17ec681f3Smrg/*
27ec681f3Smrg * Copyright © 2019 Collabora, Ltd.
37ec681f3Smrg *
47ec681f3Smrg * Permission is hereby granted, free of charge, to any person obtaining a
57ec681f3Smrg * copy of this software and associated documentation files (the "Software"),
67ec681f3Smrg * to deal in the Software without restriction, including without limitation
77ec681f3Smrg * the rights to use, copy, modify, merge, publish, distribute, sublicense,
87ec681f3Smrg * and/or sell copies of the Software, and to permit persons to whom the
97ec681f3Smrg * Software is furnished to do so, subject to the following conditions:
107ec681f3Smrg *
117ec681f3Smrg * The above copyright notice and this permission notice (including the next
127ec681f3Smrg * paragraph) shall be included in all copies or substantial portions of the
137ec681f3Smrg * Software.
147ec681f3Smrg *
157ec681f3Smrg * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
167ec681f3Smrg * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
177ec681f3Smrg * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
187ec681f3Smrg * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
197ec681f3Smrg * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
207ec681f3Smrg * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
217ec681f3Smrg * IN THE SOFTWARE.
227ec681f3Smrg *
237ec681f3Smrg * Authors (Collabora):
247ec681f3Smrg *    Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
257ec681f3Smrg */
267ec681f3Smrg
277ec681f3Smrg#include "nir.h"
287ec681f3Smrg#include "nir_builder.h"
297ec681f3Smrg
307ec681f3Smrg/*
317ec681f3Smrg * Lowers SSBOs to globals, for hardware that lack native SSBO support. When
327ec681f3Smrg * lowering, *_ssbo_* instructions will become *_global_* instructions,
337ec681f3Smrg * augmented with load_ssbo_address.
347ec681f3Smrg *
357ec681f3Smrg * DOES NOT PERFORM BOUNDS CHECKING. DO NOT USE IN PRODUCTION ON UNTRUSTED
367ec681f3Smrg * CONTEXTS INCLUDING WEBGL 2.
377ec681f3Smrg */
387ec681f3Smrg
397ec681f3Smrgstatic nir_intrinsic_op
407ec681f3Smrglower_ssbo_op(nir_intrinsic_op op)
417ec681f3Smrg{
427ec681f3Smrg   switch (op) {
437ec681f3Smrg   case nir_intrinsic_load_ssbo:
447ec681f3Smrg      return nir_intrinsic_load_global;
457ec681f3Smrg
467ec681f3Smrg   case nir_intrinsic_store_ssbo:
477ec681f3Smrg      return nir_intrinsic_store_global;
487ec681f3Smrg
497ec681f3Smrg   case nir_intrinsic_ssbo_atomic_add:
507ec681f3Smrg      return nir_intrinsic_global_atomic_add;
517ec681f3Smrg   case nir_intrinsic_ssbo_atomic_imin:
527ec681f3Smrg      return nir_intrinsic_global_atomic_imin;
537ec681f3Smrg   case nir_intrinsic_ssbo_atomic_umin:
547ec681f3Smrg      return nir_intrinsic_global_atomic_umin;
557ec681f3Smrg   case nir_intrinsic_ssbo_atomic_imax:
567ec681f3Smrg      return nir_intrinsic_global_atomic_imax;
577ec681f3Smrg   case nir_intrinsic_ssbo_atomic_umax:
587ec681f3Smrg      return nir_intrinsic_global_atomic_umax;
597ec681f3Smrg   case nir_intrinsic_ssbo_atomic_and:
607ec681f3Smrg      return nir_intrinsic_global_atomic_and;
617ec681f3Smrg   case nir_intrinsic_ssbo_atomic_or:
627ec681f3Smrg      return nir_intrinsic_global_atomic_or;
637ec681f3Smrg   case nir_intrinsic_ssbo_atomic_xor:
647ec681f3Smrg      return nir_intrinsic_global_atomic_xor;
657ec681f3Smrg   case nir_intrinsic_ssbo_atomic_exchange:
667ec681f3Smrg      return nir_intrinsic_global_atomic_exchange;
677ec681f3Smrg   case nir_intrinsic_ssbo_atomic_comp_swap:
687ec681f3Smrg      return nir_intrinsic_global_atomic_comp_swap;
697ec681f3Smrg
707ec681f3Smrg   case nir_intrinsic_ssbo_atomic_fadd:
717ec681f3Smrg      return nir_intrinsic_global_atomic_fadd;
727ec681f3Smrg   case nir_intrinsic_ssbo_atomic_fmin:
737ec681f3Smrg      return nir_intrinsic_global_atomic_fmin;
747ec681f3Smrg   case nir_intrinsic_ssbo_atomic_fmax:
757ec681f3Smrg      return nir_intrinsic_global_atomic_fmax;
767ec681f3Smrg   case nir_intrinsic_ssbo_atomic_fcomp_swap:
777ec681f3Smrg      return nir_intrinsic_global_atomic_fcomp_swap;
787ec681f3Smrg
797ec681f3Smrg   default:
807ec681f3Smrg      unreachable("Invalid SSBO op");
817ec681f3Smrg   }
827ec681f3Smrg}
837ec681f3Smrg
847ec681f3Smrg/* Like SSBO property sysvals, though SSBO index may be indirect. C.f.
857ec681f3Smrg * nir_load_system_value */
867ec681f3Smrg
877ec681f3Smrgstatic inline nir_ssa_def *
887ec681f3Smrgnir_load_ssbo_prop(nir_builder *b, nir_intrinsic_op op,
897ec681f3Smrg      nir_src *idx, unsigned bitsize)
907ec681f3Smrg{
917ec681f3Smrg   nir_intrinsic_instr *load = nir_intrinsic_instr_create(b->shader, op);
927ec681f3Smrg   load->num_components = 1;
937ec681f3Smrg   nir_src_copy(&load->src[0], idx);
947ec681f3Smrg   nir_ssa_dest_init(&load->instr, &load->dest, 1, bitsize, NULL);
957ec681f3Smrg   nir_builder_instr_insert(b, &load->instr);
967ec681f3Smrg   return &load->dest.ssa;
977ec681f3Smrg}
987ec681f3Smrg
997ec681f3Smrg#define nir_ssbo_prop(b, prop, index, bitsize) \
1007ec681f3Smrg   nir_load_ssbo_prop(b, nir_intrinsic_##prop, index, bitsize)
1017ec681f3Smrg
1027ec681f3Smrgstatic nir_ssa_def *
1037ec681f3Smrglower_ssbo_instr(nir_builder *b, nir_intrinsic_instr *intr)
1047ec681f3Smrg{
1057ec681f3Smrg   nir_intrinsic_op op = lower_ssbo_op(intr->intrinsic);
1067ec681f3Smrg   bool is_store = op == nir_intrinsic_store_global;
1077ec681f3Smrg   bool is_atomic = !is_store && op != nir_intrinsic_load_global;
1087ec681f3Smrg
1097ec681f3Smrg   /* We have to calculate the address:
1107ec681f3Smrg    *
1117ec681f3Smrg    * &(SSBO[offset]) = &SSBO + offset
1127ec681f3Smrg    */
1137ec681f3Smrg
1147ec681f3Smrg   nir_src index = intr->src[is_store ? 1 : 0];
1157ec681f3Smrg   nir_src *offset_src = nir_get_io_offset_src(intr);
1167ec681f3Smrg   nir_ssa_def *offset = nir_ssa_for_src(b, *offset_src, 1);
1177ec681f3Smrg
1187ec681f3Smrg   nir_ssa_def *address =
1197ec681f3Smrg      nir_iadd(b,
1207ec681f3Smrg            nir_ssbo_prop(b, load_ssbo_address, &index, 64),
1217ec681f3Smrg            nir_u2u64(b, offset));
1227ec681f3Smrg
1237ec681f3Smrg   /* Create the replacement intrinsic */
1247ec681f3Smrg
1257ec681f3Smrg   nir_intrinsic_instr *global =
1267ec681f3Smrg      nir_intrinsic_instr_create(b->shader, op);
1277ec681f3Smrg
1287ec681f3Smrg   global->num_components = intr->num_components;
1297ec681f3Smrg   global->src[is_store ? 1 : 0] = nir_src_for_ssa(address);
1307ec681f3Smrg
1317ec681f3Smrg   if (!is_atomic) {
1327ec681f3Smrg      nir_intrinsic_set_align_mul(global, nir_intrinsic_align_mul(intr));
1337ec681f3Smrg      nir_intrinsic_set_align_offset(global, nir_intrinsic_align_offset(intr));
1347ec681f3Smrg   }
1357ec681f3Smrg
1367ec681f3Smrg   if (is_store) {
1377ec681f3Smrg      nir_src_copy(&global->src[0], &intr->src[0]);
1387ec681f3Smrg      nir_intrinsic_set_write_mask(global, nir_intrinsic_write_mask(intr));
1397ec681f3Smrg   } else {
1407ec681f3Smrg      nir_ssa_dest_init(&global->instr, &global->dest,
1417ec681f3Smrg                        intr->dest.ssa.num_components,
1427ec681f3Smrg                        intr->dest.ssa.bit_size, NULL);
1437ec681f3Smrg
1447ec681f3Smrg      if (is_atomic) {
1457ec681f3Smrg         nir_src_copy(&global->src[1], &intr->src[2]);
1467ec681f3Smrg         if (nir_intrinsic_infos[op].num_srcs > 2)
1477ec681f3Smrg            nir_src_copy(&global->src[2], &intr->src[3]);
1487ec681f3Smrg      }
1497ec681f3Smrg   }
1507ec681f3Smrg
1517ec681f3Smrg   nir_builder_instr_insert(b, &global->instr);
1527ec681f3Smrg   return is_store ? NULL : &global->dest.ssa;
1537ec681f3Smrg}
1547ec681f3Smrg
1557ec681f3Smrgstatic bool
1567ec681f3Smrgshould_lower_ssbo_instr(const nir_instr *instr)
1577ec681f3Smrg{
1587ec681f3Smrg   if (instr->type != nir_instr_type_intrinsic)
1597ec681f3Smrg      return false;
1607ec681f3Smrg
1617ec681f3Smrg   const nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
1627ec681f3Smrg
1637ec681f3Smrg   switch (intr->intrinsic) {
1647ec681f3Smrg   case nir_intrinsic_load_ssbo:
1657ec681f3Smrg   case nir_intrinsic_store_ssbo:
1667ec681f3Smrg   case nir_intrinsic_ssbo_atomic_add:
1677ec681f3Smrg   case nir_intrinsic_ssbo_atomic_imin:
1687ec681f3Smrg   case nir_intrinsic_ssbo_atomic_umin:
1697ec681f3Smrg   case nir_intrinsic_ssbo_atomic_imax:
1707ec681f3Smrg   case nir_intrinsic_ssbo_atomic_umax:
1717ec681f3Smrg   case nir_intrinsic_ssbo_atomic_and:
1727ec681f3Smrg   case nir_intrinsic_ssbo_atomic_or:
1737ec681f3Smrg   case nir_intrinsic_ssbo_atomic_xor:
1747ec681f3Smrg   case nir_intrinsic_ssbo_atomic_exchange:
1757ec681f3Smrg   case nir_intrinsic_ssbo_atomic_comp_swap:
1767ec681f3Smrg   case nir_intrinsic_ssbo_atomic_fadd:
1777ec681f3Smrg   case nir_intrinsic_ssbo_atomic_fmin:
1787ec681f3Smrg   case nir_intrinsic_ssbo_atomic_fmax:
1797ec681f3Smrg   case nir_intrinsic_ssbo_atomic_fcomp_swap:
1807ec681f3Smrg      return true;
1817ec681f3Smrg   default:
1827ec681f3Smrg      return false;
1837ec681f3Smrg   }
1847ec681f3Smrg
1857ec681f3Smrg   return false;
1867ec681f3Smrg}
1877ec681f3Smrg
1887ec681f3Smrgbool
1897ec681f3Smrgnir_lower_ssbo(nir_shader *shader)
1907ec681f3Smrg{
1917ec681f3Smrg   bool progress = false;
1927ec681f3Smrg
1937ec681f3Smrg   nir_foreach_function(function, shader) {
1947ec681f3Smrg      nir_function_impl *impl = function->impl;
1957ec681f3Smrg      nir_builder b;
1967ec681f3Smrg      nir_builder_init(&b, impl);
1977ec681f3Smrg
1987ec681f3Smrg      nir_foreach_block(block, impl) {
1997ec681f3Smrg         nir_foreach_instr_safe(instr, block) {
2007ec681f3Smrg            if (!should_lower_ssbo_instr(instr)) continue;
2017ec681f3Smrg            progress = true;
2027ec681f3Smrg            b.cursor = nir_before_instr(instr);
2037ec681f3Smrg
2047ec681f3Smrg            nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
2057ec681f3Smrg            nir_ssa_def *replace = lower_ssbo_instr(&b, intr);
2067ec681f3Smrg
2077ec681f3Smrg            if (replace)  {
2087ec681f3Smrg               nir_ssa_def_rewrite_uses(&intr->dest.ssa,
2097ec681f3Smrg                                     replace);
2107ec681f3Smrg            }
2117ec681f3Smrg
2127ec681f3Smrg            nir_instr_remove(instr);
2137ec681f3Smrg         }
2147ec681f3Smrg      }
2157ec681f3Smrg   }
2167ec681f3Smrg
2177ec681f3Smrg   return progress;
2187ec681f3Smrg}
219