| /dports/math/libxsmm/libxsmm-1.16.3/src/template/ |
| H A D | libxsmm_dnn_rnncell_st_rnn_bwdupd_nc_ck_generic.tpl.c | 96 const libxsmm_blasint thr_end_kk = ((ltid + 1) * chunksize_kk < work_kk) ? ((ltid + 1) * chunksize_… variable
135 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
219 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
335 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
|
| H A D | libxsmm_dnn_rnncell_st_rnn_bwdupd_nc_kcck.tpl.c | 109 const libxsmm_blasint thr_end_kk = ((ltid + 1) * chunksize_kk < work_kk) ? ((ltid + 1) * chunksize_… variable
162 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
243 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
372 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
|
| H A D | libxsmm_dnn_rnncell_st_lstm_bwdupd_nc_kcck_bf16.tpl.c | 219 const libxsmm_blasint thr_end_kk = ((ltid + 1) * chunksize_kk < work_kk) ? ((ltid + 1) * chunksize_… variable
298 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
346 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
394 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
427 LIBXSMM_UNUSED(thr_end_kk);
|
| H A D | libxsmm_dnn_rnncell_st_gru_bwdupd_nc_ck_generic.tpl.c | 183 const libxsmm_blasint thr_end_kk = ((ltid + 1) * chunksize_kk < work_kk) ? ((ltid + 1) * chunksize_… variable
242 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
447 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
494 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
541 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
623 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
|
| H A D | libxsmm_dnn_rnncell_st_lstm_fwd_nc_ck_generic.tpl.c | 124 const libxsmm_blasint thr_end_kk = ((ltid + 1) * chunksize_kk < work_kk) ? ((ltid + 1) * chunksize_… variable
177 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
|
| H A D | libxsmm_dnn_rnncell_st_lstm_bwdupd_nc_ck_generic.tpl.c | 208 const libxsmm_blasint thr_end_kk = ((ltid + 1) * chunksize_kk < work_kk) ? ((ltid + 1) * chunksize_… variable
282 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
323 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
|
| H A D | libxsmm_dnn_rnncell_st_lstm_bwdupd_nc_ck_generic_bf16.tpl.c | 211 const libxsmm_blasint thr_end_kk = ((ltid + 1) * chunksize_kk < work_kk) ? ((ltid + 1) * chunksize_… variable
286 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
328 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
|
| H A D | libxsmm_dnn_rnncell_st_gru_bwdupd_nc_kcck.tpl.c | 168 const libxsmm_blasint thr_end_kk = ((ltid + 1) * chunksize_kk < work_kk) ? ((ltid + 1) * chunksize_… variable
224 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
467 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
514 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
561 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
|
| H A D | libxsmm_dnn_rnncell_st_lstm_fwd_nc_ck_generic_bf16.tpl.c | 181 const libxsmm_blasint thr_end_kk = ((ltid + 1) * chunksize_kk < work_kk) ? ((ltid + 1) * chunksize_… variable
234 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
|
| H A D | libxsmm_dnn_rnncell_st_lstm_bwdupd_nc_kcck_core_bf16.tpl.c | 205 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
245 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
|
| H A D | libxsmm_dnn_rnncell_st_gru_fwd_nc_ck_generic.tpl.c | 107 const libxsmm_blasint thr_end_kk = ((ltid + 1) * chunksize_kk < work_kk) ? ((ltid + 1) * chunksize_… variable
151 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
|
| H A D | libxsmm_dnn_rnncell_st_lstm_bwdupd_nc_kcck.tpl.c | 195 const libxsmm_blasint thr_end_kk = ((ltid + 1) * chunksize_kk < work_kk) ? ((ltid + 1) * chunksize_… variable
270 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
|
| H A D | libxsmm_dnn_rnncell_st_lstm_bwdupd_nc_kcck_core.tpl.c | 287 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
345 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
403 for (ikic = thr_begin_kk; ikic < thr_end_kk; ++ikic ) {
|
| H A D | libxsmm_dnn_rnncell_st_lstm_fwd_nc_kcck_bf16.tpl.c | 165 const libxsmm_blasint thr_end_kk = ((ltid + 1) * chunksize_kk < work_kk) ? ((ltid + 1) * chunksize_…
|