vc4: Add a QIR helper for whether the op is a MUL type.
[mesa.git] / src / gallium / drivers / vc4 / vc4_qpu_emit.c
1 /*
2 * Copyright © 2014 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 #include <inttypes.h>
25
26 #include "vc4_context.h"
27 #include "vc4_qir.h"
28 #include "vc4_qpu.h"
29 #include "util/ralloc.h"
30
31 static void
32 vc4_dump_program(struct vc4_compile *c)
33 {
34 fprintf(stderr, "%s prog %d/%d QPU:\n",
35 qir_get_stage_name(c->stage),
36 c->program_id, c->variant_id);
37
38 for (int i = 0; i < c->qpu_inst_count; i++) {
39 fprintf(stderr, "0x%016"PRIx64" ", c->qpu_insts[i]);
40 vc4_qpu_disasm(&c->qpu_insts[i], 1);
41 fprintf(stderr, "\n");
42 }
43 }
44
45 static void
46 queue(struct vc4_compile *c, uint64_t inst)
47 {
48 struct queued_qpu_inst *q = rzalloc(c, struct queued_qpu_inst);
49 q->inst = inst;
50 list_addtail(&q->link, &c->qpu_inst_list);
51 }
52
53 static uint64_t *
54 last_inst(struct vc4_compile *c)
55 {
56 struct queued_qpu_inst *q =
57 (struct queued_qpu_inst *)c->qpu_inst_list.prev;
58 return &q->inst;
59 }
60
61 static void
62 set_last_cond_add(struct vc4_compile *c, uint32_t cond)
63 {
64 *last_inst(c) = qpu_set_cond_add(*last_inst(c), cond);
65 }
66
67 /**
68 * Some special registers can be read from either file, which lets us resolve
69 * raddr conflicts without extra MOVs.
70 */
71 static bool
72 swap_file(struct qpu_reg *src)
73 {
74 switch (src->addr) {
75 case QPU_R_UNIF:
76 case QPU_R_VARY:
77 if (src->mux == QPU_MUX_SMALL_IMM) {
78 return false;
79 } else {
80 if (src->mux == QPU_MUX_A)
81 src->mux = QPU_MUX_B;
82 else
83 src->mux = QPU_MUX_A;
84 return true;
85 }
86
87 default:
88 return false;
89 }
90 }
91
92 /**
93 * This is used to resolve the fact that we might register-allocate two
94 * different operands of an instruction to the same physical register file
95 * even though instructions have only one field for the register file source
96 * address.
97 *
98 * In that case, we need to move one to a temporary that can be used in the
99 * instruction, instead. We reserve ra31/rb31 for this purpose.
100 */
101 static void
102 fixup_raddr_conflict(struct vc4_compile *c,
103 struct qpu_reg dst,
104 struct qpu_reg *src0, struct qpu_reg *src1)
105 {
106 uint32_t mux0 = src0->mux == QPU_MUX_SMALL_IMM ? QPU_MUX_B : src0->mux;
107 uint32_t mux1 = src1->mux == QPU_MUX_SMALL_IMM ? QPU_MUX_B : src1->mux;
108
109 if (mux0 <= QPU_MUX_R5 ||
110 mux0 != mux1 ||
111 (src0->addr == src1->addr &&
112 src0->mux == src1->mux)) {
113 return;
114 }
115
116 if (swap_file(src0) || swap_file(src1))
117 return;
118
119 if (mux0 == QPU_MUX_A) {
120 queue(c, qpu_a_MOV(qpu_rb(31), *src0));
121 *src0 = qpu_rb(31);
122 } else {
123 queue(c, qpu_a_MOV(qpu_ra(31), *src0));
124 *src0 = qpu_ra(31);
125 }
126 }
127
128 void
129 vc4_generate_code(struct vc4_context *vc4, struct vc4_compile *c)
130 {
131 struct qpu_reg *temp_registers = vc4_register_allocate(vc4, c);
132 bool discard = false;
133 uint32_t inputs_remaining = c->num_inputs;
134 uint32_t vpm_read_fifo_count = 0;
135 uint32_t vpm_read_offset = 0;
136 int last_vpm_read_index = -1;
137 /* Map from the QIR ops enum order to QPU unpack bits. */
138 static const uint32_t unpack_map[] = {
139 QPU_UNPACK_8A,
140 QPU_UNPACK_8B,
141 QPU_UNPACK_8C,
142 QPU_UNPACK_8D,
143 QPU_UNPACK_16A_TO_F32,
144 QPU_UNPACK_16B_TO_F32,
145 };
146
147 list_inithead(&c->qpu_inst_list);
148
149 switch (c->stage) {
150 case QSTAGE_VERT:
151 case QSTAGE_COORD:
152 /* There's a 4-entry FIFO for VPMVCD reads, each of which can
153 * load up to 16 dwords (4 vec4s) per vertex.
154 */
155 while (inputs_remaining) {
156 uint32_t num_entries = MIN2(inputs_remaining, 16);
157 queue(c, qpu_load_imm_ui(qpu_vrsetup(),
158 vpm_read_offset |
159 0x00001a00 |
160 ((num_entries & 0xf) << 20)));
161 inputs_remaining -= num_entries;
162 vpm_read_offset += num_entries;
163 vpm_read_fifo_count++;
164 }
165 assert(vpm_read_fifo_count <= 4);
166
167 queue(c, qpu_load_imm_ui(qpu_vwsetup(), 0x00001a00));
168 break;
169 case QSTAGE_FRAG:
170 break;
171 }
172
173 list_for_each_entry(struct qinst, qinst, &c->instructions, link) {
174 #if 0
175 fprintf(stderr, "translating qinst to qpu: ");
176 qir_dump_inst(qinst);
177 fprintf(stderr, "\n");
178 #endif
179
180 static const struct {
181 uint32_t op;
182 } translate[] = {
183 #define A(name) [QOP_##name] = {QPU_A_##name}
184 #define M(name) [QOP_##name] = {QPU_M_##name}
185 A(FADD),
186 A(FSUB),
187 A(FMIN),
188 A(FMAX),
189 A(FMINABS),
190 A(FMAXABS),
191 A(FTOI),
192 A(ITOF),
193 A(ADD),
194 A(SUB),
195 A(SHL),
196 A(SHR),
197 A(ASR),
198 A(MIN),
199 A(MAX),
200 A(AND),
201 A(OR),
202 A(XOR),
203 A(NOT),
204
205 M(FMUL),
206 M(MUL24),
207 };
208
209 struct qpu_reg src[4];
210 for (int i = 0; i < qir_get_op_nsrc(qinst->op); i++) {
211 int index = qinst->src[i].index;
212 switch (qinst->src[i].file) {
213 case QFILE_NULL:
214 src[i] = qpu_rn(0);
215 break;
216 case QFILE_TEMP:
217 src[i] = temp_registers[index];
218 break;
219 case QFILE_UNIF:
220 src[i] = qpu_unif();
221 break;
222 case QFILE_VARY:
223 src[i] = qpu_vary();
224 break;
225 case QFILE_SMALL_IMM:
226 src[i].mux = QPU_MUX_SMALL_IMM;
227 src[i].addr = qpu_encode_small_immediate(qinst->src[i].index);
228 /* This should only have returned a valid
229 * small immediate field, not ~0 for failure.
230 */
231 assert(src[i].addr <= 47);
232 break;
233 case QFILE_VPM:
234 assert((int)qinst->src[i].index >=
235 last_vpm_read_index);
236 (void)last_vpm_read_index;
237 last_vpm_read_index = qinst->src[i].index;
238 src[i] = qpu_ra(QPU_R_VPM);
239 break;
240 }
241 }
242
243 struct qpu_reg dst;
244 switch (qinst->dst.file) {
245 case QFILE_NULL:
246 dst = qpu_ra(QPU_W_NOP);
247 break;
248 case QFILE_TEMP:
249 dst = temp_registers[qinst->dst.index];
250 break;
251 case QFILE_VPM:
252 dst = qpu_ra(QPU_W_VPM);
253 break;
254 case QFILE_VARY:
255 case QFILE_UNIF:
256 case QFILE_SMALL_IMM:
257 assert(!"not reached");
258 break;
259 }
260
261 switch (qinst->op) {
262 case QOP_MOV:
263 /* Skip emitting the MOV if it's a no-op. */
264 if (dst.mux == QPU_MUX_A || dst.mux == QPU_MUX_B ||
265 dst.mux != src[0].mux || dst.addr != src[0].addr) {
266 queue(c, qpu_a_MOV(dst, src[0]));
267 }
268 break;
269
270 case QOP_SEL_X_0_ZS:
271 case QOP_SEL_X_0_ZC:
272 case QOP_SEL_X_0_NS:
273 case QOP_SEL_X_0_NC:
274 queue(c, qpu_a_MOV(dst, src[0]));
275 set_last_cond_add(c, qinst->op - QOP_SEL_X_0_ZS +
276 QPU_COND_ZS);
277
278 queue(c, qpu_a_XOR(dst, qpu_r0(), qpu_r0()));
279 set_last_cond_add(c, ((qinst->op - QOP_SEL_X_0_ZS) ^
280 1) + QPU_COND_ZS);
281 break;
282
283 case QOP_SEL_X_Y_ZS:
284 case QOP_SEL_X_Y_ZC:
285 case QOP_SEL_X_Y_NS:
286 case QOP_SEL_X_Y_NC:
287 queue(c, qpu_a_MOV(dst, src[0]));
288 set_last_cond_add(c, qinst->op - QOP_SEL_X_Y_ZS +
289 QPU_COND_ZS);
290
291 queue(c, qpu_a_MOV(dst, src[1]));
292 set_last_cond_add(c, ((qinst->op - QOP_SEL_X_Y_ZS) ^
293 1) + QPU_COND_ZS);
294
295 break;
296
297 case QOP_RCP:
298 case QOP_RSQ:
299 case QOP_EXP2:
300 case QOP_LOG2:
301 switch (qinst->op) {
302 case QOP_RCP:
303 queue(c, qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIP),
304 src[0]));
305 break;
306 case QOP_RSQ:
307 queue(c, qpu_a_MOV(qpu_rb(QPU_W_SFU_RECIPSQRT),
308 src[0]));
309 break;
310 case QOP_EXP2:
311 queue(c, qpu_a_MOV(qpu_rb(QPU_W_SFU_EXP),
312 src[0]));
313 break;
314 case QOP_LOG2:
315 queue(c, qpu_a_MOV(qpu_rb(QPU_W_SFU_LOG),
316 src[0]));
317 break;
318 default:
319 abort();
320 }
321
322 if (dst.mux != QPU_MUX_R4)
323 queue(c, qpu_a_MOV(dst, qpu_r4()));
324
325 break;
326
327 case QOP_PACK_8888_F:
328 queue(c, qpu_m_MOV(dst, src[0]));
329 *last_inst(c) |= QPU_PM;
330 *last_inst(c) |= QPU_SET_FIELD(QPU_PACK_MUL_8888,
331 QPU_PACK);
332 break;
333
334 case QOP_PACK_8A_F:
335 case QOP_PACK_8B_F:
336 case QOP_PACK_8C_F:
337 case QOP_PACK_8D_F:
338 queue(c,
339 qpu_m_MOV(dst, src[0]) |
340 QPU_PM |
341 QPU_SET_FIELD(QPU_PACK_MUL_8A +
342 qinst->op - QOP_PACK_8A_F,
343 QPU_PACK));
344 break;
345
346 case QOP_FRAG_X:
347 queue(c, qpu_a_ITOF(dst,
348 qpu_ra(QPU_R_XY_PIXEL_COORD)));
349 break;
350
351 case QOP_FRAG_Y:
352 queue(c, qpu_a_ITOF(dst,
353 qpu_rb(QPU_R_XY_PIXEL_COORD)));
354 break;
355
356 case QOP_FRAG_REV_FLAG:
357 queue(c, qpu_a_ITOF(dst,
358 qpu_rb(QPU_R_MS_REV_FLAGS)));
359 break;
360
361 case QOP_FRAG_Z:
362 case QOP_FRAG_W:
363 /* QOP_FRAG_Z/W don't emit instructions, just allocate
364 * the register to the Z/W payload.
365 */
366 break;
367
368 case QOP_TLB_DISCARD_SETUP:
369 discard = true;
370 queue(c, qpu_a_MOV(src[0], src[0]));
371 *last_inst(c) |= QPU_SF;
372 break;
373
374 case QOP_TLB_STENCIL_SETUP:
375 queue(c, qpu_a_MOV(qpu_ra(QPU_W_TLB_STENCIL_SETUP), src[0]));
376 break;
377
378 case QOP_TLB_Z_WRITE:
379 queue(c, qpu_a_MOV(qpu_ra(QPU_W_TLB_Z), src[0]));
380 if (discard) {
381 set_last_cond_add(c, QPU_COND_ZS);
382 }
383 break;
384
385 case QOP_TLB_COLOR_READ:
386 queue(c, qpu_NOP());
387 *last_inst(c) = qpu_set_sig(*last_inst(c),
388 QPU_SIG_COLOR_LOAD);
389
390 if (dst.mux != QPU_MUX_R4)
391 queue(c, qpu_a_MOV(dst, qpu_r4()));
392 break;
393
394 case QOP_TLB_COLOR_WRITE:
395 queue(c, qpu_a_MOV(qpu_tlbc(), src[0]));
396 if (discard) {
397 set_last_cond_add(c, QPU_COND_ZS);
398 }
399 break;
400
401 case QOP_VARY_ADD_C:
402 queue(c, qpu_a_FADD(dst, src[0], qpu_r5()));
403 break;
404
405 case QOP_PACK_16A_I:
406 case QOP_PACK_16B_I:
407 queue(c,
408 qpu_a_MOV(dst, src[0]) |
409 QPU_SET_FIELD(qinst->op == QOP_PACK_16A_I ?
410 QPU_PACK_A_16A : QPU_PACK_A_16B,
411 QPU_PACK));
412 break;
413
414 case QOP_TEX_S:
415 case QOP_TEX_T:
416 case QOP_TEX_R:
417 case QOP_TEX_B:
418 queue(c, qpu_a_MOV(qpu_rb(QPU_W_TMU0_S +
419 (qinst->op - QOP_TEX_S)),
420 src[0]));
421 break;
422
423 case QOP_TEX_DIRECT:
424 fixup_raddr_conflict(c, dst, &src[0], &src[1]);
425 queue(c, qpu_a_ADD(qpu_rb(QPU_W_TMU0_S), src[0], src[1]));
426 break;
427
428 case QOP_TEX_RESULT:
429 queue(c, qpu_NOP());
430 *last_inst(c) = qpu_set_sig(*last_inst(c),
431 QPU_SIG_LOAD_TMU0);
432 if (dst.mux != QPU_MUX_R4)
433 queue(c, qpu_a_MOV(dst, qpu_r4()));
434 break;
435
436 case QOP_UNPACK_8A_F:
437 case QOP_UNPACK_8B_F:
438 case QOP_UNPACK_8C_F:
439 case QOP_UNPACK_8D_F:
440 case QOP_UNPACK_16A_F:
441 case QOP_UNPACK_16B_F: {
442 if (src[0].mux == QPU_MUX_R4) {
443 queue(c, qpu_a_MOV(dst, src[0]));
444 *last_inst(c) |= QPU_PM;
445 *last_inst(c) |= QPU_SET_FIELD(QPU_UNPACK_8A +
446 (qinst->op -
447 QOP_UNPACK_8A_F),
448 QPU_UNPACK);
449 } else {
450 assert(src[0].mux == QPU_MUX_A);
451
452 /* Since we're setting the pack bits, if the
453 * destination is in A it would get re-packed.
454 */
455 queue(c, qpu_a_FMAX((dst.mux == QPU_MUX_A ?
456 qpu_rb(31) : dst),
457 src[0], src[0]));
458 *last_inst(c) |=
459 QPU_SET_FIELD(unpack_map[qinst->op -
460 QOP_UNPACK_8A_F],
461 QPU_UNPACK);
462
463 if (dst.mux == QPU_MUX_A) {
464 queue(c, qpu_a_MOV(dst, qpu_rb(31)));
465 }
466 }
467 }
468 break;
469
470 case QOP_UNPACK_8A_I:
471 case QOP_UNPACK_8B_I:
472 case QOP_UNPACK_8C_I:
473 case QOP_UNPACK_8D_I:
474 case QOP_UNPACK_16A_I:
475 case QOP_UNPACK_16B_I: {
476 assert(src[0].mux == QPU_MUX_A);
477
478 /* Since we're setting the pack bits, if the
479 * destination is in A it would get re-packed.
480 */
481 queue(c, qpu_a_MOV((dst.mux == QPU_MUX_A ?
482 qpu_rb(31) : dst), src[0]));
483 *last_inst(c) |= QPU_SET_FIELD(unpack_map[qinst->op -
484 QOP_UNPACK_8A_I],
485 QPU_UNPACK);
486
487 if (dst.mux == QPU_MUX_A) {
488 queue(c, qpu_a_MOV(dst, qpu_rb(31)));
489 }
490 }
491 break;
492
493 default:
494 assert(qinst->op < ARRAY_SIZE(translate));
495 assert(translate[qinst->op].op != 0); /* NOPs */
496
497 /* If we have only one source, put it in the second
498 * argument slot as well so that we don't take up
499 * another raddr just to get unused data.
500 */
501 if (qir_get_op_nsrc(qinst->op) == 1)
502 src[1] = src[0];
503
504 fixup_raddr_conflict(c, dst, &src[0], &src[1]);
505
506 if (qir_is_mul(qinst)) {
507 queue(c, qpu_m_alu2(translate[qinst->op].op,
508 dst,
509 src[0], src[1]));
510 } else {
511 queue(c, qpu_a_alu2(translate[qinst->op].op,
512 dst,
513 src[0], src[1]));
514 }
515
516 break;
517 }
518
519 if (qinst->sf) {
520 assert(!qir_is_multi_instruction(qinst));
521 *last_inst(c) |= QPU_SF;
522 }
523 }
524
525 qpu_schedule_instructions(c);
526
527 /* thread end can't have VPM write or read */
528 if (QPU_GET_FIELD(c->qpu_insts[c->qpu_inst_count - 1],
529 QPU_WADDR_ADD) == QPU_W_VPM ||
530 QPU_GET_FIELD(c->qpu_insts[c->qpu_inst_count - 1],
531 QPU_WADDR_MUL) == QPU_W_VPM ||
532 QPU_GET_FIELD(c->qpu_insts[c->qpu_inst_count - 1],
533 QPU_RADDR_A) == QPU_R_VPM ||
534 QPU_GET_FIELD(c->qpu_insts[c->qpu_inst_count - 1],
535 QPU_RADDR_B) == QPU_R_VPM) {
536 qpu_serialize_one_inst(c, qpu_NOP());
537 }
538
539 /* thread end can't have uniform read */
540 if (QPU_GET_FIELD(c->qpu_insts[c->qpu_inst_count - 1],
541 QPU_RADDR_A) == QPU_R_UNIF ||
542 QPU_GET_FIELD(c->qpu_insts[c->qpu_inst_count - 1],
543 QPU_RADDR_B) == QPU_R_UNIF) {
544 qpu_serialize_one_inst(c, qpu_NOP());
545 }
546
547 /* thread end can't have TLB operations */
548 if (qpu_inst_is_tlb(c->qpu_insts[c->qpu_inst_count - 1]))
549 qpu_serialize_one_inst(c, qpu_NOP());
550
551 c->qpu_insts[c->qpu_inst_count - 1] =
552 qpu_set_sig(c->qpu_insts[c->qpu_inst_count - 1],
553 QPU_SIG_PROG_END);
554 qpu_serialize_one_inst(c, qpu_NOP());
555 qpu_serialize_one_inst(c, qpu_NOP());
556
557 switch (c->stage) {
558 case QSTAGE_VERT:
559 case QSTAGE_COORD:
560 break;
561 case QSTAGE_FRAG:
562 c->qpu_insts[c->qpu_inst_count - 1] =
563 qpu_set_sig(c->qpu_insts[c->qpu_inst_count - 1],
564 QPU_SIG_SCOREBOARD_UNLOCK);
565 break;
566 }
567
568 if (vc4_debug & VC4_DEBUG_QPU)
569 vc4_dump_program(c);
570
571 vc4_qpu_validate(c->qpu_insts, c->qpu_inst_count);
572
573 free(temp_registers);
574 }