4 // A reimplementation of parts of the TurboPFor codecs, using the same
5 // storage format. These are not as fast as the reference implementation
6 // (about 80% of the performance, averaged over a real plocate corpus),
7 // and do not support the same breadth of codecs (in particular, only
8 // delta-plus-1 is implemented, and only 32-bit docids are tested),
9 // but aim to be more portable and (ideally) easier-to-understand.
10 // In particular, they will compile on x86 without SSE4.1 or AVX support.
12 // The main reference is https://michael.stapelberg.ch/posts/2019-02-05-turbopfor-analysis/,
13 // although some implementation details have been worked out by studying the
24 #if defined(__i386__) || defined(__x86_64__)
25 #define COULD_HAVE_SSE2
26 #include <immintrin.h>
29 // Forward declarations to declare to the template code below that they exist.
30 // (These must seemingly be non-templates for function multiversioning to work.)
31 __attribute__((target("default")))
32 const unsigned char *decode_for_interleaved_128_32(const unsigned char *in, uint32_t *out);
33 __attribute__((target("default")))
34 const unsigned char *decode_pfor_bitmap_interleaved_128_32(const unsigned char *in, uint32_t *out);
35 __attribute__((target("default")))
36 const unsigned char *decode_pfor_vb_interleaved_128_32(const unsigned char *in, uint32_t *out);
38 #ifdef COULD_HAVE_SSE2
39 __attribute__((target("sse2")))
40 const unsigned char *decode_for_interleaved_128_32(const unsigned char *in, uint32_t *out);
41 __attribute__((target("sse2")))
42 const unsigned char *decode_pfor_bitmap_interleaved_128_32(const unsigned char *in, uint32_t *out);
43 __attribute__((target("sse2")))
44 const unsigned char *decode_pfor_vb_interleaved_128_32(const unsigned char *in, uint32_t *out);
47 constexpr uint32_t mask_for_bits(unsigned bit_width)
49 if (bit_width == 32) {
52 return (1U << bit_width) - 1;
57 Docid read_le(const void *in)
60 memcpy(&val, in, sizeof(val));
61 if constexpr (sizeof(Docid) == 8) {
63 } else if constexpr (sizeof(Docid) == 4) {
65 } else if constexpr (sizeof(Docid) == 2) {
67 } else if constexpr (sizeof(Docid) == 1) {
74 // Reads a single value with an encoding that looks a bit like PrefixVarint.
75 // It's unclear why this doesn't use the varbyte encoding.
77 const unsigned char *read_baseval(const unsigned char *in, Docid *out)
79 //fprintf(stderr, "baseval: 0x%02x 0x%02x 0x%02x 0x%02x\n", in[0], in[1], in[2], in[3]);
83 } else if (*in < 192) {
84 *out = ((uint32_t(in[0]) << 8) | uint32_t(in[1])) & 0x3fff;
86 } else if (*in < 224) {
87 *out = ((uint32_t(in[0]) << 16) |
88 (uint32_t(in[2]) << 8) |
89 (uint32_t(in[1]))) & 0x1fffff;
92 assert(false); // Not implemented.
97 const unsigned char *read_vb(const unsigned char *in, Docid *out)
102 } else if (*in <= 240) {
103 *out = ((uint32_t(in[0] - 177) << 8) | uint32_t(in[1])) + 177;
105 } else if (*in <= 248) {
106 *out = ((uint32_t(in[0] - 241) << 16) | read_le<uint16_t>(in + 1)) + 16561;
108 } else if (*in == 249) {
109 *out = (uint32_t(in[1])) |
110 (uint32_t(in[2]) << 8) |
111 (uint32_t(in[3]) << 16);
113 } else if (*in == 250) {
114 *out = read_le<uint32_t>(in + 1);
123 BitReader(const unsigned char *in, unsigned bits)
124 : in(in), bits(bits), mask(mask_for_bits(bits)) {}
127 uint32_t val = (read_le<uint32_t>(in) >> bits_used) & mask;
137 const unsigned char *in;
140 unsigned bits_used = 0;
143 template<unsigned NumStreams>
144 struct InterleavedBitReader {
146 InterleavedBitReader(const unsigned char *in, unsigned bits)
147 : in(in), bits(bits), mask(mask_for_bits(bits)) {}
151 if (bits_used + bits > 32) {
152 val = (read_le<uint32_t>(in) >> bits_used) | (read_le<uint32_t>(in + Stride) << (32 - bits_used));
154 val = (read_le<uint32_t>(in) >> bits_used);
158 in += Stride * (bits_used / 32);
165 static constexpr unsigned Stride = NumStreams * sizeof(uint32_t);
166 const unsigned char *in;
169 unsigned bits_used = 0;
172 #ifdef COULD_HAVE_SSE2
173 struct InterleavedBitReaderSSE2 {
175 InterleavedBitReaderSSE2(const unsigned char *in, unsigned bits)
176 : in(reinterpret_cast<const __m128i *>(in)), bits(bits), mask(_mm_set1_epi32(mask_for_bits(bits))) {}
178 __m128i val = _mm_srli_epi32(_mm_loadu_si128(in), bits_used);
179 if (bits_used + bits > 32) {
180 __m128i val_upper = _mm_slli_epi32(_mm_loadu_si128(in + 1), 32 - bits_used);
181 val = _mm_or_si128(val, val_upper);
183 val = _mm_and_si128(val, mask);
186 in += bits_used / 32;
195 unsigned bits_used = 0;
199 // Does not properly account for overflow.
200 inline unsigned div_round_up(unsigned val, unsigned div)
202 return (val + div - 1) / div;
205 inline unsigned bytes_for_packed_bits(unsigned num, unsigned bit_width)
207 return div_round_up(num * bit_width, CHAR_BIT);
210 // Constant block. Layout:
212 // - Bit width (6 bits) | type << 6
213 // - Base values (<bits> bits, rounded up to nearest byte)
214 template<class Docid>
215 const unsigned char *decode_constant(const unsigned char *in, unsigned num, Docid *out)
217 const unsigned bit_width = *in++ & 0x3f;
218 Docid val = read_le<Docid>(in);
219 if (bit_width < sizeof(Docid) * 8) {
220 val &= mask_for_bits(bit_width);
223 Docid prev_val = out[-1];
224 for (unsigned i = 0; i < num; ++i) {
225 out[i] = prev_val = val + prev_val + 1;
227 return in + div_round_up(bit_width, 8);
230 // FOR block (ie., PFor without exceptions). Layout:
232 // - Bit width (6 bits) | type << 6
233 // - Base values (<num> values of <bits> bits, rounded up to a multiple of 32 values)
234 template<class Docid>
235 const unsigned char *decode_for(const unsigned char *in, unsigned num, Docid *out)
237 const unsigned bit_width = *in++ & 0x3f;
239 Docid prev_val = out[-1];
240 BitReader bs(in, bit_width);
241 for (unsigned i = 0; i < num; ++i) {
242 prev_val = out[i] = bs.read() + prev_val + 1;
244 return in + bytes_for_packed_bits(num, bit_width);
247 #ifdef COULD_HAVE_SSE2
248 class DeltaDecoderSSE2 {
250 DeltaDecoderSSE2(uint32_t prev_val) : prev_val(_mm_set1_epi32(prev_val)) {}
251 __m128i decode(__m128i val) {
252 val = _mm_add_epi32(val, _mm_slli_si128(val, 4));
253 val = _mm_add_epi32(val, _mm_slli_si128(val, 8));
254 val = _mm_add_epi32(val, _mm_add_epi32(prev_val, delta));
255 prev_val = _mm_shuffle_epi32(val, _MM_SHUFFLE(3, 3, 3, 3));
260 // Use 4/3/2/1 as delta instead of fixed 1, so that we can do the prev_val + delta
261 // in parallel with something else.
262 const __m128i delta = _mm_set_epi32(4, 3, 2, 1);
267 template<unsigned BlockSize>
268 __attribute__((target("sse2")))
269 inline void delta_decode_sse2(uint32_t *out)
271 DeltaDecoderSSE2 delta(out[-1]);
272 __m128i *outvec = reinterpret_cast<__m128i *>(out);
273 for (unsigned i = 0; i < BlockSize / 4; ++i) {
274 __m128i val = _mm_loadu_si128(outvec + i);
275 _mm_storeu_si128(outvec + i, delta.decode(val));
279 template<unsigned BlockSize, bool OrWithExisting, bool DeltaDecode, unsigned bit_width>
280 __attribute__((target("sse2")))
281 const unsigned char *decode_bitmap_sse2_unrolled(const unsigned char *in, uint32_t *out)
283 __m128i *outvec = reinterpret_cast<__m128i *>(out);
284 DeltaDecoderSSE2 delta(out[-1]);
285 InterleavedBitReaderSSE2 bs(in, bit_width);
286 #pragma GCC unroll 32
287 for (unsigned i = 0; i < BlockSize / 4; ++i) {
288 __m128i val = bs.read();
289 if constexpr (OrWithExisting) {
290 val = _mm_or_si128(val, _mm_slli_epi32(_mm_loadu_si128(outvec + i), bit_width));
292 if constexpr (DeltaDecode) {
293 val = delta.decode(val);
295 _mm_storeu_si128(outvec + i, val);
297 in += bytes_for_packed_bits(BlockSize, bit_width);
301 template<unsigned BlockSize, bool OrWithExisting, bool DeltaDecode>
302 __attribute__((target("sse2")))
303 const unsigned char *decode_bitmap_sse2(const unsigned char *in, unsigned bit_width, uint32_t *out)
306 case 0: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 0>(in, out);
307 case 1: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 1>(in, out);
308 case 2: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 2>(in, out);
309 case 3: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 3>(in, out);
310 case 4: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 4>(in, out);
311 case 5: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 5>(in, out);
312 case 6: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 6>(in, out);
313 case 7: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 7>(in, out);
314 case 8: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 8>(in, out);
315 case 9: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 9>(in, out);
316 case 10: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 10>(in, out);
317 case 11: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 11>(in, out);
318 case 12: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 12>(in, out);
319 case 13: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 13>(in, out);
320 case 14: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 14>(in, out);
321 case 15: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 15>(in, out);
322 case 16: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 16>(in, out);
323 case 17: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 17>(in, out);
324 case 18: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 18>(in, out);
325 case 19: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 19>(in, out);
326 case 20: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 20>(in, out);
327 case 21: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 21>(in, out);
328 case 22: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 22>(in, out);
329 case 23: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 23>(in, out);
330 case 24: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 24>(in, out);
331 case 25: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 25>(in, out);
332 case 26: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 26>(in, out);
333 case 27: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 27>(in, out);
334 case 28: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 28>(in, out);
335 case 29: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 29>(in, out);
336 case 30: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 30>(in, out);
337 case 31: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 31>(in, out);
338 case 32: return decode_bitmap_sse2_unrolled<BlockSize, OrWithExisting, DeltaDecode, 32>(in, out);
344 // Like decode_for(), but the values are organized in four independent streams,
345 // for SIMD (presumably SSE2). Supports a whole block only.
346 template<unsigned BlockSize, class Docid>
347 const unsigned char *decode_for_interleaved_generic(const unsigned char *in, Docid *out)
349 const unsigned bit_width = *in++ & 0x3f;
351 InterleavedBitReader<4> bs0(in + 0 * sizeof(uint32_t), bit_width);
352 InterleavedBitReader<4> bs1(in + 1 * sizeof(uint32_t), bit_width);
353 InterleavedBitReader<4> bs2(in + 2 * sizeof(uint32_t), bit_width);
354 InterleavedBitReader<4> bs3(in + 3 * sizeof(uint32_t), bit_width);
355 for (unsigned i = 0; i < BlockSize / 4; ++i) {
356 out[i * 4 + 0] = bs0.read();
357 out[i * 4 + 1] = bs1.read();
358 out[i * 4 + 2] = bs2.read();
359 out[i * 4 + 3] = bs3.read();
361 Docid prev_val = out[-1];
362 for (unsigned i = 0; i < BlockSize; ++i) {
363 out[i] = prev_val = out[i] + prev_val + 1;
365 return in + bytes_for_packed_bits(BlockSize, bit_width);
368 template<unsigned BlockSize, class Docid>
369 const unsigned char *decode_for_interleaved(const unsigned char *in, Docid *out)
371 if constexpr (BlockSize == 128 && sizeof(Docid) == sizeof(uint32_t)) {
372 return decode_for_interleaved_128_32(in, out);
374 return decode_for_interleaved_generic(in, out);
378 __attribute__((target("default")))
379 const unsigned char *decode_for_interleaved_128_32(const unsigned char *in, uint32_t *out)
381 return decode_for_interleaved_generic<128>(in, out);
384 #ifdef COULD_HAVE_SSE2
385 // Specialized version for SSE2.
386 __attribute__((target("sse2")))
387 const unsigned char *decode_for_interleaved_128_32(const unsigned char *in, uint32_t *out)
389 constexpr unsigned BlockSize = 128;
391 const unsigned bit_width = *in++ & 0x3f;
393 in = decode_bitmap_sse2<BlockSize, /*OrWithExisting=*/false, /*DeltaDecode=*/true>(in, bit_width, out);
399 template<class Docid>
400 const unsigned char *decode_pfor_bitmap_exceptions(const unsigned char *in, unsigned num, Docid *out)
402 const unsigned exception_bit_width = *in++;
403 const uint64_t *exception_bitmap_ptr = reinterpret_cast<const uint64_t *>(in);
404 in += div_round_up(num, 8);
406 int num_exceptions = 0;
408 BitReader bs(in, exception_bit_width);
409 for (unsigned i = 0; i < num; i += 64, ++exception_bitmap_ptr) {
410 uint64_t exceptions = read_le<uint64_t>(exception_bitmap_ptr);
412 // We've read some bytes past the end, so clear out the junk bits.
413 exceptions &= (1ULL << (num - i)) - 1;
415 for (; exceptions != 0; exceptions &= exceptions - 1, ++num_exceptions) {
416 unsigned idx = (ffsll(exceptions) - 1) + i;
417 out[idx] = bs.read();
420 in += bytes_for_packed_bits(num_exceptions, exception_bit_width);
424 // PFor block with bitmap exceptions. Layout:
426 // - Bit width (6 bits) | type << 6
427 // - Exception bit width (8 bits)
428 // - Bitmap of which values have exceptions (<num> bits, rounded up to a byte)
429 // - Exceptions (<num_exc> values of <bits_exc> bits, rounded up to a byte)
430 // - Base values (<num> values of <bits> bits, rounded up to a byte)
431 template<class Docid>
432 const unsigned char *decode_pfor_bitmap(const unsigned char *in, unsigned num, Docid *out)
434 memset(out, 0, num * sizeof(Docid));
436 const unsigned bit_width = *in++ & 0x3f;
438 in = decode_pfor_bitmap_exceptions(in, num, out);
440 // Decode the base values, and delta-decode.
441 Docid prev_val = out[-1];
442 BitReader bs(in, bit_width);
443 for (unsigned i = 0; i < num; ++i) {
444 out[i] = prev_val = ((out[i] << bit_width) | bs.read()) + prev_val + 1;
446 return in + bytes_for_packed_bits(num, bit_width);
449 // Like decode_pfor_bitmap(), but the base values are organized in four
450 // independent streams, for SIMD (presumably SSE2). Supports a whole block only.
451 template<unsigned BlockSize, class Docid>
452 const unsigned char *decode_pfor_bitmap_interleaved_generic(const unsigned char *in, Docid *out)
454 memset(out, 0, BlockSize * sizeof(Docid));
456 const unsigned bit_width = *in++ & 0x3f;
458 in = decode_pfor_bitmap_exceptions(in, BlockSize, out);
460 // Decode the base values.
461 InterleavedBitReader<4> bs0(in + 0 * sizeof(uint32_t), bit_width);
462 InterleavedBitReader<4> bs1(in + 1 * sizeof(uint32_t), bit_width);
463 InterleavedBitReader<4> bs2(in + 2 * sizeof(uint32_t), bit_width);
464 InterleavedBitReader<4> bs3(in + 3 * sizeof(uint32_t), bit_width);
465 for (unsigned i = 0; i < BlockSize / 4; ++i) {
466 out[i * 4 + 0] = bs0.read() | (out[i * 4 + 0] << bit_width);
467 out[i * 4 + 1] = bs1.read() | (out[i * 4 + 1] << bit_width);
468 out[i * 4 + 2] = bs2.read() | (out[i * 4 + 2] << bit_width);
469 out[i * 4 + 3] = bs3.read() | (out[i * 4 + 3] << bit_width);
473 Docid prev_val = out[-1];
474 for (unsigned i = 0; i < BlockSize; ++i) {
475 out[i] = prev_val = out[i] + prev_val + 1;
477 return in + bytes_for_packed_bits(BlockSize, bit_width);
480 template<unsigned BlockSize, class Docid>
481 const unsigned char *decode_pfor_bitmap_interleaved(const unsigned char *in, Docid *out)
483 if constexpr (BlockSize == 128 && sizeof(Docid) == sizeof(uint32_t)) {
484 return decode_pfor_bitmap_interleaved_128_32(in, out);
486 return decode_pfor_bitmap_interleaved_generic(in, out);
490 __attribute__((target("default")))
491 const unsigned char *decode_pfor_bitmap_interleaved_128_32(const unsigned char *in, uint32_t *out)
493 return decode_pfor_bitmap_interleaved_generic<128>(in, out);
496 #ifdef COULD_HAVE_SSE2
497 // Specialized version for SSE2.
498 __attribute__((target("sse2")))
499 const unsigned char *decode_pfor_bitmap_interleaved_128_32(const unsigned char *in, uint32_t *out)
501 constexpr unsigned BlockSize = 128;
502 using Docid = uint32_t;
504 memset(out, 0, BlockSize * sizeof(Docid));
506 const unsigned bit_width = *in++ & 0x3f;
508 in = decode_pfor_bitmap_exceptions(in, BlockSize, out);
509 in = decode_bitmap_sse2<BlockSize, /*OrWithExisting=*/true, /*DeltaDecode=*/true>(in, bit_width, out);
515 // PFor block with variable-byte exceptions. Layout:
517 // - Bit width (6 bits) | type << 6
518 // - Number of exceptions (8 bits)
519 // - Base values (<num> values of <bits> bits, rounded up to a byte)
521 // - If first byte is 255, <num_exc> 32-bit values (does not include the 255 byte)
522 // - Else, <num_exc> varbyte-encoded values (includes the non-255 byte)
523 // - Indexes of exceptions (<num_exc> bytes).
524 template<unsigned BlockSize, class Docid>
525 const unsigned char *decode_pfor_vb(const unsigned char *in, unsigned num, Docid *out)
527 //fprintf(stderr, "in=%p out=%p num=%u\n", in, out, num);
529 const unsigned bit_width = *in++ & 0x3f;
530 unsigned num_exceptions = *in++;
532 // Decode the base values.
533 BitReader bs(in, bit_width);
534 for (unsigned i = 0; i < num; ++i) {
537 in += bytes_for_packed_bits(num, bit_width);
539 // Decode exceptions.
540 Docid exceptions[BlockSize];
543 for (unsigned i = 0; i < num_exceptions; ++i) {
544 exceptions[i] = read_le<Docid>(in);
548 for (unsigned i = 0; i < num_exceptions; ++i) {
549 in = read_vb(in, &exceptions[i]);
553 for (unsigned i = 0; i < num_exceptions; ++i) {
554 unsigned idx = *in++;
555 out[idx] |= exceptions[i] << bit_width;
559 Docid prev_val = out[-1];
560 for (unsigned i = 0; i < num; ++i) {
561 out[i] = prev_val = out[i] + prev_val + 1;
567 // Like decode_pfor_vb(), but the base values are organized in four
568 // independent streams, for SIMD (presumably SSE2). Supports a whole block only.
569 template<unsigned BlockSize, class Docid>
570 const unsigned char *decode_pfor_vb_interleaved_generic(const unsigned char *in, Docid *out)
572 const unsigned bit_width = *in++ & 0x3f;
573 unsigned num_exceptions = *in++;
575 // Decode the base values.
576 InterleavedBitReader<4> bs0(in + 0 * sizeof(uint32_t), bit_width);
577 InterleavedBitReader<4> bs1(in + 1 * sizeof(uint32_t), bit_width);
578 InterleavedBitReader<4> bs2(in + 2 * sizeof(uint32_t), bit_width);
579 InterleavedBitReader<4> bs3(in + 3 * sizeof(uint32_t), bit_width);
580 for (unsigned i = 0; i < BlockSize / 4; ++i) {
581 out[i * 4 + 0] = bs0.read();
582 out[i * 4 + 1] = bs1.read();
583 out[i * 4 + 2] = bs2.read();
584 out[i * 4 + 3] = bs3.read();
586 in += bytes_for_packed_bits(BlockSize, bit_width);
588 // Decode exceptions.
589 Docid exceptions[BlockSize];
592 for (unsigned i = 0; i < num_exceptions; ++i) {
593 exceptions[i] = read_le<Docid>(in);
597 for (unsigned i = 0; i < num_exceptions; ++i) {
598 in = read_vb(in, &exceptions[i]);
603 for (unsigned i = 0; i < num_exceptions; ++i) {
604 unsigned idx = *in++;
605 out[idx] |= exceptions[i] << bit_width;
609 Docid prev_val = out[-1];
610 for (unsigned i = 0; i < BlockSize; ++i) {
611 out[i] = prev_val = out[i] + prev_val + 1;
617 template<unsigned BlockSize, class Docid>
618 const unsigned char *decode_pfor_vb_interleaved(const unsigned char *in, Docid *out)
620 if constexpr (BlockSize == 128 && sizeof(Docid) == sizeof(uint32_t)) {
621 return decode_pfor_vb_interleaved_128_32(in, out);
623 return decode_pfor_vb_interleaved_generic(in, out);
627 __attribute__((target("default")))
628 const unsigned char *decode_pfor_vb_interleaved_128_32(const unsigned char *in, uint32_t *out)
630 return decode_pfor_vb_interleaved_generic<128>(in, out);
633 // Specialized version for SSE2.
634 __attribute__((target("sse2")))
635 const unsigned char *decode_pfor_vb_interleaved_128_32(const unsigned char *in, uint32_t *out)
637 constexpr unsigned BlockSize = 128;
638 using Docid = uint32_t;
640 const unsigned bit_width = *in++ & 0x3f;
641 unsigned num_exceptions = *in++;
643 // Decode the base values.
644 in = decode_bitmap_sse2<BlockSize, /*OrWithExisting=*/false, /*DeltaDecode=*/false>(in, bit_width, out);
646 // Decode exceptions.
647 Docid exceptions[BlockSize];
650 for (unsigned i = 0; i < num_exceptions; ++i) {
651 exceptions[i] = read_le<Docid>(in);
655 for (unsigned i = 0; i < num_exceptions; ++i) {
656 in = read_vb(in, &exceptions[i]);
661 for (unsigned i = 0; i < num_exceptions; ++i) {
662 unsigned idx = *in++;
663 out[idx] |= exceptions[i] << bit_width;
666 delta_decode_sse2<BlockSize>(out);
678 template<unsigned BlockSize, class Docid>
679 const unsigned char *decode_pfor_delta1(const unsigned char *in, unsigned num, bool interleaved, Docid *out)
684 in = read_baseval(in, out++);
686 for (unsigned i = 1; i < num; i += BlockSize, out += BlockSize) {
687 const unsigned num_this_block = std::min<unsigned>(num - i, BlockSize);
688 switch (in[0] >> 6) {
690 if (interleaved && num_this_block == BlockSize) {
691 dprintf("%d+%d: blocktype=%d (for, interleaved), bitwidth=%d\n", i, num_this_block, in[0] >> 6, in[0] & 0x3f);
692 in = decode_for_interleaved<BlockSize>(in, out);
694 dprintf("%d+%d: blocktype=%d (for), bitwidth=%d\n", i, num_this_block, in[0] >> 6, in[0] & 0x3f);
695 in = decode_for(in, num_this_block, out);
698 case BlockType::PFOR_VB:
699 if (interleaved && num_this_block == BlockSize) {
700 dprintf("%d+%d: blocktype=%d (pfor + vb, interleaved), bitwidth=%d\n", i, num_this_block, in[0] >> 6, in[0] & 0x3f);
701 in = decode_pfor_vb_interleaved<BlockSize>(in, out);
703 dprintf("%d+%d: blocktype=%d (pfor + vb), bitwidth=%d\n", i, num_this_block, in[0] >> 6, in[0] & 0x3f);
704 in = decode_pfor_vb<BlockSize>(in, num_this_block, out);
707 case BlockType::PFOR_BITMAP:
708 if (interleaved && num_this_block == BlockSize) {
709 dprintf("%d+%d: blocktype=%d (pfor + bitmap, interleaved), bitwidth=%d\n", i, num_this_block, in[0] >> 6, in[0] & 0x3f);
710 in = decode_pfor_bitmap_interleaved<BlockSize>(in, out);
712 dprintf("%d+%d: blocktype=%d (pfor + bitmap), bitwidth=%d\n", i, num_this_block, in[0] >> 6, in[0] & 0x3f);
713 in = decode_pfor_bitmap(in, num_this_block, out);
716 case BlockType::CONSTANT:
717 dprintf("%d+%d: blocktype=%d (constant), bitwidth=%d\n", i, num_this_block, in[0] >> 6, in[0] & 0x3f);
718 in = decode_constant(in, num_this_block, out);
726 #endif // !defined(_TURBOPFOR_H)