search

ROCm / MIVisionX

MIVisionX toolkit is a set of comprehensive computer vision and machine intelligence libraries, utilities, and applications bundled into a single toolkit. AMD MIVisionX also delivers a highly optimized open-source implementation of the Khronos OpenVX™ and OpenVX™ Extensions.
https://rocm.docs.amd.com/projects/MIVisionX/en/latest/
MIT License
186 stars 73 forks source link

Model Compiler - NNEF Fuse Op Warning #430

Closed kiritigowda closed 3 years ago

kiritigowda commented 3 years ago

Using the -fuse operation has the following warning

 nnef2nnir2openvx with FUSED Operations -- mnist-nnef

mnist-nnef - Batch size 1
reading NNEF model from /home/kiriti/nn-test/MIVisionX/tests/neural_network_tests/models/mnist-nnef...
OK: creating IR description in ./graph.nnir ...
OK: creating IR binaries in ./binary ...
Done
reading IR model from . ...
OK: reading IR description from ./graph.nnir ...
OK: reading IR binaries from ./binary ...
Traceback (most recent call last):
  File "/opt/rocm/mivisionx/model_compiler/python/nnir_update.py", line 85, in <module>
    main()
  File "/opt/rocm/mivisionx/model_compiler/python/nnir_update.py", line 76, in main
    graph.fuseOps()
  File "/opt/rocm-3.7.0/mivisionx/model_compiler/python/nnir.py", line 1112, in fuseOps
    self.addBinary(prevNode.inputs[1], np.getbuffer(weight))
AttributeError: module 'numpy' has no attribute 'getbuffer'
reading IR model from . ...
OK: reading IR description from ./graph.nnir ...
OK: reading IR binaries from ./binary ...
#OUTPUT-TENSOR: output 3 10 1 1 
creating C code in . ...
creating ./CMakeLists.txt ...
creating ./cmake/FindOpenCL.cmake ...
creating ./annmodule.h ...
creating ./annmodule.cpp ...
creating ./weights.bin ...
creating ./anntest.cpp ...
creating ./annpython.h ...
creating ./annpython.cpp ...
creating ./anntest.py ...
-- The C compiler identification is GNU 7.5.0
-- The CXX compiler identification is GNU 7.5.0
-- Check for working C compiler: /usr/bin/cc
-- Check for working C compiler: /usr/bin/cc -- works
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Detecting C compile features
-- Detecting C compile features - done
-- Check for working CXX compiler: /usr/bin/c++
-- Check for working CXX compiler: /usr/bin/c++ -- works
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Found OPENCL: /opt/rocm/opencl/lib/libOpenCL.so  
-- Configuring done
-- Generating done
-- Build files have been written to: /home/kiriti/nn-test/MIVisionX/tests/neural_network_tests/models/develop/nnefFuse/nnir_build_1
Scanning dependencies of target annmodule
[ 16%] Building CXX object CMakeFiles/annmodule.dir/annmodule.cpp.o
[ 33%] Linking CXX shared library libannmodule.so
[ 33%] Built target annmodule
Scanning dependencies of target annpython
[ 50%] Building CXX object CMakeFiles/annpython.dir/annpython.cpp.o
[ 66%] Linking CXX shared library libannpython.so
[ 66%] Built target annpython
Scanning dependencies of target anntest
[ 83%] Building CXX object CMakeFiles/anntest.dir/anntest.cpp.o
[100%] Linking CXX executable anntest
[100%] Built target anntest
mnist-nnef - Batch size 1
OK: loaded 37 kernels from libvx_nn.so
OK: OpenVX using GPU device#0 (gfx906) [OpenCL 2.0 ] [SvmCaps 0 0]
OK: graph initialization with annAddToGraph() took 683.092 msec
OK: vxProcessGraph() took 12.196 msec (1st iteration)
OK: vxProcessGraph() took 0.233 msec (average over 100 iterations)
OK: OpenCL buffer usage: 184128, 12/12
OK: successful
kiritigowda commented 3 years ago

@hansely changes to model compiler have to pass these tests - https://github.com/GPUOpen-ProfessionalCompute-Libraries/MIVisionX/tree/master/tests/neural_network_tests#mivisionx-neural-network-tests

kiritigowda commented 3 years ago

@hansely any updates on this?

hansely commented 3 years ago

@kiritigowda The fuse-ops flow is working on my machine. Can you try installing numpy on python3?

pip3 install numpy

kiritigowda commented 3 years ago

@hansely the script installs all deps -

https://github.com/GPUOpen-ProfessionalCompute-Libraries/MIVisionX/blob/ced14e8eee0cb7e3c822790e2962b70c50097aa0/tests/neural_network_tests/runNeuralNetworkTests.py#L151

kiritigowda commented 3 years ago

@hansely any updates on this?

hansely commented 3 years ago

@kiritigowda Here is the output of the script on my system

python runNeuralNetworkTests.py --profiler_mode 8

Model Name Batch Size Time/Batch (ms) Time/Image (ms)
nnef-mnist 1 0.185 0.185
nnef-mnist 2 0.183 0.091
nnef-mnist 4 0.186 0.046
nnef-mnist 8 0.185 0.023
nnef-mnist 16 0.184 0.011
nnef-mnist 32 0.184 0.006
nnef-mnist 64 0.182 0.003

python3 runNeuralNetworkTests.py --profiler_mode 8

Model Name Batch Size Time/Batch (ms) Time/Image (ms)
nnef-mnist 1 0.182 0.182
nnef-mnist 2 0.182 0.091
nnef-mnist 4 0.182 0.045
nnef-mnist 8 0.186 0.023
nnef-mnist 16 0.186 0.012
nnef-mnist 32 0.188 0.006
nnef-mnist 64 0.187 0.003
hansely commented 3 years ago

@kiritigowda Did you tried it on your system? Or was it on a docker?

kiritigowda commented 3 years ago

@hansely this issue is no more observed on TOT.

CI LOG for reference- http://math-ci.rocm.amd.com/blue/organizations/jenkins/compute-rocm-dkms-no-npi-hipclang%2FShort-GPUOpen%2FMIVisionX/detail/master/55/pipeline

reading NNEF model from /var/jenkins_home/workspace/g_Short-GPUOpen_MIVisionX_master/LMZNbarS1/mivisionx/tests/neural_network_tests/models/nnef-mnist...
OK: creating IR description in ./graph.nnir ...
OK: creating IR binaries in ./binary ...
Done
reading IR model from . ...
OK: reading IR description from ./graph.nnir ...
OK: reading IR binaries from ./binary ...
writing IR model into . ...
OK: creating IR description in ./graph.nnir ...
OK: creating IR binaries in ./binary ...
reading IR model from . ...
OK: reading IR description from ./graph.nnir ...
OK: reading IR binaries from ./binary ...
#OUTPUT-TENSOR: output 3 10 1 1 
creating C code in . ...
creating ./CMakeLists.txt ...
creating ./cmake/FindOpenCL.cmake ...
creating ./annmodule.h ...
creating ./annmodule.cpp ...
creating ./weights.bin ...
creating ./anntest.cpp ...
creating ./annpython.h ...
creating ./annpython.cpp ...
creating ./anntest.py ...
-- The C compiler identification is GNU 7.5.0
-- The CXX compiler identification is GNU 7.5.0
-- Check for working C compiler: /usr/bin/cc
-- Check for working C compiler: /usr/bin/cc -- works
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Detecting C compile features
-- Detecting C compile features - done
-- Check for working CXX compiler: /usr/bin/c++
-- Check for working CXX compiler: /usr/bin/c++ -- works
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Found OPENCL: /opt/rocm/opencl/lib/libOpenCL.so  
-- Configuring done
-- Generating done
-- Build files have been written to: /var/jenkins_home/workspace/g_Short-GPUOpen_MIVisionX_master/LMZNbarS1/mivisionx/tests/neural_network_tests/models/develop/nnefFuse/nnir_build_64
Scanning dependencies of target annmodule
[ 16%] Building CXX object CMakeFiles/annmodule.dir/annmodule.cpp.o
[ 33%] Linking CXX shared library libannmodule.so
[ 33%] Built target annmodule
Scanning dependencies of target annpython
[ 50%] Building CXX object CMakeFiles/annpython.dir/annpython.cpp.o
[ 66%] Linking CXX shared library libannpython.so
[ 66%] Built target annpython
Scanning dependencies of target anntest
[ 83%] Building CXX object CMakeFiles/anntest.dir/anntest.cpp.o
[100%] Linking CXX executable anntest
[100%] Built target anntest
nnef-mnist - Batch size 64
OK: loaded 38 kernels from libvx_nn.so
OK: OpenVX using GPU device#0 (gfx908:sramecc+:xnack-) [OpenCL 2.0 ] [SvmCaps 0 0]
MIOpen(OpenCL): Warning [ParseAndLoadDb] File is unreadable: /opt/rocm/miopen/share/miopen/db/gfx90878.OpenCL.fdb.txt
OK: graph initialization with annAddToGraph() took 713.874 msec
OK: vxProcessGraph() took 101.052 msec (1st iteration)
OK: vxProcessGraph() took 0.175 msec (average over 100 iterations)
OK: OpenCL buffer usage: 127584, 10/10
OK: successful