static inline void hyperloglog_count_avx2(const uint8_t *registers,
uint32_t n_registers,
float *inverse_sum, uint32_t *n_zeros)
{
const __m256i ones = (__m256i)_mm256_set1_ps(1.0f);
__m256 agg = _mm256_set1_ps(0.0f);
for (size_t i = 0; i < n_registers / sizeof(__m256i); ++i) {
const __m256i simd = _mm256_load_si256((__m256i *)registers + i);
/* For some reason, VPSRLDQ works on lane of 128bits instead of 256. */
const __m128i low = _mm256_extracti128_si256(simd, 0);
const __m128i high = _mm256_extracti128_si256(simd, 1);
__m256i sums = inverse_power_avx2(low);
agg = _mm256_add_ps(agg, (__m256)sums);
sums = inverse_power_avx2(_mm_srli_si128(low, 8));
agg = _mm256_add_ps(agg, (__m256)sums);
sums = inverse_power_avx2(high);
agg = _mm256_add_ps(agg, (__m256)sums);
sums = inverse_power_avx2(_mm_srli_si128(high, 8));
agg = _mm256_add_ps(agg, (__m256)sums);
*n_zeros += _mm256_cntz_epi8(simd);
}
*inverse_sum = horizontal_sum_avx2(agg);
}
int32_t avx2_sumsignedbytes(int8_t* array, size_t size) {
__m256i accumulator = _mm256_setzero_si256();
for (size_t i=0; i < size; i += 32) {
const __m256i v = _mm256_loadu_si256((__m256i*)(array + i));
const __m128i lo = _mm256_extracti128_si256(v, 0);
const __m128i hi = _mm256_extracti128_si256(v, 1);
const __m256i t0 = _mm256_cvtepi8_epi32(lo);
const __m256i t1 = _mm256_cvtepi8_epi32(hi);
const __m256i t2 = _mm256_cvtepi8_epi32(_mm_bsrli_si128(lo, 8));
const __m256i t3 = _mm256_cvtepi8_epi32(_mm_bsrli_si128(hi, 8));
accumulator = _mm256_add_epi32(accumulator, t0);
accumulator = _mm256_add_epi32(accumulator, t1);
accumulator = _mm256_add_epi32(accumulator, t2);
accumulator = _mm256_add_epi32(accumulator, t3);
}
return int32_t(_mm256_extract_epi32(accumulator, 0)) +
int32_t(_mm256_extract_epi32(accumulator, 1)) +
int32_t(_mm256_extract_epi32(accumulator, 2)) +
int32_t(_mm256_extract_epi32(accumulator, 3)) +
int32_t(_mm256_extract_epi32(accumulator, 4)) +
int32_t(_mm256_extract_epi32(accumulator, 5)) +
int32_t(_mm256_extract_epi32(accumulator, 6)) +
int32_t(_mm256_extract_epi32(accumulator, 7));
}
// credit: Harold Aptroot
uint32_t maskedvectorsum(uint32_t * z, uint32_t N, uint32_t * accesses,
uint32_t nmbr) {
__m256i Nvec = _mm256_set1_epi32(N - 1);
__m256i sum = _mm256_setzero_si256();
for(uint32_t j = 0; j < nmbr ; j += 8) {
__m256i indexes = _mm256_loadu_si256((__m256i*)(accesses + j));
indexes = _mm256_and_si256(indexes, Nvec);
__m256i fi = _mm256_i32gather_epi32((int*)z, indexes, 4);
sum = _mm256_add_epi32(sum, fi);
}
__m128i sum128 = _mm_add_epi32(_mm256_extracti128_si256(sum, 0), _mm256_extracti128_si256(sum, 1));
sum128 = _mm_hadd_epi32(sum128, sum128);
return _mm_extract_epi32(sum128, 0) + _mm_extract_epi32(sum128, 1);
}
static unsigned int sad_w64_avg_avx2(const uint8_t *src_ptr, int src_stride,
const uint8_t *ref_ptr, int ref_stride,
const int h, const uint8_t *second_pred,
const int second_pred_stride) {
int i, res;
__m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg;
__m256i sum_sad = _mm256_setzero_si256();
__m256i sum_sad_h;
__m128i sum_sad128;
for (i = 0; i < h; i++) {
ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr);
ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + 32));
ref1_reg = _mm256_avg_epu8(
ref1_reg, _mm256_loadu_si256((__m256i const *)second_pred));
ref2_reg = _mm256_avg_epu8(
ref2_reg, _mm256_loadu_si256((__m256i const *)(second_pred + 32)));
sad1_reg =
_mm256_sad_epu8(ref1_reg, _mm256_loadu_si256((__m256i const *)src_ptr));
sad2_reg = _mm256_sad_epu8(
ref2_reg, _mm256_loadu_si256((__m256i const *)(src_ptr + 32)));
sum_sad = _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg));
ref_ptr += ref_stride;
src_ptr += src_stride;
second_pred += second_pred_stride;
}
sum_sad_h = _mm256_srli_si256(sum_sad, 8);
sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h);
sum_sad128 = _mm256_extracti128_si256(sum_sad, 1);
sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128);
res = _mm_cvtsi128_si32(sum_sad128);
return res;
}
INLINE static void sum_block_dual_avx2(__m256i *ver_row, unsigned *sum0, unsigned *sum1)
{
__m256i sad = _mm256_setzero_si256();
haddwd_accumulate_dual_avx2(&sad, ver_row + 0);
haddwd_accumulate_dual_avx2(&sad, ver_row + 1);
haddwd_accumulate_dual_avx2(&sad, ver_row + 2);
haddwd_accumulate_dual_avx2(&sad, ver_row + 3);
haddwd_accumulate_dual_avx2(&sad, ver_row + 4);
haddwd_accumulate_dual_avx2(&sad, ver_row + 5);
haddwd_accumulate_dual_avx2(&sad, ver_row + 6);
haddwd_accumulate_dual_avx2(&sad, ver_row + 7);
sad = _mm256_add_epi32(sad, _mm256_shuffle_epi32(sad, KVZ_PERMUTE(2, 3, 0, 1)));
sad = _mm256_add_epi32(sad, _mm256_shuffle_epi32(sad, KVZ_PERMUTE(1, 0, 1, 0)));
*sum0 = _mm_cvtsi128_si32(_mm256_extracti128_si256(sad, 0));
*sum1 = _mm_cvtsi128_si32(_mm256_extracti128_si256(sad, 1));
}
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;
}
static inline void do_encode_12bytes(const char (*alphabet)[2], char *out, __m256i chunk)
{
const __m256i shufflemask = _mm256_set_epi8(
-1, 9, 10, 11,
-1, 9, 10, 11,
-1, 6, 7, 8,
-1, 6, 7, 8,
-1, 3, 4, 5,
-1, 3, 4, 5,
-1, 0, 1, 2,
-1, 0, 1, 2
);
const __m256i shifts = _mm256_set_epi32(0, 12, 0, 12, 0, 12, 0, 12);
const __m256i masks = _mm256_set1_epi32(4095);
// convert from big endian and rearrange the bytes
chunk = _mm256_shuffle_epi8(chunk, shufflemask);
chunk = _mm256_srlv_epi32(chunk, shifts);
chunk = _mm256_and_si256(chunk, masks);
// write the two halves to memory
do_encode_6bytes(alphabet, out + 0, _mm256_extracti128_si256(chunk, 0));
do_encode_6bytes(alphabet, out + 8, _mm256_extracti128_si256(chunk, 1));
}
static unsigned int sad32x32(const uint8_t *src_ptr, int src_stride,
const uint8_t *ref_ptr, int ref_stride) {
__m256i s1, s2, r1, r2;
__m256i sum = _mm256_setzero_si256();
__m128i sum_i128;
int i;
for (i = 0; i < 16; ++i) {
r1 = _mm256_loadu_si256((__m256i const *)ref_ptr);
r2 = _mm256_loadu_si256((__m256i const *)(ref_ptr + ref_stride));
s1 = _mm256_sad_epu8(r1, _mm256_loadu_si256((__m256i const *)src_ptr));
s2 = _mm256_sad_epu8(
r2, _mm256_loadu_si256((__m256i const *)(src_ptr + src_stride)));
sum = _mm256_add_epi32(sum, _mm256_add_epi32(s1, s2));
ref_ptr += ref_stride << 1;
src_ptr += src_stride << 1;
}
sum = _mm256_add_epi32(sum, _mm256_srli_si256(sum, 8));
sum_i128 = _mm_add_epi32(_mm256_extracti128_si256(sum, 1),
_mm256_castsi256_si128(sum));
return _mm_cvtsi128_si32(sum_i128);
}
void calculate_fma_double (unsigned char * out, double X0, double Y0, double scale, unsigned YSTART, unsigned SX, unsigned SY)
{
__m256d dd = _mm256_set1_pd (scale);
__m256d XX0 = _mm256_set1_pd (X0);
for (unsigned j = YSTART; j < SY; j++) {
__m256d y0 = _mm256_set1_pd (j*scale + Y0);
for (unsigned i = 0; i < SX; i += 4) {
__m128i ind = _mm_setr_epi32 (i, i + 1, i + 2, i + 3);
__m256d x0 = _mm256_fmadd_pd (dd, _mm256_cvtepi32_pd (ind), XX0);
__m256d x = x0;
__m256d y = y0;
__m256i counts = _mm256_setzero_si256 ();
__m256i cmp_mask = _mm256_set1_epi32 (0xFFFFFFFFu);
for (unsigned n = 0; n < 255; n++) {
__m256d x2 = _mm256_mul_pd (x, x);
__m256d y2 = _mm256_mul_pd (y, y);
__m256d abs = _mm256_add_pd (x2, y2);
__m256i cmp = _mm256_castpd_si256 (_mm256_cmp_pd (abs, _mm256_set1_pd (4), 1));
cmp_mask = _mm256_and_si256 (cmp_mask, cmp);
if (_mm256_testz_si256 (cmp_mask, cmp_mask)) {
break;
}
counts = _mm256_sub_epi64 (counts, cmp_mask);
__m256d t = _mm256_add_pd (x, x);
y = _mm256_fmadd_pd (t, y, y0);
x = _mm256_add_pd (_mm256_sub_pd (x2, y2), x0);
}
__m256i result = _mm256_shuffle_epi8 (counts, _mm256_setr_epi8 (0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8));
*(uint32_t*) out = _mm_extract_epi16 (_mm256_extracti128_si256 (result, 0), 0) | (_mm_extract_epi16 (_mm256_extracti128_si256 (result, 1), 0) << 16);
out += 4;
}
}
}
__m128i test1_mm256_extracti128_si256_1(__m256i a) {
// CHECK-LABEL: test1_mm256_extracti128_si256
// CHECK: shufflevector <4 x i64> %{{.*}}, <4 x i64> %{{.*}}, <2 x i32> <i32 2, i32 3>
return _mm256_extracti128_si256(a, 1);
}
void av1_build_compound_diffwtd_mask_avx2(uint8_t *mask,
DIFFWTD_MASK_TYPE mask_type,
const uint8_t *src0, int stride0,
const uint8_t *src1, int stride1,
int h, int w) {
const int mb = (mask_type == DIFFWTD_38_INV) ? AOM_BLEND_A64_MAX_ALPHA : 0;
const __m256i y_mask_base = _mm256_set1_epi16(38 - mb);
int i = 0;
if (4 == w) {
do {
const __m128i s0A = xx_loadl_32(src0);
const __m128i s0B = xx_loadl_32(src0 + stride0);
const __m128i s0C = xx_loadl_32(src0 + stride0 * 2);
const __m128i s0D = xx_loadl_32(src0 + stride0 * 3);
const __m128i s0AB = _mm_unpacklo_epi32(s0A, s0B);
const __m128i s0CD = _mm_unpacklo_epi32(s0C, s0D);
const __m128i s0ABCD = _mm_unpacklo_epi64(s0AB, s0CD);
const __m256i s0ABCD_w = _mm256_cvtepu8_epi16(s0ABCD);
const __m128i s1A = xx_loadl_32(src1);
const __m128i s1B = xx_loadl_32(src1 + stride1);
const __m128i s1C = xx_loadl_32(src1 + stride1 * 2);
const __m128i s1D = xx_loadl_32(src1 + stride1 * 3);
const __m128i s1AB = _mm_unpacklo_epi32(s1A, s1B);
const __m128i s1CD = _mm_unpacklo_epi32(s1C, s1D);
const __m128i s1ABCD = _mm_unpacklo_epi64(s1AB, s1CD);
const __m256i s1ABCD_w = _mm256_cvtepu8_epi16(s1ABCD);
const __m256i m16 = calc_mask_avx2(y_mask_base, s0ABCD_w, s1ABCD_w);
const __m256i m8 = _mm256_packus_epi16(m16, _mm256_setzero_si256());
const __m128i x_m8 =
_mm256_castsi256_si128(_mm256_permute4x64_epi64(m8, 0xd8));
xx_storeu_128(mask, x_m8);
src0 += (stride0 << 2);
src1 += (stride1 << 2);
mask += 16;
i += 4;
} while (i < h);
} else if (8 == w) {
do {
const __m128i s0A = xx_loadl_64(src0);
const __m128i s0B = xx_loadl_64(src0 + stride0);
const __m128i s0C = xx_loadl_64(src0 + stride0 * 2);
const __m128i s0D = xx_loadl_64(src0 + stride0 * 3);
const __m256i s0AC_w = _mm256_cvtepu8_epi16(_mm_unpacklo_epi64(s0A, s0C));
const __m256i s0BD_w = _mm256_cvtepu8_epi16(_mm_unpacklo_epi64(s0B, s0D));
const __m128i s1A = xx_loadl_64(src1);
const __m128i s1B = xx_loadl_64(src1 + stride1);
const __m128i s1C = xx_loadl_64(src1 + stride1 * 2);
const __m128i s1D = xx_loadl_64(src1 + stride1 * 3);
const __m256i s1AB_w = _mm256_cvtepu8_epi16(_mm_unpacklo_epi64(s1A, s1C));
const __m256i s1CD_w = _mm256_cvtepu8_epi16(_mm_unpacklo_epi64(s1B, s1D));
const __m256i m16AC = calc_mask_avx2(y_mask_base, s0AC_w, s1AB_w);
const __m256i m16BD = calc_mask_avx2(y_mask_base, s0BD_w, s1CD_w);
const __m256i m8 = _mm256_packus_epi16(m16AC, m16BD);
yy_storeu_256(mask, m8);
src0 += stride0 << 2;
src1 += stride1 << 2;
mask += 32;
i += 4;
} while (i < h);
} else if (16 == w) {
do {
const __m128i s0A = xx_load_128(src0);
const __m128i s0B = xx_load_128(src0 + stride0);
const __m128i s1A = xx_load_128(src1);
const __m128i s1B = xx_load_128(src1 + stride1);
const __m256i s0AL = _mm256_cvtepu8_epi16(s0A);
const __m256i s0BL = _mm256_cvtepu8_epi16(s0B);
const __m256i s1AL = _mm256_cvtepu8_epi16(s1A);
const __m256i s1BL = _mm256_cvtepu8_epi16(s1B);
const __m256i m16AL = calc_mask_avx2(y_mask_base, s0AL, s1AL);
const __m256i m16BL = calc_mask_avx2(y_mask_base, s0BL, s1BL);
const __m256i m8 =
_mm256_permute4x64_epi64(_mm256_packus_epi16(m16AL, m16BL), 0xd8);
yy_storeu_256(mask, m8);
src0 += stride0 << 1;
src1 += stride1 << 1;
mask += 32;
i += 2;
} while (i < h);
} else {
do {
int j = 0;
do {
const __m256i s0 = yy_loadu_256(src0 + j);
const __m256i s1 = yy_loadu_256(src1 + j);
const __m256i s0L = _mm256_cvtepu8_epi16(_mm256_castsi256_si128(s0));
const __m256i s1L = _mm256_cvtepu8_epi16(_mm256_castsi256_si128(s1));
const __m256i s0H =
_mm256_cvtepu8_epi16(_mm256_extracti128_si256(s0, 1));
const __m256i s1H =
_mm256_cvtepu8_epi16(_mm256_extracti128_si256(s1, 1));
const __m256i m16L = calc_mask_avx2(y_mask_base, s0L, s1L);
const __m256i m16H = calc_mask_avx2(y_mask_base, s0H, s1H);
const __m256i m8 =
_mm256_permute4x64_epi64(_mm256_packus_epi16(m16L, m16H), 0xd8);
yy_storeu_256(mask + j, m8);
j += 32;
} while (j < w);
src0 += stride0;
src1 += stride1;
mask += w;
i += 1;
} while (i < h);
}
}
void Viterbi::AlignWithOutCellOff(HMMSimd* q, HMMSimd* t,ViterbiMatrix * viterbiMatrix,
int maxres, ViterbiResult* result)
#endif
#endif
{
// Linear topology of query (and template) HMM:
// 1. The HMM HMM has L+2 columns. Columns 1 to L contain
// a match state, a delete state and an insert state each.
// 2. The Start state is M0, the virtual match state in column i=0 (j=0). (Therefore X[k][0]=ANY)
// This column has only a match state and it has only a transitions to the next match state.
// 3. The End state is M(L+1), the virtual match state in column i=L+1.(j=L+1) (Therefore X[k][L+1]=ANY)
// Column L has no transitions to the delete state: tr[L][M2D]=tr[L][D2D]=0.
// 4. Transitions I->D and D->I are ignored, since they do not appear in PsiBlast alignments
// (as long as the gap opening penalty d is higher than the best match score S(a,b)).
// Pairwise alignment of two HMMs:
// 1. Pair-states for the alignment of two HMMs are
// MM (Q:Match T:Match) , GD (Q:Gap T:Delete), IM (Q:Insert T:Match), DG (Q:Delelte, T:Match) , MI (Q:Match T:Insert)
// 2. Transitions are allowed only between the MM-state and each of the four other states.
// Saving space:
// The best score ending in pair state XY sXY[i][j] is calculated from left to right (j=1->t->L)
// and top to bottom (i=1->q->L). To save space, only the last row of scores calculated is kept in memory.
// (The backtracing matrices are kept entirely in memory [O(t->L*q->L)]).
// When the calculation has proceeded up to the point where the scores for cell (i,j) are caculated,
// sXY[i-1][j'] = sXY[j'] for j'>=j (A below)
// sXY[i][j'] = sXY[j'] for j'<j (B below)
// sXY[i-1][j-1]= sXY_i_1_j_1 (C below)
// sXY[i][j] = sXY_i_j (D below)
// j-1
// j
// i-1: CAAAAAAAAAAAAAAAAAA
// i : BBBBBBBBBBBBBD
// Variable declarations
const float smin = (this->local ? 0 : -FLT_MAX); //used to distinguish between SW and NW algorithms in maximization
const simd_float smin_vec = simdf32_set(smin);
const simd_float shift_vec = simdf32_set(shift);
// const simd_float one_vec = simdf32_set(1); // 00000001
const simd_int mm_vec = simdi32_set(2); //MM 00000010
const simd_int gd_vec = simdi32_set(3); //GD 00000011
const simd_int im_vec = simdi32_set(4); //IM 00000100
const simd_int dg_vec = simdi32_set(5); //DG 00000101
const simd_int mi_vec = simdi32_set(6); //MI 00000110
const simd_int gd_mm_vec = simdi32_set(8); // 00001000
const simd_int im_mm_vec = simdi32_set(16);// 00010000
const simd_int dg_mm_vec = simdi32_set(32);// 00100000
const simd_int mi_mm_vec = simdi32_set(64);// 01000000
#ifdef VITERBI_SS_SCORE
HMM * q_s = q->GetHMM(0);
const unsigned char * t_index;
if(ss_hmm_mode == HMM::PRED_PRED || ss_hmm_mode == HMM::DSSP_PRED ){
t_index = t->pred_index;
}else if(ss_hmm_mode == HMM::PRED_DSSP){
t_index = t->dssp_index;
}
simd_float * ss_score_vec = (simd_float *) ss_score;
#endif
#ifdef AVX2
const simd_int shuffle_mask_extract = _mm256_setr_epi8(0, 4, 8, 12, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1,
-1, -1, -1, -1, 0, 4, 8, 12, -1, -1, -1, -1, -1, -1, -1, -1);
#endif
#ifdef VITERBI_CELLOFF
const __m128i tmp_vec = _mm_set_epi32(0x40000000,0x00400000,0x00004000,0x00000040);//01000000010000000100000001000000
#ifdef AVX2
const simd_int co_vec = _mm256_inserti128_si256(_mm256_castsi128_si256(tmp_vec), tmp_vec, 1);
const simd_int float_min_vec = (simd_int) _mm256_set1_ps(-FLT_MAX);
const simd_int shuffle_mask_celloff = _mm256_set_epi8(
15, 14, 13, 12,
15, 14, 13, 12,
15, 14, 13, 12,
15, 14, 13, 12,
3, 2, 1, 0,
3, 2, 1, 0,
3, 2, 1, 0,
3, 2, 1, 0);
#else // SSE case
const simd_int co_vec = tmp_vec;
const simd_int float_min_vec = (simd_int) simdf32_set(-FLT_MAX);
#endif
#endif // AVX2 end
int i,j; //query and template match state indices
simd_int i2_vec = simdi32_set(0);
simd_int j2_vec = simdi32_set(0);
simd_float sMM_i_j = simdf32_set(0);
simd_float sMI_i_j,sIM_i_j,sGD_i_j,sDG_i_j;
simd_float Si_vec;
simd_float sMM_i_1_j_1;
simd_float sMI_i_1_j_1;
simd_float sIM_i_1_j_1;
simd_float sGD_i_1_j_1;
simd_float sDG_i_1_j_1;
simd_float score_vec = simdf32_set(-FLT_MAX);
simd_int byte_result_vec = simdi32_set(0);
// Initialization of top row, i.e. cells (0,j)
for (j=0; j <= t->L; ++j)
{
const unsigned int index_pos_j = j * 5;
sMM_DG_MI_GD_IM_vec[index_pos_j + 0] = simdf32_set(-j*penalty_gap_template);
sMM_DG_MI_GD_IM_vec[index_pos_j + 1] = simdf32_set(-FLT_MAX);
sMM_DG_MI_GD_IM_vec[index_pos_j + 2] = simdf32_set(-FLT_MAX);
sMM_DG_MI_GD_IM_vec[index_pos_j + 3] = simdf32_set(-FLT_MAX);
sMM_DG_MI_GD_IM_vec[index_pos_j + 4] = simdf32_set(-FLT_MAX);
}
// Viterbi algorithm
const int queryLength = q->L;
for (i=1; i <= queryLength; ++i) // Loop through query positions i
{
// If q is compared to t, exclude regions where overlap of q with t < min_overlap residues
// Initialize cells
sMM_i_1_j_1 = simdf32_set(-(i - 1) * penalty_gap_query); // initialize at (i-1,0)
sIM_i_1_j_1 = simdf32_set(-FLT_MAX); // initialize at (i-1,jmin-1)
sMI_i_1_j_1 = simdf32_set(-FLT_MAX);
sDG_i_1_j_1 = simdf32_set(-FLT_MAX);
sGD_i_1_j_1 = simdf32_set(-FLT_MAX);
// initialize at (i,jmin-1)
const unsigned int index_pos_i = 0 * 5;
sMM_DG_MI_GD_IM_vec[index_pos_i + 0] = simdf32_set(-i * penalty_gap_query); // initialize at (i,0)
sMM_DG_MI_GD_IM_vec[index_pos_i + 1] = simdf32_set(-FLT_MAX);
sMM_DG_MI_GD_IM_vec[index_pos_i + 2] = simdf32_set(-FLT_MAX);
sMM_DG_MI_GD_IM_vec[index_pos_i + 3] = simdf32_set(-FLT_MAX);
sMM_DG_MI_GD_IM_vec[index_pos_i + 4] = simdf32_set(-FLT_MAX);
#ifdef AVX2
unsigned long long * sCO_MI_DG_IM_GD_MM_vec = (unsigned long long *) viterbiMatrix->getRow(i);
#else
unsigned int *sCO_MI_DG_IM_GD_MM_vec = (unsigned int *) viterbiMatrix->getRow(i);
#endif
const unsigned int start_pos_tr_i_1 = (i - 1) * 7;
const unsigned int start_pos_tr_i = (i) * 7;
const simd_float q_m2m = simdf32_load((float *) (q->tr + start_pos_tr_i_1 + 2)); // M2M
const simd_float q_m2d = simdf32_load((float *) (q->tr + start_pos_tr_i_1 + 3)); // M2D
const simd_float q_d2m = simdf32_load((float *) (q->tr + start_pos_tr_i_1 + 4)); // D2M
const simd_float q_d2d = simdf32_load((float *) (q->tr + start_pos_tr_i_1 + 5)); // D2D
const simd_float q_i2m = simdf32_load((float *) (q->tr + start_pos_tr_i_1 + 6)); // I2m
const simd_float q_i2i = simdf32_load((float *) (q->tr + start_pos_tr_i)); // I2I
const simd_float q_m2i = simdf32_load((float *) (q->tr + start_pos_tr_i + 1)); // M2I
// Find maximum score; global alignment: maxize only over last row and last column
const bool findMaxInnerLoop = (local || i == queryLength);
const int targetLength = t->L;
#ifdef VITERBI_SS_SCORE
if(ss_hmm_mode == HMM::NO_SS_INFORMATION){
// set all to log(1.0) = 0.0
memset(ss_score, 0, (targetLength+1)*VECSIZE_FLOAT*sizeof(float));
}else {
const float * score;
if(ss_hmm_mode == HMM::PRED_PRED){
score = &S33[ (int)q_s->ss_pred[i]][ (int)q_s->ss_conf[i]][0][0];
}else if (ss_hmm_mode == HMM::DSSP_PRED){
score = &S73[ (int)q_s->ss_dssp[i]][0][0];
}else{
score = &S37[ (int)q_s->ss_pred[i]][ (int)q_s->ss_conf[i]][0];
}
// access SS scores and write them to the ss_score array
for (j = 0; j <= (targetLength*VECSIZE_FLOAT); j++) // Loop through template positions j
{
ss_score[j] = ssw * score[t_index[j]];
}
}
#endif
for (j=1; j <= targetLength; ++j) // Loop through template positions j
{
simd_int index_vec;
simd_int res_gt_vec;
// cache line optimized reading
const unsigned int start_pos_tr_j_1 = (j-1) * 7;
const unsigned int start_pos_tr_j = (j) * 7;
const simd_float t_m2m = simdf32_load((float *) (t->tr+start_pos_tr_j_1+2)); // M2M
const simd_float t_m2d = simdf32_load((float *) (t->tr+start_pos_tr_j_1+3)); // M2D
const simd_float t_d2m = simdf32_load((float *) (t->tr+start_pos_tr_j_1+4)); // D2M
const simd_float t_d2d = simdf32_load((float *) (t->tr+start_pos_tr_j_1+5)); // D2D
const simd_float t_i2m = simdf32_load((float *) (t->tr+start_pos_tr_j_1+6)); // I2m
const simd_float t_i2i = simdf32_load((float *) (t->tr+start_pos_tr_j)); // I2i
const simd_float t_m2i = simdf32_load((float *) (t->tr+start_pos_tr_j+1)); // M2I
// Find max value
// CALCULATE_MAX6( sMM_i_j,
// smin,
// sMM_i_1_j_1 + q->tr[i-1][M2M] + t->tr[j-1][M2M],
// sGD_i_1_j_1 + q->tr[i-1][M2M] + t->tr[j-1][D2M],
// sIM_i_1_j_1 + q->tr[i-1][I2M] + t->tr[j-1][M2M],
// sDG_i_1_j_1 + q->tr[i-1][D2M] + t->tr[j-1][M2M],
// sMI_i_1_j_1 + q->tr[i-1][M2M] + t->tr[j-1][I2M],
// bMM[i][j]
// );
// same as sMM_i_1_j_1 + q->tr[i-1][M2M] + t->tr[j-1][M2M]
simd_float mm_m2m_m2m_vec = simdf32_add( simdf32_add(sMM_i_1_j_1, q_m2m), t_m2m);
// if mm > min { 2 }
res_gt_vec = (simd_int)simdf32_gt(mm_m2m_m2m_vec, smin_vec);
byte_result_vec = simdi_and(res_gt_vec, mm_vec);
sMM_i_j = simdf32_max(smin_vec, mm_m2m_m2m_vec);
// same as sGD_i_1_j_1 + q->tr[i-1][M2M] + t->tr[j-1][D2M]
simd_float gd_m2m_d2m_vec = simdf32_add( simdf32_add(sGD_i_1_j_1, q_m2m), t_d2m);
// if gd > max { 3 }
res_gt_vec = (simd_int)simdf32_gt(gd_m2m_d2m_vec, sMM_i_j);
index_vec = simdi_and( res_gt_vec, gd_vec);
byte_result_vec = simdi_or( index_vec, byte_result_vec);
sMM_i_j = simdf32_max(sMM_i_j, gd_m2m_d2m_vec);
// same as sIM_i_1_j_1 + q->tr[i-1][I2M] + t->tr[j-1][M2M]
simd_float im_m2m_d2m_vec = simdf32_add( simdf32_add(sIM_i_1_j_1, q_i2m), t_m2m);
// if im > max { 4 }
MAX2(im_m2m_d2m_vec, sMM_i_j, im_vec,byte_result_vec);
sMM_i_j = simdf32_max(sMM_i_j, im_m2m_d2m_vec);
// same as sDG_i_1_j_1 + q->tr[i-1][D2M] + t->tr[j-1][M2M]
simd_float dg_m2m_d2m_vec = simdf32_add( simdf32_add(sDG_i_1_j_1, q_d2m), t_m2m);
// if dg > max { 5 }
MAX2(dg_m2m_d2m_vec, sMM_i_j, dg_vec,byte_result_vec);
sMM_i_j = simdf32_max(sMM_i_j, dg_m2m_d2m_vec);
// same as sMI_i_1_j_1 + q->tr[i-1][M2M] + t->tr[j-1][I2M],
simd_float mi_m2m_d2m_vec = simdf32_add( simdf32_add(sMI_i_1_j_1, q_m2m), t_i2m);
// if mi > max { 6 }
MAX2(mi_m2m_d2m_vec, sMM_i_j, mi_vec, byte_result_vec);
sMM_i_j = simdf32_max(sMM_i_j, mi_m2m_d2m_vec);
// TODO add secondary structure score
// calculate amino acid profile-profile scores
Si_vec = log2f4(ScalarProd20Vec((simd_float *) q->p[i],(simd_float *) t->p[j]));
#ifdef VITERBI_SS_SCORE
Si_vec = simdf32_add(ss_score_vec[j], Si_vec);
#endif
Si_vec = simdf32_add(Si_vec, shift_vec);
sMM_i_j = simdf32_add(sMM_i_j, Si_vec);
//+ ScoreSS(q,t,i,j) + shift + (Sstruc==NULL? 0: Sstruc[i][j]);
const unsigned int index_pos_j = (j * 5);
const unsigned int index_pos_j_1 = (j - 1) * 5;
const simd_float sMM_j_1 = simdf32_load((float *) (sMM_DG_MI_GD_IM_vec + index_pos_j_1 + 0));
const simd_float sGD_j_1 = simdf32_load((float *) (sMM_DG_MI_GD_IM_vec + index_pos_j_1 + 3));
const simd_float sIM_j_1 = simdf32_load((float *) (sMM_DG_MI_GD_IM_vec + index_pos_j_1 + 4));
const simd_float sMM_j = simdf32_load((float *) (sMM_DG_MI_GD_IM_vec + index_pos_j + 0));
const simd_float sDG_j = simdf32_load((float *) (sMM_DG_MI_GD_IM_vec + index_pos_j + 1));
const simd_float sMI_j = simdf32_load((float *) (sMM_DG_MI_GD_IM_vec + index_pos_j + 2));
sMM_i_1_j_1 = simdf32_load((float *)(sMM_DG_MI_GD_IM_vec + index_pos_j + 0));
sDG_i_1_j_1 = simdf32_load((float *)(sMM_DG_MI_GD_IM_vec + index_pos_j + 1));
sMI_i_1_j_1 = simdf32_load((float *)(sMM_DG_MI_GD_IM_vec + index_pos_j + 2));
sGD_i_1_j_1 = simdf32_load((float *)(sMM_DG_MI_GD_IM_vec + index_pos_j + 3));
sIM_i_1_j_1 = simdf32_load((float *)(sMM_DG_MI_GD_IM_vec + index_pos_j + 4));
// sGD_i_j = max2
// (
// sMM[j-1] + t->tr[j-1][M2D], // MM->GD gap opening in query
// sGD[j-1] + t->tr[j-1][D2D], // GD->GD gap extension in query
// bGD[i][j]
// );
//sMM_DG_GD_MI_IM_vec
simd_float mm_gd_vec = simdf32_add(sMM_j_1, t_m2d); // MM->GD gap opening in query
simd_float gd_gd_vec = simdf32_add(sGD_j_1, t_d2d); // GD->GD gap extension in query
// if mm_gd > gd_dg { 8 }
MAX2_SET_MASK(mm_gd_vec, gd_gd_vec,gd_mm_vec, byte_result_vec);
sGD_i_j = simdf32_max(
mm_gd_vec,
gd_gd_vec
);
// sIM_i_j = max2
// (
// sMM[j-1] + q->tr[i][M2I] + t->tr[j-1][M2M] ,
// sIM[j-1] + q->tr[i][I2I] + t->tr[j-1][M2M], // IM->IM gap extension in query
// bIM[i][j]
// );
simd_float mm_mm_vec = simdf32_add(simdf32_add(sMM_j_1, q_m2i), t_m2m);
simd_float im_im_vec = simdf32_add(simdf32_add(sIM_j_1, q_i2i), t_m2m); // IM->IM gap extension in query
// if mm_mm > im_im { 16 }
MAX2_SET_MASK(mm_mm_vec,im_im_vec, im_mm_vec, byte_result_vec);
sIM_i_j = simdf32_max(
mm_mm_vec,
im_im_vec
);
// sDG_i_j = max2
// (
// sMM[j] + q->tr[i-1][M2D],
// sDG[j] + q->tr[i-1][D2D], //gap extension (DD) in query
// bDG[i][j]
// );
simd_float mm_dg_vec = simdf32_add(sMM_j, q_m2d);
simd_float dg_dg_vec = simdf32_add(sDG_j, q_d2d); //gap extension (DD) in query
// if mm_dg > dg_dg { 32 }
MAX2_SET_MASK(mm_dg_vec,dg_dg_vec, dg_mm_vec, byte_result_vec);
sDG_i_j = simdf32_max( mm_dg_vec
,
dg_dg_vec
);
// sMI_i_j = max2
// (
// sMM[j] + q->tr[i-1][M2M] + t->tr[j][M2I], // MM->MI gap opening M2I in template
// sMI[j] + q->tr[i-1][M2M] + t->tr[j][I2I], // MI->MI gap extension I2I in template
// bMI[i][j]
// );
simd_float mm_mi_vec = simdf32_add( simdf32_add(sMM_j, q_m2m), t_m2i); // MM->MI gap opening M2I in template
simd_float mi_mi_vec = simdf32_add( simdf32_add(sMI_j, q_m2m), t_i2i); // MI->MI gap extension I2I in template
// if mm_mi > mi_mi { 64 }
MAX2_SET_MASK(mm_mi_vec, mi_mi_vec,mi_mm_vec, byte_result_vec);
sMI_i_j = simdf32_max(
mm_mi_vec,
mi_mi_vec
);
// Cell of logic
// if (cell_off[i][j])
//shift 10000000100000001000000010000000 -> 01000000010000000100000001000000
//because 10000000000000000000000000000000 = -2147483648 kills cmplt
#ifdef VITERBI_CELLOFF
#ifdef AVX2
simd_int matrix_vec = _mm256_set1_epi64x(sCO_MI_DG_IM_GD_MM_vec[j]>>1);
matrix_vec = _mm256_shuffle_epi8(matrix_vec,shuffle_mask_celloff);
#else
// if(((sCO_MI_DG_IM_GD_MM_vec[j] >>1) & 0x40404040) > 0){
// std::cout << ((sCO_MI_DG_IM_GD_MM_vec[j] >>1) & 0x40404040 ) << std::endl;
// }
simd_int matrix_vec = simdi32_set(sCO_MI_DG_IM_GD_MM_vec[j]>>1);
#endif
simd_int cell_off_vec = simdi_and(matrix_vec, co_vec);
simd_int res_eq_co_vec = simdi32_gt(co_vec, cell_off_vec ); // shift is because signed can't be checked here
simd_float cell_off_float_min_vec = (simd_float) simdi_andnot(res_eq_co_vec, float_min_vec); // inverse
sMM_i_j = simdf32_add(sMM_i_j,cell_off_float_min_vec); // add the cell off vec to sMM_i_j. Set -FLT_MAX to cell off
sGD_i_j = simdf32_add(sGD_i_j,cell_off_float_min_vec);
sIM_i_j = simdf32_add(sIM_i_j,cell_off_float_min_vec);
sDG_i_j = simdf32_add(sDG_i_j,cell_off_float_min_vec);
sMI_i_j = simdf32_add(sMI_i_j,cell_off_float_min_vec);
#endif
simdf32_store((float *)(sMM_DG_MI_GD_IM_vec+index_pos_j + 0), sMM_i_j);
simdf32_store((float *)(sMM_DG_MI_GD_IM_vec+index_pos_j + 1), sDG_i_j);
simdf32_store((float *)(sMM_DG_MI_GD_IM_vec+index_pos_j + 2), sMI_i_j);
simdf32_store((float *)(sMM_DG_MI_GD_IM_vec+index_pos_j + 3), sGD_i_j);
simdf32_store((float *)(sMM_DG_MI_GD_IM_vec+index_pos_j + 4), sIM_i_j);
// write values back to ViterbiMatrix
#ifdef AVX2
/* byte_result_vec 000H 000G 000F 000E 000D 000C 000B 000A */
/* abcdefgh 0000 0000 HGFE 0000 0000 0000 0000 DCBA */
const __m256i abcdefgh = _mm256_shuffle_epi8(byte_result_vec, shuffle_mask_extract);
/* abcd 0000 0000 0000 DCBA */
const __m128i abcd = _mm256_castsi256_si128(abcdefgh);
/* efgh 0000 0000 HGFE 0000 */
const __m128i efgh = _mm256_extracti128_si256(abcdefgh, 1);
_mm_storel_epi64((__m128i*)&sCO_MI_DG_IM_GD_MM_vec[j], _mm_or_si128(abcd, efgh));
#else
byte_result_vec = _mm_packs_epi32(byte_result_vec, byte_result_vec);
byte_result_vec = _mm_packus_epi16(byte_result_vec, byte_result_vec);
int int_result = _mm_cvtsi128_si32(byte_result_vec);
sCO_MI_DG_IM_GD_MM_vec[j] = int_result;
#endif
// Find maximum score; global alignment: maxize only over last row and last column
// if(sMM_i_j>score && (par.loc || i==q->L)) { i2=i; j2=j; score=sMM_i_j; }
if (findMaxInnerLoop){
// new score is higer
// output
// 0 0 0 MAX
simd_int lookup_mask_hi = (simd_int) simdf32_gt(sMM_i_j,score_vec);
// old score is higher
// output
// MAX MAX MAX 0
simd_int lookup_mask_lo = (simd_int) simdf32_lt(sMM_i_j,score_vec);
simd_int curr_pos_j = simdi32_set(j);
simd_int new_j_pos_hi = simdi_and(lookup_mask_hi,curr_pos_j);
simd_int old_j_pos_lo = simdi_and(lookup_mask_lo,j2_vec);
j2_vec = simdi32_add(new_j_pos_hi,old_j_pos_lo);
simd_int curr_pos_i = simdi32_set(i);
simd_int new_i_pos_hi = simdi_and(lookup_mask_hi,curr_pos_i);
simd_int old_i_pos_lo = simdi_and(lookup_mask_lo,i2_vec);
i2_vec = simdi32_add(new_i_pos_hi,old_i_pos_lo);
score_vec=simdf32_max(sMM_i_j,score_vec);
}
} //end for j
// if global alignment: look for best cell in last column
if (!local){
// new score is higer
// output
// 0 0 0 MAX
simd_int lookup_mask_hi = (simd_int) simdf32_gt(sMM_i_j,score_vec);
// old score is higher
// output
// MAX MAX MAX 0
simd_int lookup_mask_lo = (simd_int) simdf32_lt(sMM_i_j,score_vec);
simd_int curr_pos_j = simdi32_set(j);
simd_int new_j_pos_hi = simdi_and(lookup_mask_hi,curr_pos_j);
simd_int old_j_pos_lo = simdi_and(lookup_mask_lo,j2_vec);
j2_vec = simdi32_add(new_j_pos_hi,old_j_pos_lo);
simd_int curr_pos_i = simdi32_set(i);
simd_int new_i_pos_hi = simdi_and(lookup_mask_hi,curr_pos_i);
simd_int old_i_pos_lo = simdi_and(lookup_mask_lo,i2_vec);
i2_vec = simdi32_add(new_i_pos_hi,old_i_pos_lo);
score_vec = simdf32_max(sMM_i_j,score_vec);
} // end for j
} // end for i
for(int seq_index=0; seq_index < maxres; seq_index++){
result->score[seq_index]=((float*)&score_vec)[seq_index];
result->i[seq_index] = ((int*)&i2_vec)[seq_index];
result->j[seq_index] = ((int*)&j2_vec)[seq_index];
// std::cout << seq_index << "\t" << result->score[seq_index] << "\t" << result->i[seq_index] <<"\t" << result->j[seq_index] << std::endl;
}
// printf("Template=%-12.12s i=%-4i j=%-4i score=%6.3f\n",t->name,i2,j2,score);
}
// Immediate should be truncated to one bit.
__m128i test_mm256_extracti128_si256_2(__m256i a) {
// CHECK-LABEL: @test_mm256_extracti128_si256_2
// CHECK: shufflevector{{.*}}<i32 0, i32 1>
return _mm256_extracti128_si256(a, 2);
}
/**
* \brief quantize transformed coefficents
*
*/
void kvz_quant_flat_avx2(const encoder_state_t * const state, coeff_t *coef, coeff_t *q_coef, int32_t width,
int32_t height, int8_t type, int8_t scan_idx, int8_t block_type)
{
const encoder_control_t * const encoder = state->encoder_control;
const uint32_t log2_block_size = kvz_g_convert_to_bit[width] + 2;
const uint32_t * const scan = kvz_g_sig_last_scan[scan_idx][log2_block_size - 1];
int32_t qp_scaled = kvz_get_scaled_qp(type, state->global->QP, (encoder->bitdepth - 8) * 6);
const uint32_t log2_tr_size = kvz_g_convert_to_bit[width] + 2;
const int32_t scalinglist_type = (block_type == CU_INTRA ? 0 : 3) + (int8_t)("\0\3\1\2"[type]);
const int32_t *quant_coeff = encoder->scaling_list.quant_coeff[log2_tr_size - 2][scalinglist_type][qp_scaled % 6];
const int32_t transform_shift = MAX_TR_DYNAMIC_RANGE - encoder->bitdepth - log2_tr_size; //!< Represents scaling through forward transform
const int32_t q_bits = QUANT_SHIFT + qp_scaled / 6 + transform_shift;
const int32_t add = ((state->global->slicetype == KVZ_SLICE_I) ? 171 : 85) << (q_bits - 9);
const int32_t q_bits8 = q_bits - 8;
assert(quant_coeff[0] <= (1 << 15) - 1 && quant_coeff[0] >= -(1 << 15)); //Assuming flat values to fit int16_t
uint32_t ac_sum = 0;
__m256i v_ac_sum = _mm256_setzero_si256();
__m256i v_quant_coeff = _mm256_set1_epi16(quant_coeff[0]);
for (int32_t n = 0; n < width * height; n += 16) {
__m256i v_level = _mm256_loadu_si256((__m256i*)&(coef[n]));
__m256i v_sign = _mm256_cmpgt_epi16(_mm256_setzero_si256(), v_level);
v_sign = _mm256_or_si256(v_sign, _mm256_set1_epi16(1));
v_level = _mm256_abs_epi16(v_level);
__m256i low_a = _mm256_unpacklo_epi16(v_level, _mm256_set1_epi16(0));
__m256i high_a = _mm256_unpackhi_epi16(v_level, _mm256_set1_epi16(0));
__m256i low_b = _mm256_unpacklo_epi16(v_quant_coeff, _mm256_set1_epi16(0));
__m256i high_b = _mm256_unpackhi_epi16(v_quant_coeff, _mm256_set1_epi16(0));
__m256i v_level32_a = _mm256_madd_epi16(low_a, low_b);
__m256i v_level32_b = _mm256_madd_epi16(high_a, high_b);
v_level32_a = _mm256_add_epi32(v_level32_a, _mm256_set1_epi32(add));
v_level32_b = _mm256_add_epi32(v_level32_b, _mm256_set1_epi32(add));
v_level32_a = _mm256_srai_epi32(v_level32_a, q_bits);
v_level32_b = _mm256_srai_epi32(v_level32_b, q_bits);
v_level = _mm256_packs_epi32(v_level32_a, v_level32_b);
v_level = _mm256_sign_epi16(v_level, v_sign);
_mm256_storeu_si256((__m256i*)&(q_coef[n]), v_level);
v_ac_sum = _mm256_add_epi32(v_ac_sum, v_level32_a);
v_ac_sum = _mm256_add_epi32(v_ac_sum, v_level32_b);
}
__m128i temp = _mm_add_epi32(_mm256_castsi256_si128(v_ac_sum), _mm256_extracti128_si256(v_ac_sum, 1));
temp = _mm_add_epi32(temp, _mm_shuffle_epi32(temp, KVZ_PERMUTE(2, 3, 0, 1)));
temp = _mm_add_epi32(temp, _mm_shuffle_epi32(temp, KVZ_PERMUTE(1, 0, 1, 0)));
ac_sum += _mm_cvtsi128_si32(temp);
if (!(encoder->sign_hiding && ac_sum >= 2)) return;
int32_t delta_u[LCU_WIDTH*LCU_WIDTH >> 2];
for (int32_t n = 0; n < width * height; n += 16) {
__m256i v_level = _mm256_loadu_si256((__m256i*)&(coef[n]));
v_level = _mm256_abs_epi16(v_level);
__m256i low_a = _mm256_unpacklo_epi16(v_level, _mm256_set1_epi16(0));
__m256i high_a = _mm256_unpackhi_epi16(v_level, _mm256_set1_epi16(0));
__m256i low_b = _mm256_unpacklo_epi16(v_quant_coeff, _mm256_set1_epi16(0));
__m256i high_b = _mm256_unpackhi_epi16(v_quant_coeff, _mm256_set1_epi16(0));
__m256i v_level32_a = _mm256_madd_epi16(low_a, low_b);
__m256i v_level32_b = _mm256_madd_epi16(high_a, high_b);
v_level32_a = _mm256_add_epi32(v_level32_a, _mm256_set1_epi32(add));
v_level32_b = _mm256_add_epi32(v_level32_b, _mm256_set1_epi32(add));
v_level32_a = _mm256_srai_epi32(v_level32_a, q_bits);
v_level32_b = _mm256_srai_epi32(v_level32_b, q_bits);
v_level = _mm256_packs_epi32(v_level32_a, v_level32_b);
__m256i v_coef = _mm256_loadu_si256((__m256i*)&(coef[n]));
__m256i v_coef_a = _mm256_unpacklo_epi16(_mm256_abs_epi16(v_coef), _mm256_set1_epi16(0));
__m256i v_coef_b = _mm256_unpackhi_epi16(_mm256_abs_epi16(v_coef), _mm256_set1_epi16(0));
__m256i v_quant_coeff_a = _mm256_unpacklo_epi16(v_quant_coeff, _mm256_set1_epi16(0));
__m256i v_quant_coeff_b = _mm256_unpackhi_epi16(v_quant_coeff, _mm256_set1_epi16(0));
v_coef_a = _mm256_madd_epi16(v_coef_a, v_quant_coeff_a);
v_coef_b = _mm256_madd_epi16(v_coef_b, v_quant_coeff_b);
v_coef_a = _mm256_sub_epi32(v_coef_a, _mm256_slli_epi32(_mm256_unpacklo_epi16(v_level, _mm256_set1_epi16(0)), q_bits) );
v_coef_b = _mm256_sub_epi32(v_coef_b, _mm256_slli_epi32(_mm256_unpackhi_epi16(v_level, _mm256_set1_epi16(0)), q_bits) );
v_coef_a = _mm256_srai_epi32(v_coef_a, q_bits8);
v_coef_b = _mm256_srai_epi32(v_coef_b, q_bits8);
_mm_storeu_si128((__m128i*)&(delta_u[n+0*4]), _mm256_castsi256_si128(v_coef_a));
_mm_storeu_si128((__m128i*)&(delta_u[n+2*4]), _mm256_extracti128_si256(v_coef_a, 1));
_mm_storeu_si128((__m128i*)&(delta_u[n+1*4]), _mm256_castsi256_si128(v_coef_b));
_mm_storeu_si128((__m128i*)&(delta_u[n+3*4]), _mm256_extracti128_si256(v_coef_b, 1));
}
if (ac_sum >= 2) {
#define SCAN_SET_SIZE 16
#define LOG2_SCAN_SET_SIZE 4
int32_t n, last_cg = -1, abssum = 0, subset, subpos;
for (subset = (width*height - 1) >> LOG2_SCAN_SET_SIZE; subset >= 0; subset--) {
int32_t first_nz_pos_in_cg = SCAN_SET_SIZE, last_nz_pos_in_cg = -1;
subpos = subset << LOG2_SCAN_SET_SIZE;
abssum = 0;
// Find last coeff pos
for (n = SCAN_SET_SIZE - 1; n >= 0; n--) {
if (q_coef[scan[n + subpos]]) {
last_nz_pos_in_cg = n;
break;
}
}
// First coeff pos
for (n = 0; n <SCAN_SET_SIZE; n++) {
if (q_coef[scan[n + subpos]]) {
first_nz_pos_in_cg = n;
break;
}
}
// Sum all kvz_quant coeffs between first and last
for (n = first_nz_pos_in_cg; n <= last_nz_pos_in_cg; n++) {
abssum += q_coef[scan[n + subpos]];
}
if (last_nz_pos_in_cg >= 0 && last_cg == -1) {
last_cg = 1;
}
if (last_nz_pos_in_cg - first_nz_pos_in_cg >= 4) {
int32_t signbit = (q_coef[scan[subpos + first_nz_pos_in_cg]] > 0 ? 0 : 1);
if (signbit != (abssum & 0x1)) { // compare signbit with sum_parity
int32_t min_cost_inc = 0x7fffffff, min_pos = -1, cur_cost = 0x7fffffff;
int16_t final_change = 0, cur_change = 0;
for (n = (last_cg == 1 ? last_nz_pos_in_cg : SCAN_SET_SIZE - 1); n >= 0; n--) {
uint32_t blkPos = scan[n + subpos];
if (q_coef[blkPos] != 0) {
if (delta_u[blkPos] > 0) {
cur_cost = -delta_u[blkPos];
cur_change = 1;
}
else if (n == first_nz_pos_in_cg && abs(q_coef[blkPos]) == 1) {
cur_cost = 0x7fffffff;
}
else {
cur_cost = delta_u[blkPos];
cur_change = -1;
}
}
else if (n < first_nz_pos_in_cg && ((coef[blkPos] >= 0) ? 0 : 1) != signbit) {
cur_cost = 0x7fffffff;
}
else {
cur_cost = -delta_u[blkPos];
cur_change = 1;
}
if (cur_cost < min_cost_inc) {
min_cost_inc = cur_cost;
final_change = cur_change;
min_pos = blkPos;
}
} // CG loop
if (q_coef[min_pos] == 32767 || q_coef[min_pos] == -32768) {
final_change = -1;
}
if (coef[min_pos] >= 0) q_coef[min_pos] += final_change;
else q_coef[min_pos] -= final_change;
} // Hide
}
if (last_cg == 1) last_cg = 0;
}
#undef SCAN_SET_SIZE
#undef LOG2_SCAN_SET_SIZE
}
__m128i test_mm256_extracti128_si256_1(__m256i a) {
// CHECK-LABEL: @test_mm256_extracti128_si256_1
// CHECK: shufflevector{{.*}}<i32 2, i32 3>
return _mm256_extracti128_si256(a, 1);
}
void FLAC__precompute_partition_info_sums_intrin_avx2(const FLAC__int32 residual[], FLAC__uint64 abs_residual_partition_sums[],
uint32_t residual_samples, uint32_t predictor_order, uint32_t min_partition_order, uint32_t max_partition_order, uint32_t bps)
{
const uint32_t default_partition_samples = (residual_samples + predictor_order) >> max_partition_order;
uint32_t partitions = 1u << max_partition_order;
FLAC__ASSERT(default_partition_samples > predictor_order);
/* first do max_partition_order */
{
const uint32_t threshold = 32 - FLAC__bitmath_ilog2(default_partition_samples);
uint32_t partition, residual_sample, end = (uint32_t)(-(int32_t)predictor_order);
if(bps + FLAC__MAX_EXTRA_RESIDUAL_BPS < threshold) {
for(partition = residual_sample = 0; partition < partitions; partition++) {
__m256i sum256 = _mm256_setzero_si256();
__m128i sum128;
end += default_partition_samples;
for( ; (int)residual_sample < (int)end-7; residual_sample+=8) {
__m256i res256 = _mm256_abs_epi32(_mm256_loadu_si256((const __m256i*)(residual+residual_sample)));
sum256 = _mm256_add_epi32(sum256, res256);
}
sum128 = _mm_add_epi32(_mm256_extracti128_si256(sum256, 1), _mm256_castsi256_si128(sum256));
for( ; (int)residual_sample < (int)end-3; residual_sample+=4) {
__m128i res128 = _mm_abs_epi32(_mm_loadu_si128((const __m128i*)(residual+residual_sample)));
sum128 = _mm_add_epi32(sum128, res128);
}
for( ; residual_sample < end; residual_sample++) {
__m128i res128 = _mm_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample]));
sum128 = _mm_add_epi32(sum128, res128);
}
sum128 = _mm_add_epi32(sum128, _mm_shuffle_epi32(sum128, _MM_SHUFFLE(1,0,3,2)));
sum128 = _mm_add_epi32(sum128, _mm_shufflelo_epi16(sum128, _MM_SHUFFLE(1,0,3,2)));
abs_residual_partition_sums[partition] = (FLAC__uint32)_mm_cvtsi128_si32(sum128);
/* workaround for MSVC bugs (at least versions 2015 and 2017 are affected) */
#if (defined _MSC_VER) && (defined FLAC__CPU_X86_64)
abs_residual_partition_sums[partition] &= 0xFFFFFFFF; /**/
#endif
}
}
else { /* have to pessimistically use 64 bits for accumulator */
for(partition = residual_sample = 0; partition < partitions; partition++) {
__m256i sum256 = _mm256_setzero_si256();
__m128i sum128;
end += default_partition_samples;
for( ; (int)residual_sample < (int)end-3; residual_sample+=4) {
__m128i res128 = _mm_abs_epi32(_mm_loadu_si128((const __m128i*)(residual+residual_sample)));
__m256i res256 = _mm256_cvtepu32_epi64(res128);
sum256 = _mm256_add_epi64(sum256, res256);
}
sum128 = _mm_add_epi64(_mm256_extracti128_si256(sum256, 1), _mm256_castsi256_si128(sum256));
for( ; (int)residual_sample < (int)end-1; residual_sample+=2) {
__m128i res128 = _mm_abs_epi32(_mm_loadl_epi64((const __m128i*)(residual+residual_sample)));
res128 = _mm_cvtepu32_epi64(res128);
sum128 = _mm_add_epi64(sum128, res128);
}
for( ; residual_sample < end; residual_sample++) {
__m128i res128 = _mm_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample]));
sum128 = _mm_add_epi64(sum128, res128);
}
sum128 = _mm_add_epi64(sum128, _mm_srli_si128(sum128, 8));
_mm_storel_epi64((__m128i*)(abs_residual_partition_sums+partition), sum128);
}
}
}
/* now merge partitions for lower orders */
{
uint32_t from_partition = 0, to_partition = partitions;
int partition_order;
for(partition_order = (int)max_partition_order - 1; partition_order >= (int)min_partition_order; partition_order--) {
uint32_t i;
partitions >>= 1;
for(i = 0; i < partitions; i++) {
abs_residual_partition_sums[to_partition++] =
abs_residual_partition_sums[from_partition ] +
abs_residual_partition_sums[from_partition+1];
from_partition += 2;
}
}
}
_mm256_zeroupper();
}
_mm256_storeu2_m128i(__m128i* const hiaddr, __m128i* const loaddr, const __m256i a)
{
_mm_storeu_si128(loaddr, _mm256_castsi256_si128(a));
_mm_storeu_si128(hiaddr, _mm256_extracti128_si256(a, 1));
}
void FLAC__precompute_partition_info_sums_intrin_avx2(const FLAC__int32 residual[], FLAC__uint64 abs_residual_partition_sums[],
uint32_t residual_samples, uint32_t predictor_order, uint32_t min_partition_order, uint32_t max_partition_order, uint32_t bps)
{
const uint32_t default_partition_samples = (residual_samples + predictor_order) >> max_partition_order;
uint32_t partitions = 1u << max_partition_order;
FLAC__ASSERT(default_partition_samples > predictor_order);
/* first do max_partition_order */
{
const uint32_t threshold = 32 - FLAC__bitmath_ilog2(default_partition_samples);
uint32_t partition, residual_sample, end = (uint32_t)(-(int32_t)predictor_order);
if(bps + FLAC__MAX_EXTRA_RESIDUAL_BPS < threshold) {
for(partition = residual_sample = 0; partition < partitions; partition++) {
__m256i sum256 = _mm256_setzero_si256();
__m128i sum128;
end += default_partition_samples;
for( ; (int)residual_sample < (int)end-7; residual_sample+=8) {
__m256i res256 = _mm256_abs_epi32(_mm256_loadu_si256((const __m256i*)(residual+residual_sample)));
sum256 = _mm256_add_epi32(sum256, res256);
}
sum128 = _mm_add_epi32(_mm256_extracti128_si256(sum256, 1), _mm256_castsi256_si128(sum256));
for( ; (int)residual_sample < (int)end-3; residual_sample+=4) {
__m128i res128 = _mm_abs_epi32(_mm_loadu_si128((const __m128i*)(residual+residual_sample)));
sum128 = _mm_add_epi32(sum128, res128);
}
for( ; residual_sample < end; residual_sample++) {
__m128i res128 = _mm_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample]));
sum128 = _mm_add_epi32(sum128, res128);
}
sum128 = _mm_hadd_epi32(sum128, sum128);
sum128 = _mm_hadd_epi32(sum128, sum128);
abs_residual_partition_sums[partition] = (FLAC__uint32)_mm_cvtsi128_si32(sum128);
/* workaround for a bug in MSVC2015U2 - see https://connect.microsoft.com/VisualStudio/feedback/details/2659191/incorrect-code-generation-for-x86-64 */
#if (defined _MSC_VER) && (_MSC_FULL_VER == 190023918) && (defined FLAC__CPU_X86_64)
abs_residual_partition_sums[partition] &= 0xFFFFFFFF; /**/
#endif
}
}
else { /* have to pessimistically use 64 bits for accumulator */
for(partition = residual_sample = 0; partition < partitions; partition++) {
__m256i sum256 = _mm256_setzero_si256();
__m128i sum128;
end += default_partition_samples;
for( ; (int)residual_sample < (int)end-3; residual_sample+=4) {
__m128i res128 = _mm_abs_epi32(_mm_loadu_si128((const __m128i*)(residual+residual_sample)));
__m256i res256 = _mm256_cvtepu32_epi64(res128);
sum256 = _mm256_add_epi64(sum256, res256);
}
sum128 = _mm_add_epi64(_mm256_extracti128_si256(sum256, 1), _mm256_castsi256_si128(sum256));
for( ; (int)residual_sample < (int)end-1; residual_sample+=2) {
__m128i res128 = _mm_abs_epi32(_mm_loadl_epi64((const __m128i*)(residual+residual_sample)));
res128 = _mm_cvtepu32_epi64(res128);
sum128 = _mm_add_epi64(sum128, res128);
}
for( ; residual_sample < end; residual_sample++) {
__m128i res128 = _mm_abs_epi32(_mm_cvtsi32_si128(residual[residual_sample]));
sum128 = _mm_add_epi64(sum128, res128);
}
sum128 = _mm_add_epi64(sum128, _mm_srli_si128(sum128, 8));
_mm_storel_epi64((__m128i*)(abs_residual_partition_sums+partition), sum128);
}
}
}
/* now merge partitions for lower orders */
{
uint32_t from_partition = 0, to_partition = partitions;
int partition_order;
for(partition_order = (int)max_partition_order - 1; partition_order >= (int)min_partition_order; partition_order--) {
uint32_t i;
partitions >>= 1;
for(i = 0; i < partitions; i++) {
abs_residual_partition_sums[to_partition++] =
abs_residual_partition_sums[from_partition ] +
abs_residual_partition_sums[from_partition+1];
from_partition += 2;
}
}
}
_mm256_zeroupper();
}
// Immediate should be truncated to one bit.
__m128i test2_mm256_extracti128_si256(__m256i a) {
// CHECK-LABEL: test2_mm256_extracti128_si256
// CHECK: shufflevector <4 x i64> %{{.*}}, <4 x i64> %{{.*}}, <2 x i32> <i32 0, i32 1>
return _mm256_extracti128_si256(a, 2);
}
__m128i test_mm256_extracti128_si256(__m256i a) {
// CHECK: @llvm.x86.avx2.vextracti128
return _mm256_extracti128_si256(a, 1);
}
void
test1bit (void)
{
i1 = _mm256_extracti128_si256 (l1, 2); /* { dg-error "the last argument must be an 1-bit immediate" } */
l1 = _mm256_inserti128_si256 (l1, i2, 2); /* { dg-error "the last argument must be an 1-bit immediate" } */
}
