int32_t vp8_block_error_msa(int16_t *coeff_ptr, int16_t *dq_coeff_ptr) { int32_t err = 0; uint32_t loop_cnt; v8i16 coeff, dq_coeff, coeff0, coeff1; v4i32 diff0, diff1; v2i64 err0 = { 0 }; v2i64 err1 = { 0 }; for (loop_cnt = 2; loop_cnt--;) { coeff = LD_SH(coeff_ptr); dq_coeff = LD_SH(dq_coeff_ptr); ILVRL_H2_SH(coeff, dq_coeff, coeff0, coeff1); HSUB_UH2_SW(coeff0, coeff1, diff0, diff1); DPADD_SD2_SD(diff0, diff1, err0, err1); coeff_ptr += 8; dq_coeff_ptr += 8; } err0 += __msa_splati_d(err0, 1); err1 += __msa_splati_d(err1, 1); err = __msa_copy_s_d(err0, 0); err += __msa_copy_s_d(err1, 0); return err; }
src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1); src_c8 -= src_a0 * src_b; src_c9 -= src_a1 * src_b; src_c10 -= src_a2 * src_b; src_c11 -= src_a3 * src_b; src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1); src_c12 -= src_a0 * src_b; src_c13 -= src_a1 * src_b; src_c14 -= src_a2 * src_b; src_c15 -= src_a3 * src_b; } } src_b0 = LD_DP(b + 0); src_b1 = (v2f64) __msa_splati_d((v2i64) src_b0, 1); src_b0 = (v2f64) __msa_splati_d((v2i64) src_b0, 0); src_b2 = LD_DP(b + 2); src_b3 = (v2f64) __msa_splati_d((v2i64) src_b2, 1); src_b2 = (v2f64) __msa_splati_d((v2i64) src_b2, 0); src_b5 = __msa_cast_to_vector_double(*(b + 5)); src_b5 = (v2f64) __msa_splati_d((v2i64) src_b5, 0); src_b6 = LD_DP(b + 6); src_b7 = (v2f64) __msa_splati_d((v2i64) src_b6, 1); src_b6 = (v2f64) __msa_splati_d((v2i64) src_b6, 0); src_b10 = LD_DP(b + 10); src_b11 = (v2f64) __msa_splati_d((v2i64) src_b10, 1); src_b10 = (v2f64) __msa_splati_d((v2i64) src_b10, 0); src_b15 = __msa_cast_to_vector_double(*(b + 15)); src_b15 = (v2f64) __msa_splati_d((v2i64) src_b15, 0);
static void dsolve_4x2_ln_msa(FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG bk) { v2f64 src_c0, src_c1, src_c2, src_c3, res_c0, res_c1, res_c2, res_c3; v2f64 src_a0, src_a4, src_a5, src_a8, src_a9, src_a10, src_a12, src_a13; v2f64 src_a14, src_a15; LD_DP2(c, 2, src_c0, src_c1); LD_DP2(c + ldc, 2, src_c2, src_c3); if (bk > 0) { BLASLONG i; FLOAT *aa = a, *bb = b; v2f64 src_a0, src_a1, src_b, src_b0; for (i = bk; i--;) { LD_DP2(aa, 2, src_a0, src_a1); src_b0 = LD_DP(bb); src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0); src_c0 -= src_a0 * src_b; src_c1 -= src_a1 * src_b; src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0); src_c2 -= src_a0 * src_b; src_c3 -= src_a1 * src_b; aa += 4; bb += 2; } } a -= 16; b -= 8; ILVRL_D2_DP(src_c2, src_c0, res_c0, res_c1); ILVRL_D2_DP(src_c3, src_c1, res_c2, res_c3); src_a14 = LD_DP(a + 14); src_a15 = (v2f64) __msa_splati_d((v2i64) src_a14, 1); src_a14 = (v2f64) __msa_splati_d((v2i64) src_a14, 0); src_a12 = LD_DP(a + 12); src_a13 = (v2f64) __msa_splati_d((v2i64) src_a12, 1); src_a12 = (v2f64) __msa_splati_d((v2i64) src_a12, 0); src_a9 = LD_DP(a + 9); src_a10 = (v2f64) __msa_splati_d((v2i64) src_a9, 1); src_a9 = (v2f64) __msa_splati_d((v2i64) src_a9, 0); src_a8 = __msa_cast_to_vector_double(*(a + 8)); src_a0 = __msa_cast_to_vector_double(*(a + 0)); src_a8 = (v2f64) __msa_splati_d((v2i64) src_a8, 0); src_a0 = (v2f64) __msa_splati_d((v2i64) src_a0, 0); src_a4 = LD_DP(a + 4); src_a5 = (v2f64) __msa_splati_d((v2i64) src_a4, 1); src_a4 = (v2f64) __msa_splati_d((v2i64) src_a4, 0); res_c3 *= src_a15; res_c2 -= res_c3 * src_a14; res_c2 *= src_a10; res_c1 -= res_c3 * src_a13; res_c1 -= res_c2 * src_a9; res_c1 *= src_a5; res_c0 -= res_c3 * src_a12; res_c0 -= res_c2 * src_a8; res_c0 -= res_c1 * src_a4; res_c0 *= src_a0; ST_DP(res_c3, b + 6); ST_DP(res_c2, b + 4); ST_DP(res_c1, b + 2); ST_DP(res_c0, b + 0); ILVRL_D2_DP(res_c1, res_c0, src_c0, src_c2); ILVRL_D2_DP(res_c3, res_c2, src_c1, src_c3); ST_DP2(src_c0, src_c1, c, 2); ST_DP2(src_c2, src_c3, c + ldc, 2); }
static void dsolve_8x2_ln_msa(FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG bk) { v2f64 src_c0, src_c1, src_c2, src_c3, src_c4, src_c5, src_c6, src_c7; v2f64 res_c0, res_c1, res_c2, res_c3, res_c4, res_c5, res_c6, res_c7; v2f64 src_a0, src_a1, src_a2, src_a3, src_a8, src_a9, src_a16, src_a17; v2f64 src_a18, src_a24, src_a25, src_a26, src_a27, src_a32, src_a33; v2f64 src_a34, src_a35, src_a36, src_a40, src_a41, src_a42, src_a43; v2f64 src_a44, src_a45, src_a48, src_a49, src_a50, src_a51, src_a52; v2f64 src_a53, src_a54, src_a56, src_a57, src_a58, src_a59, src_a60; v2f64 src_a61, src_a62, src_a63; LD_DP4(c, 2, src_c0, src_c1, src_c2, src_c3); LD_DP4(c + ldc, 2, src_c4, src_c5, src_c6, src_c7); if (bk > 0) { BLASLONG i; FLOAT *pba = a, *pbb = b; v2f64 src_b, src_b0, src_b1; LD_DP4(pba, 2, src_a0, src_a1, src_a2, src_a3); src_b0 = LD_DP(pbb); for (i = bk - 1; i--;) { pba += 8; pbb += 2; LD_DP4(pba, 2, src_a8, src_a9, src_a16, src_a17); src_b1 = LD_DP(pbb); src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0); src_c0 -= src_a0 * src_b; src_c1 -= src_a1 * src_b; src_c2 -= src_a2 * src_b; src_c3 -= src_a3 * src_b; src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0); src_c4 -= src_a0 * src_b; src_c5 -= src_a1 * src_b; src_c6 -= src_a2 * src_b; src_c7 -= src_a3 * src_b; src_a0 = src_a8; src_a1 = src_a9; src_a2 = src_a16; src_a3 = src_a17; src_b0 = src_b1; } src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0); src_c0 -= src_a0 * src_b; src_c1 -= src_a1 * src_b; src_c2 -= src_a2 * src_b; src_c3 -= src_a3 * src_b; src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0); src_c4 -= src_a0 * src_b; src_c5 -= src_a1 * src_b; src_c6 -= src_a2 * src_b; src_c7 -= src_a3 * src_b; } ILVRL_D2_DP(src_c4, src_c0, res_c0, res_c1); ILVRL_D2_DP(src_c5, src_c1, res_c2, res_c3); ILVRL_D2_DP(src_c6, src_c2, res_c4, res_c5); ILVRL_D2_DP(src_c7, src_c3, res_c6, res_c7); src_a56 = LD_DP(a - 8); src_a57 = (v2f64) __msa_splati_d((v2i64) src_a56, 1); src_a56 = (v2f64) __msa_splati_d((v2i64) src_a56, 0); src_a58 = LD_DP(a - 6); src_a59 = (v2f64) __msa_splati_d((v2i64) src_a58, 1); src_a58 = (v2f64) __msa_splati_d((v2i64) src_a58, 0); src_a60 = LD_DP(a - 4); src_a61 = (v2f64) __msa_splati_d((v2i64) src_a60, 1); src_a60 = (v2f64) __msa_splati_d((v2i64) src_a60, 0); src_a62 = LD_DP(a - 2); src_a63 = (v2f64) __msa_splati_d((v2i64) src_a62, 1); src_a62 = (v2f64) __msa_splati_d((v2i64) src_a62, 0); res_c7 *= src_a63; res_c6 -= res_c7 * src_a62; res_c5 -= res_c7 * src_a61; res_c4 -= res_c7 * src_a60; res_c3 -= res_c7 * src_a59; res_c2 -= res_c7 * src_a58; res_c1 -= res_c7 * src_a57; res_c0 -= res_c7 * src_a56; src_a48 = LD_DP(a - 16); src_a49 = (v2f64) __msa_splati_d((v2i64) src_a48, 1); src_a48 = (v2f64) __msa_splati_d((v2i64) src_a48, 0); src_a50 = LD_DP(a - 14); src_a51 = (v2f64) __msa_splati_d((v2i64) src_a50, 1); src_a50 = (v2f64) __msa_splati_d((v2i64) src_a50, 0); src_a52 = LD_DP(a - 12); src_a53 = (v2f64) __msa_splati_d((v2i64) src_a52, 1); src_a52 = (v2f64) __msa_splati_d((v2i64) src_a52, 0); src_a54 = __msa_cast_to_vector_double(*(a - 10)); src_a54 = (v2f64) __msa_splati_d((v2i64) src_a54, 0); src_a40 = LD_DP(a - 24); src_a41 = (v2f64) __msa_splati_d((v2i64) src_a40, 1); src_a40 = (v2f64) __msa_splati_d((v2i64) src_a40, 0); src_a42 = LD_DP(a - 22); src_a43 = (v2f64) __msa_splati_d((v2i64) src_a42, 1); src_a42 = (v2f64) __msa_splati_d((v2i64) src_a42, 0); src_a44 = LD_DP(a - 20); src_a45 = (v2f64) __msa_splati_d((v2i64) src_a44, 1); src_a44 = (v2f64) __msa_splati_d((v2i64) src_a44, 0); res_c6 *= src_a54; res_c5 -= res_c6 * src_a53; res_c4 -= res_c6 * src_a52; res_c3 -= res_c6 * src_a51; res_c2 -= res_c6 * src_a50; res_c1 -= res_c6 * src_a49; res_c0 -= res_c6 * src_a48; res_c5 *= src_a45; res_c4 -= res_c5 * src_a44; res_c3 -= res_c5 * src_a43; res_c2 -= res_c5 * src_a42; res_c1 -= res_c5 * src_a41; res_c0 -= res_c5 * src_a40; ST_DP(res_c7, b - 2); ST_DP(res_c6, b - 4); ST_DP(res_c5, b - 6); src_a32 = LD_DP(a - 32); src_a33 = (v2f64) __msa_splati_d((v2i64) src_a32, 1); src_a32 = (v2f64) __msa_splati_d((v2i64) src_a32, 0); src_a34 = LD_DP(a - 30); src_a35 = (v2f64) __msa_splati_d((v2i64) src_a34, 1); src_a34 = (v2f64) __msa_splati_d((v2i64) src_a34, 0); src_a36 = __msa_cast_to_vector_double(*(a - 28)); src_a36 = (v2f64) __msa_splati_d((v2i64) src_a36, 0); res_c4 *= src_a36; res_c3 -= res_c4 * src_a35; res_c2 -= res_c4 * src_a34; res_c1 -= res_c4 * src_a33; res_c0 -= res_c4 * src_a32; src_a24 = LD_DP(a - 40); src_a25 = (v2f64) __msa_splati_d((v2i64) src_a24, 1); src_a24 = (v2f64) __msa_splati_d((v2i64) src_a24, 0); src_a26 = LD_DP(a - 38); src_a27 = (v2f64) __msa_splati_d((v2i64) src_a26, 1); src_a26 = (v2f64) __msa_splati_d((v2i64) src_a26, 0); src_a16 = LD_DP(a - 48); src_a17 = (v2f64) __msa_splati_d((v2i64) src_a16, 1); src_a16 = (v2f64) __msa_splati_d((v2i64) src_a16, 0); src_a18 = __msa_cast_to_vector_double(*(a - 46)); src_a18 = (v2f64) __msa_splati_d((v2i64) src_a18, 0); src_a0 = __msa_cast_to_vector_double(*(a - 64)); src_a0 = (v2f64) __msa_splati_d((v2i64) src_a0, 0); src_a8 = LD_DP(a - 56); src_a9 = (v2f64) __msa_splati_d((v2i64) src_a8, 1); src_a8 = (v2f64) __msa_splati_d((v2i64) src_a8, 0); res_c3 *= src_a27; res_c2 -= res_c3 * src_a26; res_c1 -= res_c3 * src_a25; res_c0 -= res_c3 * src_a24; res_c2 *= src_a18; res_c1 -= res_c2 * src_a17; res_c0 -= res_c2 * src_a16; res_c1 *= src_a9; res_c0 -= res_c1 * src_a8; res_c0 *= src_a0; ST_DP(res_c4, b - 8); ST_DP(res_c3, b - 10); ST_DP(res_c2, b - 12); ST_DP(res_c1, b - 14); ST_DP(res_c0, b - 16); ILVRL_D2_DP(res_c1, res_c0, src_c0, src_c4); ILVRL_D2_DP(res_c3, res_c2, src_c1, src_c5); ILVRL_D2_DP(res_c5, res_c4, src_c2, src_c6); ILVRL_D2_DP(res_c7, res_c6, src_c3, src_c7); ST_DP4(src_c0, src_c1, src_c2, src_c3, c, 2); ST_DP4(src_c4, src_c5, src_c6, src_c7, c + ldc, 2); }
static void dsolve_8x4_ln_msa(FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG bk) { v2f64 src_c0, src_c1, src_c2, src_c3, src_c4, src_c5, src_c6, src_c7; v2f64 res_c0, res_c1, res_c2, res_c3, res_c4, res_c5, res_c6, res_c7; v2f64 src_c8, src_c9, src_c10, src_c11, src_c12, src_c13, src_c14, src_c15; v2f64 res_c8, res_c9, res_c10, res_c11, res_c12, res_c13, res_c14, res_c15; v2f64 src_a0, src_a1, src_a2, src_a3, src_a8, src_a9, src_a16, src_a17; v2f64 src_a18, src_a24, src_a25, src_a26, src_a27, src_a32, src_a33; v2f64 src_a34, src_a35, src_a36, src_a40, src_a41, src_a42, src_a43; v2f64 src_a44, src_a45, src_a48, src_a49, src_a50, src_a51, src_a52; v2f64 src_a53, src_a54, src_a56, src_a57, src_a58, src_a59, src_a60; v2f64 src_a61, src_a62, src_a63; FLOAT *c_nxt1line = c + ldc; FLOAT *c_nxt2line = c + 2 * ldc; FLOAT *c_nxt3line = c + 3 * ldc; LD_DP4(c, 2, src_c0, src_c1, src_c2, src_c3); LD_DP4(c_nxt1line, 2, src_c4, src_c5, src_c6, src_c7); LD_DP4(c_nxt2line, 2, src_c8, src_c9, src_c10, src_c11); LD_DP4(c_nxt3line, 2, src_c12, src_c13, src_c14, src_c15); if (bk > 0) { BLASLONG i; FLOAT *pba = a, *pbb = b; v2f64 src_b, src_b0, src_b1, src_b2, src_b3; LD_DP4(pba, 2, src_a0, src_a1, src_a2, src_a3); LD_DP2(pbb, 2, src_b0, src_b1); for (i = (bk - 1); i--;) { pba += 8; pbb += 4; LD_DP4(pba, 2, src_a8, src_a9, src_a16, src_a17); LD_DP2(pbb, 2, src_b2, src_b3); src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0); src_c0 -= src_a0 * src_b; src_c1 -= src_a1 * src_b; src_c2 -= src_a2 * src_b; src_c3 -= src_a3 * src_b; src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0); src_c4 -= src_a0 * src_b; src_c5 -= src_a1 * src_b; src_c6 -= src_a2 * src_b; src_c7 -= src_a3 * src_b; src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1); src_c8 -= src_a0 * src_b; src_c9 -= src_a1 * src_b; src_c10 -= src_a2 * src_b; src_c11 -= src_a3 * src_b; src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1); src_c12 -= src_a0 * src_b; src_c13 -= src_a1 * src_b; src_c14 -= src_a2 * src_b; src_c15 -= src_a3 * src_b; src_a0 = src_a8; src_a1 = src_a9; src_a2 = src_a16; src_a3 = src_a17; src_b0 = src_b2; src_b1 = src_b3; } src_b = (v2f64) __msa_ilvr_d((v2i64) src_b0, (v2i64) src_b0); src_c0 -= src_a0 * src_b; src_c1 -= src_a1 * src_b; src_c2 -= src_a2 * src_b; src_c3 -= src_a3 * src_b; src_b = (v2f64) __msa_ilvl_d((v2i64) src_b0, (v2i64) src_b0); src_c4 -= src_a0 * src_b; src_c5 -= src_a1 * src_b; src_c6 -= src_a2 * src_b; src_c7 -= src_a3 * src_b; src_b = (v2f64) __msa_ilvr_d((v2i64) src_b1, (v2i64) src_b1); src_c8 -= src_a0 * src_b; src_c9 -= src_a1 * src_b; src_c10 -= src_a2 * src_b; src_c11 -= src_a3 * src_b; src_b = (v2f64) __msa_ilvl_d((v2i64) src_b1, (v2i64) src_b1); src_c12 -= src_a0 * src_b; src_c13 -= src_a1 * src_b; src_c14 -= src_a2 * src_b; src_c15 -= src_a3 * src_b; } a -= 64; b -= 32; ILVRL_D2_DP(src_c4, src_c0, res_c0, res_c1); ILVRL_D2_DP(src_c5, src_c1, res_c2, res_c3); ILVRL_D2_DP(src_c6, src_c2, res_c4, res_c5); ILVRL_D2_DP(src_c7, src_c3, res_c6, res_c7); ILVRL_D2_DP(src_c12, src_c8, res_c8, res_c9); ILVRL_D2_DP(src_c13, src_c9, res_c10, res_c11); ILVRL_D2_DP(src_c14, src_c10, res_c12, res_c13); ILVRL_D2_DP(src_c15, src_c11, res_c14, res_c15); src_a54 = __msa_cast_to_vector_double(*(a + 54)); src_a54 = (v2f64) __msa_splati_d((v2i64) src_a54, 0); src_a62 = LD_DP(a + 62); src_a63 = (v2f64) __msa_splati_d((v2i64) src_a62, 1); src_a62 = (v2f64) __msa_splati_d((v2i64) src_a62, 0); src_a60 = LD_DP(a + 60); src_a61 = (v2f64) __msa_splati_d((v2i64) src_a60, 1); src_a60 = (v2f64) __msa_splati_d((v2i64) src_a60, 0); src_a52 = LD_DP(a + 52); src_a53 = (v2f64) __msa_splati_d((v2i64) src_a52, 1); src_a52 = (v2f64) __msa_splati_d((v2i64) src_a52, 0); src_a44 = LD_DP(a + 44); src_a45 = (v2f64) __msa_splati_d((v2i64) src_a44, 1); src_a44 = (v2f64) __msa_splati_d((v2i64) src_a44, 0); src_a36 = __msa_cast_to_vector_double(*(a + 36)); src_a36 = (v2f64) __msa_splati_d((v2i64) src_a36, 0); res_c7 *= src_a63; res_c6 -= res_c7 * src_a62; res_c6 *= src_a54; res_c15 *= src_a63; res_c14 -= res_c15 * src_a62; res_c14 *= src_a54; ST_DP(res_c7, b + 28); ST_DP(res_c6, b + 24); ST_DP(res_c15, b + 30); ST_DP(res_c14, b + 26); ILVRL_D2_DP(res_c7, res_c6, src_c3, src_c7); ILVRL_D2_DP(res_c15, res_c14, src_c11, src_c15); ST_DP(src_c3, c + 6); ST_DP(src_c7, c_nxt1line + 6); ST_DP(src_c11, c_nxt2line + 6); ST_DP(src_c15, c_nxt3line + 6); res_c5 -= res_c7 * src_a61; res_c5 -= res_c6 * src_a53; res_c5 *= src_a45; res_c4 -= res_c7 * src_a60; res_c4 -= res_c6 * src_a52; res_c4 -= res_c5 * src_a44; res_c4 *= src_a36; res_c13 -= res_c15 * src_a61; res_c13 -= res_c14 * src_a53; res_c13 *= src_a45; res_c12 -= res_c15 * src_a60; res_c12 -= res_c14 * src_a52; res_c12 -= res_c13 * src_a44; res_c12 *= src_a36; src_a56 = LD_DP(a + 56); src_a57 = (v2f64) __msa_splati_d((v2i64) src_a56, 1); src_a56 = (v2f64) __msa_splati_d((v2i64) src_a56, 0); src_a58 = LD_DP(a + 58); src_a59 = (v2f64) __msa_splati_d((v2i64) src_a58, 1); src_a58 = (v2f64) __msa_splati_d((v2i64) src_a58, 0); ST_DP(res_c4, b + 16); ST_DP(res_c5, b + 20); ST_DP(res_c12, b + 18); ST_DP(res_c13, b + 22); ILVRL_D2_DP(res_c5, res_c4, src_c2, src_c6); ILVRL_D2_DP(res_c13, res_c12, src_c10, src_c14); ST_DP(src_c2, c + 4); ST_DP(src_c6, c_nxt1line + 4); ST_DP(src_c10, c_nxt2line + 4); ST_DP(src_c14, c_nxt3line + 4); src_a50 = LD_DP(a + 50); src_a51 = (v2f64) __msa_splati_d((v2i64) src_a50, 1); src_a50 = (v2f64) __msa_splati_d((v2i64) src_a50, 0); src_a42 = LD_DP(a + 42); src_a43 = (v2f64) __msa_splati_d((v2i64) src_a42, 1); src_a42 = (v2f64) __msa_splati_d((v2i64) src_a42, 0); src_a34 = LD_DP(a + 34); src_a35 = (v2f64) __msa_splati_d((v2i64) src_a34, 1); src_a34 = (v2f64) __msa_splati_d((v2i64) src_a34, 0); src_a26 = LD_DP(a + 26); src_a27 = (v2f64) __msa_splati_d((v2i64) src_a26, 1); src_a26 = (v2f64) __msa_splati_d((v2i64) src_a26, 0); src_a18 = __msa_cast_to_vector_double(*(a + 18)); src_a18 = (v2f64) __msa_splati_d((v2i64) src_a18, 0); res_c3 -= res_c7 * src_a59; res_c2 -= res_c7 * src_a58; res_c1 -= res_c7 * src_a57; res_c0 -= res_c7 * src_a56; res_c11 -= res_c15 * src_a59; res_c10 -= res_c15 * src_a58; res_c9 -= res_c15 * src_a57; res_c8 -= res_c15 * src_a56; res_c3 -= res_c6 * src_a51; res_c3 -= res_c5 * src_a43; res_c3 -= res_c4 * src_a35; res_c3 *= src_a27; res_c2 -= res_c6 * src_a50; res_c2 -= res_c5 * src_a42; res_c2 -= res_c4 * src_a34; res_c2 -= res_c3 * src_a26; res_c2 *= src_a18; res_c11 -= res_c14 * src_a51; res_c11 -= res_c13 * src_a43; res_c11 -= res_c12 * src_a35; res_c11 *= src_a27; res_c10 -= res_c14 * src_a50; res_c10 -= res_c13 * src_a42; res_c10 -= res_c12 * src_a34; res_c10 -= res_c11 * src_a26; res_c10 *= src_a18; src_a48 = LD_DP(a + 48); src_a49 = (v2f64) __msa_splati_d((v2i64) src_a48, 1); src_a48 = (v2f64) __msa_splati_d((v2i64) src_a48, 0); src_a40 = LD_DP(a + 40); src_a41 = (v2f64) __msa_splati_d((v2i64) src_a40, 1); src_a40 = (v2f64) __msa_splati_d((v2i64) src_a40, 0); ST_DP(res_c2, b + 8); ST_DP(res_c3, b + 12); ST_DP(res_c10, b + 10); ST_DP(res_c11, b + 14); src_a32 = LD_DP(a + 32); src_a33 = (v2f64) __msa_splati_d((v2i64) src_a32, 1); src_a32 = (v2f64) __msa_splati_d((v2i64) src_a32, 0); src_a24 = LD_DP(a + 24); src_a25 = (v2f64) __msa_splati_d((v2i64) src_a24, 1); src_a24 = (v2f64) __msa_splati_d((v2i64) src_a24, 0); ILVRL_D2_DP(res_c3, res_c2, src_c1, src_c5); ILVRL_D2_DP(res_c11, res_c10, src_c9, src_c13); ST_DP(src_c1, c + 2); ST_DP(src_c5, c_nxt1line + 2); ST_DP(src_c9, c_nxt2line + 2); ST_DP(src_c13, c_nxt3line + 2); res_c1 -= res_c6 * src_a49; res_c1 -= res_c5 * src_a41; res_c1 -= res_c4 * src_a33; res_c1 -= res_c3 * src_a25; res_c0 -= res_c6 * src_a48; res_c0 -= res_c5 * src_a40; res_c0 -= res_c4 * src_a32; res_c0 -= res_c3 * src_a24; res_c9 -= res_c14 * src_a49; res_c9 -= res_c13 * src_a41; res_c9 -= res_c12 * src_a33; res_c9 -= res_c11 * src_a25; res_c8 -= res_c14 * src_a48; res_c8 -= res_c13 * src_a40; res_c8 -= res_c12 * src_a32; res_c8 -= res_c11 * src_a24; src_a16 = LD_DP(a + 16); src_a17 = (v2f64) __msa_splati_d((v2i64) src_a16, 1); src_a16 = (v2f64) __msa_splati_d((v2i64) src_a16, 0); src_a8 = LD_DP(a + 8); src_a9 = (v2f64) __msa_splati_d((v2i64) src_a8, 1); src_a8 = (v2f64) __msa_splati_d((v2i64) src_a8, 0); src_a0 = __msa_cast_to_vector_double(*(a + 0)); src_a0 = (v2f64) __msa_splati_d((v2i64) src_a0, 0); res_c1 -= res_c2 * src_a17; res_c1 *= src_a9; res_c9 -= res_c10 * src_a17; res_c9 *= src_a9; res_c0 -= res_c2 * src_a16; res_c0 -= res_c1 * src_a8; res_c0 *= src_a0; res_c8 -= res_c10 * src_a16; res_c8 -= res_c9 * src_a8; res_c8 *= src_a0; ST_DP(res_c0, b + 0); ST_DP(res_c8, b + 2); ST_DP(res_c1, b + 4); ST_DP(res_c9, b + 6); ILVRL_D2_DP(res_c1, res_c0, src_c0, src_c4); ILVRL_D2_DP(res_c9, res_c8, src_c8, src_c12); ST_DP(src_c0, c); ST_DP(src_c4, c_nxt1line); ST_DP(src_c8, c_nxt2line); ST_DP(src_c12, c_nxt3line); }
int32_t vp8_mbblock_error_msa(MACROBLOCK *mb, int32_t dc) { BLOCK *be; BLOCKD *bd; int16_t *coeff_ptr, *dq_coeff_ptr; int32_t err = 0; uint32_t loop_cnt; v8i16 coeff, coeff0, coeff1, coeff2, coeff3, coeff4; v8i16 dq_coeff, dq_coeff2, dq_coeff3, dq_coeff4; v4i32 diff0, diff1; v2i64 err0, err1; v16u8 zero = { 0 }; v16u8 mask0 = (v16u8)__msa_ldi_b(255); if (1 == dc) { mask0 = (v16u8)__msa_insve_w((v4i32)mask0, 0, (v4i32)zero); } for (loop_cnt = 0; loop_cnt < 8; loop_cnt++) { be = &mb->block[2 * loop_cnt]; bd = &mb->e_mbd.block[2 * loop_cnt]; coeff_ptr = be->coeff; dq_coeff_ptr = bd->dqcoeff; coeff = LD_SH(coeff_ptr); dq_coeff = LD_SH(dq_coeff_ptr); coeff_ptr += 8; dq_coeff_ptr += 8; coeff2 = LD_SH(coeff_ptr); dq_coeff2 = LD_SH(dq_coeff_ptr); be = &mb->block[2 * loop_cnt + 1]; bd = &mb->e_mbd.block[2 * loop_cnt + 1]; coeff_ptr = be->coeff; dq_coeff_ptr = bd->dqcoeff; coeff3 = LD_SH(coeff_ptr); dq_coeff3 = LD_SH(dq_coeff_ptr); coeff_ptr += 8; dq_coeff_ptr += 8; coeff4 = LD_SH(coeff_ptr); dq_coeff4 = LD_SH(dq_coeff_ptr); ILVRL_H2_SH(coeff, dq_coeff, coeff0, coeff1); HSUB_UH2_SW(coeff0, coeff1, diff0, diff1); diff0 = (v4i32)__msa_bmnz_v(zero, (v16u8)diff0, mask0); DOTP_SW2_SD(diff0, diff1, diff0, diff1, err0, err1); ILVRL_H2_SH(coeff2, dq_coeff2, coeff0, coeff1); HSUB_UH2_SW(coeff0, coeff1, diff0, diff1); DPADD_SD2_SD(diff0, diff1, err0, err1); err0 += __msa_splati_d(err0, 1); err1 += __msa_splati_d(err1, 1); err += __msa_copy_s_d(err0, 0); err += __msa_copy_s_d(err1, 0); ILVRL_H2_SH(coeff3, dq_coeff3, coeff0, coeff1); HSUB_UH2_SW(coeff0, coeff1, diff0, diff1); diff0 = (v4i32)__msa_bmnz_v(zero, (v16u8)diff0, mask0); DOTP_SW2_SD(diff0, diff1, diff0, diff1, err0, err1); ILVRL_H2_SH(coeff4, dq_coeff4, coeff0, coeff1); HSUB_UH2_SW(coeff0, coeff1, diff0, diff1); DPADD_SD2_SD(diff0, diff1, err0, err1); err0 += __msa_splati_d(err0, 1); err1 += __msa_splati_d(err1, 1); err += __msa_copy_s_d(err0, 0); err += __msa_copy_s_d(err1, 0); } return err; }
int32_t vp8_mbuverror_msa(MACROBLOCK *mb) { BLOCK *be; BLOCKD *bd; int16_t *coeff_ptr, *dq_coeff_ptr; int32_t err = 0; uint32_t loop_cnt; v8i16 coeff, coeff0, coeff1, coeff2, coeff3, coeff4; v8i16 dq_coeff, dq_coeff2, dq_coeff3, dq_coeff4; v4i32 diff0, diff1; v2i64 err0, err1, err_dup0, err_dup1; for (loop_cnt = 16; loop_cnt < 24; loop_cnt += 2) { be = &mb->block[loop_cnt]; bd = &mb->e_mbd.block[loop_cnt]; coeff_ptr = be->coeff; dq_coeff_ptr = bd->dqcoeff; coeff = LD_SH(coeff_ptr); dq_coeff = LD_SH(dq_coeff_ptr); coeff_ptr += 8; dq_coeff_ptr += 8; coeff2 = LD_SH(coeff_ptr); dq_coeff2 = LD_SH(dq_coeff_ptr); be = &mb->block[loop_cnt + 1]; bd = &mb->e_mbd.block[loop_cnt + 1]; coeff_ptr = be->coeff; dq_coeff_ptr = bd->dqcoeff; coeff3 = LD_SH(coeff_ptr); dq_coeff3 = LD_SH(dq_coeff_ptr); coeff_ptr += 8; dq_coeff_ptr += 8; coeff4 = LD_SH(coeff_ptr); dq_coeff4 = LD_SH(dq_coeff_ptr); ILVRL_H2_SH(coeff, dq_coeff, coeff0, coeff1); HSUB_UH2_SW(coeff0, coeff1, diff0, diff1); DOTP_SW2_SD(diff0, diff1, diff0, diff1, err0, err1); ILVRL_H2_SH(coeff2, dq_coeff2, coeff0, coeff1); HSUB_UH2_SW(coeff0, coeff1, diff0, diff1); DPADD_SD2_SD(diff0, diff1, err0, err1); err_dup0 = __msa_splati_d(err0, 1); err_dup1 = __msa_splati_d(err1, 1); ADD2(err0, err_dup0, err1, err_dup1, err0, err1); err += __msa_copy_s_d(err0, 0); err += __msa_copy_s_d(err1, 0); ILVRL_H2_SH(coeff3, dq_coeff3, coeff0, coeff1); HSUB_UH2_SW(coeff0, coeff1, diff0, diff1); DOTP_SW2_SD(diff0, diff1, diff0, diff1, err0, err1); ILVRL_H2_SH(coeff4, dq_coeff4, coeff0, coeff1); HSUB_UH2_SW(coeff0, coeff1, diff0, diff1); DPADD_SD2_SD(diff0, diff1, err0, err1); err_dup0 = __msa_splati_d(err0, 1); err_dup1 = __msa_splati_d(err1, 1); ADD2(err0, err_dup0, err1, err_dup1, err0, err1); err += __msa_copy_s_d(err0, 0); err += __msa_copy_s_d(err1, 0); } return err; }