v3d: Add support for CS workgroup/invocation id intrinsics.
[mesa.git] / src / broadcom / compiler / v3d_compiler.h
1 /*
2 * Copyright © 2016 Broadcom
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 */
23
24 #ifndef V3D_COMPILER_H
25 #define V3D_COMPILER_H
26
27 #include <assert.h>
28 #include <stdio.h>
29 #include <stdlib.h>
30 #include <stdbool.h>
31 #include <stdint.h>
32 #include <string.h>
33
34 #include "util/macros.h"
35 #include "common/v3d_debug.h"
36 #include "common/v3d_device_info.h"
37 #include "compiler/nir/nir.h"
38 #include "util/list.h"
39 #include "util/u_math.h"
40
41 #include "qpu/qpu_instr.h"
42 #include "pipe/p_state.h"
43
44 #define V3D_MAX_TEXTURE_SAMPLERS 32
45 #define V3D_MAX_SAMPLES 4
46 #define V3D_MAX_FS_INPUTS 64
47 #define V3D_MAX_VS_INPUTS 64
48
49 struct nir_builder;
50
51 struct v3d_fs_inputs {
52 /**
53 * Array of the meanings of the VPM inputs this shader needs.
54 *
55 * It doesn't include those that aren't part of the VPM, like
56 * point/line coordinates.
57 */
58 struct v3d_varying_slot *input_slots;
59 uint32_t num_inputs;
60 };
61
62 enum qfile {
63 /** An unused source or destination register. */
64 QFILE_NULL,
65
66 /** A physical register, such as the W coordinate payload. */
67 QFILE_REG,
68 /** One of the regsiters for fixed function interactions. */
69 QFILE_MAGIC,
70
71 /**
72 * A virtual register, that will be allocated to actual accumulator
73 * or physical registers later.
74 */
75 QFILE_TEMP,
76 QFILE_UNIF,
77 QFILE_TLB,
78 QFILE_TLBU,
79
80 /**
81 * VPM reads use this with an index value to say what part of the VPM
82 * is being read.
83 */
84 QFILE_VPM,
85
86 /**
87 * Stores an immediate value in the index field that will be used
88 * directly by qpu_load_imm().
89 */
90 QFILE_LOAD_IMM,
91
92 /**
93 * Stores an immediate value in the index field that can be turned
94 * into a small immediate field by qpu_encode_small_immediate().
95 */
96 QFILE_SMALL_IMM,
97 };
98
99 /**
100 * A reference to a QPU register or a virtual temp register.
101 */
102 struct qreg {
103 enum qfile file;
104 uint32_t index;
105 };
106
107 static inline struct qreg vir_reg(enum qfile file, uint32_t index)
108 {
109 return (struct qreg){file, index};
110 }
111
112 /**
113 * A reference to an actual register at the QPU level, for register
114 * allocation.
115 */
116 struct qpu_reg {
117 bool magic;
118 bool smimm;
119 int index;
120 };
121
122 struct qinst {
123 /** Entry in qblock->instructions */
124 struct list_head link;
125
126 /**
127 * The instruction being wrapped. Its condition codes, pack flags,
128 * signals, etc. will all be used, with just the register references
129 * being replaced by the contents of qinst->dst and qinst->src[].
130 */
131 struct v3d_qpu_instr qpu;
132
133 /* Pre-register-allocation references to src/dst registers */
134 struct qreg dst;
135 struct qreg src[3];
136 bool cond_is_exec_mask;
137 bool has_implicit_uniform;
138 bool is_last_thrsw;
139
140 /* After vir_to_qpu.c: If instr reads a uniform, which uniform from
141 * the uncompiled stream it is.
142 */
143 int uniform;
144 };
145
146 enum quniform_contents {
147 /**
148 * Indicates that a constant 32-bit value is copied from the program's
149 * uniform contents.
150 */
151 QUNIFORM_CONSTANT,
152 /**
153 * Indicates that the program's uniform contents are used as an index
154 * into the GL uniform storage.
155 */
156 QUNIFORM_UNIFORM,
157
158 /** @{
159 * Scaling factors from clip coordinates to relative to the viewport
160 * center.
161 *
162 * This is used by the coordinate and vertex shaders to produce the
163 * 32-bit entry consisting of 2 16-bit fields with 12.4 signed fixed
164 * point offsets from the viewport ccenter.
165 */
166 QUNIFORM_VIEWPORT_X_SCALE,
167 QUNIFORM_VIEWPORT_Y_SCALE,
168 /** @} */
169
170 QUNIFORM_VIEWPORT_Z_OFFSET,
171 QUNIFORM_VIEWPORT_Z_SCALE,
172
173 QUNIFORM_USER_CLIP_PLANE,
174
175 /**
176 * A reference to a V3D 3.x texture config parameter 0 uniform.
177 *
178 * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
179 * defines texture type, miplevels, and such. It will be found as a
180 * parameter to the first QOP_TEX_[STRB] instruction in a sequence.
181 */
182 QUNIFORM_TEXTURE_CONFIG_P0_0,
183 QUNIFORM_TEXTURE_CONFIG_P0_1,
184 QUNIFORM_TEXTURE_CONFIG_P0_2,
185 QUNIFORM_TEXTURE_CONFIG_P0_3,
186 QUNIFORM_TEXTURE_CONFIG_P0_4,
187 QUNIFORM_TEXTURE_CONFIG_P0_5,
188 QUNIFORM_TEXTURE_CONFIG_P0_6,
189 QUNIFORM_TEXTURE_CONFIG_P0_7,
190 QUNIFORM_TEXTURE_CONFIG_P0_8,
191 QUNIFORM_TEXTURE_CONFIG_P0_9,
192 QUNIFORM_TEXTURE_CONFIG_P0_10,
193 QUNIFORM_TEXTURE_CONFIG_P0_11,
194 QUNIFORM_TEXTURE_CONFIG_P0_12,
195 QUNIFORM_TEXTURE_CONFIG_P0_13,
196 QUNIFORM_TEXTURE_CONFIG_P0_14,
197 QUNIFORM_TEXTURE_CONFIG_P0_15,
198 QUNIFORM_TEXTURE_CONFIG_P0_16,
199 QUNIFORM_TEXTURE_CONFIG_P0_17,
200 QUNIFORM_TEXTURE_CONFIG_P0_18,
201 QUNIFORM_TEXTURE_CONFIG_P0_19,
202 QUNIFORM_TEXTURE_CONFIG_P0_20,
203 QUNIFORM_TEXTURE_CONFIG_P0_21,
204 QUNIFORM_TEXTURE_CONFIG_P0_22,
205 QUNIFORM_TEXTURE_CONFIG_P0_23,
206 QUNIFORM_TEXTURE_CONFIG_P0_24,
207 QUNIFORM_TEXTURE_CONFIG_P0_25,
208 QUNIFORM_TEXTURE_CONFIG_P0_26,
209 QUNIFORM_TEXTURE_CONFIG_P0_27,
210 QUNIFORM_TEXTURE_CONFIG_P0_28,
211 QUNIFORM_TEXTURE_CONFIG_P0_29,
212 QUNIFORM_TEXTURE_CONFIG_P0_30,
213 QUNIFORM_TEXTURE_CONFIG_P0_31,
214 QUNIFORM_TEXTURE_CONFIG_P0_32,
215
216 /**
217 * A reference to a V3D 3.x texture config parameter 1 uniform.
218 *
219 * This is a uniform implicitly loaded with a QPU_W_TMU* write, which
220 * has the pointer to the indirect texture state. Our data[] field
221 * will have a packed p1 value, but the address field will be just
222 * which texture unit's texture should be referenced.
223 */
224 QUNIFORM_TEXTURE_CONFIG_P1,
225
226 /* A V3D 4.x texture config parameter. The high 8 bits will be
227 * which texture or sampler is being sampled, and the driver must
228 * replace the address field with the appropriate address.
229 */
230 QUNIFORM_TMU_CONFIG_P0,
231 QUNIFORM_TMU_CONFIG_P1,
232
233 QUNIFORM_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 static inline uint32_t v3d_tmu_config_data_create(uint32_t unit, uint32_t value)
275 {
276 return unit << 24 | value;
277 }
278
279 static inline uint32_t v3d_tmu_config_data_get_unit(uint32_t data)
280 {
281 return data >> 24;
282 }
283
284 static inline uint32_t v3d_tmu_config_data_get_value(uint32_t data)
285 {
286 return data & 0xffffff;
287 }
288
289 struct v3d_varying_slot {
290 uint8_t slot_and_component;
291 };
292
293 static inline struct v3d_varying_slot
294 v3d_slot_from_slot_and_component(uint8_t slot, uint8_t component)
295 {
296 assert(slot < 255 / 4);
297 return (struct v3d_varying_slot){ (slot << 2) + component };
298 }
299
300 static inline uint8_t v3d_slot_get_slot(struct v3d_varying_slot slot)
301 {
302 return slot.slot_and_component >> 2;
303 }
304
305 static inline uint8_t v3d_slot_get_component(struct v3d_varying_slot slot)
306 {
307 return slot.slot_and_component & 3;
308 }
309
310 struct v3d_ubo_range {
311 /**
312 * offset in bytes from the start of the ubo where this range is
313 * uploaded.
314 *
315 * Only set once used is set.
316 */
317 uint32_t dst_offset;
318
319 /**
320 * offset in bytes from the start of the gallium uniforms where the
321 * data comes from.
322 */
323 uint32_t src_offset;
324
325 /** size in bytes of this ubo range */
326 uint32_t size;
327 };
328
329 struct v3d_key {
330 void *shader_state;
331 struct {
332 uint8_t swizzle[4];
333 uint8_t return_size;
334 uint8_t return_channels;
335 bool clamp_s:1;
336 bool clamp_t:1;
337 bool clamp_r:1;
338 } tex[V3D_MAX_TEXTURE_SAMPLERS];
339 uint8_t ucp_enables;
340 };
341
342 struct v3d_fs_key {
343 struct v3d_key base;
344 bool depth_enabled;
345 bool is_points;
346 bool is_lines;
347 bool alpha_test;
348 bool point_coord_upper_left;
349 bool light_twoside;
350 bool msaa;
351 bool sample_coverage;
352 bool sample_alpha_to_coverage;
353 bool sample_alpha_to_one;
354 bool clamp_color;
355 bool shade_model_flat;
356 uint8_t nr_cbufs;
357 uint8_t swap_color_rb;
358 /* Mask of which render targets need to be written as 32-bit floats */
359 uint8_t f32_color_rb;
360 /* Masks of which render targets need to be written as ints/uints.
361 * Used by gallium to work around lost information in TGSI.
362 */
363 uint8_t int_color_rb;
364 uint8_t uint_color_rb;
365 uint8_t alpha_test_func;
366 uint8_t logicop_func;
367 uint32_t point_sprite_mask;
368
369 struct pipe_rt_blend_state blend;
370 };
371
372 struct v3d_vs_key {
373 struct v3d_key base;
374
375 struct v3d_varying_slot fs_inputs[V3D_MAX_FS_INPUTS];
376 uint8_t num_fs_inputs;
377
378 bool is_coord;
379 bool per_vertex_point_size;
380 bool clamp_color;
381 };
382
383 /** A basic block of VIR intructions. */
384 struct qblock {
385 struct list_head link;
386
387 struct list_head instructions;
388
389 struct set *predecessors;
390 struct qblock *successors[2];
391
392 int index;
393
394 /* Instruction IPs for the first and last instruction of the block.
395 * Set by qpu_schedule.c.
396 */
397 uint32_t start_qpu_ip;
398 uint32_t end_qpu_ip;
399
400 /* Instruction IP for the branch instruction of the block. Set by
401 * qpu_schedule.c.
402 */
403 uint32_t branch_qpu_ip;
404
405 /** Offset within the uniform stream at the start of the block. */
406 uint32_t start_uniform;
407 /** Offset within the uniform stream of the branch instruction */
408 uint32_t branch_uniform;
409
410 /** @{ used by v3d_vir_live_variables.c */
411 BITSET_WORD *def;
412 BITSET_WORD *use;
413 BITSET_WORD *live_in;
414 BITSET_WORD *live_out;
415 int start_ip, end_ip;
416 /** @} */
417 };
418
419 /** Which util/list.h add mode we should use when inserting an instruction. */
420 enum vir_cursor_mode {
421 vir_cursor_add,
422 vir_cursor_addtail,
423 };
424
425 /**
426 * Tracking structure for where new instructions should be inserted. Create
427 * with one of the vir_after_inst()-style helper functions.
428 *
429 * This does not protect against removal of the block or instruction, so we
430 * have an assert in instruction removal to try to catch it.
431 */
432 struct vir_cursor {
433 enum vir_cursor_mode mode;
434 struct list_head *link;
435 };
436
437 static inline struct vir_cursor
438 vir_before_inst(struct qinst *inst)
439 {
440 return (struct vir_cursor){ vir_cursor_addtail, &inst->link };
441 }
442
443 static inline struct vir_cursor
444 vir_after_inst(struct qinst *inst)
445 {
446 return (struct vir_cursor){ vir_cursor_add, &inst->link };
447 }
448
449 static inline struct vir_cursor
450 vir_before_block(struct qblock *block)
451 {
452 return (struct vir_cursor){ vir_cursor_add, &block->instructions };
453 }
454
455 static inline struct vir_cursor
456 vir_after_block(struct qblock *block)
457 {
458 return (struct vir_cursor){ vir_cursor_addtail, &block->instructions };
459 }
460
461 /**
462 * Compiler state saved across compiler invocations, for any expensive global
463 * setup.
464 */
465 struct v3d_compiler {
466 const struct v3d_device_info *devinfo;
467 struct ra_regs *regs;
468 unsigned int reg_class_phys[3];
469 unsigned int reg_class_phys_or_acc[3];
470 };
471
472 struct v3d_compile {
473 const struct v3d_device_info *devinfo;
474 nir_shader *s;
475 nir_function_impl *impl;
476 struct exec_list *cf_node_list;
477 const struct v3d_compiler *compiler;
478
479 void (*debug_output)(const char *msg,
480 void *debug_output_data);
481 void *debug_output_data;
482
483 /**
484 * Mapping from nir_register * or nir_ssa_def * to array of struct
485 * qreg for the values.
486 */
487 struct hash_table *def_ht;
488
489 /* For each temp, the instruction generating its value. */
490 struct qinst **defs;
491 uint32_t defs_array_size;
492
493 /**
494 * Inputs to the shader, arranged by TGSI declaration order.
495 *
496 * Not all fragment shader QFILE_VARY reads are present in this array.
497 */
498 struct qreg *inputs;
499 struct qreg *outputs;
500 bool msaa_per_sample_output;
501 struct qreg color_reads[V3D_MAX_SAMPLES];
502 struct qreg sample_colors[V3D_MAX_SAMPLES];
503 uint32_t inputs_array_size;
504 uint32_t outputs_array_size;
505 uint32_t uniforms_array_size;
506
507 /* Booleans for whether the corresponding QFILE_VARY[i] is
508 * flat-shaded. This includes gl_FragColor flat-shading, which is
509 * customized based on the shademodel_flat shader key.
510 */
511 uint32_t flat_shade_flags[BITSET_WORDS(V3D_MAX_FS_INPUTS)];
512
513 uint32_t noperspective_flags[BITSET_WORDS(V3D_MAX_FS_INPUTS)];
514
515 uint32_t centroid_flags[BITSET_WORDS(V3D_MAX_FS_INPUTS)];
516
517 bool uses_center_w;
518
519 struct v3d_ubo_range *ubo_ranges;
520 bool *ubo_range_used;
521 uint32_t ubo_ranges_array_size;
522 /** Number of uniform areas tracked in ubo_ranges. */
523 uint32_t num_ubo_ranges;
524 uint32_t next_ubo_dst_offset;
525
526 /* State for whether we're executing on each channel currently. 0 if
527 * yes, otherwise a block number + 1 that the channel jumped to.
528 */
529 struct qreg execute;
530
531 struct qreg line_x, point_x, point_y;
532
533 /**
534 * Instance ID, which comes in before the vertex attribute payload if
535 * the shader record requests it.
536 */
537 struct qreg iid;
538
539 /**
540 * Vertex ID, which comes in before the vertex attribute payload
541 * (after Instance ID) if the shader record requests it.
542 */
543 struct qreg vid;
544
545 /* Fragment shader payload regs. */
546 struct qreg payload_w, payload_w_centroid, payload_z;
547
548 struct qreg cs_payload[2];
549 int local_invocation_index_bits;
550
551 uint8_t vattr_sizes[V3D_MAX_VS_INPUTS];
552 uint32_t num_vpm_writes;
553
554 /* Size in bytes of registers that have been spilled. This is how much
555 * space needs to be available in the spill BO per thread per QPU.
556 */
557 uint32_t spill_size;
558 /* Shader-db stats */
559 uint32_t spills, fills, loops;
560 /**
561 * Register spilling's per-thread base address, shared between each
562 * spill/fill's addressing calculations.
563 */
564 struct qreg spill_base;
565 /* Bit vector of which temps may be spilled */
566 BITSET_WORD *spillable;
567
568 /**
569 * Array of the VARYING_SLOT_* of all FS QFILE_VARY reads.
570 *
571 * This includes those that aren't part of the VPM varyings, like
572 * point/line coordinates.
573 */
574 struct v3d_varying_slot input_slots[V3D_MAX_FS_INPUTS];
575
576 /**
577 * An entry per outputs[] in the VS indicating what the VARYING_SLOT_*
578 * of the output is. Used to emit from the VS in the order that the
579 * FS needs.
580 */
581 struct v3d_varying_slot *output_slots;
582
583 struct pipe_shader_state *shader_state;
584 struct v3d_key *key;
585 struct v3d_fs_key *fs_key;
586 struct v3d_vs_key *vs_key;
587
588 /* Live ranges of temps. */
589 int *temp_start, *temp_end;
590 bool live_intervals_valid;
591
592 uint32_t *uniform_data;
593 enum quniform_contents *uniform_contents;
594 uint32_t uniform_array_size;
595 uint32_t num_uniforms;
596 uint32_t num_outputs;
597 uint32_t output_position_index;
598 nir_variable *output_color_var[4];
599 uint32_t output_point_size_index;
600 uint32_t output_sample_mask_index;
601
602 struct qreg undef;
603 uint32_t num_temps;
604
605 struct vir_cursor cursor;
606 struct list_head blocks;
607 int next_block_index;
608 struct qblock *cur_block;
609 struct qblock *loop_cont_block;
610 struct qblock *loop_break_block;
611
612 uint64_t *qpu_insts;
613 uint32_t qpu_inst_count;
614 uint32_t qpu_inst_size;
615
616 /* For the FS, the number of varying inputs not counting the
617 * point/line varyings payload
618 */
619 uint32_t num_inputs;
620
621 /**
622 * Number of inputs from num_inputs remaining to be queued to the read
623 * FIFO in the VS/CS.
624 */
625 uint32_t num_inputs_remaining;
626
627 /* Number of inputs currently in the read FIFO for the VS/CS */
628 uint32_t num_inputs_in_fifo;
629
630 /** Next offset in the VPM to read from in the VS/CS */
631 uint32_t vpm_read_offset;
632
633 uint32_t program_id;
634 uint32_t variant_id;
635
636 /* Set to compile program in in 1x, 2x, or 4x threaded mode, where
637 * SIG_THREAD_SWITCH is used to hide texturing latency at the cost of
638 * limiting ourselves to the part of the physical reg space.
639 *
640 * On V3D 3.x, 2x or 4x divide the physical reg space by 2x or 4x. On
641 * V3D 4.x, all shaders are 2x threaded, and 4x only divides the
642 * physical reg space in half.
643 */
644 uint8_t threads;
645 struct qinst *last_thrsw;
646 bool last_thrsw_at_top_level;
647
648 bool failed;
649 };
650
651 struct v3d_uniform_list {
652 enum quniform_contents *contents;
653 uint32_t *data;
654 uint32_t count;
655 };
656
657 struct v3d_prog_data {
658 struct v3d_uniform_list uniforms;
659
660 struct v3d_ubo_range *ubo_ranges;
661 uint32_t num_ubo_ranges;
662 uint32_t ubo_size;
663 uint32_t spill_size;
664
665 uint8_t num_inputs;
666 uint8_t threads;
667
668 /* For threads > 1, whether the program should be dispatched in the
669 * after-final-THRSW state.
670 */
671 bool single_seg;
672 };
673
674 struct v3d_vs_prog_data {
675 struct v3d_prog_data base;
676
677 bool uses_iid, uses_vid;
678
679 /* Number of components read from each vertex attribute. */
680 uint8_t vattr_sizes[32];
681
682 /* Total number of components read, for the shader state record. */
683 uint32_t vpm_input_size;
684
685 /* Total number of components written, for the shader state record. */
686 uint32_t vpm_output_size;
687
688 /* Set if there should be separate VPM segments for input and output.
689 * If unset, vpm_input_size will be 0.
690 */
691 bool separate_segments;
692
693 /* Value to be programmed in VCM_CACHE_SIZE. */
694 uint8_t vcm_cache_size;
695 };
696
697 struct v3d_fs_prog_data {
698 struct v3d_prog_data base;
699
700 struct v3d_varying_slot input_slots[V3D_MAX_FS_INPUTS];
701
702 /* Array of flat shade flags.
703 *
704 * Each entry is only 24 bits (high 8 bits 0), to match the hardware
705 * packet layout.
706 */
707 uint32_t flat_shade_flags[((V3D_MAX_FS_INPUTS - 1) / 24) + 1];
708
709 uint32_t noperspective_flags[((V3D_MAX_FS_INPUTS - 1) / 24) + 1];
710
711 uint32_t centroid_flags[((V3D_MAX_FS_INPUTS - 1) / 24) + 1];
712
713 bool writes_z;
714 bool discard;
715 bool uses_center_w;
716 };
717
718 /* Special nir_load_input intrinsic index for loading the current TLB
719 * destination color.
720 */
721 #define V3D_NIR_TLB_COLOR_READ_INPUT 2000000000
722
723 #define V3D_NIR_MS_MASK_OUTPUT 2000000000
724
725 extern const nir_shader_compiler_options v3d_nir_options;
726
727 const struct v3d_compiler *v3d_compiler_init(const struct v3d_device_info *devinfo);
728 void v3d_compiler_free(const struct v3d_compiler *compiler);
729 void v3d_optimize_nir(struct nir_shader *s);
730
731 uint64_t *v3d_compile(const struct v3d_compiler *compiler,
732 struct v3d_key *key,
733 struct v3d_prog_data **prog_data,
734 nir_shader *s,
735 void (*debug_output)(const char *msg,
736 void *debug_output_data),
737 void *debug_output_data,
738 int program_id, int variant_id,
739 uint32_t *final_assembly_size);
740
741 void v3d_nir_to_vir(struct v3d_compile *c);
742
743 void vir_compile_destroy(struct v3d_compile *c);
744 const char *vir_get_stage_name(struct v3d_compile *c);
745 struct qblock *vir_new_block(struct v3d_compile *c);
746 void vir_set_emit_block(struct v3d_compile *c, struct qblock *block);
747 void vir_link_blocks(struct qblock *predecessor, struct qblock *successor);
748 struct qblock *vir_entry_block(struct v3d_compile *c);
749 struct qblock *vir_exit_block(struct v3d_compile *c);
750 struct qinst *vir_add_inst(enum v3d_qpu_add_op op, struct qreg dst,
751 struct qreg src0, struct qreg src1);
752 struct qinst *vir_mul_inst(enum v3d_qpu_mul_op op, struct qreg dst,
753 struct qreg src0, struct qreg src1);
754 struct qinst *vir_branch_inst(enum v3d_qpu_branch_cond cond, struct qreg src0);
755 void vir_remove_instruction(struct v3d_compile *c, struct qinst *qinst);
756 struct qreg vir_uniform(struct v3d_compile *c,
757 enum quniform_contents contents,
758 uint32_t data);
759 void vir_schedule_instructions(struct v3d_compile *c);
760 struct v3d_qpu_instr v3d_qpu_nop(void);
761
762 struct qreg vir_emit_def(struct v3d_compile *c, struct qinst *inst);
763 struct qinst *vir_emit_nondef(struct v3d_compile *c, struct qinst *inst);
764 void vir_set_cond(struct qinst *inst, enum v3d_qpu_cond cond);
765 void vir_set_pf(struct qinst *inst, enum v3d_qpu_pf pf);
766 void vir_set_uf(struct qinst *inst, enum v3d_qpu_uf uf);
767 void vir_set_unpack(struct qinst *inst, int src,
768 enum v3d_qpu_input_unpack unpack);
769
770 struct qreg vir_get_temp(struct v3d_compile *c);
771 void vir_emit_last_thrsw(struct v3d_compile *c);
772 void vir_calculate_live_intervals(struct v3d_compile *c);
773 bool vir_has_implicit_uniform(struct qinst *inst);
774 int vir_get_implicit_uniform_src(struct qinst *inst);
775 int vir_get_non_sideband_nsrc(struct qinst *inst);
776 int vir_get_nsrc(struct qinst *inst);
777 bool vir_has_side_effects(struct v3d_compile *c, struct qinst *inst);
778 bool vir_get_add_op(struct qinst *inst, enum v3d_qpu_add_op *op);
779 bool vir_get_mul_op(struct qinst *inst, enum v3d_qpu_mul_op *op);
780 bool vir_is_raw_mov(struct qinst *inst);
781 bool vir_is_tex(struct qinst *inst);
782 bool vir_is_add(struct qinst *inst);
783 bool vir_is_mul(struct qinst *inst);
784 bool vir_is_float_input(struct qinst *inst);
785 bool vir_writes_r3(const struct v3d_device_info *devinfo, struct qinst *inst);
786 bool vir_writes_r4(const struct v3d_device_info *devinfo, struct qinst *inst);
787 struct qreg vir_follow_movs(struct v3d_compile *c, struct qreg reg);
788 uint8_t vir_channels_written(struct qinst *inst);
789 struct qreg ntq_get_src(struct v3d_compile *c, nir_src src, int i);
790 void ntq_store_dest(struct v3d_compile *c, nir_dest *dest, int chan,
791 struct qreg result);
792 void vir_emit_thrsw(struct v3d_compile *c);
793
794 void vir_dump(struct v3d_compile *c);
795 void vir_dump_inst(struct v3d_compile *c, struct qinst *inst);
796 void vir_dump_uniform(enum quniform_contents contents, uint32_t data);
797
798 void vir_validate(struct v3d_compile *c);
799
800 void vir_optimize(struct v3d_compile *c);
801 bool vir_opt_algebraic(struct v3d_compile *c);
802 bool vir_opt_constant_folding(struct v3d_compile *c);
803 bool vir_opt_copy_propagate(struct v3d_compile *c);
804 bool vir_opt_dead_code(struct v3d_compile *c);
805 bool vir_opt_peephole_sf(struct v3d_compile *c);
806 bool vir_opt_small_immediates(struct v3d_compile *c);
807 bool vir_opt_vpm(struct v3d_compile *c);
808 void v3d_nir_lower_blend(nir_shader *s, struct v3d_compile *c);
809 void v3d_nir_lower_io(nir_shader *s, struct v3d_compile *c);
810 void v3d_nir_lower_txf_ms(nir_shader *s, struct v3d_compile *c);
811 void v3d_nir_lower_image_load_store(nir_shader *s);
812 void vir_lower_uniforms(struct v3d_compile *c);
813
814 void v3d33_vir_vpm_read_setup(struct v3d_compile *c, int num_components);
815 void v3d33_vir_vpm_write_setup(struct v3d_compile *c);
816 void v3d33_vir_emit_tex(struct v3d_compile *c, nir_tex_instr *instr);
817 void v3d40_vir_emit_tex(struct v3d_compile *c, nir_tex_instr *instr);
818 void v3d40_vir_emit_image_load_store(struct v3d_compile *c,
819 nir_intrinsic_instr *instr);
820
821 void v3d_vir_to_qpu(struct v3d_compile *c, struct qpu_reg *temp_registers);
822 uint32_t v3d_qpu_schedule_instructions(struct v3d_compile *c);
823 void qpu_validate(struct v3d_compile *c);
824 struct qpu_reg *v3d_register_allocate(struct v3d_compile *c, bool *spilled);
825 bool vir_init_reg_sets(struct v3d_compiler *compiler);
826
827 bool v3d_gl_format_is_return_32(GLenum format);
828
829 void vir_PF(struct v3d_compile *c, struct qreg src, enum v3d_qpu_pf pf);
830
831 static inline bool
832 quniform_contents_is_texture_p0(enum quniform_contents contents)
833 {
834 return (contents >= QUNIFORM_TEXTURE_CONFIG_P0_0 &&
835 contents < (QUNIFORM_TEXTURE_CONFIG_P0_0 +
836 V3D_MAX_TEXTURE_SAMPLERS));
837 }
838
839 static inline struct qreg
840 vir_uniform_ui(struct v3d_compile *c, uint32_t ui)
841 {
842 return vir_uniform(c, QUNIFORM_CONSTANT, ui);
843 }
844
845 static inline struct qreg
846 vir_uniform_f(struct v3d_compile *c, float f)
847 {
848 return vir_uniform(c, QUNIFORM_CONSTANT, fui(f));
849 }
850
851 #define VIR_ALU0(name, vir_inst, op) \
852 static inline struct qreg \
853 vir_##name(struct v3d_compile *c) \
854 { \
855 return vir_emit_def(c, vir_inst(op, c->undef, \
856 c->undef, c->undef)); \
857 } \
858 static inline struct qinst * \
859 vir_##name##_dest(struct v3d_compile *c, struct qreg dest) \
860 { \
861 return vir_emit_nondef(c, vir_inst(op, dest, \
862 c->undef, c->undef)); \
863 }
864
865 #define VIR_ALU1(name, vir_inst, op) \
866 static inline struct qreg \
867 vir_##name(struct v3d_compile *c, struct qreg a) \
868 { \
869 return vir_emit_def(c, vir_inst(op, c->undef, \
870 a, c->undef)); \
871 } \
872 static inline struct qinst * \
873 vir_##name##_dest(struct v3d_compile *c, struct qreg dest, \
874 struct qreg a) \
875 { \
876 return vir_emit_nondef(c, vir_inst(op, dest, a, \
877 c->undef)); \
878 }
879
880 #define VIR_ALU2(name, vir_inst, op) \
881 static inline struct qreg \
882 vir_##name(struct v3d_compile *c, struct qreg a, struct qreg b) \
883 { \
884 return vir_emit_def(c, vir_inst(op, c->undef, a, b)); \
885 } \
886 static inline struct qinst * \
887 vir_##name##_dest(struct v3d_compile *c, struct qreg dest, \
888 struct qreg a, struct qreg b) \
889 { \
890 return vir_emit_nondef(c, vir_inst(op, dest, a, b)); \
891 }
892
893 #define VIR_NODST_0(name, vir_inst, op) \
894 static inline struct qinst * \
895 vir_##name(struct v3d_compile *c) \
896 { \
897 return vir_emit_nondef(c, vir_inst(op, c->undef, \
898 c->undef, c->undef)); \
899 }
900
901 #define VIR_NODST_1(name, vir_inst, op) \
902 static inline struct qinst * \
903 vir_##name(struct v3d_compile *c, struct qreg a) \
904 { \
905 return vir_emit_nondef(c, vir_inst(op, c->undef, \
906 a, c->undef)); \
907 }
908
909 #define VIR_NODST_2(name, vir_inst, op) \
910 static inline struct qinst * \
911 vir_##name(struct v3d_compile *c, struct qreg a, struct qreg b) \
912 { \
913 return vir_emit_nondef(c, vir_inst(op, c->undef, \
914 a, b)); \
915 }
916
917 #define VIR_SFU(name) \
918 static inline struct qreg \
919 vir_##name(struct v3d_compile *c, struct qreg a) \
920 { \
921 if (c->devinfo->ver >= 41) { \
922 return vir_emit_def(c, vir_add_inst(V3D_QPU_A_##name, \
923 c->undef, \
924 a, c->undef)); \
925 } else { \
926 vir_FMOV_dest(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_##name), a); \
927 return vir_FMOV(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_R4)); \
928 } \
929 } \
930 static inline struct qinst * \
931 vir_##name##_dest(struct v3d_compile *c, struct qreg dest, \
932 struct qreg a) \
933 { \
934 if (c->devinfo->ver >= 41) { \
935 return vir_emit_nondef(c, vir_add_inst(V3D_QPU_A_##name, \
936 dest, \
937 a, c->undef)); \
938 } else { \
939 vir_FMOV_dest(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_##name), a); \
940 return vir_FMOV_dest(c, dest, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_R4)); \
941 } \
942 }
943
944 #define VIR_A_ALU2(name) VIR_ALU2(name, vir_add_inst, V3D_QPU_A_##name)
945 #define VIR_M_ALU2(name) VIR_ALU2(name, vir_mul_inst, V3D_QPU_M_##name)
946 #define VIR_A_ALU1(name) VIR_ALU1(name, vir_add_inst, V3D_QPU_A_##name)
947 #define VIR_M_ALU1(name) VIR_ALU1(name, vir_mul_inst, V3D_QPU_M_##name)
948 #define VIR_A_ALU0(name) VIR_ALU0(name, vir_add_inst, V3D_QPU_A_##name)
949 #define VIR_M_ALU0(name) VIR_ALU0(name, vir_mul_inst, V3D_QPU_M_##name)
950 #define VIR_A_NODST_2(name) VIR_NODST_2(name, vir_add_inst, V3D_QPU_A_##name)
951 #define VIR_M_NODST_2(name) VIR_NODST_2(name, vir_mul_inst, V3D_QPU_M_##name)
952 #define VIR_A_NODST_1(name) VIR_NODST_1(name, vir_add_inst, V3D_QPU_A_##name)
953 #define VIR_M_NODST_1(name) VIR_NODST_1(name, vir_mul_inst, V3D_QPU_M_##name)
954 #define VIR_A_NODST_0(name) VIR_NODST_0(name, vir_add_inst, V3D_QPU_A_##name)
955
956 VIR_A_ALU2(FADD)
957 VIR_A_ALU2(VFPACK)
958 VIR_A_ALU2(FSUB)
959 VIR_A_ALU2(FMIN)
960 VIR_A_ALU2(FMAX)
961
962 VIR_A_ALU2(ADD)
963 VIR_A_ALU2(SUB)
964 VIR_A_ALU2(SHL)
965 VIR_A_ALU2(SHR)
966 VIR_A_ALU2(ASR)
967 VIR_A_ALU2(ROR)
968 VIR_A_ALU2(MIN)
969 VIR_A_ALU2(MAX)
970 VIR_A_ALU2(UMIN)
971 VIR_A_ALU2(UMAX)
972 VIR_A_ALU2(AND)
973 VIR_A_ALU2(OR)
974 VIR_A_ALU2(XOR)
975 VIR_A_ALU2(VADD)
976 VIR_A_ALU2(VSUB)
977 VIR_A_NODST_2(STVPMV)
978 VIR_A_ALU1(NOT)
979 VIR_A_ALU1(NEG)
980 VIR_A_ALU1(FLAPUSH)
981 VIR_A_ALU1(FLBPUSH)
982 VIR_A_ALU1(FLPOP)
983 VIR_A_ALU1(SETMSF)
984 VIR_A_ALU1(SETREVF)
985 VIR_A_ALU0(TIDX)
986 VIR_A_ALU0(EIDX)
987 VIR_A_ALU1(LDVPMV_IN)
988 VIR_A_ALU1(LDVPMV_OUT)
989 VIR_A_ALU0(TMUWT)
990
991 VIR_A_ALU0(FXCD)
992 VIR_A_ALU0(XCD)
993 VIR_A_ALU0(FYCD)
994 VIR_A_ALU0(YCD)
995 VIR_A_ALU0(MSF)
996 VIR_A_ALU0(REVF)
997 VIR_A_NODST_1(VPMSETUP)
998 VIR_A_NODST_0(VPMWT)
999 VIR_A_ALU2(FCMP)
1000 VIR_A_ALU2(VFMAX)
1001
1002 VIR_A_ALU1(FROUND)
1003 VIR_A_ALU1(FTOIN)
1004 VIR_A_ALU1(FTRUNC)
1005 VIR_A_ALU1(FTOIZ)
1006 VIR_A_ALU1(FFLOOR)
1007 VIR_A_ALU1(FTOUZ)
1008 VIR_A_ALU1(FCEIL)
1009 VIR_A_ALU1(FTOC)
1010
1011 VIR_A_ALU1(FDX)
1012 VIR_A_ALU1(FDY)
1013
1014 VIR_A_ALU1(ITOF)
1015 VIR_A_ALU1(CLZ)
1016 VIR_A_ALU1(UTOF)
1017
1018 VIR_M_ALU2(UMUL24)
1019 VIR_M_ALU2(FMUL)
1020 VIR_M_ALU2(SMUL24)
1021 VIR_M_NODST_2(MULTOP)
1022
1023 VIR_M_ALU1(MOV)
1024 VIR_M_ALU1(FMOV)
1025
1026 VIR_SFU(RECIP)
1027 VIR_SFU(RSQRT)
1028 VIR_SFU(EXP)
1029 VIR_SFU(LOG)
1030 VIR_SFU(SIN)
1031 VIR_SFU(RSQRT2)
1032
1033 static inline struct qinst *
1034 vir_MOV_cond(struct v3d_compile *c, enum v3d_qpu_cond cond,
1035 struct qreg dest, struct qreg src)
1036 {
1037 struct qinst *mov = vir_MOV_dest(c, dest, src);
1038 vir_set_cond(mov, cond);
1039 return mov;
1040 }
1041
1042 static inline struct qreg
1043 vir_SEL(struct v3d_compile *c, enum v3d_qpu_cond cond,
1044 struct qreg src0, struct qreg src1)
1045 {
1046 struct qreg t = vir_get_temp(c);
1047 vir_MOV_dest(c, t, src1);
1048 vir_MOV_cond(c, cond, t, src0);
1049 return t;
1050 }
1051
1052 static inline struct qinst *
1053 vir_NOP(struct v3d_compile *c)
1054 {
1055 return vir_emit_nondef(c, vir_add_inst(V3D_QPU_A_NOP,
1056 c->undef, c->undef, c->undef));
1057 }
1058
1059 static inline struct qreg
1060 vir_LDTMU(struct v3d_compile *c)
1061 {
1062 if (c->devinfo->ver >= 41) {
1063 struct qinst *ldtmu = vir_add_inst(V3D_QPU_A_NOP, c->undef,
1064 c->undef, c->undef);
1065 ldtmu->qpu.sig.ldtmu = true;
1066
1067 return vir_emit_def(c, ldtmu);
1068 } else {
1069 vir_NOP(c)->qpu.sig.ldtmu = true;
1070 return vir_MOV(c, vir_reg(QFILE_MAGIC, V3D_QPU_WADDR_R4));
1071 }
1072 }
1073
1074 static inline struct qreg
1075 vir_UMUL(struct v3d_compile *c, struct qreg src0, struct qreg src1)
1076 {
1077 vir_MULTOP(c, src0, src1);
1078 return vir_UMUL24(c, src0, src1);
1079 }
1080
1081 /*
1082 static inline struct qreg
1083 vir_LOAD_IMM(struct v3d_compile *c, uint32_t val)
1084 {
1085 return vir_emit_def(c, vir_inst(QOP_LOAD_IMM, c->undef,
1086 vir_reg(QFILE_LOAD_IMM, val), c->undef));
1087 }
1088
1089 static inline struct qreg
1090 vir_LOAD_IMM_U2(struct v3d_compile *c, uint32_t val)
1091 {
1092 return vir_emit_def(c, vir_inst(QOP_LOAD_IMM_U2, c->undef,
1093 vir_reg(QFILE_LOAD_IMM, val),
1094 c->undef));
1095 }
1096 static inline struct qreg
1097 vir_LOAD_IMM_I2(struct v3d_compile *c, uint32_t val)
1098 {
1099 return vir_emit_def(c, vir_inst(QOP_LOAD_IMM_I2, c->undef,
1100 vir_reg(QFILE_LOAD_IMM, val),
1101 c->undef));
1102 }
1103 */
1104
1105 static inline struct qinst *
1106 vir_BRANCH(struct v3d_compile *c, enum v3d_qpu_branch_cond cond)
1107 {
1108 /* The actual uniform_data value will be set at scheduling time */
1109 return vir_emit_nondef(c, vir_branch_inst(cond, vir_uniform_ui(c, 0)));
1110 }
1111
1112 #define vir_for_each_block(block, c) \
1113 list_for_each_entry(struct qblock, block, &c->blocks, link)
1114
1115 #define vir_for_each_block_rev(block, c) \
1116 list_for_each_entry_rev(struct qblock, block, &c->blocks, link)
1117
1118 /* Loop over the non-NULL members of the successors array. */
1119 #define vir_for_each_successor(succ, block) \
1120 for (struct qblock *succ = block->successors[0]; \
1121 succ != NULL; \
1122 succ = (succ == block->successors[1] ? NULL : \
1123 block->successors[1]))
1124
1125 #define vir_for_each_inst(inst, block) \
1126 list_for_each_entry(struct qinst, inst, &block->instructions, link)
1127
1128 #define vir_for_each_inst_rev(inst, block) \
1129 list_for_each_entry_rev(struct qinst, inst, &block->instructions, link)
1130
1131 #define vir_for_each_inst_safe(inst, block) \
1132 list_for_each_entry_safe(struct qinst, inst, &block->instructions, link)
1133
1134 #define vir_for_each_inst_inorder(inst, c) \
1135 vir_for_each_block(_block, c) \
1136 vir_for_each_inst(inst, _block)
1137
1138 #endif /* V3D_COMPILER_H */