Hello,
I am trying to learn the basics of intrinsic optimizations for native Xeon Phi Applications.
The selected target application is libethereum as it implements strict memory hard hashing functions.
https://bitslog.files.wordpress.com/2013/12/memohash-v0-3.pdf
The code-base makes use of nodes composed of 512 bits!
This seems ideal for targeting the native Xeon Phi.
However, I cannot figure out the appropriate syntax for the union typedef.
I have tried numerous combinations, all of which compile and produce a segfault.
The only enabled change is to the structure of the union.
The segfault occurs any time a node is accessed via uint32_t words.
The specifics are outlined in this pull request:
https://github.com/ethereum/libethereum/pull/251/files
Any idea whats causing the segfaults here?
typedef union node {
uint8_t bytes[NODE_WORDS * 4] __attribute__((aligned(64)));
uint32_t words[NODE_WORDS] __attribute__((aligned(64)));
uint64_t double_words[NODE_WORDS / 2] __attribute__((aligned(64)));
#if defined(_M_X64) && ENABLE_SSE
__m128i xmm[NODE_WORDS/4] __attribute__((aligned(64)));
#endif
#if defined(__MIC__)
__m512i zmm;
#endif
} node;
#define FNV_PRIME 0x01000193
static inline uint32_t fnv_hash(uint32_t const x, uint32_t const y)
{
return x * FNV_PRIME ^ y;
}
#if defined(_M_X64) && ENABLE_SSE
{
__m128i fnv_prime = _mm_set1_epi32(FNV_PRIME);
__m128i xmm0 = _mm_mullo_epi32(fnv_prime, mix.xmm[0]);
__m128i xmm1 = _mm_mullo_epi32(fnv_prime, mix.xmm[1]);
__m128i xmm2 = _mm_mullo_epi32(fnv_prime, mix.xmm[2]);
__m128i xmm3 = _mm_mullo_epi32(fnv_prime, mix.xmm[3]);
mix.xmm[0] = _mm_xor_si128(xmm0, dag_node->xmm[0]);
mix.xmm[1] = _mm_xor_si128(xmm1, dag_node->xmm[1]);
mix.xmm[2] = _mm_xor_si128(xmm2, dag_node->xmm[2]);
mix.xmm[3] = _mm_xor_si128(xmm3, dag_node->xmm[3]);
}
#elif defined(__MIC__) && 0
{
// TODO: __m512i implementation
// Each vector register (zmm) can store sixteen 32-bit integer numbers
// <--- Begin Critical Section --->
__m512i fnv_prime = _mm512_set1_epi32(FNV_PRIME)
__m512i zmm0 = _mm512_mullo_epi32(fnv_prime, mix.zmm);
mix.zmm = _mm512_xor_si512(zmm0, dag_node->zmm);
// <--- End Critical Section --->
}
#else
{
for (unsigned w = 0; w != NODE_WORDS; ++w) {
mix.words = fnv_hash(mix.words, dag_node->words);
}
}
#endif
