1 /*
2 * Copyright © 2016-2017 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 "broadcom/common/v3d_device_info.h"
25 #include "v3d_compiler.h"
26 #include "util/u_prim.h"
27 #include "compiler/nir/nir_schedule.h"
28
29 int
vir_get_nsrc(struct qinst * inst)30 vir_get_nsrc(struct qinst *inst)
31 {
32 switch (inst->qpu.type) {
33 case V3D_QPU_INSTR_TYPE_BRANCH:
34 return 0;
35 case V3D_QPU_INSTR_TYPE_ALU:
36 if (inst->qpu.alu.add.op != V3D_QPU_A_NOP)
37 return v3d_qpu_add_op_num_src(inst->qpu.alu.add.op);
38 else
39 return v3d_qpu_mul_op_num_src(inst->qpu.alu.mul.op);
40 }
41
42 return 0;
43 }
44
45 /**
46 * Returns whether the instruction has any side effects that must be
47 * preserved.
48 */
49 bool
vir_has_side_effects(struct v3d_compile * c,struct qinst * inst)50 vir_has_side_effects(struct v3d_compile *c, struct qinst *inst)
51 {
52 switch (inst->qpu.type) {
53 case V3D_QPU_INSTR_TYPE_BRANCH:
54 return true;
55 case V3D_QPU_INSTR_TYPE_ALU:
56 switch (inst->qpu.alu.add.op) {
57 case V3D_QPU_A_SETREVF:
58 case V3D_QPU_A_SETMSF:
59 case V3D_QPU_A_VPMSETUP:
60 case V3D_QPU_A_STVPMV:
61 case V3D_QPU_A_STVPMD:
62 case V3D_QPU_A_STVPMP:
63 case V3D_QPU_A_VPMWT:
64 case V3D_QPU_A_TMUWT:
65 return true;
66 default:
67 break;
68 }
69
70 switch (inst->qpu.alu.mul.op) {
71 case V3D_QPU_M_MULTOP:
72 return true;
73 default:
74 break;
75 }
76 }
77
78 if (inst->qpu.sig.ldtmu ||
79 inst->qpu.sig.ldvary ||
80 inst->qpu.sig.ldtlbu ||
81 inst->qpu.sig.ldtlb ||
82 inst->qpu.sig.wrtmuc ||
83 inst->qpu.sig.thrsw) {
84 return true;
85 }
86
87 return false;
88 }
89
90 bool
vir_is_raw_mov(struct qinst * inst)91 vir_is_raw_mov(struct qinst *inst)
92 {
93 if (inst->qpu.type != V3D_QPU_INSTR_TYPE_ALU ||
94 (inst->qpu.alu.mul.op != V3D_QPU_M_FMOV &&
95 inst->qpu.alu.mul.op != V3D_QPU_M_MOV)) {
96 return false;
97 }
98
99 if (inst->qpu.alu.add.output_pack != V3D_QPU_PACK_NONE ||
100 inst->qpu.alu.mul.output_pack != V3D_QPU_PACK_NONE) {
101 return false;
102 }
103
104 if (inst->qpu.alu.add.a_unpack != V3D_QPU_UNPACK_NONE ||
105 inst->qpu.alu.add.b_unpack != V3D_QPU_UNPACK_NONE ||
106 inst->qpu.alu.mul.a_unpack != V3D_QPU_UNPACK_NONE ||
107 inst->qpu.alu.mul.b_unpack != V3D_QPU_UNPACK_NONE) {
108 return false;
109 }
110
111 if (inst->qpu.flags.ac != V3D_QPU_COND_NONE ||
112 inst->qpu.flags.mc != V3D_QPU_COND_NONE)
113 return false;
114
115 return true;
116 }
117
118 bool
vir_is_add(struct qinst * inst)119 vir_is_add(struct qinst *inst)
120 {
121 return (inst->qpu.type == V3D_QPU_INSTR_TYPE_ALU &&
122 inst->qpu.alu.add.op != V3D_QPU_A_NOP);
123 }
124
125 bool
vir_is_mul(struct qinst * inst)126 vir_is_mul(struct qinst *inst)
127 {
128 return (inst->qpu.type == V3D_QPU_INSTR_TYPE_ALU &&
129 inst->qpu.alu.mul.op != V3D_QPU_M_NOP);
130 }
131
132 bool
vir_is_tex(struct qinst * inst)133 vir_is_tex(struct qinst *inst)
134 {
135 if (inst->dst.file == QFILE_MAGIC)
136 return v3d_qpu_magic_waddr_is_tmu(inst->dst.index);
137
138 if (inst->qpu.type == V3D_QPU_INSTR_TYPE_ALU &&
139 inst->qpu.alu.add.op == V3D_QPU_A_TMUWT) {
140 return true;
141 }
142
143 return false;
144 }
145
146 bool
vir_writes_r3(const struct v3d_device_info * devinfo,struct qinst * inst)147 vir_writes_r3(const struct v3d_device_info *devinfo, struct qinst *inst)
148 {
149 for (int i = 0; i < vir_get_nsrc(inst); i++) {
150 switch (inst->src[i].file) {
151 case QFILE_VPM:
152 return true;
153 default:
154 break;
155 }
156 }
157
158 if (devinfo->ver < 41 && (inst->qpu.sig.ldvary ||
159 inst->qpu.sig.ldtlb ||
160 inst->qpu.sig.ldtlbu ||
161 inst->qpu.sig.ldvpm)) {
162 return true;
163 }
164
165 return false;
166 }
167
168 bool
vir_writes_r4(const struct v3d_device_info * devinfo,struct qinst * inst)169 vir_writes_r4(const struct v3d_device_info *devinfo, struct qinst *inst)
170 {
171 switch (inst->dst.file) {
172 case QFILE_MAGIC:
173 switch (inst->dst.index) {
174 case V3D_QPU_WADDR_RECIP:
175 case V3D_QPU_WADDR_RSQRT:
176 case V3D_QPU_WADDR_EXP:
177 case V3D_QPU_WADDR_LOG:
178 case V3D_QPU_WADDR_SIN:
179 return true;
180 }
181 break;
182 default:
183 break;
184 }
185
186 if (devinfo->ver < 41 && inst->qpu.sig.ldtmu)
187 return true;
188
189 return false;
190 }
191
192 void
vir_set_unpack(struct qinst * inst,int src,enum v3d_qpu_input_unpack unpack)193 vir_set_unpack(struct qinst *inst, int src,
194 enum v3d_qpu_input_unpack unpack)
195 {
196 assert(src == 0 || src == 1);
197
198 if (vir_is_add(inst)) {
199 if (src == 0)
200 inst->qpu.alu.add.a_unpack = unpack;
201 else
202 inst->qpu.alu.add.b_unpack = unpack;
203 } else {
204 assert(vir_is_mul(inst));
205 if (src == 0)
206 inst->qpu.alu.mul.a_unpack = unpack;
207 else
208 inst->qpu.alu.mul.b_unpack = unpack;
209 }
210 }
211
212 void
vir_set_cond(struct qinst * inst,enum v3d_qpu_cond cond)213 vir_set_cond(struct qinst *inst, enum v3d_qpu_cond cond)
214 {
215 if (vir_is_add(inst)) {
216 inst->qpu.flags.ac = cond;
217 } else {
218 assert(vir_is_mul(inst));
219 inst->qpu.flags.mc = cond;
220 }
221 }
222
223 void
vir_set_pf(struct qinst * inst,enum v3d_qpu_pf pf)224 vir_set_pf(struct qinst *inst, enum v3d_qpu_pf pf)
225 {
226 if (vir_is_add(inst)) {
227 inst->qpu.flags.apf = pf;
228 } else {
229 assert(vir_is_mul(inst));
230 inst->qpu.flags.mpf = pf;
231 }
232 }
233
234 void
vir_set_uf(struct qinst * inst,enum v3d_qpu_uf uf)235 vir_set_uf(struct qinst *inst, enum v3d_qpu_uf uf)
236 {
237 if (vir_is_add(inst)) {
238 inst->qpu.flags.auf = uf;
239 } else {
240 assert(vir_is_mul(inst));
241 inst->qpu.flags.muf = uf;
242 }
243 }
244
245 #if 0
246 uint8_t
247 vir_channels_written(struct qinst *inst)
248 {
249 if (vir_is_mul(inst)) {
250 switch (inst->dst.pack) {
251 case QPU_PACK_MUL_NOP:
252 case QPU_PACK_MUL_8888:
253 return 0xf;
254 case QPU_PACK_MUL_8A:
255 return 0x1;
256 case QPU_PACK_MUL_8B:
257 return 0x2;
258 case QPU_PACK_MUL_8C:
259 return 0x4;
260 case QPU_PACK_MUL_8D:
261 return 0x8;
262 }
263 } else {
264 switch (inst->dst.pack) {
265 case QPU_PACK_A_NOP:
266 case QPU_PACK_A_8888:
267 case QPU_PACK_A_8888_SAT:
268 case QPU_PACK_A_32_SAT:
269 return 0xf;
270 case QPU_PACK_A_8A:
271 case QPU_PACK_A_8A_SAT:
272 return 0x1;
273 case QPU_PACK_A_8B:
274 case QPU_PACK_A_8B_SAT:
275 return 0x2;
276 case QPU_PACK_A_8C:
277 case QPU_PACK_A_8C_SAT:
278 return 0x4;
279 case QPU_PACK_A_8D:
280 case QPU_PACK_A_8D_SAT:
281 return 0x8;
282 case QPU_PACK_A_16A:
283 case QPU_PACK_A_16A_SAT:
284 return 0x3;
285 case QPU_PACK_A_16B:
286 case QPU_PACK_A_16B_SAT:
287 return 0xc;
288 }
289 }
290 unreachable("Bad pack field");
291 }
292 #endif
293
294 struct qreg
vir_get_temp(struct v3d_compile * c)295 vir_get_temp(struct v3d_compile *c)
296 {
297 struct qreg reg;
298
299 reg.file = QFILE_TEMP;
300 reg.index = c->num_temps++;
301
302 if (c->num_temps > c->defs_array_size) {
303 uint32_t old_size = c->defs_array_size;
304 c->defs_array_size = MAX2(old_size * 2, 16);
305
306 c->defs = reralloc(c, c->defs, struct qinst *,
307 c->defs_array_size);
308 memset(&c->defs[old_size], 0,
309 sizeof(c->defs[0]) * (c->defs_array_size - old_size));
310
311 c->spillable = reralloc(c, c->spillable,
312 BITSET_WORD,
313 BITSET_WORDS(c->defs_array_size));
314 for (int i = old_size; i < c->defs_array_size; i++)
315 BITSET_SET(c->spillable, i);
316 }
317
318 return reg;
319 }
320
321 struct qinst *
vir_add_inst(enum v3d_qpu_add_op op,struct qreg dst,struct qreg src0,struct qreg src1)322 vir_add_inst(enum v3d_qpu_add_op op, struct qreg dst, struct qreg src0, struct qreg src1)
323 {
324 struct qinst *inst = calloc(1, sizeof(*inst));
325
326 inst->qpu = v3d_qpu_nop();
327 inst->qpu.alu.add.op = op;
328
329 inst->dst = dst;
330 inst->src[0] = src0;
331 inst->src[1] = src1;
332 inst->uniform = ~0;
333
334 return inst;
335 }
336
337 struct qinst *
vir_mul_inst(enum v3d_qpu_mul_op op,struct qreg dst,struct qreg src0,struct qreg src1)338 vir_mul_inst(enum v3d_qpu_mul_op op, struct qreg dst, struct qreg src0, struct qreg src1)
339 {
340 struct qinst *inst = calloc(1, sizeof(*inst));
341
342 inst->qpu = v3d_qpu_nop();
343 inst->qpu.alu.mul.op = op;
344
345 inst->dst = dst;
346 inst->src[0] = src0;
347 inst->src[1] = src1;
348 inst->uniform = ~0;
349
350 return inst;
351 }
352
353 struct qinst *
vir_branch_inst(struct v3d_compile * c,enum v3d_qpu_branch_cond cond)354 vir_branch_inst(struct v3d_compile *c, enum v3d_qpu_branch_cond cond)
355 {
356 struct qinst *inst = calloc(1, sizeof(*inst));
357
358 inst->qpu = v3d_qpu_nop();
359 inst->qpu.type = V3D_QPU_INSTR_TYPE_BRANCH;
360 inst->qpu.branch.cond = cond;
361 inst->qpu.branch.msfign = V3D_QPU_MSFIGN_NONE;
362 inst->qpu.branch.bdi = V3D_QPU_BRANCH_DEST_REL;
363 inst->qpu.branch.ub = true;
364 inst->qpu.branch.bdu = V3D_QPU_BRANCH_DEST_REL;
365
366 inst->dst = vir_nop_reg();
367 inst->uniform = vir_get_uniform_index(c, QUNIFORM_CONSTANT, 0);
368
369 return inst;
370 }
371
372 static void
vir_emit(struct v3d_compile * c,struct qinst * inst)373 vir_emit(struct v3d_compile *c, struct qinst *inst)
374 {
375 switch (c->cursor.mode) {
376 case vir_cursor_add:
377 list_add(&inst->link, c->cursor.link);
378 break;
379 case vir_cursor_addtail:
380 list_addtail(&inst->link, c->cursor.link);
381 break;
382 }
383
384 c->cursor = vir_after_inst(inst);
385 c->live_intervals_valid = false;
386 }
387
388 /* Updates inst to write to a new temporary, emits it, and notes the def. */
389 struct qreg
vir_emit_def(struct v3d_compile * c,struct qinst * inst)390 vir_emit_def(struct v3d_compile *c, struct qinst *inst)
391 {
392 assert(inst->dst.file == QFILE_NULL);
393
394 /* If we're emitting an instruction that's a def, it had better be
395 * writing a register.
396 */
397 if (inst->qpu.type == V3D_QPU_INSTR_TYPE_ALU) {
398 assert(inst->qpu.alu.add.op == V3D_QPU_A_NOP ||
399 v3d_qpu_add_op_has_dst(inst->qpu.alu.add.op));
400 assert(inst->qpu.alu.mul.op == V3D_QPU_M_NOP ||
401 v3d_qpu_mul_op_has_dst(inst->qpu.alu.mul.op));
402 }
403
404 inst->dst = vir_get_temp(c);
405
406 if (inst->dst.file == QFILE_TEMP)
407 c->defs[inst->dst.index] = inst;
408
409 vir_emit(c, inst);
410
411 return inst->dst;
412 }
413
414 struct qinst *
vir_emit_nondef(struct v3d_compile * c,struct qinst * inst)415 vir_emit_nondef(struct v3d_compile *c, struct qinst *inst)
416 {
417 if (inst->dst.file == QFILE_TEMP)
418 c->defs[inst->dst.index] = NULL;
419
420 vir_emit(c, inst);
421
422 return inst;
423 }
424
425 struct qblock *
vir_new_block(struct v3d_compile * c)426 vir_new_block(struct v3d_compile *c)
427 {
428 struct qblock *block = rzalloc(c, struct qblock);
429
430 list_inithead(&block->instructions);
431
432 block->predecessors = _mesa_set_create(block,
433 _mesa_hash_pointer,
434 _mesa_key_pointer_equal);
435
436 block->index = c->next_block_index++;
437
438 return block;
439 }
440
441 void
vir_set_emit_block(struct v3d_compile * c,struct qblock * block)442 vir_set_emit_block(struct v3d_compile *c, struct qblock *block)
443 {
444 c->cur_block = block;
445 c->cursor = vir_after_block(block);
446 list_addtail(&block->link, &c->blocks);
447 }
448
449 struct qblock *
vir_entry_block(struct v3d_compile * c)450 vir_entry_block(struct v3d_compile *c)
451 {
452 return list_first_entry(&c->blocks, struct qblock, link);
453 }
454
455 struct qblock *
vir_exit_block(struct v3d_compile * c)456 vir_exit_block(struct v3d_compile *c)
457 {
458 return list_last_entry(&c->blocks, struct qblock, link);
459 }
460
461 void
vir_link_blocks(struct qblock * predecessor,struct qblock * successor)462 vir_link_blocks(struct qblock *predecessor, struct qblock *successor)
463 {
464 _mesa_set_add(successor->predecessors, predecessor);
465 if (predecessor->successors[0]) {
466 assert(!predecessor->successors[1]);
467 predecessor->successors[1] = successor;
468 } else {
469 predecessor->successors[0] = successor;
470 }
471 }
472
473 const struct v3d_compiler *
v3d_compiler_init(const struct v3d_device_info * devinfo)474 v3d_compiler_init(const struct v3d_device_info *devinfo)
475 {
476 struct v3d_compiler *compiler = rzalloc(NULL, struct v3d_compiler);
477 if (!compiler)
478 return NULL;
479
480 compiler->devinfo = devinfo;
481
482 if (!vir_init_reg_sets(compiler)) {
483 ralloc_free(compiler);
484 return NULL;
485 }
486
487 return compiler;
488 }
489
490 void
v3d_compiler_free(const struct v3d_compiler * compiler)491 v3d_compiler_free(const struct v3d_compiler *compiler)
492 {
493 ralloc_free((void *)compiler);
494 }
495
496 static struct v3d_compile *
vir_compile_init(const struct v3d_compiler * compiler,struct v3d_key * key,nir_shader * s,void (* debug_output)(const char * msg,void * debug_output_data),void * debug_output_data,int program_id,int variant_id,bool fallback_scheduler)497 vir_compile_init(const struct v3d_compiler *compiler,
498 struct v3d_key *key,
499 nir_shader *s,
500 void (*debug_output)(const char *msg,
501 void *debug_output_data),
502 void *debug_output_data,
503 int program_id, int variant_id,
504 bool fallback_scheduler)
505 {
506 struct v3d_compile *c = rzalloc(NULL, struct v3d_compile);
507
508 c->compiler = compiler;
509 c->devinfo = compiler->devinfo;
510 c->key = key;
511 c->program_id = program_id;
512 c->variant_id = variant_id;
513 c->threads = 4;
514 c->debug_output = debug_output;
515 c->debug_output_data = debug_output_data;
516 c->compilation_result = V3D_COMPILATION_SUCCEEDED;
517 c->fallback_scheduler = fallback_scheduler;
518
519 s = nir_shader_clone(c, s);
520 c->s = s;
521
522 list_inithead(&c->blocks);
523 vir_set_emit_block(c, vir_new_block(c));
524
525 c->output_position_index = -1;
526 c->output_sample_mask_index = -1;
527
528 c->def_ht = _mesa_hash_table_create(c, _mesa_hash_pointer,
529 _mesa_key_pointer_equal);
530
531 return c;
532 }
533
534 static int
type_size_vec4(const struct glsl_type * type,bool bindless)535 type_size_vec4(const struct glsl_type *type, bool bindless)
536 {
537 return glsl_count_attribute_slots(type, false);
538 }
539
540 static void
v3d_lower_nir(struct v3d_compile * c)541 v3d_lower_nir(struct v3d_compile *c)
542 {
543 struct nir_lower_tex_options tex_options = {
544 .lower_txd = true,
545 .lower_tg4_broadcom_swizzle = true,
546
547 .lower_rect = false, /* XXX: Use this on V3D 3.x */
548 .lower_txp = ~0,
549 /* Apply swizzles to all samplers. */
550 .swizzle_result = ~0,
551 };
552
553 /* Lower the format swizzle and (for 32-bit returns)
554 * ARB_texture_swizzle-style swizzle.
555 */
556 for (int i = 0; i < ARRAY_SIZE(c->key->tex); i++) {
557 for (int j = 0; j < 4; j++)
558 tex_options.swizzles[i][j] = c->key->tex[i].swizzle[j];
559
560 if (c->key->tex[i].clamp_s)
561 tex_options.saturate_s |= 1 << i;
562 if (c->key->tex[i].clamp_t)
563 tex_options.saturate_t |= 1 << i;
564 if (c->key->tex[i].clamp_r)
565 tex_options.saturate_r |= 1 << i;
566 if (c->key->tex[i].return_size == 16) {
567 tex_options.lower_tex_packing[i] =
568 nir_lower_tex_packing_16;
569 }
570 }
571
572 /* CS textures may not have return_size reflecting the shadow state. */
573 nir_foreach_uniform_variable(var, c->s) {
574 const struct glsl_type *type = glsl_without_array(var->type);
575 unsigned array_len = MAX2(glsl_get_length(var->type), 1);
576
577 if (!glsl_type_is_sampler(type) ||
578 !glsl_sampler_type_is_shadow(type))
579 continue;
580
581 for (int i = 0; i < array_len; i++) {
582 tex_options.lower_tex_packing[var->data.binding + i] =
583 nir_lower_tex_packing_16;
584 }
585 }
586
587 NIR_PASS_V(c->s, nir_lower_tex, &tex_options);
588 NIR_PASS_V(c->s, nir_lower_system_values);
589
590 NIR_PASS_V(c->s, nir_lower_vars_to_scratch,
591 nir_var_function_temp,
592 0,
593 glsl_get_natural_size_align_bytes);
594 NIR_PASS_V(c->s, v3d_nir_lower_scratch);
595 }
596
597 static void
v3d_set_prog_data_uniforms(struct v3d_compile * c,struct v3d_prog_data * prog_data)598 v3d_set_prog_data_uniforms(struct v3d_compile *c,
599 struct v3d_prog_data *prog_data)
600 {
601 int count = c->num_uniforms;
602 struct v3d_uniform_list *ulist = &prog_data->uniforms;
603
604 ulist->count = count;
605 ulist->data = ralloc_array(prog_data, uint32_t, count);
606 memcpy(ulist->data, c->uniform_data,
607 count * sizeof(*ulist->data));
608 ulist->contents = ralloc_array(prog_data, enum quniform_contents, count);
609 memcpy(ulist->contents, c->uniform_contents,
610 count * sizeof(*ulist->contents));
611 }
612
613 static void
v3d_vs_set_prog_data(struct v3d_compile * c,struct v3d_vs_prog_data * prog_data)614 v3d_vs_set_prog_data(struct v3d_compile *c,
615 struct v3d_vs_prog_data *prog_data)
616 {
617 /* The vertex data gets format converted by the VPM so that
618 * each attribute channel takes up a VPM column. Precompute
619 * the sizes for the shader record.
620 */
621 for (int i = 0; i < ARRAY_SIZE(prog_data->vattr_sizes); i++) {
622 prog_data->vattr_sizes[i] = c->vattr_sizes[i];
623 prog_data->vpm_input_size += c->vattr_sizes[i];
624 }
625
626 prog_data->uses_vid = (c->s->info.system_values_read &
627 (1ull << SYSTEM_VALUE_VERTEX_ID));
628 prog_data->uses_iid = (c->s->info.system_values_read &
629 (1ull << SYSTEM_VALUE_INSTANCE_ID));
630
631 if (prog_data->uses_vid)
632 prog_data->vpm_input_size++;
633 if (prog_data->uses_iid)
634 prog_data->vpm_input_size++;
635
636 /* Input/output segment size are in sectors (8 rows of 32 bits per
637 * channel).
638 */
639 prog_data->vpm_input_size = align(prog_data->vpm_input_size, 8) / 8;
640 prog_data->vpm_output_size = align(c->vpm_output_size, 8) / 8;
641
642 /* Set us up for shared input/output segments. This is apparently
643 * necessary for our VCM setup to avoid varying corruption.
644 */
645 prog_data->separate_segments = false;
646 prog_data->vpm_output_size = MAX2(prog_data->vpm_output_size,
647 prog_data->vpm_input_size);
648 prog_data->vpm_input_size = 0;
649
650 /* Compute VCM cache size. We set up our program to take up less than
651 * half of the VPM, so that any set of bin and render programs won't
652 * run out of space. We need space for at least one input segment,
653 * and then allocate the rest to output segments (one for the current
654 * program, the rest to VCM). The valid range of the VCM cache size
655 * field is 1-4 16-vertex batches, but GFXH-1744 limits us to 2-4
656 * batches.
657 */
658 assert(c->devinfo->vpm_size);
659 int sector_size = V3D_CHANNELS * sizeof(uint32_t) * 8;
660 int vpm_size_in_sectors = c->devinfo->vpm_size / sector_size;
661 int half_vpm = vpm_size_in_sectors / 2;
662 int vpm_output_sectors = half_vpm - prog_data->vpm_input_size;
663 int vpm_output_batches = vpm_output_sectors / prog_data->vpm_output_size;
664 assert(vpm_output_batches >= 2);
665 prog_data->vcm_cache_size = CLAMP(vpm_output_batches - 1, 2, 4);
666 }
667
668 static void
v3d_gs_set_prog_data(struct v3d_compile * c,struct v3d_gs_prog_data * prog_data)669 v3d_gs_set_prog_data(struct v3d_compile *c,
670 struct v3d_gs_prog_data *prog_data)
671 {
672 prog_data->num_inputs = c->num_inputs;
673 memcpy(prog_data->input_slots, c->input_slots,
674 c->num_inputs * sizeof(*c->input_slots));
675
676 /* gl_PrimitiveIdIn is written by the GBG into the first word of the
677 * VPM output header automatically and the shader will overwrite
678 * it after reading it if necessary, so it doesn't add to the VPM
679 * size requirements.
680 */
681 prog_data->uses_pid = (c->s->info.system_values_read &
682 (1ull << SYSTEM_VALUE_PRIMITIVE_ID));
683
684 /* Output segment size is in sectors (8 rows of 32 bits per channel) */
685 prog_data->vpm_output_size = align(c->vpm_output_size, 8) / 8;
686
687 /* Compute SIMD dispatch width and update VPM output size accordingly
688 * to ensure we can fit our program in memory. Available widths are
689 * 16, 8, 4, 1.
690 *
691 * Notice that at draw time we will have to consider VPM memory
692 * requirements from other stages and choose a smaller dispatch
693 * width if needed to fit the program in VPM memory.
694 */
695 prog_data->simd_width = 16;
696 while ((prog_data->simd_width > 1 && prog_data->vpm_output_size > 16) ||
697 prog_data->simd_width == 2) {
698 prog_data->simd_width >>= 1;
699 prog_data->vpm_output_size =
700 align(prog_data->vpm_output_size, 2) / 2;
701 }
702 assert(prog_data->vpm_output_size <= 16);
703 assert(prog_data->simd_width != 2);
704
705 prog_data->out_prim_type = c->s->info.gs.output_primitive;
706 prog_data->num_invocations = c->s->info.gs.invocations;
707 }
708
709 static void
v3d_set_fs_prog_data_inputs(struct v3d_compile * c,struct v3d_fs_prog_data * prog_data)710 v3d_set_fs_prog_data_inputs(struct v3d_compile *c,
711 struct v3d_fs_prog_data *prog_data)
712 {
713 prog_data->num_inputs = c->num_inputs;
714 memcpy(prog_data->input_slots, c->input_slots,
715 c->num_inputs * sizeof(*c->input_slots));
716
717 STATIC_ASSERT(ARRAY_SIZE(prog_data->flat_shade_flags) >
718 (V3D_MAX_FS_INPUTS - 1) / 24);
719 for (int i = 0; i < V3D_MAX_FS_INPUTS; i++) {
720 if (BITSET_TEST(c->flat_shade_flags, i))
721 prog_data->flat_shade_flags[i / 24] |= 1 << (i % 24);
722
723 if (BITSET_TEST(c->noperspective_flags, i))
724 prog_data->noperspective_flags[i / 24] |= 1 << (i % 24);
725
726 if (BITSET_TEST(c->centroid_flags, i))
727 prog_data->centroid_flags[i / 24] |= 1 << (i % 24);
728 }
729 }
730
731 static void
v3d_fs_set_prog_data(struct v3d_compile * c,struct v3d_fs_prog_data * prog_data)732 v3d_fs_set_prog_data(struct v3d_compile *c,
733 struct v3d_fs_prog_data *prog_data)
734 {
735 v3d_set_fs_prog_data_inputs(c, prog_data);
736 prog_data->writes_z = c->writes_z;
737 prog_data->disable_ez = !c->s->info.fs.early_fragment_tests;
738 prog_data->uses_center_w = c->uses_center_w;
739 prog_data->uses_implicit_point_line_varyings =
740 c->uses_implicit_point_line_varyings;
741 prog_data->lock_scoreboard_on_first_thrsw =
742 c->lock_scoreboard_on_first_thrsw;
743 }
744
745 static void
v3d_cs_set_prog_data(struct v3d_compile * c,struct v3d_compute_prog_data * prog_data)746 v3d_cs_set_prog_data(struct v3d_compile *c,
747 struct v3d_compute_prog_data *prog_data)
748 {
749 prog_data->shared_size = c->s->info.cs.shared_size;
750 }
751
752 static void
v3d_set_prog_data(struct v3d_compile * c,struct v3d_prog_data * prog_data)753 v3d_set_prog_data(struct v3d_compile *c,
754 struct v3d_prog_data *prog_data)
755 {
756 prog_data->threads = c->threads;
757 prog_data->single_seg = !c->last_thrsw;
758 prog_data->spill_size = c->spill_size;
759 prog_data->tmu_dirty_rcl = c->tmu_dirty_rcl;
760
761 v3d_set_prog_data_uniforms(c, prog_data);
762
763 switch (c->s->info.stage) {
764 case MESA_SHADER_VERTEX:
765 v3d_vs_set_prog_data(c, (struct v3d_vs_prog_data *)prog_data);
766 break;
767 case MESA_SHADER_GEOMETRY:
768 v3d_gs_set_prog_data(c, (struct v3d_gs_prog_data *)prog_data);
769 break;
770 case MESA_SHADER_FRAGMENT:
771 v3d_fs_set_prog_data(c, (struct v3d_fs_prog_data *)prog_data);
772 break;
773 case MESA_SHADER_COMPUTE:
774 v3d_cs_set_prog_data(c, (struct v3d_compute_prog_data *)prog_data);
775 break;
776 default:
777 unreachable("unsupported shader stage");
778 }
779 }
780
781 static uint64_t *
v3d_return_qpu_insts(struct v3d_compile * c,uint32_t * final_assembly_size)782 v3d_return_qpu_insts(struct v3d_compile *c, uint32_t *final_assembly_size)
783 {
784 *final_assembly_size = c->qpu_inst_count * sizeof(uint64_t);
785
786 uint64_t *qpu_insts = malloc(*final_assembly_size);
787 if (!qpu_insts)
788 return NULL;
789
790 memcpy(qpu_insts, c->qpu_insts, *final_assembly_size);
791
792 vir_compile_destroy(c);
793
794 return qpu_insts;
795 }
796
797 static void
v3d_nir_lower_vs_early(struct v3d_compile * c)798 v3d_nir_lower_vs_early(struct v3d_compile *c)
799 {
800 /* Split our I/O vars and dead code eliminate the unused
801 * components.
802 */
803 NIR_PASS_V(c->s, nir_lower_io_to_scalar_early,
804 nir_var_shader_in | nir_var_shader_out);
805 uint64_t used_outputs[4] = {0};
806 for (int i = 0; i < c->vs_key->num_used_outputs; i++) {
807 int slot = v3d_slot_get_slot(c->vs_key->used_outputs[i]);
808 int comp = v3d_slot_get_component(c->vs_key->used_outputs[i]);
809 used_outputs[comp] |= 1ull << slot;
810 }
811 NIR_PASS_V(c->s, nir_remove_unused_io_vars,
812 nir_var_shader_out, used_outputs, NULL); /* demotes to globals */
813 NIR_PASS_V(c->s, nir_lower_global_vars_to_local);
814 v3d_optimize_nir(c->s);
815 NIR_PASS_V(c->s, nir_remove_dead_variables, nir_var_shader_in, NULL);
816
817 /* This must go before nir_lower_io */
818 if (c->vs_key->per_vertex_point_size)
819 NIR_PASS_V(c->s, nir_lower_point_size, 1.0f, 0.0f);
820
821 NIR_PASS_V(c->s, nir_lower_io, nir_var_shader_in | nir_var_shader_out,
822 type_size_vec4,
823 (nir_lower_io_options)0);
824 /* clean up nir_lower_io's deref_var remains */
825 NIR_PASS_V(c->s, nir_opt_dce);
826 }
827
828 static void
v3d_nir_lower_gs_early(struct v3d_compile * c)829 v3d_nir_lower_gs_early(struct v3d_compile *c)
830 {
831 /* Split our I/O vars and dead code eliminate the unused
832 * components.
833 */
834 NIR_PASS_V(c->s, nir_lower_io_to_scalar_early,
835 nir_var_shader_in | nir_var_shader_out);
836 uint64_t used_outputs[4] = {0};
837 for (int i = 0; i < c->gs_key->num_used_outputs; i++) {
838 int slot = v3d_slot_get_slot(c->gs_key->used_outputs[i]);
839 int comp = v3d_slot_get_component(c->gs_key->used_outputs[i]);
840 used_outputs[comp] |= 1ull << slot;
841 }
842 NIR_PASS_V(c->s, nir_remove_unused_io_vars,
843 nir_var_shader_out, used_outputs, NULL); /* demotes to globals */
844 NIR_PASS_V(c->s, nir_lower_global_vars_to_local);
845 v3d_optimize_nir(c->s);
846 NIR_PASS_V(c->s, nir_remove_dead_variables, nir_var_shader_in, NULL);
847
848 /* This must go before nir_lower_io */
849 if (c->gs_key->per_vertex_point_size)
850 NIR_PASS_V(c->s, nir_lower_point_size, 1.0f, 0.0f);
851
852 NIR_PASS_V(c->s, nir_lower_io, nir_var_shader_in | nir_var_shader_out,
853 type_size_vec4,
854 (nir_lower_io_options)0);
855 /* clean up nir_lower_io's deref_var remains */
856 NIR_PASS_V(c->s, nir_opt_dce);
857 }
858
859 static void
v3d_fixup_fs_output_types(struct v3d_compile * c)860 v3d_fixup_fs_output_types(struct v3d_compile *c)
861 {
862 nir_foreach_shader_out_variable(var, c->s) {
863 uint32_t mask = 0;
864
865 switch (var->data.location) {
866 case FRAG_RESULT_COLOR:
867 mask = ~0;
868 break;
869 case FRAG_RESULT_DATA0:
870 case FRAG_RESULT_DATA1:
871 case FRAG_RESULT_DATA2:
872 case FRAG_RESULT_DATA3:
873 mask = 1 << (var->data.location - FRAG_RESULT_DATA0);
874 break;
875 }
876
877 if (c->fs_key->int_color_rb & mask) {
878 var->type =
879 glsl_vector_type(GLSL_TYPE_INT,
880 glsl_get_components(var->type));
881 } else if (c->fs_key->uint_color_rb & mask) {
882 var->type =
883 glsl_vector_type(GLSL_TYPE_UINT,
884 glsl_get_components(var->type));
885 }
886 }
887 }
888
889 static void
v3d_nir_lower_fs_early(struct v3d_compile * c)890 v3d_nir_lower_fs_early(struct v3d_compile *c)
891 {
892 if (c->fs_key->int_color_rb || c->fs_key->uint_color_rb)
893 v3d_fixup_fs_output_types(c);
894
895 NIR_PASS_V(c->s, v3d_nir_lower_logic_ops, c);
896
897 if (c->fs_key->line_smoothing) {
898 v3d_nir_lower_line_smooth(c->s);
899 NIR_PASS_V(c->s, nir_lower_global_vars_to_local);
900 /* The lowering pass can introduce new sysval reads */
901 nir_shader_gather_info(c->s, nir_shader_get_entrypoint(c->s));
902 }
903
904 /* If the shader has no non-TLB side effects, we can promote it to
905 * enabling early_fragment_tests even if the user didn't.
906 */
907 if (!(c->s->info.num_images ||
908 c->s->info.num_ssbos)) {
909 c->s->info.fs.early_fragment_tests = true;
910 }
911 }
912
913 static void
v3d_nir_lower_gs_late(struct v3d_compile * c)914 v3d_nir_lower_gs_late(struct v3d_compile *c)
915 {
916 if (c->key->ucp_enables) {
917 NIR_PASS_V(c->s, nir_lower_clip_gs, c->key->ucp_enables,
918 false, NULL);
919 }
920
921 /* Note: GS output scalarizing must happen after nir_lower_clip_gs. */
922 NIR_PASS_V(c->s, nir_lower_io_to_scalar, nir_var_shader_out);
923 }
924
925 static void
v3d_nir_lower_vs_late(struct v3d_compile * c)926 v3d_nir_lower_vs_late(struct v3d_compile *c)
927 {
928 if (c->vs_key->clamp_color)
929 NIR_PASS_V(c->s, nir_lower_clamp_color_outputs);
930
931 if (c->key->ucp_enables) {
932 NIR_PASS_V(c->s, nir_lower_clip_vs, c->key->ucp_enables,
933 false, false, NULL);
934 NIR_PASS_V(c->s, nir_lower_io_to_scalar,
935 nir_var_shader_out);
936 }
937
938 /* Note: VS output scalarizing must happen after nir_lower_clip_vs. */
939 NIR_PASS_V(c->s, nir_lower_io_to_scalar, nir_var_shader_out);
940 }
941
942 static void
v3d_nir_lower_fs_late(struct v3d_compile * c)943 v3d_nir_lower_fs_late(struct v3d_compile *c)
944 {
945 if (c->fs_key->light_twoside)
946 NIR_PASS_V(c->s, nir_lower_two_sided_color, true);
947
948 if (c->fs_key->clamp_color)
949 NIR_PASS_V(c->s, nir_lower_clamp_color_outputs);
950
951 if (c->fs_key->alpha_test) {
952 NIR_PASS_V(c->s, nir_lower_alpha_test,
953 c->fs_key->alpha_test_func,
954 false, NULL);
955 }
956
957 /* In OpenGL the fragment shader can't read gl_ClipDistance[], but
958 * Vulkan allows it, in which case the SPIR-V compiler will declare
959 * VARING_SLOT_CLIP_DIST0 as compact array variable. Pass true as
960 * the last parameter to always operate with a compact array in both
961 * OpenGL and Vulkan so we do't have to care about the API we
962 * are using.
963 */
964 if (c->key->ucp_enables)
965 NIR_PASS_V(c->s, nir_lower_clip_fs, c->key->ucp_enables, true);
966
967 /* Note: FS input scalarizing must happen after
968 * nir_lower_two_sided_color, which only handles a vec4 at a time.
969 */
970 NIR_PASS_V(c->s, nir_lower_io_to_scalar, nir_var_shader_in);
971 }
972
973 static uint32_t
vir_get_max_temps(struct v3d_compile * c)974 vir_get_max_temps(struct v3d_compile *c)
975 {
976 int max_ip = 0;
977 vir_for_each_inst_inorder(inst, c)
978 max_ip++;
979
980 uint32_t *pressure = rzalloc_array(NULL, uint32_t, max_ip);
981
982 for (int t = 0; t < c->num_temps; t++) {
983 for (int i = c->temp_start[t]; (i < c->temp_end[t] &&
984 i < max_ip); i++) {
985 if (i > max_ip)
986 break;
987 pressure[i]++;
988 }
989 }
990
991 uint32_t max_temps = 0;
992 for (int i = 0; i < max_ip; i++)
993 max_temps = MAX2(max_temps, pressure[i]);
994
995 ralloc_free(pressure);
996
997 return max_temps;
998 }
999
1000 enum v3d_dependency_class {
1001 V3D_DEPENDENCY_CLASS_GS_VPM_OUTPUT_0
1002 };
1003
1004 static bool
v3d_intrinsic_dependency_cb(nir_intrinsic_instr * intr,nir_schedule_dependency * dep,void * user_data)1005 v3d_intrinsic_dependency_cb(nir_intrinsic_instr *intr,
1006 nir_schedule_dependency *dep,
1007 void *user_data)
1008 {
1009 struct v3d_compile *c = user_data;
1010
1011 switch (intr->intrinsic) {
1012 case nir_intrinsic_store_output:
1013 /* Writing to location 0 overwrites the value passed in for
1014 * gl_PrimitiveID on geometry shaders
1015 */
1016 if (c->s->info.stage != MESA_SHADER_GEOMETRY ||
1017 nir_intrinsic_base(intr) != 0)
1018 break;
1019
1020 nir_const_value *const_value =
1021 nir_src_as_const_value(intr->src[1]);
1022
1023 if (const_value == NULL)
1024 break;
1025
1026 uint64_t offset =
1027 nir_const_value_as_uint(*const_value,
1028 nir_src_bit_size(intr->src[1]));
1029 if (offset != 0)
1030 break;
1031
1032 dep->klass = V3D_DEPENDENCY_CLASS_GS_VPM_OUTPUT_0;
1033 dep->type = NIR_SCHEDULE_WRITE_DEPENDENCY;
1034 return true;
1035
1036 case nir_intrinsic_load_primitive_id:
1037 if (c->s->info.stage != MESA_SHADER_GEOMETRY)
1038 break;
1039
1040 dep->klass = V3D_DEPENDENCY_CLASS_GS_VPM_OUTPUT_0;
1041 dep->type = NIR_SCHEDULE_READ_DEPENDENCY;
1042 return true;
1043
1044 default:
1045 break;
1046 }
1047
1048 return false;
1049 }
1050
1051 static void
v3d_attempt_compile(struct v3d_compile * c)1052 v3d_attempt_compile(struct v3d_compile *c)
1053 {
1054 switch (c->s->info.stage) {
1055 case MESA_SHADER_VERTEX:
1056 c->vs_key = (struct v3d_vs_key *) c->key;
1057 break;
1058 case MESA_SHADER_GEOMETRY:
1059 c->gs_key = (struct v3d_gs_key *) c->key;
1060 break;
1061 case MESA_SHADER_FRAGMENT:
1062 c->fs_key = (struct v3d_fs_key *) c->key;
1063 break;
1064 case MESA_SHADER_COMPUTE:
1065 break;
1066 default:
1067 unreachable("unsupported shader stage");
1068 }
1069
1070 switch (c->s->info.stage) {
1071 case MESA_SHADER_VERTEX:
1072 v3d_nir_lower_vs_early(c);
1073 break;
1074 case MESA_SHADER_GEOMETRY:
1075 v3d_nir_lower_gs_early(c);
1076 break;
1077 case MESA_SHADER_FRAGMENT:
1078 v3d_nir_lower_fs_early(c);
1079 break;
1080 default:
1081 break;
1082 }
1083
1084 v3d_lower_nir(c);
1085
1086 switch (c->s->info.stage) {
1087 case MESA_SHADER_VERTEX:
1088 v3d_nir_lower_vs_late(c);
1089 break;
1090 case MESA_SHADER_GEOMETRY:
1091 v3d_nir_lower_gs_late(c);
1092 break;
1093 case MESA_SHADER_FRAGMENT:
1094 v3d_nir_lower_fs_late(c);
1095 break;
1096 default:
1097 break;
1098 }
1099
1100 NIR_PASS_V(c->s, v3d_nir_lower_io, c);
1101 NIR_PASS_V(c->s, v3d_nir_lower_txf_ms, c);
1102 NIR_PASS_V(c->s, v3d_nir_lower_image_load_store);
1103 NIR_PASS_V(c->s, nir_lower_idiv, nir_lower_idiv_fast);
1104
1105 v3d_optimize_nir(c->s);
1106
1107 /* Do late algebraic optimization to turn add(a, neg(b)) back into
1108 * subs, then the mandatory cleanup after algebraic. Note that it may
1109 * produce fnegs, and if so then we need to keep running to squash
1110 * fneg(fneg(a)).
1111 */
1112 bool more_late_algebraic = true;
1113 while (more_late_algebraic) {
1114 more_late_algebraic = false;
1115 NIR_PASS(more_late_algebraic, c->s, nir_opt_algebraic_late);
1116 NIR_PASS_V(c->s, nir_opt_constant_folding);
1117 NIR_PASS_V(c->s, nir_copy_prop);
1118 NIR_PASS_V(c->s, nir_opt_dce);
1119 NIR_PASS_V(c->s, nir_opt_cse);
1120 }
1121
1122 NIR_PASS_V(c->s, nir_lower_bool_to_int32);
1123 NIR_PASS_V(c->s, nir_convert_from_ssa, true);
1124
1125 struct nir_schedule_options schedule_options = {
1126 /* Schedule for about half our register space, to enable more
1127 * shaders to hit 4 threads.
1128 */
1129 .threshold = 24,
1130
1131 /* Vertex shaders share the same memory for inputs and outputs,
1132 * fragement and geometry shaders do not.
1133 */
1134 .stages_with_shared_io_memory =
1135 (((1 << MESA_ALL_SHADER_STAGES) - 1) &
1136 ~((1 << MESA_SHADER_FRAGMENT) |
1137 (1 << MESA_SHADER_GEOMETRY))),
1138
1139 .fallback = c->fallback_scheduler,
1140
1141 .intrinsic_cb = v3d_intrinsic_dependency_cb,
1142 .intrinsic_cb_data = c,
1143 };
1144 NIR_PASS_V(c->s, nir_schedule, &schedule_options);
1145
1146 v3d_nir_to_vir(c);
1147 }
1148
v3d_compile(const struct v3d_compiler * compiler,struct v3d_key * key,struct v3d_prog_data ** out_prog_data,nir_shader * s,void (* debug_output)(const char * msg,void * debug_output_data),void * debug_output_data,int program_id,int variant_id,uint32_t * final_assembly_size)1149 uint64_t *v3d_compile(const struct v3d_compiler *compiler,
1150 struct v3d_key *key,
1151 struct v3d_prog_data **out_prog_data,
1152 nir_shader *s,
1153 void (*debug_output)(const char *msg,
1154 void *debug_output_data),
1155 void *debug_output_data,
1156 int program_id, int variant_id,
1157 uint32_t *final_assembly_size)
1158 {
1159 struct v3d_compile *c;
1160
1161 for (int i = 0; true; i++) {
1162 c = vir_compile_init(compiler, key, s,
1163 debug_output, debug_output_data,
1164 program_id, variant_id,
1165 i > 0 /* fallback_scheduler */);
1166
1167 v3d_attempt_compile(c);
1168
1169 if (i > 0 ||
1170 c->compilation_result !=
1171 V3D_COMPILATION_FAILED_REGISTER_ALLOCATION)
1172 break;
1173
1174 char *debug_msg;
1175 int ret = asprintf(&debug_msg,
1176 "Using fallback scheduler for %s",
1177 vir_get_stage_name(c));
1178
1179 if (ret >= 0) {
1180 if (unlikely(V3D_DEBUG & V3D_DEBUG_PERF))
1181 fprintf(stderr, "%s\n", debug_msg);
1182
1183 c->debug_output(debug_msg, c->debug_output_data);
1184 free(debug_msg);
1185 }
1186
1187 vir_compile_destroy(c);
1188 }
1189
1190 struct v3d_prog_data *prog_data;
1191
1192 static const int prog_data_size[] = {
1193 [MESA_SHADER_VERTEX] = sizeof(struct v3d_vs_prog_data),
1194 [MESA_SHADER_GEOMETRY] = sizeof(struct v3d_gs_prog_data),
1195 [MESA_SHADER_FRAGMENT] = sizeof(struct v3d_fs_prog_data),
1196 [MESA_SHADER_COMPUTE] = sizeof(struct v3d_compute_prog_data),
1197 };
1198
1199 assert(c->s->info.stage >= 0 &&
1200 c->s->info.stage < ARRAY_SIZE(prog_data_size) &&
1201 prog_data_size[c->s->info.stage]);
1202
1203 prog_data = rzalloc_size(NULL, prog_data_size[c->s->info.stage]);
1204
1205 v3d_set_prog_data(c, prog_data);
1206
1207 *out_prog_data = prog_data;
1208
1209 char *shaderdb;
1210 int ret = asprintf(&shaderdb,
1211 "%s shader: %d inst, %d threads, %d loops, "
1212 "%d uniforms, %d max-temps, %d:%d spills:fills, "
1213 "%d sfu-stalls, %d inst-and-stalls",
1214 vir_get_stage_name(c),
1215 c->qpu_inst_count,
1216 c->threads,
1217 c->loops,
1218 c->num_uniforms,
1219 vir_get_max_temps(c),
1220 c->spills,
1221 c->fills,
1222 c->qpu_inst_stalled_count,
1223 c->qpu_inst_count + c->qpu_inst_stalled_count);
1224 if (ret >= 0) {
1225 if (V3D_DEBUG & V3D_DEBUG_SHADERDB)
1226 fprintf(stderr, "SHADER-DB: %s\n", shaderdb);
1227
1228 c->debug_output(shaderdb, c->debug_output_data);
1229 free(shaderdb);
1230 }
1231
1232 return v3d_return_qpu_insts(c, final_assembly_size);
1233 }
1234
1235 void
vir_remove_instruction(struct v3d_compile * c,struct qinst * qinst)1236 vir_remove_instruction(struct v3d_compile *c, struct qinst *qinst)
1237 {
1238 if (qinst->dst.file == QFILE_TEMP)
1239 c->defs[qinst->dst.index] = NULL;
1240
1241 assert(&qinst->link != c->cursor.link);
1242
1243 list_del(&qinst->link);
1244 free(qinst);
1245
1246 c->live_intervals_valid = false;
1247 }
1248
1249 struct qreg
vir_follow_movs(struct v3d_compile * c,struct qreg reg)1250 vir_follow_movs(struct v3d_compile *c, struct qreg reg)
1251 {
1252 /* XXX
1253 int pack = reg.pack;
1254
1255 while (reg.file == QFILE_TEMP &&
1256 c->defs[reg.index] &&
1257 (c->defs[reg.index]->op == QOP_MOV ||
1258 c->defs[reg.index]->op == QOP_FMOV) &&
1259 !c->defs[reg.index]->dst.pack &&
1260 !c->defs[reg.index]->src[0].pack) {
1261 reg = c->defs[reg.index]->src[0];
1262 }
1263
1264 reg.pack = pack;
1265 */
1266 return reg;
1267 }
1268
1269 void
vir_compile_destroy(struct v3d_compile * c)1270 vir_compile_destroy(struct v3d_compile *c)
1271 {
1272 /* Defuse the assert that we aren't removing the cursor's instruction.
1273 */
1274 c->cursor.link = NULL;
1275
1276 vir_for_each_block(block, c) {
1277 while (!list_is_empty(&block->instructions)) {
1278 struct qinst *qinst =
1279 list_first_entry(&block->instructions,
1280 struct qinst, link);
1281 vir_remove_instruction(c, qinst);
1282 }
1283 }
1284
1285 ralloc_free(c);
1286 }
1287
1288 uint32_t
vir_get_uniform_index(struct v3d_compile * c,enum quniform_contents contents,uint32_t data)1289 vir_get_uniform_index(struct v3d_compile *c,
1290 enum quniform_contents contents,
1291 uint32_t data)
1292 {
1293 for (int i = 0; i < c->num_uniforms; i++) {
1294 if (c->uniform_contents[i] == contents &&
1295 c->uniform_data[i] == data) {
1296 return i;
1297 }
1298 }
1299
1300 uint32_t uniform = c->num_uniforms++;
1301
1302 if (uniform >= c->uniform_array_size) {
1303 c->uniform_array_size = MAX2(MAX2(16, uniform + 1),
1304 c->uniform_array_size * 2);
1305
1306 c->uniform_data = reralloc(c, c->uniform_data,
1307 uint32_t,
1308 c->uniform_array_size);
1309 c->uniform_contents = reralloc(c, c->uniform_contents,
1310 enum quniform_contents,
1311 c->uniform_array_size);
1312 }
1313
1314 c->uniform_contents[uniform] = contents;
1315 c->uniform_data[uniform] = data;
1316
1317 return uniform;
1318 }
1319
1320 struct qreg
vir_uniform(struct v3d_compile * c,enum quniform_contents contents,uint32_t data)1321 vir_uniform(struct v3d_compile *c,
1322 enum quniform_contents contents,
1323 uint32_t data)
1324 {
1325 struct qinst *inst = vir_NOP(c);
1326 inst->qpu.sig.ldunif = true;
1327 inst->uniform = vir_get_uniform_index(c, contents, data);
1328 inst->dst = vir_get_temp(c);
1329 c->defs[inst->dst.index] = inst;
1330 return inst->dst;
1331 }
1332
1333 #define OPTPASS(func) \
1334 do { \
1335 bool stage_progress = func(c); \
1336 if (stage_progress) { \
1337 progress = true; \
1338 if (print_opt_debug) { \
1339 fprintf(stderr, \
1340 "VIR opt pass %2d: %s progress\n", \
1341 pass, #func); \
1342 } \
1343 /*XXX vir_validate(c);*/ \
1344 } \
1345 } while (0)
1346
1347 void
vir_optimize(struct v3d_compile * c)1348 vir_optimize(struct v3d_compile *c)
1349 {
1350 bool print_opt_debug = false;
1351 int pass = 1;
1352
1353 while (true) {
1354 bool progress = false;
1355
1356 OPTPASS(vir_opt_copy_propagate);
1357 OPTPASS(vir_opt_redundant_flags);
1358 OPTPASS(vir_opt_dead_code);
1359 OPTPASS(vir_opt_small_immediates);
1360
1361 if (!progress)
1362 break;
1363
1364 pass++;
1365 }
1366 }
1367
1368 const char *
vir_get_stage_name(struct v3d_compile * c)1369 vir_get_stage_name(struct v3d_compile *c)
1370 {
1371 if (c->vs_key && c->vs_key->is_coord)
1372 return "MESA_SHADER_VERTEX_BIN";
1373 else if (c->gs_key && c->gs_key->is_coord)
1374 return "MESA_SHADER_GEOMETRY_BIN";
1375 else
1376 return gl_shader_stage_name(c->s->info.stage);
1377 }
1378