1 /* 2 * Copyright 2016 Red Hat 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 FROM, 20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 21 * SOFTWARE. 22 */ 23 24 /* Lowering pass that lowers accesses to built-in uniform variables. 25 * Built-in uniforms are not necessarily packed the same way that 26 * normal uniform structs are, for example: 27 * 28 * struct gl_FogParameters { 29 * vec4 color; 30 * float density; 31 * float start; 32 * float end; 33 * float scale; 34 * }; 35 * 36 * is packed into vec4[2], whereas the same struct would be packed 37 * (by gallium), as vec4[5] if it where not built-in. Because of 38 * this, we need to replace (for example) access like: 39 * 40 * vec1 ssa_1 = intrinsic load_var () (gl_Fog.start) () 41 * 42 * with: 43 * 44 * vec4 ssa_2 = intrinsic load_var () (fog.params) () 45 * vec1 ssa_1 = ssa_2.y 46 * 47 * with appropriate substitutions in the uniform variables list: 48 * 49 * decl_var uniform INTERP_MODE_NONE gl_FogParameters gl_Fog (0, 0) 50 * 51 * would become: 52 * 53 * decl_var uniform INTERP_MODE_NONE vec4 state.fog.color (0, 0) 54 * decl_var uniform INTERP_MODE_NONE vec4 state.fog.params (0, 1) 55 * 56 * See in particular 'struct gl_builtin_uniform_element'. 57 */ 58 59 #include "compiler/nir/nir.h" 60 #include "compiler/nir/nir_builder.h" 61 #include "st_nir.h" 62 #include "compiler/glsl/ir.h" 63 #include "uniforms.h" 64 #include "program/prog_instruction.h" 65 66 typedef struct { 67 nir_shader *shader; 68 nir_builder builder; 69 void *mem_ctx; 70 } lower_builtin_state; 71 72 static const struct gl_builtin_uniform_element * 73 get_element(const struct gl_builtin_uniform_desc *desc, nir_deref_var *deref) 74 { 75 nir_deref *tail = &deref->deref; 76 77 if ((desc->num_elements == 1) && (desc->elements[0].field == NULL)) 78 return NULL; 79 80 /* we handle arrays in get_variable(): */ 81 if (tail->child->deref_type == nir_deref_type_array) 82 tail = tail->child; 83 84 /* don't need to deal w/ non-struct or array of non-struct: */ 85 if (!tail->child) 86 return NULL; 87 88 if (tail->child->deref_type != nir_deref_type_struct) 89 return NULL; 90 91 nir_deref_struct *deref_struct = nir_deref_as_struct(tail->child); 92 93 assert(deref_struct->index < desc->num_elements); 94 95 return &desc->elements[deref_struct->index]; 96 } 97 98 static nir_variable * 99 get_variable(lower_builtin_state *state, nir_deref_var *deref, 100 const struct gl_builtin_uniform_element *element) 101 { 102 nir_shader *shader = state->shader; 103 int tokens[STATE_LENGTH]; 104 105 memcpy(tokens, element->tokens, sizeof(tokens)); 106 107 if (deref->deref.child->deref_type == nir_deref_type_array) { 108 nir_deref_array *darr = nir_deref_as_array(deref->deref.child); 109 110 assert(darr->deref_array_type == nir_deref_array_type_direct); 111 112 /* we need to fixup the array index slot: */ 113 switch (tokens[0]) { 114 case STATE_MODELVIEW_MATRIX: 115 case STATE_PROJECTION_MATRIX: 116 case STATE_MVP_MATRIX: 117 case STATE_TEXTURE_MATRIX: 118 case STATE_PROGRAM_MATRIX: 119 case STATE_LIGHT: 120 case STATE_LIGHTPROD: 121 case STATE_TEXGEN: 122 case STATE_TEXENV_COLOR: 123 case STATE_CLIPPLANE: 124 tokens[1] = darr->base_offset; 125 break; 126 } 127 } 128 129 char *name = _mesa_program_state_string((gl_state_index *)tokens); 130 131 nir_foreach_variable(var, &shader->uniforms) { 132 if (strcmp(var->name, name) == 0) { 133 free(name); 134 return var; 135 } 136 } 137 138 /* variable doesn't exist yet, so create it: */ 139 nir_variable *var = 140 nir_variable_create(shader, nir_var_uniform, glsl_vec4_type(), name); 141 142 var->num_state_slots = 1; 143 var->state_slots = ralloc_array(var, nir_state_slot, 1); 144 memcpy(var->state_slots[0].tokens, tokens, 145 sizeof(var->state_slots[0].tokens)); 146 147 free(name); 148 149 return var; 150 } 151 152 static bool 153 lower_builtin_block(lower_builtin_state *state, nir_block *block) 154 { 155 nir_builder *b = &state->builder; 156 157 nir_foreach_instr_safe(instr, block) { 158 if (instr->type != nir_instr_type_intrinsic) 159 continue; 160 161 nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr); 162 163 if (intrin->intrinsic != nir_intrinsic_load_var) 164 continue; 165 166 nir_variable *var = intrin->variables[0]->var; 167 if (var->data.mode != nir_var_uniform) 168 continue; 169 170 /* built-in's will always start with "gl_" */ 171 if (strncmp(var->name, "gl_", 3) != 0) 172 continue; 173 174 const struct gl_builtin_uniform_desc *desc = 175 _mesa_glsl_get_builtin_uniform_desc(var->name); 176 177 /* if no descriptor, it isn't something we need to handle specially: */ 178 if (!desc) 179 continue; 180 181 const struct gl_builtin_uniform_element *element = 182 get_element(desc, intrin->variables[0]); 183 184 /* matrix elements (array_deref) do not need special handling: */ 185 if (!element) 186 continue; 187 188 /* remove existing var from uniform list: */ 189 exec_node_remove(&var->node); 190 /* the _self_link() ensures we can remove multiple times, rather than 191 * trying to keep track of what we have already removed: 192 */ 193 exec_node_self_link(&var->node); 194 195 nir_variable *new_var = 196 get_variable(state, intrin->variables[0], element); 197 198 b->cursor = nir_before_instr(instr); 199 200 nir_ssa_def *def = nir_load_var(b, new_var); 201 202 /* swizzle the result: */ 203 unsigned swiz[4]; 204 for (unsigned i = 0; i < 4; i++) { 205 swiz[i] = GET_SWZ(element->swizzle, i); 206 assert(swiz[i] <= SWIZZLE_W); 207 } 208 def = nir_swizzle(b, def, swiz, intrin->num_components, true); 209 210 /* and rewrite uses of original instruction: */ 211 assert(intrin->dest.is_ssa); 212 nir_ssa_def_rewrite_uses(&intrin->dest.ssa, nir_src_for_ssa(def)); 213 214 /* at this point intrin should be unused. We need to remove it 215 * (rather than waiting for DCE pass) to avoid dangling reference 216 * to remove'd var. And we have to remove the original uniform 217 * var since we don't want it to get uniform space allocated. 218 */ 219 exec_node_remove(&intrin->instr.node); 220 } 221 222 return true; 223 } 224 225 static void 226 lower_builtin_impl(lower_builtin_state *state, nir_function_impl *impl) 227 { 228 nir_builder_init(&state->builder, impl); 229 state->mem_ctx = ralloc_parent(impl); 230 231 nir_foreach_block(block, impl) { 232 lower_builtin_block(state, block); 233 } 234 235 nir_metadata_preserve(impl, nir_metadata_block_index | 236 nir_metadata_dominance); 237 } 238 239 void 240 st_nir_lower_builtin(nir_shader *shader) 241 { 242 lower_builtin_state state; 243 state.shader = shader; 244 nir_foreach_function(function, shader) { 245 if (function->impl) 246 lower_builtin_impl(&state, function->impl); 247 } 248 } 249