// There's no equivalent in libc, you'd think so ... std::mismatch exists, but it's not optimized at all. :(
static inline size_t find_change(const uint16_t * a, const uint16_t * b)
{
const __m128i * a128=(const __m128i*)a;
const __m128i * b128=(const __m128i*)b;
while (true)
{
__m128i v0 = _mm_loadu_si128(a128);
__m128i v1 = _mm_loadu_si128(b128);
__m128i c = _mm_cmpeq_epi32(v0, v1);
uint32_t mask = _mm_movemask_epi8(c);
a128++;
b128++;
__m128i v0b = _mm_loadu_si128(a128);
__m128i v1b = _mm_loadu_si128(b128);
__m128i cb = _mm_cmpeq_epi32(v0b, v1b);
uint32_t maskb = _mm_movemask_epi8(cb);
if (mask != 0xffff || maskb != 0xffff) // Something has changed, figure out where.
{
if (mask == 0xffff) mask=maskb;
else a128--;//ignore b128 since we'll return anyways
size_t ret=(((char*)a128-(char*)a) | (compat_ctz(~mask))) >> 1;
return (ret | (a[ret]==b[ret]));
}
a128++;
b128++;
}
}
QT_BEGIN_NAMESPACE
bool convert_ARGB_to_ARGB_PM_inplace_sse2(QImageData *data, Qt::ImageConversionFlags)
{
Q_ASSERT(data->format == QImage::Format_ARGB32);
// extra pixels on each line
const int spare = data->width & 3;
// width in pixels of the pad at the end of each line
const int pad = (data->bytes_per_line >> 2) - data->width;
const int iter = data->width >> 2;
int height = data->height;
const __m128i alphaMask = _mm_set1_epi32(0xff000000);
const __m128i nullVector = _mm_setzero_si128();
const __m128i half = _mm_set1_epi16(0x80);
const __m128i colorMask = _mm_set1_epi32(0x00ff00ff);
__m128i *d = reinterpret_cast<__m128i*>(data->data);
while (height--) {
const __m128i *end = d + iter;
for (; d != end; ++d) {
const __m128i srcVector = _mm_loadu_si128(d);
const __m128i srcVectorAlpha = _mm_and_si128(srcVector, alphaMask);
if (_mm_movemask_epi8(_mm_cmpeq_epi32(srcVectorAlpha, alphaMask)) == 0xffff) {
// opaque, data is unchanged
} else if (_mm_movemask_epi8(_mm_cmpeq_epi32(srcVectorAlpha, nullVector)) == 0xffff) {
// fully transparent
_mm_storeu_si128(d, nullVector);
} else {
__m128i alphaChannel = _mm_srli_epi32(srcVector, 24);
alphaChannel = _mm_or_si128(alphaChannel, _mm_slli_epi32(alphaChannel, 16));
__m128i result;
BYTE_MUL_SSE2(result, srcVector, alphaChannel, colorMask, half);
result = _mm_or_si128(_mm_andnot_si128(alphaMask, result), srcVectorAlpha);
_mm_storeu_si128(d, result);
}
}
QRgb *p = reinterpret_cast<QRgb*>(d);
QRgb *pe = p+spare;
for (; p != pe; ++p) {
if (*p < 0x00ffffff)
*p = 0;
else if (*p < 0xff000000)
*p = PREMUL(*p);
}
d = reinterpret_cast<__m128i*>(p+pad);
}
data->format = QImage::Format_ARGB32_Premultiplied;
return true;
}
const char *ssechr(const char *s, char ch)
{
__m128i zero = _mm_setzero_si128();
__m128i cx16 = _mm_set1_epi8(ch); // (ch) replicated 16 times.
while (1) {
__m128i x = _mm_loadu_si128((__m128i const *)s);
unsigned u = _mm_movemask_epi8(_mm_cmpeq_epi8(zero, x));
unsigned v = _mm_movemask_epi8(_mm_cmpeq_epi8(cx16, x))
& ~u & (u - 1);
if (v) return s + __builtin_ctz(v) - 1;
if (u) return NULL;
s += 16;
}
}
static int VectorMismatch_SSE2(const uint32_t* const array1,
const uint32_t* const array2, int length) {
int match_len;
if (length >= 12) {
__m128i A0 = _mm_loadu_si128((const __m128i*)&array1[0]);
__m128i A1 = _mm_loadu_si128((const __m128i*)&array2[0]);
match_len = 0;
do {
// Loop unrolling and early load both provide a speedup of 10% for the
// current function. Also, max_limit can be MAX_LENGTH=4096 at most.
const __m128i cmpA = _mm_cmpeq_epi32(A0, A1);
const __m128i B0 =
_mm_loadu_si128((const __m128i*)&array1[match_len + 4]);
const __m128i B1 =
_mm_loadu_si128((const __m128i*)&array2[match_len + 4]);
if (_mm_movemask_epi8(cmpA) != 0xffff) break;
match_len += 4;
{
const __m128i cmpB = _mm_cmpeq_epi32(B0, B1);
A0 = _mm_loadu_si128((const __m128i*)&array1[match_len + 4]);
A1 = _mm_loadu_si128((const __m128i*)&array2[match_len + 4]);
if (_mm_movemask_epi8(cmpB) != 0xffff) break;
match_len += 4;
}
} while (match_len + 12 < length);
} else {
match_len = 0;
// Unroll the potential first two loops.
if (length >= 4 &&
_mm_movemask_epi8(_mm_cmpeq_epi32(
_mm_loadu_si128((const __m128i*)&array1[0]),
_mm_loadu_si128((const __m128i*)&array2[0]))) == 0xffff) {
match_len = 4;
if (length >= 8 &&
_mm_movemask_epi8(_mm_cmpeq_epi32(
_mm_loadu_si128((const __m128i*)&array1[4]),
_mm_loadu_si128((const __m128i*)&array2[4]))) == 0xffff) {
match_len = 8;
}
}
}
while (match_len < length && array1[match_len] == array2[match_len]) {
++match_len;
}
return match_len;
}
/*__forceinline*/ bool Cmp_ClutBuffer_GSMem<u32>(u32* GSmem, u32 csa, u32 clutsize)
{
u64* _GSmem = (u64*) GSmem;
u64* clut = (u64*)GetClutBufferAddress<u32>(csa);
while(clutsize > 0) {
#ifdef ZEROGS_SSE2
// Note: local memory datas are swizzles
__m128i GSmem_0 = _mm_load_si128((__m128i*)_GSmem); // 9 8 1 0
__m128i GSmem_1 = _mm_load_si128((__m128i*)_GSmem+1); // 11 10 3 2
__m128i GSmem_2 = _mm_load_si128((__m128i*)_GSmem+2); // 13 12 5 4
__m128i GSmem_3 = _mm_load_si128((__m128i*)_GSmem+3); // 15 14 7 6
__m128i clut_0 = _mm_load_si128((__m128i*)clut);
__m128i clut_1 = _mm_load_si128((__m128i*)clut+1);
__m128i clut_2 = _mm_load_si128((__m128i*)clut+2);
__m128i clut_3 = _mm_load_si128((__m128i*)clut+3);
__m128i result = _mm_cmpeq_epi32(_mm_unpacklo_epi64(GSmem_0, GSmem_1), clut_0);
__m128i result_tmp = _mm_cmpeq_epi32(_mm_unpacklo_epi64(GSmem_2, GSmem_3), clut_1);
result = _mm_and_si128(result, result_tmp);
result_tmp = _mm_cmpeq_epi32(_mm_unpackhi_epi64(GSmem_0, GSmem_1), clut_2);
result = _mm_and_si128(result, result_tmp);
result_tmp = _mm_cmpeq_epi32(_mm_unpackhi_epi64(GSmem_2, GSmem_3), clut_3);
result = _mm_and_si128(result, result_tmp);
u32 result_int = _mm_movemask_epi8(result);
if (result_int != 0xFFFF)
return true;
#else
// I see no point to keep an mmx version. SSE2 versions is probably faster.
// Keep a slow portable C version for reference/debug
// Note: local memory datas are swizzles
if (clut[0] != _GSmem[0] || clut[1] != _GSmem[2] || clut[2] != _GSmem[4] || clut[3] != _GSmem[6]
|| clut[4] != _GSmem[1] || clut[5] != _GSmem[3] || clut[6] != _GSmem[5] || clut[7] != _GSmem[7])
return true;
#endif
// go to the next memory block
_GSmem += 32;
// go back to the previous memory block then down one memory column
if (clutsize & 0x40) {
_GSmem -= (64-8);
}
// In case previous operation (down one column) cross the block boundary
// Go to the next block
if (clutsize == 0x240) {
_GSmem += 32;
}
clut += 8;
clutsize -= 64;
}
return false;
}
bool CPathUtils::ContainsEscapedChars(const char * psz, size_t length)
{
// most of our strings will be tens of bytes long
// -> affort some minor overhead to handle the main part very fast
const char* end = psz + length;
if (sse2supported)
{
__m128i mask = _mm_set_epi8 ( '%', '%', '%', '%', '%', '%', '%', '%'
, '%', '%', '%', '%', '%', '%', '%', '%');
for (; psz + sizeof (mask) <= end; psz += sizeof (mask))
{
// fetch the next 16 bytes from the source
__m128i chunk = _mm_loadu_si128 ((const __m128i*)psz);
// check for non-ASCII
int flags = _mm_movemask_epi8 (_mm_cmpeq_epi8 (chunk, mask));
if (flags != 0)
return true;
};
}
// return odd bytes at the end of the string
for (; psz < end; ++psz)
if (*psz == '%')
return true;
return false;
}
/* Not so fussy on its alignment */
static void
threshold_16_SSE_unaligned(byte *contone_ptr, byte *thresh_ptr, byte *ht_data)
{
__m128i input1;
__m128i input2;
int result_int;
byte *sse_data;
const unsigned int mask1 = 0x80808080;
__m128i sign_fix = _mm_set_epi32(mask1, mask1, mask1, mask1);
sse_data = (byte*) &(result_int);
/* Load */
input1 = _mm_loadu_si128((const __m128i *)contone_ptr);
input2 = _mm_loadu_si128((const __m128i *) thresh_ptr);
/* Unsigned subtraction does Unsigned saturation so we
have to use the signed operation */
input1 = _mm_xor_si128(input1, sign_fix);
input2 = _mm_xor_si128(input2, sign_fix);
/* Subtract the two */
input2 = _mm_subs_epi8(input1, input2);
/* Grab the sign mask */
result_int = _mm_movemask_epi8(input2);
/* bit wise reversal on 16 bit word */
ht_data[0] = bitreverse[sse_data[0]];
ht_data[1] = bitreverse[sse_data[1]];
}
/* Shuffle bits within the bytes of eight element blocks. */
int64_t bshuf_shuffle_bit_eightelem_sse2(void* in, void* out, const size_t size,
const size_t elem_size) {
/* With a bit of care, this could be written such that such that it is */
/* in_buf = out_buf safe. */
char* in_b = (char*) in;
uint16_t* out_ui16 = (uint16_t*) out;
size_t nbyte = elem_size * size;
__m128i xmm;
int32_t bt;
size_t ii, jj, kk;
size_t ind;
CHECK_MULT_EIGHT(size);
if (elem_size % 2) {
bshuf_shuffle_bit_eightelem_scal(in, out, size, elem_size);
} else {
for (ii = 0; ii + 8 * elem_size - 1 < nbyte;
ii += 8 * elem_size) {
for (jj = 0; jj + 15 < 8 * elem_size; jj += 16) {
xmm = _mm_loadu_si128((__m128i *) &in_b[ii + jj]);
for (kk = 0; kk < 8; kk++) {
bt = _mm_movemask_epi8(xmm);
xmm = _mm_slli_epi16(xmm, 1);
ind = (ii + jj / 8 + (7 - kk) * elem_size);
out_ui16[ind / 2] = bt;
}
}
}
}
return size * elem_size;
}
static INLINE unsigned
build_mask_linear(int c, int dcdx, int dcdy)
{
__m128i cstep0 = _mm_setr_epi32(c, c+dcdx, c+dcdx*2, c+dcdx*3);
__m128i xdcdy = _mm_set1_epi32(dcdy);
/* Get values across the quad
*/
__m128i cstep1 = _mm_add_epi32(cstep0, xdcdy);
__m128i cstep2 = _mm_add_epi32(cstep1, xdcdy);
__m128i cstep3 = _mm_add_epi32(cstep2, xdcdy);
/* pack pairs of results into epi16
*/
__m128i cstep01 = _mm_packs_epi32(cstep0, cstep1);
__m128i cstep23 = _mm_packs_epi32(cstep2, cstep3);
/* pack into epi8, preserving sign bits
*/
__m128i result = _mm_packs_epi16(cstep01, cstep23);
/* extract sign bits to create mask
*/
return _mm_movemask_epi8(result);
}
int norx_aead_decrypt(
unsigned char *m, size_t *mlen,
const unsigned char *a, size_t alen,
const unsigned char *c, size_t clen,
const unsigned char *z, size_t zlen,
const unsigned char *nonce,
const unsigned char *key
)
{
const __m128i K = LOADU(key);
__m128i S[4];
if (clen < BYTES(NORX_T)) { return -1; }
*mlen = clen - BYTES(NORX_T);
INITIALISE(S, nonce, K);
ABSORB_DATA(S, a, alen, HEADER_TAG);
DECRYPT_DATA(S, m, c, clen - BYTES(NORX_T));
ABSORB_DATA(S, z, zlen, TRAILER_TAG);
FINALISE(S, K);
/* Verify tag */
S[3] = _mm_cmpeq_epi8(S[3], LOADU(c + clen - BYTES(NORX_T)));
return (((_mm_movemask_epi8(S[3]) & 0xFFFFU) + 1) >> 16) - 1;
}
/* Transpose bits within bytes. */
int64_t bshuf_trans_bit_byte_sse2(void* in, void* out, const size_t size,
const size_t elem_size) {
char* in_b = (char*) in;
char* out_b = (char*) out;
uint16_t* out_ui16;
int64_t count;
size_t nbyte = elem_size * size;
__m128i xmm;
int32_t bt;
size_t ii, kk;
CHECK_MULT_EIGHT(nbyte);
for (ii = 0; ii + 15 < nbyte; ii += 16) {
xmm = _mm_loadu_si128((__m128i *) &in_b[ii]);
for (kk = 0; kk < 8; kk++) {
bt = _mm_movemask_epi8(xmm);
xmm = _mm_slli_epi16(xmm, 1);
out_ui16 = (uint16_t*) &out_b[((7 - kk) * nbyte + ii) / 8];
*out_ui16 = bt;
}
}
count = bshuf_trans_bit_byte_remainder(in, out, size, elem_size,
nbyte - nbyte % 16);
return count;
}
void
mandel_sse2(unsigned char *image, const struct spec *s)
{
__m128 xmin = _mm_set_ps1(s->xlim[0]);
__m128 ymin = _mm_set_ps1(s->ylim[0]);
__m128 xscale = _mm_set_ps1((s->xlim[1] - s->xlim[0]) / s->width);
__m128 yscale = _mm_set_ps1((s->ylim[1] - s->ylim[0]) / s->height);
__m128 threshold = _mm_set_ps1(4);
__m128 one = _mm_set_ps1(1);
__m128i zero = _mm_setzero_si128();
__m128 iter_scale = _mm_set_ps1(1.0f / s->iterations);
__m128 depth_scale = _mm_set_ps1(s->depth - 1);
#pragma omp parallel for schedule(dynamic, 1)
for (int y = 0; y < s->height; y++) {
for (int x = 0; x < s->width; x += 4) {
__m128 mx = _mm_set_ps(x + 3, x + 2, x + 1, x + 0);
__m128 my = _mm_set_ps1(y);
__m128 cr = _mm_add_ps(_mm_mul_ps(mx, xscale), xmin);
__m128 ci = _mm_add_ps(_mm_mul_ps(my, yscale), ymin);
__m128 zr = cr;
__m128 zi = ci;
int k = 1;
__m128 mk = _mm_set_ps1(k);
while (++k < s->iterations) {
/* Compute z1 from z0 */
__m128 zr2 = _mm_mul_ps(zr, zr);
__m128 zi2 = _mm_mul_ps(zi, zi);
__m128 zrzi = _mm_mul_ps(zr, zi);
/* zr1 = zr0 * zr0 - zi0 * zi0 + cr */
/* zi1 = zr0 * zi0 + zr0 * zi0 + ci */
zr = _mm_add_ps(_mm_sub_ps(zr2, zi2), cr);
zi = _mm_add_ps(_mm_add_ps(zrzi, zrzi), ci);
/* Increment k */
zr2 = _mm_mul_ps(zr, zr);
zi2 = _mm_mul_ps(zi, zi);
__m128 mag2 = _mm_add_ps(zr2, zi2);
__m128 mask = _mm_cmplt_ps(mag2, threshold);
mk = _mm_add_ps(_mm_and_ps(mask, one), mk);
/* Early bailout? */
__m128i maski = _mm_castps_si128(mask);
if (0xFFFF == _mm_movemask_epi8(_mm_cmpeq_epi8(maski, zero)))
break;
}
mk = _mm_mul_ps(mk, iter_scale);
mk = _mm_sqrt_ps(mk);
mk = _mm_mul_ps(mk, depth_scale);
__m128i pixels = _mm_cvtps_epi32(mk);
unsigned char *dst = image + y * s->width * 3 + x * 3;
unsigned char *src = (unsigned char *)&pixels;
for (int i = 0; i < 4; i++) {
dst[i * 3 + 0] = src[i * 4];
dst[i * 3 + 1] = src[i * 4];
dst[i * 3 + 2] = src[i * 4];
}
}
}
}
size_t sse4_strstr_unrolled_max20(const char* s, size_t n, const char* needle, size_t needle_size) {
const __m128i zeros = _mm_setzero_si128();
const __m128i prefix = sse::load(needle);
const __m128i suffix = sse::load(needle + 4);
const __m128i suff_mask = sse::mask_lower_bytes(needle_size - 4);
for (size_t i = 0; i < n; i += 8) {
const __m128i data = sse::load(s + i);
const __m128i result = _mm_mpsadbw_epu8(data, prefix, 0);
const __m128i cmp = _mm_cmpeq_epi16(result, zeros);
unsigned mask = _mm_movemask_epi8(cmp) & 0x5555;
while (mask != 0) {
const auto bitpos = bits::get_first_bit_set(mask)/2;
const __m128i str = sse::load(s + i + bitpos + 4);
const __m128i cmp = _mm_cmpeq_epi8(str, suffix);
if (_mm_testc_si128(cmp, suff_mask)) {
return i + bitpos;
}
mask = bits::clear_leftmost_set(mask);
}
}
return std::string::npos;
}
void WriteBufferValidUTF8::nextImpl()
{
char * p = memory.data();
char * valid_start = p;
while (p < pos)
{
#ifdef __SSE2__
/// Fast skip of ASCII
static constexpr size_t SIMD_BYTES = 16;
const char * simd_end = p + (pos - p) / SIMD_BYTES * SIMD_BYTES;
while (p < simd_end && !_mm_movemask_epi8(_mm_loadu_si128(reinterpret_cast<const __m128i*>(p))))
p += SIMD_BYTES;
if (!(p < pos))
break;
#endif
size_t len = length_of_utf8_sequence[static_cast<unsigned char>(*p)];
if (len > 4)
{
/// Invalid start of sequence. Skip one byte.
putValid(valid_start, p - valid_start);
putReplacement();
++p;
valid_start = p;
}
else if (p + len > pos)
{
/// Sequence was not fully written to this buffer.
break;
}
else if (Poco::UTF8Encoding::isLegal(reinterpret_cast<unsigned char *>(p), len))
{
/// Valid sequence.
p += len;
}
else
{
/// Invalid sequence. Skip just first byte.
putValid(valid_start, p - valid_start);
putReplacement();
++p;
valid_start = p;
}
}
putValid(valid_start, p - valid_start);
size_t cnt = pos - p;
/// Shift unfinished sequence to start of buffer.
for (size_t i = 0; i < cnt; ++i)
memory[i] = p[i];
working_buffer = Buffer(&memory[cnt], memory.data() + memory.size());
}
int test_mm_movemask_epi8(__m128i A) {
// DAG-LABEL: test_mm_movemask_epi8
// DAG: call i32 @llvm.x86.sse2.pmovmskb.128(<16 x i8> %{{.*}})
//
// ASM-LABEL: test_mm_movemask_epi8
// ASM: pmovmskb
return _mm_movemask_epi8(A);
}
void WriteBufferValidUTF8::nextImpl()
{
char *p = &memory[0];
char *valid_start = p;
while (p < pos)
{
#ifdef __x86_64__
/// Быстрый пропуск ASCII
static constexpr size_t SIMD_BYTES = 16;
const char * simd_end = p + (pos - p) / SIMD_BYTES * SIMD_BYTES;
while (p < simd_end && !_mm_movemask_epi8(_mm_loadu_si128(reinterpret_cast<const __m128i*>(p))))
p += SIMD_BYTES;
if (!(p < pos))
break;
#endif
size_t len = 1 + static_cast<size_t>(trailingBytesForUTF8[static_cast<unsigned char>(*p)]);
if (len > 4)
{
/// Невалидное начало последовательности. Пропустим один байт.
putValid(valid_start, p - valid_start);
putReplacement();
++p;
valid_start = p;
}
else if (p + len > pos)
{
/// Еще не вся последовательность записана.
break;
}
else if (Poco::UTF8Encoding::isLegal(reinterpret_cast<unsigned char*>(p), len))
{
/// Валидная последовательность.
p += len;
}
else
{
/// Невалидная последовательность. Пропустим только первый байт.
putValid(valid_start, p - valid_start);
putReplacement();
++p;
valid_start = p;
}
}
putValid(valid_start, p - valid_start);
size_t cnt = pos - p;
/// Сдвинем незаконченную последовательность в начало буфера.
for (size_t i = 0; i < cnt; ++i)
memory[i] = p[i];
working_buffer = Buffer(&memory[cnt], &memory[0] + memory.size());
}
static INLINE void
build_masks(int c,
int cdiff,
int dcdx,
int dcdy,
unsigned *outmask,
unsigned *partmask)
{
__m128i cstep0 = _mm_setr_epi32(c, c+dcdx, c+dcdx*2, c+dcdx*3);
__m128i xdcdy = _mm_set1_epi32(dcdy);
/* Get values across the quad
*/
__m128i cstep1 = _mm_add_epi32(cstep0, xdcdy);
__m128i cstep2 = _mm_add_epi32(cstep1, xdcdy);
__m128i cstep3 = _mm_add_epi32(cstep2, xdcdy);
{
__m128i cstep01, cstep23, result;
cstep01 = _mm_packs_epi32(cstep0, cstep1);
cstep23 = _mm_packs_epi32(cstep2, cstep3);
result = _mm_packs_epi16(cstep01, cstep23);
*outmask |= _mm_movemask_epi8(result);
}
{
__m128i cio4 = _mm_set1_epi32(cdiff);
__m128i cstep01, cstep23, result;
cstep0 = _mm_add_epi32(cstep0, cio4);
cstep1 = _mm_add_epi32(cstep1, cio4);
cstep2 = _mm_add_epi32(cstep2, cio4);
cstep3 = _mm_add_epi32(cstep3, cio4);
cstep01 = _mm_packs_epi32(cstep0, cstep1);
cstep23 = _mm_packs_epi32(cstep2, cstep3);
result = _mm_packs_epi16(cstep01, cstep23);
*partmask |= _mm_movemask_epi8(result);
}
}
int64_t av1_highbd_block_error_sse2(tran_low_t *coeff, tran_low_t *dqcoeff,
intptr_t block_size, int64_t *ssz,
int bps) {
int i, j, test;
uint32_t temp[4];
__m128i max, min, cmp0, cmp1, cmp2, cmp3;
int64_t error = 0, sqcoeff = 0;
const int shift = 2 * (bps - 8);
const int rounding = shift > 0 ? 1 << (shift - 1) : 0;
for (i = 0; i < block_size; i += 8) {
// Load the data into xmm registers
__m128i mm_coeff = _mm_load_si128((__m128i *)(coeff + i));
__m128i mm_coeff2 = _mm_load_si128((__m128i *)(coeff + i + 4));
__m128i mm_dqcoeff = _mm_load_si128((__m128i *)(dqcoeff + i));
__m128i mm_dqcoeff2 = _mm_load_si128((__m128i *)(dqcoeff + i + 4));
// Check if any values require more than 15 bit
max = _mm_set1_epi32(0x3fff);
min = _mm_set1_epi32(0xffffc000);
cmp0 = _mm_xor_si128(_mm_cmpgt_epi32(mm_coeff, max),
_mm_cmplt_epi32(mm_coeff, min));
cmp1 = _mm_xor_si128(_mm_cmpgt_epi32(mm_coeff2, max),
_mm_cmplt_epi32(mm_coeff2, min));
cmp2 = _mm_xor_si128(_mm_cmpgt_epi32(mm_dqcoeff, max),
_mm_cmplt_epi32(mm_dqcoeff, min));
cmp3 = _mm_xor_si128(_mm_cmpgt_epi32(mm_dqcoeff2, max),
_mm_cmplt_epi32(mm_dqcoeff2, min));
test = _mm_movemask_epi8(
_mm_or_si128(_mm_or_si128(cmp0, cmp1), _mm_or_si128(cmp2, cmp3)));
if (!test) {
__m128i mm_diff, error_sse2, sqcoeff_sse2;
mm_coeff = _mm_packs_epi32(mm_coeff, mm_coeff2);
mm_dqcoeff = _mm_packs_epi32(mm_dqcoeff, mm_dqcoeff2);
mm_diff = _mm_sub_epi16(mm_coeff, mm_dqcoeff);
error_sse2 = _mm_madd_epi16(mm_diff, mm_diff);
sqcoeff_sse2 = _mm_madd_epi16(mm_coeff, mm_coeff);
_mm_storeu_si128((__m128i *)temp, error_sse2);
error = error + temp[0] + temp[1] + temp[2] + temp[3];
_mm_storeu_si128((__m128i *)temp, sqcoeff_sse2);
sqcoeff += temp[0] + temp[1] + temp[2] + temp[3];
} else {
for (j = 0; j < 8; j++) {
const int64_t diff = coeff[i + j] - dqcoeff[i + j];
error += diff * diff;
sqcoeff += (int64_t)coeff[i + j] * (int64_t)coeff[i + j];
}
}
}
assert(error >= 0 && sqcoeff >= 0);
error = (error + rounding) >> shift;
sqcoeff = (sqcoeff + rounding) >> shift;
*ssz = sqcoeff;
return error;
}
void* memrchr(void *dst, int c, size_t len) {
/* Backwards */
uint8_t* a = dst;
if(!len)
return NULL;
int i = len;
int aligned_a = 0;
aligned_a = ((uintptr_t)a & (sizeof(__m128i) - 1));
/* aligned */
if(aligned_a) {
while(i && ((uintptr_t) &a[i] & ( sizeof(__m128i)-1))) {
i--;
if(a[i] == (char)c) {
return a + i;
}
}
}
if(i >= 16) {
uint32_t buf_32 = c;
buf_32 |= (buf_32 << 8);
buf_32 |= (buf_32 << 16);
__m128i r1 = _mm_set_epi32(buf_32, buf_32, buf_32, buf_32);
while(i >= 16) {
i -= 16;
__m128i x = _mm_loadu_si128((__m128i*)&(a[i])); //16byte
__m128i cmp = _mm_cmpeq_epi8(x, r1);
uint16_t result = (uint16_t)_mm_movemask_epi8(cmp);
if(result != 0x0000U) {
i += 15;
while(!(result & 0x8000)) {
result = result << 1;
i--;
}
return a + i;
}
}
}
while(i) {
i--;
if(a[i] == (char)c) {
return a + i;
}
}
return NULL;
}
int SSEBinSearchBlock::search(uint32_t key) const {
const __m128i keys = _mm_set1_epi32(key);
__m128i v;
int limit = data.size() - 1;
int a = 0;
int b = limit;
while (a <= b) {
const int c = (a + b)/2;
if (data[c] == key) {
return c;
}
if (key < data[c]) {
b = c - 1;
if (b >= 4) {
v = _mm_loadu_si128(reinterpret_cast<const __m128i*>(&data[b - 4]));
v = _mm_cmpeq_epi32(v, keys);
const uint16_t mask = _mm_movemask_epi8(v);
if (mask) {
return b - 4 + __builtin_ctz(mask)/4;
}
}
} else {
a = c + 1;
if (a + 4 < limit) {
v = _mm_loadu_si128(reinterpret_cast<const __m128i*>(&data[a]));
v = _mm_cmpeq_epi32(v, keys);
const uint16_t mask = _mm_movemask_epi8(v);
if (mask) {
return a + __builtin_ctz(mask)/4;
}
}
}
}
return -1;
}
virtual size_t match(const char* data, size_t size)
{
__m128i firstLetter = _mm_loadu_si128(reinterpret_cast<const __m128i*>(this->firstLetter));
__m128i patternData = _mm_loadu_si128(reinterpret_cast<const __m128i*>(this->patternData));
__m128i patternMask = _mm_loadu_si128(reinterpret_cast<const __m128i*>(this->patternMask));
size_t offset = firstLetterPos;
while (offset + 32 <= size)
{
__m128i value = _mm_loadu_si128(reinterpret_cast<const __m128i*>(data + offset));
unsigned int mask = _mm_movemask_epi8(_mm_cmpeq_epi8(value, firstLetter));
// advance offset regardless of match results to reduce number of live values
offset += 16;
while (mask != 0)
{
unsigned int pos = re2::countTrailingZeros(mask);
size_t dataOffset = offset - 16 + pos - firstLetterOffset;
mask &= ~(1 << pos);
// check if we have a match
__m128i patternMatch = _mm_loadu_si128(reinterpret_cast<const __m128i*>(data + dataOffset));
__m128i matchMask = _mm_or_si128(patternMask, _mm_cmpeq_epi8(patternMatch, patternData));
if (_mm_movemask_epi8(matchMask) == 0xffff)
{
size_t matchOffset = dataOffset + firstLetterOffset - firstLetterPos;
// final check for full pattern
if (matchOffset + pattern.size() < size && memcmp(data + matchOffset, pattern.c_str(), pattern.size()) == 0)
{
return matchOffset;
}
}
}
}
return findMatch(pattern.c_str(), pattern.size(), data, size, offset - firstLetterPos);
}
inline void matrix16x8::transpose(square128& output, int x, int y)
{
for (int j = 0; j < 8; j++)
{
int row = _mm_movemask_epi8(whole);
whole = _mm_slli_epi64(whole, 1);
// _mm_movemask_epi8 uses most significant bit, hence +7-j
output.doublebytes[8*x+7-j][y] = row;
}
}
int valuesOK(cl_float8* to, cl_float8* from, size_t length) {
#ifdef DEBUG
printf("Checking data of size: %lu\n", length);
#endif
for(int i = 0; i < length; ++i) {
#ifdef __SSE__
__cl_float4 __hostFirstValue = to->v4[0];
__cl_float4 __hostSecondValue = to->v4[1];
__cl_float4 __deviceFirstValue = from->v4[0];
__cl_float4 __deviceSecondValue = from->v4[1];
__m128i vcmp = (__m128i) _mm_cmpneq_ps(__hostFirstValue, __deviceFirstValue);
uint16_t test = _mm_movemask_epi8(vcmp);
__m128i vcmp_2 = (__m128i) _mm_cmpneq_ps(__hostSecondValue, __deviceSecondValue);
uint16_t test_2 = _mm_movemask_epi8(vcmp_2);
if( (test|test_2) != 0 ) return 0; // indicative that the result failed
#else
#error "SSE not supported, which is required for example code to work!"
#endif
}
return 1;
}
static int DispatchAlpha(const uint8_t* alpha, int alpha_stride,
int width, int height,
uint8_t* dst, int dst_stride) {
// alpha_and stores an 'and' operation of all the alpha[] values. The final
// value is not 0xff if any of the alpha[] is not equal to 0xff.
uint32_t alpha_and = 0xff;
int i, j;
const __m128i zero = _mm_setzero_si128();
const __m128i rgb_mask = _mm_set1_epi32(0xffffff00u); // to preserve RGB
const __m128i all_0xff = _mm_set_epi32(0, 0, ~0u, ~0u);
__m128i all_alphas = all_0xff;
// We must be able to access 3 extra bytes after the last written byte
// 'dst[4 * width - 4]', because we don't know if alpha is the first or the
// last byte of the quadruplet.
const int limit = (width - 1) & ~7;
for (j = 0; j < height; ++j) {
__m128i* out = (__m128i*)dst;
for (i = 0; i < limit; i += 8) {
// load 8 alpha bytes
const __m128i a0 = _mm_loadl_epi64((const __m128i*)&alpha[i]);
const __m128i a1 = _mm_unpacklo_epi8(a0, zero);
const __m128i a2_lo = _mm_unpacklo_epi16(a1, zero);
const __m128i a2_hi = _mm_unpackhi_epi16(a1, zero);
// load 8 dst pixels (32 bytes)
const __m128i b0_lo = _mm_loadu_si128(out + 0);
const __m128i b0_hi = _mm_loadu_si128(out + 1);
// mask dst alpha values
const __m128i b1_lo = _mm_and_si128(b0_lo, rgb_mask);
const __m128i b1_hi = _mm_and_si128(b0_hi, rgb_mask);
// combine
const __m128i b2_lo = _mm_or_si128(b1_lo, a2_lo);
const __m128i b2_hi = _mm_or_si128(b1_hi, a2_hi);
// store
_mm_storeu_si128(out + 0, b2_lo);
_mm_storeu_si128(out + 1, b2_hi);
// accumulate eight alpha 'and' in parallel
all_alphas = _mm_and_si128(all_alphas, a0);
out += 2;
}
for (; i < width; ++i) {
const uint32_t alpha_value = alpha[i];
dst[4 * i] = alpha_value;
alpha_and &= alpha_value;
}
alpha += alpha_stride;
dst += dst_stride;
}
// Combine the eight alpha 'and' into a 8-bit mask.
alpha_and &= _mm_movemask_epi8(_mm_cmpeq_epi8(all_alphas, all_0xff));
return (alpha_and != 0xff);
}
static void SetResidualCoeffsSSE2(const int16_t* const coeffs,
VP8Residual* const res) {
const __m128i c0 = _mm_loadu_si128((const __m128i*)coeffs);
const __m128i c1 = _mm_loadu_si128((const __m128i*)(coeffs + 8));
// Use SSE to compare 8 values with a single instruction.
const __m128i zero = _mm_setzero_si128();
const __m128i m0 = _mm_cmpeq_epi16(c0, zero);
const __m128i m1 = _mm_cmpeq_epi16(c1, zero);
// Get the comparison results as a bitmask, consisting of two times 16 bits:
// two identical bits for each result. Concatenate both bitmasks to get a
// single 32 bit value. Negate the mask to get the position of entries that
// are not equal to zero. We don't need to mask out least significant bits
// according to res->first, since coeffs[0] is 0 if res->first > 0
const uint32_t mask =
~(((uint32_t)_mm_movemask_epi8(m1) << 16) | _mm_movemask_epi8(m0));
// The position of the most significant non-zero bit indicates the position of
// the last non-zero value. Divide the result by two because __movemask_epi8
// operates on 8 bit values instead of 16 bit values.
assert(res->first == 0 || coeffs[0] == 0);
res->last = mask ? (BitsLog2Floor(mask) >> 1) : -1;
res->coeffs = coeffs;
}
static float CombinedShannonEntropy(const int X[256], const int Y[256]) {
int i;
double retval = 0.;
int sumX, sumXY;
int32_t tmp[4];
__m128i zero = _mm_setzero_si128();
// Sums up X + Y, 4 ints at a time (and will merge it at the end for sumXY).
__m128i sumXY_128 = zero;
__m128i sumX_128 = zero;
for (i = 0; i < 256; i += 4) {
const __m128i x = _mm_loadu_si128((const __m128i*)(X + i));
const __m128i y = _mm_loadu_si128((const __m128i*)(Y + i));
// Check if any X is non-zero: this actually provides a speedup as X is
// usually sparse.
if (_mm_movemask_epi8(_mm_cmpeq_epi32(x, zero)) != 0xFFFF) {
const __m128i xy_128 = _mm_add_epi32(x, y);
sumXY_128 = _mm_add_epi32(sumXY_128, xy_128);
sumX_128 = _mm_add_epi32(sumX_128, x);
// Analyze the different X + Y.
_mm_storeu_si128((__m128i*)tmp, xy_128);
ANALYZE_XY(0);
ANALYZE_XY(1);
ANALYZE_XY(2);
ANALYZE_XY(3);
} else {
// X is fully 0, so only deal with Y.
sumXY_128 = _mm_add_epi32(sumXY_128, y);
ANALYZE_X_OR_Y(Y, 0);
ANALYZE_X_OR_Y(Y, 1);
ANALYZE_X_OR_Y(Y, 2);
ANALYZE_X_OR_Y(Y, 3);
}
}
// Sum up sumX_128 to get sumX.
_mm_storeu_si128((__m128i*)tmp, sumX_128);
sumX = tmp[3] + tmp[2] + tmp[1] + tmp[0];
// Sum up sumXY_128 to get sumXY.
_mm_storeu_si128((__m128i*)tmp, sumXY_128);
sumXY = tmp[3] + tmp[2] + tmp[1] + tmp[0];
retval += VP8LFastSLog2(sumX) + VP8LFastSLog2(sumXY);
return (float)retval;
}
virtual size_t match(const char* data, size_t size)
{
__m128i firstLetter = _mm_loadu_si128(reinterpret_cast<const __m128i*>(this->firstLetter));
__m128i patternData = _mm_loadu_si128(reinterpret_cast<const __m128i*>(this->patternData));
__m128i patternMask = _mm_loadu_si128(reinterpret_cast<const __m128i*>(this->patternMask));
size_t offset = firstLetterPos;
while (offset + 32 <= size)
{
__m128i value = _mm_loadu_si128(reinterpret_cast<const __m128i*>(data + offset));
int mask = _mm_movemask_epi8(_mm_cmpeq_epi8(value, firstLetter));
if (mask == 0)
offset += 16;
else
{
offset += re2::countTrailingZeros(mask);
// check if we have a match
__m128i patternMatch = _mm_loadu_si128(reinterpret_cast<const __m128i*>(data + offset - firstLetterOffset));
__m128i matchMask = _mm_or_si128(patternMask, _mm_cmpeq_epi8(patternMatch, patternData));
if (_mm_movemask_epi8(matchMask) == 0xffff)
{
// final check for full pattern
if (memcmp(data + offset - firstLetterPos, pattern.c_str(), pattern.size()) == 0)
{
return offset - firstLetterPos;
}
}
offset += 1;
}
}
return findMatch(pattern.c_str(), pattern.size(), data, size, offset - firstLetterPos);
}
size_t sse4_strstr_unrolled_max36(const char* s, size_t n, const char* needle, size_t needle_size) {
const __m128i zeros = _mm_setzero_si128();
const __m128i prefix = sse::load(needle);
const __m128i suffix1 = sse::load(needle + 4);
const __m128i suffix2 = sse::load(needle + 16 + 4);
const __m128i suff_mask = sse::mask_higher_bytes(needle_size - (16 + 4));
for (size_t i = 0; i < n; i += 8) {
const __m128i data = sse::load(s + i);
const __m128i result = _mm_mpsadbw_epu8(data, prefix, 0);
const __m128i cmp = _mm_cmpeq_epi16(result, zeros);
unsigned mask = _mm_movemask_epi8(cmp) & 0x5555;
while (mask != 0) {
const auto bitpos = bits::get_first_bit_set(mask)/2;
const __m128i c1 = _mm_cmpeq_epi8(sse::load(s + i + bitpos + 4), suffix1);
const __m128i c2 = _mm_cmpeq_epi8(sse::load(s + i + bitpos + 16 + 4), suffix2);
const __m128i c3 = _mm_or_si128(c2, suff_mask);
const __m128i tmp = _mm_and_si128(c1, c3);
if (_mm_movemask_epi8(tmp) == 0xffff) {
return i + bitpos;
}
mask = bits::clear_leftmost_set(mask);
}
}
return std::string::npos;
}
bool test(const index_t & kmer) const
{
__m128 __attribute__ ((aligned (16))) zero = _mm_setzero_si128();
const size_t BitsPerElement = sizeof(block_t) * 8;
const Hash hashfunction = Hash();
kmer_t hashvalue = kmer;
for (int hcount = this->h; hcount > 0; hcount--)
{
hashvalue = hashfunction(hashvalue);
// we expect the compiler to automatically turn this into a shift because it's a const power of two
size_t offset = (hashvalue % this->m) / BitsPerElement;
if (_mm_movemask_epi8(
_mm_cmpeq_epi32(
_mm_and_ps(bitarray[offset],masks[hashvalue & (BitsPerElement-1)]),zero)) != 0xFFFF) return false;
}
return true;
}
int countZeroBytes_SSE(char* values, int length) {
int zeroCount = 0;
__m128i zero16 = _mm_set1_epi8(0);
__m128i and16 = _mm_set1_epi8(1);
for(int i=0; i<length; i+=16) {
__m128i values16 = _mm_loadu_si128((__m128i*)&values[i]);
__m128i cmp = _mm_cmpeq_epi8(values16, zero16);
if(_mm_movemask_epi8(cmp)) {
cmp = _mm_and_si128(and16, cmp); //change -1 values to 1
//hortiontal sum of 16 bytes
__m128i sum1 = _mm_sad_epu8(cmp,zero16);
__m128i sum2 = _mm_shuffle_epi32(sum1,2);
__m128i sum3 = _mm_add_epi16(sum1,sum2);
zeroCount += _mm_cvtsi128_si32(sum3);
}
}
return zeroCount;
}
