GPU Matrix Addition

In [1]:

    

using CUDA


    

    




CUDA Driver Initialized

In [2]:

    

println("CUDA driver version: $(CUDA.DriverVersion)")


    

    




CUDA driver version: 7000

In [3]:

    

println("Devices")
list_devices()
println()


    

    




Devices
device[0]: GeForce GTX 970, capability 5.2, total mem = 4095 MB

In [4]:

    

dev = CuDevice(0)
# create context 
ctx = create_context(dev)


    

    




create context






    
Out[4]:





CuContext(Ptr{Void} @0x0000000008102ab0)

In [22]:

    

# device API functions
println("Device memory (GB): $(totalmem(dev)/1e9)")
println("Device capability: $(capability(dev))")


    

    




Device memory (GB): 4.2942464
Device capability: CuCapability(5,2)

In [5]:

    

# compile kernel
run(`cat kernels/vadd.cu`)
run(`nvcc -ptx kernels/vadd.cu`)


    

    




// filename: vadd.cu
// simple CUDA kernel to add 2 vectors

extern "C"
{
  __global__ void vadd(const float *a, const float *b, float *c)
  {
    int i= threadIdx.x + blockIdx.x * blockDim.x;
    c[i]=a[i]+b[i];
  }
}

In [6]:

    

println("load module from vadd.ptx")
md = CuModule("vadd.ptx")


    

    




load module from vadd.ptx






    
Out[6]:





CuModule(Ptr{Void} @0x000000000a4c5970)

In [7]:

    

println("get function vadd")
f = CuFunction(md, "vadd")


    

    




get function vadd






    
Out[7]:





CuFunction(Ptr{Void} @0x000000000b0adc70)

In [8]:

    

siz = (3, 4)
len = prod(siz)


    

    
Out[8]:





12

In [9]:

    

# load array a to GPU
a = round(rand(Float32, siz) * 100)
ga = CuArray(a)
# load array b to GPU
b = round(rand(Float32, siz) * 100)
gb = CuArray(b)
# create array c on GPU
gc = CuArray(Float32, siz)


    

    
Out[9]:





CuArray{Float32,2}(CuPtr(0x0000000503e40400),(3,4),12)

In [10]:

    

# launch kernel
launch(f, len, 1, (ga, gb, gc))


    

In [11]:

    

# fetch results from GPU
c = to_host(gc)


    

    
Out[11]:





3x4 Array{Float32,2}:
 107.0   91.0  57.0   31.0
  49.0  100.0  86.0  121.0
 120.0   88.0  98.0   68.0

In [12]:

    

# free GPU memory
free(ga)
free(gb)
free(gc)


    

    
Out[12]:





CuPtr(0x0000000000000000)

In [13]:

    

println("Results:")
println("a = \n$a")
println("b = \n$b")
println("c = \n$c")


    

    




Results:
a = 
Float32[7.0 47.0 3.0 3.0
        29.0 59.0 80.0 66.0
        47.0 6.0 62.0 13.0]
b = 
Float32[100.0 44.0 54.0 28.0
        20.0 41.0 6.0 55.0
        73.0 82.0 36.0 55.0]
c = 
Float32[107.0 91.0 57.0 31.0
        49.0 100.0 86.0 121.0
        120.0 88.0 98.0 68.0]

In [14]:

    

# unload module - this is sensitive, will crash kernel if unload twice
unload(md)
# destroy context
destroy(ctx)