1 /*
2 * Copyright (c) 2014 Scott Mansell
3 * Copyright © 2014 Broadcom
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice (including the next
13 * paragraph) shall be included in all copies or substantial portions of the
14 * Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
22 * IN THE SOFTWARE.
23 */
24
25 #include <inttypes.h>
26 #include "util/u_format.h"
27 #include "util/crc32.h"
28 #include "util/u_math.h"
29 #include "util/u_memory.h"
30 #include "util/ralloc.h"
31 #include "util/hash_table.h"
32 #include "tgsi/tgsi_dump.h"
33 #include "tgsi/tgsi_parse.h"
34 #include "compiler/nir/nir.h"
35 #include "compiler/nir/nir_builder.h"
36 #include "compiler/nir_types.h"
37 #include "nir/tgsi_to_nir.h"
38 #include "vc4_context.h"
39 #include "vc4_qpu.h"
40 #include "vc4_qir.h"
41
42 static struct qreg
43 ntq_get_src(struct vc4_compile *c, nir_src src, int i);
44 static void
45 ntq_emit_cf_list(struct vc4_compile *c, struct exec_list *list);
46
47 static int
type_size(const struct glsl_type * type)48 type_size(const struct glsl_type *type)
49 {
50 return glsl_count_attribute_slots(type, false);
51 }
52
53 static void
resize_qreg_array(struct vc4_compile * c,struct qreg ** regs,uint32_t * size,uint32_t decl_size)54 resize_qreg_array(struct vc4_compile *c,
55 struct qreg **regs,
56 uint32_t *size,
57 uint32_t decl_size)
58 {
59 if (*size >= decl_size)
60 return;
61
62 uint32_t old_size = *size;
63 *size = MAX2(*size * 2, decl_size);
64 *regs = reralloc(c, *regs, struct qreg, *size);
65 if (!*regs) {
66 fprintf(stderr, "Malloc failure\n");
67 abort();
68 }
69
70 for (uint32_t i = old_size; i < *size; i++)
71 (*regs)[i] = c->undef;
72 }
73
74 static void
ntq_emit_thrsw(struct vc4_compile * c)75 ntq_emit_thrsw(struct vc4_compile *c)
76 {
77 if (!c->fs_threaded)
78 return;
79
80 /* Always thread switch after each texture operation for now.
81 *
82 * We could do better by batching a bunch of texture fetches up and
83 * then doing one thread switch and collecting all their results
84 * afterward.
85 */
86 qir_emit_nondef(c, qir_inst(QOP_THRSW, c->undef,
87 c->undef, c->undef));
88 c->last_thrsw_at_top_level = (c->execute.file == QFILE_NULL);
89 }
90
91 static struct qreg
indirect_uniform_load(struct vc4_compile * c,nir_intrinsic_instr * intr)92 indirect_uniform_load(struct vc4_compile *c, nir_intrinsic_instr *intr)
93 {
94 struct qreg indirect_offset = ntq_get_src(c, intr->src[0], 0);
95 uint32_t offset = nir_intrinsic_base(intr);
96 struct vc4_compiler_ubo_range *range = NULL;
97 unsigned i;
98 for (i = 0; i < c->num_uniform_ranges; i++) {
99 range = &c->ubo_ranges[i];
100 if (offset >= range->src_offset &&
101 offset < range->src_offset + range->size) {
102 break;
103 }
104 }
105 /* The driver-location-based offset always has to be within a declared
106 * uniform range.
107 */
108 assert(range);
109 if (!range->used) {
110 range->used = true;
111 range->dst_offset = c->next_ubo_dst_offset;
112 c->next_ubo_dst_offset += range->size;
113 c->num_ubo_ranges++;
114 }
115
116 offset -= range->src_offset;
117
118 /* Adjust for where we stored the TGSI register base. */
119 indirect_offset = qir_ADD(c, indirect_offset,
120 qir_uniform_ui(c, (range->dst_offset +
121 offset)));
122
123 /* Clamp to [0, array size). Note that MIN/MAX are signed. */
124 indirect_offset = qir_MAX(c, indirect_offset, qir_uniform_ui(c, 0));
125 indirect_offset = qir_MIN_NOIMM(c, indirect_offset,
126 qir_uniform_ui(c, (range->dst_offset +
127 range->size - 4)));
128
129 qir_ADD_dest(c, qir_reg(QFILE_TEX_S_DIRECT, 0),
130 indirect_offset,
131 qir_uniform(c, QUNIFORM_UBO_ADDR, 0));
132
133 c->num_texture_samples++;
134
135 ntq_emit_thrsw(c);
136
137 return qir_TEX_RESULT(c);
138 }
139
140 nir_ssa_def *
vc4_nir_get_swizzled_channel(nir_builder * b,nir_ssa_def ** srcs,int swiz)141 vc4_nir_get_swizzled_channel(nir_builder *b, nir_ssa_def **srcs, int swiz)
142 {
143 switch (swiz) {
144 default:
145 case PIPE_SWIZZLE_NONE:
146 fprintf(stderr, "warning: unknown swizzle\n");
147 /* FALLTHROUGH */
148 case PIPE_SWIZZLE_0:
149 return nir_imm_float(b, 0.0);
150 case PIPE_SWIZZLE_1:
151 return nir_imm_float(b, 1.0);
152 case PIPE_SWIZZLE_X:
153 case PIPE_SWIZZLE_Y:
154 case PIPE_SWIZZLE_Z:
155 case PIPE_SWIZZLE_W:
156 return srcs[swiz];
157 }
158 }
159
160 static struct qreg *
ntq_init_ssa_def(struct vc4_compile * c,nir_ssa_def * def)161 ntq_init_ssa_def(struct vc4_compile *c, nir_ssa_def *def)
162 {
163 struct qreg *qregs = ralloc_array(c->def_ht, struct qreg,
164 def->num_components);
165 _mesa_hash_table_insert(c->def_ht, def, qregs);
166 return qregs;
167 }
168
169 /**
170 * This function is responsible for getting QIR results into the associated
171 * storage for a NIR instruction.
172 *
173 * If it's a NIR SSA def, then we just set the associated hash table entry to
174 * the new result.
175 *
176 * If it's a NIR reg, then we need to update the existing qreg assigned to the
177 * NIR destination with the incoming value. To do that without introducing
178 * new MOVs, we require that the incoming qreg either be a uniform, or be
179 * SSA-defined by the previous QIR instruction in the block and rewritable by
180 * this function. That lets us sneak ahead and insert the SF flag beforehand
181 * (knowing that the previous instruction doesn't depend on flags) and rewrite
182 * its destination to be the NIR reg's destination
183 */
184 static void
ntq_store_dest(struct vc4_compile * c,nir_dest * dest,int chan,struct qreg result)185 ntq_store_dest(struct vc4_compile *c, nir_dest *dest, int chan,
186 struct qreg result)
187 {
188 struct qinst *last_inst = NULL;
189 if (!list_empty(&c->cur_block->instructions))
190 last_inst = (struct qinst *)c->cur_block->instructions.prev;
191
192 assert(result.file == QFILE_UNIF ||
193 (result.file == QFILE_TEMP &&
194 last_inst && last_inst == c->defs[result.index]));
195
196 if (dest->is_ssa) {
197 assert(chan < dest->ssa.num_components);
198
199 struct qreg *qregs;
200 struct hash_entry *entry =
201 _mesa_hash_table_search(c->def_ht, &dest->ssa);
202
203 if (entry)
204 qregs = entry->data;
205 else
206 qregs = ntq_init_ssa_def(c, &dest->ssa);
207
208 qregs[chan] = result;
209 } else {
210 nir_register *reg = dest->reg.reg;
211 assert(dest->reg.base_offset == 0);
212 assert(reg->num_array_elems == 0);
213 struct hash_entry *entry =
214 _mesa_hash_table_search(c->def_ht, reg);
215 struct qreg *qregs = entry->data;
216
217 /* Insert a MOV if the source wasn't an SSA def in the
218 * previous instruction.
219 */
220 if (result.file == QFILE_UNIF) {
221 result = qir_MOV(c, result);
222 last_inst = c->defs[result.index];
223 }
224
225 /* We know they're both temps, so just rewrite index. */
226 c->defs[last_inst->dst.index] = NULL;
227 last_inst->dst.index = qregs[chan].index;
228
229 /* If we're in control flow, then make this update of the reg
230 * conditional on the execution mask.
231 */
232 if (c->execute.file != QFILE_NULL) {
233 last_inst->dst.index = qregs[chan].index;
234
235 /* Set the flags to the current exec mask. To insert
236 * the SF, we temporarily remove our SSA instruction.
237 */
238 list_del(&last_inst->link);
239 qir_SF(c, c->execute);
240 list_addtail(&last_inst->link,
241 &c->cur_block->instructions);
242
243 last_inst->cond = QPU_COND_ZS;
244 last_inst->cond_is_exec_mask = true;
245 }
246 }
247 }
248
249 static struct qreg *
ntq_get_dest(struct vc4_compile * c,nir_dest * dest)250 ntq_get_dest(struct vc4_compile *c, nir_dest *dest)
251 {
252 if (dest->is_ssa) {
253 struct qreg *qregs = ntq_init_ssa_def(c, &dest->ssa);
254 for (int i = 0; i < dest->ssa.num_components; i++)
255 qregs[i] = c->undef;
256 return qregs;
257 } else {
258 nir_register *reg = dest->reg.reg;
259 assert(dest->reg.base_offset == 0);
260 assert(reg->num_array_elems == 0);
261 struct hash_entry *entry =
262 _mesa_hash_table_search(c->def_ht, reg);
263 return entry->data;
264 }
265 }
266
267 static struct qreg
ntq_get_src(struct vc4_compile * c,nir_src src,int i)268 ntq_get_src(struct vc4_compile *c, nir_src src, int i)
269 {
270 struct hash_entry *entry;
271 if (src.is_ssa) {
272 entry = _mesa_hash_table_search(c->def_ht, src.ssa);
273 assert(i < src.ssa->num_components);
274 } else {
275 nir_register *reg = src.reg.reg;
276 entry = _mesa_hash_table_search(c->def_ht, reg);
277 assert(reg->num_array_elems == 0);
278 assert(src.reg.base_offset == 0);
279 assert(i < reg->num_components);
280 }
281
282 struct qreg *qregs = entry->data;
283 return qregs[i];
284 }
285
286 static struct qreg
ntq_get_alu_src(struct vc4_compile * c,nir_alu_instr * instr,unsigned src)287 ntq_get_alu_src(struct vc4_compile *c, nir_alu_instr *instr,
288 unsigned src)
289 {
290 assert(util_is_power_of_two(instr->dest.write_mask));
291 unsigned chan = ffs(instr->dest.write_mask) - 1;
292 struct qreg r = ntq_get_src(c, instr->src[src].src,
293 instr->src[src].swizzle[chan]);
294
295 assert(!instr->src[src].abs);
296 assert(!instr->src[src].negate);
297
298 return r;
299 };
300
301 static inline struct qreg
qir_SAT(struct vc4_compile * c,struct qreg val)302 qir_SAT(struct vc4_compile *c, struct qreg val)
303 {
304 return qir_FMAX(c,
305 qir_FMIN(c, val, qir_uniform_f(c, 1.0)),
306 qir_uniform_f(c, 0.0));
307 }
308
309 static struct qreg
ntq_rcp(struct vc4_compile * c,struct qreg x)310 ntq_rcp(struct vc4_compile *c, struct qreg x)
311 {
312 struct qreg r = qir_RCP(c, x);
313
314 /* Apply a Newton-Raphson step to improve the accuracy. */
315 r = qir_FMUL(c, r, qir_FSUB(c,
316 qir_uniform_f(c, 2.0),
317 qir_FMUL(c, x, r)));
318
319 return r;
320 }
321
322 static struct qreg
ntq_rsq(struct vc4_compile * c,struct qreg x)323 ntq_rsq(struct vc4_compile *c, struct qreg x)
324 {
325 struct qreg r = qir_RSQ(c, x);
326
327 /* Apply a Newton-Raphson step to improve the accuracy. */
328 r = qir_FMUL(c, r, qir_FSUB(c,
329 qir_uniform_f(c, 1.5),
330 qir_FMUL(c,
331 qir_uniform_f(c, 0.5),
332 qir_FMUL(c, x,
333 qir_FMUL(c, r, r)))));
334
335 return r;
336 }
337
338 static struct qreg
ntq_umul(struct vc4_compile * c,struct qreg src0,struct qreg src1)339 ntq_umul(struct vc4_compile *c, struct qreg src0, struct qreg src1)
340 {
341 struct qreg src0_hi = qir_SHR(c, src0,
342 qir_uniform_ui(c, 24));
343 struct qreg src1_hi = qir_SHR(c, src1,
344 qir_uniform_ui(c, 24));
345
346 struct qreg hilo = qir_MUL24(c, src0_hi, src1);
347 struct qreg lohi = qir_MUL24(c, src0, src1_hi);
348 struct qreg lolo = qir_MUL24(c, src0, src1);
349
350 return qir_ADD(c, lolo, qir_SHL(c,
351 qir_ADD(c, hilo, lohi),
352 qir_uniform_ui(c, 24)));
353 }
354
355 static struct qreg
ntq_scale_depth_texture(struct vc4_compile * c,struct qreg src)356 ntq_scale_depth_texture(struct vc4_compile *c, struct qreg src)
357 {
358 struct qreg depthf = qir_ITOF(c, qir_SHR(c, src,
359 qir_uniform_ui(c, 8)));
360 return qir_FMUL(c, depthf, qir_uniform_f(c, 1.0f/0xffffff));
361 }
362
363 /**
364 * Emits a lowered TXF_MS from an MSAA texture.
365 *
366 * The addressing math has been lowered in NIR, and now we just need to read
367 * it like a UBO.
368 */
369 static void
ntq_emit_txf(struct vc4_compile * c,nir_tex_instr * instr)370 ntq_emit_txf(struct vc4_compile *c, nir_tex_instr *instr)
371 {
372 uint32_t tile_width = 32;
373 uint32_t tile_height = 32;
374 uint32_t tile_size = (tile_height * tile_width *
375 VC4_MAX_SAMPLES * sizeof(uint32_t));
376
377 unsigned unit = instr->texture_index;
378 uint32_t w = align(c->key->tex[unit].msaa_width, tile_width);
379 uint32_t w_tiles = w / tile_width;
380 uint32_t h = align(c->key->tex[unit].msaa_height, tile_height);
381 uint32_t h_tiles = h / tile_height;
382 uint32_t size = w_tiles * h_tiles * tile_size;
383
384 struct qreg addr;
385 assert(instr->num_srcs == 1);
386 assert(instr->src[0].src_type == nir_tex_src_coord);
387 addr = ntq_get_src(c, instr->src[0].src, 0);
388
389 /* Perform the clamping required by kernel validation. */
390 addr = qir_MAX(c, addr, qir_uniform_ui(c, 0));
391 addr = qir_MIN_NOIMM(c, addr, qir_uniform_ui(c, size - 4));
392
393 qir_ADD_dest(c, qir_reg(QFILE_TEX_S_DIRECT, 0),
394 addr, qir_uniform(c, QUNIFORM_TEXTURE_MSAA_ADDR, unit));
395
396 ntq_emit_thrsw(c);
397
398 struct qreg tex = qir_TEX_RESULT(c);
399 c->num_texture_samples++;
400
401 enum pipe_format format = c->key->tex[unit].format;
402 if (util_format_is_depth_or_stencil(format)) {
403 struct qreg scaled = ntq_scale_depth_texture(c, tex);
404 for (int i = 0; i < 4; i++)
405 ntq_store_dest(c, &instr->dest, i, qir_MOV(c, scaled));
406 } else {
407 for (int i = 0; i < 4; i++)
408 ntq_store_dest(c, &instr->dest, i,
409 qir_UNPACK_8_F(c, tex, i));
410 }
411 }
412
413 static void
ntq_emit_tex(struct vc4_compile * c,nir_tex_instr * instr)414 ntq_emit_tex(struct vc4_compile *c, nir_tex_instr *instr)
415 {
416 struct qreg s, t, r, lod, compare;
417 bool is_txb = false, is_txl = false;
418 unsigned unit = instr->texture_index;
419
420 if (instr->op == nir_texop_txf) {
421 ntq_emit_txf(c, instr);
422 return;
423 }
424
425 for (unsigned i = 0; i < instr->num_srcs; i++) {
426 switch (instr->src[i].src_type) {
427 case nir_tex_src_coord:
428 s = ntq_get_src(c, instr->src[i].src, 0);
429 if (instr->sampler_dim == GLSL_SAMPLER_DIM_1D)
430 t = qir_uniform_f(c, 0.5);
431 else
432 t = ntq_get_src(c, instr->src[i].src, 1);
433 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE)
434 r = ntq_get_src(c, instr->src[i].src, 2);
435 break;
436 case nir_tex_src_bias:
437 lod = ntq_get_src(c, instr->src[i].src, 0);
438 is_txb = true;
439 break;
440 case nir_tex_src_lod:
441 lod = ntq_get_src(c, instr->src[i].src, 0);
442 is_txl = true;
443 break;
444 case nir_tex_src_comparator:
445 compare = ntq_get_src(c, instr->src[i].src, 0);
446 break;
447 default:
448 unreachable("unknown texture source");
449 }
450 }
451
452 if (c->stage != QSTAGE_FRAG && !is_txl) {
453 /* From the GLSL 1.20 spec:
454 *
455 * "If it is mip-mapped and running on the vertex shader,
456 * then the base texture is used."
457 */
458 is_txl = true;
459 lod = qir_uniform_ui(c, 0);
460 }
461
462 if (c->key->tex[unit].force_first_level) {
463 lod = qir_uniform(c, QUNIFORM_TEXTURE_FIRST_LEVEL, unit);
464 is_txl = true;
465 is_txb = false;
466 }
467
468 struct qreg texture_u[] = {
469 qir_uniform(c, QUNIFORM_TEXTURE_CONFIG_P0, unit),
470 qir_uniform(c, QUNIFORM_TEXTURE_CONFIG_P1, unit),
471 qir_uniform(c, QUNIFORM_CONSTANT, 0),
472 qir_uniform(c, QUNIFORM_CONSTANT, 0),
473 };
474 uint32_t next_texture_u = 0;
475
476 /* There is no native support for GL texture rectangle coordinates, so
477 * we have to rescale from ([0, width], [0, height]) to ([0, 1], [0,
478 * 1]).
479 */
480 if (instr->sampler_dim == GLSL_SAMPLER_DIM_RECT) {
481 s = qir_FMUL(c, s,
482 qir_uniform(c, QUNIFORM_TEXRECT_SCALE_X, unit));
483 t = qir_FMUL(c, t,
484 qir_uniform(c, QUNIFORM_TEXRECT_SCALE_Y, unit));
485 }
486
487 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE || is_txl) {
488 texture_u[2] = qir_uniform(c, QUNIFORM_TEXTURE_CONFIG_P2,
489 unit | (is_txl << 16));
490 }
491
492 struct qinst *tmu;
493 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
494 tmu = qir_MOV_dest(c, qir_reg(QFILE_TEX_R, 0), r);
495 tmu->src[qir_get_tex_uniform_src(tmu)] =
496 texture_u[next_texture_u++];
497 } else if (c->key->tex[unit].wrap_s == PIPE_TEX_WRAP_CLAMP_TO_BORDER ||
498 c->key->tex[unit].wrap_s == PIPE_TEX_WRAP_CLAMP ||
499 c->key->tex[unit].wrap_t == PIPE_TEX_WRAP_CLAMP_TO_BORDER ||
500 c->key->tex[unit].wrap_t == PIPE_TEX_WRAP_CLAMP) {
501 tmu = qir_MOV_dest(c, qir_reg(QFILE_TEX_R, 0),
502 qir_uniform(c, QUNIFORM_TEXTURE_BORDER_COLOR,
503 unit));
504 tmu->src[qir_get_tex_uniform_src(tmu)] =
505 texture_u[next_texture_u++];
506 }
507
508 if (c->key->tex[unit].wrap_s == PIPE_TEX_WRAP_CLAMP) {
509 s = qir_SAT(c, s);
510 }
511
512 if (c->key->tex[unit].wrap_t == PIPE_TEX_WRAP_CLAMP) {
513 t = qir_SAT(c, t);
514 }
515
516 tmu = qir_MOV_dest(c, qir_reg(QFILE_TEX_T, 0), t);
517 tmu->src[qir_get_tex_uniform_src(tmu)] =
518 texture_u[next_texture_u++];
519
520 if (is_txl || is_txb) {
521 tmu = qir_MOV_dest(c, qir_reg(QFILE_TEX_B, 0), lod);
522 tmu->src[qir_get_tex_uniform_src(tmu)] =
523 texture_u[next_texture_u++];
524 }
525
526 tmu = qir_MOV_dest(c, qir_reg(QFILE_TEX_S, 0), s);
527 tmu->src[qir_get_tex_uniform_src(tmu)] = texture_u[next_texture_u++];
528
529 c->num_texture_samples++;
530
531 ntq_emit_thrsw(c);
532
533 struct qreg tex = qir_TEX_RESULT(c);
534
535 enum pipe_format format = c->key->tex[unit].format;
536
537 struct qreg *dest = ntq_get_dest(c, &instr->dest);
538 if (util_format_is_depth_or_stencil(format)) {
539 struct qreg normalized = ntq_scale_depth_texture(c, tex);
540 struct qreg depth_output;
541
542 struct qreg u0 = qir_uniform_f(c, 0.0f);
543 struct qreg u1 = qir_uniform_f(c, 1.0f);
544 if (c->key->tex[unit].compare_mode) {
545 /* From the GL_ARB_shadow spec:
546 *
547 * "Let Dt (D subscript t) be the depth texture
548 * value, in the range [0, 1]. Let R be the
549 * interpolated texture coordinate clamped to the
550 * range [0, 1]."
551 */
552 compare = qir_SAT(c, compare);
553
554 switch (c->key->tex[unit].compare_func) {
555 case PIPE_FUNC_NEVER:
556 depth_output = qir_uniform_f(c, 0.0f);
557 break;
558 case PIPE_FUNC_ALWAYS:
559 depth_output = u1;
560 break;
561 case PIPE_FUNC_EQUAL:
562 qir_SF(c, qir_FSUB(c, compare, normalized));
563 depth_output = qir_SEL(c, QPU_COND_ZS, u1, u0);
564 break;
565 case PIPE_FUNC_NOTEQUAL:
566 qir_SF(c, qir_FSUB(c, compare, normalized));
567 depth_output = qir_SEL(c, QPU_COND_ZC, u1, u0);
568 break;
569 case PIPE_FUNC_GREATER:
570 qir_SF(c, qir_FSUB(c, compare, normalized));
571 depth_output = qir_SEL(c, QPU_COND_NC, u1, u0);
572 break;
573 case PIPE_FUNC_GEQUAL:
574 qir_SF(c, qir_FSUB(c, normalized, compare));
575 depth_output = qir_SEL(c, QPU_COND_NS, u1, u0);
576 break;
577 case PIPE_FUNC_LESS:
578 qir_SF(c, qir_FSUB(c, compare, normalized));
579 depth_output = qir_SEL(c, QPU_COND_NS, u1, u0);
580 break;
581 case PIPE_FUNC_LEQUAL:
582 qir_SF(c, qir_FSUB(c, normalized, compare));
583 depth_output = qir_SEL(c, QPU_COND_NC, u1, u0);
584 break;
585 }
586 } else {
587 depth_output = normalized;
588 }
589
590 for (int i = 0; i < 4; i++)
591 dest[i] = depth_output;
592 } else {
593 for (int i = 0; i < 4; i++)
594 dest[i] = qir_UNPACK_8_F(c, tex, i);
595 }
596 }
597
598 /**
599 * Computes x - floor(x), which is tricky because our FTOI truncates (rounds
600 * to zero).
601 */
602 static struct qreg
ntq_ffract(struct vc4_compile * c,struct qreg src)603 ntq_ffract(struct vc4_compile *c, struct qreg src)
604 {
605 struct qreg trunc = qir_ITOF(c, qir_FTOI(c, src));
606 struct qreg diff = qir_FSUB(c, src, trunc);
607 qir_SF(c, diff);
608
609 qir_FADD_dest(c, diff,
610 diff, qir_uniform_f(c, 1.0))->cond = QPU_COND_NS;
611
612 return qir_MOV(c, diff);
613 }
614
615 /**
616 * Computes floor(x), which is tricky because our FTOI truncates (rounds to
617 * zero).
618 */
619 static struct qreg
ntq_ffloor(struct vc4_compile * c,struct qreg src)620 ntq_ffloor(struct vc4_compile *c, struct qreg src)
621 {
622 struct qreg result = qir_ITOF(c, qir_FTOI(c, src));
623
624 /* This will be < 0 if we truncated and the truncation was of a value
625 * that was < 0 in the first place.
626 */
627 qir_SF(c, qir_FSUB(c, src, result));
628
629 struct qinst *sub = qir_FSUB_dest(c, result,
630 result, qir_uniform_f(c, 1.0));
631 sub->cond = QPU_COND_NS;
632
633 return qir_MOV(c, result);
634 }
635
636 /**
637 * Computes ceil(x), which is tricky because our FTOI truncates (rounds to
638 * zero).
639 */
640 static struct qreg
ntq_fceil(struct vc4_compile * c,struct qreg src)641 ntq_fceil(struct vc4_compile *c, struct qreg src)
642 {
643 struct qreg result = qir_ITOF(c, qir_FTOI(c, src));
644
645 /* This will be < 0 if we truncated and the truncation was of a value
646 * that was > 0 in the first place.
647 */
648 qir_SF(c, qir_FSUB(c, result, src));
649
650 qir_FADD_dest(c, result,
651 result, qir_uniform_f(c, 1.0))->cond = QPU_COND_NS;
652
653 return qir_MOV(c, result);
654 }
655
656 static struct qreg
ntq_fsin(struct vc4_compile * c,struct qreg src)657 ntq_fsin(struct vc4_compile *c, struct qreg src)
658 {
659 float coeff[] = {
660 -2.0 * M_PI,
661 pow(2.0 * M_PI, 3) / (3 * 2 * 1),
662 -pow(2.0 * M_PI, 5) / (5 * 4 * 3 * 2 * 1),
663 pow(2.0 * M_PI, 7) / (7 * 6 * 5 * 4 * 3 * 2 * 1),
664 -pow(2.0 * M_PI, 9) / (9 * 8 * 7 * 6 * 5 * 4 * 3 * 2 * 1),
665 };
666
667 struct qreg scaled_x =
668 qir_FMUL(c,
669 src,
670 qir_uniform_f(c, 1.0 / (M_PI * 2.0)));
671
672 struct qreg x = qir_FADD(c,
673 ntq_ffract(c, scaled_x),
674 qir_uniform_f(c, -0.5));
675 struct qreg x2 = qir_FMUL(c, x, x);
676 struct qreg sum = qir_FMUL(c, x, qir_uniform_f(c, coeff[0]));
677 for (int i = 1; i < ARRAY_SIZE(coeff); i++) {
678 x = qir_FMUL(c, x, x2);
679 sum = qir_FADD(c,
680 sum,
681 qir_FMUL(c,
682 x,
683 qir_uniform_f(c, coeff[i])));
684 }
685 return sum;
686 }
687
688 static struct qreg
ntq_fcos(struct vc4_compile * c,struct qreg src)689 ntq_fcos(struct vc4_compile *c, struct qreg src)
690 {
691 float coeff[] = {
692 -1.0f,
693 pow(2.0 * M_PI, 2) / (2 * 1),
694 -pow(2.0 * M_PI, 4) / (4 * 3 * 2 * 1),
695 pow(2.0 * M_PI, 6) / (6 * 5 * 4 * 3 * 2 * 1),
696 -pow(2.0 * M_PI, 8) / (8 * 7 * 6 * 5 * 4 * 3 * 2 * 1),
697 pow(2.0 * M_PI, 10) / (10 * 9 * 8 * 7 * 6 * 5 * 4 * 3 * 2 * 1),
698 };
699
700 struct qreg scaled_x =
701 qir_FMUL(c, src,
702 qir_uniform_f(c, 1.0f / (M_PI * 2.0f)));
703 struct qreg x_frac = qir_FADD(c,
704 ntq_ffract(c, scaled_x),
705 qir_uniform_f(c, -0.5));
706
707 struct qreg sum = qir_uniform_f(c, coeff[0]);
708 struct qreg x2 = qir_FMUL(c, x_frac, x_frac);
709 struct qreg x = x2; /* Current x^2, x^4, or x^6 */
710 for (int i = 1; i < ARRAY_SIZE(coeff); i++) {
711 if (i != 1)
712 x = qir_FMUL(c, x, x2);
713
714 struct qreg mul = qir_FMUL(c,
715 x,
716 qir_uniform_f(c, coeff[i]));
717 if (i == 0)
718 sum = mul;
719 else
720 sum = qir_FADD(c, sum, mul);
721 }
722 return sum;
723 }
724
725 static struct qreg
ntq_fsign(struct vc4_compile * c,struct qreg src)726 ntq_fsign(struct vc4_compile *c, struct qreg src)
727 {
728 struct qreg t = qir_get_temp(c);
729
730 qir_SF(c, src);
731 qir_MOV_dest(c, t, qir_uniform_f(c, 0.0));
732 qir_MOV_dest(c, t, qir_uniform_f(c, 1.0))->cond = QPU_COND_ZC;
733 qir_MOV_dest(c, t, qir_uniform_f(c, -1.0))->cond = QPU_COND_NS;
734 return qir_MOV(c, t);
735 }
736
737 static void
emit_vertex_input(struct vc4_compile * c,int attr)738 emit_vertex_input(struct vc4_compile *c, int attr)
739 {
740 enum pipe_format format = c->vs_key->attr_formats[attr];
741 uint32_t attr_size = util_format_get_blocksize(format);
742
743 c->vattr_sizes[attr] = align(attr_size, 4);
744 for (int i = 0; i < align(attr_size, 4) / 4; i++) {
745 c->inputs[attr * 4 + i] =
746 qir_MOV(c, qir_reg(QFILE_VPM, attr * 4 + i));
747 c->num_inputs++;
748 }
749 }
750
751 static void
emit_fragcoord_input(struct vc4_compile * c,int attr)752 emit_fragcoord_input(struct vc4_compile *c, int attr)
753 {
754 c->inputs[attr * 4 + 0] = qir_ITOF(c, qir_reg(QFILE_FRAG_X, 0));
755 c->inputs[attr * 4 + 1] = qir_ITOF(c, qir_reg(QFILE_FRAG_Y, 0));
756 c->inputs[attr * 4 + 2] =
757 qir_FMUL(c,
758 qir_ITOF(c, qir_FRAG_Z(c)),
759 qir_uniform_f(c, 1.0 / 0xffffff));
760 c->inputs[attr * 4 + 3] = qir_RCP(c, qir_FRAG_W(c));
761 }
762
763 static struct qreg
emit_fragment_varying(struct vc4_compile * c,gl_varying_slot slot,uint8_t swizzle)764 emit_fragment_varying(struct vc4_compile *c, gl_varying_slot slot,
765 uint8_t swizzle)
766 {
767 uint32_t i = c->num_input_slots++;
768 struct qreg vary = {
769 QFILE_VARY,
770 i
771 };
772
773 if (c->num_input_slots >= c->input_slots_array_size) {
774 c->input_slots_array_size =
775 MAX2(4, c->input_slots_array_size * 2);
776
777 c->input_slots = reralloc(c, c->input_slots,
778 struct vc4_varying_slot,
779 c->input_slots_array_size);
780 }
781
782 c->input_slots[i].slot = slot;
783 c->input_slots[i].swizzle = swizzle;
784
785 return qir_VARY_ADD_C(c, qir_FMUL(c, vary, qir_FRAG_W(c)));
786 }
787
788 static void
emit_fragment_input(struct vc4_compile * c,int attr,gl_varying_slot slot)789 emit_fragment_input(struct vc4_compile *c, int attr, gl_varying_slot slot)
790 {
791 for (int i = 0; i < 4; i++) {
792 c->inputs[attr * 4 + i] =
793 emit_fragment_varying(c, slot, i);
794 c->num_inputs++;
795 }
796 }
797
798 static void
add_output(struct vc4_compile * c,uint32_t decl_offset,uint8_t slot,uint8_t swizzle)799 add_output(struct vc4_compile *c,
800 uint32_t decl_offset,
801 uint8_t slot,
802 uint8_t swizzle)
803 {
804 uint32_t old_array_size = c->outputs_array_size;
805 resize_qreg_array(c, &c->outputs, &c->outputs_array_size,
806 decl_offset + 1);
807
808 if (old_array_size != c->outputs_array_size) {
809 c->output_slots = reralloc(c,
810 c->output_slots,
811 struct vc4_varying_slot,
812 c->outputs_array_size);
813 }
814
815 c->output_slots[decl_offset].slot = slot;
816 c->output_slots[decl_offset].swizzle = swizzle;
817 }
818
819 static void
declare_uniform_range(struct vc4_compile * c,uint32_t start,uint32_t size)820 declare_uniform_range(struct vc4_compile *c, uint32_t start, uint32_t size)
821 {
822 unsigned array_id = c->num_uniform_ranges++;
823 if (array_id >= c->ubo_ranges_array_size) {
824 c->ubo_ranges_array_size = MAX2(c->ubo_ranges_array_size * 2,
825 array_id + 1);
826 c->ubo_ranges = reralloc(c, c->ubo_ranges,
827 struct vc4_compiler_ubo_range,
828 c->ubo_ranges_array_size);
829 }
830
831 c->ubo_ranges[array_id].dst_offset = 0;
832 c->ubo_ranges[array_id].src_offset = start;
833 c->ubo_ranges[array_id].size = size;
834 c->ubo_ranges[array_id].used = false;
835 }
836
837 static bool
ntq_src_is_only_ssa_def_user(nir_src * src)838 ntq_src_is_only_ssa_def_user(nir_src *src)
839 {
840 if (!src->is_ssa)
841 return false;
842
843 if (!list_empty(&src->ssa->if_uses))
844 return false;
845
846 return (src->ssa->uses.next == &src->use_link &&
847 src->ssa->uses.next->next == &src->ssa->uses);
848 }
849
850 /**
851 * In general, emits a nir_pack_unorm_4x8 as a series of MOVs with the pack
852 * bit set.
853 *
854 * However, as an optimization, it tries to find the instructions generating
855 * the sources to be packed and just emit the pack flag there, if possible.
856 */
857 static void
ntq_emit_pack_unorm_4x8(struct vc4_compile * c,nir_alu_instr * instr)858 ntq_emit_pack_unorm_4x8(struct vc4_compile *c, nir_alu_instr *instr)
859 {
860 struct qreg result = qir_get_temp(c);
861 struct nir_alu_instr *vec4 = NULL;
862
863 /* If packing from a vec4 op (as expected), identify it so that we can
864 * peek back at what generated its sources.
865 */
866 if (instr->src[0].src.is_ssa &&
867 instr->src[0].src.ssa->parent_instr->type == nir_instr_type_alu &&
868 nir_instr_as_alu(instr->src[0].src.ssa->parent_instr)->op ==
869 nir_op_vec4) {
870 vec4 = nir_instr_as_alu(instr->src[0].src.ssa->parent_instr);
871 }
872
873 /* If the pack is replicating the same channel 4 times, use the 8888
874 * pack flag. This is common for blending using the alpha
875 * channel.
876 */
877 if (instr->src[0].swizzle[0] == instr->src[0].swizzle[1] &&
878 instr->src[0].swizzle[0] == instr->src[0].swizzle[2] &&
879 instr->src[0].swizzle[0] == instr->src[0].swizzle[3]) {
880 struct qreg rep = ntq_get_src(c,
881 instr->src[0].src,
882 instr->src[0].swizzle[0]);
883 ntq_store_dest(c, &instr->dest.dest, 0, qir_PACK_8888_F(c, rep));
884 return;
885 }
886
887 for (int i = 0; i < 4; i++) {
888 int swiz = instr->src[0].swizzle[i];
889 struct qreg src;
890 if (vec4) {
891 src = ntq_get_src(c, vec4->src[swiz].src,
892 vec4->src[swiz].swizzle[0]);
893 } else {
894 src = ntq_get_src(c, instr->src[0].src, swiz);
895 }
896
897 if (vec4 &&
898 ntq_src_is_only_ssa_def_user(&vec4->src[swiz].src) &&
899 src.file == QFILE_TEMP &&
900 c->defs[src.index] &&
901 qir_is_mul(c->defs[src.index]) &&
902 !c->defs[src.index]->dst.pack) {
903 struct qinst *rewrite = c->defs[src.index];
904 c->defs[src.index] = NULL;
905 rewrite->dst = result;
906 rewrite->dst.pack = QPU_PACK_MUL_8A + i;
907 continue;
908 }
909
910 qir_PACK_8_F(c, result, src, i);
911 }
912
913 ntq_store_dest(c, &instr->dest.dest, 0, qir_MOV(c, result));
914 }
915
916 /** Handles sign-extended bitfield extracts for 16 bits. */
917 static struct qreg
ntq_emit_ibfe(struct vc4_compile * c,struct qreg base,struct qreg offset,struct qreg bits)918 ntq_emit_ibfe(struct vc4_compile *c, struct qreg base, struct qreg offset,
919 struct qreg bits)
920 {
921 assert(bits.file == QFILE_UNIF &&
922 c->uniform_contents[bits.index] == QUNIFORM_CONSTANT &&
923 c->uniform_data[bits.index] == 16);
924
925 assert(offset.file == QFILE_UNIF &&
926 c->uniform_contents[offset.index] == QUNIFORM_CONSTANT);
927 int offset_bit = c->uniform_data[offset.index];
928 assert(offset_bit % 16 == 0);
929
930 return qir_UNPACK_16_I(c, base, offset_bit / 16);
931 }
932
933 /** Handles unsigned bitfield extracts for 8 bits. */
934 static struct qreg
ntq_emit_ubfe(struct vc4_compile * c,struct qreg base,struct qreg offset,struct qreg bits)935 ntq_emit_ubfe(struct vc4_compile *c, struct qreg base, struct qreg offset,
936 struct qreg bits)
937 {
938 assert(bits.file == QFILE_UNIF &&
939 c->uniform_contents[bits.index] == QUNIFORM_CONSTANT &&
940 c->uniform_data[bits.index] == 8);
941
942 assert(offset.file == QFILE_UNIF &&
943 c->uniform_contents[offset.index] == QUNIFORM_CONSTANT);
944 int offset_bit = c->uniform_data[offset.index];
945 assert(offset_bit % 8 == 0);
946
947 return qir_UNPACK_8_I(c, base, offset_bit / 8);
948 }
949
950 /**
951 * If compare_instr is a valid comparison instruction, emits the
952 * compare_instr's comparison and returns the sel_instr's return value based
953 * on the compare_instr's result.
954 */
955 static bool
ntq_emit_comparison(struct vc4_compile * c,struct qreg * dest,nir_alu_instr * compare_instr,nir_alu_instr * sel_instr)956 ntq_emit_comparison(struct vc4_compile *c, struct qreg *dest,
957 nir_alu_instr *compare_instr,
958 nir_alu_instr *sel_instr)
959 {
960 enum qpu_cond cond;
961
962 switch (compare_instr->op) {
963 case nir_op_feq:
964 case nir_op_ieq:
965 case nir_op_seq:
966 cond = QPU_COND_ZS;
967 break;
968 case nir_op_fne:
969 case nir_op_ine:
970 case nir_op_sne:
971 cond = QPU_COND_ZC;
972 break;
973 case nir_op_fge:
974 case nir_op_ige:
975 case nir_op_uge:
976 case nir_op_sge:
977 cond = QPU_COND_NC;
978 break;
979 case nir_op_flt:
980 case nir_op_ilt:
981 case nir_op_slt:
982 cond = QPU_COND_NS;
983 break;
984 default:
985 return false;
986 }
987
988 struct qreg src0 = ntq_get_alu_src(c, compare_instr, 0);
989 struct qreg src1 = ntq_get_alu_src(c, compare_instr, 1);
990
991 unsigned unsized_type =
992 nir_alu_type_get_base_type(nir_op_infos[compare_instr->op].input_types[0]);
993 if (unsized_type == nir_type_float)
994 qir_SF(c, qir_FSUB(c, src0, src1));
995 else
996 qir_SF(c, qir_SUB(c, src0, src1));
997
998 switch (sel_instr->op) {
999 case nir_op_seq:
1000 case nir_op_sne:
1001 case nir_op_sge:
1002 case nir_op_slt:
1003 *dest = qir_SEL(c, cond,
1004 qir_uniform_f(c, 1.0), qir_uniform_f(c, 0.0));
1005 break;
1006
1007 case nir_op_bcsel:
1008 *dest = qir_SEL(c, cond,
1009 ntq_get_alu_src(c, sel_instr, 1),
1010 ntq_get_alu_src(c, sel_instr, 2));
1011 break;
1012
1013 default:
1014 *dest = qir_SEL(c, cond,
1015 qir_uniform_ui(c, ~0), qir_uniform_ui(c, 0));
1016 break;
1017 }
1018
1019 /* Make the temporary for nir_store_dest(). */
1020 *dest = qir_MOV(c, *dest);
1021
1022 return true;
1023 }
1024
1025 /**
1026 * Attempts to fold a comparison generating a boolean result into the
1027 * condition code for selecting between two values, instead of comparing the
1028 * boolean result against 0 to generate the condition code.
1029 */
ntq_emit_bcsel(struct vc4_compile * c,nir_alu_instr * instr,struct qreg * src)1030 static struct qreg ntq_emit_bcsel(struct vc4_compile *c, nir_alu_instr *instr,
1031 struct qreg *src)
1032 {
1033 if (!instr->src[0].src.is_ssa)
1034 goto out;
1035 if (instr->src[0].src.ssa->parent_instr->type != nir_instr_type_alu)
1036 goto out;
1037 nir_alu_instr *compare =
1038 nir_instr_as_alu(instr->src[0].src.ssa->parent_instr);
1039 if (!compare)
1040 goto out;
1041
1042 struct qreg dest;
1043 if (ntq_emit_comparison(c, &dest, compare, instr))
1044 return dest;
1045
1046 out:
1047 qir_SF(c, src[0]);
1048 return qir_MOV(c, qir_SEL(c, QPU_COND_NS, src[1], src[2]));
1049 }
1050
1051 static struct qreg
ntq_fddx(struct vc4_compile * c,struct qreg src)1052 ntq_fddx(struct vc4_compile *c, struct qreg src)
1053 {
1054 /* Make sure that we have a bare temp to use for MUL rotation, so it
1055 * can be allocated to an accumulator.
1056 */
1057 if (src.pack || src.file != QFILE_TEMP)
1058 src = qir_MOV(c, src);
1059
1060 struct qreg from_left = qir_ROT_MUL(c, src, 1);
1061 struct qreg from_right = qir_ROT_MUL(c, src, 15);
1062
1063 /* Distinguish left/right pixels of the quad. */
1064 qir_SF(c, qir_AND(c, qir_reg(QFILE_QPU_ELEMENT, 0),
1065 qir_uniform_ui(c, 1)));
1066
1067 return qir_MOV(c, qir_SEL(c, QPU_COND_ZS,
1068 qir_FSUB(c, from_right, src),
1069 qir_FSUB(c, src, from_left)));
1070 }
1071
1072 static struct qreg
ntq_fddy(struct vc4_compile * c,struct qreg src)1073 ntq_fddy(struct vc4_compile *c, struct qreg src)
1074 {
1075 if (src.pack || src.file != QFILE_TEMP)
1076 src = qir_MOV(c, src);
1077
1078 struct qreg from_bottom = qir_ROT_MUL(c, src, 2);
1079 struct qreg from_top = qir_ROT_MUL(c, src, 14);
1080
1081 /* Distinguish top/bottom pixels of the quad. */
1082 qir_SF(c, qir_AND(c,
1083 qir_reg(QFILE_QPU_ELEMENT, 0),
1084 qir_uniform_ui(c, 2)));
1085
1086 return qir_MOV(c, qir_SEL(c, QPU_COND_ZS,
1087 qir_FSUB(c, from_top, src),
1088 qir_FSUB(c, src, from_bottom)));
1089 }
1090
1091 static void
ntq_emit_alu(struct vc4_compile * c,nir_alu_instr * instr)1092 ntq_emit_alu(struct vc4_compile *c, nir_alu_instr *instr)
1093 {
1094 /* This should always be lowered to ALU operations for VC4. */
1095 assert(!instr->dest.saturate);
1096
1097 /* Vectors are special in that they have non-scalarized writemasks,
1098 * and just take the first swizzle channel for each argument in order
1099 * into each writemask channel.
1100 */
1101 if (instr->op == nir_op_vec2 ||
1102 instr->op == nir_op_vec3 ||
1103 instr->op == nir_op_vec4) {
1104 struct qreg srcs[4];
1105 for (int i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
1106 srcs[i] = ntq_get_src(c, instr->src[i].src,
1107 instr->src[i].swizzle[0]);
1108 for (int i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
1109 ntq_store_dest(c, &instr->dest.dest, i,
1110 qir_MOV(c, srcs[i]));
1111 return;
1112 }
1113
1114 if (instr->op == nir_op_pack_unorm_4x8) {
1115 ntq_emit_pack_unorm_4x8(c, instr);
1116 return;
1117 }
1118
1119 if (instr->op == nir_op_unpack_unorm_4x8) {
1120 struct qreg src = ntq_get_src(c, instr->src[0].src,
1121 instr->src[0].swizzle[0]);
1122 for (int i = 0; i < 4; i++) {
1123 if (instr->dest.write_mask & (1 << i))
1124 ntq_store_dest(c, &instr->dest.dest, i,
1125 qir_UNPACK_8_F(c, src, i));
1126 }
1127 return;
1128 }
1129
1130 /* General case: We can just grab the one used channel per src. */
1131 struct qreg src[nir_op_infos[instr->op].num_inputs];
1132 for (int i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
1133 src[i] = ntq_get_alu_src(c, instr, i);
1134 }
1135
1136 struct qreg result;
1137
1138 switch (instr->op) {
1139 case nir_op_fmov:
1140 case nir_op_imov:
1141 result = qir_MOV(c, src[0]);
1142 break;
1143 case nir_op_fmul:
1144 result = qir_FMUL(c, src[0], src[1]);
1145 break;
1146 case nir_op_fadd:
1147 result = qir_FADD(c, src[0], src[1]);
1148 break;
1149 case nir_op_fsub:
1150 result = qir_FSUB(c, src[0], src[1]);
1151 break;
1152 case nir_op_fmin:
1153 result = qir_FMIN(c, src[0], src[1]);
1154 break;
1155 case nir_op_fmax:
1156 result = qir_FMAX(c, src[0], src[1]);
1157 break;
1158
1159 case nir_op_f2i32:
1160 case nir_op_f2u32:
1161 result = qir_FTOI(c, src[0]);
1162 break;
1163 case nir_op_i2f32:
1164 case nir_op_u2f32:
1165 result = qir_ITOF(c, src[0]);
1166 break;
1167 case nir_op_b2f:
1168 result = qir_AND(c, src[0], qir_uniform_f(c, 1.0));
1169 break;
1170 case nir_op_b2i:
1171 result = qir_AND(c, src[0], qir_uniform_ui(c, 1));
1172 break;
1173 case nir_op_i2b:
1174 case nir_op_f2b:
1175 qir_SF(c, src[0]);
1176 result = qir_MOV(c, qir_SEL(c, QPU_COND_ZC,
1177 qir_uniform_ui(c, ~0),
1178 qir_uniform_ui(c, 0)));
1179 break;
1180
1181 case nir_op_iadd:
1182 result = qir_ADD(c, src[0], src[1]);
1183 break;
1184 case nir_op_ushr:
1185 result = qir_SHR(c, src[0], src[1]);
1186 break;
1187 case nir_op_isub:
1188 result = qir_SUB(c, src[0], src[1]);
1189 break;
1190 case nir_op_ishr:
1191 result = qir_ASR(c, src[0], src[1]);
1192 break;
1193 case nir_op_ishl:
1194 result = qir_SHL(c, src[0], src[1]);
1195 break;
1196 case nir_op_imin:
1197 result = qir_MIN(c, src[0], src[1]);
1198 break;
1199 case nir_op_imax:
1200 result = qir_MAX(c, src[0], src[1]);
1201 break;
1202 case nir_op_iand:
1203 result = qir_AND(c, src[0], src[1]);
1204 break;
1205 case nir_op_ior:
1206 result = qir_OR(c, src[0], src[1]);
1207 break;
1208 case nir_op_ixor:
1209 result = qir_XOR(c, src[0], src[1]);
1210 break;
1211 case nir_op_inot:
1212 result = qir_NOT(c, src[0]);
1213 break;
1214
1215 case nir_op_imul:
1216 result = ntq_umul(c, src[0], src[1]);
1217 break;
1218
1219 case nir_op_seq:
1220 case nir_op_sne:
1221 case nir_op_sge:
1222 case nir_op_slt:
1223 case nir_op_feq:
1224 case nir_op_fne:
1225 case nir_op_fge:
1226 case nir_op_flt:
1227 case nir_op_ieq:
1228 case nir_op_ine:
1229 case nir_op_ige:
1230 case nir_op_uge:
1231 case nir_op_ilt:
1232 if (!ntq_emit_comparison(c, &result, instr, instr)) {
1233 fprintf(stderr, "Bad comparison instruction\n");
1234 }
1235 break;
1236
1237 case nir_op_bcsel:
1238 result = ntq_emit_bcsel(c, instr, src);
1239 break;
1240 case nir_op_fcsel:
1241 qir_SF(c, src[0]);
1242 result = qir_MOV(c, qir_SEL(c, QPU_COND_ZC, src[1], src[2]));
1243 break;
1244
1245 case nir_op_frcp:
1246 result = ntq_rcp(c, src[0]);
1247 break;
1248 case nir_op_frsq:
1249 result = ntq_rsq(c, src[0]);
1250 break;
1251 case nir_op_fexp2:
1252 result = qir_EXP2(c, src[0]);
1253 break;
1254 case nir_op_flog2:
1255 result = qir_LOG2(c, src[0]);
1256 break;
1257
1258 case nir_op_ftrunc:
1259 result = qir_ITOF(c, qir_FTOI(c, src[0]));
1260 break;
1261 case nir_op_fceil:
1262 result = ntq_fceil(c, src[0]);
1263 break;
1264 case nir_op_ffract:
1265 result = ntq_ffract(c, src[0]);
1266 break;
1267 case nir_op_ffloor:
1268 result = ntq_ffloor(c, src[0]);
1269 break;
1270
1271 case nir_op_fsin:
1272 result = ntq_fsin(c, src[0]);
1273 break;
1274 case nir_op_fcos:
1275 result = ntq_fcos(c, src[0]);
1276 break;
1277
1278 case nir_op_fsign:
1279 result = ntq_fsign(c, src[0]);
1280 break;
1281
1282 case nir_op_fabs:
1283 result = qir_FMAXABS(c, src[0], src[0]);
1284 break;
1285 case nir_op_iabs:
1286 result = qir_MAX(c, src[0],
1287 qir_SUB(c, qir_uniform_ui(c, 0), src[0]));
1288 break;
1289
1290 case nir_op_ibitfield_extract:
1291 result = ntq_emit_ibfe(c, src[0], src[1], src[2]);
1292 break;
1293
1294 case nir_op_ubitfield_extract:
1295 result = ntq_emit_ubfe(c, src[0], src[1], src[2]);
1296 break;
1297
1298 case nir_op_usadd_4x8:
1299 result = qir_V8ADDS(c, src[0], src[1]);
1300 break;
1301
1302 case nir_op_ussub_4x8:
1303 result = qir_V8SUBS(c, src[0], src[1]);
1304 break;
1305
1306 case nir_op_umin_4x8:
1307 result = qir_V8MIN(c, src[0], src[1]);
1308 break;
1309
1310 case nir_op_umax_4x8:
1311 result = qir_V8MAX(c, src[0], src[1]);
1312 break;
1313
1314 case nir_op_umul_unorm_4x8:
1315 result = qir_V8MULD(c, src[0], src[1]);
1316 break;
1317
1318 case nir_op_fddx:
1319 case nir_op_fddx_coarse:
1320 case nir_op_fddx_fine:
1321 result = ntq_fddx(c, src[0]);
1322 break;
1323
1324 case nir_op_fddy:
1325 case nir_op_fddy_coarse:
1326 case nir_op_fddy_fine:
1327 result = ntq_fddy(c, src[0]);
1328 break;
1329
1330 default:
1331 fprintf(stderr, "unknown NIR ALU inst: ");
1332 nir_print_instr(&instr->instr, stderr);
1333 fprintf(stderr, "\n");
1334 abort();
1335 }
1336
1337 /* We have a scalar result, so the instruction should only have a
1338 * single channel written to.
1339 */
1340 assert(util_is_power_of_two(instr->dest.write_mask));
1341 ntq_store_dest(c, &instr->dest.dest,
1342 ffs(instr->dest.write_mask) - 1, result);
1343 }
1344
1345 static void
emit_frag_end(struct vc4_compile * c)1346 emit_frag_end(struct vc4_compile *c)
1347 {
1348 struct qreg color;
1349 if (c->output_color_index != -1) {
1350 color = c->outputs[c->output_color_index];
1351 } else {
1352 color = qir_uniform_ui(c, 0);
1353 }
1354
1355 uint32_t discard_cond = QPU_COND_ALWAYS;
1356 if (c->s->info.fs.uses_discard) {
1357 qir_SF(c, c->discard);
1358 discard_cond = QPU_COND_ZS;
1359 }
1360
1361 if (c->fs_key->stencil_enabled) {
1362 qir_MOV_dest(c, qir_reg(QFILE_TLB_STENCIL_SETUP, 0),
1363 qir_uniform(c, QUNIFORM_STENCIL, 0));
1364 if (c->fs_key->stencil_twoside) {
1365 qir_MOV_dest(c, qir_reg(QFILE_TLB_STENCIL_SETUP, 0),
1366 qir_uniform(c, QUNIFORM_STENCIL, 1));
1367 }
1368 if (c->fs_key->stencil_full_writemasks) {
1369 qir_MOV_dest(c, qir_reg(QFILE_TLB_STENCIL_SETUP, 0),
1370 qir_uniform(c, QUNIFORM_STENCIL, 2));
1371 }
1372 }
1373
1374 if (c->output_sample_mask_index != -1) {
1375 qir_MS_MASK(c, c->outputs[c->output_sample_mask_index]);
1376 }
1377
1378 if (c->fs_key->depth_enabled) {
1379 if (c->output_position_index != -1) {
1380 qir_FTOI_dest(c, qir_reg(QFILE_TLB_Z_WRITE, 0),
1381 qir_FMUL(c,
1382 c->outputs[c->output_position_index],
1383 qir_uniform_f(c, 0xffffff)))->cond = discard_cond;
1384 } else {
1385 qir_MOV_dest(c, qir_reg(QFILE_TLB_Z_WRITE, 0),
1386 qir_FRAG_Z(c))->cond = discard_cond;
1387 }
1388 }
1389
1390 if (!c->msaa_per_sample_output) {
1391 qir_MOV_dest(c, qir_reg(QFILE_TLB_COLOR_WRITE, 0),
1392 color)->cond = discard_cond;
1393 } else {
1394 for (int i = 0; i < VC4_MAX_SAMPLES; i++) {
1395 qir_MOV_dest(c, qir_reg(QFILE_TLB_COLOR_WRITE_MS, 0),
1396 c->sample_colors[i])->cond = discard_cond;
1397 }
1398 }
1399 }
1400
1401 static void
emit_scaled_viewport_write(struct vc4_compile * c,struct qreg rcp_w)1402 emit_scaled_viewport_write(struct vc4_compile *c, struct qreg rcp_w)
1403 {
1404 struct qreg packed = qir_get_temp(c);
1405
1406 for (int i = 0; i < 2; i++) {
1407 struct qreg scale =
1408 qir_uniform(c, QUNIFORM_VIEWPORT_X_SCALE + i, 0);
1409
1410 struct qreg packed_chan = packed;
1411 packed_chan.pack = QPU_PACK_A_16A + i;
1412
1413 qir_FTOI_dest(c, packed_chan,
1414 qir_FMUL(c,
1415 qir_FMUL(c,
1416 c->outputs[c->output_position_index + i],
1417 scale),
1418 rcp_w));
1419 }
1420
1421 qir_VPM_WRITE(c, packed);
1422 }
1423
1424 static void
emit_zs_write(struct vc4_compile * c,struct qreg rcp_w)1425 emit_zs_write(struct vc4_compile *c, struct qreg rcp_w)
1426 {
1427 struct qreg zscale = qir_uniform(c, QUNIFORM_VIEWPORT_Z_SCALE, 0);
1428 struct qreg zoffset = qir_uniform(c, QUNIFORM_VIEWPORT_Z_OFFSET, 0);
1429
1430 qir_VPM_WRITE(c, qir_FADD(c, qir_FMUL(c, qir_FMUL(c,
1431 c->outputs[c->output_position_index + 2],
1432 zscale),
1433 rcp_w),
1434 zoffset));
1435 }
1436
1437 static void
emit_rcp_wc_write(struct vc4_compile * c,struct qreg rcp_w)1438 emit_rcp_wc_write(struct vc4_compile *c, struct qreg rcp_w)
1439 {
1440 qir_VPM_WRITE(c, rcp_w);
1441 }
1442
1443 static void
emit_point_size_write(struct vc4_compile * c)1444 emit_point_size_write(struct vc4_compile *c)
1445 {
1446 struct qreg point_size;
1447
1448 if (c->output_point_size_index != -1)
1449 point_size = c->outputs[c->output_point_size_index];
1450 else
1451 point_size = qir_uniform_f(c, 1.0);
1452
1453 /* Workaround: HW-2726 PTB does not handle zero-size points (BCM2835,
1454 * BCM21553).
1455 */
1456 point_size = qir_FMAX(c, point_size, qir_uniform_f(c, .125));
1457
1458 qir_VPM_WRITE(c, point_size);
1459 }
1460
1461 /**
1462 * Emits a VPM read of the stub vertex attribute set up by vc4_draw.c.
1463 *
1464 * The simulator insists that there be at least one vertex attribute, so
1465 * vc4_draw.c will emit one if it wouldn't have otherwise. The simulator also
1466 * insists that all vertex attributes loaded get read by the VS/CS, so we have
1467 * to consume it here.
1468 */
1469 static void
emit_stub_vpm_read(struct vc4_compile * c)1470 emit_stub_vpm_read(struct vc4_compile *c)
1471 {
1472 if (c->num_inputs)
1473 return;
1474
1475 c->vattr_sizes[0] = 4;
1476 (void)qir_MOV(c, qir_reg(QFILE_VPM, 0));
1477 c->num_inputs++;
1478 }
1479
1480 static void
emit_vert_end(struct vc4_compile * c,struct vc4_varying_slot * fs_inputs,uint32_t num_fs_inputs)1481 emit_vert_end(struct vc4_compile *c,
1482 struct vc4_varying_slot *fs_inputs,
1483 uint32_t num_fs_inputs)
1484 {
1485 struct qreg rcp_w = ntq_rcp(c, c->outputs[c->output_position_index + 3]);
1486
1487 emit_stub_vpm_read(c);
1488
1489 emit_scaled_viewport_write(c, rcp_w);
1490 emit_zs_write(c, rcp_w);
1491 emit_rcp_wc_write(c, rcp_w);
1492 if (c->vs_key->per_vertex_point_size)
1493 emit_point_size_write(c);
1494
1495 for (int i = 0; i < num_fs_inputs; i++) {
1496 struct vc4_varying_slot *input = &fs_inputs[i];
1497 int j;
1498
1499 for (j = 0; j < c->num_outputs; j++) {
1500 struct vc4_varying_slot *output =
1501 &c->output_slots[j];
1502
1503 if (input->slot == output->slot &&
1504 input->swizzle == output->swizzle) {
1505 qir_VPM_WRITE(c, c->outputs[j]);
1506 break;
1507 }
1508 }
1509 /* Emit padding if we didn't find a declared VS output for
1510 * this FS input.
1511 */
1512 if (j == c->num_outputs)
1513 qir_VPM_WRITE(c, qir_uniform_f(c, 0.0));
1514 }
1515 }
1516
1517 static void
emit_coord_end(struct vc4_compile * c)1518 emit_coord_end(struct vc4_compile *c)
1519 {
1520 struct qreg rcp_w = ntq_rcp(c, c->outputs[c->output_position_index + 3]);
1521
1522 emit_stub_vpm_read(c);
1523
1524 for (int i = 0; i < 4; i++)
1525 qir_VPM_WRITE(c, c->outputs[c->output_position_index + i]);
1526
1527 emit_scaled_viewport_write(c, rcp_w);
1528 emit_zs_write(c, rcp_w);
1529 emit_rcp_wc_write(c, rcp_w);
1530 if (c->vs_key->per_vertex_point_size)
1531 emit_point_size_write(c);
1532 }
1533
1534 static void
vc4_optimize_nir(struct nir_shader * s)1535 vc4_optimize_nir(struct nir_shader *s)
1536 {
1537 bool progress;
1538
1539 do {
1540 progress = false;
1541
1542 NIR_PASS_V(s, nir_lower_vars_to_ssa);
1543 NIR_PASS(progress, s, nir_lower_alu_to_scalar);
1544 NIR_PASS(progress, s, nir_lower_phis_to_scalar);
1545 NIR_PASS(progress, s, nir_copy_prop);
1546 NIR_PASS(progress, s, nir_opt_remove_phis);
1547 NIR_PASS(progress, s, nir_opt_dce);
1548 NIR_PASS(progress, s, nir_opt_dead_cf);
1549 NIR_PASS(progress, s, nir_opt_cse);
1550 NIR_PASS(progress, s, nir_opt_peephole_select, 8);
1551 NIR_PASS(progress, s, nir_opt_algebraic);
1552 NIR_PASS(progress, s, nir_opt_constant_folding);
1553 NIR_PASS(progress, s, nir_opt_undef);
1554 NIR_PASS(progress, s, nir_opt_loop_unroll,
1555 nir_var_shader_in |
1556 nir_var_shader_out |
1557 nir_var_local);
1558 } while (progress);
1559 }
1560
1561 static int
driver_location_compare(const void * in_a,const void * in_b)1562 driver_location_compare(const void *in_a, const void *in_b)
1563 {
1564 const nir_variable *const *a = in_a;
1565 const nir_variable *const *b = in_b;
1566
1567 return (*a)->data.driver_location - (*b)->data.driver_location;
1568 }
1569
1570 static void
ntq_setup_inputs(struct vc4_compile * c)1571 ntq_setup_inputs(struct vc4_compile *c)
1572 {
1573 unsigned num_entries = 0;
1574 nir_foreach_variable(var, &c->s->inputs)
1575 num_entries++;
1576
1577 nir_variable *vars[num_entries];
1578
1579 unsigned i = 0;
1580 nir_foreach_variable(var, &c->s->inputs)
1581 vars[i++] = var;
1582
1583 /* Sort the variables so that we emit the input setup in
1584 * driver_location order. This is required for VPM reads, whose data
1585 * is fetched into the VPM in driver_location (TGSI register index)
1586 * order.
1587 */
1588 qsort(&vars, num_entries, sizeof(*vars), driver_location_compare);
1589
1590 for (unsigned i = 0; i < num_entries; i++) {
1591 nir_variable *var = vars[i];
1592 unsigned array_len = MAX2(glsl_get_length(var->type), 1);
1593 unsigned loc = var->data.driver_location;
1594
1595 assert(array_len == 1);
1596 (void)array_len;
1597 resize_qreg_array(c, &c->inputs, &c->inputs_array_size,
1598 (loc + 1) * 4);
1599
1600 if (c->stage == QSTAGE_FRAG) {
1601 if (var->data.location == VARYING_SLOT_POS) {
1602 emit_fragcoord_input(c, loc);
1603 } else if (var->data.location == VARYING_SLOT_PNTC ||
1604 (var->data.location >= VARYING_SLOT_VAR0 &&
1605 (c->fs_key->point_sprite_mask &
1606 (1 << (var->data.location -
1607 VARYING_SLOT_VAR0))))) {
1608 c->inputs[loc * 4 + 0] = c->point_x;
1609 c->inputs[loc * 4 + 1] = c->point_y;
1610 } else {
1611 emit_fragment_input(c, loc, var->data.location);
1612 }
1613 } else {
1614 emit_vertex_input(c, loc);
1615 }
1616 }
1617 }
1618
1619 static void
ntq_setup_outputs(struct vc4_compile * c)1620 ntq_setup_outputs(struct vc4_compile *c)
1621 {
1622 nir_foreach_variable(var, &c->s->outputs) {
1623 unsigned array_len = MAX2(glsl_get_length(var->type), 1);
1624 unsigned loc = var->data.driver_location * 4;
1625
1626 assert(array_len == 1);
1627 (void)array_len;
1628
1629 for (int i = 0; i < 4; i++)
1630 add_output(c, loc + i, var->data.location, i);
1631
1632 if (c->stage == QSTAGE_FRAG) {
1633 switch (var->data.location) {
1634 case FRAG_RESULT_COLOR:
1635 case FRAG_RESULT_DATA0:
1636 c->output_color_index = loc;
1637 break;
1638 case FRAG_RESULT_DEPTH:
1639 c->output_position_index = loc;
1640 break;
1641 case FRAG_RESULT_SAMPLE_MASK:
1642 c->output_sample_mask_index = loc;
1643 break;
1644 }
1645 } else {
1646 switch (var->data.location) {
1647 case VARYING_SLOT_POS:
1648 c->output_position_index = loc;
1649 break;
1650 case VARYING_SLOT_PSIZ:
1651 c->output_point_size_index = loc;
1652 break;
1653 }
1654 }
1655 }
1656 }
1657
1658 static void
ntq_setup_uniforms(struct vc4_compile * c)1659 ntq_setup_uniforms(struct vc4_compile *c)
1660 {
1661 nir_foreach_variable(var, &c->s->uniforms) {
1662 uint32_t vec4_count = type_size(var->type);
1663 unsigned vec4_size = 4 * sizeof(float);
1664
1665 declare_uniform_range(c, var->data.driver_location * vec4_size,
1666 vec4_count * vec4_size);
1667
1668 }
1669 }
1670
1671 /**
1672 * Sets up the mapping from nir_register to struct qreg *.
1673 *
1674 * Each nir_register gets a struct qreg per 32-bit component being stored.
1675 */
1676 static void
ntq_setup_registers(struct vc4_compile * c,struct exec_list * list)1677 ntq_setup_registers(struct vc4_compile *c, struct exec_list *list)
1678 {
1679 foreach_list_typed(nir_register, nir_reg, node, list) {
1680 unsigned array_len = MAX2(nir_reg->num_array_elems, 1);
1681 struct qreg *qregs = ralloc_array(c->def_ht, struct qreg,
1682 array_len *
1683 nir_reg->num_components);
1684
1685 _mesa_hash_table_insert(c->def_ht, nir_reg, qregs);
1686
1687 for (int i = 0; i < array_len * nir_reg->num_components; i++)
1688 qregs[i] = qir_get_temp(c);
1689 }
1690 }
1691
1692 static void
ntq_emit_load_const(struct vc4_compile * c,nir_load_const_instr * instr)1693 ntq_emit_load_const(struct vc4_compile *c, nir_load_const_instr *instr)
1694 {
1695 struct qreg *qregs = ntq_init_ssa_def(c, &instr->def);
1696 for (int i = 0; i < instr->def.num_components; i++)
1697 qregs[i] = qir_uniform_ui(c, instr->value.u32[i]);
1698
1699 _mesa_hash_table_insert(c->def_ht, &instr->def, qregs);
1700 }
1701
1702 static void
ntq_emit_ssa_undef(struct vc4_compile * c,nir_ssa_undef_instr * instr)1703 ntq_emit_ssa_undef(struct vc4_compile *c, nir_ssa_undef_instr *instr)
1704 {
1705 struct qreg *qregs = ntq_init_ssa_def(c, &instr->def);
1706
1707 /* QIR needs there to be *some* value, so pick 0 (same as for
1708 * ntq_setup_registers().
1709 */
1710 for (int i = 0; i < instr->def.num_components; i++)
1711 qregs[i] = qir_uniform_ui(c, 0);
1712 }
1713
1714 static void
ntq_emit_color_read(struct vc4_compile * c,nir_intrinsic_instr * instr)1715 ntq_emit_color_read(struct vc4_compile *c, nir_intrinsic_instr *instr)
1716 {
1717 assert(nir_src_as_const_value(instr->src[0])->u32[0] == 0);
1718
1719 /* Reads of the per-sample color need to be done in
1720 * order.
1721 */
1722 int sample_index = (nir_intrinsic_base(instr) -
1723 VC4_NIR_TLB_COLOR_READ_INPUT);
1724 for (int i = 0; i <= sample_index; i++) {
1725 if (c->color_reads[i].file == QFILE_NULL) {
1726 c->color_reads[i] =
1727 qir_TLB_COLOR_READ(c);
1728 }
1729 }
1730 ntq_store_dest(c, &instr->dest, 0,
1731 qir_MOV(c, c->color_reads[sample_index]));
1732 }
1733
1734 static void
ntq_emit_load_input(struct vc4_compile * c,nir_intrinsic_instr * instr)1735 ntq_emit_load_input(struct vc4_compile *c, nir_intrinsic_instr *instr)
1736 {
1737 assert(instr->num_components == 1);
1738
1739 nir_const_value *const_offset = nir_src_as_const_value(instr->src[0]);
1740 assert(const_offset && "vc4 doesn't support indirect inputs");
1741
1742 if (c->stage == QSTAGE_FRAG &&
1743 nir_intrinsic_base(instr) >= VC4_NIR_TLB_COLOR_READ_INPUT) {
1744 ntq_emit_color_read(c, instr);
1745 return;
1746 }
1747
1748 uint32_t offset = nir_intrinsic_base(instr) + const_offset->u32[0];
1749 int comp = nir_intrinsic_component(instr);
1750 ntq_store_dest(c, &instr->dest, 0,
1751 qir_MOV(c, c->inputs[offset * 4 + comp]));
1752 }
1753
1754 static void
ntq_emit_intrinsic(struct vc4_compile * c,nir_intrinsic_instr * instr)1755 ntq_emit_intrinsic(struct vc4_compile *c, nir_intrinsic_instr *instr)
1756 {
1757 nir_const_value *const_offset;
1758 unsigned offset;
1759
1760 switch (instr->intrinsic) {
1761 case nir_intrinsic_load_uniform:
1762 assert(instr->num_components == 1);
1763 const_offset = nir_src_as_const_value(instr->src[0]);
1764 if (const_offset) {
1765 offset = nir_intrinsic_base(instr) + const_offset->u32[0];
1766 assert(offset % 4 == 0);
1767 /* We need dwords */
1768 offset = offset / 4;
1769 ntq_store_dest(c, &instr->dest, 0,
1770 qir_uniform(c, QUNIFORM_UNIFORM,
1771 offset));
1772 } else {
1773 ntq_store_dest(c, &instr->dest, 0,
1774 indirect_uniform_load(c, instr));
1775 }
1776 break;
1777
1778 case nir_intrinsic_load_user_clip_plane:
1779 for (int i = 0; i < instr->num_components; i++) {
1780 ntq_store_dest(c, &instr->dest, i,
1781 qir_uniform(c, QUNIFORM_USER_CLIP_PLANE,
1782 nir_intrinsic_ucp_id(instr) *
1783 4 + i));
1784 }
1785 break;
1786
1787 case nir_intrinsic_load_blend_const_color_r_float:
1788 case nir_intrinsic_load_blend_const_color_g_float:
1789 case nir_intrinsic_load_blend_const_color_b_float:
1790 case nir_intrinsic_load_blend_const_color_a_float:
1791 ntq_store_dest(c, &instr->dest, 0,
1792 qir_uniform(c, QUNIFORM_BLEND_CONST_COLOR_X +
1793 (instr->intrinsic -
1794 nir_intrinsic_load_blend_const_color_r_float),
1795 0));
1796 break;
1797
1798 case nir_intrinsic_load_blend_const_color_rgba8888_unorm:
1799 ntq_store_dest(c, &instr->dest, 0,
1800 qir_uniform(c, QUNIFORM_BLEND_CONST_COLOR_RGBA,
1801 0));
1802 break;
1803
1804 case nir_intrinsic_load_blend_const_color_aaaa8888_unorm:
1805 ntq_store_dest(c, &instr->dest, 0,
1806 qir_uniform(c, QUNIFORM_BLEND_CONST_COLOR_AAAA,
1807 0));
1808 break;
1809
1810 case nir_intrinsic_load_alpha_ref_float:
1811 ntq_store_dest(c, &instr->dest, 0,
1812 qir_uniform(c, QUNIFORM_ALPHA_REF, 0));
1813 break;
1814
1815 case nir_intrinsic_load_sample_mask_in:
1816 ntq_store_dest(c, &instr->dest, 0,
1817 qir_uniform(c, QUNIFORM_SAMPLE_MASK, 0));
1818 break;
1819
1820 case nir_intrinsic_load_front_face:
1821 /* The register contains 0 (front) or 1 (back), and we need to
1822 * turn it into a NIR bool where true means front.
1823 */
1824 ntq_store_dest(c, &instr->dest, 0,
1825 qir_ADD(c,
1826 qir_uniform_ui(c, -1),
1827 qir_reg(QFILE_FRAG_REV_FLAG, 0)));
1828 break;
1829
1830 case nir_intrinsic_load_input:
1831 ntq_emit_load_input(c, instr);
1832 break;
1833
1834 case nir_intrinsic_store_output:
1835 const_offset = nir_src_as_const_value(instr->src[1]);
1836 assert(const_offset && "vc4 doesn't support indirect outputs");
1837 offset = nir_intrinsic_base(instr) + const_offset->u32[0];
1838
1839 /* MSAA color outputs are the only case where we have an
1840 * output that's not lowered to being a store of a single 32
1841 * bit value.
1842 */
1843 if (c->stage == QSTAGE_FRAG && instr->num_components == 4) {
1844 assert(offset == c->output_color_index);
1845 for (int i = 0; i < 4; i++) {
1846 c->sample_colors[i] =
1847 qir_MOV(c, ntq_get_src(c, instr->src[0],
1848 i));
1849 }
1850 } else {
1851 offset = offset * 4 + nir_intrinsic_component(instr);
1852 assert(instr->num_components == 1);
1853 c->outputs[offset] =
1854 qir_MOV(c, ntq_get_src(c, instr->src[0], 0));
1855 c->num_outputs = MAX2(c->num_outputs, offset + 1);
1856 }
1857 break;
1858
1859 case nir_intrinsic_discard:
1860 if (c->execute.file != QFILE_NULL) {
1861 qir_SF(c, c->execute);
1862 qir_MOV_cond(c, QPU_COND_ZS, c->discard,
1863 qir_uniform_ui(c, ~0));
1864 } else {
1865 qir_MOV_dest(c, c->discard, qir_uniform_ui(c, ~0));
1866 }
1867 break;
1868
1869 case nir_intrinsic_discard_if: {
1870 /* true (~0) if we're discarding */
1871 struct qreg cond = ntq_get_src(c, instr->src[0], 0);
1872
1873 if (c->execute.file != QFILE_NULL) {
1874 /* execute == 0 means the channel is active. Invert
1875 * the condition so that we can use zero as "executing
1876 * and discarding."
1877 */
1878 qir_SF(c, qir_AND(c, c->execute, qir_NOT(c, cond)));
1879 qir_MOV_cond(c, QPU_COND_ZS, c->discard, cond);
1880 } else {
1881 qir_OR_dest(c, c->discard, c->discard,
1882 ntq_get_src(c, instr->src[0], 0));
1883 }
1884
1885 break;
1886 }
1887
1888 default:
1889 fprintf(stderr, "Unknown intrinsic: ");
1890 nir_print_instr(&instr->instr, stderr);
1891 fprintf(stderr, "\n");
1892 break;
1893 }
1894 }
1895
1896 /* Clears (activates) the execute flags for any channels whose jump target
1897 * matches this block.
1898 */
1899 static void
ntq_activate_execute_for_block(struct vc4_compile * c)1900 ntq_activate_execute_for_block(struct vc4_compile *c)
1901 {
1902 qir_SF(c, qir_SUB(c,
1903 c->execute,
1904 qir_uniform_ui(c, c->cur_block->index)));
1905 qir_MOV_cond(c, QPU_COND_ZS, c->execute, qir_uniform_ui(c, 0));
1906 }
1907
1908 static void
ntq_emit_if(struct vc4_compile * c,nir_if * if_stmt)1909 ntq_emit_if(struct vc4_compile *c, nir_if *if_stmt)
1910 {
1911 if (!c->vc4->screen->has_control_flow) {
1912 fprintf(stderr,
1913 "IF statement support requires updated kernel.\n");
1914 return;
1915 }
1916
1917 nir_block *nir_else_block = nir_if_first_else_block(if_stmt);
1918 bool empty_else_block =
1919 (nir_else_block == nir_if_last_else_block(if_stmt) &&
1920 exec_list_is_empty(&nir_else_block->instr_list));
1921
1922 struct qblock *then_block = qir_new_block(c);
1923 struct qblock *after_block = qir_new_block(c);
1924 struct qblock *else_block;
1925 if (empty_else_block)
1926 else_block = after_block;
1927 else
1928 else_block = qir_new_block(c);
1929
1930 bool was_top_level = false;
1931 if (c->execute.file == QFILE_NULL) {
1932 c->execute = qir_MOV(c, qir_uniform_ui(c, 0));
1933 was_top_level = true;
1934 }
1935
1936 /* Set ZS for executing (execute == 0) and jumping (if->condition ==
1937 * 0) channels, and then update execute flags for those to point to
1938 * the ELSE block.
1939 */
1940 qir_SF(c, qir_OR(c,
1941 c->execute,
1942 ntq_get_src(c, if_stmt->condition, 0)));
1943 qir_MOV_cond(c, QPU_COND_ZS, c->execute,
1944 qir_uniform_ui(c, else_block->index));
1945
1946 /* Jump to ELSE if nothing is active for THEN, otherwise fall
1947 * through.
1948 */
1949 qir_SF(c, c->execute);
1950 qir_BRANCH(c, QPU_COND_BRANCH_ALL_ZC);
1951 qir_link_blocks(c->cur_block, else_block);
1952 qir_link_blocks(c->cur_block, then_block);
1953
1954 /* Process the THEN block. */
1955 qir_set_emit_block(c, then_block);
1956 ntq_emit_cf_list(c, &if_stmt->then_list);
1957
1958 if (!empty_else_block) {
1959 /* Handle the end of the THEN block. First, all currently
1960 * active channels update their execute flags to point to
1961 * ENDIF
1962 */
1963 qir_SF(c, c->execute);
1964 qir_MOV_cond(c, QPU_COND_ZS, c->execute,
1965 qir_uniform_ui(c, after_block->index));
1966
1967 /* If everything points at ENDIF, then jump there immediately. */
1968 qir_SF(c, qir_SUB(c, c->execute, qir_uniform_ui(c, after_block->index)));
1969 qir_BRANCH(c, QPU_COND_BRANCH_ALL_ZS);
1970 qir_link_blocks(c->cur_block, after_block);
1971 qir_link_blocks(c->cur_block, else_block);
1972
1973 qir_set_emit_block(c, else_block);
1974 ntq_activate_execute_for_block(c);
1975 ntq_emit_cf_list(c, &if_stmt->else_list);
1976 }
1977
1978 qir_link_blocks(c->cur_block, after_block);
1979
1980 qir_set_emit_block(c, after_block);
1981 if (was_top_level) {
1982 c->execute = c->undef;
1983 c->last_top_block = c->cur_block;
1984 } else {
1985 ntq_activate_execute_for_block(c);
1986 }
1987 }
1988
1989 static void
ntq_emit_jump(struct vc4_compile * c,nir_jump_instr * jump)1990 ntq_emit_jump(struct vc4_compile *c, nir_jump_instr *jump)
1991 {
1992 struct qblock *jump_block;
1993 switch (jump->type) {
1994 case nir_jump_break:
1995 jump_block = c->loop_break_block;
1996 break;
1997 case nir_jump_continue:
1998 jump_block = c->loop_cont_block;
1999 break;
2000 default:
2001 unreachable("Unsupported jump type\n");
2002 }
2003
2004 qir_SF(c, c->execute);
2005 qir_MOV_cond(c, QPU_COND_ZS, c->execute,
2006 qir_uniform_ui(c, jump_block->index));
2007
2008 /* Jump to the destination block if everyone has taken the jump. */
2009 qir_SF(c, qir_SUB(c, c->execute, qir_uniform_ui(c, jump_block->index)));
2010 qir_BRANCH(c, QPU_COND_BRANCH_ALL_ZS);
2011 struct qblock *new_block = qir_new_block(c);
2012 qir_link_blocks(c->cur_block, jump_block);
2013 qir_link_blocks(c->cur_block, new_block);
2014 qir_set_emit_block(c, new_block);
2015 }
2016
2017 static void
ntq_emit_instr(struct vc4_compile * c,nir_instr * instr)2018 ntq_emit_instr(struct vc4_compile *c, nir_instr *instr)
2019 {
2020 switch (instr->type) {
2021 case nir_instr_type_alu:
2022 ntq_emit_alu(c, nir_instr_as_alu(instr));
2023 break;
2024
2025 case nir_instr_type_intrinsic:
2026 ntq_emit_intrinsic(c, nir_instr_as_intrinsic(instr));
2027 break;
2028
2029 case nir_instr_type_load_const:
2030 ntq_emit_load_const(c, nir_instr_as_load_const(instr));
2031 break;
2032
2033 case nir_instr_type_ssa_undef:
2034 ntq_emit_ssa_undef(c, nir_instr_as_ssa_undef(instr));
2035 break;
2036
2037 case nir_instr_type_tex:
2038 ntq_emit_tex(c, nir_instr_as_tex(instr));
2039 break;
2040
2041 case nir_instr_type_jump:
2042 ntq_emit_jump(c, nir_instr_as_jump(instr));
2043 break;
2044
2045 default:
2046 fprintf(stderr, "Unknown NIR instr type: ");
2047 nir_print_instr(instr, stderr);
2048 fprintf(stderr, "\n");
2049 abort();
2050 }
2051 }
2052
2053 static void
ntq_emit_block(struct vc4_compile * c,nir_block * block)2054 ntq_emit_block(struct vc4_compile *c, nir_block *block)
2055 {
2056 nir_foreach_instr(instr, block) {
2057 ntq_emit_instr(c, instr);
2058 }
2059 }
2060
2061 static void ntq_emit_cf_list(struct vc4_compile *c, struct exec_list *list);
2062
2063 static void
ntq_emit_loop(struct vc4_compile * c,nir_loop * loop)2064 ntq_emit_loop(struct vc4_compile *c, nir_loop *loop)
2065 {
2066 if (!c->vc4->screen->has_control_flow) {
2067 fprintf(stderr,
2068 "loop support requires updated kernel.\n");
2069 ntq_emit_cf_list(c, &loop->body);
2070 return;
2071 }
2072
2073 bool was_top_level = false;
2074 if (c->execute.file == QFILE_NULL) {
2075 c->execute = qir_MOV(c, qir_uniform_ui(c, 0));
2076 was_top_level = true;
2077 }
2078
2079 struct qblock *save_loop_cont_block = c->loop_cont_block;
2080 struct qblock *save_loop_break_block = c->loop_break_block;
2081
2082 c->loop_cont_block = qir_new_block(c);
2083 c->loop_break_block = qir_new_block(c);
2084
2085 qir_link_blocks(c->cur_block, c->loop_cont_block);
2086 qir_set_emit_block(c, c->loop_cont_block);
2087 ntq_activate_execute_for_block(c);
2088
2089 ntq_emit_cf_list(c, &loop->body);
2090
2091 /* If anything had explicitly continued, or is here at the end of the
2092 * loop, then we need to loop again. SF updates are masked by the
2093 * instruction's condition, so we can do the OR of the two conditions
2094 * within SF.
2095 */
2096 qir_SF(c, c->execute);
2097 struct qinst *cont_check =
2098 qir_SUB_dest(c,
2099 c->undef,
2100 c->execute,
2101 qir_uniform_ui(c, c->loop_cont_block->index));
2102 cont_check->cond = QPU_COND_ZC;
2103 cont_check->sf = true;
2104
2105 qir_BRANCH(c, QPU_COND_BRANCH_ANY_ZS);
2106 qir_link_blocks(c->cur_block, c->loop_cont_block);
2107 qir_link_blocks(c->cur_block, c->loop_break_block);
2108
2109 qir_set_emit_block(c, c->loop_break_block);
2110 if (was_top_level) {
2111 c->execute = c->undef;
2112 c->last_top_block = c->cur_block;
2113 } else {
2114 ntq_activate_execute_for_block(c);
2115 }
2116
2117 c->loop_break_block = save_loop_break_block;
2118 c->loop_cont_block = save_loop_cont_block;
2119 }
2120
2121 static void
ntq_emit_function(struct vc4_compile * c,nir_function_impl * func)2122 ntq_emit_function(struct vc4_compile *c, nir_function_impl *func)
2123 {
2124 fprintf(stderr, "FUNCTIONS not handled.\n");
2125 abort();
2126 }
2127
2128 static void
ntq_emit_cf_list(struct vc4_compile * c,struct exec_list * list)2129 ntq_emit_cf_list(struct vc4_compile *c, struct exec_list *list)
2130 {
2131 foreach_list_typed(nir_cf_node, node, node, list) {
2132 switch (node->type) {
2133 case nir_cf_node_block:
2134 ntq_emit_block(c, nir_cf_node_as_block(node));
2135 break;
2136
2137 case nir_cf_node_if:
2138 ntq_emit_if(c, nir_cf_node_as_if(node));
2139 break;
2140
2141 case nir_cf_node_loop:
2142 ntq_emit_loop(c, nir_cf_node_as_loop(node));
2143 break;
2144
2145 case nir_cf_node_function:
2146 ntq_emit_function(c, nir_cf_node_as_function(node));
2147 break;
2148
2149 default:
2150 fprintf(stderr, "Unknown NIR node type\n");
2151 abort();
2152 }
2153 }
2154 }
2155
2156 static void
ntq_emit_impl(struct vc4_compile * c,nir_function_impl * impl)2157 ntq_emit_impl(struct vc4_compile *c, nir_function_impl *impl)
2158 {
2159 ntq_setup_registers(c, &impl->registers);
2160 ntq_emit_cf_list(c, &impl->body);
2161 }
2162
2163 static void
nir_to_qir(struct vc4_compile * c)2164 nir_to_qir(struct vc4_compile *c)
2165 {
2166 if (c->stage == QSTAGE_FRAG && c->s->info.fs.uses_discard)
2167 c->discard = qir_MOV(c, qir_uniform_ui(c, 0));
2168
2169 ntq_setup_inputs(c);
2170 ntq_setup_outputs(c);
2171 ntq_setup_uniforms(c);
2172 ntq_setup_registers(c, &c->s->registers);
2173
2174 /* Find the main function and emit the body. */
2175 nir_foreach_function(function, c->s) {
2176 assert(strcmp(function->name, "main") == 0);
2177 assert(function->impl);
2178 ntq_emit_impl(c, function->impl);
2179 }
2180 }
2181
2182 static const nir_shader_compiler_options nir_options = {
2183 .lower_extract_byte = true,
2184 .lower_extract_word = true,
2185 .lower_ffma = true,
2186 .lower_flrp32 = true,
2187 .lower_fpow = true,
2188 .lower_fsat = true,
2189 .lower_fsqrt = true,
2190 .lower_negate = true,
2191 .native_integers = true,
2192 .max_unroll_iterations = 32,
2193 };
2194
2195 const void *
vc4_screen_get_compiler_options(struct pipe_screen * pscreen,enum pipe_shader_ir ir,enum pipe_shader_type shader)2196 vc4_screen_get_compiler_options(struct pipe_screen *pscreen,
2197 enum pipe_shader_ir ir,
2198 enum pipe_shader_type shader)
2199 {
2200 return &nir_options;
2201 }
2202
2203 static int
count_nir_instrs(nir_shader * nir)2204 count_nir_instrs(nir_shader *nir)
2205 {
2206 int count = 0;
2207 nir_foreach_function(function, nir) {
2208 if (!function->impl)
2209 continue;
2210 nir_foreach_block(block, function->impl) {
2211 nir_foreach_instr(instr, block)
2212 count++;
2213 }
2214 }
2215 return count;
2216 }
2217
2218 static struct vc4_compile *
vc4_shader_ntq(struct vc4_context * vc4,enum qstage stage,struct vc4_key * key,bool fs_threaded)2219 vc4_shader_ntq(struct vc4_context *vc4, enum qstage stage,
2220 struct vc4_key *key, bool fs_threaded)
2221 {
2222 struct vc4_compile *c = qir_compile_init();
2223
2224 c->vc4 = vc4;
2225 c->stage = stage;
2226 c->shader_state = &key->shader_state->base;
2227 c->program_id = key->shader_state->program_id;
2228 c->variant_id =
2229 p_atomic_inc_return(&key->shader_state->compiled_variant_count);
2230 c->fs_threaded = fs_threaded;
2231
2232 c->key = key;
2233 switch (stage) {
2234 case QSTAGE_FRAG:
2235 c->fs_key = (struct vc4_fs_key *)key;
2236 if (c->fs_key->is_points) {
2237 c->point_x = emit_fragment_varying(c, ~0, 0);
2238 c->point_y = emit_fragment_varying(c, ~0, 0);
2239 } else if (c->fs_key->is_lines) {
2240 c->line_x = emit_fragment_varying(c, ~0, 0);
2241 }
2242 break;
2243 case QSTAGE_VERT:
2244 c->vs_key = (struct vc4_vs_key *)key;
2245 break;
2246 case QSTAGE_COORD:
2247 c->vs_key = (struct vc4_vs_key *)key;
2248 break;
2249 }
2250
2251 c->s = nir_shader_clone(c, key->shader_state->base.ir.nir);
2252
2253 if (stage == QSTAGE_FRAG) {
2254 if (c->fs_key->alpha_test_func != COMPARE_FUNC_ALWAYS) {
2255 NIR_PASS_V(c->s, nir_lower_alpha_test,
2256 c->fs_key->alpha_test_func,
2257 c->fs_key->sample_alpha_to_one &&
2258 c->fs_key->msaa);
2259 }
2260 NIR_PASS_V(c->s, vc4_nir_lower_blend, c);
2261 }
2262
2263 struct nir_lower_tex_options tex_options = {
2264 /* We would need to implement txs, but we don't want the
2265 * int/float conversions
2266 */
2267 .lower_rect = false,
2268
2269 .lower_txp = ~0,
2270
2271 /* Apply swizzles to all samplers. */
2272 .swizzle_result = ~0,
2273 };
2274
2275 /* Lower the format swizzle and ARB_texture_swizzle-style swizzle.
2276 * The format swizzling applies before sRGB decode, and
2277 * ARB_texture_swizzle is the last thing before returning the sample.
2278 */
2279 for (int i = 0; i < ARRAY_SIZE(key->tex); i++) {
2280 enum pipe_format format = c->key->tex[i].format;
2281
2282 if (!format)
2283 continue;
2284
2285 const uint8_t *format_swizzle = vc4_get_format_swizzle(format);
2286
2287 for (int j = 0; j < 4; j++) {
2288 uint8_t arb_swiz = c->key->tex[i].swizzle[j];
2289
2290 if (arb_swiz <= 3) {
2291 tex_options.swizzles[i][j] =
2292 format_swizzle[arb_swiz];
2293 } else {
2294 tex_options.swizzles[i][j] = arb_swiz;
2295 }
2296 }
2297
2298 if (util_format_is_srgb(format))
2299 tex_options.lower_srgb |= (1 << i);
2300 }
2301
2302 NIR_PASS_V(c->s, nir_lower_tex, &tex_options);
2303
2304 if (c->fs_key && c->fs_key->light_twoside)
2305 NIR_PASS_V(c->s, nir_lower_two_sided_color);
2306
2307 if (c->vs_key && c->vs_key->clamp_color)
2308 NIR_PASS_V(c->s, nir_lower_clamp_color_outputs);
2309
2310 if (c->key->ucp_enables) {
2311 if (stage == QSTAGE_FRAG) {
2312 NIR_PASS_V(c->s, nir_lower_clip_fs, c->key->ucp_enables);
2313 } else {
2314 NIR_PASS_V(c->s, nir_lower_clip_vs, c->key->ucp_enables);
2315 NIR_PASS_V(c->s, nir_lower_io_to_scalar,
2316 nir_var_shader_out);
2317 }
2318 }
2319
2320 /* FS input scalarizing must happen after nir_lower_two_sided_color,
2321 * which only handles a vec4 at a time. Similarly, VS output
2322 * scalarizing must happen after nir_lower_clip_vs.
2323 */
2324 if (c->stage == QSTAGE_FRAG)
2325 NIR_PASS_V(c->s, nir_lower_io_to_scalar, nir_var_shader_in);
2326 else
2327 NIR_PASS_V(c->s, nir_lower_io_to_scalar, nir_var_shader_out);
2328
2329 NIR_PASS_V(c->s, vc4_nir_lower_io, c);
2330 NIR_PASS_V(c->s, vc4_nir_lower_txf_ms, c);
2331 NIR_PASS_V(c->s, nir_lower_idiv);
2332
2333 vc4_optimize_nir(c->s);
2334
2335 NIR_PASS_V(c->s, nir_convert_from_ssa, true);
2336
2337 if (vc4_debug & VC4_DEBUG_SHADERDB) {
2338 fprintf(stderr, "SHADER-DB: %s prog %d/%d: %d NIR instructions\n",
2339 qir_get_stage_name(c->stage),
2340 c->program_id, c->variant_id,
2341 count_nir_instrs(c->s));
2342 }
2343
2344 if (vc4_debug & VC4_DEBUG_NIR) {
2345 fprintf(stderr, "%s prog %d/%d NIR:\n",
2346 qir_get_stage_name(c->stage),
2347 c->program_id, c->variant_id);
2348 nir_print_shader(c->s, stderr);
2349 }
2350
2351 nir_to_qir(c);
2352
2353 switch (stage) {
2354 case QSTAGE_FRAG:
2355 /* FS threading requires that the thread execute
2356 * QPU_SIG_LAST_THREAD_SWITCH exactly once before terminating
2357 * (with no other THRSW afterwards, obviously). If we didn't
2358 * fetch a texture at a top level block, this wouldn't be
2359 * true.
2360 */
2361 if (c->fs_threaded && !c->last_thrsw_at_top_level) {
2362 c->failed = true;
2363 return c;
2364 }
2365
2366 emit_frag_end(c);
2367 break;
2368 case QSTAGE_VERT:
2369 emit_vert_end(c,
2370 c->vs_key->fs_inputs->input_slots,
2371 c->vs_key->fs_inputs->num_inputs);
2372 break;
2373 case QSTAGE_COORD:
2374 emit_coord_end(c);
2375 break;
2376 }
2377
2378 if (vc4_debug & VC4_DEBUG_QIR) {
2379 fprintf(stderr, "%s prog %d/%d pre-opt QIR:\n",
2380 qir_get_stage_name(c->stage),
2381 c->program_id, c->variant_id);
2382 qir_dump(c);
2383 fprintf(stderr, "\n");
2384 }
2385
2386 qir_optimize(c);
2387 qir_lower_uniforms(c);
2388
2389 qir_schedule_instructions(c);
2390 qir_emit_uniform_stream_resets(c);
2391
2392 if (vc4_debug & VC4_DEBUG_QIR) {
2393 fprintf(stderr, "%s prog %d/%d QIR:\n",
2394 qir_get_stage_name(c->stage),
2395 c->program_id, c->variant_id);
2396 qir_dump(c);
2397 fprintf(stderr, "\n");
2398 }
2399
2400 qir_reorder_uniforms(c);
2401 vc4_generate_code(vc4, c);
2402
2403 if (vc4_debug & VC4_DEBUG_SHADERDB) {
2404 fprintf(stderr, "SHADER-DB: %s prog %d/%d: %d instructions\n",
2405 qir_get_stage_name(c->stage),
2406 c->program_id, c->variant_id,
2407 c->qpu_inst_count);
2408 fprintf(stderr, "SHADER-DB: %s prog %d/%d: %d uniforms\n",
2409 qir_get_stage_name(c->stage),
2410 c->program_id, c->variant_id,
2411 c->num_uniforms);
2412 }
2413
2414 ralloc_free(c->s);
2415
2416 return c;
2417 }
2418
2419 static void *
vc4_shader_state_create(struct pipe_context * pctx,const struct pipe_shader_state * cso)2420 vc4_shader_state_create(struct pipe_context *pctx,
2421 const struct pipe_shader_state *cso)
2422 {
2423 struct vc4_context *vc4 = vc4_context(pctx);
2424 struct vc4_uncompiled_shader *so = CALLOC_STRUCT(vc4_uncompiled_shader);
2425 if (!so)
2426 return NULL;
2427
2428 so->program_id = vc4->next_uncompiled_program_id++;
2429
2430 nir_shader *s;
2431
2432 if (cso->type == PIPE_SHADER_IR_NIR) {
2433 /* The backend takes ownership of the NIR shader on state
2434 * creation.
2435 */
2436 s = cso->ir.nir;
2437
2438 NIR_PASS_V(s, nir_lower_io, nir_var_all, type_size,
2439 (nir_lower_io_options)0);
2440 } else {
2441 assert(cso->type == PIPE_SHADER_IR_TGSI);
2442
2443 if (vc4_debug & VC4_DEBUG_TGSI) {
2444 fprintf(stderr, "prog %d TGSI:\n",
2445 so->program_id);
2446 tgsi_dump(cso->tokens, 0);
2447 fprintf(stderr, "\n");
2448 }
2449 s = tgsi_to_nir(cso->tokens, &nir_options);
2450 }
2451
2452 NIR_PASS_V(s, nir_opt_global_to_local);
2453 NIR_PASS_V(s, nir_lower_regs_to_ssa);
2454 NIR_PASS_V(s, nir_normalize_cubemap_coords);
2455
2456 NIR_PASS_V(s, nir_lower_load_const_to_scalar);
2457
2458 vc4_optimize_nir(s);
2459
2460 NIR_PASS_V(s, nir_remove_dead_variables, nir_var_local);
2461
2462 /* Garbage collect dead instructions */
2463 nir_sweep(s);
2464
2465 so->base.type = PIPE_SHADER_IR_NIR;
2466 so->base.ir.nir = s;
2467
2468 if (vc4_debug & VC4_DEBUG_NIR) {
2469 fprintf(stderr, "%s prog %d NIR:\n",
2470 gl_shader_stage_name(s->info.stage),
2471 so->program_id);
2472 nir_print_shader(s, stderr);
2473 fprintf(stderr, "\n");
2474 }
2475
2476 return so;
2477 }
2478
2479 static void
copy_uniform_state_to_shader(struct vc4_compiled_shader * shader,struct vc4_compile * c)2480 copy_uniform_state_to_shader(struct vc4_compiled_shader *shader,
2481 struct vc4_compile *c)
2482 {
2483 int count = c->num_uniforms;
2484 struct vc4_shader_uniform_info *uinfo = &shader->uniforms;
2485
2486 uinfo->count = count;
2487 uinfo->data = ralloc_array(shader, uint32_t, count);
2488 memcpy(uinfo->data, c->uniform_data,
2489 count * sizeof(*uinfo->data));
2490 uinfo->contents = ralloc_array(shader, enum quniform_contents, count);
2491 memcpy(uinfo->contents, c->uniform_contents,
2492 count * sizeof(*uinfo->contents));
2493 uinfo->num_texture_samples = c->num_texture_samples;
2494
2495 vc4_set_shader_uniform_dirty_flags(shader);
2496 }
2497
2498 static void
vc4_setup_compiled_fs_inputs(struct vc4_context * vc4,struct vc4_compile * c,struct vc4_compiled_shader * shader)2499 vc4_setup_compiled_fs_inputs(struct vc4_context *vc4, struct vc4_compile *c,
2500 struct vc4_compiled_shader *shader)
2501 {
2502 struct vc4_fs_inputs inputs;
2503
2504 memset(&inputs, 0, sizeof(inputs));
2505 inputs.input_slots = ralloc_array(shader,
2506 struct vc4_varying_slot,
2507 c->num_input_slots);
2508
2509 bool input_live[c->num_input_slots];
2510
2511 memset(input_live, 0, sizeof(input_live));
2512 qir_for_each_inst_inorder(inst, c) {
2513 for (int i = 0; i < qir_get_nsrc(inst); i++) {
2514 if (inst->src[i].file == QFILE_VARY)
2515 input_live[inst->src[i].index] = true;
2516 }
2517 }
2518
2519 for (int i = 0; i < c->num_input_slots; i++) {
2520 struct vc4_varying_slot *slot = &c->input_slots[i];
2521
2522 if (!input_live[i])
2523 continue;
2524
2525 /* Skip non-VS-output inputs. */
2526 if (slot->slot == (uint8_t)~0)
2527 continue;
2528
2529 if (slot->slot == VARYING_SLOT_COL0 ||
2530 slot->slot == VARYING_SLOT_COL1 ||
2531 slot->slot == VARYING_SLOT_BFC0 ||
2532 slot->slot == VARYING_SLOT_BFC1) {
2533 shader->color_inputs |= (1 << inputs.num_inputs);
2534 }
2535
2536 inputs.input_slots[inputs.num_inputs] = *slot;
2537 inputs.num_inputs++;
2538 }
2539 shader->num_inputs = inputs.num_inputs;
2540
2541 /* Add our set of inputs to the set of all inputs seen. This way, we
2542 * can have a single pointer that identifies an FS inputs set,
2543 * allowing VS to avoid recompiling when the FS is recompiled (or a
2544 * new one is bound using separate shader objects) but the inputs
2545 * don't change.
2546 */
2547 struct set_entry *entry = _mesa_set_search(vc4->fs_inputs_set, &inputs);
2548 if (entry) {
2549 shader->fs_inputs = entry->key;
2550 ralloc_free(inputs.input_slots);
2551 } else {
2552 struct vc4_fs_inputs *alloc_inputs;
2553
2554 alloc_inputs = rzalloc(vc4->fs_inputs_set, struct vc4_fs_inputs);
2555 memcpy(alloc_inputs, &inputs, sizeof(inputs));
2556 ralloc_steal(alloc_inputs, inputs.input_slots);
2557 _mesa_set_add(vc4->fs_inputs_set, alloc_inputs);
2558
2559 shader->fs_inputs = alloc_inputs;
2560 }
2561 }
2562
2563 static struct vc4_compiled_shader *
vc4_get_compiled_shader(struct vc4_context * vc4,enum qstage stage,struct vc4_key * key)2564 vc4_get_compiled_shader(struct vc4_context *vc4, enum qstage stage,
2565 struct vc4_key *key)
2566 {
2567 struct hash_table *ht;
2568 uint32_t key_size;
2569 bool try_threading;
2570
2571 if (stage == QSTAGE_FRAG) {
2572 ht = vc4->fs_cache;
2573 key_size = sizeof(struct vc4_fs_key);
2574 try_threading = vc4->screen->has_threaded_fs;
2575 } else {
2576 ht = vc4->vs_cache;
2577 key_size = sizeof(struct vc4_vs_key);
2578 try_threading = false;
2579 }
2580
2581 struct vc4_compiled_shader *shader;
2582 struct hash_entry *entry = _mesa_hash_table_search(ht, key);
2583 if (entry)
2584 return entry->data;
2585
2586 struct vc4_compile *c = vc4_shader_ntq(vc4, stage, key, try_threading);
2587 /* If the FS failed to compile threaded, fall back to single threaded. */
2588 if (try_threading && c->failed) {
2589 qir_compile_destroy(c);
2590 c = vc4_shader_ntq(vc4, stage, key, false);
2591 }
2592
2593 shader = rzalloc(NULL, struct vc4_compiled_shader);
2594
2595 shader->program_id = vc4->next_compiled_program_id++;
2596 if (stage == QSTAGE_FRAG) {
2597 vc4_setup_compiled_fs_inputs(vc4, c, shader);
2598
2599 /* Note: the temporary clone in c->s has been freed. */
2600 nir_shader *orig_shader = key->shader_state->base.ir.nir;
2601 if (orig_shader->info.outputs_written & (1 << FRAG_RESULT_DEPTH))
2602 shader->disable_early_z = true;
2603 } else {
2604 shader->num_inputs = c->num_inputs;
2605
2606 shader->vattr_offsets[0] = 0;
2607 for (int i = 0; i < 8; i++) {
2608 shader->vattr_offsets[i + 1] =
2609 shader->vattr_offsets[i] + c->vattr_sizes[i];
2610
2611 if (c->vattr_sizes[i])
2612 shader->vattrs_live |= (1 << i);
2613 }
2614 }
2615
2616 shader->failed = c->failed;
2617 if (c->failed) {
2618 shader->failed = true;
2619 } else {
2620 copy_uniform_state_to_shader(shader, c);
2621 shader->bo = vc4_bo_alloc_shader(vc4->screen, c->qpu_insts,
2622 c->qpu_inst_count *
2623 sizeof(uint64_t));
2624 }
2625
2626 shader->fs_threaded = c->fs_threaded;
2627
2628 /* Copy the compiler UBO range state to the compiled shader, dropping
2629 * out arrays that were never referenced by an indirect load.
2630 *
2631 * (Note that QIR dead code elimination of an array access still
2632 * leaves that array alive, though)
2633 */
2634 if (c->num_ubo_ranges) {
2635 shader->num_ubo_ranges = c->num_ubo_ranges;
2636 shader->ubo_ranges = ralloc_array(shader, struct vc4_ubo_range,
2637 c->num_ubo_ranges);
2638 uint32_t j = 0;
2639 for (int i = 0; i < c->num_uniform_ranges; i++) {
2640 struct vc4_compiler_ubo_range *range =
2641 &c->ubo_ranges[i];
2642 if (!range->used)
2643 continue;
2644
2645 shader->ubo_ranges[j].dst_offset = range->dst_offset;
2646 shader->ubo_ranges[j].src_offset = range->src_offset;
2647 shader->ubo_ranges[j].size = range->size;
2648 shader->ubo_size += c->ubo_ranges[i].size;
2649 j++;
2650 }
2651 }
2652 if (shader->ubo_size) {
2653 if (vc4_debug & VC4_DEBUG_SHADERDB) {
2654 fprintf(stderr, "SHADER-DB: %s prog %d/%d: %d UBO uniforms\n",
2655 qir_get_stage_name(c->stage),
2656 c->program_id, c->variant_id,
2657 shader->ubo_size / 4);
2658 }
2659 }
2660
2661 if ((vc4_debug & VC4_DEBUG_SHADERDB) && stage == QSTAGE_FRAG) {
2662 fprintf(stderr, "SHADER-DB: %s prog %d/%d: %d FS threads\n",
2663 qir_get_stage_name(c->stage),
2664 c->program_id, c->variant_id,
2665 1 + shader->fs_threaded);
2666 }
2667
2668 qir_compile_destroy(c);
2669
2670 struct vc4_key *dup_key;
2671 dup_key = rzalloc_size(shader, key_size); /* TODO: don't use rzalloc */
2672 memcpy(dup_key, key, key_size);
2673 _mesa_hash_table_insert(ht, dup_key, shader);
2674
2675 return shader;
2676 }
2677
2678 static void
vc4_setup_shared_key(struct vc4_context * vc4,struct vc4_key * key,struct vc4_texture_stateobj * texstate)2679 vc4_setup_shared_key(struct vc4_context *vc4, struct vc4_key *key,
2680 struct vc4_texture_stateobj *texstate)
2681 {
2682 for (int i = 0; i < texstate->num_textures; i++) {
2683 struct pipe_sampler_view *sampler = texstate->textures[i];
2684 struct vc4_sampler_view *vc4_sampler = vc4_sampler_view(sampler);
2685 struct pipe_sampler_state *sampler_state =
2686 texstate->samplers[i];
2687
2688 if (!sampler)
2689 continue;
2690
2691 key->tex[i].format = sampler->format;
2692 key->tex[i].swizzle[0] = sampler->swizzle_r;
2693 key->tex[i].swizzle[1] = sampler->swizzle_g;
2694 key->tex[i].swizzle[2] = sampler->swizzle_b;
2695 key->tex[i].swizzle[3] = sampler->swizzle_a;
2696
2697 if (sampler->texture->nr_samples > 1) {
2698 key->tex[i].msaa_width = sampler->texture->width0;
2699 key->tex[i].msaa_height = sampler->texture->height0;
2700 } else if (sampler){
2701 key->tex[i].compare_mode = sampler_state->compare_mode;
2702 key->tex[i].compare_func = sampler_state->compare_func;
2703 key->tex[i].wrap_s = sampler_state->wrap_s;
2704 key->tex[i].wrap_t = sampler_state->wrap_t;
2705 key->tex[i].force_first_level =
2706 vc4_sampler->force_first_level;
2707 }
2708 }
2709
2710 key->ucp_enables = vc4->rasterizer->base.clip_plane_enable;
2711 }
2712
2713 static void
vc4_update_compiled_fs(struct vc4_context * vc4,uint8_t prim_mode)2714 vc4_update_compiled_fs(struct vc4_context *vc4, uint8_t prim_mode)
2715 {
2716 struct vc4_job *job = vc4->job;
2717 struct vc4_fs_key local_key;
2718 struct vc4_fs_key *key = &local_key;
2719
2720 if (!(vc4->dirty & (VC4_DIRTY_PRIM_MODE |
2721 VC4_DIRTY_BLEND |
2722 VC4_DIRTY_FRAMEBUFFER |
2723 VC4_DIRTY_ZSA |
2724 VC4_DIRTY_RASTERIZER |
2725 VC4_DIRTY_SAMPLE_MASK |
2726 VC4_DIRTY_FRAGTEX |
2727 VC4_DIRTY_UNCOMPILED_FS))) {
2728 return;
2729 }
2730
2731 memset(key, 0, sizeof(*key));
2732 vc4_setup_shared_key(vc4, &key->base, &vc4->fragtex);
2733 key->base.shader_state = vc4->prog.bind_fs;
2734 key->is_points = (prim_mode == PIPE_PRIM_POINTS);
2735 key->is_lines = (prim_mode >= PIPE_PRIM_LINES &&
2736 prim_mode <= PIPE_PRIM_LINE_STRIP);
2737 key->blend = vc4->blend->rt[0];
2738 if (vc4->blend->logicop_enable) {
2739 key->logicop_func = vc4->blend->logicop_func;
2740 } else {
2741 key->logicop_func = PIPE_LOGICOP_COPY;
2742 }
2743 if (job->msaa) {
2744 key->msaa = vc4->rasterizer->base.multisample;
2745 key->sample_coverage = (vc4->sample_mask != (1 << VC4_MAX_SAMPLES) - 1);
2746 key->sample_alpha_to_coverage = vc4->blend->alpha_to_coverage;
2747 key->sample_alpha_to_one = vc4->blend->alpha_to_one;
2748 }
2749
2750 if (vc4->framebuffer.cbufs[0])
2751 key->color_format = vc4->framebuffer.cbufs[0]->format;
2752
2753 key->stencil_enabled = vc4->zsa->stencil_uniforms[0] != 0;
2754 key->stencil_twoside = vc4->zsa->stencil_uniforms[1] != 0;
2755 key->stencil_full_writemasks = vc4->zsa->stencil_uniforms[2] != 0;
2756 key->depth_enabled = (vc4->zsa->base.depth.enabled ||
2757 key->stencil_enabled);
2758 if (vc4->zsa->base.alpha.enabled)
2759 key->alpha_test_func = vc4->zsa->base.alpha.func;
2760 else
2761 key->alpha_test_func = COMPARE_FUNC_ALWAYS;
2762
2763 if (key->is_points) {
2764 key->point_sprite_mask =
2765 vc4->rasterizer->base.sprite_coord_enable;
2766 key->point_coord_upper_left =
2767 (vc4->rasterizer->base.sprite_coord_mode ==
2768 PIPE_SPRITE_COORD_UPPER_LEFT);
2769 }
2770
2771 key->light_twoside = vc4->rasterizer->base.light_twoside;
2772
2773 struct vc4_compiled_shader *old_fs = vc4->prog.fs;
2774 vc4->prog.fs = vc4_get_compiled_shader(vc4, QSTAGE_FRAG, &key->base);
2775 if (vc4->prog.fs == old_fs)
2776 return;
2777
2778 vc4->dirty |= VC4_DIRTY_COMPILED_FS;
2779
2780 if (vc4->rasterizer->base.flatshade &&
2781 (!old_fs || vc4->prog.fs->color_inputs != old_fs->color_inputs)) {
2782 vc4->dirty |= VC4_DIRTY_FLAT_SHADE_FLAGS;
2783 }
2784
2785 if (!old_fs || vc4->prog.fs->fs_inputs != old_fs->fs_inputs)
2786 vc4->dirty |= VC4_DIRTY_FS_INPUTS;
2787 }
2788
2789 static void
vc4_update_compiled_vs(struct vc4_context * vc4,uint8_t prim_mode)2790 vc4_update_compiled_vs(struct vc4_context *vc4, uint8_t prim_mode)
2791 {
2792 struct vc4_vs_key local_key;
2793 struct vc4_vs_key *key = &local_key;
2794
2795 if (!(vc4->dirty & (VC4_DIRTY_PRIM_MODE |
2796 VC4_DIRTY_RASTERIZER |
2797 VC4_DIRTY_VERTTEX |
2798 VC4_DIRTY_VTXSTATE |
2799 VC4_DIRTY_UNCOMPILED_VS |
2800 VC4_DIRTY_FS_INPUTS))) {
2801 return;
2802 }
2803
2804 memset(key, 0, sizeof(*key));
2805 vc4_setup_shared_key(vc4, &key->base, &vc4->verttex);
2806 key->base.shader_state = vc4->prog.bind_vs;
2807 key->fs_inputs = vc4->prog.fs->fs_inputs;
2808 key->clamp_color = vc4->rasterizer->base.clamp_vertex_color;
2809
2810 for (int i = 0; i < ARRAY_SIZE(key->attr_formats); i++)
2811 key->attr_formats[i] = vc4->vtx->pipe[i].src_format;
2812
2813 key->per_vertex_point_size =
2814 (prim_mode == PIPE_PRIM_POINTS &&
2815 vc4->rasterizer->base.point_size_per_vertex);
2816
2817 struct vc4_compiled_shader *vs =
2818 vc4_get_compiled_shader(vc4, QSTAGE_VERT, &key->base);
2819 if (vs != vc4->prog.vs) {
2820 vc4->prog.vs = vs;
2821 vc4->dirty |= VC4_DIRTY_COMPILED_VS;
2822 }
2823
2824 key->is_coord = true;
2825 /* Coord shaders don't care what the FS inputs are. */
2826 key->fs_inputs = NULL;
2827 struct vc4_compiled_shader *cs =
2828 vc4_get_compiled_shader(vc4, QSTAGE_COORD, &key->base);
2829 if (cs != vc4->prog.cs) {
2830 vc4->prog.cs = cs;
2831 vc4->dirty |= VC4_DIRTY_COMPILED_CS;
2832 }
2833 }
2834
2835 bool
vc4_update_compiled_shaders(struct vc4_context * vc4,uint8_t prim_mode)2836 vc4_update_compiled_shaders(struct vc4_context *vc4, uint8_t prim_mode)
2837 {
2838 vc4_update_compiled_fs(vc4, prim_mode);
2839 vc4_update_compiled_vs(vc4, prim_mode);
2840
2841 return !(vc4->prog.cs->failed ||
2842 vc4->prog.vs->failed ||
2843 vc4->prog.fs->failed);
2844 }
2845
2846 static uint32_t
fs_cache_hash(const void * key)2847 fs_cache_hash(const void *key)
2848 {
2849 return _mesa_hash_data(key, sizeof(struct vc4_fs_key));
2850 }
2851
2852 static uint32_t
vs_cache_hash(const void * key)2853 vs_cache_hash(const void *key)
2854 {
2855 return _mesa_hash_data(key, sizeof(struct vc4_vs_key));
2856 }
2857
2858 static bool
fs_cache_compare(const void * key1,const void * key2)2859 fs_cache_compare(const void *key1, const void *key2)
2860 {
2861 return memcmp(key1, key2, sizeof(struct vc4_fs_key)) == 0;
2862 }
2863
2864 static bool
vs_cache_compare(const void * key1,const void * key2)2865 vs_cache_compare(const void *key1, const void *key2)
2866 {
2867 return memcmp(key1, key2, sizeof(struct vc4_vs_key)) == 0;
2868 }
2869
2870 static uint32_t
fs_inputs_hash(const void * key)2871 fs_inputs_hash(const void *key)
2872 {
2873 const struct vc4_fs_inputs *inputs = key;
2874
2875 return _mesa_hash_data(inputs->input_slots,
2876 sizeof(*inputs->input_slots) *
2877 inputs->num_inputs);
2878 }
2879
2880 static bool
fs_inputs_compare(const void * key1,const void * key2)2881 fs_inputs_compare(const void *key1, const void *key2)
2882 {
2883 const struct vc4_fs_inputs *inputs1 = key1;
2884 const struct vc4_fs_inputs *inputs2 = key2;
2885
2886 return (inputs1->num_inputs == inputs2->num_inputs &&
2887 memcmp(inputs1->input_slots,
2888 inputs2->input_slots,
2889 sizeof(*inputs1->input_slots) *
2890 inputs1->num_inputs) == 0);
2891 }
2892
2893 static void
delete_from_cache_if_matches(struct hash_table * ht,struct vc4_compiled_shader ** last_compile,struct hash_entry * entry,struct vc4_uncompiled_shader * so)2894 delete_from_cache_if_matches(struct hash_table *ht,
2895 struct vc4_compiled_shader **last_compile,
2896 struct hash_entry *entry,
2897 struct vc4_uncompiled_shader *so)
2898 {
2899 const struct vc4_key *key = entry->key;
2900
2901 if (key->shader_state == so) {
2902 struct vc4_compiled_shader *shader = entry->data;
2903 _mesa_hash_table_remove(ht, entry);
2904 vc4_bo_unreference(&shader->bo);
2905
2906 if (shader == *last_compile)
2907 *last_compile = NULL;
2908
2909 ralloc_free(shader);
2910 }
2911 }
2912
2913 static void
vc4_shader_state_delete(struct pipe_context * pctx,void * hwcso)2914 vc4_shader_state_delete(struct pipe_context *pctx, void *hwcso)
2915 {
2916 struct vc4_context *vc4 = vc4_context(pctx);
2917 struct vc4_uncompiled_shader *so = hwcso;
2918
2919 struct hash_entry *entry;
2920 hash_table_foreach(vc4->fs_cache, entry) {
2921 delete_from_cache_if_matches(vc4->fs_cache, &vc4->prog.fs,
2922 entry, so);
2923 }
2924 hash_table_foreach(vc4->vs_cache, entry) {
2925 delete_from_cache_if_matches(vc4->vs_cache, &vc4->prog.vs,
2926 entry, so);
2927 }
2928
2929 ralloc_free(so->base.ir.nir);
2930 free(so);
2931 }
2932
2933 static void
vc4_fp_state_bind(struct pipe_context * pctx,void * hwcso)2934 vc4_fp_state_bind(struct pipe_context *pctx, void *hwcso)
2935 {
2936 struct vc4_context *vc4 = vc4_context(pctx);
2937 vc4->prog.bind_fs = hwcso;
2938 vc4->dirty |= VC4_DIRTY_UNCOMPILED_FS;
2939 }
2940
2941 static void
vc4_vp_state_bind(struct pipe_context * pctx,void * hwcso)2942 vc4_vp_state_bind(struct pipe_context *pctx, void *hwcso)
2943 {
2944 struct vc4_context *vc4 = vc4_context(pctx);
2945 vc4->prog.bind_vs = hwcso;
2946 vc4->dirty |= VC4_DIRTY_UNCOMPILED_VS;
2947 }
2948
2949 void
vc4_program_init(struct pipe_context * pctx)2950 vc4_program_init(struct pipe_context *pctx)
2951 {
2952 struct vc4_context *vc4 = vc4_context(pctx);
2953
2954 pctx->create_vs_state = vc4_shader_state_create;
2955 pctx->delete_vs_state = vc4_shader_state_delete;
2956
2957 pctx->create_fs_state = vc4_shader_state_create;
2958 pctx->delete_fs_state = vc4_shader_state_delete;
2959
2960 pctx->bind_fs_state = vc4_fp_state_bind;
2961 pctx->bind_vs_state = vc4_vp_state_bind;
2962
2963 vc4->fs_cache = _mesa_hash_table_create(pctx, fs_cache_hash,
2964 fs_cache_compare);
2965 vc4->vs_cache = _mesa_hash_table_create(pctx, vs_cache_hash,
2966 vs_cache_compare);
2967 vc4->fs_inputs_set = _mesa_set_create(pctx, fs_inputs_hash,
2968 fs_inputs_compare);
2969 }
2970
2971 void
vc4_program_fini(struct pipe_context * pctx)2972 vc4_program_fini(struct pipe_context *pctx)
2973 {
2974 struct vc4_context *vc4 = vc4_context(pctx);
2975
2976 struct hash_entry *entry;
2977 hash_table_foreach(vc4->fs_cache, entry) {
2978 struct vc4_compiled_shader *shader = entry->data;
2979 vc4_bo_unreference(&shader->bo);
2980 ralloc_free(shader);
2981 _mesa_hash_table_remove(vc4->fs_cache, entry);
2982 }
2983
2984 hash_table_foreach(vc4->vs_cache, entry) {
2985 struct vc4_compiled_shader *shader = entry->data;
2986 vc4_bo_unreference(&shader->bo);
2987 ralloc_free(shader);
2988 _mesa_hash_table_remove(vc4->vs_cache, entry);
2989 }
2990 }
2991