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);
}
static int32_t avc_quant_8x8_msa( int16_t *p_dct, uint16_t *p_mf,
uint16_t *p_bias )
{
int32_t non_zero = 0;
v8i16 dct0, dct1, dct2, dct3;
v8i16 zero = { 0 };
v8i16 dct0_mask, dct1_mask, dct2_mask, dct3_mask;
v8i16 dct_h0, dct_h1, dct_h2, dct_h3, mf_h0, mf_h1, mf_h2, mf_h3;
v8i16 bias_h0, bias_h1, bias_h2, bias_h3;
v4i32 dct_w0, dct_w1, dct_w2, dct_w3, dct_w4, dct_w5, dct_w6, dct_w7;
v4i32 dct_signed_w0, dct_signed_w1, dct_signed_w2, dct_signed_w3;
v4i32 dct_signed_w4, dct_signed_w5, dct_signed_w6, dct_signed_w7;
v4i32 mf_vec0, mf_vec1, mf_vec2, mf_vec3;
v4i32 mf_vec4, mf_vec5, mf_vec6, mf_vec7;
v4i32 bias0, bias1, bias2, bias3, bias4, bias5, bias6, bias7;
LD_SH4( p_dct, 8, dct0, dct1, dct2, dct3 );
dct0_mask = __msa_clei_s_h( dct0, 0 );
dct1_mask = __msa_clei_s_h( dct1, 0 );
dct2_mask = __msa_clei_s_h( dct2, 0 );
dct3_mask = __msa_clei_s_h( dct3, 0 );
UNPCK_SH_SW( dct0, dct_signed_w0, dct_signed_w1 );
UNPCK_SH_SW( dct1, dct_signed_w2, dct_signed_w3 );
UNPCK_SH_SW( dct2, dct_signed_w4, dct_signed_w5 );
UNPCK_SH_SW( dct3, dct_signed_w6, dct_signed_w7 );
LD_SH4( p_bias, 8, bias_h0, bias_h1, bias_h2, bias_h3 );
ILVR_H4_SW( zero, bias_h0, zero, bias_h1, zero, bias_h2, zero, bias_h3,
bias0, bias2, bias4, bias6 );
ILVL_H4_SW( zero, bias_h0, zero, bias_h1, zero, bias_h2, zero, bias_h3,
bias1, bias3, bias5, bias7 );
LD_SH4( p_mf, 8, mf_h0, mf_h1, mf_h2, mf_h3 );
ILVR_H4_SW( zero, mf_h0, zero, mf_h1, zero, mf_h2, zero, mf_h3,
mf_vec0, mf_vec2, mf_vec4, mf_vec6 );
ILVL_H4_SW( zero, mf_h0, zero, mf_h1, zero, mf_h2, zero, mf_h3,
mf_vec1, mf_vec3, mf_vec5, mf_vec7 );
dct_w0 = __msa_add_a_w( dct_signed_w0, bias0 );
dct_w1 = __msa_add_a_w( dct_signed_w1, bias1 );
dct_w2 = __msa_add_a_w( dct_signed_w2, bias2 );
dct_w3 = __msa_add_a_w( dct_signed_w3, bias3 );
dct_w4 = __msa_add_a_w( dct_signed_w4, bias4 );
dct_w5 = __msa_add_a_w( dct_signed_w5, bias5 );
dct_w6 = __msa_add_a_w( dct_signed_w6, bias6 );
dct_w7 = __msa_add_a_w( dct_signed_w7, bias7 );
dct_w0 *= mf_vec0;
dct_w1 *= mf_vec1;
dct_w2 *= mf_vec2;
dct_w3 *= mf_vec3;
dct_w4 *= mf_vec4;
dct_w5 *= mf_vec5;
dct_w6 *= mf_vec6;
dct_w7 *= mf_vec7;
SRA_4V( dct_w0, dct_w1, dct_w2, dct_w3, 16 );
SRA_4V( dct_w4, dct_w5, dct_w6, dct_w7, 16 );
PCKEV_H4_SH( dct_w1, dct_w0, dct_w3, dct_w2, dct_w5, dct_w4, dct_w7, dct_w6,
dct_h0, dct_h1, dct_h2, dct_h3 );
SUB4( zero, dct_h0, zero, dct_h1, zero, dct_h2, zero, dct_h3,
dct0, dct1, dct2, dct3 );
dct0 = ( v8i16 ) __msa_bmnz_v( ( v16u8 ) dct_h0,
( v16u8 ) dct0, ( v16u8 ) dct0_mask );
dct1 = ( v8i16 ) __msa_bmnz_v( ( v16u8 ) dct_h1,
( v16u8 ) dct1, ( v16u8 ) dct1_mask );
dct2 = ( v8i16 ) __msa_bmnz_v( ( v16u8 ) dct_h2,
( v16u8 ) dct2, ( v16u8 ) dct2_mask );
dct3 = ( v8i16 ) __msa_bmnz_v( ( v16u8 ) dct_h3,
( v16u8 ) dct3, ( v16u8 ) dct3_mask );
non_zero = HADD_SW_S32( ( v4u32 )( dct_h0 + dct_h1 + dct_h2 + dct_h3 ) );
ST_SH4( dct0, dct1, dct2, dct3, p_dct, 8 );
LD_SH4( p_dct + 32, 8, dct0, dct1, dct2, dct3 );
dct0_mask = __msa_clei_s_h( dct0, 0 );
dct1_mask = __msa_clei_s_h( dct1, 0 );
dct2_mask = __msa_clei_s_h( dct2, 0 );
dct3_mask = __msa_clei_s_h( dct3, 0 );
UNPCK_SH_SW( dct0, dct_signed_w0, dct_signed_w1 );
UNPCK_SH_SW( dct1, dct_signed_w2, dct_signed_w3 );
UNPCK_SH_SW( dct2, dct_signed_w4, dct_signed_w5 );
UNPCK_SH_SW( dct3, dct_signed_w6, dct_signed_w7 );
LD_SH4( p_bias + 32, 8, bias_h0, bias_h1, bias_h2, bias_h3 );
ILVR_H4_SW( zero, bias_h0, zero, bias_h1, zero, bias_h2, zero, bias_h3,
bias0, bias2, bias4, bias6 );
ILVL_H4_SW( zero, bias_h0, zero, bias_h1, zero, bias_h2, zero, bias_h3,
bias1, bias3, bias5, bias7 );
LD_SH4( p_mf + 32, 8, mf_h0, mf_h1, mf_h2, mf_h3 );
ILVR_H4_SW( zero, mf_h0, zero, mf_h1, zero, mf_h2, zero, mf_h3,
mf_vec0, mf_vec2, mf_vec4, mf_vec6 );
ILVL_H4_SW( zero, mf_h0, zero, mf_h1, zero, mf_h2, zero, mf_h3,
mf_vec1, mf_vec3, mf_vec5, mf_vec7 );
dct_w0 = __msa_add_a_w( dct_signed_w0, bias0 );
dct_w1 = __msa_add_a_w( dct_signed_w1, bias1 );
dct_w2 = __msa_add_a_w( dct_signed_w2, bias2 );
dct_w3 = __msa_add_a_w( dct_signed_w3, bias3 );
dct_w4 = __msa_add_a_w( dct_signed_w4, bias4 );
dct_w5 = __msa_add_a_w( dct_signed_w5, bias5 );
dct_w6 = __msa_add_a_w( dct_signed_w6, bias6 );
dct_w7 = __msa_add_a_w( dct_signed_w7, bias7 );
dct_w0 *= mf_vec0;
dct_w1 *= mf_vec1;
dct_w2 *= mf_vec2;
dct_w3 *= mf_vec3;
dct_w4 *= mf_vec4;
dct_w5 *= mf_vec5;
dct_w6 *= mf_vec6;
dct_w7 *= mf_vec7;
SRA_4V( dct_w0, dct_w1, dct_w2, dct_w3, 16 );
SRA_4V( dct_w4, dct_w5, dct_w6, dct_w7, 16 );
PCKEV_H2_SH( dct_w1, dct_w0, dct_w3, dct_w2, dct_h0, dct_h1 );
PCKEV_H2_SH( dct_w5, dct_w4, dct_w7, dct_w6, dct_h2, dct_h3 );
SUB4( zero, dct_h0, zero, dct_h1, zero, dct_h2, zero, dct_h3,
dct0, dct1, dct2, dct3 );
dct0 = ( v8i16 ) __msa_bmnz_v( ( v16u8 ) dct_h0,
( v16u8 ) dct0, ( v16u8 ) dct0_mask );
dct1 = ( v8i16 ) __msa_bmnz_v( ( v16u8 ) dct_h1,
( v16u8 ) dct1, ( v16u8 ) dct1_mask );
dct2 = ( v8i16 ) __msa_bmnz_v( ( v16u8 ) dct_h2,
( v16u8 ) dct2, ( v16u8 ) dct2_mask );
dct3 = ( v8i16 ) __msa_bmnz_v( ( v16u8 ) dct_h3,
( v16u8 ) dct3, ( v16u8 ) dct3_mask );
non_zero += HADD_SW_S32( ( v4u32 ) ( dct_h0 + dct_h1 + dct_h2 + dct_h3 ) );
ST_SH4( dct0, dct1, dct2, dct3, p_dct + 32, 8 );
return !!non_zero;
}
