1/* 2 * Copyright © 2020 Google, Inc. 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#include "nir.h" 25#include "nir_builder.h" 26 27/* A pass to split intrinsics with discontinuous writemasks into ones 28 * with contiguous writemasks starting with .x, ie: 29 * 30 * vec4 32 ssa_76 = vec4 ssa_35, ssa_35, ssa_35, ssa_35 31 * intrinsic store_ssbo (ssa_76, ssa_105, ssa_106) (2, 0, 4, 0) // wrmask=y 32 * 33 * is turned into: 34 * 35 * vec4 32 ssa_76 = vec4 ssa_35, ssa_35, ssa_35, ssa_35 36 * vec1 32 ssa_107 = load_const (0x00000001) 37 * vec1 32 ssa_108 = iadd ssa_106, ssa_107 38 * vec1 32 ssa_109 = mov ssa_76.y 39 * intrinsic store_ssbo (ssa_109, ssa_105, ssa_108) (1, 0, 4, 0) // wrmask=x 40 * 41 * and likewise: 42 * 43 * vec4 32 ssa_76 = vec4 ssa_35, ssa_35, ssa_35, ssa_35 44 * intrinsic store_ssbo (ssa_76, ssa_105, ssa_106) (15, 0, 4, 0) // wrmask=xzw 45 * 46 * is split into: 47 * 48 * // .x component: 49 * vec4 32 ssa_76 = vec4 ssa_35, ssa_35, ssa_35, ssa_35 50 * vec1 32 ssa_107 = load_const (0x00000000) 51 * vec1 32 ssa_108 = iadd ssa_106, ssa_107 52 * vec1 32 ssa_109 = mov ssa_76.x 53 * intrinsic store_ssbo (ssa_109, ssa_105, ssa_108) (1, 0, 4, 0) // wrmask=x 54 * // .zw components: 55 * vec1 32 ssa_110 = load_const (0x00000002) 56 * vec1 32 ssa_111 = iadd ssa_106, ssa_110 57 * vec2 32 ssa_112 = mov ssa_76.zw 58 * intrinsic store_ssbo (ssa_112, ssa_105, ssa_111) (3, 0, 4, 0) // wrmask=xy 59 */ 60 61static int 62value_src(nir_intrinsic_op intrinsic) 63{ 64 switch (intrinsic) { 65 case nir_intrinsic_store_output: 66 case nir_intrinsic_store_per_vertex_output: 67 case nir_intrinsic_store_ssbo: 68 case nir_intrinsic_store_shared: 69 case nir_intrinsic_store_global: 70 case nir_intrinsic_store_scratch: 71 return 0; 72 default: 73 return -1; 74 } 75} 76 77static int 78offset_src(nir_intrinsic_op intrinsic) 79{ 80 switch (intrinsic) { 81 case nir_intrinsic_store_output: 82 case nir_intrinsic_store_shared: 83 case nir_intrinsic_store_global: 84 case nir_intrinsic_store_scratch: 85 return 1; 86 case nir_intrinsic_store_per_vertex_output: 87 case nir_intrinsic_store_ssbo: 88 return 2; 89 default: 90 return -1; 91 } 92} 93 94static void 95split_wrmask(nir_builder *b, nir_intrinsic_instr *intr) 96{ 97 const nir_intrinsic_info *info = &nir_intrinsic_infos[intr->intrinsic]; 98 99 b->cursor = nir_before_instr(&intr->instr); 100 101 assert(!info->has_dest); /* expecting only store intrinsics */ 102 103 unsigned num_srcs = info->num_srcs; 104 unsigned value_idx = value_src(intr->intrinsic); 105 unsigned offset_idx = offset_src(intr->intrinsic); 106 unsigned num_comp = nir_intrinsic_src_components(intr, value_idx); 107 108 unsigned wrmask = nir_intrinsic_write_mask(intr); 109 while (wrmask) { 110 unsigned first_component = ffs(wrmask) - 1; 111 unsigned length = ffs(~(wrmask >> first_component)) - 1; 112 113 nir_ssa_def *value = nir_ssa_for_src(b, intr->src[value_idx], num_comp); 114 nir_ssa_def *offset = nir_ssa_for_src(b, intr->src[offset_idx], 1); 115 116 /* swizzle out the consecutive components that we'll store 117 * in this iteration: 118 */ 119 unsigned cur_mask = (BITFIELD_MASK(length) << first_component); 120 value = nir_channels(b, value, cur_mask); 121 122 /* and create the replacement intrinsic: */ 123 nir_intrinsic_instr *new_intr = 124 nir_intrinsic_instr_create(b->shader, intr->intrinsic); 125 126 nir_intrinsic_copy_const_indices(new_intr, intr); 127 nir_intrinsic_set_write_mask(new_intr, BITFIELD_MASK(length)); 128 129 const int offset_units = value->bit_size / 8; 130 131 if (nir_intrinsic_has_align_mul(intr)) { 132 assert(nir_intrinsic_has_align_offset(intr)); 133 unsigned align_mul = nir_intrinsic_align_mul(intr); 134 unsigned align_off = nir_intrinsic_align_offset(intr); 135 136 align_off += offset_units * first_component; 137 align_off = align_off % align_mul; 138 139 nir_intrinsic_set_align(new_intr, align_mul, align_off); 140 } 141 142 /* if the instruction has a BASE, fold the offset adjustment 143 * into that instead of adding alu instructions, otherwise add 144 * instructions 145 */ 146 unsigned offset_adj = offset_units * first_component; 147 if (nir_intrinsic_has_base(intr)) { 148 nir_intrinsic_set_base(new_intr, 149 nir_intrinsic_base(intr) + offset_adj); 150 } else { 151 offset = nir_iadd(b, offset, 152 nir_imm_intN_t(b, offset_adj, offset->bit_size)); 153 } 154 155 new_intr->num_components = length; 156 157 /* Copy the sources, replacing value/offset, and passing everything 158 * else through to the new instrution: 159 */ 160 for (unsigned i = 0; i < num_srcs; i++) { 161 if (i == value_idx) { 162 new_intr->src[i] = nir_src_for_ssa(value); 163 } else if (i == offset_idx) { 164 new_intr->src[i] = nir_src_for_ssa(offset); 165 } else { 166 new_intr->src[i] = intr->src[i]; 167 } 168 } 169 170 nir_builder_instr_insert(b, &new_intr->instr); 171 172 /* Clear the bits in the writemask that we just wrote, then try 173 * again to see if more channels are left. 174 */ 175 wrmask &= ~cur_mask; 176 } 177 178 /* Finally remove the original intrinsic. */ 179 nir_instr_remove(&intr->instr); 180} 181 182struct nir_lower_wrmasks_state { 183 nir_instr_filter_cb cb; 184 const void *data; 185}; 186 187static bool 188nir_lower_wrmasks_instr(nir_builder *b, nir_instr *instr, void *data) 189{ 190 struct nir_lower_wrmasks_state *state = data; 191 192 if (instr->type != nir_instr_type_intrinsic) 193 return false; 194 195 nir_intrinsic_instr *intr = nir_instr_as_intrinsic(instr); 196 197 /* if no wrmask, then skip it: */ 198 if (!nir_intrinsic_has_write_mask(intr)) 199 return false; 200 201 /* if wrmask is already contiguous, then nothing to do: */ 202 if (nir_intrinsic_write_mask(intr) == BITFIELD_MASK(intr->num_components)) 203 return false; 204 205 /* do we know how to lower this instruction? */ 206 if (value_src(intr->intrinsic) < 0) 207 return false; 208 209 assert(offset_src(intr->intrinsic) >= 0); 210 211 /* does backend need us to lower this intrinsic? */ 212 if (state->cb && !state->cb(instr, state->data)) 213 return false; 214 215 split_wrmask(b, intr); 216 217 return true; 218} 219 220bool 221nir_lower_wrmasks(nir_shader *shader, nir_instr_filter_cb cb, const void *data) 222{ 223 struct nir_lower_wrmasks_state state = { 224 .cb = cb, 225 .data = data, 226 }; 227 228 return nir_shader_instructions_pass(shader, 229 nir_lower_wrmasks_instr, 230 nir_metadata_block_index | 231 nir_metadata_dominance, 232 &state); 233} 234