def __init__(self, bit_file_path):
self.ol = Overlay(bit_file_path)
self.dma = self.ol.axi_dma_0
# declare input/output types and shapes for the accelerator
# input FINN DataType
self.idt = DataType.UINT2
# normal, folded and packed input shapes
self.ishape_normal = (1, 8)
self.ishape_folded = (1, 1, 8)
self.ishape_packed = (1, 1, 2)
# output FINN DataType
self.odt = DataType.INT32
# normal, folded and packed output shapes
self.oshape_normal = (1, 3)
self.oshape_folded = (1, 1, 3)
self.oshape_packed = (1, 1, 12)
# allocate a PYNQ buffer for the packed input buffer
self.ibuf_packed_device = allocate(shape=self.ishape_packed,
dtype=np.uint8)
# allocate a PYNQ buffer for the packed output buffer
self.obuf_packed = allocate(shape=self.oshape_packed, dtype=np.uint8)
self.mt_node_thresholds = np.asarray(
[[-0.01027827151119709, 0.01027827151119709]])
self.multiply_node_const = 0.014718746766448021
self.add_node_mat = np.asarray(
[-0.08831246197223663, 0.17662496864795685, -0.058874987065792084])
def __init__(self, fs=1000000, N=6000):
# Input paramters
self._fs = fs
self._N = N
# Cache timeseries and reusable random data
self._random_data = [rnd.randint(0, 3) for _ in range(N)]
self._t = np.array([i / fs for i in range(N)])
# Overlay config
bit_name = os.path.dirname(__file__) + '/agc_loopback.bit'
ol = Overlay(bit_name)
self._ol = ol
#Avoid PYNQ's get_attr overhead by aliasing IPs
self._agc = ol.agc
self._dma_in_i = ol.dma_in_i
self._dma_in_q = ol.dma_in_q
self._dma_agc_i = ol.dma_agc_i
self._dma_agc_q = ol.dma_agc_q
self._dma_agc_g = ol.dma_agc_g
# Allocate buffers
self._buf_in_i = allocate(shape=(N, ), dtype=np.int16)
self._buf_in_q = allocate(shape=(N, ), dtype=np.int16)
self._buf_agc_i = allocate(shape=(N, ), dtype=np.int16)
self._buf_agc_q = allocate(shape=(N, ), dtype=np.int16)
self._buf_agc_g = allocate(shape=(N, ), dtype=np.uint32)
def hw_inference(self, input_data_processed):
packed_input = self.pack(input_data_processed.image_array)
directions = 2 if self.bidirectional_enabled else 1
self.accel_input_buffer = allocate(shape=packed_input.shape,
dtype=np.uint64)
self.accel_output_buffer = allocate(shape=(128, ), dtype=np.uint64)
np.copyto(self.accel_input_buffer, packed_input)
bytes_read = np.ceil((self.input_bitwidth * self.input_size * 8) / 64)
bytes_read = int(bytes_read * directions * input_data_processed.width)
self.BLSTM_CTC.numberColumns_V = input_data_processed.width
self.BLSTM_CTC.numberColumnsTwice_V = directions * input_data_processed.width
self.BLSTM_CTC.numberBytesRead_V = bytes_read
self.BLSTM_CTC.input_buffer_V_1 = self.accel_input_buffer.physical_address & 0xffffffff
self.BLSTM_CTC.input_buffer_V_2 = (
self.accel_input_buffer.physical_address >> 32) & 0xffffffff
self.BLSTM_CTC.output_buffer_V_1 = self.accel_output_buffer.physical_address & 0xffffffff
self.BLSTM_CTC.output_buffer_V_2 = (
self.accel_output_buffer.physical_address >> 32) & 0xffffffff
start = time.time()
self.ExecAccel()
end = time.time()
print("Inference took = {} ms...".format((end - start) * 1000))
print("Frames per second = {}".format(1 / (end - start)))
predictions = np.copy(
np.frombuffer(self.accel_output_buffer, dtype=np.uint64))
return predictions
def fpga_evaluate_dance(self, imu_data):
global dma
input_buffer0 = allocate(shape=(100,), dtype=np.float32)
output_buffer0 = allocate(shape=(12,), dtype=np.float32)
segment = self.preprocess_segment(imu_data)
segment = (segment - self.scaler_mean) / self.scaler_scale
self.choose_model[0] = 0
dma.sendchannel.transfer(self.choose_model)
dma.sendchannel.wait()
for j in range(100):
input_buffer0[j] = segment[j]
dma.sendchannel.transfer(input_buffer0)
dma.recvchannel.transfer(output_buffer0)
dma.sendchannel.wait()
dma.recvchannel.wait()
raw_output = [output_buffer0[i] for i in range(12)]
index = np.argmax(raw_output, axis=-1)
#print(f"Prediction: {self.dance_labels[index]} | {raw_output[index]}")
return self.dance_labels[index]
def __init__(self, bit_file_path):
self.ol = Overlay(bit_file_path)
self.dma = self.ol.axi_dma_0
# declare input/output types and shapes for the accelerator
# input FINN DataType
self.idt = DataType.UINT2
# normal, folded and packed input shapes
self.ishape_normal = (1, 256)
self.ishape_folded = (1, 4, 64)
self.ishape_packed = (1, 4, 16)
# output FINN DataType
self.odt = DataType.INT32
# normal, folded and packed output shapes
self.oshape_normal = (1, 6)
self.oshape_folded = (1, 1, 6)
self.oshape_packed = (1, 1, 24)
# allocate a PYNQ buffer for the packed input buffer
self.ibuf_packed_device = allocate(shape=self.ishape_packed,
dtype=np.uint8)
# allocate a PYNQ buffer for the packed output buffer
self.obuf_packed = allocate(shape=self.oshape_packed, dtype=np.uint8)
self.mt_node_thresholds = np.asarray(
[[-0.00000762939453125, 0.00000762939453125]])
self.multiply_node_const = 0.19533488154411316
self.add_node_mat = np.asarray([
0.7813395261764526, 0, 0, 0.7813395261764526, -0.3906697630882263,
0
])
def __init__(self, N, bitfile):
"""Instantiate the FINN accelerator driver.
Gets batchsize (N) as integer and path to bitfile as string."""
self.N = N
# input FINN DataType
self.idt = DataType.UINT8
# output FINN DataType
self.odt = DataType.UINT32
# input and output shapes
self.ishape_normal = (N, 256)
self.oshape_normal = (N, 9)
self.ishape_folded = (N, 8, 32)
self.oshape_folded = (N, 1, 9)
self.ishape_packed = (N, 8, 32) # datatype np.uint8
self.oshape_packed = (N, 1, 36) # datatype np.uint8
# load bitfile and set up accelerator
self.ol = Overlay(bitfile)
self.dma = self.ol.axi_dma_0
self.ctrl_regs = self.ol.resize_accel_0
# neuron folding factor of output = iterations per sample
self.itersPerSample = self.oshape_packed[-2]
# AXI lite register offset for number of iterations
# used by TLastMarker to signal end of transmission for AXI CDMA
self.REG_OFFSET_NUM_ITERS = 0x10
# set up TLastMarker with correct num. samples
self.ctrl_regs.write(self.REG_OFFSET_NUM_ITERS,
self.N * self.itersPerSample)
# allocate a PYNQ buffer for the packed input and buffer
self.ibuf_packed_device = allocate(shape=self.ishape_packed,
dtype=np.uint8)
self.obuf_packed_device = allocate(shape=self.oshape_packed,
dtype=np.uint8)
def benchmark_synthetic(bs, nreps):
ibuf = pynq.allocate((bs, 3, 224, 224), dtype=np.int8, target=ol.bank0)
obuf = pynq.allocate((bs, 5), dtype=np.uint32, target=ol.bank0)
# Start power monitoring
pwr_rec = setup_power_recording()
pwr_rec.record(0.1)
total_duration = time.monotonic()
for i in range(nreps):
accelerator.call(ibuf, obuf, fcbuf, bs)
total_duration = time.monotonic() - total_duration
# Stop the power monitoring
pwr_rec.stop()
latency = total_duration / nreps
fps = int((nreps / total_duration) * bs)
# Aggregate board/fpga power into a Pandas dataframe
f = pwr_rec.frame
powers = pd.DataFrame(index=f.index)
powers['board_power'] = f['12v_aux_power'] + f['12v_pex_power']
powers['fpga_power'] = f['vccint_power']
return fps, latency, powers
def predict(self, X, y_shape, dtype=np.float32, debug=None, profile=False, encode=None, decode=None):
"""
Obtain the predictions of the NN implemented in the FPGA.
Parameters:
- X : the input vector. Should be numpy ndarray.
- y_shape : the shape of the output vector. Needed to the accelerator to set the TLAST bit properly and
for sizing the output vector shape.
- dtype : the data type of the elements of the input/output vectors.
Note: it should be set depending on the interface of the accelerator; if it uses 'float'
types for the 'data' AXI-Stream field, 'np.float32' dtype is the correct one to use.
Instead if it uses 'ap_fixed<A,B>', 'np.intA' is the correct one to use (note that A cannot
any integer value, but it can assume {..., 8, 16, 32, ...} values. Check `numpy`
doc for more info).
In this case the encoding/decoding has to be computed by the PS. For example for
'ap_fixed<16,6>' type the following 2 functions are the correct one to use for encode/decode
'float' -> 'ap_fixed<16,6>':
```
def encode(xi):
return np.int16(round(xi * 2**10)) # note 2**10 = 2**(A-B)
def decode(yi):
return yi * 2**-10
encode_v = np.vectorize(encode) # to apply them element-wise
decode_v = np.vectorize(decode)
```
- profile : boolean. Set it to `True` to print the performance of the algorithm in term of `inference/s`.
- encode/decode: function pointers. See `dtype` section for more information.
- return: an output array based on `np.ndarray` with a shape equal to `y_shape` and a `dtype` equal to
the namesake parameter.
"""
if profile:
timea = datetime.now()
if encode is not None:
X = encode(X)
with allocate(shape=X.shape, dtype=dtype) as input_buffer, \
allocate(shape=y_shape, dtype=dtype) as output_buffer:
input_buffer[:] = X
self.hier_0.axi_dma_0.sendchannel.transfer(input_buffer)
self.hier_0.axi_dma_0.recvchannel.transfer(output_buffer)
if debug:
print("Transfer OK")
self.hier_0.axi_dma_0.sendchannel.wait()
if debug:
print("Send OK")
self.hier_0.axi_dma_0.recvchannel.wait()
if debug:
print("Receive OK")
result = output_buffer.copy()
if decode is not None:
result = decode(result)
if profile:
timeb = datetime.now()
dts, rate = self._print_dt(timea, timeb, len(X))
return result, dts, rate
return result
def __init__(self,
N,
bitfile,
platform="alveo",
device_name="xilinx_u50_gen3x16_xdma_201920_3"):
"""Instantiate the FINN accelerator driver.
Gets batchsize (N) as integer and path to bitfile as string."""
self.platform = platform
self.N = N
# input FINN DataType
self.idt = DataType.UINT8
# output FINN DataType
self.odt = DataType.UINT8
# input and output shapes
self.ishape_normal = (N, 32, 32, 3)
self.oshape_normal = (N, 1)
self.ishape_folded = (N, 32, 32, 1, 3)
self.oshape_folded = (N, 1, 1)
self.ishape_packed = (N, 32, 32, 1, 3) # datatype np.uint8
self.oshape_packed = (N, 1, 1) # datatype np.uint8
# load bitfile and set up accelerator
self.device = [i for i in Device.devices if i.name == device_name][0]
Device.active_device = self.device
self.ol = Overlay(bitfile)
# neuron folding factor of output = iterations per sample
self.itersPerSample = self.oshape_packed[-2]
# clock frequency as specified by user
self.fclk_mhz = 100.0
if self.platform == "alveo":
self.idma = self.ol.idma0
self.odma = self.ol.odma0
elif self.platform == "zynq-iodma":
self.idma = self.ol.idma0
self.odma = self.ol.odma0
# set the clock frequency as specified by user during transformations
if self.fclk_mhz > 0:
Clocks.fclk0_mhz = self.fclk_mhz
else:
raise ValueError("Supported platforms are zynq-iodma alveo")
# allocate a PYNQ buffer for the packed input and buffer
if self.platform == "alveo":
self.ibuf_packed_device = allocate(shape=self.ishape_packed,
dtype=np.uint8)
self.obuf_packed_device = allocate(shape=self.oshape_packed,
dtype=np.uint8)
else:
self.ibuf_packed_device = allocate(shape=self.ishape_packed,
dtype=np.uint8,
cacheable=True)
self.obuf_packed_device = allocate(shape=self.oshape_packed,
dtype=np.uint8,
cacheable=True)
def _create_buffer(
self,
data=np.array(
[72, 101, 108, 108, 111, 32, 87, 111, 114, 108, 100, 33],
dtype=np.uint8),
eof=1,
padding=0):
"""Create a buffer that is loaded user data. Append the Extended Barker sequence
to the user data and then pad with zeros
"""
self._flags = eof
if data.size == 0:
raise ValueError('Message size should be greater than 0.')
msg = np.array(data, dtype=np.uint8)
# Append Barker and Random Data
bkr = np.array([
0, 0, 63, 112, 28,
len(msg) + 5, self._frame_number, self._flags, 5,
len(msg), padding
],
dtype=np.uint8)
rnd = np.array([randint(0, 255) for p in range(0, self.random_size)],
dtype=np.uint8)
seq = np.append(bkr, msg)
seq = np.append(rnd, seq)
pad = np.append(
seq,
np.zeros(int(
np.ceil((len(rnd) + len(bkr) + len(msg)) / 32) * 32 -
(len(rnd) + len(bkr) + len(msg))),
dtype=np.uint8))
buf = allocate(shape=(len(pad), ), dtype=np.uint8)
buf[:] = pad[:]
return buf
def TestStreamMaster(baseval):
# Set DMA to write SS2M data to a specifc physical address
output_buffer = allocate(shape=(8, ), dtype='u4')
output_buffer[:] = 0
dma.recvchannel.start()
dma.recvchannel.transfer(output_buffer)
# Confiugre the peripheral to send baseval to baseval + 7
ol.AXIMasterSlaveStream_0.write(0x00, 0x02)
ol.AXIMasterSlaveStream_0.write(0x00, 0x00)
ol.AXIMasterSlaveStream_0.write(0x04, baseval)
ol.AXIMasterSlaveStream_0.write(0x00, 0x01)
dma.recvchannel.wait()
dma.recvchannel.stop()
# Check the data matches.
for i in range(8):
expected = baseval + i
if not (output_buffer[i] == expected):
print("TestStreamMaster: Error on output buffer " + str(i) + \
" Expected " + "{:x}".format(expected) + \
" Actual " + str(output_buffer[i]))
return 0
return 1
def render_frame(ipRender, ipIDMA, ipODMA, mesh_buffers, transforms, texture_ids, output_buffer):
mbuf_trans = []
for i, mbuf in enumerate(mesh_buffers):
num_faces = len(mbuf) // 24
render_set_num_faces(ipRender, num_faces)
mbuf_trans.append(allocate(shape=len(mbuf), dtype=np.float32))
render_set_transform(ipRender, transforms[i])
render_transform(ipRender, ipIDMA, ipODMA, mbuf, mbuf_trans[-1])
for i in range(8):
fh0 = 60 * i
fh1 = 60 * (i + 1)
ipRender.write(0x00C0, fh0)
ipRender.write(0x00C4, fh1)
render_reset_buffer(ipRender)
for j, mbuf in enumerate(mbuf_trans):
num_faces = len(mbuf) // 24
render_set_num_faces(ipRender, num_faces)
render_set_texture_id(ipRender, texture_ids[j])
ipRender.write(0x0010, 2)
ipRender.write(0x0000, 1)
ipIDMA.sendchannel.transfer(mbuf)
ipIDMA.sendchannel.wait()
render_wait_done(ipRender)
ipRender.write(0x0010, 3)
ipRender.write(0x0000, 1)
ipODMA.recvchannel.transfer(output_buffer[fh0*640:fh1*640])
ipODMA.recvchannel.wait()
render_wait_done(ipRender)
def test_withfree(device):
cache = MockCache()
with device.check_memops(allocates=[(10, False, BUFFER_ADDRESS)]):
with pynq.allocate((1024, 1024), 'u4', target=device) as buf:
buf.pointer = 1234
buf.return_to = cache
assert cache.returns == [1234]
def getframe(self):
"""Retrieve a frame from the cache or create a new frame if the
cache is empty. The freebuffer method of the returned array is
overriden to return the object to the cache rather than freeing
the object.
"""
if self._cache:
frame = allocate(
shape=self._mode.shape, dtype='u1', cacheable=self._cacheable,
pointer=self._cache.pop(), cache=self)
else:
frame = allocate(
shape=self._mode.shape, dtype=np.uint8,
cacheable=self._cacheable, cache=self)
return frame
def allocate_buffer(self, name, num_samples, data_type="unsigned int"):
"""This method allocates the source or the destination buffers.
Usually, the source buffer stores 32-bit samples, while the
destination buffer stores 64-bit samples.
Note that the numpy array has to be deep-copied before users can
free the buffer.
Parameters
----------
name : str
The name of the string, used for indexing the buffers.
num_samples : int
The number of samples that needs to be generated or captured.
data_type : str
The type of the data.
Returns
-------
int
The address of the source or destination buffer.
"""
dtype = data_type
for k, v in BYTE_WIDTH_TO_CTYPE.items():
if v == data_type:
dtype = BYTE_WIDTH_TO_NPTYPE[k]
break
buf = allocate(num_samples, dtype=dtype)
self.buffers[name] = buf
return buf.physical_address
def window(self, window):
window_size = self._window_packetsize
buffer = allocate(shape=(window_size, ), dtype=np.int32)
self._window_address = buffer.device_address
if window == 'rectangular':
buffer[:] = np.int32(np.ones(window_size)[:] * 2**14)
elif window == 'bartlett':
buffer[:] = np.int32(np.bartlett(window_size)[:] * 2**14)
elif window == 'blackman':
buffer[:] = np.int32(np.blackman(window_size)[:] * 2**14)
elif window == 'hamming':
buffer[:] = np.int32(np.hamming(window_size)[:] * 2**14)
elif window == 'hanning':
buffer[:] = np.int32(np.hanning(window_size)[:] * 2**14)
else:
buffer[:] = np.int32(np.ones(window_size)[:] * 2**14)
window = 'rectangular'
self._window_transfer = 1
while not self.window_ready:
pass
self._window_transfer = 0
self._window_type = window
self._window_squaresum = np.sum(
(np.array(buffer, dtype=np.single) * 2**-14)**2)
self._window_sum = np.sum((np.array(buffer, dtype=np.single)))
buffer.freebuffer()
self._spectrum_typescale = \
int(struct.unpack('!i',struct.pack('!f',float((self._sample_frequency/self._decimation_factor)/(self._number_samples))))[0])
self._spectrum_powerscale = \
int(struct.unpack('!i',struct.pack('!f',float(1/((self._sample_frequency/self._decimation_factor)*self._window_squaresum))))[0])
def _transfer(self):
buff_len = self.controller.receive_size
if buff_len > 0:
# Create new receive buffer for message
self._rx_buff.freebuffer()
self._rx_buff = allocate(shape=(buff_len, ), dtype=np.uint8)
# Prepare to receive the message
self._dma_transfer(self._rx_buff)
# Obtain the message
self._message = np.array(self._rx_buff.astype(np.uint32), \
dtype = np.uint8)[5:len(self._rx_buff)]
# Set frame to allow the user to read the frame data
self.frame = {
"number": self._rx_buff[0],
"flags": self._rx_buff[1],
"length": {
"total": buff_len,
"header": self._rx_buff[2],
"payload": self._rx_buff[3],
"padding": self._rx_buff[4]
},
"payload": self._message
}
def fft_size(self, fft_size):
if fft_size in [8192, 4096, 2048, 1024, 512, 256, 128, 64]:
running = False
if self.dma_enable:
self.dma_enable = 0
running = True
self.ssr_packetsize = 0
time.sleep(0.2)
self._dma_length = fft_size
[self._buffer[x].freebuffer() for x in range(0, 3)]
self._buffer = [
allocate(shape=(self._dma_length, ), dtype=np.single)
for x in range(0, 3)
]
self._dma_bufferaddress_0 = self._buffer[0].device_address
self._dma_bufferaddress_1 = self._buffer[0].device_address
self._dma_bufferaddress_2 = self._buffer[0].device_address
self._spectrum_fftselector = int(np.log2(fft_size) - 6)
self._window_packetsize = fft_size
self.window = self._window_type
self.ssr_packetsize = int(fft_size / 8)
self.plot.number_samples = fft_size
self._number_samples = int(2**(self._spectrum_fftselector + 6))
self._spectrum_typescale = \
int(struct.unpack('!i',struct.pack('!f',float((self._sample_frequency/self._decimation_factor)/(self._number_samples))))[0])
self._spectrum_powerscale = \
int(struct.unpack('!i',struct.pack('!f',float(1/((self._sample_frequency/self._decimation_factor)*self._window_squaresum))))[0])
if running:
self.dma_enable = 1
def test_allocate(device):
with device.check_memops(allocates=[(10, False, BUFFER_ADDRESS)]):
buf = pynq.allocate((1024, 1024), 'u4', target=device)
assert buf.device_address == BUFFER_ADDRESS
assert buf.physical_address == BUFFER_ADDRESS
assert buf.bo == 10
assert buf.coherent is False
assert buf.cacheable is True
def __init__(self, width, height, frequency, bitfile_name=None, **kwargs):
"""Construct a new HoughEvaluation
width: Required. The width of the image to be processed.
height: Required. The height of the image to be processed.
frequency: Required. The system frequency of the Hough architecture.
bitfile_name: Optional. If left None, the 'zcu104_hep.bit' bundled
with the hough_evaluation_platform package will be used.
"""
# Generate default bitfile name
if bitfile_name is None:
this_dir = os.path.dirname(__file__)
bitfile_name = os.path.join(this_dir, 'bitstream',
'zcu104_hep.bit')
# Set FPD and LPD interface widths
from pynq import MMIO
fpd_cfg = MMIO(0xfd615000, 4)
fpd_cfg.write(0, 0x00000A00)
lpd_cfg = MMIO(0xff419000, 4)
lpd_cfg.write(0, 0x00000000)
# Create Overlay
super().__init__(bitfile_name, **kwargs)
# Set system parameters
self.width = width
self.height = height
self.frequency = frequency
self.nrho = int(
np.ceil(np.sqrt((self.width / 2)**2 + (self.height / 2)**2)) * 2)
self.ntheta = np.size(np.arange(0, 180))
# Initialise system outputs
self.time = 0
self.pmvr = 0
# Extract ipcore from the overlay with friendly name and initialise
self.hpa = self.hpa_module
self.hpa.frequency = self.frequency
# Set up dma buffers
self.inarray = allocate(shape=(self.height, self.width),
dtype=np.uint32)
self.outarray = allocate(shape=(self.ntheta, self.nrho),
dtype=np.uint32)
def allocate_mem(self,
X_shape,
y_shape,
dtype=np.float32,
trg_in=None,
trg_out=None):
"""
Buffer allocation in the card memory
Parameters
----------
X_shape : input buffer shape.
y_shape : output buffer shape.
dtype : the data type of the elements of the input/output vectors.
Note: it should be set depending on the interface of the accelerator; if it uses 'float'
types for the 'data' AXI-Stream field, 'np.float32' dtype is the correct one to use.
Instead if it uses 'ap_fixed<A,B>', 'np.intA' is the correct one to use (note that A cannot
any integer value, but it can assume {..., 8, 16, 32, ...} values. Check `numpy`
doc for more info).
In this case the encoding/decoding has to be computed by the host machine. For example for
'ap_fixed<16,6>' type the following 2 functions are the correct one to use for encode/decode
'float' -> 'ap_fixed<16,6>':
```
def encode(xi):
return np.int16(round(xi * 2**10)) # note 2**10 = 2**(A-B)
def decode(yi):
return yi * 2**-10
encode_v = np.vectorize(encode) # to apply them element-wise
decode_v = np.vectorize(decode)
```
trg_in : input buffer target memory. By default the v++ command
set it to HBM[0] for alveo-u50.
trg_out : output buffer target memory.By default the v++ command
set it to HBM[0] for alveo-u50.
Assigns
-------
input_buffer : input PYNQ buffer, must be allocated first and just once.
output_buffer : output PYNQ buffer, must be allocated first and just once.
input_buffer, output_buffer : input and output PYNQ buffers
"""
self.input_buffer = allocate(shape=X_shape, dtype=dtype, target=trg_in)
self.output_buffer = allocate(shape=y_shape,
dtype=dtype,
target=trg_out)
def __init__(self,
bitfile_name,
x_shape,
y_shape,
dtype=np.float32,
dtbo=None,
download=True,
ignore_version=False,
device=None):
super().__init__(bitfile_name,
dtbo=None,
download=True,
ignore_version=False,
device=None)
self.sendchannel = self.hier_0.axi_dma_0.sendchannel
self.recvchannel = self.hier_0.axi_dma_0.recvchannel
self.input_buffer = allocate(shape=x_shape, dtype=dtype)
self.output_buffer = allocate(shape=y_shape, dtype=dtype)
def TestStreamSlave(depth):
# Send MM2S data from CPU Memory to Stream by DMA.
sendBuffer = allocate(shape=(depth, ), dtype='u4')
sendBuffer[:] = [100 + i for i in range(0, depth)]
recvBuffer = allocate(shape=(depth, ), dtype='u4')
recvBuffer[:] = 0
dma.recvchannel.start()
dma.recvchannel.transfer(recvBuffer)
dma.sendchannel.start()
dma.sendchannel.transfer(sendBuffer)
dma.sendchannel.wait()
dma.sendchannel.stop()
dma.recvchannel.wait()
dma.recvchannel.stop()
CompareLists(sendBuffer, recvBuffer)
def __init__(self, description):
super().__init__(description)
# Init config register
self.pkt_size = 16
# Init buffer
self.buf_dtype = np.uint32
self.buf = allocate(shape=(16, ), dtype=np.uint32)
def batch_size(self, value):
self._batch_size = value
# free the old buffers by setting to None
# (reference counting should care of it)
if self.ibuf_packed_device is not None:
self.ibuf_packed_device = None
if self.obuf_packed_device is not None:
self.obuf_packed_device = None
if self.platform == "alveo":
self.ibuf_packed_device = allocate(shape=self.ishape_packed, dtype=np.uint8)
self.obuf_packed_device = allocate(shape=self.oshape_packed, dtype=np.uint8)
else:
self.ibuf_packed_device = allocate(
shape=self.ishape_packed, dtype=np.uint8, cacheable=True
)
self.obuf_packed_device = allocate(
shape=self.oshape_packed, dtype=np.uint8, cacheable=True
)
self.obuf_packed = np.empty_like(self.obuf_packed_device)
def test_default(device):
pynq.Device.active_device = device
with device.check_memops(allocates=[(10, False, BUFFER_ADDRESS)]):
buf = pynq.allocate((1024, 1024), 'u4')
assert buf.device_address == BUFFER_ADDRESS
assert buf.physical_address == BUFFER_ADDRESS
assert buf.bo == 10
assert buf.coherent is False
assert buf.cacheable is True
pynq.Device.active_device = None
def setup_accelerator(xclbin, wfile):
ol = pynq.Overlay(xclbin)
fcweights = np.genfromtxt(wfile, delimiter=',', dtype=np.int8)
fcbuf = pynq.allocate((1000, 2048), dtype=np.int8, target=ol.PLRAM0)
#csv reader erroneously adds one extra element to the end, so remove, then reshape
fcweights = fcweights[:-1].reshape(1000, 2048)
fcbuf[:] = fcweights
fcbuf.sync_to_device()
return ol, fcbuf
def __init__(self, description, pkt_size, buf_dtype=np.int16, buf_words_per_pkt=2):
super().__init__(description)
# Init config register
self.reset = 1
self.enable = 1
self.pkt_size = pkt_size-1
self.auto_restart = 0
self.reset = 0
# Init buffer
self.buf = allocate(shape=(pkt_size * buf_words_per_pkt, ), dtype=np.int16)
def transfer(self, packetsize):
"""Returns a numpy array with inspected data of length packetsize.
"""
transfersize = int(np.ceil(packetsize / 8))
if transfersize > 4096 or transfersize < 2:
raise ValueError(
'Packet size incorrect, should be in range 16 to 32768.')
self._pgen.packetsize = transfersize
buffer_re = allocate(shape=(transfersize * 8, ), dtype=np.int16)
buffer_im = allocate(shape=(transfersize * 8, ), dtype=np.int16)
self._dma_real.recvchannel.transfer(buffer_re)
self._dma_imag.recvchannel.transfer(buffer_im)
self._pgen.transfer = 1
self._dma_real.recvchannel.wait()
self._dma_imag.recvchannel.wait()
self._pgen.transfer = 0
re_data = np.array(buffer_re) * 2**-15
im_data = np.array(buffer_im) * 2**-15
buffer_re.freebuffer()
buffer_im.freebuffer()
c_data = re_data.astype('double') + 1j * im_data.astype('double')
return c_data[0:packetsize]
def set_shape(self, shape):
"""Set the buffer shape by first freeing the existing buffer
and then allocating a new buffer with the given tuple. Obtain the
tuple product to set the packetsize of the data_inspector_module.
"""
self.buffer.freebuffer()
lshape = list(shape)
lshape[0] = lshape[0] * 2
tshape = tuple(lshape)
self.buffer = allocate(shape=tshape, dtype=np.int16)
product = 1
for i in shape:
product *= i
self.data_inspector_module.packetsize = product
