nir_lower_io.c revision 01e04c3f 
1/*
2 * Copyright © 2014 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 *
23 * Authors:
24 *    Connor Abbott (cwabbott0@gmail.com)
25 *    Jason Ekstrand (jason@jlekstrand.net)
26 *
27 */
28
29/*
30 * This lowering pass converts references to input/output variables with
31 * loads/stores to actual input/output intrinsics.
32 */
33
34#include "nir.h"
35#include "nir_builder.h"
36#include "nir_deref.h"
37
38struct lower_io_state {
39   void *dead_ctx;
40   nir_builder builder;
41   int (*type_size)(const struct glsl_type *type);
42   nir_variable_mode modes;
43   nir_lower_io_options options;
44};
45
46void
47nir_assign_var_locations(struct exec_list *var_list, unsigned *size,
48                         int (*type_size)(const struct glsl_type *))
49{
50   unsigned location = 0;
51
52   nir_foreach_variable(var, var_list) {
53      /*
54       * UBOs have their own address spaces, so don't count them towards the
55       * number of global uniforms
56       */
57      if ((var->data.mode == nir_var_uniform || var->data.mode == nir_var_shader_storage) &&
58          var->interface_type != NULL)
59         continue;
60
61      var->data.driver_location = location;
62      location += type_size(var->type);
63   }
64
65   *size = location;
66}
67
68/**
69 * Return true if the given variable is a per-vertex input/output array.
70 * (such as geometry shader inputs).
71 */
72bool
73nir_is_per_vertex_io(const nir_variable *var, gl_shader_stage stage)
74{
75   if (var->data.patch || !glsl_type_is_array(var->type))
76      return false;
77
78   if (var->data.mode == nir_var_shader_in)
79      return stage == MESA_SHADER_GEOMETRY ||
80             stage == MESA_SHADER_TESS_CTRL ||
81             stage == MESA_SHADER_TESS_EVAL;
82
83   if (var->data.mode == nir_var_shader_out)
84      return stage == MESA_SHADER_TESS_CTRL;
85
86   return false;
87}
88
89static nir_ssa_def *
90get_io_offset(nir_builder *b, nir_deref_instr *deref,
91              nir_ssa_def **vertex_index,
92              int (*type_size)(const struct glsl_type *),
93              unsigned *component)
94{
95   nir_deref_path path;
96   nir_deref_path_init(&path, deref, NULL);
97
98   assert(path.path[0]->deref_type == nir_deref_type_var);
99   nir_deref_instr **p = &path.path[1];
100
101   /* For per-vertex input arrays (i.e. geometry shader inputs), keep the
102    * outermost array index separate.  Process the rest normally.
103    */
104   if (vertex_index != NULL) {
105      assert((*p)->deref_type == nir_deref_type_array);
106      *vertex_index = nir_ssa_for_src(b, (*p)->arr.index, 1);
107      p++;
108   }
109
110   if (path.path[0]->var->data.compact) {
111      assert((*p)->deref_type == nir_deref_type_array);
112      assert(glsl_type_is_scalar((*p)->type));
113
114      /* We always lower indirect dereferences for "compact" array vars. */
115      const unsigned index = nir_src_as_uint((*p)->arr.index);
116      const unsigned total_offset = *component + index;
117      const unsigned slot_offset = total_offset / 4;
118      *component = total_offset % 4;
119      return nir_imm_int(b, type_size(glsl_vec4_type()) * slot_offset);
120   }
121
122   /* Just emit code and let constant-folding go to town */
123   nir_ssa_def *offset = nir_imm_int(b, 0);
124
125   for (; *p; p++) {
126      if ((*p)->deref_type == nir_deref_type_array) {
127         unsigned size = type_size((*p)->type);
128
129         nir_ssa_def *mul =
130            nir_imul(b, nir_imm_int(b, size),
131                     nir_ssa_for_src(b, (*p)->arr.index, 1));
132
133         offset = nir_iadd(b, offset, mul);
134      } else if ((*p)->deref_type == nir_deref_type_struct) {
135         /* p starts at path[1], so this is safe */
136         nir_deref_instr *parent = *(p - 1);
137
138         unsigned field_offset = 0;
139         for (unsigned i = 0; i < (*p)->strct.index; i++) {
140            field_offset += type_size(glsl_get_struct_field(parent->type, i));
141         }
142         offset = nir_iadd(b, offset, nir_imm_int(b, field_offset));
143      } else {
144         unreachable("Unsupported deref type");
145      }
146   }
147
148   nir_deref_path_finish(&path);
149
150   return offset;
151}
152
153static nir_intrinsic_instr *
154lower_load(nir_intrinsic_instr *intrin, struct lower_io_state *state,
155           nir_ssa_def *vertex_index, nir_variable *var, nir_ssa_def *offset,
156           unsigned component)
157{
158   const nir_shader *nir = state->builder.shader;
159   nir_variable_mode mode = var->data.mode;
160   nir_ssa_def *barycentric = NULL;
161
162   nir_intrinsic_op op;
163   switch (mode) {
164   case nir_var_shader_in:
165      if (nir->info.stage == MESA_SHADER_FRAGMENT &&
166          nir->options->use_interpolated_input_intrinsics &&
167          var->data.interpolation != INTERP_MODE_FLAT) {
168         assert(vertex_index == NULL);
169
170         nir_intrinsic_op bary_op;
171         if (var->data.sample ||
172             (state->options & nir_lower_io_force_sample_interpolation))
173            bary_op = nir_intrinsic_load_barycentric_sample;
174         else if (var->data.centroid)
175            bary_op = nir_intrinsic_load_barycentric_centroid;
176         else
177            bary_op = nir_intrinsic_load_barycentric_pixel;
178
179         barycentric = nir_load_barycentric(&state->builder, bary_op,
180                                            var->data.interpolation);
181         op = nir_intrinsic_load_interpolated_input;
182      } else {
183         op = vertex_index ? nir_intrinsic_load_per_vertex_input :
184                             nir_intrinsic_load_input;
185      }
186      break;
187   case nir_var_shader_out:
188      op = vertex_index ? nir_intrinsic_load_per_vertex_output :
189                          nir_intrinsic_load_output;
190      break;
191   case nir_var_uniform:
192      op = nir_intrinsic_load_uniform;
193      break;
194   case nir_var_shared:
195      op = nir_intrinsic_load_shared;
196      break;
197   default:
198      unreachable("Unknown variable mode");
199   }
200
201   nir_intrinsic_instr *load =
202      nir_intrinsic_instr_create(state->builder.shader, op);
203   load->num_components = intrin->num_components;
204
205   nir_intrinsic_set_base(load, var->data.driver_location);
206   if (mode == nir_var_shader_in || mode == nir_var_shader_out)
207      nir_intrinsic_set_component(load, component);
208
209   if (load->intrinsic == nir_intrinsic_load_uniform)
210      nir_intrinsic_set_range(load, state->type_size(var->type));
211
212   if (vertex_index) {
213      load->src[0] = nir_src_for_ssa(vertex_index);
214      load->src[1] = nir_src_for_ssa(offset);
215   } else if (barycentric) {
216      load->src[0] = nir_src_for_ssa(barycentric);
217      load->src[1] = nir_src_for_ssa(offset);
218   } else {
219      load->src[0] = nir_src_for_ssa(offset);
220   }
221
222   return load;
223}
224
225static nir_intrinsic_instr *
226lower_store(nir_intrinsic_instr *intrin, struct lower_io_state *state,
227            nir_ssa_def *vertex_index, nir_variable *var, nir_ssa_def *offset,
228            unsigned component)
229{
230   nir_variable_mode mode = var->data.mode;
231
232   nir_intrinsic_op op;
233   if (mode == nir_var_shared) {
234      op = nir_intrinsic_store_shared;
235   } else {
236      assert(mode == nir_var_shader_out);
237      op = vertex_index ? nir_intrinsic_store_per_vertex_output :
238                          nir_intrinsic_store_output;
239   }
240
241   nir_intrinsic_instr *store =
242      nir_intrinsic_instr_create(state->builder.shader, op);
243   store->num_components = intrin->num_components;
244
245   nir_src_copy(&store->src[0], &intrin->src[1], store);
246
247   nir_intrinsic_set_base(store, var->data.driver_location);
248
249   if (mode == nir_var_shader_out)
250      nir_intrinsic_set_component(store, component);
251
252   nir_intrinsic_set_write_mask(store, nir_intrinsic_write_mask(intrin));
253
254   if (vertex_index)
255      store->src[1] = nir_src_for_ssa(vertex_index);
256
257   store->src[vertex_index ? 2 : 1] = nir_src_for_ssa(offset);
258
259   return store;
260}
261
262static nir_intrinsic_instr *
263lower_atomic(nir_intrinsic_instr *intrin, struct lower_io_state *state,
264             nir_variable *var, nir_ssa_def *offset)
265{
266   assert(var->data.mode == nir_var_shared);
267
268   nir_intrinsic_op op;
269   switch (intrin->intrinsic) {
270#define OP(O) case nir_intrinsic_deref_##O: op = nir_intrinsic_shared_##O; break;
271   OP(atomic_exchange)
272   OP(atomic_comp_swap)
273   OP(atomic_add)
274   OP(atomic_imin)
275   OP(atomic_umin)
276   OP(atomic_imax)
277   OP(atomic_umax)
278   OP(atomic_and)
279   OP(atomic_or)
280   OP(atomic_xor)
281   OP(atomic_fadd)
282   OP(atomic_fmin)
283   OP(atomic_fmax)
284   OP(atomic_fcomp_swap)
285#undef OP
286   default:
287      unreachable("Invalid atomic");
288   }
289
290   nir_intrinsic_instr *atomic =
291      nir_intrinsic_instr_create(state->builder.shader, op);
292
293   nir_intrinsic_set_base(atomic, var->data.driver_location);
294
295   atomic->src[0] = nir_src_for_ssa(offset);
296   assert(nir_intrinsic_infos[intrin->intrinsic].num_srcs ==
297          nir_intrinsic_infos[op].num_srcs);
298   for (unsigned i = 1; i < nir_intrinsic_infos[op].num_srcs; i++) {
299      nir_src_copy(&atomic->src[i], &intrin->src[i], atomic);
300   }
301
302   return atomic;
303}
304
305static nir_intrinsic_instr *
306lower_interpolate_at(nir_intrinsic_instr *intrin, struct lower_io_state *state,
307                     nir_variable *var, nir_ssa_def *offset, unsigned component)
308{
309   assert(var->data.mode == nir_var_shader_in);
310
311   /* Ignore interpolateAt() for flat variables - flat is flat. */
312   if (var->data.interpolation == INTERP_MODE_FLAT)
313      return lower_load(intrin, state, NULL, var, offset, component);
314
315   nir_intrinsic_op bary_op;
316   switch (intrin->intrinsic) {
317   case nir_intrinsic_interp_deref_at_centroid:
318      bary_op = (state->options & nir_lower_io_force_sample_interpolation) ?
319                nir_intrinsic_load_barycentric_sample :
320                nir_intrinsic_load_barycentric_centroid;
321      break;
322   case nir_intrinsic_interp_deref_at_sample:
323      bary_op = nir_intrinsic_load_barycentric_at_sample;
324      break;
325   case nir_intrinsic_interp_deref_at_offset:
326      bary_op = nir_intrinsic_load_barycentric_at_offset;
327      break;
328   default:
329      unreachable("Bogus interpolateAt() intrinsic.");
330   }
331
332   nir_intrinsic_instr *bary_setup =
333      nir_intrinsic_instr_create(state->builder.shader, bary_op);
334
335   nir_ssa_dest_init(&bary_setup->instr, &bary_setup->dest, 2, 32, NULL);
336   nir_intrinsic_set_interp_mode(bary_setup, var->data.interpolation);
337
338   if (intrin->intrinsic == nir_intrinsic_interp_deref_at_sample ||
339       intrin->intrinsic == nir_intrinsic_interp_deref_at_offset)
340      nir_src_copy(&bary_setup->src[0], &intrin->src[1], bary_setup);
341
342   nir_builder_instr_insert(&state->builder, &bary_setup->instr);
343
344   nir_intrinsic_instr *load =
345      nir_intrinsic_instr_create(state->builder.shader,
346                                 nir_intrinsic_load_interpolated_input);
347   load->num_components = intrin->num_components;
348
349   nir_intrinsic_set_base(load, var->data.driver_location);
350   nir_intrinsic_set_component(load, component);
351
352   load->src[0] = nir_src_for_ssa(&bary_setup->dest.ssa);
353   load->src[1] = nir_src_for_ssa(offset);
354
355   return load;
356}
357
358static bool
359nir_lower_io_block(nir_block *block,
360                   struct lower_io_state *state)
361{
362   nir_builder *b = &state->builder;
363   const nir_shader_compiler_options *options = b->shader->options;
364   bool progress = false;
365
366   nir_foreach_instr_safe(instr, block) {
367      if (instr->type != nir_instr_type_intrinsic)
368         continue;
369
370      nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
371
372      switch (intrin->intrinsic) {
373      case nir_intrinsic_load_deref:
374      case nir_intrinsic_store_deref:
375      case nir_intrinsic_deref_atomic_add:
376      case nir_intrinsic_deref_atomic_imin:
377      case nir_intrinsic_deref_atomic_umin:
378      case nir_intrinsic_deref_atomic_imax:
379      case nir_intrinsic_deref_atomic_umax:
380      case nir_intrinsic_deref_atomic_and:
381      case nir_intrinsic_deref_atomic_or:
382      case nir_intrinsic_deref_atomic_xor:
383      case nir_intrinsic_deref_atomic_exchange:
384      case nir_intrinsic_deref_atomic_comp_swap:
385      case nir_intrinsic_deref_atomic_fadd:
386      case nir_intrinsic_deref_atomic_fmin:
387      case nir_intrinsic_deref_atomic_fmax:
388      case nir_intrinsic_deref_atomic_fcomp_swap:
389         /* We can lower the io for this nir instrinsic */
390         break;
391      case nir_intrinsic_interp_deref_at_centroid:
392      case nir_intrinsic_interp_deref_at_sample:
393      case nir_intrinsic_interp_deref_at_offset:
394         /* We can optionally lower these to load_interpolated_input */
395         if (options->use_interpolated_input_intrinsics)
396            break;
397      default:
398         /* We can't lower the io for this nir instrinsic, so skip it */
399         continue;
400      }
401
402      nir_deref_instr *deref = nir_src_as_deref(intrin->src[0]);
403
404      nir_variable *var = nir_deref_instr_get_variable(deref);
405      nir_variable_mode mode = var->data.mode;
406
407      if ((state->modes & mode) == 0)
408         continue;
409
410      if (mode != nir_var_shader_in &&
411          mode != nir_var_shader_out &&
412          mode != nir_var_shared &&
413          mode != nir_var_uniform)
414         continue;
415
416      b->cursor = nir_before_instr(instr);
417
418      const bool per_vertex = nir_is_per_vertex_io(var, b->shader->info.stage);
419
420      nir_ssa_def *offset;
421      nir_ssa_def *vertex_index = NULL;
422      unsigned component_offset = var->data.location_frac;
423
424      offset = get_io_offset(b, deref, per_vertex ? &vertex_index : NULL,
425                             state->type_size, &component_offset);
426
427      nir_intrinsic_instr *replacement;
428
429      switch (intrin->intrinsic) {
430      case nir_intrinsic_load_deref:
431         replacement = lower_load(intrin, state, vertex_index, var, offset,
432                                  component_offset);
433         break;
434
435      case nir_intrinsic_store_deref:
436         replacement = lower_store(intrin, state, vertex_index, var, offset,
437                                   component_offset);
438         break;
439
440      case nir_intrinsic_deref_atomic_add:
441      case nir_intrinsic_deref_atomic_imin:
442      case nir_intrinsic_deref_atomic_umin:
443      case nir_intrinsic_deref_atomic_imax:
444      case nir_intrinsic_deref_atomic_umax:
445      case nir_intrinsic_deref_atomic_and:
446      case nir_intrinsic_deref_atomic_or:
447      case nir_intrinsic_deref_atomic_xor:
448      case nir_intrinsic_deref_atomic_exchange:
449      case nir_intrinsic_deref_atomic_comp_swap:
450      case nir_intrinsic_deref_atomic_fadd:
451      case nir_intrinsic_deref_atomic_fmin:
452      case nir_intrinsic_deref_atomic_fmax:
453      case nir_intrinsic_deref_atomic_fcomp_swap:
454         assert(vertex_index == NULL);
455         replacement = lower_atomic(intrin, state, var, offset);
456         break;
457
458      case nir_intrinsic_interp_deref_at_centroid:
459      case nir_intrinsic_interp_deref_at_sample:
460      case nir_intrinsic_interp_deref_at_offset:
461         assert(vertex_index == NULL);
462         replacement = lower_interpolate_at(intrin, state, var, offset,
463                                            component_offset);
464         break;
465
466      default:
467         continue;
468      }
469
470      if (nir_intrinsic_infos[intrin->intrinsic].has_dest) {
471         if (intrin->dest.is_ssa) {
472            nir_ssa_dest_init(&replacement->instr, &replacement->dest,
473                              intrin->dest.ssa.num_components,
474                              intrin->dest.ssa.bit_size, NULL);
475            nir_ssa_def_rewrite_uses(&intrin->dest.ssa,
476                                     nir_src_for_ssa(&replacement->dest.ssa));
477         } else {
478            nir_dest_copy(&replacement->dest, &intrin->dest, &intrin->instr);
479         }
480      }
481
482      nir_instr_insert_before(&intrin->instr, &replacement->instr);
483      nir_instr_remove(&intrin->instr);
484      progress = true;
485   }
486
487   return progress;
488}
489
490static bool
491nir_lower_io_impl(nir_function_impl *impl,
492                  nir_variable_mode modes,
493                  int (*type_size)(const struct glsl_type *),
494                  nir_lower_io_options options)
495{
496   struct lower_io_state state;
497   bool progress = false;
498
499   nir_builder_init(&state.builder, impl);
500   state.dead_ctx = ralloc_context(NULL);
501   state.modes = modes;
502   state.type_size = type_size;
503   state.options = options;
504
505   nir_foreach_block(block, impl) {
506      progress |= nir_lower_io_block(block, &state);
507   }
508
509   ralloc_free(state.dead_ctx);
510
511   nir_metadata_preserve(impl, nir_metadata_block_index |
512                               nir_metadata_dominance);
513   return progress;
514}
515
516bool
517nir_lower_io(nir_shader *shader, nir_variable_mode modes,
518             int (*type_size)(const struct glsl_type *),
519             nir_lower_io_options options)
520{
521   bool progress = false;
522
523   nir_foreach_function(function, shader) {
524      if (function->impl) {
525         progress |= nir_lower_io_impl(function->impl, modes,
526                                       type_size, options);
527      }
528   }
529
530   return progress;
531}
532
533/**
534 * Return the offset source for a load/store intrinsic.
535 */
536nir_src *
537nir_get_io_offset_src(nir_intrinsic_instr *instr)
538{
539   switch (instr->intrinsic) {
540   case nir_intrinsic_load_input:
541   case nir_intrinsic_load_output:
542   case nir_intrinsic_load_uniform:
543      return &instr->src[0];
544   case nir_intrinsic_load_ubo:
545   case nir_intrinsic_load_ssbo:
546   case nir_intrinsic_load_per_vertex_input:
547   case nir_intrinsic_load_per_vertex_output:
548   case nir_intrinsic_load_interpolated_input:
549   case nir_intrinsic_store_output:
550      return &instr->src[1];
551   case nir_intrinsic_store_ssbo:
552   case nir_intrinsic_store_per_vertex_output:
553      return &instr->src[2];
554   default:
555      return NULL;
556   }
557}
558
559/**
560 * Return the vertex index source for a load/store per_vertex intrinsic.
561 */
562nir_src *
563nir_get_io_vertex_index_src(nir_intrinsic_instr *instr)
564{
565   switch (instr->intrinsic) {
566   case nir_intrinsic_load_per_vertex_input:
567   case nir_intrinsic_load_per_vertex_output:
568      return &instr->src[0];
569   case nir_intrinsic_store_per_vertex_output:
570      return &instr->src[1];
571   default:
572      return NULL;
573   }
574}
575