BatchSVD

Functions

NVCVStatus pvaBatchSVDCreate(NVCVOperatorHandle *handle, const NVCVTensorRequirements *srcParams)

Constructs an instance of the BatchSVD operator.

NVCVStatus pvaBatchSVDSubmit(NVCVOperatorHandle handle, cupvaStream_t stream, const NVCVTensorHandle src, const NVCVTensorHandle U, const NVCVTensorHandle S, const NVCVTensorHandle V, const int32_t maxIters, const float tol)

Executes the BatchSVD operation.

Functions

NVCVStatus pvaBatchSVDCreate(

NVCVOperatorHandle *handle,

const NVCVTensorRequirements *srcParams,

)

Constructs an instance of the BatchSVD operator.

Limitations:

Data Layout: [HWC]
    C: Number of matrices
    H: Matrix height
    W: Matrix width
    Due to VMEM capacity, max matrix width = 31
    The formula for memory constraint on matrix width and height is:
    width * height * PITCH * 2 + (width + 1) * BATCH) > 1024 * 128 (128KB one VMEM superbank)
    PITCH and BATCH are set to 16, as defined in batch_svd_params.h
Per matrix SVD: A = USV^T, economy SVD
    A and U size: H x W, H >= W >= 2
    S size: 1 x W, W >= 2, H = 1
    V size: W x W, W >= 2, H = W
    Please note that the S output is not in a diagonal matrix format.
    Instead, it is a vector containing W singular values.

Data Type      | Allowed
-------------- | -------------
8bit  Unsigned | No
8bit  Signed   | No
16bit Unsigned | No
16bit Signed   | No
32bit Unsigned | No
32bit Signed   | No
32bit Float    | Yes
64bit Float    | No

Parameters:

• handle – [out] Where the operator instance handle will be written to.

  • Must not be NULL.

• srcParams – [in] Pointer to the NVCVTensorRequirements structure which contains input matrices tensor rank, shape, layout and data type information.

Return values:

• NVCV_ERROR_INVALID_ARGUMENT – Handle is null.

• NVCV_ERROR_OUT_OF_MEMORY – Not enough memory to create the operator.

• NVCV_SUCCESS – Operation executed successfully.

NVCVStatus pvaBatchSVDSubmit(

NVCVOperatorHandle handle,

cupvaStream_t stream,

const NVCVTensorHandle src,

const NVCVTensorHandle U,

const NVCVTensorHandle S,

const NVCVTensorHandle V,

const int32_t maxIters,

const float tol,

)

Executes the BatchSVD operation.

Parameters:

• handle – [in] Handle to the operator.

  • Must not be NULL.

• stream – [in] Handle to a valid CUPVA stream.

• src – [in] Input matrix tensor handle. Data Type: NVCV_DATA_TYPE_F32 Data Layout: [HWC]

• U – [out] Output U matrix tensor handle. Data Type: NVCV_DATA_TYPE_F32 Data Layout: [HWC]

• S – [out] Output S matrix tensor handle. Data Type: NVCV_DATA_TYPE_F32 Data Layout: [HWC]

• V – [out] Output V matrix tensor handle. Data Type: NVCV_DATA_TYPE_F32 Data Layout: [HWC]

• maxIters – [in] The process of Jacobi rotations stops after maxIters sweeps.

• tol – [in] Converge criterion to check if the off-diagonal elements are sufficiently small compared to the diagonal elements. Note that if the value is less than 1e-7f, convergence may not occur until the maximum number of iterations is reached.

Return values:

• NVCV_ERROR_INVALID_ARGUMENT – Some parameter is outside the valid range.

• NVCV_ERROR_INTERNAL – Internal error in the operator, invalid types passed in.

• NVCV_SUCCESS – Operation executed successfully.