Lines Matching refs:s

105     dup       v14.4s, w4                // populate the u4_qp_div_6
109     saddl     v4.4s, v0.4h, v3.4h       //x0 = x4 + x7
110     saddl     v5.4s, v1.4h, v2.4h       //x1 = x5 + x6
111     ssubl     v6.4s, v1.4h, v2.4h       //x2 = x5 - x6
112     ssubl     v7.4s, v0.4h, v3.4h       //x3 = x4 - x7
114     add       v0.4s, v4.4s, v5.4s       //pi4_tmp_ptr[0] = x0 + x1
115     add       v1.4s, v7.4s, v6.4s       //pi4_tmp_ptr[1] = x3 + x2
116     sub       v2.4s, v4.4s, v5.4s       //pi4_tmp_ptr[2] = x0 - x1
117     sub       v3.4s, v7.4s, v6.4s       //pi4_tmp_ptr[3] = x3 - x2
119     umull     v15.4s, v15.4h, v16.4h
120     dup       v15.4s, v15.s[0]          //pu2_weigh_mat[0]*pu2_iscal_mat[0]
123     trn1      v4.4s, v0.4s, v1.4s
124     trn2      v5.4s, v0.4s, v1.4s
125     trn1      v6.4s, v2.4s, v3.4s
126     trn2      v7.4s, v2.4s, v3.4s
134     add       v4.4s, v0.4s, v3.4s       //x0 = x4+x7
135     add       v5.4s, v1.4s, v2.4s       //x1 = x5+x6
136     sub       v6.4s, v1.4s, v2.4s       //x2 = x5-x6
137     sub       v7.4s, v0.4s, v3.4s       //x3 = x4-x7
139     add       v0.4s, v4.4s, v5.4s       //pi4_tmp_ptr[0] = x0 + x1
140     add       v1.4s, v7.4s, v6.4s       //pi4_tmp_ptr[1] = x3 + x2
141     sub       v2.4s, v4.4s, v5.4s       //pi4_tmp_ptr[2] = x0 - x1
142     sub       v3.4s, v7.4s, v6.4s       //pi4_tmp_ptr[3] = x3 - x2
144     mul       v0.4s, v0.4s, v15.4s      // q0  = p[i] = (x[i] * trns_coeff[i]) where i = 0..3
145     mul       v1.4s, v1.4s, v15.4s      // q1  = p[i] = (x[i] * trns_coeff[i]) where i = 4..7
146     mul       v2.4s, v2.4s, v15.4s      // q2  = p[i] = (x[i] * trns_coeff[i]) where i = 8..11
147     mul       v3.4s, v3.4s, v15.4s      // q3  = p[i] = (x[i] * trns_coeff[i]) where i = 12..15
149     sshl      v0.4s, v0.4s, v14.4s      // q0  = q[i] = (p[i] << (qp/6)) where i = 0..3
150     sshl      v1.4s, v1.4s, v14.4s      // q1  = q[i] = (p[i] << (qp/6)) where i = 4..7
151     sshl      v2.4s, v2.4s, v14.4s      // q2  = q[i] = (p[i] << (qp/6)) where i = 8..11
152     sshl      v3.4s, v3.4s, v14.4s      // q3  = q[i] = (p[i] << (qp/6)) where i = 12..15
154     sqrshrn   v0.4h, v0.4s, #6          // d0  = c[i] = ((q[i] + 32) >> 4) where i = 0..3
155     sqrshrn   v1.4h, v1.4s, #6          // d1  = c[i] = ((q[i] + 32) >> 4) where i = 4..7
156     sqrshrn   v2.4h, v2.4s, #6          // d2  = c[i] = ((q[i] + 32) >> 4) where i = 8..11
157     sqrshrn   v3.4h, v3.4s, #6          // d3  = c[i] = ((q[i] + 32) >> 4) where i = 12..15
218     dup       v28.4s, w4                //load qp/6
224     saddl     v2.4s, v0.4h, v1.4h       //i4_x0 = i4_x4 + i4_x5;...x2
225     ssubl     v4.4s, v0.4h, v1.4h       //i4_x1 = i4_x4 - i4_x5;...x3
227     umull     v30.4s, v26.4h, v27.4h    //pu2_iscal_mat[0]*pu2_weigh_mat[0]
228     dup       v30.4s, v30.s[0]
230     trn1      v0.4s, v2.4s, v4.4s
231     trn2      v1.4s, v2.4s, v4.4s       //i4_x0 i4_x1 -> q1
233     add       v2.4s, v0.4s, v1.4s       //i4_x4 = i4_x0+i4_x2;.. i4_x5
234     sub       v3.4s, v0.4s, v1.4s       //i4_x6 = i4_x0-i4_x2;.. i4_x7
236     mul       v2.4s, v2.4s, v30.4s
237     mul       v3.4s, v3.4s, v30.4s
239     sshl      v2.4s, v2.4s, v28.4s
240     sshl      v3.4s, v3.4s, v28.4s
242     xtn       v0.4h, v2.4s              //i4_x4 i4_x5 i4_y4 i4_y5
243     xtn       v1.4h, v3.4s              //i4_x6 i4_x7 i4_y6 i4_y7
245     st2       {v0.4s-v1.4s}, [x1]