Vector Dot Product Workload

Table of Contents

• Operation Description
  • Summary
  • Operation Parameters
  • Result:
  • Details
• Workloads
  • Workload Parameters
• Categories
  • Category Parameters
• Data Type
• Data Layout
  • 1. Vectors
  • 2. Scalars
• Dataset

Operation Description

Summary

This operation is defined as:

c = op(A, B)

where

op(A, B) = A . B

and operation . is the standard vector dot product.

Operation Parameters

Input: 2 parameters

Parameter	Description
0	A is a vector with n components.
1	B is a vector with n components.

Result:

Output: 1 output

Output	Description
0	c is a scalar. It is the result of dot product between A and B.

Details

If Z[i] denotes the element at component i in vector Z, then, the standard vector dot product operation is defined as:

c = A[0] * B[0] + A[1] * B[1] + ... + A[n - 1] * B[n - 1]

Workloads

This document applies to the following workloads:

// From hebench/api_bridge/types.h
 
hebench:APIBridge::Workload::DotProduct

Workload Parameters

Required workload parameters: 1

Index	Name	Type	Description
0	n	uint64_t	Number of components in a vector.

Above parameters are required for the workload in the specified order. A backend must specify, at least, a set of default arguments for these parameters.

Backends can require extra parameters beyond the base requirements. If a backend requires extra parameters, these must have default values in every set of default arguments for the workload parameters.

Categories

This workload supports the following categories:

hebench:APIBridge::Category::Latency
hebench:APIBridge::Category::Offline

Category Parameters

Latency

See Latency Category .

Offline

See Offline Category .

Value ranges for elements in CategoryParams::offline::data_count. Default value is used when the backend implementation sets the data_count for the corresponding operand to 0, but user specified 0 or no value at run-time.

Parameter	Lower bound	Upper bound	Default
0	1	none	5
1	1	none	5

Data Type

This workload is defined for the following data types:

hebench::APIBridge::DataType::Int32
hebench::APIBridge::DataType::Int64
hebench::APIBridge::DataType::Float32
hebench::APIBridge::DataType::Float64

Data Layout

1. Vectors

All scalar elements in a vector with elements of type T (where T is any of the supported types) lie contiguous in memory.

For example, given the following vector A with n = 3 components:

A = [ a0, a1, a2 ]

The elements will be stored in memory as:

Offset:	0	1	2
A	a0	a1	a2

Backends should expect this layout for their raw, clear text inputs, and must generate this layout for their decoded outputs.

Notes

If several vectors will be pointed at by a single pointer, consecutive vectors will follow each other in memory.

2. Scalars

All scalars will be represented as vectors with a single element.

Dataset

Supported modes:

Generate	External
yes	yes

Data generation for vectors used as input for this workload occurs during workload initialization by Test Harness. Ground truths are pre-computed during data generation. There is no standard dataset.

During data generation, all vector elements are extracted from a pseudo-random normal distribution of mean 0 and standard deviation of 10: n(0, 10).