static INLINE unsigned int hbd_obmc_sad_w8n(const uint8_t *pre8,
const int pre_stride,
const int32_t *wsrc,
const int32_t *mask,
const int width, const int height) {
const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8);
const int pre_step = pre_stride - width;
int n = 0;
__m128i v_sad_d = _mm_setzero_si128();
assert(width >= 8);
assert(IS_POWER_OF_TWO(width));
do {
const __m128i v_p1_w = xx_loadl_64(pre + n + 4);
const __m128i v_m1_d = xx_load_128(mask + n + 4);
const __m128i v_w1_d = xx_load_128(wsrc + n + 4);
const __m128i v_p0_w = xx_loadl_64(pre + n);
const __m128i v_m0_d = xx_load_128(mask + n);
const __m128i v_w0_d = xx_load_128(wsrc + n);
const __m128i v_p0_d = _mm_cvtepu16_epi32(v_p0_w);
const __m128i v_p1_d = _mm_cvtepu16_epi32(v_p1_w);
// Values in both pre and mask fit in 15 bits, and are packed at 32 bit
// boundaries. We use pmaddwd, as it has lower latency on Haswell
// than pmulld but produces the same result with these inputs.
const __m128i v_pm0_d = _mm_madd_epi16(v_p0_d, v_m0_d);
const __m128i v_pm1_d = _mm_madd_epi16(v_p1_d, v_m1_d);
const __m128i v_diff0_d = _mm_sub_epi32(v_w0_d, v_pm0_d);
const __m128i v_diff1_d = _mm_sub_epi32(v_w1_d, v_pm1_d);
const __m128i v_absdiff0_d = _mm_abs_epi32(v_diff0_d);
const __m128i v_absdiff1_d = _mm_abs_epi32(v_diff1_d);
// Rounded absolute difference
const __m128i v_rad0_d = xx_roundn_epu32(v_absdiff0_d, 12);
const __m128i v_rad1_d = xx_roundn_epu32(v_absdiff1_d, 12);
v_sad_d = _mm_add_epi32(v_sad_d, v_rad0_d);
v_sad_d = _mm_add_epi32(v_sad_d, v_rad1_d);
n += 8;
if (n % width == 0) pre += pre_step;
} while (n < width * height);
return xx_hsum_epi32_si32(v_sad_d);
}
// Add 'sum_u16' to 'count'. Multiply by 'pred' and add to 'accumulator.'
static void accumulate_and_store_8(const __m128i sum_u16, const uint8_t *pred,
uint16_t *count, uint32_t *accumulator) {
const __m128i pred_u8 = _mm_loadl_epi64((const __m128i *)pred);
const __m128i zero = _mm_setzero_si128();
__m128i count_u16 = _mm_loadu_si128((const __m128i *)count);
__m128i pred_u16 = _mm_cvtepu8_epi16(pred_u8);
__m128i pred_0_u32, pred_1_u32;
__m128i accum_0_u32, accum_1_u32;
count_u16 = _mm_adds_epu16(count_u16, sum_u16);
_mm_storeu_si128((__m128i *)count, count_u16);
pred_u16 = _mm_mullo_epi16(sum_u16, pred_u16);
pred_0_u32 = _mm_cvtepu16_epi32(pred_u16);
pred_1_u32 = _mm_unpackhi_epi16(pred_u16, zero);
accum_0_u32 = _mm_loadu_si128((const __m128i *)accumulator);
accum_1_u32 = _mm_loadu_si128((const __m128i *)(accumulator + 4));
accum_0_u32 = _mm_add_epi32(pred_0_u32, accum_0_u32);
accum_1_u32 = _mm_add_epi32(pred_1_u32, accum_1_u32);
_mm_storeu_si128((__m128i *)accumulator, accum_0_u32);
_mm_storeu_si128((__m128i *)(accumulator + 4), accum_1_u32);
}
inline __m128i load_aligned_int32(const uint16_t* src)
{
__m128i tmp = _mm_loadl_epi64((const __m128i*)src);
#if XSIMD_X86_INSTR_SET >= XSIMD_X86_SSE4_1_VERSION
__m128i res = _mm_cvtepu16_epi32(tmp);
#else
__m128i res = _mm_unpacklo_epi16(tmp, _mm_set1_epi16(0));
#endif
return res;
}
void ahd_interpolate_tile(int top, char * buffer)
{
int row, col, tr, tc, c, val;
const int dir[4] = { -1, 1, -width, width };
__m128i ldiff[2], abdiff[2];
union hvrgbpix (*rgb)[width] = (union hvrgbpix (*)[width])buffer;
union hvrgbpix *rix;
union rgbpix * pix;
union hvrgbpix (*lab)[width];
short (*lix)[8];
char (*h**o)[width][2];
lab = (union hvrgbpix (*)[width])(buffer + 16*width*TS);
h**o = (char (*)[width][2])(buffer + 32*width*TS);
const int left=2;
if ((uintptr_t)(image+top*width)&0xf || (uintptr_t)buffer&0xf) {
fprintf(stderr, "unaligned buffers defeat speed!\n"); abort();
}
/* Interpolate gren horz&vert, red and blue, and convert to CIELab: */
//do the first two rows of green first.
//then one green, and rgb through the tile.. this because R/B needs down-right green value
for (row=top; row < top+2 && row < height-2; row++) {
col = left + (FC(row,left) & 1);
for (c = FC(row,col); col < width-2; col+=2) {
pix = (union rgbpix*)image + row*width+col;
val = ((pix[-1].g + pix[0].c[c] + pix[1].g) * 2 - pix[-2].c[c] - pix[2].c[c]) >> 2;
rgb[row-top][col-left].h.g = ULIM(val,pix[-1].g,pix[1].g);
val = ((pix[-width].g + pix[0].c[c] + pix[width].g) * 2 - pix[-2*width].c[c] - pix[2*width].c[c]) >> 2;
rgb[row-top][col-left].v.g = ULIM(val,pix[-width].g,pix[width].g);
}
}
for (; row < top+TS && row < height-2; row++) {
int rowx = row-1;
if (FC(rowx,left+1)==1) {
int c1 = FC(rowx+1,left+1),
c2 = FC(rowx,left+2);
pix = (union rgbpix*)image + row*width+left+1;
rix = &rgb[row-top][1];
val = ((pix[-1].g + pix[0].c[c1] + pix[1].g) * 2 - pix[-2].c[c1] - pix[2].c[c1]) >> 2;
rix[0].h.g = ULIM(val,pix[-1].g,pix[1].g);
val = ((pix[-width].g + pix[0].c[c1] + pix[width].g) * 2 - pix[-2*width].c[c1] - pix[2*width].c[c1]) >> 2;
rix[0].v.g = ULIM(val,pix[-width].g,pix[width].g);
for (col=left+1; col < width-3; col+=2) {
pix = (union rgbpix*)image + rowx*width+col+1;
union hvrgbpix rixr, rix0;
rix = &rgb[rowx-top][col-left]+1;
signed pix_diag = pix[-width-1].c[c1] + pix[-width+1].c[c1];
signed pix_ul = pix[-width-1].c[c1];
rixr.vec = _mm_set1_epi16(pix[-1].g);
signed pix_lr = pix[-2].c[c2] + pix[0].c[c2];
rix0.h.c[c2] = rix0.v.c[c2] = pix[0].c[c2];
pix_diag += pix[width-1].c[c1] + pix[width+1].c[c1] + 1;
signed pix_dl = pix[width-1].c[c1];
//fully loaded
__m128i rix_dr = _mm_setr_epi32(pix[width].g, pix[width-1].c[c1], pix[1].g, pix[-width+1].c[c1]);
rix_dr = _mm_add_epi32(rix_dr,_mm_setr_epi32(pix[width+1].c[c1], pix[width+3].c[c1], pix[width+1].c[c1], 0));
rix_dr = _mm_add_epi32(rix_dr,_mm_setr_epi32(pix[width+2].g, 0, pix[2*width+1].g, pix[3*width+1].c[c1]));
rix_dr = _mm_mullo_epi32(rix_dr,_mm_setr_epi32(2,1,2,1));
//half loaded
rix_dr = _mm_hsub_epi32(rix_dr,_mm_setzero_si128());
rix_dr = _mm_srai_epi32(rix_dr,2);
__m128i a = _mm_setr_epi32(pix[width].g,pix[1].g,0,0);
__m128i b = _mm_setr_epi32(pix[width+2].g,pix[2*width+1].g,0,0);
__m128i m = _mm_min_epi32(a,b);
__m128i M = _mm_max_epi32(a,b);
rix_dr = _mm_min_epi32(rix_dr,M);
rix_dr = _mm_max_epi32(rix_dr,m);
signed pix_udr = pix_ul + pix_dl;
signed rix0_ul = rix[-width-1].h.g;
signed rix1_ul = rix[-width-1].v.g;
__m128i rix_ur = _mm_setr_epi32(rix[-width+1].h.g, rix[-width+1].v.g, 0, 0);
signed rix0_rr = rix[-2].h.g;
signed rix1_rr = rix[-2].v.g;
rix0.h.g = rix[0].h.g;
rix0.v.g = rix[0].v.g;
signed rix0_dl = rix[width-1].h.g;
signed rix1_dl = rix[width-1].v.g;
// fully loaded
__m128i rix_udr = _mm_setr_epi32(rix0_ul, rix1_ul, rix0_rr, rix1_rr);
rix_udr = _mm_add_epi32(rix_udr, _mm_setr_epi32(rix0_dl, rix1_dl, rix0.h.g, rix0.v.g));
__m128i v2 = _mm_set_epi32(pix_lr, pix_lr, pix_udr, pix_udr);
v2 = _mm_sub_epi32(v2, rix_udr);
v2 = _mm_srai_epi32(v2,1);
v2 = _mm_add_epi32(v2,_mm_cvtepu16_epi32(rixr.vec));
v2 = _mm_max_epi32(v2, _mm_setzero_si128());
v2 = _mm_min_epi32(v2, _mm_set1_epi32(0xffff));
rixr.h.c[c2] = _mm_extract_epi32(v2,2);
rixr.v.c[c2] = _mm_extract_epi32(v2,3);
rixr.h.c[c1] = _mm_extract_epi32(v2,0);
rixr.v.c[c1] = _mm_extract_epi32(v2,1);
// following only uses 64 bit
__m128i v1 = _mm_set1_epi32(pix_diag);
v1 = _mm_sub_epi32(v1, rix_ur);
v1 = _mm_sub_epi32(v1, rix_dr);
v1 = _mm_sub_epi32(v1, rix_udr);
v1 = _mm_srai_epi32(v1,2);
v1 = _mm_add_epi32(v1, _mm_setr_epi32(rix0.h.g, rix0.v.g, 0, 0));
v1 = _mm_max_epi32(v1, _mm_setzero_si128());
v1 = _mm_min_epi32(v1, _mm_set1_epi32(0xffff));
rix0.h.c[c1] = _mm_extract_epi32(v1,0);
rix0.v.c[c1] = _mm_extract_epi32(v1,1);
lab[rowx-top][col-left].vec = cielabv(rixr);
lab[rowx-top][col-left+1].vec = cielabv(rix0);
_mm_store_si128(&rix[-1].vec,rixr.vec);
_mm_store_si128(&rix[0].vec,rix0.vec);
rix[width+1].h.g = _mm_extract_epi32(rix_dr,0);
rix[width+1].v.g = _mm_extract_epi32(rix_dr,1);
}
} else {
__m128i test_mm_cvtepu16_epi32(__m128i a) {
// CHECK-LABEL: test_mm_cvtepu16_epi32
// CHECK: call <4 x i32> @llvm.x86.sse41.pmovzxwd(<8 x i16> {{.*}})
// CHECK-ASM: pmovzxwd %xmm{{.*}}, %xmm{{.*}}
return _mm_cvtepu16_epi32(a);
}
__m128i test_mm_cvtepu16_epi32(__m128i a) {
// CHECK-LABEL: test_mm_cvtepu16_epi32
// CHECK: shufflevector <8 x i16> {{.*}}, <8 x i16> {{.*}}, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
// CHECK: zext <4 x i16> {{.*}} to <4 x i32>
return _mm_cvtepu16_epi32(a);
}
void av1_highbd_jnt_convolve_2d_sse4_1(
const uint16_t *src, int src_stride, uint16_t *dst0, int dst_stride0, int w,
int h, const InterpFilterParams *filter_params_x,
const InterpFilterParams *filter_params_y, const int subpel_x_q4,
const int subpel_y_q4, ConvolveParams *conv_params, int bd) {
DECLARE_ALIGNED(16, int16_t,
im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE]);
CONV_BUF_TYPE *dst = conv_params->dst;
int dst_stride = conv_params->dst_stride;
int im_h = h + filter_params_y->taps - 1;
int im_stride = MAX_SB_SIZE;
int i, j;
const int do_average = conv_params->do_average;
const int use_jnt_comp_avg = conv_params->use_jnt_comp_avg;
const int fo_vert = filter_params_y->taps / 2 - 1;
const int fo_horiz = filter_params_x->taps / 2 - 1;
const uint16_t *const src_ptr = src - fo_vert * src_stride - fo_horiz;
const int w0 = conv_params->fwd_offset;
const int w1 = conv_params->bck_offset;
const __m128i wt0 = _mm_set1_epi32(w0);
const __m128i wt1 = _mm_set1_epi32(w1);
const int offset_0 =
bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
const int offset = (1 << offset_0) + (1 << (offset_0 - 1));
const __m128i offset_const = _mm_set1_epi32(offset);
const int rounding_shift =
2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
const __m128i rounding_const = _mm_set1_epi32((1 << rounding_shift) >> 1);
const __m128i clip_pixel_to_bd =
_mm_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
// Check that, even with 12-bit input, the intermediate values will fit
// into an unsigned 16-bit intermediate array.
assert(bd + FILTER_BITS + 2 - conv_params->round_0 <= 16);
/* Horizontal filter */
{
const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
filter_params_x, subpel_x_q4 & SUBPEL_MASK);
const __m128i coeffs_x = _mm_loadu_si128((__m128i *)x_filter);
// coeffs 0 1 0 1 2 3 2 3
const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_x, coeffs_x);
// coeffs 4 5 4 5 6 7 6 7
const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_x, coeffs_x);
// coeffs 0 1 0 1 0 1 0 1
const __m128i coeff_01 = _mm_unpacklo_epi64(tmp_0, tmp_0);
// coeffs 2 3 2 3 2 3 2 3
const __m128i coeff_23 = _mm_unpackhi_epi64(tmp_0, tmp_0);
// coeffs 4 5 4 5 4 5 4 5
const __m128i coeff_45 = _mm_unpacklo_epi64(tmp_1, tmp_1);
// coeffs 6 7 6 7 6 7 6 7
const __m128i coeff_67 = _mm_unpackhi_epi64(tmp_1, tmp_1);
const __m128i round_const = _mm_set1_epi32(
((1 << conv_params->round_0) >> 1) + (1 << (bd + FILTER_BITS - 1)));
const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_0);
for (i = 0; i < im_h; ++i) {
for (j = 0; j < w; j += 8) {
const __m128i data =
_mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]);
const __m128i data2 =
_mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j + 8]);
// Filter even-index pixels
const __m128i res_0 = _mm_madd_epi16(data, coeff_01);
const __m128i res_2 =
_mm_madd_epi16(_mm_alignr_epi8(data2, data, 4), coeff_23);
const __m128i res_4 =
_mm_madd_epi16(_mm_alignr_epi8(data2, data, 8), coeff_45);
const __m128i res_6 =
_mm_madd_epi16(_mm_alignr_epi8(data2, data, 12), coeff_67);
__m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_4),
_mm_add_epi32(res_2, res_6));
res_even =
_mm_sra_epi32(_mm_add_epi32(res_even, round_const), round_shift);
// Filter odd-index pixels
const __m128i res_1 =
_mm_madd_epi16(_mm_alignr_epi8(data2, data, 2), coeff_01);
const __m128i res_3 =
_mm_madd_epi16(_mm_alignr_epi8(data2, data, 6), coeff_23);
const __m128i res_5 =
_mm_madd_epi16(_mm_alignr_epi8(data2, data, 10), coeff_45);
const __m128i res_7 =
_mm_madd_epi16(_mm_alignr_epi8(data2, data, 14), coeff_67);
__m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_5),
_mm_add_epi32(res_3, res_7));
res_odd =
_mm_sra_epi32(_mm_add_epi32(res_odd, round_const), round_shift);
// Pack in the column order 0, 2, 4, 6, 1, 3, 5, 7
__m128i res = _mm_packs_epi32(res_even, res_odd);
_mm_storeu_si128((__m128i *)&im_block[i * im_stride + j], res);
}
}
}
/* Vertical filter */
{
const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
filter_params_y, subpel_y_q4 & SUBPEL_MASK);
const __m128i coeffs_y = _mm_loadu_si128((__m128i *)y_filter);
// coeffs 0 1 0 1 2 3 2 3
const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_y, coeffs_y);
// coeffs 4 5 4 5 6 7 6 7
const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_y, coeffs_y);
// coeffs 0 1 0 1 0 1 0 1
const __m128i coeff_01 = _mm_unpacklo_epi64(tmp_0, tmp_0);
// coeffs 2 3 2 3 2 3 2 3
const __m128i coeff_23 = _mm_unpackhi_epi64(tmp_0, tmp_0);
// coeffs 4 5 4 5 4 5 4 5
const __m128i coeff_45 = _mm_unpacklo_epi64(tmp_1, tmp_1);
// coeffs 6 7 6 7 6 7 6 7
const __m128i coeff_67 = _mm_unpackhi_epi64(tmp_1, tmp_1);
const __m128i round_const = _mm_set1_epi32(
((1 << conv_params->round_1) >> 1) -
(1 << (bd + 2 * FILTER_BITS - conv_params->round_0 - 1)));
const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_1);
for (i = 0; i < h; ++i) {
for (j = 0; j < w; j += 8) {
// Filter even-index pixels
const int16_t *data = &im_block[i * im_stride + j];
const __m128i src_0 =
_mm_unpacklo_epi16(*(__m128i *)(data + 0 * im_stride),
*(__m128i *)(data + 1 * im_stride));
const __m128i src_2 =
_mm_unpacklo_epi16(*(__m128i *)(data + 2 * im_stride),
*(__m128i *)(data + 3 * im_stride));
const __m128i src_4 =
_mm_unpacklo_epi16(*(__m128i *)(data + 4 * im_stride),
*(__m128i *)(data + 5 * im_stride));
const __m128i src_6 =
_mm_unpacklo_epi16(*(__m128i *)(data + 6 * im_stride),
*(__m128i *)(data + 7 * im_stride));
const __m128i res_0 = _mm_madd_epi16(src_0, coeff_01);
const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23);
const __m128i res_4 = _mm_madd_epi16(src_4, coeff_45);
const __m128i res_6 = _mm_madd_epi16(src_6, coeff_67);
const __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_2),
_mm_add_epi32(res_4, res_6));
// Filter odd-index pixels
const __m128i src_1 =
_mm_unpackhi_epi16(*(__m128i *)(data + 0 * im_stride),
*(__m128i *)(data + 1 * im_stride));
const __m128i src_3 =
_mm_unpackhi_epi16(*(__m128i *)(data + 2 * im_stride),
*(__m128i *)(data + 3 * im_stride));
const __m128i src_5 =
_mm_unpackhi_epi16(*(__m128i *)(data + 4 * im_stride),
*(__m128i *)(data + 5 * im_stride));
const __m128i src_7 =
_mm_unpackhi_epi16(*(__m128i *)(data + 6 * im_stride),
*(__m128i *)(data + 7 * im_stride));
const __m128i res_1 = _mm_madd_epi16(src_1, coeff_01);
const __m128i res_3 = _mm_madd_epi16(src_3, coeff_23);
const __m128i res_5 = _mm_madd_epi16(src_5, coeff_45);
const __m128i res_7 = _mm_madd_epi16(src_7, coeff_67);
const __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_3),
_mm_add_epi32(res_5, res_7));
// Rearrange pixels back into the order 0 ... 7
const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd);
const __m128i res_hi = _mm_unpackhi_epi32(res_even, res_odd);
const __m128i res_lo_round =
_mm_sra_epi32(_mm_add_epi32(res_lo, round_const), round_shift);
const __m128i res_unsigned_lo =
_mm_add_epi32(res_lo_round, offset_const);
if (w < 8) {
if (do_average) {
const __m128i data_0 =
_mm_loadl_epi64((__m128i *)(&dst[i * dst_stride + j]));
const __m128i data_ref_0 = _mm_cvtepu16_epi32(data_0);
const __m128i comp_avg_res = highbd_comp_avg_sse4_1(
&data_ref_0, &res_unsigned_lo, &wt0, &wt1, use_jnt_comp_avg);
const __m128i round_result = highbd_convolve_rounding_sse2(
&comp_avg_res, &offset_const, &rounding_const, rounding_shift);
const __m128i res_16b =
_mm_packus_epi32(round_result, round_result);
const __m128i res_clip = _mm_min_epi16(res_16b, clip_pixel_to_bd);
_mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_clip);
} else {
const __m128i res_16b =
_mm_packus_epi32(res_unsigned_lo, res_unsigned_lo);
_mm_storel_epi64((__m128i *)(&dst[i * dst_stride + j]), res_16b);
}
} else {
const __m128i res_hi_round =
_mm_sra_epi32(_mm_add_epi32(res_hi, round_const), round_shift);
const __m128i res_unsigned_hi =
_mm_add_epi32(res_hi_round, offset_const);
if (do_average) {
const __m128i data_lo =
_mm_loadl_epi64((__m128i *)(&dst[i * dst_stride + j]));
const __m128i data_hi =
_mm_loadl_epi64((__m128i *)(&dst[i * dst_stride + j + 4]));
const __m128i data_ref_0_lo = _mm_cvtepu16_epi32(data_lo);
const __m128i data_ref_0_hi = _mm_cvtepu16_epi32(data_hi);
const __m128i comp_avg_res_lo = highbd_comp_avg_sse4_1(
&data_ref_0_lo, &res_unsigned_lo, &wt0, &wt1, use_jnt_comp_avg);
const __m128i comp_avg_res_hi = highbd_comp_avg_sse4_1(
&data_ref_0_hi, &res_unsigned_hi, &wt0, &wt1, use_jnt_comp_avg);
const __m128i round_result_lo =
highbd_convolve_rounding_sse2(&comp_avg_res_lo, &offset_const,
&rounding_const, rounding_shift);
const __m128i round_result_hi =
highbd_convolve_rounding_sse2(&comp_avg_res_hi, &offset_const,
&rounding_const, rounding_shift);
const __m128i res_16b =
_mm_packus_epi32(round_result_lo, round_result_hi);
const __m128i res_clip = _mm_min_epi16(res_16b, clip_pixel_to_bd);
_mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j]), res_clip);
} else {
const __m128i res_16b =
_mm_packus_epi32(res_unsigned_lo, res_unsigned_hi);
_mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_16b);
}
}
}
}
}
}
