1bf215546Sopenharmony_ci/*
2bf215546Sopenharmony_ci * Copyright © 2019 Collabora, Ltd.
3bf215546Sopenharmony_ci *
4bf215546Sopenharmony_ci * Permission is hereby granted, free of charge, to any person obtaining a
5bf215546Sopenharmony_ci * copy of this software and associated documentation files (the "Software"),
6bf215546Sopenharmony_ci * to deal in the Software without restriction, including without limitation
7bf215546Sopenharmony_ci * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8bf215546Sopenharmony_ci * and/or sell copies of the Software, and to permit persons to whom the
9bf215546Sopenharmony_ci * Software is furnished to do so, subject to the following conditions:
10bf215546Sopenharmony_ci *
11bf215546Sopenharmony_ci * The above copyright notice and this permission notice (including the next
12bf215546Sopenharmony_ci * paragraph) shall be included in all copies or substantial portions of the
13bf215546Sopenharmony_ci * Software.
14bf215546Sopenharmony_ci *
15bf215546Sopenharmony_ci * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16bf215546Sopenharmony_ci * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17bf215546Sopenharmony_ci * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18bf215546Sopenharmony_ci * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19bf215546Sopenharmony_ci * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20bf215546Sopenharmony_ci * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21bf215546Sopenharmony_ci * IN THE SOFTWARE.
22bf215546Sopenharmony_ci *
23bf215546Sopenharmony_ci * Authors (Collabora):
24bf215546Sopenharmony_ci *    Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
25bf215546Sopenharmony_ci */
26bf215546Sopenharmony_ci
27bf215546Sopenharmony_ci#include "nir.h"
28bf215546Sopenharmony_ci#include "nir_builder.h"
29bf215546Sopenharmony_ci
30bf215546Sopenharmony_ci/*
31bf215546Sopenharmony_ci * Lowers SSBOs to globals, for hardware that lack native SSBO support. When
32bf215546Sopenharmony_ci * lowering, *_ssbo_* instructions will become *_global_* instructions,
33bf215546Sopenharmony_ci * augmented with load_ssbo_address.
34bf215546Sopenharmony_ci *
35bf215546Sopenharmony_ci * DOES NOT PERFORM BOUNDS CHECKING. DO NOT USE IN PRODUCTION ON UNTRUSTED
36bf215546Sopenharmony_ci * CONTEXTS INCLUDING WEBGL 2.
37bf215546Sopenharmony_ci */
38bf215546Sopenharmony_ci
39bf215546Sopenharmony_cistatic nir_intrinsic_op
40bf215546Sopenharmony_cilower_ssbo_op(nir_intrinsic_op op)
41bf215546Sopenharmony_ci{
42bf215546Sopenharmony_ci   switch (op) {
43bf215546Sopenharmony_ci   case nir_intrinsic_load_ssbo:
44bf215546Sopenharmony_ci      return nir_intrinsic_load_global;
45bf215546Sopenharmony_ci
46bf215546Sopenharmony_ci   case nir_intrinsic_store_ssbo:
47bf215546Sopenharmony_ci      return nir_intrinsic_store_global;
48bf215546Sopenharmony_ci
49bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_add:
50bf215546Sopenharmony_ci      return nir_intrinsic_global_atomic_add;
51bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_imin:
52bf215546Sopenharmony_ci      return nir_intrinsic_global_atomic_imin;
53bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_umin:
54bf215546Sopenharmony_ci      return nir_intrinsic_global_atomic_umin;
55bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_imax:
56bf215546Sopenharmony_ci      return nir_intrinsic_global_atomic_imax;
57bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_umax:
58bf215546Sopenharmony_ci      return nir_intrinsic_global_atomic_umax;
59bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_and:
60bf215546Sopenharmony_ci      return nir_intrinsic_global_atomic_and;
61bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_or:
62bf215546Sopenharmony_ci      return nir_intrinsic_global_atomic_or;
63bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_xor:
64bf215546Sopenharmony_ci      return nir_intrinsic_global_atomic_xor;
65bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_exchange:
66bf215546Sopenharmony_ci      return nir_intrinsic_global_atomic_exchange;
67bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_comp_swap:
68bf215546Sopenharmony_ci      return nir_intrinsic_global_atomic_comp_swap;
69bf215546Sopenharmony_ci
70bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_fadd:
71bf215546Sopenharmony_ci      return nir_intrinsic_global_atomic_fadd;
72bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_fmin:
73bf215546Sopenharmony_ci      return nir_intrinsic_global_atomic_fmin;
74bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_fmax:
75bf215546Sopenharmony_ci      return nir_intrinsic_global_atomic_fmax;
76bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_fcomp_swap:
77bf215546Sopenharmony_ci      return nir_intrinsic_global_atomic_fcomp_swap;
78bf215546Sopenharmony_ci
79bf215546Sopenharmony_ci   default:
80bf215546Sopenharmony_ci      unreachable("Invalid SSBO op");
81bf215546Sopenharmony_ci   }
82bf215546Sopenharmony_ci}
83bf215546Sopenharmony_ci
84bf215546Sopenharmony_ci/* Like SSBO property sysvals, though SSBO index may be indirect. C.f.
85bf215546Sopenharmony_ci * nir_load_system_value */
86bf215546Sopenharmony_ci
87bf215546Sopenharmony_cistatic inline nir_ssa_def *
88bf215546Sopenharmony_cinir_load_ssbo_prop(nir_builder *b, nir_intrinsic_op op,
89bf215546Sopenharmony_ci      nir_src *idx, unsigned bitsize)
90bf215546Sopenharmony_ci{
91bf215546Sopenharmony_ci   nir_intrinsic_instr *load = nir_intrinsic_instr_create(b->shader, op);
92bf215546Sopenharmony_ci   load->num_components = 1;
93bf215546Sopenharmony_ci   nir_src_copy(&load->src[0], idx);
94bf215546Sopenharmony_ci   nir_ssa_dest_init(&load->instr, &load->dest, 1, bitsize, NULL);
95bf215546Sopenharmony_ci   nir_builder_instr_insert(b, &load->instr);
96bf215546Sopenharmony_ci   return &load->dest.ssa;
97bf215546Sopenharmony_ci}
98bf215546Sopenharmony_ci
99bf215546Sopenharmony_ci#define nir_ssbo_prop(b, prop, index, bitsize) \
100bf215546Sopenharmony_ci   nir_load_ssbo_prop(b, nir_intrinsic_##prop, index, bitsize)
101bf215546Sopenharmony_ci
102bf215546Sopenharmony_cistatic nir_ssa_def *
103bf215546Sopenharmony_cilower_ssbo_instr(nir_builder *b, nir_intrinsic_instr *intr)
104bf215546Sopenharmony_ci{
105bf215546Sopenharmony_ci   nir_intrinsic_op op = lower_ssbo_op(intr->intrinsic);
106bf215546Sopenharmony_ci   bool is_store = op == nir_intrinsic_store_global;
107bf215546Sopenharmony_ci   bool is_atomic = !is_store && op != nir_intrinsic_load_global;
108bf215546Sopenharmony_ci
109bf215546Sopenharmony_ci   /* We have to calculate the address:
110bf215546Sopenharmony_ci    *
111bf215546Sopenharmony_ci    * &(SSBO[offset]) = &SSBO + offset
112bf215546Sopenharmony_ci    */
113bf215546Sopenharmony_ci
114bf215546Sopenharmony_ci   nir_src index = intr->src[is_store ? 1 : 0];
115bf215546Sopenharmony_ci   nir_src *offset_src = nir_get_io_offset_src(intr);
116bf215546Sopenharmony_ci   nir_ssa_def *offset = nir_ssa_for_src(b, *offset_src, 1);
117bf215546Sopenharmony_ci
118bf215546Sopenharmony_ci   nir_ssa_def *address =
119bf215546Sopenharmony_ci      nir_iadd(b,
120bf215546Sopenharmony_ci            nir_ssbo_prop(b, load_ssbo_address, &index, 64),
121bf215546Sopenharmony_ci            nir_u2u64(b, offset));
122bf215546Sopenharmony_ci
123bf215546Sopenharmony_ci   /* Create the replacement intrinsic */
124bf215546Sopenharmony_ci
125bf215546Sopenharmony_ci   nir_intrinsic_instr *global =
126bf215546Sopenharmony_ci      nir_intrinsic_instr_create(b->shader, op);
127bf215546Sopenharmony_ci
128bf215546Sopenharmony_ci   global->num_components = intr->num_components;
129bf215546Sopenharmony_ci   global->src[is_store ? 1 : 0] = nir_src_for_ssa(address);
130bf215546Sopenharmony_ci
131bf215546Sopenharmony_ci   if (!is_atomic) {
132bf215546Sopenharmony_ci      nir_intrinsic_set_align_mul(global, nir_intrinsic_align_mul(intr));
133bf215546Sopenharmony_ci      nir_intrinsic_set_align_offset(global, nir_intrinsic_align_offset(intr));
134bf215546Sopenharmony_ci   }
135bf215546Sopenharmony_ci
136bf215546Sopenharmony_ci   if (is_store) {
137bf215546Sopenharmony_ci      nir_src_copy(&global->src[0], &intr->src[0]);
138bf215546Sopenharmony_ci      nir_intrinsic_set_write_mask(global, nir_intrinsic_write_mask(intr));
139bf215546Sopenharmony_ci   } else {
140bf215546Sopenharmony_ci      nir_ssa_dest_init(&global->instr, &global->dest,
141bf215546Sopenharmony_ci                        intr->dest.ssa.num_components,
142bf215546Sopenharmony_ci                        intr->dest.ssa.bit_size, NULL);
143bf215546Sopenharmony_ci
144bf215546Sopenharmony_ci      if (is_atomic) {
145bf215546Sopenharmony_ci         nir_src_copy(&global->src[1], &intr->src[2]);
146bf215546Sopenharmony_ci         if (nir_intrinsic_infos[op].num_srcs > 2)
147bf215546Sopenharmony_ci            nir_src_copy(&global->src[2], &intr->src[3]);
148bf215546Sopenharmony_ci      }
149bf215546Sopenharmony_ci   }
150bf215546Sopenharmony_ci
151bf215546Sopenharmony_ci   nir_builder_instr_insert(b, &global->instr);
152bf215546Sopenharmony_ci   return is_store ? NULL : &global->dest.ssa;
153bf215546Sopenharmony_ci}
154bf215546Sopenharmony_ci
155bf215546Sopenharmony_cistatic bool
156bf215546Sopenharmony_cishould_lower_ssbo_instr(const nir_instr *instr)
157bf215546Sopenharmony_ci{
158bf215546Sopenharmony_ci   if (instr->type != nir_instr_type_intrinsic)
159bf215546Sopenharmony_ci      return false;
160bf215546Sopenharmony_ci
161bf215546Sopenharmony_ci   const nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
162bf215546Sopenharmony_ci
163bf215546Sopenharmony_ci   switch (intr->intrinsic) {
164bf215546Sopenharmony_ci   case nir_intrinsic_load_ssbo:
165bf215546Sopenharmony_ci   case nir_intrinsic_store_ssbo:
166bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_add:
167bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_imin:
168bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_umin:
169bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_imax:
170bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_umax:
171bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_and:
172bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_or:
173bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_xor:
174bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_exchange:
175bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_comp_swap:
176bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_fadd:
177bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_fmin:
178bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_fmax:
179bf215546Sopenharmony_ci   case nir_intrinsic_ssbo_atomic_fcomp_swap:
180bf215546Sopenharmony_ci      return true;
181bf215546Sopenharmony_ci   default:
182bf215546Sopenharmony_ci      return false;
183bf215546Sopenharmony_ci   }
184bf215546Sopenharmony_ci
185bf215546Sopenharmony_ci   return false;
186bf215546Sopenharmony_ci}
187bf215546Sopenharmony_ci
188bf215546Sopenharmony_cibool
189bf215546Sopenharmony_cinir_lower_ssbo(nir_shader *shader)
190bf215546Sopenharmony_ci{
191bf215546Sopenharmony_ci   bool progress = false;
192bf215546Sopenharmony_ci
193bf215546Sopenharmony_ci   nir_foreach_function(function, shader) {
194bf215546Sopenharmony_ci      nir_function_impl *impl = function->impl;
195bf215546Sopenharmony_ci      nir_builder b;
196bf215546Sopenharmony_ci      nir_builder_init(&b, impl);
197bf215546Sopenharmony_ci
198bf215546Sopenharmony_ci      nir_foreach_block(block, impl) {
199bf215546Sopenharmony_ci         nir_foreach_instr_safe(instr, block) {
200bf215546Sopenharmony_ci            if (!should_lower_ssbo_instr(instr)) continue;
201bf215546Sopenharmony_ci            progress = true;
202bf215546Sopenharmony_ci            b.cursor = nir_before_instr(instr);
203bf215546Sopenharmony_ci
204bf215546Sopenharmony_ci            nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
205bf215546Sopenharmony_ci            nir_ssa_def *replace = lower_ssbo_instr(&b, intr);
206bf215546Sopenharmony_ci
207bf215546Sopenharmony_ci            if (replace)  {
208bf215546Sopenharmony_ci               nir_ssa_def_rewrite_uses(&intr->dest.ssa,
209bf215546Sopenharmony_ci                                     replace);
210bf215546Sopenharmony_ci            }
211bf215546Sopenharmony_ci
212bf215546Sopenharmony_ci            nir_instr_remove(instr);
213bf215546Sopenharmony_ci         }
214bf215546Sopenharmony_ci      }
215bf215546Sopenharmony_ci   }
216bf215546Sopenharmony_ci
217bf215546Sopenharmony_ci   return progress;
218bf215546Sopenharmony_ci}
219