brw_nir_analyze_ubo_ranges.c 11.7 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126
/*
 * Copyright © 2015 Intel Corporation
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
 * IN THE SOFTWARE.
 */

#include "brw_nir.h"
#include "compiler/nir/nir.h"
#include "util/u_dynarray.h"

/**
 * \file brw_nir_analyze_ubo_ranges.c
 *
 * This pass decides which portions of UBOs to upload as push constants,
 * so shaders can access them as part of the thread payload, rather than
 * having to issue expensive memory reads to pull the data.
 *
 * The 3DSTATE_CONSTANT_* mechanism can push data from up to 4 different
 * buffers, in GRF (256-bit/32-byte) units.
 *
 * To do this, we examine NIR load_ubo intrinsics, recording the number of
 * loads at each offset.  We track offsets at a 32-byte granularity, so even
 * fields with a bit of padding between them tend to fall into contiguous
 * ranges.  We build a list of these ranges, tracking their "cost" (number
 * of registers required) and "benefit" (number of pull loads eliminated
 * by pushing the range).  We then sort the list to obtain the four best
 * ranges (most benefit for the least cost).
 */

struct ubo_range_entry
{
   struct brw_ubo_range range;
   int benefit;
};

static int
score(const struct ubo_range_entry *entry)
{
   return 2 * entry->benefit - entry->range.length;
}

/**
 * Compares score for two UBO range entries.
 *
 * For a descending qsort().
 */
static int
cmp_ubo_range_entry(const void *va, const void *vb)
{
   const struct ubo_range_entry *a = va;
   const struct ubo_range_entry *b = vb;

   /* Rank based on scores */
   int delta = score(b) - score(a);

   /* Then use the UBO block index as a tie-breaker */
   if (delta == 0)
      delta = b->range.block - a->range.block;

   /* Finally use the UBO offset as a second tie-breaker */
   if (delta == 0)
      delta = b->range.block - a->range.block;

   return delta;
}

struct ubo_block_info
{
   /* Each bit in the offsets bitfield represents a 32-byte section of data.
    * If it's set to one, there is interesting UBO data at that offset.  If
    * not, there's a "hole" - padding between data - or just nothing at all.
    */
   uint64_t offsets;
   uint8_t uses[64];
};

struct ubo_analysis_state
{
   struct hash_table *blocks;
   bool uses_regular_uniforms;
};

static struct ubo_block_info *
get_block_info(struct ubo_analysis_state *state, int block)
{
   uint32_t hash = block + 1;
   void *key = (void *) (uintptr_t) hash;

   struct hash_entry *entry =
      _mesa_hash_table_search_pre_hashed(state->blocks, hash, key);

   if (entry)
      return (struct ubo_block_info *) entry->data;

   struct ubo_block_info *info =
      rzalloc(state->blocks, struct ubo_block_info);
   _mesa_hash_table_insert_pre_hashed(state->blocks, hash, key, info);

   return info;
}

static void
analyze_ubos_block(struct ubo_analysis_state *state, nir_block *block)
{
   nir_foreach_instr(instr, block) {
      if (instr->type != nir_instr_type_intrinsic)
         continue;

      nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
127 128 129 130 131 132 133 134 135 136 137 138 139
      switch (intrin->intrinsic) {
      case nir_intrinsic_load_uniform:
      case nir_intrinsic_image_deref_load:
      case nir_intrinsic_image_deref_store:
      case nir_intrinsic_image_deref_atomic_add:
      case nir_intrinsic_image_deref_atomic_min:
      case nir_intrinsic_image_deref_atomic_max:
      case nir_intrinsic_image_deref_atomic_and:
      case nir_intrinsic_image_deref_atomic_or:
      case nir_intrinsic_image_deref_atomic_xor:
      case nir_intrinsic_image_deref_atomic_exchange:
      case nir_intrinsic_image_deref_atomic_comp_swap:
      case nir_intrinsic_image_deref_size:
140 141 142
         state->uses_regular_uniforms = true;
         continue;

143 144 145 146 147 148 149
      case nir_intrinsic_load_ubo:
         break; /* Fall through to the analysis below */

      default:
         continue; /* Not a uniform or UBO intrinsic */
      }

150 151 152 153 154
      if (nir_src_is_const(intrin->src[0]) &&
          nir_src_is_const(intrin->src[1])) {
         const int block = nir_src_as_uint(intrin->src[0]);
         const unsigned byte_offset = nir_src_as_uint(intrin->src[1]);
         const int offset = byte_offset / 32;
155

156 157 158 159 160 161
         /* Avoid shifting by larger than the width of our bitfield, as this
          * is undefined in C.  Even if we require multiple bits to represent
          * the entire value, it's OK to record a partial value - the backend
          * is capable of falling back to pull loads for later components of
          * vectors, as it has to shrink ranges for other reasons anyway.
          */
162 163 164
         if (offset >= 64)
            continue;

165 166 167
         /* The value might span multiple 32-byte chunks. */
         const int bytes = nir_intrinsic_dest_components(intrin) *
                           (nir_dest_bit_size(intrin->dest) / 8);
168 169
         const int start = ROUND_DOWN_TO(byte_offset, 32);
         const int end = ALIGN(byte_offset + bytes, 32);
170 171
         const int chunks = (end - start) / 32;

172 173 174
         /* TODO: should we count uses in loops as higher benefit? */

         struct ubo_block_info *info = get_block_info(state, block);
175
         info->offsets |= ((1ull << chunks) - 1) << offset;
176 177 178 179 180 181 182 183 184 185 186 187 188
         info->uses[offset]++;
      }
   }
}

static void
print_ubo_entry(FILE *file,
                const struct ubo_range_entry *entry,
                struct ubo_analysis_state *state)
{
   struct ubo_block_info *info = get_block_info(state, entry->range.block);

   fprintf(file,
189
           "block %2d, start %2d, length %2d, bits = %"PRIx64", "
190 191 192 193 194 195 196 197
           "benefit %2d, cost %2d, score = %2d\n",
           entry->range.block, entry->range.start, entry->range.length,
           info->offsets, entry->benefit, entry->range.length, score(entry));
}

void
brw_nir_analyze_ubo_ranges(const struct brw_compiler *compiler,
                           nir_shader *nir,
198
                           const struct brw_vs_prog_key *vs_key,
199 200 201 202 203
                           struct brw_ubo_range out_ranges[4])
{
   const struct gen_device_info *devinfo = compiler->devinfo;

   if ((devinfo->gen <= 7 && !devinfo->is_haswell) ||
204
       !compiler->scalar_stage[nir->info.stage]) {
205 206 207 208 209 210 211 212 213 214 215 216
      memset(out_ranges, 0, 4 * sizeof(struct brw_ubo_range));
      return;
   }

   void *mem_ctx = ralloc_context(NULL);

   struct ubo_analysis_state state = {
      .uses_regular_uniforms = false,
      .blocks =
         _mesa_hash_table_create(mem_ctx, NULL, _mesa_key_pointer_equal),
   };

217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233
   switch (nir->info.stage) {
   case MESA_SHADER_VERTEX:
      if (vs_key && vs_key->nr_userclip_plane_consts > 0)
         state.uses_regular_uniforms = true;
      break;

   case MESA_SHADER_COMPUTE:
      /* Compute shaders use push constants to get the subgroup ID so it's
       * best to just assume some system values are pushed.
       */
      state.uses_regular_uniforms = true;
      break;

   default:
      break;
   }

234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343
   /* Walk the IR, recording how many times each UBO block/offset is used. */
   nir_foreach_function(function, nir) {
      if (function->impl) {
         nir_foreach_block(block, function->impl) {
            analyze_ubos_block(&state, block);
         }
      }
   }

   /* Find ranges: a block, starting 32-byte offset, and length. */
   struct util_dynarray ranges;
   util_dynarray_init(&ranges, mem_ctx);

   hash_table_foreach(state.blocks, entry) {
      const int b = entry->hash - 1;
      const struct ubo_block_info *info = entry->data;
      uint64_t offsets = info->offsets;

      /* Walk through the offsets bitfield, finding contiguous regions of
       * set bits:
       *
       *   0000000001111111111111000000000000111111111111110000000011111100
       *            ^^^^^^^^^^^^^            ^^^^^^^^^^^^^^        ^^^^^^
       *
       * Each of these will become a UBO range.
       */
      while (offsets != 0) {
         /* Find the first 1 in the offsets bitfield.  This represents the
          * start of a range of interesting UBO data.  Make it zero-indexed.
          */
         int first_bit = ffsll(offsets) - 1;

         /* Find the first 0 bit in offsets beyond first_bit.  To find the
          * first zero bit, we find the first 1 bit in the complement.  In
          * order to ignore bits before first_bit, we mask off those bits.
          */
         int first_hole = ffsll(~offsets & ~((1ull << first_bit) - 1)) - 1;

         if (first_hole == -1) {
            /* If we didn't find a hole, then set it to the end of the
             * bitfield.  There are no more ranges to process.
             */
            first_hole = 64;
            offsets = 0;
         } else {
            /* We've processed all bits before first_hole.  Mask them off. */
            offsets &= ~((1ull << first_hole) - 1);
         }

         struct ubo_range_entry *entry =
            util_dynarray_grow(&ranges, sizeof(struct ubo_range_entry));

         entry->range.block = b;
         entry->range.start = first_bit;
         /* first_hole is one beyond the end, so we don't need to add 1 */
         entry->range.length = first_hole - first_bit;
         entry->benefit = 0;

         for (int i = 0; i < entry->range.length; i++)
            entry->benefit += info->uses[first_bit + i];
      }
   }

   int nr_entries = ranges.size / sizeof(struct ubo_range_entry);

   if (0) {
      util_dynarray_foreach(&ranges, struct ubo_range_entry, entry) {
         print_ubo_entry(stderr, entry, &state);
      }
   }

   /* TODO: Consider combining ranges.
    *
    * We can only push 3-4 ranges via 3DSTATE_CONSTANT_XS.  If there are
    * more ranges, and two are close by with only a small hole, it may be
    * worth combining them.  The holes will waste register space, but the
    * benefit of removing pulls may outweigh that cost.
    */

   /* Sort the list so the most beneficial ranges are at the front. */
   qsort(ranges.data, nr_entries, sizeof(struct ubo_range_entry),
         cmp_ubo_range_entry);

   struct ubo_range_entry *entries = ranges.data;

   /* Return the top 4 or so.  We drop by one if regular uniforms are in
    * use, assuming one push buffer will be dedicated to those.  We may
    * also only get 3 on Haswell if we can't write INSTPM.
    *
    * The backend may need to shrink these ranges to ensure that they
    * don't exceed the maximum push constant limits.  It can simply drop
    * the tail of the list, as that's the least valuable portion.  We
    * unfortunately can't truncate it here, because we don't know what
    * the backend is planning to do with regular uniforms.
    */
   const int max_ubos = (compiler->constant_buffer_0_is_relative ? 3 : 4) -
                        state.uses_regular_uniforms;
   nr_entries = MIN2(nr_entries, max_ubos);

   for (int i = 0; i < nr_entries; i++) {
      out_ranges[i] = entries[i].range;
   }
   for (int i = nr_entries; i < 4; i++) {
      out_ranges[i].block = 0;
      out_ranges[i].start = 0;
      out_ranges[i].length = 0;
   }

   ralloc_free(ranges.mem_ctx);
}