summary | shortlog | log | commit | commitdiff | tree
raw | patch | inline | side by side (parent: f85a37e)
raw | patch | inline | side by side (parent: f85a37e)
author | Jianzhong Xu <a0869574@ti.com> | |
Fri, 13 May 2016 20:24:48 +0000 (16:24 -0400) | ||
committer | Jianzhong Xu <a0869574@ti.com> | |
Fri, 13 May 2016 20:24:48 +0000 (16:24 -0400) |
diff --git a/src/ti/linalg/blis/frame/1m/packm/bli_packm_blk_var2.c b/src/ti/linalg/blis/frame/1m/packm/bli_packm_blk_var2.c
index 75e4091b37c9be50b5163a0cfa1efce40583700c..b1ddf84870f4269e24e174e88f07a63e6f7a9868 100644 (file)
kappa_p = &BLIS_ONE;
}
}
+ else {
+ kappa_p = NULL;
+ }
kappa_p = thread_obroadcast( t, kappa_p );
diff --git a/src/ti/linalg/blis/frame/3/gemm/bli_gemm_blk_var1f.c b/src/ti/linalg/blis/frame/3/gemm/bli_gemm_blk_var1f.c
index f739d56fcdff38e6e247dd089f30539c4ef03c76..fd421d8f9dabbaefeef4418b20f0d8995ed5de72 100644 (file)
cntl_sub_unpackm_c( cntl ),
gemm_thread_sub_ipackm( thread ) );
- bli_obj_alias_to(c2, c1);
+ if(i < end) // Only if i < end do we want to DMA the next block
+ {
+ bli_obj_alias_to(c2, c1);
+ }
}
}
diff --git a/src/ti/linalg/blis/frame/3/herk/bli_herk_blk_var1f.c b/src/ti/linalg/blis/frame/3/herk/bli_herk_blk_var1f.c
index 7817434a7ac1b1d9e480a526f2e9240247cb2799..492d24611b353260ada30c50ed39a8760a8aa636 100644 (file)
cntl_sub_unpackm_c( cntl ),
herk_thread_sub_ipackm( thread ) );
- bli_obj_alias_to(c2, c1);
-
+ if(i < end) // Only if i < end do we want to DMA the next block
+ {
+ bli_obj_alias_to(c2, c1);
+ }
}
}
#else
diff --git a/src/ti/linalg/blis/frame/3/herk/bli_herk_l_ker_var2.c b/src/ti/linalg/blis/frame/3/herk/bli_herk_l_ker_var2.c
index 0ffd2b4b6cd6ba1c831039b878c3e75bd3f16f97..a04670861e6d345e2ee8f64f4130654689037551 100644 (file)
ptr_dest += cs_c; \
} \
} \
- c1_new = c1_next; \
- a1_new = a1_next; \
- mc_new = mc_next; \
- m_iter_new = m_iter_next; \
- m_left_new = m_left_next; \
- diagoffc_j = diagoffc_j_next; \
- cs_c11 = cs_c11_next; \
+ if(j < (n_iter-jr_num_threads)) /* no transfer for last iteration */ \
+ {\
+ c1_new = c1_next; \
+ a1_new = a1_next; \
+ mc_new = mc_next; \
+ m_iter_new = m_iter_next; \
+ m_left_new = m_left_next; \
+ diagoffc_j = diagoffc_j_next; \
+ cs_c11 = cs_c11_next; \
+ }\
} \
if (BLIS_ENABLE_PROFILE_KERVAR2 == 1) \
{ \
diff --git a/src/ti/linalg/blis/frame/3/trmm/bli_trmm_blk_var1f.c b/src/ti/linalg/blis/frame/3/trmm/bli_trmm_blk_var1f.c
index 32ae4af9930bb8d4e0a8215f9644467f5a9b7429..10033137c04be3d6304bbe9db0793046de6a9a2c 100644 (file)
cntl_sub_unpackm_c( cntl ),
trmm_thread_sub_ipackm( thread ) );
- bli_obj_alias_to(c2, c1);
+ if(i < end) // Only if i < end do we want to DMA the next block
+ {
+ bli_obj_alias_to(c2, c1);
+ }
}
}
diff --git a/src/ti/linalg/blis/frame/3/trsm/bli_trsm_blk_var1b.c b/src/ti/linalg/blis/frame/3/trsm/bli_trsm_blk_var1b.c
index c6154d12a717428956ad12bbc5812e3510643566..c85eee256b55d0790568057449210597912bcc2a 100644 (file)
}
#endif
#ifdef BLIS_ENABLE_C66X_EDMA
- bli_obj_alias_to(c2, c1);
+ if( (i+b_alg) < end ) //Making sure we aren't at the end of the matrix
+ {
+ bli_obj_alias_to(c2, c1);
+ }
#endif
}
diff --git a/src/ti/linalg/blis/frame/3/trsm/bli_trsm_blk_var1f.c b/src/ti/linalg/blis/frame/3/trsm/bli_trsm_blk_var1f.c
index 8296455ea8ee682cd642f279391b4a59f586edbf..8c7d159580568e9724d04c33c06cec8b772c8406 100644 (file)
trsm_thread_sub_trsm( thread ) );
#ifdef BLIS_ENABLE_C66X_EDMA
- bli_obj_alias_to(c2, c1);
+ if( (i+b_alg) < end ) //Making sure we aren't at the end of the matrix
+ {
+ bli_obj_alias_to(c2, c1);
+ }
#endif
}
index f1ee815792b243b45024469d9e30cd50672219bb..be509d9b4fd8fed7d9a38cc38e5e2e0733d6ad9b 100755 (executable)
uplo_p = bli_obj_uplo( *p );
diagoff_p = bli_obj_diag_offset( *p );
+ m_transfer = m_p;
+ n_transfer = n_p;
if(!bli_obj_root_is_general( *p ) && bli_obj_intersects_diag( *p ) && !(bli_is_triangular( struc_p )))
{
//printf("not general and diag intersects\n");
}
}
}
- else
- {
- m_transfer = m_p;
- n_transfer = n_p;
- }
-
//If source is column major
if( bli_is_col_stored( rs_source, cs_source ) )