PyCUDA. An Introduction

Transcription

1 PyCUDA An Introduction

2 Scripting GPUs with PyCUDA

3 Why do Scripting for GPUs? GPUs are everything that scripting languages are not: Highly parallel Very architecture-sensitive Built for maximum FP/memory throughput Complement each other CPU is largely restricted to control tasks (~1000/sec) Scripting fast enough Python + CUDA = PyCUDA Python + OpenCL = PyOpenCL

4 Why Python Scripting Mature Large and active community Emphasises readability Written in widely-portable C Supports multiple programming paradigms Rich ecosystem of scientific computing related software

5 Traditional Program Workflow

6 Scripting: Interpreted, not Compiled

7 PyCUDA Workflow

8 How are HPC Applications Constructed? Traditional C, C++ or Fortran Libraries Alternative Scripting for setup or admin GPUs for inner loops Play to the strengths of each programming environment

9 PyCUDA Philosophy Provide complete access Automatically manage resources Provide abstractions Check for and report errors automatically Full documentation Integrate tightly with numpy

10 What is numpy? Numpy: package for large, multi-dimensional arrays Vectors, matrices, A+B, sin(a), dot(a,b) la.solve(a,b), la.eig(a) All much faster than functional equivalents in Python Python's Matlab: basis for scipy, plotting,...

11 CUDA C Example #include <stdio.h> global void twice(float *a) { int index = threadidx.x + threadidx.y*4; } a[index] *= 2; int main() { int width = 4; int N = width * width; int size = sizeof(float) * N; float* a = (float*) malloc(size); float* a_double = (float*) malloc(size); for(unsigned int i = 0; i < N; i++) a[i] = float(i);

12 CUDA C Example } float* d_a; cudamalloc((void**)&d_a, size); cudamemcpy(d_a, a, size, cudamemcpyhosttodevice); dim3 grid(1, 1, 1); dim3 block(width, width, 1); twice<<<grid, block>>>(d_a); cudamemcpy(a_double, d_a, size, cudamemcpydevicetohost); for(int i = 0; i < 16; i++) printf("%.2f ", a[i]); printf("\n"); for(int i = 0; i < 16; i++) printf("%.2f ", a_double[i]); printf("\n"); cudafree(d_a); free(a_double); free(a); return 0;

13 PyCUDA Example import pycuda.driver as cuda import pycuda.autoinit, pycuda.compiler import numpy a = numpy.random.randn(4,4).astype(numpy.float32) a_gpu = cuda.mem_alloc(a.nbytes) cuda.memcpy_htod(a_gpu, a) mod = pycuda.compiler.sourcemodule(""" global void twice(float *a) { int index = threadidx.x + threadidx.y*4; a[index] *= 2; } """) func = mod.get_function("twice") func(a_gpu, block=(4,4,1)) a_doubled = numpy.empty_like(a) cuda.memcpy_dtoh(a_doubled, a_gpu) print a_doubled print a

14 PyCUDA Vital Information Available for download Complete documentation MIT License (no warranty, free for all use) Requires: numpy, Python 2.4+, CUDA SDK Cross platform Mailing list for support

15 gpuarray: Simple Linear Algebra Meant to look and feel just like numpy gpuarray.to_gpu(numpy_array) numpy_array = gpuarray.get() +, -, *, /, fill, sin, exp, rand, basic indexing, norm, inner product, Mixed types (int32 + float32 = float64) print gpuarray for debugging Allows access to raw bits Use as kernel arguments, textures,...

16 gpuarray Example import pycuda.autoinit import pycuda.gpuarray as gpuarray import numpy a = numpy.random.randn(4,4).astype(numpy.float32) a_gpu = gpuarray.to_gpu(a) a_doubled = (2*a_gpu).get() print a_doubled print a_gpu

17 Constructing gpuarray Instances gpuarray.to_gpu(ary, ) gpuarray.empty(shape, dtype, ) gpuarray.zeros(shape, dtype, ) gpuarray.empty_like(other_ary) gpuarray.zeros_like(other_ary) gpuarray.arange(start,stop,step, )