 |
|
|
|
.. (parent) |
 |
- |
rw-r--r-- |
3,814 |
CMakeLists.txt
|
|
- |
rw-r--r-- |
0 |
__init__.py
|
|
- |
rw-r--r-- |
3,600 |
adagrad.cc
|
|
- |
rw-r--r-- |
6,104 |
adagrad.h
|
|
- |
rw-r--r-- |
3,885 |
adagrad_avx2.cc
|
|
- |
rw-r--r-- |
5,408 |
common.h
|
|
- |
rw-r--r-- |
665 |
common_avx.cc
|
|
- |
rw-r--r-- |
680 |
common_avx2.cc
|
|
- |
rw-r--r-- |
974 |
common_avx512.cc
|
|
- |
rw-r--r-- |
2,062 |
cvtsh_ss_bugfix.h
|
|
- |
rw-r--r-- |
16,146 |
embedding_lookup.cc
|
|
- |
rw-r--r-- |
1,539 |
embedding_lookup.h
|
|
- |
rw-r--r-- |
127,741 |
embedding_lookup_avx2.cc
|
|
- |
rw-r--r-- |
128,563 |
embedding_lookup_fused_8bit_rowwise_avx2.cc
|
|
- |
rw-r--r-- |
131,323 |
embedding_lookup_fused_8bit_rowwise_idx_avx2.cc
|
|
- |
rw-r--r-- |
16,729 |
embedding_lookup_idx.cc
|
|
- |
rw-r--r-- |
1,674 |
embedding_lookup_idx.h
|
|
- |
rw-r--r-- |
130,459 |
embedding_lookup_idx_avx2.cc
|
|
- |
rw-r--r-- |
13,758 |
fused_8bit_rowwise_embedding_lookup.cc
|
|
- |
rw-r--r-- |
1,690 |
fused_8bit_rowwise_embedding_lookup.h
|
|
- |
rw-r--r-- |
14,199 |
fused_8bit_rowwise_embedding_lookup_idx.cc
|
|
- |
rw-r--r-- |
1,807 |
fused_8bit_rowwise_embedding_lookup_idx.h
|
|
- |
rw-r--r-- |
7,442 |
fused_nbit_rowwise_conversion.cc
|
|
- |
rw-r--r-- |
789 |
fused_nbit_rowwise_conversion.h
|
|
- |
rw-r--r-- |
19,261 |
hp_emblookup_codegen.py
|
|
- |
rw-r--r-- |
4,320 |
lstm_unit_cpu-impl.h
|
|
- |
rw-r--r-- |
1,446 |
lstm_unit_cpu.h
|
|
- |
rw-r--r-- |
2,568 |
lstm_unit_cpu_avx2.cc
|
|
- |
rw-r--r-- |
2,394 |
lstm_unit_cpu_common.cc
|
|
- |
rw-r--r-- |
1,442 |
lstm_unit_cpu_common.h
|
|
- |
rw-r--r-- |
1,103 |
math.h
|
|
- |
rw-r--r-- |
11,364 |
math_cpu_avx2.cc
|
|
- |
rw-r--r-- |
5,734 |
math_cpu_base.cc
|
|
- |
rw-r--r-- |
2,653 |
typed_axpy.cc
|
|
- |
rw-r--r-- |
355 |
typed_axpy.h
|
|
- |
rw-r--r-- |
1,924 |
typed_axpy_avx.cc
|
|
- |
rw-r--r-- |
3,024 |
typed_axpy_avx2.cc
|
