Intro to Cuda
Download as PPTX, PDF0 likes186 views
This document discusses GPU algorithms and graphics processing units (GPUs). It provides an overview of using GPUs for general purpose tasks by copying data to the GPU memory, running executable kernels on the GPU, and copying results back to the CPU memory. It also discusses GPU memory management and optimizations like using shared memory and avoiding serialization. An example k-means clustering algorithm is mentioned to illustrate a potential GPU algorithm.
![int index = threadIdx.x;
c[index] =
a[index] +
b[index];](https://image.slidesharecdn.com/introtocuda-170113131256/85/Intro-to-Cuda-39-320.jpg)
Intro to Cuda
- 5. Why?
- 9. T E M S R
- 10. HOST
- 11. DEVICE
- 13. PCI Bus Copy initial data to DEVICE
- 15. PCI Bus Copy Results Back To HOST
- 16. Still Running on CPU
- 17. Still Running on CPU GPU is a Resource
- 21. int *a;
- 24. arr = malloc(size); cudaMalloc(&d_arr, size);
- 25. free(arr);
- 28. memcpy(dest, source, size); cudaMemcpy(&dest, src, size, …);
- 29. 1: HOST DEVICE 2: EXECUTE 3: DEVICE HOST
- 30. 1: HOST DEVICE 3: DEVICE HOST cudaMemcpy();
- 31. 1: HOST DEVICE cudaMemcpy( &dest, source, size, ..hostToDevice);
- 32. EXECUTION
- 33. __global__ void myKernel(int *a){ }
- 35. Let’s do an example
- 39. int index = threadIdx.x; c[index] = a[index] + b[index];
- 55. CODE
- 56. Shared Memory • ~48k • Multiple GB device memory (100x higher latency) • Access memory in order • 1 2 3 • 4 5 6 • 7 8 9
- 57. Considerations • Transistors are allocated to arithmetic, not memory. Sometimes better to recompute rather than cache • Copying to/from host takes a while. Sometimes sequential operations can stay on gpu • Avoid serialization (shared memory bank conflicts) • Asynchronous memory operations