这 以库为先的工程原则 代表了一种从手动内核开发向系统架构方法的范式转变。在 ROCm 生态中，这一理念要求工程师将资源集中在应用层逻辑上，而将设备特定的调优工作交由专门的 AMD 库来完成。
1. 哲学层面的转变
一位成熟的 GPU 工程师不会问： “我能写出这个内核吗？” 而是会问： “我应该写这个内核吗？” 自定义内核常常成为技术债务；像 rocBLAS 或 rocFFT 这样的库代表了数千小时的汇编级优化，单个开发者几乎无法达到同等水平。
2. 积极使用库
通过选择 积极使用库你就能确保应用程序获得‘免费’的性能提升。当 AMD 发布新架构（例如 CDNA 3）时，库的更新即可带来即时优化，无需修改一行主机代码。
main.py
TERMINALbash — 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.
✅ Correct!
Defaulting to libraries ensures you benefit from vendor-tuned performance and reduces technical debt.❌ Incorrect
Writing custom kernels by default is considered inefficient in a 'Library-First' philosophy.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.
✅ Correct!
Custom kernels require manual maintenance for every new GPU generation, whereas libraries handle this abstraction for you.❌ Incorrect
The principle views custom code as a maintenance burden unless it provides a unique competitive advantage.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.
✅ Correct!
AMD tunes the libraries for specific silicon; updating the library package often boosts performance without source code changes.❌ Incorrect
One of the greatest strengths of libraries is hardware abstraction.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?"
✅ Correct!
A mature engineer prioritizes efficiency, maintenance, and performance over the pride of writing custom code.❌ Incorrect
Just because you 'can' write something doesn't mean it is the best use of project resources.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
✅ Correct!
Many stencil operations can be mapped to frequency domain transforms or sparse matrix operations already optimized in these libraries.❌ Incorrect
SMI is for management; hipInfo doesn't exist; rocAL is for augmentation. rocSPARSE and rocFFT are the compute engines.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.
Q
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.
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.
Q
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.
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.