Recipe for running simple CUDA code on a GPU based

Rocks cluster

GPU workshopUR June 20,

2008Alice Quillen

Head node: sneth.pas.rochester.edu

http://astro.pas.rochester.edu/~aquillen/gpuworkshop.html

Outline

• The Kernel .cu CUDA files

• Calling __global__ functions

• Setting grids and threads

• Allocating memory on device

• Compiling in emulate mode

• Submitting a job to the queue using SGE

• Timing your job for benchmarks

CUDA C, CU files

• Routines that call the device must be in plain C -- with extension .cu

• Often 2 files1) The kernel -- .cu file containing routines that are

running on the device.

2) A .cu file that calls the routines from kernel. #includes the kernel

• Optional additional .cpp or .c files with other routines can be linked

Kernel, Threads + Blocks

Kernel has two different types of functions__global__ called from host, executed on device__device__ called from device, run on device

__global__ void addone(float* d_array) // kernel is addone{

int index = blockIdx.x * blockDim.x + threadIdx.x;d_array[index] += 1.0;

}

d_array must be allocated on devicethe above code loops over all threads and blocks

automaticallyon a CPU this would be inside a loop

Calling a __global__ routinethreads and blocks

Calling the routine from the host to run on deviceint p = 256; // power of 2dim3 threads = (p,1,1); // number of threads per block is pdim3 grid = (arraylength/p,1,1); // number of blocks is arraylength divided by the

// number of threads// arraylength is assumed a multiple of p

int sharedMemsize = 0;addone <<< grid, threads, sharedMemsize >>> (d_array); // call the kernel

Inside the routine blockDim.x is p (number of threads), threadIdx.x ranges from 0 to p-1, (covers all threads)blockIdx.x ranges from 0 to arraylength/p-1 (covers all blocks)__global__ void addone(float * d_array){

int index = blockIdx.x * blockDim.x + threadIdx.x;d_array[index] += 1.0;

}

Memory allocation on deviceint lengthvector = 10240; // note multiple of 256, pad with zeros if your vector

// isn’t a multipleunsigned int memSize = lengtharray*sizeof(float);// allocate array on hostfloat* h_array = (float*) malloc(memSize); // array on host

float* d_array; // array on device// allocate array on deviceCUDA_SAFE_CALL(cudaMalloc((void**)&d_array, memSize )); // CUDA_SAFE_CALL utility Tool. If an error is found, it prints out and error // message, file name, and line number in file where the error can be found

// Copy array from host to device CUDA_SAFE_CALL(cudaMemcpy(d_array, h_array, memSize,

cudaMemcpyHostToDevice));// Copy array from device to hostCUDA_SAFE_CALL(cudaMemcpy(h_array, d_array, memSize,

cudaMemcpyDeviceToHost));

Global Device Memory

• Keep track of what data is on device and on host

• If you address host memory from device then the code doesn’t work

A code that adds 1 to every element of a large array on the GPU

#include<cutil.h> // CUDA defined stuff#include<stdio.h> // usual io #include<addone_kernel.cu> // contains kernelint main(int argc, char ** argv){

CUT_DEVICE_INIT(); // device init, also will tell you if multiple devices // present

Allocate host array, fill it with somethingAllocate device arrayCopy array from host to deviceLaunch kernelCopy array back from device to hostPrint out some of it out so you can see that the device did

somethingCUDA_SAFE_CALL(cudaFree(d_vector)); // free memory

CUT_EXIT(argc, argv); // Use CUDA Utility Tool to exit cleanly}

Compiling

Sample head of Makefile# name of executableEXECUTABLE := addone# Cuda source files (compiled with cudacc) don't put kernel

in as is included in other fileCUFILES := addone.cuCU_DEPS := addone_kernel.cu # C/C++ source files (compiled with gcc / c++)CCFILES :=C_DEPS := ROOTDIR = /usr/local/NVIDIA_CUDA_SDK/common/include $(ROOTDIR)/common.mk # SDK make

Emulator[aquillen@sneth ~]$ make emu=1Makes executable code that runs on the CPU not on deviceexecutable can be run on head node (which lacks GPUs) not necessary to send a job to queuein addone_kernel.cu:__global__ void addone(float* d_array){

int index = blockIdx.x * blockDim.x + threadIdx.x;d_array[index] += 1.0;printf(“%.2e\n”,d_array[index]); // emulator only!

}Will not run or compile for device. Will run and compile in emulator.Prints for every block and thread. If (threadIdx.x == 0) printf(“%.2e\n”,d_array[index]); // much more manageable!

Debugging

• Code that works in emulator mode may not work on device (particularly memory access, synch problems)

• However debugging in emulator mode is a pretty good way to get closer to working code

• Profiler recently available, but we have not yet had a chance to play with it.

Queue system

• SGE = Sun Grid Engine is an open source batch-queuing system supported by SUN and it comes as a roll with ROCKS

• Rocks is an open-source Linux cluster distribution that enables end users to easily build computational clusters. CUDA comes as a roll.

To submit a job

[aquillen@sneth ~]$ qsub sge.bash

where sge.bash is a text file that looks like …

Example sge.bash file#!/bin/bash remind me that this uses bash shell##$ -cwd use current directory#$ -o std.out stdout goes into this file#$ -e std.err stderr goes into this file#$ -S /bin/bash specify bash shell #$ -m be email at begin and end of job#$ -M alice.quillen@gmail.com where to send email#$ -N gal10 name of job#$ -l h_rt=24:00:00 let it run no more than 24 hours#$ -V something about exporting env variables

date +%s time in seconds for simple bench mark# execute some code comment/home/aquillen/galint/exe/release/mkgal.out executables/home/aquillen/galint/exe/release/galintdate print datedate +%secho “complete”

queue commands

• qstat look at what jobs are running

• qdel 701 kill that job• qhost check all the nodes• qstat –j 703 explain to me what is going on

with that job

Workshop http://astro.pas.rochester.edu/~aquillen/gpuworkshop/assigna.html

Assignment #1 Submitting a job to the queue using SGE qsub • Here is a sample sge (Grid Engine) file that is written in bash shell. sge.bash • Transfer this file to your home directory on our rocks cluster sneth.pas.rochester.edu. • Modify the file sge.bash so that it will print "hello, World" into a file called std.out and will sleep for 60 seconds or

so. • Submit it as a job to the queue (qsub sge.bash). • Check the queue status to make sure your job is run (qstat). Assignment #2 Compiling and running the routine addone that adds 1 to every element in a large array • Here is a sample kernel addone_kernel.cu

Here is a routine that calls this kernel addone.cu A Makefile to compile it Makefile A sge bash file to send it to the queue cudasge.bash

• Transfer the above 4 files to your home directory on sneth. They can also be copied directly from the directory ~aquillen/addone/ on sneth.

• Edit the cudasge.bash file so it is appropriate for you. • Compile the code with make. The executable will be in a subdirectory called release . Run it by submitting the

cudasge.bash file to the queue. • Compile the code in emulate model (make emu=1). The executable will be in a subdirectory called emurelease .

Run this code. It will run on the head node so you don't need to submit it to the queue. Assignment #3 Write your own routine by modifying the above addone routines. • Modify the kernel so that it computes the derivative of the array using a finite difference (f[i+1]-f[i]). • You will need an input and output arrays as arguments to your __global__ routine in the kernel as the order that

threads and blocks executed is not specified. • You do not want to address memory outside the array as this will crash the code. If you check the addressing with

an if statement your routine will take twice as long to run. Try to write the routine without an if statement. Hint: you can trade an irrelevant amount of device global memory for the if statement.

Assignment #4 Discuss how you could implement a box smooth or 1d convolution routine

BenchmarkingThere are many ways to time routines. Here is the one used in the SDK. I think it uses sys/math.h gettimeofday() in linux

// initialize a timerunsigned int timer = 0; CUT_SAFE_CALL(cutCreateTimer(&timer));// start itCUT_SAFE_CALL(cutStartTimer(timer));//// run your code here //// Stop the timerCUT_SAFE_CALL(cutStopTimer(timer));printf( "Processing time: %f (ms)\n", cutGetTimerValue(timer));// time in milliseconds, returned as a float// Delete the timerCUT_SAFE_CALL(cutDeleteTimer(timer));

Links

http://astro.pas.rochester.edu/~aquillen/gpuworkshop.htmlIncludes links to the CUDA manual, the SDK (software devel kit, lots of

good programming examples) and other on line tutorials