def runTest(vlength = 128,loops = 1):
n2 = vlength*vlength
alfa = c_float(.5)
cublasInit()
h_X = (c_float*n2)()
h_Y = (c_float*n2)()
g_Y = (c_float*n2)()
vectorInit(h_X)
vectorInit(h_Y)
d_X = c_void_p()
d_Y = c_void_p()
cublasAlloc(n2, sizeof(c_float), byref(d_X))
cublasAlloc(n2, sizeof(c_float), byref(d_Y))
cublasSetVector(n2, sizeof(c_float), h_X, 1, d_X, 1)
cublasSetVector(n2, sizeof(c_float), h_Y, 1, d_Y, 1)
flops = (2.e-9*n2)*float(loops)
t0 = time()
for i in range(loops):
cublasSaxpy(n2, alfa, d_X, 1, d_Y, 1)
cudaThreadSynchronize()
t0 = time()-t0
print "Processing time: %.3g sec" % t0
print "Gigaflops GPU: %.2f (%d)" % (flops/t0,n2)
t1 = time()
for i in range(loops):
cpuSAXPY(alfa,h_X,h_Y)
t1 = time()-t1
print "\nProcessing time: %.3g sec" % t1
print "Gigaflops CPU: %.2f" % (flops/t1)
print "GPU vs. CPU : %.2f" % (t1/t0)
cublasGetVector(n2, sizeof(c_float), d_Y, 1, g_Y, 1)
err,mxe = checkError(h_Y,g_Y)
print "\nAvg and max rel error = %.2e %.2e" % (err,mxe)
cublasFree(d_X)
cublasFree(d_Y)
cublasShutdown()
def runTest(vlength = 128,loops = 1):
n2 = vlength*vlength
alfa = c_float(.5)
cublasInit()
h_X = (c_float*n2)()
h_Y = (c_float*n2)()
vectorInit(h_X)
vectorInit(h_Y)
d_X = c_void_p()
d_Y = c_void_p()
cublasAlloc(n2, sizeof(c_float), byref(d_X))
cublasAlloc(n2, sizeof(c_float), byref(d_Y))
cublasSetVector(n2, sizeof(c_float), h_X, 1, d_X, 1)
cublasSetVector(n2, sizeof(c_float), h_Y, 1, d_Y, 1)
flops = (2.e-9*n2)*float(loops)
s0 = 0.
t0 = time()
for i in range(loops):
s0 += cublasSdot(n2, d_X, 1, d_Y, 1)
cudaThreadSynchronize()
t0 = time()-t0
print "Processing time: %.3g sec" % t0
print "Gigaflops GPU: %.2f (%d)" % (flops/t0,n2)
s1 = 0.
t1 = time()
for i in range(loops):
s1 += cpuSDOT(h_X,h_Y)
t1 = time()-t1
print "\nProcessing time: %.3g sec" % t1
print "Gigaflops CPU: %.2f" % (flops/t1)
print "GPU vs. CPU : %.2f" % (t1/t0)
sx = max(1.e-7,max(abs(s0),abs(s1)))
err = abs(s1-s0)/sx
print "\nError = %.2e" % err
cublasFree(d_X)
cublasFree(d_Y)
cublasShutdown()
def main(N = 1024,L = 100):
M = N
K = N >> 1
N = N << 1
flops = (2.e-9*M*N)*float(K*L)
print "M = %d, N = %d, K = %d, L = %d; GFlops = %.1f\n" % (M,N,K,L,flops)
na,nb,nc,alfa,beta = M*K,K*N,M*N,1.,0.
t0 = time()
cublasInit()
h_A = (c_float*na)()
h_B = (c_float*nb)()
h_C = (c_float*nc)()
g_C = (c_float*nc)()
arrayInit(h_A,na)
arrayInit(h_B,nb)
arrayInit(h_C,nc)
d_A = c_void_p()
d_B = c_void_p()
d_C = c_void_p()
cublasAlloc(na, sizeof(c_float), byref(d_A))
cublasAlloc(nb, sizeof(c_float), byref(d_B))
cublasAlloc(nc, sizeof(c_float), byref(d_C))
cublasSetVector(na, sizeof(c_float), h_A, 1, d_A, 1)
cublasSetVector(nb, sizeof(c_float), h_B, 1, d_B, 1)
cublasSetVector(nc, sizeof(c_float), h_C, 1, d_C, 1)
tt = t0 = time()-t0
print "Overhead CUBLAS: %.3f sec\n" % t0
t0 = time()
for i in range(L):
cublasSgemm('n', 'n', M, N, K, alfa, d_A, M, d_B, K, beta, d_C, M)
cudaThreadSynchronize()
t0 = time()-t0
tt += t0
print "Processing time: %.3g (%.3g) sec" % (t0,tt)
print "Gigaflops GPU: %.2f (%.2f)" % (flops/t0,flops/tt)
t1 = time()
for i in range(L):
sgemm(h_C,h_A,h_B,M,N,K)
t1 = time()-t1
print "\nProcessing time: %.3g sec" % t1
print "Gigaflops CPU : %.2f" % (flops/t1)
print "Speedup GPU/CPU: %.2f" % (t1/t0)
cublasGetVector(nc, sizeof(c_float), d_C, 1, g_C, 1)
err,mxe = checkError(h_C,g_C,nc)
print "\nAvg and max rel error = %.2e %.2e" % (err,mxe)
cublasFree(d_A)
cublasFree(d_B)
cublasFree(d_C)
cublasShutdown()
