static void common_hv_8ht_8vt_and_aver_dst_4w_msa(const uint8_t *src,
int32_t src_stride,
uint8_t *dst,
int32_t dst_stride,
int8_t *filter_horiz,
int8_t *filter_vert,
int32_t height) {
uint32_t loop_cnt;
v16i8 src0, src1, src2, src3, src4, src5, src6, src7, src8, src9, src10;
v16u8 dst0, dst1, dst2, dst3, mask0, mask1, mask2, mask3, tmp0, tmp1;
v16i8 filt_hz0, filt_hz1, filt_hz2, filt_hz3;
v8i16 hz_out0, hz_out1, hz_out2, hz_out3, hz_out4, hz_out5, hz_out6;
v8i16 hz_out7, hz_out8, hz_out9, res0, res1, vec0, vec1, vec2, vec3, vec4;
v8i16 filt, filt_vt0, filt_vt1, filt_vt2, filt_vt3;
mask0 = LD_UB(&mc_filt_mask_arr[16]);
src -= (3 + 3 * src_stride);
/* rearranging filter */
filt = LD_SH(filter_horiz);
SPLATI_H4_SB(filt, 0, 1, 2, 3, filt_hz0, filt_hz1, filt_hz2, filt_hz3);
mask1 = mask0 + 2;
mask2 = mask0 + 4;
mask3 = mask0 + 6;
LD_SB7(src, src_stride, src0, src1, src2, src3, src4, src5, src6);
XORI_B7_128_SB(src0, src1, src2, src3, src4, src5, src6);
src += (7 * src_stride);
hz_out0 = HORIZ_8TAP_FILT(src0, src1, mask0, mask1, mask2, mask3, filt_hz0,
filt_hz1, filt_hz2, filt_hz3);
hz_out2 = HORIZ_8TAP_FILT(src2, src3, mask0, mask1, mask2, mask3, filt_hz0,
filt_hz1, filt_hz2, filt_hz3);
hz_out4 = HORIZ_8TAP_FILT(src4, src5, mask0, mask1, mask2, mask3, filt_hz0,
filt_hz1, filt_hz2, filt_hz3);
hz_out5 = HORIZ_8TAP_FILT(src5, src6, mask0, mask1, mask2, mask3, filt_hz0,
filt_hz1, filt_hz2, filt_hz3);
SLDI_B2_SH(hz_out2, hz_out4, hz_out0, hz_out2, hz_out1, hz_out3, 8);
filt = LD_SH(filter_vert);
SPLATI_H4_SH(filt, 0, 1, 2, 3, filt_vt0, filt_vt1, filt_vt2, filt_vt3);
ILVEV_B2_SH(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
vec2 = (v8i16)__msa_ilvev_b((v16i8)hz_out5, (v16i8)hz_out4);
for (loop_cnt = (height >> 2); loop_cnt--;) {
filt = LD_SH(filter_vert);
SPLATI_H4_SH(filt, 0, 1, 2, 3, filt_vt0, filt_vt1, filt_vt2, filt_vt3);
ILVEV_B2_SH(hz_out0, hz_out1, hz_out2, hz_out3, out0, out1);
out2 = (v8i16)__msa_ilvev_b((v16i8)hz_out5, (v16i8)hz_out4);
for (loop_cnt = (height >> 2); loop_cnt--;) {
LD_SB4(src, src_stride, src7, src8, src9, src10);
XORI_B4_128_SB(src7, src8, src9, src10);
src += (4 * src_stride);
hz_out7 = HORIZ_8TAP_FILT(src7, src8, mask0, mask1, mask2, mask3,
filt_hz0, filt_hz1, filt_hz2, filt_hz3);
hz_out6 = (v8i16)__msa_sldi_b((v16i8)hz_out7, (v16i8)hz_out5, 8);
out3 = (v8i16)__msa_ilvev_b((v16i8)hz_out7, (v16i8)hz_out6);
tmp0 = FILT_8TAP_DPADD_S_H(out0, out1, out2, out3, filt_vt0, filt_vt1,
filt_vt2, filt_vt3);
hz_out9 = HORIZ_8TAP_FILT(src9, src10, mask0, mask1, mask2, mask3,
filt_hz0, filt_hz1, filt_hz2, filt_hz3);
hz_out8 = (v8i16)__msa_sldi_b((v16i8)hz_out9, (v16i8)hz_out7, 8);
out4 = (v8i16)__msa_ilvev_b((v16i8)hz_out9, (v16i8)hz_out8);
tmp1 = FILT_8TAP_DPADD_S_H(out1, out2, out3, out4, filt_vt0, filt_vt1,
filt_vt2, filt_vt3);
SRARI_H2_SH(tmp0, tmp1, FILTER_BITS);
SAT_SH2_SH(tmp0, tmp1, 7);
out = PCKEV_XORI128_UB(tmp0, tmp1);
ST4x4_UB(out, out, 0, 1, 2, 3, dst, dst_stride);
dst += (4 * dst_stride);
filt = LD_SH(filter_vert);
SPLATI_H4_SH(filt, 0, 1, 2, 3, filt_vt0, filt_vt1, filt_vt2, filt_vt3);
ILVEV_B2_SH(hz_out0, hz_out1, hz_out2, hz_out3, vec0, vec1);
vec2 = (v8i16)__msa_ilvev_b((v16i8)hz_out5, (v16i8)hz_out4);
for (loop_cnt = (height >> 2); loop_cnt--;) {
LD_SB4(src, src_stride, src7, src8, src9, src10);
XORI_B4_128_SB(src7, src8, src9, src10);
src += (4 * src_stride);
LD_UB4(dst, dst_stride, dst0, dst1, dst2, dst3);
hz_out7 = HORIZ_8TAP_FILT(src7, src8, mask0, mask1, mask2, mask3,
filt_hz0, filt_hz1, filt_hz2, filt_hz3);
hz_out6 = (v8i16)__msa_sldi_b((v16i8)hz_out7, (v16i8)hz_out5, 8);
vec3 = (v8i16)__msa_ilvev_b((v16i8)hz_out7, (v16i8)hz_out6);
res0 = FILT_8TAP_DPADD_S_H(vec0, vec1, vec2, vec3, filt_vt0, filt_vt1,
filt_vt2, filt_vt3);
hz_out9 = HORIZ_8TAP_FILT(src9, src10, mask0, mask1, mask2, mask3,
filt_hz0, filt_hz1, filt_hz2, filt_hz3);
hz_out8 = (v8i16)__msa_sldi_b((v16i8)hz_out9, (v16i8)hz_out7, 8);
vec4 = (v8i16)__msa_ilvev_b((v16i8)hz_out9, (v16i8)hz_out8);
res1 = FILT_8TAP_DPADD_S_H(vec1, vec2, vec3, vec4, filt_vt0, filt_vt1,
filt_vt2, filt_vt3);
ILVR_W2_UB(dst1, dst0, dst3, dst2, dst0, dst2);
SRARI_H2_SH(res0, res1, FILTER_BITS);
SAT_SH2_SH(res0, res1, 7);
PCKEV_B2_UB(res0, res0, res1, res1, tmp0, tmp1);
XORI_B2_128_UB(tmp0, tmp1);
