jasperzhong
commented
2 years ago https://www.youtube.com/watch?v=4APkMJdiudU&list=PLC6u37oFvF40BAm7gwVP7uDdzmW83yHPe
CPU和GPU区别总结的挺精辟的.
- CPU: designed to minimize latency. 所以主要的silicon area用作advanced control logic和large cache.
- GPU: desigend to maximize throughput. 所以主要silicon area用作massive number of cores.
如何利用massive number of GPU cores.
- kernels: 运行在GPU上的函数
- threads: kernel以一堆并行的线程执行
概念:
- Host vs Device => heterogenous => where CUDA comes in (C with extensions)
- Host与Device通过PICe通信 (但PCIe通常很慢).
- CUDA threads以SIMD方式执行. NVIDIA把这个叫做SIMT (其实是一回事)...
线程的组织:
- thread -> one GPU core
- block
- grid -> entire GPU
grid/block有dimension，可以是1D/2D/3D.
https://www.youtube.com/watch?v=OSpy-HoR0ac&list=PLC6u37oFvF40BAm7gwVP7uDdzmW83yHPe&index=5
Memory model
这节讲的很好...excellent!
注意local memory是很慢的..比shared memory慢很多. 这是因为local memory是off-chip的.
图中绿色的部分是NVIDIA cores所在的地方，上面的memory是on-chip的. 蓝色部分是DRAM，是off-chip.
https://www.youtube.com/watch?v=PJCISyoGpug&list=PLC6u37oFvF40BAm7gwVP7uDdzmW83yHPe&index=6
讲了synchronization primitives.
- thread级别的是_syncthreads().
- kernel级别的是cudaDeviceSynchronize() - Blocks until the device has completed all preceding requested task.
https://github.com/NVIDIA/cuda-samples
简直是宝藏.