def __init__(self):
self.ffi = cffi.FFI()
if IS_PYNQ:
bitfile = "../pynq_chainer/HLS/bitstream.bit"
libfile = "../pynq_chainer/HLS/src/libaccel.so"
#self.ffi.cdef("int _Z18_p0_mmult_accel1_0PfS_S_iii(float*, float*, float*, int, int, int);")
self.ffi.cdef(
"int _Z17_p0_mmult_accel_0PfS_S_iii(float*, float*, float*, int, int, int);"
)
self.lib = self.ffi.dlopen(libfile)
#self.accel_fn = self.lib._Z18_p0_mmult_accel1_0PfS_S_iii
self.accel_fn = self.lib._Z17_p0_mmult_accel_0PfS_S_iii
overlay = Overlay(bitfile)
if not overlay.is_loaded():
overlay.download()
print("load Overlay")
else:
self.accel_fn = pcsim.mmult_accel
def __init__(self):
self.ffi = cffi.FFI()
if IS_PYNQ:
bitfile = "../pynq_chainer/HLS/bitstream.bit"
libfile = "../pynq_chainer/HLS/src/libaccel.so"
#self.ffi.cdef("void _Z17_p0_mmult_accel_0PiS_S_ii(int*, int*, int*, int, int);")
#self.ffi.cdef("void _Z20_p0_binary_connect_0tPiS_S_tt(unsigned short, int*, int*, int*, unsigned short, unsigned short);")
self.ffi.cdef(
"void _Z17_p0_BlackBoxJam_0P7ap_uintILi32EES1_bjjjS0_(unsigned int*, unsigned int*, bool, unsigned int, unsigned int, unsigned int, unsigned int);"
)
self.lib = self.ffi.dlopen(libfile)
print(self.lib.__dict__)
#self.accel_fn = self.lib._Z17_p0_mmult_accel_0PiS_S_ii
#self.accel_fn = self.lib._Z20_p0_binary_connect_0tPiS_S_tt
self.accel_fn = self.lib._Z17_p0_BlackBoxJam_0P7ap_uintILi32EES1_bjjjS0_
overlay = Overlay(bitfile)
if not overlay.is_loaded():
overlay.download()
print("load Overlay")
else:
self.accel_fn = pcsim.mmult_accel
def __init__(self):
#Set attributes
self.bitfile = BITFILE
self.libfile = LIBFILE
self._ffi = cffi.FFI()
#Accelerator functions
self._ffi.cdef("void _p0_rgb_2_gray_0(uint8_t * input," +
"uint8_t * output);")
self._ffi.cdef("void _p0_sobel_filter_0(uint8_t * input," +
"uint8_t * output);")
self._ffilib = self._ffi.dlopen(LIBFILE)
self.overlay = Overlay(self.bitfile)
#XLNK functions
self._ffi.cdef("void *cma_alloc(uint32_t len," +
"uint32_t cacheable);")
self._ffi.cdef("void cma_free(void *buf);")
#Allocate memory for gray frame
self.frame_gray = self._ffi.cast(
"uint8_t *", self._ffilib.cma_alloc(1920 * 1080, 0))
#Check if bitstream is loaded
if not Overlay.is_loaded(self.overlay):
self.overlay.download()
MB = args.d # MB of the arry
NUM_TEST = args.n # Number of benchmark tests
VERBOSE = args.v # Verbose logging
logging.basicConfig(format='%(levelname)s: %(message)s', level = logging.DEBUG if VERBOSE else logging.INFO)
logging.debug("Configuration: MB={}, NUM_TEST={}".format(MB, NUM_TEST))
# Load the overlay
logging.info("Loading Overlay...");
start = time.time()
ol = Overlay("dma_benchmark.bit")
ol.download()
end = time.time()
if not ol.is_loaded():
logging.error("The overlay can not be loaded!")
exit(-1)
logging.info("Overlay loaded in {0:.4f}s".format(end-start));
logging.debug("Overlay info: [{}, {}]".format(PL.bitfile_name, ol.bitstream.timestamp))
# DMA buffer for Transferring from PS-PL
TXdma = DMA(0x40400000, direction=0)
# DMA buffer for Transferring from PL-PS
RXdma = DMA(0x40400000, direction=1)
logging.debug("DMA configuration:")
logging.debug(TXdma.Configuration)
logging.debug(RXdma.Configuration)
from pynq import Overlay
overlay = Overlay('/home/xilinx/pynq/tensorcore.bit')
######################################################################################################
overlay.is_loaded()
help(overlay)
######################################################################################################
tensor_stat = overlay.tensorcore_0
help(tensor_stat)
######################################################################################################
# Configure Register (Address 0x00) Format
# ______________________________________________________________________________________________
# | 31 -- 28 | 27 -- 26 | 25 -- | 24 -- 21 |20 -- 17 | 16 | 15 --12 | 11 -- 1 | 0 |
# | Feat Sel, X-4b,Y-2b | Bias/Zero | Psum Index-4b |Out Shift| Relu | Ouput Sel*| Wgt Sel** | NA|
# ----------------------------------------------------------------------------------------------
# for version 0.1, Ouput Sel should set as same as Psum Index
# for version 0.1, Weight Sel only use two bit and 4 index are utilized
# for version 0.1, Feat Sel X should be <12
tensor_stat.write(0x00, 0b0000_00_0_0001_0000_0_0001_00000000000_1)
######################################################################################################
# Write the Feature map, a 6 row 12 column table with each element of 8 bit
# Address Space Mapping Format (a 32-bit word includes 4 elements)
# ____________________________________________________
