/// <summary>
/// 1GPUでの共役勾配法を生成する
/// </summary>
/// <param name="count">要素数</param>
/// <param name="maxNonZeroCount"></param>
/// <param name="_minIteration"></param>
/// <param name="_maxIteration"></param>
/// <param name="_allowableResidual"></param>
public ConjugateGradientSingleGpu(int count, int maxNonZeroCount, int _minIteration, int _maxIteration, double allowableResidual)
: base(count, maxNonZeroCount, _minIteration, _maxIteration, allowableResidual)
{
// cudaの使用準備
cublas = CreateBlas();
cusparse = CreateSparse();
matDescr = CreateMatDescr();
// 行列を初期化
vectorA = new VectorDouble(count * maxNonZeroCount);
vectorColumnIndeces = new VectorInt(count * maxNonZeroCount);
vectorRowOffsets = new VectorInt(count + 1);
// ベクトルを初期化
vectorX = new VectorDouble(count);
vectorB = new VectorDouble(count);
vectorAp = new VectorDouble(count);
vectorP = new VectorDouble(count);
vectorR = new VectorDouble(count);
}
/// <summary>
/// 複数GPUでの共役勾配法を生成する
/// </summary>
/// <param name="count">要素数</param>
/// <param name="maxNonZeroCount"></param>
/// <param name="_minIteration"></param>
/// <param name="_maxIteration"></param>
/// <param name="_allowableResidual"></param>
public ConjugateGradientParallelGpu(int count, int maxNonZeroCount, int _minIteration, int _maxIteration, double allowableResidual)
: base(count, maxNonZeroCount, _minIteration, _maxIteration, allowableResidual)
{
// デバイス数を取得
deviceCount = GetDeviceCount();
// デバイスが計算する要素の先頭位置を生成
offsetsForDevice = new int[deviceCount + 1];
offsetsForDevice[0] = 0;
for(int i = 1; i < deviceCount; i++)
{
offsetsForDevice[i] = offsetsForDevice[i - 1] + (int)Math.Floor((double)Count / deviceCount);
}
offsetsForDevice[deviceCount] = Count;
// 内積の結果を初期化
resultsDot = new double[deviceCount];
bufferHost = new double[Count];
minJ = new int[deviceCount];
maxJ = new int[deviceCount];
// 配列を初期化
cublas = new IntPtr[deviceCount];
cusparse = new IntPtr[deviceCount];
matDescr = new IntPtr[deviceCount];
vectorElements = new VectorDouble[deviceCount];
vectorColumnIndeces = new VectorInt[deviceCount];
vectorRowOffsets = new VectorInt[deviceCount];
vectorX = new VectorDouble[deviceCount];
vectorB = new VectorDouble[deviceCount];
vectorAp = new VectorDouble[deviceCount];
vectorP = new VectorDouble[deviceCount];
vectorR = new VectorDouble[deviceCount];
// 全デバイスで
Parallel.For(0, deviceCount, deviceID =>
{
SetDevice(deviceID);
var countForDevice = CountForDevice(deviceID);
// cudaの使用準備
cublas[deviceID] = CreateBlas();
cusparse[deviceID] = CreateSparse();
matDescr[deviceID] = CreateMatDescr();
// 行列を初期化
vectorElements[deviceID] = new VectorDouble(countForDevice * maxNonZeroCount);
vectorColumnIndeces[deviceID] = new VectorInt(countForDevice * maxNonZeroCount);
vectorRowOffsets[deviceID] = new VectorInt(countForDevice + 1);
// ベクトルを初期化
vectorX[deviceID] = new VectorDouble(countForDevice);
vectorB[deviceID] = new VectorDouble(countForDevice);
vectorAp[deviceID] = new VectorDouble(countForDevice);
vectorP[deviceID] = new VectorDouble(count);
vectorR[deviceID] = new VectorDouble(countForDevice);
});
}
