def __init__(self, numcoarsechannels, numfinechannels, accumulation_length, bandwidth, input_bitwidth, fft_coarse_out_bitwidth, antennas=1):
self.maxdesigns=0
self.blocks = {}
self.blockalgs = {}
self.totalblocks = 0
self.maxdesigns = 1
self.singleimplementation = 1
#add the platform array
self.platforms = {}
# add platforms: cost, inputbw, outputbw, resources
self.platforms['ROACH'] = Platform('ROACH',6700,40,40,['registers','luts','dsp','bram'])
self.platforms['GTX580'] = Platform('GTX580',3500,10,1,['time'])
for i in range(0,antennas):
# add the ADC
adc_bw = bandwidth*input_bitwidth
self.blocks['ADC'+`i`] = CBlock('ADC', CBlock.getADCModel(self.platforms, bandwidth, input_bitwidth),-1,0,0,'PFB'+`i`,0,adc_bw,1)
self.totalblocks += 1
# add the PFB
pfb_bw = bandwidth*32
self.blocks['PFB'+`i`] = CBlock('PFB',CBlock.getPFBWModel(self.platforms, bandwidth, pfb_bw, numcoarsechannels),'ADC'+`i`,0,adc_bw,'FFT_coarse'+`i`,0,adc_bw,1)
self.totalblocks += 1
# add the FFT
#print CBlock.getFFTModel(self.platforms, bandwidth, input_bitwidth, numchannels)
fft_coarse_out_bandwidth = bandwidth* fft_coarse_out_bitwidth*2
self.blocks['FFT_coarse'+`i`] = CBlock('FFT_coarse',CBlock.getFFTWModel(self.platforms, bandwidth, numcoarsechannels),'PFB'+`i`,0,pfb_bw,'Transpose'+`i`,0,fft_coarse_out_bandwidth,1)
self.totalblocks += 1
# adjust to ensure the block fits on the gpu
fft_fine_in_bandwidth = fft_coarse_out_bandwidth/numcoarsechannels
finemodel = CBlock.getFFTModel(self.platforms, fft_fine_in_bandwidth, numfinechannels)
if(finemodel['GTX580']['time']<0.1):
multiplier = pow(2,int(log(0.1/finemodel['GTX580']['time'],2)))
else:
multiplier = 1
finemodel['GTX580']['time'] = finemodel['GTX580']['time']*multiplier
fine_blocks = int(numcoarsechannels/multiplier)
fine_sky_bandwidth = bandwidth/fineblocks
fine_block_bandwidth = fft_coarse_out_bandwidth/fine_blocks
self.blocks['Transpose'+`i`] = CBlock('Transpose', CBlock.getTransposeModel(self.platforms, bandwidth, numcoarsechannels, numfinechannels), 'FFT_coarse'+`i`,0,fft_coarse_out_bandwidth,'FFT_fine'+`i`,1,fft_coarse_out_bandwidth,1)
self.totalblocks += 1
self.blocks['FFT_fine'+`i`] = CBlock('FFT_fine',finemodel,'Transpose'+`i`,0,fine_block_bandwidth,'VAcc'+`i`,0,fft_fine_in_bandwidth,fine_blocks)
self.totalblocks += fine_blocks
self.blocks['VAcc'+`i`] = CBlock('VAcc',{'ROACH': {'registers': 0.2, 'luts': 0.1, 'dsp': 0, 'bram':0.4}, 'GTX580': {'time': 0.001}},'FFT_fine'+`i`,0,fine_block_bandwidth,-1,0,0,fine_blocks)
self.totalblocks += fine_blocks
def __init__(self,
numcoarsechannels,
numfinechannels,
accumulation_length,
bandwidth,
input_bitwidth,
fft_coarse_out_bitwidth,
antennas=1):
self.maxdesigns = 0
self.blocks = {}
self.blockalgs = {}
self.totalblocks = 0
self.maxdesigns = 1
self.singleimplementation = 1
#add the platform array
self.platforms = {}
# add platforms: cost, inputbw, outputbw, resources
self.platforms['ROACH'] = Platform(
'ROACH', 6700, 40, 40, ['registers', 'luts', 'dsp', 'bram'])
self.platforms['GTX580'] = Platform('GTX580', 3500, 10, 1, ['time'])
for i in range(0, antennas):
# add the ADC
adc_bw = bandwidth * input_bitwidth
self.blocks['ADC' + ` i `] = CBlock(
'ADC',
CBlock.getADCModel(self.platforms, bandwidth, input_bitwidth),
-1, 0, 0, 'PFB' + ` i `, 0, adc_bw, 1)
self.totalblocks += 1
# add the PFB
pfb_bw = bandwidth * 32
self.blocks['PFB' + ` i `] = CBlock(
'PFB',
CBlock.getPFBWModel(self.platforms, bandwidth, pfb_bw,
numcoarsechannels), 'ADC' + ` i `, 0,
adc_bw, 'FFT_coarse' + ` i `, 0, adc_bw, 1)
self.totalblocks += 1
# add the FFT
#print CBlock.getFFTModel(self.platforms, bandwidth, input_bitwidth, numchannels)
fft_coarse_out_bandwidth = bandwidth * fft_coarse_out_bitwidth * 2
self.blocks['FFT_coarse' + ` i `] = CBlock(
'FFT_coarse',
CBlock.getFFTWModel(self.platforms, bandwidth,
numcoarsechannels), 'PFB' + ` i `, 0,
pfb_bw, 'Transpose' + ` i `, 0, fft_coarse_out_bandwidth, 1)
self.totalblocks += 1
# adjust to ensure the block fits on the gpu
fft_fine_in_bandwidth = fft_coarse_out_bandwidth / numcoarsechannels
finemodel = CBlock.getFFTModel(self.platforms,
fft_fine_in_bandwidth,
numfinechannels)
if (finemodel['GTX580']['time'] < 0.1):
multiplier = pow(
2, int(log(0.1 / finemodel['GTX580']['time'], 2)))
else:
multiplier = 1
finemodel['GTX580'][
'time'] = finemodel['GTX580']['time'] * multiplier
fine_blocks = int(numcoarsechannels / multiplier)
fine_sky_bandwidth = bandwidth / fineblocks
fine_block_bandwidth = fft_coarse_out_bandwidth / fine_blocks
self.blocks['Transpose' + ` i `] = CBlock(
'Transpose',
CBlock.getTransposeModel(self.platforms, bandwidth,
numcoarsechannels, numfinechannels),
'FFT_coarse' + ` i `, 0, fft_coarse_out_bandwidth,
'FFT_fine' + ` i `, 1, fft_coarse_out_bandwidth, 1)
self.totalblocks += 1
self.blocks['FFT_fine' + ` i `] = CBlock('FFT_fine', finemodel,
'Transpose' + ` i `, 0,
fine_block_bandwidth,
'VAcc' + ` i `, 0,
fft_fine_in_bandwidth,
fine_blocks)
self.totalblocks += fine_blocks
self.blocks['VAcc' + ` i `] = CBlock(
'VAcc', {
'ROACH': {
'registers': 0.2,
'luts': 0.1,
'dsp': 0,
'bram': 0.4
},
'GTX580': {
'time': 0.001
}
}, 'FFT_fine' + ` i `, 0, fine_block_bandwidth, -1, 0, 0,
fine_blocks)
self.totalblocks += fine_blocks
def __init__(self, numchannels, numantpol, accumulation_length, skybandwidth, input_bitwidth, fft_out_bitwidth):
self.blocks = {}
self.totalblocks = 0
self.maxdesigns = 1
self.singleimplementation = 1
self.windowsize = 1024
cost = 'dollars'
#cost = 'power'
#add the platforms
self.platforms = {}
#self.platforms['GTX580'] = Platform.createGTX580Server(cost)
#self.platforms['ROACH'] = Platform.createRoach(cost)
self.platforms['DualGTX690'] = Platform.createDualGTX690Server(cost)
self.platforms['ROACH2'] = Platform.createRoach2(cost)
# add the ADC
adc_bw = skybandwidth*2*input_bitwidth
#self.blocks['ADC'] = CBlock('ADC',CBlock.getADCModel(self.platforms, skybandwidth, input_bitwidth),-1,0,0,'FIR',0,4*adc_bw,numantpol/4)
# we are using a 16 input adc board
# multiplier needs to be a multiple of 4 because the benchmarks do 4 parallel firs and ffts
adcmultiplier = 4 # process 4 streams at a time
self.blocks['ADC'] = CBlock('ADC',CBlock.getADCModel(self.platforms, skybandwidth, input_bitwidth), -1,0,0,'PFB',0,adcmultiplier*adc_bw,numantpol/adcmultiplier, CBlock.getADCMaximums(self.platforms, adcmultiplier))
self.totalblocks += numantpol/adcmultiplier
#use pfb to process 4 channels at a time
fft_out_bandwidth = skybandwidth * 2 * fft_out_bitwidth
pfb_model = CBlock.getPFBModel(self.platforms, skybandwidth, input_bitwidth, numchannels)
fft_model = CBlock.getFFTRealModel(self.platforms, skybandwidth, numchannels)
firfftmodel = CBlock.combineModels(pfb_model, fft_model)
#print firfftmodel
#firfft2xmodel = CBlock.combineModels(firfftmodel, firfftmodel)
#print firfft2xmodel
self.blocks['PFB'] = CBlock('PFB',firfftmodel,'ADC',0,adcmultiplier*adc_bw,'Transpose',0,adcmultiplier*fft_out_bandwidth,numantpol/adcmultiplier)
self.totalblocks += numantpol/adcmultiplier
transposemodel = CBlock.getTransposeModel(self.platforms, skybandwidth, numchannels, self.windowsize)
self.blocks['Transpose'] = CBlock('Transpose',transposemodel,'PFB',0,adcmultiplier*fft_out_bandwidth,'XEng',0,adcmultiplier*fft_out_bandwidth,numantpol/adcmultiplier)
self.totalblocks += numantpol/adcmultiplier
# add the PFBTranspose
#fft_out_bandwidth = skybandwidth * 2 * fft_out_bitwidth
#firmodel = CBlock.getPFBModel(self.platforms, skybandwidth, input_bitwidth, numchannels)
#fftmodel = CBlock.getFFTRealModel(self.platforms, skybandwidth, numchannels)
#transposemodel = CBlock.getTransposeModel(self.platforms, skybandwidth, numchannels, self.windowsize)
#combinedmodel = CBlock.combineModels(firmodel, CBlock.combineModels(fftmodel, transposemodel))
#we need 4 of these for our 16 input adc
#self.blocks['PFBTranspose'] = CBlock('PFBTranspose',combinedmodel,'ADC',0,adcmultiplier*adc_bw,'XEng',1,adcmultiplier * fft_out_bandwidth,numantpol/adcmultiplier)
#self.totalblocks += numantpol/adcmultiplier
# add the XEngines
gtx580_max_bw = {32:0.06914, 64:0.03095, 96:0.01748, 128:0.01069, 192:0.00536, 256:0.00318, 512:0.00087, 1024:0.00023}
# the minimum number of xengines we need
# if we use any fewer, they will not fit on the gpu
mingpuxengines = int(numpy.power(2,numpy.ceil(numpy.log2(skybandwidth/gtx580_max_bw[numantpol]))))
# assume xengine is running at 200MHz, takes nantpol clock cycles to get the data out for a single frequency channel
# maximum bandwidth it can process is 200MHz/nantpol
maxfpgaxengbw = .2/numantpol
#the maximum amount of bandwidth we can process in an xengine and still support our platforms
maxxenginebw = min(maxfpgaxengbw,gtx580_max_bw[numantpol])
# we need to create this many xengines to meet the spec
minxengines = int(skybandwidth/maxxenginebw)
#note: this needs to be a power of 2
numxengines = 8*minxengines
#numxengines = 4*mingpuxengines
#print 'Num xengines is: ' + `numxengines`
#numxengines = mingpuxengines*4
xengine_sky_bandwidth = skybandwidth/numxengines
#print 'Sky bw is: ' + `xengine_sky_bandwidth`
#print xengine_sky_bandwidth
xengine_in_bandwidth = numantpol*fft_out_bandwidth/numxengines
#print CBlock.getXEngModel(self.platforms, xengine_sky_bandwidth, numantpol)
self.blocks['XEng'] = CBlock('XEng',CBlock.getXEngModel(self.platforms, xengine_sky_bandwidth, numantpol) ,'Transpose', 1,xengine_in_bandwidth,-1,0,0,numxengines)
self.totalblocks += numxengines
def __init__(self, numchannels, numantpol, accumulation_length,
skybandwidth, input_bitwidth, fft_out_bitwidth):
self.blocks = {}
self.totalblocks = 0
self.maxdesigns = 1
self.singleimplementation = 1
self.windowsize = 1024
cost = 'dollars'
#cost = 'power'
#add the platforms
self.platforms = {}
#self.platforms['GTX580'] = Platform.createGTX580Server(cost)
#self.platforms['ROACH'] = Platform.createRoach(cost)
self.platforms['DualGTX690'] = Platform.createDualGTX690Server(cost)
self.platforms['ROACH2'] = Platform.createRoach2(cost)
# add the ADC
adc_bw = skybandwidth * 2 * input_bitwidth
#self.blocks['ADC'] = CBlock('ADC',CBlock.getADCModel(self.platforms, skybandwidth, input_bitwidth),-1,0,0,'FIR',0,4*adc_bw,numantpol/4)
# we are using a 16 input adc board
# multiplier needs to be a multiple of 4 because the benchmarks do 4 parallel firs and ffts
adcmultiplier = 4 # process 4 streams at a time
self.blocks['ADC'] = CBlock(
'ADC',
CBlock.getADCModel(self.platforms, skybandwidth,
input_bitwidth), -1, 0, 0, 'PFB', 0,
adcmultiplier * adc_bw, numantpol / adcmultiplier,
CBlock.getADCMaximums(self.platforms, adcmultiplier))
self.totalblocks += numantpol / adcmultiplier
#use pfb to process 4 channels at a time
fft_out_bandwidth = skybandwidth * 2 * fft_out_bitwidth
pfb_model = CBlock.getPFBModel(self.platforms, skybandwidth,
input_bitwidth, numchannels)
fft_model = CBlock.getFFTRealModel(self.platforms, skybandwidth,
numchannels)
firfftmodel = CBlock.combineModels(pfb_model, fft_model)
#print firfftmodel
#firfft2xmodel = CBlock.combineModels(firfftmodel, firfftmodel)
#print firfft2xmodel
self.blocks['PFB'] = CBlock('PFB', firfftmodel, 'ADC', 0,
adcmultiplier * adc_bw, 'Transpose', 0,
adcmultiplier * fft_out_bandwidth,
numantpol / adcmultiplier)
self.totalblocks += numantpol / adcmultiplier
transposemodel = CBlock.getTransposeModel(self.platforms, skybandwidth,
numchannels, self.windowsize)
self.blocks['Transpose'] = CBlock('Transpose', transposemodel, 'PFB',
0, adcmultiplier * fft_out_bandwidth,
'XEng', 0,
adcmultiplier * fft_out_bandwidth,
numantpol / adcmultiplier)
self.totalblocks += numantpol / adcmultiplier
# add the PFBTranspose
#fft_out_bandwidth = skybandwidth * 2 * fft_out_bitwidth
#firmodel = CBlock.getPFBModel(self.platforms, skybandwidth, input_bitwidth, numchannels)
#fftmodel = CBlock.getFFTRealModel(self.platforms, skybandwidth, numchannels)
#transposemodel = CBlock.getTransposeModel(self.platforms, skybandwidth, numchannels, self.windowsize)
#combinedmodel = CBlock.combineModels(firmodel, CBlock.combineModels(fftmodel, transposemodel))
#we need 4 of these for our 16 input adc
#self.blocks['PFBTranspose'] = CBlock('PFBTranspose',combinedmodel,'ADC',0,adcmultiplier*adc_bw,'XEng',1,adcmultiplier * fft_out_bandwidth,numantpol/adcmultiplier)
#self.totalblocks += numantpol/adcmultiplier
# add the XEngines
gtx580_max_bw = {
32: 0.06914,
64: 0.03095,
96: 0.01748,
128: 0.01069,
192: 0.00536,
256: 0.00318,
512: 0.00087,
1024: 0.00023
}
# the minimum number of xengines we need
# if we use any fewer, they will not fit on the gpu
mingpuxengines = int(
numpy.power(
2,
numpy.ceil(numpy.log2(skybandwidth /
gtx580_max_bw[numantpol]))))
# assume xengine is running at 200MHz, takes nantpol clock cycles to get the data out for a single frequency channel
# maximum bandwidth it can process is 200MHz/nantpol
maxfpgaxengbw = .2 / numantpol
#the maximum amount of bandwidth we can process in an xengine and still support our platforms
maxxenginebw = min(maxfpgaxengbw, gtx580_max_bw[numantpol])
# we need to create this many xengines to meet the spec
minxengines = int(skybandwidth / maxxenginebw)
#note: this needs to be a power of 2
numxengines = 8 * minxengines
#numxengines = 4*mingpuxengines
#print 'Num xengines is: ' + `numxengines`
#numxengines = mingpuxengines*4
xengine_sky_bandwidth = skybandwidth / numxengines
#print 'Sky bw is: ' + `xengine_sky_bandwidth`
#print xengine_sky_bandwidth
xengine_in_bandwidth = numantpol * fft_out_bandwidth / numxengines
#print CBlock.getXEngModel(self.platforms, xengine_sky_bandwidth, numantpol)
self.blocks['XEng'] = CBlock(
'XEng',
CBlock.getXEngModel(self.platforms, xengine_sky_bandwidth,
numantpol), 'Transpose', 1,
xengine_in_bandwidth, -1, 0, 0, numxengines)
self.totalblocks += numxengines