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 static void
  78 lower_ubo_load_to_uniform(nir_intrinsic_instr *instr, nir_builder *b,
  79                                                   struct ir3_ubo_analysis_state *state)
  80 {
  81         /* We don't lower dynamic block index UBO loads to load_uniform, but we
  82          * could probably with some effort determine a block stride in number of
  83          * registers.
  84          */
  85         if (!nir_src_is_const(instr->src[0]))
  86                 return;
  87
  88         const uint32_t block = nir_src_as_uint(instr->src[0]);
  89
  90         if (block > 0) {
  91                 /* We don't lower dynamic array indexing either, but we definitely should.
  92                  * We don't have a good way of determining the range of the dynamic
  93                  * access, so for now just fall back to pulling.
  94                  */
  95                 if (!nir_src_is_const(instr->src[1]))
  96                         return;
  97
  98                 /* After gathering the UBO access ranges, we limit the total
  99                  * upload. Reject if we're now outside the range.
 100                  */
 101                 const struct ir3_ubo_range r = get_ubo_load_range(instr);
 102                 if (!(state->range[block].start <= r.start &&
 103                           r.end <= state->range[block].end))
 104                         return;
 105         }
 106
 107         b->cursor = nir_before_instr(&instr->instr);
 108
 109         nir_ssa_def *ubo_offset = nir_ssa_for_src(b, instr->src[1], 1);
 110         nir_ssa_def *new_offset = ir3_nir_try_propagate_bit_shift(b, ubo_offset, -2);
 111         nir_ssa_def *uniform_offset;
 112         if (new_offset)
 113                 uniform_offset = new_offset;
 114         else
 115                 uniform_offset = nir_ushr(b, ubo_offset, nir_imm_int(b, 2));
 116
 117         const int range_offset =
 118                 (state->range[block].offset - state->range[block].start) / 4;
 119
 120         nir_intrinsic_instr *uniform =
 121                 nir_intrinsic_instr_create(b->shader, nir_intrinsic_load_uniform);
 122         uniform->num_components = instr->num_components;
 123         uniform->src[0] = nir_src_for_ssa(uniform_offset);
 124         nir_intrinsic_set_base(uniform, range_offset);
 125         nir_ssa_dest_init(&uniform->instr, &uniform->dest,
 126                                           uniform->num_components, instr->dest.ssa.bit_size,
 127                                           instr->dest.ssa.name);
 128         nir_builder_instr_insert(b, &uniform->instr);
 129         nir_ssa_def_rewrite_uses(&instr->dest.ssa,
 130                                                          nir_src_for_ssa(&uniform->dest.ssa));
 131
 132         nir_instr_remove(&instr->instr);
 133
 134         state->lower_count++;
 135 }
 136
 137 bool
 138 ir3_nir_analyze_ubo_ranges(nir_shader *nir, struct ir3_shader *shader)
 139 {
 140         struct ir3_ubo_analysis_state *state = &shader->ubo_state;
 141
 142         memset(state, 0, sizeof(*state));
 143
 144         nir_foreach_function(function, nir) {
 145                 if (function->impl) {
 146                         nir_foreach_block(block, function->impl) {
 147                                 nir_foreach_instr(instr, block) {
 148                                         if (instr->type == nir_instr_type_intrinsic &&
 149                                                 nir_instr_as_intrinsic(instr)->intrinsic == nir_intrinsic_load_ubo)
 150                                                 gather_ubo_ranges(nir, nir_instr_as_intrinsic(instr), state);
 151                                 }
 152                         }
 153                 }
 154         }
 155
 156         /* For now, everything we upload is accessed statically and thus will be
 157          * used by the shader. Once we can upload dynamically indexed data, we may
 158          * upload sparsely accessed arrays, at which point we probably want to
 159          * give priority to smaller UBOs, on the assumption that big UBOs will be
 160          * accessed dynamically.  Alternatively, we can track statically and
 161          * dynamically accessed ranges separately and upload static rangtes
 162          * first.
 163          */
 164         const uint32_t max_upload = 16 * 1024;
 165         uint32_t offset = 0;
 166         for (uint32_t i = 0; i < ARRAY_SIZE(state->range); i++) {
 167                 uint32_t range_size = state->range[i].end - state->range[i].start;
 168
 169                 debug_assert(offset <= max_upload);
 170                 state->range[i].offset = offset;
 171                 if (offset + range_size > max_upload) {
 172                         range_size = max_upload - offset;
 173                         state->range[i].end = state->range[i].start + range_size;
 174                 }
 175                 offset += range_size;
 176         }
 177         state->size = offset;
 178
 179         nir_foreach_function(function, nir) {
 180                 if (function->impl) {
 181                         nir_builder builder;
 182                         nir_builder_init(&builder, function->impl);
 183                         nir_foreach_block(block, function->impl) {
 184                                 nir_foreach_instr_safe(instr, block) {
 185                                         if (instr->type == nir_instr_type_intrinsic &&
 186                                                 nir_instr_as_intrinsic(instr)->intrinsic == nir_intrinsic_load_ubo)
 187                                                 lower_ubo_load_to_uniform(nir_instr_as_intrinsic(instr), &builder, state);
 188                                 }
 189                         }
 190
 191                         nir_metadata_preserve(function->impl, nir_metadata_block_index |
 192                                                                   nir_metadata_dominance);
 193                 }
 194         }
 195
 196         return state->lower_count > 0;
 197 }