• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (C) 2018-2020 Collabora, Ltd.
3  * Copyright (C) 2019-2020 Icecream95
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 FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 
25 #include "pan_ir.h"
26 #include "compiler/nir/nir_builder.h"
27 
28 /* Midgard can write all of color, depth and stencil in a single writeout
29  * operation, so we merge depth/stencil stores with color stores.
30  * If there are no color stores, we add a write to the "depth RT".
31  *
32  * For Bifrost, we want these combined so we can properly order
33  * +ZS_EMIT with respect to +ATEST and +BLEND, as well as combining
34  * depth/stencil stores into a single +ZS_EMIT op.
35  */
36 
37 /*
38  * Get the type to report for a piece of a combined store, given the store it
39  * is combining from. If there is no store to render target #0, a dummy <0.0,
40  * 0.0, 0.0, 0.0> write is used, so report a matching float32 type.
41  */
42 static nir_alu_type
pan_nir_rt_store_type(nir_intrinsic_instr * store)43 pan_nir_rt_store_type(nir_intrinsic_instr *store)
44 {
45         return store ? nir_intrinsic_src_type(store) : nir_type_float32;
46 }
47 
48 static void
pan_nir_emit_combined_store(nir_builder * b,nir_intrinsic_instr * rt0_store,unsigned writeout,nir_intrinsic_instr ** stores)49 pan_nir_emit_combined_store(nir_builder *b,
50                             nir_intrinsic_instr *rt0_store,
51                             unsigned writeout,
52                             nir_intrinsic_instr **stores)
53 {
54         nir_intrinsic_instr *intr = nir_intrinsic_instr_create(b->shader, nir_intrinsic_store_combined_output_pan);
55 
56         intr->num_components = rt0_store ? rt0_store->src[0].ssa->num_components : 4;
57 
58         if (rt0_store)
59                 nir_intrinsic_set_base(intr, nir_intrinsic_base(rt0_store));
60         nir_intrinsic_set_src_type(intr, pan_nir_rt_store_type(rt0_store));
61         nir_intrinsic_set_dest_type(intr, pan_nir_rt_store_type(stores[2]));
62         nir_intrinsic_set_component(intr, writeout);
63 
64         nir_ssa_def *zero = nir_imm_int(b, 0);
65         nir_ssa_def *zero4 = nir_imm_ivec4(b, 0, 0, 0, 0);
66 
67         nir_ssa_def *src[] = {
68                 rt0_store ? rt0_store->src[0].ssa : zero4,
69                 rt0_store ? rt0_store->src[1].ssa : zero,
70                 stores[0] ? stores[0]->src[0].ssa : zero,
71                 stores[1] ? stores[1]->src[0].ssa : zero,
72                 stores[2] ? stores[2]->src[0].ssa : zero4,
73         };
74 
75         for (int i = 0; i < ARRAY_SIZE(src); ++i)
76                 intr->src[i] = nir_src_for_ssa(src[i]);
77 
78         nir_builder_instr_insert(b, &intr->instr);
79 }
80 bool
pan_nir_lower_zs_store(nir_shader * nir)81 pan_nir_lower_zs_store(nir_shader *nir)
82 {
83         if (nir->info.stage != MESA_SHADER_FRAGMENT)
84                 return false;
85 
86         nir_variable *vars[3] = { NULL };
87 
88         nir_foreach_shader_out_variable(var, nir) {
89                 if (var->data.location == FRAG_RESULT_DEPTH)
90                         vars[0] = var;
91                 else if (var->data.location == FRAG_RESULT_STENCIL)
92                         vars[1] = var;
93                 else if (var->data.index)
94                         vars[2] = var;
95         }
96 
97         if (!vars[0] && !vars[1] && !vars[2])
98                 return false;
99 
100         bool progress = false;
101 
102         nir_foreach_function(function, nir) {
103                 if (!function->impl) continue;
104 
105                 nir_intrinsic_instr *stores[3] = { NULL };
106 
107                 nir_foreach_block(block, function->impl) {
108                         nir_foreach_instr_safe(instr, block) {
109                                 if (instr->type != nir_instr_type_intrinsic)
110                                         continue;
111 
112                                 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
113                                 if (intr->intrinsic != nir_intrinsic_store_output)
114                                         continue;
115 
116                                 for (unsigned i = 0; i < ARRAY_SIZE(vars); ++i) {
117                                         if (vars[i] && nir_intrinsic_base(intr) == vars[i]->data.driver_location) {
118                                                 assert(!stores[i]);
119                                                 stores[i] = intr;
120                                         }
121                                 }
122                         }
123                 }
124 
125                 if (!stores[0] && !stores[1] && !stores[2]) continue;
126 
127                 nir_block *common_block = NULL;
128 
129                 /* Ensure all stores are in the same block */
130                 for (unsigned i = 0; i < ARRAY_SIZE(stores); ++i) {
131                         if (!stores[i])
132                                 continue;
133 
134                         nir_block *block = stores[i]->instr.block;
135 
136                         if (common_block)
137                                 assert(common_block == block);
138                         else
139                                 common_block = block;
140                 }
141 
142                 unsigned writeout = 0;
143                 if (stores[0])
144                         writeout |= PAN_WRITEOUT_Z;
145                 if (stores[1])
146                         writeout |= PAN_WRITEOUT_S;
147                 if (stores[2])
148                         writeout |= PAN_WRITEOUT_2;
149 
150                 bool replaced = false;
151 
152                 nir_foreach_block(block, function->impl) {
153                         nir_foreach_instr_safe(instr, block) {
154                                 if (instr->type != nir_instr_type_intrinsic)
155                                         continue;
156 
157                                 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr);
158                                 if (intr->intrinsic != nir_intrinsic_store_output)
159                                         continue;
160 
161                                 const nir_variable *var = nir_find_variable_with_driver_location(nir, nir_var_shader_out, nir_intrinsic_base(intr));
162                                 assert(var);
163 
164                                 if (var->data.location < FRAG_RESULT_DATA0)
165                                         continue;
166 
167                                 if (var->data.index)
168                                         continue;
169 
170                                 assert(nir_src_is_const(intr->src[1]) && "no indirect outputs");
171 
172                                 nir_builder b;
173                                 nir_builder_init(&b, function->impl);
174                                 b.cursor = nir_after_block_before_jump(instr->block);
175 
176                                 /* Trying to write depth twice results in the
177                                  * wrong blend shader being executed on
178                                  * Midgard */
179                                 unsigned this_store = PAN_WRITEOUT_C | (replaced ? 0 : writeout);
180 
181                                 pan_nir_emit_combined_store(&b, intr, this_store, stores);
182 
183                                 nir_instr_remove(instr);
184 
185                                 replaced = true;
186                         }
187                 }
188 
189                 /* Insert a store to the depth RT (0xff) if needed */
190                 if (!replaced) {
191                         nir_builder b;
192                         nir_builder_init(&b, function->impl);
193                         b.cursor = nir_after_block_before_jump(common_block);
194 
195                         pan_nir_emit_combined_store(&b, NULL, writeout, stores);
196                 }
197 
198                 for (unsigned i = 0; i < ARRAY_SIZE(stores); ++i) {
199                         if (stores[i])
200                                 nir_instr_remove(&stores[i]->instr);
201                 }
202 
203                 nir_metadata_preserve(function->impl, nir_metadata_block_index | nir_metadata_dominance);
204                 progress = true;
205         }
206 
207         return progress;
208 }
209