57         for (CeedInt jj = 0; jj < JJ; jj++) {  in CeedTensorContract_Avx_Blocked()  local
61           for (CeedInt jj = 0; jj < JJ; jj++) {  // unroll  in CeedTensorContract_Avx_Blocked()  local
68         for (CeedInt jj = 0; jj < JJ; jj++) {  in CeedTensorContract_Avx_Blocked()  local
80         for (CeedInt jj = 0; jj < J - j; jj++) {  in CeedTensorContract_Avx_Blocked()  local
84           for (CeedInt jj = 0; jj < J - j; jj++) {  // doesn't unroll  in CeedTensorContract_Avx_Blocked()  local
92         for (CeedInt jj = 0; jj < J - j; jj++) {  in CeedTensorContract_Avx_Blocked()  local
123         for (CeedInt jj = 0; jj < JJ; jj++) vv[jj] = loadu(&v[(a * J + j + jj) * C + c]);  in CeedTensorContract_Avx_Remainder()  local
131           for (CeedInt jj = 0; jj < JJ; jj++) {  // unroll  in CeedTensorContract_Avx_Remainder()  local
135         for (CeedInt jj = 0; jj < JJ; jj++) storeu(&v[(a * J + j + jj) * C + c], vv[jj]);  in CeedTensorContract_Avx_Remainder()  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()  local
172         for (CeedInt jj = 0; jj < JJ / 4; jj++) {  // unroll  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()  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()  local
196       for (CeedInt jj = 0; jj < JJ / 4; jj++) {  // unroll  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()  local