projects / mesa.git / blob
? search:


84add5253ff780099d117e64ec1dae65a27006b0
[mesa.git] / src / gallium / drivers / vc4 / vc4_program.c
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/u_hash.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 "nir/tgsi_to_nir.h"
37 #include "vc4_context.h"
38 #include "vc4_qpu.h"
39 #include "vc4_qir.h"
40 #include "mesa/state_tracker/st_glsl_types.h"
41 #ifdef USE_VC4_SIMULATOR
42 #include "simpenrose/simpenrose.h"
43 #endif
44
45 static struct qreg
46 ntq_get_src(struct vc4_compile *c, nir_src src, int i);
47 static void
48 ntq_emit_cf_list(struct vc4_compile *c, struct exec_list *list);
49
50 static void
51 resize_qreg_array(struct vc4_compile *c,
52 struct qreg **regs,
53 uint32_t *size,
54 uint32_t decl_size)
55 {
56 if (*size >= decl_size)
57 return;
58
59 uint32_t old_size = *size;
60 *size = MAX2(*size * 2, decl_size);
61 *regs = reralloc(c, *regs, struct qreg, *size);
62 if (!*regs) {
63 fprintf(stderr, "Malloc failure\n");
64 abort();
65 }
66
67 for (uint32_t i = old_size; i < *size; i++)
68 (*regs)[i] = c->undef;
69 }
70
71 static struct qreg
72 indirect_uniform_load(struct vc4_compile *c, nir_intrinsic_instr *intr)
73 {
74 struct qreg indirect_offset = ntq_get_src(c, intr->src[0], 0);
75 uint32_t offset = nir_intrinsic_base(intr);
76 struct vc4_compiler_ubo_range *range = NULL;
77 unsigned i;
78 for (i = 0; i < c->num_uniform_ranges; i++) {
79 range = &c->ubo_ranges[i];
80 if (offset >= range->src_offset &&
81 offset < range->src_offset + range->size) {
82 break;
83 }
84 }
85 /* The driver-location-based offset always has to be within a declared
86 * uniform range.
87 */
88 assert(range);
89 if (!range->used) {
90 range->used = true;
91 range->dst_offset = c->next_ubo_dst_offset;
92 c->next_ubo_dst_offset += range->size;
93 c->num_ubo_ranges++;
94 }
95
96 offset -= range->src_offset;
97
98 /* Adjust for where we stored the TGSI register base. */
99 indirect_offset = qir_ADD(c, indirect_offset,
100 qir_uniform_ui(c, (range->dst_offset +
101 offset)));
102
103 /* Clamp to [0, array size). Note that MIN/MAX are signed. */
104 indirect_offset = qir_MAX(c, indirect_offset, qir_uniform_ui(c, 0));
105 indirect_offset = qir_MIN(c, indirect_offset,
106 qir_uniform_ui(c, (range->dst_offset +
107 range->size - 4)));
108
109 qir_TEX_DIRECT(c, indirect_offset, qir_uniform(c, QUNIFORM_UBO_ADDR, 0));
110 c->num_texture_samples++;
111 return qir_TEX_RESULT(c);
112 }
113
114 nir_ssa_def *
115 vc4_nir_get_swizzled_channel(nir_builder *b, nir_ssa_def **srcs, int swiz)
116 {
117 switch (swiz) {
118 default:
119 case PIPE_SWIZZLE_NONE:
120 fprintf(stderr, "warning: unknown swizzle\n");
121 /* FALLTHROUGH */
122 case PIPE_SWIZZLE_0:
123 return nir_imm_float(b, 0.0);
124 case PIPE_SWIZZLE_1:
125 return nir_imm_float(b, 1.0);
126 case PIPE_SWIZZLE_X:
127 case PIPE_SWIZZLE_Y:
128 case PIPE_SWIZZLE_Z:
129 case PIPE_SWIZZLE_W:
130 return srcs[swiz];
131 }
132 }
133
134 static struct qreg *
135 ntq_init_ssa_def(struct vc4_compile *c, nir_ssa_def *def)
136 {
137 struct qreg *qregs = ralloc_array(c->def_ht, struct qreg,
138 def->num_components);
139 _mesa_hash_table_insert(c->def_ht, def, qregs);
140 return qregs;
141 }
142
143 static void
144 ntq_store_dest(struct vc4_compile *c, nir_dest *dest, int chan,
145 struct qreg result)
146 {
147 if (dest->is_ssa) {
148 assert(chan < dest->ssa.num_components);
149
150 struct qreg *qregs;
151 struct hash_entry *entry =
152 _mesa_hash_table_search(c->def_ht, &dest->ssa);
153
154 if (entry)
155 qregs = entry->data;
156 else
157 qregs = ntq_init_ssa_def(c, &dest->ssa);
158
159 qregs[chan] = result;
160 } else {
161 nir_register *reg = dest->reg.reg;
162 assert(dest->reg.base_offset == 0);
163 assert(reg->num_array_elems == 0);
164 struct hash_entry *entry =
165 _mesa_hash_table_search(c->def_ht, reg);
166 struct qreg *qregs = entry->data;
167
168 /* Conditionally move the result to the destination if the
169 * channel is active.
170 */
171 if (c->execute.file != QFILE_NULL) {
172 qir_SF(c, c->execute);
173 qir_MOV_cond(c, QPU_COND_ZS, qregs[chan], result);
174 } else {
175 qir_MOV_dest(c, qregs[chan], result);
176 }
177 }
178 }
179
180 static struct qreg *
181 ntq_get_dest(struct vc4_compile *c, nir_dest *dest)
182 {
183 if (dest->is_ssa) {
184 struct qreg *qregs = ntq_init_ssa_def(c, &dest->ssa);
185 for (int i = 0; i < dest->ssa.num_components; i++)
186 qregs[i] = c->undef;
187 return qregs;
188 } else {
189 nir_register *reg = dest->reg.reg;
190 assert(dest->reg.base_offset == 0);
191 assert(reg->num_array_elems == 0);
192 struct hash_entry *entry =
193 _mesa_hash_table_search(c->def_ht, reg);
194 return entry->data;
195 }
196 }
197
198 static struct qreg
199 ntq_get_src(struct vc4_compile *c, nir_src src, int i)
200 {
201 struct hash_entry *entry;
202 if (src.is_ssa) {
203 entry = _mesa_hash_table_search(c->def_ht, src.ssa);
204 assert(i < src.ssa->num_components);
205 } else {
206 nir_register *reg = src.reg.reg;
207 entry = _mesa_hash_table_search(c->def_ht, reg);
208 assert(reg->num_array_elems == 0);
209 assert(src.reg.base_offset == 0);
210 assert(i < reg->num_components);
211 }
212
213 struct qreg *qregs = entry->data;
214 return qregs[i];
215 }
216
217 static struct qreg
218 ntq_get_alu_src(struct vc4_compile *c, nir_alu_instr *instr,
219 unsigned src)
220 {
221 assert(util_is_power_of_two(instr->dest.write_mask));
222 unsigned chan = ffs(instr->dest.write_mask) - 1;
223 struct qreg r = ntq_get_src(c, instr->src[src].src,
224 instr->src[src].swizzle[chan]);
225
226 assert(!instr->src[src].abs);
227 assert(!instr->src[src].negate);
228
229 return r;
230 };
231
232 static inline struct qreg
233 qir_SAT(struct vc4_compile *c, struct qreg val)
234 {
235 return qir_FMAX(c,
236 qir_FMIN(c, val, qir_uniform_f(c, 1.0)),
237 qir_uniform_f(c, 0.0));
238 }
239
240 static struct qreg
241 ntq_rcp(struct vc4_compile *c, struct qreg x)
242 {
243 struct qreg r = qir_RCP(c, x);
244
245 /* Apply a Newton-Raphson step to improve the accuracy. */
246 r = qir_FMUL(c, r, qir_FSUB(c,
247 qir_uniform_f(c, 2.0),
248 qir_FMUL(c, x, r)));
249
250 return r;
251 }
252
253 static struct qreg
254 ntq_rsq(struct vc4_compile *c, struct qreg x)
255 {
256 struct qreg r = qir_RSQ(c, x);
257
258 /* Apply a Newton-Raphson step to improve the accuracy. */
259 r = qir_FMUL(c, r, qir_FSUB(c,
260 qir_uniform_f(c, 1.5),
261 qir_FMUL(c,
262 qir_uniform_f(c, 0.5),
263 qir_FMUL(c, x,
264 qir_FMUL(c, r, r)))));
265
266 return r;
267 }
268
269 static struct qreg
270 ntq_umul(struct vc4_compile *c, struct qreg src0, struct qreg src1)
271 {
272 struct qreg src0_hi = qir_SHR(c, src0,
273 qir_uniform_ui(c, 24));
274 struct qreg src1_hi = qir_SHR(c, src1,
275 qir_uniform_ui(c, 24));
276
277 struct qreg hilo = qir_MUL24(c, src0_hi, src1);
278 struct qreg lohi = qir_MUL24(c, src0, src1_hi);
279 struct qreg lolo = qir_MUL24(c, src0, src1);
280
281 return qir_ADD(c, lolo, qir_SHL(c,
282 qir_ADD(c, hilo, lohi),
283 qir_uniform_ui(c, 24)));
284 }
285
286 static struct qreg
287 ntq_scale_depth_texture(struct vc4_compile *c, struct qreg src)
288 {
289 struct qreg depthf = qir_ITOF(c, qir_SHR(c, src,
290 qir_uniform_ui(c, 8)));
291 return qir_FMUL(c, depthf, qir_uniform_f(c, 1.0f/0xffffff));
292 }
293
294 /**
295 * Emits a lowered TXF_MS from an MSAA texture.
296 *
297 * The addressing math has been lowered in NIR, and now we just need to read
298 * it like a UBO.
299 */
300 static void
301 ntq_emit_txf(struct vc4_compile *c, nir_tex_instr *instr)
302 {
303 uint32_t tile_width = 32;
304 uint32_t tile_height = 32;
305 uint32_t tile_size = (tile_height * tile_width *
306 VC4_MAX_SAMPLES * sizeof(uint32_t));
307
308 unsigned unit = instr->texture_index;
309 uint32_t w = align(c->key->tex[unit].msaa_width, tile_width);
310 uint32_t w_tiles = w / tile_width;
311 uint32_t h = align(c->key->tex[unit].msaa_height, tile_height);
312 uint32_t h_tiles = h / tile_height;
313 uint32_t size = w_tiles * h_tiles * tile_size;
314
315 struct qreg addr;
316 assert(instr->num_srcs == 1);
317 assert(instr->src[0].src_type == nir_tex_src_coord);
318 addr = ntq_get_src(c, instr->src[0].src, 0);
319
320 /* Perform the clamping required by kernel validation. */
321 addr = qir_MAX(c, addr, qir_uniform_ui(c, 0));
322 addr = qir_MIN(c, addr, qir_uniform_ui(c, size - 4));
323
324 qir_TEX_DIRECT(c, addr, qir_uniform(c, QUNIFORM_TEXTURE_MSAA_ADDR, unit));
325
326 struct qreg tex = qir_TEX_RESULT(c);
327 c->num_texture_samples++;
328
329 struct qreg dest[4];
330 enum pipe_format format = c->key->tex[unit].format;
331 if (util_format_is_depth_or_stencil(format)) {
332 struct qreg scaled = ntq_scale_depth_texture(c, tex);
333 for (int i = 0; i < 4; i++)
334 dest[i] = scaled;
335 } else {
336 for (int i = 0; i < 4; i++)
337 dest[i] = qir_UNPACK_8_F(c, tex, i);
338 }
339
340 for (int i = 0; i < 4; i++)
341 ntq_store_dest(c, &instr->dest, i, dest[i]);
342 }
343
344 static void
345 ntq_emit_tex(struct vc4_compile *c, nir_tex_instr *instr)
346 {
347 struct qreg s, t, r, lod, compare;
348 bool is_txb = false, is_txl = false;
349 unsigned unit = instr->texture_index;
350
351 if (instr->op == nir_texop_txf) {
352 ntq_emit_txf(c, instr);
353 return;
354 }
355
356 for (unsigned i = 0; i < instr->num_srcs; i++) {
357 switch (instr->src[i].src_type) {
358 case nir_tex_src_coord:
359 s = ntq_get_src(c, instr->src[i].src, 0);
360 if (instr->sampler_dim == GLSL_SAMPLER_DIM_1D)
361 t = qir_uniform_f(c, 0.5);
362 else
363 t = ntq_get_src(c, instr->src[i].src, 1);
364 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE)
365 r = ntq_get_src(c, instr->src[i].src, 2);
366 break;
367 case nir_tex_src_bias:
368 lod = ntq_get_src(c, instr->src[i].src, 0);
369 is_txb = true;
370 break;
371 case nir_tex_src_lod:
372 lod = ntq_get_src(c, instr->src[i].src, 0);
373 is_txl = true;
374 break;
375 case nir_tex_src_comparitor:
376 compare = ntq_get_src(c, instr->src[i].src, 0);
377 break;
378 default:
379 unreachable("unknown texture source");
380 }
381 }
382
383 if (c->key->tex[unit].force_first_level) {
384 lod = qir_uniform(c, QUNIFORM_TEXTURE_FIRST_LEVEL, unit);
385 is_txl = true;
386 is_txb = false;
387 }
388
389 struct qreg texture_u[] = {
390 qir_uniform(c, QUNIFORM_TEXTURE_CONFIG_P0, unit),
391 qir_uniform(c, QUNIFORM_TEXTURE_CONFIG_P1, unit),
392 qir_uniform(c, QUNIFORM_CONSTANT, 0),
393 qir_uniform(c, QUNIFORM_CONSTANT, 0),
394 };
395 uint32_t next_texture_u = 0;
396
397 /* There is no native support for GL texture rectangle coordinates, so
398 * we have to rescale from ([0, width], [0, height]) to ([0, 1], [0,
399 * 1]).
400 */
401 if (instr->sampler_dim == GLSL_SAMPLER_DIM_RECT) {
402 s = qir_FMUL(c, s,
403 qir_uniform(c, QUNIFORM_TEXRECT_SCALE_X, unit));
404 t = qir_FMUL(c, t,
405 qir_uniform(c, QUNIFORM_TEXRECT_SCALE_Y, unit));
406 }
407
408 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE || is_txl) {
409 texture_u[2] = qir_uniform(c, QUNIFORM_TEXTURE_CONFIG_P2,
410 unit | (is_txl << 16));
411 }
412
413 if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
414 qir_TEX_R(c, r, texture_u[next_texture_u++]);
415 } else if (c->key->tex[unit].wrap_s == PIPE_TEX_WRAP_CLAMP_TO_BORDER ||
416 c->key->tex[unit].wrap_s == PIPE_TEX_WRAP_CLAMP ||
417 c->key->tex[unit].wrap_t == PIPE_TEX_WRAP_CLAMP_TO_BORDER ||
418 c->key->tex[unit].wrap_t == PIPE_TEX_WRAP_CLAMP) {
419 qir_TEX_R(c, qir_uniform(c, QUNIFORM_TEXTURE_BORDER_COLOR, unit),
420 texture_u[next_texture_u++]);
421 }
422
423 if (c->key->tex[unit].wrap_s == PIPE_TEX_WRAP_CLAMP) {
424 s = qir_SAT(c, s);
425 }
426
427 if (c->key->tex[unit].wrap_t == PIPE_TEX_WRAP_CLAMP) {
428 t = qir_SAT(c, t);
429 }
430
431 qir_TEX_T(c, t, texture_u[next_texture_u++]);
432
433 if (is_txl || is_txb)
434 qir_TEX_B(c, lod, texture_u[next_texture_u++]);
435
436 qir_TEX_S(c, s, texture_u[next_texture_u++]);
437
438 c->num_texture_samples++;
439 struct qreg tex = qir_TEX_RESULT(c);
440
441 enum pipe_format format = c->key->tex[unit].format;
442
443 struct qreg *dest = ntq_get_dest(c, &instr->dest);
444 if (util_format_is_depth_or_stencil(format)) {
445 struct qreg normalized = ntq_scale_depth_texture(c, tex);
446 struct qreg depth_output;
447
448 struct qreg u0 = qir_uniform_f(c, 0.0f);
449 struct qreg u1 = qir_uniform_f(c, 1.0f);
450 if (c->key->tex[unit].compare_mode) {
451 switch (c->key->tex[unit].compare_func) {
452 case PIPE_FUNC_NEVER:
453 depth_output = qir_uniform_f(c, 0.0f);
454 break;
455 case PIPE_FUNC_ALWAYS:
456 depth_output = u1;
457 break;
458 case PIPE_FUNC_EQUAL:
459 qir_SF(c, qir_FSUB(c, compare, normalized));
460 depth_output = qir_SEL(c, QPU_COND_ZS, u1, u0);
461 break;
462 case PIPE_FUNC_NOTEQUAL:
463 qir_SF(c, qir_FSUB(c, compare, normalized));
464 depth_output = qir_SEL(c, QPU_COND_ZC, u1, u0);
465 break;
466 case PIPE_FUNC_GREATER:
467 qir_SF(c, qir_FSUB(c, compare, normalized));
468 depth_output = qir_SEL(c, QPU_COND_NC, u1, u0);
469 break;
470 case PIPE_FUNC_GEQUAL:
471 qir_SF(c, qir_FSUB(c, normalized, compare));
472 depth_output = qir_SEL(c, QPU_COND_NS, u1, u0);
473 break;
474 case PIPE_FUNC_LESS:
475 qir_SF(c, qir_FSUB(c, compare, normalized));
476 depth_output = qir_SEL(c, QPU_COND_NS, u1, u0);
477 break;
478 case PIPE_FUNC_LEQUAL:
479 qir_SF(c, qir_FSUB(c, normalized, compare));
480 depth_output = qir_SEL(c, QPU_COND_NC, u1, u0);
481 break;
482 }
483 } else {
484 depth_output = normalized;
485 }
486
487 for (int i = 0; i < 4; i++)
488 dest[i] = depth_output;
489 } else {
490 for (int i = 0; i < 4; i++)
491 dest[i] = qir_UNPACK_8_F(c, tex, i);
492 }
493 }
494
495 /**
496 * Computes x - floor(x), which is tricky because our FTOI truncates (rounds
497 * to zero).
498 */
499 static struct qreg
500 ntq_ffract(struct vc4_compile *c, struct qreg src)
501 {
502 struct qreg trunc = qir_ITOF(c, qir_FTOI(c, src));
503 struct qreg diff = qir_FSUB(c, src, trunc);
504 qir_SF(c, diff);
505 return qir_SEL(c, QPU_COND_NS,
506 qir_FADD(c, diff, qir_uniform_f(c, 1.0)), diff);
507 }
508
509 /**
510 * Computes floor(x), which is tricky because our FTOI truncates (rounds to
511 * zero).
512 */
513 static struct qreg
514 ntq_ffloor(struct vc4_compile *c, struct qreg src)
515 {
516 struct qreg trunc = qir_ITOF(c, qir_FTOI(c, src));
517
518 /* This will be < 0 if we truncated and the truncation was of a value
519 * that was < 0 in the first place.
520 */
521 qir_SF(c, qir_FSUB(c, src, trunc));
522
523 return qir_SEL(c, QPU_COND_NS,
524 qir_FSUB(c, trunc, qir_uniform_f(c, 1.0)), trunc);
525 }
526
527 /**
528 * Computes ceil(x), which is tricky because our FTOI truncates (rounds to
529 * zero).
530 */
531 static struct qreg
532 ntq_fceil(struct vc4_compile *c, struct qreg src)
533 {
534 struct qreg trunc = qir_ITOF(c, qir_FTOI(c, src));
535
536 /* This will be < 0 if we truncated and the truncation was of a value
537 * that was > 0 in the first place.
538 */
539 qir_SF(c, qir_FSUB(c, trunc, src));
540
541 return qir_SEL(c, QPU_COND_NS,
542 qir_FADD(c, trunc, qir_uniform_f(c, 1.0)), trunc);
543 }
544
545 static struct qreg
546 ntq_fsin(struct vc4_compile *c, struct qreg src)
547 {
548 float coeff[] = {
549 -2.0 * M_PI,
550 pow(2.0 * M_PI, 3) / (3 * 2 * 1),
551 -pow(2.0 * M_PI, 5) / (5 * 4 * 3 * 2 * 1),
552 pow(2.0 * M_PI, 7) / (7 * 6 * 5 * 4 * 3 * 2 * 1),
553 -pow(2.0 * M_PI, 9) / (9 * 8 * 7 * 6 * 5 * 4 * 3 * 2 * 1),
554 };
555
556 struct qreg scaled_x =
557 qir_FMUL(c,
558 src,
559 qir_uniform_f(c, 1.0 / (M_PI * 2.0)));
560
561 struct qreg x = qir_FADD(c,
562 ntq_ffract(c, scaled_x),
563 qir_uniform_f(c, -0.5));
564 struct qreg x2 = qir_FMUL(c, x, x);
565 struct qreg sum = qir_FMUL(c, x, qir_uniform_f(c, coeff[0]));
566 for (int i = 1; i < ARRAY_SIZE(coeff); i++) {
567 x = qir_FMUL(c, x, x2);
568 sum = qir_FADD(c,
569 sum,
570 qir_FMUL(c,
571 x,
572 qir_uniform_f(c, coeff[i])));
573 }
574 return sum;
575 }
576
577 static struct qreg
578 ntq_fcos(struct vc4_compile *c, struct qreg src)
579 {
580 float coeff[] = {
581 -1.0f,
582 pow(2.0 * M_PI, 2) / (2 * 1),
583 -pow(2.0 * M_PI, 4) / (4 * 3 * 2 * 1),
584 pow(2.0 * M_PI, 6) / (6 * 5 * 4 * 3 * 2 * 1),
585 -pow(2.0 * M_PI, 8) / (8 * 7 * 6 * 5 * 4 * 3 * 2 * 1),
586 pow(2.0 * M_PI, 10) / (10 * 9 * 8 * 7 * 6 * 5 * 4 * 3 * 2 * 1),
587 };
588
589 struct qreg scaled_x =
590 qir_FMUL(c, src,
591 qir_uniform_f(c, 1.0f / (M_PI * 2.0f)));
592 struct qreg x_frac = qir_FADD(c,
593 ntq_ffract(c, scaled_x),
594 qir_uniform_f(c, -0.5));
595
596 struct qreg sum = qir_uniform_f(c, coeff[0]);
597 struct qreg x2 = qir_FMUL(c, x_frac, x_frac);
598 struct qreg x = x2; /* Current x^2, x^4, or x^6 */
599 for (int i = 1; i < ARRAY_SIZE(coeff); i++) {
600 if (i != 1)
601 x = qir_FMUL(c, x, x2);
602
603 struct qreg mul = qir_FMUL(c,
604 x,
605 qir_uniform_f(c, coeff[i]));
606 if (i == 0)
607 sum = mul;
608 else
609 sum = qir_FADD(c, sum, mul);
610 }
611 return sum;
612 }
613
614 static struct qreg
615 ntq_fsign(struct vc4_compile *c, struct qreg src)
616 {
617 struct qreg t = qir_get_temp(c);
618
619 qir_SF(c, src);
620 qir_MOV_dest(c, t, qir_uniform_f(c, 0.0));
621 qir_MOV_dest(c, t, qir_uniform_f(c, 1.0))->cond = QPU_COND_ZC;
622 qir_MOV_dest(c, t, qir_uniform_f(c, -1.0))->cond = QPU_COND_NS;
623 return t;
624 }
625
626 static void
627 emit_vertex_input(struct vc4_compile *c, int attr)
628 {
629 enum pipe_format format = c->vs_key->attr_formats[attr];
630 uint32_t attr_size = util_format_get_blocksize(format);
631
632 c->vattr_sizes[attr] = align(attr_size, 4);
633 for (int i = 0; i < align(attr_size, 4) / 4; i++) {
634 c->inputs[attr * 4 + i] =
635 qir_MOV(c, qir_reg(QFILE_VPM, attr * 4 + i));
636 c->num_inputs++;
637 }
638 }
639
640 static void
641 emit_fragcoord_input(struct vc4_compile *c, int attr)
642 {
643 c->inputs[attr * 4 + 0] = qir_ITOF(c, qir_reg(QFILE_FRAG_X, 0));
644 c->inputs[attr * 4 + 1] = qir_ITOF(c, qir_reg(QFILE_FRAG_Y, 0));
645 c->inputs[attr * 4 + 2] =
646 qir_FMUL(c,
647 qir_ITOF(c, qir_FRAG_Z(c)),
648 qir_uniform_f(c, 1.0 / 0xffffff));
649 c->inputs[attr * 4 + 3] = qir_RCP(c, qir_FRAG_W(c));
650 }
651
652 static struct qreg
653 emit_fragment_varying(struct vc4_compile *c, gl_varying_slot slot,
654 uint8_t swizzle)
655 {
656 uint32_t i = c->num_input_slots++;
657 struct qreg vary = {
658 QFILE_VARY,
659 i
660 };
661
662 if (c->num_input_slots >= c->input_slots_array_size) {
663 c->input_slots_array_size =
664 MAX2(4, c->input_slots_array_size * 2);
665
666 c->input_slots = reralloc(c, c->input_slots,
667 struct vc4_varying_slot,
668 c->input_slots_array_size);
669 }
670
671 c->input_slots[i].slot = slot;
672 c->input_slots[i].swizzle = swizzle;
673
674 return qir_VARY_ADD_C(c, qir_FMUL(c, vary, qir_FRAG_W(c)));
675 }
676
677 static void
678 emit_fragment_input(struct vc4_compile *c, int attr, gl_varying_slot slot)
679 {
680 for (int i = 0; i < 4; i++) {
681 c->inputs[attr * 4 + i] =
682 emit_fragment_varying(c, slot, i);
683 c->num_inputs++;
684 }
685 }
686
687 static void
688 add_output(struct vc4_compile *c,
689 uint32_t decl_offset,
690 uint8_t slot,
691 uint8_t swizzle)
692 {
693 uint32_t old_array_size = c->outputs_array_size;
694 resize_qreg_array(c, &c->outputs, &c->outputs_array_size,
695 decl_offset + 1);
696
697 if (old_array_size != c->outputs_array_size) {
698 c->output_slots = reralloc(c,
699 c->output_slots,
700 struct vc4_varying_slot,
701 c->outputs_array_size);
702 }
703
704 c->output_slots[decl_offset].slot = slot;
705 c->output_slots[decl_offset].swizzle = swizzle;
706 }
707
708 static void
709 declare_uniform_range(struct vc4_compile *c, uint32_t start, uint32_t size)
710 {
711 unsigned array_id = c->num_uniform_ranges++;
712 if (array_id >= c->ubo_ranges_array_size) {
713 c->ubo_ranges_array_size = MAX2(c->ubo_ranges_array_size * 2,
714 array_id + 1);
715 c->ubo_ranges = reralloc(c, c->ubo_ranges,
716 struct vc4_compiler_ubo_range,
717 c->ubo_ranges_array_size);
718 }
719
720 c->ubo_ranges[array_id].dst_offset = 0;
721 c->ubo_ranges[array_id].src_offset = start;
722 c->ubo_ranges[array_id].size = size;
723 c->ubo_ranges[array_id].used = false;
724 }
725
726 static bool
727 ntq_src_is_only_ssa_def_user(nir_src *src)
728 {
729 if (!src->is_ssa)
730 return false;
731
732 if (!list_empty(&src->ssa->if_uses))
733 return false;
734
735 return (src->ssa->uses.next == &src->use_link &&
736 src->ssa->uses.next->next == &src->ssa->uses);
737 }
738
739 /**
740 * In general, emits a nir_pack_unorm_4x8 as a series of MOVs with the pack
741 * bit set.
742 *
743 * However, as an optimization, it tries to find the instructions generating
744 * the sources to be packed and just emit the pack flag there, if possible.
745 */
746 static void
747 ntq_emit_pack_unorm_4x8(struct vc4_compile *c, nir_alu_instr *instr)
748 {
749 struct qreg result = qir_get_temp(c);
750 struct nir_alu_instr *vec4 = NULL;
751
752 /* If packing from a vec4 op (as expected), identify it so that we can
753 * peek back at what generated its sources.
754 */
755 if (instr->src[0].src.is_ssa &&
756 instr->src[0].src.ssa->parent_instr->type == nir_instr_type_alu &&
757 nir_instr_as_alu(instr->src[0].src.ssa->parent_instr)->op ==
758 nir_op_vec4) {
759 vec4 = nir_instr_as_alu(instr->src[0].src.ssa->parent_instr);
760 }
761
762 /* If the pack is replicating the same channel 4 times, use the 8888
763 * pack flag. This is common for blending using the alpha
764 * channel.
765 */
766 if (instr->src[0].swizzle[0] == instr->src[0].swizzle[1] &&
767 instr->src[0].swizzle[0] == instr->src[0].swizzle[2] &&
768 instr->src[0].swizzle[0] == instr->src[0].swizzle[3]) {
769 struct qreg rep = ntq_get_src(c,
770 instr->src[0].src,
771 instr->src[0].swizzle[0]);
772 ntq_store_dest(c, &instr->dest.dest, 0, qir_PACK_8888_F(c, rep));
773 return;
774 }
775
776 for (int i = 0; i < 4; i++) {
777 int swiz = instr->src[0].swizzle[i];
778 struct qreg src;
779 if (vec4) {
780 src = ntq_get_src(c, vec4->src[swiz].src,
781 vec4->src[swiz].swizzle[0]);
782 } else {
783 src = ntq_get_src(c, instr->src[0].src, swiz);
784 }
785
786 if (vec4 &&
787 ntq_src_is_only_ssa_def_user(&vec4->src[swiz].src) &&
788 src.file == QFILE_TEMP &&
789 c->defs[src.index] &&
790 qir_is_mul(c->defs[src.index]) &&
791 !c->defs[src.index]->dst.pack) {
792 struct qinst *rewrite = c->defs[src.index];
793 c->defs[src.index] = NULL;
794 rewrite->dst = result;
795 rewrite->dst.pack = QPU_PACK_MUL_8A + i;
796 continue;
797 }
798
799 qir_PACK_8_F(c, result, src, i);
800 }
801
802 ntq_store_dest(c, &instr->dest.dest, 0, result);
803 }
804
805 /** Handles sign-extended bitfield extracts for 16 bits. */
806 static struct qreg
807 ntq_emit_ibfe(struct vc4_compile *c, struct qreg base, struct qreg offset,
808 struct qreg bits)
809 {
810 assert(bits.file == QFILE_UNIF &&
811 c->uniform_contents[bits.index] == QUNIFORM_CONSTANT &&
812 c->uniform_data[bits.index] == 16);
813
814 assert(offset.file == QFILE_UNIF &&
815 c->uniform_contents[offset.index] == QUNIFORM_CONSTANT);
816 int offset_bit = c->uniform_data[offset.index];
817 assert(offset_bit % 16 == 0);
818
819 return qir_UNPACK_16_I(c, base, offset_bit / 16);
820 }
821
822 /** Handles unsigned bitfield extracts for 8 bits. */
823 static struct qreg
824 ntq_emit_ubfe(struct vc4_compile *c, struct qreg base, struct qreg offset,
825 struct qreg bits)
826 {
827 assert(bits.file == QFILE_UNIF &&
828 c->uniform_contents[bits.index] == QUNIFORM_CONSTANT &&
829 c->uniform_data[bits.index] == 8);
830
831 assert(offset.file == QFILE_UNIF &&
832 c->uniform_contents[offset.index] == QUNIFORM_CONSTANT);
833 int offset_bit = c->uniform_data[offset.index];
834 assert(offset_bit % 8 == 0);
835
836 return qir_UNPACK_8_I(c, base, offset_bit / 8);
837 }
838
839 /**
840 * If compare_instr is a valid comparison instruction, emits the
841 * compare_instr's comparison and returns the sel_instr's return value based
842 * on the compare_instr's result.
843 */
844 static bool
845 ntq_emit_comparison(struct vc4_compile *c, struct qreg *dest,
846 nir_alu_instr *compare_instr,
847 nir_alu_instr *sel_instr)
848 {
849 enum qpu_cond cond;
850
851 switch (compare_instr->op) {
852 case nir_op_feq:
853 case nir_op_ieq:
854 case nir_op_seq:
855 cond = QPU_COND_ZS;
856 break;
857 case nir_op_fne:
858 case nir_op_ine:
859 case nir_op_sne:
860 cond = QPU_COND_ZC;
861 break;
862 case nir_op_fge:
863 case nir_op_ige:
864 case nir_op_uge:
865 case nir_op_sge:
866 cond = QPU_COND_NC;
867 break;
868 case nir_op_flt:
869 case nir_op_ilt:
870 case nir_op_slt:
871 cond = QPU_COND_NS;
872 break;
873 default:
874 return false;
875 }
876
877 struct qreg src0 = ntq_get_alu_src(c, compare_instr, 0);
878 struct qreg src1 = ntq_get_alu_src(c, compare_instr, 1);
879
880 unsigned unsized_type =
881 nir_alu_type_get_base_type(nir_op_infos[compare_instr->op].input_types[0]);
882 if (unsized_type == nir_type_float)
883 qir_SF(c, qir_FSUB(c, src0, src1));
884 else
885 qir_SF(c, qir_SUB(c, src0, src1));
886
887 switch (sel_instr->op) {
888 case nir_op_seq:
889 case nir_op_sne:
890 case nir_op_sge:
891 case nir_op_slt:
892 *dest = qir_SEL(c, cond,
893 qir_uniform_f(c, 1.0), qir_uniform_f(c, 0.0));
894 break;
895
896 case nir_op_bcsel:
897 *dest = qir_SEL(c, cond,
898 ntq_get_alu_src(c, sel_instr, 1),
899 ntq_get_alu_src(c, sel_instr, 2));
900 break;
901
902 default:
903 *dest = qir_SEL(c, cond,
904 qir_uniform_ui(c, ~0), qir_uniform_ui(c, 0));
905 break;
906 }
907
908 return true;
909 }
910
911 /**
912 * Attempts to fold a comparison generating a boolean result into the
913 * condition code for selecting between two values, instead of comparing the
914 * boolean result against 0 to generate the condition code.
915 */
916 static struct qreg ntq_emit_bcsel(struct vc4_compile *c, nir_alu_instr *instr,
917 struct qreg *src)
918 {
919 if (!instr->src[0].src.is_ssa)
920 goto out;
921 nir_alu_instr *compare =
922 nir_instr_as_alu(instr->src[0].src.ssa->parent_instr);
923 if (!compare)
924 goto out;
925
926 struct qreg dest;
927 if (ntq_emit_comparison(c, &dest, compare, instr))
928 return dest;
929
930 out:
931 qir_SF(c, src[0]);
932 return qir_SEL(c, QPU_COND_NS, src[1], src[2]);
933 }
934
935 static struct qreg
936 ntq_fddx(struct vc4_compile *c, struct qreg src)
937 {
938 /* Make sure that we have a bare temp to use for MUL rotation, so it
939 * can be allocated to an accumulator.
940 */
941 if (src.pack || src.file != QFILE_TEMP)
942 src = qir_MOV(c, src);
943
944 struct qreg from_left = qir_ROT_MUL(c, src, 1);
945 struct qreg from_right = qir_ROT_MUL(c, src, 15);
946
947 /* Distinguish left/right pixels of the quad. */
948 qir_SF(c, qir_AND(c, qir_reg(QFILE_QPU_ELEMENT, 0),
949 qir_uniform_ui(c, 1)));
950
951 return qir_SEL(c, QPU_COND_ZS,
952 qir_FSUB(c, from_right, src),
953 qir_FSUB(c, src, from_left));
954 }
955
956 static struct qreg
957 ntq_fddy(struct vc4_compile *c, struct qreg src)
958 {
959 if (src.pack || src.file != QFILE_TEMP)
960 src = qir_MOV(c, src);
961
962 struct qreg from_bottom = qir_ROT_MUL(c, src, 2);
963 struct qreg from_top = qir_ROT_MUL(c, src, 14);
964
965 /* Distinguish top/bottom pixels of the quad. */
966 qir_SF(c, qir_AND(c,
967 qir_reg(QFILE_QPU_ELEMENT, 0),
968 qir_uniform_ui(c, 2)));
969
970 return qir_SEL(c, QPU_COND_ZS,
971 qir_FSUB(c, from_top, src),
972 qir_FSUB(c, src, from_bottom));
973 }
974
975 static void
976 ntq_emit_alu(struct vc4_compile *c, nir_alu_instr *instr)
977 {
978 /* This should always be lowered to ALU operations for VC4. */
979 assert(!instr->dest.saturate);
980
981 /* Vectors are special in that they have non-scalarized writemasks,
982 * and just take the first swizzle channel for each argument in order
983 * into each writemask channel.
984 */
985 if (instr->op == nir_op_vec2 ||
986 instr->op == nir_op_vec3 ||
987 instr->op == nir_op_vec4) {
988 struct qreg srcs[4];
989 for (int i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
990 srcs[i] = ntq_get_src(c, instr->src[i].src,
991 instr->src[i].swizzle[0]);
992 for (int i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
993 ntq_store_dest(c, &instr->dest.dest, i, srcs[i]);
994 return;
995 }
996
997 if (instr->op == nir_op_pack_unorm_4x8) {
998 ntq_emit_pack_unorm_4x8(c, instr);
999 return;
1000 }
1001
1002 if (instr->op == nir_op_unpack_unorm_4x8) {
1003 struct qreg src = ntq_get_src(c, instr->src[0].src,
1004 instr->src[0].swizzle[0]);
1005 for (int i = 0; i < 4; i++) {
1006 if (instr->dest.write_mask & (1 << i))
1007 ntq_store_dest(c, &instr->dest.dest, i,
1008 qir_UNPACK_8_F(c, src, i));
1009 }
1010 return;
1011 }
1012
1013 /* General case: We can just grab the one used channel per src. */
1014 struct qreg src[nir_op_infos[instr->op].num_inputs];
1015 for (int i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
1016 src[i] = ntq_get_alu_src(c, instr, i);
1017 }
1018
1019 struct qreg result;
1020
1021 switch (instr->op) {
1022 case nir_op_fmov:
1023 case nir_op_imov:
1024 result = qir_MOV(c, src[0]);
1025 break;
1026 case nir_op_fmul:
1027 result = qir_FMUL(c, src[0], src[1]);
1028 break;
1029 case nir_op_fadd:
1030 result = qir_FADD(c, src[0], src[1]);
1031 break;
1032 case nir_op_fsub:
1033 result = qir_FSUB(c, src[0], src[1]);
1034 break;
1035 case nir_op_fmin:
1036 result = qir_FMIN(c, src[0], src[1]);
1037 break;
1038 case nir_op_fmax:
1039 result = qir_FMAX(c, src[0], src[1]);
1040 break;
1041
1042 case nir_op_f2i:
1043 case nir_op_f2u:
1044 result = qir_FTOI(c, src[0]);
1045 break;
1046 case nir_op_i2f:
1047 case nir_op_u2f:
1048 result = qir_ITOF(c, src[0]);
1049 break;
1050 case nir_op_b2f:
1051 result = qir_AND(c, src[0], qir_uniform_f(c, 1.0));
1052 break;
1053 case nir_op_b2i:
1054 result = qir_AND(c, src[0], qir_uniform_ui(c, 1));
1055 break;
1056 case nir_op_i2b:
1057 case nir_op_f2b:
1058 qir_SF(c, src[0]);
1059 result = qir_SEL(c, QPU_COND_ZC,
1060 qir_uniform_ui(c, ~0),
1061 qir_uniform_ui(c, 0));
1062 break;
1063
1064 case nir_op_iadd:
1065 result = qir_ADD(c, src[0], src[1]);
1066 break;
1067 case nir_op_ushr:
1068 result = qir_SHR(c, src[0], src[1]);
1069 break;
1070 case nir_op_isub:
1071 result = qir_SUB(c, src[0], src[1]);
1072 break;
1073 case nir_op_ishr:
1074 result = qir_ASR(c, src[0], src[1]);
1075 break;
1076 case nir_op_ishl:
1077 result = qir_SHL(c, src[0], src[1]);
1078 break;
1079 case nir_op_imin:
1080 result = qir_MIN(c, src[0], src[1]);
1081 break;
1082 case nir_op_imax:
1083 result = qir_MAX(c, src[0], src[1]);
1084 break;
1085 case nir_op_iand:
1086 result = qir_AND(c, src[0], src[1]);
1087 break;
1088 case nir_op_ior:
1089 result = qir_OR(c, src[0], src[1]);
1090 break;
1091 case nir_op_ixor:
1092 result = qir_XOR(c, src[0], src[1]);
1093 break;
1094 case nir_op_inot:
1095 result = qir_NOT(c, src[0]);
1096 break;
1097
1098 case nir_op_imul:
1099 result = ntq_umul(c, src[0], src[1]);
1100 break;
1101
1102 case nir_op_seq:
1103 case nir_op_sne:
1104 case nir_op_sge:
1105 case nir_op_slt:
1106 case nir_op_feq:
1107 case nir_op_fne:
1108 case nir_op_fge:
1109 case nir_op_flt:
1110 case nir_op_ieq:
1111 case nir_op_ine:
1112 case nir_op_ige:
1113 case nir_op_uge:
1114 case nir_op_ilt:
1115 if (!ntq_emit_comparison(c, &result, instr, instr)) {
1116 fprintf(stderr, "Bad comparison instruction\n");
1117 }
1118 break;
1119
1120 case nir_op_bcsel:
1121 result = ntq_emit_bcsel(c, instr, src);
1122 break;
1123 case nir_op_fcsel:
1124 qir_SF(c, src[0]);
1125 result = qir_SEL(c, QPU_COND_ZC, src[1], src[2]);
1126 break;
1127
1128 case nir_op_frcp:
1129 result = ntq_rcp(c, src[0]);
1130 break;
1131 case nir_op_frsq:
1132 result = ntq_rsq(c, src[0]);
1133 break;
1134 case nir_op_fexp2:
1135 result = qir_EXP2(c, src[0]);
1136 break;
1137 case nir_op_flog2:
1138 result = qir_LOG2(c, src[0]);
1139 break;
1140
1141 case nir_op_ftrunc:
1142 result = qir_ITOF(c, qir_FTOI(c, src[0]));
1143 break;
1144 case nir_op_fceil:
1145 result = ntq_fceil(c, src[0]);
1146 break;
1147 case nir_op_ffract:
1148 result = ntq_ffract(c, src[0]);
1149 break;
1150 case nir_op_ffloor:
1151 result = ntq_ffloor(c, src[0]);
1152 break;
1153
1154 case nir_op_fsin:
1155 result = ntq_fsin(c, src[0]);
1156 break;
1157 case nir_op_fcos:
1158 result = ntq_fcos(c, src[0]);
1159 break;
1160
1161 case nir_op_fsign:
1162 result = ntq_fsign(c, src[0]);
1163 break;
1164
1165 case nir_op_fabs:
1166 result = qir_FMAXABS(c, src[0], src[0]);
1167 break;
1168 case nir_op_iabs:
1169 result = qir_MAX(c, src[0],
1170 qir_SUB(c, qir_uniform_ui(c, 0), src[0]));
1171 break;
1172
1173 case nir_op_ibitfield_extract:
1174 result = ntq_emit_ibfe(c, src[0], src[1], src[2]);
1175 break;
1176
1177 case nir_op_ubitfield_extract:
1178 result = ntq_emit_ubfe(c, src[0], src[1], src[2]);
1179 break;
1180
1181 case nir_op_usadd_4x8:
1182 result = qir_V8ADDS(c, src[0], src[1]);
1183 break;
1184
1185 case nir_op_ussub_4x8:
1186 result = qir_V8SUBS(c, src[0], src[1]);
1187 break;
1188
1189 case nir_op_umin_4x8:
1190 result = qir_V8MIN(c, src[0], src[1]);
1191 break;
1192
1193 case nir_op_umax_4x8:
1194 result = qir_V8MAX(c, src[0], src[1]);
1195 break;
1196
1197 case nir_op_umul_unorm_4x8:
1198 result = qir_V8MULD(c, src[0], src[1]);
1199 break;
1200
1201 case nir_op_fddx:
1202 case nir_op_fddx_coarse:
1203 case nir_op_fddx_fine:
1204 result = ntq_fddx(c, src[0]);
1205 break;
1206
1207 case nir_op_fddy:
1208 case nir_op_fddy_coarse:
1209 case nir_op_fddy_fine:
1210 result = ntq_fddy(c, src[0]);
1211 break;
1212
1213 default:
1214 fprintf(stderr, "unknown NIR ALU inst: ");
1215 nir_print_instr(&instr->instr, stderr);
1216 fprintf(stderr, "\n");
1217 abort();
1218 }
1219
1220 /* We have a scalar result, so the instruction should only have a
1221 * single channel written to.
1222 */
1223 assert(util_is_power_of_two(instr->dest.write_mask));
1224 ntq_store_dest(c, &instr->dest.dest,
1225 ffs(instr->dest.write_mask) - 1, result);
1226 }
1227
1228 static void
1229 emit_frag_end(struct vc4_compile *c)
1230 {
1231 struct qreg color;
1232 if (c->output_color_index != -1) {
1233 color = c->outputs[c->output_color_index];
1234 } else {
1235 color = qir_uniform_ui(c, 0);
1236 }
1237
1238 uint32_t discard_cond = QPU_COND_ALWAYS;
1239 if (c->s->info.fs.uses_discard) {
1240 qir_SF(c, c->discard);
1241 discard_cond = QPU_COND_ZS;
1242 }
1243
1244 if (c->fs_key->stencil_enabled) {
1245 qir_MOV_dest(c, qir_reg(QFILE_TLB_STENCIL_SETUP, 0),
1246 qir_uniform(c, QUNIFORM_STENCIL, 0));
1247 if (c->fs_key->stencil_twoside) {
1248 qir_MOV_dest(c, qir_reg(QFILE_TLB_STENCIL_SETUP, 0),
1249 qir_uniform(c, QUNIFORM_STENCIL, 1));
1250 }
1251 if (c->fs_key->stencil_full_writemasks) {
1252 qir_MOV_dest(c, qir_reg(QFILE_TLB_STENCIL_SETUP, 0),
1253 qir_uniform(c, QUNIFORM_STENCIL, 2));
1254 }
1255 }
1256
1257 if (c->output_sample_mask_index != -1) {
1258 qir_MS_MASK(c, c->outputs[c->output_sample_mask_index]);
1259 }
1260
1261 if (c->fs_key->depth_enabled) {
1262 if (c->output_position_index != -1) {
1263 qir_FTOI_dest(c, qir_reg(QFILE_TLB_Z_WRITE, 0),
1264 qir_FMUL(c,
1265 c->outputs[c->output_position_index],
1266 qir_uniform_f(c, 0xffffff)))->cond = discard_cond;
1267 } else {
1268 qir_MOV_dest(c, qir_reg(QFILE_TLB_Z_WRITE, 0),
1269 qir_FRAG_Z(c))->cond = discard_cond;
1270 }
1271 }
1272
1273 if (!c->msaa_per_sample_output) {
1274 qir_MOV_dest(c, qir_reg(QFILE_TLB_COLOR_WRITE, 0),
1275 color)->cond = discard_cond;
1276 } else {
1277 for (int i = 0; i < VC4_MAX_SAMPLES; i++) {
1278 qir_MOV_dest(c, qir_reg(QFILE_TLB_COLOR_WRITE_MS, 0),
1279 c->sample_colors[i])->cond = discard_cond;
1280 }
1281 }
1282 }
1283
1284 static void
1285 emit_scaled_viewport_write(struct vc4_compile *c, struct qreg rcp_w)
1286 {
1287 struct qreg packed = qir_get_temp(c);
1288
1289 for (int i = 0; i < 2; i++) {
1290 struct qreg scale =
1291 qir_uniform(c, QUNIFORM_VIEWPORT_X_SCALE + i, 0);
1292
1293 struct qreg packed_chan = packed;
1294 packed_chan.pack = QPU_PACK_A_16A + i;
1295
1296 qir_FTOI_dest(c, packed_chan,
1297 qir_FMUL(c,
1298 qir_FMUL(c,
1299 c->outputs[c->output_position_index + i],
1300 scale),
1301 rcp_w));
1302 }
1303
1304 qir_VPM_WRITE(c, packed);
1305 }
1306
1307 static void
1308 emit_zs_write(struct vc4_compile *c, struct qreg rcp_w)
1309 {
1310 struct qreg zscale = qir_uniform(c, QUNIFORM_VIEWPORT_Z_SCALE, 0);
1311 struct qreg zoffset = qir_uniform(c, QUNIFORM_VIEWPORT_Z_OFFSET, 0);
1312
1313 qir_VPM_WRITE(c, qir_FADD(c, qir_FMUL(c, qir_FMUL(c,
1314 c->outputs[c->output_position_index + 2],
1315 zscale),
1316 rcp_w),
1317 zoffset));
1318 }
1319
1320 static void
1321 emit_rcp_wc_write(struct vc4_compile *c, struct qreg rcp_w)
1322 {
1323 qir_VPM_WRITE(c, rcp_w);
1324 }
1325
1326 static void
1327 emit_point_size_write(struct vc4_compile *c)
1328 {
1329 struct qreg point_size;
1330
1331 if (c->output_point_size_index != -1)
1332 point_size = c->outputs[c->output_point_size_index];
1333 else
1334 point_size = qir_uniform_f(c, 1.0);
1335
1336 /* Workaround: HW-2726 PTB does not handle zero-size points (BCM2835,
1337 * BCM21553).
1338 */
1339 point_size = qir_FMAX(c, point_size, qir_uniform_f(c, .125));
1340
1341 qir_VPM_WRITE(c, point_size);
1342 }
1343
1344 /**
1345 * Emits a VPM read of the stub vertex attribute set up by vc4_draw.c.
1346 *
1347 * The simulator insists that there be at least one vertex attribute, so
1348 * vc4_draw.c will emit one if it wouldn't have otherwise. The simulator also
1349 * insists that all vertex attributes loaded get read by the VS/CS, so we have
1350 * to consume it here.
1351 */
1352 static void
1353 emit_stub_vpm_read(struct vc4_compile *c)
1354 {
1355 if (c->num_inputs)
1356 return;
1357
1358 c->vattr_sizes[0] = 4;
1359 (void)qir_MOV(c, qir_reg(QFILE_VPM, 0));
1360 c->num_inputs++;
1361 }
1362
1363 static void
1364 emit_vert_end(struct vc4_compile *c,
1365 struct vc4_varying_slot *fs_inputs,
1366 uint32_t num_fs_inputs)
1367 {
1368 struct qreg rcp_w = qir_RCP(c, c->outputs[c->output_position_index + 3]);
1369
1370 emit_stub_vpm_read(c);
1371
1372 emit_scaled_viewport_write(c, rcp_w);
1373 emit_zs_write(c, rcp_w);
1374 emit_rcp_wc_write(c, rcp_w);
1375 if (c->vs_key->per_vertex_point_size)
1376 emit_point_size_write(c);
1377
1378 for (int i = 0; i < num_fs_inputs; i++) {
1379 struct vc4_varying_slot *input = &fs_inputs[i];
1380 int j;
1381
1382 for (j = 0; j < c->num_outputs; j++) {
1383 struct vc4_varying_slot *output =
1384 &c->output_slots[j];
1385
1386 if (input->slot == output->slot &&
1387 input->swizzle == output->swizzle) {
1388 qir_VPM_WRITE(c, c->outputs[j]);
1389 break;
1390 }
1391 }
1392 /* Emit padding if we didn't find a declared VS output for
1393 * this FS input.
1394 */
1395 if (j == c->num_outputs)
1396 qir_VPM_WRITE(c, qir_uniform_f(c, 0.0));
1397 }
1398 }
1399
1400 static void
1401 emit_coord_end(struct vc4_compile *c)
1402 {
1403 struct qreg rcp_w = qir_RCP(c, c->outputs[c->output_position_index + 3]);
1404
1405 emit_stub_vpm_read(c);
1406
1407 for (int i = 0; i < 4; i++)
1408 qir_VPM_WRITE(c, c->outputs[c->output_position_index + i]);
1409
1410 emit_scaled_viewport_write(c, rcp_w);
1411 emit_zs_write(c, rcp_w);
1412 emit_rcp_wc_write(c, rcp_w);
1413 if (c->vs_key->per_vertex_point_size)
1414 emit_point_size_write(c);
1415 }
1416
1417 static void
1418 vc4_optimize_nir(struct nir_shader *s)
1419 {
1420 bool progress;
1421
1422 do {
1423 progress = false;
1424
1425 NIR_PASS_V(s, nir_lower_vars_to_ssa);
1426 NIR_PASS(progress, s, nir_lower_alu_to_scalar);
1427 NIR_PASS(progress, s, nir_lower_phis_to_scalar);
1428 NIR_PASS(progress, s, nir_copy_prop);
1429 NIR_PASS(progress, s, nir_opt_remove_phis);
1430 NIR_PASS(progress, s, nir_opt_dce);
1431 NIR_PASS(progress, s, nir_opt_dead_cf);
1432 NIR_PASS(progress, s, nir_opt_cse);
1433 NIR_PASS(progress, s, nir_opt_peephole_select, 8);
1434 NIR_PASS(progress, s, nir_opt_algebraic);
1435 NIR_PASS(progress, s, nir_opt_constant_folding);
1436 NIR_PASS(progress, s, nir_opt_undef);
1437 } while (progress);
1438 }
1439
1440 static int
1441 driver_location_compare(const void *in_a, const void *in_b)
1442 {
1443 const nir_variable *const *a = in_a;
1444 const nir_variable *const *b = in_b;
1445
1446 return (*a)->data.driver_location - (*b)->data.driver_location;
1447 }
1448
1449 static void
1450 ntq_setup_inputs(struct vc4_compile *c)
1451 {
1452 unsigned num_entries = 0;
1453 nir_foreach_variable(var, &c->s->inputs)
1454 num_entries++;
1455
1456 nir_variable *vars[num_entries];
1457
1458 unsigned i = 0;
1459 nir_foreach_variable(var, &c->s->inputs)
1460 vars[i++] = var;
1461
1462 /* Sort the variables so that we emit the input setup in
1463 * driver_location order. This is required for VPM reads, whose data
1464 * is fetched into the VPM in driver_location (TGSI register index)
1465 * order.
1466 */
1467 qsort(&vars, num_entries, sizeof(*vars), driver_location_compare);
1468
1469 for (unsigned i = 0; i < num_entries; i++) {
1470 nir_variable *var = vars[i];
1471 unsigned array_len = MAX2(glsl_get_length(var->type), 1);
1472 unsigned loc = var->data.driver_location;
1473
1474 assert(array_len == 1);
1475 (void)array_len;
1476 resize_qreg_array(c, &c->inputs, &c->inputs_array_size,
1477 (loc + 1) * 4);
1478
1479 if (c->stage == QSTAGE_FRAG) {
1480 if (var->data.location == VARYING_SLOT_POS) {
1481 emit_fragcoord_input(c, loc);
1482 } else if (var->data.location == VARYING_SLOT_PNTC ||
1483 (var->data.location >= VARYING_SLOT_VAR0 &&
1484 (c->fs_key->point_sprite_mask &
1485 (1 << (var->data.location -
1486 VARYING_SLOT_VAR0))))) {
1487 c->inputs[loc * 4 + 0] = c->point_x;
1488 c->inputs[loc * 4 + 1] = c->point_y;
1489 } else {
1490 emit_fragment_input(c, loc, var->data.location);
1491 }
1492 } else {
1493 emit_vertex_input(c, loc);
1494 }
1495 }
1496 }
1497
1498 static void
1499 ntq_setup_outputs(struct vc4_compile *c)
1500 {
1501 nir_foreach_variable(var, &c->s->outputs) {
1502 unsigned array_len = MAX2(glsl_get_length(var->type), 1);
1503 unsigned loc = var->data.driver_location * 4;
1504
1505 assert(array_len == 1);
1506 (void)array_len;
1507
1508 for (int i = 0; i < 4; i++)
1509 add_output(c, loc + i, var->data.location, i);
1510
1511 if (c->stage == QSTAGE_FRAG) {
1512 switch (var->data.location) {
1513 case FRAG_RESULT_COLOR:
1514 case FRAG_RESULT_DATA0:
1515 c->output_color_index = loc;
1516 break;
1517 case FRAG_RESULT_DEPTH:
1518 c->output_position_index = loc;
1519 break;
1520 case FRAG_RESULT_SAMPLE_MASK:
1521 c->output_sample_mask_index = loc;
1522 break;
1523 }
1524 } else {
1525 switch (var->data.location) {
1526 case VARYING_SLOT_POS:
1527 c->output_position_index = loc;
1528 break;
1529 case VARYING_SLOT_PSIZ:
1530 c->output_point_size_index = loc;
1531 break;
1532 }
1533 }
1534 }
1535 }
1536
1537 static void
1538 ntq_setup_uniforms(struct vc4_compile *c)
1539 {
1540 nir_foreach_variable(var, &c->s->uniforms) {
1541 uint32_t vec4_count = st_glsl_type_size(var->type);
1542 unsigned vec4_size = 4 * sizeof(float);
1543
1544 declare_uniform_range(c, var->data.driver_location * vec4_size,
1545 vec4_count * vec4_size);
1546
1547 }
1548 }
1549
1550 /**
1551 * Sets up the mapping from nir_register to struct qreg *.
1552 *
1553 * Each nir_register gets a struct qreg per 32-bit component being stored.
1554 */
1555 static void
1556 ntq_setup_registers(struct vc4_compile *c, struct exec_list *list)
1557 {
1558 foreach_list_typed(nir_register, nir_reg, node, list) {
1559 unsigned array_len = MAX2(nir_reg->num_array_elems, 1);
1560 struct qreg *qregs = ralloc_array(c->def_ht, struct qreg,
1561 array_len *
1562 nir_reg->num_components);
1563
1564 _mesa_hash_table_insert(c->def_ht, nir_reg, qregs);
1565
1566 for (int i = 0; i < array_len * nir_reg->num_components; i++)
1567 qregs[i] = qir_get_temp(c);
1568 }
1569 }
1570
1571 static void
1572 ntq_emit_load_const(struct vc4_compile *c, nir_load_const_instr *instr)
1573 {
1574 struct qreg *qregs = ntq_init_ssa_def(c, &instr->def);
1575 for (int i = 0; i < instr->def.num_components; i++)
1576 qregs[i] = qir_uniform_ui(c, instr->value.u32[i]);
1577
1578 _mesa_hash_table_insert(c->def_ht, &instr->def, qregs);
1579 }
1580
1581 static void
1582 ntq_emit_ssa_undef(struct vc4_compile *c, nir_ssa_undef_instr *instr)
1583 {
1584 struct qreg *qregs = ntq_init_ssa_def(c, &instr->def);
1585
1586 /* QIR needs there to be *some* value, so pick 0 (same as for
1587 * ntq_setup_registers().
1588 */
1589 for (int i = 0; i < instr->def.num_components; i++)
1590 qregs[i] = qir_uniform_ui(c, 0);
1591 }
1592
1593 static void
1594 ntq_emit_intrinsic(struct vc4_compile *c, nir_intrinsic_instr *instr)
1595 {
1596 nir_const_value *const_offset;
1597 unsigned offset;
1598
1599 switch (instr->intrinsic) {
1600 case nir_intrinsic_load_uniform:
1601 assert(instr->num_components == 1);
1602 const_offset = nir_src_as_const_value(instr->src[0]);
1603 if (const_offset) {
1604 offset = nir_intrinsic_base(instr) + const_offset->u32[0];
1605 assert(offset % 4 == 0);
1606 /* We need dwords */
1607 offset = offset / 4;
1608 ntq_store_dest(c, &instr->dest, 0,
1609 qir_uniform(c, QUNIFORM_UNIFORM,
1610 offset));
1611 } else {
1612 ntq_store_dest(c, &instr->dest, 0,
1613 indirect_uniform_load(c, instr));
1614 }
1615 break;
1616
1617 case nir_intrinsic_load_user_clip_plane:
1618 for (int i = 0; i < instr->num_components; i++) {
1619 ntq_store_dest(c, &instr->dest, i,
1620 qir_uniform(c, QUNIFORM_USER_CLIP_PLANE,
1621 nir_intrinsic_ucp_id(instr) *
1622 4 + i));
1623 }
1624 break;
1625
1626 case nir_intrinsic_load_blend_const_color_r_float:
1627 case nir_intrinsic_load_blend_const_color_g_float:
1628 case nir_intrinsic_load_blend_const_color_b_float:
1629 case nir_intrinsic_load_blend_const_color_a_float:
1630 ntq_store_dest(c, &instr->dest, 0,
1631 qir_uniform(c, QUNIFORM_BLEND_CONST_COLOR_X +
1632 (instr->intrinsic -
1633 nir_intrinsic_load_blend_const_color_r_float),
1634 0));
1635 break;
1636
1637 case nir_intrinsic_load_blend_const_color_rgba8888_unorm:
1638 ntq_store_dest(c, &instr->dest, 0,
1639 qir_uniform(c, QUNIFORM_BLEND_CONST_COLOR_RGBA,
1640 0));
1641 break;
1642
1643 case nir_intrinsic_load_blend_const_color_aaaa8888_unorm:
1644 ntq_store_dest(c, &instr->dest, 0,
1645 qir_uniform(c, QUNIFORM_BLEND_CONST_COLOR_AAAA,
1646 0));
1647 break;
1648
1649 case nir_intrinsic_load_alpha_ref_float:
1650 ntq_store_dest(c, &instr->dest, 0,
1651 qir_uniform(c, QUNIFORM_ALPHA_REF, 0));
1652 break;
1653
1654 case nir_intrinsic_load_sample_mask_in:
1655 ntq_store_dest(c, &instr->dest, 0,
1656 qir_uniform(c, QUNIFORM_SAMPLE_MASK, 0));
1657 break;
1658
1659 case nir_intrinsic_load_front_face:
1660 /* The register contains 0 (front) or 1 (back), and we need to
1661 * turn it into a NIR bool where true means front.
1662 */
1663 ntq_store_dest(c, &instr->dest, 0,
1664 qir_ADD(c,
1665 qir_uniform_ui(c, -1),
1666 qir_reg(QFILE_FRAG_REV_FLAG, 0)));
1667 break;
1668
1669 case nir_intrinsic_load_input:
1670 assert(instr->num_components == 1);
1671 const_offset = nir_src_as_const_value(instr->src[0]);
1672 assert(const_offset && "vc4 doesn't support indirect inputs");
1673 if (c->stage == QSTAGE_FRAG &&
1674 nir_intrinsic_base(instr) >= VC4_NIR_TLB_COLOR_READ_INPUT) {
1675 assert(const_offset->u32[0] == 0);
1676 /* Reads of the per-sample color need to be done in
1677 * order.
1678 */
1679 int sample_index = (nir_intrinsic_base(instr) -
1680 VC4_NIR_TLB_COLOR_READ_INPUT);
1681 for (int i = 0; i <= sample_index; i++) {
1682 if (c->color_reads[i].file == QFILE_NULL) {
1683 c->color_reads[i] =
1684 qir_TLB_COLOR_READ(c);
1685 }
1686 }
1687 ntq_store_dest(c, &instr->dest, 0,
1688 c->color_reads[sample_index]);
1689 } else {
1690 offset = nir_intrinsic_base(instr) + const_offset->u32[0];
1691 int comp = nir_intrinsic_component(instr);
1692 ntq_store_dest(c, &instr->dest, 0,
1693 c->inputs[offset * 4 + comp]);
1694 }
1695 break;
1696
1697 case nir_intrinsic_store_output:
1698 const_offset = nir_src_as_const_value(instr->src[1]);
1699 assert(const_offset && "vc4 doesn't support indirect outputs");
1700 offset = nir_intrinsic_base(instr) + const_offset->u32[0];
1701
1702 /* MSAA color outputs are the only case where we have an
1703 * output that's not lowered to being a store of a single 32
1704 * bit value.
1705 */
1706 if (c->stage == QSTAGE_FRAG && instr->num_components == 4) {
1707 assert(offset == c->output_color_index);
1708 for (int i = 0; i < 4; i++) {
1709 c->sample_colors[i] =
1710 qir_MOV(c, ntq_get_src(c, instr->src[0],
1711 i));
1712 }
1713 } else {
1714 offset = offset * 4 + nir_intrinsic_component(instr);
1715 assert(instr->num_components == 1);
1716 c->outputs[offset] =
1717 qir_MOV(c, ntq_get_src(c, instr->src[0], 0));
1718 c->num_outputs = MAX2(c->num_outputs, offset + 1);
1719 }
1720 break;
1721
1722 case nir_intrinsic_discard:
1723 if (c->execute.file != QFILE_NULL) {
1724 qir_SF(c, c->execute);
1725 qir_MOV_cond(c, QPU_COND_ZS, c->discard,
1726 qir_uniform_ui(c, ~0));
1727 } else {
1728 qir_MOV_dest(c, c->discard, qir_uniform_ui(c, ~0));
1729 }
1730 break;
1731
1732 case nir_intrinsic_discard_if: {
1733 /* true (~0) if we're discarding */
1734 struct qreg cond = ntq_get_src(c, instr->src[0], 0);
1735
1736 if (c->execute.file != QFILE_NULL) {
1737 /* execute == 0 means the channel is active. Invert
1738 * the condition so that we can use zero as "executing
1739 * and discarding."
1740 */
1741 qir_SF(c, qir_AND(c, c->execute, qir_NOT(c, cond)));
1742 qir_MOV_cond(c, QPU_COND_ZS, c->discard, cond);
1743 } else {
1744 qir_OR_dest(c, c->discard, c->discard,
1745 ntq_get_src(c, instr->src[0], 0));
1746 }
1747
1748 break;
1749 }
1750
1751 default:
1752 fprintf(stderr, "Unknown intrinsic: ");
1753 nir_print_instr(&instr->instr, stderr);
1754 fprintf(stderr, "\n");
1755 break;
1756 }
1757 }
1758
1759 /* Clears (activates) the execute flags for any channels whose jump target
1760 * matches this block.
1761 */
1762 static void
1763 ntq_activate_execute_for_block(struct vc4_compile *c)
1764 {
1765 qir_SF(c, qir_SUB(c,
1766 c->execute,
1767 qir_uniform_ui(c, c->cur_block->index)));
1768 qir_MOV_cond(c, QPU_COND_ZS, c->execute, qir_uniform_ui(c, 0));
1769 }
1770
1771 static void
1772 ntq_emit_if(struct vc4_compile *c, nir_if *if_stmt)
1773 {
1774 if (!c->vc4->screen->has_control_flow) {
1775 fprintf(stderr,
1776 "IF statement support requires updated kernel.\n");
1777 return;
1778 }
1779
1780 nir_block *nir_else_block = nir_if_first_else_block(if_stmt);
1781 bool empty_else_block =
1782 (nir_else_block == nir_if_last_else_block(if_stmt) &&
1783 exec_list_is_empty(&nir_else_block->instr_list));
1784
1785 struct qblock *then_block = qir_new_block(c);
1786 struct qblock *after_block = qir_new_block(c);
1787 struct qblock *else_block;
1788 if (empty_else_block)
1789 else_block = after_block;
1790 else
1791 else_block = qir_new_block(c);
1792
1793 bool was_top_level = false;
1794 if (c->execute.file == QFILE_NULL) {
1795 c->execute = qir_MOV(c, qir_uniform_ui(c, 0));
1796 was_top_level = true;
1797 }
1798
1799 /* Set ZS for executing (execute == 0) and jumping (if->condition ==
1800 * 0) channels, and then update execute flags for those to point to
1801 * the ELSE block.
1802 */
1803 qir_SF(c, qir_OR(c,
1804 c->execute,
1805 ntq_get_src(c, if_stmt->condition, 0)));
1806 qir_MOV_cond(c, QPU_COND_ZS, c->execute,
1807 qir_uniform_ui(c, else_block->index));
1808
1809 /* Jump to ELSE if nothing is active for THEN, otherwise fall
1810 * through.
1811 */
1812 qir_SF(c, c->execute);
1813 qir_BRANCH(c, QPU_COND_BRANCH_ALL_ZC);
1814 qir_link_blocks(c->cur_block, else_block);
1815 qir_link_blocks(c->cur_block, then_block);
1816
1817 /* Process the THEN block. */
1818 qir_set_emit_block(c, then_block);
1819 ntq_emit_cf_list(c, &if_stmt->then_list);
1820
1821 if (!empty_else_block) {
1822 /* Handle the end of the THEN block. First, all currently
1823 * active channels update their execute flags to point to
1824 * ENDIF
1825 */
1826 qir_SF(c, c->execute);
1827 qir_MOV_cond(c, QPU_COND_ZS, c->execute,
1828 qir_uniform_ui(c, after_block->index));
1829
1830 /* If everything points at ENDIF, then jump there immediately. */
1831 qir_SF(c, qir_SUB(c, c->execute, qir_uniform_ui(c, after_block->index)));
1832 qir_BRANCH(c, QPU_COND_BRANCH_ALL_ZS);
1833 qir_link_blocks(c->cur_block, after_block);
1834 qir_link_blocks(c->cur_block, else_block);
1835
1836 qir_set_emit_block(c, else_block);
1837 ntq_activate_execute_for_block(c);
1838 ntq_emit_cf_list(c, &if_stmt->else_list);
1839 }
1840
1841 qir_link_blocks(c->cur_block, after_block);
1842
1843 qir_set_emit_block(c, after_block);
1844 if (was_top_level)
1845 c->execute = c->undef;
1846 else
1847 ntq_activate_execute_for_block(c);
1848
1849 }
1850
1851 static void
1852 ntq_emit_jump(struct vc4_compile *c, nir_jump_instr *jump)
1853 {
1854 switch (jump->type) {
1855 case nir_jump_break:
1856 qir_SF(c, c->execute);
1857 qir_MOV_cond(c, QPU_COND_ZS, c->execute,
1858 qir_uniform_ui(c, c->loop_break_block->index));
1859 break;
1860
1861 case nir_jump_continue:
1862 qir_SF(c, c->execute);
1863 qir_MOV_cond(c, QPU_COND_ZS, c->execute,
1864 qir_uniform_ui(c, c->loop_cont_block->index));
1865 break;
1866
1867 case nir_jump_return:
1868 unreachable("All returns shouold be lowered\n");
1869 }
1870 }
1871
1872 static void
1873 ntq_emit_instr(struct vc4_compile *c, nir_instr *instr)
1874 {
1875 switch (instr->type) {
1876 case nir_instr_type_alu:
1877 ntq_emit_alu(c, nir_instr_as_alu(instr));
1878 break;
1879
1880 case nir_instr_type_intrinsic:
1881 ntq_emit_intrinsic(c, nir_instr_as_intrinsic(instr));
1882 break;
1883
1884 case nir_instr_type_load_const:
1885 ntq_emit_load_const(c, nir_instr_as_load_const(instr));
1886 break;
1887
1888 case nir_instr_type_ssa_undef:
1889 ntq_emit_ssa_undef(c, nir_instr_as_ssa_undef(instr));
1890 break;
1891
1892 case nir_instr_type_tex:
1893 ntq_emit_tex(c, nir_instr_as_tex(instr));
1894 break;
1895
1896 case nir_instr_type_jump:
1897 ntq_emit_jump(c, nir_instr_as_jump(instr));
1898 break;
1899
1900 default:
1901 fprintf(stderr, "Unknown NIR instr type: ");
1902 nir_print_instr(instr, stderr);
1903 fprintf(stderr, "\n");
1904 abort();
1905 }
1906 }
1907
1908 static void
1909 ntq_emit_block(struct vc4_compile *c, nir_block *block)
1910 {
1911 nir_foreach_instr(instr, block) {
1912 ntq_emit_instr(c, instr);
1913 }
1914 }
1915
1916 static void ntq_emit_cf_list(struct vc4_compile *c, struct exec_list *list);
1917
1918 static void
1919 ntq_emit_loop(struct vc4_compile *c, nir_loop *loop)
1920 {
1921 if (!c->vc4->screen->has_control_flow) {
1922 fprintf(stderr,
1923 "loop support requires updated kernel.\n");
1924 ntq_emit_cf_list(c, &loop->body);
1925 return;
1926 }
1927
1928 bool was_top_level = false;
1929 if (c->execute.file == QFILE_NULL) {
1930 c->execute = qir_MOV(c, qir_uniform_ui(c, 0));
1931 was_top_level = true;
1932 }
1933
1934 struct qblock *save_loop_cont_block = c->loop_cont_block;
1935 struct qblock *save_loop_break_block = c->loop_break_block;
1936
1937 c->loop_cont_block = qir_new_block(c);
1938 c->loop_break_block = qir_new_block(c);
1939
1940 qir_link_blocks(c->cur_block, c->loop_cont_block);
1941 qir_set_emit_block(c, c->loop_cont_block);
1942 ntq_activate_execute_for_block(c);
1943
1944 ntq_emit_cf_list(c, &loop->body);
1945
1946 /* If anything had explicitly continued, or is here at the end of the
1947 * loop, then we need to loop again. SF updates are masked by the
1948 * instruction's condition, so we can do the OR of the two conditions
1949 * within SF.
1950 */
1951 qir_SF(c, c->execute);
1952 struct qinst *cont_check =
1953 qir_SUB_dest(c,
1954 c->undef,
1955 c->execute,
1956 qir_uniform_ui(c, c->loop_cont_block->index));
1957 cont_check->cond = QPU_COND_ZC;
1958 cont_check->sf = true;
1959
1960 qir_BRANCH(c, QPU_COND_BRANCH_ANY_ZS);
1961 qir_link_blocks(c->cur_block, c->loop_cont_block);
1962 qir_link_blocks(c->cur_block, c->loop_break_block);
1963
1964 qir_set_emit_block(c, c->loop_break_block);
1965 if (was_top_level)
1966 c->execute = c->undef;
1967 else
1968 ntq_activate_execute_for_block(c);
1969
1970 c->loop_break_block = save_loop_break_block;
1971 c->loop_cont_block = save_loop_cont_block;
1972 }
1973
1974 static void
1975 ntq_emit_function(struct vc4_compile *c, nir_function_impl *func)
1976 {
1977 fprintf(stderr, "FUNCTIONS not handled.\n");
1978 abort();
1979 }
1980
1981 static void
1982 ntq_emit_cf_list(struct vc4_compile *c, struct exec_list *list)
1983 {
1984 foreach_list_typed(nir_cf_node, node, node, list) {
1985 switch (node->type) {
1986 case nir_cf_node_block:
1987 ntq_emit_block(c, nir_cf_node_as_block(node));
1988 break;
1989
1990 case nir_cf_node_if:
1991 ntq_emit_if(c, nir_cf_node_as_if(node));
1992 break;
1993
1994 case nir_cf_node_loop:
1995 ntq_emit_loop(c, nir_cf_node_as_loop(node));
1996 break;
1997
1998 case nir_cf_node_function:
1999 ntq_emit_function(c, nir_cf_node_as_function(node));
2000 break;
2001
2002 default:
2003 fprintf(stderr, "Unknown NIR node type\n");
2004 abort();
2005 }
2006 }
2007 }
2008
2009 static void
2010 ntq_emit_impl(struct vc4_compile *c, nir_function_impl *impl)
2011 {
2012 ntq_setup_registers(c, &impl->registers);
2013 ntq_emit_cf_list(c, &impl->body);
2014 }
2015
2016 static void
2017 nir_to_qir(struct vc4_compile *c)
2018 {
2019 if (c->stage == QSTAGE_FRAG && c->s->info.fs.uses_discard)
2020 c->discard = qir_MOV(c, qir_uniform_ui(c, 0));
2021
2022 ntq_setup_inputs(c);
2023 ntq_setup_outputs(c);
2024 ntq_setup_uniforms(c);
2025 ntq_setup_registers(c, &c->s->registers);
2026
2027 /* Find the main function and emit the body. */
2028 nir_foreach_function(function, c->s) {
2029 assert(strcmp(function->name, "main") == 0);
2030 assert(function->impl);
2031 ntq_emit_impl(c, function->impl);
2032 }
2033 }
2034
2035 static const nir_shader_compiler_options nir_options = {
2036 .lower_extract_byte = true,
2037 .lower_extract_word = true,
2038 .lower_ffma = true,
2039 .lower_flrp32 = true,
2040 .lower_fpow = true,
2041 .lower_fsat = true,
2042 .lower_fsqrt = true,
2043 .lower_negate = true,
2044 .native_integers = true,
2045 };
2046
2047 const void *
2048 vc4_screen_get_compiler_options(struct pipe_screen *pscreen,
2049 enum pipe_shader_ir ir, unsigned shader)
2050 {
2051 return &nir_options;
2052 }
2053
2054 static int
2055 count_nir_instrs(nir_shader *nir)
2056 {
2057 int count = 0;
2058 nir_foreach_function(function, nir) {
2059 if (!function->impl)
2060 continue;
2061 nir_foreach_block(block, function->impl) {
2062 nir_foreach_instr(instr, block)
2063 count++;
2064 }
2065 }
2066 return count;
2067 }
2068
2069 static struct vc4_compile *
2070 vc4_shader_ntq(struct vc4_context *vc4, enum qstage stage,
2071 struct vc4_key *key)
2072 {
2073 struct vc4_compile *c = qir_compile_init();
2074
2075 c->vc4 = vc4;
2076 c->stage = stage;
2077 c->shader_state = &key->shader_state->base;
2078 c->program_id = key->shader_state->program_id;
2079 c->variant_id =
2080 p_atomic_inc_return(&key->shader_state->compiled_variant_count);
2081
2082 c->key = key;
2083 switch (stage) {
2084 case QSTAGE_FRAG:
2085 c->fs_key = (struct vc4_fs_key *)key;
2086 if (c->fs_key->is_points) {
2087 c->point_x = emit_fragment_varying(c, ~0, 0);
2088 c->point_y = emit_fragment_varying(c, ~0, 0);
2089 } else if (c->fs_key->is_lines) {
2090 c->line_x = emit_fragment_varying(c, ~0, 0);
2091 }
2092 break;
2093 case QSTAGE_VERT:
2094 c->vs_key = (struct vc4_vs_key *)key;
2095 break;
2096 case QSTAGE_COORD:
2097 c->vs_key = (struct vc4_vs_key *)key;
2098 break;
2099 }
2100
2101 c->s = nir_shader_clone(c, key->shader_state->base.ir.nir);
2102
2103 if (stage == QSTAGE_FRAG)
2104 NIR_PASS_V(c->s, vc4_nir_lower_blend, c);
2105
2106 struct nir_lower_tex_options tex_options = {
2107 /* We would need to implement txs, but we don't want the
2108 * int/float conversions
2109 */
2110 .lower_rect = false,
2111
2112 .lower_txp = ~0,
2113
2114 /* Apply swizzles to all samplers. */
2115 .swizzle_result = ~0,
2116 };
2117
2118 /* Lower the format swizzle and ARB_texture_swizzle-style swizzle.
2119 * The format swizzling applies before sRGB decode, and
2120 * ARB_texture_swizzle is the last thing before returning the sample.
2121 */
2122 for (int i = 0; i < ARRAY_SIZE(key->tex); i++) {
2123 enum pipe_format format = c->key->tex[i].format;
2124
2125 if (!format)
2126 continue;
2127
2128 const uint8_t *format_swizzle = vc4_get_format_swizzle(format);
2129
2130 for (int j = 0; j < 4; j++) {
2131 uint8_t arb_swiz = c->key->tex[i].swizzle[j];
2132
2133 if (arb_swiz <= 3) {
2134 tex_options.swizzles[i][j] =
2135 format_swizzle[arb_swiz];
2136 } else {
2137 tex_options.swizzles[i][j] = arb_swiz;
2138 }
2139 }
2140
2141 if (util_format_is_srgb(format))
2142 tex_options.lower_srgb |= (1 << i);
2143 }
2144
2145 NIR_PASS_V(c->s, nir_lower_tex, &tex_options);
2146
2147 if (c->fs_key && c->fs_key->light_twoside)
2148 NIR_PASS_V(c->s, nir_lower_two_sided_color);
2149
2150 if (c->vs_key && c->vs_key->clamp_color)
2151 NIR_PASS_V(c->s, nir_lower_clamp_color_outputs);
2152
2153 if (c->key->ucp_enables) {
2154 if (stage == QSTAGE_FRAG) {
2155 NIR_PASS_V(c->s, nir_lower_clip_fs, c->key->ucp_enables);
2156 } else {
2157 NIR_PASS_V(c->s, nir_lower_clip_vs, c->key->ucp_enables);
2158 NIR_PASS_V(c->s, nir_lower_io_to_scalar,
2159 nir_var_shader_out);
2160 }
2161 }
2162
2163 /* FS input scalarizing must happen after nir_lower_two_sided_color,
2164 * which only handles a vec4 at a time. Similarly, VS output
2165 * scalarizing must happen after nir_lower_clip_vs.
2166 */
2167 if (c->stage == QSTAGE_FRAG)
2168 NIR_PASS_V(c->s, nir_lower_io_to_scalar, nir_var_shader_in);
2169 else
2170 NIR_PASS_V(c->s, nir_lower_io_to_scalar, nir_var_shader_out);
2171
2172 NIR_PASS_V(c->s, vc4_nir_lower_io, c);
2173 NIR_PASS_V(c->s, vc4_nir_lower_txf_ms, c);
2174 NIR_PASS_V(c->s, nir_lower_idiv);
2175
2176 vc4_optimize_nir(c->s);
2177
2178 NIR_PASS_V(c->s, nir_convert_from_ssa, true);
2179
2180 if (vc4_debug & VC4_DEBUG_SHADERDB) {
2181 fprintf(stderr, "SHADER-DB: %s prog %d/%d: %d NIR instructions\n",
2182 qir_get_stage_name(c->stage),
2183 c->program_id, c->variant_id,
2184 count_nir_instrs(c->s));
2185 }
2186
2187 if (vc4_debug & VC4_DEBUG_NIR) {
2188 fprintf(stderr, "%s prog %d/%d NIR:\n",
2189 qir_get_stage_name(c->stage),
2190 c->program_id, c->variant_id);
2191 nir_print_shader(c->s, stderr);
2192 }
2193
2194 nir_to_qir(c);
2195
2196 switch (stage) {
2197 case QSTAGE_FRAG:
2198 emit_frag_end(c);
2199 break;
2200 case QSTAGE_VERT:
2201 emit_vert_end(c,
2202 c->vs_key->fs_inputs->input_slots,
2203 c->vs_key->fs_inputs->num_inputs);
2204 break;
2205 case QSTAGE_COORD:
2206 emit_coord_end(c);
2207 break;
2208 }
2209
2210 if (vc4_debug & VC4_DEBUG_QIR) {
2211 fprintf(stderr, "%s prog %d/%d pre-opt QIR:\n",
2212 qir_get_stage_name(c->stage),
2213 c->program_id, c->variant_id);
2214 qir_dump(c);
2215 fprintf(stderr, "\n");
2216 }
2217
2218 qir_optimize(c);
2219 qir_lower_uniforms(c);
2220
2221 qir_schedule_instructions(c);
2222 qir_emit_uniform_stream_resets(c);
2223
2224 if (vc4_debug & VC4_DEBUG_QIR) {
2225 fprintf(stderr, "%s prog %d/%d QIR:\n",
2226 qir_get_stage_name(c->stage),
2227 c->program_id, c->variant_id);
2228 qir_dump(c);
2229 fprintf(stderr, "\n");
2230 }
2231
2232 qir_reorder_uniforms(c);
2233 vc4_generate_code(vc4, c);
2234
2235 if (vc4_debug & VC4_DEBUG_SHADERDB) {
2236 fprintf(stderr, "SHADER-DB: %s prog %d/%d: %d instructions\n",
2237 qir_get_stage_name(c->stage),
2238 c->program_id, c->variant_id,
2239 c->qpu_inst_count);
2240 fprintf(stderr, "SHADER-DB: %s prog %d/%d: %d uniforms\n",
2241 qir_get_stage_name(c->stage),
2242 c->program_id, c->variant_id,
2243 c->num_uniforms);
2244 }
2245
2246 ralloc_free(c->s);
2247
2248 return c;
2249 }
2250
2251 static void *
2252 vc4_shader_state_create(struct pipe_context *pctx,
2253 const struct pipe_shader_state *cso)
2254 {
2255 struct vc4_context *vc4 = vc4_context(pctx);
2256 struct vc4_uncompiled_shader *so = CALLOC_STRUCT(vc4_uncompiled_shader);
2257 if (!so)
2258 return NULL;
2259
2260 so->program_id = vc4->next_uncompiled_program_id++;
2261
2262 nir_shader *s;
2263
2264 if (cso->type == PIPE_SHADER_IR_NIR) {
2265 /* The backend takes ownership of the NIR shader on state
2266 * creation.
2267 */
2268 s = cso->ir.nir;
2269 } else {
2270 assert(cso->type == PIPE_SHADER_IR_TGSI);
2271
2272 if (vc4_debug & VC4_DEBUG_TGSI) {
2273 fprintf(stderr, "prog %d TGSI:\n",
2274 so->program_id);
2275 tgsi_dump(cso->tokens, 0);
2276 fprintf(stderr, "\n");
2277 }
2278 s = tgsi_to_nir(cso->tokens, &nir_options);
2279 }
2280
2281 NIR_PASS_V(s, nir_opt_global_to_local);
2282 NIR_PASS_V(s, nir_convert_to_ssa);
2283 NIR_PASS_V(s, nir_normalize_cubemap_coords);
2284
2285 NIR_PASS_V(s, nir_lower_load_const_to_scalar);
2286
2287 vc4_optimize_nir(s);
2288
2289 NIR_PASS_V(s, nir_remove_dead_variables, nir_var_local);
2290
2291 /* Garbage collect dead instructions */
2292 nir_sweep(s);
2293
2294 so->base.type = PIPE_SHADER_IR_NIR;
2295 so->base.ir.nir = s;
2296
2297 if (vc4_debug & VC4_DEBUG_NIR) {
2298 fprintf(stderr, "%s prog %d NIR:\n",
2299 gl_shader_stage_name(s->stage),
2300 so->program_id);
2301 nir_print_shader(s, stderr);
2302 fprintf(stderr, "\n");
2303 }
2304
2305 return so;
2306 }
2307
2308 static void
2309 copy_uniform_state_to_shader(struct vc4_compiled_shader *shader,
2310 struct vc4_compile *c)
2311 {
2312 int count = c->num_uniforms;
2313 struct vc4_shader_uniform_info *uinfo = &shader->uniforms;
2314
2315 uinfo->count = count;
2316 uinfo->data = ralloc_array(shader, uint32_t, count);
2317 memcpy(uinfo->data, c->uniform_data,
2318 count * sizeof(*uinfo->data));
2319 uinfo->contents = ralloc_array(shader, enum quniform_contents, count);
2320 memcpy(uinfo->contents, c->uniform_contents,
2321 count * sizeof(*uinfo->contents));
2322 uinfo->num_texture_samples = c->num_texture_samples;
2323
2324 vc4_set_shader_uniform_dirty_flags(shader);
2325 }
2326
2327 static void
2328 vc4_setup_compiled_fs_inputs(struct vc4_context *vc4, struct vc4_compile *c,
2329 struct vc4_compiled_shader *shader)
2330 {
2331 struct vc4_fs_inputs inputs;
2332
2333 memset(&inputs, 0, sizeof(inputs));
2334 inputs.input_slots = ralloc_array(shader,
2335 struct vc4_varying_slot,
2336 c->num_input_slots);
2337
2338 bool input_live[c->num_input_slots];
2339
2340 memset(input_live, 0, sizeof(input_live));
2341 qir_for_each_inst_inorder(inst, c) {
2342 for (int i = 0; i < qir_get_op_nsrc(inst->op); i++) {
2343 if (inst->src[i].file == QFILE_VARY)
2344 input_live[inst->src[i].index] = true;
2345 }
2346 }
2347
2348 for (int i = 0; i < c->num_input_slots; i++) {
2349 struct vc4_varying_slot *slot = &c->input_slots[i];
2350
2351 if (!input_live[i])
2352 continue;
2353
2354 /* Skip non-VS-output inputs. */
2355 if (slot->slot == (uint8_t)~0)
2356 continue;
2357
2358 if (slot->slot == VARYING_SLOT_COL0 ||
2359 slot->slot == VARYING_SLOT_COL1 ||
2360 slot->slot == VARYING_SLOT_BFC0 ||
2361 slot->slot == VARYING_SLOT_BFC1) {
2362 shader->color_inputs |= (1 << inputs.num_inputs);
2363 }
2364
2365 inputs.input_slots[inputs.num_inputs] = *slot;
2366 inputs.num_inputs++;
2367 }
2368 shader->num_inputs = inputs.num_inputs;
2369
2370 /* Add our set of inputs to the set of all inputs seen. This way, we
2371 * can have a single pointer that identifies an FS inputs set,
2372 * allowing VS to avoid recompiling when the FS is recompiled (or a
2373 * new one is bound using separate shader objects) but the inputs
2374 * don't change.
2375 */
2376 struct set_entry *entry = _mesa_set_search(vc4->fs_inputs_set, &inputs);
2377 if (entry) {
2378 shader->fs_inputs = entry->key;
2379 ralloc_free(inputs.input_slots);
2380 } else {
2381 struct vc4_fs_inputs *alloc_inputs;
2382
2383 alloc_inputs = rzalloc(vc4->fs_inputs_set, struct vc4_fs_inputs);
2384 memcpy(alloc_inputs, &inputs, sizeof(inputs));
2385 ralloc_steal(alloc_inputs, inputs.input_slots);
2386 _mesa_set_add(vc4->fs_inputs_set, alloc_inputs);
2387
2388 shader->fs_inputs = alloc_inputs;
2389 }
2390 }
2391
2392 static struct vc4_compiled_shader *
2393 vc4_get_compiled_shader(struct vc4_context *vc4, enum qstage stage,
2394 struct vc4_key *key)
2395 {
2396 struct hash_table *ht;
2397 uint32_t key_size;
2398 if (stage == QSTAGE_FRAG) {
2399 ht = vc4->fs_cache;
2400 key_size = sizeof(struct vc4_fs_key);
2401 } else {
2402 ht = vc4->vs_cache;
2403 key_size = sizeof(struct vc4_vs_key);
2404 }
2405
2406 struct vc4_compiled_shader *shader;
2407 struct hash_entry *entry = _mesa_hash_table_search(ht, key);
2408 if (entry)
2409 return entry->data;
2410
2411 struct vc4_compile *c = vc4_shader_ntq(vc4, stage, key);
2412 shader = rzalloc(NULL, struct vc4_compiled_shader);
2413
2414 shader->program_id = vc4->next_compiled_program_id++;
2415 if (stage == QSTAGE_FRAG) {
2416 vc4_setup_compiled_fs_inputs(vc4, c, shader);
2417
2418 /* Note: the temporary clone in c->s has been freed. */
2419 nir_shader *orig_shader = key->shader_state->base.ir.nir;
2420 if (orig_shader->info.outputs_written & (1 << FRAG_RESULT_DEPTH))
2421 shader->disable_early_z = true;
2422 } else {
2423 shader->num_inputs = c->num_inputs;
2424
2425 shader->vattr_offsets[0] = 0;
2426 for (int i = 0; i < 8; i++) {
2427 shader->vattr_offsets[i + 1] =
2428 shader->vattr_offsets[i] + c->vattr_sizes[i];
2429
2430 if (c->vattr_sizes[i])
2431 shader->vattrs_live |= (1 << i);
2432 }
2433 }
2434
2435 copy_uniform_state_to_shader(shader, c);
2436 shader->bo = vc4_bo_alloc_shader(vc4->screen, c->qpu_insts,
2437 c->qpu_inst_count * sizeof(uint64_t));
2438
2439 /* Copy the compiler UBO range state to the compiled shader, dropping
2440 * out arrays that were never referenced by an indirect load.
2441 *
2442 * (Note that QIR dead code elimination of an array access still
2443 * leaves that array alive, though)
2444 */
2445 if (c->num_ubo_ranges) {
2446 shader->num_ubo_ranges = c->num_ubo_ranges;
2447 shader->ubo_ranges = ralloc_array(shader, struct vc4_ubo_range,
2448 c->num_ubo_ranges);
2449 uint32_t j = 0;
2450 for (int i = 0; i < c->num_uniform_ranges; i++) {
2451 struct vc4_compiler_ubo_range *range =
2452 &c->ubo_ranges[i];
2453 if (!range->used)
2454 continue;
2455
2456 shader->ubo_ranges[j].dst_offset = range->dst_offset;
2457 shader->ubo_ranges[j].src_offset = range->src_offset;
2458 shader->ubo_ranges[j].size = range->size;
2459 shader->ubo_size += c->ubo_ranges[i].size;
2460 j++;
2461 }
2462 }
2463 if (shader->ubo_size) {
2464 if (vc4_debug & VC4_DEBUG_SHADERDB) {
2465 fprintf(stderr, "SHADER-DB: %s prog %d/%d: %d UBO uniforms\n",
2466 qir_get_stage_name(c->stage),
2467 c->program_id, c->variant_id,
2468 shader->ubo_size / 4);
2469 }
2470 }
2471
2472 qir_compile_destroy(c);
2473
2474 struct vc4_key *dup_key;
2475 dup_key = ralloc_size(shader, key_size);
2476 memcpy(dup_key, key, key_size);
2477 _mesa_hash_table_insert(ht, dup_key, shader);
2478
2479 return shader;
2480 }
2481
2482 static void
2483 vc4_setup_shared_key(struct vc4_context *vc4, struct vc4_key *key,
2484 struct vc4_texture_stateobj *texstate)
2485 {
2486 for (int i = 0; i < texstate->num_textures; i++) {
2487 struct pipe_sampler_view *sampler = texstate->textures[i];
2488 struct vc4_sampler_view *vc4_sampler = vc4_sampler_view(sampler);
2489 struct pipe_sampler_state *sampler_state =
2490 texstate->samplers[i];
2491
2492 if (!sampler)
2493 continue;
2494
2495 key->tex[i].format = sampler->format;
2496 key->tex[i].swizzle[0] = sampler->swizzle_r;
2497 key->tex[i].swizzle[1] = sampler->swizzle_g;
2498 key->tex[i].swizzle[2] = sampler->swizzle_b;
2499 key->tex[i].swizzle[3] = sampler->swizzle_a;
2500
2501 if (sampler->texture->nr_samples > 1) {
2502 key->tex[i].msaa_width = sampler->texture->width0;
2503 key->tex[i].msaa_height = sampler->texture->height0;
2504 } else if (sampler){
2505 key->tex[i].compare_mode = sampler_state->compare_mode;
2506 key->tex[i].compare_func = sampler_state->compare_func;
2507 key->tex[i].wrap_s = sampler_state->wrap_s;
2508 key->tex[i].wrap_t = sampler_state->wrap_t;
2509 key->tex[i].force_first_level =
2510 vc4_sampler->force_first_level;
2511 }
2512 }
2513
2514 key->ucp_enables = vc4->rasterizer->base.clip_plane_enable;
2515 }
2516
2517 static void
2518 vc4_update_compiled_fs(struct vc4_context *vc4, uint8_t prim_mode)
2519 {
2520 struct vc4_job *job = vc4->job;
2521 struct vc4_fs_key local_key;
2522 struct vc4_fs_key *key = &local_key;
2523
2524 if (!(vc4->dirty & (VC4_DIRTY_PRIM_MODE |
2525 VC4_DIRTY_BLEND |
2526 VC4_DIRTY_FRAMEBUFFER |
2527 VC4_DIRTY_ZSA |
2528 VC4_DIRTY_RASTERIZER |
2529 VC4_DIRTY_SAMPLE_MASK |
2530 VC4_DIRTY_FRAGTEX |
2531 VC4_DIRTY_UNCOMPILED_FS))) {
2532 return;
2533 }
2534
2535 memset(key, 0, sizeof(*key));
2536 vc4_setup_shared_key(vc4, &key->base, &vc4->fragtex);
2537 key->base.shader_state = vc4->prog.bind_fs;
2538 key->is_points = (prim_mode == PIPE_PRIM_POINTS);
2539 key->is_lines = (prim_mode >= PIPE_PRIM_LINES &&
2540 prim_mode <= PIPE_PRIM_LINE_STRIP);
2541 key->blend = vc4->blend->rt[0];
2542 if (vc4->blend->logicop_enable) {
2543 key->logicop_func = vc4->blend->logicop_func;
2544 } else {
2545 key->logicop_func = PIPE_LOGICOP_COPY;
2546 }
2547 if (job->msaa) {
2548 key->msaa = vc4->rasterizer->base.multisample;
2549 key->sample_coverage = (vc4->rasterizer->base.multisample &&
2550 vc4->sample_mask != (1 << VC4_MAX_SAMPLES) - 1);
2551 key->sample_alpha_to_coverage = vc4->blend->alpha_to_coverage;
2552 key->sample_alpha_to_one = vc4->blend->alpha_to_one;
2553 }
2554
2555 if (vc4->framebuffer.cbufs[0])
2556 key->color_format = vc4->framebuffer.cbufs[0]->format;
2557
2558 key->stencil_enabled = vc4->zsa->stencil_uniforms[0] != 0;
2559 key->stencil_twoside = vc4->zsa->stencil_uniforms[1] != 0;
2560 key->stencil_full_writemasks = vc4->zsa->stencil_uniforms[2] != 0;
2561 key->depth_enabled = (vc4->zsa->base.depth.enabled ||
2562 key->stencil_enabled);
2563 if (vc4->zsa->base.alpha.enabled) {
2564 key->alpha_test = true;
2565 key->alpha_test_func = vc4->zsa->base.alpha.func;
2566 }
2567
2568 if (key->is_points) {
2569 key->point_sprite_mask =
2570 vc4->rasterizer->base.sprite_coord_enable;
2571 key->point_coord_upper_left =
2572 (vc4->rasterizer->base.sprite_coord_mode ==
2573 PIPE_SPRITE_COORD_UPPER_LEFT);
2574 }
2575
2576 key->light_twoside = vc4->rasterizer->base.light_twoside;
2577
2578 struct vc4_compiled_shader *old_fs = vc4->prog.fs;
2579 vc4->prog.fs = vc4_get_compiled_shader(vc4, QSTAGE_FRAG, &key->base);
2580 if (vc4->prog.fs == old_fs)
2581 return;
2582
2583 vc4->dirty |= VC4_DIRTY_COMPILED_FS;
2584
2585 if (vc4->rasterizer->base.flatshade &&
2586 old_fs && vc4->prog.fs->color_inputs != old_fs->color_inputs) {
2587 vc4->dirty |= VC4_DIRTY_FLAT_SHADE_FLAGS;
2588 }
2589
2590 if (old_fs && vc4->prog.fs->fs_inputs != old_fs->fs_inputs)
2591 vc4->dirty |= VC4_DIRTY_FS_INPUTS;
2592 }
2593
2594 static void
2595 vc4_update_compiled_vs(struct vc4_context *vc4, uint8_t prim_mode)
2596 {
2597 struct vc4_vs_key local_key;
2598 struct vc4_vs_key *key = &local_key;
2599
2600 if (!(vc4->dirty & (VC4_DIRTY_PRIM_MODE |
2601 VC4_DIRTY_RASTERIZER |
2602 VC4_DIRTY_VERTTEX |
2603 VC4_DIRTY_VTXSTATE |
2604 VC4_DIRTY_UNCOMPILED_VS |
2605 VC4_DIRTY_FS_INPUTS))) {
2606 return;
2607 }
2608
2609 memset(key, 0, sizeof(*key));
2610 vc4_setup_shared_key(vc4, &key->base, &vc4->verttex);
2611 key->base.shader_state = vc4->prog.bind_vs;
2612 key->fs_inputs = vc4->prog.fs->fs_inputs;
2613 key->clamp_color = vc4->rasterizer->base.clamp_vertex_color;
2614
2615 for (int i = 0; i < ARRAY_SIZE(key->attr_formats); i++)
2616 key->attr_formats[i] = vc4->vtx->pipe[i].src_format;
2617
2618 key->per_vertex_point_size =
2619 (prim_mode == PIPE_PRIM_POINTS &&
2620 vc4->rasterizer->base.point_size_per_vertex);
2621
2622 struct vc4_compiled_shader *vs =
2623 vc4_get_compiled_shader(vc4, QSTAGE_VERT, &key->base);
2624 if (vs != vc4->prog.vs) {
2625 vc4->prog.vs = vs;
2626 vc4->dirty |= VC4_DIRTY_COMPILED_VS;
2627 }
2628
2629 key->is_coord = true;
2630 /* Coord shaders don't care what the FS inputs are. */
2631 key->fs_inputs = NULL;
2632 struct vc4_compiled_shader *cs =
2633 vc4_get_compiled_shader(vc4, QSTAGE_COORD, &key->base);
2634 if (cs != vc4->prog.cs) {
2635 vc4->prog.cs = cs;
2636 vc4->dirty |= VC4_DIRTY_COMPILED_CS;
2637 }
2638 }
2639
2640 void
2641 vc4_update_compiled_shaders(struct vc4_context *vc4, uint8_t prim_mode)
2642 {
2643 vc4_update_compiled_fs(vc4, prim_mode);
2644 vc4_update_compiled_vs(vc4, prim_mode);
2645 }
2646
2647 static uint32_t
2648 fs_cache_hash(const void *key)
2649 {
2650 return _mesa_hash_data(key, sizeof(struct vc4_fs_key));
2651 }
2652
2653 static uint32_t
2654 vs_cache_hash(const void *key)
2655 {
2656 return _mesa_hash_data(key, sizeof(struct vc4_vs_key));
2657 }
2658
2659 static bool
2660 fs_cache_compare(const void *key1, const void *key2)
2661 {
2662 return memcmp(key1, key2, sizeof(struct vc4_fs_key)) == 0;
2663 }
2664
2665 static bool
2666 vs_cache_compare(const void *key1, const void *key2)
2667 {
2668 return memcmp(key1, key2, sizeof(struct vc4_vs_key)) == 0;
2669 }
2670
2671 static uint32_t
2672 fs_inputs_hash(const void *key)
2673 {
2674 const struct vc4_fs_inputs *inputs = key;
2675
2676 return _mesa_hash_data(inputs->input_slots,
2677 sizeof(*inputs->input_slots) *
2678 inputs->num_inputs);
2679 }
2680
2681 static bool
2682 fs_inputs_compare(const void *key1, const void *key2)
2683 {
2684 const struct vc4_fs_inputs *inputs1 = key1;
2685 const struct vc4_fs_inputs *inputs2 = key2;
2686
2687 return (inputs1->num_inputs == inputs2->num_inputs &&
2688 memcmp(inputs1->input_slots,
2689 inputs2->input_slots,
2690 sizeof(*inputs1->input_slots) *
2691 inputs1->num_inputs) == 0);
2692 }
2693
2694 static void
2695 delete_from_cache_if_matches(struct hash_table *ht,
2696 struct hash_entry *entry,
2697 struct vc4_uncompiled_shader *so)
2698 {
2699 const struct vc4_key *key = entry->key;
2700
2701 if (key->shader_state == so) {
2702 struct vc4_compiled_shader *shader = entry->data;
2703 _mesa_hash_table_remove(ht, entry);
2704 vc4_bo_unreference(&shader->bo);
2705 ralloc_free(shader);
2706 }
2707 }
2708
2709 static void
2710 vc4_shader_state_delete(struct pipe_context *pctx, void *hwcso)
2711 {
2712 struct vc4_context *vc4 = vc4_context(pctx);
2713 struct vc4_uncompiled_shader *so = hwcso;
2714
2715 struct hash_entry *entry;
2716 hash_table_foreach(vc4->fs_cache, entry)
2717 delete_from_cache_if_matches(vc4->fs_cache, entry, so);
2718 hash_table_foreach(vc4->vs_cache, entry)
2719 delete_from_cache_if_matches(vc4->vs_cache, entry, so);
2720
2721 ralloc_free(so->base.ir.nir);
2722 free(so);
2723 }
2724
2725 static void
2726 vc4_fp_state_bind(struct pipe_context *pctx, void *hwcso)
2727 {
2728 struct vc4_context *vc4 = vc4_context(pctx);
2729 vc4->prog.bind_fs = hwcso;
2730 vc4->dirty |= VC4_DIRTY_UNCOMPILED_FS;
2731 }
2732
2733 static void
2734 vc4_vp_state_bind(struct pipe_context *pctx, void *hwcso)
2735 {
2736 struct vc4_context *vc4 = vc4_context(pctx);
2737 vc4->prog.bind_vs = hwcso;
2738 vc4->dirty |= VC4_DIRTY_UNCOMPILED_VS;
2739 }
2740
2741 void
2742 vc4_program_init(struct pipe_context *pctx)
2743 {
2744 struct vc4_context *vc4 = vc4_context(pctx);
2745
2746 pctx->create_vs_state = vc4_shader_state_create;
2747 pctx->delete_vs_state = vc4_shader_state_delete;
2748
2749 pctx->create_fs_state = vc4_shader_state_create;
2750 pctx->delete_fs_state = vc4_shader_state_delete;
2751
2752 pctx->bind_fs_state = vc4_fp_state_bind;
2753 pctx->bind_vs_state = vc4_vp_state_bind;
2754
2755 vc4->fs_cache = _mesa_hash_table_create(pctx, fs_cache_hash,
2756 fs_cache_compare);
2757 vc4->vs_cache = _mesa_hash_table_create(pctx, vs_cache_hash,
2758 vs_cache_compare);
2759 vc4->fs_inputs_set = _mesa_set_create(pctx, fs_inputs_hash,
2760 fs_inputs_compare);
2761 }
2762
2763 void
2764 vc4_program_fini(struct pipe_context *pctx)
2765 {
2766 struct vc4_context *vc4 = vc4_context(pctx);
2767
2768 struct hash_entry *entry;
2769 hash_table_foreach(vc4->fs_cache, entry) {
2770 struct vc4_compiled_shader *shader = entry->data;
2771 vc4_bo_unreference(&shader->bo);
2772 ralloc_free(shader);
2773 _mesa_hash_table_remove(vc4->fs_cache, entry);
2774 }
2775
2776 hash_table_foreach(vc4->vs_cache, entry) {
2777 struct vc4_compiled_shader *shader = entry->data;
2778 vc4_bo_unreference(&shader->bo);
2779 ralloc_free(shader);
2780 _mesa_hash_table_remove(vc4->vs_cache, entry);
2781 }
2782 }