#include "CudaMatrix.h"
#include "AuroraDefs.h"
#include "Function1D.cuh"
#include "Function1D.h"
#include "Matrix.h"

#include <cmath>
#include <cstddef>
#include <cstdlib>
#include <thrust/device_vector.h>
#include <thrust/transform.h>
#include <thrust/iterator/constant_iterator.h>
#include <thrust/iterator/counting_iterator.h>
#include <thrust/complex.h>
#include <cuda_runtime.h>
#include <cusolverDn.h>

using namespace Aurora;
using namespace thrust::placeholders;

namespace
{
    const int THREADS_PER_BLOCK = 256;
    const int THREADS_PER_BLOCK_DIM2_X = 16;
    const int THREADS_PER_BLOCK_DIM2_Y = 16;
    const int THREADS_PER_BLOCK_DIM3_X = 8;
    const int THREADS_PER_BLOCK_DIM3_Y = 8;
    const int THREADS_PER_BLOCK_DIM3_Z = 8;
}

__global__ void complexKernel(float* aInputData, float* aOutput, unsigned int aSize)
{
    unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < aSize)
    {
        aOutput[2*idx] = aInputData[idx];
        aOutput[2*idx + 1] = 0;
    }
}

CudaMatrix Aurora::complex(const CudaMatrix& aMatrix)
{
    if(aMatrix.isComplex())
    {
        return CudaMatrix();
    }

    size_t size = aMatrix.getDataSize();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * aMatrix.getDataSize() * Aurora::Complex);
    int blocksPerGrid = (size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    complexKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(aMatrix.getData(), data, size);
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0), aMatrix.getDimSize(1), aMatrix.getDimSize(2), Aurora::Complex);
}

__global__ void realKernel(float* aInputData, float* aOutput, unsigned int aSize)
{
    unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < aSize)
    {
        aOutput[idx] = aInputData[idx*2];
    }
}

CudaMatrix Aurora::real(const CudaMatrix& aMatrix)
{
    if(!aMatrix.isComplex())
    {
        return CudaMatrix();
    }

    size_t size = aMatrix.getDataSize();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * aMatrix.getDataSize());
    int blocksPerGrid = (size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    realKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(aMatrix.getData(), data, size);
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0), aMatrix.getDimSize(1), aMatrix.getDimSize(2), Aurora::Normal);
}

__global__ void imageKernel(float* aInputData, float* aOutput, unsigned int aSize)
{
    unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < aSize)
    {
        aOutput[idx] = aInputData[idx*2 + 1];
    }
}

CudaMatrix Aurora::imag(const CudaMatrix& aMatrix)
{
    if(!aMatrix.isComplex())
    {
        return CudaMatrix();
    }

    size_t size = aMatrix.getDataSize();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * aMatrix.getDataSize());
    int blocksPerGrid = (size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    imageKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(aMatrix.getData(), data, size);
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0), aMatrix.getDimSize(1), aMatrix.getDimSize(2), Aurora::Normal);
}

__global__ void ceilKernel(float* aInputData, float* aOutput, unsigned int aSize)
{
    unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < aSize)
    {
        aOutput[idx] = std::ceil(aInputData[idx]);
    }
}

CudaMatrix Aurora::ceil(const CudaMatrix& aMatrix)
{
    size_t size = aMatrix.getDataSize() * aMatrix.getValueType();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size);
    int blocksPerGrid = (size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    ceilKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(aMatrix.getData(), data, size);
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0), aMatrix.getDimSize(1), aMatrix.getDimSize(2), aMatrix.getValueType());
}

CudaMatrix Aurora::ceil(const CudaMatrix&& aMatrix)
{
    size_t size = aMatrix.getDataSize() * aMatrix.getValueType();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size);
    int blocksPerGrid = (size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    ceilKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(aMatrix.getData(), data, size);
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0), aMatrix.getDimSize(1), aMatrix.getDimSize(2), aMatrix.getValueType());
}

__global__ void roundKernel(float* aInputData, float* aOutput, unsigned int aSize)
{
    unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < aSize)
    {
        aOutput[idx] = std::round(aInputData[idx]);
    }
}

CudaMatrix Aurora::round(const CudaMatrix& aMatrix)
{
    size_t size = aMatrix.getDataSize() * aMatrix.getValueType();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size);
    int blocksPerGrid = (size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    roundKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(aMatrix.getData(), data, size);
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0), aMatrix.getDimSize(1), aMatrix.getDimSize(2), aMatrix.getValueType());
}

CudaMatrix Aurora::round(const CudaMatrix&& aMatrix)
{
    size_t size = aMatrix.getDataSize() * aMatrix.getValueType();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size);
    int blocksPerGrid = (size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    roundKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(aMatrix.getData(), data, size);
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0), aMatrix.getDimSize(1), aMatrix.getDimSize(2), aMatrix.getValueType());
}

__global__ void floorKernel(float* aInputData, float* aOutput, unsigned int aSize)
{
    unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < aSize)
    {
        aOutput[idx] = std::floor(aInputData[idx]);
    }
}

CudaMatrix Aurora::floor(const CudaMatrix& aMatrix)
{
    size_t size = aMatrix.getDataSize() * aMatrix.getValueType();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size);
    int blocksPerGrid = (size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    floorKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(aMatrix.getData(), data, size);
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0), aMatrix.getDimSize(1), aMatrix.getDimSize(2), aMatrix.getValueType());
}

CudaMatrix Aurora::floor(const CudaMatrix&& aMatrix)
{
    size_t size = aMatrix.getDataSize() * aMatrix.getValueType();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size);
    int blocksPerGrid = (size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    floorKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(aMatrix.getData(), data, size);
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0), aMatrix.getDimSize(1), aMatrix.getDimSize(2), aMatrix.getValueType());
}

__global__ void sqrtKernel(float* aInputData, float* aOutput, unsigned int aSize)
{
    unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < aSize)
    {
        aOutput[idx] = std::sqrt(aInputData[idx]);
    }
}

CudaMatrix Aurora::sqrt(const CudaMatrix& aMatrix)
{
    if(aMatrix.getValueType() == Aurora::Complex)
    {
        std::cerr<<"sqrt not support complex"<<std::endl;
        return CudaMatrix();
    }
    size_t size = aMatrix.getDataSize() * aMatrix.getValueType();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size);
    int blocksPerGrid = (size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    sqrtKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(aMatrix.getData(), data, size);
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0), aMatrix.getDimSize(1), aMatrix.getDimSize(2), aMatrix.getValueType());
}

CudaMatrix Aurora::sqrt(const CudaMatrix&& aMatrix)
{
    if(aMatrix.getValueType() == Aurora::Complex)
    {
        std::cerr<<"sqrt not support complex"<<std::endl;
        return CudaMatrix();
    }
    size_t size = aMatrix.getDataSize() * aMatrix.getValueType();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size);
    int blocksPerGrid = (size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    sqrtKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(aMatrix.getData(), data, size);
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0), aMatrix.getDimSize(1), aMatrix.getDimSize(2), aMatrix.getValueType());
}

__global__ void absKernel(float* aInputData, float* aOutput, unsigned int aSize, bool aIsComplex)
{
    unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < aSize)
    {
        if(aIsComplex)
        {
            aOutput[idx] = sqrt(aInputData[2*idx] * aInputData[2*idx] +  aInputData[2*idx+1] * aInputData[2*idx+1]);
        }
        else
        {
            aOutput[idx] = abs(aInputData[idx]);
        }       
    }
}

CudaMatrix Aurora::abs(const CudaMatrix& aMatrix)
{
    size_t size = aMatrix.getDataSize();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size);
    int blocksPerGrid = (size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    absKernel<<<blocksPerGrid,THREADS_PER_BLOCK >>>(aMatrix.getData(), data, size, aMatrix.isComplex());
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0), aMatrix.getDimSize(1), aMatrix.getDimSize(2));
}

CudaMatrix Aurora::abs(const CudaMatrix&& aMatrix)
{
    size_t size = aMatrix.getDataSize();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size);
    int blocksPerGrid = (size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    absKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(aMatrix.getData(), data, size, aMatrix.isComplex());
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0), aMatrix.getDimSize(1), aMatrix.getDimSize(2));
}

__global__ void signKernel(float* aInputData, float* aOutput, unsigned int aSize, bool aIsComplex)
{
    unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < aSize)
    {
        if(aIsComplex)
        {
            float absValue = sqrt(aInputData[2*idx] * aInputData[2*idx] + aInputData[2*idx + 1] * aInputData[2*idx + 1]);
            aOutput[2*idx] = aInputData[2*idx] / absValue;
            aOutput[2*idx + 1] = aInputData[2*idx + 1] / absValue;
            return;
        }

        if(aInputData[idx] < 0)
        {
            aOutput[idx] = -1;
        }
        else if(aInputData[idx] > 0)
        {
            aOutput[idx] = 1;
        }
        else
        {
            aOutput[idx] = 0;
        }
    }
}

CudaMatrix Aurora::sign(const CudaMatrix& aMatrix)
{
    size_t size = aMatrix.getDataSize() * aMatrix.getValueType();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size);
    int blocksPerGrid = (size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    signKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(aMatrix.getData(), data, aMatrix.getDataSize(), aMatrix.isComplex());
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0), aMatrix.getDimSize(1), aMatrix.getDimSize(2), aMatrix.getValueType());
}

CudaMatrix Aurora::sign(const CudaMatrix&& aMatrix)
{
    size_t size = aMatrix.getDataSize() * aMatrix.getValueType();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size);
    int blocksPerGrid = (size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    signKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(aMatrix.getData(), data, aMatrix.getDataSize(), aMatrix.isComplex());
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0), aMatrix.getDimSize(1), aMatrix.getDimSize(2), aMatrix.getValueType());
}

void Aurora::nantoval(CudaMatrix& aMatrix,float val){
    auto lambda = [=] __host__ __device__ (const float& x){
        return ::isnan(x)?val:x;
    };
    thrust::transform(thrust::device,aMatrix.getData(),
        aMatrix.getData()+aMatrix.getDataSize(),aMatrix.getData(),lambda);
}

CudaMatrix Aurora::isnan(const CudaMatrix& aMatrix){
    size_t size = aMatrix.getDataSize() * aMatrix.getValueType();
    float* data = nullptr;
    
    cudaMalloc((void**)&data, sizeof(float) * size);
        auto lambda = [=] __host__ __device__ (const float& x){
        return ::isnan(x)?1.0:0;
    };
    thrust::transform(thrust::device,aMatrix.getData(),aMatrix.getData()+aMatrix.getDataSize(),
        data,lambda);
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0), 
        aMatrix.getDimSize(1), aMatrix.getDimSize(2), aMatrix.getValueType());    
}

CudaMatrix Aurora::isfinite(const CudaMatrix& aMatrix){
    size_t size = aMatrix.getDataSize() * aMatrix.getValueType();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size);
        auto lambda = [=] __host__ __device__ (const float& x){
        return ::isfinite(x)?1.0:0;
    };
    thrust::transform(thrust::device,aMatrix.getData(),aMatrix.getData()+aMatrix.getDataSize(),
        data,lambda);
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0), 
        aMatrix.getDimSize(1), aMatrix.getDimSize(2), aMatrix.getValueType());  
}

void Aurora::padding(CudaMatrix& aMatrix, int aIndex, float aValue){
    if(aMatrix.isNull() || !aMatrix.isVector() || aMatrix.isComplex())
    {
        std::cerr<<"padding only support real vector"<<std::endl;
        return;
    }
    if (aMatrix.getDataSize()>aIndex){
        aMatrix.setValue(aIndex,  aValue);
        return;
    }
    //长度不足需补齐
    size_t size = (aIndex+1) ;
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size);
    cudaMemcpy(data, aMatrix.getData(), aMatrix.getDataSize(), cudaMemcpyDeviceToDevice);
    thrust::fill_n(thrust::device, data+aMatrix.getDataSize(),size-aMatrix.getDataSize(),aValue);
    aMatrix=CudaMatrix::fromRawData(data,size,1,1,aMatrix.getValueType());
}


CudaMatrix Aurora::auroraNot(const CudaMatrix& aMatrix){
    return auroraNot(std::forward<CudaMatrix&&>(aMatrix.deepCopy()));
}

CudaMatrix Aurora::auroraNot(CudaMatrix&& aMatrix){
    size_t size = aMatrix.getDataSize() * aMatrix.getValueType();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size);
    auto lambda = [=] __host__ __device__ (const float& x){
        return x<=0?1.0:0;
    };
    thrust::transform(thrust::device,aMatrix.getData(),aMatrix.getData()+aMatrix.getDataSize(),
        data,lambda);
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0), 
        aMatrix.getDimSize(1), aMatrix.getDimSize(2), aMatrix.getValueType());    
}


void Aurora::compareSet(CudaMatrix& aValueMatrix,float compareValue, float newValue,CompareOp op)
{
    switch (op)
    {
    case GT:
    {
        auto lambda = [=] __host__ __device__ (const float& x){
            return x>compareValue?newValue:x;
        };
        thrust::transform(thrust::device,aValueMatrix.getData(),aValueMatrix.getData()+aValueMatrix.getDataSize(),aValueMatrix.getData(),lambda);
        break;
    }
    case NG:{
        auto lambda = [=] __host__ __device__ (const float& x){
            return x<=compareValue?newValue:x;
        };
        thrust::transform(thrust::device,aValueMatrix.getData(),aValueMatrix.getData()+aValueMatrix.getDataSize(),aValueMatrix.getData(),lambda);
        break;
    }
    case EQ:{
        auto lambda = [=] __host__ __device__ (const float& x){
            return x==compareValue?newValue:x;
        };
        thrust::transform(thrust::device,aValueMatrix.getData(),aValueMatrix.getData()+aValueMatrix.getDataSize(),aValueMatrix.getData(),lambda);
        break;
    }
    case NE:{
        auto lambda = [=] __host__ __device__ (const float& x){
            return x!=compareValue?newValue:x;
        };
        thrust::transform(thrust::device,aValueMatrix.getData(),aValueMatrix.getData()+aValueMatrix.getDataSize(),aValueMatrix.getData(),lambda);
        break;
    }
    case NL:{
        auto lambda = [=] __host__ __device__ (const float& x){
            return x>=compareValue?newValue:x;
        };
        thrust::transform(thrust::device,aValueMatrix.getData(),aValueMatrix.getData()+aValueMatrix.getDataSize(),aValueMatrix.getData(),lambda);
        break;
    }
    case LT:{
        auto lambda = [=] __host__ __device__ (const float& x){
            return x<compareValue?newValue:x;
        };
        thrust::transform(thrust::device,aValueMatrix.getData(),aValueMatrix.getData()+aValueMatrix.getDataSize(),aValueMatrix.getData(),lambda);
        break;
    }
    default:
        break;
    }

}

void Aurora::compareSet(CudaMatrix& aValueMatrix,CudaMatrix& aCompareMatrix,float compareValue, float newValue,CompareOp op){
    switch (op)
    {
    case GT:
    {
        auto lambda = [=] __host__ __device__ (const float& x, const float& y){
            return x>compareValue?newValue:y;
        };
        thrust::transform(thrust::device,aCompareMatrix.getData(),
            aCompareMatrix.getData()+aValueMatrix.getDataSize(),
            aValueMatrix.getData(), aValueMatrix.getData(),
            lambda);
        break;
    }
    case NG:{
        auto lambda = [=] __host__ __device__ (const float& x, const float& y){
            return x<=compareValue?newValue:y;
        };
        thrust::transform(thrust::device,aCompareMatrix.getData(),
            aCompareMatrix.getData()+aValueMatrix.getDataSize(),
            aValueMatrix.getData(), aValueMatrix.getData(),
            lambda);
        break;
    }
    case EQ:{
        auto lambda = [=] __host__ __device__ (const float& x, const float& y){
            return x==compareValue?newValue:y;
        };
        thrust::transform(thrust::device,aCompareMatrix.getData(),
            aCompareMatrix.getData()+aValueMatrix.getDataSize(),
            aValueMatrix.getData(), aValueMatrix.getData(),
            lambda);
        break;
    }
    case NE:{
        auto lambda = [=] __host__ __device__ (const float& x, const float& y){
            return x!=compareValue?newValue:y;
        };
        thrust::transform(thrust::device,aCompareMatrix.getData(),
            aCompareMatrix.getData()+aValueMatrix.getDataSize(),
            aValueMatrix.getData(), aValueMatrix.getData(),
            lambda);
        break;
    }
    case NL:{
        auto lambda = [=] __host__ __device__ (const float& x, const float& y){
            return x>=compareValue?newValue:y;
        };
        thrust::transform(thrust::device,aCompareMatrix.getData(),
            aCompareMatrix.getData()+aValueMatrix.getDataSize(),
            aValueMatrix.getData(), aValueMatrix.getData(),
            lambda);
        break;
    }
    case LT:{
        auto lambda = [=] __host__ __device__ (const float& x, const float& y){
            return x<compareValue?newValue:y;
        };
        thrust::transform(thrust::device,aCompareMatrix.getData(),
            aCompareMatrix.getData()+aValueMatrix.getDataSize(),
            aValueMatrix.getData(), aValueMatrix.getData(),
            lambda);
        break;
    }
    default:
        break;
    }
}

void Aurora::compareSet(CudaMatrix& aDesAndCompareMatrix,CudaMatrix& aOtherCompareMatrix, float newValue,CompareOp op){
    switch (op)
    {
    case GT:
    {
        auto lambda = [=] __host__ __device__ (const float& x, const float& y){
            return x>y?newValue:x;
        };
        thrust::transform(thrust::device,aDesAndCompareMatrix.getData(),
            aDesAndCompareMatrix.getData()+aDesAndCompareMatrix.getDataSize(),
            aOtherCompareMatrix.getData(), aDesAndCompareMatrix.getData(),
            lambda);
        break;
    }
    case NG:{
        auto lambda = [=] __host__ __device__ (const float& x, const float& y){
            return x<=y?newValue:x;
        };
        thrust::transform(thrust::device,aDesAndCompareMatrix.getData(),
            aDesAndCompareMatrix.getData()+aDesAndCompareMatrix.getDataSize(),
            aOtherCompareMatrix.getData(), aDesAndCompareMatrix.getData(),
            lambda);
        break;
    }
    case EQ:{
        auto lambda = [=] __host__ __device__ (const float& x, const float& y){
            return x==y?newValue:x;
        };
        thrust::transform(thrust::device,aDesAndCompareMatrix.getData(),
            aDesAndCompareMatrix.getData()+aDesAndCompareMatrix.getDataSize(),
            aOtherCompareMatrix.getData(), aDesAndCompareMatrix.getData(),
            lambda);
        break;
    }
    case NE:{
        auto lambda = [=] __host__ __device__ (const float& x, const float& y){
            return x!=y?newValue:x;
        };
        thrust::transform(thrust::device,aDesAndCompareMatrix.getData(),
            aDesAndCompareMatrix.getData()+aDesAndCompareMatrix.getDataSize(),
            aOtherCompareMatrix.getData(), aDesAndCompareMatrix.getData(),
            lambda);
        break;
    }
    case NL:{
        auto lambda = [=] __host__ __device__ (const float& x, const float& y){
            return x>=y?newValue:x;
        };
        thrust::transform(thrust::device,aDesAndCompareMatrix.getData(),
            aDesAndCompareMatrix.getData()+aDesAndCompareMatrix.getDataSize(),
            aOtherCompareMatrix.getData(), aDesAndCompareMatrix.getData(),
            lambda);
        break;
    }
    case LT:{
        auto lambda = [=] __host__ __device__ (const float& x, const float& y){
            return x<y?newValue:x;
        };
        thrust::transform(thrust::device,aDesAndCompareMatrix.getData(),
            aDesAndCompareMatrix.getData()+aDesAndCompareMatrix.getDataSize(),
            aOtherCompareMatrix.getData(), aDesAndCompareMatrix.getData(),
            lambda);
        break;
    }
    default:
        break;
    }
}

void Aurora::compareSet(CudaMatrix& aCompareMatrix,float compareValue, CudaMatrix& aNewValueMatrix,CompareOp op)
{
    switch (op)
    {
    case GT:
    {
        auto lambda = [=] __host__ __device__ (const float& x, const float& y){
            return x>compareValue?y:x;
        };
        thrust::transform(thrust::device,aCompareMatrix.getData(),
            aCompareMatrix.getData()+aCompareMatrix.getDataSize(),
            aNewValueMatrix.getData(), aCompareMatrix.getData(),
            lambda);
        break;
    }
    case NG:{
        auto lambda = [=] __host__ __device__ (const float& x, const float& y){
            return x<=compareValue?y:x;
        };
        thrust::transform(thrust::device,aCompareMatrix.getData(),
            aCompareMatrix.getData()+aCompareMatrix.getDataSize(),
            aNewValueMatrix.getData(), aCompareMatrix.getData(),
            lambda);
        break;
    }
    case EQ:{
        auto lambda = [=] __host__ __device__ (const float& x, const float& y){
            return x==compareValue?y:x;
        };
        thrust::transform(thrust::device,aCompareMatrix.getData(),
            aCompareMatrix.getData()+aCompareMatrix.getDataSize(),
            aNewValueMatrix.getData(), aCompareMatrix.getData(),
            lambda);
        break;
    }
    case NE:{
        auto lambda = [=] __host__ __device__ (const float& x, const float& y){
            return x!=compareValue?y:x;
        };
        thrust::transform(thrust::device,aCompareMatrix.getData(),
            aCompareMatrix.getData()+aCompareMatrix.getDataSize(),
            aNewValueMatrix.getData(), aCompareMatrix.getData(),
            lambda);
        break;
    }
    case NL:{
        auto lambda = [=] __host__ __device__ (const float& x, const float& y){
            return x>=compareValue?y:x;
        };
        thrust::transform(thrust::device,aCompareMatrix.getData(),
            aCompareMatrix.getData()+aCompareMatrix.getDataSize(),
            aNewValueMatrix.getData(), aCompareMatrix.getData(),
            lambda);
        break;
    }
    case LT:{
        auto lambda = [=] __host__ __device__ (const float& x, const float& y){
            return x<compareValue?y:x;
        };
        thrust::transform(thrust::device,aCompareMatrix.getData(),
            aCompareMatrix.getData()+aCompareMatrix.getDataSize(),
            aNewValueMatrix.getData(), aCompareMatrix.getData(),
            lambda);
        break;
    }
    default:
        break;
    }
}

__global__ void repMatKernel(float* aInputData, unsigned int aInputRows, unsigned int aInputColumns,
                             float* aOutput, unsigned int aOutputRows, unsigned int aOutputColumns, unsigned int aOutputSlices, bool aIsComplex)
{
    unsigned int idX = blockIdx.x * blockDim.x + threadIdx.x;
    unsigned int idY = blockIdx.y * blockDim.y + threadIdx.y;
    unsigned int idZ = blockIdx.z * blockDim.z + threadIdx.z;

    if(idX >= aOutputRows || idY >= aOutputColumns || idZ >= aOutputSlices)
    {
        return;
    }

    if(aIsComplex)
    {
        unsigned int outPutIndex = 2 * (idZ * aOutputRows * aOutputColumns + idY * aOutputRows + idX);
        unsigned int inPutIndex = 2 * (idY % aInputColumns * aInputRows + idX % aInputRows);
        aOutput[outPutIndex] = aInputData[inPutIndex];
        aOutput[outPutIndex + 1] = aInputData[inPutIndex + 1];
    }
    else
    {
        aOutput[idZ * aOutputRows * aOutputColumns + idY * aOutputRows + idX] = aInputData[idY % aInputColumns * aInputRows + idX % aInputRows];
    }
}

CudaMatrix Aurora::repmat(const CudaMatrix& aMatrix,int aRowTimes, int aColumnTimes)
{
    if(aRowTimes < 1 || aColumnTimes < 1 || aMatrix.getDims() > 2 || aMatrix.isNull())
    {
        return CudaMatrix();
    }
    size_t rowSize = aMatrix.getDimSize(0) * aRowTimes;
    size_t columnSize = aMatrix.getDimSize(1) * aColumnTimes;
    dim3 blockSize(THREADS_PER_BLOCK_DIM2_X, THREADS_PER_BLOCK_DIM2_Y, 1);
    dim3 gridSize((rowSize + THREADS_PER_BLOCK_DIM2_X - 1)/THREADS_PER_BLOCK_DIM2_X, 
                  (columnSize + THREADS_PER_BLOCK_DIM2_Y - 1)/THREADS_PER_BLOCK_DIM2_Y, 1);

    size_t size = aMatrix.getDataSize() * aMatrix.getValueType() * aRowTimes * aColumnTimes;
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size);
    repMatKernel<<<gridSize, blockSize>>>(aMatrix.getData(), aMatrix.getDimSize(0), aMatrix.getDimSize(1), data, rowSize, columnSize, 1, aMatrix.isComplex());
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0) * aRowTimes, aMatrix.getDimSize(1) * aColumnTimes, aMatrix.getDimSize(2), aMatrix.getValueType());
}

CudaMatrix Aurora::repmat(const CudaMatrix& aMatrix,int aRowTimes, int aColumnTimes, int aSliceTimes)
{
    if(aRowTimes < 1 || aColumnTimes < 1 || aMatrix.getDims() > 2 || aMatrix.isNull())
    {
        return CudaMatrix();
    }

    size_t rowSize = aMatrix.getDimSize(0) * aRowTimes;
    size_t columnSize = aMatrix.getDimSize(1) * aColumnTimes;
    size_t sliceSize = aMatrix.getDimSize(2) * aSliceTimes;
    dim3 blockSize(THREADS_PER_BLOCK_DIM3_X, THREADS_PER_BLOCK_DIM3_Y, THREADS_PER_BLOCK_DIM3_Z);
    dim3 gridSize((rowSize + THREADS_PER_BLOCK_DIM3_X - 1)/THREADS_PER_BLOCK_DIM3_X, 
                  (columnSize + THREADS_PER_BLOCK_DIM3_Y - 1)/THREADS_PER_BLOCK_DIM3_Y,
                  (sliceSize + THREADS_PER_BLOCK_DIM3_Z - 1)/THREADS_PER_BLOCK_DIM3_Z);

    size_t size = aMatrix.getDataSize() * aMatrix.getValueType() * aRowTimes * aColumnTimes * aSliceTimes;
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size);
    repMatKernel<<<gridSize, blockSize>>>(aMatrix.getData(), aMatrix.getDimSize(0), aMatrix.getDimSize(1), data, rowSize, columnSize, sliceSize, aMatrix.isComplex());
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0) * aRowTimes, aMatrix.getDimSize(1) * aColumnTimes, aMatrix.getDimSize(2) * aSliceTimes, aMatrix.getValueType());
}

__global__ void repMat3DKernel(float* aInputData, unsigned int aInputRows, unsigned int aInputColumns, unsigned int aInputSlices,
                               float* aOutput, unsigned int aOutputRows, unsigned int aOutputColumns, unsigned int aOutputSlices, bool aIsComplex)
{
    unsigned int idX = blockIdx.x * blockDim.x + threadIdx.x;
    unsigned int idY = blockIdx.y * blockDim.y + threadIdx.y;
    unsigned int idZ = blockIdx.z * blockDim.z + threadIdx.z;

    if(idX >= aOutputRows || idY >= aOutputColumns || idZ >= aOutputSlices)
    {
        return;
    }

    if(aIsComplex)
    {
        unsigned int outPutIndex = 2 * (idZ * aOutputRows * aOutputColumns + idY * aOutputRows + idX);
        unsigned int inPutIndex = 2 * (idZ % aInputSlices * aInputRows * aInputColumns + idY % aInputColumns * aInputRows + idX % aInputRows);
        aOutput[outPutIndex] = aInputData[inPutIndex];
        aOutput[outPutIndex + 1] = aInputData[inPutIndex + 1];
    }
    else
    {
        aOutput[idZ * aOutputRows * aOutputColumns + idY * aOutputRows + idX] = aInputData[idZ % aInputSlices * aInputRows * aInputColumns + idY % aInputColumns * aInputRows + idX % aInputRows];
    }

}

CudaMatrix Aurora::repmat3d(const CudaMatrix& aMatrix,int aRowTimes, int aColumnTimes, int aSliceTimes)
{
    if(aRowTimes < 1 || aColumnTimes < 1 || aMatrix.getDims() < 3 || aMatrix.isNull())
    {
        return CudaMatrix();
    }

    size_t rowSize = aMatrix.getDimSize(0) * aRowTimes;
    size_t columnSize = aMatrix.getDimSize(1) * aColumnTimes;
    size_t sliceSize = aMatrix.getDimSize(2) * aSliceTimes;
    dim3 blockSize(THREADS_PER_BLOCK_DIM3_X, THREADS_PER_BLOCK_DIM3_Y, THREADS_PER_BLOCK_DIM3_Z);
    dim3 gridSize((rowSize + THREADS_PER_BLOCK_DIM3_X - 1)/THREADS_PER_BLOCK_DIM3_X, 
                  (columnSize + THREADS_PER_BLOCK_DIM3_Y - 1)/THREADS_PER_BLOCK_DIM3_Y,
                  (sliceSize + THREADS_PER_BLOCK_DIM3_Z - 1)/THREADS_PER_BLOCK_DIM3_Z);

    size_t size = aMatrix.getDataSize() * aMatrix.getValueType() * aRowTimes * aColumnTimes * aSliceTimes;
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size);
    repMat3DKernel<<<gridSize, blockSize>>>(aMatrix.getData(),  aMatrix.getDimSize(0), aMatrix.getDimSize(1), aMatrix.getDimSize(2), data, rowSize, columnSize, sliceSize, aMatrix.isComplex());
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0) * aRowTimes, aMatrix.getDimSize(1) * aColumnTimes, aMatrix.getDimSize(2) * aSliceTimes, aMatrix.getValueType());
}

__global__ void logKernel(float* aInputData, float* aOutput, unsigned int aInputSize, int aBaseNum)
{
    unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < aInputSize)
    {
        if(aBaseNum == -1)
        {
            aOutput[idx] = logf(aInputData[idx]); 
        }
        else
        {
            float value = logf(aBaseNum);
            aOutput[idx] = logf(aInputData[idx]) / value; 
        }
    }
}

CudaMatrix Aurora::log(const CudaMatrix& aMatrix, int aBaseNum)
{
    if(aMatrix.getValueType() == Aurora::Complex)
    {
        std::cerr<<"log not support complex"<<std::endl;
        return CudaMatrix();
    }
    size_t size = aMatrix.getDataSize();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size);
    int blocksPerGrid = (size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    logKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(aMatrix.getData(), data, size, aBaseNum);
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0), aMatrix.getDimSize(1), aMatrix.getDimSize(2), aMatrix.getValueType());
}

__global__ void expKernel(float* aInputData, float* aOutput, unsigned int aInputSize, bool aIsComplex)
{
    unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < aInputSize)
    {
        if(aIsComplex)
        {
            unsigned int index = 2 * idx;
            float expReal = expf(aInputData[index]);
            aOutput[index] = expReal * cosf(aInputData[index + 1]);
            aOutput[index + 1] = expReal * sinf(aInputData[index + 1]);
        }
        else
        {
            aOutput[idx] =  expf(aInputData[idx]); 
        }
    }
}

CudaMatrix Aurora::exp(const CudaMatrix& aMatrix)
{
    size_t size = aMatrix.getDataSize();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size * aMatrix.getValueType());
    int blocksPerGrid = (size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    expKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(aMatrix.getData(), data, size, aMatrix.isComplex());
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0), aMatrix.getDimSize(1), aMatrix.getDimSize(2), aMatrix.getValueType());
}

__global__ void modKernel(float* aInputData, float* aOutput, unsigned int aInputSize, float aModValue)
{
    unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < aInputSize)
    {
        aOutput[idx] =  fmodf(aInputData[idx], aModValue); 
    }
}

CudaMatrix Aurora::mod(const CudaMatrix& aMatrix, float aValue)
{
    if(aMatrix.isComplex())
    {
        std::cerr<<"mod not support complex"<<std::endl;
        return CudaMatrix();
    }

    size_t size = aMatrix.getDataSize();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size);
    int blocksPerGrid = (size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    modKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(aMatrix.getData(), data, size, aValue);
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0), aMatrix.getDimSize(1), aMatrix.getDimSize(2));
}

__global__ void acosKernel(float* aInputData, float* aOutput, unsigned int aInputSize)
{
    unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < aInputSize)
    {
        aOutput[idx] =  acosf(aInputData[idx]); 
    }
}

CudaMatrix Aurora::acos(const CudaMatrix& aMatrix)
{
    if(aMatrix.isComplex())
    {
        std::cerr<<"acos not support complex"<<std::endl;
        return CudaMatrix();
    }

    size_t size = aMatrix.getDataSize();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size);
    int blocksPerGrid = (size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    acosKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(aMatrix.getData(), data, size);
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0), aMatrix.getDimSize(1), aMatrix.getDimSize(2));
}

__global__ void acosdKernel(float* aInputData, float* aOutput, unsigned int aInputSize)
{
    unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < aInputSize)
    {
        aOutput[idx] =  acosf(aInputData[idx]) * 180 / PI; 
    }
}

CudaMatrix Aurora::acosd(const CudaMatrix& aMatrix)
{
    if(aMatrix.isComplex())
    {
        std::cerr<<"acos not support complex"<<std::endl;
        return CudaMatrix();
    }

    size_t size = aMatrix.getDataSize();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size);
    int blocksPerGrid = (size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    acosdKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(aMatrix.getData(), data, size);
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0), aMatrix.getDimSize(1), aMatrix.getDimSize(2));
}

__global__ void conjKernel(float* aInputData, float* aOutput, unsigned int aInputSize)
{
    unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < aInputSize)
    {
        unsigned int index = idx * 2;
        aOutput[index] =  aInputData[index]; 
        aOutput[index + 1] = -aInputData[index + 1];
    }
}

CudaMatrix Aurora::conj(const CudaMatrix& aMatrix)
{
    if(!aMatrix.isComplex())
    {
        return CudaMatrix::copyFromRawData(aMatrix.getData(),aMatrix.getDimSize(0),aMatrix.getDimSize(1),aMatrix.getDimSize(2));
    }
    size_t size = aMatrix.getDataSize();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size * aMatrix.getValueType());
    int blocksPerGrid = (size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    conjKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(aMatrix.getData(), data, size);
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(0), aMatrix.getDimSize(1), aMatrix.getDimSize(2), aMatrix.getValueType());
}


float Aurora::norm(const CudaMatrix& aMatrix, NormMethod aNormMethod)
{
    float resultValue = 0;
    if(aMatrix.getDims() > 2)
    {
        std::cerr<<"norm not support 3d matrix"<<std::endl;
        return 0;
    }
    //for 1 dims
    if(aMatrix.getDimSize(0) == 1 || aMatrix.getDimSize(1) == 1 )
    {
        if(aNormMethod == Aurora::Norm1)
        {
            CudaMatrix result = abs(aMatrix);
            resultValue = thrust::reduce(thrust::device, result.getData(), result.getData() + result.getDataSize(),  0.0, thrust::plus<float>());
            return resultValue;
        }
        else
        {
            CudaMatrix result = aMatrix.deepCopy();
            thrust::transform(thrust::device, result.getData(), result.getData() + result.getDataSize() * result.getValueType(),  result.getData(), thrust::square<float>());
            resultValue = thrust::reduce(thrust::device, result.getData(), result.getData() + result.getDataSize() * result.getValueType(),  0.0, thrust::plus<float>());
            return std::sqrt(resultValue);
        }
    }
    //for 2 dims
    if(aNormMethod == Aurora::NormF)
    {
        CudaMatrix result = aMatrix.deepCopy();
        thrust::transform(thrust::device, result.getData(), result.getData() + result.getDataSize() * result.getValueType(),  result.getData(), thrust::square<float>());
        resultValue = thrust::reduce(thrust::device, result.getData(), result.getData() + result.getDataSize() * result.getValueType(),  0.0, thrust::plus<float>());
        return std::sqrt(resultValue);
    }
    else if(aNormMethod == Aurora::Norm1)
    {
        for(int i=0; i<aMatrix.getDimSize(1); ++i)
        {
            CudaMatrix result = abs(aMatrix.block(1, i, i));
            float tempValue = thrust::reduce(thrust::device, result.getData(), result.getData() + result.getDataSize(),  0.0, thrust::plus<float>());
            if(resultValue < tempValue)
            {
                resultValue = tempValue;
            }
        }
        return resultValue;
    }
    else
    {
        if(aMatrix.isComplex())
        {
            std::cerr<<"norm2 not support 2d complex matrix"<<std::endl;
            return 0;
        }
        cusolverDnHandle_t cusolverH = NULL;
        cudaStream_t stream = NULL;
        float* d_S = NULL;
        float* d_U = NULL;
        float* d_VT = NULL;
        int* devInfo = NULL;
        float* d_work = NULL;
        int lwork = 0;
        const int m = aMatrix.getDimSize(0);
        const int n = aMatrix.getDimSize(1);
        const int lda = m;
        const int ldu = m;
        const int ldvt = n;

        cusolverDnCreate(&cusolverH);
        cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking);
        cusolverDnSetStream(cusolverH, stream);
        cudaMalloc((void**)&d_S, sizeof(float)*n);
        cudaMalloc((void**)&d_U, sizeof(float)*ldu*m);
        cudaMalloc((void**)&d_VT, sizeof(float)*ldvt*n);
        cudaMalloc((void**)&devInfo, sizeof(int));

        cusolverDnSgesvd_bufferSize(cusolverH, m, n, &lwork);

        cudaMalloc((void**)&d_work, sizeof(float)*lwork);
        auto matrix = aMatrix.deepCopy();
        cusolverDnSgesvd(cusolverH, 'A', 'A', m, n, matrix.getData(), lda, d_S,d_U, ldu, d_VT, ldvt, d_work, lwork, NULL, devInfo);

        int devInfo_h = 0;
        cudaMemcpy(&devInfo_h, devInfo, sizeof(int), cudaMemcpyDeviceToHost);

        if (devInfo_h != 0) 
        {
            printf("Unsuccessful SVD execution\n");
            printf("Error code: %d\n", devInfo_h);
        }

        float S[n] = {0};  
        cudaMemcpy(S, d_S, sizeof(float)*n, cudaMemcpyDeviceToHost);

        float resultValue = S[0];
        for (int i = 1; i < n; i++)
        { 
            if (S[i] > resultValue)
            {
                resultValue = S[i];
            }
        }

        if (d_S ) cudaFree(d_S);
        if (d_U ) cudaFree(d_U);
        if (d_VT) cudaFree(d_VT);
        if (devInfo) cudaFree(devInfo);
        if (d_work) cudaFree(d_work);
        if (cusolverH) cusolverDnDestroy(cusolverH);
        if (stream) cudaStreamDestroy(stream);
        return resultValue;
    }
}

__global__ void transposeKernel(float* aInputData, float* aOutput, unsigned int aRowSize, unsigned int aColumnSize, bool aIsComplex)
{
    unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
    unsigned int idy = blockIdx.y * blockDim.y + threadIdx.y;
    if (idx < aRowSize && idy < aColumnSize)
    {
        unsigned int inputindex = idy * aRowSize + idx;
        unsigned int outputIndex = idx * aColumnSize + idy;
        if(aIsComplex)
        {
            aOutput[2*outputIndex] =  aInputData[2*inputindex]; 
            aOutput[2*outputIndex + 1] =  aInputData[2*inputindex + 1]; 
        }
        else
        {
            aOutput[outputIndex] =  aInputData[inputindex]; 
        }
    }
}


CudaMatrix Aurora::transpose(const CudaMatrix& aMatrix)
{
    //not surpport for 3 dims.
    if(aMatrix.isNull() || aMatrix.getDimSize(2) > 1)
    {
        std::cerr<<"transpose not support 3d complex matrix"<<std::endl;
        return CudaMatrix();
    }

    size_t size = aMatrix.getDataSize();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size * aMatrix.getValueType());
    dim3 blocksPerGrid((aMatrix.getDimSize(0) + THREADS_PER_BLOCK_DIM2_X - 1) / THREADS_PER_BLOCK_DIM2_X, (aMatrix.getDimSize(1) + THREADS_PER_BLOCK_DIM2_Y - 1) / THREADS_PER_BLOCK_DIM2_Y, 1);
    dim3 threadPerBlock(THREADS_PER_BLOCK_DIM2_X, THREADS_PER_BLOCK_DIM2_X, 1);
    transposeKernel<<<blocksPerGrid, threadPerBlock>>>(aMatrix.getData(), data, aMatrix.getDimSize(0), aMatrix.getDimSize(1), aMatrix.isComplex());
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, aMatrix.getDimSize(1), aMatrix.getDimSize(0), aMatrix.getDimSize(2), aMatrix.getValueType());
}

CudaMatrix Aurora::horzcat(const CudaMatrix& aMatrix1, const CudaMatrix& aMatrix2)
{
    if(aMatrix1.isNull() || aMatrix2.isNull() || aMatrix1.getDimSize(2) != aMatrix2.getDimSize(2) ||
       aMatrix1.getDimSize(0) != aMatrix2.getDimSize(0) || aMatrix1.getValueType() != aMatrix2.getValueType())
    {
        std::cerr<<"horzcat must have same rows and slices"<<std::endl;
        return CudaMatrix();
    }

    int column1 = aMatrix1.getDimSize(1);
    int column2 = aMatrix2.getDimSize(1);
    int slice = aMatrix1.getDimSize(2);
    int row = aMatrix1.getDimSize(0);
    size_t size1= row*column1*aMatrix1.getValueType();
    size_t size2= row*column2*aMatrix2.getValueType();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * (aMatrix1.getDataSize() + aMatrix2.getDataSize()) * aMatrix1.getValueType());
    size_t sliceStride = row*(column1+column2) * aMatrix1.getValueType();
    for (size_t i = 0; i < slice; i++)
    {
        cudaMemcpy(data + i*sliceStride, aMatrix1.getData() + i*size1, sizeof(float) * size1, cudaMemcpyDeviceToDevice);
        cudaMemcpy(data + i*sliceStride + size1, aMatrix2.getData() + i*size2, sizeof(float) * size2, cudaMemcpyDeviceToDevice);
    }
    return CudaMatrix::fromRawData(data, row, column1+column2, slice, aMatrix1.getValueType());
}

__global__ void vertcatKernel(float* aInputData1, unsigned int aInputData1RowSize, float* aInputData2, unsigned int aInputData2RowSize,
                              float* aOutput, unsigned int aOutputRowSize, unsigned int aOutputColumnSize, unsigned int aOutputSliceSize, bool aIsComplex)
{
    unsigned int idX = blockIdx.x * blockDim.x + threadIdx.x;
    unsigned int idY = blockIdx.y * blockDim.y + threadIdx.y;
    unsigned int idZ = blockIdx.z * blockDim.z + threadIdx.z;

    if(idX >= aOutputRowSize || idY >= aOutputColumnSize || idZ >= aOutputSliceSize)
    {
        return;
    }

    if(aIsComplex)
    {
        if(idX < aInputData1RowSize)
        {
            unsigned int inputIndex = idZ * aInputData1RowSize * aOutputColumnSize + idY * aInputData1RowSize + idX;
            unsigned int outputIndex = idZ * aOutputRowSize * aOutputColumnSize + idY * aOutputRowSize + idX;
            aOutput[2*outputIndex] = aInputData1[2*inputIndex];
            aOutput[2*outputIndex + 1] = aInputData1[2*inputIndex + 1];
        }
        else
        {
            unsigned int inputIndex = idZ * aInputData2RowSize * aOutputColumnSize + idY * aInputData2RowSize + idX - aInputData1RowSize;
            unsigned int outputIndex =idZ * aOutputRowSize * aOutputColumnSize + idY * aOutputRowSize + idX;
            aOutput[2*outputIndex] = aInputData2[2*inputIndex];
            aOutput[2*outputIndex + 1] = aInputData2[2*inputIndex + 1];
        }
    }
    else
    {
        if(idX < aInputData1RowSize)
        {
            aOutput[idZ * aOutputRowSize * aOutputColumnSize + idY * aOutputRowSize + idX] = aInputData1[idZ * aInputData1RowSize * aOutputColumnSize + idY * aInputData1RowSize + idX];
        }
        else
        {
            aOutput[idZ * aOutputRowSize * aOutputColumnSize + idY * aOutputRowSize + idX] = aInputData2[idZ * aInputData2RowSize * aOutputColumnSize + idY * aInputData2RowSize + idX - aInputData1RowSize];
        }
    }

}

CudaMatrix Aurora::vertcat(const CudaMatrix& aMatrix1, const CudaMatrix& aMatrix2)
{
    if(aMatrix1.isNull() || aMatrix2.isNull() || aMatrix1.getDimSize(2) != aMatrix2.getDimSize(2) ||
       aMatrix1.getDimSize(1) != aMatrix2.getDimSize(1) || aMatrix1.getValueType() != aMatrix2.getValueType())
    {
        return CudaMatrix();
    }
    size_t outputRows = aMatrix1.getDimSize(0) + aMatrix2.getDimSize(0);
    size_t outputColumns = aMatrix1.getDimSize(1);
    size_t outputSlices = aMatrix1.getDimSize(2);

    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * outputRows * outputColumns * outputSlices * aMatrix1.getValueType());

    dim3 blockSize(THREADS_PER_BLOCK_DIM3_X, THREADS_PER_BLOCK_DIM3_Y, THREADS_PER_BLOCK_DIM3_Z);
    dim3 gridSize((outputRows + THREADS_PER_BLOCK_DIM3_X - 1)/THREADS_PER_BLOCK_DIM3_X, 
                  (outputColumns + THREADS_PER_BLOCK_DIM3_Y - 1)/THREADS_PER_BLOCK_DIM3_Y,
                  (outputSlices + THREADS_PER_BLOCK_DIM3_Z - 1)/THREADS_PER_BLOCK_DIM3_Z);
    vertcatKernel<<<gridSize, blockSize>>>(aMatrix1.getData(), aMatrix1.getDimSize(0), aMatrix2.getData(), aMatrix2.getDimSize(0), 
                                           data, outputRows, outputColumns, outputSlices, aMatrix1.isComplex());
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data, outputRows, outputColumns, outputSlices, aMatrix1.getValueType());
}

__global__ void vecnorm1Kernel(float* aInputData, unsigned int aInputRowSize, float* aOutput, bool aIsComplex)
{
    __shared__ float sharedValue[THREADS_PER_BLOCK];
    sharedValue[threadIdx.x] = 0;
    if(aIsComplex)
    {
        for(unsigned int i=0; i<=aInputRowSize/blockDim.x; ++i)
        {
            unsigned int indexByRows =  i*blockDim.x + threadIdx.x;
            if(indexByRows < aInputRowSize)
            {
                unsigned int idx = blockIdx.x*aInputRowSize + indexByRows;
                sharedValue[threadIdx.x] += sqrt(aInputData[2*idx] * aInputData[2*idx] +  aInputData[2*idx+1] * aInputData[2*idx+1]);
            }
        }
    }
    else
    {
        for(unsigned int i=0; i<=aInputRowSize/blockDim.x; ++i)
        {
            unsigned int indexByRows =  i*blockDim.x + threadIdx.x;
            if(indexByRows < aInputRowSize)
            {
                sharedValue[threadIdx.x] += abs(aInputData[blockIdx.x*aInputRowSize + indexByRows]);
            }
        }
    }
    __syncthreads();
    for(unsigned int i = blockDim.x/2; i>0; i >>= 1)
    {
        if(threadIdx.x < i)
        {
            sharedValue[threadIdx.x] += sharedValue[threadIdx.x + i];
        }
        __syncthreads();
    }
    aOutput[blockIdx.x] = sharedValue[0];   
}

__global__ void vecnorm2Kernel(float* aInputData, unsigned int aInputRowSize, float* aOutput, bool aIsComplex)
{
    __shared__ float sharedValue[THREADS_PER_BLOCK];
    sharedValue[threadIdx.x] = 0;

    for(unsigned int i=0; i<=aInputRowSize/blockDim.x; ++i)
    {
        unsigned int indexByRows =  i*blockDim.x + threadIdx.x;
        if(indexByRows < aInputRowSize)
        {
            if(aIsComplex)
            {
                unsigned int idx = blockIdx.x*aInputRowSize + indexByRows;
                sharedValue[threadIdx.x] += aInputData[2 * idx] * aInputData[2 * idx];
                sharedValue[threadIdx.x] += aInputData[2 * idx + 1] * aInputData[2 * idx + 1];
            }
            else
            {
                sharedValue[threadIdx.x] += aInputData[blockIdx.x*aInputRowSize + indexByRows] * aInputData[blockIdx.x*aInputRowSize + indexByRows];
            }
        }
    }    
    __syncthreads();
    for(unsigned int i = blockDim.x/2; i>0; i >>= 1)
    {
        if(threadIdx.x < i)
        {
            sharedValue[threadIdx.x] += sharedValue[threadIdx.x + i];
        }
        __syncthreads();
    }
    aOutput[blockIdx.x] = sqrt(sharedValue[0]); 
}

CudaMatrix Aurora::vecnorm(const CudaMatrix& aMatrix, NormMethod aNormMethod, int aDim)
{
    //only surpport aDim = 1 for now.
    if(aDim != 1 || aNormMethod == NormMethod::NormF || aMatrix.isNull())
    {
        return CudaMatrix();
    }
    unsigned int column = aMatrix.getDimSize(1);
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * column);
    if(aNormMethod == Aurora::Norm1)
    {
        vecnorm1Kernel<<<column, THREADS_PER_BLOCK>>>(aMatrix.getData(), aMatrix.getDimSize(0), data, aMatrix.isComplex());
    }
    else if(aNormMethod == Aurora::Norm2)
    {
        vecnorm2Kernel<<<column, THREADS_PER_BLOCK>>>(aMatrix.getData(), aMatrix.getDimSize(0), data, aMatrix.isComplex());
    }
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data,column);
}

__global__ void linspaceKernel(float* aOutput, unsigned int aOutputSize, float aStartNum, float aStepNum)
{
    unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < aOutputSize)
    {
        aOutput[idx] = aStartNum + idx * aStepNum;
    }
}

CudaMatrix Aurora::linspaceCuda(float aStart, float aEnd, int aNum)
{
    float step = (aEnd - aStart) / (aNum - 1);
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * aNum);
    int blocksPerGrid = (aNum + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    linspaceKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(data, aNum, aStart, step);
    cudaDeviceSynchronize();
    return Aurora::CudaMatrix::fromRawData(data,aNum);
}

CudaMatrix Aurora::auroraUnion(const CudaMatrix& aMatrix1, const CudaMatrix& aMatrix2)
{
    if(aMatrix1.isNull() || aMatrix2.isNull() || aMatrix1.isComplex() || aMatrix2.isComplex())
    {
        std::cerr<<"auroraUnion not support complex cudamatrix"<<std::endl;
        return CudaMatrix();
    }

    size_t size1= aMatrix1.getDataSize();
    size_t size2= aMatrix2.getDataSize();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * (size1 + size2));
    cudaMemcpy(data, aMatrix1.getData(), sizeof(float) * size1, cudaMemcpyDeviceToDevice);
    cudaMemcpy(data + size1, aMatrix2.getData(), sizeof(float) * size2, cudaMemcpyDeviceToDevice);
    thrust::sort(thrust::device, data, data+size1+size2);
    float* endPointer = thrust::unique(thrust::device, data, data+size1+size2);

    return CudaMatrix::fromRawData(data, endPointer - data);    
}

CudaMatrix Aurora::intersect(const CudaMatrix& aMatrix1, const CudaMatrix& aMatrix2)
{
    if(aMatrix1.isNull() || aMatrix2.isNull() || aMatrix1.isComplex() || aMatrix2.isComplex())
    {
        std::cerr<<"intersect not support complex cudamatrix"<<std::endl;
        return CudaMatrix();
    }

    size_t size1= aMatrix1.getDataSize();
    size_t size2= aMatrix2.getDataSize();
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * (size1 + size2));
    cudaMemcpy(data, aMatrix1.getData(), sizeof(float) * size1, cudaMemcpyDeviceToDevice);
    cudaMemcpy(data + size1, aMatrix2.getData(), sizeof(float) * size2, cudaMemcpyDeviceToDevice);
    thrust::sort(thrust::device, data, data+size1);
    thrust::sort(thrust::device, data+size1, data+size1+size2);
    float* end = thrust::set_intersection(thrust::device, data, data+size1,data+size1, data+size1+size2,data);

    return CudaMatrix::fromRawData(data, end - data);
}

__global__ void intersectKernel(float* aMatrixData, float* aIntersectData, unsigned int aMatrixDataSize, float* aOutputData, unsigned int aOutputDataSize)
{
    unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < aOutputDataSize)
    {
        for(unsigned int i=0; i<aMatrixDataSize; ++i)
        {
            if(aMatrixData[i] == aIntersectData[idx])
            {
                aOutputData[idx] = i+1;
                return;
            }
        }
    }
}

CudaMatrix Aurora::intersect(const CudaMatrix& aMatrix1, const CudaMatrix& aMatrix2, CudaMatrix& aIa)
{
    if(aMatrix1.isNull() || aMatrix2.isNull() || aMatrix1.isComplex() || aMatrix2.isComplex())
    {
        std::cerr<<"intersect not support complex cudamatrix"<<std::endl;
        return CudaMatrix();
    }
    CudaMatrix result = intersect(aMatrix1,aMatrix2);

    size_t size = result.getDataSize();
    float* iaResult = nullptr;
    cudaMalloc((void**)&iaResult, sizeof(float) * size);
    int blocksPerGrid = (size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    intersectKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(aMatrix1.getData(), result.getData(), aMatrix1.getDataSize(), iaResult, size);
    cudaDeviceSynchronize();

    aIa = CudaMatrix::fromRawData(iaResult,size);
    return result;
}

CudaMatrix Aurora::reshape(const CudaMatrix& aMatrix, int aRows, int aColumns, int aSlices)
{
    if(aMatrix.isNull() || (aMatrix.getDataSize() != aRows * aColumns * aSlices))
    {
        std::cerr<<"reshape diffirent size with cudamatrix"<<std::endl;
        return CudaMatrix();
    }
    return CudaMatrix::copyFromRawData(aMatrix.getData(),aRows,aColumns,aSlices);
}

__global__ void xcorrKernel(float* aInputData1, float* aInputData2,unsigned int aInputSize, float* aOutput, unsigned int aOutputSize)
{
    unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < aOutputSize/2 + 1)
    {
        for(unsigned int i=0; i<=idx; ++i)
        {
            aOutput[idx] += aInputData1[i] * aInputData2[aInputSize - idx - 1 + i];
        }
        return;
    }

    if (idx < aOutputSize)
    {
        for(int i=0; i<idx-aOutputSize/2; ++i)
        {
            aOutput[aOutputSize - idx + aOutputSize/2] += aInputData1[aInputSize + i - idx + aOutputSize/2] * aInputData2[i];
        }
        return;
    }
}

CudaMatrix Aurora::xcorr(const CudaMatrix& aMatrix1, const CudaMatrix& aMatrix2)
{
    if (aMatrix1.isNull() || aMatrix2.isNull() || aMatrix1.getDataSize() != aMatrix2.getDataSize() || aMatrix1.isComplex() || aMatrix2.isComplex())
    {
        std::cerr<<"xcorr not surpport with diffirent input size or complex cudamatrix"<<std::endl;
        return CudaMatrix();
    }

    size_t size = aMatrix1.getDataSize() * 2 - 1;
    float* data = nullptr;
    cudaMalloc((void**)&data, sizeof(float) * size);
    cudaMemset(data, 0.0, size);
    int blocksPerGrid = (size + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK;
    xcorrKernel<<<blocksPerGrid, THREADS_PER_BLOCK>>>(aMatrix1.getData(), aMatrix2.getData(), aMatrix1.getDataSize(), data, size);
    cudaDeviceSynchronize();

    return CudaMatrix::fromRawData(data, size);
}