• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright © 2015 Intel Corporation
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
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21  * IN THE SOFTWARE.
22  */
23 
24 #include "brw_nir.h"
25 #include "compiler/nir/nir.h"
26 #include "util/u_dynarray.h"
27 
28 /**
29  * \file brw_nir_analyze_ubo_ranges.c
30  *
31  * This pass decides which portions of UBOs to upload as push constants,
32  * so shaders can access them as part of the thread payload, rather than
33  * having to issue expensive memory reads to pull the data.
34  *
35  * The 3DSTATE_CONSTANT_* mechanism can push data from up to 4 different
36  * buffers, in GRF (256-bit/32-byte) units.
37  *
38  * To do this, we examine NIR load_ubo intrinsics, recording the number of
39  * loads at each offset.  We track offsets at a 32-byte granularity, so even
40  * fields with a bit of padding between them tend to fall into contiguous
41  * ranges.  We build a list of these ranges, tracking their "cost" (number
42  * of registers required) and "benefit" (number of pull loads eliminated
43  * by pushing the range).  We then sort the list to obtain the four best
44  * ranges (most benefit for the least cost).
45  */
46 
47 struct ubo_range_entry
48 {
49    struct brw_ubo_range range;
50    int benefit;
51 };
52 
53 static int
score(const struct ubo_range_entry * entry)54 score(const struct ubo_range_entry *entry)
55 {
56    return 2 * entry->benefit - entry->range.length;
57 }
58 
59 /**
60  * Compares score for two UBO range entries.
61  *
62  * For a descending qsort().
63  */
64 static int
cmp_ubo_range_entry(const void * va,const void * vb)65 cmp_ubo_range_entry(const void *va, const void *vb)
66 {
67    const struct ubo_range_entry *a = va;
68    const struct ubo_range_entry *b = vb;
69 
70    /* Rank based on scores */
71    int delta = score(b) - score(a);
72 
73    /* Then use the UBO block index as a tie-breaker */
74    if (delta == 0)
75       delta = b->range.block - a->range.block;
76 
77    /* Finally use the UBO offset as a second tie-breaker */
78    if (delta == 0)
79       delta = b->range.block - a->range.block;
80 
81    return delta;
82 }
83 
84 struct ubo_block_info
85 {
86    /* Each bit in the offsets bitfield represents a 32-byte section of data.
87     * If it's set to one, there is interesting UBO data at that offset.  If
88     * not, there's a "hole" - padding between data - or just nothing at all.
89     */
90    uint64_t offsets;
91    uint8_t uses[64];
92 };
93 
94 struct ubo_analysis_state
95 {
96    struct hash_table *blocks;
97    bool uses_regular_uniforms;
98 };
99 
100 static struct ubo_block_info *
get_block_info(struct ubo_analysis_state * state,int block)101 get_block_info(struct ubo_analysis_state *state, int block)
102 {
103    uint32_t hash = block + 1;
104    void *key = (void *) (uintptr_t) hash;
105 
106    struct hash_entry *entry =
107       _mesa_hash_table_search_pre_hashed(state->blocks, hash, key);
108 
109    if (entry)
110       return (struct ubo_block_info *) entry->data;
111 
112    struct ubo_block_info *info =
113       rzalloc(state->blocks, struct ubo_block_info);
114    _mesa_hash_table_insert_pre_hashed(state->blocks, hash, key, info);
115 
116    return info;
117 }
118 
119 static void
analyze_ubos_block(struct ubo_analysis_state * state,nir_block * block)120 analyze_ubos_block(struct ubo_analysis_state *state, nir_block *block)
121 {
122    nir_foreach_instr(instr, block) {
123       if (instr->type != nir_instr_type_intrinsic)
124          continue;
125 
126       nir_intrinsic_instr *intrin = nir_instr_as_intrinsic(instr);
127       if (intrin->intrinsic == nir_intrinsic_load_uniform)
128          state->uses_regular_uniforms = true;
129 
130       if (intrin->intrinsic != nir_intrinsic_load_ubo)
131          continue;
132 
133       nir_const_value *block_const = nir_src_as_const_value(intrin->src[0]);
134       nir_const_value *offset_const = nir_src_as_const_value(intrin->src[1]);
135 
136       if (block_const && offset_const) {
137          const int block = block_const->u32[0];
138          const int offset = offset_const->u32[0] / 32;
139 
140          /* Won't fit in our bitfield */
141          if (offset >= 64)
142             continue;
143 
144          /* TODO: should we count uses in loops as higher benefit? */
145 
146          struct ubo_block_info *info = get_block_info(state, block);
147          info->offsets |= 1ull << offset;
148          info->uses[offset]++;
149       }
150    }
151 }
152 
153 static void
print_ubo_entry(FILE * file,const struct ubo_range_entry * entry,struct ubo_analysis_state * state)154 print_ubo_entry(FILE *file,
155                 const struct ubo_range_entry *entry,
156                 struct ubo_analysis_state *state)
157 {
158    struct ubo_block_info *info = get_block_info(state, entry->range.block);
159 
160    fprintf(file,
161            "block %2d, start %2d, length %2d, bits = %zx, "
162            "benefit %2d, cost %2d, score = %2d\n",
163            entry->range.block, entry->range.start, entry->range.length,
164            info->offsets, entry->benefit, entry->range.length, score(entry));
165 }
166 
167 void
brw_nir_analyze_ubo_ranges(const struct brw_compiler * compiler,nir_shader * nir,struct brw_ubo_range out_ranges[4])168 brw_nir_analyze_ubo_ranges(const struct brw_compiler *compiler,
169                            nir_shader *nir,
170                            struct brw_ubo_range out_ranges[4])
171 {
172    const struct gen_device_info *devinfo = compiler->devinfo;
173 
174    if ((devinfo->gen <= 7 && !devinfo->is_haswell) ||
175        !compiler->scalar_stage[nir->info.stage]) {
176       memset(out_ranges, 0, 4 * sizeof(struct brw_ubo_range));
177       return;
178    }
179 
180    void *mem_ctx = ralloc_context(NULL);
181 
182    struct ubo_analysis_state state = {
183       .uses_regular_uniforms = false,
184       .blocks =
185          _mesa_hash_table_create(mem_ctx, NULL, _mesa_key_pointer_equal),
186    };
187 
188    /* Walk the IR, recording how many times each UBO block/offset is used. */
189    nir_foreach_function(function, nir) {
190       if (function->impl) {
191          nir_foreach_block(block, function->impl) {
192             analyze_ubos_block(&state, block);
193          }
194       }
195    }
196 
197    /* Find ranges: a block, starting 32-byte offset, and length. */
198    struct util_dynarray ranges;
199    util_dynarray_init(&ranges, mem_ctx);
200 
201    struct hash_entry *entry;
202    hash_table_foreach(state.blocks, entry) {
203       const int b = entry->hash - 1;
204       const struct ubo_block_info *info = entry->data;
205       uint64_t offsets = info->offsets;
206 
207       /* Walk through the offsets bitfield, finding contiguous regions of
208        * set bits:
209        *
210        *   0000000001111111111111000000000000111111111111110000000011111100
211        *            ^^^^^^^^^^^^^            ^^^^^^^^^^^^^^        ^^^^^^
212        *
213        * Each of these will become a UBO range.
214        */
215       while (offsets != 0) {
216          /* Find the first 1 in the offsets bitfield.  This represents the
217           * start of a range of interesting UBO data.  Make it zero-indexed.
218           */
219          int first_bit = ffsll(offsets) - 1;
220 
221          /* Find the first 0 bit in offsets beyond first_bit.  To find the
222           * first zero bit, we find the first 1 bit in the complement.  In
223           * order to ignore bits before first_bit, we mask off those bits.
224           */
225          int first_hole = ffsll(~offsets & ~((1ull << first_bit) - 1)) - 1;
226 
227          if (first_hole == -1) {
228             /* If we didn't find a hole, then set it to the end of the
229              * bitfield.  There are no more ranges to process.
230              */
231             first_hole = 64;
232             offsets = 0;
233          } else {
234             /* We've processed all bits before first_hole.  Mask them off. */
235             offsets &= ~((1ull << first_hole) - 1);
236          }
237 
238          struct ubo_range_entry *entry =
239             util_dynarray_grow(&ranges, sizeof(struct ubo_range_entry));
240 
241          entry->range.block = b;
242          entry->range.start = first_bit;
243          /* first_hole is one beyond the end, so we don't need to add 1 */
244          entry->range.length = first_hole - first_bit;
245          entry->benefit = 0;
246 
247          for (int i = 0; i < entry->range.length; i++)
248             entry->benefit += info->uses[first_bit + i];
249       }
250    }
251 
252    int nr_entries = ranges.size / sizeof(struct ubo_range_entry);
253 
254    if (0) {
255       util_dynarray_foreach(&ranges, struct ubo_range_entry, entry) {
256          print_ubo_entry(stderr, entry, &state);
257       }
258    }
259 
260    /* TODO: Consider combining ranges.
261     *
262     * We can only push 3-4 ranges via 3DSTATE_CONSTANT_XS.  If there are
263     * more ranges, and two are close by with only a small hole, it may be
264     * worth combining them.  The holes will waste register space, but the
265     * benefit of removing pulls may outweigh that cost.
266     */
267 
268    /* Sort the list so the most beneficial ranges are at the front. */
269    qsort(ranges.data, nr_entries, sizeof(struct ubo_range_entry),
270          cmp_ubo_range_entry);
271 
272    struct ubo_range_entry *entries = ranges.data;
273 
274    /* Return the top 4 or so.  We drop by one if regular uniforms are in
275     * use, assuming one push buffer will be dedicated to those.  We may
276     * also only get 3 on Haswell if we can't write INSTPM.
277     *
278     * The backend may need to shrink these ranges to ensure that they
279     * don't exceed the maximum push constant limits.  It can simply drop
280     * the tail of the list, as that's the least valuable portion.  We
281     * unfortunately can't truncate it here, because we don't know what
282     * the backend is planning to do with regular uniforms.
283     */
284    const int max_ubos = (compiler->constant_buffer_0_is_relative ? 3 : 4) -
285                         state.uses_regular_uniforms;
286    nr_entries = MIN2(nr_entries, max_ubos);
287 
288    for (int i = 0; i < nr_entries; i++) {
289       out_ranges[i] = entries[i].range;
290    }
291    for (int i = nr_entries; i < 4; i++) {
292       out_ranges[i].block = 0;
293       out_ranges[i].start = 0;
294       out_ranges[i].length = 0;
295    }
296 
297    ralloc_free(ranges.mem_ctx);
298 }
299