projects / mesa.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
history | raw | HEAD
v3d: Add Compute Shader compilation support.
[mesa.git] / src / broadcom / compiler / vir.c
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
27 int
28 vir_get_nsrc(struct qinst *inst)
29 {
30 switch (inst->qpu.type) {
31 case V3D_QPU_INSTR_TYPE_BRANCH:
32 return 0;
33 case V3D_QPU_INSTR_TYPE_ALU:
34 if (inst->qpu.alu.add.op != V3D_QPU_A_NOP)
35 return v3d_qpu_add_op_num_src(inst->qpu.alu.add.op);
36 else
37 return v3d_qpu_mul_op_num_src(inst->qpu.alu.mul.op);
38 }
39
40 return 0;
41 }
42
43 /**
44 * Returns whether the instruction has any side effects that must be
45 * preserved.
46 */
47 bool
48 vir_has_side_effects(struct v3d_compile *c, struct qinst *inst)
49 {
50 switch (inst->qpu.type) {
51 case V3D_QPU_INSTR_TYPE_BRANCH:
52 return true;
53 case V3D_QPU_INSTR_TYPE_ALU:
54 switch (inst->qpu.alu.add.op) {
55 case V3D_QPU_A_SETREVF:
56 case V3D_QPU_A_SETMSF:
57 case V3D_QPU_A_VPMSETUP:
58 case V3D_QPU_A_STVPMV:
59 case V3D_QPU_A_STVPMD:
60 case V3D_QPU_A_STVPMP:
61 case V3D_QPU_A_VPMWT:
62 case V3D_QPU_A_TMUWT:
63 return true;
64 default:
65 break;
66 }
67
68 switch (inst->qpu.alu.mul.op) {
69 case V3D_QPU_M_MULTOP:
70 return true;
71 default:
72 break;
73 }
74 }
75
76 if (inst->qpu.sig.ldtmu ||
77 inst->qpu.sig.ldvary ||
78 inst->qpu.sig.wrtmuc ||
79 inst->qpu.sig.thrsw) {
80 return true;
81 }
82
83 return false;
84 }
85
86 bool
87 vir_is_raw_mov(struct qinst *inst)
88 {
89 if (inst->qpu.type != V3D_QPU_INSTR_TYPE_ALU ||
90 (inst->qpu.alu.mul.op != V3D_QPU_M_FMOV &&
91 inst->qpu.alu.mul.op != V3D_QPU_M_MOV)) {
92 return false;
93 }
94
95 if (inst->qpu.alu.add.output_pack != V3D_QPU_PACK_NONE ||
96 inst->qpu.alu.mul.output_pack != V3D_QPU_PACK_NONE) {
97 return false;
98 }
99
100 if (inst->qpu.alu.add.a_unpack != V3D_QPU_UNPACK_NONE ||
101 inst->qpu.alu.add.b_unpack != V3D_QPU_UNPACK_NONE ||
102 inst->qpu.alu.mul.a_unpack != V3D_QPU_UNPACK_NONE ||
103 inst->qpu.alu.mul.b_unpack != V3D_QPU_UNPACK_NONE) {
104 return false;
105 }
106
107 if (inst->qpu.flags.ac != V3D_QPU_COND_NONE ||
108 inst->qpu.flags.mc != V3D_QPU_COND_NONE)
109 return false;
110
111 return true;
112 }
113
114 bool
115 vir_is_add(struct qinst *inst)
116 {
117 return (inst->qpu.type == V3D_QPU_INSTR_TYPE_ALU &&
118 inst->qpu.alu.add.op != V3D_QPU_A_NOP);
119 }
120
121 bool
122 vir_is_mul(struct qinst *inst)
123 {
124 return (inst->qpu.type == V3D_QPU_INSTR_TYPE_ALU &&
125 inst->qpu.alu.mul.op != V3D_QPU_M_NOP);
126 }
127
128 bool
129 vir_is_tex(struct qinst *inst)
130 {
131 if (inst->dst.file == QFILE_MAGIC)
132 return v3d_qpu_magic_waddr_is_tmu(inst->dst.index);
133
134 if (inst->qpu.type == V3D_QPU_INSTR_TYPE_ALU &&
135 inst->qpu.alu.add.op == V3D_QPU_A_TMUWT) {
136 return true;
137 }
138
139 return false;
140 }
141
142 bool
143 vir_writes_r3(const struct v3d_device_info *devinfo, struct qinst *inst)
144 {
145 for (int i = 0; i < vir_get_nsrc(inst); i++) {
146 switch (inst->src[i].file) {
147 case QFILE_VPM:
148 return true;
149 default:
150 break;
151 }
152 }
153
154 if (devinfo->ver < 41 && (inst->qpu.sig.ldvary ||
155 inst->qpu.sig.ldtlb ||
156 inst->qpu.sig.ldtlbu ||
157 inst->qpu.sig.ldvpm)) {
158 return true;
159 }
160
161 return false;
162 }
163
164 bool
165 vir_writes_r4(const struct v3d_device_info *devinfo, struct qinst *inst)
166 {
167 switch (inst->dst.file) {
168 case QFILE_MAGIC:
169 switch (inst->dst.index) {
170 case V3D_QPU_WADDR_RECIP:
171 case V3D_QPU_WADDR_RSQRT:
172 case V3D_QPU_WADDR_EXP:
173 case V3D_QPU_WADDR_LOG:
174 case V3D_QPU_WADDR_SIN:
175 return true;
176 }
177 break;
178 default:
179 break;
180 }
181
182 if (devinfo->ver < 41 && inst->qpu.sig.ldtmu)
183 return true;
184
185 return false;
186 }
187
188 void
189 vir_set_unpack(struct qinst *inst, int src,
190 enum v3d_qpu_input_unpack unpack)
191 {
192 assert(src == 0 || src == 1);
193
194 if (vir_is_add(inst)) {
195 if (src == 0)
196 inst->qpu.alu.add.a_unpack = unpack;
197 else
198 inst->qpu.alu.add.b_unpack = unpack;
199 } else {
200 assert(vir_is_mul(inst));
201 if (src == 0)
202 inst->qpu.alu.mul.a_unpack = unpack;
203 else
204 inst->qpu.alu.mul.b_unpack = unpack;
205 }
206 }
207
208 void
209 vir_set_cond(struct qinst *inst, enum v3d_qpu_cond cond)
210 {
211 if (vir_is_add(inst)) {
212 inst->qpu.flags.ac = cond;
213 } else {
214 assert(vir_is_mul(inst));
215 inst->qpu.flags.mc = cond;
216 }
217 }
218
219 void
220 vir_set_pf(struct qinst *inst, enum v3d_qpu_pf pf)
221 {
222 if (vir_is_add(inst)) {
223 inst->qpu.flags.apf = pf;
224 } else {
225 assert(vir_is_mul(inst));
226 inst->qpu.flags.mpf = pf;
227 }
228 }
229
230 void
231 vir_set_uf(struct qinst *inst, enum v3d_qpu_uf uf)
232 {
233 if (vir_is_add(inst)) {
234 inst->qpu.flags.auf = uf;
235 } else {
236 assert(vir_is_mul(inst));
237 inst->qpu.flags.muf = uf;
238 }
239 }
240
241 #if 0
242 uint8_t
243 vir_channels_written(struct qinst *inst)
244 {
245 if (vir_is_mul(inst)) {
246 switch (inst->dst.pack) {
247 case QPU_PACK_MUL_NOP:
248 case QPU_PACK_MUL_8888:
249 return 0xf;
250 case QPU_PACK_MUL_8A:
251 return 0x1;
252 case QPU_PACK_MUL_8B:
253 return 0x2;
254 case QPU_PACK_MUL_8C:
255 return 0x4;
256 case QPU_PACK_MUL_8D:
257 return 0x8;
258 }
259 } else {
260 switch (inst->dst.pack) {
261 case QPU_PACK_A_NOP:
262 case QPU_PACK_A_8888:
263 case QPU_PACK_A_8888_SAT:
264 case QPU_PACK_A_32_SAT:
265 return 0xf;
266 case QPU_PACK_A_8A:
267 case QPU_PACK_A_8A_SAT:
268 return 0x1;
269 case QPU_PACK_A_8B:
270 case QPU_PACK_A_8B_SAT:
271 return 0x2;
272 case QPU_PACK_A_8C:
273 case QPU_PACK_A_8C_SAT:
274 return 0x4;
275 case QPU_PACK_A_8D:
276 case QPU_PACK_A_8D_SAT:
277 return 0x8;
278 case QPU_PACK_A_16A:
279 case QPU_PACK_A_16A_SAT:
280 return 0x3;
281 case QPU_PACK_A_16B:
282 case QPU_PACK_A_16B_SAT:
283 return 0xc;
284 }
285 }
286 unreachable("Bad pack field");
287 }
288 #endif
289
290 struct qreg
291 vir_get_temp(struct v3d_compile *c)
292 {
293 struct qreg reg;
294
295 reg.file = QFILE_TEMP;
296 reg.index = c->num_temps++;
297
298 if (c->num_temps > c->defs_array_size) {
299 uint32_t old_size = c->defs_array_size;
300 c->defs_array_size = MAX2(old_size * 2, 16);
301
302 c->defs = reralloc(c, c->defs, struct qinst *,
303 c->defs_array_size);
304 memset(&c->defs[old_size], 0,
305 sizeof(c->defs[0]) * (c->defs_array_size - old_size));
306
307 c->spillable = reralloc(c, c->spillable,
308 BITSET_WORD,
309 BITSET_WORDS(c->defs_array_size));
310 for (int i = old_size; i < c->defs_array_size; i++)
311 BITSET_SET(c->spillable, i);
312 }
313
314 return reg;
315 }
316
317 struct qinst *
318 vir_add_inst(enum v3d_qpu_add_op op, struct qreg dst, struct qreg src0, struct qreg src1)
319 {
320 struct qinst *inst = calloc(1, sizeof(*inst));
321
322 inst->qpu = v3d_qpu_nop();
323 inst->qpu.alu.add.op = op;
324
325 inst->dst = dst;
326 inst->src[0] = src0;
327 inst->src[1] = src1;
328 inst->uniform = ~0;
329
330 return inst;
331 }
332
333 struct qinst *
334 vir_mul_inst(enum v3d_qpu_mul_op op, struct qreg dst, struct qreg src0, struct qreg src1)
335 {
336 struct qinst *inst = calloc(1, sizeof(*inst));
337
338 inst->qpu = v3d_qpu_nop();
339 inst->qpu.alu.mul.op = op;
340
341 inst->dst = dst;
342 inst->src[0] = src0;
343 inst->src[1] = src1;
344 inst->uniform = ~0;
345
346 return inst;
347 }
348
349 struct qinst *
350 vir_branch_inst(struct v3d_compile *c, enum v3d_qpu_branch_cond cond)
351 {
352 struct qinst *inst = calloc(1, sizeof(*inst));
353
354 inst->qpu = v3d_qpu_nop();
355 inst->qpu.type = V3D_QPU_INSTR_TYPE_BRANCH;
356 inst->qpu.branch.cond = cond;
357 inst->qpu.branch.msfign = V3D_QPU_MSFIGN_NONE;
358 inst->qpu.branch.bdi = V3D_QPU_BRANCH_DEST_REL;
359 inst->qpu.branch.ub = true;
360 inst->qpu.branch.bdu = V3D_QPU_BRANCH_DEST_REL;
361
362 inst->dst = vir_nop_reg();
363 inst->uniform = vir_get_uniform_index(c, QUNIFORM_CONSTANT, 0);
364
365 return inst;
366 }
367
368 static void
369 vir_emit(struct v3d_compile *c, struct qinst *inst)
370 {
371 switch (c->cursor.mode) {
372 case vir_cursor_add:
373 list_add(&inst->link, c->cursor.link);
374 break;
375 case vir_cursor_addtail:
376 list_addtail(&inst->link, c->cursor.link);
377 break;
378 }
379
380 c->cursor = vir_after_inst(inst);
381 c->live_intervals_valid = false;
382 }
383
384 /* Updates inst to write to a new temporary, emits it, and notes the def. */
385 struct qreg
386 vir_emit_def(struct v3d_compile *c, struct qinst *inst)
387 {
388 assert(inst->dst.file == QFILE_NULL);
389
390 /* If we're emitting an instruction that's a def, it had better be
391 * writing a register.
392 */
393 if (inst->qpu.type == V3D_QPU_INSTR_TYPE_ALU) {
394 assert(inst->qpu.alu.add.op == V3D_QPU_A_NOP ||
395 v3d_qpu_add_op_has_dst(inst->qpu.alu.add.op));
396 assert(inst->qpu.alu.mul.op == V3D_QPU_M_NOP ||
397 v3d_qpu_mul_op_has_dst(inst->qpu.alu.mul.op));
398 }
399
400 inst->dst = vir_get_temp(c);
401
402 if (inst->dst.file == QFILE_TEMP)
403 c->defs[inst->dst.index] = inst;
404
405 vir_emit(c, inst);
406
407 return inst->dst;
408 }
409
410 struct qinst *
411 vir_emit_nondef(struct v3d_compile *c, struct qinst *inst)
412 {
413 if (inst->dst.file == QFILE_TEMP)
414 c->defs[inst->dst.index] = NULL;
415
416 vir_emit(c, inst);
417
418 return inst;
419 }
420
421 struct qblock *
422 vir_new_block(struct v3d_compile *c)
423 {
424 struct qblock *block = rzalloc(c, struct qblock);
425
426 list_inithead(&block->instructions);
427
428 block->predecessors = _mesa_set_create(block,
429 _mesa_hash_pointer,
430 _mesa_key_pointer_equal);
431
432 block->index = c->next_block_index++;
433
434 return block;
435 }
436
437 void
438 vir_set_emit_block(struct v3d_compile *c, struct qblock *block)
439 {
440 c->cur_block = block;
441 c->cursor = vir_after_block(block);
442 list_addtail(&block->link, &c->blocks);
443 }
444
445 struct qblock *
446 vir_entry_block(struct v3d_compile *c)
447 {
448 return list_first_entry(&c->blocks, struct qblock, link);
449 }
450
451 struct qblock *
452 vir_exit_block(struct v3d_compile *c)
453 {
454 return list_last_entry(&c->blocks, struct qblock, link);
455 }
456
457 void
458 vir_link_blocks(struct qblock *predecessor, struct qblock *successor)
459 {
460 _mesa_set_add(successor->predecessors, predecessor);
461 if (predecessor->successors[0]) {
462 assert(!predecessor->successors[1]);
463 predecessor->successors[1] = successor;
464 } else {
465 predecessor->successors[0] = successor;
466 }
467 }
468
469 const struct v3d_compiler *
470 v3d_compiler_init(const struct v3d_device_info *devinfo)
471 {
472 struct v3d_compiler *compiler = rzalloc(NULL, struct v3d_compiler);
473 if (!compiler)
474 return NULL;
475
476 compiler->devinfo = devinfo;
477
478 if (!vir_init_reg_sets(compiler)) {
479 ralloc_free(compiler);
480 return NULL;
481 }
482
483 return compiler;
484 }
485
486 void
487 v3d_compiler_free(const struct v3d_compiler *compiler)
488 {
489 ralloc_free((void *)compiler);
490 }
491
492 static struct v3d_compile *
493 vir_compile_init(const struct v3d_compiler *compiler,
494 struct v3d_key *key,
495 nir_shader *s,
496 void (*debug_output)(const char *msg,
497 void *debug_output_data),
498 void *debug_output_data,
499 int program_id, int variant_id)
500 {
501 struct v3d_compile *c = rzalloc(NULL, struct v3d_compile);
502
503 c->compiler = compiler;
504 c->devinfo = compiler->devinfo;
505 c->key = key;
506 c->program_id = program_id;
507 c->variant_id = variant_id;
508 c->threads = 4;
509 c->debug_output = debug_output;
510 c->debug_output_data = debug_output_data;
511
512 s = nir_shader_clone(c, s);
513 c->s = s;
514
515 list_inithead(&c->blocks);
516 vir_set_emit_block(c, vir_new_block(c));
517
518 c->output_position_index = -1;
519 c->output_sample_mask_index = -1;
520
521 c->def_ht = _mesa_hash_table_create(c, _mesa_hash_pointer,
522 _mesa_key_pointer_equal);
523
524 return c;
525 }
526
527 static int
528 type_size_vec4(const struct glsl_type *type, bool bindless)
529 {
530 return glsl_count_attribute_slots(type, false);
531 }
532
533 static void
534 v3d_lower_nir(struct v3d_compile *c)
535 {
536 struct nir_lower_tex_options tex_options = {
537 .lower_txd = true,
538 .lower_tg4_broadcom_swizzle = true,
539
540 .lower_rect = false, /* XXX: Use this on V3D 3.x */
541 .lower_txp = ~0,
542 /* Apply swizzles to all samplers. */
543 .swizzle_result = ~0,
544 };
545
546 /* Lower the format swizzle and (for 32-bit returns)
547 * ARB_texture_swizzle-style swizzle.
548 */
549 for (int i = 0; i < ARRAY_SIZE(c->key->tex); i++) {
550 for (int j = 0; j < 4; j++)
551 tex_options.swizzles[i][j] = c->key->tex[i].swizzle[j];
552
553 if (c->key->tex[i].clamp_s)
554 tex_options.saturate_s |= 1 << i;
555 if (c->key->tex[i].clamp_t)
556 tex_options.saturate_t |= 1 << i;
557 if (c->key->tex[i].clamp_r)
558 tex_options.saturate_r |= 1 << i;
559 if (c->key->tex[i].return_size == 16) {
560 tex_options.lower_tex_packing[i] =
561 nir_lower_tex_packing_16;
562 }
563 }
564
565 /* CS textures may not have return_size reflecting the shadow state. */
566 nir_foreach_variable(var, &c->s->uniforms) {
567 const struct glsl_type *type = glsl_without_array(var->type);
568 unsigned array_len = MAX2(glsl_get_length(var->type), 1);
569
570 if (!glsl_type_is_sampler(type) ||
571 !glsl_sampler_type_is_shadow(type))
572 continue;
573
574 for (int i = 0; i < array_len; i++) {
575 tex_options.lower_tex_packing[var->data.binding + i] =
576 nir_lower_tex_packing_16;
577 }
578 }
579
580 NIR_PASS_V(c->s, nir_lower_tex, &tex_options);
581 NIR_PASS_V(c->s, nir_lower_system_values);
582 }
583
584 static void
585 v3d_set_prog_data_uniforms(struct v3d_compile *c,
586 struct v3d_prog_data *prog_data)
587 {
588 int count = c->num_uniforms;
589 struct v3d_uniform_list *ulist = &prog_data->uniforms;
590
591 ulist->count = count;
592 ulist->data = ralloc_array(prog_data, uint32_t, count);
593 memcpy(ulist->data, c->uniform_data,
594 count * sizeof(*ulist->data));
595 ulist->contents = ralloc_array(prog_data, enum quniform_contents, count);
596 memcpy(ulist->contents, c->uniform_contents,
597 count * sizeof(*ulist->contents));
598 }
599
600 static void
601 v3d_vs_set_prog_data(struct v3d_compile *c,
602 struct v3d_vs_prog_data *prog_data)
603 {
604 /* The vertex data gets format converted by the VPM so that
605 * each attribute channel takes up a VPM column. Precompute
606 * the sizes for the shader record.
607 */
608 for (int i = 0; i < ARRAY_SIZE(prog_data->vattr_sizes); i++) {
609 prog_data->vattr_sizes[i] = c->vattr_sizes[i];
610 prog_data->vpm_input_size += c->vattr_sizes[i];
611 }
612
613 prog_data->uses_vid = (c->s->info.system_values_read &
614 (1ull << SYSTEM_VALUE_VERTEX_ID));
615 prog_data->uses_iid = (c->s->info.system_values_read &
616 (1ull << SYSTEM_VALUE_INSTANCE_ID));
617
618 if (prog_data->uses_vid)
619 prog_data->vpm_input_size++;
620 if (prog_data->uses_iid)
621 prog_data->vpm_input_size++;
622
623 /* Input/output segment size are in sectors (8 rows of 32 bits per
624 * channel).
625 */
626 prog_data->vpm_input_size = align(prog_data->vpm_input_size, 8) / 8;
627 prog_data->vpm_output_size = align(c->vpm_output_size, 8) / 8;
628
629 /* Set us up for shared input/output segments. This is apparently
630 * necessary for our VCM setup to avoid varying corruption.
631 */
632 prog_data->separate_segments = false;
633 prog_data->vpm_output_size = MAX2(prog_data->vpm_output_size,
634 prog_data->vpm_input_size);
635 prog_data->vpm_input_size = 0;
636
637 /* Compute VCM cache size. We set up our program to take up less than
638 * half of the VPM, so that any set of bin and render programs won't
639 * run out of space. We need space for at least one input segment,
640 * and then allocate the rest to output segments (one for the current
641 * program, the rest to VCM). The valid range of the VCM cache size
642 * field is 1-4 16-vertex batches, but GFXH-1744 limits us to 2-4
643 * batches.
644 */
645 assert(c->devinfo->vpm_size);
646 int sector_size = V3D_CHANNELS * sizeof(uint32_t) * 8;
647 int vpm_size_in_sectors = c->devinfo->vpm_size / sector_size;
648 int half_vpm = vpm_size_in_sectors / 2;
649 int vpm_output_sectors = half_vpm - prog_data->vpm_input_size;
650 int vpm_output_batches = vpm_output_sectors / prog_data->vpm_output_size;
651 assert(vpm_output_batches >= 2);
652 prog_data->vcm_cache_size = CLAMP(vpm_output_batches - 1, 2, 4);
653 }
654
655 static void
656 v3d_set_fs_prog_data_inputs(struct v3d_compile *c,
657 struct v3d_fs_prog_data *prog_data)
658 {
659 prog_data->num_inputs = c->num_inputs;
660 memcpy(prog_data->input_slots, c->input_slots,
661 c->num_inputs * sizeof(*c->input_slots));
662
663 STATIC_ASSERT(ARRAY_SIZE(prog_data->flat_shade_flags) >
664 (V3D_MAX_FS_INPUTS - 1) / 24);
665 for (int i = 0; i < V3D_MAX_FS_INPUTS; i++) {
666 if (BITSET_TEST(c->flat_shade_flags, i))
667 prog_data->flat_shade_flags[i / 24] |= 1 << (i % 24);
668
669 if (BITSET_TEST(c->noperspective_flags, i))
670 prog_data->noperspective_flags[i / 24] |= 1 << (i % 24);
671
672 if (BITSET_TEST(c->centroid_flags, i))
673 prog_data->centroid_flags[i / 24] |= 1 << (i % 24);
674 }
675 }
676
677 static void
678 v3d_fs_set_prog_data(struct v3d_compile *c,
679 struct v3d_fs_prog_data *prog_data)
680 {
681 v3d_set_fs_prog_data_inputs(c, prog_data);
682 prog_data->writes_z = c->writes_z;
683 prog_data->disable_ez = !c->s->info.fs.early_fragment_tests;
684 prog_data->uses_center_w = c->uses_center_w;
685 }
686
687 static void
688 v3d_cs_set_prog_data(struct v3d_compile *c,
689 struct v3d_compute_prog_data *prog_data)
690 {
691 prog_data->shared_size = c->s->info.cs.shared_size;
692 }
693
694 static void
695 v3d_set_prog_data(struct v3d_compile *c,
696 struct v3d_prog_data *prog_data)
697 {
698 prog_data->threads = c->threads;
699 prog_data->single_seg = !c->last_thrsw;
700 prog_data->spill_size = c->spill_size;
701
702 v3d_set_prog_data_uniforms(c, prog_data);
703
704 if (c->s->info.stage == MESA_SHADER_COMPUTE) {
705 v3d_cs_set_prog_data(c, (struct v3d_compute_prog_data *)prog_data);
706 } else if (c->s->info.stage == MESA_SHADER_VERTEX) {
707 v3d_vs_set_prog_data(c, (struct v3d_vs_prog_data *)prog_data);
708 } else {
709 assert(c->s->info.stage == MESA_SHADER_FRAGMENT);
710 v3d_fs_set_prog_data(c, (struct v3d_fs_prog_data *)prog_data);
711 }
712 }
713
714 static uint64_t *
715 v3d_return_qpu_insts(struct v3d_compile *c, uint32_t *final_assembly_size)
716 {
717 *final_assembly_size = c->qpu_inst_count * sizeof(uint64_t);
718
719 uint64_t *qpu_insts = malloc(*final_assembly_size);
720 if (!qpu_insts)
721 return NULL;
722
723 memcpy(qpu_insts, c->qpu_insts, *final_assembly_size);
724
725 vir_compile_destroy(c);
726
727 return qpu_insts;
728 }
729
730 static void
731 v3d_nir_lower_vs_early(struct v3d_compile *c)
732 {
733 /* Split our I/O vars and dead code eliminate the unused
734 * components.
735 */
736 NIR_PASS_V(c->s, nir_lower_io_to_scalar_early,
737 nir_var_shader_in | nir_var_shader_out);
738 uint64_t used_outputs[4] = {0};
739 for (int i = 0; i < c->vs_key->num_fs_inputs; i++) {
740 int slot = v3d_slot_get_slot(c->vs_key->fs_inputs[i]);
741 int comp = v3d_slot_get_component(c->vs_key->fs_inputs[i]);
742 used_outputs[comp] |= 1ull << slot;
743 }
744 NIR_PASS_V(c->s, nir_remove_unused_io_vars,
745 &c->s->outputs, used_outputs, NULL); /* demotes to globals */
746 NIR_PASS_V(c->s, nir_lower_global_vars_to_local);
747 v3d_optimize_nir(c->s);
748 NIR_PASS_V(c->s, nir_remove_dead_variables, nir_var_shader_in);
749 NIR_PASS_V(c->s, nir_lower_io, nir_var_shader_in | nir_var_shader_out,
750 type_size_vec4,
751 (nir_lower_io_options)0);
752 }
753
754 static void
755 v3d_fixup_fs_output_types(struct v3d_compile *c)
756 {
757 nir_foreach_variable(var, &c->s->outputs) {
758 uint32_t mask = 0;
759
760 switch (var->data.location) {
761 case FRAG_RESULT_COLOR:
762 mask = ~0;
763 break;
764 case FRAG_RESULT_DATA0:
765 case FRAG_RESULT_DATA1:
766 case FRAG_RESULT_DATA2:
767 case FRAG_RESULT_DATA3:
768 mask = 1 << (var->data.location - FRAG_RESULT_DATA0);
769 break;
770 }
771
772 if (c->fs_key->int_color_rb & mask) {
773 var->type =
774 glsl_vector_type(GLSL_TYPE_INT,
775 glsl_get_components(var->type));
776 } else if (c->fs_key->uint_color_rb & mask) {
777 var->type =
778 glsl_vector_type(GLSL_TYPE_UINT,
779 glsl_get_components(var->type));
780 }
781 }
782 }
783
784 static void
785 v3d_nir_lower_fs_early(struct v3d_compile *c)
786 {
787 if (c->fs_key->int_color_rb || c->fs_key->uint_color_rb)
788 v3d_fixup_fs_output_types(c);
789
790 /* If the shader has no non-TLB side effects, we can promote it to
791 * enabling early_fragment_tests even if the user didn't.
792 */
793 if (!(c->s->info.num_images ||
794 c->s->info.num_ssbos ||
795 c->s->info.num_abos)) {
796 c->s->info.fs.early_fragment_tests = true;
797 }
798 }
799
800 static void
801 v3d_nir_lower_vs_late(struct v3d_compile *c)
802 {
803 if (c->vs_key->clamp_color)
804 NIR_PASS_V(c->s, nir_lower_clamp_color_outputs);
805
806 if (c->key->ucp_enables) {
807 NIR_PASS_V(c->s, nir_lower_clip_vs, c->key->ucp_enables,
808 false);
809 NIR_PASS_V(c->s, nir_lower_io_to_scalar,
810 nir_var_shader_out);
811 }
812
813 /* Note: VS output scalarizing must happen after nir_lower_clip_vs. */
814 NIR_PASS_V(c->s, nir_lower_io_to_scalar, nir_var_shader_out);
815 }
816
817 static void
818 v3d_nir_lower_fs_late(struct v3d_compile *c)
819 {
820 if (c->fs_key->light_twoside)
821 NIR_PASS_V(c->s, nir_lower_two_sided_color);
822
823 if (c->fs_key->clamp_color)
824 NIR_PASS_V(c->s, nir_lower_clamp_color_outputs);
825
826 if (c->fs_key->alpha_test) {
827 NIR_PASS_V(c->s, nir_lower_alpha_test,
828 c->fs_key->alpha_test_func,
829 false);
830 }
831
832 if (c->key->ucp_enables)
833 NIR_PASS_V(c->s, nir_lower_clip_fs, c->key->ucp_enables);
834
835 /* Note: FS input scalarizing must happen after
836 * nir_lower_two_sided_color, which only handles a vec4 at a time.
837 */
838 NIR_PASS_V(c->s, nir_lower_io_to_scalar, nir_var_shader_in);
839 }
840
841 static uint32_t
842 vir_get_max_temps(struct v3d_compile *c)
843 {
844 int max_ip = 0;
845 vir_for_each_inst_inorder(inst, c)
846 max_ip++;
847
848 uint32_t *pressure = rzalloc_array(NULL, uint32_t, max_ip);
849
850 for (int t = 0; t < c->num_temps; t++) {
851 for (int i = c->temp_start[t]; (i < c->temp_end[t] &&
852 i < max_ip); i++) {
853 if (i > max_ip)
854 break;
855 pressure[i]++;
856 }
857 }
858
859 uint32_t max_temps = 0;
860 for (int i = 0; i < max_ip; i++)
861 max_temps = MAX2(max_temps, pressure[i]);
862
863 ralloc_free(pressure);
864
865 return max_temps;
866 }
867
868 uint64_t *v3d_compile(const struct v3d_compiler *compiler,
869 struct v3d_key *key,
870 struct v3d_prog_data **out_prog_data,
871 nir_shader *s,
872 void (*debug_output)(const char *msg,
873 void *debug_output_data),
874 void *debug_output_data,
875 int program_id, int variant_id,
876 uint32_t *final_assembly_size)
877 {
878 struct v3d_prog_data *prog_data;
879 struct v3d_compile *c = vir_compile_init(compiler, key, s,
880 debug_output, debug_output_data,
881 program_id, variant_id);
882
883 switch (c->s->info.stage) {
884 case MESA_SHADER_VERTEX:
885 c->vs_key = (struct v3d_vs_key *)key;
886 prog_data = rzalloc_size(NULL, sizeof(struct v3d_vs_prog_data));
887 break;
888 case MESA_SHADER_FRAGMENT:
889 c->fs_key = (struct v3d_fs_key *)key;
890 prog_data = rzalloc_size(NULL, sizeof(struct v3d_fs_prog_data));
891 break;
892 case MESA_SHADER_COMPUTE:
893 prog_data = rzalloc_size(NULL,
894 sizeof(struct v3d_compute_prog_data));
895 break;
896 default:
897 unreachable("unsupported shader stage");
898 }
899
900 if (c->s->info.stage == MESA_SHADER_VERTEX) {
901 v3d_nir_lower_vs_early(c);
902 } else if (c->s->info.stage != MESA_SHADER_COMPUTE) {
903 assert(c->s->info.stage == MESA_SHADER_FRAGMENT);
904 v3d_nir_lower_fs_early(c);
905 }
906
907 v3d_lower_nir(c);
908
909 if (c->s->info.stage == MESA_SHADER_VERTEX) {
910 v3d_nir_lower_vs_late(c);
911 } else if (c->s->info.stage != MESA_SHADER_COMPUTE) {
912 assert(c->s->info.stage == MESA_SHADER_FRAGMENT);
913 v3d_nir_lower_fs_late(c);
914 }
915
916 NIR_PASS_V(c->s, v3d_nir_lower_io, c);
917 NIR_PASS_V(c->s, v3d_nir_lower_txf_ms, c);
918 NIR_PASS_V(c->s, v3d_nir_lower_image_load_store);
919 NIR_PASS_V(c->s, nir_lower_idiv);
920
921 v3d_optimize_nir(c->s);
922 NIR_PASS_V(c->s, nir_lower_bool_to_int32);
923 NIR_PASS_V(c->s, nir_convert_from_ssa, true);
924
925 v3d_nir_to_vir(c);
926
927 v3d_set_prog_data(c, prog_data);
928
929 *out_prog_data = prog_data;
930
931 char *shaderdb;
932 int ret = asprintf(&shaderdb,
933 "%s shader: %d inst, %d threads, %d loops, "
934 "%d uniforms, %d max-temps, %d:%d spills:fills",
935 vir_get_stage_name(c),
936 c->qpu_inst_count,
937 c->threads,
938 c->loops,
939 c->num_uniforms,
940 vir_get_max_temps(c),
941 c->spills,
942 c->fills);
943 if (ret >= 0) {
944 if (V3D_DEBUG & V3D_DEBUG_SHADERDB)
945 fprintf(stderr, "SHADER-DB: %s\n", shaderdb);
946
947 c->debug_output(shaderdb, c->debug_output_data);
948 free(shaderdb);
949 }
950
951 return v3d_return_qpu_insts(c, final_assembly_size);
952 }
953
954 void
955 vir_remove_instruction(struct v3d_compile *c, struct qinst *qinst)
956 {
957 if (qinst->dst.file == QFILE_TEMP)
958 c->defs[qinst->dst.index] = NULL;
959
960 assert(&qinst->link != c->cursor.link);
961
962 list_del(&qinst->link);
963 free(qinst);
964
965 c->live_intervals_valid = false;
966 }
967
968 struct qreg
969 vir_follow_movs(struct v3d_compile *c, struct qreg reg)
970 {
971 /* XXX
972 int pack = reg.pack;
973
974 while (reg.file == QFILE_TEMP &&
975 c->defs[reg.index] &&
976 (c->defs[reg.index]->op == QOP_MOV ||
977 c->defs[reg.index]->op == QOP_FMOV) &&
978 !c->defs[reg.index]->dst.pack &&
979 !c->defs[reg.index]->src[0].pack) {
980 reg = c->defs[reg.index]->src[0];
981 }
982
983 reg.pack = pack;
984 */
985 return reg;
986 }
987
988 void
989 vir_compile_destroy(struct v3d_compile *c)
990 {
991 /* Defuse the assert that we aren't removing the cursor's instruction.
992 */
993 c->cursor.link = NULL;
994
995 vir_for_each_block(block, c) {
996 while (!list_empty(&block->instructions)) {
997 struct qinst *qinst =
998 list_first_entry(&block->instructions,
999 struct qinst, link);
1000 vir_remove_instruction(c, qinst);
1001 }
1002 }
1003
1004 ralloc_free(c);
1005 }
1006
1007 uint32_t
1008 vir_get_uniform_index(struct v3d_compile *c,
1009 enum quniform_contents contents,
1010 uint32_t data)
1011 {
1012 for (int i = 0; i < c->num_uniforms; i++) {
1013 if (c->uniform_contents[i] == contents &&
1014 c->uniform_data[i] == data) {
1015 return i;
1016 }
1017 }
1018
1019 uint32_t uniform = c->num_uniforms++;
1020
1021 if (uniform >= c->uniform_array_size) {
1022 c->uniform_array_size = MAX2(MAX2(16, uniform + 1),
1023 c->uniform_array_size * 2);
1024
1025 c->uniform_data = reralloc(c, c->uniform_data,
1026 uint32_t,
1027 c->uniform_array_size);
1028 c->uniform_contents = reralloc(c, c->uniform_contents,
1029 enum quniform_contents,
1030 c->uniform_array_size);
1031 }
1032
1033 c->uniform_contents[uniform] = contents;
1034 c->uniform_data[uniform] = data;
1035
1036 return uniform;
1037 }
1038
1039 struct qreg
1040 vir_uniform(struct v3d_compile *c,
1041 enum quniform_contents contents,
1042 uint32_t data)
1043 {
1044 struct qinst *inst = vir_NOP(c);
1045 inst->qpu.sig.ldunif = true;
1046 inst->uniform = vir_get_uniform_index(c, contents, data);
1047 inst->dst = vir_get_temp(c);
1048 c->defs[inst->dst.index] = inst;
1049 return inst->dst;
1050 }
1051
1052 #define OPTPASS(func) \
1053 do { \
1054 bool stage_progress = func(c); \
1055 if (stage_progress) { \
1056 progress = true; \
1057 if (print_opt_debug) { \
1058 fprintf(stderr, \
1059 "VIR opt pass %2d: %s progress\n", \
1060 pass, #func); \
1061 } \
1062 /*XXX vir_validate(c);*/ \
1063 } \
1064 } while (0)
1065
1066 void
1067 vir_optimize(struct v3d_compile *c)
1068 {
1069 bool print_opt_debug = false;
1070 int pass = 1;
1071
1072 while (true) {
1073 bool progress = false;
1074
1075 OPTPASS(vir_opt_copy_propagate);
1076 OPTPASS(vir_opt_redundant_flags);
1077 OPTPASS(vir_opt_dead_code);
1078 OPTPASS(vir_opt_small_immediates);
1079
1080 if (!progress)
1081 break;
1082
1083 pass++;
1084 }
1085 }
1086
1087 const char *
1088 vir_get_stage_name(struct v3d_compile *c)
1089 {
1090 if (c->vs_key && c->vs_key->is_coord)
1091 return "MESA_SHADER_COORD";
1092 else
1093 return gl_shader_stage_name(c->s->info.stage);
1094 }