package tensorflow.profiler.op_profile

Get desktop application:
View/edit binary Protocol Buffers messages

Measurements of an operation (or aggregated set of operations). Metrics are always "total" rather than "self". Next ID: 21

Used in: Node

• double flops = 2

  Floating point computations performed by this operation, as a fraction of peak core FLOPS * program time. This representation has useful properties: - it is proportional to the number of floating point operations performed - utilization is flops/time - wasted potential flops is proportional to time - flops - it does not reveal the peak core FLOPS of the hardware - it is capped at 1.0

• double uncapped_flops = 19

  Uncapped FLOPs.

• repeated double bandwidth_utils = 5

  The memory bandwidth used to load operands, as a fraction of thereotical memory bandwidth on the specific hardware. Index into array using MemBwType enum.

• double raw_time = 11

  The raw stats below are aggregated across all occurrences.

  Elapsed core-time in picoseconds.

• double raw_flops = 12

  Total floating-point operations performed.

• double bf16_flops = 18

  Total FLOPs Normalized to the bf16 (default) devices peak bandwidth.

• double normalized_time_ps = 20

  Total time in picoseconds normalized.

• repeated double raw_bytes_accessed_array = 15

  Total bytes accessed for each memory type. Index into array using MemBwType enum.

• uint32 occurrences = 16

  Number of executions.

• double avg_time_ps = 17

  Average "accumlated" time in picoseconds that the operation took.

An entry in the profile tree. (An instruction, or set of instructions).

Used in: Profile

• string name = 1

  Semantics depend on contents.

• optional Metrics metrics = 2

  May be omitted e.g. for fused instructions.

• repeated Node children = 3

  Subjected to pruning.

• oneof contents

  Details about what this node represents.

  • Node.InstructionCategory category = 4

  • Node.XLAInstruction xla = 5

• int32 num_children = 6

  Total number of children before pruning.

A category of XLA instructions. name is a descriptive string, like "data formatting".

Used in: Node

(message has no fields)

A single XLA instruction. name is the unique instruction id, like "%multiply.5".

Used in: Node

• string op = 1

  Opcode like %multiply

• string expression = 2

  %multiply = [shape]multiply(operand1, operand2)

• string provenance = 3

  Provenance op name, eg. TF Op name, JAX Op name

• string category = 4

• optional XLAInstruction.LayoutAnalysis layout = 5

  Describes the physical memory layout of the instruction's primary input. e.g. for a convolution, this analyzes the image and ignores the kernel.

• uint32 computation_primitive_size = 6

• uint64 fingerprint = 7

• uint64 program_id = 8

• optional SourceInfo source_info = 9

• optional string xprof_kernel_metadata = 10

  Xprof kernel metadata extracted from the HLO expression's frontend attributes.

Used in: XLAInstruction

• repeated LayoutAnalysis.Dimension dimensions = 1

  The physical data layout, from most-minor to most-major dimensions.

Used in: LayoutAnalysis

• int32 size = 1

  Size of the data in this dimension.

• int32 alignment = 2

  Data must be padded to a multiple of alignment.

• string semantics = 3

  What the dimension represents, e.g. "spatial".

Profile is the top-level data that summarizes a program.

• optional Node by_category = 1

  Root of a profile broken down by instruction category.

• optional Node by_program = 4

  Root of a profile broken down by program.

• string device_type = 5

  Device type.

• double agg_dvfs_time_scale_multiplier = 10

  DVFS time scale multiplier weighted aggregated across all ops. Multiplied it with the raw time to get estimated time as if running on default pstate.

• optional Node by_category_exclude_idle = 6

  Exclude idle ops.

• optional Node by_program_exclude_idle = 7

• optional Node by_provenance = 8

  Root of a profile broken down by program and provenance.

• optional Node by_provenance_exclude_idle = 9