stage_name(struct v3d_compile *c);
763 struct qblock *vir_new_block(struct v3d_compile *c);
764 void vir_set_emit_block(struct v3d_compile *c, struct qblock *block);
765 void vir_link_blocks(struct qblock *predecessor, struct qblock *successor);
766 struct qblock *vir_entry_block(struct v3d_compile *c);
767 struct qblock *vir_exit_block(struct v3d_compile *c);
768 struct qinst *vir_add_inst(enum v3d_qpu_add_op op, struct qreg dst,
769 struct qreg src0, struct qreg src1);
770 struct qinst *vir_mul_inst(enum v3d_qpu_mul_op op, struct qreg dst,
771 struct qreg src0, struct qreg src1);
772 struct qinst *vir_branch_inst(struct v3d_compile *c,
773 enum v3d_qpu_branch_cond cond);
774 void vir_remove_instruction(struct v3d_compile *c, struct qinst *qinst);
775 uint32_t vir_get_uniform_index(struct v3d_compile *c,
776 enum quniform_contents contents,
777 uint32_t data);
778 struct qreg vir_uniform(struct v3d_compile *c,
779 enum quniform_contents contents,
780 uint32_t data);
781 void vir_schedule_instructions(struct v3d_compile *c);
782 struct v3d_qpu_instr v3d_qpu_nop(void);
783
784 struct qreg vir_emit_def(struct v3d_compile *c, struct qinst *inst);
785 struct qinst *vir_emit_nondef(struct v3d_compile *c, struct qinst *inst);
786 void vir_set_cond(struct qinst *inst, enum v3d_qpu_cond cond);
787 void vir_set_pf(struct qinst *inst, enum v3d_qpu_pf pf);
788 void vir_set_uf(struct qinst *inst, enum v3d_qpu_uf uf);
789 void vir_set_unpack(struct qinst *inst, int src,
790 enum v3d_qpu_input_unpack unpack);
791
792 struct qreg vir_get_temp(struct v3d_compile *c);
793 void vir_emit_last_thrsw(struct v3d_compile *c);
794 void vir_calculate_live_intervals(struct v3d_compile *c);
795 int vir_get_nsrc(struct qinst *inst);
796 bool vir_has_side_effects(struct v3d_compile *c, struct qinst *inst);
797 bool vir_get_add_op(struct qinst *inst, enum v3d_qpu_add_op *op);
798 bool vir_get_mul_op(struct qinst *inst, enum v3d_qpu_mul_op *op);
799 bool vir_is_raw_mov(struct qinst *inst);
800 bool vir_is_tex(struct qinst *inst);
801 bool vir_is_add(struct qinst *inst);
802 bool vir_is_mul(struct qinst *inst);
803 bool vir_writes_r3(const struct v3d_device_info *devinfo, struct qinst *inst);
804 bool vir_writes_r4(const struct v3d_device_info *devinfo, struct qinst *inst);
805 struct qreg vir_follow_movs(struct v3d_compile *c, struct qreg reg);
806 uint8_t vir_channels_written(struct qinst *inst);
807 struct qreg ntq_get_src(struct v3d_compile *c, nir_src src, int i);
808 void ntq_store_dest(struct v3d_compile *c, nir_dest *dest, int chan,
809 struct qreg result);
810 void vir_emit_thrsw(struct v3d_compile *c);
811
812 void vir_dump(struct v3d_compile *c);
813 void vir_dump_inst(struct v3d_compile *c, struct qinst *inst);
814 void vir_dump_uniform(enum quniform_contents contents, uint32_t data);
815
816 void vir_validate(struct v3d_compile *c);
817
818 void vir_optimize(struct v3d_compile *c);
819 bool vir_opt_algebraic(struct v3d_compile *c);
820 bool vir_opt_constant_folding(struct v3d_compile *c);
821 bool vir_opt_copy_propagate(struct v3d_compile *c);
822 bool vir_opt_dead_code(struct v3d_compile *c);
823 bool vir_opt_peephole_sf(struct v3d_compile *c);
824 bool vir_opt_small_immediates(struct v3d_compile *c);
825 bool vir_opt_vpm(struct v3d_compile *c);
826 void v3d_nir_lower_blend(nir_shader *s, struct v3d_compile *c);
827 void v3d_nir_lower_io(nir_shader *s, struct v3d_compile *c);
828 void v3d_nir_lower_txf_ms(nir_shader *s, struct v3d_compile *c);
829 void v3d_nir_lower_image_load_store(nir_shader *s);
830 void vir_lower_uniforms(struct v3d_compile *c);
831
832 void v3d33_vir_vpm_read_setup(struct v3d_compile *c, int num_components);
833 void v3d33_vir_vpm_write_setup(struct v3d_compile *c);
834 void v3d33_vir_emit_tex(struct v3d_compile *c, nir_tex_instr *instr);
835 void v3d40_vir_emit_tex(struct v3d_compile *c, nir_tex_instr *instr);
836 void v3d40_vir_emit_image_load_store(struct v3d_compile *c,
837 nir_intrinsic_instr *instr);
838
839 void v3d_vir_to_qpu(struct v3d_compile *c, struct qpu_reg *temp_registers);
840 uint32_t v3d_qpu_schedule_instructions(struct v3d_compile *c);
841 void qpu_validate(struct v3d_compile *c);
842 struct qpu_reg *v3d_register_allocate(struct v3d_compile *c, bool *spilled);
843 bool vir_init_reg_sets(struct v3d_compiler *compiler);
844
845 bool v3d_gl_format_is_return_32(GLenum format);
846
847 static inline bool
848 quniform_contents_is_texture_p0(enum quniform_contents contents)
849 {
850 return (contents >= QUNIFORM_TEXTURE_CONFIG_P0_0 &&
851 contents < (QUNIFORM_TEXTURE_CONFIG_P0_0 +
852 V3D_MAX_TEXTURE_SAMPLERS));
853 }
854
855 static inline bool
856 vir_in_nonuniform_control_flow(struct v3d_compile *c)
857 {
858 return c->execute.file != QFILE_NULL;
859 }
860
861 static inline struct qreg
862 vir_uniform_ui(struct v3d_compile *c, uint32_t ui)
863 {
864 return vir_uniform(c, QUNIFORM_CONSTANT, ui);
865 }
866
867 static inline struct qreg
868 vir_uniform_f(struct v3d_compile *c, float f)
869 {
870 return vir_uniform(c, QUNIFORM_CONSTANT, fui(f));
871 }
872
873 #define VIR_ALU0(name, vir_inst, op) \
874 static inline struct qreg \
875 vir_##name(struct v3d_compile *c) \
876 { \
877 return vir_emit_def(c, vir_inst(op, c->undef, \
878 c->undef, c->undef)); \
879 } \
880 static inline struct qinst * \
881 vir_##name##_dest(struct v3d_compile *c, struct qreg dest) \
882 { \
883 return vir_emit_nondef(c, vir_inst(op, dest, \
884 c->undef, c->undef)); \
885 }
886
887 #define VIR_ALU1(name, vir_inst, op) \
888 static inline struct qreg \
889 vir_##name(struct v3d_compile *c, struct qreg a) \
890 { \
891 return vir_emit_def(c, vir_inst(op, c->undef, \
892 a, c->undef)); \
893 } \
894 static inline struct qinst * \
895 vir_##name##_dest(struct v3d_compile *c, struct qreg dest, \
896 struct qreg a) \
897 { \
898 return vir_emit_nondef(c, vir_inst(op, dest, a, \
899 c->undef)); \
900 }
901
902 #define VIR_ALU2(name, vir_inst, op) \
903 static inline struct qreg \
904 vir_##name(struct v3d_compile *c, struct qreg a, struct qreg b) \
905 { \
906 return vir_emit_def(c, vir_inst(op, c->undef, a, b)); \
907 } \
908 static inline struct qinst * \
909 vir_##name##_dest(struct v3d_compile *c, struct qreg dest, \
910 struct qreg a, struct qreg b) \
911 { \
912 return vir_emit_nondef(c, vir_inst(op, dest, a, b)); \
913 }
914
915 #define VIR_NODST_0(name, vir_inst, op) \
916 static inline struct qinst * \
917 vir_##name(struct v3d_compile *c) \
918 { \
919 return vir_emit_nondef(c, vir_inst(op, c->undef, \
920 c->undef, c->undef)); \
921 }
922
923 #define VIR_NODST_1(name, vir_inst, op) \
924 static inline struct qinst * \
925 vir_##name(struct v3d_compile *c, struct qreg a) \
926 { \
927 return vir_emit_nondef(c, vir_inst(op, c->undef, \
928 a, c->undef)); \
929 }
930
931 #define VIR_NODST_2(name, vir_inst, op) \
932 static inline struct qinst * \
933 vir_##name(struct v3d_compile *c, struct qreg a, struct qreg b) \
934 { \
935 return vir_emit_nondef(c, vir_inst(op, c->undef, \
936 a, b)); \
937 }
938
939 #define VIR_SFU(name) \
940 static inline struct qreg \
941 vir_##name(struct v3d_compile *c, struct qreg a) \
942 { \
943 if (c->devinfo->ver >= 41) { \
944 return vir_emit_def(c, vir_add_inst(V3D_QPU_A_##name, \
945 c->undef, \
946 a, c->undef)); \
947 } else { \
948 vir_FMOV_dest(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_##name), a); \
949 return vir_FMOV(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_R4)); \
950 } \
951 } \
952 static inline struct qinst * \
953 vir_##name##_dest(struct v3d_compile *c, struct qreg dest, \
954 struct qreg a) \
955 { \
956 if (c->devinfo->ver >= 41) { \
957 return vir_emit_nondef(c, vir_add_inst(V3D_QPU_A_##name, \
958 dest, \
959 a, c->undef)); \
960 } else { \
961 vir_FMOV_dest(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_##name), a); \
962 return vir_FMOV_dest(c, dest, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_R4)); \
963 } \
964 }
965
966 #define VIR_A_ALU2(name) VIR_ALU2(name, vir_add_inst, V3D_QPU_A_##name)
967 #define VIR_M_ALU2(name) VIR_ALU2(name, vir_mul_inst, V3D_QPU_M_##name)
968 #define VIR_A_ALU1(name) VIR_ALU1(name, vir_add_inst, V3D_QPU_A_##name)
969 #define VIR_M_ALU1(name) VIR_ALU1(name, vir_mul_inst, V3D_QPU_M_##name)
970 #define VIR_A_ALU0(name) VIR_ALU0(name, vir_add_inst, V3D_QPU_A_##name)
971 #define VIR_M_ALU0(name) VIR_ALU0(name, vir_mul_inst, V3D_QPU_M_##name)
972 #define VIR_A_NODST_2(name) VIR_NODST_2(name, vir_add_inst, V3D_QPU_A_##name)
973 #define VIR_M_NODST_2(name) VIR_NODST_2(name, vir_mul_inst, V3D_QPU_M_##name)
974 #define VIR_A_NODST_1(name) VIR_NODST_1(name, vir_add_inst, V3D_QPU_A_##name)
975 #define VIR_M_NODST_1(name) VIR_NODST_1(name, vir_mul_inst, V3D_QPU_M_##name)
976 #define VIR_A_NODST_0(name) VIR_NODST_0(name, vir_add_inst, V3D_QPU_A_##name)
977
978 VIR_A_ALU2(FADD)
979 VIR_A_ALU2(VFPACK)
980 VIR_A_ALU2(FSUB)
981 VIR_A_ALU2(FMIN)
982 VIR_A_ALU2(FMAX)
983
984 VIR_A_ALU2(ADD)
985 VIR_A_ALU2(SUB)
986 VIR_A_ALU2(SHL)
987 VIR_A_ALU2(SHR)
988 VIR_A_ALU2(ASR)
989 VIR_A_ALU2(ROR)
990 VIR_A_ALU2(MIN)
991 VIR_A_ALU2(MAX)
992 VIR_A_ALU2(UMIN)
993 VIR_A_ALU2(UMAX)
994 VIR_A_ALU2(AND)
995 VIR_A_ALU2(OR)
996 VIR_A_ALU2(XOR)
997 VIR_A_ALU2(VADD)
998 VIR_A_ALU2(VSUB)
999 VIR_A_NODST_2(STVPMV)
1000 VIR_A_ALU1(NOT)
1001 VIR_A_ALU1(NEG)
1002 VIR_A_ALU1(FLAPUSH)
1003 VIR_A_ALU1(FLBPUSH)
1004 VIR_A_ALU1(FLPOP)
1005 VIR_A_ALU1(SETMSF)
1006 VIR_A_ALU1(SETREVF)
1007 VIR_A_ALU0(TIDX)
1008 VIR_A_ALU0(EIDX)
1009 VIR_A_ALU1(LDVPMV_IN)
1010 VIR_A_ALU1(LDVPMV_OUT)
1011 VIR_A_ALU0(TMUWT)
1012
1013 VIR_A_ALU0(FXCD)
1014 VIR_A_ALU0(XCD)
1015 VIR_A_ALU0(FYCD)
1016 VIR_A_ALU0(YCD)
1017 VIR_A_ALU0(MSF)
1018 VIR_A_ALU0(REVF)
1019 VIR_A_ALU0(BARRIERID)
1020 VIR_A_NODST_1(VPMSETUP)
1021 VIR_A_NODST_0(VPMWT)
1022 VIR_A_ALU2(FCMP)
1023 VIR_A_ALU2(VFMAX)
1024
1025 VIR_A_ALU1(FROUND)
1026 VIR_A_ALU1(FTOIN)
1027 VIR_A_ALU1(FTRUNC)
1028 VIR_A_ALU1(FTOIZ)
1029 VIR_A_ALU1(FFLOOR)
1030 VIR_A_ALU1(FTOUZ)
1031 VIR_A_ALU1(FCEIL)
1032 VIR_A_ALU1(FTOC)
1033
1034 VIR_A_ALU1(FDX)
1035 VIR_A_ALU1(FDY)
1036
1037 VIR_A_ALU1(ITOF)
1038 VIR_A_ALU1(CLZ)
1039 VIR_A_ALU1(UTOF)
1040
1041 VIR_M_ALU2(UMUL24)
1042 VIR_M_ALU2(FMUL)
1043 VIR_M_ALU2(SMUL24)
1044 VIR_M_NODST_2(MULTOP)
1045
1046 VIR_M_ALU1(MOV)
1047 VIR_M_ALU1(FMOV)
1048
1049 VIR_SFU(RECIP)
1050 VIR_SFU(RSQRT)
1051 VIR_SFU(EXP)
1052 VIR_SFU(LOG)
1053 VIR_SFU(SIN)
1054 VIR_SFU(RSQRT2)
1055
1056 static inline struct qinst *
1057 vir_MOV_cond(struct v3d_compile *c, enum v3d_qpu_cond cond,
1058 struct qreg dest, struct qreg src)
1059 {
1060 struct qinst *mov = vir_MOV_dest(c, dest, src);
1061 vir_set_cond(mov, cond);
1062 return mov;
1063 }
1064
1065 static inline struct qreg
1066 vir_SEL(struct v3d_compile *c, enum v3d_qpu_cond cond,
1067 struct qreg src0, struct qreg src1)
1068 {
1069 struct qreg t = vir_get_temp(c);
1070 vir_MOV_dest(c, t, src1);
1071 vir_MOV_cond(c, cond, t, src0);
1072 return t;
1073 }
1074
1075 static inline struct qinst *
1076 vir_NOP(struct v3d_compile *c)
1077 {
1078 return vir_emit_nondef(c, vir_add_inst(V3D_QPU_A_NOP,
1079 c->undef, c->undef, c->undef));
1080 }
1081
1082 static inline struct qreg
1083 vir_LDTMU(struct v3d_compile *c)
1084 {
1085 if (c->devinfo->ver >= 41) {
1086 struct qinst *ldtmu = vir_add_inst(V3D_QPU_A_NOP, c->undef,
1087 c->undef, c->undef);
1088 ldtmu->qpu.sig.ldtmu = true;
1089
1090 return vir_emit_def(c, ldtmu);
1091 } else {
1092 vir_NOP(c)->qpu.sig.ldtmu = true;
1093 return vir_MOV(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_R4));
1094 }
1095 }
1096
1097 static inline struct qreg
1098 vir_UMUL(struct v3d_compile *c, struct qreg src0, struct qreg src1)
1099 {
1100 vir_MULTOP(c, src0, src1);
1101 return vir_UMUL24(c, src0, src1);
1102 }
1103
1104 /*
1105 static inline struct qreg
1106 vir_LOAD_IMM(struct v3d_compile *c, uint32_t val)
1107 {
1108 return vir_emit_def(c, vir_inst(QOP_LOAD_IMM, c->undef,
1109 vir_reg(QFILE_LOAD_IMM, val), c->undef));
1110 }
1111
1112 static inline struct qreg
1113 vir_LOAD_IMM_U2(struct v3d_compile *c, uint32_t val)
1114 {
1115 return vir_emit_def(c, vir_inst(QOP_LOAD_IMM_U2, c->undef,
1116 vir_reg(QFILE_LOAD_IMM, val),
1117 c->undef));
1118 }
1119 static inline struct qreg
1120 vir_LOAD_IMM_I2(struct v3d_compile *c, uint32_t val)
1121 {
1122 return vir_emit_def(c, vir_inst(QOP_LOAD_IMM_I2, c->undef,
1123 vir_reg(QFILE_LOAD_IMM, val),
1124 c->undef));
1125 }
1126 */
1127
1128 static inline struct qinst *
1129 vir_BRANCH(struct v3d_compile *c, enum v3d_qpu_branch_cond cond)
1130 {
1131 /* The actual uniform_data value will be set at scheduling time */
1132 return vir_emit_nondef(c, vir_branch_inst(c, cond));
1133 }
1134
1135 #define vir_for_each_block(block, c) \
1136 list_for_each_entry(struct qblock, block, &c->blocks, link)
1137
1138 #define vir_for_each_block_rev(block, c) \
1139 list_for_each_entry_rev(struct qblock, block, &c->blocks, link)
1140
1141 /* Loop over the non-NULL members of the successors array. */
1142 #define vir_for_each_successor(succ, block) \
1143 for (struct qblock *succ = block->successors[0]; \
1144 succ != NULL; \
1145 succ = (succ == block->successors[1] ? NULL : \
1146 block->successors[1]))
1147
1148 #define vir_for_each_inst(inst, block) \
1149 list_for_each_entry(struct qinst, inst, &block->instructions, link)
1150
1151 #define vir_for_each_inst_rev(inst, block) \
1152 list_for_each_entry_rev(struct qinst, inst, &block->instructions, link)
1153
1154 #define vir_for_each_inst_safe(inst, block) \
1155 list_for_each_entry_safe(struct qinst, inst, &block->instructions, link)
1156
1157 #define vir_for_each_inst_inorder(inst, c) \
1158 vir_for_each_block(_block, c) \
1159 vir_for_each_inst(inst, _block)
1160
1161 #define vir_for_each_inst_inorder_safe(inst, c) \
1162 vir_for_each_block(_block, c) \
1163 vir_for_each_inst_safe(inst, _block)
1164
1165 #endif /* V3D_COMPILER_H */