Following the online example provided by Tensorflow I am having trouble using the custom op they define under GPU kernels. The instructions to build the example list three required files:
header file
// kernel_example.h
#ifndef KERNEL_EXAMPLE_H_
#define KERNEL_EXAMPLE_H_
template <typename Device, typename T>
struct ExampleFunctor {
void operator()(const Device& d, int size, const T* in, T* out);
};
#if GOOGLE_CUDA
// Partially specialize functor for GpuDevice.
template <typename Eigen::GpuDevice, typename T>
struct ExampleFunctor {
void operator()(const Eigen::GpuDevice& d, int size, const T* in, T* out);
};
#endif
#endif //KERNEL_EXAMPLE_H_ [1] commented out
((1) Here I have commented out KERNEL_EXAMPLE_H_ on the last line, as it causes a compile error.)
.cc file
// kernel_example.cc
#include "kernel_example.h" <--------[2] replaced example.h
#include "tensorflow/core/framework/op_kernel.h"
using namespace tensorflow;
using CPUDevice = Eigen::ThreadPoolDevice;
using GPUDevice = Eigen::GpuDevice;
// CPU specialization of actual computation.
template <typename T>
struct ExampleFunctor<CPUDevice, T> {
void operator()(const CPUDevice& d, int size, const T* in, T* out) {
for (int i = 0; i < size; ++i) {
out[i] = 2 * in[i];
}
}
};
// OpKernel definition.
// template parameter <T> is the datatype of the tensors.
template <typename Device, typename T>
class ExampleOp : public OpKernel {
public:
explicit ExampleOp(OpKernelConstruction* context) : OpKernel(context) {}
void Compute(OpKernelContext* context) override {
// Grab the input tensor
const Tensor& input_tensor = context->input(0);
// Create an output tensor
Tensor* output_tensor = NULL;
OP_REQUIRES_OK(context, context->allocate_output(0, input_tensor.shape(),
&output_tensor));
// Do the computation.
OP_REQUIRES(context, input_tensor.NumElements() <= tensorflow::kint32max,
errors::InvalidArgument("Too many elements in tensor"));
ExampleFunctor<Device, T>()(
context->eigen_device<Device>(),
static_cast<int>(input_tensor.NumElements()),
input_tensor.flat<T>().data(),
output_tensor->flat<T>().data());
}
};
// Register the CPU kernels.
#define REGISTER_CPU(T) \
REGISTER_KERNEL_BUILDER( \
Name("Example").Device(DEVICE_CPU).TypeConstraint<T>("T"), \
ExampleOp<CPUDevice, T>);
REGISTER_CPU(float);
REGISTER_CPU(int32);
// Register the GPU kernels.
#ifdef GOOGLE_CUDA
#define REGISTER_GPU(T) \
/* Declare explicit instantiations in kernel_example.cu.cc. */ \
extern template ExampleFunctor<GPUDevice, T>; \
REGISTER_KERNEL_BUILDER( \
Name("Example").Device(DEVICE_GPU).TypeConstraint<T>("T"), \
ExampleOp<GPUDevice, T>);
REGISTER_GPU(float);
REGISTER_GPU(int32);
#endif // GOOGLE_CUDA
([2] Here I have changed the name of the header file to match the filename.) and
.cu.cc file
// kernel_example.cu.cc
#ifdef GOOGLE_CUDA
#define EIGEN_USE_GPU
#include "kernel_example.h" //[3] replaced example.h
#include "tensorflow/core/util/cuda_kernel_helper.h"
using namespace tensorflow;
using GPUDevice = Eigen::GpuDevice;
// Define the CUDA kernel.
template <typename T>
__global__ void ExampleCudaKernel(const int size, const T* in, T* out) {
for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < size;
i += blockDim.x * gridDim.x) {
out[i] = 2 * ldg(in + i);
}
}
// Define the GPU implementation that launches the CUDA kernel.
template <typename T>
void ExampleFunctor<GPUDevice, T>::operator()(
const GPUDevice& d, int size, const T* in, T* out) {
// Launch the cuda kernel.
//
// See core/util/cuda_kernel_helper.h for example of computing
// block count and thread_per_block count.
int block_count = 1024;
int thread_per_block = 20;
ExampleCudaKernel<T>
<<<block_count, thread_per_block, 0, d.stream()>>>(size, in, out);
}
// Explicitly instantiate functors for the types of OpKernels registered.
template struct ExampleFunctor<GPUDevice, float>;
template struct ExampleFunctor<GPUDevice, int32>;
#endif // GOOGLE_CUDA
[3] Here I have changed the name of the header file to match the filename.
The only 3 minor changes I have made are listed below each script.
Building the op library using the suggested method:
TF_CFLAGS=( $(python -c 'import tensorflow as tf; print(" ".join(tf.sysconfig.get_compile_flags()))') )
TF_LFLAGS=( $(python -c 'import tensorflow as tf; print(" ".join(tf.sysconfig.get_link_flags()))') )
g++ -std=c++11 -shared kernel_example.cc kernel_example.cu.cc -o gpu_op.so -fPIC ${TF_CFLAGS[@]} ${TF_LFLAGS[@]} -O2
appears successful. And gpu_op.so is generated. But importing this op library and trying to use it:
# run_op.py
import tensorflow as tf
import numpy as np
my_module = tf.load_op_library('./gpu_op.so')
a = np.ones((20,5,5))
in1 = tf.convert_to_tensor(a, dtype = float)
print("input1: ", in1)
with tf.Session() as sess:
ans = sess.run(my_module.example(in1))
print("output:", ans)
results in the operation not being found:
File "run_op.py", line 11, in <module>
ans = sess.run(my_module.example(in1))
AttributeError: module '33c9073b4d33739023b5757fe9acdd79' has no attribute 'example'
I am relatively new to C++ and may not be compiling this correctly. So what should I be doing to make this module importable? And was I correct to make the 3 alterations to the code mentioned above?
Turns out I had overlooked that use of CUDA code in this example necessitates the use of the nvidia compiler nvcc.
can be compiled using:
TF_CFLAGS=( $(python -c 'import tensorflow as tf; print(" ".join(tf.sysconfig.get_compile_flags()))') )
TF_LFLAGS=( $(python -c 'import tensorflow as tf; print(" ".join(tf.sysconfig.get_link_flags()))') )
nvcc -std=c++11 cuda_op_kernel.cc cuda_op_kernel.cu.cc -o cuda_op_kernel.so -shared -Xcompiler -fPIC ${TF_CFLAGS[@]} ${TF_LFLAGS[@]} -O2