Sets sparse data in tensor_impl_t. This method must be called every time changes are made to the underlying storage objects.

* @brief Size of dimension of operator * * @param dim Dimension to get size of * @return Size of dimension */

* @brief Size of dimension of operator * * @param dim Dimension to get size of * @return Size of dimension */

* Get the size of a single dimension of the tensor * * @param dim * Desired dimension * * @returns Number of elements in di

Calculate performance relative to float (single-precision). float = 1.0x (baseline) Higher values mean slower (took more time relative to

/////////////////////////////////////////////////////////////////////////// \brief fucntion wrapping NVTX management for automatic creation/deletion

Find the matx_bench executable.

* @brief Constexpr conversion from float to BF16 bits * * @param f Input float value * @return uint16_t BF16 bit representation (top 16 bits of fl

* @brief Constexpr conversion from float to FP16 bits * * @param f Input float value * @return uint16_t FP16 bit representation */

fltflt_sqrt() is the df64_sqrt() function given by Thall, which he attributes to Karp. This function implements Algorithm 7 from Thall's paper. It use

* Generic caching object for caching parameters. This class is used for * creating handles/plans on-the-fly and caching them to remove the need for

/////////////////////////////////////////////////////////////////////////// \brief Utility Function to get a unique ID for an NVTX range ///////////

Get the full path to jit_includes.h