PIM API

Overview

PIMeval provides a single public header for PIM benchmark application development, locating at libpimeval/src/libpimeval.h. Details of data types and APIs of this header are described in this wiki page.

(more details to be added)

Data Structure

• PimObjId
  • An integer value represents an allocated PIM object.
• PimDeviceEnum
  • PIM architecture.
• PimAllocEnum
  • PIM allocation strategy.
  • PIM_ALLOC_AUTO: Recommended way which can automatically determine V or H data layouts based on functional simulation target.
• PimDataType
  • PIM object element data type required for allocating a PIM object.
• PimCopyEnum
  • Data copy strategy.
• PimDeviceProperties
  • PIM device properties.

APIs

PIM Device Management

• pimCreateDevice
  • Create a PIM device of a specific PIM device type to start simulation.
  • Use PIM_FUNCTIONAL for functional simulation, so that a PIM benchmark can work for multiple PIM simulation targets.
  • Use specific PIM device type for micro-ops level simulation.
• pimCreateDeviceFromConfig
  • Read PIMeval and device parameters from a config file.
  • Alternatively, config file can be specified through environment variables.
    • PIMEVAL_TARGET
    • PIMEVAL_CONFIG_PATH
    • PIMEVAL_CONFIG_SIM
• pimGetDeviceProperties
  • Get PIM device properties.
• pimDeleteDevice
  • Delete PIM device and end simulation.

PIM Data Object Management

• pimAlloc
  • Allocate a new PIM data object.
• pimAllocAssociated
  • Allocate a new PIM data object associated with an existing PIM object.
  • Associated PIM objects have same number of elements, and element indices are aligned during allocation.
• pimFree
  • Delete a PIM object and free the memory space.

Data Transfer

• pimCopyHostToDevice
  • Copy data from CPU memory space to PIM memory space.
• pimCopyDeviceToHost
  • Copy data from PIM memory space to CPU memory space.
• pimCopyDeviceToDevice
  • Copy data between two associated PIM objects.

PIM Computation

Arithmetic

• pimAdd - C[i] = A[i] + B[i]
• pimSub - C[i] = A[i] - B[i]
• pimMul - C[i] = A[i] * B[i]
• pimDiv - C[i] = A[i] / B[i]
• pimAbs - C[i] = abs(A[i])
• pimMin - C[i] = min(A[i], B[i])
• pimMax - C[i] = max(A[i], B[i])
• pimAddScalar - C[i] = A[i] + scalar
• pimSubScalar - C[i] = A[i] - scalar
• pimMulScalar - C[i] = A[i] * scalar
• pimDivScalar - C[i] = A[i] / scalar
• pimMinScalar - C[i] = min(A[i], scalar)
• pimMaxScalar - C[i] = max(A[i], scalar)

Bitwise

• pimAnd - C[i] = A[i] & B[i]
• pimOr - C[i] = A[i] | B[i]
• pimXor - C[i] = A[i] ^ B[i]
• pimXnor - C[i] = ~(A[i] ^ B[i])
• pimAndScalar - C[i] = A[i] & scalar
• pimOrScalar - C[i] = A[i] | scalar
• pimXorScalar - C[i] = A[i] ^ scalar
• pimXnorScalar - C[i] = ~(A[i] ^ scalar)
• pimShiftBitsRight - C[i] = A[i] >> shiftAmount
• pimShiftBitsLeft - C[i] = A[i] << shiftAmount
• pimPopCount - C[i] = popcount(A[i])

Relational

• pimGT - C[i] = A[i] > B[i]
• pimLT - C[i] = A[i] < B[i]
• pimEQ - C[i] = A[i] == B[i]
• pimGTScalar - C[i] = A[i] > scalar
• pimLTScalar - C[i] = A[i] < scalar
• pimEQScalar - C[i] = A[i] == scalar

Reduction sum

• pimRedSumInt - result = sum of all elements in A
• pimRedSumUInt - result = sum of all elements in A
• pimRedSumRangedInt - result = sum of elements of a range in A
• pimRedSumRangedUInt - result = sum of elements of a range in A

Broadcast

• pimBroadcastInt - C[i] = scalar
• pimBroadcastUInt - C[i] = scalar

Element-wise shift/rotate

• pimRotateElementsRight - In place, A[i] = A[i - 1], A.first = A.last
• pimRotateElementsLeft - In place, A[i] = A[i + 1], A.last = A.first
• pimShiftElementsRight - In place, A[i] = A[i - 1]
• pimShiftElementsLeft - In place, A[i] = A[i + 1]

Micro-ops Simulation APIs

To be added.