2 * UCW Library -- Universal Sorter: Governing Routines
4 * (c) 2007 Martin Mares <mj@ucw.cz>
6 * This software may be freely distributed and used according to the terms
7 * of the GNU Lesser General Public License.
11 #include "lib/fastbuf.h"
12 #include "lib/mempool.h"
13 #include "lib/sorter/common.h"
23 gettimeofday(&tv, NULL);
24 return (u64)tv.tv_sec * 1000 + tv.tv_usec / 1000;
28 sorter_start_timer(struct sort_context *ctx)
30 ctx->start_time = sorter_clock();
34 sorter_speed(struct sort_context *ctx, u64 size)
36 u64 stop_time = sorter_clock();
39 if (stop_time <= ctx->start_time)
41 return (uns)((double)size / (1<<20) * 1000 / (stop_time-ctx->start_time));
45 sorter_presort(struct sort_context *ctx, struct sort_bucket *in, struct sort_bucket *out, struct sort_bucket *out_only)
47 sorter_alloc_buf(ctx);
48 if (in->flags & SBF_CUSTOM_PRESORT)
50 struct fastbuf *f = sbuck_write(out);
52 return ctx->custom_presort(f, ctx->big_buf, ctx->big_buf_size); // FIXME: out_only optimization?
54 return ctx->internal_sort(ctx, in, out, out_only);
57 static inline struct sort_bucket *
58 sbuck_join_to(struct sort_bucket *b)
60 if (sorter_debug & SORT_DEBUG_NO_JOIN)
63 struct sort_bucket *out = (struct sort_bucket *) b->n.prev; // Such bucket is guaranteed to exist
64 return (out->flags & SBF_FINAL) ? out : NULL;
68 sorter_join(struct sort_bucket *b)
70 struct sort_bucket *join = (struct sort_bucket *) b->n.prev;
71 ASSERT(join->flags & SBF_FINAL);
74 if (!sbuck_has_file(join))
76 // The final bucket doesn't have any file associated yet, so replace
77 // it with the new bucket.
78 SORT_XTRACE(2, "Replaced final bucket");
79 b->flags |= SBF_FINAL;
84 SORT_TRACE("Copying to output file: %s", F_BSIZE(b));
85 struct fastbuf *src = sbuck_read(b);
86 struct fastbuf *dest = sbuck_write(join);
87 bbcopy(src, dest, ~0U);
93 sorter_twoway(struct sort_context *ctx, struct sort_bucket *b)
95 struct sort_bucket *ins[3] = { NULL }, *outs[3] = { NULL };
96 cnode *list_pos = b->n.prev;
97 struct sort_bucket *join = sbuck_join_to(b);
99 if (!(sorter_debug & SORT_DEBUG_NO_PRESORT) || (b->flags & SBF_CUSTOM_PRESORT))
101 SORT_XTRACE(2, "%s", ((b->flags & SBF_CUSTOM_PRESORT) ? "Custom presorting" : "Presorting"));
102 sorter_start_timer(ctx);
103 ins[0] = sbuck_new(ctx);
104 if (!sorter_presort(ctx, b, ins[0], join ? : ins[0]))
106 SORT_XTRACE(((b->flags & SBF_SOURCE) ? 1 : 2), "Sorted in memory");
110 clist_insert_after(&ins[0]->n, list_pos);
115 ins[1] = sbuck_new(ctx);
117 while (sorter_presort(ctx, b, ins[i], ins[i]))
120 SORT_TRACE("Presorting pass (%d+%d runs, %s+%s, %dMB/s)",
121 ins[0]->runs, ins[1]->runs,
122 F_BSIZE(ins[0]), F_BSIZE(ins[1]),
123 sorter_speed(ctx, sbuck_size(ins[0]) + sbuck_size(ins[1])));
127 SORT_XTRACE(2, "Presorting disabled");
131 SORT_XTRACE(2, "Main sorting");
135 sorter_start_timer(ctx);
136 if (ins[0]->runs == 1 && ins[1]->runs == 1 && join)
138 // This is guaranteed to produce a single run, so join if possible
141 ctx->twoway_merge(ctx, ins, outs);
142 ASSERT(outs[0]->runs == 2);
144 SORT_TRACE("Mergesort pass %d (final run, %s, %dMB/s)", pass, F_BSIZE(outs[0]), sorter_speed(ctx, sbuck_size(outs[0])));
149 outs[0] = sbuck_new(ctx);
150 outs[1] = sbuck_new(ctx);
152 ctx->twoway_merge(ctx, ins, outs);
153 SORT_TRACE("Mergesort pass %d (%d+%d runs, %s+%s, %dMB/s)", pass,
154 outs[0]->runs, outs[1]->runs,
155 F_BSIZE(outs[0]), F_BSIZE(outs[1]),
156 sorter_speed(ctx, sbuck_size(outs[0]) + sbuck_size(outs[1])));
159 memcpy(ins, outs, 3*sizeof(struct sort_bucket *));
160 } while (sbuck_have(ins[1]));
163 clist_insert_after(&ins[0]->n, list_pos);
167 sorter_radix_p(struct sort_context *ctx, struct sort_bucket *b)
169 return b->hash_bits && ctx->radix_split &&
170 !(sorter_debug & SORT_DEBUG_NO_RADIX) &&
171 sbuck_size(b) > (sh_off_t)sorter_bufsize;
175 sorter_radix(struct sort_context *ctx, struct sort_bucket *b)
177 uns bits = MIN(b->hash_bits, 4); /* FIXME */
178 uns nbuck = 1 << bits;
179 SORT_XTRACE(2, "Running radix sort on %s with %d bits of %d", F_BSIZE(b), bits, b->hash_bits);
180 sorter_start_timer(ctx);
182 struct sort_bucket *outs[nbuck];
183 for (uns i=nbuck; i--; )
185 outs[i] = sbuck_new(ctx);
186 outs[i]->hash_bits = b->hash_bits - bits;
187 clist_insert_after(&outs[i]->n, &b->n);
190 ctx->radix_split(ctx, b, outs, b->hash_bits - bits, bits);
192 u64 min = ~0U, max = 0, sum = 0;
193 for (uns i=0; i<nbuck; i++)
195 u64 s = sbuck_size(outs[i]);
201 SORT_TRACE("Radix split (%d buckets, %s min, %s max, %s avg, %dMB/s)", nbuck,
202 F_SIZE(min), F_SIZE(max), F_SIZE(sum / nbuck), sorter_speed(ctx, sum));
207 sorter_run(struct sort_context *ctx)
209 ctx->pool = mp_new(4096);
210 clist_init(&ctx->bucket_list);
212 /* FIXME: Remember to test sorting of empty files */
214 // Create bucket containing the source
215 struct sort_bucket *bin = sbuck_new(ctx);
216 bin->flags = SBF_SOURCE | SBF_OPEN_READ;
217 if (ctx->custom_presort)
218 bin->flags |= SBF_CUSTOM_PRESORT;
220 bin->fb = ctx->in_fb;
222 bin->size = ctx->in_size; /* Sizes should be either sh_off_t or u64, not both; beware of ~0U */
223 bin->hash_bits = ctx->hash_bits;
224 clist_add_tail(&ctx->bucket_list, &bin->n);
225 SORT_XTRACE(2, "Input size: %s", (ctx->in_size == ~(u64)0 ? (byte*)"unknown" : F_BSIZE(bin)));
227 // Create bucket for the output
228 struct sort_bucket *bout = sbuck_new(ctx);
229 bout->flags = SBF_FINAL;
230 if (bout->fb = ctx->out_fb)
231 bout->flags |= SBF_OPEN_WRITE;
234 clist_add_head(&ctx->bucket_list, &bout->n);
236 struct sort_bucket *b;
237 while (bout = clist_head(&ctx->bucket_list), b = clist_next(&ctx->bucket_list, &bout->n))
241 else if (b->runs == 1)
243 else if (sorter_radix_p(ctx, b))
244 sorter_radix(ctx, b);
246 sorter_twoway(ctx, b);
249 sorter_free_buf(ctx);
250 sbuck_write(bout); // Force empty bucket to a file
251 SORT_XTRACE(2, "Final size: %s", F_BSIZE(bout));
252 ctx->out_fb = sbuck_read(bout);