nki.isa.tensor_tensor#

nki.isa.tensor_tensor(dst, data1, data2, op, engine=engine.unknown, name=None)[source]#

Perform an element-wise operation of input two tiles using Vector Engine or GpSimd Engine. The two tiles must have the same partition axis size and the same number of elements per partition.

The element-wise operator is specified using the op field and can be any binary operator supported by NKI (see Supported Math Operators for NKI ISA for details) that runs on the Vector Engine, or can be np.power/nl.power that runs on the GpSimd Engine. For bitvec operators, the input/output data types must be integer types and Vector Engine treats all input elements as bit patterns without any data type casting. For arithmetic operators, there is no restriction on the input/output data types, but the engine automatically casts input data types to float32 and performs the element-wise operation in float32 math. The float32 results are cast to the target data type specified in the dtype field before written into the output tile. If the dtype field is not specified, it is default to be the same as the data type of data1 or data2, whichever has the higher precision.

Since GpSimd Engine cannot access PSUM, the input or output tiles cannot be in PSUM if op is np.power/nl.power (see NeuronCore-v2 Compute Engines for details). Otherwise, the output tile can be in either SBUF or PSUM. However, the two input tiles, data1 and data2 cannot both reside in PSUM. The three legal cases are:

  1. Both data1 and data2 are in SBUF.

  2. data1 is in SBUF, while data2 is in PSUM.

  3. data1 is in PSUM, while data2 is in SBUF.

Note, if you need broadcasting capability in the free dimension for either input tile, you should consider using nki.isa.tensor_scalar API instead, which has better performance than nki.isa.tensor_tensor in general.

Parameters:

  • dst – an output tile of the element-wise operation

  • data1 – lhs input operand of the element-wise operation

  • data2 – rhs input operand of the element-wise operation

  • op – a binary math operator (see Supported Math Operators for NKI ISA for supported operators)

  • engine – (optional) the engine to use for the operation: nki.isa.vector_engine, nki.isa.gpsimd_engine or nki.isa.unknown_engine (default, let compiler select best engine based on the input tile shape).