static int HafCpu_Histogram3Thresholds_DATA_U8
(
vx_uint32 dstHist[],
vx_uint8 distThreshold0,
vx_uint8 distThreshold1,
vx_uint8 distThreshold2,
vx_uint32 srcWidth,
vx_uint32 srcHeight,
vx_uint8 * pSrcImage,
vx_uint32 srcImageStrideInBytes
)
{
// offset: to convert the range from 0..255 to -128..127, because SSE does not have compare instructions for unsigned bytes
// thresh: source threshold in -128..127 range
__m128i offset = _mm_set1_epi8((char)0x80);
__m128i T0 = _mm_set1_epi8((char)((distThreshold0 - 1) ^ 0x80));
__m128i T1 = _mm_set1_epi8((char)((distThreshold1 - 1) ^ 0x80));
__m128i T2 = _mm_set1_epi8((char)((distThreshold2 - 1) ^ 0x80));
__m128i onemask = _mm_set1_epi8((char)1);
// process one pixel row at a time that counts "pixel < srcThreshold"
__m128i count0 = _mm_set1_epi8((char)0);
__m128i count1 = _mm_set1_epi8((char)0);
__m128i count2 = _mm_set1_epi8((char)0);
vx_uint8 * srcRow = pSrcImage;
vx_uint32 width = (srcWidth + 15) >> 4;
for (unsigned int y = 0; y < srcHeight; y++) {
__m128i * src = (__m128i *)srcRow;
for (unsigned int x = 0; x < width; x++) {
__m128i pixels = _mm_load_si128(src++);
pixels = _mm_xor_si128(pixels, offset);
__m128i cmpout;
cmpout = _mm_cmpgt_epi8(pixels, T0);
cmpout = _mm_and_si128(cmpout, onemask);
cmpout = _mm_sad_epu8(cmpout, onemask);
count0 = _mm_add_epi32(count0, cmpout);
cmpout = _mm_cmpgt_epi8(pixels, T1);
cmpout = _mm_and_si128(cmpout, onemask);
cmpout = _mm_sad_epu8(cmpout, onemask);
count1 = _mm_add_epi32(count1, cmpout);
cmpout = _mm_cmpgt_epi8(pixels, T2);
cmpout = _mm_and_si128(cmpout, onemask);
cmpout = _mm_sad_epu8(cmpout, onemask);
count2 = _mm_add_epi32(count2, cmpout);
}
srcRow += srcImageStrideInBytes;
}
// extract histogram from count: special case needed when T1 == T2
dstHist[0] = M128I(count0).m128i_u32[0] + M128I(count0).m128i_u32[2];
dstHist[1] = M128I(count1).m128i_u32[0] + M128I(count1).m128i_u32[2] - dstHist[0];
dstHist[2] = M128I(count2).m128i_u32[0] + M128I(count2).m128i_u32[2] - dstHist[0] - dstHist[1];
dstHist[3] = srcWidth * srcHeight - dstHist[0] - dstHist[1] - dstHist[2];
if (M128I(T1).m128i_i8[0] == M128I(T2).m128i_i8[0]) {
dstHist[2] = dstHist[3];
dstHist[3] = 0;
}
return AGO_SUCCESS;
}
// Denoise a 16x1 vector.
static INLINE __m128i vp9_denoiser_16x1_sse2(
const uint8_t *sig, const uint8_t *mc_running_avg_y, uint8_t *running_avg_y,
const __m128i *k_0, const __m128i *k_4, const __m128i *k_8,
const __m128i *k_16, const __m128i *l3, const __m128i *l32,
const __m128i *l21, __m128i acc_diff) {
// Calculate differences
const __m128i v_sig = _mm_loadu_si128((const __m128i *)(&sig[0]));
const __m128i v_mc_running_avg_y =
_mm_loadu_si128((const __m128i *)(&mc_running_avg_y[0]));
__m128i v_running_avg_y;
const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig);
const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y);
// Obtain the sign. FF if diff is negative.
const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, *k_0);
// Clamp absolute difference to 16 to be used to get mask. Doing this
// allows us to use _mm_cmpgt_epi8, which operates on signed byte.
const __m128i clamped_absdiff =
_mm_min_epu8(_mm_or_si128(pdiff, ndiff), *k_16);
// Get masks for l2 l1 and l0 adjustments.
const __m128i mask2 = _mm_cmpgt_epi8(*k_16, clamped_absdiff);
const __m128i mask1 = _mm_cmpgt_epi8(*k_8, clamped_absdiff);
const __m128i mask0 = _mm_cmpgt_epi8(*k_4, clamped_absdiff);
// Get adjustments for l2, l1, and l0.
__m128i adj2 = _mm_and_si128(mask2, *l32);
const __m128i adj1 = _mm_and_si128(mask1, *l21);
const __m128i adj0 = _mm_and_si128(mask0, clamped_absdiff);
__m128i adj, padj, nadj;
// Combine the adjustments and get absolute adjustments.
adj2 = _mm_add_epi8(adj2, adj1);
adj = _mm_sub_epi8(*l3, adj2);
adj = _mm_andnot_si128(mask0, adj);
adj = _mm_or_si128(adj, adj0);
// Restore the sign and get positive and negative adjustments.
padj = _mm_andnot_si128(diff_sign, adj);
nadj = _mm_and_si128(diff_sign, adj);
// Calculate filtered value.
v_running_avg_y = _mm_adds_epu8(v_sig, padj);
v_running_avg_y = _mm_subs_epu8(v_running_avg_y, nadj);
_mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y);
// Adjustments <=7, and each element in acc_diff can fit in signed
// char.
acc_diff = _mm_adds_epi8(acc_diff, padj);
acc_diff = _mm_subs_epi8(acc_diff, nadj);
return acc_diff;
}
inline void casefoldRange(char* dest, const char* begin, const char* end)
{
if (end - begin < 64)
{
// short string, don't bother optimizing
for (const char* i = begin; i != end; ++i)
*dest++ = casefold(*i);
}
else
{
// Shift 'A'..'Z' range ([65..90]) to [102..127] to use one signed comparison insn
__m128i shiftAmount = _mm_set1_epi8(127 - 'Z');
__m128i lowerBound = _mm_set1_epi8(127 - ('Z' - 'A') - 1);
__m128i upperBit = _mm_set1_epi8(0x20);
const char* i = begin;
for (; i + 16 < end; i += 16)
{
__m128i v = _mm_loadu_si128(reinterpret_cast<const __m128i*>(i));
__m128i upperMask = _mm_cmpgt_epi8(_mm_add_epi8(v, shiftAmount), lowerBound);
__m128i cfv = _mm_or_si128(v, _mm_and_si128(upperMask, upperBit));
_mm_storeu_si128(reinterpret_cast<__m128i*>(dest), cfv);
dest += 16;
}
for (; i != end; ++i)
*dest++ = casefold(*i);
}
}
__m128i test_mm_cmpgt_epi8(__m128i A, __m128i B) {
// DAG-LABEL: test_mm_cmpgt_epi8
// DAG: icmp sgt <16 x i8>
//
// ASM-LABEL: test_mm_cmpgt_epi8
// ASM: pcmpgtb
return _mm_cmpgt_epi8(A, B);
}
// Shift each byte of "x" by 3 bits while preserving by the sign bit.
static WEBP_INLINE void SignedShift8b(__m128i* const x) {
const __m128i zero = _mm_setzero_si128();
const __m128i signs = _mm_cmpgt_epi8(zero, *x);
const __m128i lo_0 = _mm_unpacklo_epi8(*x, signs); // s8 -> s16 sign extend
const __m128i hi_0 = _mm_unpackhi_epi8(*x, signs);
const __m128i lo_1 = _mm_srai_epi16(lo_0, 3);
const __m128i hi_1 = _mm_srai_epi16(hi_0, 3);
*x = _mm_packs_epi16(lo_1, hi_1);
}
SIMDValue SIMDInt8x16Operation::OpGreaterThan(const SIMDValue& aValue, const SIMDValue& bValue)
{
X86SIMDValue x86Result;
X86SIMDValue tmpaValue = X86SIMDValue::ToX86SIMDValue(aValue);
X86SIMDValue tmpbValue = X86SIMDValue::ToX86SIMDValue(bValue);
x86Result.m128i_value = _mm_cmpgt_epi8(tmpaValue.m128i_value, tmpbValue.m128i_value); // compare a > b?
return X86SIMDValue::ToSIMDValue(x86Result);
}
mlib_status
__mlib_VectorConvert_U8_S8_Sat(
mlib_u8 *z,
const mlib_s8 *x,
mlib_s32 n)
{
if (n < 1)
return (MLIB_FAILURE);
mlib_s32 i, ax, az, nstep, n1, n2, n3, xval;
mlib_s8 *px = (mlib_s8 *)x;
mlib_u8 *pz = (mlib_u8 *)z;
__m128i zbuf, xbuf, zero, mask;
zero = _mm_setzero_si128();
ax = (mlib_addr)x & 15;
az = (mlib_addr)z & 15;
nstep = 16 / sizeof (mlib_u8);
n1 = ((16 - ax) & 15) / sizeof (mlib_u8);
n2 = (n - n1) / nstep;
n3 = n - n1 - n2 * nstep;
if (n2 < 1) {
for (i = 0; i < n; i++) {
xval = *px++;
if (xval < 0)
xval = 0;
*pz++ = xval;
}
} else {
for (i = 0; i < n1; i++) {
xval = *px++;
if (xval < 0)
xval = 0;
*pz++ = xval;
}
for (i = 0; i < n2; i++) {
xbuf = _mm_load_si128((__m128i *)px);
mask = _mm_cmpgt_epi8(zero, xbuf);
zbuf = _mm_andnot_si128(mask, xbuf);
_mm_storeu_si128((__m128i *)pz, zbuf);
px += nstep;
pz += nstep;
}
for (i = 0; i < n3; i++) {
xval = *px++;
if (xval < 0)
xval = 0;
*pz++ = xval;
}
}
return (MLIB_SUCCESS);
}
__m64 _m_pcmpgtb(__m64 _MM1, __m64 _MM2)
{
__m128i lhs = {0}, rhs = {0};
lhs.m128i_i64[0] = _MM1.m64_i64;
rhs.m128i_i64[0] = _MM2.m64_i64;
lhs = _mm_cmpgt_epi8(lhs, rhs);
_MM1.m64_i64 = lhs.m128i_i64[0];
return _MM1;
}
// The function assumes that the image pointers are 16 byte aligned, and the source and destination strides as well
// It processes the pixels in a width which is the next highest multiple of 16 after dstWidth
static int HafCpu_Histogram1Threshold_DATA_U8
(
vx_uint32 dstHist[],
vx_uint8 distThreshold,
vx_uint32 srcWidth,
vx_uint32 srcHeight,
vx_uint8 * pSrcImage,
vx_uint32 srcImageStrideInBytes
)
{
// offset: to convert the range from 0..255 to -128..127, because SSE does not have compare instructions for unsigned bytes
// thresh: source threshold in -128..127 range
__m128i offset = _mm_set1_epi8((char)0x80);
__m128i thresh = _mm_set1_epi8((char)((distThreshold - 1) ^ 0x80));
__m128i onemask = _mm_set1_epi8((char)1);
// process one pixel row at a time that counts "pixel < srcThreshold"
__m128i count = _mm_set1_epi8((char)0);
vx_uint8 * srcRow = pSrcImage;
vx_uint32 width = (srcWidth + 15) >> 4;
for (unsigned int y = 0; y < srcHeight; y++) {
__m128i * src = (__m128i *)srcRow;
for (unsigned int x = 0; x < width; x++) {
__m128i pixels = _mm_load_si128(src++);
pixels = _mm_xor_si128(pixels, offset);
pixels = _mm_cmpgt_epi8(pixels, thresh);
pixels = _mm_and_si128(pixels, onemask);
pixels = _mm_sad_epu8(pixels, onemask);
count = _mm_add_epi32(count, pixels);
}
srcRow += srcImageStrideInBytes;
}
// extract histogram from count
dstHist[0] = M128I(count).m128i_u32[0] + M128I(count).m128i_u32[2];
dstHist[1] = srcWidth * srcHeight - dstHist[0];
return AGO_SUCCESS;
}
static inline __m128i _mm_min_epi8_rpl(__m128i a, __m128i b) {
__m128i mask = _mm_cmpgt_epi8(b, a);
a = _mm_and_si128(a, mask);
b = _mm_andnot_si128(mask, b);
return _mm_or_si128(a, b);
}
ColumnPtr ColumnFixedString::filter(const IColumn::Filter & filt, ssize_t result_size_hint) const
{
size_t col_size = size();
if (col_size != filt.size())
throw Exception("Size of filter doesn't match size of column.", ErrorCodes::SIZES_OF_COLUMNS_DOESNT_MATCH);
auto res = ColumnFixedString::create(n);
if (result_size_hint)
res->chars.reserve(result_size_hint > 0 ? result_size_hint * n : chars.size());
const UInt8 * filt_pos = &filt[0];
const UInt8 * filt_end = filt_pos + col_size;
const UInt8 * data_pos = &chars[0];
#if __SSE2__
/** A slightly more optimized version.
* Based on the assumption that often pieces of consecutive values
* completely pass or do not pass the filter.
* Therefore, we will optimistically check the parts of `SIMD_BYTES` values.
*/
static constexpr size_t SIMD_BYTES = 16;
const __m128i zero16 = _mm_setzero_si128();
const UInt8 * filt_end_sse = filt_pos + col_size / SIMD_BYTES * SIMD_BYTES;
const size_t chars_per_simd_elements = SIMD_BYTES * n;
while (filt_pos < filt_end_sse)
{
int mask = _mm_movemask_epi8(_mm_cmpgt_epi8(_mm_loadu_si128(reinterpret_cast<const __m128i *>(filt_pos)), zero16));
if (0 == mask)
{
/// Nothing is inserted.
data_pos += chars_per_simd_elements;
}
else if (0xFFFF == mask)
{
res->chars.insert(data_pos, data_pos + chars_per_simd_elements);
data_pos += chars_per_simd_elements;
}
else
{
size_t res_chars_size = res->chars.size();
for (size_t i = 0; i < SIMD_BYTES; ++i)
{
if (filt_pos[i])
{
res->chars.resize(res_chars_size + n);
memcpySmallAllowReadWriteOverflow15(&res->chars[res_chars_size], data_pos, n);
res_chars_size += n;
}
data_pos += n;
}
}
filt_pos += SIMD_BYTES;
}
#endif
size_t res_chars_size = res->chars.size();
while (filt_pos < filt_end)
{
if (*filt_pos)
{
res->chars.resize(res_chars_size + n);
memcpySmallAllowReadWriteOverflow15(&res->chars[res_chars_size], data_pos, n);
res_chars_size += n;
}
++filt_pos;
data_pos += n;
}
return std::move(res);
}
mlib_status
__mlib_VectorSumAbsDiff_S8_Sat(
mlib_d64 *z,
const mlib_s8 *x,
const mlib_s8 *y,
mlib_s32 n)
{
if (n <= 0)
return (MLIB_FAILURE);
mlib_s32 i, nstep, ax, ay, n1, n2, n3, diff, sum = 0;
mlib_s8 *px = (mlib_s8 *)x, *py = (mlib_s8 *)y;
__m128i zero, xbuf, ybuf, zbuf, mext, mbuf;
zero = _mm_setzero_si128();
zbuf = zero;
nstep = 16 / sizeof (mlib_s8);
ax = (mlib_addr)x & 15;
ay = (mlib_addr)y & 15;
n1 = ((16 - ax) & 15) / sizeof (mlib_s8);
n2 = (n - n1) / nstep;
n3 = n - n1 - n2 * nstep;
if (n2 < 1) {
for (i = 0; i < n; i++) {
diff = (mlib_s32)(*px++) - (*py++);
sum += ABS_VALUE(diff);
}
*z = sum;
} else {
for (i = 0; i < n1; i++) {
diff = (mlib_s32)(*px++) - (*py++);
sum += ABS_VALUE(diff);
}
if (ax == ay) {
for (i = 0; i < n2; i++) {
xbuf = _mm_load_si128((__m128i *)px);
ybuf = _mm_load_si128((__m128i *)py);
mext = _mm_cmpgt_epi8(ybuf, xbuf);
mbuf = _mm_sub_epi8(xbuf, ybuf);
mbuf = _mm_xor_si128(mbuf, mext);
mbuf = _mm_sub_epi8(mbuf, mext);
mbuf = _mm_sad_epu8(mbuf, zero);
zbuf = _mm_add_epi64(zbuf, mbuf);
px += nstep;
py += nstep;
}
} else {
for (i = 0; i < n2; i++) {
xbuf = _mm_load_si128((__m128i *)px);
ybuf = _mm_loadu_si128((__m128i *)py);
mext = _mm_cmpgt_epi8(ybuf, xbuf);
mbuf = _mm_sub_epi8(xbuf, ybuf);
mbuf = _mm_xor_si128(mbuf, mext);
mbuf = _mm_sub_epi8(mbuf, mext);
mbuf = _mm_sad_epu8(mbuf, zero);
zbuf = _mm_add_epi64(zbuf, mbuf);
px += nstep;
py += nstep;
}
}
for (i = 0; i < n3; i++) {
diff = (mlib_s32)(*px++) - (*py++);
sum += ABS_VALUE(diff);
}
mlib_d64 dsum = sum;
long long pz[2];
_mm_storeu_si128((__m128i *)pz, zbuf);
dsum += pz[0];
dsum += pz[1];
*z = dsum;
}
return (MLIB_SUCCESS);
}
int vp8_denoiser_filter_sse2(unsigned char *mc_running_avg_y,
int mc_avg_y_stride, unsigned char *running_avg_y,
int avg_y_stride, unsigned char *sig,
int sig_stride, unsigned int motion_magnitude,
int increase_denoising) {
unsigned char *running_avg_y_start = running_avg_y;
unsigned char *sig_start = sig;
unsigned int sum_diff_thresh;
int r;
int shift_inc =
(increase_denoising && motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD)
? 1
: 0;
__m128i acc_diff = _mm_setzero_si128();
const __m128i k_0 = _mm_setzero_si128();
const __m128i k_4 = _mm_set1_epi8(4 + shift_inc);
const __m128i k_8 = _mm_set1_epi8(8);
const __m128i k_16 = _mm_set1_epi8(16);
/* Modify each level's adjustment according to motion_magnitude. */
const __m128i l3 = _mm_set1_epi8(
(motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 7 + shift_inc : 6);
/* Difference between level 3 and level 2 is 2. */
const __m128i l32 = _mm_set1_epi8(2);
/* Difference between level 2 and level 1 is 1. */
const __m128i l21 = _mm_set1_epi8(1);
for (r = 0; r < 16; ++r) {
/* Calculate differences */
const __m128i v_sig = _mm_loadu_si128((__m128i *)(&sig[0]));
const __m128i v_mc_running_avg_y =
_mm_loadu_si128((__m128i *)(&mc_running_avg_y[0]));
__m128i v_running_avg_y;
const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig);
const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y);
/* Obtain the sign. FF if diff is negative. */
const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
/* Clamp absolute difference to 16 to be used to get mask. Doing this
* allows us to use _mm_cmpgt_epi8, which operates on signed byte. */
const __m128i clamped_absdiff =
_mm_min_epu8(_mm_or_si128(pdiff, ndiff), k_16);
/* Get masks for l2 l1 and l0 adjustments */
const __m128i mask2 = _mm_cmpgt_epi8(k_16, clamped_absdiff);
const __m128i mask1 = _mm_cmpgt_epi8(k_8, clamped_absdiff);
const __m128i mask0 = _mm_cmpgt_epi8(k_4, clamped_absdiff);
/* Get adjustments for l2, l1, and l0 */
__m128i adj2 = _mm_and_si128(mask2, l32);
const __m128i adj1 = _mm_and_si128(mask1, l21);
const __m128i adj0 = _mm_and_si128(mask0, clamped_absdiff);
__m128i adj, padj, nadj;
/* Combine the adjustments and get absolute adjustments. */
adj2 = _mm_add_epi8(adj2, adj1);
adj = _mm_sub_epi8(l3, adj2);
adj = _mm_andnot_si128(mask0, adj);
adj = _mm_or_si128(adj, adj0);
/* Restore the sign and get positive and negative adjustments. */
padj = _mm_andnot_si128(diff_sign, adj);
nadj = _mm_and_si128(diff_sign, adj);
/* Calculate filtered value. */
v_running_avg_y = _mm_adds_epu8(v_sig, padj);
v_running_avg_y = _mm_subs_epu8(v_running_avg_y, nadj);
_mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y);
/* Adjustments <=7, and each element in acc_diff can fit in signed
* char.
*/
acc_diff = _mm_adds_epi8(acc_diff, padj);
acc_diff = _mm_subs_epi8(acc_diff, nadj);
/* Update pointers for next iteration. */
sig += sig_stride;
mc_running_avg_y += mc_avg_y_stride;
running_avg_y += avg_y_stride;
}
{
/* Compute the sum of all pixel differences of this MB. */
unsigned int abs_sum_diff = abs_sum_diff_16x1(acc_diff);
sum_diff_thresh = SUM_DIFF_THRESHOLD;
if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH;
if (abs_sum_diff > sum_diff_thresh) {
// Before returning to copy the block (i.e., apply no denoising),
// check if we can still apply some (weaker) temporal filtering to
// this block, that would otherwise not be denoised at all. Simplest
// is to apply an additional adjustment to running_avg_y to bring it
// closer to sig. The adjustment is capped by a maximum delta, and
// chosen such that in most cases the resulting sum_diff will be
// within the acceptable range given by sum_diff_thresh.
// The delta is set by the excess of absolute pixel diff over the
// threshold.
int delta = ((abs_sum_diff - sum_diff_thresh) >> 8) + 1;
// Only apply the adjustment for max delta up to 3.
if (delta < 4) {
const __m128i k_delta = _mm_set1_epi8(delta);
sig -= sig_stride * 16;
mc_running_avg_y -= mc_avg_y_stride * 16;
running_avg_y -= avg_y_stride * 16;
for (r = 0; r < 16; ++r) {
__m128i v_running_avg_y =
_mm_loadu_si128((__m128i *)(&running_avg_y[0]));
// Calculate differences.
const __m128i v_sig = _mm_loadu_si128((__m128i *)(&sig[0]));
const __m128i v_mc_running_avg_y =
_mm_loadu_si128((__m128i *)(&mc_running_avg_y[0]));
const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig);
const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y);
// Obtain the sign. FF if diff is negative.
const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
// Clamp absolute difference to delta to get the adjustment.
const __m128i adj = _mm_min_epu8(_mm_or_si128(pdiff, ndiff), k_delta);
// Restore the sign and get positive and negative adjustments.
__m128i padj, nadj;
padj = _mm_andnot_si128(diff_sign, adj);
nadj = _mm_and_si128(diff_sign, adj);
// Calculate filtered value.
v_running_avg_y = _mm_subs_epu8(v_running_avg_y, padj);
v_running_avg_y = _mm_adds_epu8(v_running_avg_y, nadj);
_mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y);
// Accumulate the adjustments.
acc_diff = _mm_subs_epi8(acc_diff, padj);
acc_diff = _mm_adds_epi8(acc_diff, nadj);
// Update pointers for next iteration.
sig += sig_stride;
mc_running_avg_y += mc_avg_y_stride;
running_avg_y += avg_y_stride;
}
abs_sum_diff = abs_sum_diff_16x1(acc_diff);
if (abs_sum_diff > sum_diff_thresh) {
return COPY_BLOCK;
}
} else {
return COPY_BLOCK;
}
}
}
// count genotype sum and number of calls, not requiring 16-aligned p
COREARRAY_DLL_DEFAULT C_UInt8* vec_u8_geno_count(C_UInt8 *p,
size_t n, C_Int32 &out_sum, C_Int32 &out_num)
{
C_Int32 sum=0, num=0;
#if defined(COREARRAY_SIMD_AVX2)
const __m256i three = _mm256_set1_epi8(3);
const __m256i zero = _mm256_setzero_si256();
__m256i sum32 = zero, num32 = zero;
size_t limit_by_U8 = 0;
for (; n >= 32; )
{
__m256i v = _mm256_loadu_si256((__m256i const*)p);
p += 32;
__m256i m = _mm256_cmpgt_epi8(three, _mm256_min_epu8(v, three));
sum32 = _mm256_add_epi8(sum32, _mm256_and_si256(v, m));
num32 = _mm256_sub_epi8(num32, m);
n -= 32;
limit_by_U8 ++;
if ((limit_by_U8 >= 127) || (n < 32))
{
// add to sum
sum32 = _mm256_sad_epu8(sum32, zero);
sum32 = _mm256_add_epi32(sum32,
_mm256_permute4x64_epi64(sum32, _MM_SHUFFLE(1,0,3,2)));
sum32 = _mm256_add_epi32(sum32,
_mm256_permute4x64_epi64(sum32, _MM_SHUFFLE(0,0,0,1)));
sum += _mm_cvtsi128_si32(_mm256_castsi256_si128(sum32));
// add to num
num32 = _mm256_sad_epu8(num32, zero);
num32 = _mm256_add_epi32(num32,
_mm256_permute4x64_epi64(num32, _MM_SHUFFLE(1,0,3,2)));
num32 = _mm256_add_epi32(num32,
_mm256_permute4x64_epi64(num32, _MM_SHUFFLE(0,0,0,1)));
num += _mm_cvtsi128_si32(_mm256_castsi256_si128(num32));
// reset
sum32 = num32 = zero;
limit_by_U8 = 0;
}
}
#elif defined(COREARRAY_SIMD_SSE2)
// header, 16-byte aligned
size_t h = (16 - ((size_t)p & 0x0F)) & 0x0F;
for (; (n > 0) && (h > 0); n--, h--, p++)
if (*p <= 2) { sum += *p; num++; }
const __m128i three = _mm_set1_epi8(3);
const __m128i zero = _mm_setzero_si128();
__m128i sum16=zero, num16=zero;
size_t limit_by_U8 = 0;
for (; n >= 16; )
{
__m128i v = _mm_load_si128((__m128i const*)p);
p += 16;
__m128i m = _mm_cmpgt_epi8(three, _mm_min_epu8(v, three));
sum16 = _mm_add_epi8(sum16, v & m);
num16 = _mm_sub_epi8(num16, m);
n -= 16;
limit_by_U8 ++;
if ((limit_by_U8 >= 127) || (n < 16))
{
// add to sum
sum16 = _mm_sad_epu8(sum16, zero);
sum += _mm_cvtsi128_si32(sum16);
sum += _mm_cvtsi128_si32(_mm_shuffle_epi32(sum16, 2));
// add to num
num16 = _mm_sad_epu8(num16, zero);
num += _mm_cvtsi128_si32(num16);
num += _mm_cvtsi128_si32(_mm_shuffle_epi32(num16, 2));
// reset
sum16 = num16 = zero;
limit_by_U8 = 0;
}
}
#endif
for (; n > 0; n--, p++)
if (*p <= 2) { sum += *p; num++; }
out_sum = sum;
out_num = num;
return p;
}
void FAST_t(InputArray _img, std::vector<KeyPoint>& keypoints, int threshold, bool nonmax_suppression)
{
Mat img = _img.getMat();
const int K = patternSize/2, N = patternSize + K + 1;
#if CV_SSE2
const int quarterPatternSize = patternSize/4;
(void)quarterPatternSize;
#endif
int i, j, k, pixel[25];
makeOffsets(pixel, (int)img.step, patternSize);
keypoints.clear();
threshold = std::min(std::max(threshold, 0), 255);
#if CV_SSE2
__m128i delta = _mm_set1_epi8(-128), t = _mm_set1_epi8((char)threshold), K16 = _mm_set1_epi8((char)K);
(void)K16;
(void)delta;
(void)t;
#endif
uchar threshold_tab[512];
for( i = -255; i <= 255; i++ )
threshold_tab[i+255] = (uchar)(i < -threshold ? 1 : i > threshold ? 2 : 0);
AutoBuffer<uchar> _buf((img.cols+16)*3*(sizeof(int) + sizeof(uchar)) + 128);
uchar* buf[3];
buf[0] = _buf; buf[1] = buf[0] + img.cols; buf[2] = buf[1] + img.cols;
int* cpbuf[3];
cpbuf[0] = (int*)alignPtr(buf[2] + img.cols, sizeof(int)) + 1;
cpbuf[1] = cpbuf[0] + img.cols + 1;
cpbuf[2] = cpbuf[1] + img.cols + 1;
memset(buf[0], 0, img.cols*3);
for(i = 3; i < img.rows-2; i++)
{
const uchar* ptr = img.ptr<uchar>(i) + 3;
uchar* curr = buf[(i - 3)%3];
int* cornerpos = cpbuf[(i - 3)%3];
memset(curr, 0, img.cols);
int ncorners = 0;
if( i < img.rows - 3 )
{
j = 3;
#if CV_SSE2
if( patternSize == 16 )
{
for(; j < img.cols - 16 - 3; j += 16, ptr += 16)
{
__m128i m0, m1;
__m128i v0 = _mm_loadu_si128((const __m128i*)ptr);
__m128i v1 = _mm_xor_si128(_mm_subs_epu8(v0, t), delta);
v0 = _mm_xor_si128(_mm_adds_epu8(v0, t), delta);
__m128i x0 = _mm_sub_epi8(_mm_loadu_si128((const __m128i*)(ptr + pixel[0])), delta);
__m128i x1 = _mm_sub_epi8(_mm_loadu_si128((const __m128i*)(ptr + pixel[quarterPatternSize])), delta);
__m128i x2 = _mm_sub_epi8(_mm_loadu_si128((const __m128i*)(ptr + pixel[2*quarterPatternSize])), delta);
__m128i x3 = _mm_sub_epi8(_mm_loadu_si128((const __m128i*)(ptr + pixel[3*quarterPatternSize])), delta);
m0 = _mm_and_si128(_mm_cmpgt_epi8(x0, v0), _mm_cmpgt_epi8(x1, v0));
m1 = _mm_and_si128(_mm_cmpgt_epi8(v1, x0), _mm_cmpgt_epi8(v1, x1));
m0 = _mm_or_si128(m0, _mm_and_si128(_mm_cmpgt_epi8(x1, v0), _mm_cmpgt_epi8(x2, v0)));
m1 = _mm_or_si128(m1, _mm_and_si128(_mm_cmpgt_epi8(v1, x1), _mm_cmpgt_epi8(v1, x2)));
m0 = _mm_or_si128(m0, _mm_and_si128(_mm_cmpgt_epi8(x2, v0), _mm_cmpgt_epi8(x3, v0)));
m1 = _mm_or_si128(m1, _mm_and_si128(_mm_cmpgt_epi8(v1, x2), _mm_cmpgt_epi8(v1, x3)));
m0 = _mm_or_si128(m0, _mm_and_si128(_mm_cmpgt_epi8(x3, v0), _mm_cmpgt_epi8(x0, v0)));
m1 = _mm_or_si128(m1, _mm_and_si128(_mm_cmpgt_epi8(v1, x3), _mm_cmpgt_epi8(v1, x0)));
m0 = _mm_or_si128(m0, m1);
int mask = _mm_movemask_epi8(m0);
if( mask == 0 )
continue;
if( (mask & 255) == 0 )
{
j -= 8;
ptr -= 8;
continue;
}
__m128i c0 = _mm_setzero_si128(), c1 = c0, max0 = c0, max1 = c0;
for( k = 0; k < N; k++ )
{
__m128i x = _mm_xor_si128(_mm_loadu_si128((const __m128i*)(ptr + pixel[k])), delta);
m0 = _mm_cmpgt_epi8(x, v0);
m1 = _mm_cmpgt_epi8(v1, x);
c0 = _mm_and_si128(_mm_sub_epi8(c0, m0), m0);
c1 = _mm_and_si128(_mm_sub_epi8(c1, m1), m1);
max0 = _mm_max_epu8(max0, c0);
max1 = _mm_max_epu8(max1, c1);
}
max0 = _mm_max_epu8(max0, max1);
int m = _mm_movemask_epi8(_mm_cmpgt_epi8(max0, K16));
for( k = 0; m > 0 && k < 16; k++, m >>= 1 )
if(m & 1)
{
cornerpos[ncorners++] = j+k;
if(nonmax_suppression)
curr[j+k] = (uchar)cornerScore<patternSize>(ptr+k, pixel, threshold);
}
}
}
#endif
for( ; j < img.cols - 3; j++, ptr++ )
{
int v = ptr[0];
const uchar* tab = &threshold_tab[0] - v + 255;
int d = tab[ptr[pixel[0]]] | tab[ptr[pixel[8]]];
if( d == 0 )
continue;
d &= tab[ptr[pixel[2]]] | tab[ptr[pixel[10]]];
d &= tab[ptr[pixel[4]]] | tab[ptr[pixel[12]]];
d &= tab[ptr[pixel[6]]] | tab[ptr[pixel[14]]];
if( d == 0 )
continue;
d &= tab[ptr[pixel[1]]] | tab[ptr[pixel[9]]];
d &= tab[ptr[pixel[3]]] | tab[ptr[pixel[11]]];
d &= tab[ptr[pixel[5]]] | tab[ptr[pixel[13]]];
d &= tab[ptr[pixel[7]]] | tab[ptr[pixel[15]]];
if( d & 1 )
{
int vt = v - threshold, count = 0;
for( k = 0; k < N; k++ )
{
int x = ptr[pixel[k]];
if(x < vt)
{
if( ++count > K )
{
cornerpos[ncorners++] = j;
if(nonmax_suppression)
curr[j] = (uchar)cornerScore<patternSize>(ptr, pixel, threshold);
break;
}
}
else
count = 0;
}
}
if( d & 2 )
{
int vt = v + threshold, count = 0;
for( k = 0; k < N; k++ )
{
int x = ptr[pixel[k]];
if(x > vt)
{
if( ++count > K )
{
cornerpos[ncorners++] = j;
if(nonmax_suppression)
curr[j] = (uchar)cornerScore<patternSize>(ptr, pixel, threshold);
break;
}
}
else
count = 0;
}
}
}
}
cornerpos[-1] = ncorners;
if( i == 3 )
continue;
const uchar* prev = buf[(i - 4 + 3)%3];
const uchar* pprev = buf[(i - 5 + 3)%3];
cornerpos = cpbuf[(i - 4 + 3)%3];
ncorners = cornerpos[-1];
for( k = 0; k < ncorners; k++ )
{
j = cornerpos[k];
int score = prev[j];
if( !nonmax_suppression ||
(score > prev[j+1] && score > prev[j-1] &&
score > pprev[j-1] && score > pprev[j] && score > pprev[j+1] &&
score > curr[j-1] && score > curr[j] && score > curr[j+1]) )
{
keypoints.push_back(KeyPoint((float)j, (float)(i-1), 7.f, -1, (float)score));
}
}
}
ColumnPtr ColumnVector<T>::filter(const IColumn::Filter & filt, ssize_t result_size_hint) const
{
size_t size = data.size();
if (size != filt.size())
throw Exception("Size of filter doesn't match size of column.", ErrorCodes::SIZES_OF_COLUMNS_DOESNT_MATCH);
auto res = this->create();
Container & res_data = res->getData();
if (result_size_hint)
res_data.reserve(result_size_hint > 0 ? result_size_hint : size);
const UInt8 * filt_pos = &filt[0];
const UInt8 * filt_end = filt_pos + size;
const T * data_pos = &data[0];
#if __SSE2__
/** A slightly more optimized version.
* Based on the assumption that often pieces of consecutive values
* completely pass or do not pass the filter.
* Therefore, we will optimistically check the parts of `SIMD_BYTES` values.
*/
static constexpr size_t SIMD_BYTES = 16;
const __m128i zero16 = _mm_setzero_si128();
const UInt8 * filt_end_sse = filt_pos + size / SIMD_BYTES * SIMD_BYTES;
while (filt_pos < filt_end_sse)
{
int mask = _mm_movemask_epi8(_mm_cmpgt_epi8(_mm_loadu_si128(reinterpret_cast<const __m128i *>(filt_pos)), zero16));
if (0 == mask)
{
/// Nothing is inserted.
}
else if (0xFFFF == mask)
{
res_data.insert(data_pos, data_pos + SIMD_BYTES);
}
else
{
for (size_t i = 0; i < SIMD_BYTES; ++i)
if (filt_pos[i])
res_data.push_back(data_pos[i]);
}
filt_pos += SIMD_BYTES;
data_pos += SIMD_BYTES;
}
#endif
while (filt_pos < filt_end)
{
if (*filt_pos)
res_data.push_back(*data_pos);
++filt_pos;
++data_pos;
}
return std::move(res);
}
/// Element-wise comparison for greater than.
inline xmm_i8 operator > (const xmm_i8 &a, const xmm_i8 &b) { return _mm_cmpgt_epi8(a, b); }
test (__m128i s1, __m128i s2)
{
return _mm_cmpgt_epi8 (s1, s2);
}
__m128i test_mm_cmpgt_epi8(__m128i A, __m128i B) {
// CHECK-LABEL: test_mm_cmpgt_epi8
// CHECK: icmp sgt <16 x i8>
return _mm_cmpgt_epi8(A, B);
}
int vp8_denoiser_filter_sse2(YV12_BUFFER_CONFIG *mc_running_avg,
YV12_BUFFER_CONFIG *running_avg,
MACROBLOCK *signal, unsigned int motion_magnitude,
int y_offset, int uv_offset)
{
unsigned char *sig = signal->thismb;
int sig_stride = 16;
unsigned char *mc_running_avg_y = mc_running_avg->y_buffer + y_offset;
int mc_avg_y_stride = mc_running_avg->y_stride;
unsigned char *running_avg_y = running_avg->y_buffer + y_offset;
int avg_y_stride = running_avg->y_stride;
int r;
(void)uv_offset;
__m128i acc_diff = _mm_setzero_si128();
const __m128i k_0 = _mm_setzero_si128();
const __m128i k_4 = _mm_set1_epi8(4);
const __m128i k_8 = _mm_set1_epi8(8);
const __m128i k_16 = _mm_set1_epi8(16);
/* Modify each level's adjustment according to motion_magnitude. */
const __m128i l3 = _mm_set1_epi8(
(motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 7 : 6);
/* Difference between level 3 and level 2 is 2. */
const __m128i l32 = _mm_set1_epi8(2);
/* Difference between level 2 and level 1 is 1. */
const __m128i l21 = _mm_set1_epi8(1);
for (r = 0; r < 16; ++r)
{
/* Calculate differences */
const __m128i v_sig = _mm_loadu_si128((__m128i *)(&sig[0]));
const __m128i v_mc_running_avg_y = _mm_loadu_si128(
(__m128i *)(&mc_running_avg_y[0]));
__m128i v_running_avg_y;
const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig);
const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y);
/* Obtain the sign. FF if diff is negative. */
const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
/* Clamp absolute difference to 16 to be used to get mask. Doing this
* allows us to use _mm_cmpgt_epi8, which operates on signed byte. */
const __m128i clamped_absdiff = _mm_min_epu8(
_mm_or_si128(pdiff, ndiff), k_16);
/* Get masks for l2 l1 and l0 adjustments */
const __m128i mask2 = _mm_cmpgt_epi8(k_16, clamped_absdiff);
const __m128i mask1 = _mm_cmpgt_epi8(k_8, clamped_absdiff);
const __m128i mask0 = _mm_cmpgt_epi8(k_4, clamped_absdiff);
/* Get adjustments for l2, l1, and l0 */
__m128i adj2 = _mm_and_si128(mask2, l32);
const __m128i adj1 = _mm_and_si128(mask1, l21);
const __m128i adj0 = _mm_and_si128(mask0, clamped_absdiff);
__m128i adj, padj, nadj;
/* Combine the adjustments and get absolute adjustments. */
adj2 = _mm_add_epi8(adj2, adj1);
adj = _mm_sub_epi8(l3, adj2);
adj = _mm_andnot_si128(mask0, adj);
adj = _mm_or_si128(adj, adj0);
/* Restore the sign and get positive and negative adjustments. */
padj = _mm_andnot_si128(diff_sign, adj);
nadj = _mm_and_si128(diff_sign, adj);
/* Calculate filtered value. */
v_running_avg_y = _mm_adds_epu8(v_sig, padj);
v_running_avg_y = _mm_subs_epu8(v_running_avg_y, nadj);
_mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y);
/* Adjustments <=7, and each element in acc_diff can fit in signed
* char.
*/
acc_diff = _mm_adds_epi8(acc_diff, padj);
acc_diff = _mm_subs_epi8(acc_diff, nadj);
/* Update pointers for next iteration. */
sig += sig_stride;
mc_running_avg_y += mc_avg_y_stride;
running_avg_y += avg_y_stride;
}
{
/* Compute the sum of all pixel differences of this MB. */
union sum_union s;
int sum_diff = 0;
s.v = acc_diff;
sum_diff = s.e[0] + s.e[1] + s.e[2] + s.e[3] + s.e[4] + s.e[5]
+ s.e[6] + s.e[7] + s.e[8] + s.e[9] + s.e[10] + s.e[11]
+ s.e[12] + s.e[13] + s.e[14] + s.e[15];
if (abs(sum_diff) > SUM_DIFF_THRESHOLD)
{
return COPY_BLOCK;
}
}
vp8_copy_mem16x16(running_avg->y_buffer + y_offset, avg_y_stride,
signal->thismb, sig_stride);
return FILTER_BLOCK;
}
static int HafCpu_Histogram16Bins_DATA_U8
(
vx_uint32 * dstHist,
vx_uint8 distOffset,
vx_uint8 distWindow,
vx_uint32 srcWidth,
vx_uint32 srcHeight,
vx_uint8 * pSrcImage,
vx_uint32 srcImageStrideInBytes
)
{
// offset: to convert the range from 0..255 to -128..127, because SSE does not have compare instructions for unsigned bytes
// thresh: source threshold in -128..127 range
__m128i offset = _mm_set1_epi8((char)0x80);
__m128i T0 = _mm_set1_epi8((char)(((distOffset ? distOffset : distWindow) - 1) ^ 0x80));
__m128i dT = _mm_set1_epi8((char)distWindow);
__m128i onemask = _mm_set1_epi8((char)1);
// process one pixel row at a time that counts "pixel < srcThreshold"
vx_uint32 count[16] = { 0 };
vx_uint8 * srcRow = pSrcImage;
vx_uint32 width = (srcWidth + 15) >> 4;
for (unsigned int y = 0; y < srcHeight; y++) {
__m128i * src = (__m128i *)srcRow;
__m128i count0 = _mm_set1_epi8((char)0);
__m128i count1 = _mm_set1_epi8((char)0);
__m128i count2 = _mm_set1_epi8((char)0);
__m128i count3 = _mm_set1_epi8((char)0);
for (unsigned int x = 0; x < width; x++) {
__m128i pixels = _mm_load_si128(src++);
pixels = _mm_xor_si128(pixels, offset);
__m128i cmpout, Tnext = T0;
// 0..3
cmpout = _mm_cmpgt_epi8(pixels, Tnext);
cmpout = _mm_and_si128(cmpout, onemask);
cmpout = _mm_sad_epu8(cmpout, onemask);
count0 = _mm_add_epi32(count0, cmpout);
Tnext = _mm_add_epi8(Tnext, dT);
cmpout = _mm_cmpgt_epi8(pixels, Tnext);
cmpout = _mm_and_si128(cmpout, onemask);
cmpout = _mm_sad_epu8(cmpout, onemask);
cmpout = _mm_slli_epi64(cmpout, 16);
count0 = _mm_add_epi32(count0, cmpout);
Tnext = _mm_add_epi8(Tnext, dT);
cmpout = _mm_cmpgt_epi8(pixels, Tnext);
cmpout = _mm_and_si128(cmpout, onemask);
cmpout = _mm_sad_epu8(cmpout, onemask);
cmpout = _mm_slli_epi64(cmpout, 32);
count0 = _mm_add_epi32(count0, cmpout);
Tnext = _mm_add_epi8(Tnext, dT);
cmpout = _mm_cmpgt_epi8(pixels, Tnext);
cmpout = _mm_and_si128(cmpout, onemask);
cmpout = _mm_sad_epu8(cmpout, onemask);
cmpout = _mm_slli_epi64(cmpout, 48);
count0 = _mm_add_epi32(count0, cmpout);
// 4..7
Tnext = _mm_add_epi8(Tnext, dT);
cmpout = _mm_cmpgt_epi8(pixels, Tnext);
cmpout = _mm_and_si128(cmpout, onemask);
cmpout = _mm_sad_epu8(cmpout, onemask);
count1 = _mm_add_epi32(count1, cmpout);
Tnext = _mm_add_epi8(Tnext, dT);
cmpout = _mm_cmpgt_epi8(pixels, Tnext);
cmpout = _mm_and_si128(cmpout, onemask);
cmpout = _mm_sad_epu8(cmpout, onemask);
cmpout = _mm_slli_epi64(cmpout, 16);
count1 = _mm_add_epi32(count1, cmpout);
Tnext = _mm_add_epi8(Tnext, dT);
cmpout = _mm_cmpgt_epi8(pixels, Tnext);
cmpout = _mm_and_si128(cmpout, onemask);
cmpout = _mm_sad_epu8(cmpout, onemask);
cmpout = _mm_slli_epi64(cmpout, 32);
count1 = _mm_add_epi32(count1, cmpout);
Tnext = _mm_add_epi8(Tnext, dT);
cmpout = _mm_cmpgt_epi8(pixels, Tnext);
cmpout = _mm_and_si128(cmpout, onemask);
cmpout = _mm_sad_epu8(cmpout, onemask);
cmpout = _mm_slli_epi64(cmpout, 48);
count1 = _mm_add_epi32(count1, cmpout);
// 8..11
Tnext = _mm_add_epi8(Tnext, dT);
cmpout = _mm_cmpgt_epi8(pixels, Tnext);
cmpout = _mm_and_si128(cmpout, onemask);
cmpout = _mm_sad_epu8(cmpout, onemask);
count2 = _mm_add_epi32(count2, cmpout);
Tnext = _mm_add_epi8(Tnext, dT);
cmpout = _mm_cmpgt_epi8(pixels, Tnext);
cmpout = _mm_and_si128(cmpout, onemask);
cmpout = _mm_sad_epu8(cmpout, onemask);
cmpout = _mm_slli_epi64(cmpout, 16);
count2 = _mm_add_epi32(count2, cmpout);
Tnext = _mm_add_epi8(Tnext, dT);
cmpout = _mm_cmpgt_epi8(pixels, Tnext);
cmpout = _mm_and_si128(cmpout, onemask);
cmpout = _mm_sad_epu8(cmpout, onemask);
cmpout = _mm_slli_epi64(cmpout, 32);
count2 = _mm_add_epi32(count2, cmpout);
Tnext = _mm_add_epi8(Tnext, dT);
cmpout = _mm_cmpgt_epi8(pixels, Tnext);
cmpout = _mm_and_si128(cmpout, onemask);
cmpout = _mm_sad_epu8(cmpout, onemask);
cmpout = _mm_slli_epi64(cmpout, 48);
count2 = _mm_add_epi32(count2, cmpout);
// 12..15
Tnext = _mm_add_epi8(Tnext, dT);
cmpout = _mm_cmpgt_epi8(pixels, Tnext);
cmpout = _mm_and_si128(cmpout, onemask);
cmpout = _mm_sad_epu8(cmpout, onemask);
count3 = _mm_add_epi32(count3, cmpout);
Tnext = _mm_add_epi8(Tnext, dT);
cmpout = _mm_cmpgt_epi8(pixels, Tnext);
cmpout = _mm_and_si128(cmpout, onemask);
cmpout = _mm_sad_epu8(cmpout, onemask);
cmpout = _mm_slli_epi64(cmpout, 16);
count3 = _mm_add_epi32(count3, cmpout);
Tnext = _mm_add_epi8(Tnext, dT);
cmpout = _mm_cmpgt_epi8(pixels, Tnext);
cmpout = _mm_and_si128(cmpout, onemask);
cmpout = _mm_sad_epu8(cmpout, onemask);
cmpout = _mm_slli_epi64(cmpout, 32);
count3 = _mm_add_epi32(count3, cmpout);
Tnext = _mm_add_epi8(Tnext, dT);
cmpout = _mm_cmpgt_epi8(pixels, Tnext);
cmpout = _mm_and_si128(cmpout, onemask);
cmpout = _mm_sad_epu8(cmpout, onemask);
cmpout = _mm_slli_epi64(cmpout, 48);
count3 = _mm_add_epi32(count3, cmpout);
}
srcRow += srcImageStrideInBytes;
// move counts from count0..2 into count[]
for (int i = 0; i < 4; i++) {
count[ 0 + i] += M128I(count0).m128i_u16[i] + M128I(count0).m128i_u16[4 + i];
count[ 4 + i] += M128I(count1).m128i_u16[i] + M128I(count1).m128i_u16[4 + i];
count[ 8 + i] += M128I(count2).m128i_u16[i] + M128I(count2).m128i_u16[4 + i];
count[12 + i] += M128I(count3).m128i_u16[i] + M128I(count3).m128i_u16[4 + i];
}
}
// extract histogram from count
if (distOffset == 0) {
vx_uint32 last = (distWindow >= 16) ? srcWidth * srcHeight : count[15];
for (int i = 14; i >= 0; i--) {
count[i] = last - count[i];
last -= count[i];
}
dstHist[0] = last;
for (int i = 1; i < 16; i++)
dstHist[i] = count[i - 1];
}
else {
vx_uint32 last = srcWidth * srcHeight;
for (int i = 15; i >= 0; i--) {
count[i] = last - count[i];
last -= count[i];
dstHist[i] = count[i];
}
}
return AGO_SUCCESS;
}
}bool validate_utf8_sse(const char *src, size_t len) {
const char *end = src + len;
while (src + 16 < end) {
__m128i chunk = _mm_loadu_si128((const __m128i *)(src));
int asciiMask = _mm_movemask_epi8(chunk);
if (!asciiMask) {
src += 16;
continue;
}
__m128i chunk_signed = _mm_add_epi8(chunk, _mm_set1_epi8(0x80));
__m128i cond2 =
_mm_cmplt_epi8(_mm_set1_epi8(0xc2 - 1 - 0x80), chunk_signed);
__m128i state = _mm_set1_epi8((char)(0x0 | 0x80));
state = _mm_blendv_epi8(state, _mm_set1_epi8((char)(0x2 | 0xc0)), cond2);
__m128i cond3 =
_mm_cmplt_epi8(_mm_set1_epi8(0xe0 - 1 - 0x80), chunk_signed);
state = _mm_blendv_epi8(state, _mm_set1_epi8((char)(0x3 | 0xe0)), cond3);
__m128i mask3 = _mm_slli_si128(cond3, 1);
__m128i cond4 =
_mm_cmplt_epi8(_mm_set1_epi8(0xf0 - 1 - 0x80), chunk_signed);
// Fall back to the scalar processing
if (_mm_movemask_epi8(cond4)) {
break;
}
__m128i count = _mm_and_si128(state, _mm_set1_epi8(0x7));
__m128i count_sub1 = _mm_subs_epu8(count, _mm_set1_epi8(0x1));
__m128i counts = _mm_add_epi8(count, _mm_slli_si128(count_sub1, 1));
__m128i shifts = count_sub1;
shifts = _mm_add_epi8(shifts, _mm_slli_si128(shifts, 1));
counts = _mm_add_epi8(
counts, _mm_slli_si128(_mm_subs_epu8(counts, _mm_set1_epi8(0x2)), 2));
shifts = _mm_add_epi8(shifts, _mm_slli_si128(shifts, 2));
if (asciiMask ^ _mm_movemask_epi8(_mm_cmpgt_epi8(counts, _mm_set1_epi8(0))))
return false; // error
shifts = _mm_add_epi8(shifts, _mm_slli_si128(shifts, 4));
if (_mm_movemask_epi8(_mm_cmpgt_epi8(
_mm_sub_epi8(_mm_slli_si128(counts, 1), counts), _mm_set1_epi8(1))))
return false; // error
shifts = _mm_add_epi8(shifts, _mm_slli_si128(shifts, 8));
__m128i mask = _mm_and_si128(state, _mm_set1_epi8(0xf8));
shifts =
_mm_and_si128(shifts, _mm_cmplt_epi8(counts, _mm_set1_epi8(2))); // <=1
chunk =
_mm_andnot_si128(mask, chunk); // from now on, we only have usefull bits
shifts = _mm_blendv_epi8(shifts, _mm_srli_si128(shifts, 1),
_mm_srli_si128(_mm_slli_epi16(shifts, 7), 1));
__m128i chunk_right = _mm_slli_si128(chunk, 1);
__m128i chunk_low = _mm_blendv_epi8(
chunk,
_mm_or_si128(chunk, _mm_and_si128(_mm_slli_epi16(chunk_right, 6),
_mm_set1_epi8(0xc0))),
_mm_cmpeq_epi8(counts, _mm_set1_epi8(1)));
__m128i chunk_high =
_mm_and_si128(chunk, _mm_cmpeq_epi8(counts, _mm_set1_epi8(2)));
shifts = _mm_blendv_epi8(shifts, _mm_srli_si128(shifts, 2),
_mm_srli_si128(_mm_slli_epi16(shifts, 6), 2));
chunk_high = _mm_srli_epi32(chunk_high, 2);
shifts = _mm_blendv_epi8(shifts, _mm_srli_si128(shifts, 4),
_mm_srli_si128(_mm_slli_epi16(shifts, 5), 4));
chunk_high = _mm_or_si128(
chunk_high, _mm_and_si128(_mm_and_si128(_mm_slli_epi32(chunk_right, 4),
_mm_set1_epi8(0xf0)),
mask3));
int c = _mm_extract_epi16(counts, 7);
int source_advance = !(c & 0x0200) ? 16 : !(c & 0x02) ? 15 : 14;
__m128i high_bits = _mm_and_si128(chunk_high, _mm_set1_epi8(0xf8));
if (!_mm_testz_si128(
mask3,
_mm_or_si128(_mm_cmpeq_epi8(high_bits, _mm_set1_epi8(0x00)),
_mm_cmpeq_epi8(high_bits, _mm_set1_epi8(0xd8)))))
return false;
shifts = _mm_blendv_epi8(shifts, _mm_srli_si128(shifts, 8),
_mm_srli_si128(_mm_slli_epi16(shifts, 4), 8));
chunk_high = _mm_slli_si128(chunk_high, 1);
__m128i shuf =
_mm_add_epi8(shifts, _mm_set_epi8(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5,
4, 3, 2, 1, 0));
chunk_low = _mm_shuffle_epi8(chunk_low, shuf);
chunk_high = _mm_shuffle_epi8(chunk_high, shuf);
__m128i utf16_low = _mm_unpacklo_epi8(chunk_low, chunk_high);
__m128i utf16_high = _mm_unpackhi_epi8(chunk_low, chunk_high);
if (_mm_cmpestrc(_mm_cvtsi64_si128(0xfdeffdd0fffffffe), 4, utf16_high, 8,
_SIDD_UWORD_OPS | _SIDD_CMP_RANGES) |
_mm_cmpestrc(_mm_cvtsi64_si128(0xfdeffdd0fffffffe), 4, utf16_low, 8,
_SIDD_UWORD_OPS | _SIDD_CMP_RANGES)) {
return false;
}
src += source_advance;
}
return validate_utf8(src, end - src);
}
static void mb_lpf_horizontal_edge_w_avx2_16(unsigned char *s, int p,
const unsigned char *_blimit,
const unsigned char *_limit,
const unsigned char *_thresh) {
__m128i mask, hev, flat, flat2;
const __m128i zero = _mm_set1_epi16(0);
const __m128i one = _mm_set1_epi8(1);
__m128i p7, p6, p5;
__m128i p4, p3, p2, p1, p0, q0, q1, q2, q3, q4;
__m128i q5, q6, q7;
__m256i p256_7, q256_7, p256_6, q256_6, p256_5, q256_5, p256_4, q256_4,
p256_3, q256_3, p256_2, q256_2, p256_1, q256_1, p256_0, q256_0;
const __m128i thresh =
_mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_thresh[0]));
const __m128i limit = _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_limit[0]));
const __m128i blimit =
_mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_blimit[0]));
p256_4 =
_mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 5 * p)));
p256_3 =
_mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 4 * p)));
p256_2 =
_mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 3 * p)));
p256_1 =
_mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 2 * p)));
p256_0 =
_mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 1 * p)));
q256_0 =
_mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 0 * p)));
q256_1 =
_mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 1 * p)));
q256_2 =
_mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 2 * p)));
q256_3 =
_mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 3 * p)));
q256_4 =
_mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 4 * p)));
p4 = _mm256_castsi256_si128(p256_4);
p3 = _mm256_castsi256_si128(p256_3);
p2 = _mm256_castsi256_si128(p256_2);
p1 = _mm256_castsi256_si128(p256_1);
p0 = _mm256_castsi256_si128(p256_0);
q0 = _mm256_castsi256_si128(q256_0);
q1 = _mm256_castsi256_si128(q256_1);
q2 = _mm256_castsi256_si128(q256_2);
q3 = _mm256_castsi256_si128(q256_3);
q4 = _mm256_castsi256_si128(q256_4);
{
const __m128i abs_p1p0 =
_mm_or_si128(_mm_subs_epu8(p1, p0), _mm_subs_epu8(p0, p1));
const __m128i abs_q1q0 =
_mm_or_si128(_mm_subs_epu8(q1, q0), _mm_subs_epu8(q0, q1));
const __m128i fe = _mm_set1_epi8(0xfe);
const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
__m128i abs_p0q0 =
_mm_or_si128(_mm_subs_epu8(p0, q0), _mm_subs_epu8(q0, p0));
__m128i abs_p1q1 =
_mm_or_si128(_mm_subs_epu8(p1, q1), _mm_subs_epu8(q1, p1));
__m128i work;
flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
hev = _mm_subs_epu8(flat, thresh);
hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
// mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
mask = _mm_max_epu8(flat, mask);
// mask |= (abs(p1 - p0) > limit) * -1;
// mask |= (abs(q1 - q0) > limit) * -1;
work = _mm_max_epu8(
_mm_or_si128(_mm_subs_epu8(p2, p1), _mm_subs_epu8(p1, p2)),
_mm_or_si128(_mm_subs_epu8(p3, p2), _mm_subs_epu8(p2, p3)));
mask = _mm_max_epu8(work, mask);
work = _mm_max_epu8(
_mm_or_si128(_mm_subs_epu8(q2, q1), _mm_subs_epu8(q1, q2)),
_mm_or_si128(_mm_subs_epu8(q3, q2), _mm_subs_epu8(q2, q3)));
mask = _mm_max_epu8(work, mask);
mask = _mm_subs_epu8(mask, limit);
mask = _mm_cmpeq_epi8(mask, zero);
}
// lp filter
{
const __m128i t4 = _mm_set1_epi8(4);
const __m128i t3 = _mm_set1_epi8(3);
const __m128i t80 = _mm_set1_epi8(0x80);
const __m128i te0 = _mm_set1_epi8(0xe0);
const __m128i t1f = _mm_set1_epi8(0x1f);
const __m128i t1 = _mm_set1_epi8(0x1);
const __m128i t7f = _mm_set1_epi8(0x7f);
__m128i ps1 = _mm_xor_si128(p1, t80);
__m128i ps0 = _mm_xor_si128(p0, t80);
__m128i qs0 = _mm_xor_si128(q0, t80);
__m128i qs1 = _mm_xor_si128(q1, t80);
__m128i filt;
__m128i work_a;
__m128i filter1, filter2;
__m128i flat2_p6, flat2_p5, flat2_p4, flat2_p3, flat2_p2, flat2_p1,
flat2_p0, flat2_q0, flat2_q1, flat2_q2, flat2_q3, flat2_q4, flat2_q5,
flat2_q6, flat_p2, flat_p1, flat_p0, flat_q0, flat_q1, flat_q2;
filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
work_a = _mm_subs_epi8(qs0, ps0);
filt = _mm_adds_epi8(filt, work_a);
filt = _mm_adds_epi8(filt, work_a);
filt = _mm_adds_epi8(filt, work_a);
/* (vpx_filter + 3 * (qs0 - ps0)) & mask */
filt = _mm_and_si128(filt, mask);
filter1 = _mm_adds_epi8(filt, t4);
filter2 = _mm_adds_epi8(filt, t3);
/* Filter1 >> 3 */
work_a = _mm_cmpgt_epi8(zero, filter1);
filter1 = _mm_srli_epi16(filter1, 3);
work_a = _mm_and_si128(work_a, te0);
filter1 = _mm_and_si128(filter1, t1f);
filter1 = _mm_or_si128(filter1, work_a);
qs0 = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
/* Filter2 >> 3 */
work_a = _mm_cmpgt_epi8(zero, filter2);
filter2 = _mm_srli_epi16(filter2, 3);
work_a = _mm_and_si128(work_a, te0);
filter2 = _mm_and_si128(filter2, t1f);
filter2 = _mm_or_si128(filter2, work_a);
ps0 = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
/* filt >> 1 */
filt = _mm_adds_epi8(filter1, t1);
work_a = _mm_cmpgt_epi8(zero, filt);
filt = _mm_srli_epi16(filt, 1);
work_a = _mm_and_si128(work_a, t80);
filt = _mm_and_si128(filt, t7f);
filt = _mm_or_si128(filt, work_a);
filt = _mm_andnot_si128(hev, filt);
ps1 = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
qs1 = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
// loopfilter done
{
__m128i work;
work = _mm_max_epu8(
_mm_or_si128(_mm_subs_epu8(p2, p0), _mm_subs_epu8(p0, p2)),
_mm_or_si128(_mm_subs_epu8(q2, q0), _mm_subs_epu8(q0, q2)));
flat = _mm_max_epu8(work, flat);
work = _mm_max_epu8(
_mm_or_si128(_mm_subs_epu8(p3, p0), _mm_subs_epu8(p0, p3)),
_mm_or_si128(_mm_subs_epu8(q3, q0), _mm_subs_epu8(q0, q3)));
flat = _mm_max_epu8(work, flat);
work = _mm_max_epu8(
_mm_or_si128(_mm_subs_epu8(p4, p0), _mm_subs_epu8(p0, p4)),
_mm_or_si128(_mm_subs_epu8(q4, q0), _mm_subs_epu8(q0, q4)));
flat = _mm_subs_epu8(flat, one);
flat = _mm_cmpeq_epi8(flat, zero);
flat = _mm_and_si128(flat, mask);
p256_5 = _mm256_castpd_si256(
_mm256_broadcast_pd((__m128d const *)(s - 6 * p)));
q256_5 = _mm256_castpd_si256(
_mm256_broadcast_pd((__m128d const *)(s + 5 * p)));
p5 = _mm256_castsi256_si128(p256_5);
q5 = _mm256_castsi256_si128(q256_5);
flat2 = _mm_max_epu8(
_mm_or_si128(_mm_subs_epu8(p5, p0), _mm_subs_epu8(p0, p5)),
_mm_or_si128(_mm_subs_epu8(q5, q0), _mm_subs_epu8(q0, q5)));
flat2 = _mm_max_epu8(work, flat2);
p256_6 = _mm256_castpd_si256(
_mm256_broadcast_pd((__m128d const *)(s - 7 * p)));
q256_6 = _mm256_castpd_si256(
_mm256_broadcast_pd((__m128d const *)(s + 6 * p)));
p6 = _mm256_castsi256_si128(p256_6);
q6 = _mm256_castsi256_si128(q256_6);
work = _mm_max_epu8(
_mm_or_si128(_mm_subs_epu8(p6, p0), _mm_subs_epu8(p0, p6)),
_mm_or_si128(_mm_subs_epu8(q6, q0), _mm_subs_epu8(q0, q6)));
flat2 = _mm_max_epu8(work, flat2);
p256_7 = _mm256_castpd_si256(
_mm256_broadcast_pd((__m128d const *)(s - 8 * p)));
q256_7 = _mm256_castpd_si256(
_mm256_broadcast_pd((__m128d const *)(s + 7 * p)));
p7 = _mm256_castsi256_si128(p256_7);
q7 = _mm256_castsi256_si128(q256_7);
work = _mm_max_epu8(
_mm_or_si128(_mm_subs_epu8(p7, p0), _mm_subs_epu8(p0, p7)),
_mm_or_si128(_mm_subs_epu8(q7, q0), _mm_subs_epu8(q0, q7)));
flat2 = _mm_max_epu8(work, flat2);
flat2 = _mm_subs_epu8(flat2, one);
flat2 = _mm_cmpeq_epi8(flat2, zero);
flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask
}
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// flat and wide flat calculations
{
const __m256i eight = _mm256_set1_epi16(8);
const __m256i four = _mm256_set1_epi16(4);
__m256i pixelFilter_p, pixelFilter_q, pixetFilter_p2p1p0,
pixetFilter_q2q1q0, sum_p7, sum_q7, sum_p3, sum_q3, res_p, res_q;
const __m256i filter =
_mm256_load_si256((__m256i const *)filt_loopfilter_avx2);
p256_7 = _mm256_shuffle_epi8(p256_7, filter);
p256_6 = _mm256_shuffle_epi8(p256_6, filter);
p256_5 = _mm256_shuffle_epi8(p256_5, filter);
p256_4 = _mm256_shuffle_epi8(p256_4, filter);
p256_3 = _mm256_shuffle_epi8(p256_3, filter);
p256_2 = _mm256_shuffle_epi8(p256_2, filter);
p256_1 = _mm256_shuffle_epi8(p256_1, filter);
p256_0 = _mm256_shuffle_epi8(p256_0, filter);
q256_0 = _mm256_shuffle_epi8(q256_0, filter);
q256_1 = _mm256_shuffle_epi8(q256_1, filter);
q256_2 = _mm256_shuffle_epi8(q256_2, filter);
q256_3 = _mm256_shuffle_epi8(q256_3, filter);
q256_4 = _mm256_shuffle_epi8(q256_4, filter);
q256_5 = _mm256_shuffle_epi8(q256_5, filter);
q256_6 = _mm256_shuffle_epi8(q256_6, filter);
q256_7 = _mm256_shuffle_epi8(q256_7, filter);
pixelFilter_p = _mm256_add_epi16(_mm256_add_epi16(p256_6, p256_5),
_mm256_add_epi16(p256_4, p256_3));
pixelFilter_q = _mm256_add_epi16(_mm256_add_epi16(q256_6, q256_5),
_mm256_add_epi16(q256_4, q256_3));
pixetFilter_p2p1p0 =
_mm256_add_epi16(p256_0, _mm256_add_epi16(p256_2, p256_1));
pixelFilter_p = _mm256_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
pixetFilter_q2q1q0 =
_mm256_add_epi16(q256_0, _mm256_add_epi16(q256_2, q256_1));
pixelFilter_q = _mm256_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
pixelFilter_p = _mm256_add_epi16(
eight, _mm256_add_epi16(pixelFilter_p, pixelFilter_q));
pixetFilter_p2p1p0 = _mm256_add_epi16(
four, _mm256_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0));
res_p = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(p256_7, p256_0)), 4);
flat2_p0 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
res_q = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(q256_7, q256_0)), 4);
flat2_q0 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
res_p =
_mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0,
_mm256_add_epi16(p256_3, p256_0)),
3);
flat_p0 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
res_q =
_mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0,
_mm256_add_epi16(q256_3, q256_0)),
3);
flat_q0 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
sum_p7 = _mm256_add_epi16(p256_7, p256_7);
sum_q7 = _mm256_add_epi16(q256_7, q256_7);
sum_p3 = _mm256_add_epi16(p256_3, p256_3);
sum_q3 = _mm256_add_epi16(q256_3, q256_3);
pixelFilter_q = _mm256_sub_epi16(pixelFilter_p, p256_6);
pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_6);
res_p = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_1)), 4);
flat2_p1 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
res_q = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_1)), 4);
flat2_q1 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
pixetFilter_q2q1q0 = _mm256_sub_epi16(pixetFilter_p2p1p0, p256_2);
pixetFilter_p2p1p0 = _mm256_sub_epi16(pixetFilter_p2p1p0, q256_2);
res_p =
_mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0,
_mm256_add_epi16(sum_p3, p256_1)),
3);
flat_p1 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
res_q =
_mm256_srli_epi16(_mm256_add_epi16(pixetFilter_q2q1q0,
_mm256_add_epi16(sum_q3, q256_1)),
3);
flat_q1 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
sum_p3 = _mm256_add_epi16(sum_p3, p256_3);
sum_q3 = _mm256_add_epi16(sum_q3, q256_3);
pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_5);
pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_5);
res_p = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_2)), 4);
flat2_p2 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
res_q = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_2)), 4);
flat2_q2 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
pixetFilter_p2p1p0 = _mm256_sub_epi16(pixetFilter_p2p1p0, q256_1);
pixetFilter_q2q1q0 = _mm256_sub_epi16(pixetFilter_q2q1q0, p256_1);
res_p =
_mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0,
_mm256_add_epi16(sum_p3, p256_2)),
3);
flat_p2 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
res_q =
_mm256_srli_epi16(_mm256_add_epi16(pixetFilter_q2q1q0,
_mm256_add_epi16(sum_q3, q256_2)),
3);
flat_q2 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_4);
pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_4);
res_p = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_3)), 4);
flat2_p3 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
res_q = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_3)), 4);
flat2_q3 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_3);
pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_3);
res_p = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_4)), 4);
flat2_p4 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
res_q = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_4)), 4);
flat2_q4 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_2);
pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_2);
res_p = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_5)), 4);
flat2_p5 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
res_q = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_5)), 4);
flat2_q5 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_1);
pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_1);
res_p = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_p, _mm256_add_epi16(sum_p7, p256_6)), 4);
flat2_p6 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
res_q = _mm256_srli_epi16(
_mm256_add_epi16(pixelFilter_q, _mm256_add_epi16(sum_q7, q256_6)), 4);
flat2_q6 = _mm256_castsi256_si128(
_mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
}
// wide flat
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
p2 = _mm_andnot_si128(flat, p2);
flat_p2 = _mm_and_si128(flat, flat_p2);
p2 = _mm_or_si128(flat_p2, p2);
p1 = _mm_andnot_si128(flat, ps1);
flat_p1 = _mm_and_si128(flat, flat_p1);
p1 = _mm_or_si128(flat_p1, p1);
p0 = _mm_andnot_si128(flat, ps0);
flat_p0 = _mm_and_si128(flat, flat_p0);
p0 = _mm_or_si128(flat_p0, p0);
q0 = _mm_andnot_si128(flat, qs0);
flat_q0 = _mm_and_si128(flat, flat_q0);
q0 = _mm_or_si128(flat_q0, q0);
q1 = _mm_andnot_si128(flat, qs1);
flat_q1 = _mm_and_si128(flat, flat_q1);
q1 = _mm_or_si128(flat_q1, q1);
q2 = _mm_andnot_si128(flat, q2);
flat_q2 = _mm_and_si128(flat, flat_q2);
q2 = _mm_or_si128(flat_q2, q2);
p6 = _mm_andnot_si128(flat2, p6);
flat2_p6 = _mm_and_si128(flat2, flat2_p6);
p6 = _mm_or_si128(flat2_p6, p6);
_mm_storeu_si128((__m128i *)(s - 7 * p), p6);
p5 = _mm_andnot_si128(flat2, p5);
flat2_p5 = _mm_and_si128(flat2, flat2_p5);
p5 = _mm_or_si128(flat2_p5, p5);
_mm_storeu_si128((__m128i *)(s - 6 * p), p5);
p4 = _mm_andnot_si128(flat2, p4);
flat2_p4 = _mm_and_si128(flat2, flat2_p4);
p4 = _mm_or_si128(flat2_p4, p4);
_mm_storeu_si128((__m128i *)(s - 5 * p), p4);
p3 = _mm_andnot_si128(flat2, p3);
flat2_p3 = _mm_and_si128(flat2, flat2_p3);
p3 = _mm_or_si128(flat2_p3, p3);
_mm_storeu_si128((__m128i *)(s - 4 * p), p3);
p2 = _mm_andnot_si128(flat2, p2);
flat2_p2 = _mm_and_si128(flat2, flat2_p2);
p2 = _mm_or_si128(flat2_p2, p2);
_mm_storeu_si128((__m128i *)(s - 3 * p), p2);
p1 = _mm_andnot_si128(flat2, p1);
flat2_p1 = _mm_and_si128(flat2, flat2_p1);
p1 = _mm_or_si128(flat2_p1, p1);
_mm_storeu_si128((__m128i *)(s - 2 * p), p1);
p0 = _mm_andnot_si128(flat2, p0);
flat2_p0 = _mm_and_si128(flat2, flat2_p0);
p0 = _mm_or_si128(flat2_p0, p0);
_mm_storeu_si128((__m128i *)(s - 1 * p), p0);
q0 = _mm_andnot_si128(flat2, q0);
flat2_q0 = _mm_and_si128(flat2, flat2_q0);
q0 = _mm_or_si128(flat2_q0, q0);
_mm_storeu_si128((__m128i *)(s - 0 * p), q0);
q1 = _mm_andnot_si128(flat2, q1);
flat2_q1 = _mm_and_si128(flat2, flat2_q1);
q1 = _mm_or_si128(flat2_q1, q1);
_mm_storeu_si128((__m128i *)(s + 1 * p), q1);
q2 = _mm_andnot_si128(flat2, q2);
flat2_q2 = _mm_and_si128(flat2, flat2_q2);
q2 = _mm_or_si128(flat2_q2, q2);
_mm_storeu_si128((__m128i *)(s + 2 * p), q2);
q3 = _mm_andnot_si128(flat2, q3);
flat2_q3 = _mm_and_si128(flat2, flat2_q3);
q3 = _mm_or_si128(flat2_q3, q3);
_mm_storeu_si128((__m128i *)(s + 3 * p), q3);
q4 = _mm_andnot_si128(flat2, q4);
flat2_q4 = _mm_and_si128(flat2, flat2_q4);
q4 = _mm_or_si128(flat2_q4, q4);
_mm_storeu_si128((__m128i *)(s + 4 * p), q4);
q5 = _mm_andnot_si128(flat2, q5);
flat2_q5 = _mm_and_si128(flat2, flat2_q5);
q5 = _mm_or_si128(flat2_q5, q5);
_mm_storeu_si128((__m128i *)(s + 5 * p), q5);
q6 = _mm_andnot_si128(flat2, q6);
flat2_q6 = _mm_and_si128(flat2, flat2_q6);
q6 = _mm_or_si128(flat2_q6, q6);
_mm_storeu_si128((__m128i *)(s + 6 * p), q6);
}
}
