static void filter_by_weight16x16_msa(const uint8_t *src_ptr,
int32_t src_stride, uint8_t *dst_ptr,
int32_t dst_stride, int32_t src_weight) {
int32_t dst_weight = (1 << MFQE_PRECISION) - src_weight;
int32_t row;
v16i8 src0, src1, src2, src3, dst0, dst1, dst2, dst3;
v8i16 src_wt, dst_wt, res_h_r, res_h_l, src_r, src_l, dst_r, dst_l;
src_wt = __msa_fill_h(src_weight);
dst_wt = __msa_fill_h(dst_weight);
for (row = 4; row--;) {
LD_SB4(src_ptr, src_stride, src0, src1, src2, src3);
src_ptr += (4 * src_stride);
LD_SB4(dst_ptr, dst_stride, dst0, dst1, dst2, dst3);
UNPCK_UB_SH(src0, src_r, src_l);
UNPCK_UB_SH(dst0, dst_r, dst_l);
res_h_r = (src_r * src_wt);
res_h_r += (dst_r * dst_wt);
res_h_l = (src_l * src_wt);
res_h_l += (dst_l * dst_wt);
SRARI_H2_SH(res_h_r, res_h_l, MFQE_PRECISION);
PCKEV_ST_SB(res_h_r, res_h_l, dst_ptr);
dst_ptr += dst_stride;
UNPCK_UB_SH(src1, src_r, src_l);
UNPCK_UB_SH(dst1, dst_r, dst_l);
res_h_r = (src_r * src_wt);
res_h_r += (dst_r * dst_wt);
res_h_l = (src_l * src_wt);
res_h_l += (dst_l * dst_wt);
SRARI_H2_SH(res_h_r, res_h_l, MFQE_PRECISION);
PCKEV_ST_SB(res_h_r, res_h_l, dst_ptr);
dst_ptr += dst_stride;
UNPCK_UB_SH(src2, src_r, src_l);
UNPCK_UB_SH(dst2, dst_r, dst_l);
res_h_r = (src_r * src_wt);
res_h_r += (dst_r * dst_wt);
res_h_l = (src_l * src_wt);
res_h_l += (dst_l * dst_wt);
SRARI_H2_SH(res_h_r, res_h_l, MFQE_PRECISION);
PCKEV_ST_SB(res_h_r, res_h_l, dst_ptr);
dst_ptr += dst_stride;
UNPCK_UB_SH(src3, src_r, src_l);
UNPCK_UB_SH(dst3, dst_r, dst_l);
res_h_r = (src_r * src_wt);
res_h_r += (dst_r * dst_wt);
res_h_l = (src_l * src_wt);
res_h_l += (dst_l * dst_wt);
SRARI_H2_SH(res_h_r, res_h_l, MFQE_PRECISION);
PCKEV_ST_SB(res_h_r, res_h_l, dst_ptr);
dst_ptr += dst_stride;
}
}
static WEBP_INLINE void ITransformOne(const uint8_t* ref, const int16_t* in,
uint8_t* dst) {
v8i16 input0, input1;
v4i32 in0, in1, in2, in3, hz0, hz1, hz2, hz3, vt0, vt1, vt2, vt3;
v4i32 res0, res1, res2, res3;
v16i8 dest0, dest1, dest2, dest3;
const v16i8 zero = { 0 };
LD_SH2(in, 8, input0, input1);
UNPCK_SH_SW(input0, in0, in1);
UNPCK_SH_SW(input1, in2, in3);
IDCT_1D_W(in0, in1, in2, in3, hz0, hz1, hz2, hz3);
TRANSPOSE4x4_SW_SW(hz0, hz1, hz2, hz3, hz0, hz1, hz2, hz3);
IDCT_1D_W(hz0, hz1, hz2, hz3, vt0, vt1, vt2, vt3);
SRARI_W4_SW(vt0, vt1, vt2, vt3, 3);
TRANSPOSE4x4_SW_SW(vt0, vt1, vt2, vt3, vt0, vt1, vt2, vt3);
LD_SB4(ref, BPS, dest0, dest1, dest2, dest3);
ILVR_B4_SW(zero, dest0, zero, dest1, zero, dest2, zero, dest3,
res0, res1, res2, res3);
ILVR_H4_SW(zero, res0, zero, res1, zero, res2, zero, res3,
res0, res1, res2, res3);
ADD4(res0, vt0, res1, vt1, res2, vt2, res3, vt3, res0, res1, res2, res3);
CLIP_SW4_0_255(res0, res1, res2, res3);
PCKEV_B2_SW(res0, res1, res2, res3, vt0, vt1);
res0 = (v4i32)__msa_pckev_b((v16i8)vt0, (v16i8)vt1);
ST4x4_UB(res0, res0, 3, 2, 1, 0, dst, BPS);
}
