Lines Matching refs:aa

168       for (CeedInt aa = 0; aa < AA; aa++) {  in CeedTensorContract_Avx_Single()  local
169 for (CeedInt jj = 0; jj < JJ / 4; jj++) vv[aa][jj] = loadu(&v[(a + aa) * J + j + jj * 4]); in CeedTensorContract_Avx_Single()
176 for (CeedInt aa = 0; aa < AA; aa++) { // unroll in CeedTensorContract_Avx_Single() local
177 fmadd(vv[aa][jj], tqv, set1(u[(a + aa) * B + b])); in CeedTensorContract_Avx_Single()
181 for (CeedInt aa = 0; aa < AA; aa++) { in CeedTensorContract_Avx_Single() local
182 for (CeedInt jj = 0; jj < JJ / 4; jj++) storeu(&v[(a + aa) * J + j + jj * 4], vv[aa][jj]); in CeedTensorContract_Avx_Single()
192 for (CeedInt aa = 0; aa < A - a; aa++) { in CeedTensorContract_Avx_Single() local
193 for (CeedInt jj = 0; jj < JJ / 4; jj++) vv[aa][jj] = loadu(&v[(a + aa) * J + j + jj * 4]); in CeedTensorContract_Avx_Single()
200 for (CeedInt aa = 0; aa < A - a; aa++) { // unroll in CeedTensorContract_Avx_Single() local
201 fmadd(vv[aa][jj], tqv, set1(u[(a + aa) * B + b])); in CeedTensorContract_Avx_Single()
205 for (CeedInt aa = 0; aa < A - a; aa++) { in CeedTensorContract_Avx_Single() local
206 for (CeedInt jj = 0; jj < JJ / 4; jj++) storeu(&v[(a + aa) * J + j + jj * 4], vv[aa][jj]); in CeedTensorContract_Avx_Single()
218 for (CeedInt aa = 0; aa < AA; aa++) vv[aa] = loadu(&v[(a + aa) * J + j]); in CeedTensorContract_Avx_Single() local
233 for (CeedInt aa = 0; aa < AA; aa++) { // unroll in CeedTensorContract_Avx_Single() local
234 fmadd(vv[aa], tqv, set1(u[(a + aa) * B + b])); in CeedTensorContract_Avx_Single()
237 for (CeedInt aa = 0; aa < AA; aa++) storeu(&v[(a + aa) * J + j], vv[aa]); in CeedTensorContract_Avx_Single() local