src/freedreno/ir3/ir3_nir_analyze_ubo_ranges.c

   1 /*
   2  * Copyright © 2019 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 "ir3_nir.h"
  25 #include "ir3_compiler.h"
  26 #include "compiler/nir/nir.h"
  27 #include "compiler/nir/nir_builder.h"
  28 #include "mesa/main/macros.h"
  29
  30 static inline struct ir3_ubo_range
  31 get_ubo_load_range(nir_intrinsic_instr *instr)
  32 {
  33         struct ir3_ubo_range r;
  34
  35         const int offset = nir_src_as_uint(instr->src[1]);
  36         const int bytes = nir_intrinsic_dest_components(instr) * 4;
  37
  38         r.start = ROUND_DOWN_TO(offset, 16 * 4);
  39         r.end = ALIGN(offset + bytes, 16 * 4);
  40
  41         return r;
  42 }
  43
  44 static void
  45 gather_ubo_ranges(nir_shader *nir, nir_intrinsic_instr *instr,
  46                                   struct ir3_ubo_analysis_state *state)
  47 {
  48         if (!nir_src_is_const(instr->src[0]))
  49                 return;
  50
  51         if (!nir_src_is_const(instr->src[1])) {
  52                 if (nir_src_as_uint(instr->src[0]) == 0) {
  53                         /* If this is an indirect on UBO 0, we'll still lower it back to
  54                          * load_uniform.  Set the range to cover all of UBO 0.
  55                          */
  56                         state->range[0].end = align(nir->num_uniforms * 16, 16 * 4);
  57                 }
  58
  59                 return;
  60         }
  61
  62         const struct ir3_ubo_range r = get_ubo_load_range(instr);
  63         const uint32_t block = nir_src_as_uint(instr->src[0]);
  64
  65         /* if UBO lowering is disabled, we still want to lower block 0
  66          * (which is normal uniforms):
  67          */
  68         if ((block > 0) && (ir3_shader_debug & IR3_DBG_NOUBOOPT))
  69                 return;
  70
  71         if (r.start < state->range[block].start)
  72                 state->range[block].start = r.start;
  73         if (state->range[block].end < r.end)
  74                 state->range[block].end = r.end;
  75 }
  76
  77 /* For indirect offset, it is common to see a pattern of multiple
  78  * loads with the same base, but different constant offset, ie:
  79  *
  80  *    vec1 32 ssa_33 = iadd ssa_base, const_offset
  81  *    vec4 32 ssa_34 = intrinsic load_uniform (ssa_33) (base=N, 0, 0)
  82  *
  83  * Detect this, and peel out the const_offset part, to end up with:
  84  *
  85  *    vec4 32 ssa_34 = intrinsic load_uniform (ssa_base) (base=N+const_offset, 0, 0)
  86  *
  87  * This gives the other opt passes something much easier to work
  88  * with (ie. not requiring value range tracking)
  89  */
  90 static void
  91 handle_partial_const(nir_ssa_def **srcp, unsigned *offp)
  92 {
  93         if ((*srcp)->parent_instr->type != nir_instr_type_alu)
  94                 return;
  95
  96         nir_alu_instr *alu = nir_instr_as_alu((*srcp)->parent_instr);
  97         if (alu->op != nir_op_iadd)
  98                 return;
  99
 100         if (!(alu->src[0].src.is_ssa && alu->src[1].src.is_ssa))
 101                 return;
 102
 103         if (nir_src_is_const(alu->src[0].src)) {
 104                 *offp += nir_src_as_uint(alu->src[0].src);
 105                 *srcp = alu->src[1].src.ssa;
 106         } else if (nir_src_is_const(alu->src[1].src)) {
 107                 *srcp = alu->src[0].src.ssa;
 108                 *offp += nir_src_as_uint(alu->src[1].src);
 109         }
 110 }
 111
 112 static void
 113 lower_ubo_load_to_uniform(nir_intrinsic_instr *instr, nir_builder *b,
 114                                                   struct ir3_ubo_analysis_state *state)
 115 {
 116         /* We don't lower dynamic block index UBO loads to load_uniform, but we
 117          * could probably with some effort determine a block stride in number of
 118          * registers.
 119          */
 120         if (!nir_src_is_const(instr->src[0]))
 121                 return;
 122
 123         const uint32_t block = nir_src_as_uint(instr->src[0]);
 124
 125         if (block > 0) {
 126                 /* We don't lower dynamic array indexing either, but we definitely should.
 127                  * We don't have a good way of determining the range of the dynamic
 128                  * access, so for now just fall back to pulling.
 129                  */
 130                 if (!nir_src_is_const(instr->src[1]))
 131                         return;
 132
 133                 /* After gathering the UBO access ranges, we limit the total
 134                  * upload. Reject if we're now outside the range.
 135                  */
 136                 const struct ir3_ubo_range r = get_ubo_load_range(instr);
 137                 if (!(state->range[block].start <= r.start &&
 138                           r.end <= state->range[block].end))
 139                         return;
 140         }
 141
 142         b->cursor = nir_before_instr(&instr->instr);
 143
 144         nir_ssa_def *ubo_offset = nir_ssa_for_src(b, instr->src[1], 1);
 145         unsigned const_offset = 0;
 146
 147         handle_partial_const(&ubo_offset, &const_offset);
 148
 149         /* UBO offset is in bytes, but uniform offset is in units of
 150          * dwords, so we need to divide by 4 (right-shift by 2).  And
 151          * also the same for the constant part of the offset:
 152          */
 153         nir_ssa_def *new_offset = ir3_nir_try_propagate_bit_shift(b, ubo_offset, -2);
 154         nir_ssa_def *uniform_offset = NULL;
 155         if (new_offset) {
 156                 uniform_offset = new_offset;
 157         } else {
 158                 uniform_offset = nir_ushr(b, ubo_offset, nir_imm_int(b, 2));
 159         }
 160
 161         debug_assert(!(const_offset & 0x3));
 162         const_offset >>= 2;
 163
 164         const int range_offset =
 165                 (state->range[block].offset - state->range[block].start) / 4;
 166         const_offset += range_offset;
 167
 168         nir_intrinsic_instr *uniform =
 169                 nir_intrinsic_instr_create(b->shader, nir_intrinsic_load_uniform);
 170         uniform->num_components = instr->num_components;
 171         uniform->src[0] = nir_src_for_ssa(uniform_offset);
 172         nir_intrinsic_set_base(uniform, const_offset);
 173         nir_ssa_dest_init(&uniform->instr, &uniform->dest,
 174                                           uniform->num_components, instr->dest.ssa.bit_size,
 175                                           instr->dest.ssa.name);
 176         nir_builder_instr_insert(b, &uniform->instr);
 177         nir_ssa_def_rewrite_uses(&instr->dest.ssa,
 178                                                          nir_src_for_ssa(&uniform->dest.ssa));
 179
 180         nir_instr_remove(&instr->instr);
 181
 182         state->lower_count++;
 183 }
 184
 185 bool
 186 ir3_nir_analyze_ubo_ranges(nir_shader *nir, struct ir3_shader *shader)
 187 {
 188         struct ir3_ubo_analysis_state *state = &shader->ubo_state;
 189
 190         memset(state, 0, sizeof(*state));
 191
 192         nir_foreach_function(function, nir) {
 193                 if (function->impl) {
 194                         nir_foreach_block(block, function->impl) {
 195                                 nir_foreach_instr(instr, block) {
 196                                         if (instr->type == nir_instr_type_intrinsic &&
 197                                                 nir_instr_as_intrinsic(instr)->intrinsic == nir_intrinsic_load_ubo)
 198                                                 gather_ubo_ranges(nir, nir_instr_as_intrinsic(instr), state);
 199                                 }
 200                         }
 201                 }
 202         }
 203
 204         /* For now, everything we upload is accessed statically and thus will be
 205          * used by the shader. Once we can upload dynamically indexed data, we may
 206          * upload sparsely accessed arrays, at which point we probably want to
 207          * give priority to smaller UBOs, on the assumption that big UBOs will be
 208          * accessed dynamically.  Alternatively, we can track statically and
 209          * dynamically accessed ranges separately and upload static rangtes
 210          * first.
 211          */
 212         const uint32_t max_upload = 16 * 1024;
 213         uint32_t offset = 0;
 214         for (uint32_t i = 0; i < ARRAY_SIZE(state->range); i++) {
 215                 uint32_t range_size = state->range[i].end - state->range[i].start;
 216
 217                 debug_assert(offset <= max_upload);
 218                 state->range[i].offset = offset;
 219                 if (offset + range_size > max_upload) {
 220                         range_size = max_upload - offset;
 221                         state->range[i].end = state->range[i].start + range_size;
 222                 }
 223                 offset += range_size;
 224         }
 225         state->size = offset;
 226
 227         nir_foreach_function(function, nir) {
 228                 if (function->impl) {
 229                         nir_builder builder;
 230                         nir_builder_init(&builder, function->impl);
 231                         nir_foreach_block(block, function->impl) {
 232                                 nir_foreach_instr_safe(instr, block) {
 233                                         if (instr->type == nir_instr_type_intrinsic &&
 234                                                 nir_instr_as_intrinsic(instr)->intrinsic == nir_intrinsic_load_ubo)
 235                                                 lower_ubo_load_to_uniform(nir_instr_as_intrinsic(instr), &builder, state);
 236                                 }
 237                         }
 238
 239                         nir_metadata_preserve(function->impl, nir_metadata_block_index |
 240                                                                   nir_metadata_dominance);
 241                 }
 242         }
 243
 244         return state->lower_count > 0;
 245 }