コード例 #1
0
ファイル: BLAS3.cs プロジェクト: rblenis/cudafy
        public void Test_BLAS3_TRSM()
        {
            // Solve AX = B
            double maxError;

            // Lower triangular, Side left, No transpose
            ClearBuffer(hiMatrixAMM);
            ClearBuffer(hiMatrixBMN);

            CreateMainDiagonalOnlyMatrix(hiMatrixAMM, M);
            FillBuffer(hiMatrixBMN);

            diMatrixA = _gpu.CopyToDevice(hiMatrixAMM);
            diMatrixB = _gpu.CopyToDevice(hiMatrixBMN);

            _blas.TRSM(M, N, Alpha, diMatrixA, diMatrixB);

            _gpu.CopyFromDevice(diMatrixB, gpuResultMN);

            // Check AX - B
            maxError = 0.0;

            for (int i = 0; i < M; i++)
            {
                for (int j = 0; j < N; j++)
                {
                    double cpuResult = 0.0;

                    for (int k = 0; k < M; k++)
                    {
                        cpuResult += hiMatrixAMM[GetIndexColumnMajor(i, k, M)] * gpuResultMN[GetIndexColumnMajor(k, j, M)];
                    }

                    double error = Math.Abs(cpuResult - Alpha * hiMatrixBMN[GetIndexColumnMajor(i, j, M)]);

                    if (maxError < error)
                    {
                        maxError = error;
                    }
                }
            }
            Console.WriteLine("Max error : {0} (Lower fill mode, Side left, No transpose)", maxError);

            _gpu.FreeAll();

            // Lower triangular, Side left, Transpose
            ClearBuffer(hiMatrixAMM);
            ClearBuffer(hiMatrixBMN);

            CreateMainDiagonalOnlyMatrix(hiMatrixAMM, M);
            FillBuffer(hiMatrixBMN);

            diMatrixA = _gpu.CopyToDevice(hiMatrixAMM);
            diMatrixB = _gpu.CopyToDevice(hiMatrixBMN);

            _blas.TRSM(M, N, Alpha, diMatrixA, diMatrixB, cublasSideMode.Left, cublasOperation.T, cublasFillMode.Lower);

            _gpu.CopyFromDevice(diMatrixB, gpuResultMN);

            maxError = 0.0;

            for (int i = 0; i < M; i++)
            {
                for (int j = 0; j < N; j++)
                {
                    double cpuResult = 0.0;

                    for (int k = 0; k < M; k++)
                    {
                        cpuResult += hiMatrixAMM[GetIndexColumnMajor(k, i, M)] * gpuResultMN[GetIndexColumnMajor(k, j, M)];
                    }

                    double error = Math.Abs(cpuResult - Alpha * hiMatrixBMN[GetIndexColumnMajor(i, j, M)]);

                    if (maxError < error)
                    {
                        maxError = error;
                    }
                }
            }
            Console.WriteLine("Max error : {0} (Lower fill mode, Side left, Transpose)", maxError);

            _gpu.FreeAll();

            // Lower triangular, Side right, No transpose
            ClearBuffer(hiMatrixANN);
            ClearBuffer(hiMatrixBMN);

            CreateMainDiagonalOnlyMatrix(hiMatrixANN, N);
            FillBuffer(hiMatrixBMN);

            diMatrixA = _gpu.CopyToDevice(hiMatrixANN);
            diMatrixB = _gpu.CopyToDevice(hiMatrixBMN);

            _blas.TRSM(M, N, Alpha, diMatrixA, diMatrixB, cublasSideMode.Right);

            _gpu.CopyFromDevice(diMatrixB, gpuResultMN);

            // Check AX - B
            maxError = 0.0;

            for (int i = 0; i < M; i++)
            {
                for (int j = 0; j < N; j++)
                {
                    double cpuResult = 0.0;

                    for (int k = 0; k < N; k++)
                    {
                        cpuResult += gpuResultMN[GetIndexColumnMajor(i, k, M)] * hiMatrixANN[GetIndexColumnMajor(k, j, N)];
                    }

                    double error = Math.Abs(cpuResult - Alpha * hiMatrixBMN[GetIndexColumnMajor(i, j, M)]);

                    if (maxError < error)
                    {
                        maxError = error;
                    }
                }
            }
            Console.WriteLine("Max error : {0} (Lower fill mode, Side right, No transpose)", maxError);

            _gpu.FreeAll();

            // Lower triangular, Side right, Transpose
            ClearBuffer(hiMatrixANN);
            ClearBuffer(hiMatrixBMN);

            CreateMainDiagonalOnlyMatrix(hiMatrixANN, N);
            FillBuffer(hiMatrixBMN);

            diMatrixA = _gpu.CopyToDevice(hiMatrixANN);
            diMatrixB = _gpu.CopyToDevice(hiMatrixBMN);

            _blas.TRSM(M, N, Alpha, diMatrixA, diMatrixB, cublasSideMode.Right, cublasOperation.T);

            _gpu.CopyFromDevice(diMatrixB, gpuResultMN);

            // Check AX - B
            maxError = 0.0;

            for (int i = 0; i < M; i++)
            {
                for (int j = 0; j < N; j++)
                {
                    double cpuResult = 0.0;

                    for (int k = 0; k < N; k++)
                    {
                        cpuResult += gpuResultMN[GetIndexColumnMajor(i, k, M)] * hiMatrixANN[GetIndexColumnMajor(j, k, N)];
                    }

                    double error = Math.Abs(cpuResult - Alpha * hiMatrixBMN[GetIndexColumnMajor(i, j, M)]);

                    if (maxError < error)
                    {
                        maxError = error;
                    }
                }
            }
            Console.WriteLine("Max error : {0} (Lower fill mode, Side right, Transpose)", maxError);

            _gpu.FreeAll();

            // Upper triangular, Side left, No transpose
            ClearBuffer(hiMatrixAMM);
            ClearBuffer(hiMatrixBMN);

            CreateMainDiagonalOnlyMatrix(hiMatrixAMM, M);
            FillBuffer(hiMatrixBMN);

            diMatrixA = _gpu.CopyToDevice(hiMatrixAMM);
            diMatrixB = _gpu.CopyToDevice(hiMatrixBMN);

            _blas.TRSM(M, N, Alpha, diMatrixA, diMatrixB, cublasSideMode.Left, cublasOperation.N, cublasFillMode.Upper);

            _gpu.CopyFromDevice(diMatrixB, gpuResultMN);

            // Check AX - B
            maxError = 0.0;

            for (int i = 0; i < M; i++)
            {
                for (int j = 0; j < N; j++)
                {
                    double cpuResult = 0.0;

                    for (int k = 0; k < M; k++)
                    {
                        cpuResult += hiMatrixAMM[GetIndexColumnMajor(i, k, M)] * gpuResultMN[GetIndexColumnMajor(k, j, M)];
                    }

                    double error = Math.Abs(cpuResult - Alpha * hiMatrixBMN[GetIndexColumnMajor(i, j, M)]);

                    if (maxError < error)
                    {
                        maxError = error;
                    }
                }
            }
            Console.WriteLine("Max error : {0} (Upper fill mode, Side left, No transpose)", maxError);

            _gpu.FreeAll();

            // Upper triangular, Side left, Transpose
            ClearBuffer(hiMatrixAMM);
            ClearBuffer(hiMatrixBMN);

            CreateMainDiagonalOnlyMatrix(hiMatrixAMM, M);
            FillBuffer(hiMatrixBMN);

            diMatrixA = _gpu.CopyToDevice(hiMatrixAMM);
            diMatrixB = _gpu.CopyToDevice(hiMatrixBMN);

            _blas.TRSM(M, N, Alpha, diMatrixA, diMatrixB, cublasSideMode.Left, cublasOperation.T, cublasFillMode.Upper);

            _gpu.CopyFromDevice(diMatrixB, gpuResultMN);

            maxError = 0.0;

            for (int i = 0; i < M; i++)
            {
                for (int j = 0; j < N; j++)
                {
                    double cpuResult = 0.0;

                    for (int k = 0; k < M; k++)
                    {
                        cpuResult += hiMatrixAMM[GetIndexColumnMajor(k, i, M)] * gpuResultMN[GetIndexColumnMajor(k, j, M)];
                    }

                    double error = Math.Abs(cpuResult - Alpha * hiMatrixBMN[GetIndexColumnMajor(i, j, M)]);

                    if (maxError < error)
                    {
                        maxError = error;
                    }
                }
            }
            Console.WriteLine("Max error : {0} (Upper fill mode, Side left, Transpose)", maxError);

            _gpu.FreeAll();

            // Upper triangular, Side right, No transpose
            ClearBuffer(hiMatrixANN);
            ClearBuffer(hiMatrixBMN);

            CreateMainDiagonalOnlyMatrix(hiMatrixANN, N);
            FillBuffer(hiMatrixBMN);

            diMatrixA = _gpu.CopyToDevice(hiMatrixANN);
            diMatrixB = _gpu.CopyToDevice(hiMatrixBMN);

            _blas.TRSM(M, N, Alpha, diMatrixA, diMatrixB, cublasSideMode.Right, cublasOperation.N, cublasFillMode.Upper);

            _gpu.CopyFromDevice(diMatrixB, gpuResultMN);

            // Check AX - B
            maxError = 0.0;

            for (int i = 0; i < M; i++)
            {
                for (int j = 0; j < N; j++)
                {
                    double cpuResult = 0.0;

                    for (int k = 0; k < N; k++)
                    {
                        cpuResult += gpuResultMN[GetIndexColumnMajor(i, k, M)] * hiMatrixANN[GetIndexColumnMajor(k, j, N)];
                    }

                    double error = Math.Abs(cpuResult - Alpha * hiMatrixBMN[GetIndexColumnMajor(i, j, M)]);

                    if (maxError < error)
                    {
                        maxError = error;
                    }
                }
            }
            Console.WriteLine("Max error : {0} (Upper fill mode, Side right, No transpose)", maxError);

            _gpu.FreeAll();

            // Upper triangular, Side right, Transpose
            ClearBuffer(hiMatrixANN);
            ClearBuffer(hiMatrixBMN);

            CreateMainDiagonalOnlyMatrix(hiMatrixANN, N);
            FillBuffer(hiMatrixBMN);

            diMatrixA = _gpu.CopyToDevice(hiMatrixANN);
            diMatrixB = _gpu.CopyToDevice(hiMatrixBMN);

            _blas.TRSM(M, N, Alpha, diMatrixA, diMatrixB, cublasSideMode.Right, cublasOperation.T, cublasFillMode.Upper);

            _gpu.CopyFromDevice(diMatrixB, gpuResultMN);

            // Check AX - B
            maxError = 0.0;

            for (int i = 0; i < M; i++)
            {
                for (int j = 0; j < N; j++)
                {
                    double cpuResult = 0.0;

                    for (int k = 0; k < N; k++)
                    {
                        cpuResult += gpuResultMN[GetIndexColumnMajor(i, k, M)] * hiMatrixANN[GetIndexColumnMajor(j, k, N)];
                    }

                    double error = Math.Abs(cpuResult - Alpha * hiMatrixBMN[GetIndexColumnMajor(i, j, M)]);

                    if (maxError < error)
                    {
                        maxError = error;
                    }
                }
            }
            Console.WriteLine("Max error : {0} (Upper fill mode, Side right, Transpose)", maxError);

            _gpu.FreeAll();
        }