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 main(device, vlength=128, loops=1):
alfa = c_float(.5)
n2 = vlength ## Vector length
gpuSAXPY = device.functions["gpuSAXPY"]
h_X = (c_float * n2)()
h_Y = (c_float * n2)()
g_Y = (c_float * n2)()
fixedInit(h_X)
d_X = getMemory(h_X)
d_Y = getMemory(h_Y)
cuFuncSetBlockShape(gpuSAXPY, BLOCK_SIZE, 1, 1)
cuParamSetf(gpuSAXPY, 0, alfa)
cuParamSeti(gpuSAXPY, 4, d_X)
cuParamSeti(gpuSAXPY, 8, d_Y)
cuParamSeti(gpuSAXPY, 12, n2)
cuParamSetSize(gpuSAXPY, 16)
cuCtxSynchronize()
t0 = time()
for i in range(loops):
cuLaunchGrid(gpuSAXPY, GRID_SIZE, 1)
cuCtxSynchronize()
t0 = time() - t0
flops = (2.e-9 * n2) * float(loops)
cuMemcpyDtoH(g_Y, d_Y, n2 * S4)
cuCtxSynchronize()
cuMemFree(d_X)
cuMemFree(d_Y)
t1 = time()
for i in range(loops):
cpuSAXPY(alfa, h_X, h_Y)
t1 = time() - t1
print "%10d%6.2f%6.2f" % (vlength, flops / t1, flops / t0)
if checkErrorFlag:
err, mxe = checkError(h_Y, g_Y)
print "Avg and max rel error = %.2e %.2e" % (err, mxe)
def main(device, vlength=128, loops=1):
alfa = c_float(0.5)
n2 = vlength ## Vector length
gpuSAXPY = device.functions["gpuSAXPY"]
h_X = (c_float * n2)()
h_Y = (c_float * n2)()
g_Y = (c_float * n2)()
fixedInit(h_X)
d_X = getMemory(h_X)
d_Y = getMemory(h_Y)
cuFuncSetBlockShape(gpuSAXPY, BLOCK_SIZE, 1, 1)
cuParamSetf(gpuSAXPY, 0, alfa)
cuParamSeti(gpuSAXPY, 4, d_X)
cuParamSeti(gpuSAXPY, 8, d_Y)
cuParamSeti(gpuSAXPY, 12, n2)
cuParamSetSize(gpuSAXPY, 16)
cuCtxSynchronize()
t0 = time()
for i in range(loops):
cuLaunchGrid(gpuSAXPY, GRID_SIZE, 1)
cuCtxSynchronize()
t0 = time() - t0
flops = (2.0e-9 * n2) * float(loops)
cuMemcpyDtoH(g_Y, d_Y, n2 * S4)
cuCtxSynchronize()
cuMemFree(d_X)
cuMemFree(d_Y)
t1 = time()
for i in range(loops):
cpuSAXPY(alfa, h_X, h_Y)
t1 = time() - t1
print "%10d%6.2f%6.2f" % (vlength, flops / t1, flops / t0)
if checkErrorFlag:
err, mxe = checkError(h_Y, g_Y)
print "Avg and max rel error = %.2e %.2e" % (err, mxe)
def main(vlength = 128,loops = 1):
alfa = c_float(.5)
n2 = vlength ## Vector length
h_X = (c_float*n2)()
h_Y = (c_float*n2)()
g_Y = (c_float*n2)()
fixedInit(h_X)
d_X = getMemory(h_X)
d_Y = getMemory(h_Y)
blockDim = dim3(BLOCK_SIZE,1,1)
gridDim = dim3(GRID_SIZE,1,1)
t0 = time()
cudaThreadSynchronize()
for i in range(loops):
cudaConfigureCall(gridDim,blockDim,0,0)
gpuSAXPY(alfa,d_X,d_Y,n2)
cudaThreadSynchronize()
t0 = time()-t0
flops = (2.e-9*n2)*float(loops)
g_Y = (c_float*n2)()
cudaMemcpy(g_Y,d_Y,S4*n2,cudaMemcpyDeviceToHost)
cudaThreadSynchronize()
cudaFree(d_X)
cudaFree(d_Y)
cudaThreadExit()
t1 = time()
for i in range(loops):
cpuSAXPY(alfa,h_X,h_Y)
t1 = time()-t1
print "%10d%6.2f%6.2f" % (vlength,flops/t1,flops/t0)
if checkErrorFlag:
err,mxe = checkError(h_Y,g_Y)
print "Avg and max rel error = %.2e %.2e" % (err,mxe)
def main(vlength = 128,loops = 1):
alfa = c_float(.5)
n2 = vlength ## Vector length
h_X = (c_float*n2)()
h_Y = (c_float*n2)()
g_Y = (c_float*n2)()
fixedInit(h_X)
d_X = getMemory(h_X)
d_Y = getMemory(h_Y)
blockDim = dim3(BLOCK_SIZE,1,1)
gridDim = dim3(GRID_SIZE,1,1)
t0 = time()
cudaThreadSynchronize()
for i in range(loops):
cudaConfigureCall(gridDim,blockDim,0,0)
gpuSAXPY(alfa,d_X,d_Y,n2)
cudaThreadSynchronize()
t0 = time()-t0
flops = (2.e-9*n2)*float(loops)
g_Y = (c_float*n2)()
cudaMemcpy(g_Y,d_Y,S4*n2,cudaMemcpyDeviceToHost)
cudaThreadSynchronize()
cudaFree(d_X)
cudaFree(d_Y)
cudaThreadExit()
t1 = time()
for i in range(loops):
cpuSAXPY(alfa,h_X,h_Y)
t1 = time()-t1
print "%10d%6.2f%6.2f" % (vlength,flops/t1,flops/t0)
if checkErrorFlag:
err,mxe = checkError(h_Y,g_Y)
print "Avg and max rel error = %.2e %.2e" % (err,mxe)