void bitmask_avx2(uint32_t* ptr, size_t n, uint32_t key, uint8_t* out) {
uint32_t* output = (uint32_t*)out;
const size_t N = 8*4; // unrolled 4 times
const size_t chunks = n / N;
const size_t tail = n % N;
const __m256i vkey = _mm256_set1_epi32(key);
for (size_t i=0; i < chunks; i++) {
const __m256i in0 = _mm256_loadu_si256((const __m256i*)(ptr + i*N + 0*8));
const __m256i in1 = _mm256_loadu_si256((const __m256i*)(ptr + i*N + 1*8));
const __m256i in2 = _mm256_loadu_si256((const __m256i*)(ptr + i*N + 2*8));
const __m256i in3 = _mm256_loadu_si256((const __m256i*)(ptr + i*N + 3*8));
const __m256i eq0 = _mm256_cmpeq_epi32(in0, vkey);
const __m256i eq1 = _mm256_cmpeq_epi32(in1, vkey);
const __m256i eq2 = _mm256_cmpeq_epi32(in2, vkey);
const __m256i eq3 = _mm256_cmpeq_epi32(in3, vkey);
// eq0 = [a0 a1 a2 a3 a4 a5 a6 a7] (packed dword)
// eq1 = [b0 b1 b2 b3 b4 b5 b6 b7] (packed dword)
// eq2 = [c0 c1 c2 c3 c4 c5 c6 c7] (packed dword)
// eq3 = [d0 d1 d2 d3 d4 d5 d6 d7] (packed dword)
// t0 = [a0 a1 a2 a3 c0 c1 c2 c3 a4 a5 a6 a7 c4 c5 c6 c7] (packed word)
const __m256i t0 = _mm256_packs_epi32(eq0, eq2);
// m02 = [a0 a1 a2 a3 a4 a5 a6 a7 c0 c1 c2 c3 c4 c5 c6 c7] (packed word)
const __m256i m02 = _mm256_permutevar8x32_epi32(t0,
_mm256_setr_epi32(0, 1, 4, 5, 2, 3, 6, 7));
// t0 = [b0 b1 b2 b3 d0 d1 d2 d3 b4 b5 b6 b7 d4 d5 d6 d7] (packed word)
const __m256i t1 = _mm256_packs_epi32(eq1, eq3);
// m13 = [b0 b1 b2 b3 b4 b5 b6 b7 d0 d1 d2 d3 d4 d5 d6 d7] (packed word)
const __m256i m13 = _mm256_permutevar8x32_epi32(t1,
_mm256_setr_epi32(0, 1, 4, 5, 2, 3, 6, 7));
// m = [a0..7 b0..7 c0..7 d0..7] (packed byte)
const __m256i m = _mm256_packs_epi16(m02, m13);
*output++ = _mm256_movemask_epi8(m);
}
if (tail > 0) {
bitmask_better_2(ptr + chunks*N, tail, key, out + chunks*N);
}
}
void static
avx2_test (void)
{
union256i_d s1, s2;
union256i_w u;
short e[16];
int i;
s1.x = _mm256_set_epi32 (1, 2, 3, 4, 65000, 20, 30, 90);
s2.x = _mm256_set_epi32 (88, 44, 33, 22, 11, 98, 76, -65000);
u.x = _mm256_packs_epi32 (s1.x, s2.x);
for (i = 0; i < 4; i++)
{
e[i] = int_to_short (s1.a[i]);
e[i + 4] = int_to_short (s2.a[i]);
e[i + 8] = int_to_short (s1.a[i + 4]);
e[i + 12] = int_to_short (s2.a[i + 4]);
}
if (check_union256i_w (u, e))
abort ();
}
__m256i Scaler::process_vect_int_avx2 (const __m256i &add_cst, int kernel_size, const __m256i coef_base_ptr [], typename SRC::PtrConst::Type pix_ptr, const __m256i &zero, int src_stride, const __m256i &sign_bit, int len)
{
typedef typename SRC::template S16 <false, (SB == 16)> SrcS16R;
__m256i sum0 = add_cst;
__m256i sum1 = add_cst;
for (int k = 0; k < kernel_size; ++k)
{
const __m256i coef = _mm256_load_si256 (coef_base_ptr + k);
const __m256i src = ReadWrapperInt <SRC, SrcS16R, PF>::read (
pix_ptr, zero, sign_bit, len
);
fstb::ToolsAvx2::mac_s16_s16_s32 (sum0, sum1, src, coef);
SRC::PtrConst::jump (pix_ptr, src_stride);
}
sum0 = _mm256_srai_epi32 (sum0, SHIFT_INT + SB - DB);
sum1 = _mm256_srai_epi32 (sum1, SHIFT_INT + SB - DB);
const __m256i val = _mm256_packs_epi32 (sum0, sum1);
return (val);
}
static INLINE unsigned int highbd_masked_sad16xh_avx2(
const uint8_t *src8, int src_stride, const uint8_t *a8, int a_stride,
const uint8_t *b8, int b_stride, const uint8_t *m_ptr, int m_stride,
int width, int height) {
const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src8);
const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(a8);
const uint16_t *b_ptr = CONVERT_TO_SHORTPTR(b8);
int x, y;
__m256i res = _mm256_setzero_si256();
const __m256i mask_max = _mm256_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS));
const __m256i round_const =
_mm256_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1);
const __m256i one = _mm256_set1_epi16(1);
for (y = 0; y < height; y++) {
for (x = 0; x < width; x += 16) {
const __m256i src = _mm256_lddqu_si256((const __m256i *)&src_ptr[x]);
const __m256i a = _mm256_lddqu_si256((const __m256i *)&a_ptr[x]);
const __m256i b = _mm256_lddqu_si256((const __m256i *)&b_ptr[x]);
// Zero-extend mask to 16 bits
const __m256i m =
_mm256_cvtepu8_epi16(_mm_lddqu_si128((const __m128i *)&m_ptr[x]));
const __m256i m_inv = _mm256_sub_epi16(mask_max, m);
const __m256i data_l = _mm256_unpacklo_epi16(a, b);
const __m256i mask_l = _mm256_unpacklo_epi16(m, m_inv);
__m256i pred_l = _mm256_madd_epi16(data_l, mask_l);
pred_l = _mm256_srai_epi32(_mm256_add_epi32(pred_l, round_const),
AOM_BLEND_A64_ROUND_BITS);
const __m256i data_r = _mm256_unpackhi_epi16(a, b);
const __m256i mask_r = _mm256_unpackhi_epi16(m, m_inv);
__m256i pred_r = _mm256_madd_epi16(data_r, mask_r);
pred_r = _mm256_srai_epi32(_mm256_add_epi32(pred_r, round_const),
AOM_BLEND_A64_ROUND_BITS);
// Note: the maximum value in pred_l/r is (2^bd)-1 < 2^15,
// so it is safe to do signed saturation here.
const __m256i pred = _mm256_packs_epi32(pred_l, pred_r);
// There is no 16-bit SAD instruction, so we have to synthesize
// an 8-element SAD. We do this by storing 4 32-bit partial SADs,
// and accumulating them at the end
const __m256i diff = _mm256_abs_epi16(_mm256_sub_epi16(pred, src));
res = _mm256_add_epi32(res, _mm256_madd_epi16(diff, one));
}
src_ptr += src_stride;
a_ptr += a_stride;
b_ptr += b_stride;
m_ptr += m_stride;
}
// At this point, we have four 32-bit partial SADs stored in 'res'.
res = _mm256_hadd_epi32(res, res);
res = _mm256_hadd_epi32(res, res);
int sad = _mm256_extract_epi32(res, 0) + _mm256_extract_epi32(res, 4);
return (sad + 31) >> 6;
}
int warpAffineBlockline(int *adelta, int *bdelta, short* xy, short* alpha, int X0, int Y0, int bw)
{
const int AB_BITS = MAX(10, (int)INTER_BITS);
int x1 = 0;
__m256i fxy_mask = _mm256_set1_epi32(INTER_TAB_SIZE - 1);
__m256i XX = _mm256_set1_epi32(X0), YY = _mm256_set1_epi32(Y0);
for (; x1 <= bw - 16; x1 += 16)
{
__m256i tx0, tx1, ty0, ty1;
tx0 = _mm256_add_epi32(_mm256_loadu_si256((const __m256i*)(adelta + x1)), XX);
ty0 = _mm256_add_epi32(_mm256_loadu_si256((const __m256i*)(bdelta + x1)), YY);
tx1 = _mm256_add_epi32(_mm256_loadu_si256((const __m256i*)(adelta + x1 + 8)), XX);
ty1 = _mm256_add_epi32(_mm256_loadu_si256((const __m256i*)(bdelta + x1 + 8)), YY);
tx0 = _mm256_srai_epi32(tx0, AB_BITS - INTER_BITS);
ty0 = _mm256_srai_epi32(ty0, AB_BITS - INTER_BITS);
tx1 = _mm256_srai_epi32(tx1, AB_BITS - INTER_BITS);
ty1 = _mm256_srai_epi32(ty1, AB_BITS - INTER_BITS);
__m256i fx_ = _mm256_packs_epi32(_mm256_and_si256(tx0, fxy_mask),
_mm256_and_si256(tx1, fxy_mask));
__m256i fy_ = _mm256_packs_epi32(_mm256_and_si256(ty0, fxy_mask),
_mm256_and_si256(ty1, fxy_mask));
tx0 = _mm256_packs_epi32(_mm256_srai_epi32(tx0, INTER_BITS),
_mm256_srai_epi32(tx1, INTER_BITS));
ty0 = _mm256_packs_epi32(_mm256_srai_epi32(ty0, INTER_BITS),
_mm256_srai_epi32(ty1, INTER_BITS));
fx_ = _mm256_adds_epi16(fx_, _mm256_slli_epi16(fy_, INTER_BITS));
fx_ = _mm256_permute4x64_epi64(fx_, (3 << 6) + (1 << 4) + (2 << 2) + 0);
_mm256_storeu_si256((__m256i*)(xy + x1 * 2), _mm256_unpacklo_epi16(tx0, ty0));
_mm256_storeu_si256((__m256i*)(xy + x1 * 2 + 16), _mm256_unpackhi_epi16(tx0, ty0));
_mm256_storeu_si256((__m256i*)(alpha + x1), fx_);
}
_mm256_zeroupper();
return x1;
}
static INLINE __m256i highbd_comp_mask_pred_line_avx2(const __m256i s0,
const __m256i s1,
const __m256i a) {
const __m256i alpha_max = _mm256_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS));
const __m256i round_const =
_mm256_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1);
const __m256i a_inv = _mm256_sub_epi16(alpha_max, a);
const __m256i s_lo = _mm256_unpacklo_epi16(s0, s1);
const __m256i a_lo = _mm256_unpacklo_epi16(a, a_inv);
const __m256i pred_lo = _mm256_madd_epi16(s_lo, a_lo);
const __m256i pred_l = _mm256_srai_epi32(
_mm256_add_epi32(pred_lo, round_const), AOM_BLEND_A64_ROUND_BITS);
const __m256i s_hi = _mm256_unpackhi_epi16(s0, s1);
const __m256i a_hi = _mm256_unpackhi_epi16(a, a_inv);
const __m256i pred_hi = _mm256_madd_epi16(s_hi, a_hi);
const __m256i pred_h = _mm256_srai_epi32(
_mm256_add_epi32(pred_hi, round_const), AOM_BLEND_A64_ROUND_BITS);
const __m256i comp = _mm256_packs_epi32(pred_l, pred_h);
return comp;
}
__m256i test_mm256_packs_epi32(__m256i a, __m256i b) {
// CHECK: @llvm.x86.avx2.packssdw
return _mm256_packs_epi32(a, b);
}
__m256i test_mm256_packs_epi32(__m256i a, __m256i b) {
// CHECK-LABEL: test_mm256_packs_epi32
// CHECK: call <16 x i16> @llvm.x86.avx2.packssdw(<8 x i32> %{{.*}}, <8 x i32> %{{.*}})
return _mm256_packs_epi32(a, b);
}
/**
* \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
}