 |
|
|
|
.. (parent) |
 |
- |
rw-r--r-- |
3,647 |
CMakeLists.txt
|
|
- |
rw-r--r-- |
0 |
__init__.py
|
|
- |
rw-r--r-- |
3,601 |
adagrad.cc
|
|
- |
rw-r--r-- |
6,068 |
adagrad.h
|
|
- |
rw-r--r-- |
3,888 |
adagrad_avx2.cc
|
|
- |
rw-r--r-- |
5,263 |
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-- |
1,951 |
cvtsh_ss_bugfix.h
|
|
- |
rw-r--r-- |
15,957 |
embedding_lookup.cc
|
|
- |
rw-r--r-- |
1,539 |
embedding_lookup.h
|
|
- |
rw-r--r-- |
123,687 |
embedding_lookup_avx2.cc
|
|
- |
rw-r--r-- |
124,509 |
embedding_lookup_fused_8bit_rowwise_avx2.cc
|
|
- |
rw-r--r-- |
127,185 |
embedding_lookup_fused_8bit_rowwise_idx_avx2.cc
|
|
- |
rw-r--r-- |
16,618 |
embedding_lookup_idx.cc
|
|
- |
rw-r--r-- |
1,672 |
embedding_lookup_idx.h
|
|
- |
rw-r--r-- |
126,321 |
embedding_lookup_idx_avx2.cc
|
|
- |
rw-r--r-- |
13,569 |
fused_8bit_rowwise_embedding_lookup.cc
|
|
- |
rw-r--r-- |
1,690 |
fused_8bit_rowwise_embedding_lookup.h
|
|
- |
rw-r--r-- |
14,087 |
fused_8bit_rowwise_embedding_lookup_idx.cc
|
|
- |
rw-r--r-- |
1,805 |
fused_8bit_rowwise_embedding_lookup_idx.h
|
|
- |
rw-r--r-- |
6,488 |
fused_nbit_rowwise_conversion.cc
|
|
- |
rw-r--r-- |
758 |
fused_nbit_rowwise_conversion.h
|
|
- |
rw-r--r-- |
19,251 |
hp_emblookup_codegen.py
|
|
- |
rw-r--r-- |
4,220 |
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-- |
7,663 |
math_cpu_avx2.cc
|
|
- |
rw-r--r-- |
5,201 |
math_cpu_base.cc
|
|
- |
rw-r--r-- |
2,169 |
typed_axpy.cc
|
|
- |
rw-r--r-- |
355 |
typed_axpy.h
|
|
- |
rw-r--r-- |
1,747 |
typed_axpy_avx.cc
|
|
- |
rw-r--r-- |
2,729 |
typed_axpy_avx2.cc
|
