v3d: Kill off vir_PF(), which is hard to use right.
[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 cond_is_exec_mask;
138 bool has_implicit_uniform;
139 bool is_last_thrsw;
140
141 /* After vir_to_qpu.c: If instr reads a uniform, which uniform from
142 * the uncompiled stream it is.
143 */
144 int uniform;
145 };
146
147 enum quniform_contents {
148 /**
149 * Indicates that a constant 32-bit value is copied from the program's
150 * uniform contents.
151 */
152 QUNIFORM_CONSTANT,
153 /**
154 * Indicates that the program's uniform contents are used as an index
155 * into the GL uniform storage.
156 */
157 QUNIFORM_UNIFORM,
158
159 /** @{
160 * Scaling factors from clip coordinates to relative to the viewport
161 * center.
162 *
163 * This is used by the coordinate and vertex shaders to produce the
164 * 32-bit entry consisting of 2 16-bit fields with 12.4 signed fixed
165 * point offsets from the viewport ccenter.
166 */
167 QUNIFORM_VIEWPORT_X_SCALE,
168 QUNIFORM_VIEWPORT_Y_SCALE,
169 /** @} */
170
171 QUNIFORM_VIEWPORT_Z_OFFSET,
172 QUNIFORM_VIEWPORT_Z_SCALE,
173
174 QUNIFORM_USER_CLIP_PLANE,
175
176 /**
177 * A reference to a V3D 3.x texture config parameter 0 uniform.
178 *
179 * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
180 * defines texture type, miplevels, and such. It will be found as a
181 * parameter to the first QOP_TEX_[STRB] instruction in a sequence.
182 */
183 QUNIFORM_TEXTURE_CONFIG_P0_0,
184 QUNIFORM_TEXTURE_CONFIG_P0_1,
185 QUNIFORM_TEXTURE_CONFIG_P0_2,
186 QUNIFORM_TEXTURE_CONFIG_P0_3,
187 QUNIFORM_TEXTURE_CONFIG_P0_4,
188 QUNIFORM_TEXTURE_CONFIG_P0_5,
189 QUNIFORM_TEXTURE_CONFIG_P0_6,
190 QUNIFORM_TEXTURE_CONFIG_P0_7,
191 QUNIFORM_TEXTURE_CONFIG_P0_8,
192 QUNIFORM_TEXTURE_CONFIG_P0_9,
193 QUNIFORM_TEXTURE_CONFIG_P0_10,
194 QUNIFORM_TEXTURE_CONFIG_P0_11,
195 QUNIFORM_TEXTURE_CONFIG_P0_12,
196 QUNIFORM_TEXTURE_CONFIG_P0_13,
197 QUNIFORM_TEXTURE_CONFIG_P0_14,
198 QUNIFORM_TEXTURE_CONFIG_P0_15,
199 QUNIFORM_TEXTURE_CONFIG_P0_16,
200 QUNIFORM_TEXTURE_CONFIG_P0_17,
201 QUNIFORM_TEXTURE_CONFIG_P0_18,
202 QUNIFORM_TEXTURE_CONFIG_P0_19,
203 QUNIFORM_TEXTURE_CONFIG_P0_20,
204 QUNIFORM_TEXTURE_CONFIG_P0_21,
205 QUNIFORM_TEXTURE_CONFIG_P0_22,
206 QUNIFORM_TEXTURE_CONFIG_P0_23,
207 QUNIFORM_TEXTURE_CONFIG_P0_24,
208 QUNIFORM_TEXTURE_CONFIG_P0_25,
209 QUNIFORM_TEXTURE_CONFIG_P0_26,
210 QUNIFORM_TEXTURE_CONFIG_P0_27,
211 QUNIFORM_TEXTURE_CONFIG_P0_28,
212 QUNIFORM_TEXTURE_CONFIG_P0_29,
213 QUNIFORM_TEXTURE_CONFIG_P0_30,
214 QUNIFORM_TEXTURE_CONFIG_P0_31,
215 QUNIFORM_TEXTURE_CONFIG_P0_32,
216
217 /**
218 * A reference to a V3D 3.x texture config parameter 1 uniform.
219 *
220 * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
221 * has the pointer to the indirect texture state. Our data[] field
222 * will have a packed p1 value, but the address field will be just
223 * which texture unit's texture should be referenced.
224 */
225 QUNIFORM_TEXTURE_CONFIG_P1,
226
227 /* A V3D 4.x texture config parameter. The high 8 bits will be
228 * which texture or sampler is being sampled, and the driver must
229 * replace the address field with the appropriate address.
230 */
231 QUNIFORM_TMU_CONFIG_P0,
232 QUNIFORM_TMU_CONFIG_P1,
233
234 QUNIFORM_IMAGE_TMU_CONFIG_P0,
235
236 QUNIFORM_TEXTURE_FIRST_LEVEL,
237
238 QUNIFORM_TEXTURE_WIDTH,
239 QUNIFORM_TEXTURE_HEIGHT,
240 QUNIFORM_TEXTURE_DEPTH,
241 QUNIFORM_TEXTURE_ARRAY_SIZE,
242 QUNIFORM_TEXTURE_LEVELS,
243
244 QUNIFORM_UBO_ADDR,
245
246 QUNIFORM_TEXRECT_SCALE_X,
247 QUNIFORM_TEXRECT_SCALE_Y,
248
249 /* Returns the base offset of the SSBO given by the data value. */
250 QUNIFORM_SSBO_OFFSET,
251
252 /* Returns the size of the SSBO given by the data value. */
253 QUNIFORM_GET_BUFFER_SIZE,
254
255 /* Sizes (in pixels) of a shader image given by the data value. */
256 QUNIFORM_IMAGE_WIDTH,
257 QUNIFORM_IMAGE_HEIGHT,
258 QUNIFORM_IMAGE_DEPTH,
259 QUNIFORM_IMAGE_ARRAY_SIZE,
260
261 QUNIFORM_ALPHA_REF,
262
263 /* Number of workgroups passed to glDispatchCompute in the dimension
264 * selected by the data value.
265 */
266 QUNIFORM_NUM_WORK_GROUPS,
267
268 /**
269 * Returns the the offset of the scratch buffer for register spilling.
270 */
271 QUNIFORM_SPILL_OFFSET,
272 QUNIFORM_SPILL_SIZE_PER_THREAD,
273
274 /**
275 * Returns the offset of the shared memory for compute shaders.
276 *
277 * This will be accessed using TMU general memory operations, so the
278 * L2T cache will effectively be the shared memory area.
279 */
280 QUNIFORM_SHARED_OFFSET,
281 };
282
283 static inline uint32_t v3d_tmu_config_data_create(uint32_t unit, uint32_t value)
284 {
285 return unit << 24 | value;
286 }
287
288 static inline uint32_t v3d_tmu_config_data_get_unit(uint32_t data)
289 {
290 return data >> 24;
291 }
292
293 static inline uint32_t v3d_tmu_config_data_get_value(uint32_t data)
294 {
295 return data & 0xffffff;
296 }
297
298 struct v3d_varying_slot {
299 uint8_t slot_and_component;
300 };
301
302 static inline struct v3d_varying_slot
303 v3d_slot_from_slot_and_component(uint8_t slot, uint8_t component)
304 {
305 assert(slot < 255 / 4);
306 return (struct v3d_varying_slot){ (slot << 2) + component };
307 }
308
309 static inline uint8_t v3d_slot_get_slot(struct v3d_varying_slot slot)
310 {
311 return slot.slot_and_component >> 2;
312 }
313
314 static inline uint8_t v3d_slot_get_component(struct v3d_varying_slot slot)
315 {
316 return slot.slot_and_component & 3;
317 }
318
319 struct v3d_ubo_range {
320 /**
321 * offset in bytes from the start of the ubo where this range is
322 * uploaded.
323 *
324 * Only set once used is set.
325 */
326 uint32_t dst_offset;
327
328 /**
329 * offset in bytes from the start of the gallium uniforms where the
330 * data comes from.
331 */
332 uint32_t src_offset;
333
334 /** size in bytes of this ubo range */
335 uint32_t size;
336 };
337
338 struct v3d_key {
339 void *shader_state;
340 struct {
341 uint8_t swizzle[4];
342 uint8_t return_size;
343 uint8_t return_channels;
344 bool clamp_s:1;
345 bool clamp_t:1;
346 bool clamp_r:1;
347 } tex[V3D_MAX_TEXTURE_SAMPLERS];
348 uint8_t ucp_enables;
349 };
350
351 struct v3d_fs_key {
352 struct v3d_key base;
353 bool depth_enabled;
354 bool is_points;
355 bool is_lines;
356 bool alpha_test;
357 bool point_coord_upper_left;
358 bool light_twoside;
359 bool msaa;
360 bool sample_coverage;
361 bool sample_alpha_to_coverage;
362 bool sample_alpha_to_one;
363 bool clamp_color;
364 bool shade_model_flat;
365 /* Mask of which color render targets are present. */
366 uint8_t cbufs;
367 uint8_t swap_color_rb;
368 /* Mask of which render targets need to be written as 32-bit floats */
369 uint8_t f32_color_rb;
370 /* Masks of which render targets need to be written as ints/uints.
371 * Used by gallium to work around lost information in TGSI.
372 */
373 uint8_t int_color_rb;
374 uint8_t uint_color_rb;
375 uint8_t alpha_test_func;
376 uint8_t logicop_func;
377 uint32_t point_sprite_mask;
378
379 struct pipe_rt_blend_state blend;
380 };
381
382 struct v3d_vs_key {
383 struct v3d_key base;
384
385 struct v3d_varying_slot fs_inputs[V3D_MAX_FS_INPUTS];
386 uint8_t num_fs_inputs;
387
388 bool is_coord;
389 bool per_vertex_point_size;
390 bool clamp_color;
391 };
392
393 /** A basic block of VIR intructions. */
394 struct qblock {
395 struct list_head link;
396
397 struct list_head instructions;
398
399 struct set *predecessors;
400 struct qblock *successors[2];
401
402 int index;
403
404 /* Instruction IPs for the first and last instruction of the block.
405 * Set by qpu_schedule.c.
406 */
407 uint32_t start_qpu_ip;
408 uint32_t end_qpu_ip;
409
410 /* Instruction IP for the branch instruction of the block. Set by
411 * qpu_schedule.c.
412 */
413 uint32_t branch_qpu_ip;
414
415 /** Offset within the uniform stream at the start of the block. */
416 uint32_t start_uniform;
417 /** Offset within the uniform stream of the branch instruction */
418 uint32_t branch_uniform;
419
420 /** @{ used by v3d_vir_live_variables.c */
421 BITSET_WORD *def;
422 BITSET_WORD *use;
423 BITSET_WORD *live_in;
424 BITSET_WORD *live_out;
425 int start_ip, end_ip;
426 /** @} */
427 };
428
429 /** Which util/list.h add mode we should use when inserting an instruction. */
430 enum vir_cursor_mode {
431 vir_cursor_add,
432 vir_cursor_addtail,
433 };
434
435 /**
436 * Tracking structure for where new instructions should be inserted. Create
437 * with one of the vir_after_inst()-style helper functions.
438 *
439 * This does not protect against removal of the block or instruction, so we
440 * have an assert in instruction removal to try to catch it.
441 */
442 struct vir_cursor {
443 enum vir_cursor_mode mode;
444 struct list_head *link;
445 };
446
447 static inline struct vir_cursor
448 vir_before_inst(struct qinst *inst)
449 {
450 return (struct vir_cursor){ vir_cursor_addtail, &inst->link };
451 }
452
453 static inline struct vir_cursor
454 vir_after_inst(struct qinst *inst)
455 {
456 return (struct vir_cursor){ vir_cursor_add, &inst->link };
457 }
458
459 static inline struct vir_cursor
460 vir_before_block(struct qblock *block)
461 {
462 return (struct vir_cursor){ vir_cursor_add, &block->instructions };
463 }
464
465 static inline struct vir_cursor
466 vir_after_block(struct qblock *block)
467 {
468 return (struct vir_cursor){ vir_cursor_addtail, &block->instructions };
469 }
470
471 /**
472 * Compiler state saved across compiler invocations, for any expensive global
473 * setup.
474 */
475 struct v3d_compiler {
476 const struct v3d_device_info *devinfo;
477 struct ra_regs *regs;
478 unsigned int reg_class_phys[3];
479 unsigned int reg_class_phys_or_acc[3];
480 };
481
482 struct v3d_compile {
483 const struct v3d_device_info *devinfo;
484 nir_shader *s;
485 nir_function_impl *impl;
486 struct exec_list *cf_node_list;
487 const struct v3d_compiler *compiler;
488
489 void (*debug_output)(const char *msg,
490 void *debug_output_data);
491 void *debug_output_data;
492
493 /**
494 * Mapping from nir_register * or nir_ssa_def * to array of struct
495 * qreg for the values.
496 */
497 struct hash_table *def_ht;
498
499 /* For each temp, the instruction generating its value. */
500 struct qinst **defs;
501 uint32_t defs_array_size;
502
503 /**
504 * Inputs to the shader, arranged by TGSI declaration order.
505 *
506 * Not all fragment shader QFILE_VARY reads are present in this array.
507 */
508 struct qreg *inputs;
509 struct qreg *outputs;
510 bool msaa_per_sample_output;
511 struct qreg color_reads[V3D_MAX_SAMPLES];
512 struct qreg sample_colors[V3D_MAX_SAMPLES];
513 uint32_t inputs_array_size;
514 uint32_t outputs_array_size;
515 uint32_t uniforms_array_size;
516
517 /* Booleans for whether the corresponding QFILE_VARY[i] is
518 * flat-shaded. This includes gl_FragColor flat-shading, which is
519 * customized based on the shademodel_flat shader key.
520 */
521 uint32_t flat_shade_flags[BITSET_WORDS(V3D_MAX_FS_INPUTS)];
522
523 uint32_t noperspective_flags[BITSET_WORDS(V3D_MAX_FS_INPUTS)];
524
525 uint32_t centroid_flags[BITSET_WORDS(V3D_MAX_FS_INPUTS)];
526
527 bool uses_center_w;
528 bool writes_z;
529
530 struct v3d_ubo_range *ubo_ranges;
531 bool *ubo_range_used;
532 uint32_t ubo_ranges_array_size;
533 /** Number of uniform areas tracked in ubo_ranges. */
534 uint32_t num_ubo_ranges;
535 uint32_t next_ubo_dst_offset;
536
537 /* State for whether we're executing on each channel currently. 0 if
538 * yes, otherwise a block number + 1 that the channel jumped to.
539 */
540 struct qreg execute;
541
542 struct qreg line_x, point_x, point_y;
543
544 /**
545 * Instance ID, which comes in before the vertex attribute payload if
546 * the shader record requests it.
547 */
548 struct qreg iid;
549
550 /**
551 * Vertex ID, which comes in before the vertex attribute payload
552 * (after Instance ID) if the shader record requests it.
553 */
554 struct qreg vid;
555
556 /* Fragment shader payload regs. */
557 struct qreg payload_w, payload_w_centroid, payload_z;
558
559 struct qreg cs_payload[2];
560 struct qreg cs_shared_offset;
561 int local_invocation_index_bits;
562
563 uint8_t vattr_sizes[V3D_MAX_VS_INPUTS / 4];
564 uint32_t num_vpm_writes;
565
566 /* Size in bytes of registers that have been spilled. This is how much
567 * space needs to be available in the spill BO per thread per QPU.
568 */
569 uint32_t spill_size;
570 /* Shader-db stats */
571 uint32_t spills, fills, loops;
572 /**
573 * Register spilling's per-thread base address, shared between each
574 * spill/fill's addressing calculations.
575 */
576 struct qreg spill_base;
577 /* Bit vector of which temps may be spilled */
578 BITSET_WORD *spillable;
579
580 /**
581 * Array of the VARYING_SLOT_* of all FS QFILE_VARY reads.
582 *
583 * This includes those that aren't part of the VPM varyings, like
584 * point/line coordinates.
585 */
586 struct v3d_varying_slot input_slots[V3D_MAX_FS_INPUTS];
587
588 /**
589 * An entry per outputs[] in the VS indicating what the VARYING_SLOT_*
590 * of the output is. Used to emit from the VS in the order that the
591 * FS needs.
592 */
593 struct v3d_varying_slot *output_slots;
594
595 struct pipe_shader_state *shader_state;
596 struct v3d_key *key;
597 struct v3d_fs_key *fs_key;
598 struct v3d_vs_key *vs_key;
599
600 /* Live ranges of temps. */
601 int *temp_start, *temp_end;
602 bool live_intervals_valid;
603
604 uint32_t *uniform_data;
605 enum quniform_contents *uniform_contents;
606 uint32_t uniform_array_size;
607 uint32_t num_uniforms;
608 uint32_t num_outputs;
609 uint32_t output_position_index;
610 nir_variable *output_color_var[4];
611 uint32_t output_point_size_index;
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 num_inputs;
678 uint8_t threads;
679
680 /* For threads > 1, whether the program should be dispatched in the
681 * after-final-THRSW state.
682 */
683 bool single_seg;
684 };
685
686 struct v3d_vs_prog_data {
687 struct v3d_prog_data base;
688
689 bool uses_iid, uses_vid;
690
691 /* Number of components read from each vertex attribute. */
692 uint8_t vattr_sizes[V3D_MAX_VS_INPUTS / 4];
693
694 /* Total number of components read, for the shader state record. */
695 uint32_t vpm_input_size;
696
697 /* Total number of components written, for the shader state record. */
698 uint32_t vpm_output_size;
699
700 /* Set if there should be separate VPM segments for input and output.
701 * If unset, vpm_input_size will be 0.
702 */
703 bool separate_segments;
704
705 /* Value to be programmed in VCM_CACHE_SIZE. */
706 uint8_t vcm_cache_size;
707 };
708
709 struct v3d_fs_prog_data {
710 struct v3d_prog_data base;
711
712 struct v3d_varying_slot input_slots[V3D_MAX_FS_INPUTS];
713
714 /* Array of flat shade flags.
715 *
716 * Each entry is only 24 bits (high 8 bits 0), to match the hardware
717 * packet layout.
718 */
719 uint32_t flat_shade_flags[((V3D_MAX_FS_INPUTS - 1) / 24) + 1];
720
721 uint32_t noperspective_flags[((V3D_MAX_FS_INPUTS - 1) / 24) + 1];
722
723 uint32_t centroid_flags[((V3D_MAX_FS_INPUTS - 1) / 24) + 1];
724
725 bool writes_z;
726 bool disable_ez;
727 bool uses_center_w;
728 };
729
730 /* Special nir_load_input intrinsic index for loading the current TLB
731 * destination color.
732 */
733 #define V3D_NIR_TLB_COLOR_READ_INPUT 2000000000
734
735 #define V3D_NIR_MS_MASK_OUTPUT 2000000000
736
737 extern const nir_shader_compiler_options v3d_nir_options;
738
739 const struct v3d_compiler *v3d_compiler_init(const struct v3d_device_info *devinfo);
740 void v3d_compiler_free(const struct v3d_compiler *compiler);
741 void v3d_optimize_nir(struct nir_shader *s);
742
743 uint64_t *v3d_compile(const struct v3d_compiler *compiler,
744 struct v3d_key *key,
745 struct v3d_prog_data **prog_data,
746 nir_shader *s,
747 void (*debug_output)(const char *msg,
748 void *debug_output_data),
749 void *debug_output_data,
750 int program_id, int variant_id,
751 uint32_t *final_assembly_size);
752
753 void v3d_nir_to_vir(struct v3d_compile *c);
754
755 void vir_compile_destroy(struct v3d_compile *c);
756 const char *vir_get_stage_name(struct v3d_compile *c);
757 struct qblock *vir_new_block(struct v3d_compile *c);
758 void vir_set_emit_block(struct v3d_compile *c, struct qblock *block);
759 void vir_link_blocks(struct qblock *predecessor, struct qblock *successor);
760 struct qblock *vir_entry_block(struct v3d_compile *c);
761 struct qblock *vir_exit_block(struct v3d_compile *c);
762 struct qinst *vir_add_inst(enum v3d_qpu_add_op op, struct qreg dst,
763 struct qreg src0, struct qreg src1);
764 struct qinst *vir_mul_inst(enum v3d_qpu_mul_op op, struct qreg dst,
765 struct qreg src0, struct qreg src1);
766 struct qinst *vir_branch_inst(enum v3d_qpu_branch_cond cond, struct qreg src0);
767 void vir_remove_instruction(struct v3d_compile *c, struct qinst *qinst);
768 struct qreg vir_uniform(struct v3d_compile *c,
769 enum quniform_contents contents,
770 uint32_t data);
771 void vir_schedule_instructions(struct v3d_compile *c);
772 struct v3d_qpu_instr v3d_qpu_nop(void);
773
774 struct qreg vir_emit_def(struct v3d_compile *c, struct qinst *inst);
775 struct qinst *vir_emit_nondef(struct v3d_compile *c, struct qinst *inst);
776 void vir_set_cond(struct qinst *inst, enum v3d_qpu_cond cond);
777 void vir_set_pf(struct qinst *inst, enum v3d_qpu_pf pf);
778 void vir_set_uf(struct qinst *inst, enum v3d_qpu_uf uf);
779 void vir_set_unpack(struct qinst *inst, int src,
780 enum v3d_qpu_input_unpack unpack);
781
782 struct qreg vir_get_temp(struct v3d_compile *c);
783 void vir_emit_last_thrsw(struct v3d_compile *c);
784 void vir_calculate_live_intervals(struct v3d_compile *c);
785 bool vir_has_implicit_uniform(struct qinst *inst);
786 int vir_get_implicit_uniform_src(struct qinst *inst);
787 int vir_get_non_sideband_nsrc(struct qinst *inst);
788 int vir_get_nsrc(struct qinst *inst);
789 bool vir_has_side_effects(struct v3d_compile *c, struct qinst *inst);
790 bool vir_get_add_op(struct qinst *inst, enum v3d_qpu_add_op *op);
791 bool vir_get_mul_op(struct qinst *inst, enum v3d_qpu_mul_op *op);
792 bool vir_is_raw_mov(struct qinst *inst);
793 bool vir_is_tex(struct qinst *inst);
794 bool vir_is_add(struct qinst *inst);
795 bool vir_is_mul(struct qinst *inst);
796 bool vir_writes_r3(const struct v3d_device_info *devinfo, struct qinst *inst);
797 bool vir_writes_r4(const struct v3d_device_info *devinfo, struct qinst *inst);
798 struct qreg vir_follow_movs(struct v3d_compile *c, struct qreg reg);
799 uint8_t vir_channels_written(struct qinst *inst);
800 struct qreg ntq_get_src(struct v3d_compile *c, nir_src src, int i);
801 void ntq_store_dest(struct v3d_compile *c, nir_dest *dest, int chan,
802 struct qreg result);
803 void vir_emit_thrsw(struct v3d_compile *c);
804
805 void vir_dump(struct v3d_compile *c);
806 void vir_dump_inst(struct v3d_compile *c, struct qinst *inst);
807 void vir_dump_uniform(enum quniform_contents contents, uint32_t data);
808
809 void vir_validate(struct v3d_compile *c);
810
811 void vir_optimize(struct v3d_compile *c);
812 bool vir_opt_algebraic(struct v3d_compile *c);
813 bool vir_opt_constant_folding(struct v3d_compile *c);
814 bool vir_opt_copy_propagate(struct v3d_compile *c);
815 bool vir_opt_dead_code(struct v3d_compile *c);
816 bool vir_opt_peephole_sf(struct v3d_compile *c);
817 bool vir_opt_small_immediates(struct v3d_compile *c);
818 bool vir_opt_vpm(struct v3d_compile *c);
819 void v3d_nir_lower_blend(nir_shader *s, struct v3d_compile *c);
820 void v3d_nir_lower_io(nir_shader *s, struct v3d_compile *c);
821 void v3d_nir_lower_txf_ms(nir_shader *s, struct v3d_compile *c);
822 void v3d_nir_lower_image_load_store(nir_shader *s);
823 void vir_lower_uniforms(struct v3d_compile *c);
824
825 void v3d33_vir_vpm_read_setup(struct v3d_compile *c, int num_components);
826 void v3d33_vir_vpm_write_setup(struct v3d_compile *c);
827 void v3d33_vir_emit_tex(struct v3d_compile *c, nir_tex_instr *instr);
828 void v3d40_vir_emit_tex(struct v3d_compile *c, nir_tex_instr *instr);
829 void v3d40_vir_emit_image_load_store(struct v3d_compile *c,
830 nir_intrinsic_instr *instr);
831
832 void v3d_vir_to_qpu(struct v3d_compile *c, struct qpu_reg *temp_registers);
833 uint32_t v3d_qpu_schedule_instructions(struct v3d_compile *c);
834 void qpu_validate(struct v3d_compile *c);
835 struct qpu_reg *v3d_register_allocate(struct v3d_compile *c, bool *spilled);
836 bool vir_init_reg_sets(struct v3d_compiler *compiler);
837
838 bool v3d_gl_format_is_return_32(GLenum format);
839
840 static inline bool
841 quniform_contents_is_texture_p0(enum quniform_contents contents)
842 {
843 return (contents >= QUNIFORM_TEXTURE_CONFIG_P0_0 &&
844 contents < (QUNIFORM_TEXTURE_CONFIG_P0_0 +
845 V3D_MAX_TEXTURE_SAMPLERS));
846 }
847
848 static inline struct qreg
849 vir_uniform_ui(struct v3d_compile *c, uint32_t ui)
850 {
851 return vir_uniform(c, QUNIFORM_CONSTANT, ui);
852 }
853
854 static inline struct qreg
855 vir_uniform_f(struct v3d_compile *c, float f)
856 {
857 return vir_uniform(c, QUNIFORM_CONSTANT, fui(f));
858 }
859
860 #define VIR_ALU0(name, vir_inst, op) \
861 static inline struct qreg \
862 vir_##name(struct v3d_compile *c) \
863 { \
864 return vir_emit_def(c, vir_inst(op, c->undef, \
865 c->undef, c->undef)); \
866 } \
867 static inline struct qinst * \
868 vir_##name##_dest(struct v3d_compile *c, struct qreg dest) \
869 { \
870 return vir_emit_nondef(c, vir_inst(op, dest, \
871 c->undef, c->undef)); \
872 }
873
874 #define VIR_ALU1(name, vir_inst, op) \
875 static inline struct qreg \
876 vir_##name(struct v3d_compile *c, struct qreg a) \
877 { \
878 return vir_emit_def(c, vir_inst(op, c->undef, \
879 a, c->undef)); \
880 } \
881 static inline struct qinst * \
882 vir_##name##_dest(struct v3d_compile *c, struct qreg dest, \
883 struct qreg a) \
884 { \
885 return vir_emit_nondef(c, vir_inst(op, dest, a, \
886 c->undef)); \
887 }
888
889 #define VIR_ALU2(name, vir_inst, op) \
890 static inline struct qreg \
891 vir_##name(struct v3d_compile *c, struct qreg a, struct qreg b) \
892 { \
893 return vir_emit_def(c, vir_inst(op, c->undef, a, b)); \
894 } \
895 static inline struct qinst * \
896 vir_##name##_dest(struct v3d_compile *c, struct qreg dest, \
897 struct qreg a, struct qreg b) \
898 { \
899 return vir_emit_nondef(c, vir_inst(op, dest, a, b)); \
900 }
901
902 #define VIR_NODST_0(name, vir_inst, op) \
903 static inline struct qinst * \
904 vir_##name(struct v3d_compile *c) \
905 { \
906 return vir_emit_nondef(c, vir_inst(op, c->undef, \
907 c->undef, c->undef)); \
908 }
909
910 #define VIR_NODST_1(name, vir_inst, op) \
911 static inline struct qinst * \
912 vir_##name(struct v3d_compile *c, struct qreg a) \
913 { \
914 return vir_emit_nondef(c, vir_inst(op, c->undef, \
915 a, c->undef)); \
916 }
917
918 #define VIR_NODST_2(name, vir_inst, op) \
919 static inline struct qinst * \
920 vir_##name(struct v3d_compile *c, struct qreg a, struct qreg b) \
921 { \
922 return vir_emit_nondef(c, vir_inst(op, c->undef, \
923 a, b)); \
924 }
925
926 #define VIR_SFU(name) \
927 static inline struct qreg \
928 vir_##name(struct v3d_compile *c, struct qreg a) \
929 { \
930 if (c->devinfo->ver >= 41) { \
931 return vir_emit_def(c, vir_add_inst(V3D_QPU_A_##name, \
932 c->undef, \
933 a, c->undef)); \
934 } else { \
935 vir_FMOV_dest(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_##name), a); \
936 return vir_FMOV(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_R4)); \
937 } \
938 } \
939 static inline struct qinst * \
940 vir_##name##_dest(struct v3d_compile *c, struct qreg dest, \
941 struct qreg a) \
942 { \
943 if (c->devinfo->ver >= 41) { \
944 return vir_emit_nondef(c, vir_add_inst(V3D_QPU_A_##name, \
945 dest, \
946 a, c->undef)); \
947 } else { \
948 vir_FMOV_dest(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_##name), a); \
949 return vir_FMOV_dest(c, dest, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_R4)); \
950 } \
951 }
952
953 #define VIR_A_ALU2(name) VIR_ALU2(name, vir_add_inst, V3D_QPU_A_##name)
954 #define VIR_M_ALU2(name) VIR_ALU2(name, vir_mul_inst, V3D_QPU_M_##name)
955 #define VIR_A_ALU1(name) VIR_ALU1(name, vir_add_inst, V3D_QPU_A_##name)
956 #define VIR_M_ALU1(name) VIR_ALU1(name, vir_mul_inst, V3D_QPU_M_##name)
957 #define VIR_A_ALU0(name) VIR_ALU0(name, vir_add_inst, V3D_QPU_A_##name)
958 #define VIR_M_ALU0(name) VIR_ALU0(name, vir_mul_inst, V3D_QPU_M_##name)
959 #define VIR_A_NODST_2(name) VIR_NODST_2(name, vir_add_inst, V3D_QPU_A_##name)
960 #define VIR_M_NODST_2(name) VIR_NODST_2(name, vir_mul_inst, V3D_QPU_M_##name)
961 #define VIR_A_NODST_1(name) VIR_NODST_1(name, vir_add_inst, V3D_QPU_A_##name)
962 #define VIR_M_NODST_1(name) VIR_NODST_1(name, vir_mul_inst, V3D_QPU_M_##name)
963 #define VIR_A_NODST_0(name) VIR_NODST_0(name, vir_add_inst, V3D_QPU_A_##name)
964
965 VIR_A_ALU2(FADD)
966 VIR_A_ALU2(VFPACK)
967 VIR_A_ALU2(FSUB)
968 VIR_A_ALU2(FMIN)
969 VIR_A_ALU2(FMAX)
970
971 VIR_A_ALU2(ADD)
972 VIR_A_ALU2(SUB)
973 VIR_A_ALU2(SHL)
974 VIR_A_ALU2(SHR)
975 VIR_A_ALU2(ASR)
976 VIR_A_ALU2(ROR)
977 VIR_A_ALU2(MIN)
978 VIR_A_ALU2(MAX)
979 VIR_A_ALU2(UMIN)
980 VIR_A_ALU2(UMAX)
981 VIR_A_ALU2(AND)
982 VIR_A_ALU2(OR)
983 VIR_A_ALU2(XOR)
984 VIR_A_ALU2(VADD)
985 VIR_A_ALU2(VSUB)
986 VIR_A_NODST_2(STVPMV)
987 VIR_A_ALU1(NOT)
988 VIR_A_ALU1(NEG)
989 VIR_A_ALU1(FLAPUSH)
990 VIR_A_ALU1(FLBPUSH)
991 VIR_A_ALU1(FLPOP)
992 VIR_A_ALU1(SETMSF)
993 VIR_A_ALU1(SETREVF)
994 VIR_A_ALU0(TIDX)
995 VIR_A_ALU0(EIDX)
996 VIR_A_ALU1(LDVPMV_IN)
997 VIR_A_ALU1(LDVPMV_OUT)
998 VIR_A_ALU0(TMUWT)
999
1000 VIR_A_ALU0(FXCD)
1001 VIR_A_ALU0(XCD)
1002 VIR_A_ALU0(FYCD)
1003 VIR_A_ALU0(YCD)
1004 VIR_A_ALU0(MSF)
1005 VIR_A_ALU0(REVF)
1006 VIR_A_ALU0(BARRIERID)
1007 VIR_A_NODST_1(VPMSETUP)
1008 VIR_A_NODST_0(VPMWT)
1009 VIR_A_ALU2(FCMP)
1010 VIR_A_ALU2(VFMAX)
1011
1012 VIR_A_ALU1(FROUND)
1013 VIR_A_ALU1(FTOIN)
1014 VIR_A_ALU1(FTRUNC)
1015 VIR_A_ALU1(FTOIZ)
1016 VIR_A_ALU1(FFLOOR)
1017 VIR_A_ALU1(FTOUZ)
1018 VIR_A_ALU1(FCEIL)
1019 VIR_A_ALU1(FTOC)
1020
1021 VIR_A_ALU1(FDX)
1022 VIR_A_ALU1(FDY)
1023
1024 VIR_A_ALU1(ITOF)
1025 VIR_A_ALU1(CLZ)
1026 VIR_A_ALU1(UTOF)
1027
1028 VIR_M_ALU2(UMUL24)
1029 VIR_M_ALU2(FMUL)
1030 VIR_M_ALU2(SMUL24)
1031 VIR_M_NODST_2(MULTOP)
1032
1033 VIR_M_ALU1(MOV)
1034 VIR_M_ALU1(FMOV)
1035
1036 VIR_SFU(RECIP)
1037 VIR_SFU(RSQRT)
1038 VIR_SFU(EXP)
1039 VIR_SFU(LOG)
1040 VIR_SFU(SIN)
1041 VIR_SFU(RSQRT2)
1042
1043 static inline struct qinst *
1044 vir_MOV_cond(struct v3d_compile *c, enum v3d_qpu_cond cond,
1045 struct qreg dest, struct qreg src)
1046 {
1047 struct qinst *mov = vir_MOV_dest(c, dest, src);
1048 vir_set_cond(mov, cond);
1049 return mov;
1050 }
1051
1052 static inline struct qreg
1053 vir_SEL(struct v3d_compile *c, enum v3d_qpu_cond cond,
1054 struct qreg src0, struct qreg src1)
1055 {
1056 struct qreg t = vir_get_temp(c);
1057 vir_MOV_dest(c, t, src1);
1058 vir_MOV_cond(c, cond, t, src0);
1059 return t;
1060 }
1061
1062 static inline struct qinst *
1063 vir_NOP(struct v3d_compile *c)
1064 {
1065 return vir_emit_nondef(c, vir_add_inst(V3D_QPU_A_NOP,
1066 c->undef, c->undef, c->undef));
1067 }
1068
1069 static inline struct qreg
1070 vir_LDTMU(struct v3d_compile *c)
1071 {
1072 if (c->devinfo->ver >= 41) {
1073 struct qinst *ldtmu = vir_add_inst(V3D_QPU_A_NOP, c->undef,
1074 c->undef, c->undef);
1075 ldtmu->qpu.sig.ldtmu = true;
1076
1077 return vir_emit_def(c, ldtmu);
1078 } else {
1079 vir_NOP(c)->qpu.sig.ldtmu = true;
1080 return vir_MOV(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_R4));
1081 }
1082 }
1083
1084 static inline struct qreg
1085 vir_UMUL(struct v3d_compile *c, struct qreg src0, struct qreg src1)
1086 {
1087 vir_MULTOP(c, src0, src1);
1088 return vir_UMUL24(c, src0, src1);
1089 }
1090
1091 /*
1092 static inline struct qreg
1093 vir_LOAD_IMM(struct v3d_compile *c, uint32_t val)
1094 {
1095 return vir_emit_def(c, vir_inst(QOP_LOAD_IMM, c->undef,
1096 vir_reg(QFILE_LOAD_IMM, val), c->undef));
1097 }
1098
1099 static inline struct qreg
1100 vir_LOAD_IMM_U2(struct v3d_compile *c, uint32_t val)
1101 {
1102 return vir_emit_def(c, vir_inst(QOP_LOAD_IMM_U2, c->undef,
1103 vir_reg(QFILE_LOAD_IMM, val),
1104 c->undef));
1105 }
1106 static inline struct qreg
1107 vir_LOAD_IMM_I2(struct v3d_compile *c, uint32_t val)
1108 {
1109 return vir_emit_def(c, vir_inst(QOP_LOAD_IMM_I2, c->undef,
1110 vir_reg(QFILE_LOAD_IMM, val),
1111 c->undef));
1112 }
1113 */
1114
1115 static inline struct qinst *
1116 vir_BRANCH(struct v3d_compile *c, enum v3d_qpu_branch_cond cond)
1117 {
1118 /* The actual uniform_data value will be set at scheduling time */
1119 return vir_emit_nondef(c, vir_branch_inst(cond, vir_uniform_ui(c, 0)));
1120 }
1121
1122 #define vir_for_each_block(block, c) \
1123 list_for_each_entry(struct qblock, block, &c->blocks, link)
1124
1125 #define vir_for_each_block_rev(block, c) \
1126 list_for_each_entry_rev(struct qblock, block, &c->blocks, link)
1127
1128 /* Loop over the non-NULL members of the successors array. */
1129 #define vir_for_each_successor(succ, block) \
1130 for (struct qblock *succ = block->successors[0]; \
1131 succ != NULL; \
1132 succ = (succ == block->successors[1] ? NULL : \
1133 block->successors[1]))
1134
1135 #define vir_for_each_inst(inst, block) \
1136 list_for_each_entry(struct qinst, inst, &block->instructions, link)
1137
1138 #define vir_for_each_inst_rev(inst, block) \
1139 list_for_each_entry_rev(struct qinst, inst, &block->instructions, link)
1140
1141 #define vir_for_each_inst_safe(inst, block) \
1142 list_for_each_entry_safe(struct qinst, inst, &block->instructions, link)
1143
1144 #define vir_for_each_inst_inorder(inst, c) \
1145 vir_for_each_block(_block, c) \
1146 vir_for_each_inst(inst, _block)
1147
1148 #endif /* V3D_COMPILER_H */