High-Level GPU Programming with Julia - 鄭景文

請從這裡開始

slides

What is GPU

How CPU DRAW

How GPU DRAW

GPU Programming Concept

• Utilize the parallel structure
• LOts of threads running at the same time

Work flow

1. GPU Memory allocation
2. copy data to device
3. call kernel <- I'm parallel
4. copy data back to host
5. free GPU memory

How GPU PROGRAMMING LOOKS LIKE IN OTHER LANGUAGES

Need to write CUDA-C in almost any language

Why JULIA

• HIgh level programming language with low-level performance
  • no need to worry about the non-gpu part is too slow
• Provide a first calss array implementation
  • No more reimplementation on array / tensor type
• Good compiler design
  • Allow us to compile native julia code??

JULIA OVER GPU

Kernel

• Code that run on GPU
• Execute on mutiple GPU threads

Writing A KERNEL

• pick a function you want to parallelize
• Spilt the workload into several subset
• …

THREADS & BLOCKS

GPU 執行有多個 block
每一 block 中有多個 thread
ThreadId : block 中的 順序？

• hardware dependent numbers of threads & blocks
• each threads / blocks ca be organized as 1/2/3 dimension grid
  • e.g. : 32 threads can be (4,8) or (4,4,2)

GPU PROGRAMMING IN JULIA

Packages

• CUDAnative.jl
• CUDAAdrv.jl
• CuArrays.jl
• GPUArrays.jl
• see github.com/JuliaGPU

GPU HELLO World

using CUDAnative

#hello world kernel 
function greeting ()
    @cuprint("This is block %ld , thread %d speaking , Hello /n")

function greeting2d()

helloGPU(3,5)

HIGH LEVEL GPU PROGRAMMING

• With CuArray.jl
  • cu()

GPU UNAWARE CODE ON GPU

GPU KERNEL EXAMPLE

MATMUL

using CuArrays

function matmul_kernel(C ,A ,B)
    i  = threadIdx().x + (blockIdx().x -1 )*blockDim().x
    j  = threadIdx().y + (blockIdx().x -y )*blockDim().y

CONV2D

GATHER

TOOLS

PROFILING WITH NVPROF

CUDAdrv.@profile func()

Info

• API call
• time : Max/Min/Avg

WHAT GET COMPILED

@device_code_typed @cuda
@device_code_llvm
@device_code_sass
@device_code_ptx @cuda

tags: COSCUP2019 Julia language IB501