• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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/format/u_format.h"
27 #include "util/crc32.h"
28 #include "util/u_helpers.h"
29 #include "util/u_math.h"
30 #include "util/u_memory.h"
31 #include "util/ralloc.h"
32 #include "util/hash_table.h"
33 #include "tgsi/tgsi_dump.h"
34 #include "compiler/glsl_types.h"
35 #include "compiler/nir/nir.h"
36 #include "compiler/nir/nir_builder.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 struct vc4_compiled_shader *
48 vc4_get_compiled_shader(struct vc4_context *vc4, enum qstage stage,
49                         struct vc4_key *key);
50 
51 static int
type_size(const struct glsl_type * type,bool bindless)52 type_size(const struct glsl_type *type, bool bindless)
53 {
54    return glsl_count_attribute_slots(type, false);
55 }
56 
57 static void
resize_qreg_array(struct vc4_compile * c,struct qreg ** regs,uint32_t * size,uint32_t decl_size)58 resize_qreg_array(struct vc4_compile *c,
59                   struct qreg **regs,
60                   uint32_t *size,
61                   uint32_t decl_size)
62 {
63         if (*size >= decl_size)
64                 return;
65 
66         uint32_t old_size = *size;
67         *size = MAX2(*size * 2, decl_size);
68         *regs = reralloc(c, *regs, struct qreg, *size);
69         if (!*regs) {
70                 fprintf(stderr, "Malloc failure\n");
71                 abort();
72         }
73 
74         for (uint32_t i = old_size; i < *size; i++)
75                 (*regs)[i] = c->undef;
76 }
77 
78 static void
ntq_emit_thrsw(struct vc4_compile * c)79 ntq_emit_thrsw(struct vc4_compile *c)
80 {
81         if (!c->fs_threaded)
82                 return;
83 
84         /* Always thread switch after each texture operation for now.
85          *
86          * We could do better by batching a bunch of texture fetches up and
87          * then doing one thread switch and collecting all their results
88          * afterward.
89          */
90         qir_emit_nondef(c, qir_inst(QOP_THRSW, c->undef,
91                                     c->undef, c->undef));
92         c->last_thrsw_at_top_level = (c->execute.file == QFILE_NULL);
93 }
94 
95 static struct qreg
indirect_uniform_load(struct vc4_compile * c,nir_intrinsic_instr * intr)96 indirect_uniform_load(struct vc4_compile *c, nir_intrinsic_instr *intr)
97 {
98         struct qreg indirect_offset = ntq_get_src(c, intr->src[0], 0);
99 
100         /* Clamp to [0, array size).  Note that MIN/MAX are signed. */
101         uint32_t range = nir_intrinsic_range(intr);
102         indirect_offset = qir_MAX(c, indirect_offset, qir_uniform_ui(c, 0));
103         indirect_offset = qir_MIN_NOIMM(c, indirect_offset,
104                                         qir_uniform_ui(c, range - 4));
105 
106         qir_ADD_dest(c, qir_reg(QFILE_TEX_S_DIRECT, 0),
107                      indirect_offset,
108                      qir_uniform(c, QUNIFORM_UBO0_ADDR,
109                                  nir_intrinsic_base(intr)));
110 
111         c->num_texture_samples++;
112 
113         ntq_emit_thrsw(c);
114 
115         return qir_TEX_RESULT(c);
116 }
117 
118 static struct qreg
vc4_ubo_load(struct vc4_compile * c,nir_intrinsic_instr * intr)119 vc4_ubo_load(struct vc4_compile *c, nir_intrinsic_instr *intr)
120 {
121         ASSERTED int buffer_index = nir_src_as_uint(intr->src[0]);
122         assert(buffer_index == 1);
123         assert(c->stage == QSTAGE_FRAG);
124 
125         struct qreg offset = ntq_get_src(c, intr->src[1], 0);
126 
127         /* Clamp to [0, array size).  Note that MIN/MAX are signed. */
128         offset = qir_MAX(c, offset, qir_uniform_ui(c, 0));
129         offset = qir_MIN_NOIMM(c, offset,
130                                qir_uniform_ui(c, c->fs_key->ubo_1_size - 4));
131 
132         qir_ADD_dest(c, qir_reg(QFILE_TEX_S_DIRECT, 0),
133                      offset,
134                      qir_uniform(c, QUNIFORM_UBO1_ADDR, 0));
135 
136         c->num_texture_samples++;
137 
138         ntq_emit_thrsw(c);
139 
140         return qir_TEX_RESULT(c);
141 }
142 
143 nir_def *
vc4_nir_get_swizzled_channel(nir_builder * b,nir_def ** srcs,int swiz)144 vc4_nir_get_swizzled_channel(nir_builder *b, nir_def **srcs, int swiz)
145 {
146         switch (swiz) {
147         default:
148         case PIPE_SWIZZLE_NONE:
149                 fprintf(stderr, "warning: unknown swizzle\n");
150                 FALLTHROUGH;
151         case PIPE_SWIZZLE_0:
152                 return nir_imm_float(b, 0.0);
153         case PIPE_SWIZZLE_1:
154                 return nir_imm_float(b, 1.0);
155         case PIPE_SWIZZLE_X:
156         case PIPE_SWIZZLE_Y:
157         case PIPE_SWIZZLE_Z:
158         case PIPE_SWIZZLE_W:
159                 return srcs[swiz];
160         }
161 }
162 
163 static struct qreg *
ntq_init_ssa_def(struct vc4_compile * c,nir_def * def)164 ntq_init_ssa_def(struct vc4_compile *c, nir_def *def)
165 {
166         struct qreg *qregs = ralloc_array(c->def_ht, struct qreg,
167                                           def->num_components);
168         _mesa_hash_table_insert(c->def_ht, def, qregs);
169         return qregs;
170 }
171 
172 /**
173  * This function is responsible for getting QIR results into the associated
174  * storage for a NIR instruction.
175  *
176  * If it's a NIR SSA def, then we just set the associated hash table entry to
177  * the new result.
178  *
179  * If it's a NIR reg, then we need to update the existing qreg assigned to the
180  * NIR destination with the incoming value.  To do that without introducing
181  * new MOVs, we require that the incoming qreg either be a uniform, or be
182  * SSA-defined by the previous QIR instruction in the block and rewritable by
183  * this function.  That lets us sneak ahead and insert the SF flag beforehand
184  * (knowing that the previous instruction doesn't depend on flags) and rewrite
185  * its destination to be the NIR reg's destination
186  */
187 static void
ntq_store_def(struct vc4_compile * c,nir_def * def,int chan,struct qreg result)188 ntq_store_def(struct vc4_compile *c, nir_def *def, int chan,
189               struct qreg result)
190 {
191         struct qinst *last_inst = NULL;
192         if (!list_is_empty(&c->cur_block->instructions))
193                 last_inst = (struct qinst *)c->cur_block->instructions.prev;
194 
195         assert(result.file == QFILE_UNIF ||
196                (result.file == QFILE_TEMP &&
197                 last_inst && last_inst == c->defs[result.index]));
198 
199         nir_intrinsic_instr *store = nir_store_reg_for_def(def);
200         if (store == NULL) {
201                 assert(chan < def->num_components);
202 
203                 struct qreg *qregs;
204                 struct hash_entry *entry =
205                         _mesa_hash_table_search(c->def_ht, def);
206 
207                 if (entry)
208                         qregs = entry->data;
209                 else
210                         qregs = ntq_init_ssa_def(c, def);
211 
212                 qregs[chan] = result;
213         } else {
214                 nir_def *reg = store->src[1].ssa;
215                 ASSERTED nir_intrinsic_instr *decl = nir_reg_get_decl(reg);
216                 assert(nir_intrinsic_base(store) == 0);
217                 assert(nir_intrinsic_num_array_elems(decl) == 0);
218                 struct hash_entry *entry =
219                         _mesa_hash_table_search(c->def_ht, reg);
220                 struct qreg *qregs = entry->data;
221 
222                 /* Insert a MOV if the source wasn't an SSA def in the
223                  * previous instruction.
224                  */
225                 if (result.file == QFILE_UNIF) {
226                         result = qir_MOV(c, result);
227                         last_inst = c->defs[result.index];
228                 }
229 
230                 /* We know they're both temps, so just rewrite index. */
231                 c->defs[last_inst->dst.index] = NULL;
232                 last_inst->dst.index = qregs[chan].index;
233 
234                 /* If we're in control flow, then make this update of the reg
235                  * conditional on the execution mask.
236                  */
237                 if (c->execute.file != QFILE_NULL) {
238                         last_inst->dst.index = qregs[chan].index;
239 
240                         /* Set the flags to the current exec mask.  To insert
241                          * the SF, we temporarily remove our SSA instruction.
242                          */
243                         list_del(&last_inst->link);
244                         qir_SF(c, c->execute);
245                         list_addtail(&last_inst->link,
246                                      &c->cur_block->instructions);
247 
248                         last_inst->cond = QPU_COND_ZS;
249                         last_inst->cond_is_exec_mask = true;
250                 }
251         }
252 }
253 
254 static struct qreg
ntq_get_src(struct vc4_compile * c,nir_src src,int i)255 ntq_get_src(struct vc4_compile *c, nir_src src, int i)
256 {
257         struct hash_entry *entry;
258 
259         nir_intrinsic_instr *load = nir_load_reg_for_def(src.ssa);
260         if (load == NULL) {
261                 entry = _mesa_hash_table_search(c->def_ht, src.ssa);
262                 assert(i < src.ssa->num_components);
263         } else {
264                 nir_def *reg = load->src[0].ssa;
265                 ASSERTED nir_intrinsic_instr *decl = nir_reg_get_decl(reg);
266                 assert(nir_intrinsic_base(load) == 0);
267                 assert(nir_intrinsic_num_array_elems(decl) == 0);
268                 entry = _mesa_hash_table_search(c->def_ht, reg);
269                 assert(i < nir_intrinsic_num_components(decl));
270         }
271 
272         struct qreg *qregs = entry->data;
273         return qregs[i];
274 }
275 
276 static struct qreg
ntq_get_alu_src(struct vc4_compile * c,nir_alu_instr * instr,unsigned src)277 ntq_get_alu_src(struct vc4_compile *c, nir_alu_instr *instr,
278                 unsigned src)
279 {
280         struct qreg r = ntq_get_src(c, instr->src[src].src,
281                                     instr->src[src].swizzle[0]);
282 
283         return r;
284 };
285 
286 static inline struct qreg
qir_SAT(struct vc4_compile * c,struct qreg val)287 qir_SAT(struct vc4_compile *c, struct qreg val)
288 {
289         return qir_FMAX(c,
290                         qir_FMIN(c, val, qir_uniform_f(c, 1.0)),
291                         qir_uniform_f(c, 0.0));
292 }
293 
294 static struct qreg
ntq_rcp(struct vc4_compile * c,struct qreg x)295 ntq_rcp(struct vc4_compile *c, struct qreg x)
296 {
297         struct qreg r = qir_RCP(c, x);
298 
299         /* Apply a Newton-Raphson step to improve the accuracy. */
300         r = qir_FMUL(c, r, qir_FSUB(c,
301                                     qir_uniform_f(c, 2.0),
302                                     qir_FMUL(c, x, r)));
303 
304         return r;
305 }
306 
307 static struct qreg
ntq_rsq(struct vc4_compile * c,struct qreg x)308 ntq_rsq(struct vc4_compile *c, struct qreg x)
309 {
310         struct qreg r = qir_RSQ(c, x);
311 
312         /* Apply a Newton-Raphson step to improve the accuracy. */
313         r = qir_FMUL(c, r, qir_FSUB(c,
314                                     qir_uniform_f(c, 1.5),
315                                     qir_FMUL(c,
316                                              qir_uniform_f(c, 0.5),
317                                              qir_FMUL(c, x,
318                                                       qir_FMUL(c, r, r)))));
319 
320         return r;
321 }
322 
323 static struct qreg
ntq_umul(struct vc4_compile * c,struct qreg src0,struct qreg src1)324 ntq_umul(struct vc4_compile *c, struct qreg src0, struct qreg src1)
325 {
326         struct qreg src0_hi = qir_SHR(c, src0,
327                                       qir_uniform_ui(c, 24));
328         struct qreg src1_hi = qir_SHR(c, src1,
329                                       qir_uniform_ui(c, 24));
330 
331         struct qreg hilo = qir_MUL24(c, src0_hi, src1);
332         struct qreg lohi = qir_MUL24(c, src0, src1_hi);
333         struct qreg lolo = qir_MUL24(c, src0, src1);
334 
335         return qir_ADD(c, lolo, qir_SHL(c,
336                                         qir_ADD(c, hilo, lohi),
337                                         qir_uniform_ui(c, 24)));
338 }
339 
340 static struct qreg
ntq_scale_depth_texture(struct vc4_compile * c,struct qreg src)341 ntq_scale_depth_texture(struct vc4_compile *c, struct qreg src)
342 {
343         struct qreg depthf = qir_ITOF(c, qir_SHR(c, src,
344                                                  qir_uniform_ui(c, 8)));
345         return qir_FMUL(c, depthf, qir_uniform_f(c, 1.0f/0xffffff));
346 }
347 
348 /**
349  * Emits a lowered TXF_MS from an MSAA texture.
350  *
351  * The addressing math has been lowered in NIR, and now we just need to read
352  * it like a UBO.
353  */
354 static void
ntq_emit_txf(struct vc4_compile * c,nir_tex_instr * instr)355 ntq_emit_txf(struct vc4_compile *c, nir_tex_instr *instr)
356 {
357         uint32_t tile_width = 32;
358         uint32_t tile_height = 32;
359         uint32_t tile_size = (tile_height * tile_width *
360                               VC4_MAX_SAMPLES * sizeof(uint32_t));
361 
362         unsigned unit = instr->texture_index;
363         uint32_t w = align(c->key->tex[unit].msaa_width, tile_width);
364         uint32_t w_tiles = w / tile_width;
365         uint32_t h = align(c->key->tex[unit].msaa_height, tile_height);
366         uint32_t h_tiles = h / tile_height;
367         uint32_t size = w_tiles * h_tiles * tile_size;
368 
369         struct qreg addr;
370         assert(instr->num_srcs == 1);
371         assert(instr->src[0].src_type == nir_tex_src_coord);
372         addr = ntq_get_src(c, instr->src[0].src, 0);
373 
374         /* Perform the clamping required by kernel validation. */
375         addr = qir_MAX(c, addr, qir_uniform_ui(c, 0));
376         addr = qir_MIN_NOIMM(c, addr, qir_uniform_ui(c, size - 4));
377 
378         qir_ADD_dest(c, qir_reg(QFILE_TEX_S_DIRECT, 0),
379                      addr, qir_uniform(c, QUNIFORM_TEXTURE_MSAA_ADDR, unit));
380 
381         ntq_emit_thrsw(c);
382 
383         struct qreg tex = qir_TEX_RESULT(c);
384         c->num_texture_samples++;
385 
386         enum pipe_format format = c->key->tex[unit].format;
387         if (util_format_is_depth_or_stencil(format)) {
388                 struct qreg scaled = ntq_scale_depth_texture(c, tex);
389                 for (int i = 0; i < 4; i++)
390                         ntq_store_def(c, &instr->def, i, qir_MOV(c, scaled));
391         } else {
392                 for (int i = 0; i < 4; i++)
393                         ntq_store_def(c, &instr->def, i,
394                                       qir_UNPACK_8_F(c, tex, i));
395         }
396 }
397 
398 static void
ntq_emit_tex(struct vc4_compile * c,nir_tex_instr * instr)399 ntq_emit_tex(struct vc4_compile *c, nir_tex_instr *instr)
400 {
401         struct qreg s, t, r, lod, compare;
402         bool is_txb = false, is_txl = false;
403         unsigned unit = instr->texture_index;
404 
405         if (instr->op == nir_texop_txf) {
406                 ntq_emit_txf(c, instr);
407                 return;
408         }
409 
410         for (unsigned i = 0; i < instr->num_srcs; i++) {
411                 switch (instr->src[i].src_type) {
412                 case nir_tex_src_coord:
413                         s = ntq_get_src(c, instr->src[i].src, 0);
414                         if (instr->sampler_dim == GLSL_SAMPLER_DIM_1D)
415                                 t = qir_uniform_f(c, 0.5);
416                         else
417                                 t = ntq_get_src(c, instr->src[i].src, 1);
418                         if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE)
419                                 r = ntq_get_src(c, instr->src[i].src, 2);
420                         break;
421                 case nir_tex_src_bias:
422                         lod = ntq_get_src(c, instr->src[i].src, 0);
423                         is_txb = true;
424                         break;
425                 case nir_tex_src_lod:
426                         lod = ntq_get_src(c, instr->src[i].src, 0);
427                         is_txl = true;
428                         break;
429                 case nir_tex_src_comparator:
430                         compare = ntq_get_src(c, instr->src[i].src, 0);
431                         break;
432                 default:
433                         unreachable("unknown texture source");
434                 }
435         }
436 
437         if (c->stage != QSTAGE_FRAG && !is_txl) {
438                 /* From the GLSL 1.20 spec:
439                  *
440                  *     "If it is mip-mapped and running on the vertex shader,
441                  *      then the base texture is used."
442                  */
443                 is_txl = true;
444                 lod = qir_uniform_ui(c, 0);
445         }
446 
447         if (c->key->tex[unit].force_first_level) {
448                 lod = qir_uniform(c, QUNIFORM_TEXTURE_FIRST_LEVEL, unit);
449                 is_txl = true;
450                 is_txb = false;
451         }
452 
453         struct qreg texture_u[] = {
454                 qir_uniform(c, QUNIFORM_TEXTURE_CONFIG_P0, unit),
455                 qir_uniform(c, QUNIFORM_TEXTURE_CONFIG_P1, unit),
456                 qir_uniform(c, QUNIFORM_CONSTANT, 0),
457                 qir_uniform(c, QUNIFORM_CONSTANT, 0),
458         };
459         uint32_t next_texture_u = 0;
460 
461         if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE || is_txl) {
462                 texture_u[2] = qir_uniform(c, QUNIFORM_TEXTURE_CONFIG_P2,
463                                            unit | (is_txl << 16));
464         }
465 
466         struct qinst *tmu;
467         if (instr->sampler_dim == GLSL_SAMPLER_DIM_CUBE) {
468                 tmu = qir_MOV_dest(c, qir_reg(QFILE_TEX_R, 0), r);
469                 tmu->src[qir_get_tex_uniform_src(tmu)] =
470                         texture_u[next_texture_u++];
471         } else if (c->key->tex[unit].wrap_s == PIPE_TEX_WRAP_CLAMP_TO_BORDER ||
472                    c->key->tex[unit].wrap_s == PIPE_TEX_WRAP_CLAMP ||
473                    c->key->tex[unit].wrap_t == PIPE_TEX_WRAP_CLAMP_TO_BORDER ||
474                    c->key->tex[unit].wrap_t == PIPE_TEX_WRAP_CLAMP) {
475                 tmu = qir_MOV_dest(c, qir_reg(QFILE_TEX_R, 0),
476                                    qir_uniform(c, QUNIFORM_TEXTURE_BORDER_COLOR,
477                                                unit));
478                 tmu->src[qir_get_tex_uniform_src(tmu)] =
479                         texture_u[next_texture_u++];
480         }
481 
482         if (c->key->tex[unit].wrap_s == PIPE_TEX_WRAP_CLAMP) {
483                 s = qir_SAT(c, s);
484         }
485 
486         if (c->key->tex[unit].wrap_t == PIPE_TEX_WRAP_CLAMP) {
487                 t = qir_SAT(c, t);
488         }
489 
490         tmu = qir_MOV_dest(c, qir_reg(QFILE_TEX_T, 0), t);
491         tmu->src[qir_get_tex_uniform_src(tmu)] =
492                 texture_u[next_texture_u++];
493 
494         if (is_txl || is_txb) {
495                 tmu = qir_MOV_dest(c, qir_reg(QFILE_TEX_B, 0), lod);
496                 tmu->src[qir_get_tex_uniform_src(tmu)] =
497                         texture_u[next_texture_u++];
498         }
499 
500         tmu = qir_MOV_dest(c, qir_reg(QFILE_TEX_S, 0), s);
501         tmu->src[qir_get_tex_uniform_src(tmu)] = texture_u[next_texture_u++];
502 
503         c->num_texture_samples++;
504 
505         ntq_emit_thrsw(c);
506 
507         struct qreg tex = qir_TEX_RESULT(c);
508 
509         enum pipe_format format = c->key->tex[unit].format;
510 
511         if (util_format_is_depth_or_stencil(format)) {
512                 struct qreg normalized = ntq_scale_depth_texture(c, tex);
513                 struct qreg depth_output;
514 
515                 struct qreg u0 = qir_uniform_f(c, 0.0f);
516                 struct qreg u1 = qir_uniform_f(c, 1.0f);
517                 if (c->key->tex[unit].compare_mode) {
518                         /* From the GL_ARB_shadow spec:
519                          *
520                          *     "Let Dt (D subscript t) be the depth texture
521                          *      value, in the range [0, 1].  Let R be the
522                          *      interpolated texture coordinate clamped to the
523                          *      range [0, 1]."
524                          */
525                         compare = qir_SAT(c, compare);
526 
527                         switch (c->key->tex[unit].compare_func) {
528                         case PIPE_FUNC_NEVER:
529                                 depth_output = qir_uniform_f(c, 0.0f);
530                                 break;
531                         case PIPE_FUNC_ALWAYS:
532                                 depth_output = u1;
533                                 break;
534                         case PIPE_FUNC_EQUAL:
535                                 qir_SF(c, qir_FSUB(c, compare, normalized));
536                                 depth_output = qir_SEL(c, QPU_COND_ZS, u1, u0);
537                                 break;
538                         case PIPE_FUNC_NOTEQUAL:
539                                 qir_SF(c, qir_FSUB(c, compare, normalized));
540                                 depth_output = qir_SEL(c, QPU_COND_ZC, u1, u0);
541                                 break;
542                         case PIPE_FUNC_GREATER:
543                                 qir_SF(c, qir_FSUB(c, compare, normalized));
544                                 depth_output = qir_SEL(c, QPU_COND_NC, u1, u0);
545                                 break;
546                         case PIPE_FUNC_GEQUAL:
547                                 qir_SF(c, qir_FSUB(c, normalized, compare));
548                                 depth_output = qir_SEL(c, QPU_COND_NS, u1, u0);
549                                 break;
550                         case PIPE_FUNC_LESS:
551                                 qir_SF(c, qir_FSUB(c, compare, normalized));
552                                 depth_output = qir_SEL(c, QPU_COND_NS, u1, u0);
553                                 break;
554                         case PIPE_FUNC_LEQUAL:
555                                 qir_SF(c, qir_FSUB(c, normalized, compare));
556                                 depth_output = qir_SEL(c, QPU_COND_NC, u1, u0);
557                                 break;
558                         }
559                 } else {
560                         depth_output = normalized;
561                 }
562 
563                 for (int i = 0; i < 4; i++)
564                         ntq_store_def(c, &instr->def, i,
565                                       qir_MOV(c, depth_output));
566         } else {
567                 for (int i = 0; i < 4; i++)
568                         ntq_store_def(c, &instr->def, i,
569                                       qir_UNPACK_8_F(c, tex, i));
570         }
571 }
572 
573 /**
574  * Computes x - floor(x), which is tricky because our FTOI truncates (rounds
575  * to zero).
576  */
577 static struct qreg
ntq_ffract(struct vc4_compile * c,struct qreg src)578 ntq_ffract(struct vc4_compile *c, struct qreg src)
579 {
580         struct qreg trunc = qir_ITOF(c, qir_FTOI(c, src));
581         struct qreg diff = qir_FSUB(c, src, trunc);
582         qir_SF(c, diff);
583 
584         qir_FADD_dest(c, diff,
585                       diff, qir_uniform_f(c, 1.0))->cond = QPU_COND_NS;
586 
587         return qir_MOV(c, diff);
588 }
589 
590 /**
591  * Computes floor(x), which is tricky because our FTOI truncates (rounds to
592  * zero).
593  */
594 static struct qreg
ntq_ffloor(struct vc4_compile * c,struct qreg src)595 ntq_ffloor(struct vc4_compile *c, struct qreg src)
596 {
597         struct qreg result = qir_ITOF(c, qir_FTOI(c, src));
598 
599         /* This will be < 0 if we truncated and the truncation was of a value
600          * that was < 0 in the first place.
601          */
602         qir_SF(c, qir_FSUB(c, src, result));
603 
604         struct qinst *sub = qir_FSUB_dest(c, result,
605                                           result, qir_uniform_f(c, 1.0));
606         sub->cond = QPU_COND_NS;
607 
608         return qir_MOV(c, result);
609 }
610 
611 /**
612  * Computes ceil(x), which is tricky because our FTOI truncates (rounds to
613  * zero).
614  */
615 static struct qreg
ntq_fceil(struct vc4_compile * c,struct qreg src)616 ntq_fceil(struct vc4_compile *c, struct qreg src)
617 {
618         struct qreg result = qir_ITOF(c, qir_FTOI(c, src));
619 
620         /* This will be < 0 if we truncated and the truncation was of a value
621          * that was > 0 in the first place.
622          */
623         qir_SF(c, qir_FSUB(c, result, src));
624 
625         qir_FADD_dest(c, result,
626                       result, qir_uniform_f(c, 1.0))->cond = QPU_COND_NS;
627 
628         return qir_MOV(c, result);
629 }
630 
631 static struct qreg
ntq_shrink_sincos_input_range(struct vc4_compile * c,struct qreg x)632 ntq_shrink_sincos_input_range(struct vc4_compile *c, struct qreg x)
633 {
634         /* Since we're using a Taylor approximation, we want to have a small
635          * number of coefficients and take advantage of sin/cos repeating
636          * every 2pi.  We keep our x as close to 0 as we can, since the series
637          * will be less accurate as |x| increases.  (Also, be careful of
638          * shifting the input x value to be tricky with sin/cos relations,
639          * because getting accurate values for x==0 is very important for SDL
640          * rendering)
641          */
642         struct qreg scaled_x =
643                 qir_FMUL(c, x,
644                          qir_uniform_f(c, 1.0f / (M_PI * 2.0f)));
645         /* Note: FTOI truncates toward 0. */
646         struct qreg x_frac = qir_FSUB(c, scaled_x,
647                                       qir_ITOF(c, qir_FTOI(c, scaled_x)));
648         /* Map [0.5, 1] to [-0.5, 0] */
649         qir_SF(c, qir_FSUB(c, x_frac, qir_uniform_f(c, 0.5)));
650         qir_FSUB_dest(c, x_frac, x_frac, qir_uniform_f(c, 1.0))->cond = QPU_COND_NC;
651         /* Map [-1, -0.5] to [0, 0.5] */
652         qir_SF(c, qir_FADD(c, x_frac, qir_uniform_f(c, 0.5)));
653         qir_FADD_dest(c, x_frac, x_frac, qir_uniform_f(c, 1.0))->cond = QPU_COND_NS;
654 
655         return x_frac;
656 }
657 
658 static struct qreg
ntq_fsin(struct vc4_compile * c,struct qreg src)659 ntq_fsin(struct vc4_compile *c, struct qreg src)
660 {
661         float coeff[] = {
662                 2.0 * M_PI,
663                 -pow(2.0 * M_PI, 3) / (3 * 2 * 1),
664                 pow(2.0 * M_PI, 5) / (5 * 4 * 3 * 2 * 1),
665                 -pow(2.0 * M_PI, 7) / (7 * 6 * 5 * 4 * 3 * 2 * 1),
666                 pow(2.0 * M_PI, 9) / (9 * 8 * 7 * 6 * 5 * 4 * 3 * 2 * 1),
667         };
668 
669         struct qreg x = ntq_shrink_sincos_input_range(c, src);
670         struct qreg x2 = qir_FMUL(c, x, x);
671         struct qreg sum = qir_FMUL(c, x, qir_uniform_f(c, coeff[0]));
672         for (int i = 1; i < ARRAY_SIZE(coeff); i++) {
673                 x = qir_FMUL(c, x, x2);
674                 sum = qir_FADD(c,
675                                sum,
676                                qir_FMUL(c,
677                                         x,
678                                         qir_uniform_f(c, coeff[i])));
679         }
680         return sum;
681 }
682 
683 static struct qreg
ntq_fcos(struct vc4_compile * c,struct qreg src)684 ntq_fcos(struct vc4_compile *c, struct qreg src)
685 {
686         float coeff[] = {
687                 1.0f,
688                 -pow(2.0 * M_PI, 2) / (2 * 1),
689                 pow(2.0 * M_PI, 4) / (4 * 3 * 2 * 1),
690                 -pow(2.0 * M_PI, 6) / (6 * 5 * 4 * 3 * 2 * 1),
691                 pow(2.0 * M_PI, 8) / (8 * 7 * 6 * 5 * 4 * 3 * 2 * 1),
692                 -pow(2.0 * M_PI, 10) / (10 * 9 * 8 * 7 * 6 * 5 * 4 * 3 * 2 * 1),
693         };
694 
695         struct qreg x_frac = ntq_shrink_sincos_input_range(c, src);
696         struct qreg sum = qir_uniform_f(c, coeff[0]);
697         struct qreg x2 = qir_FMUL(c, x_frac, x_frac);
698         struct qreg x = x2; /* Current x^2, x^4, or x^6 */
699         for (int i = 1; i < ARRAY_SIZE(coeff); i++) {
700                 if (i != 1)
701                         x = qir_FMUL(c, x, x2);
702 
703                 sum = qir_FADD(c, qir_FMUL(c,
704                                            x,
705                                            qir_uniform_f(c, coeff[i])),
706                                sum);
707         }
708         return sum;
709 }
710 
711 static struct qreg
ntq_fsign(struct vc4_compile * c,struct qreg src)712 ntq_fsign(struct vc4_compile *c, struct qreg src)
713 {
714         struct qreg t = qir_get_temp(c);
715 
716         qir_SF(c, src);
717         qir_MOV_dest(c, t, qir_uniform_f(c, 0.0));
718         qir_MOV_dest(c, t, qir_uniform_f(c, 1.0))->cond = QPU_COND_ZC;
719         qir_MOV_dest(c, t, qir_uniform_f(c, -1.0))->cond = QPU_COND_NS;
720         return qir_MOV(c, t);
721 }
722 
723 static void
emit_vertex_input(struct vc4_compile * c,int attr)724 emit_vertex_input(struct vc4_compile *c, int attr)
725 {
726         enum pipe_format format = c->vs_key->attr_formats[attr];
727         uint32_t attr_size = util_format_get_blocksize(format);
728 
729         c->vattr_sizes[attr] = align(attr_size, 4);
730         for (int i = 0; i < align(attr_size, 4) / 4; i++) {
731                 c->inputs[attr * 4 + i] =
732                         qir_MOV(c, qir_reg(QFILE_VPM, attr * 4 + i));
733                 c->num_inputs++;
734         }
735 }
736 
737 static void
emit_fragcoord_input(struct vc4_compile * c,int attr)738 emit_fragcoord_input(struct vc4_compile *c, int attr)
739 {
740         c->inputs[attr * 4 + 0] = qir_ITOF(c, qir_reg(QFILE_FRAG_X, 0));
741         c->inputs[attr * 4 + 1] = qir_ITOF(c, qir_reg(QFILE_FRAG_Y, 0));
742         c->inputs[attr * 4 + 2] =
743                 qir_FMUL(c,
744                          qir_ITOF(c, qir_FRAG_Z(c)),
745                          qir_uniform_f(c, 1.0 / 0xffffff));
746         c->inputs[attr * 4 + 3] = qir_RCP(c, qir_FRAG_W(c));
747 }
748 
749 static struct qreg
emit_fragment_varying(struct vc4_compile * c,gl_varying_slot slot,uint8_t swizzle)750 emit_fragment_varying(struct vc4_compile *c, gl_varying_slot slot,
751                       uint8_t swizzle)
752 {
753         uint32_t i = c->num_input_slots++;
754         struct qreg vary = {
755                 QFILE_VARY,
756                 i
757         };
758 
759         if (c->num_input_slots >= c->input_slots_array_size) {
760                 c->input_slots_array_size =
761                         MAX2(4, c->input_slots_array_size * 2);
762 
763                 c->input_slots = reralloc(c, c->input_slots,
764                                           struct vc4_varying_slot,
765                                           c->input_slots_array_size);
766         }
767 
768         c->input_slots[i].slot = slot;
769         c->input_slots[i].swizzle = swizzle;
770 
771         return qir_VARY_ADD_C(c, qir_FMUL(c, vary, qir_FRAG_W(c)));
772 }
773 
774 static void
emit_fragment_input(struct vc4_compile * c,int attr,gl_varying_slot slot)775 emit_fragment_input(struct vc4_compile *c, int attr, gl_varying_slot slot)
776 {
777         for (int i = 0; i < 4; i++) {
778                 c->inputs[attr * 4 + i] =
779                         emit_fragment_varying(c, slot, i);
780                 c->num_inputs++;
781         }
782 }
783 
784 static void
add_output(struct vc4_compile * c,uint32_t decl_offset,uint8_t slot,uint8_t swizzle)785 add_output(struct vc4_compile *c,
786            uint32_t decl_offset,
787            uint8_t slot,
788            uint8_t swizzle)
789 {
790         uint32_t old_array_size = c->outputs_array_size;
791         resize_qreg_array(c, &c->outputs, &c->outputs_array_size,
792                           decl_offset + 1);
793 
794         if (old_array_size != c->outputs_array_size) {
795                 c->output_slots = reralloc(c,
796                                            c->output_slots,
797                                            struct vc4_varying_slot,
798                                            c->outputs_array_size);
799         }
800 
801         c->output_slots[decl_offset].slot = slot;
802         c->output_slots[decl_offset].swizzle = swizzle;
803 }
804 
805 static bool
ntq_src_is_only_ssa_def_user(nir_src * src)806 ntq_src_is_only_ssa_def_user(nir_src *src)
807 {
808         return list_is_singular(&src->ssa->uses) &&
809                nir_load_reg_for_def(src->ssa) == NULL;
810 }
811 
812 /**
813  * In general, emits a nir_pack_unorm_4x8 as a series of MOVs with the pack
814  * bit set.
815  *
816  * However, as an optimization, it tries to find the instructions generating
817  * the sources to be packed and just emit the pack flag there, if possible.
818  */
819 static void
ntq_emit_pack_unorm_4x8(struct vc4_compile * c,nir_alu_instr * instr)820 ntq_emit_pack_unorm_4x8(struct vc4_compile *c, nir_alu_instr *instr)
821 {
822         struct qreg result = qir_get_temp(c);
823         struct nir_alu_instr *vec4 = NULL;
824 
825         /* If packing from a vec4 op (as expected), identify it so that we can
826          * peek back at what generated its sources.
827          */
828         if (instr->src[0].src.ssa->parent_instr->type == nir_instr_type_alu &&
829             nir_instr_as_alu(instr->src[0].src.ssa->parent_instr)->op ==
830             nir_op_vec4) {
831                 vec4 = nir_instr_as_alu(instr->src[0].src.ssa->parent_instr);
832         }
833 
834         /* If the pack is replicating the same channel 4 times, use the 8888
835          * pack flag.  This is common for blending using the alpha
836          * channel.
837          */
838         if (instr->src[0].swizzle[0] == instr->src[0].swizzle[1] &&
839             instr->src[0].swizzle[0] == instr->src[0].swizzle[2] &&
840             instr->src[0].swizzle[0] == instr->src[0].swizzle[3]) {
841                 struct qreg rep = ntq_get_src(c,
842                                               instr->src[0].src,
843                                               instr->src[0].swizzle[0]);
844                 ntq_store_def(c, &instr->def, 0, qir_PACK_8888_F(c, rep));
845                 return;
846         }
847 
848         for (int i = 0; i < 4; i++) {
849                 int swiz = instr->src[0].swizzle[i];
850                 struct qreg src;
851                 if (vec4) {
852                         src = ntq_get_src(c, vec4->src[swiz].src,
853                                           vec4->src[swiz].swizzle[0]);
854                 } else {
855                         src = ntq_get_src(c, instr->src[0].src, swiz);
856                 }
857 
858                 if (vec4 &&
859                     ntq_src_is_only_ssa_def_user(&vec4->src[swiz].src) &&
860                     src.file == QFILE_TEMP &&
861                     c->defs[src.index] &&
862                     qir_is_mul(c->defs[src.index]) &&
863                     !c->defs[src.index]->dst.pack) {
864                         struct qinst *rewrite = c->defs[src.index];
865                         c->defs[src.index] = NULL;
866                         rewrite->dst = result;
867                         rewrite->dst.pack = QPU_PACK_MUL_8A + i;
868                         continue;
869                 }
870 
871                 qir_PACK_8_F(c, result, src, i);
872         }
873 
874         ntq_store_def(c, &instr->def, 0, qir_MOV(c, result));
875 }
876 
877 /** Handles sign-extended bitfield extracts for 16 bits. */
878 static struct qreg
ntq_emit_ibfe(struct vc4_compile * c,struct qreg base,struct qreg offset,struct qreg bits)879 ntq_emit_ibfe(struct vc4_compile *c, struct qreg base, struct qreg offset,
880               struct qreg bits)
881 {
882         assert(bits.file == QFILE_UNIF &&
883                c->uniform_contents[bits.index] == QUNIFORM_CONSTANT &&
884                c->uniform_data[bits.index] == 16);
885 
886         assert(offset.file == QFILE_UNIF &&
887                c->uniform_contents[offset.index] == QUNIFORM_CONSTANT);
888         int offset_bit = c->uniform_data[offset.index];
889         assert(offset_bit % 16 == 0);
890 
891         return qir_UNPACK_16_I(c, base, offset_bit / 16);
892 }
893 
894 /** Handles unsigned bitfield extracts for 8 bits. */
895 static struct qreg
ntq_emit_ubfe(struct vc4_compile * c,struct qreg base,struct qreg offset,struct qreg bits)896 ntq_emit_ubfe(struct vc4_compile *c, struct qreg base, struct qreg offset,
897               struct qreg bits)
898 {
899         assert(bits.file == QFILE_UNIF &&
900                c->uniform_contents[bits.index] == QUNIFORM_CONSTANT &&
901                c->uniform_data[bits.index] == 8);
902 
903         assert(offset.file == QFILE_UNIF &&
904                c->uniform_contents[offset.index] == QUNIFORM_CONSTANT);
905         int offset_bit = c->uniform_data[offset.index];
906         assert(offset_bit % 8 == 0);
907 
908         return qir_UNPACK_8_I(c, base, offset_bit / 8);
909 }
910 
911 /**
912  * If compare_instr is a valid comparison instruction, emits the
913  * compare_instr's comparison and returns the sel_instr's return value based
914  * on the compare_instr's result.
915  */
916 static bool
ntq_emit_comparison(struct vc4_compile * c,struct qreg * dest,nir_alu_instr * compare_instr,nir_alu_instr * sel_instr)917 ntq_emit_comparison(struct vc4_compile *c, struct qreg *dest,
918                     nir_alu_instr *compare_instr,
919                     nir_alu_instr *sel_instr)
920 {
921         enum qpu_cond cond;
922 
923         switch (compare_instr->op) {
924         case nir_op_feq32:
925         case nir_op_ieq32:
926         case nir_op_seq:
927                 cond = QPU_COND_ZS;
928                 break;
929         case nir_op_fneu32:
930         case nir_op_ine32:
931         case nir_op_sne:
932                 cond = QPU_COND_ZC;
933                 break;
934         case nir_op_fge32:
935         case nir_op_ige32:
936         case nir_op_uge32:
937         case nir_op_sge:
938                 cond = QPU_COND_NC;
939                 break;
940         case nir_op_flt32:
941         case nir_op_ilt32:
942         case nir_op_slt:
943                 cond = QPU_COND_NS;
944                 break;
945         default:
946                 return false;
947         }
948 
949         struct qreg src0 = ntq_get_alu_src(c, compare_instr, 0);
950         struct qreg src1 = ntq_get_alu_src(c, compare_instr, 1);
951 
952         unsigned unsized_type =
953                 nir_alu_type_get_base_type(nir_op_infos[compare_instr->op].input_types[0]);
954         if (unsized_type == nir_type_float)
955                 qir_SF(c, qir_FSUB(c, src0, src1));
956         else
957                 qir_SF(c, qir_SUB(c, src0, src1));
958 
959         switch (sel_instr->op) {
960         case nir_op_seq:
961         case nir_op_sne:
962         case nir_op_sge:
963         case nir_op_slt:
964                 *dest = qir_SEL(c, cond,
965                                 qir_uniform_f(c, 1.0), qir_uniform_f(c, 0.0));
966                 break;
967 
968         case nir_op_b32csel:
969                 *dest = qir_SEL(c, cond,
970                                 ntq_get_alu_src(c, sel_instr, 1),
971                                 ntq_get_alu_src(c, sel_instr, 2));
972                 break;
973 
974         default:
975                 *dest = qir_SEL(c, cond,
976                                 qir_uniform_ui(c, ~0), qir_uniform_ui(c, 0));
977                 break;
978         }
979 
980         /* Make the temporary for nir_store_def(). */
981         *dest = qir_MOV(c, *dest);
982 
983         return true;
984 }
985 
986 /**
987  * Attempts to fold a comparison generating a boolean result into the
988  * condition code for selecting between two values, instead of comparing the
989  * boolean result against 0 to generate the condition code.
990  */
ntq_emit_bcsel(struct vc4_compile * c,nir_alu_instr * instr,struct qreg * src)991 static struct qreg ntq_emit_bcsel(struct vc4_compile *c, nir_alu_instr *instr,
992                                   struct qreg *src)
993 {
994         if (nir_load_reg_for_def(instr->src[0].src.ssa))
995                 goto out;
996         if (instr->src[0].src.ssa->parent_instr->type != nir_instr_type_alu)
997                 goto out;
998         nir_alu_instr *compare =
999                 nir_instr_as_alu(instr->src[0].src.ssa->parent_instr);
1000         if (!compare)
1001                 goto out;
1002 
1003         struct qreg dest;
1004         if (ntq_emit_comparison(c, &dest, compare, instr))
1005                 return dest;
1006 
1007 out:
1008         qir_SF(c, src[0]);
1009         return qir_MOV(c, qir_SEL(c, QPU_COND_NS, src[1], src[2]));
1010 }
1011 
1012 static struct qreg
ntq_fddx(struct vc4_compile * c,struct qreg src)1013 ntq_fddx(struct vc4_compile *c, struct qreg src)
1014 {
1015         /* Make sure that we have a bare temp to use for MUL rotation, so it
1016          * can be allocated to an accumulator.
1017          */
1018         if (src.pack || src.file != QFILE_TEMP)
1019                 src = qir_MOV(c, src);
1020 
1021         struct qreg from_left = qir_ROT_MUL(c, src, 1);
1022         struct qreg from_right = qir_ROT_MUL(c, src, 15);
1023 
1024         /* Distinguish left/right pixels of the quad. */
1025         qir_SF(c, qir_AND(c, qir_reg(QFILE_QPU_ELEMENT, 0),
1026                           qir_uniform_ui(c, 1)));
1027 
1028         return qir_MOV(c, qir_SEL(c, QPU_COND_ZS,
1029                                   qir_FSUB(c, from_right, src),
1030                                   qir_FSUB(c, src, from_left)));
1031 }
1032 
1033 static struct qreg
ntq_fddy(struct vc4_compile * c,struct qreg src)1034 ntq_fddy(struct vc4_compile *c, struct qreg src)
1035 {
1036         if (src.pack || src.file != QFILE_TEMP)
1037                 src = qir_MOV(c, src);
1038 
1039         struct qreg from_bottom = qir_ROT_MUL(c, src, 2);
1040         struct qreg from_top = qir_ROT_MUL(c, src, 14);
1041 
1042         /* Distinguish top/bottom pixels of the quad. */
1043         qir_SF(c, qir_AND(c,
1044                           qir_reg(QFILE_QPU_ELEMENT, 0),
1045                           qir_uniform_ui(c, 2)));
1046 
1047         return qir_MOV(c, qir_SEL(c, QPU_COND_ZS,
1048                                   qir_FSUB(c, from_top, src),
1049                                   qir_FSUB(c, src, from_bottom)));
1050 }
1051 
1052 static struct qreg
ntq_emit_cond_to_int(struct vc4_compile * c,enum qpu_cond cond)1053 ntq_emit_cond_to_int(struct vc4_compile *c, enum qpu_cond cond)
1054 {
1055         return qir_MOV(c, qir_SEL(c, cond,
1056                                   qir_uniform_ui(c, 1),
1057                                   qir_uniform_ui(c, 0)));
1058 }
1059 
1060 static void
ntq_emit_alu(struct vc4_compile * c,nir_alu_instr * instr)1061 ntq_emit_alu(struct vc4_compile *c, nir_alu_instr *instr)
1062 {
1063         /* Vectors are special in that they have non-scalarized writemasks,
1064          * and just take the first swizzle channel for each argument in order
1065          * into each writemask channel.
1066          */
1067         if (instr->op == nir_op_vec2 ||
1068             instr->op == nir_op_vec3 ||
1069             instr->op == nir_op_vec4) {
1070                 struct qreg srcs[4];
1071                 for (int i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
1072                         srcs[i] = ntq_get_src(c, instr->src[i].src,
1073                                               instr->src[i].swizzle[0]);
1074                 for (int i = 0; i < nir_op_infos[instr->op].num_inputs; i++)
1075                         ntq_store_def(c, &instr->def, i,
1076                                       qir_MOV(c, srcs[i]));
1077                 return;
1078         }
1079 
1080         if (instr->op == nir_op_pack_unorm_4x8) {
1081                 ntq_emit_pack_unorm_4x8(c, instr);
1082                 return;
1083         }
1084 
1085         if (instr->op == nir_op_unpack_unorm_4x8) {
1086                 struct qreg src = ntq_get_src(c, instr->src[0].src,
1087                                               instr->src[0].swizzle[0]);
1088                 unsigned count = instr->def.num_components;
1089                 for (int i = 0; i < count; i++) {
1090                         ntq_store_def(c, &instr->def, i,
1091                                       qir_UNPACK_8_F(c, src, i));
1092                 }
1093                 return;
1094         }
1095 
1096         /* General case: We can just grab the one used channel per src. */
1097         struct qreg src[nir_op_infos[instr->op].num_inputs];
1098         for (int i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
1099                 src[i] = ntq_get_alu_src(c, instr, i);
1100         }
1101 
1102         struct qreg result;
1103 
1104         switch (instr->op) {
1105         case nir_op_mov:
1106                 result = qir_MOV(c, src[0]);
1107                 break;
1108         case nir_op_fmul:
1109                 result = qir_FMUL(c, src[0], src[1]);
1110                 break;
1111         case nir_op_fadd:
1112                 result = qir_FADD(c, src[0], src[1]);
1113                 break;
1114         case nir_op_fsub:
1115                 result = qir_FSUB(c, src[0], src[1]);
1116                 break;
1117         case nir_op_fmin:
1118                 result = qir_FMIN(c, src[0], src[1]);
1119                 break;
1120         case nir_op_fmax:
1121                 result = qir_FMAX(c, src[0], src[1]);
1122                 break;
1123 
1124         case nir_op_f2i32:
1125         case nir_op_f2u32:
1126                 result = qir_FTOI(c, src[0]);
1127                 break;
1128         case nir_op_i2f32:
1129         case nir_op_u2f32:
1130                 result = qir_ITOF(c, src[0]);
1131                 break;
1132         case nir_op_b2f32:
1133                 result = qir_AND(c, src[0], qir_uniform_f(c, 1.0));
1134                 break;
1135         case nir_op_b2i32:
1136                 result = qir_AND(c, src[0], qir_uniform_ui(c, 1));
1137                 break;
1138 
1139         case nir_op_iadd:
1140                 result = qir_ADD(c, src[0], src[1]);
1141                 break;
1142         case nir_op_ushr:
1143                 result = qir_SHR(c, src[0], src[1]);
1144                 break;
1145         case nir_op_isub:
1146                 result = qir_SUB(c, src[0], src[1]);
1147                 break;
1148         case nir_op_ishr:
1149                 result = qir_ASR(c, src[0], src[1]);
1150                 break;
1151         case nir_op_ishl:
1152                 result = qir_SHL(c, src[0], src[1]);
1153                 break;
1154         case nir_op_imin:
1155                 result = qir_MIN(c, src[0], src[1]);
1156                 break;
1157         case nir_op_imax:
1158                 result = qir_MAX(c, src[0], src[1]);
1159                 break;
1160         case nir_op_iand:
1161                 result = qir_AND(c, src[0], src[1]);
1162                 break;
1163         case nir_op_ior:
1164                 result = qir_OR(c, src[0], src[1]);
1165                 break;
1166         case nir_op_ixor:
1167                 result = qir_XOR(c, src[0], src[1]);
1168                 break;
1169         case nir_op_inot:
1170                 result = qir_NOT(c, src[0]);
1171                 break;
1172 
1173         case nir_op_imul:
1174                 result = ntq_umul(c, src[0], src[1]);
1175                 break;
1176 
1177         case nir_op_seq:
1178         case nir_op_sne:
1179         case nir_op_sge:
1180         case nir_op_slt:
1181         case nir_op_feq32:
1182         case nir_op_fneu32:
1183         case nir_op_fge32:
1184         case nir_op_flt32:
1185         case nir_op_ieq32:
1186         case nir_op_ine32:
1187         case nir_op_ige32:
1188         case nir_op_uge32:
1189         case nir_op_ilt32:
1190                 if (!ntq_emit_comparison(c, &result, instr, instr)) {
1191                         fprintf(stderr, "Bad comparison instruction\n");
1192                 }
1193                 break;
1194 
1195         case nir_op_b32csel:
1196                 result = ntq_emit_bcsel(c, instr, src);
1197                 break;
1198         case nir_op_fcsel:
1199                 qir_SF(c, src[0]);
1200                 result = qir_MOV(c, qir_SEL(c, QPU_COND_ZC, src[1], src[2]));
1201                 break;
1202 
1203         case nir_op_frcp:
1204                 result = ntq_rcp(c, src[0]);
1205                 break;
1206         case nir_op_frsq:
1207                 result = ntq_rsq(c, src[0]);
1208                 break;
1209         case nir_op_fexp2:
1210                 result = qir_EXP2(c, src[0]);
1211                 break;
1212         case nir_op_flog2:
1213                 result = qir_LOG2(c, src[0]);
1214                 break;
1215 
1216         case nir_op_ftrunc:
1217                 result = qir_ITOF(c, qir_FTOI(c, src[0]));
1218                 break;
1219         case nir_op_fceil:
1220                 result = ntq_fceil(c, src[0]);
1221                 break;
1222         case nir_op_ffract:
1223                 result = ntq_ffract(c, src[0]);
1224                 break;
1225         case nir_op_ffloor:
1226                 result = ntq_ffloor(c, src[0]);
1227                 break;
1228 
1229         case nir_op_fsin:
1230                 result = ntq_fsin(c, src[0]);
1231                 break;
1232         case nir_op_fcos:
1233                 result = ntq_fcos(c, src[0]);
1234                 break;
1235 
1236         case nir_op_fsign:
1237                 result = ntq_fsign(c, src[0]);
1238                 break;
1239 
1240         case nir_op_fabs:
1241                 result = qir_FMAXABS(c, src[0], src[0]);
1242                 break;
1243         case nir_op_iabs:
1244                 result = qir_MAX(c, src[0],
1245                                 qir_SUB(c, qir_uniform_ui(c, 0), src[0]));
1246                 break;
1247 
1248         case nir_op_ibitfield_extract:
1249                 result = ntq_emit_ibfe(c, src[0], src[1], src[2]);
1250                 break;
1251 
1252         case nir_op_ubitfield_extract:
1253                 result = ntq_emit_ubfe(c, src[0], src[1], src[2]);
1254                 break;
1255 
1256         case nir_op_usadd_4x8_vc4:
1257                 result = qir_V8ADDS(c, src[0], src[1]);
1258                 break;
1259 
1260         case nir_op_ussub_4x8_vc4:
1261                 result = qir_V8SUBS(c, src[0], src[1]);
1262                 break;
1263 
1264         case nir_op_umin_4x8_vc4:
1265                 result = qir_V8MIN(c, src[0], src[1]);
1266                 break;
1267 
1268         case nir_op_umax_4x8_vc4:
1269                 result = qir_V8MAX(c, src[0], src[1]);
1270                 break;
1271 
1272         case nir_op_umul_unorm_4x8_vc4:
1273                 result = qir_V8MULD(c, src[0], src[1]);
1274                 break;
1275 
1276         case nir_op_fddx:
1277         case nir_op_fddx_coarse:
1278         case nir_op_fddx_fine:
1279                 result = ntq_fddx(c, src[0]);
1280                 break;
1281 
1282         case nir_op_fddy:
1283         case nir_op_fddy_coarse:
1284         case nir_op_fddy_fine:
1285                 result = ntq_fddy(c, src[0]);
1286                 break;
1287 
1288         case nir_op_uadd_carry:
1289                 qir_SF(c, qir_ADD(c, src[0], src[1]));
1290                 result = ntq_emit_cond_to_int(c, QPU_COND_CS);
1291                 break;
1292 
1293         case nir_op_usub_borrow:
1294                 qir_SF(c, qir_SUB(c, src[0], src[1]));
1295                 result = ntq_emit_cond_to_int(c, QPU_COND_CS);
1296                 break;
1297 
1298         default:
1299                 fprintf(stderr, "unknown NIR ALU inst: ");
1300                 nir_print_instr(&instr->instr, stderr);
1301                 fprintf(stderr, "\n");
1302                 abort();
1303         }
1304 
1305         ntq_store_def(c, &instr->def, 0, result);
1306 }
1307 
1308 static void
emit_frag_end(struct vc4_compile * c)1309 emit_frag_end(struct vc4_compile *c)
1310 {
1311         struct qreg color;
1312         if (c->output_color_index != -1) {
1313                 color = c->outputs[c->output_color_index];
1314         } else {
1315                 color = qir_uniform_ui(c, 0);
1316         }
1317 
1318         uint32_t discard_cond = QPU_COND_ALWAYS;
1319         if (c->s->info.fs.uses_discard) {
1320                 qir_SF(c, c->discard);
1321                 discard_cond = QPU_COND_ZS;
1322         }
1323 
1324         if (c->fs_key->stencil_enabled) {
1325                 qir_MOV_dest(c, qir_reg(QFILE_TLB_STENCIL_SETUP, 0),
1326                              qir_uniform(c, QUNIFORM_STENCIL, 0));
1327                 if (c->fs_key->stencil_twoside) {
1328                         qir_MOV_dest(c, qir_reg(QFILE_TLB_STENCIL_SETUP, 0),
1329                                      qir_uniform(c, QUNIFORM_STENCIL, 1));
1330                 }
1331                 if (c->fs_key->stencil_full_writemasks) {
1332                         qir_MOV_dest(c, qir_reg(QFILE_TLB_STENCIL_SETUP, 0),
1333                                      qir_uniform(c, QUNIFORM_STENCIL, 2));
1334                 }
1335         }
1336 
1337         if (c->output_sample_mask_index != -1) {
1338                 qir_MS_MASK(c, c->outputs[c->output_sample_mask_index]);
1339         }
1340 
1341         if (c->fs_key->depth_enabled) {
1342                 if (c->output_position_index != -1) {
1343                         qir_FTOI_dest(c, qir_reg(QFILE_TLB_Z_WRITE, 0),
1344                                       qir_FMUL(c,
1345                                                c->outputs[c->output_position_index],
1346                                                qir_uniform_f(c, 0xffffff)))->cond = discard_cond;
1347                 } else {
1348                         qir_MOV_dest(c, qir_reg(QFILE_TLB_Z_WRITE, 0),
1349                                      qir_FRAG_Z(c))->cond = discard_cond;
1350                 }
1351         }
1352 
1353         if (!c->msaa_per_sample_output) {
1354                 qir_MOV_dest(c, qir_reg(QFILE_TLB_COLOR_WRITE, 0),
1355                              color)->cond = discard_cond;
1356         } else {
1357                 for (int i = 0; i < VC4_MAX_SAMPLES; i++) {
1358                         qir_MOV_dest(c, qir_reg(QFILE_TLB_COLOR_WRITE_MS, 0),
1359                                      c->sample_colors[i])->cond = discard_cond;
1360                 }
1361         }
1362 }
1363 
1364 static void
emit_scaled_viewport_write(struct vc4_compile * c,struct qreg rcp_w)1365 emit_scaled_viewport_write(struct vc4_compile *c, struct qreg rcp_w)
1366 {
1367         struct qreg packed = qir_get_temp(c);
1368 
1369         for (int i = 0; i < 2; i++) {
1370                 struct qreg scale =
1371                         qir_uniform(c, QUNIFORM_VIEWPORT_X_SCALE + i, 0);
1372 
1373                 struct qreg packed_chan = packed;
1374                 packed_chan.pack = QPU_PACK_A_16A + i;
1375 
1376                 qir_FTOI_dest(c, packed_chan,
1377                               qir_FMUL(c,
1378                                        qir_FMUL(c,
1379                                                 c->outputs[c->output_position_index + i],
1380                                                 scale),
1381                                        rcp_w));
1382         }
1383 
1384         qir_VPM_WRITE(c, packed);
1385 }
1386 
1387 static void
emit_zs_write(struct vc4_compile * c,struct qreg rcp_w)1388 emit_zs_write(struct vc4_compile *c, struct qreg rcp_w)
1389 {
1390         struct qreg zscale = qir_uniform(c, QUNIFORM_VIEWPORT_Z_SCALE, 0);
1391         struct qreg zoffset = qir_uniform(c, QUNIFORM_VIEWPORT_Z_OFFSET, 0);
1392 
1393         qir_VPM_WRITE(c, qir_FADD(c, qir_FMUL(c, qir_FMUL(c,
1394                                                           c->outputs[c->output_position_index + 2],
1395                                                           zscale),
1396                                               rcp_w),
1397                                   zoffset));
1398 }
1399 
1400 static void
emit_rcp_wc_write(struct vc4_compile * c,struct qreg rcp_w)1401 emit_rcp_wc_write(struct vc4_compile *c, struct qreg rcp_w)
1402 {
1403         qir_VPM_WRITE(c, rcp_w);
1404 }
1405 
1406 static void
emit_point_size_write(struct vc4_compile * c)1407 emit_point_size_write(struct vc4_compile *c)
1408 {
1409         struct qreg point_size;
1410 
1411         if (c->output_point_size_index != -1)
1412                 point_size = c->outputs[c->output_point_size_index];
1413         else
1414                 point_size = qir_uniform_f(c, 1.0);
1415 
1416         qir_VPM_WRITE(c, point_size);
1417 }
1418 
1419 /**
1420  * Emits a VPM read of the stub vertex attribute set up by vc4_draw.c.
1421  *
1422  * The simulator insists that there be at least one vertex attribute, so
1423  * vc4_draw.c will emit one if it wouldn't have otherwise.  The simulator also
1424  * insists that all vertex attributes loaded get read by the VS/CS, so we have
1425  * to consume it here.
1426  */
1427 static void
emit_stub_vpm_read(struct vc4_compile * c)1428 emit_stub_vpm_read(struct vc4_compile *c)
1429 {
1430         if (c->num_inputs)
1431                 return;
1432 
1433         c->vattr_sizes[0] = 4;
1434         (void)qir_MOV(c, qir_reg(QFILE_VPM, 0));
1435         c->num_inputs++;
1436 }
1437 
1438 static void
emit_vert_end(struct vc4_compile * c,struct vc4_varying_slot * fs_inputs,uint32_t num_fs_inputs)1439 emit_vert_end(struct vc4_compile *c,
1440               struct vc4_varying_slot *fs_inputs,
1441               uint32_t num_fs_inputs)
1442 {
1443         struct qreg rcp_w = ntq_rcp(c, c->outputs[c->output_position_index + 3]);
1444 
1445         emit_stub_vpm_read(c);
1446 
1447         emit_scaled_viewport_write(c, rcp_w);
1448         emit_zs_write(c, rcp_w);
1449         emit_rcp_wc_write(c, rcp_w);
1450         if (c->vs_key->per_vertex_point_size)
1451                 emit_point_size_write(c);
1452 
1453         for (int i = 0; i < num_fs_inputs; i++) {
1454                 struct vc4_varying_slot *input = &fs_inputs[i];
1455                 int j;
1456 
1457                 for (j = 0; j < c->num_outputs; j++) {
1458                         struct vc4_varying_slot *output =
1459                                 &c->output_slots[j];
1460 
1461                         if (input->slot == output->slot &&
1462                             input->swizzle == output->swizzle) {
1463                                 qir_VPM_WRITE(c, c->outputs[j]);
1464                                 break;
1465                         }
1466                 }
1467                 /* Emit padding if we didn't find a declared VS output for
1468                  * this FS input.
1469                  */
1470                 if (j == c->num_outputs)
1471                         qir_VPM_WRITE(c, qir_uniform_f(c, 0.0));
1472         }
1473 }
1474 
1475 static void
emit_coord_end(struct vc4_compile * c)1476 emit_coord_end(struct vc4_compile *c)
1477 {
1478         struct qreg rcp_w = ntq_rcp(c, c->outputs[c->output_position_index + 3]);
1479 
1480         emit_stub_vpm_read(c);
1481 
1482         for (int i = 0; i < 4; i++)
1483                 qir_VPM_WRITE(c, c->outputs[c->output_position_index + i]);
1484 
1485         emit_scaled_viewport_write(c, rcp_w);
1486         emit_zs_write(c, rcp_w);
1487         emit_rcp_wc_write(c, rcp_w);
1488         if (c->vs_key->per_vertex_point_size)
1489                 emit_point_size_write(c);
1490 }
1491 
1492 static void
vc4_optimize_nir(struct nir_shader * s)1493 vc4_optimize_nir(struct nir_shader *s)
1494 {
1495         bool progress;
1496         unsigned lower_flrp =
1497                 (s->options->lower_flrp16 ? 16 : 0) |
1498                 (s->options->lower_flrp32 ? 32 : 0) |
1499                 (s->options->lower_flrp64 ? 64 : 0);
1500 
1501         do {
1502                 progress = false;
1503 
1504                 NIR_PASS_V(s, nir_lower_vars_to_ssa);
1505                 NIR_PASS(progress, s, nir_lower_alu_to_scalar, NULL, NULL);
1506                 NIR_PASS(progress, s, nir_lower_phis_to_scalar, false);
1507                 NIR_PASS(progress, s, nir_copy_prop);
1508                 NIR_PASS(progress, s, nir_opt_remove_phis);
1509                 NIR_PASS(progress, s, nir_opt_dce);
1510                 NIR_PASS(progress, s, nir_opt_dead_cf);
1511                 NIR_PASS(progress, s, nir_opt_cse);
1512                 NIR_PASS(progress, s, nir_opt_peephole_select, 8, true, true);
1513                 NIR_PASS(progress, s, nir_opt_algebraic);
1514                 NIR_PASS(progress, s, nir_opt_constant_folding);
1515                 if (lower_flrp != 0) {
1516                         bool lower_flrp_progress = false;
1517 
1518                         NIR_PASS(lower_flrp_progress, s, nir_lower_flrp,
1519                                  lower_flrp,
1520                                  false /* always_precise */);
1521                         if (lower_flrp_progress) {
1522                                 NIR_PASS(progress, s, nir_opt_constant_folding);
1523                                 progress = true;
1524                         }
1525 
1526                         /* Nothing should rematerialize any flrps, so we only
1527                          * need to do this lowering once.
1528                          */
1529                         lower_flrp = 0;
1530                 }
1531 
1532                 NIR_PASS(progress, s, nir_opt_undef);
1533                 NIR_PASS(progress, s, nir_opt_loop_unroll);
1534         } while (progress);
1535 }
1536 
1537 static int
driver_location_compare(const void * in_a,const void * in_b)1538 driver_location_compare(const void *in_a, const void *in_b)
1539 {
1540         const nir_variable *const *a = in_a;
1541         const nir_variable *const *b = in_b;
1542 
1543         return (*a)->data.driver_location - (*b)->data.driver_location;
1544 }
1545 
1546 static void
ntq_setup_inputs(struct vc4_compile * c)1547 ntq_setup_inputs(struct vc4_compile *c)
1548 {
1549         unsigned num_entries = 0;
1550         nir_foreach_shader_in_variable(var, c->s)
1551                 num_entries++;
1552 
1553         nir_variable *vars[num_entries];
1554 
1555         unsigned i = 0;
1556         nir_foreach_shader_in_variable(var, c->s)
1557                 vars[i++] = var;
1558 
1559         /* Sort the variables so that we emit the input setup in
1560          * driver_location order.  This is required for VPM reads, whose data
1561          * is fetched into the VPM in driver_location (TGSI register index)
1562          * order.
1563          */
1564         qsort(&vars, num_entries, sizeof(*vars), driver_location_compare);
1565 
1566         for (unsigned i = 0; i < num_entries; i++) {
1567                 nir_variable *var = vars[i];
1568                 assert(glsl_type_is_vector_or_scalar(var->type));
1569                 unsigned loc = var->data.driver_location;
1570 
1571                 resize_qreg_array(c, &c->inputs, &c->inputs_array_size,
1572                                   (loc + 1) * 4);
1573 
1574                 if (c->stage == QSTAGE_FRAG) {
1575                         if (var->data.location == VARYING_SLOT_POS) {
1576                                 emit_fragcoord_input(c, loc);
1577                         } else if (util_varying_is_point_coord(var->data.location,
1578                                                                c->fs_key->point_sprite_mask)) {
1579                                 c->inputs[loc * 4 + 0] = c->point_x;
1580                                 c->inputs[loc * 4 + 1] = c->point_y;
1581                         } else {
1582                                 emit_fragment_input(c, loc, var->data.location);
1583                         }
1584                 } else {
1585                         emit_vertex_input(c, loc);
1586                 }
1587         }
1588 }
1589 
1590 static void
ntq_setup_outputs(struct vc4_compile * c)1591 ntq_setup_outputs(struct vc4_compile *c)
1592 {
1593         nir_foreach_shader_out_variable(var, c->s) {
1594                 assert(glsl_type_is_vector_or_scalar(var->type));
1595                 unsigned loc = var->data.driver_location * 4;
1596 
1597                 for (int i = 0; i < 4; i++)
1598                         add_output(c, loc + i, var->data.location, i);
1599 
1600                 if (c->stage == QSTAGE_FRAG) {
1601                         switch (var->data.location) {
1602                         case FRAG_RESULT_COLOR:
1603                         case FRAG_RESULT_DATA0:
1604                                 c->output_color_index = loc;
1605                                 break;
1606                         case FRAG_RESULT_DEPTH:
1607                                 c->output_position_index = loc;
1608                                 break;
1609                         case FRAG_RESULT_SAMPLE_MASK:
1610                                 c->output_sample_mask_index = loc;
1611                                 break;
1612                         }
1613                 } else {
1614                         switch (var->data.location) {
1615                         case VARYING_SLOT_POS:
1616                                 c->output_position_index = loc;
1617                                 break;
1618                         case VARYING_SLOT_PSIZ:
1619                                 c->output_point_size_index = loc;
1620                                 break;
1621                         }
1622                 }
1623         }
1624 }
1625 
1626 /**
1627  * Sets up the mapping from nir_register to struct qreg *.
1628  *
1629  * Each nir_register gets a struct qreg per 32-bit component being stored.
1630  */
1631 static void
ntq_setup_registers(struct vc4_compile * c,nir_function_impl * impl)1632 ntq_setup_registers(struct vc4_compile *c, nir_function_impl *impl)
1633 {
1634         nir_foreach_reg_decl(decl, impl) {
1635                 unsigned num_components = nir_intrinsic_num_components(decl);
1636                 unsigned array_len = nir_intrinsic_num_array_elems(decl);
1637                 array_len = MAX2(array_len, 1);
1638                 struct qreg *qregs = ralloc_array(c->def_ht, struct qreg,
1639                                                   array_len * num_components);
1640 
1641                 nir_def *nir_reg = &decl->def;
1642                 _mesa_hash_table_insert(c->def_ht, nir_reg, qregs);
1643 
1644                 for (int i = 0; i < array_len * num_components; i++)
1645                         qregs[i] = qir_get_temp(c);
1646         }
1647 }
1648 
1649 static void
ntq_emit_load_const(struct vc4_compile * c,nir_load_const_instr * instr)1650 ntq_emit_load_const(struct vc4_compile *c, nir_load_const_instr *instr)
1651 {
1652         struct qreg *qregs = ntq_init_ssa_def(c, &instr->def);
1653         for (int i = 0; i < instr->def.num_components; i++)
1654                 qregs[i] = qir_uniform_ui(c, instr->value[i].u32);
1655 
1656         _mesa_hash_table_insert(c->def_ht, &instr->def, qregs);
1657 }
1658 
1659 static void
ntq_emit_ssa_undef(struct vc4_compile * c,nir_undef_instr * instr)1660 ntq_emit_ssa_undef(struct vc4_compile *c, nir_undef_instr *instr)
1661 {
1662         struct qreg *qregs = ntq_init_ssa_def(c, &instr->def);
1663 
1664         /* QIR needs there to be *some* value, so pick 0 (same as for
1665          * ntq_setup_registers().
1666          */
1667         for (int i = 0; i < instr->def.num_components; i++)
1668                 qregs[i] = qir_uniform_ui(c, 0);
1669 }
1670 
1671 static void
ntq_emit_color_read(struct vc4_compile * c,nir_intrinsic_instr * instr)1672 ntq_emit_color_read(struct vc4_compile *c, nir_intrinsic_instr *instr)
1673 {
1674         assert(nir_src_as_uint(instr->src[0]) == 0);
1675 
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_def(c, &instr->def, 0,
1688                       qir_MOV(c, c->color_reads[sample_index]));
1689 }
1690 
1691 static void
ntq_emit_load_input(struct vc4_compile * c,nir_intrinsic_instr * instr)1692 ntq_emit_load_input(struct vc4_compile *c, nir_intrinsic_instr *instr)
1693 {
1694         assert(instr->num_components == 1);
1695         assert(nir_src_is_const(instr->src[0]) &&
1696                "vc4 doesn't support indirect inputs");
1697 
1698         if (c->stage == QSTAGE_FRAG &&
1699             nir_intrinsic_base(instr) >= VC4_NIR_TLB_COLOR_READ_INPUT) {
1700                 ntq_emit_color_read(c, instr);
1701                 return;
1702         }
1703 
1704         uint32_t offset = nir_intrinsic_base(instr) +
1705                           nir_src_as_uint(instr->src[0]);
1706         int comp = nir_intrinsic_component(instr);
1707         ntq_store_def(c, &instr->def, 0,
1708                       qir_MOV(c, c->inputs[offset * 4 + comp]));
1709 }
1710 
1711 static void
ntq_emit_intrinsic(struct vc4_compile * c,nir_intrinsic_instr * instr)1712 ntq_emit_intrinsic(struct vc4_compile *c, nir_intrinsic_instr *instr)
1713 {
1714         unsigned offset;
1715 
1716         switch (instr->intrinsic) {
1717         case nir_intrinsic_decl_reg:
1718         case nir_intrinsic_load_reg:
1719         case nir_intrinsic_store_reg:
1720                 break; /* Ignore these */
1721 
1722         case nir_intrinsic_load_uniform:
1723                 assert(instr->num_components == 1);
1724                 if (nir_src_is_const(instr->src[0])) {
1725                         offset = nir_intrinsic_base(instr) +
1726                                  nir_src_as_uint(instr->src[0]);
1727                         assert(offset % 4 == 0);
1728                         /* We need dwords */
1729                         offset = offset / 4;
1730                         ntq_store_def(c, &instr->def, 0,
1731                                       qir_uniform(c, QUNIFORM_UNIFORM,
1732                                                    offset));
1733                 } else {
1734                         ntq_store_def(c, &instr->def, 0,
1735                                       indirect_uniform_load(c, instr));
1736                 }
1737                 break;
1738 
1739         case nir_intrinsic_load_ubo:
1740                 assert(instr->num_components == 1);
1741                 ntq_store_def(c, &instr->def, 0, vc4_ubo_load(c, instr));
1742                 break;
1743 
1744         case nir_intrinsic_load_user_clip_plane:
1745                 for (int i = 0; i < nir_intrinsic_dest_components(instr); i++) {
1746                         ntq_store_def(c, &instr->def, i,
1747                                       qir_uniform(c, QUNIFORM_USER_CLIP_PLANE,
1748                                                   nir_intrinsic_ucp_id(instr) *
1749                                                   4 + i));
1750                 }
1751                 break;
1752 
1753         case nir_intrinsic_load_blend_const_color_r_float:
1754         case nir_intrinsic_load_blend_const_color_g_float:
1755         case nir_intrinsic_load_blend_const_color_b_float:
1756         case nir_intrinsic_load_blend_const_color_a_float:
1757                 ntq_store_def(c, &instr->def, 0,
1758                               qir_uniform(c, QUNIFORM_BLEND_CONST_COLOR_X +
1759                                           (instr->intrinsic -
1760                                            nir_intrinsic_load_blend_const_color_r_float),
1761                                           0));
1762                 break;
1763 
1764         case nir_intrinsic_load_blend_const_color_rgba8888_unorm:
1765                 ntq_store_def(c, &instr->def, 0,
1766                               qir_uniform(c, QUNIFORM_BLEND_CONST_COLOR_RGBA,
1767                                           0));
1768                 break;
1769 
1770         case nir_intrinsic_load_blend_const_color_aaaa8888_unorm:
1771                 ntq_store_def(c, &instr->def, 0,
1772                               qir_uniform(c, QUNIFORM_BLEND_CONST_COLOR_AAAA,
1773                                           0));
1774                 break;
1775 
1776         case nir_intrinsic_load_sample_mask_in:
1777                 ntq_store_def(c, &instr->def, 0,
1778                               qir_uniform(c, QUNIFORM_SAMPLE_MASK, 0));
1779                 break;
1780 
1781         case nir_intrinsic_load_front_face:
1782                 /* The register contains 0 (front) or 1 (back), and we need to
1783                  * turn it into a NIR bool where true means front.
1784                  */
1785                 ntq_store_def(c, &instr->def, 0,
1786                               qir_ADD(c,
1787                                       qir_uniform_ui(c, -1),
1788                                       qir_reg(QFILE_FRAG_REV_FLAG, 0)));
1789                 break;
1790 
1791         case nir_intrinsic_load_input:
1792                 ntq_emit_load_input(c, instr);
1793                 break;
1794 
1795         case nir_intrinsic_store_output:
1796                 assert(nir_src_is_const(instr->src[1]) &&
1797                        "vc4 doesn't support indirect outputs");
1798                 offset = nir_intrinsic_base(instr) +
1799                          nir_src_as_uint(instr->src[1]);
1800 
1801                 /* MSAA color outputs are the only case where we have an
1802                  * output that's not lowered to being a store of a single 32
1803                  * bit value.
1804                  */
1805                 if (c->stage == QSTAGE_FRAG && instr->num_components == 4) {
1806                         assert(offset == c->output_color_index);
1807                         for (int i = 0; i < 4; i++) {
1808                                 c->sample_colors[i] =
1809                                         qir_MOV(c, ntq_get_src(c, instr->src[0],
1810                                                                i));
1811                         }
1812                 } else {
1813                         offset = offset * 4 + nir_intrinsic_component(instr);
1814                         assert(instr->num_components == 1);
1815                         c->outputs[offset] =
1816                                 qir_MOV(c, ntq_get_src(c, instr->src[0], 0));
1817                         c->num_outputs = MAX2(c->num_outputs, offset + 1);
1818                 }
1819                 break;
1820 
1821         case nir_intrinsic_discard:
1822                 if (c->execute.file != QFILE_NULL) {
1823                         qir_SF(c, c->execute);
1824                         qir_MOV_cond(c, QPU_COND_ZS, c->discard,
1825                                      qir_uniform_ui(c, ~0));
1826                 } else {
1827                         qir_MOV_dest(c, c->discard, qir_uniform_ui(c, ~0));
1828                 }
1829                 break;
1830 
1831         case nir_intrinsic_discard_if: {
1832                 /* true (~0) if we're discarding */
1833                 struct qreg cond = ntq_get_src(c, instr->src[0], 0);
1834 
1835                 if (c->execute.file != QFILE_NULL) {
1836                         /* execute == 0 means the channel is active.  Invert
1837                          * the condition so that we can use zero as "executing
1838                          * and discarding."
1839                          */
1840                         qir_SF(c, qir_AND(c, c->execute, qir_NOT(c, cond)));
1841                         qir_MOV_cond(c, QPU_COND_ZS, c->discard, cond);
1842                 } else {
1843                         qir_OR_dest(c, c->discard, c->discard,
1844                                     ntq_get_src(c, instr->src[0], 0));
1845                 }
1846 
1847                 break;
1848         }
1849 
1850         case nir_intrinsic_load_texture_scale: {
1851                 assert(nir_src_is_const(instr->src[0]));
1852                 int sampler = nir_src_as_int(instr->src[0]);
1853 
1854                 ntq_store_def(c, &instr->def, 0,
1855                               qir_uniform(c, QUNIFORM_TEXRECT_SCALE_X, sampler));
1856                 ntq_store_def(c, &instr->def, 1,
1857                               qir_uniform(c, QUNIFORM_TEXRECT_SCALE_Y, sampler));
1858                 break;
1859         }
1860 
1861         default:
1862                 fprintf(stderr, "Unknown intrinsic: ");
1863                 nir_print_instr(&instr->instr, stderr);
1864                 fprintf(stderr, "\n");
1865                 break;
1866         }
1867 }
1868 
1869 /* Clears (activates) the execute flags for any channels whose jump target
1870  * matches this block.
1871  */
1872 static void
ntq_activate_execute_for_block(struct vc4_compile * c)1873 ntq_activate_execute_for_block(struct vc4_compile *c)
1874 {
1875         qir_SF(c, qir_SUB(c,
1876                           c->execute,
1877                           qir_uniform_ui(c, c->cur_block->index)));
1878         qir_MOV_cond(c, QPU_COND_ZS, c->execute, qir_uniform_ui(c, 0));
1879 }
1880 
1881 static void
ntq_emit_if(struct vc4_compile * c,nir_if * if_stmt)1882 ntq_emit_if(struct vc4_compile *c, nir_if *if_stmt)
1883 {
1884         if (!c->vc4->screen->has_control_flow) {
1885                 fprintf(stderr,
1886                         "IF statement support requires updated kernel.\n");
1887                 return;
1888         }
1889 
1890         nir_block *nir_else_block = nir_if_first_else_block(if_stmt);
1891         bool empty_else_block =
1892                 (nir_else_block == nir_if_last_else_block(if_stmt) &&
1893                  exec_list_is_empty(&nir_else_block->instr_list));
1894 
1895         struct qblock *then_block = qir_new_block(c);
1896         struct qblock *after_block = qir_new_block(c);
1897         struct qblock *else_block;
1898         if (empty_else_block)
1899                 else_block = after_block;
1900         else
1901                 else_block = qir_new_block(c);
1902 
1903         bool was_top_level = false;
1904         if (c->execute.file == QFILE_NULL) {
1905                 c->execute = qir_MOV(c, qir_uniform_ui(c, 0));
1906                 was_top_level = true;
1907         }
1908 
1909         /* Set ZS for executing (execute == 0) and jumping (if->condition ==
1910          * 0) channels, and then update execute flags for those to point to
1911          * the ELSE block.
1912          */
1913         qir_SF(c, qir_OR(c,
1914                          c->execute,
1915                          ntq_get_src(c, if_stmt->condition, 0)));
1916         qir_MOV_cond(c, QPU_COND_ZS, c->execute,
1917                      qir_uniform_ui(c, else_block->index));
1918 
1919         /* Jump to ELSE if nothing is active for THEN, otherwise fall
1920          * through.
1921          */
1922         qir_SF(c, c->execute);
1923         qir_BRANCH(c, QPU_COND_BRANCH_ALL_ZC);
1924         qir_link_blocks(c->cur_block, else_block);
1925         qir_link_blocks(c->cur_block, then_block);
1926 
1927         /* Process the THEN block. */
1928         qir_set_emit_block(c, then_block);
1929         ntq_emit_cf_list(c, &if_stmt->then_list);
1930 
1931         if (!empty_else_block) {
1932                 /* Handle the end of the THEN block.  First, all currently
1933                  * active channels update their execute flags to point to
1934                  * ENDIF
1935                  */
1936                 qir_SF(c, c->execute);
1937                 qir_MOV_cond(c, QPU_COND_ZS, c->execute,
1938                              qir_uniform_ui(c, after_block->index));
1939 
1940                 /* If everything points at ENDIF, then jump there immediately. */
1941                 qir_SF(c, qir_SUB(c, c->execute, qir_uniform_ui(c, after_block->index)));
1942                 qir_BRANCH(c, QPU_COND_BRANCH_ALL_ZS);
1943                 qir_link_blocks(c->cur_block, after_block);
1944                 qir_link_blocks(c->cur_block, else_block);
1945 
1946                 qir_set_emit_block(c, else_block);
1947                 ntq_activate_execute_for_block(c);
1948                 ntq_emit_cf_list(c, &if_stmt->else_list);
1949         }
1950 
1951         qir_link_blocks(c->cur_block, after_block);
1952 
1953         qir_set_emit_block(c, after_block);
1954         if (was_top_level) {
1955                 c->execute = c->undef;
1956                 c->last_top_block = c->cur_block;
1957         } else {
1958                 ntq_activate_execute_for_block(c);
1959         }
1960 }
1961 
1962 static void
ntq_emit_jump(struct vc4_compile * c,nir_jump_instr * jump)1963 ntq_emit_jump(struct vc4_compile *c, nir_jump_instr *jump)
1964 {
1965         struct qblock *jump_block;
1966         switch (jump->type) {
1967         case nir_jump_break:
1968                 jump_block = c->loop_break_block;
1969                 break;
1970         case nir_jump_continue:
1971                 jump_block = c->loop_cont_block;
1972                 break;
1973         default:
1974                 unreachable("Unsupported jump type\n");
1975         }
1976 
1977         qir_SF(c, c->execute);
1978         qir_MOV_cond(c, QPU_COND_ZS, c->execute,
1979                      qir_uniform_ui(c, jump_block->index));
1980 
1981         /* Jump to the destination block if everyone has taken the jump. */
1982         qir_SF(c, qir_SUB(c, c->execute, qir_uniform_ui(c, jump_block->index)));
1983         qir_BRANCH(c, QPU_COND_BRANCH_ALL_ZS);
1984         struct qblock *new_block = qir_new_block(c);
1985         qir_link_blocks(c->cur_block, jump_block);
1986         qir_link_blocks(c->cur_block, new_block);
1987         qir_set_emit_block(c, new_block);
1988 }
1989 
1990 static void
ntq_emit_instr(struct vc4_compile * c,nir_instr * instr)1991 ntq_emit_instr(struct vc4_compile *c, nir_instr *instr)
1992 {
1993         switch (instr->type) {
1994         case nir_instr_type_alu:
1995                 ntq_emit_alu(c, nir_instr_as_alu(instr));
1996                 break;
1997 
1998         case nir_instr_type_intrinsic:
1999                 ntq_emit_intrinsic(c, nir_instr_as_intrinsic(instr));
2000                 break;
2001 
2002         case nir_instr_type_load_const:
2003                 ntq_emit_load_const(c, nir_instr_as_load_const(instr));
2004                 break;
2005 
2006         case nir_instr_type_undef:
2007                 ntq_emit_ssa_undef(c, nir_instr_as_undef(instr));
2008                 break;
2009 
2010         case nir_instr_type_tex:
2011                 ntq_emit_tex(c, nir_instr_as_tex(instr));
2012                 break;
2013 
2014         case nir_instr_type_jump:
2015                 ntq_emit_jump(c, nir_instr_as_jump(instr));
2016                 break;
2017 
2018         default:
2019                 fprintf(stderr, "Unknown NIR instr type: ");
2020                 nir_print_instr(instr, stderr);
2021                 fprintf(stderr, "\n");
2022                 abort();
2023         }
2024 }
2025 
2026 static void
ntq_emit_block(struct vc4_compile * c,nir_block * block)2027 ntq_emit_block(struct vc4_compile *c, nir_block *block)
2028 {
2029         nir_foreach_instr(instr, block) {
2030                 ntq_emit_instr(c, instr);
2031         }
2032 }
2033 
2034 static void ntq_emit_cf_list(struct vc4_compile *c, struct exec_list *list);
2035 
2036 static void
ntq_emit_loop(struct vc4_compile * c,nir_loop * loop)2037 ntq_emit_loop(struct vc4_compile *c, nir_loop *loop)
2038 {
2039         assert(!nir_loop_has_continue_construct(loop));
2040         if (!c->vc4->screen->has_control_flow) {
2041                 fprintf(stderr,
2042                         "loop support requires updated kernel.\n");
2043                 ntq_emit_cf_list(c, &loop->body);
2044                 return;
2045         }
2046 
2047         bool was_top_level = false;
2048         if (c->execute.file == QFILE_NULL) {
2049                 c->execute = qir_MOV(c, qir_uniform_ui(c, 0));
2050                 was_top_level = true;
2051         }
2052 
2053         struct qblock *save_loop_cont_block = c->loop_cont_block;
2054         struct qblock *save_loop_break_block = c->loop_break_block;
2055 
2056         c->loop_cont_block = qir_new_block(c);
2057         c->loop_break_block = qir_new_block(c);
2058 
2059         qir_link_blocks(c->cur_block, c->loop_cont_block);
2060         qir_set_emit_block(c, c->loop_cont_block);
2061         ntq_activate_execute_for_block(c);
2062 
2063         ntq_emit_cf_list(c, &loop->body);
2064 
2065         /* If anything had explicitly continued, or is here at the end of the
2066          * loop, then we need to loop again.  SF updates are masked by the
2067          * instruction's condition, so we can do the OR of the two conditions
2068          * within SF.
2069          */
2070         qir_SF(c, c->execute);
2071         struct qinst *cont_check =
2072                 qir_SUB_dest(c,
2073                              c->undef,
2074                              c->execute,
2075                              qir_uniform_ui(c, c->loop_cont_block->index));
2076         cont_check->cond = QPU_COND_ZC;
2077         cont_check->sf = true;
2078 
2079         qir_BRANCH(c, QPU_COND_BRANCH_ANY_ZS);
2080         qir_link_blocks(c->cur_block, c->loop_cont_block);
2081         qir_link_blocks(c->cur_block, c->loop_break_block);
2082 
2083         qir_set_emit_block(c, c->loop_break_block);
2084         if (was_top_level) {
2085                 c->execute = c->undef;
2086                 c->last_top_block = c->cur_block;
2087         } else {
2088                 ntq_activate_execute_for_block(c);
2089         }
2090 
2091         c->loop_break_block = save_loop_break_block;
2092         c->loop_cont_block = save_loop_cont_block;
2093 }
2094 
2095 static void
ntq_emit_function(struct vc4_compile * c,nir_function_impl * func)2096 ntq_emit_function(struct vc4_compile *c, nir_function_impl *func)
2097 {
2098         fprintf(stderr, "FUNCTIONS not handled.\n");
2099         abort();
2100 }
2101 
2102 static void
ntq_emit_cf_list(struct vc4_compile * c,struct exec_list * list)2103 ntq_emit_cf_list(struct vc4_compile *c, struct exec_list *list)
2104 {
2105         foreach_list_typed(nir_cf_node, node, node, list) {
2106                 switch (node->type) {
2107                 case nir_cf_node_block:
2108                         ntq_emit_block(c, nir_cf_node_as_block(node));
2109                         break;
2110 
2111                 case nir_cf_node_if:
2112                         ntq_emit_if(c, nir_cf_node_as_if(node));
2113                         break;
2114 
2115                 case nir_cf_node_loop:
2116                         ntq_emit_loop(c, nir_cf_node_as_loop(node));
2117                         break;
2118 
2119                 case nir_cf_node_function:
2120                         ntq_emit_function(c, nir_cf_node_as_function(node));
2121                         break;
2122 
2123                 default:
2124                         fprintf(stderr, "Unknown NIR node type\n");
2125                         abort();
2126                 }
2127         }
2128 }
2129 
2130 static void
ntq_emit_impl(struct vc4_compile * c,nir_function_impl * impl)2131 ntq_emit_impl(struct vc4_compile *c, nir_function_impl *impl)
2132 {
2133         ntq_setup_registers(c, impl);
2134         ntq_emit_cf_list(c, &impl->body);
2135 }
2136 
2137 static void
nir_to_qir(struct vc4_compile * c)2138 nir_to_qir(struct vc4_compile *c)
2139 {
2140         if (c->stage == QSTAGE_FRAG && c->s->info.fs.uses_discard)
2141                 c->discard = qir_MOV(c, qir_uniform_ui(c, 0));
2142 
2143         ntq_setup_inputs(c);
2144         ntq_setup_outputs(c);
2145 
2146         /* Find the main function and emit the body. */
2147         nir_foreach_function(function, c->s) {
2148                 assert(strcmp(function->name, "main") == 0);
2149                 assert(function->impl);
2150                 ntq_emit_impl(c, function->impl);
2151         }
2152 }
2153 
2154 static const nir_shader_compiler_options nir_options = {
2155         .lower_all_io_to_temps = true,
2156         .lower_extract_byte = true,
2157         .lower_extract_word = true,
2158         .lower_insert_byte = true,
2159         .lower_insert_word = true,
2160         .lower_fdiv = true,
2161         .lower_ffma16 = true,
2162         .lower_ffma32 = true,
2163         .lower_ffma64 = true,
2164         .lower_flrp32 = true,
2165         .lower_fmod = true,
2166         .lower_fpow = true,
2167         .lower_fsat = true,
2168         .lower_fsqrt = true,
2169         .lower_ldexp = true,
2170         .lower_fneg = true,
2171         .lower_ineg = true,
2172         .lower_to_scalar = true,
2173         .lower_umax = true,
2174         .lower_umin = true,
2175         .lower_isign = true,
2176         .has_fsub = true,
2177         .has_isub = true,
2178         .has_texture_scaling = true,
2179         .lower_mul_high = true,
2180         .max_unroll_iterations = 32,
2181         .force_indirect_unrolling = (nir_var_shader_in | nir_var_shader_out | nir_var_function_temp),
2182 };
2183 
2184 const void *
vc4_screen_get_compiler_options(struct pipe_screen * pscreen,enum pipe_shader_ir ir,enum pipe_shader_type shader)2185 vc4_screen_get_compiler_options(struct pipe_screen *pscreen,
2186                                 enum pipe_shader_ir ir,
2187                                 enum pipe_shader_type shader)
2188 {
2189         return &nir_options;
2190 }
2191 
2192 static int
count_nir_instrs(nir_shader * nir)2193 count_nir_instrs(nir_shader *nir)
2194 {
2195         int count = 0;
2196         nir_foreach_function_impl(impl, nir) {
2197                 nir_foreach_block(block, impl) {
2198                         nir_foreach_instr(instr, block)
2199                                 count++;
2200                 }
2201         }
2202         return count;
2203 }
2204 
2205 static struct vc4_compile *
vc4_shader_ntq(struct vc4_context * vc4,enum qstage stage,struct vc4_key * key,bool fs_threaded)2206 vc4_shader_ntq(struct vc4_context *vc4, enum qstage stage,
2207                struct vc4_key *key, bool fs_threaded)
2208 {
2209         struct vc4_compile *c = qir_compile_init();
2210 
2211         c->vc4 = vc4;
2212         c->stage = stage;
2213         c->shader_state = &key->shader_state->base;
2214         c->program_id = key->shader_state->program_id;
2215         c->variant_id =
2216                 p_atomic_inc_return(&key->shader_state->compiled_variant_count);
2217         c->fs_threaded = fs_threaded;
2218 
2219         c->key = key;
2220         switch (stage) {
2221         case QSTAGE_FRAG:
2222                 c->fs_key = (struct vc4_fs_key *)key;
2223                 if (c->fs_key->is_points) {
2224                         c->point_x = emit_fragment_varying(c, ~0, 0);
2225                         c->point_y = emit_fragment_varying(c, ~0, 0);
2226                 } else if (c->fs_key->is_lines) {
2227                         c->line_x = emit_fragment_varying(c, ~0, 0);
2228                 }
2229                 break;
2230         case QSTAGE_VERT:
2231                 c->vs_key = (struct vc4_vs_key *)key;
2232                 break;
2233         case QSTAGE_COORD:
2234                 c->vs_key = (struct vc4_vs_key *)key;
2235                 break;
2236         }
2237 
2238         c->s = nir_shader_clone(c, key->shader_state->base.ir.nir);
2239 
2240         if (stage == QSTAGE_FRAG) {
2241                 NIR_PASS_V(c->s, vc4_nir_lower_blend, c);
2242         }
2243 
2244         struct nir_lower_tex_options tex_options = {
2245                 .lower_txp = ~0,
2246 
2247                 /* Apply swizzles to all samplers. */
2248                 .swizzle_result = ~0,
2249                 .lower_invalid_implicit_lod = true,
2250         };
2251 
2252         /* Lower the format swizzle and ARB_texture_swizzle-style swizzle.
2253          * The format swizzling applies before sRGB decode, and
2254          * ARB_texture_swizzle is the last thing before returning the sample.
2255          */
2256         for (int i = 0; i < ARRAY_SIZE(key->tex); i++) {
2257                 enum pipe_format format = c->key->tex[i].format;
2258 
2259                 if (!format)
2260                         continue;
2261 
2262                 const uint8_t *format_swizzle = vc4_get_format_swizzle(format);
2263 
2264                 for (int j = 0; j < 4; j++) {
2265                         uint8_t arb_swiz = c->key->tex[i].swizzle[j];
2266 
2267                         if (arb_swiz <= 3) {
2268                                 tex_options.swizzles[i][j] =
2269                                         format_swizzle[arb_swiz];
2270                         } else {
2271                                 tex_options.swizzles[i][j] = arb_swiz;
2272                         }
2273                 }
2274 
2275                 if (util_format_is_srgb(format))
2276                         tex_options.lower_srgb |= (1 << i);
2277         }
2278 
2279         NIR_PASS_V(c->s, nir_lower_tex, &tex_options);
2280 
2281         if (c->key->ucp_enables) {
2282                 if (stage == QSTAGE_FRAG) {
2283                         NIR_PASS_V(c->s, nir_lower_clip_fs,
2284                                    c->key->ucp_enables, false);
2285                 } else {
2286                         NIR_PASS_V(c->s, nir_lower_clip_vs,
2287                                    c->key->ucp_enables, false, false, NULL);
2288                         NIR_PASS_V(c->s, nir_lower_io_to_scalar,
2289                                    nir_var_shader_out, NULL, NULL);
2290                 }
2291         }
2292 
2293         /* FS input scalarizing must happen after nir_lower_two_sided_color,
2294          * which only handles a vec4 at a time.  Similarly, VS output
2295          * scalarizing must happen after nir_lower_clip_vs.
2296          */
2297         if (c->stage == QSTAGE_FRAG)
2298                 NIR_PASS_V(c->s, nir_lower_io_to_scalar, nir_var_shader_in, NULL, NULL);
2299         else
2300                 NIR_PASS_V(c->s, nir_lower_io_to_scalar, nir_var_shader_out, NULL, NULL);
2301 
2302         NIR_PASS_V(c->s, vc4_nir_lower_io, c);
2303         NIR_PASS_V(c->s, vc4_nir_lower_txf_ms, c);
2304         nir_lower_idiv_options idiv_options = {
2305                 .allow_fp16 = true,
2306         };
2307         NIR_PASS_V(c->s, nir_lower_idiv, &idiv_options);
2308         NIR_PASS(_, c->s, nir_lower_alu);
2309 
2310         vc4_optimize_nir(c->s);
2311 
2312         /* Do late algebraic optimization to turn add(a, neg(b)) back into
2313          * subs, then the mandatory cleanup after algebraic.  Note that it may
2314          * produce fnegs, and if so then we need to keep running to squash
2315          * fneg(fneg(a)).
2316          */
2317         bool more_late_algebraic = true;
2318         while (more_late_algebraic) {
2319                 more_late_algebraic = false;
2320                 NIR_PASS(more_late_algebraic, c->s, nir_opt_algebraic_late);
2321                 NIR_PASS_V(c->s, nir_opt_constant_folding);
2322                 NIR_PASS_V(c->s, nir_copy_prop);
2323                 NIR_PASS_V(c->s, nir_opt_dce);
2324                 NIR_PASS_V(c->s, nir_opt_cse);
2325         }
2326 
2327         NIR_PASS_V(c->s, nir_lower_bool_to_int32);
2328 
2329         NIR_PASS_V(c->s, nir_convert_from_ssa, true);
2330         NIR_PASS_V(c->s, nir_trivialize_registers);
2331 
2332         if (VC4_DBG(NIR)) {
2333                 fprintf(stderr, "%s prog %d/%d NIR:\n",
2334                         qir_get_stage_name(c->stage),
2335                         c->program_id, c->variant_id);
2336                 nir_print_shader(c->s, stderr);
2337         }
2338 
2339         nir_to_qir(c);
2340 
2341         switch (stage) {
2342         case QSTAGE_FRAG:
2343                 /* FS threading requires that the thread execute
2344                  * QPU_SIG_LAST_THREAD_SWITCH exactly once before terminating
2345                  * (with no other THRSW afterwards, obviously).  If we didn't
2346                  * fetch a texture at a top level block, this wouldn't be
2347                  * true.
2348                  */
2349                 if (c->fs_threaded && !c->last_thrsw_at_top_level) {
2350                         c->failed = true;
2351                         return c;
2352                 }
2353 
2354                 emit_frag_end(c);
2355                 break;
2356         case QSTAGE_VERT:
2357                 emit_vert_end(c,
2358                               c->vs_key->fs_inputs->input_slots,
2359                               c->vs_key->fs_inputs->num_inputs);
2360                 break;
2361         case QSTAGE_COORD:
2362                 emit_coord_end(c);
2363                 break;
2364         }
2365 
2366         if (VC4_DBG(QIR)) {
2367                 fprintf(stderr, "%s prog %d/%d pre-opt QIR:\n",
2368                         qir_get_stage_name(c->stage),
2369                         c->program_id, c->variant_id);
2370                 qir_dump(c);
2371                 fprintf(stderr, "\n");
2372         }
2373 
2374         qir_optimize(c);
2375         qir_lower_uniforms(c);
2376 
2377         qir_schedule_instructions(c);
2378         qir_emit_uniform_stream_resets(c);
2379 
2380         if (VC4_DBG(QIR)) {
2381                 fprintf(stderr, "%s prog %d/%d QIR:\n",
2382                         qir_get_stage_name(c->stage),
2383                         c->program_id, c->variant_id);
2384                 qir_dump(c);
2385                 fprintf(stderr, "\n");
2386         }
2387 
2388         qir_reorder_uniforms(c);
2389         vc4_generate_code(vc4, c);
2390 
2391         ralloc_free(c->s);
2392 
2393         return c;
2394 }
2395 
2396 static void
vc4_setup_shared_precompile_key(struct vc4_uncompiled_shader * uncompiled,struct vc4_key * key)2397 vc4_setup_shared_precompile_key(struct vc4_uncompiled_shader *uncompiled,
2398                                 struct vc4_key *key)
2399 {
2400         nir_shader *s = uncompiled->base.ir.nir;
2401 
2402         for (int i = 0; i < s->info.num_textures; i++) {
2403                 key->tex[i].format = PIPE_FORMAT_R8G8B8A8_UNORM;
2404                 key->tex[i].swizzle[0] = PIPE_SWIZZLE_X;
2405                 key->tex[i].swizzle[1] = PIPE_SWIZZLE_Y;
2406                 key->tex[i].swizzle[2] = PIPE_SWIZZLE_Z;
2407                 key->tex[i].swizzle[3] = PIPE_SWIZZLE_W;
2408         }
2409 }
2410 
2411 static inline struct vc4_varying_slot
vc4_slot_from_slot_and_component(uint8_t slot,uint8_t component)2412 vc4_slot_from_slot_and_component(uint8_t slot, uint8_t component)
2413 {
2414         assume(slot < 255 / 4);
2415         return (struct vc4_varying_slot){ (slot << 2) + component };
2416 }
2417 
2418 static void
precompile_all_fs_inputs(nir_shader * s,struct vc4_fs_inputs * fs_inputs)2419 precompile_all_fs_inputs(nir_shader *s,
2420                          struct vc4_fs_inputs *fs_inputs)
2421 {
2422         /* Assume all VS outputs will actually be used by the FS and output
2423          * them (the two sides have to match exactly) */
2424         nir_foreach_shader_out_variable(var, s) {
2425                 const int array_len =
2426                         glsl_type_is_vector_or_scalar(var->type) ?
2427                         1 : glsl_get_length(var->type);
2428                 for (int j = 0; j < array_len; j++) {
2429                         const int slot = var->data.location + j;
2430                         const int num_components =
2431                                 glsl_get_components(var->type);
2432                         for (int i = 0; i < num_components; i++) {
2433                                 const int swiz = var->data.location_frac + i;
2434                                 fs_inputs->input_slots[fs_inputs->num_inputs++] =
2435                                         vc4_slot_from_slot_and_component(slot,
2436                                                                          swiz);
2437                         }
2438                 }
2439         }
2440 }
2441 
2442 /**
2443  * Precompiles a shader variant at shader state creation time if
2444  * VC4_DEBUG=shaderdb is set.
2445  */
2446 static void
vc4_shader_precompile(struct vc4_context * vc4,struct vc4_uncompiled_shader * so)2447 vc4_shader_precompile(struct vc4_context *vc4,
2448                       struct vc4_uncompiled_shader *so)
2449 {
2450         nir_shader *s = so->base.ir.nir;
2451 
2452         if (s->info.stage == MESA_SHADER_FRAGMENT) {
2453                 struct vc4_fs_key key = {
2454                         .base.shader_state = so,
2455                         .depth_enabled = true,
2456                         .logicop_func = PIPE_LOGICOP_COPY,
2457                         .color_format = PIPE_FORMAT_R8G8B8A8_UNORM,
2458                         .blend = {
2459                                 .blend_enable = false,
2460                                 .colormask = PIPE_MASK_RGBA,
2461                         },
2462                 };
2463 
2464                 vc4_setup_shared_precompile_key(so, &key.base);
2465                 vc4_get_compiled_shader(vc4, QSTAGE_FRAG, &key.base);
2466         } else {
2467                 assert(s->info.stage == MESA_SHADER_VERTEX);
2468                 struct vc4_varying_slot input_slots[64] = {};
2469                 struct vc4_fs_inputs fs_inputs = {
2470                         .input_slots = input_slots,
2471                         .num_inputs = 0,
2472                 };
2473                 struct vc4_vs_key key = {
2474                         .base.shader_state = so,
2475                         .fs_inputs = &fs_inputs,
2476                 };
2477 
2478                 vc4_setup_shared_precompile_key(so, &key.base);
2479                 precompile_all_fs_inputs(s, &fs_inputs);
2480                 vc4_get_compiled_shader(vc4, QSTAGE_VERT, &key.base);
2481 
2482                 /* Compile VS bin shader: only position (XXX: include TF) */
2483                 key.is_coord = true;
2484                 fs_inputs.num_inputs = 0;
2485                 precompile_all_fs_inputs(s, &fs_inputs);
2486                 for (int i = 0; i < 4; i++) {
2487                         fs_inputs.input_slots[fs_inputs.num_inputs++] =
2488                                 vc4_slot_from_slot_and_component(VARYING_SLOT_POS,
2489                                                                  i);
2490                 }
2491                 vc4_get_compiled_shader(vc4, QSTAGE_VERT, &key.base);
2492         }
2493 }
2494 
2495 static void *
vc4_shader_state_create(struct pipe_context * pctx,const struct pipe_shader_state * cso)2496 vc4_shader_state_create(struct pipe_context *pctx,
2497                         const struct pipe_shader_state *cso)
2498 {
2499         struct vc4_context *vc4 = vc4_context(pctx);
2500         struct vc4_uncompiled_shader *so = CALLOC_STRUCT(vc4_uncompiled_shader);
2501         if (!so)
2502                 return NULL;
2503 
2504         so->program_id = vc4->next_uncompiled_program_id++;
2505 
2506         nir_shader *s;
2507 
2508         if (cso->type == PIPE_SHADER_IR_NIR) {
2509                 /* The backend takes ownership of the NIR shader on state
2510                  * creation.
2511                  */
2512                 s = cso->ir.nir;
2513        } else {
2514                 assert(cso->type == PIPE_SHADER_IR_TGSI);
2515 
2516                 if (VC4_DBG(TGSI)) {
2517                         fprintf(stderr, "prog %d TGSI:\n",
2518                                 so->program_id);
2519                         tgsi_dump(cso->tokens, 0);
2520                         fprintf(stderr, "\n");
2521                 }
2522                 s = tgsi_to_nir(cso->tokens, pctx->screen, false);
2523         }
2524 
2525         if (s->info.stage == MESA_SHADER_VERTEX)
2526                 NIR_PASS_V(s, nir_lower_point_size, 1.0f, 0.0f);
2527 
2528         NIR_PASS_V(s, nir_lower_io,
2529                    nir_var_shader_in | nir_var_shader_out | nir_var_uniform,
2530                    type_size, (nir_lower_io_options)0);
2531 
2532         NIR_PASS_V(s, nir_normalize_cubemap_coords);
2533 
2534         NIR_PASS_V(s, nir_lower_load_const_to_scalar);
2535 
2536         vc4_optimize_nir(s);
2537 
2538         NIR_PASS_V(s, nir_remove_dead_variables, nir_var_function_temp, NULL);
2539 
2540         /* Garbage collect dead instructions */
2541         nir_sweep(s);
2542 
2543         so->base.type = PIPE_SHADER_IR_NIR;
2544         so->base.ir.nir = s;
2545 
2546         if (VC4_DBG(NIR)) {
2547                 fprintf(stderr, "%s prog %d NIR:\n",
2548                         gl_shader_stage_name(s->info.stage),
2549                         so->program_id);
2550                 nir_print_shader(s, stderr);
2551                 fprintf(stderr, "\n");
2552         }
2553 
2554         if (VC4_DBG(SHADERDB)) {
2555                 vc4_shader_precompile(vc4, so);
2556         }
2557 
2558         return so;
2559 }
2560 
2561 static void
copy_uniform_state_to_shader(struct vc4_compiled_shader * shader,struct vc4_compile * c)2562 copy_uniform_state_to_shader(struct vc4_compiled_shader *shader,
2563                              struct vc4_compile *c)
2564 {
2565         int count = c->num_uniforms;
2566         struct vc4_shader_uniform_info *uinfo = &shader->uniforms;
2567 
2568         uinfo->count = count;
2569         uinfo->data = ralloc_array(shader, uint32_t, count);
2570         memcpy(uinfo->data, c->uniform_data,
2571                count * sizeof(*uinfo->data));
2572         uinfo->contents = ralloc_array(shader, enum quniform_contents, count);
2573         memcpy(uinfo->contents, c->uniform_contents,
2574                count * sizeof(*uinfo->contents));
2575         uinfo->num_texture_samples = c->num_texture_samples;
2576 
2577         vc4_set_shader_uniform_dirty_flags(shader);
2578 }
2579 
2580 static void
vc4_setup_compiled_fs_inputs(struct vc4_context * vc4,struct vc4_compile * c,struct vc4_compiled_shader * shader)2581 vc4_setup_compiled_fs_inputs(struct vc4_context *vc4, struct vc4_compile *c,
2582                              struct vc4_compiled_shader *shader)
2583 {
2584         struct vc4_fs_inputs inputs;
2585 
2586         memset(&inputs, 0, sizeof(inputs));
2587         inputs.input_slots = ralloc_array(shader,
2588                                           struct vc4_varying_slot,
2589                                           c->num_input_slots);
2590 
2591         bool input_live[c->num_input_slots];
2592 
2593         memset(input_live, 0, sizeof(input_live));
2594         qir_for_each_inst_inorder(inst, c) {
2595                 for (int i = 0; i < qir_get_nsrc(inst); i++) {
2596                         if (inst->src[i].file == QFILE_VARY)
2597                                 input_live[inst->src[i].index] = true;
2598                 }
2599         }
2600 
2601         for (int i = 0; i < c->num_input_slots; i++) {
2602                 struct vc4_varying_slot *slot = &c->input_slots[i];
2603 
2604                 if (!input_live[i])
2605                         continue;
2606 
2607                 /* Skip non-VS-output inputs. */
2608                 if (slot->slot == (uint8_t)~0)
2609                         continue;
2610 
2611                 if (slot->slot == VARYING_SLOT_COL0 ||
2612                     slot->slot == VARYING_SLOT_COL1 ||
2613                     slot->slot == VARYING_SLOT_BFC0 ||
2614                     slot->slot == VARYING_SLOT_BFC1) {
2615                         shader->color_inputs |= (1 << inputs.num_inputs);
2616                 }
2617 
2618                 inputs.input_slots[inputs.num_inputs] = *slot;
2619                 inputs.num_inputs++;
2620         }
2621         shader->num_inputs = inputs.num_inputs;
2622 
2623         /* Add our set of inputs to the set of all inputs seen.  This way, we
2624          * can have a single pointer that identifies an FS inputs set,
2625          * allowing VS to avoid recompiling when the FS is recompiled (or a
2626          * new one is bound using separate shader objects) but the inputs
2627          * don't change.
2628          */
2629         struct set_entry *entry = _mesa_set_search(vc4->fs_inputs_set, &inputs);
2630         if (entry) {
2631                 shader->fs_inputs = entry->key;
2632                 ralloc_free(inputs.input_slots);
2633         } else {
2634                 struct vc4_fs_inputs *alloc_inputs;
2635 
2636                 alloc_inputs = rzalloc(vc4->fs_inputs_set, struct vc4_fs_inputs);
2637                 memcpy(alloc_inputs, &inputs, sizeof(inputs));
2638                 ralloc_steal(alloc_inputs, inputs.input_slots);
2639                 _mesa_set_add(vc4->fs_inputs_set, alloc_inputs);
2640 
2641                 shader->fs_inputs = alloc_inputs;
2642         }
2643 }
2644 
2645 static struct vc4_compiled_shader *
vc4_get_compiled_shader(struct vc4_context * vc4,enum qstage stage,struct vc4_key * key)2646 vc4_get_compiled_shader(struct vc4_context *vc4, enum qstage stage,
2647                         struct vc4_key *key)
2648 {
2649         struct hash_table *ht;
2650         uint32_t key_size;
2651         bool try_threading;
2652 
2653         if (stage == QSTAGE_FRAG) {
2654                 ht = vc4->fs_cache;
2655                 key_size = sizeof(struct vc4_fs_key);
2656                 try_threading = vc4->screen->has_threaded_fs;
2657         } else {
2658                 ht = vc4->vs_cache;
2659                 key_size = sizeof(struct vc4_vs_key);
2660                 try_threading = false;
2661         }
2662 
2663         struct vc4_compiled_shader *shader;
2664         struct hash_entry *entry = _mesa_hash_table_search(ht, key);
2665         if (entry)
2666                 return entry->data;
2667 
2668         struct vc4_compile *c = vc4_shader_ntq(vc4, stage, key, try_threading);
2669         /* If the FS failed to compile threaded, fall back to single threaded. */
2670         if (try_threading && c->failed) {
2671                 qir_compile_destroy(c);
2672                 c = vc4_shader_ntq(vc4, stage, key, false);
2673         }
2674 
2675         shader = rzalloc(NULL, struct vc4_compiled_shader);
2676 
2677         shader->program_id = vc4->next_compiled_program_id++;
2678         if (stage == QSTAGE_FRAG) {
2679                 vc4_setup_compiled_fs_inputs(vc4, c, shader);
2680 
2681                 /* Note: the temporary clone in c->s has been freed. */
2682                 nir_shader *orig_shader = key->shader_state->base.ir.nir;
2683                 if (orig_shader->info.outputs_written & (1 << FRAG_RESULT_DEPTH))
2684                         shader->disable_early_z = true;
2685         } else {
2686                 shader->num_inputs = c->num_inputs;
2687 
2688                 shader->vattr_offsets[0] = 0;
2689                 for (int i = 0; i < 8; i++) {
2690                         shader->vattr_offsets[i + 1] =
2691                                 shader->vattr_offsets[i] + c->vattr_sizes[i];
2692 
2693                         if (c->vattr_sizes[i])
2694                                 shader->vattrs_live |= (1 << i);
2695                 }
2696         }
2697 
2698         shader->failed = c->failed;
2699         if (c->failed) {
2700                 shader->failed = true;
2701         } else {
2702                 copy_uniform_state_to_shader(shader, c);
2703                 shader->bo = vc4_bo_alloc_shader(vc4->screen, c->qpu_insts,
2704                                                  c->qpu_inst_count *
2705                                                  sizeof(uint64_t));
2706         }
2707 
2708         shader->fs_threaded = c->fs_threaded;
2709 
2710         qir_compile_destroy(c);
2711 
2712         struct vc4_key *dup_key;
2713         dup_key = rzalloc_size(shader, key_size); /* TODO: don't use rzalloc */
2714         memcpy(dup_key, key, key_size);
2715         _mesa_hash_table_insert(ht, dup_key, shader);
2716 
2717         return shader;
2718 }
2719 
2720 static void
vc4_setup_shared_key(struct vc4_context * vc4,struct vc4_key * key,struct vc4_texture_stateobj * texstate)2721 vc4_setup_shared_key(struct vc4_context *vc4, struct vc4_key *key,
2722                      struct vc4_texture_stateobj *texstate)
2723 {
2724         for (int i = 0; i < texstate->num_textures; i++) {
2725                 struct pipe_sampler_view *sampler = texstate->textures[i];
2726                 struct vc4_sampler_view *vc4_sampler = vc4_sampler_view(sampler);
2727                 struct pipe_sampler_state *sampler_state =
2728                         texstate->samplers[i];
2729 
2730                 if (!sampler)
2731                         continue;
2732 
2733                 key->tex[i].format = sampler->format;
2734                 key->tex[i].swizzle[0] = sampler->swizzle_r;
2735                 key->tex[i].swizzle[1] = sampler->swizzle_g;
2736                 key->tex[i].swizzle[2] = sampler->swizzle_b;
2737                 key->tex[i].swizzle[3] = sampler->swizzle_a;
2738 
2739                 if (sampler->texture->nr_samples > 1) {
2740                         key->tex[i].msaa_width = sampler->texture->width0;
2741                         key->tex[i].msaa_height = sampler->texture->height0;
2742                 } else if (sampler){
2743                         key->tex[i].compare_mode = sampler_state->compare_mode;
2744                         key->tex[i].compare_func = sampler_state->compare_func;
2745                         key->tex[i].wrap_s = sampler_state->wrap_s;
2746                         key->tex[i].wrap_t = sampler_state->wrap_t;
2747                         key->tex[i].force_first_level =
2748                                 vc4_sampler->force_first_level;
2749                 }
2750         }
2751 
2752         key->ucp_enables = vc4->rasterizer->base.clip_plane_enable;
2753 }
2754 
2755 static void
vc4_update_compiled_fs(struct vc4_context * vc4,uint8_t prim_mode)2756 vc4_update_compiled_fs(struct vc4_context *vc4, uint8_t prim_mode)
2757 {
2758         struct vc4_job *job = vc4->job;
2759         struct vc4_fs_key local_key;
2760         struct vc4_fs_key *key = &local_key;
2761 
2762         if (!(vc4->dirty & (VC4_DIRTY_PRIM_MODE |
2763                             VC4_DIRTY_BLEND |
2764                             VC4_DIRTY_FRAMEBUFFER |
2765                             VC4_DIRTY_ZSA |
2766                             VC4_DIRTY_RASTERIZER |
2767                             VC4_DIRTY_SAMPLE_MASK |
2768                             VC4_DIRTY_FRAGTEX |
2769                             VC4_DIRTY_UNCOMPILED_FS |
2770                             VC4_DIRTY_UBO_1_SIZE))) {
2771                 return;
2772         }
2773 
2774         memset(key, 0, sizeof(*key));
2775         vc4_setup_shared_key(vc4, &key->base, &vc4->fragtex);
2776         key->base.shader_state = vc4->prog.bind_fs;
2777         key->is_points = (prim_mode == MESA_PRIM_POINTS);
2778         key->is_lines = (prim_mode >= MESA_PRIM_LINES &&
2779                          prim_mode <= MESA_PRIM_LINE_STRIP);
2780         key->blend = vc4->blend->rt[0];
2781         if (vc4->blend->logicop_enable) {
2782                 key->logicop_func = vc4->blend->logicop_func;
2783         } else {
2784                 key->logicop_func = PIPE_LOGICOP_COPY;
2785         }
2786         if (job->msaa) {
2787                 key->msaa = vc4->rasterizer->base.multisample;
2788                 key->sample_coverage = (vc4->sample_mask != (1 << VC4_MAX_SAMPLES) - 1);
2789                 key->sample_alpha_to_coverage = vc4->blend->alpha_to_coverage;
2790                 key->sample_alpha_to_one = vc4->blend->alpha_to_one;
2791         }
2792 
2793         if (vc4->framebuffer.cbufs[0])
2794                 key->color_format = vc4->framebuffer.cbufs[0]->format;
2795 
2796         key->stencil_enabled = vc4->zsa->stencil_uniforms[0] != 0;
2797         key->stencil_twoside = vc4->zsa->stencil_uniforms[1] != 0;
2798         key->stencil_full_writemasks = vc4->zsa->stencil_uniforms[2] != 0;
2799         key->depth_enabled = (vc4->zsa->base.depth_enabled ||
2800                               key->stencil_enabled);
2801 
2802         if (key->is_points) {
2803                 key->point_sprite_mask =
2804                         vc4->rasterizer->base.sprite_coord_enable;
2805                 key->point_coord_upper_left =
2806                         (vc4->rasterizer->base.sprite_coord_mode ==
2807                          PIPE_SPRITE_COORD_UPPER_LEFT);
2808         }
2809 
2810         key->ubo_1_size = vc4->constbuf[PIPE_SHADER_FRAGMENT].cb[1].buffer_size;
2811 
2812         struct vc4_compiled_shader *old_fs = vc4->prog.fs;
2813         vc4->prog.fs = vc4_get_compiled_shader(vc4, QSTAGE_FRAG, &key->base);
2814         if (vc4->prog.fs == old_fs)
2815                 return;
2816 
2817         vc4->dirty |= VC4_DIRTY_COMPILED_FS;
2818 
2819         if (vc4->rasterizer->base.flatshade &&
2820             (!old_fs || vc4->prog.fs->color_inputs != old_fs->color_inputs)) {
2821                 vc4->dirty |= VC4_DIRTY_FLAT_SHADE_FLAGS;
2822         }
2823 
2824         if (!old_fs || vc4->prog.fs->fs_inputs != old_fs->fs_inputs)
2825                 vc4->dirty |= VC4_DIRTY_FS_INPUTS;
2826 }
2827 
2828 static void
vc4_update_compiled_vs(struct vc4_context * vc4,uint8_t prim_mode)2829 vc4_update_compiled_vs(struct vc4_context *vc4, uint8_t prim_mode)
2830 {
2831         struct vc4_vs_key local_key;
2832         struct vc4_vs_key *key = &local_key;
2833 
2834         if (!(vc4->dirty & (VC4_DIRTY_PRIM_MODE |
2835                             VC4_DIRTY_RASTERIZER |
2836                             VC4_DIRTY_VERTTEX |
2837                             VC4_DIRTY_VTXSTATE |
2838                             VC4_DIRTY_UNCOMPILED_VS |
2839                             VC4_DIRTY_FS_INPUTS))) {
2840                 return;
2841         }
2842 
2843         memset(key, 0, sizeof(*key));
2844         vc4_setup_shared_key(vc4, &key->base, &vc4->verttex);
2845         key->base.shader_state = vc4->prog.bind_vs;
2846         key->fs_inputs = vc4->prog.fs->fs_inputs;
2847 
2848         for (int i = 0; i < ARRAY_SIZE(key->attr_formats); i++)
2849                 key->attr_formats[i] = vc4->vtx->pipe[i].src_format;
2850 
2851         key->per_vertex_point_size =
2852                 (prim_mode == MESA_PRIM_POINTS &&
2853                  vc4->rasterizer->base.point_size_per_vertex);
2854 
2855         struct vc4_compiled_shader *vs =
2856                 vc4_get_compiled_shader(vc4, QSTAGE_VERT, &key->base);
2857         if (vs != vc4->prog.vs) {
2858                 vc4->prog.vs = vs;
2859                 vc4->dirty |= VC4_DIRTY_COMPILED_VS;
2860         }
2861 
2862         key->is_coord = true;
2863         /* Coord shaders don't care what the FS inputs are. */
2864         key->fs_inputs = NULL;
2865         struct vc4_compiled_shader *cs =
2866                 vc4_get_compiled_shader(vc4, QSTAGE_COORD, &key->base);
2867         if (cs != vc4->prog.cs) {
2868                 vc4->prog.cs = cs;
2869                 vc4->dirty |= VC4_DIRTY_COMPILED_CS;
2870         }
2871 }
2872 
2873 bool
vc4_update_compiled_shaders(struct vc4_context * vc4,uint8_t prim_mode)2874 vc4_update_compiled_shaders(struct vc4_context *vc4, uint8_t prim_mode)
2875 {
2876         vc4_update_compiled_fs(vc4, prim_mode);
2877         vc4_update_compiled_vs(vc4, prim_mode);
2878 
2879         return !(vc4->prog.cs->failed ||
2880                  vc4->prog.vs->failed ||
2881                  vc4->prog.fs->failed);
2882 }
2883 
2884 static uint32_t
fs_cache_hash(const void * key)2885 fs_cache_hash(const void *key)
2886 {
2887         return _mesa_hash_data(key, sizeof(struct vc4_fs_key));
2888 }
2889 
2890 static uint32_t
vs_cache_hash(const void * key)2891 vs_cache_hash(const void *key)
2892 {
2893         return _mesa_hash_data(key, sizeof(struct vc4_vs_key));
2894 }
2895 
2896 static bool
fs_cache_compare(const void * key1,const void * key2)2897 fs_cache_compare(const void *key1, const void *key2)
2898 {
2899         return memcmp(key1, key2, sizeof(struct vc4_fs_key)) == 0;
2900 }
2901 
2902 static bool
vs_cache_compare(const void * key1,const void * key2)2903 vs_cache_compare(const void *key1, const void *key2)
2904 {
2905         return memcmp(key1, key2, sizeof(struct vc4_vs_key)) == 0;
2906 }
2907 
2908 static uint32_t
fs_inputs_hash(const void * key)2909 fs_inputs_hash(const void *key)
2910 {
2911         const struct vc4_fs_inputs *inputs = key;
2912 
2913         return _mesa_hash_data(inputs->input_slots,
2914                                sizeof(*inputs->input_slots) *
2915                                inputs->num_inputs);
2916 }
2917 
2918 static bool
fs_inputs_compare(const void * key1,const void * key2)2919 fs_inputs_compare(const void *key1, const void *key2)
2920 {
2921         const struct vc4_fs_inputs *inputs1 = key1;
2922         const struct vc4_fs_inputs *inputs2 = key2;
2923 
2924         return (inputs1->num_inputs == inputs2->num_inputs &&
2925                 memcmp(inputs1->input_slots,
2926                        inputs2->input_slots,
2927                        sizeof(*inputs1->input_slots) *
2928                        inputs1->num_inputs) == 0);
2929 }
2930 
2931 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)2932 delete_from_cache_if_matches(struct hash_table *ht,
2933                              struct vc4_compiled_shader **last_compile,
2934                              struct hash_entry *entry,
2935                              struct vc4_uncompiled_shader *so)
2936 {
2937         const struct vc4_key *key = entry->key;
2938 
2939         if (key->shader_state == so) {
2940                 struct vc4_compiled_shader *shader = entry->data;
2941                 _mesa_hash_table_remove(ht, entry);
2942                 vc4_bo_unreference(&shader->bo);
2943 
2944                 if (shader == *last_compile)
2945                         *last_compile = NULL;
2946 
2947                 ralloc_free(shader);
2948         }
2949 }
2950 
2951 static void
vc4_shader_state_delete(struct pipe_context * pctx,void * hwcso)2952 vc4_shader_state_delete(struct pipe_context *pctx, void *hwcso)
2953 {
2954         struct vc4_context *vc4 = vc4_context(pctx);
2955         struct vc4_uncompiled_shader *so = hwcso;
2956 
2957         hash_table_foreach(vc4->fs_cache, entry) {
2958                 delete_from_cache_if_matches(vc4->fs_cache, &vc4->prog.fs,
2959                                              entry, so);
2960         }
2961         hash_table_foreach(vc4->vs_cache, entry) {
2962                 delete_from_cache_if_matches(vc4->vs_cache, &vc4->prog.vs,
2963                                              entry, so);
2964         }
2965 
2966         ralloc_free(so->base.ir.nir);
2967         free(so);
2968 }
2969 
2970 static void
vc4_fp_state_bind(struct pipe_context * pctx,void * hwcso)2971 vc4_fp_state_bind(struct pipe_context *pctx, void *hwcso)
2972 {
2973         struct vc4_context *vc4 = vc4_context(pctx);
2974         vc4->prog.bind_fs = hwcso;
2975         vc4->dirty |= VC4_DIRTY_UNCOMPILED_FS;
2976 }
2977 
2978 static void
vc4_vp_state_bind(struct pipe_context * pctx,void * hwcso)2979 vc4_vp_state_bind(struct pipe_context *pctx, void *hwcso)
2980 {
2981         struct vc4_context *vc4 = vc4_context(pctx);
2982         vc4->prog.bind_vs = hwcso;
2983         vc4->dirty |= VC4_DIRTY_UNCOMPILED_VS;
2984 }
2985 
2986 void
vc4_program_init(struct pipe_context * pctx)2987 vc4_program_init(struct pipe_context *pctx)
2988 {
2989         struct vc4_context *vc4 = vc4_context(pctx);
2990 
2991         pctx->create_vs_state = vc4_shader_state_create;
2992         pctx->delete_vs_state = vc4_shader_state_delete;
2993 
2994         pctx->create_fs_state = vc4_shader_state_create;
2995         pctx->delete_fs_state = vc4_shader_state_delete;
2996 
2997         pctx->bind_fs_state = vc4_fp_state_bind;
2998         pctx->bind_vs_state = vc4_vp_state_bind;
2999 
3000         vc4->fs_cache = _mesa_hash_table_create(pctx, fs_cache_hash,
3001                                                 fs_cache_compare);
3002         vc4->vs_cache = _mesa_hash_table_create(pctx, vs_cache_hash,
3003                                                 vs_cache_compare);
3004         vc4->fs_inputs_set = _mesa_set_create(pctx, fs_inputs_hash,
3005                                               fs_inputs_compare);
3006 }
3007 
3008 void
vc4_program_fini(struct pipe_context * pctx)3009 vc4_program_fini(struct pipe_context *pctx)
3010 {
3011         struct vc4_context *vc4 = vc4_context(pctx);
3012 
3013         hash_table_foreach(vc4->fs_cache, entry) {
3014                 struct vc4_compiled_shader *shader = entry->data;
3015                 vc4_bo_unreference(&shader->bo);
3016                 ralloc_free(shader);
3017                 _mesa_hash_table_remove(vc4->fs_cache, entry);
3018         }
3019 
3020         hash_table_foreach(vc4->vs_cache, entry) {
3021                 struct vc4_compiled_shader *shader = entry->data;
3022                 vc4_bo_unreference(&shader->bo);
3023                 ralloc_free(shader);
3024                 _mesa_hash_table_remove(vc4->vs_cache, entry);
3025         }
3026 }
3027