ROCm and HIP: A Detailed 10-Chapter Tutorial: The Library-First Engineering Principle

The Library-First Engineering Principle represents a paradigm shift from manual kernel development to a systems-architecture approach. In the ROCm ecosystem, this philosophy dictates that engineering resources should focus on application-level logic while delegating device-specific tuning to specialized AMD libraries.

1. The Philosophical Shift

A mature GPU engineer does not ask: “Can I write this kernel?” but rather: “Should I write this kernel?” Custom kernels often become technical debt; libraries like rocBLAS or rocFFT represent thousands of hours of assembly-level tuning that a single developer rarely matches.

2. Aggressive Library Usage

By choosing to use libraries aggressively, you ensure that your application inherits "free" performance gains. When AMD releases a new architecture (e.g., CDNA 3), library updates provide instant optimization without you modifying a single line of your host code.

TERMINAL bash — 80x24

> Ready. Click "Run" to execute.

QUESTION 1

What is the primary mandate of the Library-First Engineering Principle?

To write custom HIP kernels for every operation to ensure maximum control.

To default to existing ROCm libraries before attempting custom HIP implementations.

To prioritize CPU execution over GPU acceleration.

To minimize the use of AMD-native headers.

QUESTION 2

According to the lesson, how should custom kernels be treated in a production environment?

As the primary mode of operation.

As technical debt that must be justified by extreme edge cases.

As assets that increase the value of the codebase significantly.

As temporary placeholders for library functions.

QUESTION 3

What is a major benefit of using ROCm libraries when transitioning between GPU architectures (e.g., CDNA 2 to CDNA 3)?

The developer must rewrite the kernel in assembly.

The developer receives 'free' performance gains via library updates.

The developer must manually adjust thread block sizes.

Libraries prevent the use of newer hardware features.

QUESTION 4

Which question characterizes the maturity of a GPU engineer?

"How can I maximize my line count?"

"Can I write this kernel?"

"Should I write this kernel?"

"Is there a way to avoid using handles?"

QUESTION 5

Which ROCm library would a 'Library-First' team use to replace a 3D Stencil kernel if possible?

rocSPARSE or rocFFT

hipInfo

ROCm-SMI

rocAL

Architectural Migration Challenge

Applying Library-First Principles to Legacy Systems

You are tasked with migrating a seismic imaging application that contains multiple custom-written HIP kernels for Fourier transforms and vector additions. The code currently requires manual tuning every time the hardware is upgraded from Radeon Pro to Instinct GPUs.

Identify the primary step in the migration workflow regarding kernel and host code separation.

Solution:
The developer should split the kernel and host code into separate source files. This modularity allows for the incremental replacement of custom `__global__` functions with calls to optimized libraries like rocFFT or rocBLAS without disrupting the high-level application flow or memory management logic.

Why would a 'Library-First' approach be faster to implement for a team of developers?

Solution:
By mapping operations to libraries, the team achieves 95%+ of theoretical peak performance immediately. They avoid the weeks or months typically spent on micro-architectural tuning (tiling, occupancy, shared memory bank conflicts) which are already solved within the pre-built ROCm library binaries.