tensor_impl_CUDA.h

#ifndef TENSOR_IMPL_CUDA_H
#define TENSOR_IMPL_CUDA_H
 
#include "../tensor_impl.h"
#include "nforge/core/tensor_shape.h"
 
class Tensor::CUDAImpl : public Tensor::Impl {
public:
    CUDAImpl(const Tensor::Shape& shape);
    ~CUDAImpl();
 
    void fillAll(float value) override;
    void fillRand() override;
 
    void print() const override;
    void print(const std::vector<size_t>& position) const override;
 
    size_t getNumElements() const override;
    Tensor::Shape getShape() const override;
 
    float* dataPtr() const;
    std::vector<float> toVector() const override;
    std::string toString() const override;
 
    std::unique_ptr<Tensor::Impl> clone() const override;
 
    void copyFromHost(const float* data, size_t count) override;
 
    void set(const TensorLayout& lhsLayout, const Tensor::Impl* rhsImpl,
             const TensorLayout& rhsLayout) override;
 
    bool compare(const TensorLayout& lhsLayout, const Tensor::Impl* rhsImpl,
                 const TensorLayout& rhsLayout) const override;
 
    std::unique_ptr<Tensor::Impl> add(const TensorLayout& lhsLayout, const Tensor::Impl* rhsImpl,
                                      const TensorLayout& rhsLayout,
                                      const TensorLayout& outLayout) const override;
 
    std::unique_ptr<Tensor::Impl> sub(const TensorLayout& lhsLayout, const Tensor::Impl* rhsImpl,
                                      const TensorLayout& rhsLayout,
                                      const TensorLayout& outLayout) const override;
 
    std::unique_ptr<Tensor::Impl> mul(const TensorLayout& lhsLayout, const Tensor::Impl* rhsImpl,
                                      const TensorLayout& rhsLayout,
                                      const TensorLayout& outLayout) const override;
 
    std::unique_ptr<Tensor::Impl> div(const TensorLayout& lhsLayout, const Tensor::Impl* rhsImpl,
                                      const TensorLayout& rhsLayout,
                                      const TensorLayout& outLayout) const override;
 
    void iadd(const TensorLayout& lhsLayout, const Tensor::Impl* rhsImpl,
              const TensorLayout& rhsLayout) override;
 
    void isub(const TensorLayout& lhsLayout, const Tensor::Impl* rhsImpl,
              const TensorLayout& rhsLayout) override;
 
    void imul(const TensorLayout& lhsLayout, const Tensor::Impl* rhsImpl,
              const TensorLayout& rhsLayout) override;
 
    void idiv(const TensorLayout& lhsLayout, const Tensor::Impl* rhsImpl,
              const TensorLayout& rhsLayout) override;
 
    std::unique_ptr<Tensor::Impl> sum(const TensorLayout& layout, const TensorLayout& blockLayout,
                                      const TensorLayout& outLayout) const override;
 
    std::unique_ptr<Tensor::Impl> min(const TensorLayout& layout, const TensorLayout& blockLayout,
                                      const TensorLayout& outLayout) const override;
 
    std::unique_ptr<Tensor::Impl> max(const TensorLayout& layout, const TensorLayout& blockLayout,
                                      const TensorLayout& outLayout) const override;
 
    std::unique_ptr<Tensor::Impl> prod(const TensorLayout& layout, const TensorLayout& blockLayout,
                                       const TensorLayout& outLayout) const override;
 
    std::unique_ptr<Tensor::Impl> norm(const TensorLayout& layout) const override;
 
 
    std::unique_ptr<Tensor::Impl> matmul(const TensorLayout& lhsLayout, const Tensor::Impl* rhsImpl,
                                         const TensorLayout& rhsLayout,
                                         const TensorLayout& outLayout, size_t batch, size_t m,
                                         size_t k, size_t p) const override;
 
 
    std::unique_ptr<Tensor::Impl> less(const TensorLayout& lhsLayout, const Tensor::Impl* rhsImpl,
                                       const TensorLayout& rhsLayout,
                                       const TensorLayout& outLayout) const override;
 
    std::unique_ptr<Tensor::Impl> lessEqual(const TensorLayout& lhsLayout,
                                            const Tensor::Impl* rhsImpl,
                                            const TensorLayout& rhsLayout,
                                            const TensorLayout& outLayout) const override;
 
    std::unique_ptr<Tensor::Impl> greater(const TensorLayout& lhsLayout,
                                          const Tensor::Impl* rhsImpl,
                                          const TensorLayout& rhsLayout,
                                          const TensorLayout& outLayout) const override;
 
    std::unique_ptr<Tensor::Impl> greaterEqual(const TensorLayout& lhsLayout,
                                               const Tensor::Impl* rhsImpl,
                                               const TensorLayout& rhsLayout,
                                               const TensorLayout& outLayout) const override;
 
    std::unique_ptr<Tensor::Impl> isClose(const TensorLayout& lhsLayout,
                                          const Tensor::Impl* rhsImpl,
                                          const TensorLayout& rhsLayout,
                                          const TensorLayout& outLayout,
                                          float tolerance) const override;
 
private:
    Tensor::Shape m_shape;
    float* d_data;
 
    const Tensor::CUDAImpl* cast(const Tensor::Impl* p) const;
 
    template <typename Kernel>
    std::unique_ptr<Tensor::Impl> applyKernel(const TensorLayout& lhsLayout,
                                              const Tensor::Impl* rhsImpl,
                                              const TensorLayout& rhsLayout,
                                              const TensorLayout& outLayout, Kernel kernel) const;
 
    template <typename Kernel>
    void applyInplaceKernel(const TensorLayout& lhsLayout, const Tensor::Impl* rhsImpl,
                            const TensorLayout& rhsLayout, Kernel kernel);
 
    // kernel must be associative
    template <typename Kernel>
    std::unique_ptr<Tensor::Impl> applyReductionKernel(const TensorLayout& layout,
                                                       const TensorLayout& blockLayout,
                                                       const TensorLayout& outLayout,
                                                       float initValue, Kernel kernel) const;
};
 
#endif  // TENSOR_IMPL_CUDA_H