int64_t vp9_block_error_avx2(const int16_t *coeff,
const int16_t *dqcoeff,
intptr_t block_size,
int64_t *ssz) {
__m256i sse_reg, ssz_reg, coeff_reg, dqcoeff_reg;
__m256i exp_dqcoeff_lo, exp_dqcoeff_hi, exp_coeff_lo, exp_coeff_hi;
__m256i sse_reg_64hi, ssz_reg_64hi;
__m128i sse_reg128, ssz_reg128;
int64_t sse;
int i;
const __m256i zero_reg = _mm256_set1_epi16(0);
// init sse and ssz registerd to zero
sse_reg = _mm256_set1_epi16(0);
ssz_reg = _mm256_set1_epi16(0);
for (i = 0 ; i < block_size ; i+= 16) {
// load 32 bytes from coeff and dqcoeff
coeff_reg = _mm256_loadu_si256((const __m256i *)(coeff + i));
dqcoeff_reg = _mm256_loadu_si256((const __m256i *)(dqcoeff + i));
// dqcoeff - coeff
dqcoeff_reg = _mm256_sub_epi16(dqcoeff_reg, coeff_reg);
// madd (dqcoeff - coeff)
dqcoeff_reg = _mm256_madd_epi16(dqcoeff_reg, dqcoeff_reg);
// madd coeff
coeff_reg = _mm256_madd_epi16(coeff_reg, coeff_reg);
// expand each double word of madd (dqcoeff - coeff) to quad word
exp_dqcoeff_lo = _mm256_unpacklo_epi32(dqcoeff_reg, zero_reg);
exp_dqcoeff_hi = _mm256_unpackhi_epi32(dqcoeff_reg, zero_reg);
// expand each double word of madd (coeff) to quad word
exp_coeff_lo = _mm256_unpacklo_epi32(coeff_reg, zero_reg);
exp_coeff_hi = _mm256_unpackhi_epi32(coeff_reg, zero_reg);
// add each quad word of madd (dqcoeff - coeff) and madd (coeff)
sse_reg = _mm256_add_epi64(sse_reg, exp_dqcoeff_lo);
ssz_reg = _mm256_add_epi64(ssz_reg, exp_coeff_lo);
sse_reg = _mm256_add_epi64(sse_reg, exp_dqcoeff_hi);
ssz_reg = _mm256_add_epi64(ssz_reg, exp_coeff_hi);
}
// save the higher 64 bit of each 128 bit lane
sse_reg_64hi = _mm256_srli_si256(sse_reg, 8);
ssz_reg_64hi = _mm256_srli_si256(ssz_reg, 8);
// add the higher 64 bit to the low 64 bit
sse_reg = _mm256_add_epi64(sse_reg, sse_reg_64hi);
ssz_reg = _mm256_add_epi64(ssz_reg, ssz_reg_64hi);
// add each 64 bit from each of the 128 bit lane of the 256 bit
sse_reg128 = _mm_add_epi64(_mm256_castsi256_si128(sse_reg),
_mm256_extractf128_si256(sse_reg, 1));
ssz_reg128 = _mm_add_epi64(_mm256_castsi256_si128(ssz_reg),
_mm256_extractf128_si256(ssz_reg, 1));
// store the results
_mm_storel_epi64((__m128i*)(&sse), sse_reg128);
_mm_storel_epi64((__m128i*)(ssz), ssz_reg128);
return sse;
}
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;
}
SIMD_INLINE __m256i BgraToGray32(__m256i bgra)
{
const __m256i g0a0 = _mm256_and_si256(_mm256_srli_si256(bgra, 1), K16_00FF);
const __m256i b0r0 = _mm256_and_si256(bgra, K16_00FF);
const __m256i weightedSum = _mm256_add_epi32(_mm256_madd_epi16(g0a0, K16_GREEN_0000), _mm256_madd_epi16(b0r0, K16_BLUE_RED));
return _mm256_srli_epi32(_mm256_add_epi32(weightedSum, K32_ROUND_TERM), Base::BGR_TO_GRAY_AVERAGING_SHIFT);
}
static FORCE_INLINE __m256i lookup_AVX2(const int16_t *VXFull, const int16_t *VYFull, const PixelType *pref, int w, const __m256i &dwords_time256, const __m256i &dwords_ref_pitch, const __m256i &dwords_hoffsets) {
__m256i vx = _mm256_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)&VXFull[w]));
__m256i vy = _mm256_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)&VYFull[w]));
vx = _mm256_madd_epi16(vx, dwords_time256);
vx = _mm256_srai_epi32(vx, 8);
vy = _mm256_madd_epi16(vy, dwords_time256);
vy = _mm256_srai_epi32(vy, 8);
__m256i addr = _mm256_madd_epi16(vy, dwords_ref_pitch);
addr = _mm256_add_epi32(addr, vx);
addr = _mm256_add_epi32(addr, dwords_hoffsets);
// It's okay to read two or three bytes more than needed. pref is always padded, unless the user chooses a horizontal padding of 0, which would be stupid.
__m256i gathered = _mm256_i32gather_epi32((const int *)pref, addr, sizeof(PixelType));
gathered = _mm256_and_si256(gathered, _mm256_set1_epi32((1 << (sizeof(PixelType) * 8)) - 1));
return gathered;
}
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));
}
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;
}
static FORCE_INLINE void FlowInterSimple_double_8px_AVX2(
int w, PixelType *pdst,
const PixelType *prefB, const PixelType *prefF,
const int16_t *VXFullB, const int16_t *VXFullF,
const int16_t *VYFullB, const int16_t *VYFullF,
const uint8_t *MaskB, const uint8_t *MaskF,
int nPelLog,
const __m256i &dwords_ref_pitch, const __m256i &dwords_hoffsets) {
__m256i dwords_w = _mm256_add_epi32(_mm256_set1_epi32(w << nPelLog), dwords_hoffsets); /// maybe do it another way
__m256i dstF = lookup_double_AVX2(VXFullF, VYFullF, prefF, w, dwords_ref_pitch, dwords_w);
__m256i dstB = lookup_double_AVX2(VXFullB, VYFullB, prefB, w, dwords_ref_pitch, dwords_w);
__m256i maskf = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i *)&MaskF[w]));
__m256i maskb = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i *)&MaskB[w]));
__m256i dstF_dstB = _mm256_add_epi32(dstF, dstB);
dstF_dstB = _mm256_slli_epi32(dstF_dstB, 8);
__m256i dst;
if (sizeof(PixelType) == 1) {
__m256i dstB_dstF = _mm256_sub_epi16(dstB, dstF);
__m256i maskf_maskb = _mm256_sub_epi16(maskf, maskb);
dst = _mm256_madd_epi16(dstB_dstF, maskf_maskb);
} else {
__m256i dstB_dstF = _mm256_sub_epi32(dstB, dstF);
__m256i maskf_maskb = _mm256_sub_epi32(maskf, maskb);
dst = _mm256_mullo_epi32(dstB_dstF, maskf_maskb);
}
dst = _mm256_add_epi32(dst, dstF_dstB);
dst = _mm256_srai_epi32(dst, 9);
dst = _mm256_packus_epi32(dst, dst);
dst = _mm256_permute4x64_epi64(dst, 0xe8); // 0b11101000 - copy third qword to second qword
__m128i dst128 = _mm256_castsi256_si128(dst);
if (sizeof(PixelType) == 1) {
dst128 = _mm_packus_epi16(dst128, dst128);
_mm_storel_epi64((__m128i *)&pdst[w], dst128);
} else {
_mm_storeu_si128((__m128i *)&pdst[w], dst128);
}
}
int vpx_satd_avx2(const tran_low_t *coeff, int length) {
const __m256i one = _mm256_set1_epi16(1);
__m256i accum = _mm256_setzero_si256();
int i;
for (i = 0; i < length; i += 16) {
const __m256i src_line = load_tran_low(coeff);
const __m256i abs = _mm256_abs_epi16(src_line);
const __m256i sum = _mm256_madd_epi16(abs, one);
accum = _mm256_add_epi32(accum, sum);
coeff += 16;
}
{ // 32 bit horizontal add
const __m256i a = _mm256_srli_si256(accum, 8);
const __m256i b = _mm256_add_epi32(accum, a);
const __m256i c = _mm256_srli_epi64(b, 32);
const __m256i d = _mm256_add_epi32(b, c);
const __m128i accum_128 = _mm_add_epi32(_mm256_castsi256_si128(d),
_mm256_extractf128_si256(d, 1));
return _mm_cvtsi128_si32(accum_128);
}
}
static FORCE_INLINE void FlowInterExtra_8px_AVX2(
int w, PixelType *pdst,
const PixelType *prefB, const PixelType *prefF,
const int16_t *VXFullB, const int16_t *VXFullF,
const int16_t *VYFullB, const int16_t *VYFullF,
const uint8_t *MaskB, const uint8_t *MaskF,
int nPelLog,
const int16_t *VXFullBB, const int16_t *VXFullFF,
const int16_t *VYFullBB, const int16_t *VYFullFF,
const __m256i &dwords_time256, const __m256i &dwords_256_time256,
const __m256i &dwords_ref_pitch, const __m256i &dwords_hoffsets) {
__m256i dwords_w = _mm256_add_epi32(_mm256_set1_epi32(w << nPelLog), dwords_hoffsets);
__m256i dstF = lookup_AVX2(VXFullF, VYFullF, prefF, w, dwords_time256, dwords_ref_pitch, dwords_w);
__m256i dstB = lookup_AVX2(VXFullB, VYFullB, prefB, w, dwords_256_time256, dwords_ref_pitch, dwords_w);
__m256i dstFF = lookup_AVX2(VXFullFF, VYFullFF, prefF, w, dwords_time256, dwords_ref_pitch, dwords_w);
__m256i dstBB = lookup_AVX2(VXFullBB, VYFullBB, prefB, w, dwords_256_time256, dwords_ref_pitch, dwords_w);
__m256i minfb = mm256_min_epu<PixelType>(dstF, dstB);
__m256i maxfb = mm256_max_epu<PixelType>(dstF, dstB);
__m256i medianBB = mm256_max_epu<PixelType>(minfb, mm256_min_epu<PixelType>(maxfb, dstBB));
__m256i medianFF = mm256_max_epu<PixelType>(minfb, mm256_min_epu<PixelType>(maxfb, dstFF));
__m256i maskf = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i *)&MaskF[w]));
__m256i maskb = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i *)&MaskB[w]));
const __m256i dwords_255 = _mm256_set1_epi32(255);
__m256i maskf_inv = _mm256_sub_epi32(dwords_255, maskf);
__m256i maskb_inv = _mm256_sub_epi32(dwords_255, maskb);
if (sizeof(PixelType) == 1) {
dstF = _mm256_mullo_epi16(dstF, maskf_inv);
dstB = _mm256_mullo_epi16(dstB, maskb_inv);
medianBB = _mm256_mullo_epi16(medianBB, maskf);
medianFF = _mm256_mullo_epi16(medianFF, maskb);
} else {
dstF = _mm256_mullo_epi32(dstF, maskf_inv);
dstB = _mm256_mullo_epi32(dstB, maskb_inv);
medianBB = _mm256_mullo_epi32(medianBB, maskf);
medianFF = _mm256_mullo_epi32(medianFF, maskb);
}
dstF = _mm256_add_epi32(dstF, dwords_255);
dstB = _mm256_add_epi32(dstB, dwords_255);
dstF = _mm256_add_epi32(dstF, medianBB);
dstB = _mm256_add_epi32(dstB, medianFF);
dstF = _mm256_srai_epi32(dstF, 8);
dstB = _mm256_srai_epi32(dstB, 8);
if (sizeof(PixelType) == 2) {
dstF = _mm256_sub_epi16(dstF, _mm256_set1_epi32(32768));
dstB = _mm256_sub_epi16(dstB, _mm256_set1_epi32(32768));
}
dstF = _mm256_madd_epi16(dstF, dwords_256_time256);
dstB = _mm256_madd_epi16(dstB, dwords_time256);
if (sizeof(PixelType) == 2) {
// dstF = _mm256_add_epi32(dstF, _mm256_slli_epi32(dwords_256_time256, 15));
// dstB = _mm256_add_epi32(dstB, _mm256_slli_epi32(dwords_time256, 15));
// Knowing that they add up to 256, the two additions can be combined.
dstF = _mm256_add_epi32(dstF, _mm256_set1_epi32(256 << 15));
}
__m256i dst = _mm256_add_epi32(dstF, dstB);
dst = _mm256_srai_epi32(dst, 8);
dst = _mm256_packus_epi32(dst, dst);
dst = _mm256_permute4x64_epi64(dst, 0xe8); // 0b11101000 - copy third qword to second qword
__m128i dst128 = _mm256_castsi256_si128(dst);
if (sizeof(PixelType) == 1) {
dst128 = _mm_packus_epi16(dst128, dst128);
_mm_storel_epi64((__m128i *)&pdst[w], dst128);
} else {
_mm_storeu_si128((__m128i *)&pdst[w], dst128);
}
}
static FORCE_INLINE void FlowInter_8px_AVX2(
int w, PixelType *pdst,
const PixelType *prefB, const PixelType *prefF,
const int16_t *VXFullB, const int16_t *VXFullF,
const int16_t *VYFullB, const int16_t *VYFullF,
const uint8_t *MaskB, const uint8_t *MaskF,
int nPelLog,
const __m256i &dwords_time256, const __m256i &dwords_256_time256,
const __m256i &dwords_ref_pitch, const __m256i &dwords_hoffsets) {
__m256i dwords_w = _mm256_add_epi32(_mm256_set1_epi32(w << nPelLog), dwords_hoffsets);
__m256i dstF = lookup_AVX2(VXFullF, VYFullF, prefF, w, dwords_time256, dwords_ref_pitch, dwords_w);
__m256i dstB = lookup_AVX2(VXFullB, VYFullB, prefB, w, dwords_256_time256, dwords_ref_pitch, dwords_w);
__m256i dstF0 = _mm256_i32gather_epi32((const int *)prefF, dwords_w, sizeof(PixelType));
__m256i dstB0 = _mm256_i32gather_epi32((const int *)prefB, dwords_w, sizeof(PixelType));
dstF0 = _mm256_and_si256(dstF0, _mm256_set1_epi32((1 << (sizeof(PixelType) * 8)) - 1));
dstB0 = _mm256_and_si256(dstB0, _mm256_set1_epi32((1 << (sizeof(PixelType) * 8)) - 1));
__m256i maskf = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i *)&MaskF[w]));
__m256i maskb = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i *)&MaskB[w]));
const __m256i dwords_255 = _mm256_set1_epi32(255);
__m256i maskf_inv = _mm256_sub_epi32(dwords_255, maskf);
__m256i maskb_inv = _mm256_sub_epi32(dwords_255, maskb);
__m256i dstF_maskf_inv, dstB_maskb_inv, dstF0_maskb, dstB0_maskf;
if (sizeof(PixelType) == 1) {
dstF_maskf_inv = _mm256_mullo_epi16(dstF, maskf_inv);
dstB_maskb_inv = _mm256_mullo_epi16(dstB, maskb_inv);
dstF0_maskb = _mm256_mullo_epi16(dstF0, maskb);
dstB0_maskf = _mm256_mullo_epi16(dstB0, maskf);
} else {
dstF_maskf_inv = _mm256_mullo_epi32(dstF, maskf_inv);
dstB_maskb_inv = _mm256_mullo_epi32(dstB, maskb_inv);
dstF0_maskb = _mm256_mullo_epi32(dstF0, maskb);
dstB0_maskf = _mm256_mullo_epi32(dstB0, maskf);
}
__m256i f = _mm256_add_epi32(dstF0_maskb, dstB_maskb_inv);
__m256i b = _mm256_add_epi32(dstB0_maskf, dstF_maskf_inv);
if (sizeof(PixelType) == 1) {
f = _mm256_mullo_epi32(f, maskf);
b = _mm256_mullo_epi32(b, maskb);
f = _mm256_add_epi32(f, dwords_255);
b = _mm256_add_epi32(b, dwords_255);
f = _mm256_srai_epi32(f, 8);
b = _mm256_srai_epi32(b, 8);
} else {
const __m256i qwords_255 = _mm256_set1_epi64x(255);
__m256i tempf = _mm256_mul_epu32(f, maskf);
__m256i tempb = _mm256_mul_epu32(b, maskb);
tempf = _mm256_add_epi64(tempf, qwords_255);
tempb = _mm256_add_epi64(tempb, qwords_255);
tempf = _mm256_srli_epi64(tempf, 8);
tempb = _mm256_srli_epi64(tempb, 8);
f = _mm256_srli_epi64(f, 32);
b = _mm256_srli_epi64(b, 32);
f = _mm256_mul_epu32(f, _mm256_srli_epi64(maskf, 32));
b = _mm256_mul_epu32(b, _mm256_srli_epi64(maskb, 32));
f = _mm256_add_epi64(f, qwords_255);
b = _mm256_add_epi64(b, qwords_255);
f = _mm256_srli_epi64(f, 8);
b = _mm256_srli_epi64(b, 8);
f = _mm256_or_si256(tempf, _mm256_slli_epi64(f, 32));
b = _mm256_or_si256(tempb, _mm256_slli_epi64(b, 32));
}
f = _mm256_add_epi32(f, dstF_maskf_inv);
b = _mm256_add_epi32(b, dstB_maskb_inv);
f = _mm256_add_epi32(f, dwords_255);
b = _mm256_add_epi32(b, dwords_255);
f = _mm256_srai_epi32(f, 8);
b = _mm256_srai_epi32(b, 8);
if (sizeof(PixelType) == 1) {
f = _mm256_madd_epi16(f, dwords_256_time256);
b = _mm256_madd_epi16(b, dwords_time256);
} else {
f = _mm256_mullo_epi32(f, dwords_256_time256);
b = _mm256_mullo_epi32(b, dwords_time256);
}
__m256i dst = _mm256_add_epi32(f, b);
dst = _mm256_srai_epi32(dst, 8);
dst = _mm256_packus_epi32(dst, dst);
dst = _mm256_permute4x64_epi64(dst, 0xe8); // 0b11101000 - copy third qword to second qword
__m128i dst128 = _mm256_castsi256_si128(dst);
if (sizeof(PixelType) == 1) {
dst128 = _mm_packus_epi16(dst128, dst128);
_mm_storel_epi64((__m128i *)&pdst[w], dst128);
} else {
_mm_storeu_si128((__m128i *)&pdst[w], dst128);
}
}
static FORCE_INLINE void FlowInterSimple_generic_8px_AVX2(
int w, PixelType *pdst,
const PixelType *prefB, const PixelType *prefF,
const int16_t *VXFullB, const int16_t *VXFullF,
const int16_t *VYFullB, const int16_t *VYFullF,
const uint8_t *MaskB, const uint8_t *MaskF,
int nPelLog,
const __m256i &dwords_time256, const __m256i &dwords_256_time256,
const __m256i &dwords_ref_pitch, const __m256i &dwords_hoffsets) {
__m256i dwords_w = _mm256_add_epi32(_mm256_set1_epi32(w << nPelLog), dwords_hoffsets);
__m256i dstF = lookup_AVX2(VXFullF, VYFullF, prefF, w, dwords_time256, dwords_ref_pitch, dwords_w);
__m256i dstB = lookup_AVX2(VXFullB, VYFullB, prefB, w, dwords_256_time256, dwords_ref_pitch, dwords_w);
__m256i maskf = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i *)&MaskF[w]));
__m256i maskb = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i *)&MaskB[w]));
const __m256i dwords_255 = _mm256_set1_epi32(255);
__m256i maskf_inv = _mm256_sub_epi32(dwords_255, maskf);
__m256i maskb_inv = _mm256_sub_epi32(dwords_255, maskb);
__m256i f, b;
if (sizeof(PixelType) == 1) {
__m256i dstF_dstB = _mm256_or_si256(dstF, _mm256_slli_epi32(dstB, 16));
maskf = _mm256_or_si256(_mm256_slli_epi32(maskf, 16), maskf_inv);
maskb = _mm256_or_si256(maskb, _mm256_slli_epi32(maskb_inv, 16));
f = _mm256_madd_epi16(dstF_dstB, maskf);
b = _mm256_madd_epi16(dstF_dstB, maskb);
} else {
__m256i dstF_maskf_inv = _mm256_mullo_epi32(dstF, maskf_inv);
__m256i dstB_maskb_inv = _mm256_mullo_epi32(dstB, maskb_inv);
__m256i dstB_maskf = _mm256_mullo_epi32(dstB, maskf);
__m256i dstF_maskb = _mm256_mullo_epi32(dstF, maskb);
f = _mm256_add_epi32(dstF_maskf_inv, dstB_maskf);
b = _mm256_add_epi32(dstB_maskb_inv, dstF_maskb);
}
f = _mm256_add_epi32(f, dwords_255);
b = _mm256_add_epi32(b, dwords_255);
f = _mm256_srai_epi32(f, 8);
b = _mm256_srai_epi32(b, 8);
if (sizeof(PixelType) == 1) {
f = _mm256_madd_epi16(f, dwords_256_time256);
b = _mm256_madd_epi16(b, dwords_time256);
} else {
f = _mm256_mullo_epi32(f, dwords_256_time256);
b = _mm256_mullo_epi32(b, dwords_time256);
}
__m256i dst = _mm256_add_epi32(f, b);
dst = _mm256_srai_epi32(dst, 8);
dst = _mm256_packus_epi32(dst, dst);
dst = _mm256_permute4x64_epi64(dst, 0xe8); // 0b11101000 - copy third qword to second qword
__m128i dst128 = _mm256_castsi256_si128(dst);
if (sizeof(PixelType) == 1) {
dst128 = _mm_packus_epi16(dst128, dst128);
_mm_storel_epi64((__m128i *)&pdst[w], dst128);
} else {
_mm_storeu_si128((__m128i *)&pdst[w], dst128);
}
}
SIMD_INLINE __m256i SquaredDifference(__m256i a, __m256i b)
{
const __m256i lo = SubUnpackedU8<0>(a, b);
const __m256i hi = SubUnpackedU8<1>(a, b);
return _mm256_add_epi32(_mm256_madd_epi16(lo, lo), _mm256_madd_epi16(hi, hi));
}
/**
* \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
}
void extern
avx2_test (void)
{
x = _mm256_madd_epi16 (x, x);
}
__m256i test_mm256_madd_epi16(__m256i a, __m256i b) {
// CHECK-LABEL: test_mm256_madd_epi16
// CHECK: call <8 x i32> @llvm.x86.avx2.pmadd.wd(<16 x i16> %{{.*}}, <16 x i16> %{{.*}})
return _mm256_madd_epi16(a, b);
}
__m256i test_mm256_madd_epi16(__m256i a, __m256i b) {
// CHECK: @llvm.x86.avx2.pmadd.wd
return _mm256_madd_epi16(a, b);
}
INLINE static void haddwd_accumulate_dual_avx2(__m256i *accumulate, __m256i *ver_row)
{
__m256i abs_value = _mm256_abs_epi16(*ver_row);
*accumulate = _mm256_add_epi32(*accumulate, _mm256_madd_epi16(abs_value, _mm256_set1_epi16(1)));
}
