def common_uspmv(rows,cols,datas,m_sizes,k_sizes,nnz_sizes, num_runs,vector_range):
ddrWidth = int(xclbin_opts["GEMX_ddrWidth"])
min_k = ddrWidth
min_m = ddrWidth * int(xclbin_opts["GEMX_uspmvInterleaves"])
for i in range(len(m_sizes)):
m_sizes[i] = test.get_padded_size (m_sizes[i], min_m)
k_sizes[i] = test.get_padded_size (k_sizes[i], min_m)
print ("size:",m_sizes,k_sizes,"nnz:",nnz_sizes)
B = np.zeros((num_runs, k_sizes[i]), dtype=np.float32)
test.fillMod(9, num_runs, k_sizes[i],B)
B = B.astype(np.float32)
C_list=[B]
for i in range(len(m_sizes)):
C = np.zeros ((num_runs, m_sizes[i]), dtype=np.float32)
C_list.append(C)
A = gemx.sendUSpMat(np.array(rows[i]).astype(np.uint16),
np.array(cols[i]).astype(np.uint16),
np.array(datas[i]),
np.array(m_sizes[i],dtype=np.int32),
np.array(k_sizes[i],dtype=np.int32),
np.array(nnz_sizes[i],dtype=np.int32),
np.array(1,dtype=np.float32),
xclbin_opts)
gemx.sendMat(C_list[i])
gemx.sendMat(C_list[i+1])
gemx.addUSPMVOp(A,C_list[i],C_list[i+1],num_runs)
gemx.execute()
gemx.clearInstrBuf()
gemx.getMat(C_list[-1])
test.multiply_and_cmp_uspmv(rows,cols,datas,m_sizes,k_sizes,B,C_list[-1])
def predict ( self, inp): self.out_dim = (inp.shape[0],self.out_dim[1]) inp = self.format_for_fpga(inp, 1, 1) B = inp.astype(np.float32) gemx.sendMat(B) C = np.zeros ((inp.shape[0], self.sizes[0][0]), dtype=np.float32) gemx.sendMat(C) gemx.addUSPMVOp(self.A_list[0],B,C,inp.shape[0]) gemx.execute() gemx.getMat(C) gemx.clearInstrBuf() result = C return result[:self.out_dim[0],:self.out_dim[1]]
def predict(self, inp):
"""
prepare input matrix for the engine, send all the matrices and instructions to kernel and get the result prediction matrix
Parameters
----------
inp: array
input matrix
Return
------
array
result prediction matrix
"""
stage_size = int(self.xclbin_opts["GEMX_uspmvStages"])
layer_size = len(self._qw)
if stage_size == 1:
inp = self.format_for_fpga(inp, 1, self.min_m)
C_list = [inp.astype(np.float32)]
gemx.sendMat(C_list[0])
for i in range(layer_size):
C_list.append(
np.zeros((inp.shape[0], self.sizes[i][0]),
dtype=np.float32))
gemx.sendMat(C_list[i + 1])
gemx.addUSPMVOp(self.A_list[i], C_list[i], C_list[i + 1],
inp.shape[0])
else:
inp = self.format_for_fpga(inp, 1, self.min_m)
C_list = [inp.astype(np.float32)]
gemx.sendMat(C_list[0])
C_end = np.zeros((inp.shape[0], self.sizes[-1][0]),
dtype=np.float32)
gemx.sendMat(C_end)
gemx.addUSPMVOp(self.A_list[0], C_list[0], C_list[-1],
inp.shape[0])
gemx.execute()
gemx.getMat(C_list[-1])
gemx.clearInstrBuf()
result = C_list[-1]
return result[:self.out_dim[0], :self.out_dim[1]]
gemx.sendMat(C_buf[i + 1])
else:
gemx.sendMat(C_buf[i + 1])
gemx.sendMat(bias_buf[i])
else:
gemx.sendMat(C_buf[i + 1])
time.sleep(2)
total_time = 0
for k in range(args.numiter): #interations
start_time = time.time()
gemx.sendMat(C_buf[0])
if stage_size == 1:
for i in range(num_matrix):
if args.engine == 'spmv':
gemx.addUSPMVOp(A_buf[i], C_buf[i], C_buf[i + 1],
number_runs)
else:
gemx.addFCNOp(A_buf[i], C_buf[i], C_buf[i + 1],
bias_buf[i], 1, 0, 0, 0)
else:
gemx.addUSPMVOp(A_buf[0], C_buf[0], C_buf[-1], number_runs)
gemx.execute()
gemx.getMat(C_buf[-1])
total_time += time.time() - start_time
gemx.clearInstrBuf()
exectime = total_time / args.numiter
print("Average FPGA exec time(python): ", exectime * 1000, " ms")
print("inf/s:", number_runs / exectime)
gemx.printStats()