/**
* Performs an absorb operation for a single block (BLOCK_LEN_BLAKE2_SAFE_INT64
* words of type uint64_t), using Blake2b's G function as the internal permutation
*
* @param state The current state of the sponge
* @param in The block to be absorbed (BLOCK_LEN_BLAKE2_SAFE_INT64 words)
*/
void absorbBlockBlake2Safe(uint64_t *state, const uint64_t *in)
{
//XORs the first BLOCK_LEN_BLAKE2_SAFE_INT64 words of "in" with the current state
#if defined __AVX2__
__m256i state_v[2], in_v[2];
state_v[0] = _mm256_load_si256( (__m256i*)(&state[0]) );
in_v [0] = _mm256_loadu_si256( (__m256i*)(&in[0]) );
state_v[1] = _mm256_load_si256( (__m256i*)(&state[4]) );
in_v [1] = _mm256_loadu_si256( (__m256i*)(&in[4]) );
_mm256_store_si256( (__m256i*)(&state[0]),
_mm256_xor_si256( state_v[0], in_v[0] ) );
_mm256_store_si256( (__m256i*)(&state[4]),
_mm256_xor_si256( state_v[1], in_v[1] ) );
#elif defined __AVX__
__m128i state_v[4], in_v[4];
state_v[0] = _mm_load_si128( (__m128i*)(&state[0]) );
state_v[1] = _mm_load_si128( (__m128i*)(&state[2]) );
state_v[2] = _mm_load_si128( (__m128i*)(&state[4]) );
state_v[3] = _mm_load_si128( (__m128i*)(&state[6]) );
in_v[0] = _mm_loadu_si128( (__m128i*)(&in[0]) );
in_v[1] = _mm_loadu_si128( (__m128i*)(&in[2]) );
in_v[2] = _mm_loadu_si128( (__m128i*)(&in[4]) );
in_v[3] = _mm_loadu_si128( (__m128i*)(&in[6]) );
_mm_store_si128( (__m128i*)(&state[0]),
_mm_xor_si128( state_v[0], in_v[0] ) );
_mm_store_si128( (__m128i*)(&state[2]),
_mm_xor_si128( state_v[1], in_v[1] ) );
_mm_store_si128( (__m128i*)(&state[4]),
_mm_xor_si128( state_v[2], in_v[2] ) );
_mm_store_si128( (__m128i*)(&state[6]),
_mm_xor_si128( state_v[3], in_v[3] ) );
#else
state[0] ^= in[0];
state[1] ^= in[1];
state[2] ^= in[2];
state[3] ^= in[3];
state[4] ^= in[4];
state[5] ^= in[5];
state[6] ^= in[6];
state[7] ^= in[7];
#endif
//Applies the transformation f to the sponge's state
blake2bLyra(state);
}
static fstb_FORCEINLINE void TransLut_store_avx2 (T *dst_ptr, __m256 val)
{
_mm256_store_si256 (
reinterpret_cast <__m256i *> (dst_ptr),
_mm256_cvtps_epi32 (val)
);
}
static char *
detile_xmajor(struct surface *s, __m256i alpha)
{
int height = align_u64(s->height, 8);
void *pixels;
int tile_stride = s->stride / 512;
int ret;
ret = posix_memalign(&pixels, 32, s->stride * height);
ksim_assert(ret == 0);
ksim_assert((s->stride & 511) == 0);
for (int y = 0; y < height; y++) {
int tile_y = y / 8;
int iy = y & 7;
void *src = s->pixels + tile_y * tile_stride * 4096 + iy * 512;
void *dst = pixels + y * s->stride;
for (int x = 0; x < tile_stride; x++) {
for (int c = 0; c < 512; c += 32) {
__m256i m = _mm256_load_si256(src + x * 4096 + c);
m = _mm256_or_si256(m, alpha);
_mm256_store_si256(dst + x * 512 + c, m);
}
}
}
return pixels;
}
void
mulrc16_shuffle_avx2(uint8_t *region, uint8_t constant, size_t length)
{
uint8_t *end;
register __m256i in, out, t1, t2, m1, m2, l, h;
register __m128i bc;
if (constant == 0) {
memset(region, 0, length);
return;
}
if (constant == 1)
return;
bc = _mm_load_si128((void *)tl[constant]);
t1 = __builtin_ia32_vbroadcastsi256(bc);
bc = _mm_load_si128((void *)th[constant]);
t2 = __builtin_ia32_vbroadcastsi256(bc);
m1 = _mm256_set1_epi8(0x0f);
m2 = _mm256_set1_epi8(0xf0);
for (end=region+length; region<end; region+=32) {
in = _mm256_load_si256((void *)region);
l = _mm256_and_si256(in, m1);
l = _mm256_shuffle_epi8(t1, l);
h = _mm256_and_si256(in, m2);
h = _mm256_srli_epi64(h, 4);
h = _mm256_shuffle_epi8(t2, h);
out = _mm256_xor_si256(h, l);
_mm256_store_si256((void *)region, out);
}
}
template <> SIMD_INLINE void InterpolateX<1>(const __m256i * alpha, __m256i * buffer)
{
#if defined(_MSC_VER) // Workaround for Visual Studio 2012 compiler bug in release mode:
__m256i _buffer = _mm256_or_si256(K_ZERO, _mm256_load_si256(buffer));
#else
__m256i _buffer = _mm256_load_si256(buffer);
#endif
_mm256_store_si256(buffer, _mm256_maddubs_epi16(_buffer, _mm256_load_si256(alpha)));
}
void PriorityQueue_AVX2::clear()
{
_size = _current = 0;
_maxrank = INT_MAX;
_isPopping = false;
__m256i max = _mm256_set1_epi32(_maxrank);
for (int i = 0; i < 5; ++i)
_mm256_store_si256(_rv + i, max);
}
template <> SIMD_INLINE void InterpolateX<3>(const __m256i * alpha, __m256i * buffer)
{
__m256i src[3], shuffled;
src[0] = _mm256_load_si256(buffer + 0);
src[1] = _mm256_load_si256(buffer + 1);
src[2] = _mm256_load_si256(buffer + 2);
shuffled = _mm256_shuffle_epi8(_mm256_permute2x128_si256(src[0], src[0], 0x21), K8_SHUFFLE_X3_00);
shuffled = _mm256_or_si256(shuffled, _mm256_shuffle_epi8(src[0], K8_SHUFFLE_X3_01));
shuffled = _mm256_or_si256(shuffled, _mm256_shuffle_epi8(_mm256_permute2x128_si256(src[0], src[1], 0x21), K8_SHUFFLE_X3_02));
_mm256_store_si256(buffer + 0, _mm256_maddubs_epi16(shuffled, _mm256_load_si256(alpha + 0)));
shuffled = _mm256_shuffle_epi8(_mm256_permute2x128_si256(src[0], src[1], 0x21), K8_SHUFFLE_X3_10);
shuffled = _mm256_or_si256(shuffled, _mm256_shuffle_epi8(src[1], K8_SHUFFLE_X3_11));
shuffled = _mm256_or_si256(shuffled, _mm256_shuffle_epi8(_mm256_permute2x128_si256(src[1], src[2], 0x21), K8_SHUFFLE_X3_12));
_mm256_store_si256(buffer + 1, _mm256_maddubs_epi16(shuffled, _mm256_load_si256(alpha + 1)));
shuffled = _mm256_shuffle_epi8(_mm256_permute2x128_si256(src[1], src[2], 0x21), K8_SHUFFLE_X3_20);
shuffled = _mm256_or_si256(shuffled, _mm256_shuffle_epi8(src[2], K8_SHUFFLE_X3_21));
shuffled = _mm256_or_si256(shuffled, _mm256_shuffle_epi8(_mm256_permute2x128_si256(src[2], src[2], 0x21), K8_SHUFFLE_X3_22));
_mm256_store_si256(buffer + 2, _mm256_maddubs_epi16(shuffled, _mm256_load_si256(alpha + 2)));
}
void
maddrc16_imul_avx2(uint8_t* region1, const uint8_t* region2,
uint8_t constant, size_t length)
{
uint8_t *end;
register __m256i reg1, reg2, ri[4], sp[4], mi[4];
const uint8_t *p = pt[constant];
if (constant == 0)
return;
if (constant == 1) {
xorr_avx2(region1, region2, length);
return;
}
mi[0] = _mm256_set1_epi8(0x11);
mi[1] = _mm256_set1_epi8(0x22);
mi[2] = _mm256_set1_epi8(0x44);
mi[3] = _mm256_set1_epi8(0x88);
sp[0] = _mm256_set1_epi16(p[0]);
sp[1] = _mm256_set1_epi16(p[1]);
sp[2] = _mm256_set1_epi16(p[2]);
sp[3] = _mm256_set1_epi16(p[3]);
for (end=region1+length; region1<end; region1+=32, region2+=32) {
reg2 = _mm256_load_si256((void *)region2);
reg1 = _mm256_load_si256((void *)region1);
ri[0] = _mm256_and_si256(reg2, mi[0]);
ri[1] = _mm256_and_si256(reg2, mi[1]);
ri[2] = _mm256_and_si256(reg2, mi[2]);
ri[3] = _mm256_and_si256(reg2, mi[3]);
ri[1] = _mm256_srli_epi16(ri[1], 1);
ri[2] = _mm256_srli_epi16(ri[2], 2);
ri[3] = _mm256_srli_epi16(ri[3], 3);
ri[0] = _mm256_mullo_epi16(ri[0], sp[0]);
ri[1] = _mm256_mullo_epi16(ri[1], sp[1]);
ri[2] = _mm256_mullo_epi16(ri[2], sp[2]);
ri[3] = _mm256_mullo_epi16(ri[3], sp[3]);
ri[0] = _mm256_xor_si256(ri[0], ri[1]);
ri[2] = _mm256_xor_si256(ri[2], ri[3]);
ri[0] = _mm256_xor_si256(ri[0], ri[2]);
ri[0] = _mm256_xor_si256(ri[0], reg1);
_mm256_store_si256((void *)region1, ri[0]);
}
}
int main() {
unsigned int i, bytes_read, n, written;
char buffer[BUFF_SIZE];
__m256i* mem256 = (__m256i*)&mem[0];
__m256i* final = (__m256i*)&mem[999936];
__m256i zero256 = _mm256_setzero_si256();
for (; mem256 < final; mem256 += 8) {
_mm256_store_si256(&mem256[0], zero256);
_mm256_store_si256(&mem256[1], zero256);
_mm256_store_si256(&mem256[2], zero256);
_mm256_store_si256(&mem256[3], zero256);
_mm256_store_si256(&mem256[4], zero256);
_mm256_store_si256(&mem256[5], zero256);
_mm256_store_si256(&mem256[6], zero256);
_mm256_store_si256(&mem256[7], zero256);
}
final = (__m256i*) &mem[1000000];
inline void Sort4Deg6(__m256 llrI, int pos[], int ipos[])
{
int llr[8] __attribute__((aligned(64)));
const auto v1 = _mm256_set1_ps( 67108864.0f );
const auto v2 = _mm256_mul_ps( v1, llrI );
_mm256_store_si256((__m256i *)llr, _mm256_cvttps_epi32(v2));
//register float x0,x1,x2,x3,x4,x5;
const auto x0 = llr[0];
const auto x1 = llr[1];
const auto x2 = llr[2];
const auto x3 = llr[3];
const auto x4 = llr[4];
const auto x5 = llr[5];
int o0 = (x0<x1) +(x0<x2)+(x0<x3)+(x0<x4)+(x0<x5);
int o1 = (x1<=x0)+(x1<x2)+(x1<x3)+(x1<x4)+(x1<x5);
int o2 = (x2<=x0)+(x2<=x1)+(x2<x3)+(x2<x4)+(x2<x5);
int o3 = (x3<=x0)+(x3<=x1)+(x3<=x2)+(x3<x4)+(x3<x5);
int o4 = (x4<=x0)+(x4<=x1)+(x4<=x2)+(x4<=x3)+(x4<x5);
int o5 = 15-(o0+o1+o2+o3+o4);
pos[o0] = 0; pos[o1]= 1; pos[o2]= 2; pos[o3]= 3; pos[o4]= 4; pos[o5]= 5; pos[6]=6; pos[7]=7;
ipos[ 0] = o0; ipos[ 1]=o1; ipos[ 2]=o2; ipos[ 3]=o3; ipos[ 4]=o4; ipos[ 5]=o5; ipos[6]=6; ipos[7]=7;
}
void
maddrc16_shuffle_avx2(uint8_t* region1, const uint8_t* region2,
uint8_t constant, size_t length)
{
uint8_t *end;
register __m256i in1, in2, out, t1, t2, m1, m2, l, h;
register __m128i bc;
if (constant == 0)
return;
if (constant == 1) {
xorr_avx2(region1, region2, length);
return;
}
bc = _mm_load_si128((void *)tl[constant]);
t1 = __builtin_ia32_vbroadcastsi256(bc);
bc = _mm_load_si128((void *)th[constant]);
t2 = __builtin_ia32_vbroadcastsi256(bc);
m1 = _mm256_set1_epi8(0x0f);
m2 = _mm256_set1_epi8(0xf0);
for (end=region1+length; region1<end; region1+=32, region2+=32) {
in2 = _mm256_load_si256((void *)region2);
in1 = _mm256_load_si256((void *)region1);
l = _mm256_and_si256(in2, m1);
l = _mm256_shuffle_epi8(t1, l);
h = _mm256_and_si256(in2, m2);
h = _mm256_srli_epi64(h, 4);
h = _mm256_shuffle_epi8(t2, h);
out = _mm256_xor_si256(h,l);
out = _mm256_xor_si256(out, in1);
_mm256_store_si256((void *)region1, out);
}
}
/**
* Performs a reduced duplex operation for a single row, from the highest to
* the lowest index, using the reduced-round Blake2b's G function as the
* internal permutation
*
* @param state The current state of the sponge
* @param rowIn Row to feed the sponge
* @param rowOut Row to receive the sponge's output
*/
void reducedDuplexRow1(uint64_t *state, uint64_t *rowIn, uint64_t *rowOut, const uint32_t nCols)
{
uint64_t* ptrWordIn = rowIn; //In Lyra2: pointer to prev
uint64_t* ptrWordOut = rowOut + (nCols-1)*BLOCK_LEN_INT64; //In Lyra2: pointer to row
unsigned int i;
for (i = 0; i < nCols; i++)
{
//Absorbing "M[prev][col]"
#if defined __AVX2__
__m256i state_v[3], in_v[3];
state_v[0] = _mm256_load_si256( (__m256i*)(&state[0]) );
in_v [0] = _mm256_loadu_si256( (__m256i*)(&ptrWordIn[0]) );
state_v[1] = _mm256_load_si256( (__m256i*)(&state[4]) );
in_v [1] = _mm256_loadu_si256( (__m256i*)(&ptrWordIn[4]) );
state_v[2] = _mm256_load_si256( (__m256i*)(&state[8]) );
in_v [2] = _mm256_loadu_si256( (__m256i*)(&ptrWordIn[8]) );
_mm256_store_si256( (__m256i*)(&state[0]),
_mm256_xor_si256( state_v[0], in_v[0] ) );
_mm256_store_si256( (__m256i*)(&state[4]),
_mm256_xor_si256( state_v[1], in_v[1] ) );
_mm256_store_si256( (__m256i*)(&state[8]),
_mm256_xor_si256( state_v[2], in_v[2] ) );
#elif defined __AVX__
__m128i state_v[6], in_v[6];
state_v[0] = _mm_load_si128( (__m128i*)(&state[0]) );
state_v[1] = _mm_load_si128( (__m128i*)(&state[2]) );
state_v[2] = _mm_load_si128( (__m128i*)(&state[4]) );
state_v[3] = _mm_load_si128( (__m128i*)(&state[6]) );
state_v[4] = _mm_load_si128( (__m128i*)(&state[8]) );
state_v[5] = _mm_load_si128( (__m128i*)(&state[10]) );
in_v[0] = _mm_loadu_si128( (__m128i*)(&ptrWordIn[0]) );
in_v[1] = _mm_loadu_si128( (__m128i*)(&ptrWordIn[2]) );
in_v[2] = _mm_loadu_si128( (__m128i*)(&ptrWordIn[4]) );
in_v[3] = _mm_loadu_si128( (__m128i*)(&ptrWordIn[6]) );
in_v[4] = _mm_loadu_si128( (__m128i*)(&ptrWordIn[8]) );
in_v[5] = _mm_loadu_si128( (__m128i*)(&ptrWordIn[10]) );
_mm_store_si128( (__m128i*)(&state[0]),
_mm_xor_si128( state_v[0], in_v[0] ) );
_mm_store_si128( (__m128i*)(&state[2]),
_mm_xor_si128( state_v[1], in_v[1] ) );
_mm_store_si128( (__m128i*)(&state[4]),
_mm_xor_si128( state_v[2], in_v[2] ) );
_mm_store_si128( (__m128i*)(&state[6]),
_mm_xor_si128( state_v[3], in_v[3] ) );
_mm_store_si128( (__m128i*)(&state[8]),
_mm_xor_si128( state_v[4], in_v[4] ) );
_mm_store_si128( (__m128i*)(&state[10]),
_mm_xor_si128( state_v[5], in_v[5] ) );
#else
state[0] ^= (ptrWordIn[0]);
state[1] ^= (ptrWordIn[1]);
state[2] ^= (ptrWordIn[2]);
state[3] ^= (ptrWordIn[3]);
state[4] ^= (ptrWordIn[4]);
state[5] ^= (ptrWordIn[5]);
state[6] ^= (ptrWordIn[6]);
state[7] ^= (ptrWordIn[7]);
state[8] ^= (ptrWordIn[8]);
state[9] ^= (ptrWordIn[9]);
state[10] ^= (ptrWordIn[10]);
state[11] ^= (ptrWordIn[11]);
#endif
//Applies the reduced-round transformation f to the sponge's state
reducedBlake2bLyra(state);
//M[row][C-1-col] = M[prev][col] XOR rand
#if defined __AVX2__
// in_v should not need to be reloaded, but it does and it segfaults if
// loading alogned
state_v[0] = _mm256_load_si256( (__m256i*)(&state[0]) );
in_v [0] = _mm256_loadu_si256( (__m256i*)(&ptrWordIn[0]) );
state_v[1] = _mm256_load_si256( (__m256i*)(&state[4]) );
in_v [1] = _mm256_loadu_si256( (__m256i*)(&ptrWordIn[4]) );
state_v[2] = _mm256_load_si256( (__m256i*)(&state[8]) );
in_v [2] = _mm256_loadu_si256( (__m256i*)(&ptrWordIn[8]) );
_mm256_storeu_si256( (__m256i*)(&ptrWordOut[0]),
_mm256_xor_si256( state_v[0], in_v[0] ) );
_mm256_storeu_si256( (__m256i*)(&ptrWordOut[4]),
_mm256_xor_si256( state_v[1], in_v[1] ) );
_mm256_storeu_si256( (__m256i*)(&ptrWordOut[8]),
_mm256_xor_si256( state_v[2], in_v[2] ) );
#elif defined __AVX__
state_v[0] = _mm_load_si128( (__m128i*)(&state[0]) );
state_v[1] = _mm_load_si128( (__m128i*)(&state[2]) );
state_v[2] = _mm_load_si128( (__m128i*)(&state[4]) );
state_v[3] = _mm_load_si128( (__m128i*)(&state[6]) );
state_v[4] = _mm_load_si128( (__m128i*)(&state[8]) );
state_v[5] = _mm_load_si128( (__m128i*)(&state[10]) );
in_v[0] = _mm_loadu_si128( (__m128i*)(&ptrWordIn[0]) );
in_v[1] = _mm_loadu_si128( (__m128i*)(&ptrWordIn[2]) );
in_v[2] = _mm_loadu_si128( (__m128i*)(&ptrWordIn[4]) );
in_v[3] = _mm_loadu_si128( (__m128i*)(&ptrWordIn[6]) );
in_v[4] = _mm_loadu_si128( (__m128i*)(&ptrWordIn[8]) );
in_v[5] = _mm_loadu_si128( (__m128i*)(&ptrWordIn[10]) );
_mm_storeu_si128( (__m128i*)(&ptrWordOut[0]),
_mm_xor_si128( state_v[0], in_v[0] ) );
_mm_storeu_si128( (__m128i*)(&ptrWordOut[2]),
_mm_xor_si128( state_v[1], in_v[1] ) );
_mm_storeu_si128( (__m128i*)(&ptrWordOut[4]),
_mm_xor_si128( state_v[2], in_v[2] ) );
_mm_storeu_si128( (__m128i*)(&ptrWordOut[6]),
_mm_xor_si128( state_v[3], in_v[3] ) );
_mm_storeu_si128( (__m128i*)(&ptrWordOut[8]),
_mm_xor_si128( state_v[4], in_v[4] ) );
_mm_storeu_si128( (__m128i*)(&ptrWordOut[10]),
_mm_xor_si128( state_v[5], in_v[5] ) );
#else
ptrWordOut[0] = ptrWordIn[0] ^ state[0];
ptrWordOut[1] = ptrWordIn[1] ^ state[1];
ptrWordOut[2] = ptrWordIn[2] ^ state[2];
ptrWordOut[3] = ptrWordIn[3] ^ state[3];
ptrWordOut[4] = ptrWordIn[4] ^ state[4];
ptrWordOut[5] = ptrWordIn[5] ^ state[5];
ptrWordOut[6] = ptrWordIn[6] ^ state[6];
ptrWordOut[7] = ptrWordIn[7] ^ state[7];
ptrWordOut[8] = ptrWordIn[8] ^ state[8];
ptrWordOut[9] = ptrWordIn[9] ^ state[9];
ptrWordOut[10] = ptrWordIn[10] ^ state[10];
ptrWordOut[11] = ptrWordIn[11] ^ state[11];
#endif
//Input: next column (i.e., next block in sequence)
ptrWordIn += BLOCK_LEN_INT64;
//Output: goes to previous column
ptrWordOut -= BLOCK_LEN_INT64;
}
}
EvalSum(const EvalSum& es) {
_mm256_store_si256(&mm, es.mm);
}
EvalSum& operator = (const EvalSum& rhs) {
_mm256_store_si256(&mm, rhs.mm);
return *this;
}
/*!
* \brief Aligned store of the given packed vector at the
* given memory position
*/
ETL_STATIC_INLINE(void) store(int64_t* memory, avx_simd_long value) {
_mm256_store_si256(reinterpret_cast<__m256i*>(memory), value.value);
}
SIMD_INLINE void InterpolateX4(const __m256i * alpha, __m256i * buffer)
{
__m256i src = _mm256_shuffle_epi8(_mm256_load_si256(buffer), K8_SHUFFLE_X4);
_mm256_store_si256(buffer, _mm256_maddubs_epi16(src, _mm256_load_si256(alpha)));
}
/**
* Performs a reduced duplex operation for a single row, from the highest to
* the lowest index, using the reduced-round Blake2b's G function as the
* internal permutation
*
* @param state The current state of the sponge
* @param rowIn Row to feed the sponge
* @param rowOut Row to receive the sponge's output
*/
inline void reducedDuplexRow1( uint64_t *state, uint64_t *rowIn,
uint64_t *rowOut, uint64_t nCols )
{
uint64_t* ptrWordIn = rowIn; //In Lyra2: pointer to prev
uint64_t* ptrWordOut = rowOut + (nCols-1)*BLOCK_LEN_INT64; //In Lyra2: pointer to row
int i;
#if defined __AVX2__
__m256i state_v[4], in_v[3];
state_v[0] = _mm256_load_si256( (__m256i*)(&state[0]) );
state_v[1] = _mm256_load_si256( (__m256i*)(&state[4]) );
state_v[2] = _mm256_load_si256( (__m256i*)(&state[8]) );
state_v[3] = _mm256_load_si256( (__m256i*)(&state[12]) );
#endif
for ( i = 0; i < nCols; i++ )
{
#if defined __AVX2__
in_v [0] = _mm256_loadu_si256( (__m256i*)(&ptrWordIn[0]) );
in_v [1] = _mm256_loadu_si256( (__m256i*)(&ptrWordIn[4]) );
in_v [2] = _mm256_loadu_si256( (__m256i*)(&ptrWordIn[8]) );
state_v[0] = _mm256_xor_si256( state_v[0], in_v[0] );
state_v[1] = _mm256_xor_si256( state_v[1], in_v[1] );
state_v[2] = _mm256_xor_si256( state_v[2], in_v[2] );
LYRA_ROUND_AVX2( state_v[0], state_v[1], state_v[2], state_v[3] );
_mm256_storeu_si256( (__m256i*)(&ptrWordOut[0]),
_mm256_xor_si256( state_v[0], in_v[0] ) );
_mm256_storeu_si256( (__m256i*)(&ptrWordOut[4]),
_mm256_xor_si256( state_v[1], in_v[1] ) );
_mm256_storeu_si256( (__m256i*)(&ptrWordOut[8]),
_mm256_xor_si256( state_v[2], in_v[2] ) );
#elif defined __AVX__
__m128i state_v[6], in_v[6];
state_v[0] = _mm_load_si128( (__m128i*)(&state[0]) );
state_v[1] = _mm_load_si128( (__m128i*)(&state[2]) );
state_v[2] = _mm_load_si128( (__m128i*)(&state[4]) );
state_v[3] = _mm_load_si128( (__m128i*)(&state[6]) );
state_v[4] = _mm_load_si128( (__m128i*)(&state[8]) );
state_v[5] = _mm_load_si128( (__m128i*)(&state[10]) );
in_v[0] = _mm_load_si128( (__m128i*)(&ptrWordIn[0]) );
in_v[1] = _mm_load_si128( (__m128i*)(&ptrWordIn[2]) );
in_v[2] = _mm_load_si128( (__m128i*)(&ptrWordIn[4]) );
in_v[3] = _mm_load_si128( (__m128i*)(&ptrWordIn[6]) );
in_v[4] = _mm_load_si128( (__m128i*)(&ptrWordIn[8]) );
in_v[5] = _mm_load_si128( (__m128i*)(&ptrWordIn[10]) );
_mm_store_si128( (__m128i*)(&state[0]),
_mm_xor_si128( state_v[0], in_v[0] ) );
_mm_store_si128( (__m128i*)(&state[2]),
_mm_xor_si128( state_v[1], in_v[1] ) );
_mm_store_si128( (__m128i*)(&state[4]),
_mm_xor_si128( state_v[2], in_v[2] ) );
_mm_store_si128( (__m128i*)(&state[6]),
_mm_xor_si128( state_v[3], in_v[3] ) );
_mm_store_si128( (__m128i*)(&state[8]),
_mm_xor_si128( state_v[4], in_v[4] ) );
_mm_store_si128( (__m128i*)(&state[10]),
_mm_xor_si128( state_v[5], in_v[5] ) );
//Applies the reduced-round transformation f to the sponge's state
reducedBlake2bLyra(state);
#else
//Absorbing "M[prev][col]"
state[0] ^= (ptrWordIn[0]);
state[1] ^= (ptrWordIn[1]);
state[2] ^= (ptrWordIn[2]);
state[3] ^= (ptrWordIn[3]);
state[4] ^= (ptrWordIn[4]);
state[5] ^= (ptrWordIn[5]);
state[6] ^= (ptrWordIn[6]);
state[7] ^= (ptrWordIn[7]);
state[8] ^= (ptrWordIn[8]);
state[9] ^= (ptrWordIn[9]);
state[10] ^= (ptrWordIn[10]);
state[11] ^= (ptrWordIn[11]);
//Applies the reduced-round transformation f to the sponge's state
reducedBlake2bLyra(state);
#endif
#if defined __AVX2__
/* state_v[0] = _mm256_load_si256( (__m256i*)(&state[0]) );
state_v[1] = _mm256_load_si256( (__m256i*)(&state[4]) );
state_v[2] = _mm256_load_si256( (__m256i*)(&state[8]) );
_mm256_storeu_si256( (__m256i*)(&ptrWordOut[0]),
_mm256_xor_si256( state_v[0], in_v[0] ) );
_mm256_storeu_si256( (__m256i*)(&ptrWordOut[4]),
_mm256_xor_si256( state_v[1], in_v[1] ) );
_mm256_storeu_si256( (__m256i*)(&ptrWordOut[8]),
_mm256_xor_si256( state_v[2], in_v[2] ) );
*/
#elif defined __AVX__
state_v[0] = _mm_load_si128( (__m128i*)(&state[0]) );
state_v[1] = _mm_load_si128( (__m128i*)(&state[2]) );
state_v[2] = _mm_load_si128( (__m128i*)(&state[4]) );
state_v[3] = _mm_load_si128( (__m128i*)(&state[6]) );
state_v[4] = _mm_load_si128( (__m128i*)(&state[8]) );
state_v[5] = _mm_load_si128( (__m128i*)(&state[10]) );
_mm_storeu_si128( (__m128i*)(&ptrWordOut[0]),
_mm_xor_si128( state_v[0], in_v[0] ) );
_mm_storeu_si128( (__m128i*)(&ptrWordOut[2]),
_mm_xor_si128( state_v[1], in_v[1] ) );
_mm_storeu_si128( (__m128i*)(&ptrWordOut[4]),
_mm_xor_si128( state_v[2], in_v[2] ) );
_mm_storeu_si128( (__m128i*)(&ptrWordOut[6]),
_mm_xor_si128( state_v[3], in_v[3] ) );
_mm_storeu_si128( (__m128i*)(&ptrWordOut[8]),
_mm_xor_si128( state_v[4], in_v[4] ) );
_mm_storeu_si128( (__m128i*)(&ptrWordOut[10]),
_mm_xor_si128( state_v[5], in_v[5] ) );
#else
//M[row][C-1-col] = M[prev][col] XOR rand
ptrWordOut[0] = ptrWordIn[0] ^ state[0];
ptrWordOut[1] = ptrWordIn[1] ^ state[1];
ptrWordOut[2] = ptrWordIn[2] ^ state[2];
ptrWordOut[3] = ptrWordIn[3] ^ state[3];
ptrWordOut[4] = ptrWordIn[4] ^ state[4];
ptrWordOut[5] = ptrWordIn[5] ^ state[5];
ptrWordOut[6] = ptrWordIn[6] ^ state[6];
ptrWordOut[7] = ptrWordIn[7] ^ state[7];
ptrWordOut[8] = ptrWordIn[8] ^ state[8];
ptrWordOut[9] = ptrWordIn[9] ^ state[9];
ptrWordOut[10] = ptrWordIn[10] ^ state[10];
ptrWordOut[11] = ptrWordIn[11] ^ state[11];
#endif
//Input: next column (i.e., next block in sequence)
ptrWordIn += BLOCK_LEN_INT64;
//Output: goes to previous column
ptrWordOut -= BLOCK_LEN_INT64;
}
#if defined __AVX2__
_mm256_store_si256( (__m256i*)&state[0], state_v[0] );
_mm256_store_si256( (__m256i*)&state[4], state_v[1] );
_mm256_store_si256( (__m256i*)&state[8], state_v[2] );
_mm256_store_si256( (__m256i*)&state[12], state_v[3] );
#endif
}
/**
* Performs a duplexing operation over "M[rowInOut][col] [+] M[rowIn][col]" (i.e.,
* the wordwise addition of two columns, ignoring carries between words). The
* output of this operation, "rand", is then used to make
* "M[rowOut][(N_COLS-1)-col] = M[rowIn][col] XOR rand" and
* "M[rowInOut][col] = M[rowInOut][col] XOR rotW(rand)", where rotW is a 64-bit
* rotation to the left and N_COLS is a system parameter.
*
* @param state The current state of the sponge
* @param rowIn Row used only as input
* @param rowInOut Row used as input and to receive output after rotation
* @param rowOut Row receiving the output
*
*/
inline void reducedDuplexRowSetup( uint64_t *state, uint64_t *rowIn,
uint64_t *rowInOut, uint64_t *rowOut,
uint64_t nCols )
{
uint64_t* ptrWordIn = rowIn; //In Lyra2: pointer to prev
uint64_t* ptrWordInOut = rowInOut; //In Lyra2: pointer to row*
uint64_t* ptrWordOut = rowOut + (nCols-1)*BLOCK_LEN_INT64; //In Lyra2: pointer to row
int i;
#if defined __AVX2__
__m256i state_v[4], in_v[3], inout_v[3];
#define t_state in_v
state_v[0] = _mm256_load_si256( (__m256i*)(&state[0]) );
state_v[1] = _mm256_load_si256( (__m256i*)(&state[4]) );
state_v[2] = _mm256_load_si256( (__m256i*)(&state[8]) );
state_v[3] = _mm256_load_si256( (__m256i*)(&state[12]) );
for ( i = 0; i < nCols; i++ )
{
in_v [0] = _mm256_loadu_si256( (__m256i*)(&ptrWordIn[0]) );
inout_v[0] = _mm256_loadu_si256( (__m256i*)(&ptrWordInOut[0]) );
in_v [1] = _mm256_loadu_si256( (__m256i*)(&ptrWordIn[4]) );
inout_v[1] = _mm256_loadu_si256( (__m256i*)(&ptrWordInOut[4]) );
in_v [2] = _mm256_loadu_si256( (__m256i*)(&ptrWordIn[8]) );
inout_v[2] = _mm256_loadu_si256( (__m256i*)(&ptrWordInOut[8]) );
state_v[0] = _mm256_xor_si256( state_v[0], _mm256_add_epi64( in_v[0],
inout_v[0] ) );
state_v[1] = _mm256_xor_si256( state_v[1], _mm256_add_epi64( in_v[1],
inout_v[1] ) );
state_v[2] = _mm256_xor_si256( state_v[2], _mm256_add_epi64( in_v[2],
inout_v[2] ) );
LYRA_ROUND_AVX2( state_v[0], state_v[1], state_v[2], state_v[3] );
_mm256_storeu_si256( (__m256i*)(&ptrWordOut[0]),
_mm256_xor_si256( state_v[0], in_v[0] ) );
_mm256_storeu_si256( (__m256i*)(&ptrWordOut[4]),
_mm256_xor_si256( state_v[1], in_v[1] ) );
_mm256_storeu_si256( (__m256i*)(&ptrWordOut[8]),
_mm256_xor_si256( state_v[2], in_v[2] ) );
//M[row*][col] = M[row*][col] XOR rotW(rand)
t_state[0] = _mm256_permute4x64_epi64( state_v[0], 0x93 );
t_state[1] = _mm256_permute4x64_epi64( state_v[1], 0x93 );
t_state[2] = _mm256_permute4x64_epi64( state_v[2], 0x93 );
inout_v[0] = _mm256_xor_si256( inout_v[0],
_mm256_blend_epi32( t_state[0], t_state[2], 0x03 ) );
inout_v[1] = _mm256_xor_si256( inout_v[1],
_mm256_blend_epi32( t_state[1], t_state[0], 0x03 ) );
inout_v[2] = _mm256_xor_si256( inout_v[2],
_mm256_blend_epi32( t_state[2], t_state[1], 0x03 ) );
_mm256_storeu_si256( (__m256i*)&ptrWordInOut[0], inout_v[0] );
_mm256_storeu_si256( (__m256i*)&ptrWordInOut[4], inout_v[1] );
_mm256_storeu_si256( (__m256i*)&ptrWordInOut[8], inout_v[2] );
//Inputs: next column (i.e., next block in sequence)
ptrWordInOut += BLOCK_LEN_INT64;
ptrWordIn += BLOCK_LEN_INT64;
//Output: goes to previous column
ptrWordOut -= BLOCK_LEN_INT64;
}
_mm256_store_si256( (__m256i*)&state[0], state_v[0] );
_mm256_store_si256( (__m256i*)&state[4], state_v[1] );
_mm256_store_si256( (__m256i*)&state[8], state_v[2] );
_mm256_store_si256( (__m256i*)&state[12], state_v[3] );
#undef t_state
#elif defined __AVX__
__m128i state_v[6], in_v[6], inout_v[6];
for ( i = 0; i < nCols; i++ )
{
state_v[0] = _mm_load_si128( (__m128i*)(&state[0]) );
state_v[1] = _mm_load_si128( (__m128i*)(&state[2]) );
state_v[2] = _mm_load_si128( (__m128i*)(&state[4]) );
state_v[3] = _mm_load_si128( (__m128i*)(&state[6]) );
state_v[4] = _mm_load_si128( (__m128i*)(&state[8]) );
state_v[5] = _mm_load_si128( (__m128i*)(&state[10]) );
inout_v[0] = _mm_load_si128( (__m128i*)(&ptrWordInOut[0]) );
inout_v[1] = _mm_load_si128( (__m128i*)(&ptrWordInOut[2]) );
inout_v[2] = _mm_load_si128( (__m128i*)(&ptrWordInOut[4]) );
inout_v[3] = _mm_load_si128( (__m128i*)(&ptrWordInOut[6]) );
inout_v[4] = _mm_load_si128( (__m128i*)(&ptrWordInOut[8]) );
inout_v[5] = _mm_load_si128( (__m128i*)(&ptrWordInOut[10]) );
in_v[0] = _mm_load_si128( (__m128i*)(&ptrWordIn[0]) );
in_v[1] = _mm_load_si128( (__m128i*)(&ptrWordIn[2]) );
in_v[2] = _mm_load_si128( (__m128i*)(&ptrWordIn[4]) );
in_v[3] = _mm_load_si128( (__m128i*)(&ptrWordIn[6]) );
in_v[4] = _mm_load_si128( (__m128i*)(&ptrWordIn[8]) );
in_v[5] = _mm_load_si128( (__m128i*)(&ptrWordIn[10]) );
_mm_store_si128( (__m128i*)(&state[0]),
_mm_xor_si128( state_v[0],
_mm_add_epi64( in_v[0],
inout_v[0] ) ) );
_mm_store_si128( (__m128i*)(&state[2]),
_mm_xor_si128( state_v[1],
_mm_add_epi64( in_v[1],
inout_v[1] ) ) );
_mm_store_si128( (__m128i*)(&state[4]),
_mm_xor_si128( state_v[2],
_mm_add_epi64( in_v[2],
inout_v[2] ) ) );
_mm_store_si128( (__m128i*)(&state[6]),
_mm_xor_si128( state_v[3],
_mm_add_epi64( in_v[3],
inout_v[3] ) ) );
_mm_store_si128( (__m128i*)(&state[8]),
_mm_xor_si128( state_v[4],
_mm_add_epi64( in_v[4],
inout_v[4] ) ) );
_mm_store_si128( (__m128i*)(&state[10]),
_mm_xor_si128( state_v[5],
_mm_add_epi64( in_v[5],
inout_v[5] ) ) );
//Applies the reduced-round transformation f to the sponge's state
reducedBlake2bLyra(state);
#else
for ( i = 0; i < nCols; i++ )
{
//Absorbing "M[prev] [+] M[row*]"
state[0] ^= (ptrWordIn[0] + ptrWordInOut[0]);
state[1] ^= (ptrWordIn[1] + ptrWordInOut[1]);
state[2] ^= (ptrWordIn[2] + ptrWordInOut[2]);
state[3] ^= (ptrWordIn[3] + ptrWordInOut[3]);
state[4] ^= (ptrWordIn[4] + ptrWordInOut[4]);
state[5] ^= (ptrWordIn[5] + ptrWordInOut[5]);
state[6] ^= (ptrWordIn[6] + ptrWordInOut[6]);
state[7] ^= (ptrWordIn[7] + ptrWordInOut[7]);
state[8] ^= (ptrWordIn[8] + ptrWordInOut[8]);
state[9] ^= (ptrWordIn[9] + ptrWordInOut[9]);
state[10] ^= (ptrWordIn[10] + ptrWordInOut[10]);
state[11] ^= (ptrWordIn[11] + ptrWordInOut[11]);
//Applies the reduced-round transformation f to the sponge's state
reducedBlake2bLyra(state);
//M[row][col] = M[prev][col] XOR rand
#endif
#if defined __AVX2__
#elif defined __AVX__
state_v[0] = _mm_load_si128( (__m128i*)(&state[0]) );
state_v[1] = _mm_load_si128( (__m128i*)(&state[2]) );
state_v[2] = _mm_load_si128( (__m128i*)(&state[4]) );
state_v[3] = _mm_load_si128( (__m128i*)(&state[6]) );
state_v[4] = _mm_load_si128( (__m128i*)(&state[8]) );
state_v[5] = _mm_load_si128( (__m128i*)(&state[10]) );
_mm_store_si128( (__m128i*)(&ptrWordOut[0]),
_mm_xor_si128( state_v[0], in_v[0] ) );
_mm_store_si128( (__m128i*)(&ptrWordOut[2]),
_mm_xor_si128( state_v[1], in_v[1] ) );
_mm_store_si128( (__m128i*)(&ptrWordOut[4]),
_mm_xor_si128( state_v[2], in_v[2] ) );
_mm_store_si128( (__m128i*)(&ptrWordOut[6]),
_mm_xor_si128( state_v[3], in_v[3] ) );
_mm_store_si128( (__m128i*)(&ptrWordOut[8]),
_mm_xor_si128( state_v[4], in_v[4] ) );
_mm_store_si128( (__m128i*)(&ptrWordOut[10]),
_mm_xor_si128( state_v[5], in_v[5] ) );
#else
ptrWordOut[0] = ptrWordIn[0] ^ state[0];
ptrWordOut[1] = ptrWordIn[1] ^ state[1];
ptrWordOut[2] = ptrWordIn[2] ^ state[2];
ptrWordOut[3] = ptrWordIn[3] ^ state[3];
ptrWordOut[4] = ptrWordIn[4] ^ state[4];
ptrWordOut[5] = ptrWordIn[5] ^ state[5];
ptrWordOut[6] = ptrWordIn[6] ^ state[6];
ptrWordOut[7] = ptrWordIn[7] ^ state[7];
ptrWordOut[8] = ptrWordIn[8] ^ state[8];
ptrWordOut[9] = ptrWordIn[9] ^ state[9];
ptrWordOut[10] = ptrWordIn[10] ^ state[10];
ptrWordOut[11] = ptrWordIn[11] ^ state[11];
#endif
//M[row*][col] = M[row*][col] XOR rotW(rand)
// Need to fix this before taking state load/store out of loop
#ifdef __AVX2__
#else
ptrWordInOut[0] ^= state[11];
ptrWordInOut[1] ^= state[0];
ptrWordInOut[2] ^= state[1];
ptrWordInOut[3] ^= state[2];
ptrWordInOut[4] ^= state[3];
ptrWordInOut[5] ^= state[4];
ptrWordInOut[6] ^= state[5];
ptrWordInOut[7] ^= state[6];
ptrWordInOut[8] ^= state[7];
ptrWordInOut[9] ^= state[8];
ptrWordInOut[10] ^= state[9];
ptrWordInOut[11] ^= state[10];
//Inputs: next column (i.e., next block in sequence)
ptrWordInOut += BLOCK_LEN_INT64;
ptrWordIn += BLOCK_LEN_INT64;
//Output: goes to previous column
ptrWordOut -= BLOCK_LEN_INT64;
}
#endif
}
/**
* Performs a duplexing operation over "M[rowInOut][col] [+] M[rowIn][col]" (i.e.,
* the wordwise addition of two columns, ignoring carries between words). The
* output of this operation, "rand", is then used to make
* "M[rowOut][col] = M[rowOut][col] XOR rand" and
* "M[rowInOut][col] = M[rowInOut][col] XOR rotW(rand)", where rotW is a 64-bit
* rotation to the left.
*
* @param state The current state of the sponge
* @param rowIn Row used only as input
* @param rowInOut Row used as input and to receive output after rotation
* @param rowOut Row receiving the output
*
*/
inline void reducedDuplexRow( uint64_t *state, uint64_t *rowIn,
uint64_t *rowInOut, uint64_t *rowOut,
uint64_t nCols )
{
uint64_t* ptrWordInOut = rowInOut; //In Lyra2: pointer to row*
uint64_t* ptrWordIn = rowIn; //In Lyra2: pointer to prev
uint64_t* ptrWordOut = rowOut; //In Lyra2: pointer to row
int i;
#if defined __AVX2__
for ( i = 0; i < nCols; i++)
{
//Absorbing "M[prev] [+] M[row*]"
__m256i state_v[4], in_v[3], inout_v[3];
#define out_v in_v // reuse register in next code block
#define t_state in_v
state_v[0] = _mm256_load_si256( (__m256i*)(&state[0]) );
in_v [0] = _mm256_loadu_si256( (__m256i*)(&ptrWordIn[0]) );
inout_v[0] = _mm256_loadu_si256( (__m256i*)(&ptrWordInOut[0]) );
state_v[1] = _mm256_load_si256( (__m256i*)(&state[4]) );
in_v [1] = _mm256_loadu_si256( (__m256i*)(&ptrWordIn[4]) );
inout_v[1] = _mm256_loadu_si256( (__m256i*)(&ptrWordInOut[4]) );
state_v[2] = _mm256_load_si256( (__m256i*)(&state[8]) );
in_v [2] = _mm256_loadu_si256( (__m256i*)(&ptrWordIn[8]) );
inout_v[2] = _mm256_loadu_si256( (__m256i*)(&ptrWordInOut[8]) );
state_v[3] = _mm256_load_si256( (__m256i*)(&state[12]) );
state_v[0] = _mm256_xor_si256( state_v[0], _mm256_add_epi64( in_v[0],
inout_v[0] ) );
state_v[1] = _mm256_xor_si256( state_v[1], _mm256_add_epi64( in_v[1],
inout_v[1] ) );
state_v[2] = _mm256_xor_si256( state_v[2], _mm256_add_epi64( in_v[2],
inout_v[2] ) );
out_v[0] = _mm256_loadu_si256( (__m256i*)(&ptrWordOut[0]) );
out_v[1] = _mm256_loadu_si256( (__m256i*)(&ptrWordOut[4]) );
out_v[2] = _mm256_loadu_si256( (__m256i*)(&ptrWordOut[8]) );
LYRA_ROUND_AVX2( state_v[0], state_v[1], state_v[2], state_v[3] );
_mm256_store_si256( (__m256i*)&state[0], state_v[0] );
_mm256_store_si256( (__m256i*)&state[4], state_v[1] );
_mm256_store_si256( (__m256i*)&state[8], state_v[2] );
_mm256_store_si256( (__m256i*)&state[12], state_v[3] );
_mm256_storeu_si256( (__m256i*)(&ptrWordOut[0]),
_mm256_xor_si256( state_v[0], out_v[0] ) );
_mm256_storeu_si256( (__m256i*)(&ptrWordOut[4]),
_mm256_xor_si256( state_v[1], out_v[1] ) );
_mm256_storeu_si256( (__m256i*)(&ptrWordOut[8]),
_mm256_xor_si256( state_v[2], out_v[2] ) );
/*
t_state[0] = _mm256_permute4x64_epi64( state_v[0], 0x93 );
t_state[1] = _mm256_permute4x64_epi64( state_v[1], 0x93 );
t_state[2] = _mm256_permute4x64_epi64( state_v[2], 0x93 );
inout_v[0] = _mm256_xor_si256( inout_v[0],
_mm256_blend_epi32( t_state[0], t_state[2], 0x03 ) );
inout_v[1] = _mm256_xor_si256( inout_v[1],
_mm256_blend_epi32( t_state[1], t_state[0], 0x03 ) );
inout_v[2] = _mm256_xor_si256( inout_v[2],
_mm256_blend_epi32( t_state[2], t_state[1], 0x03 ) );
_mm256_storeu_si256( (__m256i*)(&ptrWordInOut[0]), inout_v[0] );
_mm256_storeu_si256( (__m256i*)(&ptrWordInOut[4]), inout_v[1] );
_mm256_storeu_si256( (__m256i*)(&ptrWordInOut[8]), inout_v[2] );
_mm256_store_si256( (__m256i*)&state[0], state_v[0] );
_mm256_store_si256( (__m256i*)&state[4], state_v[1] );
_mm256_store_si256( (__m256i*)&state[8], state_v[2] );
_mm256_store_si256( (__m256i*)&state[12], state_v[3] );
*/
#undef out_v
#undef t_state
//M[rowInOut][col] = M[rowInOut][col] XOR rotW(rand)
ptrWordInOut[0] ^= state[11];
ptrWordInOut[1] ^= state[0];
ptrWordInOut[2] ^= state[1];
ptrWordInOut[3] ^= state[2];
ptrWordInOut[4] ^= state[3];
ptrWordInOut[5] ^= state[4];
ptrWordInOut[6] ^= state[5];
ptrWordInOut[7] ^= state[6];
ptrWordInOut[8] ^= state[7];
ptrWordInOut[9] ^= state[8];
ptrWordInOut[10] ^= state[9];
ptrWordInOut[11] ^= state[10];
//Goes to next block
ptrWordOut += BLOCK_LEN_INT64;
ptrWordInOut += BLOCK_LEN_INT64;
ptrWordIn += BLOCK_LEN_INT64;
}
#elif defined __AVX__
for ( i = 0; i < nCols; i++)
{
__m128i state_v[6], in_v[6], inout_v[6];
#define out_v in_v // reuse register in next code block
state_v[0] = _mm_load_si128( (__m128i*)(&state[0]) );
state_v[1] = _mm_load_si128( (__m128i*)(&state[2]) );
state_v[2] = _mm_load_si128( (__m128i*)(&state[4]) );
state_v[3] = _mm_load_si128( (__m128i*)(&state[6]) );
state_v[4] = _mm_load_si128( (__m128i*)(&state[8]) );
state_v[5] = _mm_load_si128( (__m128i*)(&state[10]) );
inout_v[0] = _mm_load_si128( (__m128i*)(&ptrWordInOut[0]) );
inout_v[1] = _mm_load_si128( (__m128i*)(&ptrWordInOut[2]) );
inout_v[2] = _mm_load_si128( (__m128i*)(&ptrWordInOut[4]) );
inout_v[3] = _mm_load_si128( (__m128i*)(&ptrWordInOut[6]) );
inout_v[4] = _mm_load_si128( (__m128i*)(&ptrWordInOut[8]) );
inout_v[5] = _mm_load_si128( (__m128i*)(&ptrWordInOut[10]) );
in_v[0] = _mm_load_si128( (__m128i*)(&ptrWordIn[0]) );
in_v[1] = _mm_load_si128( (__m128i*)(&ptrWordIn[2]) );
in_v[2] = _mm_load_si128( (__m128i*)(&ptrWordIn[4]) );
in_v[3] = _mm_load_si128( (__m128i*)(&ptrWordIn[6]) );
in_v[4] = _mm_load_si128( (__m128i*)(&ptrWordIn[8]) );
in_v[5] = _mm_load_si128( (__m128i*)(&ptrWordIn[10]) );
_mm_store_si128( (__m128i*)(&state[0]),
_mm_xor_si128( state_v[0],
_mm_add_epi64( in_v[0],
inout_v[0] ) ) );
_mm_store_si128( (__m128i*)(&state[2]),
_mm_xor_si128( state_v[1],
_mm_add_epi64( in_v[1],
inout_v[1] ) ) );
_mm_store_si128( (__m128i*)(&state[4]),
_mm_xor_si128( state_v[2],
_mm_add_epi64( in_v[2],
inout_v[2] ) ) );
_mm_store_si128( (__m128i*)(&state[6]),
_mm_xor_si128( state_v[3],
_mm_add_epi64( in_v[3],
inout_v[3] ) ) );
_mm_store_si128( (__m128i*)(&state[8]),
_mm_xor_si128( state_v[4],
_mm_add_epi64( in_v[4],
inout_v[4] ) ) );
_mm_store_si128( (__m128i*)(&state[10]),
_mm_xor_si128( state_v[5],
_mm_add_epi64( in_v[5],
inout_v[5] ) ) );
//Applies the reduced-round transformation f to the sponge's state
reducedBlake2bLyra(state);
#else
for ( i = 0; i < nCols; i++)
{
state[0] ^= (ptrWordIn[0] + ptrWordInOut[0]);
state[1] ^= (ptrWordIn[1] + ptrWordInOut[1]);
state[2] ^= (ptrWordIn[2] + ptrWordInOut[2]);
state[3] ^= (ptrWordIn[3] + ptrWordInOut[3]);
state[4] ^= (ptrWordIn[4] + ptrWordInOut[4]);
state[5] ^= (ptrWordIn[5] + ptrWordInOut[5]);
state[6] ^= (ptrWordIn[6] + ptrWordInOut[6]);
state[7] ^= (ptrWordIn[7] + ptrWordInOut[7]);
state[8] ^= (ptrWordIn[8] + ptrWordInOut[8]);
state[9] ^= (ptrWordIn[9] + ptrWordInOut[9]);
state[10] ^= (ptrWordIn[10] + ptrWordInOut[10]);
state[11] ^= (ptrWordIn[11] + ptrWordInOut[11]);
//Applies the reduced-round transformation f to the sponge's state
reducedBlake2bLyra(state);
#endif
//M[rowOut][col] = M[rowOut][col] XOR rand
#if defined __AVX2__
/*
state_v[0] = _mm256_load_si256( (__m256i*)(&state[0]) );
out_v [0] = _mm256_loadu_si256( (__m256i*)(&ptrWordOut[0]) );
state_v[1] = _mm256_load_si256( (__m256i*)(&state[4]) );
out_v [1] = _mm256_loadu_si256( (__m256i*)(&ptrWordOut[4]) );
state_v[2] = _mm256_load_si256( (__m256i*)(&state[8]) );
out_v [2] = _mm256_loadu_si256( (__m256i*)(&ptrWordOut[8]) );
_mm256_storeu_si256( (__m256i*)(&ptrWordOut[0]),
_mm256_xor_si256( state_v[0], out_v[0] ) );
_mm256_storeu_si256( (__m256i*)(&ptrWordOut[4]),
_mm256_xor_si256( state_v[1], out_v[1] ) );
_mm256_storeu_si256( (__m256i*)(&ptrWordOut[8]),
_mm256_xor_si256( state_v[2], out_v[2] ) );
*/
#elif defined __AVX__
state_v[0] = _mm_load_si128( (__m128i*)(&state[0]) );
state_v[1] = _mm_load_si128( (__m128i*)(&state[2]) );
state_v[2] = _mm_load_si128( (__m128i*)(&state[4]) );
state_v[3] = _mm_load_si128( (__m128i*)(&state[6]) );
state_v[4] = _mm_load_si128( (__m128i*)(&state[8]) );
state_v[5] = _mm_load_si128( (__m128i*)(&state[10]) );
out_v[0] = _mm_load_si128( (__m128i*)(&ptrWordOut[0]) );
out_v[1] = _mm_load_si128( (__m128i*)(&ptrWordOut[2]) );
out_v[2] = _mm_load_si128( (__m128i*)(&ptrWordOut[4]) );
out_v[3] = _mm_load_si128( (__m128i*)(&ptrWordOut[6]) );
out_v[4] = _mm_load_si128( (__m128i*)(&ptrWordOut[8]) );
out_v[5] = _mm_load_si128( (__m128i*)(&ptrWordOut[10]) );
_mm_store_si128( (__m128i*)(&ptrWordOut[0]),
_mm_xor_si128( state_v[0], out_v[0] ) );
_mm_store_si128( (__m128i*)(&ptrWordOut[2]),
_mm_xor_si128( state_v[1], out_v[1] ) );
_mm_store_si128( (__m128i*)(&ptrWordOut[4]),
_mm_xor_si128( state_v[2], out_v[2] ) );
_mm_store_si128( (__m128i*)(&ptrWordOut[6]),
_mm_xor_si128( state_v[3], out_v[3] ) );
_mm_store_si128( (__m128i*)(&ptrWordOut[8]),
_mm_xor_si128( state_v[4], out_v[4] ) );
_mm_store_si128( (__m128i*)(&ptrWordOut[10]),
_mm_xor_si128( state_v[5], out_v[5] ) );
//M[rowInOut][col] = M[rowInOut][col] XOR rotW(rand)
ptrWordInOut[0] ^= state[11];
ptrWordInOut[1] ^= state[0];
ptrWordInOut[2] ^= state[1];
ptrWordInOut[3] ^= state[2];
ptrWordInOut[4] ^= state[3];
ptrWordInOut[5] ^= state[4];
ptrWordInOut[6] ^= state[5];
ptrWordInOut[7] ^= state[6];
ptrWordInOut[8] ^= state[7];
ptrWordInOut[9] ^= state[8];
ptrWordInOut[10] ^= state[9];
ptrWordInOut[11] ^= state[10];
//Goes to next block
ptrWordOut += BLOCK_LEN_INT64;
ptrWordInOut += BLOCK_LEN_INT64;
ptrWordIn += BLOCK_LEN_INT64;
}
#else
ptrWordOut[0] ^= state[0];
ptrWordOut[1] ^= state[1];
ptrWordOut[2] ^= state[2];
ptrWordOut[3] ^= state[3];
ptrWordOut[4] ^= state[4];
ptrWordOut[5] ^= state[5];
ptrWordOut[6] ^= state[6];
ptrWordOut[7] ^= state[7];
ptrWordOut[8] ^= state[8];
ptrWordOut[9] ^= state[9];
ptrWordOut[10] ^= state[10];
ptrWordOut[11] ^= state[11];
//M[rowInOut][col] = M[rowInOut][col] XOR rotW(rand)
ptrWordInOut[0] ^= state[11];
ptrWordInOut[1] ^= state[0];
ptrWordInOut[2] ^= state[1];
ptrWordInOut[3] ^= state[2];
ptrWordInOut[4] ^= state[3];
ptrWordInOut[5] ^= state[4];
ptrWordInOut[6] ^= state[5];
ptrWordInOut[7] ^= state[6];
ptrWordInOut[8] ^= state[7];
ptrWordInOut[9] ^= state[8];
ptrWordInOut[10] ^= state[9];
ptrWordInOut[11] ^= state[10];
//Goes to next block
ptrWordOut += BLOCK_LEN_INT64;
ptrWordInOut += BLOCK_LEN_INT64;
ptrWordIn += BLOCK_LEN_INT64;
}
#endif
}
/**
* Performs a reduced squeeze operation for a single row, from the highest to
* the lowest index, using the reduced-round Blake2b's G function as the
* internal permutation
*
* @param state The current state of the sponge
* @param rowOut Row to receive the data squeezed
*/
inline void reducedSqueezeRow0( uint64_t* state, uint64_t* rowOut,
uint64_t nCols )
{
uint64_t* ptrWord = rowOut + (nCols-1)*BLOCK_LEN_INT64; //In Lyra2: pointer to M[0][C-1]
int i;
//M[row][C-1-col] = H.reduced_squeeze()
#if defined __AVX2__
__m256i state_v[4];
state_v[0] = _mm256_load_si256( (__m256i*)(&state[0]) );
state_v[1] = _mm256_load_si256( (__m256i*)(&state[4]) );
state_v[2] = _mm256_load_si256( (__m256i*)(&state[8]) );
state_v[3] = _mm256_load_si256( (__m256i*)(&state[12]) );
#endif
for ( i = 0; i < nCols; i++ )
{
#if defined __AVX2__
_mm256_storeu_si256( (__m256i*)&ptrWord[0], state_v[0] );
_mm256_storeu_si256( (__m256i*)&ptrWord[4], state_v[1] );
_mm256_storeu_si256( (__m256i*)&ptrWord[8], state_v[2] );
//Goes to next block (column) that will receive the squeezed data
ptrWord -= BLOCK_LEN_INT64;
LYRA_ROUND_AVX2( state_v[0], state_v[1], state_v[2], state_v[3] );
#elif defined __AVX__
_mm_store_si128( (__m128i*)(&ptrWord[0]),
_mm_load_si128( (__m128i*)(&state[0]) ) );
_mm_store_si128( (__m128i*)(&ptrWord[2]),
_mm_load_si128( (__m128i*)(&state[2]) ) );
_mm_store_si128( (__m128i*)(&ptrWord[4]),
_mm_load_si128( (__m128i*)(&state[4]) ) );
_mm_store_si128( (__m128i*)(&ptrWord[6]),
_mm_load_si128( (__m128i*)(&state[6]) ) );
_mm_store_si128( (__m128i*)(&ptrWord[8]),
_mm_load_si128( (__m128i*)(&state[8]) ) );
_mm_store_si128( (__m128i*)(&ptrWord[10]),
_mm_load_si128( (__m128i*)(&state[10]) ) );
//Goes to next block (column) that will receive the squeezed data
ptrWord -= BLOCK_LEN_INT64;
//Applies the reduced-round transformation f to the sponge's state
reducedBlake2bLyra(state);
#else
ptrWord[0] = state[0];
ptrWord[1] = state[1];
ptrWord[2] = state[2];
ptrWord[3] = state[3];
ptrWord[4] = state[4];
ptrWord[5] = state[5];
ptrWord[6] = state[6];
ptrWord[7] = state[7];
ptrWord[8] = state[8];
ptrWord[9] = state[9];
ptrWord[10] = state[10];
ptrWord[11] = state[11];
//Goes to next block (column) that will receive the squeezed data
ptrWord -= BLOCK_LEN_INT64;
//Applies the reduced-round transformation f to the sponge's state
reducedBlake2bLyra(state);
#endif
}
#if defined __AVX2__
_mm256_store_si256( (__m256i*)&state[0], state_v[0] );
_mm256_store_si256( (__m256i*)&state[4], state_v[1] );
_mm256_store_si256( (__m256i*)&state[8], state_v[2] );
_mm256_store_si256( (__m256i*)&state[12], state_v[3] );
#endif
}
void vec_i8_replace(int8_t *p, size_t n, int8_t val, int8_t substitute)
{
#ifdef COREARRAY_SIMD_SSE2
// header 1, 16-byte aligned
size_t h = (16 - ((size_t)p & 0x0F)) & 0x0F;
for (; (n > 0) && (h > 0); n--, h--, p++)
if (*p == val) *p = substitute;
// body, SSE2
const __m128i mask = _mm_set1_epi8(val);
const __m128i sub = _mm_set1_epi8(substitute);
# ifdef COREARRAY_SIMD_AVX2
// header 2, 32-byte aligned
if ((n >= 16) && ((size_t)p & 0x10))
{
__m128i v = _mm_load_si128((__m128i const*)p);
__m128i c = _mm_cmpeq_epi8(v, mask);
if (_mm_movemask_epi8(c))
{
_mm_store_si128((__m128i *)p,
_mm_or_si128(_mm_and_si128(c, sub), _mm_andnot_si128(c, v)));
}
n -= 16; p += 16;
}
const __m256i mask2 = _mm256_set1_epi8(val);
const __m256i sub32 = _mm256_set1_epi8(substitute);
const __m256i zero = _mm256_setzero_si256();
const __m256i ones = _mm256_cmpeq_epi64(zero, zero);
for (; n >= 32; n-=32, p+=32)
{
__m256i v = _mm256_load_si256((__m256i const*)p);
__m256i c = _mm256_cmpeq_epi8(v, mask2);
if (_mm256_movemask_epi8(c))
{
// TODO
_mm256_store_si256((__m256i *)p,
_mm256_or_si256(_mm256_and_si256(c, sub32),
_mm256_andnot_si256(c, v)));
}
}
# endif
for (; n >= 16; n-=16, p+=16)
{
__m128i v = _mm_load_si128((__m128i const*)p);
__m128i c = _mm_cmpeq_epi8(v, mask);
if (_mm_movemask_epi8(c))
_mm_maskmoveu_si128(sub, c, (char*)p);
}
#endif
// tail
for (; n > 0; n--, p++)
if (*p == val) *p = substitute;
}
static int make_frame_planar_yuv_stacked
(
lw_video_output_handler_t *vohp,
int height,
AVFrame *av_frame,
PVideoFrame &as_frame
)
{
as_picture_t dst_picture = { { { NULL } } };
as_picture_t src_picture = { { { NULL } } };
as_assign_planar_yuv( as_frame, &dst_picture );
lw_video_scaler_handler_t *vshp = &vohp->scaler;
as_video_output_handler_t *as_vohp = (as_video_output_handler_t *)vohp->private_handler;
if( vshp->input_pixel_format == vshp->output_pixel_format )
for( int i = 0; i < 3; i++ )
{
src_picture.data [i] = av_frame->data [i];
src_picture.linesize[i] = av_frame->linesize[i];
}
else
{
if( convert_av_pixel_format( vshp->sws_ctx, height, av_frame, &as_vohp->scaled ) < 0 )
return -1;
src_picture = as_vohp->scaled;
}
for( int i = 0; i < 3; i++ )
{
const int src_height = height >> (i ? as_vohp->sub_height : 0);
const int width = vshp->input_width >> (i ? as_vohp->sub_width : 0);
const int width16 = sse2_available > 0 ? (width & ~15) : 0;
const int width32 = avx2_available > 0 ? (width & ~31) : 0;
const int lsb_offset = src_height * dst_picture.linesize[i];
for( int j = 0; j < src_height; j++ )
{
/* Here, if available, use SIMD instructions.
* Note: There is assumption that the address of a given data can be divided by 32 or 16.
* The destination is always 32 byte alignment unless AviSynth legacy alignment is used.
* The source is not always 32 or 16 byte alignment if the frame buffer is from libavcodec directly. */
static const uint8_t LW_ALIGN(32) sp16[32] =
{
/* saturation protector
* For setting all upper 8 bits to zero so that saturation won't make sense. */
0xFF, 0x00, 0xFF, 0x00, 0xFF, 0x00, 0xFF, 0x00, 0xFF, 0x00, 0xFF, 0x00 ,0xFF, 0x00, 0xFF, 0x00,
0xFF, 0x00, 0xFF, 0x00, 0xFF, 0x00, 0xFF, 0x00, 0xFF, 0x00, 0xFF, 0x00 ,0xFF, 0x00, 0xFF, 0x00
};
uint8_t *dst = dst_picture.data[i] + j * dst_picture.linesize[i]; /* MSB: dst + k, LSB: dst + k + lsb_offset */
const uint8_t *src = src_picture.data[i] + j * src_picture.linesize[i]; /* MSB: src + 2 * k + 1, LSB: src + 2 * k */
const int _width16 = ((intptr_t)src & 15) == 0 ? width16 : 0; /* Don't use SSE2 instructions if set to 0. */
const int _width32 = ((intptr_t)src & 31) == 0 ? width32 : 0; /* Don't use AVX(2) instructions if set to 0. */
#if VC_HAS_AVX2
/* AVX, AVX2 */
for( int k = 0; k < _width32; k += 32 )
{
__m256i ymm0 = _mm256_load_si256( (__m256i *)(src + 2 * k ) );
__m256i ymm1 = _mm256_load_si256( (__m256i *)(src + 2 * k + 32) );
__m256i mask = _mm256_load_si256( (__m256i *)sp16 );
__m256i ymm2 = _mm256_packus_epi16( _mm256_and_si256 ( ymm0, mask ), _mm256_and_si256 ( ymm1, mask ) );
__m256i ymm3 = _mm256_packus_epi16( _mm256_srli_epi16( ymm0, 8 ), _mm256_srli_epi16( ymm1, 8 ) );
_mm256_store_si256( (__m256i *)(dst + k + lsb_offset), _mm256_permute4x64_epi64( ymm2, _MM_SHUFFLE( 3, 1, 2, 0 ) ) );
_mm256_store_si256( (__m256i *)(dst + k ), _mm256_permute4x64_epi64( ymm3, _MM_SHUFFLE( 3, 1, 2, 0 ) ) );
}
#endif
/* SSE2 */
for( int k = _width32; k < _width16; k += 16 )
{
__m128i xmm0 = _mm_load_si128( (__m128i *)(src + 2 * k ) );
__m128i xmm1 = _mm_load_si128( (__m128i *)(src + 2 * k + 16) );
__m128i mask = _mm_load_si128( (__m128i *)sp16 );
_mm_store_si128( (__m128i *)(dst + k + lsb_offset), _mm_packus_epi16( _mm_and_si128 ( xmm0, mask ), _mm_and_si128 ( xmm1, mask ) ) );
_mm_store_si128( (__m128i *)(dst + k ), _mm_packus_epi16( _mm_srli_epi16( xmm0, 8 ), _mm_srli_epi16( xmm1, 8 ) ) );
}
for( int k = _width16; k < width; k++ )
{
*(dst + k + lsb_offset) = *(src + 2 * k );
*(dst + k ) = *(src + 2 * k + 1);
}
}
}
return 0;
}
void vec_i8_cnt_dosage2(const int8_t *p, int8_t *out, size_t n, int8_t val,
int8_t missing, int8_t missing_substitute)
{
#ifdef COREARRAY_SIMD_SSE2
// header 1, 16-byte aligned
size_t h = (16 - ((size_t)out & 0x0F)) & 0x0F;
for (; (n > 0) && (h > 0); n--, h--, p+=2)
{
*out ++ = ((p[0] == missing) || (p[1] == missing)) ?
missing_substitute :
(p[0]==val ? 1 : 0) + (p[1]==val ? 1 : 0);
}
// body, SSE2
const __m128i val16 = _mm_set1_epi8(val);
const __m128i miss16 = _mm_set1_epi8(missing);
const __m128i sub16 = _mm_set1_epi8(missing_substitute);
const __m128i mask = _mm_set1_epi16(0x00FF);
# ifdef COREARRAY_SIMD_AVX2
// header 2, 32-byte aligned
if ((n >= 16) && ((size_t)out & 0x10))
{
__m128i w1 = MM_LOADU_128((__m128i const*)p); p += 16;
__m128i w2 = MM_LOADU_128((__m128i const*)p); p += 16;
__m128i v1 = _mm_packus_epi16(_mm_and_si128(w1, mask), _mm_and_si128(w2, mask));
__m128i v2 = _mm_packus_epi16(_mm_srli_epi16(w1, 8), _mm_srli_epi16(w2, 8));
__m128i c = _mm_setzero_si128();
c = _mm_sub_epi8(c, _mm_cmpeq_epi8(v1, val16));
c = _mm_sub_epi8(c, _mm_cmpeq_epi8(v2, val16));
w1 = _mm_cmpeq_epi8(v1, miss16);
w2 = _mm_cmpeq_epi8(v2, miss16);
__m128i w = _mm_or_si128(w1, w2);
c = _mm_or_si128(_mm_and_si128(w, sub16), _mm_andnot_si128(w, c));
_mm_store_si128((__m128i *)out, c);
n -= 16; out += 16;
}
const __m256i val32 = _mm256_set1_epi8(val);
const __m256i miss32 = _mm256_set1_epi8(missing);
const __m256i sub32 = _mm256_set1_epi8(missing_substitute);
const __m256i mask2 = _mm256_set1_epi16(0x00FF);
for (; n >= 32; n-=32)
{
__m256i w1 = MM_LOADU_256((__m256i const*)p); p += 32;
__m256i w2 = MM_LOADU_256((__m256i const*)p); p += 32;
__m256i v1 = _mm256_packus_epi16(_mm256_and_si256(w1, mask2), _mm256_and_si256(w2, mask2));
__m256i v2 = _mm256_packus_epi16(_mm256_srli_epi16(w1, 8), _mm256_srli_epi16(w2, 8));
__m256i c = _mm256_setzero_si256();
c = _mm256_sub_epi8(c, _mm256_cmpeq_epi8(v1, val32));
c = _mm256_sub_epi8(c, _mm256_cmpeq_epi8(v2, val32));
w1 = _mm256_cmpeq_epi8(v1, miss32);
w2 = _mm256_cmpeq_epi8(v2, miss32);
__m256i w = _mm256_or_si256(w1, w2);
c = _mm256_or_si256(_mm256_and_si256(w, sub32), _mm256_andnot_si256(w, c));
c = _mm256_permute4x64_epi64(c, 0xD8);
_mm256_store_si256((__m256i *)out, c);
out += 32;
}
# endif
// SSE2 only
for (; n >= 16; n-=16)
{
__m128i w1 = MM_LOADU_128((__m128i const*)p); p += 16;
__m128i w2 = MM_LOADU_128((__m128i const*)p); p += 16;
__m128i v1 = _mm_packus_epi16(_mm_and_si128(w1, mask), _mm_and_si128(w2, mask));
__m128i v2 = _mm_packus_epi16(_mm_srli_epi16(w1, 8), _mm_srli_epi16(w2, 8));
__m128i c = _mm_setzero_si128();
c = _mm_sub_epi8(c, _mm_cmpeq_epi8(v1, val16));
c = _mm_sub_epi8(c, _mm_cmpeq_epi8(v2, val16));
w1 = _mm_cmpeq_epi8(v1, miss16);
w2 = _mm_cmpeq_epi8(v2, miss16);
__m128i w = _mm_or_si128(w1, w2);
c = _mm_or_si128(_mm_and_si128(w, sub16), _mm_andnot_si128(w, c));
_mm_store_si128((__m128i *)out, c);
out += 16;
}
#endif
// tail
for (; n > 0; n--, p+=2)
{
*out ++ = ((p[0] == missing) || (p[1] == missing)) ?
missing_substitute :
(p[0]==val ? 1 : 0) + (p[1]==val ? 1 : 0);
}
}
inline static void lyra_round( uint64_t *v )
{
#ifdef __AVX2__
__m256i a = _mm256_load_si256( (__m256i*)(&v[ 0]) );
__m256i b = _mm256_load_si256( (__m256i*)(&v[ 4]) );
__m256i c = _mm256_load_si256( (__m256i*)(&v[ 8]) );
__m256i d = _mm256_load_si256( (__m256i*)(&v[12]) );
G_4X64( a, b, c, d );
// swap words
b = mm256_rotl256_1x64( b );
c = mm256_swap128( c );
d = mm256_rotr256_1x64( d );
G_4X64( a, b, c, d );
// unswap
b = mm256_rotr256_1x64( b );
c = mm256_swap128( c );
d = mm256_rotl256_1x64( d );
_mm256_store_si256( (__m256i*)(&v[ 0]), a );
_mm256_store_si256( (__m256i*)(&v[ 4]), b );
_mm256_store_si256( (__m256i*)(&v[ 8]), c );
_mm256_store_si256( (__m256i*)(&v[12]), d );
#elif defined __AVX__
__m128i a0, a1, b0, b1, c0, c1, d0, d1;
a0 = _mm_load_si128( (__m128i*)(&v[ 0]) );
a1 = _mm_load_si128( (__m128i*)(&v[ 2]) );
b0 = _mm_load_si128( (__m128i*)(&v[ 4]) );
b1 = _mm_load_si128( (__m128i*)(&v[ 6]) );
c0 = _mm_load_si128( (__m128i*)(&v[ 8]) );
c1 = _mm_load_si128( (__m128i*)(&v[10]) );
d0 = _mm_load_si128( (__m128i*)(&v[12]) );
d1 = _mm_load_si128( (__m128i*)(&v[14]) );
G_2X64( a0, b0, c0, d0 );
G_2X64( a1, b1, c1, d1 );
// swap words
mm128_rotl256_1x64( b0, b1 );
mm128_swap128( c0, c1 );
mm128_rotr256_1x64( d0, d1 );
G_2X64( a0, b0, c0, d0 );
G_2X64( a1, b1, c1, d1 );
// unswap
mm128_rotr256_1x64( b0, b1 );
mm128_swap128( c0, c1 );
mm128_rotl256_1x64( d0, d1 );
_mm_store_si128( (__m128i*)(&v[ 0]), a0 );
_mm_store_si128( (__m128i*)(&v[ 2]), a1 );
_mm_store_si128( (__m128i*)(&v[ 4]), b0 );
_mm_store_si128( (__m128i*)(&v[ 6]), b1 );
_mm_store_si128( (__m128i*)(&v[ 8]), c0 );
_mm_store_si128( (__m128i*)(&v[10]), c1 );
_mm_store_si128( (__m128i*)(&v[12]), d0 );
_mm_store_si128( (__m128i*)(&v[14]), d1 );
#else
// macro assumes v is defined
ROUND_LYRA(0);
#endif
}
/**
* Performs a duplexing operation over "M[rowInOut][col] [+] M[rowIn][col]" (i.e.,
* the wordwise addition of two columns, ignoring carries between words). The
* output of this operation, "rand", is then used to make
* "M[rowOut][col] = M[rowOut][col] XOR rand" and
* "M[rowInOut][col] = M[rowInOut][col] XOR rotW(rand)", where rotW is a 64-bit
* rotation to the left.
*
* @param state The current state of the sponge
* @param rowIn Row used only as input
* @param rowInOut Row used as input and to receive output after rotation
* @param rowOut Row receiving the output
*
*/
void reducedDuplexRow(uint64_t *state, uint64_t *rowIn, uint64_t *rowInOut, uint64_t *rowOut, const uint32_t nCols)
{
uint64_t* ptrWordInOut = rowInOut; //In Lyra2: pointer to row*
uint64_t* ptrWordIn = rowIn; //In Lyra2: pointer to prev
uint64_t* ptrWordOut = rowOut; //In Lyra2: pointer to row
unsigned int i;
for (i = 0; i < nCols; i++)
{
//Absorbing "M[prev] [+] M[row*]"
#if defined __AVX2__
__m256i state_v[3], in_v[3], inout_v[3];
#define out_v in_v // reuse register in next code block
state_v[0] = _mm256_load_si256( (__m256i*)(&state[0]) );
in_v [0] = _mm256_loadu_si256( (__m256i*)(&ptrWordIn[0]) );
inout_v[0] = _mm256_loadu_si256( (__m256i*)(&ptrWordInOut[0]) );
state_v[1] = _mm256_load_si256( (__m256i*)(&state[4]) );
in_v [1] = _mm256_loadu_si256( (__m256i*)(&ptrWordIn[4]) );
inout_v[1] = _mm256_loadu_si256( (__m256i*)(&ptrWordInOut[4]) );
state_v[2] = _mm256_load_si256( (__m256i*)(&state[8]) );
in_v [2] = _mm256_loadu_si256( (__m256i*)(&ptrWordIn[8]) );
inout_v[2] = _mm256_loadu_si256( (__m256i*)(&ptrWordInOut[8]) );
_mm256_store_si256( (__m256i*)(&state[0]),
_mm256_xor_si256( state_v[0],
_mm256_add_epi64( in_v[0],
inout_v[0] ) ) );
_mm256_store_si256( (__m256i*)(&state[4]),
_mm256_xor_si256( state_v[1],
_mm256_add_epi64( in_v[1],
inout_v[1] ) ) );
_mm256_store_si256( (__m256i*)(&state[8]),
_mm256_xor_si256( state_v[2],
_mm256_add_epi64( in_v[2],
inout_v[2] ) ) );
#elif defined __AVX__
__m128i state_v[6], in_v[6], inout_v[6];
#define out_v in_v // reuse register in next code block
state_v[0] = _mm_load_si128( (__m128i*)(&state[0]) );
state_v[1] = _mm_load_si128( (__m128i*)(&state[2]) );
state_v[2] = _mm_load_si128( (__m128i*)(&state[4]) );
state_v[3] = _mm_load_si128( (__m128i*)(&state[6]) );
state_v[4] = _mm_load_si128( (__m128i*)(&state[8]) );
state_v[5] = _mm_load_si128( (__m128i*)(&state[10]) );
inout_v[0] = _mm_loadu_si128( (__m128i*)(&ptrWordInOut[0]) );
inout_v[1] = _mm_loadu_si128( (__m128i*)(&ptrWordInOut[2]) );
inout_v[2] = _mm_loadu_si128( (__m128i*)(&ptrWordInOut[4]) );
inout_v[3] = _mm_loadu_si128( (__m128i*)(&ptrWordInOut[6]) );
inout_v[4] = _mm_loadu_si128( (__m128i*)(&ptrWordInOut[8]) );
inout_v[5] = _mm_loadu_si128( (__m128i*)(&ptrWordInOut[10]) );
in_v[0] = _mm_loadu_si128( (__m128i*)(&ptrWordIn[0]) );
in_v[1] = _mm_loadu_si128( (__m128i*)(&ptrWordIn[2]) );
in_v[2] = _mm_loadu_si128( (__m128i*)(&ptrWordIn[4]) );
in_v[3] = _mm_loadu_si128( (__m128i*)(&ptrWordIn[6]) );
in_v[4] = _mm_loadu_si128( (__m128i*)(&ptrWordIn[8]) );
in_v[5] = _mm_loadu_si128( (__m128i*)(&ptrWordIn[10]) );
_mm_store_si128( (__m128i*)(&state[0]),
_mm_xor_si128( state_v[0],
_mm_add_epi64( in_v[0],
inout_v[0] ) ) );
_mm_store_si128( (__m128i*)(&state[2]),
_mm_xor_si128( state_v[1],
_mm_add_epi64( in_v[1],
inout_v[1] ) ) );
_mm_store_si128( (__m128i*)(&state[4]),
_mm_xor_si128( state_v[2],
_mm_add_epi64( in_v[2],
inout_v[2] ) ) );
_mm_store_si128( (__m128i*)(&state[6]),
_mm_xor_si128( state_v[3],
_mm_add_epi64( in_v[3],
inout_v[3] ) ) );
_mm_store_si128( (__m128i*)(&state[8]),
_mm_xor_si128( state_v[4],
_mm_add_epi64( in_v[4],
inout_v[4] ) ) );
_mm_store_si128( (__m128i*)(&state[10]),
_mm_xor_si128( state_v[5],
_mm_add_epi64( in_v[5],
inout_v[5] ) ) );
#else
state[0] ^= (ptrWordIn[0] + ptrWordInOut[0]);
state[1] ^= (ptrWordIn[1] + ptrWordInOut[1]);
state[2] ^= (ptrWordIn[2] + ptrWordInOut[2]);
state[3] ^= (ptrWordIn[3] + ptrWordInOut[3]);
state[4] ^= (ptrWordIn[4] + ptrWordInOut[4]);
state[5] ^= (ptrWordIn[5] + ptrWordInOut[5]);
state[6] ^= (ptrWordIn[6] + ptrWordInOut[6]);
state[7] ^= (ptrWordIn[7] + ptrWordInOut[7]);
state[8] ^= (ptrWordIn[8] + ptrWordInOut[8]);
state[9] ^= (ptrWordIn[9] + ptrWordInOut[9]);
state[10] ^= (ptrWordIn[10] + ptrWordInOut[10]);
state[11] ^= (ptrWordIn[11] + ptrWordInOut[11]);
#endif
//Applies the reduced-round transformation f to the sponge's state
reducedBlake2bLyra(state);
//M[rowOut][col] = M[rowOut][col] XOR rand
#if defined __AVX2__
state_v[0] = _mm256_load_si256( (__m256i*)(&state[0]) );
out_v [0] = _mm256_loadu_si256( (__m256i*)(&ptrWordOut[0]) );
state_v[1] = _mm256_load_si256( (__m256i*)(&state[4]) );
out_v [1] = _mm256_loadu_si256( (__m256i*)(&ptrWordOut[4]) );
state_v[2] = _mm256_load_si256( (__m256i*)(&state[8]) );
out_v [2] = _mm256_loadu_si256( (__m256i*)(&ptrWordOut[8]) );
_mm256_storeu_si256( (__m256i*)(&ptrWordOut[0]),
_mm256_xor_si256( state_v[0], out_v[0] ) );
_mm256_storeu_si256( (__m256i*)(&ptrWordOut[4]),
_mm256_xor_si256( state_v[1], out_v[1] ) );
_mm256_storeu_si256( (__m256i*)(&ptrWordOut[8]),
_mm256_xor_si256( state_v[2], out_v[2] ) );
#elif defined __AVX__
state_v[0] = _mm_load_si128( (__m128i*)(&state[0]) );
state_v[1] = _mm_load_si128( (__m128i*)(&state[2]) );
state_v[2] = _mm_load_si128( (__m128i*)(&state[4]) );
state_v[3] = _mm_load_si128( (__m128i*)(&state[6]) );
state_v[4] = _mm_load_si128( (__m128i*)(&state[8]) );
state_v[5] = _mm_load_si128( (__m128i*)(&state[10]) );
out_v[0] = _mm_loadu_si128( (__m128i*)(&ptrWordOut[0]) );
out_v[1] = _mm_loadu_si128( (__m128i*)(&ptrWordOut[2]) );
out_v[2] = _mm_loadu_si128( (__m128i*)(&ptrWordOut[4]) );
out_v[3] = _mm_loadu_si128( (__m128i*)(&ptrWordOut[6]) );
out_v[4] = _mm_loadu_si128( (__m128i*)(&ptrWordOut[8]) );
out_v[5] = _mm_loadu_si128( (__m128i*)(&ptrWordOut[10]) );
_mm_storeu_si128( (__m128i*)(&ptrWordOut[0]),
_mm_xor_si128( state_v[0], out_v[0] ) );
_mm_storeu_si128( (__m128i*)(&ptrWordOut[2]),
_mm_xor_si128( state_v[1], out_v[1] ) );
_mm_storeu_si128( (__m128i*)(&ptrWordOut[4]),
_mm_xor_si128( state_v[2], out_v[2] ) );
_mm_storeu_si128( (__m128i*)(&ptrWordOut[6]),
_mm_xor_si128( state_v[3], out_v[3] ) );
_mm_storeu_si128( (__m128i*)(&ptrWordOut[8]),
_mm_xor_si128( state_v[4], out_v[4] ) );
_mm_storeu_si128( (__m128i*)(&ptrWordOut[10]),
_mm_xor_si128( state_v[5], out_v[5] ) );
#else
ptrWordOut[0] ^= state[0];
ptrWordOut[1] ^= state[1];
ptrWordOut[2] ^= state[2];
ptrWordOut[3] ^= state[3];
ptrWordOut[4] ^= state[4];
ptrWordOut[5] ^= state[5];
ptrWordOut[6] ^= state[6];
ptrWordOut[7] ^= state[7];
ptrWordOut[8] ^= state[8];
ptrWordOut[9] ^= state[9];
ptrWordOut[10] ^= state[10];
ptrWordOut[11] ^= state[11];
#endif
//M[rowInOut][col] = M[rowInOut][col] XOR rotW(rand)
ptrWordInOut[0] ^= state[11];
ptrWordInOut[1] ^= state[0];
ptrWordInOut[2] ^= state[1];
ptrWordInOut[3] ^= state[2];
ptrWordInOut[4] ^= state[3];
ptrWordInOut[5] ^= state[4];
ptrWordInOut[6] ^= state[5];
ptrWordInOut[7] ^= state[6];
ptrWordInOut[8] ^= state[7];
ptrWordInOut[9] ^= state[8];
ptrWordInOut[10] ^= state[9];
ptrWordInOut[11] ^= state[10];
//Goes to next block
ptrWordOut += BLOCK_LEN_INT64;
ptrWordInOut += BLOCK_LEN_INT64;
ptrWordIn += BLOCK_LEN_INT64;
}
}
void* xmemset(void* dest, int c, size_t n)
{
void* ret = dest;
if (n < 16) {
xmemset_lt16(dest, c, n);
return ret;
}
__m256i mm = _mm256_set1_epi8((char)c);
if (((unsigned long)dest & 31) == 0) {
for ( ; n >= 256; n -= 256) {
_mm256_store_si256((__m256i*)dest, mm);
_mm256_store_si256((__m256i*)dest + 1, mm);
_mm256_store_si256((__m256i*)dest + 2, mm);
_mm256_store_si256((__m256i*)dest + 3, mm);
_mm256_store_si256((__m256i*)dest + 4, mm);
_mm256_store_si256((__m256i*)dest + 5, mm);
_mm256_store_si256((__m256i*)dest + 6, mm);
_mm256_store_si256((__m256i*)dest + 7, mm); // 8
dest = (void*)((__m256i*)dest + 8);
}
if (n >= 128) {
_mm256_store_si256((__m256i*)dest, mm);
_mm256_store_si256((__m256i*)dest + 1, mm);
_mm256_store_si256((__m256i*)dest + 2, mm);
_mm256_store_si256((__m256i*)dest + 3, mm);
dest = (void*)((__m256i*)dest + 4);
n -= 128;
}
if (n >= 64) {
_mm256_store_si256((__m256i*)dest, mm);
_mm256_store_si256((__m256i*)dest + 1, mm);
dest = (void*)((__m256i*)dest + 2);
n -= 64;
}
if (n >= 32) {
_mm256_store_si256((__m256i*)dest, mm);
dest = (void*)((__m256i*)dest + 1);
n -= 32;
}
if (n >= 16) {
_mm_store_si128((__m128i*)dest, _mm_set1_epi8((char)c));
dest = (void*)((__m128i*)dest + 1);
n -= 16;
}
} else {
for ( ; n >= 256; n -= 256) {
_mm256_storeu_si256((__m256i*)dest, mm);
_mm256_storeu_si256((__m256i*)dest + 1, mm);
_mm256_storeu_si256((__m256i*)dest + 2, mm);
_mm256_storeu_si256((__m256i*)dest + 3, mm);
_mm256_storeu_si256((__m256i*)dest + 4, mm);
_mm256_storeu_si256((__m256i*)dest + 5, mm);
_mm256_storeu_si256((__m256i*)dest + 6, mm);
_mm256_storeu_si256((__m256i*)dest + 7, mm); // 8
dest = (void*)((__m256i*)dest + 8);
}
if (n >= 128) {
_mm256_storeu_si256((__m256i*)dest, mm);
_mm256_storeu_si256((__m256i*)dest + 1, mm);
_mm256_storeu_si256((__m256i*)dest + 2, mm);
_mm256_storeu_si256((__m256i*)dest + 3, mm);
dest = (void*)((__m256i*)dest + 4);
n -= 128;
}
if (n >= 64) {
_mm256_storeu_si256((__m256i*)dest, mm);
_mm256_storeu_si256((__m256i*)dest + 1, mm);
dest = (void*)((__m256i*)dest + 2);
n -= 64;
}
if (n >= 32) {
_mm256_storeu_si256((__m256i*)dest, mm);
dest = (void*)((__m256i*)dest + 1);
n -= 32;
}
if (n >= 16) {
_mm_storeu_si128((__m128i*)dest, _mm_set1_epi8((char)c));
dest = (void*)((__m128i*)dest + 1);
n -= 16;
}
}
xmemset_lt16(dest, c, n);
return ret;
}
/**
* Performs an absorb operation for a single block (BLOCK_LEN_INT64 words
* of type uint64_t), using Blake2b's G function as the internal permutation
*
* @param state The current state of the sponge
* @param in The block to be absorbed (BLOCK_LEN_INT64 words)
*/
inline void absorbBlock(uint64_t *state, const uint64_t *in) {
#if defined __AVX2__
__m256i state_v[4], in_v[3];
// only state is guaranteed aligned 256
state_v[0] = _mm256_load_si256( (__m256i*)(&state[0]) );
in_v [0] = _mm256_loadu_si256( (__m256i*)(&in[0]) );
state_v[1] = _mm256_load_si256( (__m256i*)(&state[4]) );
in_v [1] = _mm256_loadu_si256( (__m256i*)(&in[4]) );
state_v[2] = _mm256_load_si256( (__m256i*)(&state[8]) );
in_v [2] = _mm256_loadu_si256( (__m256i*)(&in[8]) );
state_v[3] = _mm256_load_si256( (__m256i*)(&state[12]) );
state_v[0] = _mm256_xor_si256( state_v[0], in_v[0] );
state_v[1] = _mm256_xor_si256( state_v[1], in_v[1] );
state_v[2] = _mm256_xor_si256( state_v[2], in_v[2] );
LYRA_12_ROUNDS_AVX2( state_v[0], state_v[1], state_v[2], state_v[3] );
_mm256_store_si256( (__m256i*)&state[0], state_v[0] );
_mm256_store_si256( (__m256i*)&state[4], state_v[1] );
_mm256_store_si256( (__m256i*)&state[8], state_v[2] );
_mm256_store_si256( (__m256i*)&state[12], state_v[3] );
#elif defined __AVX__
__m128i state_v[6], in_v[6];
state_v[0] = _mm_load_si128( (__m128i*)(&state[0]) );
state_v[1] = _mm_load_si128( (__m128i*)(&state[2]) );
state_v[2] = _mm_load_si128( (__m128i*)(&state[4]) );
state_v[3] = _mm_load_si128( (__m128i*)(&state[6]) );
state_v[4] = _mm_load_si128( (__m128i*)(&state[8]) );
state_v[5] = _mm_load_si128( (__m128i*)(&state[10]) );
in_v[0] = _mm_load_si128( (__m128i*)(&in[0]) );
in_v[1] = _mm_load_si128( (__m128i*)(&in[2]) );
in_v[2] = _mm_load_si128( (__m128i*)(&in[4]) );
in_v[3] = _mm_load_si128( (__m128i*)(&in[6]) );
in_v[4] = _mm_load_si128( (__m128i*)(&in[8]) );
in_v[5] = _mm_load_si128( (__m128i*)(&in[10]) );
// do blake2bLyra without init
// LYRA_ROUND_AVX2( state_v )
_mm_store_si128( (__m128i*)(&state[0]),
_mm_xor_si128( state_v[0], in_v[0] ) );
_mm_store_si128( (__m128i*)(&state[2]),
_mm_xor_si128( state_v[1], in_v[1] ) );
_mm_store_si128( (__m128i*)(&state[4]),
_mm_xor_si128( state_v[2], in_v[2] ) );
_mm_store_si128( (__m128i*)(&state[6]),
_mm_xor_si128( state_v[3], in_v[3] ) );
_mm_store_si128( (__m128i*)(&state[8]),
_mm_xor_si128( state_v[4], in_v[4] ) );
_mm_store_si128( (__m128i*)(&state[10]),
_mm_xor_si128( state_v[5], in_v[5] ) );
//Applies the transformation f to the sponge's state
blake2bLyra(state);
#else
//XORs the first BLOCK_LEN_INT64 words of "in" with the current state
state[0] ^= in[0];
state[1] ^= in[1];
state[2] ^= in[2];
state[3] ^= in[3];
state[4] ^= in[4];
state[5] ^= in[5];
state[6] ^= in[6];
state[7] ^= in[7];
state[8] ^= in[8];
state[9] ^= in[9];
state[10] ^= in[10];
state[11] ^= in[11];
//Applies the transformation f to the sponge's state
blake2bLyra(state);
#endif
}