SIMD_INLINE __m256i BinomialSum16(const __m256i & ab, const __m256i & cd)
{
#ifdef SIMD_MADDUBS_ERROR
return _mm256_add_epi16(_mm256_maddubs_epi16(_mm256_or_si256(K_ZERO, ab), K8_01_03), _mm256_maddubs_epi16(_mm256_or_si256(K_ZERO, cd), K8_03_01));
#else
return _mm256_add_epi16(_mm256_maddubs_epi16(ab, K8_01_03), _mm256_maddubs_epi16(cd, K8_03_01));
#endif
}
SIMD_INLINE void MaskSrc(const uint8_t * src, const uint8_t * mask, const __m256i & index, ptrdiff_t offset, uint16_t * dst)
{
const __m256i _src = Load<srcAlign>((__m256i*)(src + offset));
const __m256i _mask = _mm256_and_si256(_mm256_cmpeq_epi8(Load<srcAlign>((__m256i*)(mask + offset)), index), K8_01);
__m256i lo = _mm256_mullo_epi16(_mm256_add_epi16(K16_0008, UnpackU8<0>(_src)), UnpackU8<0>(_mask));
__m256i hi = _mm256_mullo_epi16(_mm256_add_epi16(K16_0008, UnpackU8<1>(_src)), UnpackU8<1>(_mask));
Store<dstAlign>((__m256i*)(dst + offset) + 0, _mm256_permute2x128_si256(lo, hi, 0x20));
Store<dstAlign>((__m256i*)(dst + offset) + 1, _mm256_permute2x128_si256(lo, hi, 0x31));
}
template<bool align> SIMD_INLINE void MainRowY5x5(__m256i odd, __m256i even, Buffer & buffer, size_t offset)
{
__m256i cp = _mm256_mullo_epi16(odd, K16_0004);
__m256i c0 = Load<align>((__m256i*)(buffer.in0 + offset));
__m256i c1 = Load<align>((__m256i*)(buffer.in1 + offset));
Store<align>((__m256i*)(buffer.dst + offset), _mm256_add_epi16(even, _mm256_add_epi16(c1, _mm256_add_epi16(cp, _mm256_mullo_epi16(c0, K16_0006)))));
Store<align>((__m256i*)(buffer.out1 + offset), _mm256_add_epi16(c0, cp));
Store<align>((__m256i*)(buffer.out0 + offset), even);
}
template <bool align, bool compensation> SIMD_INLINE __m256i MainRowX5x5(uint16_t * dst)
{
__m256i t0 = _mm256_loadu_si256((__m256i*)(dst - 2));
__m256i t1 = _mm256_loadu_si256((__m256i*)(dst - 1));
__m256i t2 = Load<align>((__m256i*)dst);
__m256i t3 = _mm256_loadu_si256((__m256i*)(dst + 1));
__m256i t4 = _mm256_loadu_si256((__m256i*)(dst + 2));
t2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_mullo_epi16(t2, K16_0006), _mm256_mullo_epi16(_mm256_add_epi16(t1, t3), K16_0004)), _mm256_add_epi16(t0, t4));
return DivideBy256<compensation>(t2);
}
static void satd_8bit_4x4_dual_avx2(
const pred_buffer preds, const kvz_pixel * const orig, unsigned num_modes, unsigned *satds_out)
{
__m256i original = _mm256_broadcastsi128_si256(_mm_cvtepu8_epi16(_mm_loadl_epi64((__m128i*)orig)));
__m256i pred = _mm256_cvtepu8_epi16(_mm_loadl_epi64((__m128i*)preds[0]));
pred = _mm256_inserti128_si256(pred, _mm_cvtepu8_epi16(_mm_loadl_epi64((__m128i*)preds[1])), 1);
__m256i diff_lo = _mm256_sub_epi16(pred, original);
original = _mm256_broadcastsi128_si256(_mm_cvtepu8_epi16(_mm_loadl_epi64((__m128i*)(orig + 8))));
pred = _mm256_cvtepu8_epi16(_mm_loadl_epi64((__m128i*)(preds[0] + 8)));
pred = _mm256_inserti128_si256(pred, _mm_cvtepu8_epi16(_mm_loadl_epi64((__m128i*)(preds[1] + 8))), 1);
__m256i diff_hi = _mm256_sub_epi16(pred, original);
//Hor
__m256i row0 = _mm256_hadd_epi16(diff_lo, diff_hi);
__m256i row1 = _mm256_hsub_epi16(diff_lo, diff_hi);
__m256i row2 = _mm256_hadd_epi16(row0, row1);
__m256i row3 = _mm256_hsub_epi16(row0, row1);
//Ver
row0 = _mm256_hadd_epi16(row2, row3);
row1 = _mm256_hsub_epi16(row2, row3);
row2 = _mm256_hadd_epi16(row0, row1);
row3 = _mm256_hsub_epi16(row0, row1);
//Abs and sum
row2 = _mm256_abs_epi16(row2);
row3 = _mm256_abs_epi16(row3);
row3 = _mm256_add_epi16(row2, row3);
row3 = _mm256_add_epi16(row3, _mm256_shuffle_epi32(row3, KVZ_PERMUTE(2, 3, 0, 1) ));
row3 = _mm256_add_epi16(row3, _mm256_shuffle_epi32(row3, KVZ_PERMUTE(1, 0, 1, 0) ));
row3 = _mm256_add_epi16(row3, _mm256_shufflelo_epi16(row3, KVZ_PERMUTE(1, 0, 1, 0) ));
unsigned sum1 = _mm_extract_epi16(_mm256_castsi256_si128(row3), 0);
sum1 = (sum1 + 1) >> 1;
unsigned sum2 = _mm_extract_epi16(_mm256_extracti128_si256(row3, 1), 0);
sum2 = (sum2 + 1) >> 1;
satds_out[0] = sum1;
satds_out[1] = sum2;
}
void fft128_2way( void *a )
{
int i;
// Temp space to help for interleaving in the end
__m256i B[8];
__m256i *A = (__m256i*) a;
// __m256i *Twiddle = (__m256i*)FFT128_Twiddle;
/* Size-2 butterflies */
for ( i = 0; i<8; i++ )
{
B[ i ] = _mm256_add_epi16( A[ i ], A[ i+8 ] );
B[ i ] = REDUCE_FULL_S( B[ i ] );
A[ i+8 ] = _mm256_sub_epi16( A[ i ], A[ i+8 ] );
A[ i+8 ] = REDUCE_FULL_S( A[ i+8 ] );
A[ i+8 ] = _mm256_mullo_epi16( A[ i+8 ], FFT128_Twiddle[i].m256i );
A[ i+8 ] = REDUCE_FULL_S( A[ i+8 ] );
}
fft64_2way( B );
fft64_2way( A+8 );
/* Transpose (i.e. interleave) */
for ( i = 0; i < 8; i++ )
{
A[ 2*i ] = _mm256_unpacklo_epi16( B[ i ], A[ i+8 ] );
A[ 2*i+1 ] = _mm256_unpackhi_epi16( B[ i ], A[ i+8 ] );
}
}
static INLINE __m256i calc_mask_avx2(const __m256i mask_base, const __m256i s0,
const __m256i s1) {
const __m256i diff = _mm256_abs_epi16(_mm256_sub_epi16(s0, s1));
return _mm256_abs_epi16(
_mm256_add_epi16(mask_base, _mm256_srli_epi16(diff, 4)));
// clamp(diff, 0, 64) can be skiped for diff is always in the range ( 38, 54)
}
void vpx_hadamard_32x32_avx2(const int16_t *src_diff, ptrdiff_t src_stride,
tran_low_t *coeff) {
#if CONFIG_VP9_HIGHBITDEPTH
// For high bitdepths, it is unnecessary to store_tran_low
// (mult/unpack/store), then load_tran_low (load/pack) the same memory in the
// next stage. Output to an intermediate buffer first, then store_tran_low()
// in the final stage.
DECLARE_ALIGNED(32, int16_t, temp_coeff[32 * 32]);
int16_t *t_coeff = temp_coeff;
#else
int16_t *t_coeff = coeff;
#endif
int idx;
for (idx = 0; idx < 4; ++idx) {
// src_diff: 9 bit, dynamic range [-255, 255]
const int16_t *src_ptr =
src_diff + (idx >> 1) * 16 * src_stride + (idx & 0x01) * 16;
hadamard_16x16_avx2(src_ptr, src_stride,
(tran_low_t *)(t_coeff + idx * 256), 0);
}
for (idx = 0; idx < 256; idx += 16) {
const __m256i coeff0 = _mm256_loadu_si256((const __m256i *)t_coeff);
const __m256i coeff1 = _mm256_loadu_si256((const __m256i *)(t_coeff + 256));
const __m256i coeff2 = _mm256_loadu_si256((const __m256i *)(t_coeff + 512));
const __m256i coeff3 = _mm256_loadu_si256((const __m256i *)(t_coeff + 768));
__m256i b0 = _mm256_add_epi16(coeff0, coeff1);
__m256i b1 = _mm256_sub_epi16(coeff0, coeff1);
__m256i b2 = _mm256_add_epi16(coeff2, coeff3);
__m256i b3 = _mm256_sub_epi16(coeff2, coeff3);
b0 = _mm256_srai_epi16(b0, 2);
b1 = _mm256_srai_epi16(b1, 2);
b2 = _mm256_srai_epi16(b2, 2);
b3 = _mm256_srai_epi16(b3, 2);
store_tran_low(_mm256_add_epi16(b0, b2), coeff);
store_tran_low(_mm256_add_epi16(b1, b3), coeff + 256);
store_tran_low(_mm256_sub_epi16(b0, b2), coeff + 512);
store_tran_low(_mm256_sub_epi16(b1, b3), coeff + 768);
coeff += 16;
t_coeff += 16;
}
}
int main() {
const ssize_t A = 3;
const size_t Awidth = 2;
const size_t Dwidth = 4;
const ssize_t Dmin = (-1) * (1ll << (Dwidth - 1));
const ssize_t Dmax = (1ll << (Dwidth - 1)) - 1;
const ssize_t Cwidth = Awidth + Dwidth;
const ssize_t AInv = ext_euklidean(A, Cwidth) & ((1ll << Cwidth) - 1);
const size_t numCodewords = (1ull << Cwidth);
std::cout << "numCodewords: " << numCodewords << std::endl;
const size_t numMasks = numCodewords / (sizeof(int) * 4); // How many masks will we generate?
int * pNonCodewordMasks = new int[numMasks];
const int16_t c = ~((1ll << (Cwidth - 1)) - 1);
std::cout << "c = 0x" << std::hex << c << std::dec << std::endl;
for (ssize_t i = 0, cw = c, posMask = 0; i < numCodewords; ++posMask) {
int tmpMask = 0;
for (ssize_t k = 0; k < 16; ++k, ++cw, ++i) {
if ((cw % A) != 0) { // we want the non-codewords
// std::cout << "cw % A != 0: " << cw << std::endl;
tmpMask |= (1ll << (k * 2)) | (1ll << (k * 2 + 1)); // expand to 32 bits, because AVX2 cannot movemask across lanes to 16 bits
}
}
pNonCodewordMasks[posMask] = tmpMask;
}
std::cout << "numMasks: " << numMasks << std::endl;
std::cout << "non-codeword-masks: 0x" << std::hex << std::setfill('0');
for (size_t posMask = 0; posMask < numMasks; ++posMask) {
std::cout << std::setw(8) << pNonCodewordMasks[posMask] << ':';
}
std::cout << std::dec << std::endl << std::setfill(' ');
auto mmCodewords = _mm256_set_epi16(c+15, c+14, c+13, c+12, c+11, c+10, c+9, c+8, c+7, c+6, c+5, c+4, c+3, c+2, c+1, c);
auto mmAddUp = _mm256_set1_epi16(16);
auto mmAinv = _mm256_set1_epi16(AInv);
auto mmDmin = _mm256_set1_epi16(Dmin);
auto mmDmax = _mm256_set1_epi16(Dmax);
const size_t posEnd = (1ull << Cwidth);
__m256i mmFillUp[] = {_mm256_set1_epi16(0), _mm256_set1_epi16(~((1ll << Cwidth) - 1))}; // fill up all non-codeword bits with 1's if necessary
std::cout << "posEnd = 0x" << std::hex << posEnd << std::dec << std::endl;
std::cout << std::setfill('0') << std::hex;
for(size_t pos = 15, posMask = 0; pos < posEnd; pos += 16, ++posMask) {
auto isNeg = 0x1 & _mm256_movemask_epi8(_mm256_cmpgt_epi16(mmFillUp[0], mmCodewords));
auto mm1 = _mm256_or_si256(_mm256_mullo_epi16(mmCodewords, mmAinv), mmFillUp[isNeg]);
auto mm2 = _mm256_cmpgt_epi16(mm1, mmDmin);
auto mm3 = _mm256_cmpgt_epi16(mmDmax, mm1);
auto mm4 = _mm256_cmpeq_epi16(mmDmax, mm1);
auto mm5 = _mm256_or_si256(mm3, mm4);
auto mm6 = _mm256_and_si256(mm2, mm5);
auto mask = _mm256_movemask_epi8(mm6);
if (mask & pNonCodewordMasks[posMask]) {
std::cout << "BAD @0x" << std::setw((Cwidth + 7) / 8) << pos << ": 0x" << mask << " & 0x" << pNonCodewordMasks[posMask] << " = 0x" << (mask & pNonCodewordMasks[posMask]) << std::endl;
} else {
std::cout << "OK @0x" << std::setw((Cwidth + 7) / 8) << pos << ": 0x" << mask << " & 0x" << pNonCodewordMasks[posMask] << " = 0x" << (mask & pNonCodewordMasks[posMask]) << std::endl;
}
mmCodewords = _mm256_add_epi16(mmCodewords, mmAddUp);
}
std::cout << std::setfill(' ') << std::dec;
}
static INLINE void variance_kernel_avx2(const __m256i src, const __m256i ref,
__m256i *const sse,
__m256i *const sum) {
const __m256i adj_sub = _mm256_set1_epi16(0xff01); // (1,-1)
// unpack into pairs of source and reference values
const __m256i src_ref0 = _mm256_unpacklo_epi8(src, ref);
const __m256i src_ref1 = _mm256_unpackhi_epi8(src, ref);
// subtract adjacent elements using src*1 + ref*-1
const __m256i diff0 = _mm256_maddubs_epi16(src_ref0, adj_sub);
const __m256i diff1 = _mm256_maddubs_epi16(src_ref1, adj_sub);
const __m256i madd0 = _mm256_madd_epi16(diff0, diff0);
const __m256i madd1 = _mm256_madd_epi16(diff1, diff1);
// add to the running totals
*sum = _mm256_add_epi16(*sum, _mm256_add_epi16(diff0, diff1));
*sse = _mm256_add_epi32(*sse, _mm256_add_epi32(madd0, madd1));
}
// 32bpp optimized for 8-bit ARGB/RGBA. rmask should be 0x00FF,0x00FF,... etc
static inline __m256i stretchblt_line_bilinear_pixel_blend_avx_argb8(const __m256i cur,const __m256i nxt,const __m256i mul,const __m256i rmask) {
__m256i rc,gc;
__m256i rn,gn;
__m256i d,sum;
rc = _mm256_and_si256( cur ,rmask);
gc = _mm256_and_si256(_mm256_srli_epi16(cur,8),rmask);
rn = _mm256_and_si256( nxt ,rmask);
gn = _mm256_and_si256(_mm256_srli_epi16(nxt,8),rmask);
d = _mm256_sub_epi16(rn,rc);
sum = _mm256_add_epi16(rc,_mm256_mulhi_epi16(_mm256_add_epi16(d,d),mul));
d = _mm256_sub_epi16(gn,gc);
sum = _mm256_add_epi16(_mm256_slli_epi16(_mm256_add_epi16(gc,_mm256_mulhi_epi16(_mm256_add_epi16(d,d),mul)),8),sum);
return sum;
}
static INLINE void hor_transform_row_dual_avx2(__m256i* row){
__m256i mask_pos = _mm256_set1_epi16(1);
__m256i mask_neg = _mm256_set1_epi16(-1);
__m256i sign_mask = _mm256_unpacklo_epi64(mask_pos, mask_neg);
__m256i temp = _mm256_shuffle_epi32(*row, KVZ_PERMUTE(2, 3, 0, 1));
*row = _mm256_sign_epi16(*row, sign_mask);
*row = _mm256_add_epi16(*row, temp);
sign_mask = _mm256_unpacklo_epi32(mask_pos, mask_neg);
temp = _mm256_shuffle_epi32(*row, KVZ_PERMUTE(1, 0, 3, 2));
*row = _mm256_sign_epi16(*row, sign_mask);
*row = _mm256_add_epi16(*row, temp);
sign_mask = _mm256_unpacklo_epi16(mask_pos, mask_neg);
temp = _mm256_shufflelo_epi16(*row, KVZ_PERMUTE(1,0,3,2));
temp = _mm256_shufflehi_epi16(temp, KVZ_PERMUTE(1,0,3,2));
*row = _mm256_sign_epi16(*row, sign_mask);
*row = _mm256_add_epi16(*row, temp);
}
static INLINE void hadamard_16x16_avx2(const int16_t *src_diff,
ptrdiff_t src_stride, tran_low_t *coeff,
int is_final) {
#if CONFIG_VP9_HIGHBITDEPTH
DECLARE_ALIGNED(32, int16_t, temp_coeff[16 * 16]);
int16_t *t_coeff = temp_coeff;
#else
int16_t *t_coeff = coeff;
#endif
int16_t *coeff16 = (int16_t *)coeff;
int idx;
for (idx = 0; idx < 2; ++idx) {
const int16_t *src_ptr = src_diff + idx * 8 * src_stride;
hadamard_8x8x2_avx2(src_ptr, src_stride, t_coeff + (idx * 64 * 2));
}
for (idx = 0; idx < 64; idx += 16) {
const __m256i coeff0 = _mm256_loadu_si256((const __m256i *)t_coeff);
const __m256i coeff1 = _mm256_loadu_si256((const __m256i *)(t_coeff + 64));
const __m256i coeff2 = _mm256_loadu_si256((const __m256i *)(t_coeff + 128));
const __m256i coeff3 = _mm256_loadu_si256((const __m256i *)(t_coeff + 192));
__m256i b0 = _mm256_add_epi16(coeff0, coeff1);
__m256i b1 = _mm256_sub_epi16(coeff0, coeff1);
__m256i b2 = _mm256_add_epi16(coeff2, coeff3);
__m256i b3 = _mm256_sub_epi16(coeff2, coeff3);
b0 = _mm256_srai_epi16(b0, 1);
b1 = _mm256_srai_epi16(b1, 1);
b2 = _mm256_srai_epi16(b2, 1);
b3 = _mm256_srai_epi16(b3, 1);
if (is_final) {
store_tran_low(_mm256_add_epi16(b0, b2), coeff);
store_tran_low(_mm256_add_epi16(b1, b3), coeff + 64);
store_tran_low(_mm256_sub_epi16(b0, b2), coeff + 128);
store_tran_low(_mm256_sub_epi16(b1, b3), coeff + 192);
coeff += 16;
} else {
_mm256_storeu_si256((__m256i *)coeff16, _mm256_add_epi16(b0, b2));
_mm256_storeu_si256((__m256i *)(coeff16 + 64), _mm256_add_epi16(b1, b3));
_mm256_storeu_si256((__m256i *)(coeff16 + 128), _mm256_sub_epi16(b0, b2));
_mm256_storeu_si256((__m256i *)(coeff16 + 192), _mm256_sub_epi16(b1, b3));
coeff16 += 16;
}
t_coeff += 16;
}
}
// 16bpp general R/G/B, usually 5/6/5 or 5/5/5
static inline __m256i stretchblt_line_bilinear_pixel_blend_avx_rgb16(const __m256i cur,const __m256i nxt,const __m256i mul,const __m256i rmask,const uint16_t rshift,const __m256i gmask,const uint16_t gshift,const __m256i bmask,const uint16_t bshift) {
__m256i rc,gc,bc;
__m256i rn,gn,bn;
__m256i d,sum;
rc = _mm256_and_si256(_mm256_srli_epi16(cur,rshift),rmask);
gc = _mm256_and_si256(_mm256_srli_epi16(cur,gshift),gmask);
bc = _mm256_and_si256(_mm256_srli_epi16(cur,bshift),bmask);
rn = _mm256_and_si256(_mm256_srli_epi16(nxt,rshift),rmask);
gn = _mm256_and_si256(_mm256_srli_epi16(nxt,gshift),gmask);
bn = _mm256_and_si256(_mm256_srli_epi16(nxt,bshift),bmask);
d = _mm256_sub_epi16(rn,rc);
sum = _mm256_slli_epi16(_mm256_add_epi16(rc,_mm256_mulhi_epi16(_mm256_add_epi16(d,d),mul)),rshift);
d = _mm256_sub_epi16(gn,gc);
sum = _mm256_add_epi16(_mm256_slli_epi16(_mm256_add_epi16(gc,_mm256_mulhi_epi16(_mm256_add_epi16(d,d),mul)),gshift),sum);
d = _mm256_sub_epi16(bn,bc);
sum = _mm256_add_epi16(_mm256_slli_epi16(_mm256_add_epi16(bc,_mm256_mulhi_epi16(_mm256_add_epi16(d,d),mul)),bshift),sum);
return sum;
}
template <> SIMD_INLINE __m256i DivideBy256<true>(__m256i value)
{
return _mm256_srli_epi16(_mm256_add_epi16(value, K16_0080), 8);
}
__m256i test_mm256_add_epi16(__m256i a, __m256i b) {
// CHECK: add <16 x i16>
return _mm256_add_epi16(a, b);
}
static INLINE void add_sub_dual_avx2(__m256i *out, __m256i *in, unsigned out_idx0, unsigned out_idx1, unsigned in_idx0, unsigned in_idx1)
{
out[out_idx0] = _mm256_add_epi16(in[in_idx0], in[in_idx1]);
out[out_idx1] = _mm256_sub_epi16(in[in_idx0], in[in_idx1]);
}
template<bool align> SIMD_INLINE __m256i AverageRow16(const Buffer & buffer, size_t offset)
{
return _mm256_mulhi_epu16(K16_DIVISION_BY_9_FACTOR, _mm256_add_epi16(
_mm256_add_epi16(K16_0005, Load<align>((__m256i*)(buffer.src0 + offset))),
_mm256_add_epi16(Load<align>((__m256i*)(buffer.src1 + offset)), Load<align>((__m256i*)(buffer.src2 + offset)))));
}
/*!
* \brief Add the two given values and return the result.
*/
ETL_STATIC_INLINE(avx_simd_short) add(avx_simd_short lhs, avx_simd_short rhs) {
return _mm256_add_epi16(lhs.value, rhs.value);
}
template <> SIMD_INLINE __m256i OperationBinary16i<SimdOperationBinary16iAddition>(const __m256i & a, const __m256i & b)
{
return _mm256_add_epi16(a, b);
}
template<int part> SIMD_INLINE __m256i SumCol(__m256i a[3])
{
return _mm256_add_epi16(_mm256_maddubs_epi16(UnpackU8<part>(a[0], a[1]), K8_01), UnpackU8<part>(a[2]));
}
SIMD_INLINE __m256i BinomialSum16(const __m256i & a, const __m256i & b, const __m256i & c, const __m256i & d)
{
return _mm256_add_epi16(_mm256_add_epi16(a, d), _mm256_mullo_epi16(_mm256_add_epi16(b, c), K16_0003));
}
SIMD_INLINE __m256i DivideBy64(__m256i value)
{
return _mm256_srli_epi16(_mm256_add_epi16(value, K16_0020), 6);
}
SIMD_INLINE __m256i Average16(const __m256i & s0, const __m256i & s1)
{
return _mm256_srli_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(s0, K8_01), _mm256_maddubs_epi16(s1, K8_01)), K16_0002), 2);
}
SIMD_INLINE __m256i Average16(const __m256i & s0, const __m256i & s1)
{
return _mm256_srli_epi16(_mm256_add_epi16(_mm256_add_epi16(
_mm256_hadd_epi16(_mm256_unpacklo_epi8(s0, K_ZERO), _mm256_unpackhi_epi8(s0, K_ZERO)),
_mm256_hadd_epi16(_mm256_unpacklo_epi8(s1, K_ZERO), _mm256_unpackhi_epi8(s1, K_ZERO))), K16_0002), 2);
}
template<bool align> SIMD_INLINE __m256i InterpolateY(const __m256i * pbx0, const __m256i * pbx1, __m256i alpha[2])
{
__m256i sum = _mm256_add_epi16(_mm256_mullo_epi16(Load<align>(pbx0), alpha[0]), _mm256_mullo_epi16(Load<align>(pbx1), alpha[1]));
return _mm256_srli_epi16(_mm256_add_epi16(sum, K16_FRACTION_ROUND_TERM), Base::BILINEAR_SHIFT);
}
static void hadamard_col8x2_avx2(__m256i *in, int iter) {
__m256i a0 = in[0];
__m256i a1 = in[1];
__m256i a2 = in[2];
__m256i a3 = in[3];
__m256i a4 = in[4];
__m256i a5 = in[5];
__m256i a6 = in[6];
__m256i a7 = in[7];
__m256i b0 = _mm256_add_epi16(a0, a1);
__m256i b1 = _mm256_sub_epi16(a0, a1);
__m256i b2 = _mm256_add_epi16(a2, a3);
__m256i b3 = _mm256_sub_epi16(a2, a3);
__m256i b4 = _mm256_add_epi16(a4, a5);
__m256i b5 = _mm256_sub_epi16(a4, a5);
__m256i b6 = _mm256_add_epi16(a6, a7);
__m256i b7 = _mm256_sub_epi16(a6, a7);
a0 = _mm256_add_epi16(b0, b2);
a1 = _mm256_add_epi16(b1, b3);
a2 = _mm256_sub_epi16(b0, b2);
a3 = _mm256_sub_epi16(b1, b3);
a4 = _mm256_add_epi16(b4, b6);
a5 = _mm256_add_epi16(b5, b7);
a6 = _mm256_sub_epi16(b4, b6);
a7 = _mm256_sub_epi16(b5, b7);
if (iter == 0) {
b0 = _mm256_add_epi16(a0, a4);
b7 = _mm256_add_epi16(a1, a5);
b3 = _mm256_add_epi16(a2, a6);
b4 = _mm256_add_epi16(a3, a7);
b2 = _mm256_sub_epi16(a0, a4);
b6 = _mm256_sub_epi16(a1, a5);
b1 = _mm256_sub_epi16(a2, a6);
b5 = _mm256_sub_epi16(a3, a7);
a0 = _mm256_unpacklo_epi16(b0, b1);
a1 = _mm256_unpacklo_epi16(b2, b3);
a2 = _mm256_unpackhi_epi16(b0, b1);
a3 = _mm256_unpackhi_epi16(b2, b3);
a4 = _mm256_unpacklo_epi16(b4, b5);
a5 = _mm256_unpacklo_epi16(b6, b7);
a6 = _mm256_unpackhi_epi16(b4, b5);
a7 = _mm256_unpackhi_epi16(b6, b7);
b0 = _mm256_unpacklo_epi32(a0, a1);
b1 = _mm256_unpacklo_epi32(a4, a5);
b2 = _mm256_unpackhi_epi32(a0, a1);
b3 = _mm256_unpackhi_epi32(a4, a5);
b4 = _mm256_unpacklo_epi32(a2, a3);
b5 = _mm256_unpacklo_epi32(a6, a7);
b6 = _mm256_unpackhi_epi32(a2, a3);
b7 = _mm256_unpackhi_epi32(a6, a7);
in[0] = _mm256_unpacklo_epi64(b0, b1);
in[1] = _mm256_unpackhi_epi64(b0, b1);
in[2] = _mm256_unpacklo_epi64(b2, b3);
in[3] = _mm256_unpackhi_epi64(b2, b3);
in[4] = _mm256_unpacklo_epi64(b4, b5);
in[5] = _mm256_unpackhi_epi64(b4, b5);
in[6] = _mm256_unpacklo_epi64(b6, b7);
in[7] = _mm256_unpackhi_epi64(b6, b7);
} else {
in[0] = _mm256_add_epi16(a0, a4);
in[7] = _mm256_add_epi16(a1, a5);
in[3] = _mm256_add_epi16(a2, a6);
in[4] = _mm256_add_epi16(a3, a7);
in[2] = _mm256_sub_epi16(a0, a4);
in[6] = _mm256_sub_epi16(a1, a5);
in[1] = _mm256_sub_epi16(a2, a6);
in[5] = _mm256_sub_epi16(a3, a7);
}
}
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);
}
}
void av1_build_compound_diffwtd_mask_highbd_avx2(
uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0,
int src0_stride, const uint8_t *src1, int src1_stride, int h, int w,
int bd) {
if (w < 16) {
av1_build_compound_diffwtd_mask_highbd_ssse3(
mask, mask_type, src0, src0_stride, src1, src1_stride, h, w, bd);
} else {
assert(mask_type == DIFFWTD_38 || mask_type == DIFFWTD_38_INV);
assert(bd >= 8);
assert((w % 16) == 0);
const __m256i y0 = _mm256_setzero_si256();
const __m256i yAOM_BLEND_A64_MAX_ALPHA =
_mm256_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
const int mask_base = 38;
const __m256i ymask_base = _mm256_set1_epi16(mask_base);
const uint16_t *ssrc0 = CONVERT_TO_SHORTPTR(src0);
const uint16_t *ssrc1 = CONVERT_TO_SHORTPTR(src1);
if (bd == 8) {
if (mask_type == DIFFWTD_38_INV) {
for (int i = 0; i < h; ++i) {
for (int j = 0; j < w; j += 16) {
__m256i s0 = _mm256_loadu_si256((const __m256i *)&ssrc0[j]);
__m256i s1 = _mm256_loadu_si256((const __m256i *)&ssrc1[j]);
__m256i diff = _mm256_srai_epi16(
_mm256_abs_epi16(_mm256_sub_epi16(s0, s1)), DIFF_FACTOR_LOG2);
__m256i m = _mm256_min_epi16(
_mm256_max_epi16(y0, _mm256_add_epi16(diff, ymask_base)),
yAOM_BLEND_A64_MAX_ALPHA);
m = _mm256_sub_epi16(yAOM_BLEND_A64_MAX_ALPHA, m);
m = _mm256_packus_epi16(m, m);
m = _mm256_permute4x64_epi64(m, _MM_SHUFFLE(0, 0, 2, 0));
__m128i m0 = _mm256_castsi256_si128(m);
_mm_storeu_si128((__m128i *)&mask[j], m0);
}
ssrc0 += src0_stride;
ssrc1 += src1_stride;
mask += w;
}
} else {
for (int i = 0; i < h; ++i) {
for (int j = 0; j < w; j += 16) {
__m256i s0 = _mm256_loadu_si256((const __m256i *)&ssrc0[j]);
__m256i s1 = _mm256_loadu_si256((const __m256i *)&ssrc1[j]);
__m256i diff = _mm256_srai_epi16(
_mm256_abs_epi16(_mm256_sub_epi16(s0, s1)), DIFF_FACTOR_LOG2);
__m256i m = _mm256_min_epi16(
_mm256_max_epi16(y0, _mm256_add_epi16(diff, ymask_base)),
yAOM_BLEND_A64_MAX_ALPHA);
m = _mm256_packus_epi16(m, m);
m = _mm256_permute4x64_epi64(m, _MM_SHUFFLE(0, 0, 2, 0));
__m128i m0 = _mm256_castsi256_si128(m);
_mm_storeu_si128((__m128i *)&mask[j], m0);
}
ssrc0 += src0_stride;
ssrc1 += src1_stride;
mask += w;
}
}
} else {
const __m128i xshift = xx_set1_64_from_32i(bd - 8 + DIFF_FACTOR_LOG2);
if (mask_type == DIFFWTD_38_INV) {
for (int i = 0; i < h; ++i) {
for (int j = 0; j < w; j += 16) {
__m256i s0 = _mm256_loadu_si256((const __m256i *)&ssrc0[j]);
__m256i s1 = _mm256_loadu_si256((const __m256i *)&ssrc1[j]);
__m256i diff = _mm256_sra_epi16(
_mm256_abs_epi16(_mm256_sub_epi16(s0, s1)), xshift);
__m256i m = _mm256_min_epi16(
_mm256_max_epi16(y0, _mm256_add_epi16(diff, ymask_base)),
yAOM_BLEND_A64_MAX_ALPHA);
m = _mm256_sub_epi16(yAOM_BLEND_A64_MAX_ALPHA, m);
m = _mm256_packus_epi16(m, m);
m = _mm256_permute4x64_epi64(m, _MM_SHUFFLE(0, 0, 2, 0));
__m128i m0 = _mm256_castsi256_si128(m);
_mm_storeu_si128((__m128i *)&mask[j], m0);
}
ssrc0 += src0_stride;
ssrc1 += src1_stride;
mask += w;
}
} else {
for (int i = 0; i < h; ++i) {
for (int j = 0; j < w; j += 16) {
__m256i s0 = _mm256_loadu_si256((const __m256i *)&ssrc0[j]);
__m256i s1 = _mm256_loadu_si256((const __m256i *)&ssrc1[j]);
__m256i diff = _mm256_sra_epi16(
_mm256_abs_epi16(_mm256_sub_epi16(s0, s1)), xshift);
__m256i m = _mm256_min_epi16(
_mm256_max_epi16(y0, _mm256_add_epi16(diff, ymask_base)),
yAOM_BLEND_A64_MAX_ALPHA);
m = _mm256_packus_epi16(m, m);
m = _mm256_permute4x64_epi64(m, _MM_SHUFFLE(0, 0, 2, 0));
__m128i m0 = _mm256_castsi256_si128(m);
_mm_storeu_si128((__m128i *)&mask[j], m0);
}
ssrc0 += src0_stride;
ssrc1 += src1_stride;
mask += w;
}
}
}
}
}
void extern
avx2_test (void)
{
x = _mm256_add_epi16 (x, x);
}
