adler32_avx2.c 5.7 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145
  1. /* adler32_avx2.c -- compute the Adler-32 checksum of a data stream
  2. * Copyright (C) 1995-2011 Mark Adler
  3. * Copyright (C) 2022 Adam Stylinski
  4. * Authors:
  5. * Brian Bockelman <bockelman@gmail.com>
  6. * Adam Stylinski <kungfujesus06@gmail.com>
  7. * For conditions of distribution and use, see copyright notice in zlib.h
  8. */
  9. #ifdef X86_AVX2
  10. #include "zbuild.h"
  11. #include <immintrin.h>
  12. #include "adler32_p.h"
  13. #include "adler32_avx2_p.h"
  14. #include "x86_intrins.h"
  15. extern uint32_t adler32_fold_copy_sse42(uint32_t adler, uint8_t *dst, const uint8_t *src, size_t len);
  16. extern uint32_t adler32_ssse3(uint32_t adler, const uint8_t *src, size_t len);
  17. static inline uint32_t adler32_fold_copy_impl(uint32_t adler, uint8_t *dst, const uint8_t *src, size_t len, const int COPY) {
  18. if (src == NULL) return 1L;
  19. if (len == 0) return adler;
  20. uint32_t adler0, adler1;
  21. adler1 = (adler >> 16) & 0xffff;
  22. adler0 = adler & 0xffff;
  23. rem_peel:
  24. if (len < 16) {
  25. if (COPY) {
  26. return adler32_copy_len_16(adler0, src, dst, len, adler1);
  27. } else {
  28. return adler32_len_16(adler0, src, len, adler1);
  29. }
  30. } else if (len < 32) {
  31. if (COPY) {
  32. return adler32_fold_copy_sse42(adler, dst, src, len);
  33. } else {
  34. return adler32_ssse3(adler, src, len);
  35. }
  36. }
  37. __m256i vs1, vs2;
  38. const __m256i dot2v = _mm256_setr_epi8(32, 31, 30, 29, 28, 27, 26, 25, 24, 23, 22, 21, 20, 19, 18, 17, 16, 15,
  39. 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1);
  40. const __m256i dot3v = _mm256_set1_epi16(1);
  41. const __m256i zero = _mm256_setzero_si256();
  42. while (len >= 32) {
  43. vs1 = _mm256_zextsi128_si256(_mm_cvtsi32_si128(adler0));
  44. vs2 = _mm256_zextsi128_si256(_mm_cvtsi32_si128(adler1));
  45. __m256i vs1_0 = vs1;
  46. __m256i vs3 = _mm256_setzero_si256();
  47. size_t k = MIN(len, NMAX);
  48. k -= k % 32;
  49. len -= k;
  50. while (k >= 32) {
  51. /*
  52. vs1 = adler + sum(c[i])
  53. vs2 = sum2 + 32 vs1 + sum( (32-i+1) c[i] )
  54. */
  55. __m256i vbuf = _mm256_loadu_si256((__m256i*)src);
  56. src += 32;
  57. k -= 32;
  58. __m256i vs1_sad = _mm256_sad_epu8(vbuf, zero); // Sum of abs diff, resulting in 2 x int32's
  59. if (COPY) {
  60. _mm256_storeu_si256((__m256i*)dst, vbuf);
  61. dst += 32;
  62. }
  63. vs1 = _mm256_add_epi32(vs1, vs1_sad);
  64. vs3 = _mm256_add_epi32(vs3, vs1_0);
  65. __m256i v_short_sum2 = _mm256_maddubs_epi16(vbuf, dot2v); // sum 32 uint8s to 16 shorts
  66. __m256i vsum2 = _mm256_madd_epi16(v_short_sum2, dot3v); // sum 16 shorts to 8 uint32s
  67. vs2 = _mm256_add_epi32(vsum2, vs2);
  68. vs1_0 = vs1;
  69. }
  70. /* Defer the multiplication with 32 to outside of the loop */
  71. vs3 = _mm256_slli_epi32(vs3, 5);
  72. vs2 = _mm256_add_epi32(vs2, vs3);
  73. /* The compiler is generating the following sequence for this integer modulus
  74. * when done the scalar way, in GPRs:
  75. adler = (s1_unpack[0] % BASE) + (s1_unpack[1] % BASE) + (s1_unpack[2] % BASE) + (s1_unpack[3] % BASE) +
  76. (s1_unpack[4] % BASE) + (s1_unpack[5] % BASE) + (s1_unpack[6] % BASE) + (s1_unpack[7] % BASE);
  77. mov $0x80078071,%edi // move magic constant into 32 bit register %edi
  78. ...
  79. vmovd %xmm1,%esi // move vector lane 0 to 32 bit register %esi
  80. mov %rsi,%rax // zero-extend this value to 64 bit precision in %rax
  81. imul %rdi,%rsi // do a signed multiplication with magic constant and vector element
  82. shr $0x2f,%rsi // shift right by 47
  83. imul $0xfff1,%esi,%esi // do a signed multiplication with value truncated to 32 bits with 0xfff1
  84. sub %esi,%eax // subtract lower 32 bits of original vector value from modified one above
  85. ...
  86. // repeats for each element with vpextract instructions
  87. This is tricky with AVX2 for a number of reasons:
  88. 1.) There's no 64 bit multiplication instruction, but there is a sequence to get there
  89. 2.) There's ways to extend vectors to 64 bit precision, but no simple way to truncate
  90. back down to 32 bit precision later (there is in AVX512)
  91. 3.) Full width integer multiplications aren't cheap
  92. We can, however, do a relatively cheap sequence for horizontal sums.
  93. Then, we simply do the integer modulus on the resulting 64 bit GPR, on a scalar value. It was
  94. previously thought that casting to 64 bit precision was needed prior to the horizontal sum, but
  95. that is simply not the case, as NMAX is defined as the maximum number of scalar sums that can be
  96. performed on the maximum possible inputs before overflow
  97. */
  98. /* In AVX2-land, this trip through GPRs will probably be unavoidable, as there's no cheap and easy
  99. * conversion from 64 bit integer to 32 bit (needed for the inexpensive modulus with a constant).
  100. * This casting to 32 bit is cheap through GPRs (just register aliasing). See above for exactly
  101. * what the compiler is doing to avoid integer divisions. */
  102. adler0 = partial_hsum256(vs1) % BASE;
  103. adler1 = hsum256(vs2) % BASE;
  104. }
  105. adler = adler0 | (adler1 << 16);
  106. if (len) {
  107. goto rem_peel;
  108. }
  109. return adler;
  110. }
  111. Z_INTERNAL uint32_t adler32_avx2(uint32_t adler, const uint8_t *src, size_t len) {
  112. return adler32_fold_copy_impl(adler, NULL, src, len, 0);
  113. }
  114. Z_INTERNAL uint32_t adler32_fold_copy_avx2(uint32_t adler, uint8_t *dst, const uint8_t *src, size_t len) {
  115. return adler32_fold_copy_impl(adler, dst, src, len, 1);
  116. }
  117. #endif