Alyssa Rosenzweig | 42319c5 | 2020-11-04 08:37:55 -0500 | [diff] [blame] | 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 | bool |
| 37 | pan_nir_lower_zs_store(nir_shader *nir) |
| 38 | { |
| 39 | if (nir->info.stage != MESA_SHADER_FRAGMENT) |
| 40 | return false; |
| 41 | |
| 42 | nir_variable *z_var = NULL, *s_var = NULL; |
| 43 | |
| 44 | nir_foreach_shader_out_variable(var, nir) { |
| 45 | if (var->data.location == FRAG_RESULT_DEPTH) |
| 46 | z_var = var; |
| 47 | else if (var->data.location == FRAG_RESULT_STENCIL) |
| 48 | s_var = var; |
| 49 | } |
| 50 | |
| 51 | if (!z_var && !s_var) |
| 52 | return false; |
| 53 | |
| 54 | bool progress = false; |
| 55 | |
| 56 | nir_foreach_function(function, nir) { |
| 57 | if (!function->impl) continue; |
| 58 | |
| 59 | nir_intrinsic_instr *z_store = NULL, *s_store = NULL; |
| 60 | |
| 61 | nir_foreach_block(block, function->impl) { |
| 62 | nir_foreach_instr_safe(instr, block) { |
| 63 | if (instr->type != nir_instr_type_intrinsic) |
| 64 | continue; |
| 65 | |
| 66 | nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr); |
| 67 | if (intr->intrinsic != nir_intrinsic_store_output) |
| 68 | continue; |
| 69 | |
| 70 | if (z_var && nir_intrinsic_base(intr) == z_var->data.driver_location) { |
| 71 | assert(!z_store); |
| 72 | z_store = intr; |
| 73 | } |
| 74 | |
| 75 | if (s_var && nir_intrinsic_base(intr) == s_var->data.driver_location) { |
| 76 | assert(!s_store); |
| 77 | s_store = intr; |
| 78 | } |
| 79 | } |
| 80 | } |
| 81 | |
| 82 | if (!z_store && !s_store) continue; |
| 83 | |
| 84 | bool replaced = false; |
| 85 | |
| 86 | nir_foreach_block(block, function->impl) { |
| 87 | nir_foreach_instr_safe(instr, block) { |
| 88 | if (instr->type != nir_instr_type_intrinsic) |
| 89 | continue; |
| 90 | |
| 91 | nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr); |
| 92 | if (intr->intrinsic != nir_intrinsic_store_output) |
| 93 | continue; |
| 94 | |
| 95 | const nir_variable *var = nir_find_variable_with_driver_location(nir, nir_var_shader_out, nir_intrinsic_base(intr)); |
| 96 | assert(var); |
| 97 | |
| 98 | if (var->data.location != FRAG_RESULT_COLOR && |
| 99 | var->data.location < FRAG_RESULT_DATA0) |
| 100 | continue; |
| 101 | |
| 102 | if (var->data.index) |
| 103 | continue; |
| 104 | |
| 105 | assert(nir_src_is_const(intr->src[1]) && "no indirect outputs"); |
| 106 | |
| 107 | nir_builder b; |
| 108 | nir_builder_init(&b, function->impl); |
| 109 | |
| 110 | assert(!z_store || z_store->instr.block == instr->block); |
| 111 | assert(!s_store || s_store->instr.block == instr->block); |
| 112 | b.cursor = nir_after_block_before_jump(instr->block); |
| 113 | |
| 114 | nir_intrinsic_instr *combined_store; |
| 115 | combined_store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_store_combined_output_pan); |
| 116 | |
| 117 | combined_store->num_components = intr->src[0].ssa->num_components; |
| 118 | |
| 119 | nir_intrinsic_set_base(combined_store, nir_intrinsic_base(intr)); |
Alyssa Rosenzweig | 4a2ecc7 | 2020-11-04 08:59:35 -0500 | [diff] [blame^] | 120 | nir_intrinsic_set_src_type(combined_store, nir_intrinsic_src_type(intr)); |
Alyssa Rosenzweig | 42319c5 | 2020-11-04 08:37:55 -0500 | [diff] [blame] | 121 | |
| 122 | unsigned writeout = PAN_WRITEOUT_C; |
| 123 | if (z_store) |
| 124 | writeout |= PAN_WRITEOUT_Z; |
| 125 | if (s_store) |
| 126 | writeout |= PAN_WRITEOUT_S; |
| 127 | |
| 128 | nir_intrinsic_set_component(combined_store, writeout); |
| 129 | |
| 130 | struct nir_ssa_def *zero = nir_imm_int(&b, 0); |
| 131 | |
| 132 | struct nir_ssa_def *src[4] = { |
| 133 | intr->src[0].ssa, |
| 134 | intr->src[1].ssa, |
| 135 | z_store ? z_store->src[0].ssa : zero, |
| 136 | s_store ? s_store->src[0].ssa : zero, |
| 137 | }; |
| 138 | |
| 139 | for (int i = 0; i < 4; ++i) |
| 140 | combined_store->src[i] = nir_src_for_ssa(src[i]); |
| 141 | |
| 142 | nir_builder_instr_insert(&b, &combined_store->instr); |
| 143 | |
| 144 | nir_instr_remove(instr); |
| 145 | |
| 146 | replaced = true; |
| 147 | } |
| 148 | } |
| 149 | |
| 150 | /* Insert a store to the depth RT (0xff) if needed */ |
| 151 | if (!replaced) { |
| 152 | nir_builder b; |
| 153 | nir_builder_init(&b, function->impl); |
| 154 | |
| 155 | nir_block *block = NULL; |
| 156 | if (z_store && s_store) |
| 157 | assert(z_store->instr.block == s_store->instr.block); |
| 158 | |
| 159 | if (z_store) |
| 160 | block = z_store->instr.block; |
| 161 | else |
| 162 | block = s_store->instr.block; |
| 163 | |
| 164 | b.cursor = nir_after_block_before_jump(block); |
| 165 | |
| 166 | nir_intrinsic_instr *combined_store; |
| 167 | combined_store = nir_intrinsic_instr_create(b.shader, nir_intrinsic_store_combined_output_pan); |
| 168 | |
| 169 | combined_store->num_components = 4; |
| 170 | |
| 171 | unsigned base; |
| 172 | if (z_store) |
| 173 | base = nir_intrinsic_base(z_store); |
| 174 | else |
| 175 | base = nir_intrinsic_base(s_store); |
| 176 | nir_intrinsic_set_base(combined_store, base); |
Alyssa Rosenzweig | 4a2ecc7 | 2020-11-04 08:59:35 -0500 | [diff] [blame^] | 177 | nir_intrinsic_set_src_type(combined_store, nir_type_float32); |
Alyssa Rosenzweig | 42319c5 | 2020-11-04 08:37:55 -0500 | [diff] [blame] | 178 | |
| 179 | unsigned writeout = 0; |
| 180 | if (z_store) |
| 181 | writeout |= PAN_WRITEOUT_Z; |
| 182 | if (s_store) |
| 183 | writeout |= PAN_WRITEOUT_S; |
| 184 | |
| 185 | nir_intrinsic_set_component(combined_store, writeout); |
| 186 | |
| 187 | struct nir_ssa_def *zero = nir_imm_int(&b, 0); |
| 188 | |
| 189 | struct nir_ssa_def *src[4] = { |
| 190 | nir_imm_vec4(&b, 0, 0, 0, 0), |
| 191 | zero, |
| 192 | z_store ? z_store->src[0].ssa : zero, |
| 193 | s_store ? s_store->src[0].ssa : zero, |
| 194 | }; |
| 195 | |
| 196 | for (int i = 0; i < 4; ++i) |
| 197 | combined_store->src[i] = nir_src_for_ssa(src[i]); |
| 198 | |
| 199 | nir_builder_instr_insert(&b, &combined_store->instr); |
| 200 | } |
| 201 | |
| 202 | if (z_store) |
| 203 | nir_instr_remove(&z_store->instr); |
| 204 | |
| 205 | if (s_store) |
| 206 | nir_instr_remove(&s_store->instr); |
| 207 | |
| 208 | nir_metadata_preserve(function->impl, nir_metadata_block_index | nir_metadata_dominance); |
| 209 | progress = true; |
| 210 | } |
| 211 | |
| 212 | return progress; |
| 213 | } |
| 214 | |
| 215 | /* Real writeout stores, which break execution, need to be moved to after |
| 216 | * dual-source stores, which are just standard register writes. */ |
| 217 | bool |
| 218 | pan_nir_reorder_writeout(nir_shader *nir) |
| 219 | { |
| 220 | bool progress = false; |
| 221 | |
| 222 | nir_foreach_function(function, nir) { |
| 223 | if (!function->impl) continue; |
| 224 | |
| 225 | nir_foreach_block(block, function->impl) { |
| 226 | nir_instr *last_writeout = NULL; |
| 227 | |
| 228 | nir_foreach_instr_reverse_safe(instr, block) { |
| 229 | if (instr->type != nir_instr_type_intrinsic) |
| 230 | continue; |
| 231 | |
| 232 | nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr); |
| 233 | if (intr->intrinsic != nir_intrinsic_store_output) |
| 234 | continue; |
| 235 | |
| 236 | const nir_variable *var = nir_find_variable_with_driver_location(nir, nir_var_shader_out, nir_intrinsic_base(intr)); |
| 237 | |
| 238 | if (var->data.index) { |
| 239 | if (!last_writeout) |
| 240 | last_writeout = instr; |
| 241 | continue; |
| 242 | } |
| 243 | |
| 244 | if (!last_writeout) |
| 245 | continue; |
| 246 | |
| 247 | /* This is a real store, so move it to after dual-source stores */ |
| 248 | exec_node_remove(&instr->node); |
| 249 | exec_node_insert_after(&last_writeout->node, &instr->node); |
| 250 | |
| 251 | progress = true; |
| 252 | } |
| 253 | } |
| 254 | } |
| 255 | |
| 256 | return progress; |
| 257 | } |