SHA256 performance optimization in C
SHA256 performance optimization in C ...
Now that the Goldmont micro-architecture has been released, it includes Intel's SHA extensions. You can get a 5x-6x speedup in the compress function using the CPU instructions. For example, proposed code for a crypto library witnessed the following (the test occurred on a Celeron J3455, which runs at 1.5 GHz, but bursts at 2.3 GHz):
- C++ implementation
$ ./botan speed --msec=3000 SHA-1 SHA-224 SHA-256
SHA-160 [base] hash 274.826 MiB/sec (824.480 MiB in 3000.009 ms)
SHA-224 [base] hash 92.349 MiB/sec (277.051 MiB in 3000.027 ms)
SHA-256 [base] hash 92.364 MiB/sec (277.094 MiB in 3000.027 ms)
- Intel SHA extensions
$ ./botan speed --msec=3000 SHA-1 SHA-224 SHA-256
SHA-160 [base] hash 1195.907 MiB/sec (3587.723 MiB in 3000.000 ms)
SHA-224 [base] hash 535.740 MiB/sec (1607.219 MiB in 3000.000 ms)
SHA-256 [base] hash 535.970 MiB/sec (1607.914 MiB in 3000.005 ms)
Here is the code for the SHA256 compress function using Intel SHA extensions with intrinsics. Its based on Sean Gulley's blog at Intel® SHA Extensions, and his sample code in mitls | hacl-star | experimental.
The compress function below only handles full blocks of 64-bytes. You need to setup the initial state, and you need to pad the last block. It looks like you have that covered in your sample code.
#include <immintrin.h>
...
void compress(uint32_t state[8], const uint8_t input[], size_t blocks)
{
__m128i STATE0, STATE1;
__m128i MSG, TMP, MASK;
__m128i TMSG0, TMSG1, TMSG2, TMSG3;
__m128i ABEF_SAVE, CDGH_SAVE;
// Load initial values
TMP = _mm_loadu_si128((__m128i*) &state[0]);
STATE1 = _mm_loadu_si128((__m128i*) &state[4]);
MASK = _mm_set_epi64x(0x0c0d0e0f08090a0bULL, 0x0405060700010203ULL);
TMP = _mm_shuffle_epi32(TMP, 0xB1); // CDAB
STATE1 = _mm_shuffle_epi32(STATE1, 0x1B); // EFGH
STATE0 = _mm_alignr_epi8(TMP, STATE1, 8); // ABEF
STATE1 = _mm_blend_epi16(STATE1, TMP, 0xF0); // CDGH
while (blocks)
{
// Save current hash
ABEF_SAVE = STATE0;
CDGH_SAVE = STATE1;
// Rounds 0-3
MSG = _mm_loadu_si128((const __m128i*) (input+0));
TMSG0 = _mm_shuffle_epi8(MSG, MASK);
MSG = _mm_add_epi32(TMSG0, _mm_set_epi64x(0xE9B5DBA5B5C0FBCFULL, 0x71374491428A2F98ULL));
STATE1 = _mm_sha256rnds2_epu32(STATE1, STATE0, MSG);
MSG = _mm_shuffle_epi32(MSG, 0x0E);
STATE0 = _mm_sha256rnds2_epu32(STATE0, STATE1, MSG);
// Rounds 4-7
TMSG1 = _mm_loadu_si128((const __m128i*) (input+16));
TMSG1 = _mm_shuffle_epi8(TMSG1, MASK);
MSG = _mm_add_epi32(TMSG1, _mm_set_epi64x(0xAB1C5ED5923F82A4ULL, 0x59F111F13956C25BULL));
STATE1 = _mm_sha256rnds2_epu32(STATE1, STATE0, MSG);
MSG = _mm_shuffle_epi32(MSG, 0x0E);
STATE0 = _mm_sha256rnds2_epu32(STATE0, STATE1, MSG);
TMSG0 = _mm_sha256msg1_epu32(TMSG0, TMSG1);
// Rounds 8-11
TMSG2 = _mm_loadu_si128((const __m128i*) (input+32));
TMSG2 = _mm_shuffle_epi8(TMSG2, MASK);
MSG = _mm_add_epi32(TMSG2, _mm_set_epi64x(0x550C7DC3243185BEULL, 0x12835B01D807AA98ULL));
STATE1 = _mm_sha256rnds2_epu32(STATE1, STATE0, MSG);
MSG = _mm_shuffle_epi32(MSG, 0x0E);
STATE0 = _mm_sha256rnds2_epu32(STATE0, STATE1, MSG);
TMSG1 = _mm_sha256msg1_epu32(TMSG1, TMSG2);
// Rounds 12-15
TMSG3 = _mm_loadu_si128((const __m128i*) (input+48));
TMSG3 = _mm_shuffle_epi8(TMSG3, MASK);
MSG = _mm_add_epi32(TMSG3, _mm_set_epi64x(0xC19BF1749BDC06A7ULL, 0x80DEB1FE72BE5D74ULL));
STATE1 = _mm_sha256rnds2_epu32(STATE1, STATE0, MSG);
TMP = _mm_alignr_epi8(TMSG3, TMSG2, 4);
TMSG0 = _mm_add_epi32(TMSG0, TMP);
TMSG0 = _mm_sha256msg2_epu32(TMSG0, TMSG3);
MSG = _mm_shuffle_epi32(MSG, 0x0E);
STATE0 = _mm_sha256rnds2_epu32(STATE0, STATE1, MSG);
TMSG2 = _mm_sha256msg1_epu32(TMSG2, TMSG3);
// Rounds 16-19
MSG = _mm_add_epi32(TMSG0, _mm_set_epi64x(0x240CA1CC0FC19DC6ULL, 0xEFBE4786E49B69C1ULL));
STATE1 = _mm_sha256rnds2_epu32(STATE1, STATE0, MSG);
TMP = _mm_alignr_epi8(TMSG0, TMSG3, 4);
TMSG1 = _mm_add_epi32(TMSG1, TMP);
TMSG1 = _mm_sha256msg2_epu32(TMSG1, TMSG0);
MSG = _mm_shuffle_epi32(MSG, 0x0E);
STATE0 = _mm_sha256rnds2_epu32(STATE0, STATE1, MSG);
TMSG3 = _mm_sha256msg1_epu32(TMSG3, TMSG0);
// Rounds 20-23
MSG = _mm_add_epi32(TMSG1, _mm_set_epi64x(0x76F988DA5CB0A9DCULL, 0x4A7484AA2DE92C6FULL));
STATE1 = _mm_sha256rnds2_epu32(STATE1, STATE0, MSG);
TMP = _mm_alignr_epi8(TMSG1, TMSG0, 4);
TMSG2 = _mm_add_epi32(TMSG2, TMP);
TMSG2 = _mm_sha256msg2_epu32(TMSG2, TMSG1);
MSG = _mm_shuffle_epi32(MSG, 0x0E);
STATE0 = _mm_sha256rnds2_epu32(STATE0, STATE1, MSG);
TMSG0 = _mm_sha256msg1_epu32(TMSG0, TMSG1);
// Rounds 24-27
MSG = _mm_add_epi32(TMSG2, _mm_set_epi64x(0xBF597FC7B00327C8ULL, 0xA831C66D983E5152ULL));
STATE1 = _mm_sha256rnds2_epu32(STATE1, STATE0, MSG);
TMP = _mm_alignr_epi8(TMSG2, TMSG1, 4);
TMSG3 = _mm_add_epi32(TMSG3, TMP);
TMSG3 = _mm_sha256msg2_epu32(TMSG3, TMSG2);
MSG = _mm_shuffle_epi32(MSG, 0x0E);
STATE0 = _mm_sha256rnds2_epu32(STATE0, STATE1, MSG);
TMSG1 = _mm_sha256msg1_epu32(TMSG1, TMSG2);
// Rounds 28-31
MSG = _mm_add_epi32(TMSG3, _mm_set_epi64x(0x1429296706CA6351ULL, 0xD5A79147C6E00BF3ULL));
STATE1 = _mm_sha256rnds2_epu32(STATE1, STATE0, MSG);
TMP = _mm_alignr_epi8(TMSG3, TMSG2, 4);
TMSG0 = _mm_add_epi32(TMSG0, TMP);
TMSG0 = _mm_sha256msg2_epu32(TMSG0, TMSG3);
MSG = _mm_shuffle_epi32(MSG, 0x0E);
STATE0 = _mm_sha256rnds2_epu32(STATE0, STATE1, MSG);
TMSG2 = _mm_sha256msg1_epu32(TMSG2, TMSG3);
// Rounds 32-35
MSG = _mm_add_epi32(TMSG0, _mm_set_epi64x(0x53380D134D2C6DFCULL, 0x2E1B213827B70A85ULL));
STATE1 = _mm_sha256rnds2_epu32(STATE1, STATE0, MSG);
TMP = _mm_alignr_epi8(TMSG0, TMSG3, 4);
TMSG1 = _mm_add_epi32(TMSG1, TMP);
TMSG1 = _mm_sha256msg2_epu32(TMSG1, TMSG0);
MSG = _mm_shuffle_epi32(MSG, 0x0E);
STATE0 = _mm_sha256rnds2_epu32(STATE0, STATE1, MSG);
TMSG3 = _mm_sha256msg1_epu32(TMSG3, TMSG0);
// Rounds 36-39
MSG = _mm_add_epi32(TMSG1, _mm_set_epi64x(0x92722C8581C2C92EULL, 0x766A0ABB650A7354ULL));
STATE1 = _mm_sha256rnds2_epu32(STATE1, STATE0, MSG);
TMP = _mm_alignr_epi8(TMSG1, TMSG0, 4);
TMSG2 = _mm_add_epi32(TMSG2, TMP);
TMSG2 = _mm_sha256msg2_epu32(TMSG2, TMSG1);
MSG = _mm_shuffle_epi32(MSG, 0x0E);
STATE0 = _mm_sha256rnds2_epu32(STATE0, STATE1, MSG);
TMSG0 = _mm_sha256msg1_epu32(TMSG0, TMSG1);
// Rounds 40-43
MSG = _mm_add_epi32(TMSG2, _mm_set_epi64x(0xC76C51A3C24B8B70ULL, 0xA81A664BA2BFE8A1ULL));
STATE1 = _mm_sha256rnds2_epu32(STATE1, STATE0, MSG);
TMP = _mm_alignr_epi8(TMSG2, TMSG1, 4);
TMSG3 = _mm_add_epi32(TMSG3, TMP);
TMSG3 = _mm_sha256msg2_epu32(TMSG3, TMSG2);
MSG = _mm_shuffle_epi32(MSG, 0x0E);
STATE0 = _mm_sha256rnds2_epu32(STATE0, STATE1, MSG);
TMSG1 = _mm_sha256msg1_epu32(TMSG1, TMSG2);
// Rounds 44-47
MSG = _mm_add_epi32(TMSG3, _mm_set_epi64x(0x106AA070F40E3585ULL, 0xD6990624D192E819ULL));
STATE1 = _mm_sha256rnds2_epu32(STATE1, STATE0, MSG);
TMP = _mm_alignr_epi8(TMSG3, TMSG2, 4);
TMSG0 = _mm_add_epi32(TMSG0, TMP);
TMSG0 = _mm_sha256msg2_epu32(TMSG0, TMSG3);
MSG = _mm_shuffle_epi32(MSG, 0x0E);
STATE0 = _mm_sha256rnds2_epu32(STATE0, STATE1, MSG);
TMSG2 = _mm_sha256msg1_epu32(TMSG2, TMSG3);
// Rounds 48-51
MSG = _mm_add_epi32(TMSG0, _mm_set_epi64x(0x34B0BCB52748774CULL, 0x1E376C0819A4C116ULL));
STATE1 = _mm_sha256rnds2_epu32(STATE1, STATE0, MSG);
TMP = _mm_alignr_epi8(TMSG0, TMSG3, 4);
TMSG1 = _mm_add_epi32(TMSG1, TMP);
TMSG1 = _mm_sha256msg2_epu32(TMSG1, TMSG0);
MSG = _mm_shuffle_epi32(MSG, 0x0E);
STATE0 = _mm_sha256rnds2_epu32(STATE0, STATE1, MSG);
TMSG3 = _mm_sha256msg1_epu32(TMSG3, TMSG0);
// Rounds 52-55
MSG = _mm_add_epi32(TMSG1, _mm_set_epi64x(0x682E6FF35B9CCA4FULL, 0x4ED8AA4A391C0CB3ULL));
STATE1 = _mm_sha256rnds2_epu32(STATE1, STATE0, MSG);
TMP = _mm_alignr_epi8(TMSG1, TMSG0, 4);
TMSG2 = _mm_add_epi32(TMSG2, TMP);
TMSG2 = _mm_sha256msg2_epu32(TMSG2, TMSG1);
MSG = _mm_shuffle_epi32(MSG, 0x0E);
STATE0 = _mm_sha256rnds2_epu32(STATE0, STATE1, MSG);
// Rounds 56-59
MSG = _mm_add_epi32(TMSG2, _mm_set_epi64x(0x8CC7020884C87814ULL, 0x78A5636F748F82EEULL));
STATE1 = _mm_sha256rnds2_epu32(STATE1, STATE0, MSG);
TMP = _mm_alignr_epi8(TMSG2, TMSG1, 4);
TMSG3 = _mm_add_epi32(TMSG3, TMP);
TMSG3 = _mm_sha256msg2_epu32(TMSG3, TMSG2);
MSG = _mm_shuffle_epi32(MSG, 0x0E);
STATE0 = _mm_sha256rnds2_epu32(STATE0, STATE1, MSG);
// Rounds 60-63
MSG = _mm_add_epi32(TMSG3, _mm_set_epi64x(0xC67178F2BEF9A3F7ULL, 0xA4506CEB90BEFFFAULL));
STATE1 = _mm_sha256rnds2_epu32(STATE1, STATE0, MSG);
MSG = _mm_shuffle_epi32(MSG, 0x0E);
STATE0 = _mm_sha256rnds2_epu32(STATE0, STATE1, MSG);
// Add values back to state
STATE0 = _mm_add_epi32(STATE0, ABEF_SAVE);
STATE1 = _mm_add_epi32(STATE1, CDGH_SAVE);
input += 64;
blocks--;
}
TMP = _mm_shuffle_epi32(STATE0, 0x1B); // FEBA
STATE1 = _mm_shuffle_epi32(STATE1, 0xB1); // DCHG
STATE0 = _mm_blend_epi16(TMP, STATE1, 0xF0); // DCBA
STATE1 = _mm_alignr_epi8(STATE1, TMP, 8); // ABEF
// Save state
_mm_storeu_si128((__m128i*) &state[0], STATE0);
_mm_storeu_si128((__m128i*) &state[4], STATE1);
}
You can find source for both Intel SHA intrinsics and ARMv8 SHA intrinsics at Noloader GitHub | SHA-Intrinsics. They are C source files, and provide the compress function for SHA-1, SHA-224 and SHA-256. The intrinsic based implementations increase throughput approximately 3x to 4x for SHA-1, and approximately 6x to 12x for SHA-224 and SHA-256.
You may want to checkout/profile this implementation of SHA256.
Being used in cgminer (a popular bitcoin mining software), it is written specifically keeping performance in mind. It includes 4-way SIMD implementations using SSE2. It follows the same approach as the bradconte sha256_transform algorithm mentioned in the question. The code is too long to reproduce here.
Also the license is fairly permissive, allowing re-use/distribution as long as the original authors are accredited.