Extract upper triangular matrix from the result of “magma_dgeqrf2_gpu” (which is a general matrix) without transferring to host

Question

I'm building a kernel which among other things uses the Magma function magma_dgeqrf2_gpu to perform a QR factorization. This outputs the upper triangular matrix R into a general matrix d_A on the GPU device.

Without transferring d_A back to host (since I need the GPU for further operations), is there a lib way to just reduce or extract the d_A into an upper triangular matrix R on the device?

Answer 1

这有点愚蠢，但我发现解决方案只是使用magmablas_dlacopy（）并设置属性以将上三角矩阵复制到设备上的另一个矩阵（已设置为0）。