singa
singa copied to clipboard
SINGA-482 tc comprehension integration
Hi this is not completely ready for merge, pending to fix a testing error.
Hi, this is the working version of tc integration for now, with following notes:
- tested in the environment specified in the dockerfile
cuda9
- it implemented
MatMul
,Relu
in TC for demonstration and usage is shown in following tests:./bin/test_singa --gtest_filter=*TC*
, and tests are passing. - for now only working in Cuda environment, and only working with type
float
P.S. softmax
is implemented but still checking why the output is not correct.
TODO:
- move tc execution code into device
DoExec
, and support CPU - package dependencies as conda package
- integer support
Integration is done except support CPU. CPU support development is done, however the behavior was not expected even in clean TC env. CPU backed TC return zero tensor as shown in https://gist.githubusercontent.com/dcslin/f3ed411012c144163f3bedbe621257f6/raw/9f6af3f5e62b5bb52a58c087967c813b3acd8872/tensor-comprehension-cpu-errors
Ready for review.