projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
pan/bi: Pack fp16 ATEST
[mesa.git] / src / panfrost / bifrost / bi_pack.c
1 /*
2 * Copyright (C) 2020 Collabora, Ltd.
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 FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 * SOFTWARE.
22 */
23
24 #include "compiler.h"
25
26 #define RETURN_PACKED(str) { \
27 uint64_t temp = 0; \
28 memcpy(&temp, &str, sizeof(str)); \
29 return temp; \
30 }
31
32 /* This file contains the final passes of the compiler. Running after
33 * scheduling and RA, the IR is now finalized, so we need to emit it to actual
34 * bits on the wire (as well as fixup branches) */
35
36 static uint64_t
37 bi_pack_header(bi_clause *clause, bi_clause *next, bool is_fragment)
38 {
39 struct bifrost_header header = {
40 .back_to_back = clause->back_to_back,
41 .no_end_of_shader = (next != NULL),
42 .elide_writes = is_fragment,
43 .branch_cond = clause->branch_conditional,
44 .datareg_writebarrier = clause->data_register_write_barrier,
45 .datareg = clause->data_register,
46 .scoreboard_deps = next ? next->dependencies : 0,
47 .scoreboard_index = clause->scoreboard_id,
48 .clause_type = clause->clause_type,
49 .next_clause_type = next ? next->clause_type : 0,
50 };
51
52 header.branch_cond |= header.back_to_back;
53
54 uint64_t u = 0;
55 memcpy(&u, &header, sizeof(header));
56 return u;
57 }
58
59 /* Represents the assignment of ports for a given bundle */
60
61 struct bi_registers {
62 /* Register to assign to each port */
63 unsigned port[4];
64
65 /* Read ports can be disabled */
66 bool enabled[2];
67
68 /* Should we write FMA? what about ADD? If only a single port is
69 * enabled it is in port 2, else ADD/FMA is 2/3 respectively */
70 bool write_fma, write_add;
71
72 /* Should we read with port 3? */
73 bool read_port3;
74
75 /* Packed uniform/constant */
76 uint8_t uniform_constant;
77
78 /* Whether writes are actually for the last instruction */
79 bool first_instruction;
80 };
81
82 static inline void
83 bi_print_ports(struct bi_registers *regs)
84 {
85 for (unsigned i = 0; i < 2; ++i) {
86 if (regs->enabled[i])
87 printf("port %u: %u\n", i, regs->port[i]);
88 }
89
90 if (regs->write_fma || regs->write_add) {
91 printf("port 2 (%s): %u\n",
92 regs->write_add ? "ADD" : "FMA",
93 regs->port[2]);
94 }
95
96 if ((regs->write_fma && regs->write_add) || regs->read_port3) {
97 printf("port 3 (%s): %u\n",
98 regs->read_port3 ? "read" : "FMA",
99 regs->port[3]);
100 }
101 }
102
103 /* The uniform/constant slot allows loading a contiguous 64-bit immediate or
104 * pushed uniform per bundle. Figure out which one we need in the bundle (the
105 * scheduler needs to ensure we only have one type per bundle), validate
106 * everything, and rewrite away the register/uniform indices to use 3-bit
107 * sources directly. */
108
109 static unsigned
110 bi_lookup_constant(bi_clause *clause, uint64_t cons, bool *hi, bool b64)
111 {
112 uint64_t want = (cons >> 4);
113
114 for (unsigned i = 0; i < clause->constant_count; ++i) {
115 /* Only check top 60-bits since that's what's actually embedded
116 * in the clause, the bottom 4-bits are bundle-inline */
117
118 uint64_t candidates[2] = {
119 clause->constants[i] >> 4,
120 clause->constants[i] >> 36
121 };
122
123 /* For <64-bit mode, we treat lo/hi separately */
124
125 if (!b64)
126 candidates[0] &= (0xFFFFFFFF >> 4);
127
128 if (candidates[0] == want)
129 return i;
130
131 if (candidates[1] == want && !b64) {
132 *hi = true;
133 return i;
134 }
135 }
136
137 unreachable("Invalid constant accessed");
138 }
139
140 static unsigned
141 bi_constant_field(unsigned idx)
142 {
143 assert(idx <= 5);
144
145 const unsigned values[] = {
146 4, 5, 6, 7, 2, 3
147 };
148
149 return values[idx] << 4;
150 }
151
152 static bool
153 bi_assign_uniform_constant_single(
154 struct bi_registers *regs,
155 bi_clause *clause,
156 bi_instruction *ins, bool assigned, bool fast_zero)
157 {
158 if (!ins)
159 return assigned;
160
161 bi_foreach_src(ins, s) {
162 if (s == 0 && (ins->type == BI_LOAD_VAR_ADDRESS || ins->type == BI_LOAD_ATTR)) continue;
163
164 if (ins->src[s] & BIR_INDEX_CONSTANT) {
165 bool hi = false;
166 bool b64 = nir_alu_type_get_type_size(ins->src_types[s]) > 32;
167 uint64_t cons = bi_get_immediate(ins, s);
168 unsigned idx = bi_lookup_constant(clause, cons, &hi, b64);
169 unsigned lo = clause->constants[idx] & 0xF;
170 unsigned f = bi_constant_field(idx) | lo;
171
172 if (assigned && regs->uniform_constant != f)
173 unreachable("Mismatched uniform/const field: imm");
174
175 regs->uniform_constant = f;
176 ins->src[s] = BIR_INDEX_PASS | (hi ? BIFROST_SRC_CONST_HI : BIFROST_SRC_CONST_LO);
177 assigned = true;
178 } else if (ins->src[s] & BIR_INDEX_ZERO && (ins->type == BI_LOAD_UNIFORM || ins->type == BI_LOAD_VAR)) {
179 /* XXX: HACK UNTIL WE HAVE HI MATCHING DUE TO OVERFLOW XXX */
180 ins->src[s] = BIR_INDEX_PASS | BIFROST_SRC_CONST_HI;
181 } else if (ins->src[s] & BIR_INDEX_ZERO && !fast_zero) {
182 /* FMAs have a fast zero port, ADD needs to use the
183 * uniform/const port's special 0 mode handled here */
184 unsigned f = 0;
185
186 if (assigned && regs->uniform_constant != f)
187 unreachable("Mismatched uniform/const field: 0");
188
189 regs->uniform_constant = f;
190 ins->src[s] = BIR_INDEX_PASS | BIFROST_SRC_CONST_LO;
191 assigned = true;
192 } else if (s & BIR_INDEX_UNIFORM) {
193 unreachable("Push uniforms not implemented yet");
194 }
195 }
196
197 return assigned;
198 }
199
200 static void
201 bi_assign_uniform_constant(
202 bi_clause *clause,
203 struct bi_registers *regs,
204 bi_bundle bundle)
205 {
206 bool assigned =
207 bi_assign_uniform_constant_single(regs, clause, bundle.fma, false, true);
208
209 bi_assign_uniform_constant_single(regs, clause, bundle.add, assigned, false);
210 }
211
212 /* Assigns a port for reading, before anything is written */
213
214 static void
215 bi_assign_port_read(struct bi_registers *regs, unsigned src)
216 {
217 /* We only assign for registers */
218 if (!(src & BIR_INDEX_REGISTER))
219 return;
220
221 unsigned reg = src & ~BIR_INDEX_REGISTER;
222
223 /* Check if we already assigned the port */
224 for (unsigned i = 0; i <= 1; ++i) {
225 if (regs->port[i] == reg && regs->enabled[i])
226 return;
227 }
228
229 if (regs->port[3] == reg && regs->read_port3)
230 return;
231
232 /* Assign it now */
233
234 for (unsigned i = 0; i <= 1; ++i) {
235 if (!regs->enabled[i]) {
236 regs->port[i] = reg;
237 regs->enabled[i] = true;
238 return;
239 }
240 }
241
242 if (!regs->read_port3) {
243 regs->port[3] = reg;
244 regs->read_port3 = true;
245 return;
246 }
247
248 bi_print_ports(regs);
249 unreachable("Failed to find a free port for src");
250 }
251
252 static struct bi_registers
253 bi_assign_ports(bi_bundle now, bi_bundle prev)
254 {
255 struct bi_registers regs = { 0 };
256
257 /* We assign ports for the main register mechanism. Special ops
258 * use the data registers, which has its own mechanism entirely
259 * and thus gets skipped over here. */
260
261 unsigned read_dreg = now.add &&
262 bi_class_props[now.add->type] & BI_DATA_REG_SRC;
263
264 unsigned write_dreg = prev.add &&
265 bi_class_props[prev.add->type] & BI_DATA_REG_DEST;
266
267 /* First, assign reads */
268
269 if (now.fma)
270 bi_foreach_src(now.fma, src)
271 bi_assign_port_read(&regs, now.fma->src[src]);
272
273 if (now.add) {
274 bi_foreach_src(now.add, src) {
275 if (!(src == 0 && read_dreg))
276 bi_assign_port_read(&regs, now.add->src[src]);
277 }
278 }
279
280 /* Next, assign writes */
281
282 if (prev.add && prev.add->dest & BIR_INDEX_REGISTER && !write_dreg) {
283 regs.port[2] = prev.add->dest & ~BIR_INDEX_REGISTER;
284 regs.write_add = true;
285 }
286
287 if (prev.fma && prev.fma->dest & BIR_INDEX_REGISTER) {
288 unsigned r = prev.fma->dest & ~BIR_INDEX_REGISTER;
289
290 if (regs.write_add) {
291 /* Scheduler constraint: cannot read 3 and write 2 */
292 assert(!regs.read_port3);
293 regs.port[3] = r;
294 } else {
295 regs.port[2] = r;
296 }
297
298 regs.write_fma = true;
299 }
300
301 /* Finally, ensure port 1 > port 0 for the 63-x trick to function */
302
303 if (regs.enabled[0] && regs.enabled[1] && regs.port[1] < regs.port[0]) {
304 unsigned temp = regs.port[0];
305 regs.port[0] = regs.port[1];
306 regs.port[1] = temp;
307 }
308
309 return regs;
310 }
311
312 /* Determines the register control field, ignoring the first? flag */
313
314 static enum bifrost_reg_control
315 bi_pack_register_ctrl_lo(struct bi_registers r)
316 {
317 if (r.write_fma) {
318 if (r.write_add) {
319 assert(!r.read_port3);
320 return BIFROST_WRITE_ADD_P2_FMA_P3;
321 } else {
322 if (r.read_port3)
323 return BIFROST_WRITE_FMA_P2_READ_P3;
324 else
325 return BIFROST_WRITE_FMA_P2;
326 }
327 } else if (r.write_add) {
328 if (r.read_port3)
329 return BIFROST_WRITE_ADD_P2_READ_P3;
330 else
331 return BIFROST_WRITE_ADD_P2;
332 } else if (r.read_port3)
333 return BIFROST_READ_P3;
334 else
335 return BIFROST_REG_NONE;
336 }
337
338 /* Ditto but account for the first? flag this time */
339
340 static enum bifrost_reg_control
341 bi_pack_register_ctrl(struct bi_registers r)
342 {
343 enum bifrost_reg_control ctrl = bi_pack_register_ctrl_lo(r);
344
345 if (r.first_instruction) {
346 if (ctrl == BIFROST_REG_NONE)
347 ctrl = BIFROST_FIRST_NONE;
348 else if (ctrl == BIFROST_WRITE_FMA_P2_READ_P3)
349 ctrl = BIFROST_FIRST_WRITE_FMA_P2_READ_P3;
350 else
351 ctrl |= BIFROST_FIRST_NONE;
352 }
353
354 return ctrl;
355 }
356
357 static uint64_t
358 bi_pack_registers(struct bi_registers regs)
359 {
360 enum bifrost_reg_control ctrl = bi_pack_register_ctrl(regs);
361 struct bifrost_regs s = { 0 };
362 uint64_t packed = 0;
363
364 if (regs.enabled[1]) {
365 /* Gotta save that bit!~ Required by the 63-x trick */
366 assert(regs.port[1] > regs.port[0]);
367 assert(regs.enabled[0]);
368
369 /* Do the 63-x trick, see docs/disasm */
370 if (regs.port[0] > 31) {
371 regs.port[0] = 63 - regs.port[0];
372 regs.port[1] = 63 - regs.port[1];
373 }
374
375 assert(regs.port[0] <= 31);
376 assert(regs.port[1] <= 63);
377
378 s.ctrl = ctrl;
379 s.reg1 = regs.port[1];
380 s.reg0 = regs.port[0];
381 } else {
382 /* Port 1 disabled, so set to zero and use port 1 for ctrl */
383 s.ctrl = 0;
384 s.reg1 = ctrl << 2;
385
386 if (regs.enabled[0]) {
387 /* Bit 0 upper bit of port 0 */
388 s.reg1 |= (regs.port[0] >> 5);
389
390 /* Rest of port 0 in usual spot */
391 s.reg0 = (regs.port[0] & 0b11111);
392 } else {
393 /* Bit 1 set if port 0 also disabled */
394 s.reg1 |= (1 << 1);
395 }
396 }
397
398 /* When port 3 isn't used, we have to set it to port 2, and vice versa,
399 * or INSTR_INVALID_ENC is raised. The reason is unknown. */
400
401 bool has_port2 = regs.write_fma || regs.write_add;
402 bool has_port3 = regs.read_port3 || (regs.write_fma && regs.write_add);
403
404 if (!has_port3)
405 regs.port[3] = regs.port[2];
406
407 if (!has_port2)
408 regs.port[2] = regs.port[3];
409
410 s.reg3 = regs.port[3];
411 s.reg2 = regs.port[2];
412 s.uniform_const = regs.uniform_constant;
413
414 memcpy(&packed, &s, sizeof(s));
415 return packed;
416 }
417
418 static void
419 bi_set_data_register(bi_clause *clause, unsigned idx)
420 {
421 assert(idx & BIR_INDEX_REGISTER);
422 unsigned reg = idx & ~BIR_INDEX_REGISTER;
423 assert(reg <= 63);
424 clause->data_register = reg;
425 }
426
427 static void
428 bi_read_data_register(bi_clause *clause, bi_instruction *ins)
429 {
430 bi_set_data_register(clause, ins->src[0]);
431 }
432
433 static void
434 bi_write_data_register(bi_clause *clause, bi_instruction *ins)
435 {
436 bi_set_data_register(clause, ins->dest);
437 }
438
439 static enum bifrost_packed_src
440 bi_get_src_reg_port(struct bi_registers *regs, unsigned src)
441 {
442 unsigned reg = src & ~BIR_INDEX_REGISTER;
443
444 if (regs->port[0] == reg && regs->enabled[0])
445 return BIFROST_SRC_PORT0;
446 else if (regs->port[1] == reg && regs->enabled[1])
447 return BIFROST_SRC_PORT1;
448 else if (regs->port[3] == reg && regs->read_port3)
449 return BIFROST_SRC_PORT3;
450 else
451 unreachable("Tried to access register with no port");
452 }
453
454 static enum bifrost_packed_src
455 bi_get_src(bi_instruction *ins, struct bi_registers *regs, unsigned s, bool is_fma)
456 {
457 unsigned src = ins->src[s];
458
459 if (src & BIR_INDEX_REGISTER)
460 return bi_get_src_reg_port(regs, src);
461 else if (src & BIR_INDEX_ZERO && is_fma)
462 return BIFROST_SRC_STAGE;
463 else if (src & BIR_INDEX_PASS)
464 return src & ~BIR_INDEX_PASS;
465 else
466 unreachable("Unknown src");
467 }
468
469 /* Constructs a packed 2-bit swizzle for a 16-bit vec2 source. Source must be
470 * 16-bit and written components must correspond to valid swizzles (component x
471 * or y). */
472
473 static unsigned
474 bi_swiz16(bi_instruction *ins, unsigned src)
475 {
476 assert(nir_alu_type_get_type_size(ins->src_types[src]) == 16);
477 unsigned swizzle = 0;
478
479 for (unsigned c = 0; c < 2; ++c) {
480 if (!bi_writes_component(ins, src)) continue;
481
482 unsigned k = ins->swizzle[src][c];
483 assert(k <= 1);
484 swizzle |= (k << c);
485 }
486
487 return swizzle;
488 }
489
490 static unsigned
491 bi_pack_fma_fma(bi_instruction *ins, struct bi_registers *regs)
492 {
493 /* (-a)(-b) = ab, so we only need one negate bit */
494 bool negate_mul = ins->src_neg[0] ^ ins->src_neg[1];
495
496 if (ins->op.mscale) {
497 assert(!(ins->src_abs[0] && ins->src_abs[1]));
498 assert(!ins->src_abs[2] || !ins->src_neg[3] || !ins->src_abs[3]);
499
500 /* We can have exactly one abs, and can flip the multiplication
501 * to make it fit if we have to */
502 bool flip_ab = ins->src_abs[1];
503
504 struct bifrost_fma_mscale pack = {
505 .src0 = bi_get_src(ins, regs, flip_ab ? 1 : 0, true),
506 .src1 = bi_get_src(ins, regs, flip_ab ? 0 : 1, true),
507 .src2 = bi_get_src(ins, regs, 2, true),
508 .src3 = bi_get_src(ins, regs, 3, true),
509 .mscale_mode = 0,
510 .mode = ins->outmod,
511 .src0_abs = ins->src_abs[0] || ins->src_abs[1],
512 .src1_neg = negate_mul,
513 .src2_neg = ins->src_neg[2],
514 .op = BIFROST_FMA_OP_MSCALE,
515 };
516
517 RETURN_PACKED(pack);
518 } else if (ins->dest_type == nir_type_float32) {
519 struct bifrost_fma_fma pack = {
520 .src0 = bi_get_src(ins, regs, 0, true),
521 .src1 = bi_get_src(ins, regs, 1, true),
522 .src2 = bi_get_src(ins, regs, 2, true),
523 .src0_abs = ins->src_abs[0],
524 .src1_abs = ins->src_abs[1],
525 .src2_abs = ins->src_abs[2],
526 .src0_neg = negate_mul,
527 .src2_neg = ins->src_neg[2],
528 .outmod = ins->outmod,
529 .roundmode = ins->roundmode,
530 .op = BIFROST_FMA_OP_FMA
531 };
532
533 RETURN_PACKED(pack);
534 } else if (ins->dest_type == nir_type_float16) {
535 struct bifrost_fma_fma16 pack = {
536 .src0 = bi_get_src(ins, regs, 0, true),
537 .src1 = bi_get_src(ins, regs, 1, true),
538 .src2 = bi_get_src(ins, regs, 2, true),
539 .swizzle_0 = bi_swiz16(ins, 0),
540 .swizzle_1 = bi_swiz16(ins, 1),
541 .swizzle_2 = bi_swiz16(ins, 2),
542 .src0_neg = negate_mul,
543 .src2_neg = ins->src_neg[2],
544 .outmod = ins->outmod,
545 .roundmode = ins->roundmode,
546 .op = BIFROST_FMA_OP_FMA16
547 };
548
549 RETURN_PACKED(pack);
550 } else {
551 unreachable("Invalid fma dest type");
552 }
553 }
554
555 static unsigned
556 bi_pack_fma_addmin_f32(bi_instruction *ins, struct bi_registers *regs)
557 {
558 unsigned op =
559 (ins->type == BI_ADD) ? BIFROST_FMA_OP_FADD32 :
560 (ins->op.minmax == BI_MINMAX_MIN) ? BIFROST_FMA_OP_FMIN32 :
561 BIFROST_FMA_OP_FMAX32;
562
563 struct bifrost_fma_add pack = {
564 .src0 = bi_get_src(ins, regs, 0, true),
565 .src1 = bi_get_src(ins, regs, 1, true),
566 .src0_abs = ins->src_abs[0],
567 .src1_abs = ins->src_abs[1],
568 .src0_neg = ins->src_neg[0],
569 .src1_neg = ins->src_neg[1],
570 .unk = 0x0,
571 .outmod = ins->outmod,
572 .roundmode = (ins->type == BI_ADD) ? ins->roundmode : ins->minmax,
573 .op = op
574 };
575
576 RETURN_PACKED(pack);
577 }
578
579 static unsigned
580 bi_pack_fmadd_min_f16(bi_instruction *ins, struct bi_registers *regs, bool FMA)
581 {
582 unsigned op =
583 (!FMA) ? ((ins->op.minmax == BI_MINMAX_MIN) ?
584 BIFROST_ADD_OP_FMIN16 : BIFROST_ADD_OP_FMAX16) :
585 (ins->type == BI_ADD) ? BIFROST_FMA_OP_FADD16 :
586 (ins->op.minmax == BI_MINMAX_MIN) ? BIFROST_FMA_OP_FMIN16 :
587 BIFROST_FMA_OP_FMAX16;
588
589 /* Absolute values are packed in a quirky way. Let k = src1 < src0. Let
590 * l be an auxiliary bit we encode. Then the hardware determines:
591 *
592 * abs0 = l || k
593 * abs1 = l && k
594 *
595 * Since add/min/max are commutative, this saves a bit by using the
596 * order of the operands as a bit (k). To pack this, first note:
597 *
598 * (l && k) implies (l || k).
599 *
600 * That is, if the second argument is abs'd, then the first argument
601 * also has abs. So there are three cases:
602 *
603 * Case 0: Neither src has absolute value. Then we have l = k = 0.
604 *
605 * Case 1: Exactly one src has absolute value. Assign that source to
606 * src0 and the other source to src1. Compute k = src1 < src0 based on
607 * that assignment. Then l = ~k.
608 *
609 * Case 2: Both sources have absolute value. Then we have l = k = 1.
610 * Note to force k = 1 requires that (src1 < src0) OR (src0 < src1).
611 * That is, this encoding is only valid if src1 and src0 are distinct.
612 * This is a scheduling restriction (XXX); if an op of this type
613 * requires both identical sources to have abs value, then we must
614 * schedule to ADD (which does not use this ordering trick).
615 */
616
617 unsigned abs_0 = ins->src_abs[0], abs_1 = ins->src_abs[1];
618 unsigned src_0 = bi_get_src(ins, regs, 0, true);
619 unsigned src_1 = bi_get_src(ins, regs, 1, true);
620 bool l = false;
621 bool flip = false;
622
623 assert(!(abs_0 && abs_1));
624
625 if (!abs_0 && !abs_1) {
626 /* Force k = 0 <===> NOT(src1 < src0) */
627 flip = (src_1 < src_0);
628 } else if (abs_0 && !abs_1) {
629 l = src_1 >= src_0;
630 } else if (abs_1 && !abs_0) {
631 flip = true;
632 l = src_0 >= src_1;
633 } else {
634 flip = (src_0 >= src_1);
635 l = true;
636 }
637
638 if (FMA) {
639 struct bifrost_fma_add_minmax16 pack = {
640 .src0 = flip ? src_1 : src_0,
641 .src1 = flip ? src_0 : src_1,
642 .src0_neg = ins->src_neg[flip ? 1 : 0],
643 .src1_neg = ins->src_neg[flip ? 0 : 1],
644 .abs1 = l,
645 .outmod = ins->outmod,
646 .mode = (ins->type == BI_ADD) ? ins->roundmode : ins->minmax,
647 .op = op
648 };
649
650 RETURN_PACKED(pack);
651 } else {
652 /* Can't have modes for fp16 */
653 assert(ins->outmod == 0);
654
655 struct bifrost_add_fmin16 pack = {
656 .src0 = flip ? src_1 : src_0,
657 .src1 = flip ? src_0 : src_1,
658 .src0_neg = ins->src_neg[flip ? 1 : 0],
659 .src1_neg = ins->src_neg[flip ? 0 : 1],
660 .abs1 = l,
661 .src0_swizzle = bi_swiz16(ins, 0),
662 .src1_swizzle = bi_swiz16(ins, 1),
663 .mode = ins->minmax,
664 .op = op
665 };
666
667 RETURN_PACKED(pack);
668 }
669 }
670
671 static unsigned
672 bi_pack_fma_addmin(bi_instruction *ins, struct bi_registers *regs)
673 {
674 if (ins->dest_type == nir_type_float32)
675 return bi_pack_fma_addmin_f32(ins, regs);
676 else if(ins->dest_type == nir_type_float16)
677 return bi_pack_fmadd_min_f16(ins, regs, true);
678 else
679 unreachable("Unknown FMA/ADD type");
680 }
681
682 static unsigned
683 bi_pack_fma_1src(bi_instruction *ins, struct bi_registers *regs, unsigned op)
684 {
685 struct bifrost_fma_inst pack = {
686 .src0 = bi_get_src(ins, regs, 0, true),
687 .op = op
688 };
689
690 RETURN_PACKED(pack);
691 }
692
693 static unsigned
694 bi_pack_fma_2src(bi_instruction *ins, struct bi_registers *regs, unsigned op)
695 {
696 struct bifrost_fma_2src pack = {
697 .src0 = bi_get_src(ins, regs, 0, true),
698 .src1 = bi_get_src(ins, regs, 1, true),
699 .op = op
700 };
701
702 RETURN_PACKED(pack);
703 }
704
705 static unsigned
706 bi_pack_add_1src(bi_instruction *ins, struct bi_registers *regs, unsigned op)
707 {
708 struct bifrost_add_inst pack = {
709 .src0 = bi_get_src(ins, regs, 0, true),
710 .op = op
711 };
712
713 RETURN_PACKED(pack);
714 }
715
716 static enum bifrost_csel_cond
717 bi_cond_to_csel(enum bi_cond cond, bool *flip, bool *invert, nir_alu_type T)
718 {
719 nir_alu_type B = nir_alu_type_get_base_type(T);
720 unsigned idx = (B == nir_type_float) ? 0 :
721 ((B == nir_type_int) ? 1 : 2);
722
723 switch (cond){
724 case BI_COND_LT:
725 *flip = true;
726 case BI_COND_GT: {
727 const enum bifrost_csel_cond ops[] = {
728 BIFROST_FGT_F,
729 BIFROST_IGT_I,
730 BIFROST_UGT_I
731 };
732
733 return ops[idx];
734 }
735 case BI_COND_LE:
736 *flip = true;
737 case BI_COND_GE: {
738 const enum bifrost_csel_cond ops[] = {
739 BIFROST_FGE_F,
740 BIFROST_IGE_I,
741 BIFROST_UGE_I
742 };
743
744 return ops[idx];
745 }
746 case BI_COND_NE:
747 *invert = true;
748 case BI_COND_EQ: {
749 const enum bifrost_csel_cond ops[] = {
750 BIFROST_FEQ_F,
751 BIFROST_IEQ_F,
752 BIFROST_IEQ_F /* sign is irrelevant */
753 };
754
755 return ops[idx];
756 }
757 default:
758 unreachable("Invalid op for csel");
759 }
760 }
761
762 static unsigned
763 bi_pack_fma_csel(bi_instruction *ins, struct bi_registers *regs)
764 {
765 /* TODO: Use csel3 as well */
766 bool flip = false, invert = false;
767
768 enum bifrost_csel_cond cond =
769 bi_cond_to_csel(ins->csel_cond, &flip, &invert, ins->src_types[0]);
770
771 unsigned size = nir_alu_type_get_type_size(ins->dest_type);
772
773 unsigned cmp_0 = (flip ? 1 : 0);
774 unsigned cmp_1 = (flip ? 0 : 1);
775 unsigned res_0 = (invert ? 3 : 2);
776 unsigned res_1 = (invert ? 2 : 3);
777
778 struct bifrost_csel4 pack = {
779 .src0 = bi_get_src(ins, regs, cmp_0, true),
780 .src1 = bi_get_src(ins, regs, cmp_1, true),
781 .src2 = bi_get_src(ins, regs, res_0, true),
782 .src3 = bi_get_src(ins, regs, res_1, true),
783 .cond = cond,
784 .op = (size == 16) ? BIFROST_FMA_OP_CSEL4_V16 :
785 BIFROST_FMA_OP_CSEL4
786 };
787
788 RETURN_PACKED(pack);
789 }
790
791 static unsigned
792 bi_pack_fma_frexp(bi_instruction *ins, struct bi_registers *regs)
793 {
794 unsigned op = BIFROST_FMA_OP_FREXPE_LOG;
795 return bi_pack_fma_1src(ins, regs, op);
796 }
797
798 static unsigned
799 bi_pack_fma_reduce(bi_instruction *ins, struct bi_registers *regs)
800 {
801 if (ins->op.reduce == BI_REDUCE_ADD_FREXPM) {
802 return bi_pack_fma_2src(ins, regs, BIFROST_FMA_OP_ADD_FREXPM);
803 } else {
804 unreachable("Invalid reduce op");
805 }
806 }
807
808 /* We have a single convert opcode in the IR but a number of opcodes that could
809 * come out. In particular we have native opcodes for:
810 *
811 * [ui]16 --> [fui]32 -- int16_to_32
812 * f16 --> f32 -- float16_to_32
813 * f32 --> f16 -- float32_to_16
814 * f32 --> [ui]32 -- float32_to_int
815 * [ui]32 --> f32 -- int_to_float32
816 * [fui]16 --> [fui]16 -- f2i_i2f16
817 */
818
819 static unsigned
820 bi_pack_convert(bi_instruction *ins, struct bi_registers *regs, bool FMA)
821 {
822 nir_alu_type from_base = nir_alu_type_get_base_type(ins->src_types[0]);
823 unsigned from_size = nir_alu_type_get_type_size(ins->src_types[0]);
824 bool from_unsigned = from_base == nir_type_uint;
825
826 nir_alu_type to_base = nir_alu_type_get_base_type(ins->dest_type);
827 unsigned to_size = nir_alu_type_get_type_size(ins->dest_type);
828 bool to_unsigned = to_base == nir_type_uint;
829 bool to_float = to_base == nir_type_float;
830
831 /* Sanity check */
832 assert((from_base != to_base) || (from_size != to_size));
833 assert((MAX2(from_size, to_size) / MIN2(from_size, to_size)) <= 2);
834
835 /* f32 to f16 is special */
836 if (from_size == 32 && to_size == 16 && from_base == nir_type_float && to_base == from_base) {
837 /* TODO: second vectorized source? */
838 struct bifrost_fma_2src pfma = {
839 .src0 = bi_get_src(ins, regs, 0, true),
840 .src1 = BIFROST_SRC_STAGE, /* 0 */
841 .op = BIFROST_FMA_FLOAT32_TO_16
842 };
843
844 struct bifrost_add_2src padd = {
845 .src0 = bi_get_src(ins, regs, 0, true),
846 .src1 = BIFROST_SRC_STAGE, /* 0 */
847 .op = BIFROST_ADD_FLOAT32_TO_16
848 };
849
850 if (FMA) {
851 RETURN_PACKED(pfma);
852 } else {
853 RETURN_PACKED(padd);
854 }
855 }
856
857 /* Otherwise, figure out the mode */
858 unsigned op = 0;
859
860 if (from_size == 16 && to_size == 32) {
861 unsigned component = ins->swizzle[0][0];
862 assert(component <= 1);
863
864 if (from_base == nir_type_float)
865 op = BIFROST_CONVERT_5(component);
866 else
867 op = BIFROST_CONVERT_4(from_unsigned, component, to_float);
868 } else {
869 unsigned mode = 0;
870 unsigned swizzle = (from_size == 16) ? bi_swiz16(ins, 0) : 0;
871 bool is_unsigned = from_unsigned;
872
873 if (from_base == nir_type_float) {
874 assert(to_base != nir_type_float);
875 is_unsigned = to_unsigned;
876
877 if (from_size == 32 && to_size == 32)
878 mode = BIFROST_CONV_F32_TO_I32;
879 else if (from_size == 16 && to_size == 16)
880 mode = BIFROST_CONV_F16_TO_I16;
881 else
882 unreachable("Invalid float conversion");
883 } else {
884 assert(to_base == nir_type_float);
885 assert(from_size == to_size);
886
887 if (to_size == 32)
888 mode = BIFROST_CONV_I32_TO_F32;
889 else if (to_size == 16)
890 mode = BIFROST_CONV_I16_TO_F16;
891 else
892 unreachable("Invalid int conversion");
893 }
894
895 /* Fixup swizzle for 32-bit only modes */
896
897 if (mode == BIFROST_CONV_I32_TO_F32)
898 swizzle = 0b11;
899 else if (mode == BIFROST_CONV_F32_TO_I32)
900 swizzle = 0b10;
901
902 op = BIFROST_CONVERT(is_unsigned, ins->roundmode, swizzle, mode);
903
904 /* Unclear what the top bit is for... maybe 16-bit related */
905 bool mode2 = mode == BIFROST_CONV_F16_TO_I16;
906 bool mode6 = mode == BIFROST_CONV_I16_TO_F16;
907
908 if (!(mode2 || mode6))
909 op |= 0x100;
910 }
911
912 if (FMA)
913 return bi_pack_fma_1src(ins, regs, BIFROST_FMA_CONVERT | op);
914 else
915 return bi_pack_add_1src(ins, regs, BIFROST_ADD_CONVERT | op);
916 }
917
918 static unsigned
919 bi_pack_fma(bi_clause *clause, bi_bundle bundle, struct bi_registers *regs)
920 {
921 if (!bundle.fma)
922 return BIFROST_FMA_NOP;
923
924 switch (bundle.fma->type) {
925 case BI_ADD:
926 return bi_pack_fma_addmin(bundle.fma, regs);
927 case BI_CMP:
928 case BI_BITWISE:
929 return BIFROST_FMA_NOP;
930 case BI_CONVERT:
931 return bi_pack_convert(bundle.fma, regs, true);
932 case BI_CSEL:
933 return bi_pack_fma_csel(bundle.fma, regs);
934 case BI_FMA:
935 return bi_pack_fma_fma(bundle.fma, regs);
936 case BI_FREXP:
937 return bi_pack_fma_frexp(bundle.fma, regs);
938 case BI_ISUB:
939 return BIFROST_FMA_NOP;
940 case BI_MINMAX:
941 return bi_pack_fma_addmin(bundle.fma, regs);
942 case BI_MOV:
943 return bi_pack_fma_1src(bundle.fma, regs, BIFROST_FMA_OP_MOV);
944 case BI_SHIFT:
945 case BI_SWIZZLE:
946 case BI_ROUND:
947 return BIFROST_FMA_NOP;
948 case BI_REDUCE_FMA:
949 return bi_pack_fma_reduce(bundle.fma, regs);
950 default:
951 unreachable("Cannot encode class as FMA");
952 }
953 }
954
955 static unsigned
956 bi_pack_add_ld_vary(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
957 {
958 unsigned size = nir_alu_type_get_type_size(ins->dest_type);
959 assert(size == 32 || size == 16);
960
961 unsigned op = (size == 32) ?
962 BIFROST_ADD_OP_LD_VAR_32 :
963 BIFROST_ADD_OP_LD_VAR_16;
964
965 unsigned cmask = bi_from_bytemask(ins->writemask, size / 8);
966 unsigned channels = util_bitcount(cmask);
967 assert(cmask == ((1 << channels) - 1));
968
969 unsigned packed_addr = 0;
970
971 if (ins->src[0] & BIR_INDEX_CONSTANT) {
972 /* Direct uses address field directly */
973 packed_addr = bi_get_immediate(ins, 0);
974 assert(packed_addr < 0b1000);
975 } else {
976 /* Indirect gets an extra source */
977 packed_addr = bi_get_src(ins, regs, 0, false) | 0b11000;
978 }
979
980 /* The destination is thrown in the data register */
981 assert(ins->dest & BIR_INDEX_REGISTER);
982 clause->data_register = ins->dest & ~BIR_INDEX_REGISTER;
983
984 assert(channels >= 1 && channels <= 4);
985
986 struct bifrost_ld_var pack = {
987 .src0 = bi_get_src(ins, regs, 1, false),
988 .addr = packed_addr,
989 .channels = MALI_POSITIVE(channels),
990 .interp_mode = ins->load_vary.interp_mode,
991 .reuse = ins->load_vary.reuse,
992 .flat = ins->load_vary.flat,
993 .op = op
994 };
995
996 RETURN_PACKED(pack);
997 }
998
999 static unsigned
1000 bi_pack_add_2src(bi_instruction *ins, struct bi_registers *regs, unsigned op)
1001 {
1002 struct bifrost_add_2src pack = {
1003 .src0 = bi_get_src(ins, regs, 0, true),
1004 .src1 = bi_get_src(ins, regs, 1, true),
1005 .op = op
1006 };
1007
1008 RETURN_PACKED(pack);
1009 }
1010
1011 static unsigned
1012 bi_pack_add_addmin_f32(bi_instruction *ins, struct bi_registers *regs)
1013 {
1014 unsigned op =
1015 (ins->type == BI_ADD) ? BIFROST_ADD_OP_FADD32 :
1016 (ins->op.minmax == BI_MINMAX_MIN) ? BIFROST_ADD_OP_FMIN32 :
1017 BIFROST_ADD_OP_FMAX32;
1018
1019 struct bifrost_add_faddmin pack = {
1020 .src0 = bi_get_src(ins, regs, 0, true),
1021 .src1 = bi_get_src(ins, regs, 1, true),
1022 .src0_abs = ins->src_abs[0],
1023 .src1_abs = ins->src_abs[1],
1024 .src0_neg = ins->src_neg[0],
1025 .src1_neg = ins->src_neg[1],
1026 .outmod = ins->outmod,
1027 .mode = (ins->type == BI_ADD) ? ins->roundmode : ins->minmax,
1028 .op = op
1029 };
1030
1031 RETURN_PACKED(pack);
1032 }
1033
1034 static unsigned
1035 bi_pack_add_add_f16(bi_instruction *ins, struct bi_registers *regs)
1036 {
1037 /* ADD.v2f16 can't have outmod */
1038 assert(ins->outmod == BIFROST_NONE);
1039
1040 struct bifrost_add_faddmin pack = {
1041 .src0 = bi_get_src(ins, regs, 0, true),
1042 .src1 = bi_get_src(ins, regs, 1, true),
1043 .src0_abs = ins->src_abs[0],
1044 .src1_abs = ins->src_abs[1],
1045 .src0_neg = ins->src_neg[0],
1046 .src1_neg = ins->src_neg[1],
1047 .select = bi_swiz16(ins, 0), /* swizzle_0 */
1048 .outmod = bi_swiz16(ins, 1), /* swizzle_1 */
1049 .mode = ins->roundmode,
1050 .op = BIFROST_ADD_OP_FADD16
1051 };
1052
1053 RETURN_PACKED(pack);
1054 }
1055
1056 static unsigned
1057 bi_pack_add_addmin(bi_instruction *ins, struct bi_registers *regs)
1058 {
1059 if (ins->dest_type == nir_type_float32)
1060 return bi_pack_add_addmin_f32(ins, regs);
1061 else if (ins->dest_type == nir_type_float16) {
1062 if (ins->type == BI_ADD)
1063 return bi_pack_add_add_f16(ins, regs);
1064 else
1065 return bi_pack_fmadd_min_f16(ins, regs, false);
1066 } else
1067 unreachable("Unknown FMA/ADD type");
1068 }
1069
1070 static unsigned
1071 bi_pack_add_ld_ubo(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
1072 {
1073 unsigned components = bi_load32_components(ins);
1074
1075 const unsigned ops[4] = {
1076 BIFROST_ADD_OP_LD_UBO_1,
1077 BIFROST_ADD_OP_LD_UBO_2,
1078 BIFROST_ADD_OP_LD_UBO_3,
1079 BIFROST_ADD_OP_LD_UBO_4
1080 };
1081
1082 bi_write_data_register(clause, ins);
1083 return bi_pack_add_2src(ins, regs, ops[components - 1]);
1084 }
1085
1086 static enum bifrost_ldst_type
1087 bi_pack_ldst_type(nir_alu_type T)
1088 {
1089 switch (T) {
1090 case nir_type_float16: return BIFROST_LDST_F16;
1091 case nir_type_float32: return BIFROST_LDST_F32;
1092 case nir_type_int32: return BIFROST_LDST_I32;
1093 case nir_type_uint32: return BIFROST_LDST_U32;
1094 default: unreachable("Invalid type loaded");
1095 }
1096 }
1097
1098 static unsigned
1099 bi_pack_add_ld_var_addr(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
1100 {
1101 struct bifrost_ld_var_addr pack = {
1102 .src0 = bi_get_src(ins, regs, 1, false),
1103 .src1 = bi_get_src(ins, regs, 2, false),
1104 .location = bi_get_immediate(ins, 0),
1105 .type = bi_pack_ldst_type(ins->src_types[3]),
1106 .op = BIFROST_ADD_OP_LD_VAR_ADDR
1107 };
1108
1109 bi_write_data_register(clause, ins);
1110 RETURN_PACKED(pack);
1111 }
1112
1113 static unsigned
1114 bi_pack_add_ld_attr(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
1115 {
1116 struct bifrost_ld_attr pack = {
1117 .src0 = bi_get_src(ins, regs, 1, false),
1118 .src1 = bi_get_src(ins, regs, 2, false),
1119 .location = bi_get_immediate(ins, 0),
1120 .channels = MALI_POSITIVE(bi_load32_components(ins)),
1121 .type = bi_pack_ldst_type(ins->dest_type),
1122 .op = BIFROST_ADD_OP_LD_ATTR
1123 };
1124
1125 bi_write_data_register(clause, ins);
1126 RETURN_PACKED(pack);
1127 }
1128
1129 static unsigned
1130 bi_pack_add_st_vary(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
1131 {
1132 assert(ins->store_channels >= 1 && ins->store_channels <= 4);
1133
1134 struct bifrost_st_vary pack = {
1135 .src0 = bi_get_src(ins, regs, 1, false),
1136 .src1 = bi_get_src(ins, regs, 2, false),
1137 .src2 = bi_get_src(ins, regs, 3, false),
1138 .channels = MALI_POSITIVE(ins->store_channels),
1139 .op = BIFROST_ADD_OP_ST_VAR
1140 };
1141
1142 bi_read_data_register(clause, ins);
1143 RETURN_PACKED(pack);
1144 }
1145
1146 static unsigned
1147 bi_pack_add_atest(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
1148 {
1149 bool fp16 = (ins->src_types[1] == nir_type_float16);
1150
1151 struct bifrost_add_atest pack = {
1152 .src0 = bi_get_src(ins, regs, 0, false),
1153 .src1 = bi_get_src(ins, regs, 1, false),
1154 .half = fp16,
1155 .component = fp16 ? ins->swizzle[1][0] : 1, /* Set for fp32 */
1156 .op = BIFROST_ADD_OP_ATEST,
1157 };
1158
1159 /* Despite *also* writing with the usual mechanism... quirky and
1160 * perhaps unnecessary, but let's match the blob */
1161 clause->data_register = ins->dest & ~BIR_INDEX_REGISTER;
1162
1163 RETURN_PACKED(pack);
1164 }
1165
1166 static unsigned
1167 bi_pack_add_blend(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
1168 {
1169 struct bifrost_add_inst pack = {
1170 .src0 = bi_get_src(ins, regs, 1, false),
1171 .op = BIFROST_ADD_OP_BLEND
1172 };
1173
1174 /* TODO: Pack location in uniform_const */
1175 assert(ins->blend_location == 0);
1176
1177 bi_read_data_register(clause, ins);
1178 RETURN_PACKED(pack);
1179 }
1180
1181 static unsigned
1182 bi_pack_add_special(bi_instruction *ins, struct bi_registers *regs)
1183 {
1184 unsigned op = 0;
1185 bool fp16 = ins->dest_type == nir_type_float16;
1186 bool Y = ins->swizzle[0][0];
1187
1188 if (ins->op.special == BI_SPECIAL_FRCP) {
1189 op = fp16 ?
1190 (Y ? BIFROST_ADD_OP_FRCP_FAST_F16_Y :
1191 BIFROST_ADD_OP_FRCP_FAST_F16_X) :
1192 BIFROST_ADD_OP_FRCP_FAST_F32;
1193 } else if (ins->op.special == BI_SPECIAL_FRSQ) {
1194 op = fp16 ?
1195 (Y ? BIFROST_ADD_OP_FRSQ_FAST_F16_Y :
1196 BIFROST_ADD_OP_FRSQ_FAST_F16_X) :
1197 BIFROST_ADD_OP_FRSQ_FAST_F32;
1198
1199 } else if (ins->op.special == BI_SPECIAL_EXP2_LOW) {
1200 assert(!fp16);
1201 op = BIFROST_ADD_OP_FEXP2_FAST;
1202 } else {
1203 unreachable("Unknown special op");
1204 }
1205
1206 return bi_pack_add_1src(ins, regs, op);
1207 }
1208
1209 static unsigned
1210 bi_pack_add_table(bi_instruction *ins, struct bi_registers *regs)
1211 {
1212 unsigned op = 0;
1213 assert(ins->dest_type == nir_type_float32);
1214
1215 op = BIFROST_ADD_OP_LOG2_HELP;
1216 return bi_pack_add_1src(ins, regs, op);
1217 }
1218 static unsigned
1219 bi_pack_add_tex_compact(bi_clause *clause, bi_instruction *ins, struct bi_registers *regs)
1220 {
1221 bool f16 = ins->dest_type == nir_type_float16;
1222
1223 struct bifrost_tex_compact pack = {
1224 .src0 = bi_get_src(ins, regs, 0, false),
1225 .src1 = bi_get_src(ins, regs, 1, false),
1226 .op = f16 ? BIFROST_ADD_OP_TEX_COMPACT_F16 :
1227 BIFROST_ADD_OP_TEX_COMPACT_F32,
1228 .unknown = 1,
1229 .tex_index = 0,
1230 .sampler_index = 0
1231 };
1232
1233 bi_write_data_register(clause, ins);
1234 RETURN_PACKED(pack);
1235 }
1236
1237 static unsigned
1238 bi_pack_add(bi_clause *clause, bi_bundle bundle, struct bi_registers *regs)
1239 {
1240 if (!bundle.add)
1241 return BIFROST_ADD_NOP;
1242
1243 switch (bundle.add->type) {
1244 case BI_ADD:
1245 return bi_pack_add_addmin(bundle.add, regs);
1246 case BI_ATEST:
1247 return bi_pack_add_atest(clause, bundle.add, regs);
1248 case BI_BRANCH:
1249 case BI_CMP:
1250 return BIFROST_ADD_NOP;
1251 case BI_BLEND:
1252 return bi_pack_add_blend(clause, bundle.add, regs);
1253 case BI_BITWISE:
1254 return BIFROST_ADD_NOP;
1255 case BI_CONVERT:
1256 return bi_pack_convert(bundle.add, regs, false);
1257 case BI_DISCARD:
1258 case BI_FREXP:
1259 case BI_ISUB:
1260 case BI_LOAD:
1261 return BIFROST_ADD_NOP;
1262 case BI_LOAD_ATTR:
1263 return bi_pack_add_ld_attr(clause, bundle.add, regs);
1264 case BI_LOAD_UNIFORM:
1265 return bi_pack_add_ld_ubo(clause, bundle.add, regs);
1266 case BI_LOAD_VAR:
1267 return bi_pack_add_ld_vary(clause, bundle.add, regs);
1268 case BI_LOAD_VAR_ADDRESS:
1269 return bi_pack_add_ld_var_addr(clause, bundle.add, regs);
1270 case BI_MINMAX:
1271 return bi_pack_add_addmin(bundle.add, regs);
1272 case BI_MOV:
1273 case BI_SHIFT:
1274 case BI_STORE:
1275 return BIFROST_ADD_NOP;
1276 case BI_STORE_VAR:
1277 return bi_pack_add_st_vary(clause, bundle.add, regs);
1278 case BI_SPECIAL:
1279 return bi_pack_add_special(bundle.add, regs);
1280 case BI_TABLE:
1281 return bi_pack_add_table(bundle.add, regs);
1282 case BI_SWIZZLE:
1283 return BIFROST_ADD_NOP;
1284 case BI_TEX:
1285 if (bundle.add->op.texture == BI_TEX_COMPACT)
1286 return bi_pack_add_tex_compact(clause, bundle.add, regs);
1287 else
1288 unreachable("Unknown tex type");
1289 case BI_ROUND:
1290 return BIFROST_ADD_NOP;
1291 default:
1292 unreachable("Cannot encode class as ADD");
1293 }
1294 }
1295
1296 struct bi_packed_bundle {
1297 uint64_t lo;
1298 uint64_t hi;
1299 };
1300
1301 static struct bi_packed_bundle
1302 bi_pack_bundle(bi_clause *clause, bi_bundle bundle, bi_bundle prev, bool first_bundle)
1303 {
1304 struct bi_registers regs = bi_assign_ports(bundle, prev);
1305 bi_assign_uniform_constant(clause, &regs, bundle);
1306 regs.first_instruction = first_bundle;
1307
1308 uint64_t reg = bi_pack_registers(regs);
1309 uint64_t fma = bi_pack_fma(clause, bundle, &regs);
1310 uint64_t add = bi_pack_add(clause, bundle, &regs);
1311
1312 struct bi_packed_bundle packed = {
1313 .lo = reg | (fma << 35) | ((add & 0b111111) << 58),
1314 .hi = add >> 6
1315 };
1316
1317 return packed;
1318 }
1319
1320 /* Packs the next two constants as a dedicated constant quadword at the end of
1321 * the clause, returning the number packed. */
1322
1323 static unsigned
1324 bi_pack_constants(bi_context *ctx, bi_clause *clause,
1325 unsigned index,
1326 struct util_dynarray *emission)
1327 {
1328 /* After these two, are we done? Determines tag */
1329 bool done = clause->constant_count <= (index + 2);
1330 bool only = clause->constant_count <= (index + 1);
1331
1332 /* TODO: Pos */
1333 assert(index == 0 && clause->bundle_count == 1);
1334 assert(only);
1335
1336 uint64_t hi = clause->constants[index + 0] >> 60ull;
1337
1338 struct bifrost_fmt_constant quad = {
1339 .pos = 0, /* TODO */
1340 .tag = done ? BIFROST_FMTC_FINAL : BIFROST_FMTC_CONSTANTS,
1341 .imm_1 = clause->constants[index + 0] >> 4,
1342 .imm_2 = ((hi < 8) ? (hi << 60ull) : 0) >> 4,
1343 };
1344
1345 /* XXX: On G71, Connor observed that the difference of the top 4 bits
1346 * of the second constant with the first must be less than 8, otherwise
1347 * we have to swap them. On G52, I'm able to reproduce a similar issue
1348 * but with a different workaround (modeled above with a single
1349 * constant, unclear how to workaround for multiple constants.) Further
1350 * investigation needed. Possibly an errata. XXX */
1351
1352 util_dynarray_append(emission, struct bifrost_fmt_constant, quad);
1353
1354 return 2;
1355 }
1356
1357 static void
1358 bi_pack_clause(bi_context *ctx, bi_clause *clause, bi_clause *next,
1359 struct util_dynarray *emission)
1360 {
1361 struct bi_packed_bundle ins_1 = bi_pack_bundle(clause, clause->bundles[0], clause->bundles[0], true);
1362 assert(clause->bundle_count == 1);
1363
1364 /* Used to decide if we elide writes */
1365 bool is_fragment = ctx->stage == MESA_SHADER_FRAGMENT;
1366
1367 /* State for packing constants throughout */
1368 unsigned constant_index = 0;
1369
1370 struct bifrost_fmt1 quad_1 = {
1371 .tag = clause->constant_count ? BIFROST_FMT1_CONSTANTS : BIFROST_FMT1_FINAL,
1372 .header = bi_pack_header(clause, next, is_fragment),
1373 .ins_1 = ins_1.lo,
1374 .ins_2 = ins_1.hi & ((1 << 11) - 1),
1375 .ins_0 = (ins_1.hi >> 11) & 0b111,
1376 };
1377
1378 util_dynarray_append(emission, struct bifrost_fmt1, quad_1);
1379
1380 /* Pack the remaining constants */
1381
1382 while (constant_index < clause->constant_count) {
1383 constant_index += bi_pack_constants(ctx, clause,
1384 constant_index, emission);
1385 }
1386 }
1387
1388 static bi_clause *
1389 bi_next_clause(bi_context *ctx, pan_block *block, bi_clause *clause)
1390 {
1391 /* Try the next clause in this block */
1392 if (clause->link.next != &((bi_block *) block)->clauses)
1393 return list_first_entry(&(clause->link), bi_clause, link);
1394
1395 /* Try the next block, or the one after that if it's empty, etc .*/
1396 pan_block *next_block = pan_next_block(block);
1397
1398 bi_foreach_block_from(ctx, next_block, block) {
1399 bi_block *blk = (bi_block *) block;
1400
1401 if (!list_is_empty(&blk->clauses))
1402 return list_first_entry(&(blk->clauses), bi_clause, link);
1403 }
1404
1405 return NULL;
1406 }
1407
1408 void
1409 bi_pack(bi_context *ctx, struct util_dynarray *emission)
1410 {
1411 util_dynarray_init(emission, NULL);
1412
1413 bi_foreach_block(ctx, _block) {
1414 bi_block *block = (bi_block *) _block;
1415
1416 bi_foreach_clause_in_block(block, clause) {
1417 bi_clause *next = bi_next_clause(ctx, _block, clause);
1418 bi_pack_clause(ctx, clause, next, emission);
1419 }
1420 }
1421 }